diff --git a/metropolis/README.md b/metropolis/README.md
new file mode 100644
index 0000000..28fbb45
--- /dev/null
+++ b/metropolis/README.md
@@ -0,0 +1,11 @@
+# Metropolis Operating System
+
+## Launch VM
+
+Launch the VM:
+
+```
+scripts/bin/bazel run //core/scripts:launch
+```
+
+Ctrl-C will kill the VM.
diff --git a/metropolis/cli/dbg/BUILD.bazel b/metropolis/cli/dbg/BUILD.bazel
new file mode 100644
index 0000000..51fa722
--- /dev/null
+++ b/metropolis/cli/dbg/BUILD.bazel
@@ -0,0 +1,24 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/cli/dbg",
+    visibility = ["//visibility:private"],
+    deps = [
+        "//metropolis/node/core/logtree:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+        "@com_github_spf13_pflag//:go_default_library",
+        "@io_k8s_component_base//cli/flag:go_default_library",
+        "@io_k8s_kubectl//pkg/cmd/plugin:go_default_library",
+        "@io_k8s_kubectl//pkg/util/logs:go_default_library",
+        "@io_k8s_kubernetes//pkg/kubectl/cmd:go_default_library",
+        "@org_golang_google_grpc//:go_default_library",
+    ],
+)
+
+go_binary(
+    name = "dbg",
+    embed = [":go_default_library"],
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/cli/dbg/main.go b/metropolis/cli/dbg/main.go
new file mode 100644
index 0000000..1bb6185
--- /dev/null
+++ b/metropolis/cli/dbg/main.go
@@ -0,0 +1,171 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"math/rand"
+	"os"
+	"time"
+
+	"github.com/spf13/pflag"
+	"google.golang.org/grpc"
+	cliflag "k8s.io/component-base/cli/flag"
+	"k8s.io/kubectl/pkg/cmd/plugin"
+	"k8s.io/kubectl/pkg/util/logs"
+	"k8s.io/kubernetes/pkg/kubectl/cmd"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+func main() {
+	ctx := context.Background()
+	// Hardcode localhost since this should never be used to interface with a production node because of missing
+	// encryption & authentication
+	grpcClient, err := grpc.Dial("localhost:7837", grpc.WithInsecure())
+	if err != nil {
+		fmt.Printf("Failed to dial debug service (is it running): %v\n", err)
+	}
+	debugClient := apb.NewNodeDebugServiceClient(grpcClient)
+	if len(os.Args) < 2 {
+		fmt.Println("Please specify a subcommand")
+		os.Exit(1)
+	}
+
+	logsCmd := flag.NewFlagSet("logs", flag.ExitOnError)
+	logsTailN := logsCmd.Int("tail", -1, "Get last n lines (-1 = whole buffer, 0 = disable)")
+	logsStream := logsCmd.Bool("follow", false, "Stream log entries live from the system")
+	logsRecursive := logsCmd.Bool("recursive", false, "Get entries from entire DN subtree")
+	logsCmd.Usage = func() {
+		fmt.Fprintf(os.Stderr, "Usage: %s %s [options] dn\n", os.Args[0], os.Args[1])
+		flag.PrintDefaults()
+
+		fmt.Fprintf(os.Stderr, "Example:\n  %s %s --tail 5 --follow init\n", os.Args[0], os.Args[1])
+	}
+	goldenticketCmd := flag.NewFlagSet("goldenticket", flag.ExitOnError)
+	conditionCmd := flag.NewFlagSet("condition", flag.ExitOnError)
+	conditionCmd.Usage = func() {
+		fmt.Fprintf(os.Stderr, "Usage: %s %s [options] component_path\n", os.Args[0], os.Args[1])
+		flag.PrintDefaults()
+
+		fmt.Fprintf(os.Stderr, "Example:\n  %s %s IPAssigned\n", os.Args[0], os.Args[1])
+	}
+
+	switch os.Args[1] {
+	case "logs":
+		logsCmd.Parse(os.Args[2:])
+		dn := logsCmd.Arg(0)
+		req := &apb.GetLogsRequest{
+			Dn:          dn,
+			BacklogMode: apb.GetLogsRequest_BACKLOG_DISABLE,
+			StreamMode:  apb.GetLogsRequest_STREAM_DISABLE,
+			Filters:     nil,
+		}
+
+		switch *logsTailN {
+		case 0:
+		case -1:
+			req.BacklogMode = apb.GetLogsRequest_BACKLOG_ALL
+		default:
+			req.BacklogMode = apb.GetLogsRequest_BACKLOG_COUNT
+			req.BacklogCount = int64(*logsTailN)
+		}
+
+		if *logsStream {
+			req.StreamMode = apb.GetLogsRequest_STREAM_UNBUFFERED
+		}
+
+		if *logsRecursive {
+			req.Filters = append(req.Filters, &apb.LogFilter{
+				Filter: &apb.LogFilter_WithChildren_{WithChildren: &apb.LogFilter_WithChildren{}},
+			})
+		}
+
+		stream, err := debugClient.GetLogs(ctx, req)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "Failed to get logs: %v\n", err)
+			os.Exit(1)
+		}
+		for {
+			res, err := stream.Recv()
+			if err != nil {
+				if err == io.EOF {
+					os.Exit(0)
+				}
+				fmt.Fprintf(os.Stderr, "Failed to stream logs: %v\n", err)
+				os.Exit(1)
+			}
+			for _, entry := range res.BacklogEntries {
+				entry, err := logtree.LogEntryFromProto(entry)
+				if err != nil {
+					fmt.Printf("error decoding entry: %v", err)
+					continue
+				}
+				fmt.Println(entry.String())
+			}
+		}
+	case "goldenticket":
+		goldenticketCmd.Parse(os.Args[2:])
+		ip := goldenticketCmd.Arg(0)
+		res, err := debugClient.GetGoldenTicket(ctx, &apb.GetGoldenTicketRequest{ExternalIp: ip})
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "Failed to get golden ticket: %v\n", err)
+			os.Exit(1)
+		}
+		fmt.Println(res.Ticket)
+	case "kubectl":
+		// Always get a kubeconfig with cluster-admin (group system:masters), kubectl itself can impersonate
+		kubeconfigFile, err := ioutil.TempFile("", "dbg_kubeconfig")
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "Failed to create kubeconfig temp file: %v\n", err)
+			os.Exit(1)
+		}
+		defer kubeconfigFile.Close()
+		defer os.Remove(kubeconfigFile.Name())
+
+		res, err := debugClient.GetDebugKubeconfig(ctx, &apb.GetDebugKubeconfigRequest{Id: "debug-user", Groups: []string{"system:masters"}})
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "Failed to get kubeconfig: %v\n", err)
+			os.Exit(1)
+		}
+		if _, err := kubeconfigFile.WriteString(res.DebugKubeconfig); err != nil {
+			fmt.Fprintf(os.Stderr, "Failed to write kubeconfig: %v\n", err)
+			os.Exit(1)
+		}
+
+		// This magic sets up everything as if this was just the kubectl binary. It sets the KUBECONFIG environment
+		// variable so that it knows where the Kubeconfig is located and forcibly overwrites the arguments so that
+		// the "wrapper" arguments are not visible to its flags parser. The base code is straight from
+		// https://github.com/kubernetes/kubernetes/blob/master/cmd/kubectl/kubectl.go
+		os.Setenv("KUBECONFIG", kubeconfigFile.Name())
+		rand.Seed(time.Now().UnixNano())
+		pflag.CommandLine.SetNormalizeFunc(cliflag.WordSepNormalizeFunc)
+		pflag.CommandLine.AddGoFlagSet(flag.CommandLine)
+		logs.InitLogs()
+		defer logs.FlushLogs()
+		command := cmd.NewDefaultKubectlCommandWithArgs(cmd.NewDefaultPluginHandler(plugin.ValidPluginFilenamePrefixes), os.Args[2:], os.Stdin, os.Stdout, os.Stderr)
+		command.SetArgs(os.Args[2:])
+		if err := command.Execute(); err != nil {
+			os.Exit(1)
+		}
+	}
+}
diff --git a/metropolis/node/BUILD.bazel b/metropolis/node/BUILD.bazel
new file mode 100644
index 0000000..48c9177
--- /dev/null
+++ b/metropolis/node/BUILD.bazel
@@ -0,0 +1,141 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//metropolis/node/build:def.bzl", "smalltown_initramfs")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["ports.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node",
+    visibility = ["//visibility:public"],
+)
+
+# debug_build checks if we're building in debug mode and enables various debug features for the image. Currently this
+# is only used for attaching a Delve debugger to init when it's enabled.
+config_setting(
+    name = "debug_build",
+    values = {
+        "compilation_mode": "dbg",
+    },
+)
+
+smalltown_initramfs(
+    name = "initramfs",
+    extra_dirs = [
+        "/kubernetes/conf/flexvolume-plugins",
+        "/containerd/run",
+    ],
+    files = {
+        "//metropolis/node/core": "/init",
+        "//third_party/xfsprogs:mkfs.xfs": "/bin/mkfs.xfs",
+
+        # CA Certificate bundle & os-release
+        "@cacerts//file": "/etc/ssl/cert.pem",
+        ":os-release-info": "/etc/os-release",
+
+        # Hyperkube
+        "//metropolis/node/kubernetes/hyperkube": "/kubernetes/bin/kube",
+
+        # CoreDNS
+        "@com_github_coredns_coredns//:coredns": "/kubernetes/bin/coredns",
+
+        # runsc/gVisor
+        "@com_github_google_gvisor//runsc": "/containerd/bin/runsc",
+        "@com_github_google_gvisor_containerd_shim//cmd/containerd-shim-runsc-v1": "/containerd/bin/containerd-shim-runsc-v1",
+
+        # runc (runtime in files_cc because of cgo)
+        "@com_github_containerd_containerd//cmd/containerd-shim-runc-v2": "/containerd/bin/containerd-shim-runc-v2",
+
+        # Containerd
+        "@com_github_containerd_containerd//cmd/containerd": "/containerd/bin/containerd",
+
+        # Containerd config files
+        "//metropolis/node/kubernetes/containerd:runsc.toml": "/containerd/conf/runsc.toml",
+        "//metropolis/node/kubernetes/containerd:config.toml": "/containerd/conf/config.toml",
+        "//metropolis/node/kubernetes/containerd:cnispec.gojson": "/containerd/conf/cnispec.gojson",
+
+        # Containerd preseed bundles
+        "//metropolis/test/e2e/preseedtest:preseedtest.tar": "/containerd/preseed/k8s.io/preseedtest.tar",
+        "//metropolis/test/e2e/k8s_cts:k8s_cts_image.tar": "/containerd/preseed/k8s.io/k8s_cts.tar",
+
+        # CNI Plugins
+        "@com_github_containernetworking_plugins//plugins/main/loopback": "/containerd/bin/cni/loopback",
+        "@com_github_containernetworking_plugins//plugins/main/ptp": "/containerd/bin/cni/ptp",
+        "@com_github_containernetworking_plugins//plugins/ipam/host-local": "/containerd/bin/cni/host-local",
+
+        # Delve
+        "@com_github_go_delve_delve//cmd/dlv:dlv": "/dlv",
+    },
+    files_cc = {
+        # runc runtime, with cgo
+        "@com_github_opencontainers_runc//:runc": "/containerd/bin/runc",
+    },
+)
+
+genrule(
+    name = "image",
+    srcs = [
+        "//third_party/linux:bzImage",
+        ":initramfs",
+    ],
+    outs = [
+        "smalltown.img",
+    ],
+    cmd = """
+    $(location //metropolis/node/build/mkimage) \
+        -efi $(location //third_party/linux:bzImage) \
+        -initramfs $(location :initramfs) \
+        -out $@
+    """,
+    tools = [
+        "//metropolis/node/build/mkimage",
+    ],
+    visibility = ["//visibility:public"],
+)
+
+genrule(
+    name = "swtpm_data",
+    outs = [
+        "tpm/tpm2-00.permall",
+        "tpm/signkey.pem",
+        "tpm/issuercert.pem",
+    ],
+    cmd = """
+    mkdir -p tpm/ca
+
+    cat <<EOF > tpm/swtpm.conf
+create_certs_tool= /usr/share/swtpm/swtpm-localca
+create_certs_tool_config = tpm/swtpm-localca.conf
+create_certs_tool_options = /etc/swtpm-localca.options
+EOF
+
+    cat <<EOF > tpm/swtpm-localca.conf
+statedir = tpm/ca
+signingkey = tpm/ca/signkey.pem
+issuercert = tpm/ca/issuercert.pem
+certserial = tpm/ca/certserial
+EOF
+
+    swtpm_setup \
+        --tpmstate tpm \
+        --create-ek-cert \
+        --create-platform-cert \
+        --allow-signing \
+        --tpm2 \
+        --display \
+        --pcr-banks sha1,sha256,sha384,sha512 \
+        --config tpm/swtpm.conf
+
+    cp tpm/tpm2-00.permall $(location tpm/tpm2-00.permall)
+    cp tpm/ca/issuercert.pem $(location tpm/issuercert.pem)
+    cp tpm/ca/signkey.pem $(location tpm/signkey.pem)
+    """,
+    visibility = ["//visibility:public"],
+)
+
+load("//metropolis/node/build/genosrelease:defs.bzl", "os_release")
+
+os_release(
+    name = "os-release-info",
+    os_id = "smalltown",
+    os_name = "Smalltown",
+    stamp_var = "STABLE_SIGNOS_version",
+)
diff --git a/metropolis/node/build/BUILD b/metropolis/node/build/BUILD
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/metropolis/node/build/BUILD
diff --git a/metropolis/node/build/def.bzl b/metropolis/node/build/def.bzl
new file mode 100644
index 0000000..e2885e5
--- /dev/null
+++ b/metropolis/node/build/def.bzl
@@ -0,0 +1,257 @@
+#  Copyright 2020 The Monogon Project Authors.
+#
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+
+def _build_pure_transition_impl(settings, attr):
+    """
+    Transition that enables pure, static build of Go binaries.
+    """
+    return {
+        "@io_bazel_rules_go//go/config:pure": True,
+        "@io_bazel_rules_go//go/config:static": True,
+    }
+
+build_pure_transition = transition(
+    implementation = _build_pure_transition_impl,
+    inputs = [],
+    outputs = [
+        "@io_bazel_rules_go//go/config:pure",
+        "@io_bazel_rules_go//go/config:static",
+    ],
+)
+
+def _build_static_transition_impl(settings, attr):
+    """
+    Transition that enables static builds with CGo and musl for Go binaries.
+    """
+    return {
+        "@io_bazel_rules_go//go/config:static": True,
+        "//command_line_option:crosstool_top": "//build/toolchain/musl-host-gcc:musl_host_cc_suite",
+    }
+
+build_static_transition = transition(
+    implementation = _build_static_transition_impl,
+    inputs = [],
+    outputs = [
+        "@io_bazel_rules_go//go/config:static",
+        "//command_line_option:crosstool_top",
+    ],
+)
+
+def _smalltown_initramfs_impl(ctx):
+    """
+    Generate an lz4-compressed initramfs based on a label/file list.
+    """
+
+    # Generate config file for gen_init_cpio that describes the initramfs to build.
+    cpio_list_name = ctx.label.name + ".cpio_list"
+    cpio_list = ctx.actions.declare_file(cpio_list_name)
+
+    # Start out with some standard initramfs device files.
+    cpio_list_content = [
+        "dir /dev 0755 0 0",
+        "nod /dev/console 0600 0 0 c 5 1",
+        "nod /dev/null 0644 0 0 c 1 3",
+        "nod /dev/kmsg 0644 0 0 c 1 11",
+        "nod /dev/ptmx 0644 0 0 c 5 2",
+    ]
+
+    # Find all directories that need to be created.
+    directories_needed = []
+    for _, p in ctx.attr.files.items():
+        if not p.startswith("/"):
+            fail("file {} invalid: must begin with /".format(p))
+
+        # Get all intermediate directories on path to file
+        parts = p.split("/")[1:-1]
+        directories_needed.append(parts)
+
+    for _, p in ctx.attr.files_cc.items():
+        if not p.startswith("/"):
+            fail("file {} invalid: must begin with /".format(p))
+
+        # Get all intermediate directories on path to file
+        parts = p.split("/")[1:-1]
+        directories_needed.append(parts)
+
+    # Extend with extra directories defined by user.
+    for p in ctx.attr.extra_dirs:
+        if not p.startswith("/"):
+            fail("directory {} invalid: must begin with /".format(p))
+
+        parts = p.split("/")[1:]
+        directories_needed.append(parts)
+
+    directories = []
+    for parts in directories_needed:
+        # Turn directory parts [usr, local, bin] into successive subpaths [/usr, /usr/local, /usr/local/bin].
+        last = ""
+        for part in parts:
+            last += "/" + part
+
+            # TODO(q3k): this is slow - this should be a set instead, but starlark doesn't implement them.
+            # For the amount of files we're dealing with this doesn't matter, but all stars are pointing towards this
+            # becoming accidentally quadratic at some point in the future.
+            if last not in directories:
+                directories.append(last)
+
+    # Append instructions to create directories.
+    # Serendipitously, the directories should already be in the right order due to us not using a set to create the
+    # list. They might not be in an elegant order (ie, if files [/foo/one/one, /bar, /foo/two/two] are request, the
+    # order will be [/foo, /foo/one, /bar, /foo/two]), but that's fine.
+    for d in directories:
+        cpio_list_content.append("dir {} 0755 0 0".format(d))
+
+    # Append instructions to add files.
+    inputs = []
+    for label, p in ctx.attr.files.items():
+        # Figure out if this is an executable.
+        is_executable = True
+
+        di = label[DefaultInfo]
+        if di.files_to_run.executable == None:
+            # Generated non-executable files will have DefaultInfo.files_to_run.executable == None
+            is_executable = False
+        elif di.files_to_run.executable.is_source:
+            # Source files will have executable.is_source == True
+            is_executable = False
+
+        # Ensure only single output is declared.
+        # If you hit this error, figure out a better logic to find what file you need, maybe looking at providers other
+        # than DefaultInfo.
+        files = di.files.to_list()
+        if len(files) > 1:
+            fail("file {} has more than one output: {}", p, files)
+        src = files[0]
+        inputs.append(src)
+
+        mode = "0755" if is_executable else "0444"
+
+        cpio_list_content.append("file {} {} {} 0 0".format(p, src.path, mode))
+
+    for label, p in ctx.attr.files_cc.items():
+        # Figure out if this is an executable.
+        is_executable = True
+
+        di = label[DefaultInfo]
+        if di.files_to_run.executable == None:
+            # Generated non-executable files will have DefaultInfo.files_to_run.executable == None
+            is_executable = False
+        elif di.files_to_run.executable.is_source:
+            # Source files will have executable.is_source == True
+            is_executable = False
+
+        # Ensure only single output is declared.
+        # If you hit this error, figure out a better logic to find what file you need, maybe looking at providers other
+        # than DefaultInfo.
+        files = di.files.to_list()
+        if len(files) > 1:
+            fail("file {} has more than one output: {}", p, files)
+        src = files[0]
+        inputs.append(src)
+
+        mode = "0755" if is_executable else "0444"
+
+        cpio_list_content.append("file {} {} {} 0 0".format(p, src.path, mode))
+
+    # Write cpio_list.
+    ctx.actions.write(cpio_list, "\n".join(cpio_list_content))
+
+    gen_init_cpio = ctx.executable._gen_init_cpio
+    savestdout = ctx.executable._savestdout
+    lz4 = ctx.executable._lz4
+
+    # Generate 'raw' (uncompressed) initramfs
+    initramfs_raw_name = ctx.label.name
+    initramfs_raw = ctx.actions.declare_file(initramfs_raw_name)
+    ctx.actions.run(
+        outputs = [initramfs_raw],
+        inputs = [cpio_list] + inputs,
+        tools = [savestdout, gen_init_cpio],
+        executable = savestdout,
+        arguments = [initramfs_raw.path, gen_init_cpio.path, cpio_list.path],
+    )
+
+    # Compress raw initramfs using lz4c.
+    initramfs_name = ctx.label.name + ".lz4"
+    initramfs = ctx.actions.declare_file(initramfs_name)
+    ctx.actions.run(
+        outputs = [initramfs],
+        inputs = [initramfs_raw],
+        tools = [savestdout, lz4],
+        executable = lz4.path,
+        arguments = ["-l", initramfs_raw.path, initramfs.path],
+    )
+
+    return [DefaultInfo(files = depset([initramfs]))]
+
+smalltown_initramfs = rule(
+    implementation = _smalltown_initramfs_impl,
+    doc = """
+        Build a Smalltown initramfs. The initramfs will contain a basic /dev directory and all the files specified by the
+        `files` attribute. Executable files will have their permissions set to 0755, non-executable files will have
+        their permissions set to 0444. All parent directories will be created with 0755 permissions.
+    """,
+    attrs = {
+        "files": attr.label_keyed_string_dict(
+            mandatory = True,
+            allow_files = True,
+            doc = """
+                Dictionary of Labels to String, placing a given Label's output file in the initramfs at the location
+                specified by the String value. The specified labels must only have a single output.
+            """,
+            # Attach pure transition to ensure all binaries added to the initramfs are pure/static binaries.
+            cfg = build_pure_transition,
+        ),
+        "files_cc": attr.label_keyed_string_dict(
+            allow_files = True,
+            doc = """
+                 Special case of 'files' for compilation targets that need to be built with the musl toolchain like
+                 go_binary targets which need cgo or cc_binary targets.
+            """,
+            # Attach static transition to all files_cc inputs to ensure they are built with musl and static.
+            cfg = build_static_transition,
+        ),
+        "extra_dirs": attr.string_list(
+            default = [],
+            doc = """
+                Extra directories to create. These will be created in addition to all the directories required to
+                contain the files specified in the `files` attribute.
+            """,
+        ),
+
+        # Tools, implicit dependencies.
+        "_gen_init_cpio": attr.label(
+            default = Label("@linux//:gen_init_cpio"),
+            executable = True,
+            cfg = "host",
+        ),
+        "_lz4": attr.label(
+            default = Label("@com_github_lz4_lz4//programs:lz4"),
+            executable = True,
+            cfg = "host",
+        ),
+        "_savestdout": attr.label(
+            default = Label("//build/savestdout"),
+            executable = True,
+            cfg = "host",
+        ),
+
+        # Allow for transitions to be attached to this rule.
+        "_whitelist_function_transition": attr.label(
+            default = "@bazel_tools//tools/whitelists/function_transition_whitelist",
+        ),
+    },
+)
diff --git a/metropolis/node/build/genosrelease/BUILD.bazel b/metropolis/node/build/genosrelease/BUILD.bazel
new file mode 100644
index 0000000..9403d72
--- /dev/null
+++ b/metropolis/node/build/genosrelease/BUILD.bazel
@@ -0,0 +1,15 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/build/genosrelease",
+    visibility = ["//visibility:private"],
+    deps = ["@com_github_joho_godotenv//:go_default_library"],
+)
+
+go_binary(
+    name = "genosrelease",
+    embed = [":go_default_library"],
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/node/build/genosrelease/defs.bzl b/metropolis/node/build/genosrelease/defs.bzl
new file mode 100644
index 0000000..61ce9e4
--- /dev/null
+++ b/metropolis/node/build/genosrelease/defs.bzl
@@ -0,0 +1,54 @@
+#  Copyright 2020 The Monogon Project Authors.
+#
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+
+def _os_release_impl(ctx):
+    ctx.actions.run(
+        mnemonic = "GenOSRelease",
+        progress_message = "Generating os-release",
+        inputs = [ctx.info_file],
+        outputs = [ctx.outputs.out],
+        executable = ctx.executable._genosrelease,
+        arguments = [
+            "-status_file",
+            ctx.info_file.path,
+            "-out_file",
+            ctx.outputs.out.path,
+            "-stamp_var",
+            ctx.attr.stamp_var,
+            "-name",
+            ctx.attr.os_name,
+            "-id",
+            ctx.attr.os_id,
+        ],
+    )
+
+os_release = rule(
+    implementation = _os_release_impl,
+    attrs = {
+        "os_name": attr.string(mandatory = True),
+        "os_id": attr.string(mandatory = True),
+        "stamp_var": attr.string(mandatory = True),
+        "_genosrelease": attr.label(
+            default = Label("//metropolis/node/build/genosrelease"),
+            cfg = "host",
+            executable = True,
+            allow_files = True,
+        ),
+    },
+    outputs = {
+        "out": "os-release",
+    },
+)
diff --git a/metropolis/node/build/genosrelease/main.go b/metropolis/node/build/genosrelease/main.go
new file mode 100644
index 0000000..2344f19
--- /dev/null
+++ b/metropolis/node/build/genosrelease/main.go
@@ -0,0 +1,78 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// genosrelease provides rudimentary support to generate os-release files following the freedesktop spec
+// (https://www.freedesktop.org/software/systemd/man/os-release.html) from arguments and stamping
+package main
+
+import (
+	"flag"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"strings"
+
+	"github.com/joho/godotenv"
+)
+
+var (
+	flagStatusFile = flag.String("status_file", "", "path to bazel workspace status file")
+	flagOutFile    = flag.String("out_file", "os-release", "path to os-release output file")
+	flagStampVar   = flag.String("stamp_var", "", "variable to use as version from the workspace status file")
+	flagName       = flag.String("name", "", "name parameter (see freedesktop spec)")
+	flagID         = flag.String("id", "", "id parameter (see freedesktop spec)")
+)
+
+func main() {
+	flag.Parse()
+	statusFileContent, err := ioutil.ReadFile(*flagStatusFile)
+	if err != nil {
+		fmt.Printf("Failed to open bazel workspace status file: %v\n", err)
+		os.Exit(1)
+	}
+	statusVars := make(map[string]string)
+	for _, line := range strings.Split(string(statusFileContent), "\n") {
+		line = strings.TrimSpace(line)
+		parts := strings.Fields(line)
+		if len(parts) != 2 {
+			continue
+		}
+		statusVars[parts[0]] = parts[1]
+	}
+
+	smalltownVersion, ok := statusVars[*flagStampVar]
+	if !ok {
+		fmt.Printf("%v key not set in bazel workspace status file\n", *flagStampVar)
+		os.Exit(1)
+	}
+	// As specified by https://www.freedesktop.org/software/systemd/man/os-release.html
+	osReleaseVars := map[string]string{
+		"NAME":        *flagName,
+		"ID":          *flagID,
+		"VERSION":     smalltownVersion,
+		"VERSION_ID":  smalltownVersion,
+		"PRETTY_NAME": *flagName + " " + smalltownVersion,
+	}
+	osReleaseContent, err := godotenv.Marshal(osReleaseVars)
+	if err != nil {
+		fmt.Printf("Failed to encode os-release file: %v\n", err)
+		os.Exit(1)
+	}
+	if err := ioutil.WriteFile(*flagOutFile, []byte(osReleaseContent), 0644); err != nil {
+		fmt.Printf("Failed to write os-release file: %v\n", err)
+		os.Exit(1)
+	}
+}
diff --git a/metropolis/node/build/kconfig-patcher/BUILD.bazel b/metropolis/node/build/kconfig-patcher/BUILD.bazel
new file mode 100644
index 0000000..55b2b52
--- /dev/null
+++ b/metropolis/node/build/kconfig-patcher/BUILD.bazel
@@ -0,0 +1,20 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/build/kconfig-patcher",
+    visibility = ["//visibility:private"],
+)
+
+go_binary(
+    name = "kconfig-patcher",
+    embed = [":go_default_library"],
+    visibility = ["//visibility:public"],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["main_test.go"],
+    embed = [":go_default_library"],
+)
diff --git a/metropolis/node/build/kconfig-patcher/kconfig-patcher.bzl b/metropolis/node/build/kconfig-patcher/kconfig-patcher.bzl
new file mode 100644
index 0000000..337642e
--- /dev/null
+++ b/metropolis/node/build/kconfig-patcher/kconfig-patcher.bzl
@@ -0,0 +1,33 @@
+#  Copyright 2020 The Monogon Project Authors.
+#
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+
+"""Override configs in a Linux kernel Kconfig
+"""
+
+def kconfig_patch(name, src, out, override_configs, **kwargs):
+    native.genrule(
+        name = name,
+        srcs = [src],
+        outs = [out],
+        tools = [
+            "//metropolis/node/build/kconfig-patcher",
+        ],
+        cmd = """
+        $(location //metropolis/node/build/kconfig-patcher) \
+            -in $< -out $@ '%s'
+        """ % struct(overrides = override_configs).to_json(),
+        **kwargs
+    )
diff --git a/metropolis/node/build/kconfig-patcher/main.go b/metropolis/node/build/kconfig-patcher/main.go
new file mode 100644
index 0000000..27c33e9
--- /dev/null
+++ b/metropolis/node/build/kconfig-patcher/main.go
@@ -0,0 +1,95 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"bufio"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+)
+
+var (
+	inPath  = flag.String("in", "", "Path to input Kconfig")
+	outPath = flag.String("out", "", "Path to output Kconfig")
+)
+
+func main() {
+	flag.Parse()
+	if *inPath == "" || *outPath == "" {
+		flag.PrintDefaults()
+		os.Exit(2)
+	}
+	inFile, err := os.Open(*inPath)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Failed to open input Kconfig: %v\n", err)
+		os.Exit(1)
+	}
+	outFile, err := os.Create(*outPath)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Failed to create output Kconfig: %v\n", err)
+		os.Exit(1)
+	}
+	var config struct {
+		Overrides map[string]string `json:"overrides"`
+	}
+	if err := json.Unmarshal([]byte(flag.Arg(0)), &config); err != nil {
+		fmt.Fprintf(os.Stderr, "Failed to parse overrides: %v\n", err)
+		os.Exit(1)
+	}
+	err = patchKconfig(inFile, outFile, config.Overrides)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Failed to patch: %v\n", err)
+		os.Exit(1)
+	}
+}
+
+func patchKconfig(inFile io.Reader, outFile io.Writer, overrides map[string]string) error {
+	scanner := bufio.NewScanner(inFile)
+	for scanner.Scan() {
+		line := scanner.Text()
+		cleanLine := strings.TrimSpace(line)
+		if strings.HasPrefix(cleanLine, "#") || cleanLine == "" {
+			// Pass through comments and empty lines
+			fmt.Fprintln(outFile, line)
+		} else {
+			// Line contains a configuration option
+			parts := strings.SplitN(line, "=", 2)
+			keyName := parts[0]
+			if overrideVal, ok := overrides[strings.TrimSpace(keyName)]; ok {
+				// Override it
+				if overrideVal == "" {
+					fmt.Fprintf(outFile, "# %v is not set\n", keyName)
+				} else {
+					fmt.Fprintf(outFile, "%v=%v\n", keyName, overrideVal)
+				}
+				delete(overrides, keyName)
+			} else {
+				// Pass through unchanged
+				fmt.Fprintln(outFile, line)
+			}
+		}
+	}
+	// Process left over overrides
+	for key, val := range overrides {
+		fmt.Fprintf(outFile, "%v=%v\n", key, val)
+	}
+	return nil
+}
diff --git a/metropolis/node/build/kconfig-patcher/main_test.go b/metropolis/node/build/kconfig-patcher/main_test.go
new file mode 100644
index 0000000..11c7d84
--- /dev/null
+++ b/metropolis/node/build/kconfig-patcher/main_test.go
@@ -0,0 +1,61 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+)
+
+func Test_patchKconfig(t *testing.T) {
+	type args struct {
+		inFile    string
+		overrides map[string]string
+	}
+	tests := []struct {
+		name        string
+		args        args
+		wantOutFile string
+		wantErr     bool
+	}{
+		{
+			"passthroughExtend",
+			args{inFile: "# TEST=y\n\n", overrides: map[string]string{"TEST": "n"}},
+			"# TEST=y\n\nTEST=n\n",
+			false,
+		},
+		{
+			"patch",
+			args{inFile: "TEST=y\nTEST_NO=n\n", overrides: map[string]string{"TEST": "n"}},
+			"TEST=n\nTEST_NO=n\n",
+			false,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			outFile := &bytes.Buffer{}
+			if err := patchKconfig(strings.NewReader(tt.args.inFile), outFile, tt.args.overrides); (err != nil) != tt.wantErr {
+				t.Errorf("patchKconfig() error = %v, wantErr %v", err, tt.wantErr)
+				return
+			}
+			if gotOutFile := outFile.String(); gotOutFile != tt.wantOutFile {
+				t.Errorf("patchKconfig() = %v, want %v", gotOutFile, tt.wantOutFile)
+			}
+		})
+	}
+}
diff --git a/metropolis/node/build/mkimage/BUILD.bazel b/metropolis/node/build/mkimage/BUILD.bazel
new file mode 100644
index 0000000..b489002
--- /dev/null
+++ b/metropolis/node/build/mkimage/BUILD.bazel
@@ -0,0 +1,20 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/build/mkimage",
+    visibility = ["//visibility:private"],
+    deps = [
+        "@com_github_diskfs_go_diskfs//:go_default_library",
+        "@com_github_diskfs_go_diskfs//disk:go_default_library",
+        "@com_github_diskfs_go_diskfs//filesystem:go_default_library",
+        "@com_github_diskfs_go_diskfs//partition/gpt:go_default_library",
+    ],
+)
+
+go_binary(
+    name = "mkimage",
+    embed = [":go_default_library"],
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/node/build/mkimage/main.go b/metropolis/node/build/mkimage/main.go
new file mode 100644
index 0000000..9f49f0a
--- /dev/null
+++ b/metropolis/node/build/mkimage/main.go
@@ -0,0 +1,141 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+// mkimage is a tool to generate a Smalltown disk image containing the given EFI payload, and optionally, a given external
+// initramfs image and enrolment credentials.
+
+import (
+	"flag"
+	"fmt"
+	"io/ioutil"
+	"log"
+	"os"
+
+	diskfs "github.com/diskfs/go-diskfs"
+	"github.com/diskfs/go-diskfs/disk"
+	"github.com/diskfs/go-diskfs/filesystem"
+	"github.com/diskfs/go-diskfs/partition/gpt"
+)
+
+var SmalltownDataPartition gpt.Type = gpt.Type("9eeec464-6885-414a-b278-4305c51f7966")
+
+var (
+	flagEFI                  string
+	flagOut                  string
+	flagInitramfs            string
+	flagEnrolmentCredentials string
+	flagDataPartitionSize    uint64
+	flagESPPartitionSize     uint64
+)
+
+func mibToSectors(size uint64) uint64 {
+	return (size * 1024 * 1024) / 512
+}
+
+func main() {
+	flag.StringVar(&flagEFI, "efi", "", "UEFI payload")
+	flag.StringVar(&flagOut, "out", "", "Output disk image")
+	flag.StringVar(&flagInitramfs, "initramfs", "", "External initramfs [optional]")
+	flag.StringVar(&flagEnrolmentCredentials, "enrolment_credentials", "", "Enrolment credentials [optional]")
+	flag.Uint64Var(&flagDataPartitionSize, "data_partition_size", 2048, "Override the data partition size (default 2048 MiB)")
+	flag.Uint64Var(&flagESPPartitionSize, "esp_partition_size", 512, "Override the ESP partition size (default: 512MiB)")
+	flag.Parse()
+
+	if flagEFI == "" || flagOut == "" {
+		log.Fatalf("efi and initramfs must be set")
+	}
+
+	_ = os.Remove(flagOut)
+	diskImg, err := diskfs.Create(flagOut, 3*1024*1024*1024, diskfs.Raw)
+	if err != nil {
+		log.Fatalf("diskfs.Create(%q): %v", flagOut, err)
+	}
+
+	table := &gpt.Table{
+		// This is appropriate at least for virtio disks. Might need to be adjusted for real ones.
+		LogicalSectorSize:  512,
+		PhysicalSectorSize: 512,
+		ProtectiveMBR:      true,
+		Partitions: []*gpt.Partition{
+			{
+				Type:  gpt.EFISystemPartition,
+				Name:  "ESP",
+				Start: mibToSectors(1),
+				End:   mibToSectors(flagESPPartitionSize) - 1,
+			},
+			{
+				Type:  SmalltownDataPartition,
+				Name:  "SIGNOS-DATA",
+				Start: mibToSectors(flagESPPartitionSize),
+				End:   mibToSectors(flagESPPartitionSize+flagDataPartitionSize) - 1,
+			},
+		},
+	}
+	if err := diskImg.Partition(table); err != nil {
+		log.Fatalf("Failed to apply partition table: %v", err)
+	}
+
+	fs, err := diskImg.CreateFilesystem(disk.FilesystemSpec{Partition: 1, FSType: filesystem.TypeFat32, VolumeLabel: "ESP"})
+	if err != nil {
+		log.Fatalf("Failed to create filesystem: %v", err)
+	}
+
+	// Create EFI partition structure.
+	for _, dir := range []string{"/EFI", "/EFI/BOOT", "/EFI/smalltown"} {
+		if err := fs.Mkdir(dir); err != nil {
+			log.Fatalf("Mkdir(%q): %v", dir, err)
+		}
+	}
+
+	put(fs, flagEFI, "/EFI/BOOT/BOOTX64.EFI")
+
+	if flagInitramfs != "" {
+		put(fs, flagInitramfs, "/EFI/smalltown/initramfs.cpio.lz4")
+	}
+
+	if flagEnrolmentCredentials != "" {
+		put(fs, flagEnrolmentCredentials, "/EFI/smalltown/enrolment.pb")
+	}
+
+	if err := diskImg.File.Close(); err != nil {
+		log.Fatalf("Failed to finalize image: %v", err)
+	}
+	log.Printf("Success! You can now boot %v", flagOut)
+}
+
+// put copies a file from the host filesystem into the target image.
+func put(fs filesystem.FileSystem, src, dst string) {
+	target, err := fs.OpenFile(dst, os.O_CREATE|os.O_RDWR)
+	if err != nil {
+		log.Fatalf("fs.OpenFile(%q): %v", dst, err)
+	}
+	source, err := os.Open(src)
+	if err != nil {
+		log.Fatalf("os.Open(%q): %v", src, err)
+	}
+	defer source.Close()
+	// If this is streamed (e.g. using io.Copy) it exposes a bug in diskfs, so do it in one go.
+	data, err := ioutil.ReadAll(source)
+	if err != nil {
+		log.Fatalf("Reading %q: %v", src, err)
+	}
+	if _, err := target.Write(data); err != nil {
+		fmt.Printf("writing file %q: %v", dst, err)
+		os.Exit(1)
+	}
+}
diff --git a/metropolis/node/common/devicemapper/BUILD.bazel b/metropolis/node/common/devicemapper/BUILD.bazel
new file mode 100644
index 0000000..12ca0b3
--- /dev/null
+++ b/metropolis/node/common/devicemapper/BUILD.bazel
@@ -0,0 +1,13 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["devicemapper.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/devicemapper",
+    visibility = ["//visibility:public"],
+    deps = [
+        "@com_github_pkg_errors//:go_default_library",
+        "@com_github_yalue_native_endian//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/node/common/devicemapper/devicemapper.go b/metropolis/node/common/devicemapper/devicemapper.go
new file mode 100644
index 0000000..2687e3a
--- /dev/null
+++ b/metropolis/node/common/devicemapper/devicemapper.go
@@ -0,0 +1,298 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// package devicemapper is a thin wrapper for the devicemapper ioctl API.
+// See: https://github.com/torvalds/linux/blob/master/include/uapi/linux/dm-ioctl.h
+package devicemapper
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"os"
+	"runtime"
+	"unsafe"
+
+	"github.com/pkg/errors"
+	"github.com/yalue/native_endian"
+	"golang.org/x/sys/unix"
+)
+
+type DMIoctl struct {
+	Version     Version
+	DataSize    uint32
+	DataStart   uint32
+	TargetCount uint32
+	OpenCount   int32
+	Flags       uint32
+	EventNumber uint32
+	_padding1   uint32
+	Dev         uint64
+	Name        [128]byte
+	UUID        [129]byte
+	_padding2   [7]byte
+	Data        [16384]byte
+}
+
+type DMTargetSpec struct {
+	SectorStart uint64
+	Length      uint64
+	Status      int32
+	Next        uint32
+	TargetType  [16]byte
+}
+
+type DMTargetDeps struct {
+	Count   uint32
+	Padding uint32
+	Dev     []uint64
+}
+
+type DMNameList struct {
+	Dev  uint64
+	Next uint32
+	Name []byte
+}
+
+type DMTargetVersions struct {
+	Next    uint32
+	Version [3]uint32
+}
+
+type DMTargetMessage struct {
+	Sector  uint64
+	Message []byte
+}
+
+type Version [3]uint32
+
+const (
+	/* Top level cmds */
+	DM_VERSION_CMD uintptr = (0xc138fd << 8) + iota
+	DM_REMOVE_ALL_CMD
+	DM_LIST_DEVICES_CMD
+
+	/* device level cmds */
+	DM_DEV_CREATE_CMD
+	DM_DEV_REMOVE_CMD
+	DM_DEV_RENAME_CMD
+	DM_DEV_SUSPEND_CMD
+	DM_DEV_STATUS_CMD
+	DM_DEV_WAIT_CMD
+
+	/* Table level cmds */
+	DM_TABLE_LOAD_CMD
+	DM_TABLE_CLEAR_CMD
+	DM_TABLE_DEPS_CMD
+	DM_TABLE_STATUS_CMD
+
+	/* Added later */
+	DM_LIST_VERSIONS_CMD
+	DM_TARGET_MSG_CMD
+	DM_DEV_SET_GEOMETRY_CMD
+	DM_DEV_ARM_POLL_CMD
+)
+
+const (
+	DM_READONLY_FLAG       = 1 << 0 /* In/Out */
+	DM_SUSPEND_FLAG        = 1 << 1 /* In/Out */
+	DM_PERSISTENT_DEV_FLAG = 1 << 3 /* In */
+)
+
+const baseDataSize = uint32(unsafe.Sizeof(DMIoctl{})) - 16384
+
+func newReq() DMIoctl {
+	return DMIoctl{
+		Version:   Version{4, 0, 0},
+		DataSize:  baseDataSize,
+		DataStart: baseDataSize,
+	}
+}
+
+// stringToDelimitedBuf copies src to dst and returns an error if len(src) > len(dst),
+// or when the string contains a null byte.
+func stringToDelimitedBuf(dst []byte, src string) error {
+	if len(src) > len(dst)-1 {
+		return fmt.Errorf("string longer than target buffer (%v > %v)", len(src), len(dst)-1)
+	}
+	for i := 0; i < len(src); i++ {
+		if src[i] == 0x00 {
+			return errors.New("string contains null byte, this is unsupported by DM")
+		}
+		dst[i] = src[i]
+	}
+	return nil
+}
+
+var fd uintptr
+
+func getFd() (uintptr, error) {
+	if fd == 0 {
+		f, err := os.Open("/dev/mapper/control")
+		if os.IsNotExist(err) {
+			_ = os.MkdirAll("/dev/mapper", 0755)
+			if err := unix.Mknod("/dev/mapper/control", unix.S_IFCHR|0600, int(unix.Mkdev(10, 236))); err != nil {
+				return 0, err
+			}
+			f, err = os.Open("/dev/mapper/control")
+			if err != nil {
+				return 0, err
+			}
+		} else if err != nil {
+			return 0, err
+		}
+		fd = f.Fd()
+		return f.Fd(), nil
+	}
+	return fd, nil
+}
+
+func GetVersion() (Version, error) {
+	req := newReq()
+	fd, err := getFd()
+	if err != nil {
+		return Version{}, err
+	}
+	if _, _, err := unix.Syscall(unix.SYS_IOCTL, fd, DM_VERSION_CMD, uintptr(unsafe.Pointer(&req))); err != 0 {
+		return Version{}, err
+	}
+	return req.Version, nil
+}
+
+func CreateDevice(name string) (uint64, error) {
+	req := newReq()
+	if err := stringToDelimitedBuf(req.Name[:], name); err != nil {
+		return 0, err
+	}
+	fd, err := getFd()
+	if err != nil {
+		return 0, err
+	}
+	if _, _, err := unix.Syscall(unix.SYS_IOCTL, fd, DM_DEV_CREATE_CMD, uintptr(unsafe.Pointer(&req))); err != 0 {
+		return 0, err
+	}
+	return req.Dev, nil
+}
+
+func RemoveDevice(name string) error {
+	req := newReq()
+	if err := stringToDelimitedBuf(req.Name[:], name); err != nil {
+		return err
+	}
+	fd, err := getFd()
+	if err != nil {
+		return err
+	}
+	if _, _, err := unix.Syscall(unix.SYS_IOCTL, fd, DM_DEV_REMOVE_CMD, uintptr(unsafe.Pointer(&req))); err != 0 {
+		return err
+	}
+	runtime.KeepAlive(req)
+	return nil
+}
+
+type Target struct {
+	StartSector uint64
+	Length      uint64
+	Type        string
+	Parameters  string
+}
+
+func LoadTable(name string, targets []Target) error {
+	req := newReq()
+	if err := stringToDelimitedBuf(req.Name[:], name); err != nil {
+		return err
+	}
+	var data bytes.Buffer
+	for _, target := range targets {
+		// Gives the size of the spec and the null-terminated params aligned to 8 bytes
+		padding := len(target.Parameters) % 8
+		targetSize := uint32(int(unsafe.Sizeof(DMTargetSpec{})) + (len(target.Parameters) + 1) + padding)
+
+		targetSpec := DMTargetSpec{
+			SectorStart: target.StartSector,
+			Length:      target.Length,
+			Next:        targetSize,
+		}
+		if err := stringToDelimitedBuf(targetSpec.TargetType[:], target.Type); err != nil {
+			return err
+		}
+		if err := binary.Write(&data, native_endian.NativeEndian(), &targetSpec); err != nil {
+			panic(err)
+		}
+		data.WriteString(target.Parameters)
+		data.WriteByte(0x00)
+		for i := 0; i < padding; i++ {
+			data.WriteByte(0x00)
+		}
+	}
+	req.TargetCount = uint32(len(targets))
+	if data.Len() >= 16384 {
+		return errors.New("table too large for allocated memory")
+	}
+	req.DataSize = baseDataSize + uint32(data.Len())
+	copy(req.Data[:], data.Bytes())
+	fd, err := getFd()
+	if err != nil {
+		return err
+	}
+	if _, _, err := unix.Syscall(unix.SYS_IOCTL, fd, DM_TABLE_LOAD_CMD, uintptr(unsafe.Pointer(&req))); err != 0 {
+		return err
+	}
+	runtime.KeepAlive(req)
+	return nil
+}
+
+func suspendResume(name string, suspend bool) error {
+	req := newReq()
+	if err := stringToDelimitedBuf(req.Name[:], name); err != nil {
+		return err
+	}
+	if suspend {
+		req.Flags = DM_SUSPEND_FLAG
+	}
+	fd, err := getFd()
+	if err != nil {
+		return err
+	}
+	if _, _, err := unix.Syscall(unix.SYS_IOCTL, fd, DM_DEV_SUSPEND_CMD, uintptr(unsafe.Pointer(&req))); err != 0 {
+		return err
+	}
+	runtime.KeepAlive(req)
+	return nil
+}
+
+func Suspend(name string) error {
+	return suspendResume(name, true)
+}
+func Resume(name string) error {
+	return suspendResume(name, false)
+}
+
+func CreateActiveDevice(name string, targets []Target) (uint64, error) {
+	dev, err := CreateDevice(name)
+	if err != nil {
+		return 0, fmt.Errorf("DM_DEV_CREATE failed: %w", err)
+	}
+	if err := LoadTable(name, targets); err != nil {
+		_ = RemoveDevice(name)
+		return 0, fmt.Errorf("DM_TABLE_LOAD failed: %w", err)
+	}
+	if err := Resume(name); err != nil {
+		_ = RemoveDevice(name)
+		return 0, fmt.Errorf("DM_DEV_SUSPEND failed: %w", err)
+	}
+	return dev, nil
+}
diff --git a/metropolis/node/common/fileargs/BUILD.bazel b/metropolis/node/common/fileargs/BUILD.bazel
new file mode 100644
index 0000000..c4fffc2
--- /dev/null
+++ b/metropolis/node/common/fileargs/BUILD.bazel
@@ -0,0 +1,9 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["fileargs.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/fileargs",
+    visibility = ["//visibility:public"],
+    deps = ["@org_golang_x_sys//unix:go_default_library"],
+)
diff --git a/metropolis/node/common/fileargs/fileargs.go b/metropolis/node/common/fileargs/fileargs.go
new file mode 100644
index 0000000..26c054b
--- /dev/null
+++ b/metropolis/node/common/fileargs/fileargs.go
@@ -0,0 +1,101 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fileargs
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+
+	"golang.org/x/sys/unix"
+)
+
+// DefaultSize is the default size limit for FileArgs
+const DefaultSize = 4 * 1024 * 1024
+
+// TempDirectory is the directory where FileArgs will mount the actual files to. Defaults to
+// os.TempDir() but can be globally overridden by the application before any FileArgs are used.
+var TempDirectory = os.TempDir()
+
+type FileArgs struct {
+	path      string
+	lastError error
+}
+
+// New initializes a new set of file-based arguments. Remember to call Close() if you're done
+// using it, otherwise this leaks memory and mounts.
+func New() (*FileArgs, error) {
+	return NewWithSize(DefaultSize)
+}
+
+// NewWthSize is the same as new, but with a custom size limit. Please be aware that this data
+// cannot be swapped out and using a size limit that's too high can deadlock your kernel.
+func NewWithSize(size uint64) (*FileArgs, error) {
+	randomNameRaw := make([]byte, 128/8)
+	if _, err := io.ReadFull(rand.Reader, randomNameRaw); err != nil {
+		return nil, err
+	}
+	tmpPath := filepath.Join(TempDirectory, hex.EncodeToString(randomNameRaw))
+	if err := os.MkdirAll(tmpPath, 0700); err != nil {
+		return nil, err
+	}
+	// This uses ramfs instead of tmpfs because we never want to swap this for security reasons
+	if err := unix.Mount("none", tmpPath, "ramfs", unix.MS_NOEXEC|unix.MS_NOSUID|unix.MS_NODEV, fmt.Sprintf("size=%v", size)); err != nil {
+		return nil, err
+	}
+	return &FileArgs{
+		path: tmpPath,
+	}, nil
+}
+
+// ArgPath returns the path of the temporary file for this argument. It names the temporary
+// file according to name.
+func (f *FileArgs) ArgPath(name string, content []byte) string {
+	if f.lastError != nil {
+		return ""
+	}
+
+	path := filepath.Join(f.path, name)
+
+	if err := ioutil.WriteFile(path, content, 0600); err != nil {
+		f.lastError = err
+		return ""
+	}
+
+	return path
+}
+
+// FileOpt returns a full option with the temporary file name already filled in.
+// Example: `FileOpt("--testopt", "test.txt", []byte("hello")) == "--testopt=/tmp/daf8ed.../test.txt"`
+func (f *FileArgs) FileOpt(optName, fileName string, content []byte) string {
+	return fmt.Sprintf("%v=%v", optName, f.ArgPath(fileName, content))
+}
+
+func (f *FileArgs) Error() error {
+	return f.lastError
+}
+
+func (f *FileArgs) Close() error {
+	if err := unix.Unmount(f.path, 0); err != nil {
+		return err
+	}
+	return os.Remove(f.path)
+}
diff --git a/metropolis/node/common/fsquota/BUILD.bazel b/metropolis/node/common/fsquota/BUILD.bazel
new file mode 100644
index 0000000..b16d39e
--- /dev/null
+++ b/metropolis/node/common/fsquota/BUILD.bazel
@@ -0,0 +1,39 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//metropolis/test/ktest:ktest.bzl", "ktest")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "fsinfo.go",
+        "fsquota.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/fsquota",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//metropolis/node/common/fsquota/fsxattrs:go_default_library",
+        "//metropolis/node/common/fsquota/quotactl:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["fsquota_test.go"],
+    embed = [":go_default_library"],
+    pure = "on",
+    deps = [
+        "@com_github_stretchr_testify//require:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+ktest(
+    tester = ":go_default_test",
+    deps = [
+        "//third_party/xfsprogs:mkfs.xfs",
+    ],
+    initramfs_extra = """
+file /mkfs.xfs $(location //third_party/xfsprogs:mkfs.xfs) 0755 0 0
+    """,
+    cmdline = "ramdisk_size=51200",
+)
diff --git a/metropolis/node/common/fsquota/fsinfo.go b/metropolis/node/common/fsquota/fsinfo.go
new file mode 100644
index 0000000..e40a533
--- /dev/null
+++ b/metropolis/node/common/fsquota/fsinfo.go
@@ -0,0 +1,59 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsquota
+
+import (
+	"fmt"
+	"os"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// This requires fsinfo() support, which is not yet in any stable kernel.
+// Our kernel has that syscall backported. This would otherwise be an extremely expensive
+// operation and also involve lots of logic from our side.
+
+// From syscall_64.tbl
+const sys_fsinfo = 441
+
+// From uapi/linux/fsinfo.h
+const fsinfo_attr_source = 0x09
+const fsinfo_flags_query_path = 0x0000
+const fsinfo_flags_query_fd = 0x0001
+
+type fsinfoParams struct {
+	resolveFlags uint64
+	atFlags      uint32
+	flags        uint32
+	request      uint32
+	nth          uint32
+	mth          uint32
+}
+
+func fsinfoGetSource(dir *os.File) (string, error) {
+	buf := make([]byte, 256)
+	params := fsinfoParams{
+		flags:   fsinfo_flags_query_fd,
+		request: fsinfo_attr_source,
+	}
+	n, _, err := unix.Syscall6(sys_fsinfo, dir.Fd(), 0, uintptr(unsafe.Pointer(&params)), unsafe.Sizeof(params), uintptr(unsafe.Pointer(&buf[0])), 128)
+	if err != unix.Errno(0) {
+		return "", fmt.Errorf("failed to call fsinfo: %w", err)
+	}
+	return string(buf[:n]), nil
+}
diff --git a/metropolis/node/common/fsquota/fsquota.go b/metropolis/node/common/fsquota/fsquota.go
new file mode 100644
index 0000000..f702d23
--- /dev/null
+++ b/metropolis/node/common/fsquota/fsquota.go
@@ -0,0 +1,146 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package fsquota provides a simplified interface to interact with Linux's filesystem qouta API.
+// It only supports setting quotas on directories, not groups or users.
+// Quotas need to be already enabled on the filesystem to be able to use them using this package.
+// See the quotactl package if you intend to use this on a filesystem where quotas need to be
+// enabled manually.
+package fsquota
+
+import (
+	"fmt"
+	"math"
+	"os"
+
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fsquota/fsxattrs"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fsquota/quotactl"
+)
+
+// SetQuota sets the quota of bytes and/or inodes in a given path. To not set a limit, set the
+// corresponding argument to zero. Setting both arguments to zero removes the quota entirely.
+// This function can only be called on an empty directory. It can't be used to create a quota
+// below a directory which already has a quota since Linux doesn't offer hierarchical quotas.
+func SetQuota(path string, maxBytes uint64, maxInodes uint64) error {
+	dir, err := os.Open(path)
+	if err != nil {
+		return err
+	}
+	defer dir.Close()
+	source, err := fsinfoGetSource(dir)
+	if err != nil {
+		return err
+	}
+	var valid uint32
+	if maxBytes > 0 {
+		valid |= quotactl.FlagBLimitsValid
+	}
+	if maxInodes > 0 {
+		valid |= quotactl.FlagILimitsValid
+	}
+
+	attrs, err := fsxattrs.Get(dir)
+	if err != nil {
+		return err
+	}
+
+	var lastID uint32 = attrs.ProjectID
+	if lastID == 0 {
+		// No project/quota exists for this directory, assign a new project quota
+		// TODO(lorenz): This is racy, but the kernel does not support atomically assigning
+		// quotas. So this needs to be added to the kernels setquota interface. Due to the short
+		// time window and infrequent calls this should not be an immediate issue.
+		for {
+			quota, err := quotactl.GetNextQuota(source, quotactl.QuotaTypeProject, lastID)
+			if err == unix.ENOENT || err == unix.ESRCH {
+				// We have enumerated all quotas, nothing exists here
+				break
+			} else if err != nil {
+				return fmt.Errorf("failed to call GetNextQuota: %w", err)
+			}
+			if quota.ID > lastID+1 {
+				// Take the first ID in the quota ID gap
+				lastID++
+				break
+			}
+			lastID++
+		}
+	}
+
+	// If both limits are zero, this is a delete operation, process it as such
+	if maxBytes == 0 && maxInodes == 0 {
+		valid = quotactl.FlagBLimitsValid | quotactl.FlagILimitsValid
+		attrs.ProjectID = 0
+		attrs.Flags &= ^fsxattrs.FlagProjectInherit
+	} else {
+		attrs.ProjectID = lastID
+		attrs.Flags |= fsxattrs.FlagProjectInherit
+	}
+
+	if err := fsxattrs.Set(dir, attrs); err != nil {
+		return err
+	}
+
+	// Always round up to the nearest block size
+	bytesLimitBlocks := uint64(math.Ceil(float64(maxBytes) / float64(1024)))
+
+	return quotactl.SetQuota(source, quotactl.QuotaTypeProject, lastID, &quotactl.Quota{
+		BHardLimit: bytesLimitBlocks,
+		BSoftLimit: bytesLimitBlocks,
+		IHardLimit: maxInodes,
+		ISoftLimit: maxInodes,
+		Valid:      valid,
+	})
+}
+
+type Quota struct {
+	Bytes      uint64
+	BytesUsed  uint64
+	Inodes     uint64
+	InodesUsed uint64
+}
+
+// GetQuota returns the current active quota and its utilization at the given path
+func GetQuota(path string) (*Quota, error) {
+	dir, err := os.Open(path)
+	if err != nil {
+		return nil, err
+	}
+	defer dir.Close()
+	source, err := fsinfoGetSource(dir)
+	if err != nil {
+		return nil, err
+	}
+	attrs, err := fsxattrs.Get(dir)
+	if err != nil {
+		return nil, err
+	}
+	if attrs.ProjectID == 0 {
+		return nil, os.ErrNotExist
+	}
+	quota, err := quotactl.GetQuota(source, quotactl.QuotaTypeProject, attrs.ProjectID)
+	if err != nil {
+		return nil, err
+	}
+	return &Quota{
+		Bytes:      quota.BHardLimit * 1024,
+		BytesUsed:  quota.CurSpace,
+		Inodes:     quota.IHardLimit,
+		InodesUsed: quota.CurInodes,
+	}, nil
+}
diff --git a/metropolis/node/common/fsquota/fsquota_test.go b/metropolis/node/common/fsquota/fsquota_test.go
new file mode 100644
index 0000000..4729dac
--- /dev/null
+++ b/metropolis/node/common/fsquota/fsquota_test.go
@@ -0,0 +1,152 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsquota
+
+import (
+	"fmt"
+	"io/ioutil"
+	"math"
+	"os"
+	"os/exec"
+	"syscall"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+	"golang.org/x/sys/unix"
+)
+
+// withinTolerance is a helper for asserting that a value is within a certain percentage of the
+// expected value. The tolerance is specified as a float between 0 (exact match)
+// and 1 (between 0 and twice the expected value).
+func withinTolerance(t *testing.T, expected uint64, actual uint64, tolerance float64, name string) {
+	t.Helper()
+	delta := uint64(math.Round(float64(expected) * tolerance))
+	lowerBound := expected - delta
+	upperBound := expected + delta
+	if actual < lowerBound {
+		t.Errorf("Value %v (%v) is too low, expected between %v and %v", name, actual, lowerBound, upperBound)
+	}
+	if actual > upperBound {
+		t.Errorf("Value %v (%v) is too high, expected between %v and %v", name, actual, lowerBound, upperBound)
+	}
+}
+
+func TestBasic(t *testing.T) {
+	if os.Getenv("IN_KTEST") != "true" {
+		t.Skip("Not in ktest")
+	}
+	mkfsCmd := exec.Command("/mkfs.xfs", "-qf", "/dev/ram0")
+	if _, err := mkfsCmd.Output(); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Mkdir("/test", 0755); err != nil {
+		t.Error(err)
+	}
+
+	if err := unix.Mount("/dev/ram0", "/test", "xfs", unix.MS_NOEXEC|unix.MS_NODEV, "prjquota"); err != nil {
+		t.Fatal(err)
+	}
+	defer unix.Unmount("/test", 0)
+	defer os.RemoveAll("/test")
+	t.Run("SetQuota", func(t *testing.T) {
+		defer func() {
+			os.RemoveAll("/test/set")
+		}()
+		if err := os.Mkdir("/test/set", 0755); err != nil {
+			t.Fatal(err)
+		}
+		if err := SetQuota("/test/set", 1024*1024, 100); err != nil {
+			t.Fatal(err)
+		}
+	})
+	t.Run("SetQuotaAndExhaust", func(t *testing.T) {
+		defer func() {
+			os.RemoveAll("/test/sizequota")
+		}()
+		if err := os.Mkdir("/test/sizequota", 0755); err != nil {
+			t.Fatal(err)
+		}
+		const bytesQuota = 1024 * 1024 // 1MiB
+		if err := SetQuota("/test/sizequota", bytesQuota, 0); err != nil {
+			t.Fatal(err)
+		}
+		testfile, err := os.Create("/test/sizequota/testfile")
+		if err != nil {
+			t.Fatal(err)
+		}
+		testdata := make([]byte, 1024)
+		var bytesWritten int
+		for {
+			n, err := testfile.Write([]byte(testdata))
+			if err != nil {
+				if pathErr, ok := err.(*os.PathError); ok {
+					if pathErr.Err == syscall.ENOSPC {
+						// Running out of space is the only acceptable error to continue execution
+						break
+					}
+				}
+				t.Fatal(err)
+			}
+			bytesWritten += n
+		}
+		if bytesWritten > bytesQuota {
+			t.Errorf("Wrote %v bytes, quota is only %v bytes", bytesWritten, bytesQuota)
+		}
+	})
+	t.Run("GetQuotaReadbackAndUtilization", func(t *testing.T) {
+		defer func() {
+			os.RemoveAll("/test/readback")
+		}()
+		if err := os.Mkdir("/test/readback", 0755); err != nil {
+			t.Fatal(err)
+		}
+		const bytesQuota = 1024 * 1024 // 1MiB
+		const inodesQuota = 100
+		if err := SetQuota("/test/readback", bytesQuota, inodesQuota); err != nil {
+			t.Fatal(err)
+		}
+		sizeFileData := make([]byte, 512*1024)
+		if err := ioutil.WriteFile("/test/readback/512kfile", sizeFileData, 0644); err != nil {
+			t.Fatal(err)
+		}
+
+		quotaUtil, err := GetQuota("/test/readback")
+		if err != nil {
+			t.Fatal(err)
+		}
+		require.Equal(t, uint64(bytesQuota), quotaUtil.Bytes, "bytes quota readback incorrect")
+		require.Equal(t, uint64(inodesQuota), quotaUtil.Inodes, "inodes quota readback incorrect")
+
+		// Give 10% tolerance for quota used values to account for metadata overhead and internal
+		// structures that are also in there. If it's out by more than that it's an issue anyways.
+		withinTolerance(t, uint64(len(sizeFileData)), quotaUtil.BytesUsed, 0.1, "BytesUsed")
+
+		// Write 50 inodes for a total of 51 (with the 512K file)
+		for i := 0; i < 50; i++ {
+			if err := ioutil.WriteFile(fmt.Sprintf("/test/readback/ifile%v", i), []byte("test"), 0644); err != nil {
+				t.Fatal(err)
+			}
+		}
+
+		quotaUtil, err = GetQuota("/test/readback")
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		withinTolerance(t, 51, quotaUtil.InodesUsed, 0.1, "InodesUsed")
+	})
+}
diff --git a/metropolis/node/common/fsquota/fsxattrs/BUILD.bazel b/metropolis/node/common/fsquota/fsxattrs/BUILD.bazel
new file mode 100644
index 0000000..066200b
--- /dev/null
+++ b/metropolis/node/common/fsquota/fsxattrs/BUILD.bazel
@@ -0,0 +1,9 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["fsxattrs.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/fsquota/fsxattrs",
+    visibility = ["//visibility:public"],
+    deps = ["@org_golang_x_sys//unix:go_default_library"],
+)
diff --git a/metropolis/node/common/fsquota/fsxattrs/fsxattrs.go b/metropolis/node/common/fsquota/fsxattrs/fsxattrs.go
new file mode 100644
index 0000000..1d455eb
--- /dev/null
+++ b/metropolis/node/common/fsquota/fsxattrs/fsxattrs.go
@@ -0,0 +1,77 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsxattrs
+
+import (
+	"fmt"
+	"os"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+type FSXAttrFlag uint32
+
+// Defined in uapi/linux/fs.h
+const (
+	FlagRealtime        FSXAttrFlag = 0x00000001
+	FlagPreallocated    FSXAttrFlag = 0x00000002
+	FlagImmutable       FSXAttrFlag = 0x00000008
+	FlagAppend          FSXAttrFlag = 0x00000010
+	FlagSync            FSXAttrFlag = 0x00000020
+	FlagNoATime         FSXAttrFlag = 0x00000040
+	FlagNoDump          FSXAttrFlag = 0x00000080
+	FlagRealtimeInherit FSXAttrFlag = 0x00000100
+	FlagProjectInherit  FSXAttrFlag = 0x00000200
+	FlagNoSymlinks      FSXAttrFlag = 0x00000400
+	FlagExtentSize      FSXAttrFlag = 0x00000800
+	FlagNoDefragment    FSXAttrFlag = 0x00002000
+	FlagFilestream      FSXAttrFlag = 0x00004000
+	FlagDAX             FSXAttrFlag = 0x00008000
+	FlagCOWExtentSize   FSXAttrFlag = 0x00010000
+	FlagHasAttribute    FSXAttrFlag = 0x80000000
+)
+
+// FS_IOC_FSGETXATTR/FS_IOC_FSSETXATTR are defined in uapi/linux/fs.h
+const FS_IOC_FSGETXATTR = 0x801c581f
+const FS_IOC_FSSETXATTR = 0x401c5820
+
+type FSXAttrs struct {
+	Flags         FSXAttrFlag
+	ExtentSize    uint32
+	ExtentCount   uint32
+	ProjectID     uint32
+	CoWExtentSize uint32
+	_pad          [8]byte
+}
+
+func Get(file *os.File) (*FSXAttrs, error) {
+	var attrs FSXAttrs
+	_, _, errno := unix.Syscall(unix.SYS_IOCTL, file.Fd(), FS_IOC_FSGETXATTR, uintptr(unsafe.Pointer(&attrs)))
+	if errno != 0 {
+		return nil, fmt.Errorf("failed to execute getFSXAttrs: %v", errno)
+	}
+	return &attrs, nil
+}
+
+func Set(file *os.File, attrs *FSXAttrs) error {
+	_, _, errno := unix.Syscall(unix.SYS_IOCTL, file.Fd(), FS_IOC_FSSETXATTR, uintptr(unsafe.Pointer(attrs)))
+	if errno != 0 {
+		return fmt.Errorf("failed to execute setFSXAttrs: %v", errno)
+	}
+	return nil
+}
diff --git a/metropolis/node/common/fsquota/quotactl/BUILD.bazel b/metropolis/node/common/fsquota/quotactl/BUILD.bazel
new file mode 100644
index 0000000..c1582ad
--- /dev/null
+++ b/metropolis/node/common/fsquota/quotactl/BUILD.bazel
@@ -0,0 +1,9 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["quotactl.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/fsquota/quotactl",
+    visibility = ["//visibility:public"],
+    deps = ["@org_golang_x_sys//unix:go_default_library"],
+)
diff --git a/metropolis/node/common/fsquota/quotactl/quotactl.go b/metropolis/node/common/fsquota/quotactl/quotactl.go
new file mode 100644
index 0000000..5ed77d7
--- /dev/null
+++ b/metropolis/node/common/fsquota/quotactl/quotactl.go
@@ -0,0 +1,233 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package quotactl implements a low-level wrapper around the modern portion of Linux's
+// quotactl() syscall. See the fsquota package for a nicer interface to the most common part
+// of this API.
+package quotactl
+
+import (
+	"fmt"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+type QuotaType uint
+
+const (
+	QuotaTypeUser QuotaType = iota
+	QuotaTypeGroup
+	QuotaTypeProject
+)
+
+const (
+	Q_SYNC uint = ((0x800001 + iota) << 8)
+	Q_QUOTAON
+	Q_QUOTAOFF
+	Q_GETFMT
+	Q_GETINFO
+	Q_SETINFO
+	Q_GETQUOTA
+	Q_SETQUOTA
+	Q_GETNEXTQUOTA
+)
+
+const (
+	FlagBLimitsValid = 1 << iota
+	FlagSpaceValid
+	FlagILimitsValid
+	FlagInodesValid
+	FlagBTimeValid
+	FlagITimeValid
+)
+
+type DQInfo struct {
+	Bgrace uint64
+	Igrace uint64
+	Flags  uint32
+	Valid  uint32
+}
+
+type Quota struct {
+	BHardLimit uint64 // Both Byte limits are prescaled by 1024 (so are in KiB), but CurSpace is in B
+	BSoftLimit uint64
+	CurSpace   uint64
+	IHardLimit uint64
+	ISoftLimit uint64
+	CurInodes  uint64
+	BTime      uint64
+	ITime      uint64
+	Valid      uint32
+}
+
+type NextDQBlk struct {
+	HardLimitBytes  uint64
+	SoftLimitBytes  uint64
+	CurrentBytes    uint64
+	HardLimitInodes uint64
+	SoftLimitInodes uint64
+	CurrentInodes   uint64
+	BTime           uint64
+	ITime           uint64
+	Valid           uint32
+	ID              uint32
+}
+
+type QuotaFormat uint32
+
+// Collected from quota_format_type structs
+const (
+	// QuotaFormatNone is a special case where all quota information is
+	// stored inside filesystem metadata and thus requires no quotaFilePath.
+	QuotaFormatNone   QuotaFormat = 0
+	QuotaFormatVFSOld QuotaFormat = 1
+	QuotaFormatVFSV0  QuotaFormat = 2
+	QuotaFormatOCFS2  QuotaFormat = 3
+	QuotaFormatVFSV1  QuotaFormat = 4
+)
+
+// QuotaOn turns quota accounting and enforcement on
+func QuotaOn(device string, qtype QuotaType, quotaFormat QuotaFormat, quotaFilePath string) error {
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return err
+	}
+	pathArg, err := unix.BytePtrFromString(quotaFilePath)
+	if err != nil {
+		return err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_QUOTAON|uint(qtype)), uintptr(unsafe.Pointer(devArg)), uintptr(quotaFormat), uintptr(unsafe.Pointer(pathArg)), 0, 0)
+	if err != unix.Errno(0) {
+		return err
+	}
+	return nil
+}
+
+// QuotaOff turns quotas off
+func QuotaOff(device string, qtype QuotaType) error {
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_QUOTAOFF|uint(qtype)), uintptr(unsafe.Pointer(devArg)), 0, 0, 0, 0)
+	if err != unix.Errno(0) {
+		return err
+	}
+	return nil
+}
+
+// GetFmt gets the quota format used on given filesystem
+func GetFmt(device string, qtype QuotaType) (uint32, error) {
+	var fmt uint32
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return 0, err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_GETFMT|uint(qtype)), uintptr(unsafe.Pointer(devArg)), 0, uintptr(unsafe.Pointer(&fmt)), 0, 0)
+	if err != unix.Errno(0) {
+		return 0, err
+	}
+	return fmt, nil
+}
+
+// GetInfo gets information about quota files
+func GetInfo(device string, qtype QuotaType) (*DQInfo, error) {
+	var info DQInfo
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return nil, err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_GETINFO|uint(qtype)), uintptr(unsafe.Pointer(devArg)), 0, uintptr(unsafe.Pointer(&info)), 0, 0)
+	if err != unix.Errno(0) {
+		return nil, err
+	}
+	return &info, nil
+}
+
+// SetInfo sets information about quota files
+func SetInfo(device string, qtype QuotaType, info *DQInfo) error {
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_SETINFO|uint(qtype)), uintptr(unsafe.Pointer(devArg)), 0, uintptr(unsafe.Pointer(info)), 0, 0)
+	if err != unix.Errno(0) {
+		return err
+	}
+	return nil
+}
+
+// GetQuota gets user quota structure
+func GetQuota(device string, qtype QuotaType, id uint32) (*Quota, error) {
+	var info Quota
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return nil, err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_GETQUOTA|uint(qtype)), uintptr(unsafe.Pointer(devArg)), uintptr(id), uintptr(unsafe.Pointer(&info)), 0, 0)
+	if err != unix.Errno(0) {
+		return nil, err
+	}
+	return &info, nil
+}
+
+// GetNextQuota gets disk limits and usage > ID
+func GetNextQuota(device string, qtype QuotaType, id uint32) (*NextDQBlk, error) {
+	var info NextDQBlk
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return nil, err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_GETNEXTQUOTA|uint(qtype)), uintptr(unsafe.Pointer(devArg)), uintptr(id), uintptr(unsafe.Pointer(&info)), 0, 0)
+	if err != unix.Errno(0) {
+		return nil, err
+	}
+	return &info, nil
+}
+
+// SetQuota sets the given quota
+func SetQuota(device string, qtype QuotaType, id uint32, quota *Quota) error {
+	devArg, err := unix.BytePtrFromString(device)
+	if err != nil {
+		return err
+	}
+	_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_SETQUOTA|uint(qtype)), uintptr(unsafe.Pointer(devArg)), uintptr(id), uintptr(unsafe.Pointer(quota)), 0, 0)
+	if err != unix.Errno(0) {
+		return fmt.Errorf("failed to set quota: %w", err)
+	}
+	return nil
+}
+
+// Sync syncs disk copy of filesystems quotas. If device is empty it syncs all filesystems.
+func Sync(device string) error {
+	if device != "" {
+		devArg, err := unix.BytePtrFromString(device)
+		if err != nil {
+			return err
+		}
+		_, _, err = unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_SYNC), uintptr(unsafe.Pointer(devArg)), 0, 0, 0, 0)
+		if err != unix.Errno(0) {
+			return err
+		}
+	} else {
+		_, _, err := unix.Syscall6(unix.SYS_QUOTACTL, uintptr(Q_SYNC), 0, 0, 0, 0, 0)
+		if err != unix.Errno(0) {
+			return err
+		}
+	}
+	return nil
+}
diff --git a/metropolis/node/common/jsonpatch/BUILD.bazel b/metropolis/node/common/jsonpatch/BUILD.bazel
new file mode 100644
index 0000000..bd77e0a
--- /dev/null
+++ b/metropolis/node/common/jsonpatch/BUILD.bazel
@@ -0,0 +1,14 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["jsonpatch.go.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/jsonpatch",
+    visibility = ["//visibility:public"],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["jsonpatch_test.go"],
+    embed = [":go_default_library"],
+)
diff --git a/metropolis/node/common/jsonpatch/jsonpatch.go.go b/metropolis/node/common/jsonpatch/jsonpatch.go.go
new file mode 100644
index 0000000..9682980
--- /dev/null
+++ b/metropolis/node/common/jsonpatch/jsonpatch.go.go
@@ -0,0 +1,44 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package jsonpatch contains data structures and encoders for JSON Patch (RFC 6902) and JSON Pointers (RFC 6901)
+package jsonpatch
+
+import "strings"
+
+// JSON Patch operation (RFC 6902 Section 4)
+type JsonPatchOp struct {
+	Operation string      `json:"op"`
+	Path      string      `json:"path"` // Technically a JSON Pointer, but called Path in the RFC
+	From      string      `json:"from,omitempty"`
+	Value     interface{} `json:"value,omitempty"`
+}
+
+// EncodeJSONRefToken encodes a JSON reference token as part of a JSON Pointer (RFC 6901 Section 2)
+func EncodeJSONRefToken(token string) string {
+	x := strings.ReplaceAll(token, "~", "~0")
+	return strings.ReplaceAll(x, "/", "~1")
+}
+
+// PointerFromParts returns an encoded JSON Pointer from parts
+func PointerFromParts(pathParts []string) string {
+	var encodedParts []string
+	encodedParts = append(encodedParts, "")
+	for _, part := range pathParts {
+		encodedParts = append(encodedParts, EncodeJSONRefToken(part))
+	}
+	return strings.Join(encodedParts, "/")
+}
diff --git a/metropolis/node/common/jsonpatch/jsonpatch_test.go b/metropolis/node/common/jsonpatch/jsonpatch_test.go
new file mode 100644
index 0000000..33a56ba
--- /dev/null
+++ b/metropolis/node/common/jsonpatch/jsonpatch_test.go
@@ -0,0 +1,66 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package jsonpatch
+
+import (
+	"testing"
+)
+
+func TestEncodeJSONRefToken(t *testing.T) {
+	tests := []struct {
+		name  string
+		token string
+		want  string
+	}{
+		{"Passes through normal characters", "asdf123", "asdf123"},
+		{"Encodes simple slashes", "a/b", "a~1b"},
+		{"Encodes tildes", "m~n", "m~0n"},
+		{"Encodes bot tildes and slashes", "a/m~n", "a~1m~0n"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := EncodeJSONRefToken(tt.token); got != tt.want {
+				t.Errorf("EncodeJSONRefToken() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestPointerFromParts(t *testing.T) {
+	type args struct {
+		pathParts []string
+	}
+	tests := []struct {
+		name string
+		args args
+		want string
+	}{
+		{"Empty path", args{[]string{}}, ""},
+		{"Single level path", args{[]string{"foo"}}, "/foo"},
+		{"Multi-level path", args{[]string{"foo", "0"}}, "/foo/0"},
+		{"Path starting with empty key", args{[]string{""}}, "/"},
+		{"Path with part containing /", args{[]string{"a/b"}}, "/a~1b"},
+		{"Path with part containing spaces", args{[]string{" "}}, "/ "},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := PointerFromParts(tt.args.pathParts); got != tt.want {
+				t.Errorf("PointerFromParts() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
diff --git a/metropolis/node/common/logbuffer/BUILD.bazel b/metropolis/node/common/logbuffer/BUILD.bazel
new file mode 100644
index 0000000..2d4650d
--- /dev/null
+++ b/metropolis/node/common/logbuffer/BUILD.bazel
@@ -0,0 +1,22 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "linebuffer.go",
+        "logbuffer.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/logbuffer",
+    visibility = ["//visibility:public"],
+    deps = ["//metropolis/proto/api:go_default_library"],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = [
+        "linebuffer_test.go",
+        "logbuffer_test.go",
+    ],
+    embed = [":go_default_library"],
+    deps = ["@com_github_stretchr_testify//require:go_default_library"],
+)
diff --git a/metropolis/node/common/logbuffer/linebuffer.go b/metropolis/node/common/logbuffer/linebuffer.go
new file mode 100644
index 0000000..246a91b
--- /dev/null
+++ b/metropolis/node/common/logbuffer/linebuffer.go
@@ -0,0 +1,160 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logbuffer
+
+import (
+	"bytes"
+	"fmt"
+	"strings"
+	"sync"
+
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+// Line is a line stored in the log buffer - a string, that has been perhaps truncated (due to exceeded limits).
+type Line struct {
+	Data           string
+	OriginalLength int
+}
+
+// Truncated returns whether this line has been truncated to fit limits.
+func (l *Line) Truncated() bool {
+	return l.OriginalLength > len(l.Data)
+}
+
+// String returns the line with an ellipsis at the end (...) if the line has been truncated, or the original line
+// otherwise.
+func (l *Line) String() string {
+	if l.Truncated() {
+		return l.Data + "..."
+	}
+	return l.Data
+}
+
+// ProtoLog returns a Logging-specific protobuf structure.
+func (l *Line) ProtoLog() *apb.LogEntry_Raw {
+	return &apb.LogEntry_Raw{
+		Data:           l.Data,
+		OriginalLength: int64(l.OriginalLength),
+	}
+}
+
+// LineFromLogProto converts a Logging-specific protobuf message back into a Line.
+func LineFromLogProto(raw *apb.LogEntry_Raw) (*Line, error) {
+	if raw.OriginalLength < int64(len(raw.Data)) {
+		return nil, fmt.Errorf("original_length smaller than length of data")
+	}
+	originalLength := int(raw.OriginalLength)
+	if int64(originalLength) < raw.OriginalLength {
+		return nil, fmt.Errorf("original_length larger than native int size")
+	}
+	return &Line{
+		Data:           raw.Data,
+		OriginalLength: originalLength,
+	}, nil
+}
+
+// LineBuffer is a io.WriteCloser that will call a given callback every time a line is completed.
+type LineBuffer struct {
+	maxLineLength int
+	cb            LineBufferCallback
+
+	mu  sync.Mutex
+	cur strings.Builder
+	// length is the length of the line currently being written - this will continue to increase, even if the string
+	// exceeds maxLineLength.
+	length int
+	closed bool
+}
+
+// LineBufferCallback is a callback that will get called any time the line is completed. The function must not cause another
+// write to the LineBuffer, or the program will deadlock.
+type LineBufferCallback func(*Line)
+
+// NewLineBuffer creates a new LineBuffer with a given line length limit and callback.
+func NewLineBuffer(maxLineLength int, cb LineBufferCallback) *LineBuffer {
+	return &LineBuffer{
+		maxLineLength: maxLineLength,
+		cb:            cb,
+	}
+}
+
+// writeLimited writes to the internal buffer, making sure that its size does not exceed the maxLineLength.
+func (l *LineBuffer) writeLimited(data []byte) {
+	l.length += len(data)
+	if l.cur.Len()+len(data) > l.maxLineLength {
+		data = data[:l.maxLineLength-l.cur.Len()]
+	}
+	l.cur.Write(data)
+}
+
+// comitLine calls the callback and resets the builder.
+func (l *LineBuffer) commitLine() {
+	l.cb(&Line{
+		Data:           l.cur.String(),
+		OriginalLength: l.length,
+	})
+	l.cur.Reset()
+	l.length = 0
+}
+
+func (l *LineBuffer) Write(data []byte) (int, error) {
+	var pos = 0
+
+	l.mu.Lock()
+	defer l.mu.Unlock()
+
+	if l.closed {
+		return 0, fmt.Errorf("closed")
+	}
+
+	for {
+		nextNewline := bytes.IndexRune(data[pos:], '\n')
+
+		// No newline in the data, write everything to the current line
+		if nextNewline == -1 {
+			l.writeLimited(data[pos:])
+			break
+		}
+
+		// Write this line and update position
+		l.writeLimited(data[pos : pos+nextNewline])
+		l.commitLine()
+		pos += nextNewline + 1
+
+		// Data ends with a newline, stop now without writing an empty line
+		if nextNewline == len(data)-1 {
+			break
+		}
+	}
+	return len(data), nil
+}
+
+// Close will emit any leftover data in the buffer to the callback. Subsequent calls to Write will fail. Subsequent calls to Close
+// will also fail.
+func (l *LineBuffer) Close() error {
+	if l.closed {
+		return fmt.Errorf("already closed")
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.closed = true
+	if l.length > 0 {
+		l.commitLine()
+	}
+	return nil
+}
diff --git a/metropolis/node/common/logbuffer/linebuffer_test.go b/metropolis/node/common/logbuffer/linebuffer_test.go
new file mode 100644
index 0000000..c821a4b
--- /dev/null
+++ b/metropolis/node/common/logbuffer/linebuffer_test.go
@@ -0,0 +1,75 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logbuffer
+
+import (
+	"fmt"
+	"testing"
+)
+
+func TestLineBuffer(t *testing.T) {
+	var lines []*Line
+	lb := NewLineBuffer(1024, func(l *Line) {
+		lines = append(lines, l)
+	})
+
+	compare := func(a []*Line, b ...string) string {
+		msg := fmt.Sprintf("want %v, got %v", a, b)
+		if len(a) != len(b) {
+			return msg
+		}
+		for i, _ := range a {
+			if a[i].String() != b[i] {
+				return msg
+			}
+		}
+		return ""
+	}
+
+	// Write some data.
+	fmt.Fprintf(lb, "foo ")
+	if diff := compare(lines); diff != "" {
+		t.Fatal(diff)
+	}
+	fmt.Fprintf(lb, "bar\n")
+	if diff := compare(lines, "foo bar"); diff != "" {
+		t.Fatal(diff)
+	}
+	fmt.Fprintf(lb, "baz")
+	if diff := compare(lines, "foo bar"); diff != "" {
+		t.Fatal(diff)
+	}
+	fmt.Fprintf(lb, " baz")
+	if diff := compare(lines, "foo bar"); diff != "" {
+		t.Fatal(diff)
+	}
+	// Close and expect flush.
+	if err := lb.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	if diff := compare(lines, "foo bar", "baz baz"); diff != "" {
+		t.Fatal(diff)
+	}
+
+	// Check behaviour after close
+	if _, err := fmt.Fprintf(lb, "nope"); err == nil {
+		t.Fatalf("Write after Close: wanted  error, got nil")
+	}
+	if err := lb.Close(); err == nil {
+		t.Fatalf("second Close: wanted error, got nil")
+	}
+}
diff --git a/metropolis/node/common/logbuffer/logbuffer.go b/metropolis/node/common/logbuffer/logbuffer.go
new file mode 100644
index 0000000..ce47816
--- /dev/null
+++ b/metropolis/node/common/logbuffer/logbuffer.go
@@ -0,0 +1,97 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package logbuffer implements a fixed-size in-memory ring buffer for line-separated logs.
+// It implements io.Writer and splits the data into lines. The lines are kept in a ring where the
+// oldest are overwritten once it's full. It allows retrieval of the last n lines. There is a built-in
+// line length limit to bound the memory usage at maxLineLength * size.
+package logbuffer
+
+import (
+	"sync"
+)
+
+// LogBuffer implements a fixed-size in-memory ring buffer for line-separated logs
+type LogBuffer struct {
+	mu      sync.RWMutex
+	content []Line
+	length  int
+	*LineBuffer
+}
+
+// New creates a new LogBuffer with a given ringbuffer size and maximum line length.
+func New(size, maxLineLength int) *LogBuffer {
+	lb := &LogBuffer{
+		content: make([]Line, size),
+	}
+	lb.LineBuffer = NewLineBuffer(maxLineLength, lb.lineCallback)
+	return lb
+}
+
+func (b *LogBuffer) lineCallback(line *Line) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+
+	b.content[b.length%len(b.content)] = *line
+	b.length++
+}
+
+// capToContentLength caps the number of requested lines to what is actually available
+func (b *LogBuffer) capToContentLength(n int) int {
+	// If there aren't enough lines to read, reduce the request size
+	if n > b.length {
+		n = b.length
+	}
+	// If there isn't enough ringbuffer space, reduce the request size
+	if n > len(b.content) {
+		n = len(b.content)
+	}
+	return n
+}
+
+// ReadLines reads the last n lines from the buffer in chronological order. If n is bigger than the
+// ring buffer or the number of available lines only the number of stored lines are returned.
+func (b *LogBuffer) ReadLines(n int) []Line {
+	b.mu.RLock()
+	defer b.mu.RUnlock()
+
+	n = b.capToContentLength(n)
+
+	// Copy references out to keep them around
+	outArray := make([]Line, n)
+	for i := 1; i <= n; i++ {
+		outArray[n-i] = b.content[(b.length-i)%len(b.content)]
+	}
+	return outArray
+}
+
+// ReadLinesTruncated works exactly the same as ReadLines, but adds an ellipsis at the end of every
+// line that was truncated because it was over MaxLineLength
+func (b *LogBuffer) ReadLinesTruncated(n int, ellipsis string) []string {
+	b.mu.RLock()
+	defer b.mu.RUnlock()
+	// This does not use ReadLines() to prevent excessive reference copying and associated GC pressure
+	// since it could process a lot of lines.
+
+	n = b.capToContentLength(n)
+
+	outArray := make([]string, n)
+	for i := 1; i <= n; i++ {
+		line := b.content[(b.length-i)%len(b.content)]
+		outArray[n-i] = line.String()
+	}
+	return outArray
+}
diff --git a/metropolis/node/common/logbuffer/logbuffer_test.go b/metropolis/node/common/logbuffer/logbuffer_test.go
new file mode 100644
index 0000000..c38d7a6
--- /dev/null
+++ b/metropolis/node/common/logbuffer/logbuffer_test.go
@@ -0,0 +1,94 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logbuffer
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestSingleLine(t *testing.T) {
+	buf := New(1, 16000)
+	buf.Write([]byte("Hello World\n"))
+	out := buf.ReadLines(1)
+	require.Len(t, out, 1, "Invalid number of lines read")
+	require.Equal(t, "Hello World", out[0].Data, "Read bad log line")
+	require.Equal(t, 11, out[0].OriginalLength, "Invalid line length")
+}
+
+func TestPartialWritesAndReads(t *testing.T) {
+	buf := New(2, 16000)
+	buf.Write([]byte("Hello "))
+	buf.Write([]byte("World\nTest "))
+	buf.Write([]byte("2\n"))
+
+	out := buf.ReadLines(1)
+	require.Len(t, out, 1, "Invalid number of lines for partial read")
+	require.Equal(t, "Test 2", out[0].Data, "Read bad log line")
+
+	out2 := buf.ReadLines(2)
+	require.Len(t, out2, 2, "Invalid number of lines read")
+	require.Equal(t, "Hello World", out2[0].Data, "Read bad log line")
+	require.Equal(t, "Test 2", out2[1].Data, "Read bad log line")
+}
+
+func TestBufferOverwrite(t *testing.T) {
+	buf := New(3, 16000)
+	buf.Write([]byte("Test1\nTest2\nTest3\nTest4\n"))
+
+	out := buf.ReadLines(3)
+	require.Equal(t, out[0].Data, "Test2", "Read bad log line")
+	require.Equal(t, out[1].Data, "Test3", "Read bad log line")
+	require.Equal(t, out[2].Data, "Test4", "Overwritten data is invalid")
+}
+
+func TestTooLargeRequests(t *testing.T) {
+	buf := New(1, 16000)
+	outEmpty := buf.ReadLines(1)
+	require.Len(t, outEmpty, 0, "Returned more data than there is")
+
+	buf.Write([]byte("1\n2\n"))
+	out := buf.ReadLines(2)
+	require.Len(t, out, 1, "Returned more data than the ring buffer can hold")
+}
+
+func TestSpecialCases(t *testing.T) {
+	buf := New(2, 16000)
+	buf.Write([]byte("Test1"))
+	buf.Write([]byte("\nTest2\n"))
+	out := buf.ReadLines(2)
+	require.Len(t, out, 2, "Too many lines written")
+	require.Equal(t, out[0].Data, "Test1", "Read bad log line")
+	require.Equal(t, out[1].Data, "Test2", "Read bad log line")
+}
+
+func TestLineLengthLimit(t *testing.T) {
+	buf := New(2, 6)
+
+	testStr := "Just Testing"
+
+	buf.Write([]byte(testStr + "\nShort\n"))
+
+	out := buf.ReadLines(2)
+	require.Equal(t, len(testStr), out[0].OriginalLength, "Line is over length limit")
+	require.Equal(t, "Just T", out[0].Data, "Log line not properly truncated")
+
+	out2 := buf.ReadLinesTruncated(2, "...")
+	require.Equal(t, out2[0], "Just T...", "Line is over length limit")
+	require.Equal(t, out2[1], "Short", "Truncated small enough line")
+}
diff --git a/metropolis/node/common/supervisor/BUILD.bazel b/metropolis/node/common/supervisor/BUILD.bazel
new file mode 100644
index 0000000..ae95892
--- /dev/null
+++ b/metropolis/node/common/supervisor/BUILD.bazel
@@ -0,0 +1,28 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "supervisor.go",
+        "supervisor_node.go",
+        "supervisor_processor.go",
+        "supervisor_support.go",
+        "supervisor_testhelpers.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor",
+    visibility = [
+        "//metropolis/node:__subpackages__",
+        "//metropolis/test:__subpackages__",
+    ],
+    deps = [
+        "//metropolis/node/core/logtree:go_default_library",
+        "@com_github_cenkalti_backoff_v4//:go_default_library",
+        "@org_golang_google_grpc//:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["supervisor_test.go"],
+    embed = [":go_default_library"],
+)
diff --git a/metropolis/node/common/supervisor/supervisor.go b/metropolis/node/common/supervisor/supervisor.go
new file mode 100644
index 0000000..df7492c
--- /dev/null
+++ b/metropolis/node/common/supervisor/supervisor.go
@@ -0,0 +1,145 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package supervisor
+
+// The service supervision library allows for writing of reliable, service-style software within Smalltown.
+// It builds upon the Erlang/OTP supervision tree system, adapted to be more Go-ish.
+// For detailed design see go/supervision.
+
+import (
+	"context"
+	"io"
+	"sync"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+)
+
+// A Runnable is a function that will be run in a goroutine, and supervised throughout its lifetime. It can in turn
+// start more runnables as its children, and those will form part of a supervision tree.
+// The context passed to a runnable is very important and needs to be handled properly. It will be live (non-errored) as
+// long as the runnable should be running, and canceled (ctx.Err() will be non-nil) when the supervisor wants it to
+// exit. This means this context is also perfectly usable for performing any blocking operations.
+type Runnable func(ctx context.Context) error
+
+// RunGroup starts a set of runnables as a group. These runnables will run together, and if any one of them quits
+// unexpectedly, the result will be canceled and restarted.
+// The context here must be an existing Runnable context, and the spawned runnables will run under the node that this
+// context represents.
+func RunGroup(ctx context.Context, runnables map[string]Runnable) error {
+	node, unlock := fromContext(ctx)
+	defer unlock()
+	return node.runGroup(runnables)
+}
+
+// Run starts a single runnable in its own group.
+func Run(ctx context.Context, name string, runnable Runnable) error {
+	return RunGroup(ctx, map[string]Runnable{
+		name: runnable,
+	})
+}
+
+// Signal tells the supervisor that the calling runnable has reached a certain state of its lifecycle. All runnables
+// should SignalHealthy when they are ready with set up, running other child runnables and are now 'serving'.
+func Signal(ctx context.Context, signal SignalType) {
+	node, unlock := fromContext(ctx)
+	defer unlock()
+	node.signal(signal)
+}
+
+type SignalType int
+
+const (
+	// The runnable is healthy, done with setup, done with spawning more Runnables, and ready to serve in a loop.
+	// The runnable needs to check the parent context and ensure that if that context is done, the runnable exits.
+	SignalHealthy SignalType = iota
+	// The runnable is done - it does not need to run any loop. This is useful for Runnables that only set up other
+	// child runnables. This runnable will be restarted if a related failure happens somewhere in the supervision tree.
+	SignalDone
+)
+
+// supervisor represents and instance of the supervision system. It keeps track of a supervision tree and a request
+// channel to its internal processor goroutine.
+type supervisor struct {
+	// mu guards the entire state of the supervisor.
+	mu sync.RWMutex
+	// root is the root node of the supervision tree, named 'root'. It represents the Runnable started with the
+	// supervisor.New call.
+	root *node
+	// logtree is the main logtree exposed to runnables and used internally.
+	logtree *logtree.LogTree
+	// ilogger is the internal logger logging to "supervisor" in the logtree.
+	ilogger logtree.LeveledLogger
+
+	// pReq is an interface channel to the lifecycle processor of the supervisor.
+	pReq chan *processorRequest
+
+	// propagate panics, ie. don't catch them.
+	propagatePanic bool
+}
+
+// SupervisorOpt are runtime configurable options for the supervisor.
+type SupervisorOpt func(s *supervisor)
+
+var (
+	// WithPropagatePanic prevents the Supervisor from catching panics in runnables and treating them as failures.
+	// This is useful to enable for testing and local debugging.
+	WithPropagatePanic = func(s *supervisor) {
+		s.propagatePanic = true
+	}
+)
+
+func WithExistingLogtree(lt *logtree.LogTree) SupervisorOpt {
+	return func(s *supervisor) {
+		s.logtree = lt
+	}
+}
+
+// New creates a new supervisor with its root running the given root runnable.
+// The given context can be used to cancel the entire supervision tree.
+func New(ctx context.Context, rootRunnable Runnable, opts ...SupervisorOpt) *supervisor {
+	sup := &supervisor{
+		logtree: logtree.New(),
+		pReq:    make(chan *processorRequest),
+	}
+
+	for _, o := range opts {
+		o(sup)
+	}
+
+	sup.ilogger = sup.logtree.MustLeveledFor("supervisor")
+	sup.root = newNode("root", rootRunnable, sup, nil)
+
+	go sup.processor(ctx)
+
+	sup.pReq <- &processorRequest{
+		schedule: &processorRequestSchedule{dn: "root"},
+	}
+
+	return sup
+}
+
+func Logger(ctx context.Context) logtree.LeveledLogger {
+	node, unlock := fromContext(ctx)
+	defer unlock()
+	return node.sup.logtree.MustLeveledFor(logtree.DN(node.dn()))
+}
+
+func RawLogger(ctx context.Context) io.Writer {
+	node, unlock := fromContext(ctx)
+	defer unlock()
+	return node.sup.logtree.MustRawFor(logtree.DN(node.dn()))
+}
diff --git a/metropolis/node/common/supervisor/supervisor_node.go b/metropolis/node/common/supervisor/supervisor_node.go
new file mode 100644
index 0000000..a7caf82
--- /dev/null
+++ b/metropolis/node/common/supervisor/supervisor_node.go
@@ -0,0 +1,282 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package supervisor
+
+import (
+	"context"
+	"fmt"
+	"regexp"
+	"strings"
+
+	"github.com/cenkalti/backoff/v4"
+)
+
+// node is a supervision tree node. It represents the state of a Runnable within this tree, its relation to other tree
+// elements, and contains supporting data needed to actually supervise it.
+type node struct {
+	// The name of this node. Opaque string. It's used to make up the 'dn' (distinguished name) of a node within
+	// the tree. When starting a runnable inside a tree, this is where that name gets used.
+	name     string
+	runnable Runnable
+
+	// The supervisor managing this tree.
+	sup *supervisor
+	// The parent, within the tree, of this node. If this is the root node of the tree, this is nil.
+	parent *node
+	// Children of this tree. This is represented by a map keyed from child node names, for easy access.
+	children map[string]*node
+	// Supervision groups. Each group is a set of names of children. Sets, and as such groups, don't overlap between
+	// each other. A supervision group indicates that if any child within that group fails, all others should be
+	// canceled and restarted together.
+	groups []map[string]bool
+
+	// The current state of the runnable in this node.
+	state nodeState
+
+	// Backoff used to keep runnables from being restarted too fast.
+	bo *backoff.ExponentialBackOff
+
+	// Context passed to the runnable, and its cancel function.
+	ctx  context.Context
+	ctxC context.CancelFunc
+}
+
+// nodeState is the state of a runnable within a node, and in a way the node itself.
+// This follows the state diagram from go/supervision.
+type nodeState int
+
+const (
+	// A node that has just been created, and whose runnable has been started already but hasn't signaled anything yet.
+	nodeStateNew nodeState = iota
+	// A node whose runnable has signaled being healthy - this means it's ready to serve/act.
+	nodeStateHealthy
+	// A node that has unexpectedly returned or panicked.
+	nodeStateDead
+	// A node that has declared that its done with its work and should not be restarted, unless a supervision tree
+	// failure requires that.
+	nodeStateDone
+	// A node that has returned after being requested to cancel.
+	nodeStateCanceled
+)
+
+func (s nodeState) String() string {
+	switch s {
+	case nodeStateNew:
+		return "NODE_STATE_NEW"
+	case nodeStateHealthy:
+		return "NODE_STATE_HEALTHY"
+	case nodeStateDead:
+		return "NODE_STATE_DEAD"
+	case nodeStateDone:
+		return "NODE_STATE_DONE"
+	case nodeStateCanceled:
+		return "NODE_STATE_CANCELED"
+	}
+	return "UNKNOWN"
+}
+
+func (n *node) String() string {
+	return fmt.Sprintf("%s (%s)", n.dn(), n.state.String())
+}
+
+// contextKey is a type used to keep data within context values.
+type contextKey string
+
+var (
+	supervisorKey = contextKey("supervisor")
+	dnKey         = contextKey("dn")
+)
+
+// fromContext retrieves a tree node from a runnable context. It takes a lock on the tree and returns an unlock
+// function. This unlock function needs to be called once mutations on the tree/supervisor/node are done.
+func fromContext(ctx context.Context) (*node, func()) {
+	sup, ok := ctx.Value(supervisorKey).(*supervisor)
+	if !ok {
+		panic("supervisor function called from non-runnable context")
+	}
+
+	sup.mu.Lock()
+
+	dnParent, ok := ctx.Value(dnKey).(string)
+	if !ok {
+		sup.mu.Unlock()
+		panic("supervisor function called from non-runnable context")
+	}
+
+	return sup.nodeByDN(dnParent), sup.mu.Unlock
+}
+
+// All the following 'internal' supervisor functions must only be called with the supervisor lock taken. Getting a lock
+// via fromContext is enough.
+
+// dn returns the distinguished name of a node. This distinguished name is a period-separated, inverse-DNS-like name.
+// For instance, the runnable 'foo' within the runnable 'bar' will be called 'root.bar.foo'. The root of the tree is
+// always named, and has the dn, 'root'.
+func (n *node) dn() string {
+	if n.parent != nil {
+		return fmt.Sprintf("%s.%s", n.parent.dn(), n.name)
+	}
+	return n.name
+}
+
+// groupSiblings is a helper function to get all runnable group siblings of a given runnable name within this node.
+// All children are always in a group, even if that group is unary.
+func (n *node) groupSiblings(name string) map[string]bool {
+	for _, m := range n.groups {
+		if _, ok := m[name]; ok {
+			return m
+		}
+	}
+	return nil
+}
+
+// newNode creates a new node with a given parent. It does not register it with the parent (as that depends on group
+// placement).
+func newNode(name string, runnable Runnable, sup *supervisor, parent *node) *node {
+	// We use exponential backoff for failed runnables, but at some point we cap at a given backoff time.
+	// To achieve this, we set MaxElapsedTime to 0, which will cap the backoff at MaxInterval.
+	bo := backoff.NewExponentialBackOff()
+	bo.MaxElapsedTime = 0
+
+	n := &node{
+		name:     name,
+		runnable: runnable,
+
+		bo: bo,
+
+		sup:    sup,
+		parent: parent,
+	}
+	n.reset()
+	return n
+}
+
+// resetNode sets up all the dynamic fields of the node, in preparation of starting a runnable. It clears the node's
+// children, groups and resets its context.
+func (n *node) reset() {
+	// Make new context. First, acquire parent context. For the root node that's Background, otherwise it's the
+	// parent's context.
+	var pCtx context.Context
+	if n.parent == nil {
+		pCtx = context.Background()
+	} else {
+		pCtx = n.parent.ctx
+	}
+	// Mark DN and supervisor in context.
+	ctx := context.WithValue(pCtx, dnKey, n.dn())
+	ctx = context.WithValue(ctx, supervisorKey, n.sup)
+	ctx, ctxC := context.WithCancel(ctx)
+	// Set context
+	n.ctx = ctx
+	n.ctxC = ctxC
+
+	// Clear children and state
+	n.state = nodeStateNew
+	n.children = make(map[string]*node)
+	n.groups = nil
+
+	// The node is now ready to be scheduled.
+}
+
+// nodeByDN returns a node by given DN from the supervisor.
+func (s *supervisor) nodeByDN(dn string) *node {
+	parts := strings.Split(dn, ".")
+	if parts[0] != "root" {
+		panic("DN does not start with root.")
+	}
+	parts = parts[1:]
+	cur := s.root
+	for {
+		if len(parts) == 0 {
+			return cur
+		}
+
+		next, ok := cur.children[parts[0]]
+		if !ok {
+			panic(fmt.Errorf("could not find %v (%s) in %s", parts, dn, cur))
+		}
+		cur = next
+		parts = parts[1:]
+	}
+}
+
+// reNodeName validates a node name against constraints.
+var reNodeName = regexp.MustCompile(`[a-z90-9_]{1,64}`)
+
+// runGroup schedules a new group of runnables to run on a node.
+func (n *node) runGroup(runnables map[string]Runnable) error {
+	// Check that the parent node is in the right state.
+	if n.state != nodeStateNew {
+		return fmt.Errorf("cannot run new runnable on non-NEW node")
+	}
+
+	// Check the requested runnable names.
+	for name, _ := range runnables {
+		if !reNodeName.MatchString(name) {
+			return fmt.Errorf("runnable name %q is invalid", name)
+		}
+		if _, ok := n.children[name]; ok {
+			return fmt.Errorf("runnable %q already exists", name)
+		}
+	}
+
+	// Create child nodes.
+	dns := make(map[string]string)
+	group := make(map[string]bool)
+	for name, runnable := range runnables {
+		if g := n.groupSiblings(name); g != nil {
+			return fmt.Errorf("duplicate child name %q", name)
+		}
+		node := newNode(name, runnable, n.sup, n)
+		n.children[name] = node
+
+		dns[name] = node.dn()
+		group[name] = true
+	}
+	// Add group.
+	n.groups = append(n.groups, group)
+
+	// Schedule execution of group members.
+	go func() {
+		for name, _ := range runnables {
+			n.sup.pReq <- &processorRequest{
+				schedule: &processorRequestSchedule{
+					dn: dns[name],
+				},
+			}
+		}
+	}()
+	return nil
+}
+
+// signal sequences state changes by signals received from runnables and updates a node's status accordingly.
+func (n *node) signal(signal SignalType) {
+	switch signal {
+	case SignalHealthy:
+		if n.state != nodeStateNew {
+			panic(fmt.Errorf("node %s signaled healthy", n))
+		}
+		n.state = nodeStateHealthy
+		n.bo.Reset()
+	case SignalDone:
+		if n.state != nodeStateHealthy {
+			panic(fmt.Errorf("node %s signaled done", n))
+		}
+		n.state = nodeStateDone
+		n.bo.Reset()
+	}
+}
diff --git a/metropolis/node/common/supervisor/supervisor_processor.go b/metropolis/node/common/supervisor/supervisor_processor.go
new file mode 100644
index 0000000..965a667
--- /dev/null
+++ b/metropolis/node/common/supervisor/supervisor_processor.go
@@ -0,0 +1,404 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package supervisor
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"runtime/debug"
+	"time"
+)
+
+// The processor maintains runnable goroutines - ie., when requested will start one, and then once it exists it will
+// record the result and act accordingly. It is also responsible for detecting and acting upon supervision subtrees that
+// need to be restarted after death (via a 'GC' process)
+
+// processorRequest is a request for the processor. Only one of the fields can be set.
+type processorRequest struct {
+	schedule    *processorRequestSchedule
+	died        *processorRequestDied
+	waitSettled *processorRequestWaitSettled
+}
+
+// processorRequestSchedule requests that a given node's runnable be started.
+type processorRequestSchedule struct {
+	dn string
+}
+
+// processorRequestDied is a signal from a runnable goroutine that the runnable has died.
+type processorRequestDied struct {
+	dn  string
+	err error
+}
+
+type processorRequestWaitSettled struct {
+	waiter chan struct{}
+}
+
+// processor is the main processing loop.
+func (s *supervisor) processor(ctx context.Context) {
+	s.ilogger.Info("supervisor processor started")
+
+	// Waiters waiting for the GC to be settled.
+	var waiters []chan struct{}
+
+	// The GC will run every millisecond if needed. Any time the processor requests a change in the supervision tree
+	// (ie a death or a new runnable) it will mark the state as dirty and run the GC on the next millisecond cycle.
+	gc := time.NewTicker(1 * time.Millisecond)
+	defer gc.Stop()
+	clean := true
+
+	// How long has the GC been clean. This is used to notify 'settled' waiters.
+	cleanCycles := 0
+
+	markDirty := func() {
+		clean = false
+		cleanCycles = 0
+	}
+
+	for {
+		select {
+		case <-ctx.Done():
+			s.ilogger.Infof("supervisor processor exiting: %v", ctx.Err())
+			s.processKill()
+			s.ilogger.Info("supervisor exited")
+			return
+		case <-gc.C:
+			if !clean {
+				s.processGC()
+			}
+			clean = true
+			cleanCycles += 1
+
+			// This threshold is somewhat arbitrary. It's a balance between test speed and test reliability.
+			if cleanCycles > 50 {
+				for _, w := range waiters {
+					close(w)
+				}
+				waiters = nil
+			}
+		case r := <-s.pReq:
+			switch {
+			case r.schedule != nil:
+				s.processSchedule(r.schedule)
+				markDirty()
+			case r.died != nil:
+				s.processDied(r.died)
+				markDirty()
+			case r.waitSettled != nil:
+				waiters = append(waiters, r.waitSettled.waiter)
+			default:
+				panic(fmt.Errorf("unhandled request %+v", r))
+			}
+		}
+	}
+}
+
+// processKill cancels all nodes in the supervision tree. This is only called right before exiting the processor, so
+// they do not get automatically restarted.
+func (s *supervisor) processKill() {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// Gather all context cancel functions.
+	var cancels []func()
+	queue := []*node{s.root}
+	for {
+		if len(queue) == 0 {
+			break
+		}
+
+		cur := queue[0]
+		queue = queue[1:]
+
+		cancels = append(cancels, cur.ctxC)
+		for _, c := range cur.children {
+			queue = append(queue, c)
+		}
+	}
+
+	// Call all context cancels.
+	for _, c := range cancels {
+		c()
+	}
+}
+
+// processSchedule starts a node's runnable in a goroutine and records its output once it's done.
+func (s *supervisor) processSchedule(r *processorRequestSchedule) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	n := s.nodeByDN(r.dn)
+	go func() {
+		if !s.propagatePanic {
+			defer func() {
+				if rec := recover(); rec != nil {
+					s.pReq <- &processorRequest{
+						died: &processorRequestDied{
+							dn:  r.dn,
+							err: fmt.Errorf("panic: %v, stacktrace: %s", rec, string(debug.Stack())),
+						},
+					}
+				}
+			}()
+		}
+
+		res := n.runnable(n.ctx)
+
+		s.pReq <- &processorRequest{
+			died: &processorRequestDied{
+				dn:  r.dn,
+				err: res,
+			},
+		}
+	}()
+}
+
+// processDied records the result from a runnable goroutine, and updates its node state accordingly. If the result
+// is a death and not an expected exit, related nodes (ie. children and group siblings) are canceled accordingly.
+func (s *supervisor) processDied(r *processorRequestDied) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// Okay, so a Runnable has quit. What now?
+	n := s.nodeByDN(r.dn)
+	ctx := n.ctx
+
+	// Simple case: it was marked as Done and quit with no error.
+	if n.state == nodeStateDone && r.err == nil {
+		// Do nothing. This was supposed to happen. Keep the process as DONE.
+		return
+	}
+
+	// Find innermost error to check if it's a context canceled error.
+	perr := r.err
+	for {
+		if inner := errors.Unwrap(perr); inner != nil {
+			perr = inner
+			continue
+		}
+		break
+	}
+
+	// Simple case: the context was canceled and the returned error is the context error.
+	if err := ctx.Err(); err != nil && perr == err {
+		// Mark the node as canceled successfully.
+		n.state = nodeStateCanceled
+		return
+	}
+
+	// Otherwise, the Runnable should not have died or quit. Handle accordingly.
+	err := r.err
+	// A lack of returned error is also an error.
+	if err == nil {
+		err = fmt.Errorf("returned when %s", n.state)
+	} else {
+		err = fmt.Errorf("returned error when %s: %w", n.state, err)
+	}
+
+	s.ilogger.Errorf("Runnable %s died: %v", n.dn(), err)
+	// Mark as dead.
+	n.state = nodeStateDead
+
+	// Cancel that node's context, just in case something still depends on it.
+	n.ctxC()
+
+	// Cancel all siblings.
+	if n.parent != nil {
+		for name, _ := range n.parent.groupSiblings(n.name) {
+			if name == n.name {
+				continue
+			}
+			sibling := n.parent.children[name]
+			// TODO(q3k): does this need to run in a goroutine, ie. can a context cancel block?
+			sibling.ctxC()
+		}
+	}
+}
+
+// processGC runs the GC process. It's not really Garbage Collection, as in, it doesn't remove unnecessary tree nodes -
+// but it does find nodes that need to be restarted, find the subset that can and then schedules them for running.
+// As such, it's less of a Garbage Collector and more of a Necromancer. However, GC is a friendlier name.
+func (s *supervisor) processGC() {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// The 'GC' serves is the main business logic of the supervision tree. It traverses a locked tree and tries to
+	// find subtrees that must be restarted (because of a DEAD/CANCELED runnable). It then finds which of these
+	// subtrees that should be restarted can be restarted, ie. which ones are fully recursively DEAD/CANCELED. It
+	// also finds the smallest set of largest subtrees that can be restarted, ie. if there's multiple DEAD runnables
+	// that can be restarted at once, it will do so.
+
+	// Phase one: Find all leaves.
+	// This is a simple DFS that finds all the leaves of the tree, ie all nodes that do not have children nodes.
+	leaves := make(map[string]bool)
+	queue := []*node{s.root}
+	for {
+		if len(queue) == 0 {
+			break
+		}
+		cur := queue[0]
+		queue = queue[1:]
+
+		for _, c := range cur.children {
+			queue = append([]*node{c}, queue...)
+		}
+
+		if len(cur.children) == 0 {
+			leaves[cur.dn()] = true
+		}
+	}
+
+	// Phase two: traverse tree from node to root and make note of all subtrees that can be restarted.
+	// A subtree is restartable/ready iff every node in that subtree is either CANCELED, DEAD or DONE.
+	// Such a 'ready' subtree can be restarted by the supervisor if needed.
+
+	// DNs that we already visited.
+	visited := make(map[string]bool)
+	// DNs whose subtrees are ready to be restarted.
+	// These are all subtrees recursively - ie., root.a.a and root.a will both be marked here.
+	ready := make(map[string]bool)
+
+	// We build a queue of nodes to visit, starting from the leaves.
+	queue = []*node{}
+	for l, _ := range leaves {
+		queue = append(queue, s.nodeByDN(l))
+	}
+
+	for {
+		if len(queue) == 0 {
+			break
+		}
+
+		cur := queue[0]
+		curDn := cur.dn()
+
+		queue = queue[1:]
+
+		// Do we have a decision about our children?
+		allVisited := true
+		for _, c := range cur.children {
+			if !visited[c.dn()] {
+				allVisited = false
+				break
+			}
+		}
+
+		// If no decision about children is available, it means we ended up in this subtree through some shorter path
+		// of a shorter/lower-order leaf. There is a path to a leaf that's longer than the one that caused this node
+		// to be enqueued. Easy solution: just push back the current element and retry later.
+		if !allVisited {
+			// Push back to queue and wait for a decision later.
+			queue = append(queue, cur)
+			continue
+		}
+
+		// All children have been visited and we have an idea about whether they're ready/restartable. All of the node's
+		// children must be restartable in order for this node to be restartable.
+		childrenReady := true
+		for _, c := range cur.children {
+			if !ready[c.dn()] {
+				childrenReady = false
+				break
+			}
+		}
+
+		// In addition to children, the node itself must be restartable (ie. DONE, DEAD or CANCELED).
+		curReady := false
+		switch cur.state {
+		case nodeStateDone:
+			curReady = true
+		case nodeStateCanceled:
+			curReady = true
+		case nodeStateDead:
+			curReady = true
+		}
+
+		// Note down that we have an opinion on this node, and note that opinion down.
+		visited[curDn] = true
+		ready[curDn] = childrenReady && curReady
+
+		// Now we can also enqueue the parent of this node for processing.
+		if cur.parent != nil && !visited[cur.parent.dn()] {
+			queue = append(queue, cur.parent)
+		}
+	}
+
+	// Phase 3: traverse tree from root to find largest subtrees that need to be restarted and are ready to be
+	// restarted.
+
+	// All DNs that need to be restarted by the GC process.
+	want := make(map[string]bool)
+	// All DNs that need to be restarted and can be restarted by the GC process - a subset of 'want' DNs.
+	can := make(map[string]bool)
+	// The set difference between 'want' and 'can' are all nodes that should be restarted but can't yet (ie. because
+	// a child is still in the process of being canceled).
+
+	// DFS from root.
+	queue = []*node{s.root}
+	for {
+		if len(queue) == 0 {
+			break
+		}
+
+		cur := queue[0]
+		queue = queue[1:]
+
+		// If this node is DEAD or CANCELED it should be restarted.
+		if cur.state == nodeStateDead || cur.state == nodeStateCanceled {
+			want[cur.dn()] = true
+		}
+
+		// If it should be restarted and is ready to be restarted...
+		if want[cur.dn()] && ready[cur.dn()] {
+			// And its parent context is valid (ie hasn't been canceled), mark it as restartable.
+			if cur.parent == nil || cur.parent.ctx.Err() == nil {
+				can[cur.dn()] = true
+				continue
+			}
+		}
+
+		// Otherwise, traverse further down the tree to see if something else needs to be done.
+		for _, c := range cur.children {
+			queue = append(queue, c)
+		}
+	}
+
+	// Reinitialize and reschedule all subtrees
+	for dn, _ := range can {
+		n := s.nodeByDN(dn)
+
+		// Only back off when the node unexpectedly died - not when it got canceled.
+		bo := time.Duration(0)
+		if n.state == nodeStateDead {
+			bo = n.bo.NextBackOff()
+		}
+
+		// Prepare node for rescheduling - remove its children, reset its state to new.
+		n.reset()
+		s.ilogger.Infof("rescheduling supervised node %s with backoff %s", dn, bo.String())
+
+		// Reschedule node runnable to run after backoff.
+		go func(n *node, bo time.Duration) {
+			time.Sleep(bo)
+			s.pReq <- &processorRequest{
+				schedule: &processorRequestSchedule{dn: n.dn()},
+			}
+		}(n, bo)
+	}
+}
diff --git a/metropolis/node/common/supervisor/supervisor_support.go b/metropolis/node/common/supervisor/supervisor_support.go
new file mode 100644
index 0000000..d54b35c
--- /dev/null
+++ b/metropolis/node/common/supervisor/supervisor_support.go
@@ -0,0 +1,62 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package supervisor
+
+// Supporting infrastructure to allow running some non-Go payloads under supervision.
+
+import (
+	"context"
+	"net"
+	"os/exec"
+
+	"google.golang.org/grpc"
+)
+
+// GRPCServer creates a Runnable that serves gRPC requests as longs as it's not canceled.
+// If graceful is set to true, the server will be gracefully stopped instead of plain stopped. This means all pending
+// RPCs will finish, but also requires streaming gRPC handlers to check their context liveliness and exit accordingly.
+// If the server code does not support this, `graceful` should be false and the server will be killed violently instead.
+func GRPCServer(srv *grpc.Server, lis net.Listener, graceful bool) Runnable {
+	return func(ctx context.Context) error {
+		Signal(ctx, SignalHealthy)
+		errC := make(chan error)
+		go func() {
+			errC <- srv.Serve(lis)
+		}()
+		select {
+		case <-ctx.Done():
+			if graceful {
+				srv.GracefulStop()
+			} else {
+				srv.Stop()
+			}
+			return ctx.Err()
+		case err := <-errC:
+			return err
+		}
+	}
+}
+
+// RunCommand will create a Runnable that starts a long-running command, whose exit is determined to be a failure.
+func RunCommand(ctx context.Context, cmd *exec.Cmd) error {
+	Signal(ctx, SignalHealthy)
+	cmd.Stdout = RawLogger(ctx)
+	cmd.Stderr = RawLogger(ctx)
+	err := cmd.Run()
+	Logger(ctx).Infof("Command returned: %v", err)
+	return err
+}
diff --git a/metropolis/node/common/supervisor/supervisor_test.go b/metropolis/node/common/supervisor/supervisor_test.go
new file mode 100644
index 0000000..9c7bdb7
--- /dev/null
+++ b/metropolis/node/common/supervisor/supervisor_test.go
@@ -0,0 +1,557 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package supervisor
+
+import (
+	"context"
+	"fmt"
+	"testing"
+	"time"
+)
+
+func runnableBecomesHealthy(healthy, done chan struct{}) Runnable {
+	return func(ctx context.Context) error {
+		Signal(ctx, SignalHealthy)
+
+		go func() {
+			if healthy != nil {
+				healthy <- struct{}{}
+			}
+		}()
+
+		<-ctx.Done()
+
+		go func() {
+			if done != nil {
+				done <- struct{}{}
+			}
+		}()
+
+		return ctx.Err()
+	}
+}
+
+func runnableSpawnsMore(healthy, done chan struct{}, levels int) Runnable {
+	return func(ctx context.Context) error {
+		if levels > 0 {
+			err := RunGroup(ctx, map[string]Runnable{
+				"a": runnableSpawnsMore(nil, nil, levels-1),
+				"b": runnableSpawnsMore(nil, nil, levels-1),
+			})
+			if err != nil {
+				return err
+			}
+		}
+
+		Signal(ctx, SignalHealthy)
+
+		go func() {
+			if healthy != nil {
+				healthy <- struct{}{}
+			}
+		}()
+
+		<-ctx.Done()
+
+		go func() {
+			if done != nil {
+				done <- struct{}{}
+			}
+		}()
+		return ctx.Err()
+	}
+}
+
+// rc is a Remote Controlled runnable. It is a generic runnable used for testing the supervisor.
+type rc struct {
+	req chan rcRunnableRequest
+}
+
+type rcRunnableRequest struct {
+	cmd    rcRunnableCommand
+	stateC chan rcRunnableState
+}
+
+type rcRunnableCommand int
+
+const (
+	rcRunnableCommandBecomeHealthy rcRunnableCommand = iota
+	rcRunnableCommandBecomeDone
+	rcRunnableCommandDie
+	rcRunnableCommandPanic
+	rcRunnableCommandState
+)
+
+type rcRunnableState int
+
+const (
+	rcRunnableStateNew rcRunnableState = iota
+	rcRunnableStateHealthy
+	rcRunnableStateDone
+)
+
+func (r *rc) becomeHealthy() {
+	r.req <- rcRunnableRequest{cmd: rcRunnableCommandBecomeHealthy}
+}
+
+func (r *rc) becomeDone() {
+	r.req <- rcRunnableRequest{cmd: rcRunnableCommandBecomeDone}
+}
+func (r *rc) die() {
+	r.req <- rcRunnableRequest{cmd: rcRunnableCommandDie}
+}
+
+func (r *rc) panic() {
+	r.req <- rcRunnableRequest{cmd: rcRunnableCommandPanic}
+}
+
+func (r *rc) state() rcRunnableState {
+	c := make(chan rcRunnableState)
+	r.req <- rcRunnableRequest{
+		cmd:    rcRunnableCommandState,
+		stateC: c,
+	}
+	return <-c
+}
+
+func (r *rc) waitState(s rcRunnableState) {
+	// This is poll based. Making it non-poll based would make the RC runnable logic a bit more complex for little gain.
+	for {
+		got := r.state()
+		if got == s {
+			return
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+}
+
+func newRC() *rc {
+	return &rc{
+		req: make(chan rcRunnableRequest),
+	}
+}
+
+// Remote Controlled Runnable
+func (r *rc) runnable() Runnable {
+	return func(ctx context.Context) error {
+		state := rcRunnableStateNew
+
+		for {
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			case r := <-r.req:
+				switch r.cmd {
+				case rcRunnableCommandBecomeHealthy:
+					Signal(ctx, SignalHealthy)
+					state = rcRunnableStateHealthy
+				case rcRunnableCommandBecomeDone:
+					Signal(ctx, SignalDone)
+					state = rcRunnableStateDone
+				case rcRunnableCommandDie:
+					return fmt.Errorf("died on request")
+				case rcRunnableCommandPanic:
+					panic("at the disco")
+				case rcRunnableCommandState:
+					r.stateC <- state
+				}
+			}
+		}
+	}
+}
+
+func TestSimple(t *testing.T) {
+	h1 := make(chan struct{})
+	d1 := make(chan struct{})
+	h2 := make(chan struct{})
+	d2 := make(chan struct{})
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+	s := New(ctx, func(ctx context.Context) error {
+		err := RunGroup(ctx, map[string]Runnable{
+			"one": runnableBecomesHealthy(h1, d1),
+			"two": runnableBecomesHealthy(h2, d2),
+		})
+		if err != nil {
+			return err
+		}
+		Signal(ctx, SignalHealthy)
+		Signal(ctx, SignalDone)
+		return nil
+	}, WithPropagatePanic)
+
+	// Expect both to start running.
+	s.waitSettleError(ctx, t)
+	select {
+	case <-h1:
+	default:
+		t.Fatalf("runnable 'one' didn't start")
+	}
+	select {
+	case <-h2:
+	default:
+		t.Fatalf("runnable 'one' didn't start")
+	}
+}
+
+func TestSimpleFailure(t *testing.T) {
+	h1 := make(chan struct{})
+	d1 := make(chan struct{})
+	two := newRC()
+
+	ctx, ctxC := context.WithTimeout(context.Background(), 10*time.Second)
+	defer ctxC()
+	s := New(ctx, func(ctx context.Context) error {
+		err := RunGroup(ctx, map[string]Runnable{
+			"one": runnableBecomesHealthy(h1, d1),
+			"two": two.runnable(),
+		})
+		if err != nil {
+			return err
+		}
+		Signal(ctx, SignalHealthy)
+		Signal(ctx, SignalDone)
+		return nil
+	}, WithPropagatePanic)
+	s.waitSettleError(ctx, t)
+
+	two.becomeHealthy()
+	s.waitSettleError(ctx, t)
+	// Expect one to start running.
+	select {
+	case <-h1:
+	default:
+		t.Fatalf("runnable 'one' didn't start")
+	}
+
+	// Kill off two, one should restart.
+	two.die()
+	s.waitSettleError(ctx, t)
+	select {
+	case <-d1:
+	default:
+		t.Fatalf("runnable 'one' didn't acknowledge cancel")
+	}
+
+	// And one should start running again.
+	s.waitSettleError(ctx, t)
+	select {
+	case <-h1:
+	default:
+		t.Fatalf("runnable 'one' didn't restart")
+	}
+}
+
+func TestDeepFailure(t *testing.T) {
+	h1 := make(chan struct{})
+	d1 := make(chan struct{})
+	two := newRC()
+
+	ctx, ctxC := context.WithTimeout(context.Background(), 10*time.Second)
+	defer ctxC()
+	s := New(ctx, func(ctx context.Context) error {
+		err := RunGroup(ctx, map[string]Runnable{
+			"one": runnableSpawnsMore(h1, d1, 5),
+			"two": two.runnable(),
+		})
+		if err != nil {
+			return err
+		}
+		Signal(ctx, SignalHealthy)
+		Signal(ctx, SignalDone)
+		return nil
+	}, WithPropagatePanic)
+
+	two.becomeHealthy()
+	s.waitSettleError(ctx, t)
+	// Expect one to start running.
+	select {
+	case <-h1:
+	default:
+		t.Fatalf("runnable 'one' didn't start")
+	}
+
+	// Kill off two, one should restart.
+	two.die()
+	s.waitSettleError(ctx, t)
+	select {
+	case <-d1:
+	default:
+		t.Fatalf("runnable 'one' didn't acknowledge cancel")
+	}
+
+	// And one should start running again.
+	s.waitSettleError(ctx, t)
+	select {
+	case <-h1:
+	default:
+		t.Fatalf("runnable 'one' didn't restart")
+	}
+}
+
+func TestPanic(t *testing.T) {
+	h1 := make(chan struct{})
+	d1 := make(chan struct{})
+	two := newRC()
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+	s := New(ctx, func(ctx context.Context) error {
+		err := RunGroup(ctx, map[string]Runnable{
+			"one": runnableBecomesHealthy(h1, d1),
+			"two": two.runnable(),
+		})
+		if err != nil {
+			return err
+		}
+		Signal(ctx, SignalHealthy)
+		Signal(ctx, SignalDone)
+		return nil
+	})
+
+	two.becomeHealthy()
+	s.waitSettleError(ctx, t)
+	// Expect one to start running.
+	select {
+	case <-h1:
+	default:
+		t.Fatalf("runnable 'one' didn't start")
+	}
+
+	// Kill off two, one should restart.
+	two.panic()
+	s.waitSettleError(ctx, t)
+	select {
+	case <-d1:
+	default:
+		t.Fatalf("runnable 'one' didn't acknowledge cancel")
+	}
+
+	// And one should start running again.
+	s.waitSettleError(ctx, t)
+	select {
+	case <-h1:
+	default:
+		t.Fatalf("runnable 'one' didn't restart")
+	}
+}
+
+func TestMultipleLevelFailure(t *testing.T) {
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+	New(ctx, func(ctx context.Context) error {
+		err := RunGroup(ctx, map[string]Runnable{
+			"one": runnableSpawnsMore(nil, nil, 4),
+			"two": runnableSpawnsMore(nil, nil, 4),
+		})
+		if err != nil {
+			return err
+		}
+		Signal(ctx, SignalHealthy)
+		Signal(ctx, SignalDone)
+		return nil
+	}, WithPropagatePanic)
+}
+
+func TestBackoff(t *testing.T) {
+	one := newRC()
+
+	ctx, ctxC := context.WithTimeout(context.Background(), 20*time.Second)
+	defer ctxC()
+
+	s := New(ctx, func(ctx context.Context) error {
+		if err := Run(ctx, "one", one.runnable()); err != nil {
+			return err
+		}
+		Signal(ctx, SignalHealthy)
+		Signal(ctx, SignalDone)
+		return nil
+	}, WithPropagatePanic)
+
+	one.becomeHealthy()
+	// Die a bunch of times in a row, this brings up the next exponential backoff to over a second.
+	for i := 0; i < 4; i += 1 {
+		one.die()
+		one.waitState(rcRunnableStateNew)
+	}
+	// Measure how long it takes for the runnable to respawn after a number of failures
+	start := time.Now()
+	one.die()
+	one.becomeHealthy()
+	one.waitState(rcRunnableStateHealthy)
+	taken := time.Since(start)
+	if taken < 1*time.Second {
+		t.Errorf("Runnable took %v to restart, wanted at least a second from backoff", taken)
+	}
+
+	s.waitSettleError(ctx, t)
+	// Now that we've become healthy, die again. Becoming healthy resets the backoff.
+	start = time.Now()
+	one.die()
+	one.becomeHealthy()
+	one.waitState(rcRunnableStateHealthy)
+	taken = time.Since(start)
+	if taken > 1*time.Second || taken < 100*time.Millisecond {
+		t.Errorf("Runnable took %v to restart, wanted at least 100ms from backoff and at most 1s from backoff reset", taken)
+	}
+}
+
+// TestResilience throws some curveballs at the supervisor - either programming errors or high load. It then ensures
+// that another runnable is running, and that it restarts on its sibling failure.
+func TestResilience(t *testing.T) {
+	// request/response channel for testing liveness of the 'one' runnable
+	req := make(chan chan struct{})
+
+	// A runnable that responds on the 'req' channel.
+	one := func(ctx context.Context) error {
+		Signal(ctx, SignalHealthy)
+		for {
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			case r := <-req:
+				r <- struct{}{}
+			}
+		}
+	}
+	oneSibling := newRC()
+
+	oneTest := func() {
+		timeout := time.NewTicker(1000 * time.Millisecond)
+		ping := make(chan struct{})
+		req <- ping
+		select {
+		case <-ping:
+		case <-timeout.C:
+			t.Fatalf("one ping response timeout")
+		}
+		timeout.Stop()
+	}
+
+	// A nasty runnable that calls Signal with the wrong context (this is a programming error)
+	two := func(ctx context.Context) error {
+		Signal(context.TODO(), SignalHealthy)
+		return nil
+	}
+
+	// A nasty runnable that calls Signal wrong (this is a programming error).
+	three := func(ctx context.Context) error {
+		Signal(ctx, SignalDone)
+		return nil
+	}
+
+	// A nasty runnable that runs in a busy loop (this is a programming error).
+	four := func(ctx context.Context) error {
+		for {
+			time.Sleep(0)
+		}
+	}
+
+	// A nasty runnable that keeps creating more runnables.
+	five := func(ctx context.Context) error {
+		i := 1
+		for {
+			err := Run(ctx, fmt.Sprintf("r%d", i), runnableSpawnsMore(nil, nil, 2))
+			if err != nil {
+				return err
+			}
+
+			time.Sleep(100 * time.Millisecond)
+			i += 1
+		}
+	}
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+	New(ctx, func(ctx context.Context) error {
+		RunGroup(ctx, map[string]Runnable{
+			"one":        one,
+			"oneSibling": oneSibling.runnable(),
+		})
+		rs := map[string]Runnable{
+			"two": two, "three": three, "four": four, "five": five,
+		}
+		for k, v := range rs {
+			if err := Run(ctx, k, v); err != nil {
+				return err
+			}
+		}
+		Signal(ctx, SignalHealthy)
+		Signal(ctx, SignalDone)
+		return nil
+	})
+
+	// Five rounds of letting one run, then restarting it.
+	for i := 0; i < 5; i += 1 {
+		oneSibling.becomeHealthy()
+		oneSibling.waitState(rcRunnableStateHealthy)
+
+		// 'one' should work for at least a second.
+		deadline := time.Now().Add(1 * time.Second)
+		for {
+			if time.Now().After(deadline) {
+				break
+			}
+
+			oneTest()
+		}
+
+		// Killing 'oneSibling' should restart one.
+		oneSibling.panic()
+	}
+	// Make sure 'one' is still okay.
+	oneTest()
+}
+
+func ExampleNew() {
+	// Minimal runnable that is immediately done.
+	childC := make(chan struct{})
+	child := func(ctx context.Context) error {
+		Signal(ctx, SignalHealthy)
+		close(childC)
+		Signal(ctx, SignalDone)
+		return nil
+	}
+
+	// Start a supervision tree with a root runnable.
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+	New(ctx, func(ctx context.Context) error {
+		err := Run(ctx, "child", child)
+		if err != nil {
+			return fmt.Errorf("could not run 'child': %w", err)
+		}
+		Signal(ctx, SignalHealthy)
+
+		t := time.NewTicker(time.Second)
+		defer t.Stop()
+
+		// Do something in the background, and exit on context cancel.
+		for {
+			select {
+			case <-t.C:
+				fmt.Printf("tick!")
+			case <-ctx.Done():
+				return ctx.Err()
+			}
+		}
+	})
+
+	// root.child will close this channel.
+	<-childC
+}
diff --git a/metropolis/node/common/supervisor/supervisor_testhelpers.go b/metropolis/node/common/supervisor/supervisor_testhelpers.go
new file mode 100644
index 0000000..771e02f
--- /dev/null
+++ b/metropolis/node/common/supervisor/supervisor_testhelpers.go
@@ -0,0 +1,50 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package supervisor
+
+import (
+	"context"
+	"testing"
+)
+
+// waitSettle waits until the supervisor reaches a 'settled' state - ie., one
+// where no actions have been performed for a number of GC cycles.
+// This is used in tests only.
+func (s *supervisor) waitSettle(ctx context.Context) error {
+	waiter := make(chan struct{})
+	s.pReq <- &processorRequest{
+		waitSettled: &processorRequestWaitSettled{
+			waiter: waiter,
+		},
+	}
+
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-waiter:
+		return nil
+	}
+}
+
+// waitSettleError wraps waitSettle to fail a test if an error occurs, eg. the
+// context is canceled.
+func (s *supervisor) waitSettleError(ctx context.Context, t *testing.T) {
+	err := s.waitSettle(ctx)
+	if err != nil {
+		t.Fatalf("waitSettle: %v", err)
+	}
+}
diff --git a/metropolis/node/common/sysfs/BUILD.bazel b/metropolis/node/common/sysfs/BUILD.bazel
new file mode 100644
index 0000000..a4c7f18
--- /dev/null
+++ b/metropolis/node/common/sysfs/BUILD.bazel
@@ -0,0 +1,8 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["uevents.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/common/sysfs",
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/node/common/sysfs/uevents.go b/metropolis/node/common/sysfs/uevents.go
new file mode 100644
index 0000000..fed4319
--- /dev/null
+++ b/metropolis/node/common/sysfs/uevents.go
@@ -0,0 +1,50 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package sysfs
+
+import (
+	"bufio"
+	"io"
+	"os"
+	"strings"
+)
+
+func ReadUevents(filename string) (map[string]string, error) {
+	f, err := os.Open(filename)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	ueventMap := make(map[string]string)
+	reader := bufio.NewReader(f)
+	for {
+		name, err := reader.ReadString(byte('='))
+		if err == io.EOF {
+			break
+		} else if err != nil {
+			return nil, err
+		}
+		value, err := reader.ReadString(byte('\n'))
+		if err == io.EOF {
+			continue
+		} else if err != nil {
+			return nil, err
+		}
+		ueventMap[strings.Trim(name, "=")] = strings.TrimSpace(value)
+	}
+	return ueventMap, nil
+}
diff --git a/metropolis/node/core/BUILD.bazel b/metropolis/node/core/BUILD.bazel
new file mode 100644
index 0000000..2398205
--- /dev/null
+++ b/metropolis/node/core/BUILD.bazel
@@ -0,0 +1,43 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+
+go_library(
+    name = "go_default_library",
+    # keep
+    srcs = [
+        "debug_service.go",
+        "main.go",
+        "switchroot.go",
+    ] + select({
+        "//metropolis/node:debug_build": ["delve_enabled.go"],
+        "//conditions:default": ["delve_disabled.go"],
+    }),
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core",
+    visibility = ["//visibility:private"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/cluster:go_default_library",
+        "//metropolis/node/core/consensus/ca:go_default_library",
+        "//metropolis/node/core/localstorage:go_default_library",
+        "//metropolis/node/core/localstorage/declarative:go_default_library",
+        "//metropolis/node/core/logtree:go_default_library",
+        "//metropolis/node/core/network:go_default_library",
+        "//metropolis/node/core/network/dns:go_default_library",
+        "//metropolis/node/core/tpm:go_default_library",
+        "//metropolis/node/kubernetes:go_default_library",
+        "//metropolis/node/kubernetes/containerd:go_default_library",
+        "//metropolis/node/kubernetes/pki:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+        "@org_golang_google_grpc//:go_default_library",
+        "@org_golang_google_grpc//codes:go_default_library",
+        "@org_golang_google_grpc//status:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+go_binary(
+    name = "core",
+    embed = [":go_default_library"],
+    pure = "on",  # keep
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/node/core/cluster/BUILD.bazel b/metropolis/node/core/cluster/BUILD.bazel
new file mode 100644
index 0000000..70daba2
--- /dev/null
+++ b/metropolis/node/core/cluster/BUILD.bazel
@@ -0,0 +1,25 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "manager.go",
+        "node.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/cluster",
+    visibility = ["//metropolis/node/core:__subpackages__"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/consensus:go_default_library",
+        "//metropolis/node/core/localstorage:go_default_library",
+        "//metropolis/node/core/localstorage/declarative:go_default_library",
+        "//metropolis/node/core/network:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+        "//metropolis/proto/internal:go_default_library",
+        "@com_github_cenkalti_backoff_v4//:go_default_library",
+        "@com_github_golang_protobuf//proto:go_default_library",
+        "@io_etcd_go_etcd//clientv3:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/cluster/manager.go b/metropolis/node/core/cluster/manager.go
new file mode 100644
index 0000000..6bb87f4
--- /dev/null
+++ b/metropolis/node/core/cluster/manager.go
@@ -0,0 +1,555 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package cluster
+
+import (
+	"context"
+	"crypto/x509"
+	"encoding/pem"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/cenkalti/backoff/v4"
+	"github.com/golang/protobuf/proto"
+	"go.etcd.io/etcd/clientv3"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/consensus"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+// Manager is a finite state machine that joins this node (ie., Smalltown instance running on a virtual/physical machine)
+// into a Smalltown cluster (ie. group of nodes that act as a single control plane for Smalltown services). It does that
+// by bringing up all required operating-system level components, including mounting the local filesystem, bringing up
+// a consensus (etcd) server/client, ...
+//
+// The Manager runs as a single-shot Runnable. It will attempt to progress its state from the initial state (New) to
+// either Running (meaning that the node is now part of a cluster), or Failed (meaning that the node couldn't become
+// part of a cluster). It is not restartable, as it mutates quite a bit of implicit operating-system level state (like
+// filesystem mounts).  As such, it's difficult to recover reliably from failures, and since these failures indicate
+// some high issues with the cluster configuration/state, a failure requires a full kernel reboot to retry (or fix/
+// reconfigure the node).
+//
+// Currently, the Manager only supports one flow for bringing up a Node: by creating a new cluster. As such, it's
+// missing the following flows:
+//  - joining a new node into an already running cluster
+//  - restarting a node into an already existing cluster
+//  - restarting a node into an already running cluster (ie. full reboot of whole cluster)
+//
+type Manager struct {
+	storageRoot    *localstorage.Root
+	networkService *network.Service
+
+	// stateLock locks all state* variables.
+	stateLock sync.RWMutex
+	// state is the FSM state of the Manager.
+	state State
+	// stateRunningNode is the Node that this Manager got from joining a cluster. It's only valid if the Manager is
+	// Running.
+	stateRunningNode *Node
+	// stateWaiters is a list of channels that wish to be notified (by sending true or false) for when the Manager
+	// reaches a final state (Running or Failed respectively).
+	stateWaiters []chan bool
+
+	// goldenTicket is the Golden Ticket present in the enrolment config, if any.
+	goldenTicket *apb.GoldenTicket
+
+	// consensus is the spawned etcd/consensus service, if the Manager brought up a Node that should run one.
+	consensus *consensus.Service
+}
+
+// NewManager creates a new cluster Manager. The given localstorage Root must be places, but not yet started (and will
+// be started as the Manager makes progress). The given network Service must already be running.
+func NewManager(storageRoot *localstorage.Root, networkService *network.Service) *Manager {
+	return &Manager{
+		storageRoot:    storageRoot,
+		networkService: networkService,
+	}
+}
+
+// State is the state of the Manager finite state machine.
+type State int
+
+const (
+	// StateNew is the initial state of the Manager. It decides how to go about joining or creating a cluster.
+	StateNew State = iota
+	// StateCreatingCluster is when the Manager attempts to create a new cluster - this happens when a node is started
+	// with no EnrolmentConfig.
+	StateCreatingCluster
+	// StateCharlie is when the Manager uses the Golden Ticket debug/stopgap system to join an already
+	// existing cluster. This mechanism will be removed before the first Smalltown release.
+	StateCharlie
+	// StateRunning is when the Manager successfully got the node to be part of a cluster. stateRunningNode is valid.
+	StateRunning
+	// StateFailed is when the Manager failed to ge the node to be part of a cluster.
+	StateFailed
+)
+
+func (s State) String() string {
+	switch s {
+	case StateNew:
+		return "New"
+	case StateCreatingCluster:
+		return "CreatingCluster"
+	case StateCharlie:
+		return "Charlie"
+	case StateRunning:
+		return "Running"
+	case StateFailed:
+		return "Failed"
+	default:
+		return "UNKNOWN"
+	}
+}
+
+// allowedTransition describes all allowed state transitions (map[From][]To).
+var allowedTransitions = map[State][]State{
+	StateNew:             {StateCreatingCluster, StateCharlie},
+	StateCreatingCluster: {StateRunning, StateFailed},
+	StateCharlie:         {StateRunning, StateFailed},
+}
+
+// allowed returns whether a transition from a state to another state is allowed (ie. is defined in allowedTransitions).
+func (m *Manager) allowed(from, to State) bool {
+	for _, allowed := range allowedTransitions[from] {
+		if to == allowed {
+			return true
+		}
+	}
+	return false
+}
+
+// next moves the Manager finite state machine from its current state to `n`, or to Failed if the transition is not
+// allowed.
+func (m *Manager) next(ctx context.Context, n State) {
+	m.stateLock.Lock()
+	defer m.stateLock.Unlock()
+
+	if !m.allowed(m.state, n) {
+		supervisor.Logger(ctx).Errorf("Attempted invalid enrolment state transition, failing enrolment; from: %s, to: %s",
+			m.state.String(), n.String())
+		m.state = StateFailed
+		return
+	}
+
+	supervisor.Logger(ctx).Infof("Enrolment state change; from: %s, to: %s", m.state.String(), n.String())
+
+	m.state = n
+}
+
+// State returns the state of the Manager. It's safe to call this from any goroutine.
+func (m *Manager) State() State {
+	m.stateLock.RLock()
+	defer m.stateLock.RUnlock()
+	return m.state
+}
+
+// WaitFinished waits until the Manager FSM reaches Running or Failed, and returns true if the FSM is Running. It's
+// safe to call this from any goroutine.
+func (m *Manager) WaitFinished() (success bool) {
+	m.stateLock.Lock()
+	switch m.state {
+	case StateFailed:
+		m.stateLock.Unlock()
+		return false
+	case StateRunning:
+		m.stateLock.Unlock()
+		return true
+	}
+
+	C := make(chan bool)
+	m.stateWaiters = append(m.stateWaiters, C)
+	m.stateLock.Unlock()
+	return <-C
+}
+
+// wakeWaiters wakes any WaitFinished waiters and lets them know about the current state of the Manager.
+// The stateLock must already been taken, and the state must have been set in the same critical section (otherwise
+// this can cause a race condition).
+func (m *Manager) wakeWaiters() {
+	state := m.state
+	waiters := m.stateWaiters
+	m.stateWaiters = nil
+
+	for _, waiter := range waiters {
+		go func(w chan bool) {
+			w <- state == StateRunning
+		}(waiter)
+	}
+}
+
+// Run is the runnable of the Manager, to be started using the Supervisor. It is one-shot, and should not be restarted.
+func (m *Manager) Run(ctx context.Context) error {
+	if state := m.State(); state != StateNew {
+		supervisor.Logger(ctx).Errorf("Manager started with non-New state %s, failing", state.String())
+		m.stateLock.Lock()
+		m.state = StateFailed
+		m.wakeWaiters()
+		m.stateLock.Unlock()
+		return nil
+	}
+
+	var err error
+	bo := backoff.NewExponentialBackOff()
+	for {
+		done := false
+		state := m.State()
+		switch state {
+		case StateNew:
+			err = m.stateNew(ctx)
+		case StateCreatingCluster:
+			err = m.stateCreatingCluster(ctx)
+		case StateCharlie:
+			err = m.stateCharlie(ctx)
+		default:
+			done = true
+			break
+		}
+
+		if err != nil || done {
+			break
+		}
+
+		if state == m.State() && !m.allowed(state, m.State()) {
+			supervisor.Logger(ctx).Errorf("Enrolment got stuck at %s, failing", m.state.String())
+			m.stateLock.Lock()
+			m.state = StateFailed
+			m.stateLock.Unlock()
+		} else {
+			bo.Reset()
+		}
+	}
+
+	m.stateLock.Lock()
+	state := m.state
+	if state != StateRunning {
+		supervisor.Logger(ctx).Errorf("Enrolment failed at %s: %v", m.state.String(), err)
+	} else {
+		supervisor.Logger(ctx).Info("Enrolment successful!")
+	}
+	m.wakeWaiters()
+	m.stateLock.Unlock()
+
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	supervisor.Signal(ctx, supervisor.SignalDone)
+	return nil
+}
+
+// stateNew is called when a Manager is New. It makes the decision on how to join this node into a cluster.
+func (m *Manager) stateNew(ctx context.Context) error {
+	supervisor.Logger(ctx).Info("Starting enrolment process...")
+
+	// Check for presence of EnrolmentConfig on ESP or in qemu firmware variables.
+	var configRaw []byte
+	configRaw, err := m.storageRoot.ESP.Enrolment.Read()
+	if err != nil && !os.IsNotExist(err) {
+		return fmt.Errorf("could not read local enrolment file: %w", err)
+	} else if err != nil {
+		configRaw, err = ioutil.ReadFile("/sys/firmware/qemu_fw_cfg/by_name/com.nexantic.smalltown/enrolment.pb/raw")
+		if err != nil && !os.IsNotExist(err) {
+			return fmt.Errorf("could not read firmware enrolment file: %w", err)
+		}
+	}
+
+	// If no enrolment file exists, we create a new cluster.
+	if configRaw == nil {
+		m.next(ctx, StateCreatingCluster)
+		return nil
+	}
+
+	// Enrolment file exists, parse it.
+
+	enrolmentConfig := apb.EnrolmentConfig{}
+	if err := proto.Unmarshal(configRaw, &enrolmentConfig); err != nil {
+		return fmt.Errorf("could not unmarshal local enrolment file: %w", err)
+	}
+
+	// If no join ticket exists, we can't do anything yet.
+	if enrolmentConfig.GoldenTicket == nil {
+		return fmt.Errorf("joining a cluster without a golden ticket not yet implemented")
+	}
+
+	m.goldenTicket = enrolmentConfig.GoldenTicket
+
+	// Otherwise, we begin enrolling with the Golden Ticket.
+	m.next(ctx, StateCharlie)
+	return nil
+}
+
+// stateCreatingCluster is called when the Manager has decided to create a new cluster.
+//
+// The process to create a new cluster is as follows:
+//   - wait for IP address
+//   - initialize new data partition, by generating local and cluster unlock keys (the local unlock key is saved to
+//     the ESP, while the cluster unlock key is returned)
+//   - create a new node certificate and Node (with new given cluster unlock key)
+//   - start up a new etcd cluster, with this node being the only member
+//   - save the new Node to the new etcd cluster (thereby saving the node's cluster unlock key to etcd)
+func (m *Manager) stateCreatingCluster(ctx context.Context) error {
+	logger := supervisor.Logger(ctx)
+	logger.Info("Creating new cluster: waiting for IP address...")
+	ip, err := m.networkService.GetIP(ctx, true)
+	if err != nil {
+		return fmt.Errorf("when getting IP address: %w", err)
+	}
+	logger.Infof("Creating new cluster: got IP address %s", ip.String())
+
+	logger.Info("Creating new cluster: initializing storage...")
+	cuk, err := m.storageRoot.Data.MountNew(&m.storageRoot.ESP.LocalUnlock)
+	if err != nil {
+		return fmt.Errorf("when making new data partition: %w", err)
+	}
+	logger.Info("Creating new cluster: storage initialized")
+
+	// Create certificate for node.
+	cert, err := m.storageRoot.Data.Node.EnsureSelfSigned(localstorage.CertificateForNode)
+	if err != nil {
+		return fmt.Errorf("failed to create new node certificate: %w", err)
+	}
+
+	node := NewNode(cuk, *ip, *cert.Leaf)
+
+	m.consensus = consensus.New(consensus.Config{
+		Data:           &m.storageRoot.Data.Etcd,
+		Ephemeral:      &m.storageRoot.Ephemeral.Consensus,
+		NewCluster:     true,
+		Name:           node.ID(),
+		InitialCluster: ip.String(),
+		ExternalHost:   ip.String(),
+		ListenHost:     ip.String(),
+	})
+	if err := supervisor.Run(ctx, "consensus", m.consensus.Run); err != nil {
+		return fmt.Errorf("when starting consensus: %w", err)
+	}
+
+	// TODO(q3k): make timeout configurable?
+	ctxT, ctxC := context.WithTimeout(ctx, 5*time.Second)
+	defer ctxC()
+
+	supervisor.Logger(ctx).Info("Creating new cluster: waiting for consensus...")
+	if err := m.consensus.WaitReady(ctxT); err != nil {
+		return fmt.Errorf("consensus service failed to become ready: %w", err)
+	}
+
+	// Configure node to be a consensus member and kubernetes worker. In the future, different nodes will have
+	// different roles, but for now they're all symmetrical.
+	_, consensusName, err := m.consensus.MemberInfo(ctx)
+	if err != nil {
+		return fmt.Errorf("could not get consensus MemberInfo: %w", err)
+	}
+	if err := node.MakeConsensusMember(consensusName); err != nil {
+		return fmt.Errorf("could not make new node into consensus member: %w", err)
+	}
+	if err := node.MakeKubernetesWorker(node.ID()); err != nil {
+		return fmt.Errorf("could not make new node into kubernetes worker: %w", err)
+	}
+
+	// Save node into etcd.
+	supervisor.Logger(ctx).Info("Creating new cluster: storing first node...")
+	if err := node.Store(ctx, m.consensus.KV("cluster", "enrolment")); err != nil {
+		return fmt.Errorf("could not save new node: %w", err)
+	}
+
+	m.stateLock.Lock()
+	m.stateRunningNode = node
+	m.stateLock.Unlock()
+
+	m.next(ctx, StateRunning)
+	return nil
+}
+
+// stateCharlie is used to join an existing cluster via the GoldenTicket mechanism. This mechanism is temporarily
+// implemented in Smalltown in order to allow for testing multi-node clusters without a TPM attestation flow implemented.
+// The Golden Ticket contains a pregenerated node certificate, etcd certificate, and other data that any node can
+// use to join the cluster.
+// Since this flow is temporary, it has a slight impedance mismatch with methods exposed by localstorage, node, etc.,
+// and the resulting sequencing is a bit odd:
+//  - the {node,etcd} certificates/keys are loaded (this already dictates the new node name, as the node name is based
+//    off of the node public key)
+//  - local storage is initialized, a local/cluster unlock keypair is generated
+//  - etcd keys are manually saved to localstorage (vs. being generated locally by CA)
+//  - an etcd/consensus member is started, knowing that the remote member was already generated when the golden ticket
+//    was generated (vs. being created now by an RPC call, via an promote-node-to-etcd-member flow)
+//  - the node is then promoted to a consensus member and kubernetes worker, its clusterunlock key is set, and then it
+//    is saved to etcd.
+// As such, in this flow, we first create an etcd member (on goldenticket generation), and then only create a new Smalltown
+// node (when the goldenticket is used).
+func (m *Manager) stateCharlie(ctx context.Context) error {
+	t := m.goldenTicket
+	nodeCert, err := x509.ParseCertificate(t.NodeCert)
+	if err != nil {
+		return fmt.Errorf("parsing node certificate from ticket: %w", err)
+	}
+
+	supervisor.Logger(ctx).Info("Joining cluster: waiting for IP address...")
+	ip, err := m.networkService.GetIP(ctx, true)
+	if err != nil {
+		return fmt.Errorf("when getting IP address: %w", err)
+	}
+	supervisor.Logger(ctx).Info("Joining cluster: got IP address %s", ip.String())
+
+	supervisor.Logger(ctx).Info("Joining cluster: initializing storage...")
+	cuk, err := m.storageRoot.Data.MountNew(&m.storageRoot.ESP.LocalUnlock)
+	if err != nil {
+		return fmt.Errorf("when making new data partition: %w", err)
+	}
+	supervisor.Logger(ctx).Info("Joining cluster: storage initialized")
+	node := NewNode(cuk, *ip, *nodeCert)
+
+	// Save etcd PKI to disk.
+	for _, f := range []struct {
+		target    declarative.FilePlacement
+		data      []byte
+		blockType string
+	}{
+		{m.storageRoot.Data.Etcd.PeerPKI.Key, t.EtcdClientKey, "PRIVATE KEY"},
+		{m.storageRoot.Data.Etcd.PeerPKI.Certificate, t.EtcdClientCert, "CERTIFICATE"},
+		{m.storageRoot.Data.Etcd.PeerPKI.CACertificate, t.EtcdCaCert, "CERTIFICATE"},
+	} {
+		if err := f.target.Write(pem.EncodeToMemory(&pem.Block{Type: f.blockType, Bytes: f.data}), 0600); err != nil {
+			return fmt.Errorf("when writing etcd PKI data: %w", err)
+		}
+	}
+	if err := m.storageRoot.Data.Etcd.PeerCRL.Write(t.EtcdCrl, 0600); err != nil {
+		return fmt.Errorf("when writing etcd CRL: %w", err)
+	}
+
+	https := func(p *apb.GoldenTicket_EtcdPeer) string {
+		return fmt.Sprintf("%s=https://%s:%d", p.Name, p.Address, common.ConsensusPort)
+	}
+	var initialCluster []string
+	for _, p := range t.Peers {
+		initialCluster = append(initialCluster, https(p))
+	}
+	initialCluster = append(initialCluster, https(t.This))
+
+	supervisor.Logger(ctx).Infof("Joining cluster: starting etcd join, name: %s, initial_cluster: %s", node.ID(), strings.Join(initialCluster, ","))
+	m.consensus = consensus.New(consensus.Config{
+		Data:           &m.storageRoot.Data.Etcd,
+		Ephemeral:      &m.storageRoot.Ephemeral.Consensus,
+		Name:           node.ID(),
+		InitialCluster: strings.Join(initialCluster, ","),
+		ExternalHost:   ip.String(),
+		ListenHost:     ip.String(),
+	})
+
+	if err := supervisor.Run(ctx, "consensus", m.consensus.Run); err != nil {
+		return fmt.Errorf("when starting consensus: %w", err)
+	}
+
+	// TODO(q3k): make timeout configurable?
+	ctxT, ctxC := context.WithTimeout(ctx, 5*time.Second)
+	defer ctxC()
+
+	supervisor.Logger(ctx).Info("Joining cluster: waiting for consensus...")
+	if err := m.consensus.WaitReady(ctxT); err != nil {
+		return fmt.Errorf("consensus service failed to become ready: %w", err)
+	}
+
+	// Configure node to be a consensus member and kubernetes worker. In the future, different nodes will have
+	// different roles, but for now they're all symmetrical.
+	_, consensusName, err := m.consensus.MemberInfo(ctx)
+	if err != nil {
+		return fmt.Errorf("could not get consensus MemberInfo: %w", err)
+	}
+	if err := node.MakeConsensusMember(consensusName); err != nil {
+		return fmt.Errorf("could not make new node into consensus member: %w", err)
+	}
+	if err := node.MakeKubernetesWorker(node.ID()); err != nil {
+		return fmt.Errorf("could not make new node into kubernetes worker: %w", err)
+	}
+
+	// Save node into etcd.
+	supervisor.Logger(ctx).Info("Creating new cluster: storing first node...")
+	if err := node.Store(ctx, m.consensus.KV("cluster", "enrolment")); err != nil {
+		return fmt.Errorf("could not save new node: %w", err)
+	}
+
+	m.stateLock.Lock()
+	m.stateRunningNode = node
+	m.stateLock.Unlock()
+
+	m.next(ctx, StateRunning)
+	return nil
+}
+
+// Node returns the Node that the Manager brought into a cluster, or nil if the Manager is not Running.
+// This is safe to call from any goroutine.
+func (m *Manager) Node() *Node {
+	m.stateLock.Lock()
+	defer m.stateLock.Unlock()
+	if m.state != StateRunning {
+		return nil
+	}
+	return m.stateRunningNode
+}
+
+// ConsensusKV returns a namespaced etcd KV client, or nil if the Manager is not Running.
+// This is safe to call from any goroutine.
+func (m *Manager) ConsensusKV(module, space string) clientv3.KV {
+	m.stateLock.Lock()
+	defer m.stateLock.Unlock()
+	if m.state != StateRunning {
+		return nil
+	}
+	if m.stateRunningNode.ConsensusMember() == nil {
+		// TODO(q3k): in this case, we should return a client to etcd even though this
+		// node is not a member of consensus. For now, all nodes are consensus members.
+		return nil
+	}
+	return m.consensus.KV(module, space)
+}
+
+// ConsensusKVRoot returns a non-namespaced etcd KV client, or nil if the Manager is not Running.
+// This is safe to call from any goroutine.
+func (m *Manager) ConsensusKVRoot() clientv3.KV {
+	m.stateLock.Lock()
+	defer m.stateLock.Unlock()
+	if m.state != StateRunning {
+		return nil
+	}
+	if m.stateRunningNode.ConsensusMember() == nil {
+		// TODO(q3k): in this case, we should return a client to etcd even though this
+		// node is not a member of consensus. For now, all nodes are consensus members.
+		return nil
+	}
+	return m.consensus.KVRoot()
+}
+
+// ConsensusCluster returns an etcd Cluster client, or nil if the Manager is not Running.
+// This is safe to call from any goroutine.
+func (m *Manager) ConsensusCluster() clientv3.Cluster {
+	m.stateLock.Lock()
+	defer m.stateLock.Unlock()
+	if m.state != StateRunning {
+		return nil
+	}
+	if m.stateRunningNode.ConsensusMember() == nil {
+		// TODO(q3k): in this case, we should return a client to etcd even though this
+		// node is not a member of consensus. For now, all nodes are consensus members.
+		return nil
+	}
+	return m.consensus.Cluster()
+}
diff --git a/metropolis/node/core/cluster/node.go b/metropolis/node/core/cluster/node.go
new file mode 100644
index 0000000..449c2ff
--- /dev/null
+++ b/metropolis/node/core/cluster/node.go
@@ -0,0 +1,219 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package cluster
+
+import (
+	"context"
+	"crypto/ed25519"
+	"crypto/x509"
+	"encoding/hex"
+	"fmt"
+	"net"
+
+	"github.com/golang/protobuf/proto"
+	"go.etcd.io/etcd/clientv3"
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	ipb "git.monogon.dev/source/nexantic.git/metropolis/proto/internal"
+)
+
+// Node is a Smalltown cluster member. A node is a virtual or physical machine running Smalltown. This object represents a
+// node only as part of a Cluster - ie., this object will never be available outside of //metropolis/node/core/cluster
+// if the Node is not part of a Cluster.
+// Nodes are inherently tied to their long term storage, which is etcd. As such, methods on this object relate heavily
+// to the Node's expected lifecycle on etcd.
+type Node struct {
+	// clusterUnlockKey is half of the unlock key required to mount the node's data partition. It's stored in etcd, and
+	// will only be provided to the Node if it can prove its identity via an integrity mechanism (ie. via TPM), or when
+	// the Node was just created (as the key is generated locally by localstorage on first format/mount).
+	// The other part of the unlock key is the LocalUnlockKey that's present on the node's ESP partition.
+	clusterUnlockKey []byte
+	// certificate is the node's TLS certificate, used to authenticate Smalltown gRPC calls/services (but not
+	// consensus/etcd). The certificate for a node is permanent (and never expires). It's self-signed by the node on
+	// startup, and contains the node's IP address in its SAN. Callers/services should check directly against the
+	// expected certificate, and not against a CA.
+	certificate x509.Certificate
+	// address is the management IP address of the node. The management IP address of a node is permanent.
+	address net.IP
+
+	// A Node can have multiple Roles. Each Role is represented by the presence of NodeRole* structures in this
+	// structure, with a nil pointer representing the lack of a role.
+
+	consensusMember  *NodeRoleConsensusMember
+	kubernetesWorker *NodeRoleKubernetesWorker
+}
+
+// NewNode creates a new Node. This is only called when a New node is supposed to be created as part of a cluster,
+// otherwise it should be loaded from Etcd.
+func NewNode(cuk []byte, address net.IP, certificate x509.Certificate) *Node {
+	if certificate.Raw == nil {
+		panic("new node must contain raw certificate")
+	}
+	return &Node{
+		clusterUnlockKey: cuk,
+		certificate:      certificate,
+		address:          address,
+	}
+}
+
+// NodeRoleConsensusMember defines that the Node is a consensus (etcd) cluster member.
+type NodeRoleConsensusMember struct {
+	// etcdMember is the name of the node in Kubernetes. This is for now usually the same as the ID() of the Node.
+	etcdMemberName string
+}
+
+// NodeRoleKubernetesWorker defines that the Node should be running the Kubernetes control and data plane.
+type NodeRoleKubernetesWorker struct {
+	// nodeName is the name of the node in Kubernetes. This is for now usually the same as the ID() of the Node.
+	nodeName string
+}
+
+// ID returns the name of this node, which is `smalltown-{pubkeyHash}`. This name should be the primary way to refer to
+// Smalltown nodes within a cluster, and is guaranteed to be unique by relying on cryptographic randomness.
+func (n *Node) ID() string {
+	return fmt.Sprintf("smalltown-%s", n.IDBare())
+}
+
+// IDBare returns the `{pubkeyHash}` part of the node ID.
+func (n Node) IDBare() string {
+	pubKey, ok := n.certificate.PublicKey.(ed25519.PublicKey)
+	if !ok {
+		panic("node has non-ed25519 public key")
+	}
+	return hex.EncodeToString(pubKey[:16])
+}
+
+func (n *Node) String() string {
+	return n.ID()
+}
+
+// ConsensusMember returns a copy of the NodeRoleConsensusMember struct if the Node is a consensus member, otherwise
+// nil.
+func (n *Node) ConsensusMember() *NodeRoleConsensusMember {
+	if n.consensusMember == nil {
+		return nil
+	}
+	cm := *n.consensusMember
+	return &cm
+}
+
+// KubernetesWorker returns a copy of the NodeRoleKubernetesWorker struct if the Node is a kubernetes worker, otherwise
+// nil.
+func (n *Node) KubernetesWorker() *NodeRoleKubernetesWorker {
+	if n.kubernetesWorker == nil {
+		return nil
+	}
+	kw := *n.kubernetesWorker
+	return &kw
+}
+
+// etcdPath builds the etcd path in which this node's protobuf-serialized state is stored in etcd.
+func (n *Node) etcdPath() string {
+	return fmt.Sprintf("/nodes/%s", n.ID())
+}
+
+// proto serializes the Node object into protobuf, to be used for saving to etcd.
+func (n *Node) proto() *ipb.Node {
+	msg := &ipb.Node{
+		Certificate:      n.certificate.Raw,
+		ClusterUnlockKey: n.clusterUnlockKey,
+		Address:          n.address.String(),
+		Roles:            &ipb.Node_Roles{},
+	}
+	if n.consensusMember != nil {
+		msg.Roles.ConsensusMember = &ipb.Node_Roles_ConsensusMember{
+			EtcdMemberName: n.consensusMember.etcdMemberName,
+		}
+	}
+	if n.kubernetesWorker != nil {
+		msg.Roles.KubernetesWorker = &ipb.Node_Roles_KubernetesWorker{
+			NodeName: n.kubernetesWorker.nodeName,
+		}
+	}
+	return msg
+}
+
+// Store saves the Node into etcd. This should be called only once per Node (ie. when the Node has been created).
+func (n *Node) Store(ctx context.Context, kv clientv3.KV) error {
+	// Currently the only flow to store a node to etcd is a write-once flow: once a node is created, it cannot be
+	// deleted or updated. In the future, flows to change cluster node roles might be introduced (ie. to promote nodes
+	// to consensus members, etc).
+	key := n.etcdPath()
+	msg := n.proto()
+	nodeRaw, err := proto.Marshal(msg)
+	if err != nil {
+		return fmt.Errorf("failed to marshal node: %w", err)
+	}
+
+	res, err := kv.Txn(ctx).If(
+		clientv3.Compare(clientv3.CreateRevision(key), "=", 0),
+	).Then(
+		clientv3.OpPut(key, string(nodeRaw)),
+	).Commit()
+	if err != nil {
+		return fmt.Errorf("failed to store node: %w", err)
+	}
+
+	if !res.Succeeded {
+		return fmt.Errorf("attempted to re-register node (unsupported flow)")
+	}
+	return nil
+}
+
+// MakeConsensusMember turns the node into a consensus member with a given name. This only configures internal fields,
+// and does not actually start any services.
+func (n *Node) MakeConsensusMember(etcdMemberName string) error {
+	if n.consensusMember != nil {
+		return fmt.Errorf("node already is consensus member")
+	}
+	n.consensusMember = &NodeRoleConsensusMember{
+		etcdMemberName: etcdMemberName,
+	}
+	return nil
+}
+
+// MakeKubernetesWorker turns the node into a kubernetes worker with a given name. This only configures internal fields,
+// and does not actually start any services.
+func (n *Node) MakeKubernetesWorker(name string) error {
+	if n.kubernetesWorker != nil {
+		return fmt.Errorf("node is already kubernetes worker")
+	}
+	n.kubernetesWorker = &NodeRoleKubernetesWorker{
+		nodeName: name,
+	}
+	return nil
+}
+
+func (n *Node) Address() net.IP {
+	return n.address
+}
+
+// ConfigureLocalHostname uses the node's ID as a hostname, and sets the current hostname, and local files like hosts
+// and machine-id accordingly.
+func (n *Node) ConfigureLocalHostname(etc *localstorage.EtcDirectory) error {
+	if err := unix.Sethostname([]byte(n.ID())); err != nil {
+		return fmt.Errorf("failed to set runtime hostname: %w", err)
+	}
+	if err := etc.Hosts.Write([]byte(fmt.Sprintf("%s %s", "127.0.0.1", n.ID())), 0644); err != nil {
+		return fmt.Errorf("failed to write /etc/hosts: %w", err)
+	}
+	if err := etc.MachineID.Write([]byte(n.IDBare()), 0644); err != nil {
+		return fmt.Errorf("failed to write /etc/machine-id: %w", err)
+	}
+	return nil
+}
diff --git a/metropolis/node/core/consensus/BUILD.bazel b/metropolis/node/core/consensus/BUILD.bazel
new file mode 100644
index 0000000..cab2c0a
--- /dev/null
+++ b/metropolis/node/core/consensus/BUILD.bazel
@@ -0,0 +1,30 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["consensus.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/consensus",
+    visibility = ["//:__subpackages__"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/consensus/ca:go_default_library",
+        "//metropolis/node/core/localstorage:go_default_library",
+        "@io_etcd_go_etcd//clientv3:go_default_library",
+        "@io_etcd_go_etcd//clientv3/namespace:go_default_library",
+        "@io_etcd_go_etcd//embed:go_default_library",
+        "@org_uber_go_atomic//:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["consensus_test.go"],
+    embed = [":go_default_library"],
+    deps = [
+        "//golibs/common:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/localstorage:go_default_library",
+        "//metropolis/node/core/localstorage/declarative:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/consensus/ca/BUILD.bazel b/metropolis/node/core/consensus/ca/BUILD.bazel
new file mode 100644
index 0000000..fecffa0
--- /dev/null
+++ b/metropolis/node/core/consensus/ca/BUILD.bazel
@@ -0,0 +1,9 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["ca.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/consensus/ca",
+    visibility = ["//:__subpackages__"],
+    deps = ["@io_etcd_go_etcd//clientv3:go_default_library"],
+)
diff --git a/metropolis/node/core/consensus/ca/ca.go b/metropolis/node/core/consensus/ca/ca.go
new file mode 100644
index 0000000..9a1b634
--- /dev/null
+++ b/metropolis/node/core/consensus/ca/ca.go
@@ -0,0 +1,440 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// package ca implements a simple standards-compliant certificate authority.
+// It only supports ed25519 keys, and does not maintain any persistent state.
+//
+// The CA is backed by etcd storage, and can also bootstrap itself without a yet running etcd storage (and commit
+// in-memory secrets to etcd at a later date).
+//
+// CA and certificates successfully pass https://github.com/zmap/zlint
+// (minus the CA/B rules that a public CA would adhere to, which requires
+// things like OCSP servers, Certificate Policies and ECDSA/RSA-only keys).
+package ca
+
+// TODO(leo): add zlint test
+
+import (
+	"context"
+	"crypto"
+	"crypto/ed25519"
+	"crypto/rand"
+	"crypto/sha1"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/asn1"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"math/big"
+	"net"
+	"time"
+
+	"go.etcd.io/etcd/clientv3"
+)
+
+const (
+	// TODO(q3k): move this to a declarative storage layer
+	pathCACertificate      = "/etcd-ca/ca.der"
+	pathCAKey              = "/etcd-ca/ca-key.der"
+	pathCACRL              = "/etcd-ca/crl.der"
+	pathIssuedCertificates = "/etcd-ca/certs/"
+)
+
+func pathIssuedCertificate(serial *big.Int) string {
+	return pathIssuedCertificates + hex.EncodeToString(serial.Bytes())
+}
+
+var (
+	// From RFC 5280 Section 4.1.2.5
+	unknownNotAfter = time.Unix(253402300799, 0)
+)
+
+type CA struct {
+	// TODO: Potentially protect the key with memguard
+	privateKey *ed25519.PrivateKey
+	CACert     *x509.Certificate
+	CACertRaw  []byte
+
+	// bootstrapIssued are certificates that have been issued by the CA before it has been successfully Saved to etcd.
+	bootstrapIssued [][]byte
+	// canBootstrapIssue is set on CAs that have been created by New and not yet stored to etcd. If not set,
+	// certificates cannot be issued in-memory.
+	canBootstrapIssue bool
+}
+
+// Workaround for https://github.com/golang/go/issues/26676 in Go's crypto/x509. Specifically Go
+// violates Section 4.2.1.2 of RFC 5280 without this.
+// Fixed for 1.15 in https://go-review.googlesource.com/c/go/+/227098/.
+//
+// Taken from https://github.com/FiloSottile/mkcert/blob/master/cert.go#L295 written by one of Go's
+// crypto engineers (BSD 3-clause).
+func calculateSKID(pubKey crypto.PublicKey) ([]byte, error) {
+	spkiASN1, err := x509.MarshalPKIXPublicKey(pubKey)
+	if err != nil {
+		return nil, err
+	}
+
+	var spki struct {
+		Algorithm        pkix.AlgorithmIdentifier
+		SubjectPublicKey asn1.BitString
+	}
+	_, err = asn1.Unmarshal(spkiASN1, &spki)
+	if err != nil {
+		return nil, err
+	}
+	skid := sha1.Sum(spki.SubjectPublicKey.Bytes)
+	return skid[:], nil
+}
+
+// New creates a new certificate authority with the given common name. The newly created CA will be stored in memory
+// until committed to etcd by calling .Save.
+func New(name string) (*CA, error) {
+	pubKey, privKey, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		panic(err)
+	}
+
+	serialNumberLimit := new(big.Int).Lsh(big.NewInt(1), 127)
+	serialNumber, err := rand.Int(rand.Reader, serialNumberLimit)
+	if err != nil {
+		return nil, fmt.Errorf("failed to generate serial number: %w", err)
+	}
+
+	skid, err := calculateSKID(pubKey)
+	if err != nil {
+		return nil, err
+	}
+
+	caCert := &x509.Certificate{
+		SerialNumber: serialNumber,
+		Subject: pkix.Name{
+			CommonName: name,
+		},
+		IsCA:                  true,
+		BasicConstraintsValid: true,
+		NotBefore:             time.Now(),
+		NotAfter:              unknownNotAfter,
+		KeyUsage:              x509.KeyUsageCertSign | x509.KeyUsageCRLSign | x509.KeyUsageDigitalSignature,
+		ExtKeyUsage:           []x509.ExtKeyUsage{x509.ExtKeyUsageClientAuth, x509.ExtKeyUsageServerAuth, x509.ExtKeyUsageOCSPSigning},
+		AuthorityKeyId:        skid,
+		SubjectKeyId:          skid,
+	}
+
+	caCertRaw, err := x509.CreateCertificate(rand.Reader, caCert, caCert, pubKey, privKey)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create root certificate: %w", err)
+	}
+
+	ca := &CA{
+		privateKey: &privKey,
+		CACertRaw:  caCertRaw,
+		CACert:     caCert,
+
+		canBootstrapIssue: true,
+	}
+
+	return ca, nil
+}
+
+// Load restores CA state from etcd.
+func Load(ctx context.Context, kv clientv3.KV) (*CA, error) {
+	resp, err := kv.Txn(ctx).Then(
+		clientv3.OpGet(pathCACertificate),
+		clientv3.OpGet(pathCAKey),
+		// We only read the CRL to ensure it exists on etcd (and early fail on inconsistency)
+		clientv3.OpGet(pathCACRL)).Commit()
+	if err != nil {
+		return nil, fmt.Errorf("failed to retrieve CA from etcd: %w", err)
+	}
+
+	var caCert, caKey, caCRL []byte
+	for _, el := range resp.Responses {
+		for _, kv := range el.GetResponseRange().GetKvs() {
+			switch string(kv.Key) {
+			case pathCACertificate:
+				caCert = kv.Value
+			case pathCAKey:
+				caKey = kv.Value
+			case pathCACRL:
+				caCRL = kv.Value
+			}
+		}
+	}
+	if caCert == nil || caKey == nil || caCRL == nil {
+		return nil, fmt.Errorf("failed to retrieve CA from etcd, missing at least one of {ca key, ca crt, ca crl}")
+	}
+
+	if len(caKey) != ed25519.PrivateKeySize {
+		return nil, errors.New("invalid CA private key size")
+	}
+	privateKey := ed25519.PrivateKey(caKey)
+
+	caCertVal, err := x509.ParseCertificate(caCert)
+	if err != nil {
+		return nil, fmt.Errorf("failed to parse CA certificate: %w", err)
+	}
+	return &CA{
+		privateKey: &privateKey,
+		CACertRaw:  caCert,
+		CACert:     caCertVal,
+	}, nil
+}
+
+// Save stores a newly created CA into etcd, committing both the CA data and any certificates issued until then.
+func (c *CA) Save(ctx context.Context, kv clientv3.KV) error {
+	crl, err := c.makeCRL(nil)
+	if err != nil {
+		return fmt.Errorf("failed to generate initial CRL: %w", err)
+	}
+
+	ops := []clientv3.Op{
+		clientv3.OpPut(pathCACertificate, string(c.CACertRaw)),
+		clientv3.OpPut(pathCAKey, string([]byte(*c.privateKey))),
+		clientv3.OpPut(pathCACRL, string(crl)),
+	}
+	for i, certRaw := range c.bootstrapIssued {
+		cert, err := x509.ParseCertificate(certRaw)
+		if err != nil {
+			return fmt.Errorf("failed to parse in-memory certificate %d", i)
+		}
+		ops = append(ops, clientv3.OpPut(pathIssuedCertificate(cert.SerialNumber), string(certRaw)))
+	}
+
+	res, err := kv.Txn(ctx).If(
+		clientv3.Compare(clientv3.CreateRevision(pathCAKey), "=", 0),
+	).Then(ops...).Commit()
+	if err != nil {
+		return fmt.Errorf("failed to store CA to etcd: %w", err)
+	}
+	if !res.Succeeded {
+		// This should pretty much never happen, but we want to catch it just in case.
+		return fmt.Errorf("failed to store CA to etcd: CA already present - cluster-level data inconsistency")
+	}
+	c.bootstrapIssued = nil
+	c.canBootstrapIssue = false
+	return nil
+}
+
+// Issue issues a certificate. If kv is non-nil, the newly issued certificate will be immediately stored to etcd,
+// otherwise it will be kept in memory (until .Save is called). Certificates can only be issued to memory on
+// newly-created CAs that have not been saved to etcd yet.
+func (c *CA) Issue(ctx context.Context, kv clientv3.KV, commonName string, externalAddress net.IP) (cert []byte, privkey []byte, err error) {
+	serialNumberLimit := new(big.Int).Lsh(big.NewInt(1), 127)
+	serialNumber, err := rand.Int(rand.Reader, serialNumberLimit)
+	if err != nil {
+		err = fmt.Errorf("failed to generate serial number: %w", err)
+		return
+	}
+
+	pubKey, privKeyRaw, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		return
+	}
+	privkey, err = x509.MarshalPKCS8PrivateKey(privKeyRaw)
+	if err != nil {
+		return
+	}
+
+	etcdCert := &x509.Certificate{
+		SerialNumber: serialNumber,
+		Subject: pkix.Name{
+			CommonName:         commonName,
+			OrganizationalUnit: []string{"etcd"},
+		},
+		IsCA:                  false,
+		BasicConstraintsValid: true,
+		NotBefore:             time.Now(),
+		NotAfter:              unknownNotAfter,
+		ExtKeyUsage:           []x509.ExtKeyUsage{x509.ExtKeyUsageClientAuth, x509.ExtKeyUsageServerAuth},
+		DNSNames:              []string{commonName},
+		IPAddresses:           []net.IP{externalAddress},
+	}
+	cert, err = x509.CreateCertificate(rand.Reader, etcdCert, c.CACert, pubKey, c.privateKey)
+	if err != nil {
+		err = fmt.Errorf("failed to sign new certificate: %w", err)
+		return
+	}
+
+	if kv != nil {
+		path := pathIssuedCertificate(serialNumber)
+		_, err = kv.Put(ctx, path, string(cert))
+		if err != nil {
+			err = fmt.Errorf("failed to commit new certificate to etcd: %w", err)
+			return
+		}
+	} else {
+		if !c.canBootstrapIssue {
+			err = fmt.Errorf("cannot issue new certificate to memory on existing, etcd-backed CA")
+			return
+		}
+		c.bootstrapIssued = append(c.bootstrapIssued, cert)
+	}
+	return
+}
+
+func (c *CA) makeCRL(revoked []pkix.RevokedCertificate) ([]byte, error) {
+	crl, err := c.CACert.CreateCRL(rand.Reader, c.privateKey, revoked, time.Now(), unknownNotAfter)
+	if err != nil {
+		return nil, fmt.Errorf("failed to generate CRL: %w", err)
+	}
+	return crl, nil
+}
+
+// Revoke revokes a certificate by hostname. The selected hostname will be added to the CRL stored in etcd. This call
+// might fail (safely) if a simultaneous revoke happened that caused the CRL to be bumped. The call can be then retried
+// safely.
+func (c *CA) Revoke(ctx context.Context, kv clientv3.KV, hostname string) error {
+	res, err := kv.Txn(ctx).Then(
+		clientv3.OpGet(pathCACRL),
+		clientv3.OpGet(pathIssuedCertificates, clientv3.WithPrefix())).Commit()
+	if err != nil {
+		return fmt.Errorf("failed to retrieve certificates and CRL from etcd: %w", err)
+	}
+
+	var certs []*x509.Certificate
+	var crlRevision int64
+	var crl *pkix.CertificateList
+	for _, el := range res.Responses {
+		for _, kv := range el.GetResponseRange().GetKvs() {
+			if string(kv.Key) == pathCACRL {
+				crl, err = x509.ParseCRL(kv.Value)
+				if err != nil {
+					return fmt.Errorf("could not parse CRL from etcd: %w", err)
+				}
+				crlRevision = kv.CreateRevision
+			} else {
+				cert, err := x509.ParseCertificate(kv.Value)
+				if err != nil {
+					return fmt.Errorf("could not parse certificate %q from etcd: %w", string(kv.Key), err)
+				}
+				certs = append(certs, cert)
+			}
+		}
+	}
+
+	if crl == nil {
+		return fmt.Errorf("could not find CRL in etcd")
+	}
+	revoked := crl.TBSCertList.RevokedCertificates
+
+	// Find requested hostname in issued certificates.
+	var serial *big.Int
+	for _, cert := range certs {
+		for _, dnsName := range cert.DNSNames {
+			if dnsName == hostname {
+				serial = cert.SerialNumber
+				break
+			}
+		}
+		if serial != nil {
+			break
+		}
+	}
+	if serial == nil {
+		return fmt.Errorf("could not find requested hostname")
+	}
+
+	// Check if certificate has already been revoked.
+	for _, revokedCert := range revoked {
+		if revokedCert.SerialNumber.Cmp(serial) == 0 {
+			return nil // Already revoked
+		}
+	}
+
+	revoked = append(revoked, pkix.RevokedCertificate{
+		SerialNumber:   serial,
+		RevocationTime: time.Now(),
+	})
+
+	crlRaw, err := c.makeCRL(revoked)
+	if err != nil {
+		return fmt.Errorf("when generating new CRL for revocation: %w", err)
+	}
+
+	res, err = kv.Txn(ctx).If(
+		clientv3.Compare(clientv3.CreateRevision(pathCACRL), "=", crlRevision),
+	).Then(
+		clientv3.OpPut(pathCACRL, string(crlRaw)),
+	).Commit()
+	if err != nil {
+		return fmt.Errorf("when saving new CRL: %w", err)
+	}
+	if !res.Succeeded {
+		return fmt.Errorf("CRL save transaction failed, retry possibly")
+	}
+
+	return nil
+}
+
+// WaitCRLChange returns a channel that will receive a CRLUpdate any time the remote CRL changed. Immediately after
+// calling this method, the current CRL is retrieved from the cluster and put into the channel.
+func (c *CA) WaitCRLChange(ctx context.Context, kv clientv3.KV, w clientv3.Watcher) <-chan CRLUpdate {
+	C := make(chan CRLUpdate)
+
+	go func(ctx context.Context) {
+		ctxC, cancel := context.WithCancel(ctx)
+		defer cancel()
+
+		fail := func(f string, args ...interface{}) {
+			C <- CRLUpdate{Err: fmt.Errorf(f, args...)}
+			close(C)
+		}
+
+		initial, err := kv.Get(ctx, pathCACRL)
+		if err != nil {
+			fail("failed to retrieve initial CRL: %w", err)
+			return
+		}
+
+		C <- CRLUpdate{CRL: initial.Kvs[0].Value}
+
+		for wr := range w.Watch(ctxC, pathCACRL, clientv3.WithRev(initial.Kvs[0].CreateRevision)) {
+			if wr.Err() != nil {
+				fail("failed watching CRL: %w", wr.Err())
+				return
+			}
+
+			for _, e := range wr.Events {
+				if string(e.Kv.Key) != pathCACRL {
+					continue
+				}
+
+				C <- CRLUpdate{CRL: e.Kv.Value}
+			}
+		}
+	}(ctx)
+
+	return C
+}
+
+// CRLUpdate is emitted for every remote CRL change, and spuriously on ever new WaitCRLChange.
+type CRLUpdate struct {
+	// The new (or existing, in the case of the first call) CRL. If nil, Err will be set.
+	CRL []byte
+	// If set, an error occurred and the WaitCRLChange call must be restarted. If set, CRL will be nil.
+	Err error
+}
+
+// GetCurrentCRL returns the current CRL for the CA. This should only be used for one-shot operations like
+// bootstrapping a new node that doesn't yet have access to etcd - otherwise, WaitCRLChange shoulde be used.
+func (c *CA) GetCurrentCRL(ctx context.Context, kv clientv3.KV) ([]byte, error) {
+	initial, err := kv.Get(ctx, pathCACRL)
+	if err != nil {
+		return nil, fmt.Errorf("failed to retrieve initial CRL: %w", err)
+	}
+	return initial.Kvs[0].Value, nil
+}
diff --git a/metropolis/node/core/consensus/consensus.go b/metropolis/node/core/consensus/consensus.go
new file mode 100644
index 0000000..8916164
--- /dev/null
+++ b/metropolis/node/core/consensus/consensus.go
@@ -0,0 +1,429 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package consensus implements a managed etcd cluster member service, with a self-hosted CA system for issuing peer
+// certificates. Currently each Smalltown node runs an etcd member, and connects to the etcd member locally over a unix
+// domain socket.
+//
+// The service supports two modes of startup:
+//  - initializing a new cluster, by bootstrapping the CA in memory, starting a cluster, committing the CA to etcd
+//    afterwards, and saving the new node's certificate to local storage
+//  - joining an existing cluster, using certificates from local storage and loading the CA from etcd. This flow is also
+//    used when the node joins a cluster for the first time (then the certificates required must be provisioned
+//    externally before starting the consensus service).
+//
+// Regardless of how the etcd member service was started, the resulting running service is further managed and used
+// in the same way.
+//
+package consensus
+
+import (
+	"context"
+	"encoding/pem"
+	"fmt"
+	"net"
+	"net/url"
+	"sync"
+	"time"
+
+	"go.etcd.io/etcd/clientv3"
+	"go.etcd.io/etcd/clientv3/namespace"
+	"go.etcd.io/etcd/embed"
+	"go.uber.org/atomic"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/consensus/ca"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+)
+
+const (
+	DefaultClusterToken = "SIGNOS"
+	DefaultLogger       = "zap"
+)
+
+// Service is the etcd cluster member service.
+type Service struct {
+	// The configuration with which the service was started. This is immutable.
+	config *Config
+
+	// stateMu guards state. This is locked internally on public methods of Service that require access to state. The
+	// state might be recreated on service restart.
+	stateMu sync.Mutex
+	state   *state
+}
+
+// state is the runtime state of a running etcd member.
+type state struct {
+	etcd  *embed.Etcd
+	ready atomic.Bool
+
+	ca *ca.CA
+	// cl is an etcd client that loops back to the localy running etcd server. This runs over the Client unix domain
+	// socket that etcd starts.
+	cl *clientv3.Client
+}
+
+type Config struct {
+	// Data directory (persistent, encrypted storage) for etcd.
+	Data *localstorage.DataEtcdDirectory
+	// Ephemeral directory for etcd.
+	Ephemeral *localstorage.EphemeralConsensusDirectory
+
+	// Name is the cluster name. This must be the same amongst all etcd members within one cluster.
+	Name string
+	// NewCluster selects whether the etcd member will start a new cluster and bootstrap a CA and the first member
+	// certificate, or load existing PKI certificates from disk.
+	NewCluster bool
+	// InitialCluster sets the initial cluster peer URLs when NewCluster is set, and is ignored otherwise. Usually this
+	// will be just the new, single server, and more members will be added later.
+	InitialCluster string
+	// ExternalHost is the IP address or hostname at which this cluster member is reachable to other cluster members.
+	ExternalHost string
+	// ListenHost is the IP address or hostname at which this cluster member will listen.
+	ListenHost string
+	// Port is the port at which this cluster member will listen for other members. If zero, defaults to the global
+	// Smalltown setting.
+	Port int
+}
+
+func New(config Config) *Service {
+	return &Service{
+		config: &config,
+	}
+}
+
+// configure transforms the service configuration into an embedded etcd configuration. This is pure and side effect
+// free.
+func (s *Service) configure(ctx context.Context) (*embed.Config, error) {
+	if err := s.config.Ephemeral.MkdirAll(0700); err != nil {
+		return nil, fmt.Errorf("failed to create ephemeral directory: %w", err)
+	}
+	if err := s.config.Data.MkdirAll(0700); err != nil {
+		return nil, fmt.Errorf("failed to create data directory: %w", err)
+	}
+
+	port := s.config.Port
+	if port == 0 {
+		port = common.ConsensusPort
+	}
+
+	cfg := embed.NewConfig()
+
+	cfg.Name = s.config.Name
+	cfg.Dir = s.config.Data.Data.FullPath()
+	cfg.InitialClusterToken = DefaultClusterToken
+
+	cfg.PeerTLSInfo.CertFile = s.config.Data.PeerPKI.Certificate.FullPath()
+	cfg.PeerTLSInfo.KeyFile = s.config.Data.PeerPKI.Key.FullPath()
+	cfg.PeerTLSInfo.TrustedCAFile = s.config.Data.PeerPKI.CACertificate.FullPath()
+	cfg.PeerTLSInfo.ClientCertAuth = true
+	cfg.PeerTLSInfo.CRLFile = s.config.Data.PeerCRL.FullPath()
+
+	cfg.LCUrls = []url.URL{{
+		Scheme: "unix",
+		Path:   s.config.Ephemeral.ClientSocket.FullPath() + ":0",
+	}}
+	cfg.ACUrls = []url.URL{}
+	cfg.LPUrls = []url.URL{{
+		Scheme: "https",
+		Host:   fmt.Sprintf("%s:%d", s.config.ListenHost, port),
+	}}
+	cfg.APUrls = []url.URL{{
+		Scheme: "https",
+		Host:   fmt.Sprintf("%s:%d", s.config.ExternalHost, port),
+	}}
+
+	if s.config.NewCluster {
+		cfg.ClusterState = "new"
+		cfg.InitialCluster = cfg.InitialClusterFromName(cfg.Name)
+	} else if s.config.InitialCluster != "" {
+		cfg.ClusterState = "existing"
+		cfg.InitialCluster = s.config.InitialCluster
+	}
+
+	// TODO(q3k): pipe logs from etcd to supervisor.RawLogger via a file.
+	cfg.Logger = DefaultLogger
+	cfg.LogOutputs = []string{"stderr"}
+
+	return cfg, nil
+}
+
+// Run is a Supervisor runnable that starts the etcd member service. It will become healthy once the member joins the
+// cluster successfully.
+func (s *Service) Run(ctx context.Context) error {
+	st := &state{
+		ready: *atomic.NewBool(false),
+	}
+	s.stateMu.Lock()
+	s.state = st
+	s.stateMu.Unlock()
+
+	if s.config.NewCluster {
+		// Expect certificate to be absent from disk.
+		absent, err := s.config.Data.PeerPKI.AllAbsent()
+		if err != nil {
+			return fmt.Errorf("checking certificate existence: %w", err)
+		}
+		if !absent {
+			return fmt.Errorf("want new cluster, but certificates already exist on disk")
+		}
+
+		// Generate CA, keep in memory, write it down in etcd later.
+		st.ca, err = ca.New("Smalltown etcd peer Root CA")
+		if err != nil {
+			return fmt.Errorf("when creating new cluster's peer CA: %w", err)
+		}
+
+		ip := net.ParseIP(s.config.ExternalHost)
+		if ip == nil {
+			return fmt.Errorf("configued external host is not an IP address (got %q)", s.config.ExternalHost)
+		}
+
+		cert, key, err := st.ca.Issue(ctx, nil, s.config.Name, ip)
+		if err != nil {
+			return fmt.Errorf("when issuing new cluster's first certificate: %w", err)
+		}
+
+		if err := s.config.Data.PeerPKI.MkdirAll(0600); err != nil {
+			return fmt.Errorf("when creating PKI directory: %w", err)
+		}
+		if err := s.config.Data.PeerPKI.CACertificate.Write(pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: st.ca.CACertRaw}), 0600); err != nil {
+			return fmt.Errorf("when writing CA certificate to disk: %w", err)
+		}
+		if err := s.config.Data.PeerPKI.Certificate.Write(pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: cert}), 0600); err != nil {
+			return fmt.Errorf("when writing certificate to disk: %w", err)
+		}
+		if err := s.config.Data.PeerPKI.Key.Write(pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: key}), 0600); err != nil {
+			return fmt.Errorf("when writing certificate to disk: %w", err)
+		}
+	} else {
+		// Expect certificate to be present on disk.
+		present, err := s.config.Data.PeerPKI.AllExist()
+		if err != nil {
+			return fmt.Errorf("checking certificate existence: %w", err)
+		}
+		if !present {
+			return fmt.Errorf("want existing cluster, but certificate is missing from disk")
+		}
+	}
+
+	if err := s.config.Data.MkdirAll(0600); err != nil {
+		return fmt.Errorf("failed to create data directory; %w", err)
+	}
+
+	cfg, err := s.configure(ctx)
+	if err != nil {
+		return fmt.Errorf("when configuring etcd: %w", err)
+	}
+
+	server, err := embed.StartEtcd(cfg)
+	keep := false
+	defer func() {
+		if !keep && server != nil {
+			server.Close()
+		}
+	}()
+	if err != nil {
+		return fmt.Errorf("failed to start etcd: %w", err)
+	}
+	st.etcd = server
+
+	supervisor.Logger(ctx).Info("waiting for etcd...")
+
+	okay := true
+	select {
+	case <-st.etcd.Server.ReadyNotify():
+	case <-ctx.Done():
+		okay = false
+	}
+
+	if !okay {
+		supervisor.Logger(ctx).Info("context done, aborting wait")
+		return ctx.Err()
+	}
+
+	socket := s.config.Ephemeral.ClientSocket.FullPath()
+	cl, err := clientv3.New(clientv3.Config{
+		Endpoints:   []string{fmt.Sprintf("unix://%s:0", socket)},
+		DialTimeout: time.Second,
+	})
+	if err != nil {
+		return fmt.Errorf("failed to connect to new etcd instance: %w", err)
+	}
+	st.cl = cl
+
+	if s.config.NewCluster {
+		if st.ca == nil {
+			panic("peerCA has not been generated")
+		}
+
+		// Save new CA into etcd.
+		err = st.ca.Save(ctx, cl.KV)
+		if err != nil {
+			return fmt.Errorf("failed to save new CA to etcd: %w", err)
+		}
+	} else {
+		// Load existing CA from etcd.
+		st.ca, err = ca.Load(ctx, cl.KV)
+		if err != nil {
+			return fmt.Errorf("failed to load CA from etcd: %w", err)
+		}
+	}
+
+	// Start CRL watcher.
+	if err := supervisor.Run(ctx, "crl", s.watchCRL); err != nil {
+		return fmt.Errorf("failed to start CRL watcher: %w", err)
+	}
+	// Start autopromoter.
+	if err := supervisor.Run(ctx, "autopromoter", s.autopromoter); err != nil {
+		return fmt.Errorf("failed to start autopromoter: %w", err)
+	}
+
+	supervisor.Logger(ctx).Info("etcd is now ready")
+	keep = true
+	st.ready.Store(true)
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+
+	<-ctx.Done()
+	st.etcd.Close()
+	return ctx.Err()
+}
+
+// watchCRL is a sub-runnable of the etcd cluster member service that updates the on-local-storage CRL to match the
+// newest available version in etcd.
+func (s *Service) watchCRL(ctx context.Context) error {
+	s.stateMu.Lock()
+	cl := s.state.cl
+	ca := s.state.ca
+	s.stateMu.Unlock()
+
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	for e := range ca.WaitCRLChange(ctx, cl.KV, cl.Watcher) {
+		if e.Err != nil {
+			return fmt.Errorf("watching CRL: %w", e.Err)
+		}
+
+		if err := s.config.Data.PeerCRL.Write(e.CRL, 0600); err != nil {
+			return fmt.Errorf("saving CRL: %w", err)
+		}
+	}
+
+	// unreachable
+	return nil
+}
+
+func (s *Service) autopromoter(ctx context.Context) error {
+	t := time.NewTicker(5 * time.Second)
+	defer t.Stop()
+
+	autopromote := func() {
+		s.stateMu.Lock()
+		st := s.state
+		s.stateMu.Unlock()
+
+		if st.etcd.Server.Leader() != st.etcd.Server.ID() {
+			return
+		}
+
+		for _, member := range st.etcd.Server.Cluster().Members() {
+			if !member.IsLearner {
+				continue
+			}
+
+			// We always call PromoteMember since the metadata necessary to decide if we should is private.
+			// Luckily etcd already does sanity checks internally and will refuse to promote nodes that aren't
+			// connected or are still behind on transactions.
+			if _, err := st.etcd.Server.PromoteMember(ctx, uint64(member.ID)); err != nil {
+				supervisor.Logger(ctx).Infof("Failed to promote consensus node %s: %v", member.Name, err)
+			} else {
+				supervisor.Logger(ctx).Infof("Promoted new consensus node %s", member.Name)
+			}
+		}
+	}
+
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-t.C:
+			autopromote()
+		}
+	}
+}
+
+// IsReady returns whether etcd is ready and synced
+func (s *Service) IsReady() bool {
+	s.stateMu.Lock()
+	defer s.stateMu.Unlock()
+	if s.state == nil {
+		return false
+	}
+	return s.state.ready.Load()
+}
+
+func (s *Service) WaitReady(ctx context.Context) error {
+	// TODO(q3k): reimplement the atomic ready flag as an event synchronization mechanism
+	if s.IsReady() {
+		return nil
+	}
+	t := time.NewTicker(100 * time.Millisecond)
+	defer t.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-t.C:
+			if s.IsReady() {
+				return nil
+			}
+		}
+	}
+}
+
+// KV returns and etcd KV client interface to the etcd member/cluster.
+func (s *Service) KV(module, space string) clientv3.KV {
+	s.stateMu.Lock()
+	defer s.stateMu.Unlock()
+	return namespace.NewKV(s.state.cl.KV, fmt.Sprintf("%s:%s", module, space))
+}
+
+func (s *Service) KVRoot() clientv3.KV {
+	s.stateMu.Lock()
+	defer s.stateMu.Unlock()
+	return s.state.cl.KV
+}
+
+func (s *Service) Cluster() clientv3.Cluster {
+	s.stateMu.Lock()
+	defer s.stateMu.Unlock()
+	return s.state.cl.Cluster
+}
+
+// MemberInfo returns information about this etcd cluster member: its ID and name. This will block until this
+// information is available (ie. the cluster status is Ready).
+func (s *Service) MemberInfo(ctx context.Context) (id uint64, name string, err error) {
+	if err = s.WaitReady(ctx); err != nil {
+		err = fmt.Errorf("when waiting for cluster readiness: %w", err)
+		return
+	}
+
+	s.stateMu.Lock()
+	defer s.stateMu.Unlock()
+	id = uint64(s.state.etcd.Server.ID())
+	name = s.config.Name
+	return
+}
diff --git a/metropolis/node/core/consensus/consensus_test.go b/metropolis/node/core/consensus/consensus_test.go
new file mode 100644
index 0000000..e08bd29
--- /dev/null
+++ b/metropolis/node/core/consensus/consensus_test.go
@@ -0,0 +1,261 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package consensus
+
+import (
+	"bytes"
+	"context"
+	"crypto/x509"
+	"io/ioutil"
+	"net"
+	"os"
+	"testing"
+	"time"
+
+	"git.monogon.dev/source/nexantic.git/golibs/common"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+)
+
+type boilerplate struct {
+	ctx    context.Context
+	ctxC   context.CancelFunc
+	root   *localstorage.Root
+	tmpdir string
+}
+
+func prep(t *testing.T) *boilerplate {
+	ctx, ctxC := context.WithCancel(context.Background())
+	root := &localstorage.Root{}
+	tmp, err := ioutil.TempDir("", "smalltown-test")
+	if err != nil {
+		t.Fatal(err)
+	}
+	err = declarative.PlaceFS(root, tmp)
+	if err != nil {
+		t.Fatal(err)
+	}
+	os.MkdirAll(root.Data.Etcd.FullPath(), 0700)
+	os.MkdirAll(root.Ephemeral.Consensus.FullPath(), 0700)
+
+	return &boilerplate{
+		ctx:    ctx,
+		ctxC:   ctxC,
+		root:   root,
+		tmpdir: tmp,
+	}
+}
+
+func (b *boilerplate) close() {
+	b.ctxC()
+	os.RemoveAll(b.tmpdir)
+}
+
+func waitEtcd(t *testing.T, s *Service) {
+	deadline := time.Now().Add(5 * time.Second)
+	for {
+		if time.Now().After(deadline) {
+			t.Fatalf("etcd did not start up on time")
+		}
+		if s.IsReady() {
+			break
+		}
+		time.Sleep(100 * time.Millisecond)
+	}
+}
+
+func TestBootstrap(t *testing.T) {
+	b := prep(t)
+	defer b.close()
+	etcd := New(Config{
+		Data:           &b.root.Data.Etcd,
+		Ephemeral:      &b.root.Ephemeral.Consensus,
+		Name:           "test",
+		NewCluster:     true,
+		InitialCluster: "127.0.0.1",
+		ExternalHost:   "127.0.0.1",
+		ListenHost:     "127.0.0.1",
+		Port:           common.MustConsume(common.AllocateTCPPort()),
+	})
+
+	supervisor.New(b.ctx, etcd.Run)
+	waitEtcd(t, etcd)
+
+	kv := etcd.KV("foo", "bar")
+	if _, err := kv.Put(b.ctx, "/foo", "bar"); err != nil {
+		t.Fatalf("test key creation failed: %v", err)
+	}
+	if _, err := kv.Get(b.ctx, "/foo"); err != nil {
+		t.Fatalf("test key retrieval failed: %v", err)
+	}
+}
+
+func TestMemberInfo(t *testing.T) {
+	b := prep(t)
+	defer b.close()
+	etcd := New(Config{
+		Data:           &b.root.Data.Etcd,
+		Ephemeral:      &b.root.Ephemeral.Consensus,
+		Name:           "test",
+		NewCluster:     true,
+		InitialCluster: "127.0.0.1",
+		ExternalHost:   "127.0.0.1",
+		ListenHost:     "127.0.0.1",
+		Port:           common.MustConsume(common.AllocateTCPPort()),
+	})
+	supervisor.New(b.ctx, etcd.Run)
+	waitEtcd(t, etcd)
+
+	id, name, err := etcd.MemberInfo(b.ctx)
+	if err != nil {
+		t.Fatalf("MemberInfo: %v", err)
+	}
+
+	// Compare name with configured name.
+	if want, got := "test", name; want != got {
+		t.Errorf("MemberInfo returned name %q, wanted %q (per config)", got, want)
+	}
+
+	// Compare name with cluster information.
+	members, err := etcd.Cluster().MemberList(b.ctx)
+	if err != nil {
+		t.Errorf("MemberList: %v", err)
+	}
+
+	if want, got := 1, len(members.Members); want != got {
+		t.Fatalf("expected one cluster member, got %d", got)
+	}
+	if want, got := id, members.Members[0].ID; want != got {
+		t.Errorf("MemberInfo returned ID %d, Cluster endpoint says %d", want, got)
+	}
+	if want, got := name, members.Members[0].Name; want != got {
+		t.Errorf("MemberInfo returned name %q, Cluster endpoint says %q", want, got)
+	}
+}
+
+func TestRestartFromDisk(t *testing.T) {
+	b := prep(t)
+	defer b.close()
+
+	startEtcd := func(new bool) (*Service, context.CancelFunc) {
+		etcd := New(Config{
+			Data:           &b.root.Data.Etcd,
+			Ephemeral:      &b.root.Ephemeral.Consensus,
+			Name:           "test",
+			NewCluster:     new,
+			InitialCluster: "127.0.0.1",
+			ExternalHost:   "127.0.0.1",
+			ListenHost:     "127.0.0.1",
+			Port:           common.MustConsume(common.AllocateTCPPort()),
+		})
+		ctx, ctxC := context.WithCancel(b.ctx)
+		supervisor.New(ctx, etcd.Run)
+		waitEtcd(t, etcd)
+		kv := etcd.KV("foo", "bar")
+		if new {
+			if _, err := kv.Put(b.ctx, "/foo", "bar"); err != nil {
+				t.Fatalf("test key creation failed: %v", err)
+			}
+		}
+		if _, err := kv.Get(b.ctx, "/foo"); err != nil {
+			t.Fatalf("test key retrieval failed: %v", err)
+		}
+
+		return etcd, ctxC
+	}
+
+	etcd, ctxC := startEtcd(true)
+	etcd.stateMu.Lock()
+	firstCA := etcd.state.ca.CACertRaw
+	etcd.stateMu.Unlock()
+	ctxC()
+
+	etcd, ctxC = startEtcd(false)
+	etcd.stateMu.Lock()
+	secondCA := etcd.state.ca.CACertRaw
+	etcd.stateMu.Unlock()
+	ctxC()
+
+	if bytes.Compare(firstCA, secondCA) != 0 {
+		t.Fatalf("wanted same, got different CAs accross runs")
+	}
+}
+
+func TestCRL(t *testing.T) {
+	b := prep(t)
+	defer b.close()
+	etcd := New(Config{
+		Data:           &b.root.Data.Etcd,
+		Ephemeral:      &b.root.Ephemeral.Consensus,
+		Name:           "test",
+		NewCluster:     true,
+		InitialCluster: "127.0.0.1",
+		ExternalHost:   "127.0.0.1",
+		ListenHost:     "127.0.0.1",
+		Port:           common.MustConsume(common.AllocateTCPPort()),
+	})
+	supervisor.New(b.ctx, etcd.Run)
+	waitEtcd(t, etcd)
+
+	etcd.stateMu.Lock()
+	ca := etcd.state.ca
+	kv := etcd.state.cl.KV
+	etcd.stateMu.Unlock()
+
+	certRaw, _, err := ca.Issue(b.ctx, kv, "revoketest", net.ParseIP("1.2.3.4"))
+	if err != nil {
+		t.Fatalf("cert issue failed: %v", err)
+	}
+	cert, err := x509.ParseCertificate(certRaw)
+	if err != nil {
+		t.Fatalf("cert parse failed: %v", err)
+	}
+
+	if err := ca.Revoke(b.ctx, kv, "revoketest"); err != nil {
+		t.Fatalf("cert revoke failed: %v", err)
+	}
+
+	deadline := time.Now().Add(5 * time.Second)
+	for {
+		if time.Now().After(deadline) {
+			t.Fatalf("CRL did not get updated in time")
+		}
+		time.Sleep(100 * time.Millisecond)
+
+		crlRaw, err := b.root.Data.Etcd.PeerCRL.Read()
+		if err != nil {
+			// That's fine. Maybe it hasn't been written yet.
+			continue
+		}
+		crl, err := x509.ParseCRL(crlRaw)
+		if err != nil {
+			// That's fine. Maybe it hasn't been written yet.
+			continue
+		}
+
+		found := false
+		for _, revoked := range crl.TBSCertList.RevokedCertificates {
+			if revoked.SerialNumber.Cmp(cert.SerialNumber) == 0 {
+				found = true
+			}
+		}
+		if found {
+			break
+		}
+	}
+}
diff --git a/metropolis/node/core/debug_service.go b/metropolis/node/core/debug_service.go
new file mode 100644
index 0000000..0155cc6
--- /dev/null
+++ b/metropolis/node/core/debug_service.go
@@ -0,0 +1,243 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"context"
+	"crypto/x509"
+	"fmt"
+	"net"
+
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/status"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/cluster"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/consensus/ca"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+const (
+	logFilterMax = 1000
+)
+
+// debugService implements the Smalltown node debug API.
+type debugService struct {
+	cluster    *cluster.Manager
+	kubernetes *kubernetes.Service
+	logtree    *logtree.LogTree
+}
+
+func (s *debugService) GetGoldenTicket(ctx context.Context, req *apb.GetGoldenTicketRequest) (*apb.GetGoldenTicketResponse, error) {
+	ip := net.ParseIP(req.ExternalIp)
+	if ip == nil {
+		return nil, status.Errorf(codes.InvalidArgument, "could not parse IP %q", req.ExternalIp)
+	}
+	this := s.cluster.Node()
+
+	certRaw, key, err := s.nodeCertificate()
+	if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "failed to generate node certificate: %v", err)
+	}
+	cert, err := x509.ParseCertificate(certRaw)
+	if err != nil {
+		panic(err)
+	}
+	kv := s.cluster.ConsensusKVRoot()
+	ca, err := ca.Load(ctx, kv)
+	if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "could not load CA: %v", err)
+	}
+	etcdCert, etcdKey, err := ca.Issue(ctx, kv, cert.Subject.CommonName, ip)
+	if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "could not generate etcd peer certificate: %v", err)
+	}
+	etcdCRL, err := ca.GetCurrentCRL(ctx, kv)
+	if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "could not get etcd CRL: %v", err)
+	}
+
+	// Add new etcd member to etcd cluster.
+	etcd := s.cluster.ConsensusCluster()
+	etcdAddr := fmt.Sprintf("https://%s:%d", ip.String(), common.ConsensusPort)
+	_, err = etcd.MemberAddAsLearner(ctx, []string{etcdAddr})
+	if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "could not add as new etcd consensus member: %v", err)
+	}
+
+	return &apb.GetGoldenTicketResponse{
+		Ticket: &apb.GoldenTicket{
+			EtcdCaCert:     ca.CACertRaw,
+			EtcdClientCert: etcdCert,
+			EtcdClientKey:  etcdKey,
+			EtcdCrl:        etcdCRL,
+			Peers: []*apb.GoldenTicket_EtcdPeer{
+				{Name: this.ID(), Address: this.Address().String()},
+			},
+			This: &apb.GoldenTicket_EtcdPeer{Name: cert.Subject.CommonName, Address: ip.String()},
+
+			NodeId:   cert.Subject.CommonName,
+			NodeCert: certRaw,
+			NodeKey:  key,
+		},
+	}, nil
+}
+
+func (s *debugService) GetDebugKubeconfig(ctx context.Context, req *apb.GetDebugKubeconfigRequest) (*apb.GetDebugKubeconfigResponse, error) {
+	return s.kubernetes.GetDebugKubeconfig(ctx, req)
+}
+
+func (s *debugService) GetLogs(req *apb.GetLogsRequest, srv apb.NodeDebugService_GetLogsServer) error {
+	if len(req.Filters) > logFilterMax {
+		return status.Errorf(codes.InvalidArgument, "requested %d filters, maximum permitted is %d", len(req.Filters), logFilterMax)
+	}
+	dn := logtree.DN(req.Dn)
+	_, err := dn.Path()
+	switch err {
+	case nil:
+	case logtree.ErrInvalidDN:
+		return status.Errorf(codes.InvalidArgument, "invalid DN")
+	default:
+		return status.Errorf(codes.Unavailable, "could not parse DN: %v", err)
+	}
+
+	var options []logtree.LogReadOption
+
+	// Turn backlog mode into logtree option(s).
+	switch req.BacklogMode {
+	case apb.GetLogsRequest_BACKLOG_DISABLE:
+	case apb.GetLogsRequest_BACKLOG_ALL:
+		options = append(options, logtree.WithBacklog(logtree.BacklogAllAvailable))
+	case apb.GetLogsRequest_BACKLOG_COUNT:
+		count := int(req.BacklogCount)
+		if count <= 0 {
+			return status.Errorf(codes.InvalidArgument, "backlog_count must be > 0 if backlog_mode is BACKLOG_COUNT")
+		}
+		options = append(options, logtree.WithBacklog(count))
+	default:
+		return status.Errorf(codes.InvalidArgument, "unknown backlog_mode %d", req.BacklogMode)
+	}
+
+	// Turn stream mode into logtree option(s).
+	streamEnable := false
+	switch req.StreamMode {
+	case apb.GetLogsRequest_STREAM_DISABLE:
+	case apb.GetLogsRequest_STREAM_UNBUFFERED:
+		streamEnable = true
+		options = append(options, logtree.WithStream())
+	}
+
+	// Parse proto filters into logtree options.
+	for i, filter := range req.Filters {
+		switch inner := filter.Filter.(type) {
+		case *apb.LogFilter_WithChildren_:
+			options = append(options, logtree.WithChildren())
+		case *apb.LogFilter_OnlyRaw_:
+			options = append(options, logtree.OnlyRaw())
+		case *apb.LogFilter_OnlyLeveled_:
+			options = append(options, logtree.OnlyLeveled())
+		case *apb.LogFilter_LeveledWithMinimumSeverity_:
+			severity, err := logtree.SeverityFromProto(inner.LeveledWithMinimumSeverity.Minimum)
+			if err != nil {
+				return status.Errorf(codes.InvalidArgument, "filter %d has invalid severity: %v", i, err)
+			}
+			options = append(options, logtree.LeveledWithMinimumSeverity(severity))
+		}
+	}
+
+	reader, err := s.logtree.Read(logtree.DN(req.Dn), options...)
+	switch err {
+	case nil:
+	case logtree.ErrRawAndLeveled:
+		return status.Errorf(codes.InvalidArgument, "requested only raw and only leveled logs simultaneously")
+	default:
+		return status.Errorf(codes.Unavailable, "could not retrieve logs: %v", err)
+	}
+	defer reader.Close()
+
+	// Default protobuf message size limit is 64MB. We want to limit ourselves
+	// to 10MB.
+	// Currently each raw log line can be at most 1024 unicode codepoints (or
+	// 4096 bytes). To cover extra metadata and proto overhead, let's round
+	// this up to 4500 bytes. This in turn means we can store a maximum of
+	// (10e6/4500) == 2222 entries.
+	// Currently each leveled log line can also be at most 1024 unicode
+	// codepoints (or 4096 bytes). To cover extra metadata and proto overhead
+	// let's round this up to 2000 bytes. This in turn means we can store a
+	// maximum of (10e6/5000) == 2000 entries.
+	// The lowever of these numbers, ie the worst case scenario, is 2000
+	// maximum entries.
+	maxChunkSize := 2000
+
+	// Serve all backlog entries in chunks.
+	chunk := make([]*apb.LogEntry, 0, maxChunkSize)
+	for _, entry := range reader.Backlog {
+		p := entry.Proto()
+		if p == nil {
+			// TODO(q3k): log this once we have logtree/gRPC compatibility.
+			continue
+		}
+		chunk = append(chunk, p)
+
+		if len(chunk) >= maxChunkSize {
+			err := srv.Send(&apb.GetLogsResponse{
+				BacklogEntries: chunk,
+			})
+			if err != nil {
+				return err
+			}
+			chunk = make([]*apb.LogEntry, 0, maxChunkSize)
+		}
+	}
+
+	// Send last chunk of backlog, if present..
+	if len(chunk) > 0 {
+		err := srv.Send(&apb.GetLogsResponse{
+			BacklogEntries: chunk,
+		})
+		if err != nil {
+			return err
+		}
+		chunk = make([]*apb.LogEntry, 0, maxChunkSize)
+	}
+
+	// Start serving streaming data, if streaming has been requested.
+	if !streamEnable {
+		return nil
+	}
+
+	for {
+		entry, ok := <-reader.Stream
+		if !ok {
+			// Streaming has been ended by logtree - tell the client and return.
+			return status.Error(codes.Unavailable, "log streaming aborted by system")
+		}
+		p := entry.Proto()
+		if p == nil {
+			// TODO(q3k): log this once we have logtree/gRPC compatibility.
+			continue
+		}
+		err := srv.Send(&apb.GetLogsResponse{
+			StreamEntries: []*apb.LogEntry{p},
+		})
+		if err != nil {
+			return err
+		}
+	}
+}
diff --git a/metropolis/node/core/delve_disabled.go b/metropolis/node/core/delve_disabled.go
new file mode 100644
index 0000000..cb0a59b
--- /dev/null
+++ b/metropolis/node/core/delve_disabled.go
@@ -0,0 +1,23 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import "git.monogon.dev/source/nexantic.git/metropolis/node/core/network"
+
+// initializeDebugger does nothing in a non-debug build
+func initializeDebugger(*network.Service) {
+}
diff --git a/metropolis/node/core/delve_enabled.go b/metropolis/node/core/delve_enabled.go
new file mode 100644
index 0000000..b3b859c
--- /dev/null
+++ b/metropolis/node/core/delve_enabled.go
@@ -0,0 +1,41 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"context"
+	"fmt"
+	"os/exec"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network"
+)
+
+// initializeDebugger attaches Delve to ourselves and exposes it on common.DebuggerPort
+// This is coupled to compilation_mode=dbg because otherwise Delve doesn't have the necessary DWARF debug info
+func initializeDebugger(networkSvc *network.Service) {
+	go func() {
+		// This is intentionally delayed until network becomes available since Delve for some reason connects to itself
+		// and in early-boot no network interface is available to do that through. Also external access isn't possible
+		// early on anyways.
+		networkSvc.GetIP(context.Background(), true)
+		dlvCmd := exec.Command("/dlv", "--headless=true", fmt.Sprintf("--listen=:%v", common.DebuggerPort),
+			"--accept-multiclient", "--only-same-user=false", "attach", "--continue", "1", "/init")
+		if err := dlvCmd.Start(); err != nil {
+			panic(err)
+		}
+	}()
+}
diff --git a/metropolis/node/core/localstorage/BUILD.bazel b/metropolis/node/core/localstorage/BUILD.bazel
new file mode 100644
index 0000000..099a380
--- /dev/null
+++ b/metropolis/node/core/localstorage/BUILD.bazel
@@ -0,0 +1,26 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "directory_data.go",
+        "directory_pki.go",
+        "directory_root.go",
+        "storage.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage",
+    visibility = ["//metropolis/node:__subpackages__"],
+    deps = [
+        "//metropolis/node/core/localstorage/crypt:go_default_library",
+        "//metropolis/node/core/localstorage/declarative:go_default_library",
+        "//metropolis/node/core/tpm:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["storage_test.go"],
+    embed = [":go_default_library"],
+    deps = ["//metropolis/node/core/localstorage/declarative:go_default_library"],
+)
diff --git a/metropolis/node/core/localstorage/crypt/BUILD.bazel b/metropolis/node/core/localstorage/crypt/BUILD.bazel
new file mode 100644
index 0000000..38e27d6
--- /dev/null
+++ b/metropolis/node/core/localstorage/crypt/BUILD.bazel
@@ -0,0 +1,17 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "blockdev.go",
+        "crypt.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/crypt",
+    visibility = ["//metropolis/node/core/localstorage:__subpackages__"],
+    deps = [
+        "//metropolis/node/common/devicemapper:go_default_library",
+        "//metropolis/node/common/sysfs:go_default_library",
+        "@com_github_rekby_gpt//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/localstorage/crypt/blockdev.go b/metropolis/node/core/localstorage/crypt/blockdev.go
new file mode 100644
index 0000000..df5f590
--- /dev/null
+++ b/metropolis/node/core/localstorage/crypt/blockdev.go
@@ -0,0 +1,95 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package crypt
+
+import (
+	"context"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"strconv"
+
+	"github.com/rekby/gpt"
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/sysfs"
+)
+
+var (
+	// EFIPartitionType is the standardized partition type value for the EFI ESP partition. The human readable GUID is C12A7328-F81F-11D2-BA4B-00A0C93EC93B.
+	EFIPartitionType = gpt.PartType{0x28, 0x73, 0x2a, 0xc1, 0x1f, 0xf8, 0xd2, 0x11, 0xba, 0x4b, 0x00, 0xa0, 0xc9, 0x3e, 0xc9, 0x3b}
+
+	// SmalltownDataPartitionType is the partition type value for a Smalltown data partition. The human-readable GUID is 9eeec464-6885-414a-b278-4305c51f7966.
+	SmalltownDataPartitionType = gpt.PartType{0x64, 0xc4, 0xee, 0x9e, 0x85, 0x68, 0x4a, 0x41, 0xb2, 0x78, 0x43, 0x05, 0xc5, 0x1f, 0x79, 0x66}
+)
+
+const (
+	ESPDevicePath       = "/dev/esp"
+	SmalltownDataCryptPath = "/dev/data-crypt"
+)
+
+// MakeBlockDevices looks for the ESP and the Smalltown data partition and maps them to ESPDevicePath and
+// SmalltownDataCryptPath respectively. This doesn't fail if it doesn't find the partitions, only if
+// something goes catastrophically wrong.
+func MakeBlockDevices(ctx context.Context) error {
+	blockdevNames, err := ioutil.ReadDir("/sys/class/block")
+	if err != nil {
+		return fmt.Errorf("failed to read sysfs block class: %w", err)
+	}
+	for _, blockdevName := range blockdevNames {
+		ueventData, err := sysfs.ReadUevents(filepath.Join("/sys/class/block", blockdevName.Name(), "uevent"))
+		if err != nil {
+			return fmt.Errorf("failed to read uevent for block device %v: %w", blockdevName.Name(), err)
+		}
+		if ueventData["DEVTYPE"] == "disk" {
+			majorDev, err := strconv.Atoi(ueventData["MAJOR"])
+			if err != nil {
+				return fmt.Errorf("failed to convert uevent: %w", err)
+			}
+			devNodeName := fmt.Sprintf("/dev/%v", ueventData["DEVNAME"])
+			blkdev, err := os.Open(devNodeName)
+			if err != nil {
+				return fmt.Errorf("failed to open block device %v: %w", devNodeName, err)
+			}
+			defer blkdev.Close()
+			blockSize, err := unix.IoctlGetUint32(int(blkdev.Fd()), unix.BLKSSZGET)
+			if err != nil {
+				continue // This is not a regular block device
+			}
+			blkdev.Seek(int64(blockSize), 0)
+			table, err := gpt.ReadTable(blkdev, uint64(blockSize))
+			if err != nil {
+				// Probably just not a GPT-partitioned disk
+				continue
+			}
+			for partNumber, part := range table.Partitions {
+				if part.Type == EFIPartitionType {
+					if err := unix.Mknod(ESPDevicePath, 0600|unix.S_IFBLK, int(unix.Mkdev(uint32(majorDev), uint32(partNumber+1)))); err != nil {
+						return fmt.Errorf("failed to create device node for ESP partition: %w", err)
+					}
+				}
+				if part.Type == SmalltownDataPartitionType {
+					if err := unix.Mknod(SmalltownDataCryptPath, 0600|unix.S_IFBLK, int(unix.Mkdev(uint32(majorDev), uint32(partNumber+1)))); err != nil {
+						return fmt.Errorf("failed to create device node for Smalltown encrypted data partition: %w", err)
+					}
+				}
+			}
+		}
+	}
+	return nil
+}
diff --git a/metropolis/node/core/localstorage/crypt/crypt.go b/metropolis/node/core/localstorage/crypt/crypt.go
new file mode 100644
index 0000000..e0a8321
--- /dev/null
+++ b/metropolis/node/core/localstorage/crypt/crypt.go
@@ -0,0 +1,149 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package crypt
+
+import (
+	"encoding/binary"
+	"encoding/hex"
+	"fmt"
+	"os"
+	"syscall"
+
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/devicemapper"
+)
+
+func readDataSectors(path string) (uint64, error) {
+	integrityPartition, err := os.Open(path)
+	if err != nil {
+		return 0, err
+	}
+	defer integrityPartition.Close()
+	// Based on structure defined in
+	// https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/md/dm-integrity.c#n59
+	if _, err := integrityPartition.Seek(16, 0); err != nil {
+		return 0, err
+	}
+	var providedDataSectors uint64
+	if err := binary.Read(integrityPartition, binary.LittleEndian, &providedDataSectors); err != nil {
+		return 0, err
+	}
+	return providedDataSectors, nil
+}
+
+// cryptMap maps an encrypted device (node) at baseName to a
+// decrypted device at /dev/$name using the given encryptionKey
+func CryptMap(name string, baseName string, encryptionKey []byte) error {
+	integritySectors, err := readDataSectors(baseName)
+	if err != nil {
+		return fmt.Errorf("failed to read the number of usable sectors on the integrity device: %w", err)
+	}
+
+	integrityDevName := fmt.Sprintf("/dev/%v-integrity", name)
+	integrityDMName := fmt.Sprintf("%v-integrity", name)
+	integrityDev, err := devicemapper.CreateActiveDevice(integrityDMName, []devicemapper.Target{
+		devicemapper.Target{
+			Length:     integritySectors,
+			Type:       "integrity",
+			Parameters: fmt.Sprintf("%v 0 28 J 1 journal_sectors:1024", baseName),
+		},
+	})
+	if err != nil {
+		return fmt.Errorf("failed to create Integrity device: %w", err)
+	}
+	if err := unix.Mknod(integrityDevName, 0600|unix.S_IFBLK, int(integrityDev)); err != nil {
+		unix.Unlink(integrityDevName)
+		devicemapper.RemoveDevice(integrityDMName)
+		return fmt.Errorf("failed to create integrity device node: %w", err)
+	}
+
+	cryptDevName := fmt.Sprintf("/dev/%v", name)
+	cryptDev, err := devicemapper.CreateActiveDevice(name, []devicemapper.Target{
+		devicemapper.Target{
+			Length:     integritySectors,
+			Type:       "crypt",
+			Parameters: fmt.Sprintf("capi:gcm(aes)-random %v 0 %v 0 1 integrity:28:aead", hex.EncodeToString(encryptionKey), integrityDevName),
+		},
+	})
+	if err != nil {
+		unix.Unlink(integrityDevName)
+		devicemapper.RemoveDevice(integrityDMName)
+		return fmt.Errorf("failed to create crypt device: %w", err)
+	}
+	if err := unix.Mknod(cryptDevName, 0600|unix.S_IFBLK, int(cryptDev)); err != nil {
+		unix.Unlink(cryptDevName)
+		devicemapper.RemoveDevice(name)
+
+		unix.Unlink(integrityDevName)
+		devicemapper.RemoveDevice(integrityDMName)
+		return fmt.Errorf("failed to create crypt device node: %w", err)
+	}
+	return nil
+}
+
+// cryptInit initializes a new encrypted block device. This can take a long
+// time since all bytes on the mapped block device need to be zeroed.
+func CryptInit(name, baseName string, encryptionKey []byte) error {
+	integrityPartition, err := os.OpenFile(baseName, os.O_WRONLY, 0)
+	if err != nil {
+		return err
+	}
+	defer integrityPartition.Close()
+	zeroed512BBuf := make([]byte, 4096)
+	if _, err := integrityPartition.Write(zeroed512BBuf); err != nil {
+		return fmt.Errorf("failed to wipe header: %w", err)
+	}
+	integrityPartition.Close()
+
+	integrityDMName := fmt.Sprintf("%v-integrity", name)
+	_, err = devicemapper.CreateActiveDevice(integrityDMName, []devicemapper.Target{
+		{
+			Length:     1,
+			Type:       "integrity",
+			Parameters: fmt.Sprintf("%v 0 28 J 1 journal_sectors:1024", baseName),
+		},
+	})
+	if err != nil {
+		return fmt.Errorf("failed to create discovery integrity device: %w", err)
+	}
+	if err := devicemapper.RemoveDevice(integrityDMName); err != nil {
+		return fmt.Errorf("failed to remove discovery integrity device: %w", err)
+	}
+
+	if err := CryptMap(name, baseName, encryptionKey); err != nil {
+		return err
+	}
+
+	blkdev, err := os.OpenFile(fmt.Sprintf("/dev/%v", name), unix.O_DIRECT|os.O_WRONLY, 0000)
+	if err != nil {
+		return fmt.Errorf("failed to open new encrypted device for zeroing: %w", err)
+	}
+	defer blkdev.Close()
+	blockSize, err := unix.IoctlGetUint32(int(blkdev.Fd()), unix.BLKSSZGET)
+	zeroedBuf := make([]byte, blockSize*100) // Make it faster
+	for {
+		_, err := blkdev.Write(zeroedBuf)
+		if e, ok := err.(*os.PathError); ok && e.Err == syscall.ENOSPC {
+			break
+		}
+		if err != nil {
+			return fmt.Errorf("failed to zero-initalize new encrypted device: %w", err)
+		}
+	}
+	return nil
+}
diff --git a/metropolis/node/core/localstorage/declarative/BUILD.bazel b/metropolis/node/core/localstorage/declarative/BUILD.bazel
new file mode 100644
index 0000000..5b51aa2
--- /dev/null
+++ b/metropolis/node/core/localstorage/declarative/BUILD.bazel
@@ -0,0 +1,13 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "declarative.go",
+        "placement.go",
+        "placement_local.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative",
+    visibility = ["//metropolis/node:__subpackages__"],
+    deps = ["@org_golang_x_sys//unix:go_default_library"],
+)
diff --git a/metropolis/node/core/localstorage/declarative/declarative.go b/metropolis/node/core/localstorage/declarative/declarative.go
new file mode 100644
index 0000000..ce82c42
--- /dev/null
+++ b/metropolis/node/core/localstorage/declarative/declarative.go
@@ -0,0 +1,199 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package declarative
+
+import (
+	"fmt"
+	"reflect"
+	"strings"
+)
+
+// Directory represents the intent of existence of a directory in a hierarchical filesystem (simplified to a tree).
+// This structure can be embedded and still be interpreted as a Directory for purposes of use within this library. Any
+// inner fields of such an embedding structure that are in turn (embedded) Directories or files will be treated as
+// children in the intent expressed by this Directory. All contained directory fields must have a `dir:"name"` struct
+// tag that names them, and all contained file fields must have a `file:"name"` struct tag.
+//
+// Creation and management of the directory at runtime is left to the implementing code. However, the DirectoryPlacement
+// implementation (set as the directory is placed onto a backing store) facilitates this management (by exposing methods
+// that mutate the backing store).
+type Directory struct {
+	DirectoryPlacement
+}
+
+// File represents the intent of existence of a file. files are usually child structures in types that embed Directory.
+// File can also be embedded in another structure, and this embedding type will still be interpreted as a File for
+// purposes of use within this library.
+//
+// As with Directory, the runtime management of a File in a backing store is left to the implementing code, and the
+// embedded FilePlacement interface facilitates access to the backing store.
+type File struct {
+	FilePlacement
+}
+
+// unpackDirectory takes a pointer to Directory or a pointer to a structure embedding Directory, and returns a
+// reflection Value that refers to the passed structure itself (not its pointer) and a plain Go pointer to the
+// (embedded) Directory.
+func unpackDirectory(d interface{}) (*reflect.Value, *Directory, error) {
+	td := reflect.TypeOf(d)
+	if td.Kind() != reflect.Ptr {
+		return nil, nil, fmt.Errorf("wanted a pointer, got %v", td.Kind())
+	}
+
+	var dir *Directory
+	id := reflect.ValueOf(d).Elem()
+	tid := id.Type()
+	switch {
+	case tid.Name() == reflect.TypeOf(Directory{}).Name():
+		dir = id.Addr().Interface().(*Directory)
+	case id.FieldByName("Directory").IsValid():
+		dir = id.FieldByName("Directory").Addr().Interface().(*Directory)
+	default:
+		return nil, nil, fmt.Errorf("not a Directory or embedding Directory (%v)", id.Type().String())
+	}
+	return &id, dir, nil
+}
+
+// unpackFile takes a pointer to a File or a pointer to a structure embedding File, and returns a reflection Value that
+// refers to the passed structure itself (not its pointer) and a plain Go pointer to the (embedded) File.
+func unpackFile(f interface{}) (*reflect.Value, *File, error) {
+	tf := reflect.TypeOf(f)
+	if tf.Kind() != reflect.Ptr {
+		return nil, nil, fmt.Errorf("wanted a pointer, got %v", tf.Kind())
+	}
+
+	var fil *File
+	id := reflect.ValueOf(f).Elem()
+	tid := id.Type()
+	switch {
+	case tid.Name() == reflect.TypeOf(File{}).Name():
+		fil = id.Addr().Interface().(*File)
+	case id.FieldByName("File").IsValid():
+		fil = id.FieldByName("File").Addr().Interface().(*File)
+	default:
+		return nil, nil, fmt.Errorf("not a File or embedding File (%v)", tid.String())
+	}
+	return &id, fil, nil
+
+}
+
+// subdirs takes a pointer to a Directory or pointer to a structure embedding Directory, and returns a pair of pointers
+// to Directory-like structures contained within that directory with corresponding names (based on struct tags).
+func subdirs(d interface{}) ([]namedDirectory, error) {
+	s, _, err := unpackDirectory(d)
+	if err != nil {
+		return nil, fmt.Errorf("argument could not be parsed as *Directory: %w", err)
+	}
+
+	var res []namedDirectory
+	for i := 0; i < s.NumField(); i++ {
+		tf := s.Type().Field(i)
+		dirTag := tf.Tag.Get("dir")
+		if dirTag == "" {
+			continue
+		}
+		sf := s.Field(i)
+		res = append(res, namedDirectory{dirTag, sf.Addr().Interface()})
+	}
+	return res, nil
+}
+
+type namedDirectory struct {
+	name      string
+	directory interface{}
+}
+
+// files takes a pointer to a File or pointer to a structure embedding File, and returns a pair of pointers
+// to Directory-like structures contained within that directory with corresponding names (based on struct tags).
+func files(d interface{}) ([]namedFile, error) {
+	s, _, err := unpackDirectory(d)
+	if err != nil {
+		return nil, fmt.Errorf("argument could not be parsed as *Directory: %w", err)
+	}
+
+	var res []namedFile
+	for i := 0; i < s.NumField(); i++ {
+		tf := s.Type().Field(i)
+		fileTag := tf.Tag.Get("file")
+		if fileTag == "" {
+			continue
+		}
+		_, f, err := unpackFile(s.Field(i).Addr().Interface())
+		if err != nil {
+			return nil, fmt.Errorf("file %q could not be parsed as *File: %w", tf.Name, err)
+		}
+		res = append(res, namedFile{fileTag, f})
+	}
+	return res, nil
+}
+
+type namedFile struct {
+	name string
+	file *File
+}
+
+// Validate checks that a given pointer to a Directory or pointer to a structure containing Directory does not contain
+// any programmer errors in its definition:
+//  - all subdirectories/files must be named
+//  - all subdirectory/file names within a directory must be unique
+//  - all subdirectory/file names within a directory must not contain the '/' character (as it is a common path
+//    delimiter)
+func Validate(d interface{}) error {
+	names := make(map[string]bool)
+
+	subs, err := subdirs(d)
+	if err != nil {
+		return fmt.Errorf("could not get subdirectories: %w", err)
+	}
+
+	for _, nd := range subs {
+		if nd.name == "" {
+			return fmt.Errorf("subdirectory with empty name")
+		}
+		if strings.Contains(nd.name, "/") {
+			return fmt.Errorf("subdirectory with invalid path: %q", nd.name)
+		}
+		if names[nd.name] {
+			return fmt.Errorf("subdirectory with duplicate name: %q", nd.name)
+		}
+		names[nd.name] = true
+
+		err := Validate(nd.directory)
+		if err != nil {
+			return fmt.Errorf("%s: %w", nd.name, err)
+		}
+	}
+
+	filelist, err := files(d)
+	if err != nil {
+		return fmt.Errorf("could not get files: %w", err)
+	}
+
+	for _, nf := range filelist {
+		if nf.name == "" {
+			return fmt.Errorf("file with empty name")
+		}
+		if strings.Contains(nf.name, "/") {
+			return fmt.Errorf("file with invalid path: %q", nf.name)
+		}
+		if names[nf.name] {
+			return fmt.Errorf("file with duplicate name: %q", nf.name)
+		}
+		names[nf.name] = true
+	}
+	return nil
+}
diff --git a/metropolis/node/core/localstorage/declarative/placement.go b/metropolis/node/core/localstorage/declarative/placement.go
new file mode 100644
index 0000000..c2ff53d
--- /dev/null
+++ b/metropolis/node/core/localstorage/declarative/placement.go
@@ -0,0 +1,95 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package declarative
+
+import (
+	"fmt"
+	"os"
+)
+
+// A declarative Directory/File tree is an abstract definition until it's 'placed' on a backing file system.
+// By convention, all abstract definitions of hierarchies are stored as copiable structs, and only turned to pointers
+// when placed (ie., implementations like PlaceFS takes a *Directory, but Root as a declarative definition is defined as
+// non-pointer).
+
+// Placement is an interface available on Placed files and Directories. All *Placement interfaces on files/Directories
+// are only available on placed trees - eg., after a PlaceFS call. This is unfortunately not typesafe, callers need to
+// either be sure about placement, or check the interface for null.
+type Placement interface {
+	FullPath() string
+	RootRef() interface{}
+}
+
+// FilePlacement is an interface available on Placed files. It is implemented by different placement backends, and
+// set on all files during placement by a given backend.
+type FilePlacement interface {
+	Placement
+	Exists() (bool, error)
+	Read() ([]byte, error)
+	Write([]byte, os.FileMode) error
+}
+
+// DirectoryPlacement is an interface available on Placed Directories. It is implemented by different placement
+// backends, and set on all directories during placement by a given backend.
+type DirectoryPlacement interface {
+	Placement
+	// MkdirAll creates this directory and all its parents on backing stores that have a physical directory
+	// structure.
+	MkdirAll(file os.FileMode) error
+}
+
+// DirectoryPlacer is a placement backend-defined function that, given the path returned by the parent of a directory,
+// and the path to a directory, returns a DirectoryPlacement implementation for this directory. The new placement's
+// path (via .FullPath()) will be used for placement of directories/files within the new directory.
+type DirectoryPlacer func(parent, this string) DirectoryPlacement
+
+// FilePlacer is analogous to DirectoryPlacer, but for files.
+type FilePlacer func(parent, this string) FilePlacement
+
+// place recursively places a pointer to a Directory or pointer to a structure embedding Directory into a given backend,
+// by calling DirectoryPlacer and FilePlacer where appropriate. This is done recursively across a declarative tree until
+// all children are placed.
+func place(d interface{}, parent, this string, dpl DirectoryPlacer, fpl FilePlacer) error {
+	_, dir, err := unpackDirectory(d)
+	if err != nil {
+		return err
+	}
+
+	if dir.DirectoryPlacement != nil {
+		return fmt.Errorf("already placed")
+	}
+	dir.DirectoryPlacement = dpl(parent, this)
+
+	dirlist, err := subdirs(d)
+	if err != nil {
+		return fmt.Errorf("could not list subdirectories: %w", err)
+	}
+	for _, nd := range dirlist {
+		err := place(nd.directory, dir.FullPath(), nd.name, dpl, fpl)
+		if err != nil {
+			return fmt.Errorf("%v: %w", nd.name, err)
+		}
+	}
+	filelist, err := files(d)
+	if err != nil {
+		return fmt.Errorf("could not list files: %w", err)
+	}
+	for _, nf := range filelist {
+		nf.file.FilePlacement = fpl(dir.FullPath(), nf.name)
+	}
+	return nil
+}
diff --git a/metropolis/node/core/localstorage/declarative/placement_local.go b/metropolis/node/core/localstorage/declarative/placement_local.go
new file mode 100644
index 0000000..82b6a71
--- /dev/null
+++ b/metropolis/node/core/localstorage/declarative/placement_local.go
@@ -0,0 +1,115 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package declarative
+
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"sync"
+
+	"golang.org/x/sys/unix"
+)
+
+// FSRoot is a root of a storage backend that resides on the local filesystem.
+type FSRoot struct {
+	// The local path at which the declarative directory structure is located (eg. "/").
+	root string
+}
+
+type FSPlacement struct {
+	root      *FSRoot
+	path      string
+	writeLock sync.Mutex
+}
+
+func (f *FSPlacement) FullPath() string {
+	return f.path
+}
+
+func (f *FSPlacement) RootRef() interface{} {
+	return f.root
+}
+
+func (f *FSPlacement) Exists() (bool, error) {
+	_, err := os.Stat(f.FullPath())
+	if err == nil {
+		return true, nil
+	}
+	if os.IsNotExist(err) {
+		return false, nil
+	}
+	return false, err
+}
+
+func (f *FSPlacement) Read() ([]byte, error) {
+	return ioutil.ReadFile(f.FullPath())
+}
+
+// Write performs an atomic file write, via a temporary file.
+func (f *FSPlacement) Write(d []byte, mode os.FileMode) error {
+	f.writeLock.Lock()
+	defer f.writeLock.Unlock()
+
+	// TODO(q3k): ensure that these do not collide with an existing sibling file, or generate this suffix randomly.
+	tmp := f.FullPath() + ".__smalltown_tmp"
+	defer os.Remove(tmp)
+	if err := ioutil.WriteFile(tmp, d, mode); err != nil {
+		return fmt.Errorf("temporary file write failed: %w", err)
+	}
+
+	if err := unix.Rename(tmp, f.FullPath()); err != nil {
+		return fmt.Errorf("renaming target file failed: %w", err)
+	}
+
+	return nil
+}
+
+func (f *FSPlacement) MkdirAll(perm os.FileMode) error {
+	return os.MkdirAll(f.FullPath(), perm)
+}
+
+// PlaceFS takes a pointer to a Directory or a pointer to a structure embedding Directory and places it at a given
+// filesystem root. From this point on the given structure pointer has valid Placement interfaces.
+func PlaceFS(dd interface{}, root string) error {
+	r := &FSRoot{root}
+	pathFor := func(parent, this string) string {
+		var np string
+		switch {
+		case parent == "" && this == "":
+			np = "/"
+		case parent == "/":
+			np = "/" + this
+		default:
+			np = fmt.Sprintf("%s/%s", parent, this)
+		}
+		return np
+	}
+	dp := func(parent, this string) DirectoryPlacement {
+		np := pathFor(parent, this)
+		return &FSPlacement{path: np, root: r}
+	}
+	fp := func(parent, this string) FilePlacement {
+		np := pathFor(parent, this)
+		return &FSPlacement{path: np, root: r}
+	}
+	err := place(dd, r.root, "", dp, fp)
+	if err != nil {
+		return fmt.Errorf("could not place: %w", err)
+	}
+	return nil
+}
diff --git a/metropolis/node/core/localstorage/directory_data.go b/metropolis/node/core/localstorage/directory_data.go
new file mode 100644
index 0000000..e90dc48
--- /dev/null
+++ b/metropolis/node/core/localstorage/directory_data.go
@@ -0,0 +1,150 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package localstorage
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/crypt"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/tpm"
+)
+
+var keySize uint16 = 256 / 8
+
+// MountData mounts the Smalltown data partition with the given global unlock key. It automatically
+// unseals the local unlock key from the TPM.
+func (d *DataDirectory) MountExisting(unlock *ESPLocalUnlockFile, globalUnlockKey []byte) error {
+	d.flagLock.Lock()
+	defer d.flagLock.Unlock()
+
+	if !d.canMount {
+		return fmt.Errorf("cannot mount yet (root not ready?)")
+	}
+	if d.mounted {
+		return fmt.Errorf("already mounted")
+	}
+	d.mounted = true
+
+	localUnlockBlob, err := unlock.Read()
+	if err != nil {
+		return fmt.Errorf("reading local unlock file from ESP: %w", err)
+	}
+	localUnlockKey, err := tpm.Unseal(localUnlockBlob)
+	if err != nil {
+		return fmt.Errorf("unsealing local unlock key: %w", err)
+	}
+
+	key := make([]byte, keySize)
+	for i := uint16(0); i < keySize; i++ {
+		key[i] = localUnlockKey[i] ^ globalUnlockKey[i]
+	}
+
+	if err := crypt.CryptMap("data", crypt.SmalltownDataCryptPath, key); err != nil {
+		return err
+	}
+	if err := d.mount(); err != nil {
+		return err
+	}
+	return nil
+}
+
+// InitializeData initializes the Smalltown data partition and returns the global unlock key. It seals
+// the local portion into the TPM and stores the blob on the ESP. This is a potentially slow
+// operation since it touches the whole partition.
+func (d *DataDirectory) MountNew(unlock *ESPLocalUnlockFile) ([]byte, error) {
+	d.flagLock.Lock()
+	defer d.flagLock.Unlock()
+	if !d.canMount {
+		return nil, fmt.Errorf("cannot mount yet (root not ready?)")
+	}
+	if d.mounted {
+		return nil, fmt.Errorf("already mounted")
+	}
+	d.mounted = true
+
+	localUnlockKey, err := tpm.GenerateSafeKey(keySize)
+	if err != nil {
+		return nil, fmt.Errorf("generating local unlock key: %w", err)
+	}
+	globalUnlockKey, err := tpm.GenerateSafeKey(keySize)
+	if err != nil {
+		return nil, fmt.Errorf("generating global unlock key: %w", err)
+	}
+
+	localUnlockBlob, err := tpm.Seal(localUnlockKey, tpm.SecureBootPCRs)
+	if err != nil {
+		return nil, fmt.Errorf("sealing lock unlock key: %w", err)
+	}
+
+	// The actual key is generated by XORing together the localUnlockKey and the globalUnlockKey
+	// This provides us with a mathematical guarantee that the resulting key cannot be recovered
+	// whithout knowledge of both parts.
+	key := make([]byte, keySize)
+	for i := uint16(0); i < keySize; i++ {
+		key[i] = localUnlockKey[i] ^ globalUnlockKey[i]
+	}
+
+	if err := crypt.CryptInit("data", crypt.SmalltownDataCryptPath, key); err != nil {
+		return nil, fmt.Errorf("initializing encrypted block device: %w", err)
+	}
+	mkfsCmd := exec.Command("/bin/mkfs.xfs", "-qf", "/dev/data")
+	if _, err := mkfsCmd.Output(); err != nil {
+		return nil, fmt.Errorf("formatting encrypted block device: %w", err)
+	}
+
+	if err := d.mount(); err != nil {
+		return nil, fmt.Errorf("mounting: %w", err)
+	}
+
+	// TODO(q3k): do this automatically?
+	for _, d := range []declarative.DirectoryPlacement{
+		d.Etcd, d.Etcd.Data, d.Etcd.PeerPKI,
+		d.Containerd,
+		d.Kubernetes,
+		d.Kubernetes.Kubelet, d.Kubernetes.Kubelet.PKI, d.Kubernetes.Kubelet.Plugins, d.Kubernetes.Kubelet.PluginsRegistry,
+		d.Kubernetes.ClusterNetworking,
+		d.Node,
+		d.Volumes,
+	} {
+		err := d.MkdirAll(0700)
+		if err != nil {
+			return nil, fmt.Errorf("creating directory failed: %w", err)
+		}
+	}
+
+	if err := unlock.Write(localUnlockBlob, 0600); err != nil {
+		return nil, fmt.Errorf("writing unlock blob: %w", err)
+	}
+
+	return globalUnlockKey, nil
+}
+
+func (d *DataDirectory) mount() error {
+	if err := os.Mkdir(d.FullPath(), 0755); err != nil {
+		return fmt.Errorf("making data directory: %w", err)
+	}
+
+	if err := unix.Mount("/dev/data", d.FullPath(), "xfs", unix.MS_NOEXEC|unix.MS_NODEV, "pquota"); err != nil {
+		return fmt.Errorf("mounting data directory: %w", err)
+	}
+	return nil
+}
diff --git a/metropolis/node/core/localstorage/directory_pki.go b/metropolis/node/core/localstorage/directory_pki.go
new file mode 100644
index 0000000..6bdebff
--- /dev/null
+++ b/metropolis/node/core/localstorage/directory_pki.go
@@ -0,0 +1,167 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package localstorage
+
+import (
+	"crypto/ed25519"
+	"crypto/rand"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/hex"
+	"fmt"
+	"math/big"
+	"time"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+)
+
+var (
+	// From RFC 5280 Section 4.1.2.5
+	unknownNotAfter = time.Unix(253402300799, 0)
+)
+
+type CertificateTemplateNamer func(pubkey []byte) x509.Certificate
+
+func CertificateForNode(pubkey []byte) x509.Certificate {
+	name := "smalltown-" + hex.EncodeToString([]byte(pubkey[:16]))
+
+	// This has no SANs because it authenticates by public key, not by name
+	return x509.Certificate{
+		Subject: pkix.Name{
+			// We identify nodes by their ID public keys (not hashed since a strong hash is longer and serves no benefit)
+			CommonName: name,
+		},
+		IsCA:                  false,
+		BasicConstraintsValid: true,
+		NotBefore:             time.Now(),
+		NotAfter:              unknownNotAfter,
+		// Certificate is used both as server & client
+		ExtKeyUsage: []x509.ExtKeyUsage{x509.ExtKeyUsageClientAuth, x509.ExtKeyUsageServerAuth},
+	}
+}
+
+func (p *PKIDirectory) EnsureSelfSigned(namer CertificateTemplateNamer) (*tls.Certificate, error) {
+	create := false
+	for _, f := range []*declarative.File{&p.Certificate, &p.Key} {
+		exists, err := f.Exists()
+		if err != nil {
+			return nil, fmt.Errorf("could not check existence of file %q: %w", f.FullPath(), err)
+		}
+		if !exists {
+			create = true
+			break
+		}
+	}
+
+	if !create {
+		certRaw, err := p.Certificate.Read()
+		if err != nil {
+			return nil, fmt.Errorf("could not read certificate: %w", err)
+		}
+		privKeyRaw, err := p.Key.Read()
+		if err != nil {
+			return nil, fmt.Errorf("could not read key: %w", err)
+		}
+		cert, err := x509.ParseCertificate(certRaw)
+		if err != nil {
+			return nil, fmt.Errorf("could not parse certificate: %w", err)
+		}
+		privKey, err := x509.ParsePKCS8PrivateKey(privKeyRaw)
+		if err != nil {
+			return nil, fmt.Errorf("could not parse key: %w", err)
+		}
+		return &tls.Certificate{
+			Certificate: [][]byte{certRaw},
+			PrivateKey:  privKey,
+			Leaf:        cert,
+		}, nil
+	}
+
+	pubKey, privKeyRaw, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		return nil, fmt.Errorf("failed to generate key: %w", err)
+	}
+
+	privKey, err := x509.MarshalPKCS8PrivateKey(privKeyRaw)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal key: %w", err)
+	}
+
+	if err := p.Key.Write(privKey, 0600); err != nil {
+		return nil, fmt.Errorf("failed to write new private key: %w", err)
+	}
+
+	serialNumberLimit := new(big.Int).Lsh(big.NewInt(1), 127)
+	serialNumber, err := rand.Int(rand.Reader, serialNumberLimit)
+	if err != nil {
+		return nil, fmt.Errorf("failed to generate serial number: %w", err)
+	}
+
+	template := namer(pubKey)
+	template.SerialNumber = serialNumber
+
+	certRaw, err := x509.CreateCertificate(rand.Reader, &template, &template, pubKey, privKeyRaw)
+	if err != nil {
+		return nil, fmt.Errorf("could not sign certificate: %w", err)
+	}
+
+	cert, err := x509.ParseCertificate(certRaw)
+	if err != nil {
+		return nil, fmt.Errorf("could not parse newly created certificate: %w", err)
+	}
+
+	if err := p.Certificate.Write(certRaw, 0600); err != nil {
+		return nil, fmt.Errorf("failed to write new certificate: %w", err)
+	}
+
+	return &tls.Certificate{
+		Certificate: [][]byte{certRaw},
+		PrivateKey:  privKey,
+		Leaf:        cert,
+	}, nil
+}
+
+// AllExist returns true if all PKI files (cert, key, CA cert) are present on the backing
+// store.
+func (p *PKIDirectory) AllExist() (bool, error) {
+	for _, d := range []*declarative.File{&p.CACertificate, &p.Certificate, &p.Key} {
+		exists, err := d.Exists()
+		if err != nil {
+			return false, fmt.Errorf("failed to check %q: %v", d.FullPath(), err)
+		}
+		if !exists {
+			return false, nil
+		}
+	}
+	return true, nil
+}
+
+// AllAbsent returns true if all PKI files (cert, key, CA cert) are missing from the backing
+// store.
+func (p *PKIDirectory) AllAbsent() (bool, error) {
+	for _, d := range []*declarative.File{&p.CACertificate, &p.Certificate, &p.Key} {
+		exists, err := d.Exists()
+		if err != nil {
+			return false, fmt.Errorf("failed to check %q: %v", d.FullPath(), err)
+		}
+		if exists {
+			return false, nil
+		}
+	}
+	return true, nil
+}
diff --git a/metropolis/node/core/localstorage/directory_root.go b/metropolis/node/core/localstorage/directory_root.go
new file mode 100644
index 0000000..883d1e2
--- /dev/null
+++ b/metropolis/node/core/localstorage/directory_root.go
@@ -0,0 +1,83 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package localstorage
+
+import (
+	"context"
+	"fmt"
+	"os"
+
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/crypt"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+)
+
+func (r *Root) Start(ctx context.Context) error {
+	r.Data.flagLock.Lock()
+	defer r.Data.flagLock.Unlock()
+	if r.Data.canMount {
+		return fmt.Errorf("cannot re-start root storage")
+	}
+	// TODO(q3k): turn this into an Ensure call
+	err := crypt.MakeBlockDevices(ctx)
+	if err != nil {
+		return fmt.Errorf("MakeBlockDevices: %w", err)
+	}
+
+	if err := os.Mkdir(r.ESP.FullPath(), 0755); err != nil {
+		return fmt.Errorf("making ESP directory: %w", err)
+	}
+
+	if err := unix.Mount(crypt.ESPDevicePath, r.ESP.FullPath(), "vfat", unix.MS_NOEXEC|unix.MS_NODEV|unix.MS_SYNC, ""); err != nil {
+		return fmt.Errorf("mounting ESP partition: %w", err)
+	}
+
+	r.Data.canMount = true
+
+	if err := os.Mkdir(r.Tmp.FullPath(), 0777); err != nil {
+		return fmt.Errorf("making /tmp directory: %w", err)
+	}
+
+	if err := unix.Mount("tmpfs", r.Tmp.FullPath(), "tmpfs", unix.MS_NOEXEC|unix.MS_NODEV, ""); err != nil {
+		return fmt.Errorf("mounting /tmp: %w", err)
+	}
+
+	// TODO(q3k): do this automatically?
+	for _, d := range []declarative.DirectoryPlacement{
+		r.Etc,
+		r.Ephemeral,
+		r.Ephemeral.Consensus,
+		r.Ephemeral.Containerd, r.Ephemeral.Containerd.Tmp, r.Ephemeral.Containerd.RunSC, r.Ephemeral.Containerd.IPAM,
+		r.Ephemeral.FlexvolumePlugins,
+	} {
+		err := d.MkdirAll(0700)
+		if err != nil {
+			return fmt.Errorf("creating directory failed: %w", err)
+		}
+	}
+
+	for _, d := range []declarative.DirectoryPlacement{
+		r.Ephemeral, r.Ephemeral.Containerd, r.Ephemeral.Containerd.Tmp,
+	} {
+		if err := os.Chmod(d.FullPath(), 0755); err != nil {
+			return fmt.Errorf("failed to chmod containerd tmp path: %w", err)
+		}
+	}
+
+	return nil
+}
diff --git a/metropolis/node/core/localstorage/storage.go b/metropolis/node/core/localstorage/storage.go
new file mode 100644
index 0000000..8cc291f
--- /dev/null
+++ b/metropolis/node/core/localstorage/storage.go
@@ -0,0 +1,161 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package localstorage
+
+// Localstorage is a replacement for the old 'storage' internal library. It is currently unused, but will become
+// so as the node code gets rewritten.
+
+// The library is centered around the idea of a declarative filesystem tree defined as mutually recursive Go structs.
+// This structure is then Placed onto an abstract real filesystem (eg. a local POSIX filesystem at /), and a handle
+// to that placed filesystem is then used by the consumers of this library to refer to subsets of the tree (that now
+// correspond to locations on a filesystem).
+//
+// Every member of the storage hierarchy must either be, or inherit from Directory or File. In order to be placed
+// correctly, Directory embedding structures must use `dir:` or `file:` tags for child Directories and files
+// respectively. The content of the tag specifies the path part that this element will be placed at.
+//
+// Full placement path(available via FullPath()) format is placement implementation-specific. However, they're always
+// strings.
+
+import (
+	"sync"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+)
+
+type Root struct {
+	declarative.Directory
+	// UEFI ESP partition, mounted from plaintext storage.
+	ESP ESPDirectory `dir:"esp"`
+	// Persistent Data partition, mounted from encrypted and authenticated storage.
+	Data DataDirectory `dir:"data"`
+	// FHS-standard /etc directory, containes /etc/hosts, /etc/machine-id, and other compatibility files.
+	Etc EtcDirectory `dir:"etc"`
+	// Ephemeral data, used by runtime, stored in tmpfs. Things like sockets, temporary config files, etc.
+	Ephemeral EphemeralDirectory `dir:"ephemeral"`
+	// FHS-standard /tmp directory, used by ioutil.TempFile.
+	Tmp TmpDirectory `dir:"tmp"`
+}
+
+type PKIDirectory struct {
+	declarative.Directory
+	CACertificate declarative.File `file:"ca.pem"`
+	Certificate   declarative.File `file:"cert.pem"`
+	Key           declarative.File `file:"cert-key.pem"`
+}
+
+// ESPDirectory is the EFI System Partition.
+type ESPDirectory struct {
+	declarative.Directory
+	LocalUnlock ESPLocalUnlockFile `file:"local_unlock.bin"`
+	// Enrolment is the configuration/provisioning file for this node, containing information required to begin
+	// joining the cluster.
+	Enrolment declarative.File `file:"enrolment.pb"`
+}
+
+// ESPLocalUnlockFile is the localUnlock file, encrypted by the TPM of this node. After decrypting by the TPM it is used
+// in conjunction with the globalUnlock key (retrieved from the existing cluster) to decrypt the local data partition.
+type ESPLocalUnlockFile struct {
+	declarative.File
+}
+
+// DataDirectory is an xfs partition mounted via cryptsetup/LUKS, with a key derived from {global,local}Unlock keys.
+type DataDirectory struct {
+	declarative.Directory
+
+	// flagLock locks canMount and mounted.
+	flagLock sync.Mutex
+	// canMount is set by Root when it is initialized. It is required to be set for mounting the data directory.
+	canMount bool
+	// mounted is set by DataDirectory when it is mounted. It ensures it's only mounted once.
+	mounted bool
+
+	Containerd declarative.Directory   `dir:"containerd"`
+	Etcd       DataEtcdDirectory       `dir:"etcd"`
+	Kubernetes DataKubernetesDirectory `dir:"kubernetes"`
+	Node       PKIDirectory            `dir:"node_pki"`
+	Volumes    DataVolumesDirectory    `dir:"volumes"`
+}
+
+type DataEtcdDirectory struct {
+	declarative.Directory
+	PeerPKI PKIDirectory          `dir:"peer_pki"`
+	PeerCRL declarative.File      `file:"peer_crl"`
+	Data    declarative.Directory `dir:"data"`
+}
+
+type DataKubernetesDirectory struct {
+	declarative.Directory
+	ClusterNetworking DataKubernetesClusterNetworkingDirectory `dir:"clusternet"`
+	Kubelet           DataKubernetesKubeletDirectory           `dir:"kubelet"`
+}
+
+type DataKubernetesClusterNetworkingDirectory struct {
+	declarative.Directory
+	Key declarative.File `file:"private.key"`
+}
+
+type DataKubernetesKubeletDirectory struct {
+	declarative.Directory
+	Kubeconfig declarative.File `file:"kubeconfig"`
+	PKI        PKIDirectory     `dir:"pki"`
+
+	Plugins struct {
+		declarative.Directory
+		VFS declarative.File `file:"com.smalltown.vfs.sock"`
+	} `dir:"plugins"`
+
+	PluginsRegistry struct {
+		declarative.Directory
+		VFSReg declarative.File `file:"com.smalltown.vfs-reg.sock"`
+	} `dir:"plugins_registry"`
+}
+
+type DataVolumesDirectory struct {
+	declarative.Directory
+}
+
+type EtcDirectory struct {
+	declarative.Directory
+	Hosts     declarative.File `file:"hosts"`
+	MachineID declarative.File `file:"machine-id"`
+}
+
+type EphemeralDirectory struct {
+	declarative.Directory
+	Consensus         EphemeralConsensusDirectory  `dir:"consensus"`
+	Containerd        EphemeralContainerdDirectory `dir:"containerd"`
+	FlexvolumePlugins declarative.Directory        `dir:"flexvolume_plugins"`
+}
+
+type EphemeralConsensusDirectory struct {
+	declarative.Directory
+	ClientSocket declarative.File `file:"client.sock"`
+}
+
+type EphemeralContainerdDirectory struct {
+	declarative.Directory
+	ClientSocket  declarative.File      `file:"client.sock"`
+	RunSCLogsFIFO declarative.File      `file:"runsc-logs.fifo"`
+	Tmp           declarative.Directory `dir:"tmp"`
+	RunSC         declarative.Directory `dir:"runsc"`
+	IPAM          declarative.Directory `dir:"ipam"`
+}
+
+type TmpDirectory struct {
+	declarative.Directory
+}
diff --git a/metropolis/node/core/localstorage/storage_test.go b/metropolis/node/core/localstorage/storage_test.go
new file mode 100644
index 0000000..8029d02
--- /dev/null
+++ b/metropolis/node/core/localstorage/storage_test.go
@@ -0,0 +1,58 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package localstorage
+
+import (
+	"testing"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+)
+
+func TestValidateAll(t *testing.T) {
+	r := Root{}
+	if err := declarative.Validate(&r); err != nil {
+		t.Errorf("Validation failed: %v", err)
+	}
+}
+
+func TestPlaceFS(t *testing.T) {
+	rr := Root{}
+	err := declarative.PlaceFS(&rr, "")
+	if err != nil {
+		t.Errorf("Placement failed: %v", err)
+	}
+
+	// Re-placing should fail.
+	err = declarative.PlaceFS(&rr, "/foo")
+	if err == nil {
+		t.Errorf("Re-placement didn't fail")
+	}
+
+	// Check some absolute paths.
+	for i, te := range []struct {
+		pl   declarative.Placement
+		want string
+	}{
+		{rr.ESP, "/esp"},
+		{rr.Data.Etcd, "/data/etcd"},
+		{rr.Data.Node.Certificate, "/data/node_pki/cert.pem"},
+	} {
+		if got, want := te.pl.FullPath(), te.want; got != want {
+			t.Errorf("test %d: wanted path %q, got %q", i, want, got)
+		}
+	}
+}
diff --git a/metropolis/node/core/logtree/BUILD.bazel b/metropolis/node/core/logtree/BUILD.bazel
new file mode 100644
index 0000000..120bf9f
--- /dev/null
+++ b/metropolis/node/core/logtree/BUILD.bazel
@@ -0,0 +1,32 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "doc.go",
+        "journal.go",
+        "journal_entry.go",
+        "journal_subscriber.go",
+        "leveled.go",
+        "leveled_payload.go",
+        "logtree.go",
+        "logtree_access.go",
+        "logtree_entry.go",
+        "logtree_publisher.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//metropolis/node/common/logbuffer:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = [
+        "journal_test.go",
+        "logtree_test.go",
+    ],
+    embed = [":go_default_library"],
+)
diff --git a/metropolis/node/core/logtree/doc.go b/metropolis/node/core/logtree/doc.go
new file mode 100644
index 0000000..ab3c537
--- /dev/null
+++ b/metropolis/node/core/logtree/doc.go
@@ -0,0 +1,116 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+/*
+Package logtree implements a tree-shaped logger for debug events. It provides log publishers (ie. Go code) with a
+glog-like API and io.Writer API, with loggers placed in a hierarchical structure defined by a dot-delimited path
+(called a DN, short for Distinguished Name).
+
+    tree.MustLeveledFor("foo.bar.baz").Warningf("Houston, we have a problem: %v", err)
+    fmt.Fprintf(tree.MustRawFor("foo.bar.baz"), "some\nunstructured\ndata\n")
+
+Logs in this context are unstructured, operational and developer-centric human readable text messages presented as lines
+of text to consumers, with some attached metadata. Logtree does not deal with 'structured' logs as some parts of the
+industry do, and instead defers any machine-readable logs to either be handled by metrics systems like Prometheus or
+event sourcing systems like Kafka.
+
+Tree Structure
+
+As an example, consider an application that produces logs with the following DNs:
+
+    listener.http
+    listener.grpc
+    svc
+    svc.cache
+    svc.cache.gc
+
+This would correspond to a tree as follows:
+
+                          .------.
+                         |   ""   |
+                         | (root) |
+                          '------'
+           .----------------'   '------.
+    .--------------.           .---------------.
+    |     svc      |           |    listener   |
+    '--------------'           '---------------'
+           |                   .----'      '----.
+    .--------------.  .---------------.  .---------------.
+    |  svc.cache   |  | listener.http |  | listener.grpc |
+    '--------------'  '---------------'  '---------------'
+           |
+    .--------------.
+    | svc.cache.gc |
+    '--------------'
+
+In this setup, every DN acts as a separate logging target, each with its own retention policy and quota. Logging to a DN
+under foo.bar does NOT automatically log to foo - all tree mechanisms are applied on log access by consumers. Loggers
+are automatically created on first use, and importantly, can be created at any time, and will automatically be created
+if a sub-DN is created that requires a parent DN to exist first. Note, for instance, that a `listener` logging node was
+created even though the example application only logged to `listener.http` and `listener.grpc`.
+
+An implicit root node is always present in the tree, accessed by DN "" (an empty string). All other logger nodes are
+children (or transitive children) of the root node.
+
+Log consumers (application code that reads the log and passes them on to operators, or ships them off for aggregation in
+other systems) to select subtrees of logs for readout. In the example tree, a consumer could select to either read all
+logs of the entire tree, just a single DN (like svc), or a subtree (like everything under listener, ie. messages emitted
+to listener.http and listener.grpc).
+
+Leveled Log Producer API
+
+As part of the glog-like logging API available to producers, the following metadata is attached to emitted logs in
+addition to the DN of the logger to which the log entry was emitted:
+
+ - timestamp at which the entry was emitted
+ - a severity level (one of FATAL, ERROR, WARN or INFO)
+ - a source of the message (file name and line number)
+
+In addition, the logger mechanism supports a variable verbosity level (so-called 'V-logging') that can be set at every
+node of the tree. For more information about the producer-facing logging API, see the documentation of the LeveledLogger
+interface, which is the main interface exposed to log producers.
+
+If the submitted message contains newlines, it will be split accordingly into a single log entry that contains multiple
+string lines. This allows for log producers to submit long, multi-line messages that are guaranteed to be non-interleaved
+with other entries, and allows for access API consumers to maintain semantic linking between multiple lines being emitted
+as a single atomic entry.
+
+Raw Log Producer API
+
+In addition to leveled, glog-like logging, LogTree supports 'raw logging'. This is implemented as an io.Writer that will
+split incoming bytes into newline-delimited lines, and log them into that logtree's DN. This mechanism is primarily
+intended to support storage of unstructured log data from external processes - for example binaries running with redirected
+stdout/stderr.
+
+Log Access API
+
+The Log Access API is mostly exposed via a single function on the LogTree struct: Read. It allows access to log entries
+that have been already buffered inside LogTree and to subscribe to receive future entries over a channel. As outlined
+earlier, any access can specify whether it is just interested in a single logger (addressed by DN), or a subtree of
+loggers.
+
+Due to the current implementation of the logtree, subtree accesses of backlogged data is significantly slower than
+accessing data of just one DN, or the whole tree (as every subtree backlog access performs a scan on all logged data).
+Thus, log consumers should be aware that it is much better to stream and buffer logs specific to some long-standing
+logging request on their own, rather than repeatedly perform reads of a subtree backlog.
+
+The data returned from the log access API is a LogEntry, which itself can contain either a raw logging entry, or a leveled
+logging entry. Helper functions are available on LogEntry that allow canonical string representations to be returned, for
+easy use in consuming tools/interfaces. Alternatively, the consumer can itself access the internal raw/leveled entries and
+print them according to their own preferred format.
+
+*/
+package logtree
diff --git a/metropolis/node/core/logtree/journal.go b/metropolis/node/core/logtree/journal.go
new file mode 100644
index 0000000..78c55a1
--- /dev/null
+++ b/metropolis/node/core/logtree/journal.go
@@ -0,0 +1,218 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"errors"
+	"strings"
+	"sync"
+)
+
+// DN is the Distinguished Name, a dot-delimited path used to address loggers within a LogTree. For example, "foo.bar"
+// designates the 'bar' logger node under the 'foo' logger node under the root node of the logger. An empty string is
+// the root node of the tree.
+type DN string
+
+var (
+	ErrInvalidDN = errors.New("invalid DN")
+)
+
+// Path return the parts of a DN, ie. all the elements of the dot-delimited DN path. For the root node, an empty list
+// will be returned. An error will be returned if the DN is invalid (contains empty parts, eg. `foo..bar`, `.foo` or
+// `foo.`.
+func (d DN) Path() ([]string, error) {
+	if d == "" {
+		return nil, nil
+	}
+	parts := strings.Split(string(d), ".")
+	for _, p := range parts {
+		if p == "" {
+			return nil, ErrInvalidDN
+		}
+	}
+	return parts, nil
+}
+
+// journal is the main log recording structure of logtree. It manages linked lists containing the actual log entries,
+// and implements scans across them. It does not understand the hierarchical nature of logtree, and instead sees all
+// entries as part of a global linked list and a local linked list for a given DN.
+//
+// The global linked list is represented by the head/tail pointers in journal and nextGlobal/prevGlobal pointers in
+// entries. The local linked lists are represented by heads[DN]/tails[DN] pointers in journal and nextLocal/prevLocal
+// pointers in entries:
+//
+//       .------------.        .------------.        .------------.
+//       | dn: A.B    |        | dn: Z      |        | dn: A.B    |
+//       | time: 1    |        | time: 2    |        | time: 3    |
+//       |------------|        |------------|        |------------|
+//       | nextGlobal :------->| nextGlobal :------->| nextGlobal :--> nil
+// nil <-: prevGlobal |<-------: prevGlobal |<-------| prevGlobal |
+//       |------------|        |------------|  n     |------------|
+//       | nextLocal  :---. n  | nextLocal  :->i .-->| nextLocal  :--> nil
+// nil <-: prevLocal  |<--: i<-: prevLocal  |  l :---| prevLocal  |
+//       '------------'   | l  '------------'    |   '------------'
+//            ^           '----------------------'         ^
+//            |                      ^                     |
+//            |                      |                     |
+//         ( head )             ( tails[Z] )            ( tail )
+//      ( heads[A.B] )          ( heads[Z] )         ( tails[A.B] )
+//
+type journal struct {
+	// mu locks the rest of the structure. It must be taken during any operation on the journal.
+	mu sync.RWMutex
+
+	// tail is the side of the global linked list that contains the newest log entry, ie. the one that has been pushed
+	// the most recently. It can be nil when no log entry has yet been pushed. The global linked list contains all log
+	// entries pushed to the journal.
+	tail *entry
+	// head is the side of the global linked list that contains the oldest log entry. It can be nil when no log entry
+	// has yet been pushed.
+	head *entry
+
+	// tails are the tail sides of a local linked list for a given DN, ie. the sides that contain the newest entry. They
+	// are nil if there are no log entries for that DN.
+	tails map[DN]*entry
+	// heads are the head sides of a local linked list for a given DN, ie. the sides that contain the oldest entry. They
+	// are nil if there are no log entries for that DN.
+	heads map[DN]*entry
+
+	// quota is a map from DN to quota structure, representing the quota policy of a particular DN-designated logger.
+	quota map[DN]*quota
+
+	// subscribers are observer to logs. New log entries get emitted to channels present in the subscriber structure,
+	// after filtering them through subscriber-provided filters (eg. to limit events to subtrees that interest that
+	// particular subscriber).
+	subscribers []*subscriber
+}
+
+// newJournal creates a new empty journal. All journals are independent from eachother, and as such, all LogTrees are
+// also independent.
+func newJournal() *journal {
+	return &journal{
+		tails: make(map[DN]*entry),
+		heads: make(map[DN]*entry),
+
+		quota: make(map[DN]*quota),
+	}
+}
+
+// filter is a predicate that returns true if a log subscriber or reader is interested in a given log entry.
+type filter func(*entry) bool
+
+// filterAll returns a filter that accepts all log entries.
+func filterAll() filter {
+	return func(*entry) bool { return true }
+}
+
+// filterExact returns a filter that accepts only log entries at a given exact DN. This filter should not be used in
+// conjunction with journal.scanEntries - instead, journal.getEntries should be used, as it is much faster.
+func filterExact(dn DN) filter {
+	return func(e *entry) bool {
+		return e.origin == dn
+	}
+}
+
+// filterSubtree returns a filter that accepts all log entries at a given DN and sub-DNs. For example, filterSubtree at
+// "foo.bar" would allow entries at "foo.bar", "foo.bar.baz", but not "foo" or "foo.barr".
+func filterSubtree(root DN) filter {
+	if root == "" {
+		return filterAll()
+	}
+
+	rootParts := strings.Split(string(root), ".")
+	return func(e *entry) bool {
+		parts := strings.Split(string(e.origin), ".")
+		if len(parts) < len(rootParts) {
+			return false
+		}
+
+		for i, p := range rootParts {
+			if parts[i] != p {
+				return false
+			}
+		}
+
+		return true
+	}
+}
+
+// filterSeverity returns a filter that accepts log entries at a given severity level or above. See the Severity type
+// for more information about severity levels.
+func filterSeverity(atLeast Severity) filter {
+	return func(e *entry) bool {
+		return e.leveled != nil && e.leveled.severity.AtLeast(atLeast)
+	}
+}
+
+func filterOnlyRaw(e *entry) bool {
+	return e.raw != nil
+}
+
+func filterOnlyLeveled(e *entry) bool {
+	return e.leveled != nil
+}
+
+// scanEntries does a linear scan through the global entry list and returns all entries that match the given filters. If
+// retrieving entries for an exact event, getEntries should be used instead, as it will leverage DN-local linked lists
+// to retrieve them faster.
+// journal.mu must be taken at R or RW level when calling this function.
+func (j *journal) scanEntries(filters ...filter) (res []*entry) {
+	cur := j.tail
+	for {
+		if cur == nil {
+			return
+		}
+
+		passed := true
+		for _, filter := range filters {
+			if !filter(cur) {
+				passed = false
+				break
+			}
+		}
+		if passed {
+			res = append(res, cur)
+		}
+		cur = cur.nextGlobal
+	}
+}
+
+// getEntries returns all entries at a given DN. This is faster than a scanEntries(filterExact), as it uses the special
+// local linked list pointers to traverse the journal. Additional filters can be passed to further limit the entries
+// returned, but a scan through this DN's local linked list will be performed regardless.
+// journal.mu must be taken at R or RW level when calling this function.
+func (j *journal) getEntries(exact DN, filters ...filter) (res []*entry) {
+	cur := j.tails[exact]
+	for {
+		if cur == nil {
+			return
+		}
+
+		passed := true
+		for _, filter := range filters {
+			if !filter(cur) {
+				passed = false
+				break
+			}
+		}
+		if passed {
+			res = append(res, cur)
+		}
+		cur = cur.nextLocal
+	}
+
+}
diff --git a/metropolis/node/core/logtree/journal_entry.go b/metropolis/node/core/logtree/journal_entry.go
new file mode 100644
index 0000000..61619b3
--- /dev/null
+++ b/metropolis/node/core/logtree/journal_entry.go
@@ -0,0 +1,169 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import "git.monogon.dev/source/nexantic.git/metropolis/node/common/logbuffer"
+
+// entry is a journal entry, representing a single log event (encompassed in a Payload) at a given DN.
+// See the journal struct for more information about the global/local linked lists.
+type entry struct {
+	// origin is the DN at which the log entry was recorded, or conversely, in which DN it will be available at.
+	origin DN
+	// journal is the parent journal of this entry. An entry can belong only to a single journal. This pointer is used
+	// to mutate the journal's head/tail pointers when unlinking an entry.
+	journal *journal
+	// leveled is the leveled log entry for this entry, if this log entry was emitted by leveled logging. Otherwise it
+	// is nil.
+	leveled *LeveledPayload
+	// raw is the raw log entry for this entry, if this log entry was emitted by raw logging. Otherwise it is nil.
+	raw *logbuffer.Line
+
+	// prevGlobal is the previous entry in the global linked list, or nil if this entry is the oldest entry in the
+	// global linked list.
+	prevGlobal *entry
+	// nextGlobal is the next entry in the global linked list, or nil if this entry is the newest entry in the global
+	// linked list.
+	nextGlobal *entry
+
+	// prevLocal is the previous entry in this entry DN's local linked list, or nil if this entry is the oldest entry in
+	// this local linked list.
+	prevLocal *entry
+	// prevLocal is the next entry in this entry DN's local linked list, or nil if this entry is the newest entry in
+	// this local linked list.
+	nextLocal *entry
+
+	// seqLocal is a counter within a local linked list that increases by one each time a new log entry is added. It is
+	// used to quickly establish local linked list sizes (by subtracting seqLocal from both ends). This setup allows for
+	// O(1) length calculation for local linked lists as long as entries are only unlinked from the head or tail (which
+	// is the case in the current implementation).
+	seqLocal uint64
+}
+
+// external returns a LogEntry object for this entry, ie. the public version of this object, without fields relating to
+// the parent journal, linked lists, sequences, etc. These objects are visible to library consumers.
+func (e *entry) external() *LogEntry {
+	return &LogEntry{
+		DN:      e.origin,
+		Leveled: e.leveled,
+		Raw:     e.raw,
+	}
+}
+
+// unlink removes this entry from both global and local linked lists, updating the journal's head/tail pointers if
+// needed.
+// journal.mu must be taken as RW
+func (e *entry) unlink() {
+	// Unlink from the global linked list.
+	if e.prevGlobal != nil {
+		e.prevGlobal.nextGlobal = e.nextGlobal
+	}
+	if e.nextGlobal != nil {
+		e.nextGlobal.prevGlobal = e.prevGlobal
+	}
+	// Update journal head/tail pointers.
+	if e.journal.head == e {
+		e.journal.head = e.prevGlobal
+	}
+	if e.journal.tail == e {
+		e.journal.tail = e.nextGlobal
+	}
+
+	// Unlink from the local linked list.
+	if e.prevLocal != nil {
+		e.prevLocal.nextLocal = e.nextLocal
+	}
+	if e.nextLocal != nil {
+		e.nextLocal.prevLocal = e.prevLocal
+	}
+	// Update journal head/tail pointers.
+	if e.journal.heads[e.origin] == e {
+		e.journal.heads[e.origin] = e.prevLocal
+	}
+	if e.journal.tails[e.origin] == e {
+		e.journal.tails[e.origin] = e.nextLocal
+	}
+}
+
+// quota describes the quota policy for logging at a given DN.
+type quota struct {
+	// origin is the exact DN that this quota applies to.
+	origin DN
+	// max is the maximum count of log entries permitted for this DN - ie, the maximum size of the local linked list.
+	max uint64
+}
+
+// append adds an entry at the head of the global and local linked lists.
+func (j *journal) append(e *entry) {
+	j.mu.Lock()
+	defer j.mu.Unlock()
+
+	e.journal = j
+
+	// Insert at head in global linked list, set pointers.
+	e.nextGlobal = nil
+	e.prevGlobal = j.head
+	if j.head != nil {
+		j.head.nextGlobal = e
+	}
+	j.head = e
+	if j.tail == nil {
+		j.tail = e
+	}
+
+	// Create quota if necessary.
+	if _, ok := j.quota[e.origin]; !ok {
+		j.quota[e.origin] = &quota{origin: e.origin, max: 8192}
+	}
+
+	// Insert at head in local linked list, calculate seqLocal, set pointers.
+	e.nextLocal = nil
+	e.prevLocal = j.heads[e.origin]
+	if j.heads[e.origin] != nil {
+		j.heads[e.origin].nextLocal = e
+		e.seqLocal = e.prevLocal.seqLocal + 1
+	} else {
+		e.seqLocal = 0
+	}
+	j.heads[e.origin] = e
+	if j.tails[e.origin] == nil {
+		j.tails[e.origin] = e
+	}
+
+	// Apply quota to the local linked list that this entry got inserted to, ie. remove elements in excess of the
+	// quota.max count.
+	quota := j.quota[e.origin]
+	count := (j.heads[e.origin].seqLocal - j.tails[e.origin].seqLocal) + 1
+	if count > quota.max {
+		// Keep popping elements off the tail of the local linked list until quota is not violated.
+		left := count - quota.max
+		cur := j.tails[e.origin]
+		for {
+			// This shouldn't happen if quota.max >= 1.
+			if cur == nil {
+				break
+			}
+			if left == 0 {
+				break
+			}
+			el := cur
+			cur = el.nextLocal
+			// Unlinking the entry unlinks it from both the global and local linked lists.
+			el.unlink()
+			left -= 1
+		}
+	}
+}
diff --git a/metropolis/node/core/logtree/journal_subscriber.go b/metropolis/node/core/logtree/journal_subscriber.go
new file mode 100644
index 0000000..e6c7c62
--- /dev/null
+++ b/metropolis/node/core/logtree/journal_subscriber.go
@@ -0,0 +1,69 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"sync/atomic"
+)
+
+// subscriber is an observer for new entries that are appended to the journal.
+type subscriber struct {
+	// filters that entries need to pass through in order to be sent to the subscriber.
+	filters []filter
+	// dataC is the channel to which entries that pass filters will be sent. The channel must be drained regularly in
+	// order to prevent accumulation of goroutines and possible reordering of messages.
+	dataC chan *LogEntry
+	// doneC is a channel that is closed once the subscriber wishes to stop receiving notifications.
+	doneC chan struct{}
+	// missed is the amount of messages missed by the subscriber by not receiving from dataC fast enough
+	missed uint64
+}
+
+// subscribe attaches a subscriber to the journal.
+// mu must be taken in W mode
+func (j *journal) subscribe(sub *subscriber) {
+	j.subscribers = append(j.subscribers, sub)
+}
+
+// notify sends an entry to all subscribers that wish to receive it.
+func (j *journal) notify(e *entry) {
+	j.mu.Lock()
+	defer j.mu.Unlock()
+
+	newSub := make([]*subscriber, 0, len(j.subscribers))
+	for _, sub := range j.subscribers {
+		select {
+		case <-sub.doneC:
+			close(sub.dataC)
+			continue
+		default:
+			newSub = append(newSub, sub)
+		}
+
+		for _, filter := range sub.filters {
+			if !filter(e) {
+				continue
+			}
+		}
+		select {
+		case sub.dataC <- e.external():
+		default:
+			atomic.AddUint64(&sub.missed, 1)
+		}
+	}
+	j.subscribers = newSub
+}
diff --git a/metropolis/node/core/logtree/journal_test.go b/metropolis/node/core/logtree/journal_test.go
new file mode 100644
index 0000000..474748a
--- /dev/null
+++ b/metropolis/node/core/logtree/journal_test.go
@@ -0,0 +1,148 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+)
+
+func testPayload(msg string) *LeveledPayload {
+	return &LeveledPayload{
+		messages:  []string{msg},
+		timestamp: time.Now(),
+		severity:  INFO,
+		file:      "main.go",
+		line:      1337,
+	}
+}
+
+func TestJournalRetention(t *testing.T) {
+	j := newJournal()
+
+	for i := 0; i < 9000; i += 1 {
+		e := &entry{
+			origin:  "main",
+			leveled: testPayload(fmt.Sprintf("test %d", i)),
+		}
+		j.append(e)
+	}
+
+	entries := j.getEntries("main")
+	if want, got := 8192, len(entries); want != got {
+		t.Fatalf("wanted %d entries, got %d", want, got)
+	}
+	for i, entry := range entries {
+		want := fmt.Sprintf("test %d", (9000-8192)+i)
+		got := strings.Join(entry.leveled.messages, "\n")
+		if want != got {
+			t.Fatalf("wanted entry %q, got %q", want, got)
+		}
+	}
+}
+
+func TestJournalQuota(t *testing.T) {
+	j := newJournal()
+
+	for i := 0; i < 9000; i += 1 {
+		j.append(&entry{
+			origin:  "chatty",
+			leveled: testPayload(fmt.Sprintf("chatty %d", i)),
+		})
+		if i%10 == 0 {
+			j.append(&entry{
+				origin:  "solemn",
+				leveled: testPayload(fmt.Sprintf("solemn %d", i)),
+			})
+		}
+	}
+
+	entries := j.getEntries("chatty")
+	if want, got := 8192, len(entries); want != got {
+		t.Fatalf("wanted %d chatty entries, got %d", want, got)
+	}
+	entries = j.getEntries("solemn")
+	if want, got := 900, len(entries); want != got {
+		t.Fatalf("wanted %d solemn entries, got %d", want, got)
+	}
+	entries = j.getEntries("absent")
+	if want, got := 0, len(entries); want != got {
+		t.Fatalf("wanted %d absent entries, got %d", want, got)
+	}
+
+	entries = j.scanEntries(filterAll())
+	if want, got := 8192+900, len(entries); want != got {
+		t.Fatalf("wanted %d total entries, got %d", want, got)
+	}
+	setMessages := make(map[string]bool)
+	for _, entry := range entries {
+		setMessages[strings.Join(entry.leveled.messages, "\n")] = true
+	}
+
+	for i := 0; i < 900; i += 1 {
+		want := fmt.Sprintf("solemn %d", i*10)
+		if !setMessages[want] {
+			t.Fatalf("could not find entry %q in journal", want)
+		}
+	}
+	for i := 0; i < 8192; i += 1 {
+		want := fmt.Sprintf("chatty %d", i+(9000-8192))
+		if !setMessages[want] {
+			t.Fatalf("could not find entry %q in journal", want)
+		}
+	}
+}
+
+func TestJournalSubtree(t *testing.T) {
+	j := newJournal()
+	j.append(&entry{origin: "a", leveled: testPayload("a")})
+	j.append(&entry{origin: "a.b", leveled: testPayload("a.b")})
+	j.append(&entry{origin: "a.b.c", leveled: testPayload("a.b.c")})
+	j.append(&entry{origin: "a.b.d", leveled: testPayload("a.b.d")})
+	j.append(&entry{origin: "e.f", leveled: testPayload("e.f")})
+	j.append(&entry{origin: "e.g", leveled: testPayload("e.g")})
+
+	expect := func(f filter, msgs ...string) string {
+		res := j.scanEntries(f)
+		set := make(map[string]bool)
+		for _, entry := range res {
+			set[strings.Join(entry.leveled.messages, "\n")] = true
+		}
+
+		for _, want := range msgs {
+			if !set[want] {
+				return fmt.Sprintf("missing entry %q", want)
+			}
+		}
+		return ""
+	}
+
+	if res := expect(filterAll(), "a", "a.b", "a.b.c", "a.b.d", "e.f", "e.g"); res != "" {
+		t.Fatalf("All: %s", res)
+	}
+	if res := expect(filterSubtree("a"), "a", "a.b", "a.b.c", "a.b.d"); res != "" {
+		t.Fatalf("Subtree(a): %s", res)
+	}
+	if res := expect(filterSubtree("a.b"), "a.b", "a.b.c", "a.b.d"); res != "" {
+		t.Fatalf("Subtree(a.b): %s", res)
+	}
+	if res := expect(filterSubtree("e"), "e.f", "e.g"); res != "" {
+		t.Fatalf("Subtree(a.b): %s", res)
+	}
+}
diff --git a/metropolis/node/core/logtree/leveled.go b/metropolis/node/core/logtree/leveled.go
new file mode 100644
index 0000000..c24357e
--- /dev/null
+++ b/metropolis/node/core/logtree/leveled.go
@@ -0,0 +1,144 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"fmt"
+
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+// LeveledLogger is a generic interface for glog-style logging. There are four hardcoded log severities, in increasing
+// order: INFO, WARNING, ERROR, FATAL. Logging at a certain severity level logs not only to consumers expecting data
+// at that severity level, but also all lower severity levels. For example, an ERROR log will also be passed to
+// consumers looking at INFO or WARNING logs.
+type LeveledLogger interface {
+	// Info logs at the INFO severity. Arguments are handled in the manner of fmt.Print, a terminating newline is added
+	// if missing.
+	Info(args ...interface{})
+	// Infof logs at the INFO severity. Arguments are handled in the manner of fmt.Printf, a terminating newline is
+	// added if missing.
+	Infof(format string, args ...interface{})
+
+	// Warning logs at the WARNING severity. Arguments are handled in the manner of fmt.Print, a terminating newline is
+	// added if missing.
+	Warning(args ...interface{})
+	// Warningf logs at the WARNING severity. Arguments are handled in the manner of fmt.Printf, a terminating newline
+	// is added if missing.
+	Warningf(format string, args ...interface{})
+
+	// Error logs at the ERROR severity. Arguments are handled in the manner of fmt.Print, a terminating newline is
+	// added if missing.
+	Error(args ...interface{})
+	// Errorf logs at the ERROR severity. Arguments are handled in the manner of fmt.Printf, a terminating newline is
+	// added if missing.
+	Errorf(format string, args ...interface{})
+
+	// Fatal logs at the FATAL severity and aborts the current program. Arguments are handled in the manner of
+	// fmt.Print, a terminating newline is added if missing.
+	Fatal(args ...interface{})
+	// Fatalf logs at the FATAL severity and aborts the current program. Arguments are handled in the manner of
+	// fmt.Printf, a terminating newline is added if missing.
+	Fatalf(format string, args ...interface{})
+
+	// V returns a VerboseLeveledLogger at a given verbosity level. These verbosity levels can be dynamically set and
+	// unset on a package-granular level by consumers of the LeveledLogger logs. The returned value represents whether
+	// logging at the given verbosity level was active at that time, and as such should not be a long-lived object
+	// in programs.
+	// This construct is further refered to as 'V-logs'.
+	V(level VerbosityLevel) VerboseLeveledLogger
+}
+
+// VerbosityLevel is a verbosity level defined for V-logs. This can be changed programmatically per Go package. When
+// logging at a given VerbosityLevel V, the current level must be equal or higher to V for the logs to be recorded.
+// Conversely, enabling a V-logging at a VerbosityLevel V also enables all logging at lower levels [Int32Min .. (V-1)].
+type VerbosityLevel int32
+
+type VerboseLeveledLogger interface {
+	// Enabled returns if this level was enabled. If not enabled, all logging into this logger will be discarded
+	// immediately.
+	// Thus, Enabled() can be used to check the verbosity level before performing any logging:
+	//    if l.V(3).Enabled() { l.Info("V3 is enabled") }
+	// or, in simple cases, the convenience function .Info can be used:
+	//    l.V(3).Info("V3 is enabled")
+	// The second form is shorter and more convenient, but more expensive, as its arguments are always evaluated.
+	Enabled() bool
+	// Info is the equivalent of a LeveledLogger's Info call, guarded by whether this VerboseLeveledLogger is enabled.
+	Info(args ...interface{})
+	// Infof is the equivalent of a LeveledLogger's Infof call, guarded by whether this VerboseLeveledLogger is enabled.
+	Infof(format string, args ...interface{})
+}
+
+// Severity is one of the severities as described in LeveledLogger.
+type Severity string
+
+const (
+	INFO    Severity = "I"
+	WARNING Severity = "W"
+	ERROR   Severity = "E"
+	FATAL   Severity = "F"
+)
+
+var (
+	// SeverityAtLeast maps a given severity to a list of severities that at that severity or higher. In other words,
+	// SeverityAtLeast[X] returns a list of severities that might be seen in a log at severity X.
+	SeverityAtLeast = map[Severity][]Severity{
+		INFO:    {INFO, WARNING, ERROR, FATAL},
+		WARNING: {WARNING, ERROR, FATAL},
+		ERROR:   {ERROR, FATAL},
+		FATAL:   {FATAL},
+	}
+)
+
+func (s Severity) AtLeast(other Severity) bool {
+	for _, el := range SeverityAtLeast[other] {
+		if el == s {
+			return true
+		}
+	}
+	return false
+}
+
+func SeverityFromProto(s apb.LeveledLogSeverity) (Severity, error) {
+	switch s {
+	case apb.LeveledLogSeverity_INFO:
+		return INFO, nil
+	case apb.LeveledLogSeverity_WARNING:
+		return WARNING, nil
+	case apb.LeveledLogSeverity_ERROR:
+		return ERROR, nil
+	case apb.LeveledLogSeverity_FATAL:
+		return FATAL, nil
+	default:
+		return "", fmt.Errorf("unknown severity value %d", s)
+	}
+}
+
+func (s Severity) ToProto() apb.LeveledLogSeverity {
+	switch s {
+	case INFO:
+		return apb.LeveledLogSeverity_INFO
+	case WARNING:
+		return apb.LeveledLogSeverity_WARNING
+	case ERROR:
+		return apb.LeveledLogSeverity_ERROR
+	case FATAL:
+		return apb.LeveledLogSeverity_FATAL
+	default:
+		return apb.LeveledLogSeverity_INVALID
+	}
+}
diff --git a/metropolis/node/core/logtree/leveled_payload.go b/metropolis/node/core/logtree/leveled_payload.go
new file mode 100644
index 0000000..fad42e3
--- /dev/null
+++ b/metropolis/node/core/logtree/leveled_payload.go
@@ -0,0 +1,142 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+	"time"
+
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+// LeveledPayload is a log entry for leveled logs (as per leveled.go). It contains the input to these calls (severity and
+// message split into newline-delimited messages) and additional metadata that would be usually seen in a text
+// representation of a leveled log entry.
+type LeveledPayload struct {
+	// messages is the list of messages contained in this payload. This list is built from splitting up the given message
+	// from the user by newline.
+	messages []string
+	// timestamp is the time at which this message was emitted.
+	timestamp time.Time
+	// severity is the leveled Severity at which this message was emitted.
+	severity Severity
+	// file is the filename of the caller that emitted this message.
+	file string
+	// line is the line number within the file of the caller that emitted this message.
+	line int
+}
+
+// String returns a canonical representation of this payload as a single string prefixed with metadata. If the original
+// message was logged with newlines, this representation will also contain newlines, with each original message part
+// prefixed by the metadata.
+// For an alternative call that will instead return a canonical prefix and a list of lines in the message, see Strings().
+func (p *LeveledPayload) String() string {
+	prefix, lines := p.Strings()
+	res := make([]string, len(p.messages))
+	for i, line := range lines {
+		res[i] = fmt.Sprintf("%s%s", prefix, line)
+	}
+	return strings.Join(res, "\n")
+}
+
+// Strings returns the canonical representation of this payload split into a prefix and all lines that were contained in
+// the original message. This is meant to be displayed to the user by showing the prefix before each line, concatenated
+// together - possibly in a table form with the prefixes all unified with a rowspan-like mechanism.
+//
+// For example, this function can return:
+//   prefix = "I1102 17:20:06.921395 foo.go:42] "
+//   lines = []string{"current tags:", " - one", " - two"}
+//
+// With this data, the result should be presented to users this way in text form:
+// I1102 17:20:06.921395 foo.go:42] current tags:
+// I1102 17:20:06.921395 foo.go:42]  - one
+// I1102 17:20:06.921395 foo.go:42]  - two
+//
+// Or, in a table layout:
+// .-----------------------------------------------------------.
+// | I1102 17:20:06.921395     0 foo.go:42] : current tags:    |
+// |                                        :------------------|
+// |                                        :  - one           |
+// |                                        :------------------|
+// |                                        :  - two           |
+// '-----------------------------------------------------------'
+//
+func (p *LeveledPayload) Strings() (prefix string, lines []string) {
+	_, month, day := p.timestamp.Date()
+	hour, minute, second := p.timestamp.Clock()
+	nsec := p.timestamp.Nanosecond() / 1000
+
+	// Same format as in glog, but without treadid.
+	// Lmmdd hh:mm:ss.uuuuuu file:line]
+	// TODO(q3k): rewrite this to printf-less code.
+	prefix = fmt.Sprintf("%s%02d%02d %02d:%02d:%02d.%06d %s:%d] ", p.severity, month, day, hour, minute, second, nsec, p.file, p.line)
+
+	lines = p.messages
+	return
+}
+
+// Message returns the inner message lines of this entry, ie. what was passed to the actual logging method, but split by
+// newlines.
+func (p *LeveledPayload) Messages() []string { return p.messages }
+
+func (p *LeveledPayload) MessagesJoined() string { return strings.Join(p.messages, "\n") }
+
+// Timestamp returns the time at which this entry was logged.
+func (p *LeveledPayload) Timestamp() time.Time { return p.timestamp }
+
+// Location returns a string in the form of file_name:line_number that shows the origin of the log entry in the
+// program source.
+func (p *LeveledPayload) Location() string { return fmt.Sprintf("%s:%d", p.file, p.line) }
+
+// Severity returns the Severity with which this entry was logged.
+func (p *LeveledPayload) Severity() Severity { return p.severity }
+
+// Proto converts a LeveledPayload to protobuf format.
+func (p *LeveledPayload) Proto() *apb.LogEntry_Leveled {
+	return &apb.LogEntry_Leveled{
+		Lines:     p.Messages(),
+		Timestamp: p.Timestamp().UnixNano(),
+		Severity:  p.Severity().ToProto(),
+		Location:  p.Location(),
+	}
+}
+
+// LeveledPayloadFromProto parses a protobuf message into the internal format.
+func LeveledPayloadFromProto(p *apb.LogEntry_Leveled) (*LeveledPayload, error) {
+	severity, err := SeverityFromProto(p.Severity)
+	if err != nil {
+		return nil, fmt.Errorf("could not convert severity: %w", err)
+	}
+	parts := strings.Split(p.Location, ":")
+	if len(parts) != 2 {
+		return nil, fmt.Errorf("invalid location, must be two :-delimited parts, is %d parts", len(parts))
+	}
+	file := parts[0]
+	line, err := strconv.Atoi(parts[1])
+	if err != nil {
+		return nil, fmt.Errorf("invalid location line number: %w", err)
+	}
+	return &LeveledPayload{
+		messages:  p.Lines,
+		timestamp: time.Unix(0, p.Timestamp),
+		severity:  severity,
+		file:      file,
+		line:      line,
+	}, nil
+}
diff --git a/metropolis/node/core/logtree/logtree.go b/metropolis/node/core/logtree/logtree.go
new file mode 100644
index 0000000..fab72ba
--- /dev/null
+++ b/metropolis/node/core/logtree/logtree.go
@@ -0,0 +1,147 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"fmt"
+	"strings"
+	"sync"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/logbuffer"
+)
+
+// LogTree is a tree-shaped logging system. For more information, see the package-level documentation.
+type LogTree struct {
+	// journal is the tree's journal, storing all log data and managing subscribers.
+	journal *journal
+	// root is the root node of the actual tree of the log tree. The nodes contain per-DN configuration options, notably
+	// the current verbosity level of that DN.
+	root *node
+}
+
+func New() *LogTree {
+	lt := &LogTree{
+		journal: newJournal(),
+	}
+	lt.root = newNode(lt, "")
+	return lt
+}
+
+// node represents a given DN as a discrete 'logger'. It implements the LeveledLogger interface for log publishing,
+// entries from which it passes over to the logtree's journal.
+type node struct {
+	// dn is the DN which this node represents (or "" if this is the root node).
+	dn DN
+	// tree is the LogTree to which this node belongs.
+	tree *LogTree
+	// verbosity is the current verbosity level of this DN/node, affecting .V(n) LeveledLogger calls
+	verbosity     VerbosityLevel
+	rawLineBuffer *logbuffer.LineBuffer
+
+	// mu guards children.
+	mu sync.Mutex
+	// children is a map of DN-part to a children node in the logtree. A DN-part is a string representing a part of the
+	// DN between the deliming dots, as returned by DN.Path.
+	children map[string]*node
+}
+
+// newNode returns a node at a given DN in the LogTree - but doesn't set up the LogTree to insert it accordingly.
+func newNode(tree *LogTree, dn DN) *node {
+	n := &node{
+		dn:       dn,
+		tree:     tree,
+		children: make(map[string]*node),
+	}
+	// TODO(q3k): make this limit configurable. If this happens, or the default (1024) gets changes, max chunk size
+	// calculations when serving the logs (eg. in NodeDebugService) must reflect this.
+	n.rawLineBuffer = logbuffer.NewLineBuffer(1024, n.logRaw)
+	return n
+}
+
+// nodeByDN returns the LogTree node corresponding to a given DN. If either the node or some of its parents do not
+// exist they will be created as needed.
+func (l *LogTree) nodeByDN(dn DN) (*node, error) {
+	traversal, err := newTraversal(dn)
+	if err != nil {
+		return nil, fmt.Errorf("traversal failed: %w", err)
+	}
+	return traversal.execute(l.root), nil
+}
+
+// nodeTraversal represents a request to traverse the LogTree in search of a given node by DN.
+type nodeTraversal struct {
+	// want is the DN of the node's that requested to be found.
+	want DN
+	// current is the path already taken to find the node, in the form of DN parts. It starts out as want.Parts() and
+	// progresses to become empty as the traversal continues.
+	current []string
+	// left is the path that's still needed to be taken in order to find the node, in the form of DN parts. It starts
+	// out empty and progresses to become wants.Parts() as the traversal continues.
+	left []string
+}
+
+// next adjusts the traversal's current/left slices to the next element of the traversal, returns the part that's now
+// being looked for (or "" if the traveral is done) and the full DN of the element that's being looked for.
+//
+// For example, a traversal of foo.bar.baz will cause .next() to return the following on each invocation:
+//  - part: foo, full: foo
+//  - part: bar, full: foo.bar
+//  - part: baz, full: foo.bar.baz
+//  - part: "",  full: foo.bar.baz
+func (t *nodeTraversal) next() (part string, full DN) {
+	if len(t.left) == 0 {
+		return "", t.want
+	}
+	part = t.left[0]
+	t.current = append(t.current, part)
+	t.left = t.left[1:]
+	full = DN(strings.Join(t.current, "."))
+	return
+}
+
+// newTraversal returns a nodeTraversal fora a given wanted DN.
+func newTraversal(dn DN) (*nodeTraversal, error) {
+	parts, err := dn.Path()
+	if err != nil {
+		return nil, err
+	}
+	return &nodeTraversal{
+		want: dn,
+		left: parts,
+	}, nil
+}
+
+// execute the traversal in order to find the node. This can only be called once per traversal.
+// Nodes will be created within the tree until the target node is reached. Existing nodes will be reused.
+// This is effectively an idempotent way of accessing a node in the tree based on a traversal.
+func (t *nodeTraversal) execute(n *node) *node {
+	cur := n
+	for {
+		part, full := t.next()
+		if part == "" {
+			return cur
+		}
+
+		mu := &cur.mu
+		mu.Lock()
+		if _, ok := cur.children[part]; !ok {
+			cur.children[part] = newNode(n.tree, DN(full))
+		}
+		cur = cur.children[part]
+		mu.Unlock()
+	}
+}
diff --git a/metropolis/node/core/logtree/logtree_access.go b/metropolis/node/core/logtree/logtree_access.go
new file mode 100644
index 0000000..fed202e
--- /dev/null
+++ b/metropolis/node/core/logtree/logtree_access.go
@@ -0,0 +1,183 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"errors"
+	"sync/atomic"
+)
+
+// LogReadOption describes options for the LogTree.Read call.
+type LogReadOption struct {
+	withChildren               bool
+	withStream                 bool
+	withBacklog                int
+	onlyLeveled                bool
+	onlyRaw                    bool
+	leveledWithMinimumSeverity Severity
+}
+
+// WithChildren makes Read return/stream data for both a given DN and all its children.
+func WithChildren() LogReadOption { return LogReadOption{withChildren: true} }
+
+// WithStream makes Read return a stream of data. This works alongside WithBacklog to create a read-and-stream
+// construct.
+func WithStream() LogReadOption { return LogReadOption{withStream: true} }
+
+// WithBacklog makes Read return already recorded log entries, up to count elements.
+func WithBacklog(count int) LogReadOption { return LogReadOption{withBacklog: count} }
+
+// BacklogAllAvailable makes WithBacklog return all backlogged log data that logtree possesses.
+const BacklogAllAvailable int = -1
+
+func OnlyRaw() LogReadOption { return LogReadOption{onlyRaw: true} }
+
+func OnlyLeveled() LogReadOption { return LogReadOption{onlyLeveled: true} }
+
+// LeveledWithMinimumSeverity makes Read return only log entries that are at least at a given Severity. If only leveled
+// entries are needed, OnlyLeveled must be used. This is a no-op when OnlyRaw is used.
+func LeveledWithMinimumSeverity(s Severity) LogReadOption {
+	return LogReadOption{leveledWithMinimumSeverity: s}
+}
+
+// LogReader permits reading an already existing backlog of log entries and to stream further ones.
+type LogReader struct {
+	// Backlog are the log entries already logged by LogTree. This will only be set if WithBacklog has been passed to
+	// Read.
+	Backlog []*LogEntry
+	// Stream is a channel of new entries as received live by LogTree. This will only be set if WithStream has been
+	// passed to Read. In this case, entries from this channel must be read as fast as possible by the consumer in order
+	// to prevent missing entries.
+	Stream <-chan *LogEntry
+	// done is channel used to signal (by closing) that the log consumer is not interested in more Stream data.
+	done chan<- struct{}
+	// missed is an atomic integer pointer that tells the subscriber how many messages in Stream they missed. This
+	// pointer is nil if no streaming has been requested.
+	missed *uint64
+}
+
+// Missed returns the amount of entries that were missed from Stream (as the channel was not drained fast enough).
+func (l *LogReader) Missed() uint64 {
+	// No Stream.
+	if l.missed == nil {
+		return 0
+	}
+	return atomic.LoadUint64(l.missed)
+}
+
+// Close closes the LogReader's Stream. This must be called once the Reader does not wish to receive streaming messages
+// anymore.
+func (l *LogReader) Close() {
+	if l.done != nil {
+		close(l.done)
+	}
+}
+
+var (
+	ErrRawAndLeveled = errors.New("cannot return logs that are simultaneously OnlyRaw and OnlyLeveled")
+)
+
+// Read and/or stream entries from a LogTree. The returned LogReader is influenced by the LogReadOptions passed, which
+// influence whether the Read will return existing entries, a stream, or both. In addition the options also dictate
+// whether only entries for that particular DN are returned, or for all sub-DNs as well.
+func (l *LogTree) Read(dn DN, opts ...LogReadOption) (*LogReader, error) {
+	l.journal.mu.RLock()
+	defer l.journal.mu.RUnlock()
+
+	var backlog int
+	var stream bool
+	var recursive bool
+	var leveledSeverity Severity
+	var onlyRaw, onlyLeveled bool
+
+	for _, opt := range opts {
+		if opt.withBacklog > 0 || opt.withBacklog == BacklogAllAvailable {
+			backlog = opt.withBacklog
+		}
+		if opt.withStream {
+			stream = true
+		}
+		if opt.withChildren {
+			recursive = true
+		}
+		if opt.leveledWithMinimumSeverity != "" {
+			leveledSeverity = opt.leveledWithMinimumSeverity
+		}
+		if opt.onlyLeveled {
+			onlyLeveled = true
+		}
+		if opt.onlyRaw {
+			onlyRaw = true
+		}
+	}
+
+	if onlyLeveled && onlyRaw {
+		return nil, ErrRawAndLeveled
+	}
+
+	var filters []filter
+	if onlyLeveled {
+		filters = append(filters, filterOnlyLeveled)
+	}
+	if onlyRaw {
+		filters = append(filters, filterOnlyRaw)
+	}
+	if recursive {
+		filters = append(filters, filterSubtree(dn))
+	} else {
+		filters = append(filters, filterExact(dn))
+	}
+	if leveledSeverity != "" {
+		filters = append(filters, filterSeverity(leveledSeverity))
+	}
+
+	var entries []*entry
+	if backlog > 0 || backlog == BacklogAllAvailable {
+		// TODO(q3k): pass over the backlog count to scanEntries/getEntries, instead of discarding them here.
+		if recursive {
+			entries = l.journal.scanEntries(filters...)
+		} else {
+			entries = l.journal.getEntries(dn, filters...)
+		}
+		if backlog != BacklogAllAvailable && len(entries) > backlog {
+			entries = entries[:backlog]
+		}
+	}
+
+	var sub *subscriber
+	if stream {
+		sub = &subscriber{
+			// TODO(q3k): make buffer size configurable
+			dataC:   make(chan *LogEntry, 128),
+			doneC:   make(chan struct{}),
+			filters: filters,
+		}
+		l.journal.subscribe(sub)
+	}
+
+	lr := &LogReader{}
+	lr.Backlog = make([]*LogEntry, len(entries))
+	for i, entry := range entries {
+		lr.Backlog[i] = entry.external()
+	}
+	if stream {
+		lr.Stream = sub.dataC
+		lr.done = sub.doneC
+		lr.missed = &sub.missed
+	}
+	return lr, nil
+}
diff --git a/metropolis/node/core/logtree/logtree_entry.go b/metropolis/node/core/logtree/logtree_entry.go
new file mode 100644
index 0000000..635e5a8
--- /dev/null
+++ b/metropolis/node/core/logtree/logtree_entry.go
@@ -0,0 +1,141 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"fmt"
+	"strings"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/logbuffer"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+// LogEntry contains a log entry, combining both leveled and raw logging into a single stream of events. A LogEntry
+// will contain exactly one of either LeveledPayload or RawPayload.
+type LogEntry struct {
+	// If non-nil, this is a leveled logging entry.
+	Leveled *LeveledPayload
+	// If non-nil, this is a raw logging entry line.
+	Raw *logbuffer.Line
+	// DN from which this entry was logged.
+	DN DN
+}
+
+// String returns a canonical representation of this payload as a single string prefixed with metadata. If the entry is
+// a leveled log entry that originally was logged with newlines this representation will also contain newlines, with
+// each original message part prefixed by the metadata.
+// For an alternative call that will instead return a canonical prefix and a list of lines in the message, see Strings().
+func (l *LogEntry) String() string {
+	if l.Leveled != nil {
+		prefix, messages := l.Leveled.Strings()
+		res := make([]string, len(messages))
+		for i, m := range messages {
+			res[i] = fmt.Sprintf("%-32s %s%s", l.DN, prefix, m)
+		}
+		return strings.Join(res, "\n")
+	}
+	if l.Raw != nil {
+		return fmt.Sprintf("%-32s R %s", l.DN, l.Raw)
+	}
+	return "INVALID"
+}
+
+// Strings returns the canonical representation of this payload split into a prefix and all lines that were contained in
+// the original message. This is meant to be displayed to the user by showing the prefix before each line, concatenated
+// together - possibly in a table form with the prefixes all unified with a rowspan-like mechanism.
+//
+// For example, this function can return:
+//   prefix = "root.foo.bar                    I1102 17:20:06.921395     0 foo.go:42] "
+//   lines = []string{"current tags:", " - one", " - two"}
+//
+// With this data, the result should be presented to users this way in text form:
+// root.foo.bar                    I1102 17:20:06.921395 foo.go:42] current tags:
+// root.foo.bar                    I1102 17:20:06.921395 foo.go:42]  - one
+// root.foo.bar                    I1102 17:20:06.921395 foo.go:42]  - two
+//
+// Or, in a table layout:
+// .-------------------------------------------------------------------------------------.
+// | root.foo.bar                    I1102 17:20:06.921395 foo.go:42] : current tags:    |
+// |                                                                  :------------------|
+// |                                                                  :  - one           |
+// |                                                                  :------------------|
+// |                                                                  :  - two           |
+// '-------------------------------------------------------------------------------------'
+//
+func (l *LogEntry) Strings() (prefix string, lines []string) {
+	if l.Leveled != nil {
+		prefix, messages := l.Leveled.Strings()
+		prefix = fmt.Sprintf("%-32s %s", l.DN, prefix)
+		return prefix, messages
+	}
+	if l.Raw != nil {
+		return fmt.Sprintf("%-32s R ", l.DN), []string{l.Raw.Data}
+	}
+	return "INVALID ", []string{"INVALID"}
+}
+
+// Convert this LogEntry to proto. Returned value may be nil if given LogEntry is invalid, eg. contains neither a Raw
+// nor Leveled entry.
+func (l *LogEntry) Proto() *apb.LogEntry {
+	p := &apb.LogEntry{
+		Dn: string(l.DN),
+	}
+	switch {
+	case l.Leveled != nil:
+		leveled := l.Leveled
+		p.Kind = &apb.LogEntry_Leveled_{
+			Leveled: leveled.Proto(),
+		}
+	case l.Raw != nil:
+		raw := l.Raw
+		p.Kind = &apb.LogEntry_Raw_{
+			Raw: raw.ProtoLog(),
+		}
+	default:
+		return nil
+	}
+	return p
+}
+
+// Parse a proto LogEntry back into internal structure. This can be used in log proto API consumers to easily print
+// received log entries.
+func LogEntryFromProto(l *apb.LogEntry) (*LogEntry, error) {
+	dn := DN(l.Dn)
+	if _, err := dn.Path(); err != nil {
+		return nil, fmt.Errorf("could not convert DN: %w", err)
+	}
+	res := &LogEntry{
+		DN: dn,
+	}
+	switch inner := l.Kind.(type) {
+	case *apb.LogEntry_Leveled_:
+		leveled, err := LeveledPayloadFromProto(inner.Leveled)
+		if err != nil {
+			return nil, fmt.Errorf("could not convert leveled entry: %w", err)
+		}
+		res.Leveled = leveled
+	case *apb.LogEntry_Raw_:
+		line, err := logbuffer.LineFromLogProto(inner.Raw)
+		if err != nil {
+			return nil, fmt.Errorf("could not convert raw entry: %w", err)
+		}
+		res.Raw = line
+	default:
+		return nil, fmt.Errorf("proto has neither Leveled nor Raw set")
+	}
+	return res, nil
+}
diff --git a/metropolis/node/core/logtree/logtree_publisher.go b/metropolis/node/core/logtree/logtree_publisher.go
new file mode 100644
index 0000000..c4880bc
--- /dev/null
+++ b/metropolis/node/core/logtree/logtree_publisher.go
@@ -0,0 +1,185 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"fmt"
+	"io"
+	"runtime"
+	"strings"
+	"time"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/logbuffer"
+)
+
+// LeveledFor returns a LeveledLogger publishing interface for a given DN. An error may be returned if the DN is
+// malformed.
+func (l *LogTree) LeveledFor(dn DN) (LeveledLogger, error) {
+	return l.nodeByDN(dn)
+}
+
+func (l *LogTree) RawFor(dn DN) (io.Writer, error) {
+	node, err := l.nodeByDN(dn)
+	if err != nil {
+		return nil, fmt.Errorf("could not retrieve raw logger: %w", err)
+	}
+	return node.rawLineBuffer, nil
+}
+
+// MustLeveledFor returns a LeveledLogger publishing interface for a given DN, or panics if the given DN is invalid.
+func (l *LogTree) MustLeveledFor(dn DN) LeveledLogger {
+	leveled, err := l.LeveledFor(dn)
+	if err != nil {
+		panic(fmt.Errorf("LeveledFor returned: %w", err))
+	}
+	return leveled
+}
+
+func (l *LogTree) MustRawFor(dn DN) io.Writer {
+	raw, err := l.RawFor(dn)
+	if err != nil {
+		panic(fmt.Errorf("RawFor returned: %w", err))
+	}
+	return raw
+}
+
+// SetVerbosity sets the verbosity for a given DN (non-recursively, ie. for that DN only, not its children).
+func (l *LogTree) SetVerbosity(dn DN, level VerbosityLevel) error {
+	node, err := l.nodeByDN(dn)
+	if err != nil {
+		return err
+	}
+	node.verbosity = level
+	return nil
+}
+
+// logRaw is called by this node's LineBuffer any time a raw log line is completed. It will create a new entry, append
+// it to the journal, and notify all pertinent subscribers.
+func (n *node) logRaw(line *logbuffer.Line) {
+	e := &entry{
+		origin: n.dn,
+		raw:    line,
+	}
+	n.tree.journal.append(e)
+	n.tree.journal.notify(e)
+}
+
+// log builds a LeveledPayload and entry for a given message, including all related metadata. It will create a new
+// entry append it to the journal, and notify all pertinent subscribers.
+func (n *node) logLeveled(depth int, severity Severity, msg string) {
+	_, file, line, ok := runtime.Caller(2 + depth)
+	if !ok {
+		file = "???"
+		line = 1
+	} else {
+		slash := strings.LastIndex(file, "/")
+		if slash >= 0 {
+			file = file[slash+1:]
+		}
+	}
+
+	// Remove leading/trailing newlines and split.
+	messages := strings.Split(strings.Trim(msg, "\n"), "\n")
+
+	p := &LeveledPayload{
+		timestamp: time.Now(),
+		severity:  severity,
+		messages:  messages,
+		file:      file,
+		line:      line,
+	}
+	e := &entry{
+		origin:  n.dn,
+		leveled: p,
+	}
+	n.tree.journal.append(e)
+	n.tree.journal.notify(e)
+}
+
+// Info implements the LeveledLogger interface.
+func (n *node) Info(args ...interface{}) {
+	n.logLeveled(0, INFO, fmt.Sprint(args...))
+}
+
+// Infof implements the LeveledLogger interface.
+func (n *node) Infof(format string, args ...interface{}) {
+	n.logLeveled(0, INFO, fmt.Sprintf(format, args...))
+}
+
+// Warning implements the LeveledLogger interface.
+func (n *node) Warning(args ...interface{}) {
+	n.logLeveled(0, WARNING, fmt.Sprint(args...))
+}
+
+// Warningf implements the LeveledLogger interface.
+func (n *node) Warningf(format string, args ...interface{}) {
+	n.logLeveled(0, WARNING, fmt.Sprintf(format, args...))
+}
+
+// Error implements the LeveledLogger interface.
+func (n *node) Error(args ...interface{}) {
+	n.logLeveled(0, ERROR, fmt.Sprint(args...))
+}
+
+// Errorf implements the LeveledLogger interface.
+func (n *node) Errorf(format string, args ...interface{}) {
+	n.logLeveled(0, ERROR, fmt.Sprintf(format, args...))
+}
+
+// Fatal implements the LeveledLogger interface.
+func (n *node) Fatal(args ...interface{}) {
+	n.logLeveled(0, FATAL, fmt.Sprint(args...))
+}
+
+// Fatalf implements the LeveledLogger interface.
+func (n *node) Fatalf(format string, args ...interface{}) {
+	n.logLeveled(0, FATAL, fmt.Sprintf(format, args...))
+}
+
+// V implements the LeveledLogger interface.
+func (n *node) V(v VerbosityLevel) VerboseLeveledLogger {
+	return &verbose{
+		node:    n,
+		enabled: n.verbosity >= v,
+	}
+}
+
+// verbose implements the VerboseLeveledLogger interface. It is a thin wrapper around node, with an 'enabled' bool. This
+// means that V(n)-returned VerboseLeveledLoggers must be short lived, as a changed in verbosity will not affect all
+// already existing VerboseLeveledLoggers.
+type verbose struct {
+	node    *node
+	enabled bool
+}
+
+func (v *verbose) Enabled() bool {
+	return v.enabled
+}
+
+func (v *verbose) Info(args ...interface{}) {
+	if !v.enabled {
+		return
+	}
+	v.node.logLeveled(0, INFO, fmt.Sprint(args...))
+}
+
+func (v *verbose) Infof(format string, args ...interface{}) {
+	if !v.enabled {
+		return
+	}
+	v.node.logLeveled(0, INFO, fmt.Sprintf(format, args...))
+}
diff --git a/metropolis/node/core/logtree/logtree_test.go b/metropolis/node/core/logtree/logtree_test.go
new file mode 100644
index 0000000..b900201
--- /dev/null
+++ b/metropolis/node/core/logtree/logtree_test.go
@@ -0,0 +1,211 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package logtree
+
+import (
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+)
+
+func expect(tree *LogTree, t *testing.T, dn DN, entries ...string) string {
+	t.Helper()
+	res, err := tree.Read(dn, WithChildren(), WithBacklog(BacklogAllAvailable))
+	if err != nil {
+		t.Fatalf("Read: %v", err)
+	}
+	if want, got := len(entries), len(res.Backlog); want != got {
+		t.Fatalf("wanted %v backlog entries, got %v", want, got)
+	}
+	got := make(map[string]bool)
+	for _, entry := range res.Backlog {
+		if entry.Leveled != nil {
+			got[entry.Leveled.MessagesJoined()] = true
+		}
+		if entry.Raw != nil {
+			got[entry.Raw.Data] = true
+		}
+	}
+	for _, entry := range entries {
+		if !got[entry] {
+			return fmt.Sprintf("missing entry %q", entry)
+		}
+	}
+	return ""
+}
+
+func TestMultiline(t *testing.T) {
+	tree := New()
+	// Two lines in a single message.
+	tree.MustLeveledFor("main").Info("foo\nbar")
+	// Two lines in a single message with a hanging newline that should get stripped.
+	tree.MustLeveledFor("main").Info("one\ntwo\n")
+
+	if res := expect(tree, t, "main", "foo\nbar", "one\ntwo"); res != "" {
+		t.Errorf("retrieval at main failed: %s", res)
+	}
+}
+
+func TestBacklog(t *testing.T) {
+	tree := New()
+	tree.MustLeveledFor("main").Info("hello, main!")
+	tree.MustLeveledFor("main.foo").Info("hello, main.foo!")
+	tree.MustLeveledFor("main.bar").Info("hello, main.bar!")
+	tree.MustLeveledFor("aux").Info("hello, aux!")
+	// No newline at the last entry - shouldn't get propagated to the backlog.
+	fmt.Fprintf(tree.MustRawFor("aux.process"), "processing foo\nprocessing bar\nbaz")
+
+	if res := expect(tree, t, "main", "hello, main!", "hello, main.foo!", "hello, main.bar!"); res != "" {
+		t.Errorf("retrieval at main failed: %s", res)
+	}
+	if res := expect(tree, t, "", "hello, main!", "hello, main.foo!", "hello, main.bar!", "hello, aux!", "processing foo", "processing bar"); res != "" {
+		t.Errorf("retrieval at root failed: %s", res)
+	}
+	if res := expect(tree, t, "aux", "hello, aux!", "processing foo", "processing bar"); res != "" {
+		t.Errorf("retrieval at aux failed: %s", res)
+	}
+}
+
+func TestStream(t *testing.T) {
+	tree := New()
+	tree.MustLeveledFor("main").Info("hello, backlog")
+	fmt.Fprintf(tree.MustRawFor("main.process"), "hello, raw backlog\n")
+
+	res, err := tree.Read("", WithBacklog(BacklogAllAvailable), WithChildren(), WithStream())
+	if err != nil {
+		t.Fatalf("Read: %v", err)
+	}
+	defer res.Close()
+	if want, got := 2, len(res.Backlog); want != got {
+		t.Errorf("wanted %d backlog item, got %d", want, got)
+	}
+
+	tree.MustLeveledFor("main").Info("hello, stream")
+	fmt.Fprintf(tree.MustRawFor("main.raw"), "hello, raw stream\n")
+
+	entries := make(map[string]bool)
+	timeout := time.After(time.Second * 1)
+	for {
+		done := false
+		select {
+		case <-timeout:
+			done = true
+		case p := <-res.Stream:
+			if p.Leveled != nil {
+				entries[p.Leveled.MessagesJoined()] = true
+			}
+			if p.Raw != nil {
+				entries[p.Raw.Data] = true
+			}
+		}
+		if done {
+			break
+		}
+	}
+	if entry := "hello, stream"; !entries[entry] {
+		t.Errorf("Missing entry %q", entry)
+	}
+	if entry := "hello, raw stream"; !entries[entry] {
+		t.Errorf("Missing entry %q", entry)
+	}
+}
+
+func TestVerbose(t *testing.T) {
+	tree := New()
+
+	tree.MustLeveledFor("main").V(10).Info("this shouldn't get logged")
+
+	reader, err := tree.Read("", WithBacklog(BacklogAllAvailable), WithChildren())
+	if err != nil {
+		t.Fatalf("Read: %v", err)
+	}
+	if want, got := 0, len(reader.Backlog); want != got {
+		t.Fatalf("expected nothing to be logged, got %+v", reader.Backlog)
+	}
+
+	tree.SetVerbosity("main", 10)
+	tree.MustLeveledFor("main").V(10).Info("this should get logged")
+
+	reader, err = tree.Read("", WithBacklog(BacklogAllAvailable), WithChildren())
+	if err != nil {
+		t.Fatalf("Read: %v", err)
+	}
+	if want, got := 1, len(reader.Backlog); want != got {
+		t.Fatalf("expected %d entries to get logged, got %d", want, got)
+	}
+}
+
+func TestMetadata(t *testing.T) {
+	tree := New()
+	tree.MustLeveledFor("main").Error("i am an error")
+	tree.MustLeveledFor("main").Warning("i am a warning")
+	tree.MustLeveledFor("main").Info("i am informative")
+	tree.MustLeveledFor("main").V(0).Info("i am a zero-level debug")
+
+	reader, err := tree.Read("", WithChildren(), WithBacklog(BacklogAllAvailable))
+	if err != nil {
+		t.Fatalf("Read: %v", err)
+	}
+	if want, got := 4, len(reader.Backlog); want != got {
+		t.Fatalf("expected %d entries, got %d", want, got)
+	}
+
+	for _, te := range []struct {
+		ix       int
+		severity Severity
+		message  string
+	}{
+		{0, ERROR, "i am an error"},
+		{1, WARNING, "i am a warning"},
+		{2, INFO, "i am informative"},
+		{3, INFO, "i am a zero-level debug"},
+	} {
+		p := reader.Backlog[te.ix]
+		if want, got := te.severity, p.Leveled.Severity(); want != got {
+			t.Errorf("wanted element %d to have severity %s, got %s", te.ix, want, got)
+		}
+		if want, got := te.message, p.Leveled.MessagesJoined(); want != got {
+			t.Errorf("wanted element %d to have message %q, got %q", te.ix, want, got)
+		}
+		if want, got := "logtree_test.go", strings.Split(p.Leveled.Location(), ":")[0]; want != got {
+			t.Errorf("wanted element %d to have file %q, got %q", te.ix, want, got)
+		}
+	}
+}
+
+func TestSeverity(t *testing.T) {
+	tree := New()
+	tree.MustLeveledFor("main").Error("i am an error")
+	tree.MustLeveledFor("main").Warning("i am a warning")
+	tree.MustLeveledFor("main").Info("i am informative")
+	tree.MustLeveledFor("main").V(0).Info("i am a zero-level debug")
+
+	reader, err := tree.Read("main", WithBacklog(BacklogAllAvailable), LeveledWithMinimumSeverity(WARNING))
+	if err != nil {
+		t.Fatalf("Read: %v", err)
+	}
+	if want, got := 2, len(reader.Backlog); want != got {
+		t.Fatalf("wanted %d entries, got %d", want, got)
+	}
+	if want, got := "i am an error", reader.Backlog[0].Leveled.MessagesJoined(); want != got {
+		t.Fatalf("wanted entry %q, got %q", want, got)
+	}
+	if want, got := "i am a warning", reader.Backlog[1].Leveled.MessagesJoined(); want != got {
+		t.Fatalf("wanted entry %q, got %q", want, got)
+	}
+}
diff --git a/metropolis/node/core/main.go b/metropolis/node/core/main.go
new file mode 100644
index 0000000..54d09a4
--- /dev/null
+++ b/metropolis/node/core/main.go
@@ -0,0 +1,321 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"context"
+	"crypto/ed25519"
+	"crypto/rand"
+	"crypto/x509"
+	"fmt"
+	"log"
+	"math/big"
+	"net"
+	"os"
+	"os/signal"
+	"runtime/debug"
+
+	"golang.org/x/sys/unix"
+	"google.golang.org/grpc"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/cluster"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dns"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/tpm"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/containerd"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/pki"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+var (
+	// kubernetesConfig is the static/global part of the Kubernetes service configuration. In the future, this might
+	// be configurable by loading it from the EnrolmentConfig. Fow now, it's static and same across all clusters.
+	kubernetesConfig = kubernetes.Config{
+		ServiceIPRange: net.IPNet{ // TODO(q3k): Decide if configurable / final value
+			IP:   net.IP{10, 0, 255, 1},
+			Mask: net.IPMask{0xff, 0xff, 0xff, 0x00}, // /24, but Go stores as a literal mask
+		},
+		ClusterNet: net.IPNet{
+			IP:   net.IP{10, 0, 0, 0},
+			Mask: net.IPMask{0xff, 0xff, 0x00, 0x00}, // /16
+		},
+	}
+)
+
+func main() {
+	defer func() {
+		if r := recover(); r != nil {
+			fmt.Println("Init panicked:", r)
+			debug.PrintStack()
+		}
+		unix.Sync()
+		// TODO(lorenz): Switch this to Reboot when init panics are less likely
+		// Best effort, nothing we can do if this fails except printing the error to the console.
+		if err := unix.Reboot(unix.LINUX_REBOOT_CMD_POWER_OFF); err != nil {
+			panic(fmt.Sprintf("failed to halt node: %v\n", err))
+		}
+	}()
+
+	// Set up logger for Smalltown. Currently logs everything to stderr.
+	lt := logtree.New()
+	reader, err := lt.Read("", logtree.WithChildren(), logtree.WithStream())
+	if err != nil {
+		panic(fmt.Errorf("could not set up root log reader: %v", err))
+	}
+	go func() {
+		for {
+			p := <-reader.Stream
+			fmt.Fprintf(os.Stderr, "%s\n", p.String())
+		}
+	}()
+
+	// Initial logger. Used until we get to a supervisor.
+	logger := lt.MustLeveledFor("init")
+
+	// Remount onto a tmpfs and re-exec if needed. Otherwise, keep running.
+	err = switchRoot(logger)
+	if err != nil {
+		panic(fmt.Errorf("could not remount root: %w", err))
+	}
+
+	// Linux kernel default is 4096 which is far too low. Raise it to 1M which is what gVisor suggests.
+	if err := unix.Setrlimit(unix.RLIMIT_NOFILE, &unix.Rlimit{Cur: 1048576, Max: 1048576}); err != nil {
+		logger.Fatalf("Failed to raise rlimits: %v", err)
+	}
+
+	logger.Info("Starting Smalltown Init")
+
+	signalChannel := make(chan os.Signal, 2)
+	signal.Notify(signalChannel)
+
+	if err := tpm.Initialize(logger); err != nil {
+		logger.Fatalf("Failed to initialize TPM 2.0: %v", err)
+	}
+
+	corednsRegistrationChan := make(chan *dns.ExtraDirective)
+
+	networkSvc := network.New(network.Config{CorednsRegistrationChan: corednsRegistrationChan})
+
+	// This function initializes a headless Delve if this is a debug build or does nothing if it's not
+	initializeDebugger(networkSvc)
+
+	// Prepare local storage.
+	root := &localstorage.Root{}
+	if err := declarative.PlaceFS(root, "/"); err != nil {
+		panic(fmt.Errorf("when placing root FS: %w", err))
+	}
+
+	// trapdoor is a channel used to signal to the init service that a very low-level, unrecoverable failure
+	// occured. This causes a GURU MEDITATION ERROR visible to the end user.
+	trapdoor := make(chan struct{})
+
+	// Make context for supervisor. We cancel it when we reach the trapdoor.
+	ctxS, ctxC := context.WithCancel(context.Background())
+
+	// Start root initialization code as a supervisor one-shot runnable. This means waiting for the network, starting
+	// the cluster manager, and then starting all services related to the node's roles.
+	// TODO(q3k): move this to a separate 'init' service.
+	supervisor.New(ctxS, func(ctx context.Context) error {
+		logger := supervisor.Logger(ctx)
+
+		// Start storage and network - we need this to get anything else done.
+		if err := root.Start(ctx); err != nil {
+			return fmt.Errorf("cannot start root FS: %w", err)
+		}
+		if err := supervisor.Run(ctx, "network", networkSvc.Run); err != nil {
+			return fmt.Errorf("when starting network: %w", err)
+		}
+
+		// Wait for IP address from network.
+		ip, err := networkSvc.GetIP(ctx, true)
+		if err != nil {
+			return fmt.Errorf("when waiting for IP address: %w", err)
+		}
+
+		// Start cluster manager. This kicks off cluster membership machinery, which will either start
+		// a new cluster, enroll into one or join one.
+		m := cluster.NewManager(root, networkSvc)
+		if err := supervisor.Run(ctx, "enrolment", m.Run); err != nil {
+			return fmt.Errorf("when starting enrolment: %w", err)
+		}
+
+		// Wait until the cluster manager settles.
+		success := m.WaitFinished()
+		if !success {
+			close(trapdoor)
+			return fmt.Errorf("enrolment failed, aborting")
+		}
+
+		// We are now in a cluster. We can thus access our 'node' object and start all services that
+		// we should be running.
+
+		node := m.Node()
+		if err := node.ConfigureLocalHostname(&root.Etc); err != nil {
+			close(trapdoor)
+			return fmt.Errorf("failed to set local hostname: %w", err)
+		}
+
+		logger.Info("Enrolment success, continuing startup.")
+		logger.Info(fmt.Sprintf("This node (%s) has roles:", node.String()))
+		if cm := node.ConsensusMember(); cm != nil {
+			// There's no need to start anything for when we are a consensus member - the cluster
+			// manager does this for us if necessary (as creating/enrolling/joining a cluster is
+			// pretty tied into cluster lifecycle management).
+			logger.Info(fmt.Sprintf(" - etcd consensus member"))
+		}
+		if kw := node.KubernetesWorker(); kw != nil {
+			logger.Info(fmt.Sprintf(" - kubernetes worker"))
+		}
+
+		// If we're supposed to be a kubernetes worker, start kubernetes services and containerd.
+		// In the future, this might be split further into kubernetes control plane and data plane
+		// roles.
+		var containerdSvc *containerd.Service
+		var kubeSvc *kubernetes.Service
+		if kw := node.KubernetesWorker(); kw != nil {
+			logger.Info("Starting Kubernetes worker services...")
+
+			// Ensure Kubernetes PKI objects exist in etcd.
+			kpkiKV := m.ConsensusKV("cluster", "kpki")
+			kpki := pki.NewKubernetes(lt.MustLeveledFor("pki.kubernetes"), kpkiKV)
+			if err := kpki.EnsureAll(ctx); err != nil {
+				return fmt.Errorf("failed to ensure kubernetes PKI present: %w", err)
+			}
+
+			containerdSvc = &containerd.Service{
+				EphemeralVolume: &root.Ephemeral.Containerd,
+			}
+			if err := supervisor.Run(ctx, "containerd", containerdSvc.Run); err != nil {
+				return fmt.Errorf("failed to start containerd service: %w", err)
+			}
+
+			kubernetesConfig.KPKI = kpki
+			kubernetesConfig.Root = root
+			kubernetesConfig.AdvertiseAddress = *ip
+			kubernetesConfig.CorednsRegistrationChan = corednsRegistrationChan
+			kubeSvc = kubernetes.New(kubernetesConfig)
+			if err := supervisor.Run(ctx, "kubernetes", kubeSvc.Run); err != nil {
+				return fmt.Errorf("failed to start kubernetes service: %w", err)
+			}
+
+		}
+
+		// Start the node debug service.
+		dbg := &debugService{
+			cluster:    m,
+			logtree:    lt,
+			kubernetes: kubeSvc,
+		}
+		dbgSrv := grpc.NewServer()
+		apb.RegisterNodeDebugServiceServer(dbgSrv, dbg)
+		dbgLis, err := net.Listen("tcp", fmt.Sprintf(":%d", common.DebugServicePort))
+		if err != nil {
+			return fmt.Errorf("failed to listen on debug service: %w", err)
+		}
+		if err := supervisor.Run(ctx, "debug", supervisor.GRPCServer(dbgSrv, dbgLis, false)); err != nil {
+			return fmt.Errorf("failed to start debug service: %w", err)
+		}
+
+		supervisor.Signal(ctx, supervisor.SignalHealthy)
+		supervisor.Signal(ctx, supervisor.SignalDone)
+		return nil
+	}, supervisor.WithExistingLogtree(lt))
+
+	// We're PID1, so orphaned processes get reparented to us to clean up
+	for {
+		select {
+		case <-trapdoor:
+			// If the trapdoor got closed, we got stuck early enough in the boot process that we can't do anything about
+			// it. Display a generic error message until we handle error conditions better.
+			ctxC()
+			log.Printf("                  ########################")
+			log.Printf("                  # GURU MEDIATION ERROR #")
+			log.Printf("                  ########################")
+			log.Printf("")
+			log.Printf("Smalltown encountered an uncorrectable error and must be restarted.")
+			log.Printf("(Error condition: init trapdoor closed)")
+			log.Printf("")
+			select {}
+
+		case sig := <-signalChannel:
+			switch sig {
+			case unix.SIGCHLD:
+				var status unix.WaitStatus
+				var rusage unix.Rusage
+				for {
+					res, err := unix.Wait4(-1, &status, unix.WNOHANG, &rusage)
+					if err != nil && err != unix.ECHILD {
+						logger.Errorf("Failed to wait on orphaned child: %v", err)
+						break
+					}
+					if res <= 0 {
+						break
+					}
+				}
+			case unix.SIGURG:
+				// Go 1.14 introduced asynchronous preemption, which uses SIGURG.
+				// In order not to break backwards compatibility in the unlikely case
+				// of an application actually using SIGURG on its own, they're not filtering them.
+				// (https://github.com/golang/go/issues/37942)
+				logger.V(5).Info("Ignoring SIGURG")
+			// TODO(lorenz): We can probably get more than just SIGCHLD as init, but I can't think
+			// of any others right now, just log them in case we hit any of them.
+			default:
+				logger.Warningf("Got unexpected signal %s", sig.String())
+			}
+		}
+	}
+}
+
+// nodeCertificate creates a node key/certificate for a foreign node. This is duplicated code with localstorage's
+// PKIDirectory EnsureSelfSigned, but is temporary (and specific to 'golden tickets').
+func (s *debugService) nodeCertificate() (cert, key []byte, err error) {
+	pubKey, privKey, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		err = fmt.Errorf("failed to generate key: %w", err)
+		return
+	}
+
+	key, err = x509.MarshalPKCS8PrivateKey(privKey)
+	if err != nil {
+		err = fmt.Errorf("failed to marshal key: %w", err)
+		return
+	}
+
+	serialNumberLimit := new(big.Int).Lsh(big.NewInt(1), 127)
+	serialNumber, err := rand.Int(rand.Reader, serialNumberLimit)
+	if err != nil {
+		err = fmt.Errorf("failed to generate serial number: %w", err)
+		return
+	}
+
+	template := localstorage.CertificateForNode(pubKey)
+	template.SerialNumber = serialNumber
+
+	cert, err = x509.CreateCertificate(rand.Reader, &template, &template, pubKey, privKey)
+	if err != nil {
+		err = fmt.Errorf("could not sign certificate: %w", err)
+		return
+	}
+	return
+}
diff --git a/metropolis/node/core/network/BUILD.bazel b/metropolis/node/core/network/BUILD.bazel
new file mode 100644
index 0000000..9ba56a9
--- /dev/null
+++ b/metropolis/node/core/network/BUILD.bazel
@@ -0,0 +1,20 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/network",
+    visibility = ["//:__subpackages__"],
+    deps = [
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/logtree:go_default_library",
+        "//metropolis/node/core/network/dhcp4c:go_default_library",
+        "//metropolis/node/core/network/dhcp4c/callback:go_default_library",
+        "//metropolis/node/core/network/dns:go_default_library",
+        "@com_github_google_nftables//:go_default_library",
+        "@com_github_google_nftables//expr:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4:go_default_library",
+        "@com_github_vishvananda_netlink//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/network/dhcp4c/BUILD.bazel b/metropolis/node/core/network/dhcp4c/BUILD.bazel
new file mode 100644
index 0000000..19b4c70
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/BUILD.bazel
@@ -0,0 +1,35 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "dhcpc.go",
+        "doc.go",
+        "lease.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/network/dhcp4c/transport:go_default_library",
+        "@com_github_cenkalti_backoff_v4//:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4:go_default_library",
+        "@com_github_insomniacslk_dhcp//iana:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = [
+        "dhcpc_test.go",
+        "lease_test.go",
+    ],
+    embed = [":go_default_library"],
+    pure = "on",
+    deps = [
+        "//metropolis/node/core/network/dhcp4c/transport:go_default_library",
+        "@com_github_cenkalti_backoff_v4//:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4:go_default_library",
+        "@com_github_stretchr_testify//assert:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/network/dhcp4c/callback/BUILD.bazel b/metropolis/node/core/network/dhcp4c/callback/BUILD.bazel
new file mode 100644
index 0000000..ed6f330
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/callback/BUILD.bazel
@@ -0,0 +1,36 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//metropolis/test/ktest:ktest.bzl", "ktest")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["callback.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c/callback",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//metropolis/node/core/network/dhcp4c:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4:go_default_library",
+        "@com_github_vishvananda_netlink//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["callback_test.go"],
+    embed = [":go_default_library"],
+    pure = "on",
+    deps = [
+        "//metropolis/node/core/network/dhcp4c:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4:go_default_library",
+        "@com_github_stretchr_testify//require:go_default_library",
+        "@com_github_vishvananda_netlink//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+ktest(
+    cmdline = "",
+    initramfs_extra = "",
+    tester = ":go_default_test",
+    deps = [],
+)
diff --git a/metropolis/node/core/network/dhcp4c/callback/callback.go b/metropolis/node/core/network/dhcp4c/callback/callback.go
new file mode 100644
index 0000000..10eb6ba
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/callback/callback.go
@@ -0,0 +1,149 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package callback contains minimal callbacks for configuring the kernel with options received over DHCP.
+//
+// These directly configure the relevant kernel subsytems and need to own certain parts of them as documented on a per-
+// callback basis to make sure that they can recover from restarts and crashes of the DHCP client.
+// The callbacks in here are not suitable for use in advanced network scenarios like running multiple DHCP clients
+// per interface via ClientIdentifier or when running an external FIB manager. In these cases it's advised to extract
+// the necessary information from the lease in your own callback and communicate it directly to the responsible entity.
+package callback
+
+import (
+	"fmt"
+	"math"
+	"net"
+	"os"
+	"time"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c"
+
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/vishvananda/netlink"
+	"golang.org/x/sys/unix"
+)
+
+// Compose can be used to chain multiple callbacks
+func Compose(callbacks ...dhcp4c.LeaseCallback) dhcp4c.LeaseCallback {
+	return func(old, new *dhcp4c.Lease) error {
+		for _, cb := range callbacks {
+			if err := cb(old, new); err != nil {
+				return err
+			}
+		}
+		return nil
+	}
+}
+
+func isIPNetEqual(a, b *net.IPNet) bool {
+	if a == b {
+		return true
+	}
+	if a == nil || b == nil {
+		return false
+	}
+	aOnes, aBits := a.Mask.Size()
+	bOnes, bBits := b.Mask.Size()
+	return a.IP.Equal(b.IP) && aOnes == bOnes && aBits == bBits
+}
+
+// ManageIP sets up and tears down the assigned IP address. It takes exclusive ownership of all IPv4 addresses
+// on the given interface which do not have IFA_F_PERMANENT set, so it's not possible to run multiple dynamic addressing
+// clients on a single interface.
+func ManageIP(iface netlink.Link) dhcp4c.LeaseCallback {
+	return func(old, new *dhcp4c.Lease) error {
+		newNet := new.IPNet()
+
+		addrs, err := netlink.AddrList(iface, netlink.FAMILY_V4)
+		if err != nil {
+			return fmt.Errorf("netlink failed to list addresses: %w", err)
+		}
+
+		for _, addr := range addrs {
+			if addr.Flags&unix.IFA_F_PERMANENT == 0 {
+				// Linux identifies addreses by IP, mask and peer (see net/ipv4/devinet.find_matching_ifa in Linux 5.10)
+				// So don't touch addresses which match on these properties as AddrReplace will atomically reconfigure
+				// them anyways without interrupting things.
+				if isIPNetEqual(addr.IPNet, newNet) && addr.Peer == nil && new != nil {
+					continue
+				}
+
+				if err := netlink.AddrDel(iface, &addr); !os.IsNotExist(err) && err != nil {
+					return fmt.Errorf("failed to delete address: %w", err)
+				}
+			}
+		}
+
+		if new != nil {
+			remainingLifetimeSecs := int(math.Ceil(new.ExpiresAt.Sub(time.Now()).Seconds()))
+			newBroadcastIP := dhcpv4.GetIP(dhcpv4.OptionBroadcastAddress, new.Options)
+			if err := netlink.AddrReplace(iface, &netlink.Addr{
+				IPNet:       newNet,
+				ValidLft:    remainingLifetimeSecs,
+				PreferedLft: remainingLifetimeSecs,
+				Broadcast:   newBroadcastIP,
+			}); err != nil {
+				return fmt.Errorf("failed to update address: %w", err)
+			}
+		}
+		return nil
+	}
+}
+
+// ManageDefaultRoute manages a default route through the first router offered by DHCP. It does nothing if DHCP
+// doesn't provide any routers. It takes ownership of all RTPROTO_DHCP routes on the given interface, so it's not
+// possible to run multiple DHCP clients on the given interface.
+func ManageDefaultRoute(iface netlink.Link) dhcp4c.LeaseCallback {
+	return func(old, new *dhcp4c.Lease) error {
+		newRouter := new.Router()
+
+		dhcpRoutes, err := netlink.RouteListFiltered(netlink.FAMILY_V4, &netlink.Route{
+			Protocol:  unix.RTPROT_DHCP,
+			LinkIndex: iface.Attrs().Index,
+		}, netlink.RT_FILTER_OIF|netlink.RT_FILTER_PROTOCOL)
+		if err != nil {
+			return fmt.Errorf("netlink failed to list routes: %w", err)
+		}
+		ipv4DefaultRoute := net.IPNet{IP: net.IPv4zero, Mask: net.CIDRMask(0, 32)}
+		for _, route := range dhcpRoutes {
+			// Don't remove routes which can be atomically replaced by RouteReplace to prevent potential traffic
+			// disruptions.
+			if !isIPNetEqual(&ipv4DefaultRoute, route.Dst) && newRouter != nil {
+				continue
+			}
+			err := netlink.RouteDel(&route)
+			if !os.IsNotExist(err) && err != nil {
+				return fmt.Errorf("failed to delete DHCP route: %w", err)
+			}
+		}
+
+		if newRouter != nil {
+			err := netlink.RouteReplace(&netlink.Route{
+				Protocol:  unix.RTPROT_DHCP,
+				Dst:       &ipv4DefaultRoute,
+				Gw:        newRouter,
+				Src:       new.AssignedIP,
+				LinkIndex: iface.Attrs().Index,
+				Scope:     netlink.SCOPE_UNIVERSE,
+			})
+			if err != nil {
+				return fmt.Errorf("failed to add default route via %s: %w", newRouter, err)
+			}
+		}
+		return nil
+	}
+}
diff --git a/metropolis/node/core/network/dhcp4c/callback/callback_test.go b/metropolis/node/core/network/dhcp4c/callback/callback_test.go
new file mode 100644
index 0000000..d533044
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/callback/callback_test.go
@@ -0,0 +1,313 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package callback
+
+import (
+	"fmt"
+	"math"
+	"net"
+	"os"
+	"testing"
+	"time"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c"
+
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/stretchr/testify/require"
+	"github.com/vishvananda/netlink"
+	"golang.org/x/sys/unix"
+)
+
+func trivialLeaseFromNet(ipnet net.IPNet) *dhcp4c.Lease {
+	opts := make(dhcpv4.Options)
+	opts.Update(dhcpv4.OptSubnetMask(ipnet.Mask))
+	return &dhcp4c.Lease{
+		AssignedIP: ipnet.IP,
+		ExpiresAt:  time.Now().Add(1 * time.Second),
+		Options:    opts,
+	}
+}
+
+var (
+	testNet1          = net.IPNet{IP: net.IP{10, 0, 1, 2}, Mask: net.CIDRMask(24, 32)}
+	testNet1Broadcast = net.IP{10, 0, 1, 255}
+	testNet1Router    = net.IP{10, 0, 1, 1}
+	testNet2          = net.IPNet{IP: net.IP{10, 0, 2, 2}, Mask: net.CIDRMask(24, 32)}
+	testNet2Broadcast = net.IP{10, 0, 2, 255}
+	testNet2Router    = net.IP{10, 0, 2, 1}
+	mainRoutingTable  = 254 // Linux automatically puts all routes into this table unless specified
+)
+
+func TestAssignedIPCallback(t *testing.T) {
+	if os.Getenv("IN_KTEST") != "true" {
+		t.Skip("Not in ktest")
+	}
+
+	var tests = []struct {
+		name               string
+		initialAddrs       []netlink.Addr
+		oldLease, newLease *dhcp4c.Lease
+		expectedAddrs      []netlink.Addr
+	}{
+		{ // Lifetimes are necessary, otherwise the Kernel sets the IFA_F_PERMANENT flag behind our back
+			name:          "RemoveOldIPs",
+			initialAddrs:  []netlink.Addr{{IPNet: &testNet1, ValidLft: 60}, {IPNet: &testNet2, ValidLft: 60}},
+			oldLease:      nil,
+			newLease:      nil,
+			expectedAddrs: nil,
+		},
+		{
+			name:         "IgnoresPermanentIPs",
+			initialAddrs: []netlink.Addr{{IPNet: &testNet1, Flags: unix.IFA_F_PERMANENT}, {IPNet: &testNet2, ValidLft: 60}},
+			oldLease:     nil,
+			newLease:     trivialLeaseFromNet(testNet2),
+			expectedAddrs: []netlink.Addr{
+				{IPNet: &testNet1, Flags: unix.IFA_F_PERMANENT, ValidLft: math.MaxUint32, PreferedLft: math.MaxUint32, Broadcast: testNet1Broadcast},
+				{IPNet: &testNet2, ValidLft: 1, PreferedLft: 1, Broadcast: testNet2Broadcast},
+			},
+		},
+		{
+			name:         "AssignsNewIP",
+			initialAddrs: []netlink.Addr{},
+			oldLease:     nil,
+			newLease:     trivialLeaseFromNet(testNet2),
+			expectedAddrs: []netlink.Addr{
+				{IPNet: &testNet2, ValidLft: 1, PreferedLft: 1, Broadcast: testNet2Broadcast},
+			},
+		},
+		{
+			name:         "UpdatesIP",
+			initialAddrs: []netlink.Addr{},
+			oldLease:     trivialLeaseFromNet(testNet2),
+			newLease:     trivialLeaseFromNet(testNet1),
+			expectedAddrs: []netlink.Addr{
+				{IPNet: &testNet1, ValidLft: 1, PreferedLft: 1, Broadcast: testNet1Broadcast},
+			},
+		},
+		{
+			name:          "RemovesIPOnRelease",
+			initialAddrs:  []netlink.Addr{{IPNet: &testNet1, ValidLft: 60, PreferedLft: 60}},
+			oldLease:      trivialLeaseFromNet(testNet1),
+			newLease:      nil,
+			expectedAddrs: nil,
+		},
+	}
+	for i, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			testLink := &netlink.Dummy{
+				LinkAttrs: netlink.LinkAttrs{
+					Name:  fmt.Sprintf("aipcb-test-%d", i),
+					Flags: unix.IFF_UP,
+				},
+			}
+			if err := netlink.LinkAdd(testLink); err != nil {
+				t.Fatalf("test cannot set up network interface: %v", err)
+			}
+			defer netlink.LinkDel(testLink)
+			for _, addr := range test.initialAddrs {
+				if err := netlink.AddrAdd(testLink, &addr); err != nil {
+					t.Fatalf("test cannot set up initial addrs: %v", err)
+				}
+			}
+			// Associate dynamically-generated interface name for later comparison
+			for i := range test.expectedAddrs {
+				test.expectedAddrs[i].Label = testLink.Name
+			}
+			cb := ManageIP(testLink)
+			if err := cb(test.oldLease, test.newLease); err != nil {
+				t.Fatalf("callback returned an error: %v", err)
+			}
+			addrs, err := netlink.AddrList(testLink, netlink.FAMILY_V4)
+			if err != nil {
+				t.Fatalf("test cannot read back addrs from interface: %v", err)
+			}
+			require.Equal(t, test.expectedAddrs, addrs, "Wrong IPs on interface")
+		})
+	}
+}
+
+func leaseAddRouter(lease *dhcp4c.Lease, router net.IP) *dhcp4c.Lease {
+	lease.Options.Update(dhcpv4.OptRouter(router))
+	return lease
+}
+
+func TestDefaultRouteCallback(t *testing.T) {
+	if os.Getenv("IN_KTEST") != "true" {
+		t.Skip("Not in ktest")
+	}
+	// testRoute is only used as a route destination and not configured on any interface.
+	testRoute := net.IPNet{IP: net.IP{10, 0, 3, 0}, Mask: net.CIDRMask(24, 32)}
+
+	// A test interface is set up for each test and assigned testNet1 and testNet2 so that testNet1Router and
+	// testNet2Router are valid gateways for routes in this environment. A LinkIndex of -1 is replaced by the correct
+	// link index for this test interface at runtime for both initialRoutes and expectedRoutes.
+	var tests = []struct {
+		name               string
+		initialRoutes      []netlink.Route
+		oldLease, newLease *dhcp4c.Lease
+		expectedRoutes     []netlink.Route
+	}{
+		{
+			name:          "AddsDefaultRoute",
+			initialRoutes: []netlink.Route{},
+			oldLease:      nil,
+			newLease:      leaseAddRouter(trivialLeaseFromNet(testNet1), testNet1Router),
+			expectedRoutes: []netlink.Route{{
+				Protocol:  unix.RTPROT_DHCP,
+				Dst:       nil, // Linux weirdly retuns no RTA_DST for default routes, but one for everything else
+				Gw:        testNet1Router,
+				Src:       testNet1.IP,
+				Table:     mainRoutingTable,
+				LinkIndex: -1, // Filled in dynamically with test interface
+				Type:      unix.RTN_UNICAST,
+			}},
+		},
+		{
+			name:           "IgnoresLeasesWithoutRouter",
+			initialRoutes:  []netlink.Route{},
+			oldLease:       nil,
+			newLease:       trivialLeaseFromNet(testNet1),
+			expectedRoutes: nil,
+		},
+		{
+			name: "RemovesUnrelatedOldRoutes",
+			initialRoutes: []netlink.Route{{
+				Dst:       &testRoute,
+				LinkIndex: -1, // Filled in dynamically with test interface
+				Protocol:  unix.RTPROT_DHCP,
+				Gw:        testNet2Router,
+				Scope:     netlink.SCOPE_UNIVERSE,
+			}},
+			oldLease:       nil,
+			newLease:       nil,
+			expectedRoutes: nil,
+		},
+		{
+			name: "IgnoresNonDHCPRoutes",
+			initialRoutes: []netlink.Route{{
+				Dst:       &testRoute,
+				LinkIndex: -1, // Filled in dynamically with test interface
+				Protocol:  unix.RTPROT_BIRD,
+				Gw:        testNet2Router,
+			}},
+			oldLease: trivialLeaseFromNet(testNet1),
+			newLease: nil,
+			expectedRoutes: []netlink.Route{{
+				Protocol:  unix.RTPROT_BIRD,
+				Dst:       &testRoute,
+				Gw:        testNet2Router,
+				Table:     mainRoutingTable,
+				LinkIndex: -1, // Filled in dynamically with test interface
+				Type:      unix.RTN_UNICAST,
+			}},
+		},
+		{
+			name: "RemovesRoute",
+			initialRoutes: []netlink.Route{{
+				Dst:       nil,
+				LinkIndex: -1, // Filled in dynamically with test interface
+				Protocol:  unix.RTPROT_DHCP,
+				Gw:        testNet2Router,
+			}},
+			oldLease:       leaseAddRouter(trivialLeaseFromNet(testNet2), testNet2Router),
+			newLease:       nil,
+			expectedRoutes: nil,
+		},
+		{
+			name: "UpdatesRoute",
+			initialRoutes: []netlink.Route{{
+				Dst:       nil,
+				LinkIndex: -1, // Filled in dynamically with test interface
+				Protocol:  unix.RTPROT_DHCP,
+				Src:       testNet1.IP,
+				Gw:        testNet1Router,
+			}},
+			oldLease: leaseAddRouter(trivialLeaseFromNet(testNet1), testNet1Router),
+			newLease: leaseAddRouter(trivialLeaseFromNet(testNet2), testNet2Router),
+			expectedRoutes: []netlink.Route{{
+				Protocol:  unix.RTPROT_DHCP,
+				Dst:       nil,
+				Gw:        testNet2Router,
+				Src:       testNet2.IP,
+				Table:     mainRoutingTable,
+				LinkIndex: -1, // Filled in dynamically with test interface
+				Type:      unix.RTN_UNICAST,
+			}},
+		},
+	}
+	for i, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			testLink := &netlink.Dummy{
+				LinkAttrs: netlink.LinkAttrs{
+					Name:  fmt.Sprintf("drcb-test-%d", i),
+					Flags: unix.IFF_UP,
+				},
+			}
+			if err := netlink.LinkAdd(testLink); err != nil {
+				t.Fatalf("test cannot set up network interface: %v", err)
+			}
+			defer func() { // Clean up after each test
+				routes, err := netlink.RouteListFiltered(netlink.FAMILY_V4, &netlink.Route{}, 0)
+				if err == nil {
+					for _, route := range routes {
+						netlink.RouteDel(&route)
+					}
+				}
+			}()
+			defer netlink.LinkDel(testLink)
+			if err := netlink.AddrAdd(testLink, &netlink.Addr{
+				IPNet: &testNet1,
+			}); err != nil {
+				t.Fatalf("test cannot set up test addrs: %v", err)
+			}
+			if err := netlink.AddrAdd(testLink, &netlink.Addr{
+				IPNet: &testNet2,
+			}); err != nil {
+				t.Fatalf("test cannot set up test addrs: %v", err)
+			}
+			for _, route := range test.initialRoutes {
+				if route.LinkIndex == -1 {
+					route.LinkIndex = testLink.Index
+				}
+				if err := netlink.RouteAdd(&route); err != nil {
+					t.Fatalf("test cannot set up initial routes: %v", err)
+				}
+			}
+			for i := range test.expectedRoutes {
+				if test.expectedRoutes[i].LinkIndex == -1 {
+					test.expectedRoutes[i].LinkIndex = testLink.Index
+				}
+			}
+
+			cb := ManageDefaultRoute(testLink)
+			if err := cb(test.oldLease, test.newLease); err != nil {
+				t.Fatalf("callback returned an error: %v", err)
+			}
+			routes, err := netlink.RouteListFiltered(netlink.FAMILY_V4, &netlink.Route{}, 0)
+			if err != nil {
+				t.Fatalf("test cannot read back routes: %v", err)
+			}
+			var notKernelRoutes []netlink.Route
+			for _, route := range routes {
+				if route.Protocol != unix.RTPROT_KERNEL { // Filter kernel-managed routes
+					notKernelRoutes = append(notKernelRoutes, route)
+				}
+			}
+			require.Equal(t, test.expectedRoutes, notKernelRoutes, "Wrong Routes")
+		})
+	}
+}
diff --git a/metropolis/node/core/network/dhcp4c/dhcpc.go b/metropolis/node/core/network/dhcp4c/dhcpc.go
new file mode 100644
index 0000000..4352506
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/dhcpc.go
@@ -0,0 +1,677 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package dhcp4c implements a DHCPv4 Client as specified in RFC2131 (with some notable deviations).
+// It implements only the DHCP state machine itself, any configuration other than the interface IP
+// address (which is always assigned in DHCP and necessary for the protocol to work) is exposed
+// as [informers/observables/watchable variables/???] to consumers who then deal with it.
+package dhcp4c
+
+import (
+	"context"
+	"crypto/rand"
+	"errors"
+	"fmt"
+	"io"
+	"math"
+	"net"
+	"time"
+
+	"github.com/cenkalti/backoff/v4"
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/insomniacslk/dhcp/iana"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c/transport"
+)
+
+type state int
+
+const (
+	// stateDiscovering sends broadcast DHCPDISCOVER messages to the network and waits for either a DHCPOFFER or
+	// (in case of Rapid Commit) DHCPACK.
+	stateDiscovering state = iota
+	// stateRequesting sends broadcast DHCPREQUEST messages containing the server identifier for the selected lease and
+	// waits for a DHCPACK or a DHCPNAK. If it doesn't get either it transitions back into discovering.
+	stateRequesting
+	// stateBound just waits until RenewDeadline (derived from RenewTimeValue, half the lifetime by default) expires.
+	stateBound
+	// stateRenewing sends unicast DHCPREQUEST messages to the currently-selected server and waits for either a DHCPACK
+	// or DHCPNAK message. On DHCPACK it transitions to bound, otherwise to discovering.
+	stateRenewing
+	// stateRebinding sends broadcast DHCPREQUEST messages to the network and waits for either a DHCPACK or DHCPNAK from
+	// any server. Response processing is identical to stateRenewing.
+	stateRebinding
+)
+
+func (s state) String() string {
+	switch s {
+	case stateDiscovering:
+		return "DISCOVERING"
+	case stateRequesting:
+		return "REQUESTING"
+	case stateBound:
+		return "BOUND"
+	case stateRenewing:
+		return "RENEWING"
+	case stateRebinding:
+		return "REBINDING"
+	default:
+		return "INVALID"
+	}
+}
+
+// This only requests SubnetMask and IPAddressLeaseTime as renewal and rebinding times are fine if
+// they are just defaulted. They are respected (if valid, otherwise they are clamped to the nearest
+// valid value) if sent by the server.
+var internalOptions = dhcpv4.OptionCodeList{dhcpv4.OptionSubnetMask, dhcpv4.OptionIPAddressLeaseTime}
+
+// Transport represents a mechanism over which DHCP messages can be exchanged with a server.
+type Transport interface {
+	// Send attempts to send the given DHCP payload message to the transport target once. An empty return value
+	// does not indicate that the message was successfully received.
+	Send(payload *dhcpv4.DHCPv4) error
+	// SetReceiveDeadline sets a deadline for Receive() calls after which they return with DeadlineExceededErr
+	SetReceiveDeadline(time.Time) error
+	// Receive waits for a DHCP message to arrive and returns it. If the deadline expires without a message arriving
+	// it will return DeadlineExceededErr. If the message is completely malformed it will an instance of
+	// InvalidMessageError.
+	Receive() (*dhcpv4.DHCPv4, error)
+	// Close closes the given transport. Calls to any of the above methods will fail if the transport is closed.
+	// Specific transports can be reopened after being closed.
+	Close() error
+}
+
+// UnicastTransport represents a mechanism over which DHCP messages can be exchanged with a single server over an
+// arbitrary IPv4-based network. Implementers need to support servers running outside the local network via a router.
+type UnicastTransport interface {
+	Transport
+	// Open connects the transport to a new unicast target. Can only be called after calling Close() or after creating
+	// a new transport.
+	Open(serverIP, bindIP net.IP) error
+}
+
+// BroadcastTransport represents a mechanism over which DHCP messages can be exchanged with all servers on a Layer 2
+// broadcast domain. Implementers need to support sending and receiving messages without any IP being configured on
+// the interface.
+type BroadcastTransport interface {
+	Transport
+	// Open connects the transport. Can only be called after calling Close() or after creating a new transport.
+	Open() error
+}
+
+type LeaseCallback func(old, new *Lease) error
+
+// Client implements a DHCPv4 client.
+//
+// Note that the size of all data sent to the server (RequestedOptions, ClientIdentifier,
+// VendorClassIdentifier and ExtraRequestOptions) should be kept reasonably small (<500 bytes) in
+// order to maximize the chance that requests can be properly transmitted.
+type Client struct {
+	// RequestedOptions contains a list of extra options this client is interested in
+	RequestedOptions dhcpv4.OptionCodeList
+
+	// ClientIdentifier is used by the DHCP server to identify this client.
+	// If empty, on Ethernet the MAC address is used instead.
+	ClientIdentifier []byte
+
+	// VendorClassIdentifier is used by the DHCP server to identify options specific to this type of
+	// clients and to populate the vendor-specific option (43).
+	VendorClassIdentifier string
+
+	// ExtraRequestOptions are extra options sent to the server.
+	ExtraRequestOptions dhcpv4.Options
+
+	// Backoff strategies for each state. These all have sane defaults, override them only if
+	// necessary.
+	DiscoverBackoff    backoff.BackOff
+	AcceptOfferBackoff backoff.BackOff
+	RenewBackoff       backoff.BackOff
+	RebindBackoff      backoff.BackOff
+
+	state state
+
+	lastBoundTransition time.Time
+
+	iface *net.Interface
+
+	// now can be used to override time for testing
+	now func() time.Time
+
+	// LeaseCallback is called every time a lease is aquired, renewed or lost
+	LeaseCallback LeaseCallback
+
+	// Valid in states Discovering, Requesting, Rebinding
+	broadcastConn BroadcastTransport
+
+	// Valid in states Requesting
+	offer *dhcpv4.DHCPv4
+
+	// Valid in states Bound, Renewing
+	unicastConn UnicastTransport
+
+	// Valid in states Bound, Renewing, Rebinding
+	lease              *dhcpv4.DHCPv4
+	leaseDeadline      time.Time
+	leaseBoundDeadline time.Time
+	leaseRenewDeadline time.Time
+}
+
+// newDefaultBackoff returns an infinitely-retrying randomized exponential backoff with a
+// DHCP-appropriate InitialInterval
+func newDefaultBackoff() *backoff.ExponentialBackOff {
+	b := backoff.NewExponentialBackOff()
+	b.MaxElapsedTime = 0 // No Timeout
+	// Lots of servers wait 1s for existing users of an IP. Wait at least for that and keep some
+	// slack for randomization, communication and processing overhead.
+	b.InitialInterval = 1400 * time.Millisecond
+	b.MaxInterval = 30 * time.Second
+	b.RandomizationFactor = 0.2
+	return b
+}
+
+// NewClient instantiates (but doesn't start) a new DHCPv4 client.
+// To have a working client it's required to set LeaseCallback to something that is capable of configuring the IP
+// address on the given interface. Unless managed through external means like a routing protocol, setting the default
+// route is also required. A simple example with the callback package thus looks like this:
+//  c := dhcp4c.NewClient(yourInterface)
+//  c.LeaseCallback = callback.Compose(callback.ManageIP(yourInterface), callback.ManageDefaultRoute(yourInterface))
+//  c.Run(ctx)
+func NewClient(iface *net.Interface) (*Client, error) {
+	broadcastConn := transport.NewBroadcastTransport(iface)
+
+	// broadcastConn needs to be open in stateDiscovering
+	if err := broadcastConn.Open(); err != nil {
+		return nil, fmt.Errorf("failed to create DHCP broadcast transport: %w", err)
+	}
+
+	discoverBackoff := newDefaultBackoff()
+
+	acceptOfferBackoff := newDefaultBackoff()
+	// Abort after 30s and go back to discovering
+	acceptOfferBackoff.MaxElapsedTime = 30 * time.Second
+
+	renewBackoff := newDefaultBackoff()
+	// Increase maximum interval to reduce chatter when the server is down
+	renewBackoff.MaxInterval = 5 * time.Minute
+
+	rebindBackoff := newDefaultBackoff()
+	// Increase maximum interval to reduce chatter when the server is down
+	renewBackoff.MaxInterval = 5 * time.Minute
+
+	return &Client{
+		state:               stateDiscovering,
+		broadcastConn:       broadcastConn,
+		unicastConn:         transport.NewUnicastTransport(iface),
+		iface:               iface,
+		RequestedOptions:    dhcpv4.OptionCodeList{},
+		lastBoundTransition: time.Now(),
+		now:                 time.Now,
+		DiscoverBackoff:     discoverBackoff,
+		AcceptOfferBackoff:  acceptOfferBackoff,
+		RenewBackoff:        renewBackoff,
+		RebindBackoff:       rebindBackoff,
+	}, nil
+}
+
+// acceptableLease checks if the given lease is valid enough to even be processed. This is
+// intentionally not exposed to users because under certain cirumstances it can end up acquiring all
+// available IP addresses from a server.
+func (c *Client) acceptableLease(offer *dhcpv4.DHCPv4) bool {
+	// RFC2131 Section 4.3.1 Table 3
+	if offer.ServerIdentifier() == nil || offer.ServerIdentifier().To4() == nil {
+		return false
+	}
+	// RFC2131 Section 4.3.1 Table 3
+	// Minimum representable lease time is 1s (Section 1.1)
+	if offer.IPAddressLeaseTime(0) < 1*time.Second {
+		return false
+	}
+
+	// Ignore IPs that are in no way valid for an interface (multicast, loopback, ...)
+	if offer.YourIPAddr.To4() == nil || (!offer.YourIPAddr.IsGlobalUnicast() && !offer.YourIPAddr.IsLinkLocalUnicast()) {
+		return false
+	}
+
+	// Technically the options Requested IP address, Parameter request list, Client identifier
+	// and Maximum message size should be refused (MUST NOT), but in the interest of interopatibilty
+	// let's simply remove them if they are present.
+	delete(offer.Options, dhcpv4.OptionRequestedIPAddress.Code())
+	delete(offer.Options, dhcpv4.OptionParameterRequestList.Code())
+	delete(offer.Options, dhcpv4.OptionClientIdentifier.Code())
+	delete(offer.Options, dhcpv4.OptionMaximumDHCPMessageSize.Code())
+
+	// Clamp rebindinding times longer than the lease time. Otherwise the state machine might misbehave.
+	if offer.IPAddressRebindingTime(0) > offer.IPAddressLeaseTime(0) {
+		offer.UpdateOption(dhcpv4.OptGeneric(dhcpv4.OptionRebindingTimeValue, dhcpv4.Duration(offer.IPAddressLeaseTime(0)).ToBytes()))
+	}
+	// Clamp renewal times longer than the rebinding time. Otherwise the state machine might misbehave.
+	if offer.IPAddressRenewalTime(0) > offer.IPAddressRebindingTime(0) {
+		offer.UpdateOption(dhcpv4.OptGeneric(dhcpv4.OptionRenewTimeValue, dhcpv4.Duration(offer.IPAddressRebindingTime(0)).ToBytes()))
+	}
+
+	// Normalize two options that can be represented either inline or as options.
+	if len(offer.ServerHostName) > 0 {
+		offer.Options[uint8(dhcpv4.OptionTFTPServerName)] = []byte(offer.ServerHostName)
+	}
+	if len(offer.BootFileName) > 0 {
+		offer.Options[uint8(dhcpv4.OptionBootfileName)] = []byte(offer.BootFileName)
+	}
+
+	// Normalize siaddr to option 150 (see RFC5859)
+	if len(offer.GetOneOption(dhcpv4.OptionTFTPServerAddress)) == 0 {
+		if offer.ServerIPAddr.To4() != nil && (offer.ServerIPAddr.IsGlobalUnicast() || offer.ServerIPAddr.IsLinkLocalUnicast()) {
+			offer.Options[uint8(dhcpv4.OptionTFTPServerAddress)] = offer.ServerIPAddr.To4()
+		}
+	}
+
+	return true
+}
+
+func earliestDeadline(dl1, dl2 time.Time) time.Time {
+	if dl1.Before(dl2) {
+		return dl1
+	} else {
+		return dl2
+	}
+}
+
+// newXID generates a new transaction ID
+func (c *Client) newXID() (dhcpv4.TransactionID, error) {
+	var xid dhcpv4.TransactionID
+	if _, err := io.ReadFull(rand.Reader, xid[:]); err != nil {
+		return xid, fmt.Errorf("cannot read randomness for transaction ID: %w", err)
+	}
+	return xid, nil
+}
+
+// As most servers out there cannot do reassembly, let's just hope for the best and
+// provide the local interface MTU. If the packet is too big it won't work anyways.
+// Also clamp to the biggest representable MTU in DHCPv4 (2 bytes unsigned int).
+func (c *Client) maxMsgSize() uint16 {
+	if c.iface.MTU < math.MaxUint16 {
+		return uint16(c.iface.MTU)
+	} else {
+		return math.MaxUint16
+	}
+}
+
+// newMsg creates a new DHCP message of a given type and adds common options.
+func (c *Client) newMsg(t dhcpv4.MessageType) (*dhcpv4.DHCPv4, error) {
+	xid, err := c.newXID()
+	if err != nil {
+		return nil, err
+	}
+	opts := make(dhcpv4.Options)
+	opts.Update(dhcpv4.OptMessageType(t))
+	if len(c.ClientIdentifier) > 0 {
+		opts.Update(dhcpv4.OptClientIdentifier(c.ClientIdentifier))
+	}
+	if t == dhcpv4.MessageTypeDiscover || t == dhcpv4.MessageTypeRequest || t == dhcpv4.MessageTypeInform {
+		opts.Update(dhcpv4.OptParameterRequestList(append(c.RequestedOptions, internalOptions...)...))
+		opts.Update(dhcpv4.OptMaxMessageSize(c.maxMsgSize()))
+		if c.VendorClassIdentifier != "" {
+			opts.Update(dhcpv4.OptClassIdentifier(c.VendorClassIdentifier))
+		}
+		for opt, val := range c.ExtraRequestOptions {
+			opts[opt] = val
+		}
+	}
+	return &dhcpv4.DHCPv4{
+		OpCode:        dhcpv4.OpcodeBootRequest,
+		HWType:        iana.HWTypeEthernet,
+		ClientHWAddr:  c.iface.HardwareAddr,
+		HopCount:      0,
+		TransactionID: xid,
+		NumSeconds:    0,
+		Flags:         0,
+		ClientIPAddr:  net.IPv4zero,
+		YourIPAddr:    net.IPv4zero,
+		ServerIPAddr:  net.IPv4zero,
+		GatewayIPAddr: net.IPv4zero,
+		Options:       opts,
+	}, nil
+}
+
+// transactionStateSpec describes a state which is driven by a DHCP message transaction (sending a
+// specific message and then transitioning into a different state depending on the received messages)
+type transactionStateSpec struct {
+	// ctx is a context for canceling the process
+	ctx context.Context
+
+	// transport is used to send and receive messages in this state
+	transport Transport
+
+	// stateDeadline is a fixed external deadline for how long the FSM can remain in this state.
+	// If it's exceeded the stateDeadlineExceeded callback is called and responsible for
+	// transitioning out of this state. It can be left empty to signal that there's no external
+	// deadline for the state.
+	stateDeadline time.Time
+
+	// backoff controls how long to wait for answers until handing control back to the FSM.
+	// Since the FSM hasn't advanced until then this means we just get called again and retransmit.
+	backoff backoff.BackOff
+
+	// requestType is the type of DHCP request sent out in this state. This is used to populate
+	// the default options for the message.
+	requestType dhcpv4.MessageType
+
+	// setExtraOptions can modify the request and set extra options before transmitting. Returning
+	// an error here aborts the FSM an can be used to terminate when no valid request can be
+	// constructed.
+	setExtraOptions func(msg *dhcpv4.DHCPv4) error
+
+	// handleMessage gets called for every parseable (not necessarily valid) DHCP message received
+	// by the transport. It should return an error for every message that doesn't advance the
+	// state machine and no error for every one that does. It is responsible for advancing the FSM
+	// if the required information is present.
+	handleMessage func(msg *dhcpv4.DHCPv4, sentTime time.Time) error
+
+	// stateDeadlineExceeded gets called if either the backoff returns backoff.Stop or the
+	// stateDeadline runs out. It is responsible for advancing the FSM into the next state.
+	stateDeadlineExceeded func() error
+}
+
+func (c *Client) runTransactionState(s transactionStateSpec) error {
+	sentTime := c.now()
+	msg, err := c.newMsg(s.requestType)
+	if err != nil {
+		return fmt.Errorf("failed to get new DHCP message: %w", err)
+	}
+	if err := s.setExtraOptions(msg); err != nil {
+		return fmt.Errorf("failed to create DHCP message: %w", err)
+	}
+
+	wait := s.backoff.NextBackOff()
+	if wait == backoff.Stop {
+		return s.stateDeadlineExceeded()
+	}
+
+	receiveDeadline := sentTime.Add(wait)
+	if !s.stateDeadline.IsZero() {
+		receiveDeadline = earliestDeadline(s.stateDeadline, receiveDeadline)
+	}
+
+	// Jump out if deadline expires in less than 10ms. Minimum lease time is 1s and if we have less
+	// than 10ms to wait for an answer before switching state it makes no sense to send out another
+	// request. This nearly eliminates the problem of sending two different requests back-to-back.
+	if receiveDeadline.Add(-10 * time.Millisecond).Before(sentTime) {
+		return s.stateDeadlineExceeded()
+	}
+
+	if err := s.transport.Send(msg); err != nil {
+		return fmt.Errorf("failed to send message: %w", err)
+	}
+
+	if err := s.transport.SetReceiveDeadline(receiveDeadline); err != nil {
+		return fmt.Errorf("failed to set deadline: %w", err)
+	}
+
+	for {
+		offer, err := s.transport.Receive()
+		select {
+		case <-s.ctx.Done():
+			c.cleanup()
+			return s.ctx.Err()
+		default:
+		}
+		if errors.Is(err, transport.DeadlineExceededErr) {
+			return nil
+		}
+		var e transport.InvalidMessageError
+		if errors.As(err, &e) {
+			// Packet couldn't be read. Maybe log at some point in the future.
+			continue
+		}
+		if err != nil {
+			return fmt.Errorf("failed to receive packet: %w", err)
+		}
+		if offer.TransactionID != msg.TransactionID { // Not our transaction
+			continue
+		}
+		err = s.handleMessage(offer, sentTime)
+		if err == nil {
+			return nil
+		} else if !errors.Is(err, InvalidMsgErr) {
+			return err
+		}
+	}
+}
+
+var InvalidMsgErr = errors.New("invalid message")
+
+func (c *Client) runState(ctx context.Context) error {
+	switch c.state {
+	case stateDiscovering:
+		return c.runTransactionState(transactionStateSpec{
+			ctx:         ctx,
+			transport:   c.broadcastConn,
+			backoff:     c.DiscoverBackoff,
+			requestType: dhcpv4.MessageTypeDiscover,
+			setExtraOptions: func(msg *dhcpv4.DHCPv4) error {
+				msg.UpdateOption(dhcpv4.OptGeneric(dhcpv4.OptionRapidCommit, []byte{}))
+				return nil
+			},
+			handleMessage: func(offer *dhcpv4.DHCPv4, sentTime time.Time) error {
+				switch offer.MessageType() {
+				case dhcpv4.MessageTypeOffer:
+					if c.acceptableLease(offer) {
+						c.offer = offer
+						c.AcceptOfferBackoff.Reset()
+						c.state = stateRequesting
+						return nil
+					}
+				case dhcpv4.MessageTypeAck:
+					if c.acceptableLease(offer) {
+						return c.transitionToBound(offer, sentTime)
+					}
+				}
+				return InvalidMsgErr
+			},
+		})
+	case stateRequesting:
+		return c.runTransactionState(transactionStateSpec{
+			ctx:         ctx,
+			transport:   c.broadcastConn,
+			backoff:     c.AcceptOfferBackoff,
+			requestType: dhcpv4.MessageTypeRequest,
+			setExtraOptions: func(msg *dhcpv4.DHCPv4) error {
+				msg.UpdateOption(dhcpv4.OptServerIdentifier(c.offer.ServerIdentifier()))
+				msg.TransactionID = c.offer.TransactionID
+				msg.UpdateOption(dhcpv4.OptRequestedIPAddress(c.offer.YourIPAddr))
+				return nil
+			},
+			handleMessage: func(msg *dhcpv4.DHCPv4, sentTime time.Time) error {
+				switch msg.MessageType() {
+				case dhcpv4.MessageTypeAck:
+					if c.acceptableLease(msg) {
+						return c.transitionToBound(msg, sentTime)
+					}
+				case dhcpv4.MessageTypeNak:
+					c.requestingToDiscovering()
+					return nil
+				}
+				return InvalidMsgErr
+			},
+			stateDeadlineExceeded: func() error {
+				c.requestingToDiscovering()
+				return nil
+			},
+		})
+	case stateBound:
+		select {
+		case <-time.After(c.leaseBoundDeadline.Sub(c.now())):
+			c.state = stateRenewing
+			c.RenewBackoff.Reset()
+			return nil
+		case <-ctx.Done():
+			c.cleanup()
+			return ctx.Err()
+		}
+	case stateRenewing:
+		return c.runTransactionState(transactionStateSpec{
+			ctx:           ctx,
+			transport:     c.unicastConn,
+			backoff:       c.RenewBackoff,
+			requestType:   dhcpv4.MessageTypeRequest,
+			stateDeadline: c.leaseRenewDeadline,
+			setExtraOptions: func(msg *dhcpv4.DHCPv4) error {
+				msg.ClientIPAddr = c.lease.YourIPAddr
+				return nil
+			},
+			handleMessage: func(ack *dhcpv4.DHCPv4, sentTime time.Time) error {
+				switch ack.MessageType() {
+				case dhcpv4.MessageTypeAck:
+					if c.acceptableLease(ack) {
+						return c.transitionToBound(ack, sentTime)
+					}
+				case dhcpv4.MessageTypeNak:
+					return c.leaseToDiscovering()
+				}
+				return InvalidMsgErr
+			},
+			stateDeadlineExceeded: func() error {
+				c.state = stateRebinding
+				if err := c.switchToBroadcast(); err != nil {
+					return fmt.Errorf("failed to switch to broadcast: %w", err)
+				}
+				c.RebindBackoff.Reset()
+				return nil
+			},
+		})
+	case stateRebinding:
+		return c.runTransactionState(transactionStateSpec{
+			ctx:           ctx,
+			transport:     c.broadcastConn,
+			backoff:       c.RebindBackoff,
+			stateDeadline: c.leaseDeadline,
+			requestType:   dhcpv4.MessageTypeRequest,
+			setExtraOptions: func(msg *dhcpv4.DHCPv4) error {
+				msg.ClientIPAddr = c.lease.YourIPAddr
+				return nil
+			},
+			handleMessage: func(ack *dhcpv4.DHCPv4, sentTime time.Time) error {
+				switch ack.MessageType() {
+				case dhcpv4.MessageTypeAck:
+					if c.acceptableLease(ack) {
+						return c.transitionToBound(ack, sentTime)
+					}
+				case dhcpv4.MessageTypeNak:
+					return c.leaseToDiscovering()
+				}
+				return InvalidMsgErr
+			},
+			stateDeadlineExceeded: func() error {
+				return c.leaseToDiscovering()
+			},
+		})
+	}
+	return errors.New("state machine in invalid state")
+}
+
+func (c *Client) Run(ctx context.Context) error {
+	if c.LeaseCallback == nil {
+		panic("LeaseCallback must be set before calling Run")
+	}
+	logger := supervisor.Logger(ctx)
+	for {
+		oldState := c.state
+		if err := c.runState(ctx); err != nil {
+			return err
+		}
+		if c.state != oldState {
+			logger.Infof("%s => %s", oldState, c.state)
+		}
+	}
+}
+
+func (c *Client) cleanup() {
+	c.unicastConn.Close()
+	if c.lease != nil {
+		c.LeaseCallback(leaseFromAck(c.lease, c.leaseDeadline), nil)
+	}
+	c.broadcastConn.Close()
+}
+
+func (c *Client) requestingToDiscovering() {
+	c.offer = nil
+	c.DiscoverBackoff.Reset()
+	c.state = stateDiscovering
+}
+
+func (c *Client) leaseToDiscovering() error {
+	if c.state == stateRenewing {
+		if err := c.switchToBroadcast(); err != nil {
+			return err
+		}
+	}
+	c.state = stateDiscovering
+	c.DiscoverBackoff.Reset()
+	if err := c.LeaseCallback(leaseFromAck(c.lease, c.leaseDeadline), nil); err != nil {
+		return fmt.Errorf("lease callback failed: %w", err)
+	}
+	c.lease = nil
+	return nil
+}
+
+func leaseFromAck(ack *dhcpv4.DHCPv4, expiresAt time.Time) *Lease {
+	if ack == nil {
+		return nil
+	}
+	return &Lease{Options: ack.Options, AssignedIP: ack.YourIPAddr, ExpiresAt: expiresAt}
+}
+
+func (c *Client) transitionToBound(ack *dhcpv4.DHCPv4, sentTime time.Time) error {
+	// Guaranteed to exist, leases without a lease time are filtered
+	leaseTime := ack.IPAddressLeaseTime(0)
+	origLeaseDeadline := c.leaseDeadline
+	c.leaseDeadline = sentTime.Add(leaseTime)
+	c.leaseBoundDeadline = sentTime.Add(ack.IPAddressRenewalTime(time.Duration(float64(leaseTime) * 0.5)))
+	c.leaseRenewDeadline = sentTime.Add(ack.IPAddressRebindingTime(time.Duration(float64(leaseTime) * 0.85)))
+
+	if err := c.LeaseCallback(leaseFromAck(c.lease, origLeaseDeadline), leaseFromAck(ack, c.leaseDeadline)); err != nil {
+		return fmt.Errorf("lease callback failed: %w", err)
+	}
+
+	if c.state != stateRenewing {
+		if err := c.switchToUnicast(ack.ServerIdentifier(), ack.YourIPAddr); err != nil {
+			return fmt.Errorf("failed to switch transports: %w", err)
+		}
+	}
+	c.state = stateBound
+	c.lease = ack
+	return nil
+}
+
+func (c *Client) switchToUnicast(serverIP, bindIP net.IP) error {
+	if err := c.broadcastConn.Close(); err != nil {
+		return fmt.Errorf("failed to close broadcast transport: %w", err)
+	}
+	if err := c.unicastConn.Open(serverIP, bindIP); err != nil {
+		return fmt.Errorf("failed to open unicast transport: %w", err)
+	}
+	return nil
+}
+
+func (c *Client) switchToBroadcast() error {
+	if err := c.unicastConn.Close(); err != nil {
+		return fmt.Errorf("failed to close unicast transport: %w", err)
+	}
+	if err := c.broadcastConn.Open(); err != nil {
+		return fmt.Errorf("failed to open broadcast transport: %w", err)
+	}
+	return nil
+}
diff --git a/metropolis/node/core/network/dhcp4c/dhcpc_test.go b/metropolis/node/core/network/dhcp4c/dhcpc_test.go
new file mode 100644
index 0000000..6988da2
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/dhcpc_test.go
@@ -0,0 +1,514 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dhcp4c
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"testing"
+	"time"
+
+	"github.com/cenkalti/backoff/v4"
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/stretchr/testify/assert"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c/transport"
+)
+
+type fakeTime struct {
+	time time.Time
+}
+
+func newFakeTime(t time.Time) *fakeTime {
+	return &fakeTime{
+		time: t,
+	}
+}
+
+func (ft *fakeTime) Now() time.Time {
+	return ft.time
+}
+
+func (ft *fakeTime) Advance(d time.Duration) {
+	ft.time = ft.time.Add(d)
+}
+
+type mockTransport struct {
+	sentPacket     *dhcpv4.DHCPv4
+	sendError      error
+	setDeadline    time.Time
+	receivePackets []*dhcpv4.DHCPv4
+	receiveError   error
+	receiveIdx     int
+	closed         bool
+}
+
+func (mt *mockTransport) sendPackets(pkts ...*dhcpv4.DHCPv4) {
+	mt.receiveIdx = 0
+	mt.receivePackets = pkts
+}
+
+func (mt *mockTransport) Open() error {
+	mt.closed = false
+	return nil
+}
+
+func (mt *mockTransport) Send(payload *dhcpv4.DHCPv4) error {
+	mt.sentPacket = payload
+	return mt.sendError
+}
+
+func (mt *mockTransport) Receive() (*dhcpv4.DHCPv4, error) {
+	if mt.receiveError != nil {
+		return nil, mt.receiveError
+	}
+	if len(mt.receivePackets) > mt.receiveIdx {
+		packet := mt.receivePackets[mt.receiveIdx]
+		packet, err := dhcpv4.FromBytes(packet.ToBytes()) // Clone packet
+		if err != nil {
+			panic("ToBytes => FromBytes failed")
+		}
+		packet.TransactionID = mt.sentPacket.TransactionID
+		mt.receiveIdx++
+		return packet, nil
+	}
+	return nil, transport.DeadlineExceededErr
+}
+
+func (mt *mockTransport) SetReceiveDeadline(t time.Time) error {
+	mt.setDeadline = t
+	return nil
+}
+
+func (mt *mockTransport) Close() error {
+	mt.closed = true
+	return nil
+}
+
+type unicastMockTransport struct {
+	mockTransport
+	serverIP net.IP
+	bindIP   net.IP
+}
+
+func (umt *unicastMockTransport) Open(serverIP, bindIP net.IP) error {
+	if umt.serverIP != nil {
+		panic("double-open of unicast transport")
+	}
+	umt.serverIP = serverIP
+	umt.bindIP = bindIP
+	return nil
+}
+
+func (umt *unicastMockTransport) Close() error {
+	umt.serverIP = nil
+	umt.bindIP = nil
+	return umt.mockTransport.Close()
+}
+
+type mockBackoff struct {
+	indefinite bool
+	values     []time.Duration
+	idx        int
+}
+
+func newMockBackoff(vals []time.Duration, indefinite bool) *mockBackoff {
+	return &mockBackoff{values: vals, indefinite: indefinite}
+}
+
+func (mb *mockBackoff) NextBackOff() time.Duration {
+	if mb.idx < len(mb.values) || mb.indefinite {
+		val := mb.values[mb.idx%len(mb.values)]
+		mb.idx++
+		return val
+	}
+	return backoff.Stop
+}
+
+func (mb *mockBackoff) Reset() {
+	mb.idx = 0
+}
+
+func TestClient_runTransactionState(t *testing.T) {
+	ft := newFakeTime(time.Date(2020, 10, 28, 15, 02, 32, 352, time.UTC))
+	c := Client{
+		now:   ft.Now,
+		iface: &net.Interface{MTU: 9324, HardwareAddr: net.HardwareAddr{0x12, 0x23, 0x34, 0x45, 0x56, 0x67}},
+	}
+	mt := &mockTransport{}
+	err := c.runTransactionState(transactionStateSpec{
+		ctx:         context.Background(),
+		transport:   mt,
+		backoff:     newMockBackoff([]time.Duration{1 * time.Second}, true),
+		requestType: dhcpv4.MessageTypeDiscover,
+		setExtraOptions: func(msg *dhcpv4.DHCPv4) error {
+			msg.UpdateOption(dhcpv4.OptDomainName("just.testing.invalid"))
+			return nil
+		},
+		handleMessage: func(msg *dhcpv4.DHCPv4, sentTime time.Time) error {
+			return nil
+		},
+		stateDeadlineExceeded: func() error {
+			panic("shouldn't be called")
+		},
+	})
+	assert.NoError(t, err)
+	assert.Equal(t, "just.testing.invalid", mt.sentPacket.DomainName())
+	assert.Equal(t, dhcpv4.MessageTypeDiscover, mt.sentPacket.MessageType())
+}
+
+// TestAcceptableLease tests if a minimal valid lease is accepted by acceptableLease
+func TestAcceptableLease(t *testing.T) {
+	c := Client{}
+	offer := &dhcpv4.DHCPv4{
+		OpCode: dhcpv4.OpcodeBootReply,
+	}
+	offer.UpdateOption(dhcpv4.OptMessageType(dhcpv4.MessageTypeOffer))
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(net.IP{192, 0, 2, 1}))
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(10 * time.Second))
+	offer.YourIPAddr = net.IP{192, 0, 2, 2}
+	assert.True(t, c.acceptableLease(offer), "Valid lease is not acceptable")
+}
+
+type dhcpClientPuppet struct {
+	ft  *fakeTime
+	bmt *mockTransport
+	umt *unicastMockTransport
+	c   *Client
+}
+
+func newPuppetClient(initState state) *dhcpClientPuppet {
+	ft := newFakeTime(time.Date(2020, 10, 28, 15, 02, 32, 352, time.UTC))
+	bmt := &mockTransport{}
+	umt := &unicastMockTransport{}
+	c := &Client{
+		state:              initState,
+		now:                ft.Now,
+		iface:              &net.Interface{MTU: 9324, HardwareAddr: net.HardwareAddr{0x12, 0x23, 0x34, 0x45, 0x56, 0x67}},
+		broadcastConn:      bmt,
+		unicastConn:        umt,
+		DiscoverBackoff:    newMockBackoff([]time.Duration{1 * time.Second}, true),
+		AcceptOfferBackoff: newMockBackoff([]time.Duration{1 * time.Second, 2 * time.Second}, false),
+		RenewBackoff:       newMockBackoff([]time.Duration{1 * time.Second}, true),
+		RebindBackoff:      newMockBackoff([]time.Duration{1 * time.Second}, true),
+	}
+	return &dhcpClientPuppet{
+		ft:  ft,
+		bmt: bmt,
+		umt: umt,
+		c:   c,
+	}
+}
+
+func newResponse(m dhcpv4.MessageType) *dhcpv4.DHCPv4 {
+	o := &dhcpv4.DHCPv4{
+		OpCode: dhcpv4.OpcodeBootReply,
+	}
+	o.UpdateOption(dhcpv4.OptMessageType(m))
+	return o
+}
+
+// TestDiscoverOffer tests if the DHCP state machine in discovering state properly selects the first valid lease
+// and transitions to requesting state
+func TestDiscoverRequesting(t *testing.T) {
+	p := newPuppetClient(stateDiscovering)
+
+	// A minimal valid lease
+	offer := newResponse(dhcpv4.MessageTypeOffer)
+	testIP := net.IP{192, 0, 2, 2}
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(net.IP{192, 0, 2, 1}))
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(10 * time.Second))
+	offer.YourIPAddr = testIP
+
+	// Intentionally bad offer with no lease time.
+	terribleOffer := newResponse(dhcpv4.MessageTypeOffer)
+	terribleOffer.UpdateOption(dhcpv4.OptServerIdentifier(net.IP{192, 0, 2, 2}))
+	terribleOffer.YourIPAddr = net.IPv4(192, 0, 2, 2)
+
+	// Send the bad offer first, then the valid offer
+	p.bmt.sendPackets(terribleOffer, offer)
+
+	if err := p.c.runState(context.Background()); err != nil {
+		t.Error(err)
+	}
+	assert.Equal(t, stateRequesting, p.c.state, "DHCP client didn't process offer")
+	assert.Equal(t, testIP, p.c.offer.YourIPAddr, "DHCP client requested invalid offer")
+}
+
+// TestOfferBound tests if the DHCP state machine in requesting state processes a valid DHCPACK and transitions to
+// bound state.
+func TestRequestingBound(t *testing.T) {
+	p := newPuppetClient(stateRequesting)
+
+	offer := newResponse(dhcpv4.MessageTypeAck)
+	testIP := net.IP{192, 0, 2, 2}
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(net.IP{192, 0, 2, 1}))
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(10 * time.Second))
+	offer.YourIPAddr = testIP
+
+	p.bmt.sendPackets(offer)
+	p.c.offer = offer
+	p.c.LeaseCallback = func(old, new *Lease) error {
+		assert.Nil(t, old, "old lease is not nil for new lease")
+		assert.Equal(t, testIP, new.AssignedIP, "new lease has wrong IP")
+		return nil
+	}
+
+	if err := p.c.runState(context.Background()); err != nil {
+		t.Error(err)
+	}
+	assert.Equal(t, stateBound, p.c.state, "DHCP client didn't process offer")
+	assert.Equal(t, testIP, p.c.lease.YourIPAddr, "DHCP client requested invalid offer")
+}
+
+// TestRequestingDiscover tests if the DHCP state machine in requesting state transitions back to discovering if it
+// takes too long to get a valid DHCPACK.
+func TestRequestingDiscover(t *testing.T) {
+	p := newPuppetClient(stateRequesting)
+
+	offer := newResponse(dhcpv4.MessageTypeOffer)
+	testIP := net.IP{192, 0, 2, 2}
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(net.IP{192, 0, 2, 1}))
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(10 * time.Second))
+	offer.YourIPAddr = testIP
+	p.c.offer = offer
+
+	for i := 0; i < 10; i++ {
+		p.bmt.sendPackets()
+		if err := p.c.runState(context.Background()); err != nil {
+			t.Error(err)
+		}
+		assert.Equal(t, dhcpv4.MessageTypeRequest, p.bmt.sentPacket.MessageType(), "Invalid message type for requesting")
+		if p.c.state == stateDiscovering {
+			break
+		}
+		p.ft.time = p.bmt.setDeadline
+
+		if i == 9 {
+			t.Fatal("Too many tries while requesting, backoff likely wrong")
+		}
+	}
+	assert.Equal(t, stateDiscovering, p.c.state, "DHCP client didn't switch back to offer after requesting expired")
+}
+
+// TestDiscoverRapidCommit tests if the DHCP state machine in discovering state transitions directly to bound if a
+// rapid commit response (DHCPACK) is received.
+func TestDiscoverRapidCommit(t *testing.T) {
+	testIP := net.IP{192, 0, 2, 2}
+	offer := newResponse(dhcpv4.MessageTypeAck)
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(net.IP{192, 0, 2, 1}))
+	leaseTime := 10 * time.Second
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(leaseTime))
+	offer.YourIPAddr = testIP
+
+	p := newPuppetClient(stateDiscovering)
+	p.c.LeaseCallback = func(old, new *Lease) error {
+		assert.Nil(t, old, "old is not nil")
+		assert.Equal(t, testIP, new.AssignedIP, "callback called with wrong IP")
+		assert.Equal(t, p.ft.Now().Add(leaseTime), new.ExpiresAt, "invalid ExpiresAt")
+		return nil
+	}
+	p.bmt.sendPackets(offer)
+	if err := p.c.runState(context.Background()); err != nil {
+		t.Error(err)
+	}
+	assert.Equal(t, stateBound, p.c.state, "DHCP client didn't process offer")
+	assert.Equal(t, testIP, p.c.lease.YourIPAddr, "DHCP client requested invalid offer")
+	assert.Equal(t, 5*time.Second, p.c.leaseBoundDeadline.Sub(p.ft.Now()), "Renewal time was incorrectly defaulted")
+}
+
+type TestOption uint8
+
+func (o TestOption) Code() uint8 {
+	return uint8(o) + 224 // Private options
+}
+func (o TestOption) String() string {
+	return fmt.Sprintf("Test Option %d", uint8(o))
+}
+
+// TestBoundRenewingBound tests if the DHCP state machine in bound correctly transitions to renewing after
+// leaseBoundDeadline expires, sends a DHCPREQUEST and after it gets a DHCPACK response calls LeaseCallback and
+// transitions back to bound with correct new deadlines.
+func TestBoundRenewingBound(t *testing.T) {
+	offer := newResponse(dhcpv4.MessageTypeAck)
+	testIP := net.IP{192, 0, 2, 2}
+	serverIP := net.IP{192, 0, 2, 1}
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(serverIP))
+	leaseTime := 10 * time.Second
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(leaseTime))
+	offer.YourIPAddr = testIP
+
+	p := newPuppetClient(stateBound)
+	p.umt.Open(serverIP, testIP)
+	p.c.lease, _ = dhcpv4.FromBytes(offer.ToBytes())
+	// Other deadlines are intentionally empty to make sure they aren't used
+	p.c.leaseRenewDeadline = p.ft.Now().Add(8500 * time.Millisecond)
+	p.c.leaseBoundDeadline = p.ft.Now().Add(5000 * time.Millisecond)
+
+	p.ft.Advance(5*time.Second - 5*time.Millisecond)
+	if err := p.c.runState(context.Background()); err != nil {
+		t.Error(err)
+	}
+	p.ft.Advance(5 * time.Millisecond) // We cannot intercept time.After so we just advance the clock by the time slept
+	assert.Equal(t, stateRenewing, p.c.state, "DHCP client not renewing")
+	offer.UpdateOption(dhcpv4.OptGeneric(TestOption(1), []byte{0x12}))
+	p.umt.sendPackets(offer)
+	p.c.LeaseCallback = func(old, new *Lease) error {
+		assert.Equal(t, testIP, old.AssignedIP, "callback called with wrong old IP")
+		assert.Equal(t, testIP, new.AssignedIP, "callback called with wrong IP")
+		assert.Equal(t, p.ft.Now().Add(leaseTime), new.ExpiresAt, "invalid ExpiresAt")
+		assert.Empty(t, old.Options.Get(TestOption(1)), "old contains options from new")
+		assert.Equal(t, []byte{0x12}, new.Options.Get(TestOption(1)), "renewal didn't add new option")
+		return nil
+	}
+	if err := p.c.runState(context.Background()); err != nil {
+		t.Error(err)
+	}
+	assert.Equal(t, stateBound, p.c.state, "DHCP client didn't renew")
+	assert.Equal(t, p.ft.Now().Add(leaseTime), p.c.leaseDeadline, "lease deadline not updated")
+	assert.Equal(t, dhcpv4.MessageTypeRequest, p.umt.sentPacket.MessageType(), "Invalid message type for renewal")
+}
+
+// TestRenewingRebinding tests if the DHCP state machine in renewing state correctly sends DHCPREQUESTs and transitions
+// to the rebinding state when it hasn't received a valid response until the deadline expires.
+func TestRenewingRebinding(t *testing.T) {
+	offer := newResponse(dhcpv4.MessageTypeAck)
+	testIP := net.IP{192, 0, 2, 2}
+	serverIP := net.IP{192, 0, 2, 1}
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(serverIP))
+	leaseTime := 10 * time.Second
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(leaseTime))
+	offer.YourIPAddr = testIP
+
+	p := newPuppetClient(stateRenewing)
+	p.umt.Open(serverIP, testIP)
+	p.c.lease, _ = dhcpv4.FromBytes(offer.ToBytes())
+	// Other deadlines are intentionally empty to make sure they aren't used
+	p.c.leaseRenewDeadline = p.ft.Now().Add(8500 * time.Millisecond)
+	p.c.leaseDeadline = p.ft.Now().Add(10000 * time.Millisecond)
+
+	startTime := p.ft.Now()
+	p.ft.Advance(5 * time.Second)
+
+	p.c.LeaseCallback = func(old, new *Lease) error {
+		t.Fatal("Lease callback called without valid offer")
+		return nil
+	}
+
+	for i := 0; i < 10; i++ {
+		p.umt.sendPackets()
+		if err := p.c.runState(context.Background()); err != nil {
+			t.Error(err)
+		}
+		assert.Equal(t, dhcpv4.MessageTypeRequest, p.umt.sentPacket.MessageType(), "Invalid message type for renewal")
+		p.ft.time = p.umt.setDeadline
+
+		if p.c.state == stateRebinding {
+			break
+		}
+		if i == 9 {
+			t.Fatal("Too many tries while renewing, backoff likely wrong")
+		}
+	}
+	assert.Equal(t, startTime.Add(8500*time.Millisecond), p.umt.setDeadline, "wrong listen deadline when renewing")
+	assert.Equal(t, stateRebinding, p.c.state, "DHCP client not renewing")
+	assert.False(t, p.bmt.closed)
+	assert.True(t, p.umt.closed)
+}
+
+// TestRebindingBound tests if the DHCP state machine in rebinding state sends DHCPREQUESTs to the network and if
+// it receives a valid DHCPACK correctly transitions back to bound state.
+func TestRebindingBound(t *testing.T) {
+	offer := newResponse(dhcpv4.MessageTypeAck)
+	testIP := net.IP{192, 0, 2, 2}
+	serverIP := net.IP{192, 0, 2, 1}
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(serverIP))
+	leaseTime := 10 * time.Second
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(leaseTime))
+	offer.YourIPAddr = testIP
+
+	p := newPuppetClient(stateRebinding)
+	p.c.lease, _ = dhcpv4.FromBytes(offer.ToBytes())
+	// Other deadlines are intentionally empty to make sure they aren't used
+	p.c.leaseDeadline = p.ft.Now().Add(10000 * time.Millisecond)
+
+	p.ft.Advance(9 * time.Second)
+	if err := p.c.runState(context.Background()); err != nil {
+		t.Error(err)
+	}
+	assert.Equal(t, dhcpv4.MessageTypeRequest, p.bmt.sentPacket.MessageType(), "DHCP rebind sent invalid message type")
+	assert.Equal(t, stateRebinding, p.c.state, "DHCP client transferred out of rebinding state without trigger")
+	offer.UpdateOption(dhcpv4.OptGeneric(TestOption(1), []byte{0x12})) // Mark answer
+	p.bmt.sendPackets(offer)
+	p.bmt.sentPacket = nil
+	p.c.LeaseCallback = func(old, new *Lease) error {
+		assert.Equal(t, testIP, old.AssignedIP, "callback called with wrong old IP")
+		assert.Equal(t, testIP, new.AssignedIP, "callback called with wrong IP")
+		assert.Equal(t, p.ft.Now().Add(leaseTime), new.ExpiresAt, "invalid ExpiresAt")
+		assert.Empty(t, old.Options.Get(TestOption(1)), "old contains options from new")
+		assert.Equal(t, []byte{0x12}, new.Options.Get(TestOption(1)), "renewal didn't add new option")
+		return nil
+	}
+	if err := p.c.runState(context.Background()); err != nil {
+		t.Error(err)
+	}
+	assert.Equal(t, dhcpv4.MessageTypeRequest, p.bmt.sentPacket.MessageType())
+	assert.Equal(t, stateBound, p.c.state, "DHCP client didn't go back to bound")
+}
+
+// TestRebindingBound tests if the DHCP state machine in rebinding state transitions to discovering state if
+// leaseDeadline expires and calls LeaseCallback with an empty new lease.
+func TestRebindingDiscovering(t *testing.T) {
+	offer := newResponse(dhcpv4.MessageTypeAck)
+	testIP := net.IP{192, 0, 2, 2}
+	serverIP := net.IP{192, 0, 2, 1}
+	offer.UpdateOption(dhcpv4.OptServerIdentifier(serverIP))
+	leaseTime := 10 * time.Second
+	offer.UpdateOption(dhcpv4.OptIPAddressLeaseTime(leaseTime))
+	offer.YourIPAddr = testIP
+
+	p := newPuppetClient(stateRebinding)
+	p.c.lease, _ = dhcpv4.FromBytes(offer.ToBytes())
+	// Other deadlines are intentionally empty to make sure they aren't used
+	p.c.leaseDeadline = p.ft.Now().Add(10000 * time.Millisecond)
+
+	p.ft.Advance(9 * time.Second)
+	p.c.LeaseCallback = func(old, new *Lease) error {
+		assert.Equal(t, testIP, old.AssignedIP, "callback called with wrong old IP")
+		assert.Nil(t, new, "transition to discovering didn't clear new lease on callback")
+		return nil
+	}
+	for i := 0; i < 10; i++ {
+		p.bmt.sendPackets()
+		p.bmt.sentPacket = nil
+		if err := p.c.runState(context.Background()); err != nil {
+			t.Error(err)
+		}
+		if p.c.state == stateDiscovering {
+			assert.Nil(t, p.bmt.sentPacket)
+			break
+		}
+		assert.Equal(t, dhcpv4.MessageTypeRequest, p.bmt.sentPacket.MessageType(), "Invalid message type for rebind")
+		p.ft.time = p.bmt.setDeadline
+		if i == 9 {
+			t.Fatal("Too many tries while rebinding, backoff likely wrong")
+		}
+	}
+	assert.Nil(t, p.c.lease, "Lease not zeroed on transition to discovering")
+	assert.Equal(t, stateDiscovering, p.c.state, "DHCP client didn't transition to discovering after loosing lease")
+}
diff --git a/metropolis/node/core/network/dhcp4c/doc.go b/metropolis/node/core/network/dhcp4c/doc.go
new file mode 100644
index 0000000..b270c7b
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/doc.go
@@ -0,0 +1,53 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package dhcp4c provides a client implementation of the DHCPv4 protocol (RFC2131) and a few extensions for Linux-based
+// systems.
+// The code is split into three main parts:
+// - The core DHCP state machine, which lives in dhcpc.go
+// - Mechanisms to send and receive DHCP messages, which live in transport/
+// - Standard callbacks which implement necessary kernel configuration steps in a simple and standalone way living in
+//   callback/
+//
+// Since the DHCP protocol is ugly and underspecified (see https://tools.ietf.org/html/draft-ietf-dhc-implementation-02
+// for a subset of known issues), this client slightly bends the specification in the following cases:
+// - IP fragmentation for DHCP messages is not supported for both sending and receiving messages
+//   This is because the major servers (ISC, dnsmasq, ...) do not implement it and just drop fragmented packets, so it
+//   would be counterproductive to try to send them. The client just attempts to send the full message and hopes it
+//   passes through to the server.
+// - The suggested timeouts and wait periods have been tightened significantly. When the standard was written 10Mbps
+//   Ethernet with hubs was a common interconnect. Using these would make the client extremely slow on today's
+//   1Gbps+ networks.
+// - Wrong data in DHCP responses is fixed up if possible. This fixing includes dropping prohibited options, clamping
+//   semantically invalid data and defaulting not set options as far as it's possible. Non-recoverable responses
+//   (for example because a non-Unicast IP is handed out or lease time is not set or zero) are still ignored.
+//   All data which can be stored in both DHCP fields and options is also normalized to the corresponding option.
+// - Duplicate Address Detection is not implemented by default. It's slow, hard to implement correctly and generally
+//   not necessary on modern networks as the servers already waste time checking for duplicate addresses. It's possible
+//   to hook it in via a LeaseCallback if necessary in a given application.
+//
+// Operationally, there's one known caveat to using this client: If the lease offered during the select phase (in a
+// DHCPOFFER) is not the same as the one sent in the following DHCPACK the first one might be acceptable, but the second
+// one might not be. This can cause pathological behavior where the client constantly switches between discovering and
+// requesting states. Depending on the reuse policies on the DHCP server this can cause the client to consume all
+// available IP addresses. Sadly there's no good way of fixing this within the boundaries of the protocol. A DHCPRELEASE
+// for the adresse would need to be unicasted so the unaccepable address would need to be configured which can be either
+// impossible if it's not valid or not acceptable from a security standpoint (for example because it overlaps with a
+// prefix used internally) and a DHCPDECLINE would cause the server to blacklist the IP thus also depleting the IP pool.
+// This could be potentially avoided by originating DHCPRELEASE packages from a userspace transport, but said transport
+// would need to be routing- and PMTU-aware which would make it even more complicated than the existing
+// BroadcastTransport.
+package dhcp4c
diff --git a/metropolis/node/core/network/dhcp4c/lease.go b/metropolis/node/core/network/dhcp4c/lease.go
new file mode 100644
index 0000000..c56270c
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/lease.go
@@ -0,0 +1,124 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dhcp4c
+
+import (
+	"encoding/binary"
+	"net"
+	"time"
+
+	"github.com/insomniacslk/dhcp/dhcpv4"
+)
+
+// Lease represents a DHCPv4 lease. It only consists of an IP, an expiration timestamp and options as all other
+// relevant parts of the message have been normalized into their respective options. It also contains some smart
+// getters for commonly-used options which extract only valid information from options.
+type Lease struct {
+	AssignedIP net.IP
+	ExpiresAt  time.Time
+	Options    dhcpv4.Options
+}
+
+// SubnetMask returns the SubnetMask option or the default mask if not set or invalid.
+// It returns nil if the lease is nil.
+func (l *Lease) SubnetMask() net.IPMask {
+	if l == nil {
+		return nil
+	}
+	mask := net.IPMask(dhcpv4.GetIP(dhcpv4.OptionSubnetMask, l.Options))
+	if _, bits := mask.Size(); bits != 32 { // If given mask is not valid, use the default mask
+		mask = l.AssignedIP.DefaultMask()
+	}
+	return mask
+}
+
+// IPNet returns an IPNet from the assigned IP and subnet mask.
+// It returns nil if the lease is nil.
+func (l *Lease) IPNet() *net.IPNet {
+	if l == nil {
+		return nil
+	}
+	return &net.IPNet{
+		IP:   l.AssignedIP,
+		Mask: l.SubnetMask(),
+	}
+}
+
+// Router returns the first valid router from the DHCP router option or nil if none such exists.
+// It returns nil if the lease is nil.
+func (l *Lease) Router() net.IP {
+	if l == nil {
+		return nil
+	}
+	routers := dhcpv4.GetIPs(dhcpv4.OptionRouter, l.Options)
+	for _, r := range routers {
+		if r.IsGlobalUnicast() || r.IsLinkLocalUnicast() {
+			return r
+		}
+	}
+	// No (valid) router found
+	return nil
+}
+
+// DNSServers represents an ordered collection of DNS servers
+type DNSServers []net.IP
+
+func (a DNSServers) Equal(b DNSServers) bool {
+	if len(a) == len(b) {
+		if len(a) == 0 {
+			return true // both are empty or nil
+		}
+		for i, aVal := range a {
+			if !aVal.Equal(b[i]) {
+				return false
+			}
+		}
+		return true
+	}
+	return false
+
+}
+
+func ip4toInt(ip net.IP) uint32 {
+	ip4 := ip.To4()
+	if ip4 == nil {
+		return 0
+	}
+	return binary.BigEndian.Uint32(ip4)
+}
+
+// DNSServers returns all unique valid DNS servers from the DHCP DomainNameServers options.
+// It returns nil if the lease is nil.
+func (l *Lease) DNSServers() DNSServers {
+	if l == nil {
+		return nil
+	}
+	rawServers := dhcpv4.GetIPs(dhcpv4.OptionDomainNameServer, l.Options)
+	var servers DNSServers
+	serversSeenMap := make(map[uint32]struct{})
+	for _, s := range rawServers {
+		ip4Num := ip4toInt(s)
+		if s.IsGlobalUnicast() || s.IsLinkLocalUnicast() || ip4Num != 0 {
+			if _, ok := serversSeenMap[ip4Num]; ok {
+				continue
+			}
+			serversSeenMap[ip4Num] = struct{}{}
+			servers = append(servers, s)
+		}
+	}
+	return servers
+}
diff --git a/metropolis/node/core/network/dhcp4c/lease_test.go b/metropolis/node/core/network/dhcp4c/lease_test.go
new file mode 100644
index 0000000..823656f
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/lease_test.go
@@ -0,0 +1,55 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dhcp4c
+
+import (
+	"net"
+	"testing"
+
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestLeaseDHCPServers(t *testing.T) {
+	var tests = []struct {
+		name     string
+		lease    *Lease
+		expected DNSServers
+	}{{
+		name:     "ReturnsNilWithNoLease",
+		lease:    nil,
+		expected: nil,
+	}, {
+		name: "DiscardsInvalidIPs",
+		lease: &Lease{
+			Options: dhcpv4.OptionsFromList(dhcpv4.OptDNS(net.IP{0, 0, 0, 0})),
+		},
+		expected: nil,
+	}, {
+		name: "DeduplicatesIPs",
+		lease: &Lease{
+			Options: dhcpv4.OptionsFromList(dhcpv4.OptDNS(net.IP{192, 0, 2, 1}, net.IP{192, 0, 2, 2}, net.IP{192, 0, 2, 1})),
+		},
+		expected: DNSServers{net.IP{192, 0, 2, 1}, net.IP{192, 0, 2, 2}},
+	}}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			res := test.lease.DNSServers()
+			assert.Equal(t, test.expected, res)
+		})
+	}
+}
diff --git a/metropolis/node/core/network/dhcp4c/transport/BUILD.bazel b/metropolis/node/core/network/dhcp4c/transport/BUILD.bazel
new file mode 100644
index 0000000..edd47a1
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/transport/BUILD.bazel
@@ -0,0 +1,20 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "transport.go",
+        "transport_broadcast.go",
+        "transport_unicast.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c/transport",
+    visibility = ["//visibility:public"],
+    deps = [
+        "@com_github_google_gopacket//:go_default_library",
+        "@com_github_google_gopacket//layers:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4:go_default_library",
+        "@com_github_mdlayher_raw//:go_default_library",
+        "@org_golang_x_net//bpf:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/network/dhcp4c/transport/transport.go b/metropolis/node/core/network/dhcp4c/transport/transport.go
new file mode 100644
index 0000000..8f5f791
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/transport/transport.go
@@ -0,0 +1,50 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package transport contains Linux-based transports for the DHCP broadcast and unicast
+// specifications.
+package transport
+
+import (
+	"errors"
+	"fmt"
+	"net"
+)
+
+var DeadlineExceededErr = errors.New("deadline exceeded")
+
+func NewInvalidMessageError(internalErr error) error {
+	return &InvalidMessageError{internalErr: internalErr}
+}
+
+type InvalidMessageError struct {
+	internalErr error
+}
+
+func (i InvalidMessageError) Error() string {
+	return fmt.Sprintf("received invalid packet: %v", i.internalErr.Error())
+}
+
+func (i InvalidMessageError) Unwrap() error {
+	return i.internalErr
+}
+
+func deadlineFromTimeout(err error) error {
+	if timeoutErr, ok := err.(net.Error); ok && timeoutErr.Timeout() {
+		return DeadlineExceededErr
+	}
+	return err
+}
diff --git a/metropolis/node/core/network/dhcp4c/transport/transport_broadcast.go b/metropolis/node/core/network/dhcp4c/transport/transport_broadcast.go
new file mode 100644
index 0000000..79fad7d
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/transport/transport_broadcast.go
@@ -0,0 +1,207 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package transport
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"net"
+	"time"
+
+	"github.com/google/gopacket"
+	"github.com/google/gopacket/layers"
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/mdlayher/raw"
+	"golang.org/x/net/bpf"
+)
+
+const (
+	// RFC2474 Section 4.2.2.1 with reference to RFC791 Section 3.1 (Network Control Precedence)
+	dscpCS7 = 0x7 << 3
+
+	// IPv4 MTU
+	maxIPv4MTU = math.MaxUint16 // IPv4 "Total Length" field is an unsigned 16 bit integer
+)
+
+// mustAssemble calls bpf.Assemble and panics if it retuns an error.
+func mustAssemble(insns []bpf.Instruction) []bpf.RawInstruction {
+	rawInsns, err := bpf.Assemble(insns)
+	if err != nil {
+		panic("mustAssemble failed to assemble BPF: " + err.Error())
+	}
+	return rawInsns
+}
+
+// BPF filter for UDP in IPv4 with destination port 68 (DHCP Client)
+//
+// This is used to make the kernel drop non-DHCP traffic for us so that we don't have to handle
+// excessive unrelated traffic flowing on a given link which might overwhelm the single-threaded
+// receiver.
+var bpfFilterInstructions = []bpf.Instruction{
+	// Check IP protocol version equals 4 (first 4 bits of the first byte)
+	// With Ethernet II framing, this is more of a sanity check. We already request the kernel
+	// to only return EtherType 0x0800 (IPv4) frames.
+	bpf.LoadAbsolute{Off: 0, Size: 1},
+	bpf.ALUOpConstant{Op: bpf.ALUOpAnd, Val: 0xf0}, // SubnetMask second 4 bits
+	bpf.JumpIf{Cond: bpf.JumpEqual, Val: 4 << 4, SkipTrue: 1},
+	bpf.RetConstant{Val: 0}, // Discard
+
+	// Check IPv4 Protocol byte (offset 9) equals UDP
+	bpf.LoadAbsolute{Off: 9, Size: 1},
+	bpf.JumpIf{Cond: bpf.JumpEqual, Val: uint32(layers.IPProtocolUDP), SkipTrue: 1},
+	bpf.RetConstant{Val: 0}, // Discard
+
+	// Check if IPv4 fragment offset is all-zero (this is not a fragment)
+	bpf.LoadAbsolute{Off: 6, Size: 2},
+	bpf.JumpIf{Cond: bpf.JumpBitsSet, Val: 0x1fff, SkipFalse: 1},
+	bpf.RetConstant{Val: 0}, // Discard
+
+	// Load IPv4 header size from offset zero and store it into X
+	bpf.LoadMemShift{Off: 0},
+
+	// Check if UDP header destination port equals 68
+	bpf.LoadIndirect{Off: 2, Size: 2}, // Offset relative to header size in register X
+	bpf.JumpIf{Cond: bpf.JumpEqual, Val: 68, SkipTrue: 1},
+	bpf.RetConstant{Val: 0}, // Discard
+
+	// Accept packet and pass through up maximum IP packet size
+	bpf.RetConstant{Val: maxIPv4MTU},
+}
+
+var bpfFilter = mustAssemble(bpfFilterInstructions)
+
+// BroadcastTransport implements a DHCP transport based on a custom IP/UDP stack fulfilling the
+// specific requirements for broadcasting DHCP packets (like all-zero source address, no ARP, ...)
+type BroadcastTransport struct {
+	rawConn *raw.Conn
+	iface   *net.Interface
+}
+
+func NewBroadcastTransport(iface *net.Interface) *BroadcastTransport {
+	return &BroadcastTransport{iface: iface}
+}
+
+func (t *BroadcastTransport) Open() error {
+	if t.rawConn != nil {
+		return errors.New("broadcast transport already open")
+	}
+	rawConn, err := raw.ListenPacket(t.iface, uint16(layers.EthernetTypeIPv4), &raw.Config{
+		LinuxSockDGRAM: true,
+		Filter:         bpfFilter,
+	})
+	if err != nil {
+		return fmt.Errorf("failed to create raw listener: %w", err)
+	}
+	t.rawConn = rawConn
+	return nil
+}
+
+func (t *BroadcastTransport) Send(payload *dhcpv4.DHCPv4) error {
+	if t.rawConn == nil {
+		return errors.New("broadcast transport closed")
+	}
+	packet := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{
+		ComputeChecksums: true,
+		FixLengths:       true,
+	}
+
+	ipLayer := &layers.IPv4{
+		Version:  4,
+		TOS:      dscpCS7 << 2, // Shift left of ECN field
+		TTL:      1,            // These packets should never be routed (their IP headers contain garbage)
+		Protocol: layers.IPProtocolUDP,
+		Flags:    layers.IPv4DontFragment, // Most DHCP servers don't support fragmented packets
+		DstIP:    net.IPv4bcast,
+		SrcIP:    net.IPv4zero,
+	}
+	udpLayer := &layers.UDP{
+		DstPort: 67,
+		SrcPort: 68,
+	}
+	if err := udpLayer.SetNetworkLayerForChecksum(ipLayer); err != nil {
+		panic("Invalid layer stackup encountered")
+	}
+
+	err := gopacket.SerializeLayers(packet, opts,
+		ipLayer,
+		udpLayer,
+		gopacket.Payload(payload.ToBytes()))
+
+	if err != nil {
+		return fmt.Errorf("failed to assemble packet: %w", err)
+	}
+
+	_, err = t.rawConn.WriteTo(packet.Bytes(), &raw.Addr{HardwareAddr: layers.EthernetBroadcast})
+	if err != nil {
+		return fmt.Errorf("failed to transmit broadcast packet: %w", err)
+	}
+	return nil
+}
+
+func (t *BroadcastTransport) Receive() (*dhcpv4.DHCPv4, error) {
+	if t.rawConn == nil {
+		return nil, errors.New("broadcast transport closed")
+	}
+	buf := make([]byte, math.MaxUint16) // Maximum IP packet size
+	n, _, err := t.rawConn.ReadFrom(buf)
+	if err != nil {
+		return nil, deadlineFromTimeout(err)
+	}
+	respPacket := gopacket.NewPacket(buf[:n], layers.LayerTypeIPv4, gopacket.Default)
+	ipLayer := respPacket.Layer(layers.LayerTypeIPv4)
+	if ipLayer == nil {
+		return nil, NewInvalidMessageError(errors.New("got invalid IP packet"))
+	}
+	ip := ipLayer.(*layers.IPv4)
+	if ip.Flags&layers.IPv4MoreFragments != 0 {
+		return nil, NewInvalidMessageError(errors.New("got fragmented message"))
+	}
+
+	udpLayer := respPacket.Layer(layers.LayerTypeUDP)
+	if udpLayer == nil {
+		return nil, NewInvalidMessageError(errors.New("got non-UDP packet"))
+	}
+	udp := udpLayer.(*layers.UDP)
+	if udp.DstPort != 68 {
+		return nil, NewInvalidMessageError(errors.New("message not for DHCP client port"))
+	}
+	msg, err := dhcpv4.FromBytes(udp.Payload)
+	if err != nil {
+		return nil, NewInvalidMessageError(fmt.Errorf("failed to decode DHCPv4 message: %w", err))
+	}
+	return msg, nil
+}
+
+func (t *BroadcastTransport) Close() error {
+	if t.rawConn == nil {
+		return nil
+	}
+	if err := t.rawConn.Close(); err != nil {
+		return err
+	}
+	t.rawConn = nil
+	return nil
+}
+
+func (t *BroadcastTransport) SetReceiveDeadline(deadline time.Time) error {
+	if t.rawConn == nil {
+		return errors.New("broadcast transport closed")
+	}
+	return t.rawConn.SetReadDeadline(deadline)
+}
diff --git a/metropolis/node/core/network/dhcp4c/transport/transport_unicast.go b/metropolis/node/core/network/dhcp4c/transport/transport_unicast.go
new file mode 100644
index 0000000..bf2b3a4
--- /dev/null
+++ b/metropolis/node/core/network/dhcp4c/transport/transport_unicast.go
@@ -0,0 +1,122 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package transport
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"net"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"golang.org/x/sys/unix"
+)
+
+// UnicastTransport implements a DHCP transport based on a normal Linux UDP socket with some custom
+// socket options to influence DSCP and routing.
+type UnicastTransport struct {
+	udpConn  *net.UDPConn
+	targetIP net.IP
+	iface    *net.Interface
+}
+
+func NewUnicastTransport(iface *net.Interface) *UnicastTransport {
+	return &UnicastTransport{
+		iface: iface,
+	}
+}
+
+func (t *UnicastTransport) Open(serverIP, bindIP net.IP) error {
+	if t.udpConn != nil {
+		return errors.New("unicast transport already open")
+	}
+	rawFd, err := unix.Socket(unix.AF_INET, unix.SOCK_DGRAM, 0)
+	if err != nil {
+		return fmt.Errorf("failed to get socket: %w", err)
+	}
+	if err := unix.BindToDevice(rawFd, t.iface.Name); err != nil {
+		return fmt.Errorf("failed to bind UDP interface to device: %w", err)
+	}
+	if err := unix.SetsockoptByte(rawFd, unix.SOL_IP, unix.IP_TOS, dscpCS7<<2); err != nil {
+		return fmt.Errorf("failed to set DSCP CS7: %w", err)
+	}
+	var addr [4]byte
+	copy(addr[:], bindIP.To4())
+	if err := unix.Bind(rawFd, &unix.SockaddrInet4{Addr: addr, Port: 68}); err != nil {
+		return fmt.Errorf("failed to bind UDP unicast interface: %w", err)
+	}
+	filePtr := os.NewFile(uintptr(rawFd), "dhcp-udp")
+	defer filePtr.Close()
+	conn, err := net.FileConn(filePtr)
+	if err != nil {
+		return fmt.Errorf("failed to initialize runtime-supported UDP connection: %w", err)
+	}
+	realConn, ok := conn.(*net.UDPConn)
+	if !ok {
+		panic("UDP socket imported into Go runtime is no longer a UDP socket")
+	}
+	t.udpConn = realConn
+	t.targetIP = serverIP
+	return nil
+}
+
+func (t *UnicastTransport) Send(payload *dhcpv4.DHCPv4) error {
+	if t.udpConn == nil {
+		return errors.New("unicast transport closed")
+	}
+	_, _, err := t.udpConn.WriteMsgUDP(payload.ToBytes(), []byte{}, &net.UDPAddr{
+		IP:   t.targetIP,
+		Port: 67,
+	})
+	return err
+}
+
+func (t *UnicastTransport) SetReceiveDeadline(deadline time.Time) error {
+	return t.udpConn.SetReadDeadline(deadline)
+}
+
+func (t *UnicastTransport) Receive() (*dhcpv4.DHCPv4, error) {
+	if t.udpConn == nil {
+		return nil, errors.New("unicast transport closed")
+	}
+	receiveBuf := make([]byte, math.MaxUint16)
+	_, _, err := t.udpConn.ReadFromUDP(receiveBuf)
+	if err != nil {
+		return nil, deadlineFromTimeout(err)
+	}
+	msg, err := dhcpv4.FromBytes(receiveBuf)
+	if err != nil {
+		return nil, NewInvalidMessageError(err)
+	}
+	return msg, nil
+}
+
+func (t *UnicastTransport) Close() error {
+	if t.udpConn == nil {
+		return nil
+	}
+	err := t.udpConn.Close()
+	t.udpConn = nil
+	// TODO(lorenz): Move to net.ErrClosed once Go 1.16 lands
+	if err != nil && strings.Contains(err.Error(), "use of closed network connection") {
+		return nil
+	}
+	return err
+}
diff --git a/metropolis/node/core/network/dns/BUILD.bazel b/metropolis/node/core/network/dns/BUILD.bazel
new file mode 100644
index 0000000..862d4cf
--- /dev/null
+++ b/metropolis/node/core/network/dns/BUILD.bazel
@@ -0,0 +1,15 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "coredns.go",
+        "directives.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dns",
+    visibility = ["//metropolis/node:__subpackages__"],
+    deps = [
+        "//metropolis/node/common/fileargs:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/network/dns/coredns.go b/metropolis/node/core/network/dns/coredns.go
new file mode 100644
index 0000000..b6400f7
--- /dev/null
+++ b/metropolis/node/core/network/dns/coredns.go
@@ -0,0 +1,148 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package DNS provides a DNS server using CoreDNS.
+package dns
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"os/exec"
+	"strings"
+	"sync"
+	"syscall"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fileargs"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+)
+
+const corefileBase = `
+.:53 {
+    errors
+	hosts {
+		fallthrough
+	}
+	
+    cache 30
+    loadbalance
+`
+
+type Service struct {
+	directiveRegistration chan *ExtraDirective
+	directives            map[string]ExtraDirective
+	cmd                   *exec.Cmd
+	args                  *fileargs.FileArgs
+	// stateMu guards access to the directives, cmd and args fields
+	stateMu sync.Mutex
+}
+
+// New creates a new CoreDNS service.
+// The given channel can then be used to dynamically register and unregister directives in the configuaration.
+// To register a new directive, send an ExtraDirective on the channel. To remove it again, use CancelDirective()
+// to create a removal message.
+func New(directiveRegistration chan *ExtraDirective) *Service {
+	return &Service{
+		directives:            map[string]ExtraDirective{},
+		directiveRegistration: directiveRegistration,
+	}
+}
+
+func (s *Service) makeCorefile(fargs *fileargs.FileArgs) []byte {
+	corefile := bytes.Buffer{}
+	corefile.WriteString(corefileBase)
+	for _, dir := range s.directives {
+		resolvedDir := dir.directive
+		for fname, fcontent := range dir.files {
+			resolvedDir = strings.ReplaceAll(resolvedDir, fmt.Sprintf("$FILE(%v)", fname), fargs.ArgPath(fname, fcontent))
+		}
+		corefile.WriteString(resolvedDir)
+		corefile.WriteString("\n")
+	}
+	corefile.WriteString("\n}")
+	return corefile.Bytes()
+}
+
+// CancelDirective creates a message to cancel the given directive.
+func CancelDirective(d *ExtraDirective) *ExtraDirective {
+	return &ExtraDirective{
+		ID: d.ID,
+	}
+}
+
+// Run runs the DNS service consisting of the CoreDNS process and the directive registration process
+func (s *Service) Run(ctx context.Context) error {
+	supervisor.Run(ctx, "coredns", s.runCoreDNS)
+	supervisor.Run(ctx, "registration", s.runRegistration)
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	supervisor.Signal(ctx, supervisor.SignalDone)
+	return nil
+}
+
+// runCoreDNS runs the CoreDNS proceess
+func (s *Service) runCoreDNS(ctx context.Context) error {
+	s.stateMu.Lock()
+	args, err := fileargs.New()
+	if err != nil {
+		s.stateMu.Unlock()
+		return fmt.Errorf("failed to create fileargs: %w", err)
+	}
+	defer args.Close()
+	s.args = args
+
+	s.cmd = exec.CommandContext(ctx, "/kubernetes/bin/coredns",
+		args.FileOpt("-conf", "Corefile", s.makeCorefile(args)),
+	)
+
+	if args.Error() != nil {
+		s.stateMu.Unlock()
+		return fmt.Errorf("failed to use fileargs: %w", err)
+	}
+
+	s.stateMu.Unlock()
+	return supervisor.RunCommand(ctx, s.cmd)
+}
+
+// runRegistration runs the background registration runnable which has a different lifecycle from the CoreDNS
+// runnable. It is responsible for managing dynamic directives.
+func (s *Service) runRegistration(ctx context.Context) error {
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	for {
+		select {
+		case <-ctx.Done():
+			return nil
+		case d := <-s.directiveRegistration:
+			s.processRegistration(ctx, d)
+		}
+	}
+}
+
+func (s *Service) processRegistration(ctx context.Context, d *ExtraDirective) {
+	s.stateMu.Lock()
+	defer s.stateMu.Unlock()
+	if d.directive == "" {
+		delete(s.directives, d.ID)
+	} else {
+		s.directives[d.ID] = *d
+	}
+	// If the process is not currenty running we're relying on corefile regeneration on startup
+	if s.cmd != nil && s.cmd.Process != nil && s.cmd.ProcessState == nil {
+		s.args.ArgPath("Corefile", s.makeCorefile(s.args))
+		if err := s.cmd.Process.Signal(syscall.SIGUSR1); err != nil {
+			supervisor.Logger(ctx).Warningf("Failed to send SIGUSR1 to CoreDNS for reload: %v", err)
+		}
+	}
+}
diff --git a/metropolis/node/core/network/dns/directives.go b/metropolis/node/core/network/dns/directives.go
new file mode 100644
index 0000000..72c4f29
--- /dev/null
+++ b/metropolis/node/core/network/dns/directives.go
@@ -0,0 +1,73 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dns
+
+import (
+	"fmt"
+	"net"
+	"strings"
+)
+
+// Type ExtraDirective contains additional config directives for CoreDNS.
+type ExtraDirective struct {
+	// ID is the identifier of this directive. There can only be one directive with a given ID active at once.
+	// The ID is also used to identify which directive to purge.
+	ID string
+	// directive contains a full CoreDNS directive as a string. It can also use the $FILE(<filename>) macro,
+	// which will be expanded to the path of a file from the files field.
+	directive string
+	// files contains additional files used in the configuration. The map key is used as the filename.
+	files map[string][]byte
+}
+
+// NewUpstreamDirective creates a forward with no fallthrough that forwards all requests not yet matched to the given
+// upstream DNS servers.
+func NewUpstreamDirective(dnsServers []net.IP) *ExtraDirective {
+	strb := strings.Builder{}
+	if len(dnsServers) > 0 {
+		strb.WriteString("forward .")
+		for _, ip := range dnsServers {
+			strb.WriteString(" ")
+			strb.WriteString(ip.String())
+		}
+	}
+	return &ExtraDirective{
+		directive: strb.String(),
+	}
+}
+
+var kubernetesDirective = `
+kubernetes %v in-addr.arpa ip6.arpa {
+	kubeconfig $FILE(kubeconfig) default
+	pods insecure
+	fallthrough in-addr.arpa ip6.arpa
+	ttl 30
+}
+`
+
+// NewKubernetesDirective creates a directive running a "Kubernetes DNS-Based Service Discovery" [1] compliant service
+// under clusterDomain. The given kubeconfig needs at least read access to services, endpoints and endpointslices.
+// [1] https://github.com/kubernetes/dns/blob/master/docs/specification.md
+func NewKubernetesDirective(clusterDomain string, kubeconfig []byte) *ExtraDirective {
+	return &ExtraDirective{
+		ID:        "k8s-clusterdns",
+		directive: fmt.Sprintf(kubernetesDirective, clusterDomain),
+		files: map[string][]byte{
+			"kubeconfig": kubeconfig,
+		},
+	}
+}
diff --git a/metropolis/node/core/network/main.go b/metropolis/node/core/network/main.go
new file mode 100644
index 0000000..29e757d
--- /dev/null
+++ b/metropolis/node/core/network/main.go
@@ -0,0 +1,260 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package network
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io/ioutil"
+	"net"
+	"os"
+	"sync"
+	"time"
+
+	"github.com/google/nftables"
+	"github.com/google/nftables/expr"
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/vishvananda/netlink"
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c"
+	dhcpcb "git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c/callback"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dns"
+)
+
+const (
+	resolvConfPath     = "/etc/resolv.conf"
+	resolvConfSwapPath = "/etc/resolv.conf.new"
+)
+
+type Service struct {
+	config Config
+	dhcp   *dhcp4c.Client
+
+	// nftConn is a shared file descriptor handle to nftables, automatically initialized on first use.
+	nftConn             nftables.Conn
+	natTable            *nftables.Table
+	natPostroutingChain *nftables.Chain
+
+	// These are a temporary hack pending the removal of the GetIP interface
+	ipLock       sync.Mutex
+	currentIPTmp net.IP
+
+	logger logtree.LeveledLogger
+}
+
+type Config struct {
+	CorednsRegistrationChan chan *dns.ExtraDirective
+}
+
+func New(config Config) *Service {
+	return &Service{
+		config: config,
+	}
+}
+
+func setResolvconf(nameservers []net.IP, searchDomains []string) error {
+	_ = os.Mkdir("/etc", 0755)
+	newResolvConf, err := os.Create(resolvConfSwapPath)
+	if err != nil {
+		return err
+	}
+	defer newResolvConf.Close()
+	defer os.Remove(resolvConfSwapPath)
+	for _, ns := range nameservers {
+		if _, err := newResolvConf.WriteString(fmt.Sprintf("nameserver %v\n", ns)); err != nil {
+			return err
+		}
+	}
+	for _, searchDomain := range searchDomains {
+		if _, err := newResolvConf.WriteString(fmt.Sprintf("search %v", searchDomain)); err != nil {
+			return err
+		}
+	}
+	newResolvConf.Close()
+	// Atomically swap in new config
+	return unix.Rename(resolvConfSwapPath, resolvConfPath)
+}
+
+// nfifname converts an interface name into 16 bytes padded with zeroes (for nftables)
+func nfifname(n string) []byte {
+	b := make([]byte, 16)
+	copy(b, []byte(n+"\x00"))
+	return b
+}
+
+func (s *Service) dhcpDNSCallback(old, new *dhcp4c.Lease) error {
+	oldServers := old.DNSServers()
+	newServers := new.DNSServers()
+	if newServers.Equal(oldServers) {
+		return nil // nothing to do
+	}
+	s.logger.Infof("Setting upstream DNS servers to %v", newServers)
+	s.config.CorednsRegistrationChan <- dns.NewUpstreamDirective(newServers)
+	return nil
+}
+
+// TODO(lorenz): Get rid of this once we have robust node resolution
+func (s *Service) getIPCallbackHack(old, new *dhcp4c.Lease) error {
+	if old == nil && new != nil {
+		s.ipLock.Lock()
+		s.currentIPTmp = new.AssignedIP
+		s.ipLock.Unlock()
+	}
+	return nil
+}
+
+func (s *Service) useInterface(ctx context.Context, iface netlink.Link) error {
+	netIface, err := net.InterfaceByIndex(iface.Attrs().Index)
+	if err != nil {
+		return fmt.Errorf("cannot create Go net.Interface from netlink.Link: %w", err)
+	}
+	s.dhcp, err = dhcp4c.NewClient(netIface)
+	if err != nil {
+		return fmt.Errorf("failed to create DHCP client on interface %v: %w", iface.Attrs().Name, err)
+	}
+	s.dhcp.VendorClassIdentifier = "com.nexantic.smalltown.v1"
+	s.dhcp.RequestedOptions = []dhcpv4.OptionCode{dhcpv4.OptionRouter, dhcpv4.OptionNameServer}
+	s.dhcp.LeaseCallback = dhcpcb.Compose(dhcpcb.ManageIP(iface), dhcpcb.ManageDefaultRoute(iface), s.dhcpDNSCallback, s.getIPCallbackHack)
+	err = supervisor.Run(ctx, "dhcp", s.dhcp.Run)
+	if err != nil {
+		return err
+	}
+
+	// Masquerade/SNAT all traffic going out of the external interface
+	s.nftConn.AddRule(&nftables.Rule{
+		Table: s.natTable,
+		Chain: s.natPostroutingChain,
+		Exprs: []expr.Any{
+			&expr.Meta{Key: expr.MetaKeyOIFNAME, Register: 1},
+			&expr.Cmp{
+				Op:       expr.CmpOpEq,
+				Register: 1,
+				Data:     nfifname(iface.Attrs().Name),
+			},
+			&expr.Masq{},
+		},
+	})
+
+	if err := s.nftConn.Flush(); err != nil {
+		panic(err)
+	}
+
+	return nil
+}
+
+// GetIP returns the current IP (and optionally waits for one to be assigned)
+func (s *Service) GetIP(ctx context.Context, wait bool) (*net.IP, error) {
+	for {
+		var currentIP net.IP
+		s.ipLock.Lock()
+		currentIP = s.currentIPTmp
+		s.ipLock.Unlock()
+		if currentIP == nil {
+			if !wait {
+				return nil, errors.New("no IP available")
+			}
+			select {
+			case <-ctx.Done():
+				return nil, ctx.Err()
+			case <-time.After(1 * time.Second):
+				continue
+			}
+		}
+		return &currentIP, nil
+	}
+}
+
+func (s *Service) Run(ctx context.Context) error {
+	logger := supervisor.Logger(ctx)
+	dnsSvc := dns.New(s.config.CorednsRegistrationChan)
+	supervisor.Run(ctx, "dns", dnsSvc.Run)
+	supervisor.Run(ctx, "interfaces", s.runInterfaces)
+
+	s.natTable = s.nftConn.AddTable(&nftables.Table{
+		Family: nftables.TableFamilyIPv4,
+		Name:   "nat",
+	})
+
+	s.natPostroutingChain = s.nftConn.AddChain(&nftables.Chain{
+		Name:     "postrouting",
+		Hooknum:  nftables.ChainHookPostrouting,
+		Priority: nftables.ChainPriorityNATSource,
+		Table:    s.natTable,
+		Type:     nftables.ChainTypeNAT,
+	})
+	if err := s.nftConn.Flush(); err != nil {
+		logger.Fatalf("Failed to set up nftables base chains: %v", err)
+	}
+
+	if err := ioutil.WriteFile("/proc/sys/net/ipv4/ip_forward", []byte("1\n"), 0644); err != nil {
+		logger.Fatalf("Failed to enable IPv4 forwarding: %v", err)
+	}
+
+	// We're handling all DNS requests with CoreDNS, including local ones
+	if err := setResolvconf([]net.IP{{127, 0, 0, 1}}, []string{}); err != nil {
+		logger.Fatalf("Failed to set resolv.conf: %v", err)
+	}
+
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	supervisor.Signal(ctx, supervisor.SignalDone)
+	return nil
+}
+
+func (s *Service) runInterfaces(ctx context.Context) error {
+	s.logger = supervisor.Logger(ctx)
+	s.logger.Info("Starting network interface management")
+
+	links, err := netlink.LinkList()
+	if err != nil {
+		s.logger.Fatalf("Failed to list network links: %s", err)
+	}
+
+	var ethernetLinks []netlink.Link
+	for _, link := range links {
+		attrs := link.Attrs()
+		if link.Type() == "device" && len(attrs.HardwareAddr) > 0 {
+			if len(attrs.HardwareAddr) == 6 { // Ethernet
+				if attrs.Flags&net.FlagUp != net.FlagUp {
+					netlink.LinkSetUp(link) // Attempt to take up all ethernet links
+				}
+				ethernetLinks = append(ethernetLinks, link)
+			} else {
+				s.logger.Infof("Ignoring non-Ethernet interface %s", attrs.Name)
+			}
+		} else if link.Attrs().Name == "lo" {
+			if err := netlink.LinkSetUp(link); err != nil {
+				s.logger.Errorf("Failed to bring up loopback interface: %v", err)
+			}
+		}
+	}
+	if len(ethernetLinks) != 1 {
+		s.logger.Warningf("Network service needs exactly one link, bailing")
+	} else {
+		link := ethernetLinks[0]
+		if err := s.useInterface(ctx, link); err != nil {
+			return fmt.Errorf("failed to bring up link %s: %w", link.Attrs().Name, err)
+		}
+	}
+
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	supervisor.Signal(ctx, supervisor.SignalDone)
+	return nil
+}
diff --git a/metropolis/node/core/switchroot.go b/metropolis/node/core/switchroot.go
new file mode 100644
index 0000000..5865225
--- /dev/null
+++ b/metropolis/node/core/switchroot.go
@@ -0,0 +1,213 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"strings"
+	"syscall"
+
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+)
+
+// switchRoot moves the root from initramfs into a tmpfs
+// This is necessary because you cannot pivot_root from a initramfs (and runsc wants to do that).
+// In the future, we should instead use something like squashfs instead of an initramfs and just nuke this.
+func switchRoot(log logtree.LeveledLogger) error {
+	// We detect the need to remount to tmpfs over env vars.
+	// The first run of /init (from initramfs) will not have this var, and will be re-exec'd from a new tmpfs root with
+	// that variable set.
+	witness := "SIGNOS_REMOUNTED"
+
+	// If the witness env var is found in the environment, it means we are ready to go.
+	environ := os.Environ()
+	for _, env := range environ {
+		if strings.HasPrefix(env, witness+"=") {
+			log.Info("Smalltown running in tmpfs root")
+			return nil
+		}
+	}
+
+	// Otherwise, we need to remount to a tmpfs.
+	environ = append(environ, witness+"=yes")
+	log.Info("Smalltown running in initramfs, remounting to tmpfs...")
+
+	// Make note of all directories we have to make and files that we have to copy.
+	paths := []string{}
+	dirs := []string{}
+	err := filepath.Walk("/", func(path string, info os.FileInfo, err error) error {
+		if err != nil {
+			return err
+		}
+		if path == "/" {
+			return nil
+		}
+		// /dev is prepopulated by the initramfs, skip that. The target root uses devtmpfs.
+		if path == "/dev" || strings.HasPrefix(path, "/dev/") {
+			return nil
+		}
+
+		if info.IsDir() {
+			dirs = append(dirs, path)
+		} else {
+			paths = append(paths, path)
+		}
+
+		return nil
+	})
+	if err != nil {
+		return fmt.Errorf("could not list root files: %w", err)
+	}
+
+	log.Info("Copying paths to tmpfs:")
+	for _, p := range paths {
+		log.Infof(" - %s", p)
+	}
+
+	// Make new root at /mnt
+	if err := os.Mkdir("/mnt", 0755); err != nil {
+		return fmt.Errorf("could not make /mnt: %w", err)
+	}
+	// And mount a tmpfs on it
+	if err := unix.Mount("tmpfs", "/mnt", "tmpfs", 0, ""); err != nil {
+		return fmt.Errorf("could not mount tmpfs on /mnt: %w", err)
+	}
+
+	// Make all directories. Since filepath.Walk is lexicographically ordered, we don't need to ensure that the parent
+	// exists.
+	for _, src := range dirs {
+		stat, err := os.Stat(src)
+		if err != nil {
+			return fmt.Errorf("Stat(%q): %w", src, err)
+		}
+		dst := "/mnt" + src
+		err = os.Mkdir(dst, stat.Mode())
+		if err != nil {
+			return fmt.Errorf("Mkdir(%q): %w", dst, err)
+		}
+	}
+
+	// Move all files over. Parent directories will exist by now.
+	for _, src := range paths {
+		stat, err := os.Stat(src)
+		if err != nil {
+			return fmt.Errorf("Stat(%q): %w", src, err)
+		}
+		dst := "/mnt" + src
+
+		// Copy file.
+		sfd, err := os.Open(src)
+		if err != nil {
+			return fmt.Errorf("Open(%q): %w", src, err)
+		}
+		dfd, err := os.OpenFile(dst, os.O_WRONLY|os.O_CREATE, stat.Mode())
+		if err != nil {
+			sfd.Close()
+			return fmt.Errorf("OpenFile(%q): %w", dst, err)
+		}
+		_, err = io.Copy(dfd, sfd)
+
+		sfd.Close()
+		dfd.Close()
+		if err != nil {
+			return fmt.Errorf("Copying %q failed: %w", src, err)
+		}
+
+		// Remove the old file.
+		err = unix.Unlink(src)
+		if err != nil {
+			return fmt.Errorf("Unlink(%q): %w", src, err)
+		}
+	}
+
+	// Set up target filesystems.
+	for _, el := range []struct {
+		dir   string
+		fs    string
+		flags uintptr
+	}{
+		{"/sys", "sysfs", unix.MS_NOEXEC | unix.MS_NOSUID | unix.MS_NODEV},
+		{"/proc", "proc", unix.MS_NOEXEC | unix.MS_NOSUID | unix.MS_NODEV},
+		{"/dev", "devtmpfs", unix.MS_NOEXEC | unix.MS_NOSUID},
+		{"/dev/pts", "devpts", unix.MS_NOEXEC | unix.MS_NOSUID},
+	} {
+		if err := os.Mkdir("/mnt"+el.dir, 0755); err != nil {
+			return fmt.Errorf("could not make /mnt%s: %w", el.dir, err)
+		}
+		if err := unix.Mount(el.fs, "/mnt"+el.dir, el.fs, el.flags, ""); err != nil {
+			return fmt.Errorf("could not mount %s on /mnt%s: %w", el.fs, el.dir, err)
+		}
+	}
+
+	// Mount all available CGroups for v1 (v2 uses a single unified hierarchy and is not supported by our runtimes yet)
+	if unix.Mount("tmpfs", "/mnt/sys/fs/cgroup", "tmpfs", unix.MS_NOEXEC|unix.MS_NOSUID|unix.MS_NODEV, ""); err != nil {
+		panic(err)
+	}
+	cgroupsRaw, err := ioutil.ReadFile("/mnt/proc/cgroups")
+	if err != nil {
+		panic(err)
+	}
+
+	cgroupLines := strings.Split(string(cgroupsRaw), "\n")
+	for _, cgroupLine := range cgroupLines {
+		if cgroupLine == "" || strings.HasPrefix(cgroupLine, "#") {
+			continue
+		}
+		cgroupParts := strings.Split(cgroupLine, "\t")
+		cgroupName := cgroupParts[0]
+		if err := os.Mkdir("/mnt/sys/fs/cgroup/"+cgroupName, 0755); err != nil {
+			panic(err)
+		}
+		if err := unix.Mount("cgroup", "/mnt/sys/fs/cgroup/"+cgroupName, "cgroup", unix.MS_NOEXEC|unix.MS_NOSUID|unix.MS_NODEV, cgroupName); err != nil {
+			panic(err)
+		}
+	}
+
+	// Enable hierarchical memory accounting
+	useMemoryHierarchy, err := os.OpenFile("/mnt/sys/fs/cgroup/memory/memory.use_hierarchy", os.O_RDWR, 0)
+	if err != nil {
+		panic(err)
+	}
+	if _, err := useMemoryHierarchy.WriteString("1"); err != nil {
+		panic(err)
+	}
+	useMemoryHierarchy.Close()
+
+	// Chroot to new root.
+	// This is adapted from util-linux's switch_root.
+	err = os.Chdir("/mnt")
+	if err != nil {
+		return fmt.Errorf("could not chdir to /mnt: %w", err)
+	}
+	err = syscall.Mount("/mnt", "/", "", syscall.MS_MOVE, "")
+	if err != nil {
+		return fmt.Errorf("could not remount /mnt to /: %w", err)
+	}
+	err = syscall.Chroot(".")
+	if err != nil {
+		return fmt.Errorf("could not chroot to new root: %w", err)
+	}
+
+	// Re-exec into new init with new environment
+	return unix.Exec("/init", os.Args, environ)
+}
diff --git a/metropolis/node/core/tpm/BUILD.bazel b/metropolis/node/core/tpm/BUILD.bazel
new file mode 100644
index 0000000..fd42681
--- /dev/null
+++ b/metropolis/node/core/tpm/BUILD.bazel
@@ -0,0 +1,22 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "credactivation_compat.go",
+        "tpm.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/tpm",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//metropolis/node/common/sysfs:go_default_library",
+        "//metropolis/node/core/logtree:go_default_library",
+        "@com_github_gogo_protobuf//proto:go_default_library",
+        "@com_github_google_go_tpm//tpm2:go_default_library",
+        "@com_github_google_go_tpm//tpmutil:go_default_library",
+        "@com_github_google_go_tpm_tools//proto:go_default_library",
+        "@com_github_google_go_tpm_tools//tpm2tools:go_default_library",
+        "@com_github_pkg_errors//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/tpm/credactivation_compat.go b/metropolis/node/core/tpm/credactivation_compat.go
new file mode 100644
index 0000000..039f8d5
--- /dev/null
+++ b/metropolis/node/core/tpm/credactivation_compat.go
@@ -0,0 +1,123 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tpm
+
+// This file is adapted from github.com/google/go-tpm/tpm2/credactivation which outputs broken
+// challenges for unknown reasons. They use u16 length-delimited outputs for the challenge blobs
+// which is incorrect. Rather than rewriting the routine, we only applied minimal fixes to it
+// and skip the ECC part of the issue (because we would rather trust the proprietary RSA implementation).
+//
+// TODO(lorenz): I'll eventually deal with this upstream, but for now just fix it here (it's not that)
+// much code after all (https://github.com/google/go-tpm/issues/121)
+
+import (
+	"crypto/aes"
+	"crypto/cipher"
+	"crypto/hmac"
+	"crypto/rsa"
+	"fmt"
+	"io"
+
+	"github.com/google/go-tpm/tpm2"
+	"github.com/google/go-tpm/tpmutil"
+)
+
+const (
+	labelIdentity  = "IDENTITY"
+	labelStorage   = "STORAGE"
+	labelIntegrity = "INTEGRITY"
+)
+
+func generateRSA(aik *tpm2.HashValue, pub *rsa.PublicKey, symBlockSize int, secret []byte, rnd io.Reader) ([]byte, []byte, error) {
+	newAIKHash, err := aik.Alg.HashConstructor()
+	if err != nil {
+		return nil, nil, err
+	}
+
+	// The seed length should match the keysize used by the EKs symmetric cipher.
+	// For typical RSA EKs, this will be 128 bits (16 bytes).
+	// Spec: TCG 2.0 EK Credential Profile revision 14, section 2.1.5.1.
+	seed := make([]byte, symBlockSize)
+	if _, err := io.ReadFull(rnd, seed); err != nil {
+		return nil, nil, fmt.Errorf("generating seed: %v", err)
+	}
+
+	// Encrypt the seed value using the provided public key.
+	// See annex B, section 10.4 of the TPM specification revision 2 part 1.
+	label := append([]byte(labelIdentity), 0)
+	encSecret, err := rsa.EncryptOAEP(newAIKHash(), rnd, pub, seed, label)
+	if err != nil {
+		return nil, nil, fmt.Errorf("generating encrypted seed: %v", err)
+	}
+
+	// Generate the encrypted credential by convolving the seed with the digest of
+	// the AIK, and using the result as the key to encrypt the secret.
+	// See section 24.4 of TPM 2.0 specification, part 1.
+	aikNameEncoded, err := aik.Encode()
+	if err != nil {
+		return nil, nil, fmt.Errorf("encoding aikName: %v", err)
+	}
+	symmetricKey, err := tpm2.KDFa(aik.Alg, seed, labelStorage, aikNameEncoded, nil, len(seed)*8)
+	if err != nil {
+		return nil, nil, fmt.Errorf("generating symmetric key: %v", err)
+	}
+	c, err := aes.NewCipher(symmetricKey)
+	if err != nil {
+		return nil, nil, fmt.Errorf("symmetric cipher setup: %v", err)
+	}
+	cv, err := tpmutil.Pack(tpmutil.U16Bytes(secret))
+	if err != nil {
+		return nil, nil, fmt.Errorf("generating cv (TPM2B_Digest): %v", err)
+	}
+
+	// IV is all null bytes. encIdentity represents the encrypted credential.
+	encIdentity := make([]byte, len(cv))
+	cipher.NewCFBEncrypter(c, make([]byte, len(symmetricKey))).XORKeyStream(encIdentity, cv)
+
+	// Generate the integrity HMAC, which is used to protect the integrity of the
+	// encrypted structure.
+	// See section 24.5 of the TPM specification revision 2 part 1.
+	macKey, err := tpm2.KDFa(aik.Alg, seed, labelIntegrity, nil, nil, newAIKHash().Size()*8)
+	if err != nil {
+		return nil, nil, fmt.Errorf("generating HMAC key: %v", err)
+	}
+
+	mac := hmac.New(newAIKHash, macKey)
+	mac.Write(encIdentity)
+	mac.Write(aikNameEncoded)
+	integrityHMAC := mac.Sum(nil)
+
+	idObject := &tpm2.IDObject{
+		IntegrityHMAC: integrityHMAC,
+		EncIdentity:   encIdentity,
+	}
+	id, err := tpmutil.Pack(idObject)
+	if err != nil {
+		return nil, nil, fmt.Errorf("encoding IDObject: %v", err)
+	}
+
+	packedID, err := tpmutil.Pack(id)
+	if err != nil {
+		return nil, nil, fmt.Errorf("packing id: %v", err)
+	}
+	packedEncSecret, err := tpmutil.Pack(encSecret)
+	if err != nil {
+		return nil, nil, fmt.Errorf("packing encSecret: %v", err)
+	}
+
+	return packedID, packedEncSecret, nil
+}
diff --git a/metropolis/node/core/tpm/eventlog/BUILD.bazel b/metropolis/node/core/tpm/eventlog/BUILD.bazel
new file mode 100644
index 0000000..64fa1ff
--- /dev/null
+++ b/metropolis/node/core/tpm/eventlog/BUILD.bazel
@@ -0,0 +1,17 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "compat.go",
+        "eventlog.go",
+        "secureboot.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/tpm/eventlog",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//metropolis/node/core/tpm/eventlog/internal:go_default_library",
+        "@com_github_google_certificate_transparency_go//x509:go_default_library",
+        "@com_github_google_go_tpm//tpm2:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/tpm/eventlog/LICENSE-3RD-PARTY.txt b/metropolis/node/core/tpm/eventlog/LICENSE-3RD-PARTY.txt
new file mode 100644
index 0000000..2d3298c
--- /dev/null
+++ b/metropolis/node/core/tpm/eventlog/LICENSE-3RD-PARTY.txt
@@ -0,0 +1,12 @@
+Copyright 2020 Google Inc.
+Licensed under the Apache License, Version 2.0 (the "License"); you may not
+use this file except in compliance with the License. You may obtain a copy of
+the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+License for the specific language governing permissions and limitations under
+the License.
\ No newline at end of file
diff --git a/metropolis/node/core/tpm/eventlog/compat.go b/metropolis/node/core/tpm/eventlog/compat.go
new file mode 100644
index 0000000..f83972b
--- /dev/null
+++ b/metropolis/node/core/tpm/eventlog/compat.go
@@ -0,0 +1,32 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package eventlog
+
+// This file contains compatibility functions for our TPM library
+
+import (
+	"crypto"
+)
+
+// ConvertRawPCRs converts from raw PCRs to eventlog PCR structures
+func ConvertRawPCRs(pcrs [][]byte) []PCR {
+	var evPCRs []PCR
+	for i, digest := range pcrs {
+		evPCRs = append(evPCRs, PCR{DigestAlg: crypto.SHA256, Index: i, Digest: digest})
+	}
+	return evPCRs
+}
diff --git a/metropolis/node/core/tpm/eventlog/eventlog.go b/metropolis/node/core/tpm/eventlog/eventlog.go
new file mode 100644
index 0000000..49a8a26
--- /dev/null
+++ b/metropolis/node/core/tpm/eventlog/eventlog.go
@@ -0,0 +1,646 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Taken and pruned from go-attestation revision 2453c8f39a4ff46009f6a9db6fb7c6cca789d9a1 under Apache 2.0
+
+package eventlog
+
+import (
+	"bytes"
+	"crypto"
+	"crypto/sha1"
+	"crypto/sha256"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"sort"
+
+	// Ensure hashes are available.
+	_ "crypto/sha256"
+
+	"github.com/google/go-tpm/tpm2"
+)
+
+// HashAlg identifies a hashing Algorithm.
+type HashAlg uint8
+
+// Valid hash algorithms.
+var (
+	HashSHA1   = HashAlg(tpm2.AlgSHA1)
+	HashSHA256 = HashAlg(tpm2.AlgSHA256)
+)
+
+func (a HashAlg) cryptoHash() crypto.Hash {
+	switch a {
+	case HashSHA1:
+		return crypto.SHA1
+	case HashSHA256:
+		return crypto.SHA256
+	}
+	return 0
+}
+
+func (a HashAlg) goTPMAlg() tpm2.Algorithm {
+	switch a {
+	case HashSHA1:
+		return tpm2.AlgSHA1
+	case HashSHA256:
+		return tpm2.AlgSHA256
+	}
+	return 0
+}
+
+// String returns a human-friendly representation of the hash algorithm.
+func (a HashAlg) String() string {
+	switch a {
+	case HashSHA1:
+		return "SHA1"
+	case HashSHA256:
+		return "SHA256"
+	}
+	return fmt.Sprintf("HashAlg<%d>", int(a))
+}
+
+// ReplayError describes the parsed events that failed to verify against
+// a particular PCR.
+type ReplayError struct {
+	Events      []Event
+	invalidPCRs []int
+}
+
+func (e ReplayError) affected(pcr int) bool {
+	for _, p := range e.invalidPCRs {
+		if p == pcr {
+			return true
+		}
+	}
+	return false
+}
+
+// Error returns a human-friendly description of replay failures.
+func (e ReplayError) Error() string {
+	return fmt.Sprintf("event log failed to verify: the following registers failed to replay: %v", e.invalidPCRs)
+}
+
+// TPM algorithms. See the TPM 2.0 specification section 6.3.
+//
+// https://trustedcomputinggroup.org/wp-content/uploads/TPM-Rev-2.0-Part-2-Structures-01.38.pdf#page=42
+const (
+	algSHA1   uint16 = 0x0004
+	algSHA256 uint16 = 0x000B
+)
+
+// EventType indicates what kind of data an event is reporting.
+type EventType uint32
+
+// Event is a single event from a TCG event log. This reports descrete items such
+// as BIOs measurements or EFI states.
+type Event struct {
+	// order of the event in the event log.
+	sequence int
+
+	// PCR index of the event.
+	Index int
+	// Type of the event.
+	Type EventType
+
+	// Data of the event. For certain kinds of events, this must match the event
+	// digest to be valid.
+	Data []byte
+	// Digest is the verified digest of the event data. While an event can have
+	// multiple for different hash values, this is the one that was matched to the
+	// PCR value.
+	Digest []byte
+
+	// TODO(ericchiang): Provide examples or links for which event types must
+	// match their data to their digest.
+}
+
+func (e *Event) digestEquals(b []byte) error {
+	if len(e.Digest) == 0 {
+		return errors.New("no digests present")
+	}
+
+	switch len(e.Digest) {
+	case crypto.SHA256.Size():
+		s := sha256.Sum256(b)
+		if bytes.Equal(s[:], e.Digest) {
+			return nil
+		}
+	case crypto.SHA1.Size():
+		s := sha1.Sum(b)
+		if bytes.Equal(s[:], e.Digest) {
+			return nil
+		}
+	default:
+		return fmt.Errorf("cannot compare hash of length %d", len(e.Digest))
+	}
+
+	return fmt.Errorf("digest (len %d) does not match", len(e.Digest))
+}
+
+// EventLog is a parsed measurement log. This contains unverified data representing
+// boot events that must be replayed against PCR values to determine authenticity.
+type EventLog struct {
+	// Algs holds the set of algorithms that the event log uses.
+	Algs []HashAlg
+
+	rawEvents []rawEvent
+}
+
+func (e *EventLog) clone() *EventLog {
+	out := EventLog{
+		Algs:      make([]HashAlg, len(e.Algs)),
+		rawEvents: make([]rawEvent, len(e.rawEvents)),
+	}
+	copy(out.Algs, e.Algs)
+	copy(out.rawEvents, e.rawEvents)
+	return &out
+}
+
+type elWorkaround struct {
+	id          string
+	affectedPCR int
+	apply       func(e *EventLog) error
+}
+
+// inject3 appends two new events into the event log.
+func inject3(e *EventLog, pcr int, data1, data2, data3 string) error {
+	if err := inject(e, pcr, data1); err != nil {
+		return err
+	}
+	if err := inject(e, pcr, data2); err != nil {
+		return err
+	}
+	return inject(e, pcr, data3)
+}
+
+// inject2 appends two new events into the event log.
+func inject2(e *EventLog, pcr int, data1, data2 string) error {
+	if err := inject(e, pcr, data1); err != nil {
+		return err
+	}
+	return inject(e, pcr, data2)
+}
+
+// inject appends a new event into the event log.
+func inject(e *EventLog, pcr int, data string) error {
+	evt := rawEvent{
+		data:     []byte(data),
+		index:    pcr,
+		sequence: e.rawEvents[len(e.rawEvents)-1].sequence + 1,
+	}
+	for _, alg := range e.Algs {
+		h := alg.cryptoHash().New()
+		h.Write([]byte(data))
+		evt.digests = append(evt.digests, digest{hash: alg.cryptoHash(), data: h.Sum(nil)})
+	}
+	e.rawEvents = append(e.rawEvents, evt)
+	return nil
+}
+
+const (
+	ebsInvocation = "Exit Boot Services Invocation"
+	ebsSuccess    = "Exit Boot Services Returned with Success"
+	ebsFailure    = "Exit Boot Services Returned with Failure"
+)
+
+var eventlogWorkarounds = []elWorkaround{
+	{
+		id:          "EBS Invocation + Success",
+		affectedPCR: 5,
+		apply: func(e *EventLog) error {
+			return inject2(e, 5, ebsInvocation, ebsSuccess)
+		},
+	},
+	{
+		id:          "EBS Invocation + Failure",
+		affectedPCR: 5,
+		apply: func(e *EventLog) error {
+			return inject2(e, 5, ebsInvocation, ebsFailure)
+		},
+	},
+	{
+		id:          "EBS Invocation + Failure + Success",
+		affectedPCR: 5,
+		apply: func(e *EventLog) error {
+			return inject3(e, 5, ebsInvocation, ebsFailure, ebsSuccess)
+		},
+	},
+}
+
+// Verify replays the event log against a TPM's PCR values, returning the
+// events which could be matched to a provided PCR value.
+// An error is returned if the replayed digest for events with a given PCR
+// index do not match any provided value for that PCR index.
+func (e *EventLog) Verify(pcrs []PCR) ([]Event, error) {
+	events, err := e.verify(pcrs)
+	// If there were any issues replaying the PCRs, try each of the workarounds
+	// in turn.
+	// TODO(jsonp): Allow workarounds to be combined.
+	if rErr, isReplayErr := err.(ReplayError); isReplayErr {
+		for _, wkrd := range eventlogWorkarounds {
+			if !rErr.affected(wkrd.affectedPCR) {
+				continue
+			}
+			el := e.clone()
+			if err := wkrd.apply(el); err != nil {
+				return nil, fmt.Errorf("failed applying workaround %q: %v", wkrd.id, err)
+			}
+			if events, err := el.verify(pcrs); err == nil {
+				return events, nil
+			}
+		}
+	}
+
+	return events, err
+}
+
+// PCR encapsulates the value of a PCR at a point in time.
+type PCR struct {
+	Index     int
+	Digest    []byte
+	DigestAlg crypto.Hash
+}
+
+func (e *EventLog) verify(pcrs []PCR) ([]Event, error) {
+	events, err := replayEvents(e.rawEvents, pcrs)
+	if err != nil {
+		if _, isReplayErr := err.(ReplayError); isReplayErr {
+			return nil, err
+		}
+		return nil, fmt.Errorf("pcrs failed to replay: %v", err)
+	}
+	return events, nil
+}
+
+func extend(pcr PCR, replay []byte, e rawEvent) (pcrDigest []byte, eventDigest []byte, err error) {
+	h := pcr.DigestAlg
+
+	for _, digest := range e.digests {
+		if digest.hash != pcr.DigestAlg {
+			continue
+		}
+		if len(digest.data) != len(pcr.Digest) {
+			return nil, nil, fmt.Errorf("digest data length (%d) doesn't match PCR digest length (%d)", len(digest.data), len(pcr.Digest))
+		}
+		hash := h.New()
+		if len(replay) != 0 {
+			hash.Write(replay)
+		} else {
+			b := make([]byte, h.Size())
+			hash.Write(b)
+		}
+		hash.Write(digest.data)
+		return hash.Sum(nil), digest.data, nil
+	}
+	return nil, nil, fmt.Errorf("no event digest matches pcr algorithm: %v", pcr.DigestAlg)
+}
+
+// replayPCR replays the event log for a specific PCR, using pcr and
+// event digests with the algorithm in pcr. An error is returned if the
+// replayed values do not match the final PCR digest, or any event tagged
+// with that PCR does not posess an event digest with the specified algorithm.
+func replayPCR(rawEvents []rawEvent, pcr PCR) ([]Event, bool) {
+	var (
+		replay    []byte
+		outEvents []Event
+	)
+
+	for _, e := range rawEvents {
+		if e.index != pcr.Index {
+			continue
+		}
+
+		replayValue, digest, err := extend(pcr, replay, e)
+		if err != nil {
+			return nil, false
+		}
+		replay = replayValue
+		outEvents = append(outEvents, Event{sequence: e.sequence, Data: e.data, Digest: digest, Index: pcr.Index, Type: e.typ})
+	}
+
+	if len(outEvents) > 0 && !bytes.Equal(replay, pcr.Digest) {
+		return nil, false
+	}
+	return outEvents, true
+}
+
+type pcrReplayResult struct {
+	events     []Event
+	successful bool
+}
+
+func replayEvents(rawEvents []rawEvent, pcrs []PCR) ([]Event, error) {
+	var (
+		invalidReplays []int
+		verifiedEvents []Event
+		allPCRReplays  = map[int][]pcrReplayResult{}
+	)
+
+	// Replay the event log for every PCR and digest algorithm combination.
+	for _, pcr := range pcrs {
+		events, ok := replayPCR(rawEvents, pcr)
+		allPCRReplays[pcr.Index] = append(allPCRReplays[pcr.Index], pcrReplayResult{events, ok})
+	}
+
+	// Record PCR indices which do not have any successful replay. Record the
+	// events for a successful replay.
+pcrLoop:
+	for i, replaysForPCR := range allPCRReplays {
+		for _, replay := range replaysForPCR {
+			if replay.successful {
+				// We consider the PCR verified at this stage: The replay of values with
+				// one digest algorithm matched a provided value.
+				// As such, we save the PCR's events, and proceed to the next PCR.
+				verifiedEvents = append(verifiedEvents, replay.events...)
+				continue pcrLoop
+			}
+		}
+		invalidReplays = append(invalidReplays, i)
+	}
+
+	if len(invalidReplays) > 0 {
+		events := make([]Event, 0, len(rawEvents))
+		for _, e := range rawEvents {
+			events = append(events, Event{e.sequence, e.index, e.typ, e.data, nil})
+		}
+		return nil, ReplayError{
+			Events:      events,
+			invalidPCRs: invalidReplays,
+		}
+	}
+
+	sort.Slice(verifiedEvents, func(i int, j int) bool {
+		return verifiedEvents[i].sequence < verifiedEvents[j].sequence
+	})
+	return verifiedEvents, nil
+}
+
+// EV_NO_ACTION is a special event type that indicates information to the parser
+// instead of holding a measurement. For TPM 2.0, this event type is used to signal
+// switching from SHA1 format to a variable length digest.
+//
+// https://trustedcomputinggroup.org/wp-content/uploads/TCG_PCClientSpecPlat_TPM_2p0_1p04_pub.pdf#page=110
+const eventTypeNoAction = 0x03
+
+// ParseEventLog parses an unverified measurement log.
+func ParseEventLog(measurementLog []byte) (*EventLog, error) {
+	var specID *specIDEvent
+	r := bytes.NewBuffer(measurementLog)
+	parseFn := parseRawEvent
+	var el EventLog
+	e, err := parseFn(r, specID)
+	if err != nil {
+		return nil, fmt.Errorf("parse first event: %v", err)
+	}
+	if e.typ == eventTypeNoAction {
+		specID, err = parseSpecIDEvent(e.data)
+		if err != nil {
+			return nil, fmt.Errorf("failed to parse spec ID event: %v", err)
+		}
+		for _, alg := range specID.algs {
+			switch tpm2.Algorithm(alg.ID) {
+			case tpm2.AlgSHA1:
+				el.Algs = append(el.Algs, HashSHA1)
+			case tpm2.AlgSHA256:
+				el.Algs = append(el.Algs, HashSHA256)
+			}
+		}
+		if len(el.Algs) == 0 {
+			return nil, fmt.Errorf("measurement log didn't use sha1 or sha256 digests")
+		}
+		// Switch to parsing crypto agile events. Don't include this in the
+		// replayed events since it intentionally doesn't extend the PCRs.
+		//
+		// Note that this doesn't actually guarentee that events have SHA256
+		// digests.
+		parseFn = parseRawEvent2
+	} else {
+		el.Algs = []HashAlg{HashSHA1}
+		el.rawEvents = append(el.rawEvents, e)
+	}
+	sequence := 1
+	for r.Len() != 0 {
+		e, err := parseFn(r, specID)
+		if err != nil {
+			return nil, err
+		}
+		e.sequence = sequence
+		sequence++
+		el.rawEvents = append(el.rawEvents, e)
+	}
+	return &el, nil
+}
+
+type specIDEvent struct {
+	algs []specAlgSize
+}
+
+type specAlgSize struct {
+	ID   uint16
+	Size uint16
+}
+
+// Expected values for various Spec ID Event fields.
+// https://trustedcomputinggroup.org/wp-content/uploads/EFI-Protocol-Specification-rev13-160330final.pdf#page=19
+var wantSignature = [16]byte{0x53, 0x70,
+	0x65, 0x63, 0x20, 0x49,
+	0x44, 0x20, 0x45, 0x76,
+	0x65, 0x6e, 0x74, 0x30,
+	0x33, 0x00} // "Spec ID Event03\0"
+
+const (
+	wantMajor  = 2
+	wantMinor  = 0
+	wantErrata = 0
+)
+
+// parseSpecIDEvent parses a TCG_EfiSpecIDEventStruct structure from the reader.
+//
+// https://trustedcomputinggroup.org/wp-content/uploads/EFI-Protocol-Specification-rev13-160330final.pdf#page=18
+func parseSpecIDEvent(b []byte) (*specIDEvent, error) {
+	r := bytes.NewReader(b)
+	var header struct {
+		Signature     [16]byte
+		PlatformClass uint32
+		VersionMinor  uint8
+		VersionMajor  uint8
+		Errata        uint8
+		UintnSize     uint8
+		NumAlgs       uint32
+	}
+	if err := binary.Read(r, binary.LittleEndian, &header); err != nil {
+		return nil, fmt.Errorf("reading event header: %v", err)
+	}
+	if header.Signature != wantSignature {
+		return nil, fmt.Errorf("invalid spec id signature: %x", header.Signature)
+	}
+	if header.VersionMajor != wantMajor {
+		return nil, fmt.Errorf("invalid spec major version, got %02x, wanted %02x",
+			header.VersionMajor, wantMajor)
+	}
+	if header.VersionMinor != wantMinor {
+		return nil, fmt.Errorf("invalid spec minor version, got %02x, wanted %02x",
+			header.VersionMajor, wantMinor)
+	}
+
+	// TODO(ericchiang): Check errata? Or do we expect that to change in ways
+	// we're okay with?
+
+	specAlg := specAlgSize{}
+	e := specIDEvent{}
+	for i := 0; i < int(header.NumAlgs); i++ {
+		if err := binary.Read(r, binary.LittleEndian, &specAlg); err != nil {
+			return nil, fmt.Errorf("reading algorithm: %v", err)
+		}
+		e.algs = append(e.algs, specAlg)
+	}
+
+	var vendorInfoSize uint8
+	if err := binary.Read(r, binary.LittleEndian, &vendorInfoSize); err != nil {
+		return nil, fmt.Errorf("reading vender info size: %v", err)
+	}
+	if r.Len() != int(vendorInfoSize) {
+		return nil, fmt.Errorf("reading vendor info, expected %d remaining bytes, got %d", vendorInfoSize, r.Len())
+	}
+	return &e, nil
+}
+
+type digest struct {
+	hash crypto.Hash
+	data []byte
+}
+
+type rawEvent struct {
+	sequence int
+	index    int
+	typ      EventType
+	data     []byte
+	digests  []digest
+}
+
+// TPM 1.2 event log format. See "5.1 SHA1 Event Log Entry Format"
+// https://trustedcomputinggroup.org/wp-content/uploads/EFI-Protocol-Specification-rev13-160330final.pdf#page=15
+type rawEventHeader struct {
+	PCRIndex  uint32
+	Type      uint32
+	Digest    [20]byte
+	EventSize uint32
+}
+
+type eventSizeErr struct {
+	eventSize uint32
+	logSize   int
+}
+
+func (e *eventSizeErr) Error() string {
+	return fmt.Sprintf("event data size (%d bytes) is greater than remaining measurement log (%d bytes)", e.eventSize, e.logSize)
+}
+
+func parseRawEvent(r *bytes.Buffer, specID *specIDEvent) (event rawEvent, err error) {
+	var h rawEventHeader
+	if err = binary.Read(r, binary.LittleEndian, &h); err != nil {
+		return event, err
+	}
+	if h.EventSize == 0 {
+		return event, errors.New("event data size is 0")
+	}
+	if h.EventSize > uint32(r.Len()) {
+		return event, &eventSizeErr{h.EventSize, r.Len()}
+	}
+
+	data := make([]byte, int(h.EventSize))
+	if _, err := io.ReadFull(r, data); err != nil {
+		return event, err
+	}
+
+	digests := []digest{{hash: crypto.SHA1, data: h.Digest[:]}}
+
+	return rawEvent{
+		typ:     EventType(h.Type),
+		data:    data,
+		index:   int(h.PCRIndex),
+		digests: digests,
+	}, nil
+}
+
+// TPM 2.0 event log format. See "5.2 Crypto Agile Log Entry Format"
+// https://trustedcomputinggroup.org/wp-content/uploads/EFI-Protocol-Specification-rev13-160330final.pdf#page=15
+type rawEvent2Header struct {
+	PCRIndex uint32
+	Type     uint32
+}
+
+func parseRawEvent2(r *bytes.Buffer, specID *specIDEvent) (event rawEvent, err error) {
+	var h rawEvent2Header
+
+	if err = binary.Read(r, binary.LittleEndian, &h); err != nil {
+		return event, err
+	}
+	event.typ = EventType(h.Type)
+	event.index = int(h.PCRIndex)
+
+	// parse the event digests
+	var numDigests uint32
+	if err := binary.Read(r, binary.LittleEndian, &numDigests); err != nil {
+		return event, err
+	}
+
+	for i := 0; i < int(numDigests); i++ {
+		var algID uint16
+		if err := binary.Read(r, binary.LittleEndian, &algID); err != nil {
+			return event, err
+		}
+		var digest digest
+
+		for _, alg := range specID.algs {
+			if alg.ID != algID {
+				continue
+			}
+			if uint16(r.Len()) < alg.Size {
+				return event, fmt.Errorf("reading digest: %v", io.ErrUnexpectedEOF)
+			}
+			digest.data = make([]byte, alg.Size)
+			digest.hash = HashAlg(alg.ID).cryptoHash()
+		}
+		if len(digest.data) == 0 {
+			return event, fmt.Errorf("unknown algorithm ID %x", algID)
+		}
+		if _, err := io.ReadFull(r, digest.data); err != nil {
+			return event, err
+		}
+		event.digests = append(event.digests, digest)
+	}
+
+	// parse event data
+	var eventSize uint32
+	if err = binary.Read(r, binary.LittleEndian, &eventSize); err != nil {
+		return event, err
+	}
+	if eventSize == 0 {
+		return event, errors.New("event data size is 0")
+	}
+	if eventSize > uint32(r.Len()) {
+		return event, &eventSizeErr{eventSize, r.Len()}
+	}
+	event.data = make([]byte, int(eventSize))
+	if _, err := io.ReadFull(r, event.data); err != nil {
+		return event, err
+	}
+	return event, err
+}
diff --git a/metropolis/node/core/tpm/eventlog/internal/BUILD.bazel b/metropolis/node/core/tpm/eventlog/internal/BUILD.bazel
new file mode 100644
index 0000000..48e1e81
--- /dev/null
+++ b/metropolis/node/core/tpm/eventlog/internal/BUILD.bazel
@@ -0,0 +1,12 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["events.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/core/tpm/eventlog/internal",
+    visibility = ["//metropolis/node/core/tpm/eventlog:__subpackages__"],
+    deps = [
+        "@com_github_google_certificate_transparency_go//asn1:go_default_library",
+        "@com_github_google_certificate_transparency_go//x509:go_default_library",
+    ],
+)
diff --git a/metropolis/node/core/tpm/eventlog/internal/events.go b/metropolis/node/core/tpm/eventlog/internal/events.go
new file mode 100644
index 0000000..d9b933b
--- /dev/null
+++ b/metropolis/node/core/tpm/eventlog/internal/events.go
@@ -0,0 +1,403 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Taken from go-attestation under Apache 2.0
+package internal
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"unicode/utf16"
+
+	"github.com/google/certificate-transparency-go/asn1"
+	"github.com/google/certificate-transparency-go/x509"
+)
+
+const (
+	// maxNameLen is the maximum accepted byte length for a name field.
+	// This value should be larger than any reasonable value.
+	maxNameLen = 2048
+	// maxDataLen is the maximum size in bytes of a variable data field.
+	// This value should be larger than any reasonable value.
+	maxDataLen = 1024 * 1024 // 1 Megabyte.
+)
+
+// GUIDs representing the contents of an UEFI_SIGNATURE_LIST.
+var (
+	hashSHA256SigGUID        = efiGUID{0xc1c41626, 0x504c, 0x4092, [8]byte{0xac, 0xa9, 0x41, 0xf9, 0x36, 0x93, 0x43, 0x28}}
+	hashSHA1SigGUID          = efiGUID{0x826ca512, 0xcf10, 0x4ac9, [8]byte{0xb1, 0x87, 0xbe, 0x01, 0x49, 0x66, 0x31, 0xbd}}
+	hashSHA224SigGUID        = efiGUID{0x0b6e5233, 0xa65c, 0x44c9, [8]byte{0x94, 0x07, 0xd9, 0xab, 0x83, 0xbf, 0xc8, 0xbd}}
+	hashSHA384SigGUID        = efiGUID{0xff3e5307, 0x9fd0, 0x48c9, [8]byte{0x85, 0xf1, 0x8a, 0xd5, 0x6c, 0x70, 0x1e, 0x01}}
+	hashSHA512SigGUID        = efiGUID{0x093e0fae, 0xa6c4, 0x4f50, [8]byte{0x9f, 0x1b, 0xd4, 0x1e, 0x2b, 0x89, 0xc1, 0x9a}}
+	keyRSA2048SigGUID        = efiGUID{0x3c5766e8, 0x269c, 0x4e34, [8]byte{0xaa, 0x14, 0xed, 0x77, 0x6e, 0x85, 0xb3, 0xb6}}
+	certRSA2048SHA256SigGUID = efiGUID{0xe2b36190, 0x879b, 0x4a3d, [8]byte{0xad, 0x8d, 0xf2, 0xe7, 0xbb, 0xa3, 0x27, 0x84}}
+	certRSA2048SHA1SigGUID   = efiGUID{0x67f8444f, 0x8743, 0x48f1, [8]byte{0xa3, 0x28, 0x1e, 0xaa, 0xb8, 0x73, 0x60, 0x80}}
+	certX509SigGUID          = efiGUID{0xa5c059a1, 0x94e4, 0x4aa7, [8]byte{0x87, 0xb5, 0xab, 0x15, 0x5c, 0x2b, 0xf0, 0x72}}
+	certHashSHA256SigGUID    = efiGUID{0x3bd2a492, 0x96c0, 0x4079, [8]byte{0xb4, 0x20, 0xfc, 0xf9, 0x8e, 0xf1, 0x03, 0xed}}
+	certHashSHA384SigGUID    = efiGUID{0x7076876e, 0x80c2, 0x4ee6, [8]byte{0xaa, 0xd2, 0x28, 0xb3, 0x49, 0xa6, 0x86, 0x5b}}
+	certHashSHA512SigGUID    = efiGUID{0x446dbf63, 0x2502, 0x4cda, [8]byte{0xbc, 0xfa, 0x24, 0x65, 0xd2, 0xb0, 0xfe, 0x9d}}
+)
+
+// EventType describes the type of event signalled in the event log.
+type EventType uint32
+
+// 	BIOS Events (TCG PC Client Specific Implementation Specification for Conventional BIOS 1.21)
+const (
+	PrebootCert          EventType = 0x00000000
+	PostCode             EventType = 0x00000001
+	unused               EventType = 0x00000002
+	NoAction             EventType = 0x00000003
+	Separator            EventType = 0x00000004
+	Action               EventType = 0x00000005
+	EventTag             EventType = 0x00000006
+	SCRTMContents        EventType = 0x00000007
+	SCRTMVersion         EventType = 0x00000008
+	CpuMicrocode         EventType = 0x00000009
+	PlatformConfigFlags  EventType = 0x0000000A
+	TableOfDevices       EventType = 0x0000000B
+	CompactHash          EventType = 0x0000000C
+	Ipl                  EventType = 0x0000000D
+	IplPartitionData     EventType = 0x0000000E
+	NonhostCode          EventType = 0x0000000F
+	NonhostConfig        EventType = 0x00000010
+	NonhostInfo          EventType = 0x00000011
+	OmitBootDeviceEvents EventType = 0x00000012
+)
+
+// EFI Events (TCG EFI Platform Specification Version 1.22)
+const (
+	EFIEventBase               EventType = 0x80000000
+	EFIVariableDriverConfig    EventType = 0x80000001
+	EFIVariableBoot            EventType = 0x80000002
+	EFIBootServicesApplication EventType = 0x80000003
+	EFIBootServicesDriver      EventType = 0x80000004
+	EFIRuntimeServicesDriver   EventType = 0x80000005
+	EFIGPTEvent                EventType = 0x80000006
+	EFIAction                  EventType = 0x80000007
+	EFIPlatformFirmwareBlob    EventType = 0x80000008
+	EFIHandoffTables           EventType = 0x80000009
+	EFIHCRTMEvent              EventType = 0x80000010
+	EFIVariableAuthority       EventType = 0x800000e0
+)
+
+// ErrSigMissingGUID is returned if an EFI_SIGNATURE_DATA structure was parsed
+// successfully, however was missing the SignatureOwner GUID. This case is
+// handled specially as a workaround for a bug relating to authority events.
+var ErrSigMissingGUID = errors.New("signature data was missing owner GUID")
+
+var eventTypeNames = map[EventType]string{
+	PrebootCert:          "Preboot Cert",
+	PostCode:             "POST Code",
+	unused:               "Unused",
+	NoAction:             "No Action",
+	Separator:            "Separator",
+	Action:               "Action",
+	EventTag:             "Event Tag",
+	SCRTMContents:        "S-CRTM Contents",
+	SCRTMVersion:         "S-CRTM Version",
+	CpuMicrocode:         "CPU Microcode",
+	PlatformConfigFlags:  "Platform Config Flags",
+	TableOfDevices:       "Table of Devices",
+	CompactHash:          "Compact Hash",
+	Ipl:                  "IPL",
+	IplPartitionData:     "IPL Partition Data",
+	NonhostCode:          "Non-Host Code",
+	NonhostConfig:        "Non-HostConfig",
+	NonhostInfo:          "Non-Host Info",
+	OmitBootDeviceEvents: "Omit Boot Device Events",
+
+	EFIEventBase:               "EFI Event Base",
+	EFIVariableDriverConfig:    "EFI Variable Driver Config",
+	EFIVariableBoot:            "EFI Variable Boot",
+	EFIBootServicesApplication: "EFI Boot Services Application",
+	EFIBootServicesDriver:      "EFI Boot Services Driver",
+	EFIRuntimeServicesDriver:   "EFI Runtime Services Driver",
+	EFIGPTEvent:                "EFI GPT Event",
+	EFIAction:                  "EFI Action",
+	EFIPlatformFirmwareBlob:    "EFI Platform Firmware Blob",
+	EFIVariableAuthority:       "EFI Variable Authority",
+	EFIHandoffTables:           "EFI Handoff Tables",
+	EFIHCRTMEvent:              "EFI H-CRTM Event",
+}
+
+func (e EventType) String() string {
+	if s, ok := eventTypeNames[e]; ok {
+		return s
+	}
+	return fmt.Sprintf("EventType(0x%x)", uint32(e))
+}
+
+// UntrustedParseEventType returns the event type indicated by
+// the provided value.
+func UntrustedParseEventType(et uint32) (EventType, error) {
+	// "The value associated with a UEFI specific platform event type MUST be in
+	// the range between 0x80000000 and 0x800000FF, inclusive."
+	if (et < 0x80000000 && et > 0x800000FF) || (et < 0x0 && et > 0x12) {
+		return EventType(0), fmt.Errorf("event type not between [0x0, 0x12] or [0x80000000, 0x800000FF]: got %#x", et)
+	}
+	if _, ok := eventTypeNames[EventType(et)]; !ok {
+		return EventType(0), fmt.Errorf("unknown event type %#x", et)
+	}
+	return EventType(et), nil
+}
+
+// efiGUID represents the EFI_GUID type.
+// See section "2.3.1 Data Types" in the specification for more information.
+// type efiGUID [16]byte
+type efiGUID struct {
+	Data1 uint32
+	Data2 uint16
+	Data3 uint16
+	Data4 [8]byte
+}
+
+func (d efiGUID) String() string {
+	var u [8]byte
+	binary.BigEndian.PutUint32(u[:4], d.Data1)
+	binary.BigEndian.PutUint16(u[4:6], d.Data2)
+	binary.BigEndian.PutUint16(u[6:8], d.Data3)
+	return fmt.Sprintf("%x-%x-%x-%x-%x", u[:4], u[4:6], u[6:8], d.Data4[:2], d.Data4[2:])
+}
+
+// UEFIVariableDataHeader represents the leading fixed-size fields
+// within UEFI_VARIABLE_DATA.
+type UEFIVariableDataHeader struct {
+	VariableName       efiGUID
+	UnicodeNameLength  uint64 // uintN
+	VariableDataLength uint64 // uintN
+}
+
+// UEFIVariableData represents the UEFI_VARIABLE_DATA structure.
+type UEFIVariableData struct {
+	Header       UEFIVariableDataHeader
+	UnicodeName  []uint16
+	VariableData []byte // []int8
+}
+
+// ParseUEFIVariableData parses the data section of an event structured as
+// a UEFI variable.
+//
+// https://trustedcomputinggroup.org/wp-content/uploads/TCG_PCClient_Specific_Platform_Profile_for_TPM_2p0_1p04_PUBLIC.pdf#page=100
+func ParseUEFIVariableData(r io.Reader) (ret UEFIVariableData, err error) {
+	err = binary.Read(r, binary.LittleEndian, &ret.Header)
+	if err != nil {
+		return
+	}
+	if ret.Header.UnicodeNameLength > maxNameLen {
+		return UEFIVariableData{}, fmt.Errorf("unicode name too long: %d > %d", ret.Header.UnicodeNameLength, maxNameLen)
+	}
+	ret.UnicodeName = make([]uint16, ret.Header.UnicodeNameLength)
+	for i := 0; uint64(i) < ret.Header.UnicodeNameLength; i++ {
+		err = binary.Read(r, binary.LittleEndian, &ret.UnicodeName[i])
+		if err != nil {
+			return
+		}
+	}
+	if ret.Header.VariableDataLength > maxDataLen {
+		return UEFIVariableData{}, fmt.Errorf("variable data too long: %d > %d", ret.Header.VariableDataLength, maxDataLen)
+	}
+	ret.VariableData = make([]byte, ret.Header.VariableDataLength)
+	_, err = io.ReadFull(r, ret.VariableData)
+	return
+}
+
+func (v *UEFIVariableData) VarName() string {
+	return string(utf16.Decode(v.UnicodeName))
+}
+
+func (v *UEFIVariableData) SignatureData() (certs []x509.Certificate, hashes [][]byte, err error) {
+	return parseEfiSignatureList(v.VariableData)
+}
+
+// UEFIVariableAuthority describes the contents of a UEFI variable authority
+// event.
+type UEFIVariableAuthority struct {
+	Certs []x509.Certificate
+}
+
+// ParseUEFIVariableAuthority parses the data section of an event structured as
+// a UEFI variable authority.
+//
+// https://uefi.org/sites/default/files/resources/UEFI_Spec_2_8_final.pdf#page=1789
+func ParseUEFIVariableAuthority(r io.Reader) (UEFIVariableAuthority, error) {
+	v, err := ParseUEFIVariableData(r)
+	if err != nil {
+		return UEFIVariableAuthority{}, err
+	}
+	certs, err := parseEfiSignature(v.VariableData)
+	return UEFIVariableAuthority{Certs: certs}, err
+}
+
+// efiSignatureData represents the EFI_SIGNATURE_DATA type.
+// See section "31.4.1 Signature Database" in the specification for more information.
+type efiSignatureData struct {
+	SignatureOwner efiGUID
+	SignatureData  []byte // []int8
+}
+
+// efiSignatureList represents the EFI_SIGNATURE_LIST type.
+// See section "31.4.1 Signature Database" in the specification for more information.
+type efiSignatureListHeader struct {
+	SignatureType       efiGUID
+	SignatureListSize   uint32
+	SignatureHeaderSize uint32
+	SignatureSize       uint32
+}
+
+type efiSignatureList struct {
+	Header        efiSignatureListHeader
+	SignatureData []byte
+	Signatures    []byte
+}
+
+// parseEfiSignatureList parses a EFI_SIGNATURE_LIST structure.
+// The structure and related GUIDs are defined at:
+// https://uefi.org/sites/default/files/resources/UEFI_Spec_2_8_final.pdf#page=1790
+func parseEfiSignatureList(b []byte) ([]x509.Certificate, [][]byte, error) {
+	if len(b) < 28 {
+		// Being passed an empty signature list here appears to be valid
+		return nil, nil, nil
+	}
+	signatures := efiSignatureList{}
+	buf := bytes.NewReader(b)
+	certificates := []x509.Certificate{}
+	hashes := [][]byte{}
+
+	for buf.Len() > 0 {
+		err := binary.Read(buf, binary.LittleEndian, &signatures.Header)
+		if err != nil {
+			return nil, nil, err
+		}
+
+		if signatures.Header.SignatureHeaderSize > maxDataLen {
+			return nil, nil, fmt.Errorf("signature header too large: %d > %d", signatures.Header.SignatureHeaderSize, maxDataLen)
+		}
+		if signatures.Header.SignatureListSize > maxDataLen {
+			return nil, nil, fmt.Errorf("signature list too large: %d > %d", signatures.Header.SignatureListSize, maxDataLen)
+		}
+
+		signatureType := signatures.Header.SignatureType
+		switch signatureType {
+		case certX509SigGUID: // X509 certificate
+			for sigOffset := 0; uint32(sigOffset) < signatures.Header.SignatureListSize-28; {
+				signature := efiSignatureData{}
+				signature.SignatureData = make([]byte, signatures.Header.SignatureSize-16)
+				err := binary.Read(buf, binary.LittleEndian, &signature.SignatureOwner)
+				if err != nil {
+					return nil, nil, err
+				}
+				err = binary.Read(buf, binary.LittleEndian, &signature.SignatureData)
+				if err != nil {
+					return nil, nil, err
+				}
+				cert, err := x509.ParseCertificate(signature.SignatureData)
+				if err != nil {
+					return nil, nil, err
+				}
+				sigOffset += int(signatures.Header.SignatureSize)
+				certificates = append(certificates, *cert)
+			}
+		case hashSHA256SigGUID: // SHA256
+			for sigOffset := 0; uint32(sigOffset) < signatures.Header.SignatureListSize-28; {
+				signature := efiSignatureData{}
+				signature.SignatureData = make([]byte, signatures.Header.SignatureSize-16)
+				err := binary.Read(buf, binary.LittleEndian, &signature.SignatureOwner)
+				if err != nil {
+					return nil, nil, err
+				}
+				err = binary.Read(buf, binary.LittleEndian, &signature.SignatureData)
+				if err != nil {
+					return nil, nil, err
+				}
+				hashes = append(hashes, signature.SignatureData)
+				sigOffset += int(signatures.Header.SignatureSize)
+			}
+		case keyRSA2048SigGUID:
+			err = errors.New("unhandled RSA2048 key")
+		case certRSA2048SHA256SigGUID:
+			err = errors.New("unhandled RSA2048-SHA256 key")
+		case hashSHA1SigGUID:
+			err = errors.New("unhandled SHA1 hash")
+		case certRSA2048SHA1SigGUID:
+			err = errors.New("unhandled RSA2048-SHA1 key")
+		case hashSHA224SigGUID:
+			err = errors.New("unhandled SHA224 hash")
+		case hashSHA384SigGUID:
+			err = errors.New("unhandled SHA384 hash")
+		case hashSHA512SigGUID:
+			err = errors.New("unhandled SHA512 hash")
+		case certHashSHA256SigGUID:
+			err = errors.New("unhandled X509-SHA256 hash metadata")
+		case certHashSHA384SigGUID:
+			err = errors.New("unhandled X509-SHA384 hash metadata")
+		case certHashSHA512SigGUID:
+			err = errors.New("unhandled X509-SHA512 hash metadata")
+		default:
+			err = fmt.Errorf("unhandled signature type %s", signatureType)
+		}
+		if err != nil {
+			return nil, nil, err
+		}
+	}
+	return certificates, hashes, nil
+}
+
+// EFISignatureData represents the EFI_SIGNATURE_DATA type.
+// See section "31.4.1 Signature Database" in the specification
+// for more information.
+type EFISignatureData struct {
+	SignatureOwner efiGUID
+	SignatureData  []byte // []int8
+}
+
+func parseEfiSignature(b []byte) ([]x509.Certificate, error) {
+	certificates := []x509.Certificate{}
+
+	if len(b) < 16 {
+		return nil, fmt.Errorf("invalid signature: buffer smaller than header (%d < %d)", len(b), 16)
+	}
+
+	buf := bytes.NewReader(b)
+	signature := EFISignatureData{}
+	signature.SignatureData = make([]byte, len(b)-16)
+
+	if err := binary.Read(buf, binary.LittleEndian, &signature.SignatureOwner); err != nil {
+		return certificates, err
+	}
+	if err := binary.Read(buf, binary.LittleEndian, &signature.SignatureData); err != nil {
+		return certificates, err
+	}
+
+	cert, err := x509.ParseCertificate(signature.SignatureData)
+	if err == nil {
+		certificates = append(certificates, *cert)
+	} else {
+		// A bug in shim may cause an event to be missing the SignatureOwner GUID.
+		// We handle this, but signal back to the caller using ErrSigMissingGUID.
+		if _, isStructuralErr := err.(asn1.StructuralError); isStructuralErr {
+			var err2 error
+			cert, err2 = x509.ParseCertificate(b)
+			if err2 == nil {
+				certificates = append(certificates, *cert)
+				err = ErrSigMissingGUID
+			}
+		}
+	}
+	return certificates, err
+}
diff --git a/metropolis/node/core/tpm/eventlog/secureboot.go b/metropolis/node/core/tpm/eventlog/secureboot.go
new file mode 100644
index 0000000..f117d30
--- /dev/null
+++ b/metropolis/node/core/tpm/eventlog/secureboot.go
@@ -0,0 +1,210 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Taken and pruned from go-attestation under Apache 2.0
+package eventlog
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+
+	"github.com/google/certificate-transparency-go/x509"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/tpm/eventlog/internal"
+)
+
+// SecurebootState describes the secure boot status of a machine, as determined
+// by processing its event log.
+type SecurebootState struct {
+	Enabled bool
+
+	// PlatformKeys enumerates keys which can sign a key exchange key.
+	PlatformKeys []x509.Certificate
+	// PlatformKeys enumerates key hashes which can sign a key exchange key.
+	PlatformKeyHashes [][]byte
+
+	// ExchangeKeys enumerates keys which can sign a database of permitted or
+	// forbidden keys.
+	ExchangeKeys []x509.Certificate
+	// ExchangeKeyHashes enumerates key hashes which can sign a database or
+	// permitted or forbidden keys.
+	ExchangeKeyHashes [][]byte
+
+	// PermittedKeys enumerates keys which may sign binaries to run.
+	PermittedKeys []x509.Certificate
+	// PermittedHashes enumerates hashes which permit binaries to run.
+	PermittedHashes [][]byte
+
+	// ForbiddenKeys enumerates keys which must not permit a binary to run.
+	ForbiddenKeys []x509.Certificate
+	// ForbiddenKeys enumerates hashes which must not permit a binary to run.
+	ForbiddenHashes [][]byte
+
+	// PreSeparatorAuthority describes the use of a secure-boot key to authorize
+	// the execution of a binary before the separator.
+	PreSeparatorAuthority []x509.Certificate
+	// PostSeparatorAuthority describes the use of a secure-boot key to authorize
+	// the execution of a binary after the separator.
+	PostSeparatorAuthority []x509.Certificate
+}
+
+// ParseSecurebootState parses a series of events to determine the
+// configuration of secure boot on a device. An error is returned if
+// the state cannot be determined, or if the event log is structured
+// in such a way that it may have been tampered post-execution of
+// platform firmware.
+func ParseSecurebootState(events []Event) (*SecurebootState, error) {
+	// This algorithm verifies the following:
+	// - All events in PCR 7 have event types which are expected in PCR 7.
+	// - All events are parsable according to their event type.
+	// - All events have digests values corresponding to their data/event type.
+	// - No unverifiable events were present.
+	// - All variables are specified before the separator and never duplicated.
+	// - The SecureBoot variable has a value of 0 or 1.
+	// - If SecureBoot was 1 (enabled), authority events were present indicating
+	//   keys were used to perform verification.
+	// - If SecureBoot was 1 (enabled), platform + exchange + database keys
+	//   were specified.
+	// - No UEFI debugger was attached.
+
+	var (
+		out           SecurebootState
+		seenSeparator bool
+		seenAuthority bool
+		seenVars      = map[string]bool{}
+	)
+
+	for _, e := range events {
+		if e.Index != 7 {
+			continue
+		}
+
+		et, err := internal.UntrustedParseEventType(uint32(e.Type))
+		if err != nil {
+			return nil, fmt.Errorf("unrecognised event type: %v", err)
+		}
+
+		digestVerify := e.digestEquals(e.Data)
+		switch et {
+		case internal.Separator:
+			if seenSeparator {
+				return nil, fmt.Errorf("duplicate separator at event %d", e.sequence)
+			}
+			seenSeparator = true
+			if !bytes.Equal(e.Data, []byte{0, 0, 0, 0}) {
+				return nil, fmt.Errorf("invalid separator data at event %d: %v", e.sequence, e.Data)
+			}
+			if digestVerify != nil {
+				return nil, fmt.Errorf("invalid separator digest at event %d: %v", e.sequence, digestVerify)
+			}
+
+		case internal.EFIAction:
+			if string(e.Data) == "UEFI Debug Mode" {
+				return nil, errors.New("a UEFI debugger was present during boot")
+			}
+			return nil, fmt.Errorf("event %d: unexpected EFI action event", e.sequence)
+
+		case internal.EFIVariableDriverConfig:
+			v, err := internal.ParseUEFIVariableData(bytes.NewReader(e.Data))
+			if err != nil {
+				return nil, fmt.Errorf("failed parsing EFI variable at event %d: %v", e.sequence, err)
+			}
+			if _, seenBefore := seenVars[v.VarName()]; seenBefore {
+				return nil, fmt.Errorf("duplicate EFI variable %q at event %d", v.VarName(), e.sequence)
+			}
+			seenVars[v.VarName()] = true
+			if seenSeparator {
+				return nil, fmt.Errorf("event %d: variable %q specified after separator", e.sequence, v.VarName())
+			}
+
+			if digestVerify != nil {
+				return nil, fmt.Errorf("invalid digest for variable %q on event %d: %v", v.VarName(), e.sequence, digestVerify)
+			}
+
+			switch v.VarName() {
+			case "SecureBoot":
+				if len(v.VariableData) != 1 {
+					return nil, fmt.Errorf("event %d: SecureBoot data len is %d, expected 1", e.sequence, len(v.VariableData))
+				}
+				out.Enabled = v.VariableData[0] == 1
+			case "PK":
+				if out.PlatformKeys, out.PlatformKeyHashes, err = v.SignatureData(); err != nil {
+					return nil, fmt.Errorf("event %d: failed parsing platform keys: %v", e.sequence, err)
+				}
+			case "KEK":
+				if out.ExchangeKeys, out.ExchangeKeyHashes, err = v.SignatureData(); err != nil {
+					return nil, fmt.Errorf("event %d: failed parsing key exchange keys: %v", e.sequence, err)
+				}
+			case "db":
+				if out.PermittedKeys, out.PermittedHashes, err = v.SignatureData(); err != nil {
+					return nil, fmt.Errorf("event %d: failed parsing signature database: %v", e.sequence, err)
+				}
+			case "dbx":
+				if out.ForbiddenKeys, out.ForbiddenHashes, err = v.SignatureData(); err != nil {
+					return nil, fmt.Errorf("event %d: failed parsing forbidden signature database: %v", e.sequence, err)
+				}
+			}
+
+		case internal.EFIVariableAuthority:
+			a, err := internal.ParseUEFIVariableAuthority(bytes.NewReader(e.Data))
+			if err != nil {
+				// Workaround for: https://github.com/google/go-attestation/issues/157
+				if err == internal.ErrSigMissingGUID {
+					// Versions of shim which do not carry
+					// https://github.com/rhboot/shim/commit/8a27a4809a6a2b40fb6a4049071bf96d6ad71b50
+					// have an erroneous additional byte in the event, which breaks digest
+					// verification. If verification failed, we try removing the last byte.
+					if digestVerify != nil {
+						digestVerify = e.digestEquals(e.Data[:len(e.Data)-1])
+					}
+				} else {
+					return nil, fmt.Errorf("failed parsing EFI variable authority at event %d: %v", e.sequence, err)
+				}
+			}
+			seenAuthority = true
+			if digestVerify != nil {
+				return nil, fmt.Errorf("invalid digest for authority on event %d: %v", e.sequence, digestVerify)
+			}
+			if !seenSeparator {
+				out.PreSeparatorAuthority = append(out.PreSeparatorAuthority, a.Certs...)
+			} else {
+				out.PostSeparatorAuthority = append(out.PostSeparatorAuthority, a.Certs...)
+			}
+
+		default:
+			return nil, fmt.Errorf("unexpected event type: %v", et)
+		}
+	}
+
+	if !out.Enabled {
+		return &out, nil
+	}
+
+	if !seenAuthority {
+		return nil, errors.New("secure boot was enabled but no key was used")
+	}
+	if len(out.PlatformKeys) == 0 && len(out.PlatformKeyHashes) == 0 {
+		return nil, errors.New("secure boot was enabled but no platform keys were known")
+	}
+	if len(out.ExchangeKeys) == 0 && len(out.ExchangeKeyHashes) == 0 {
+		return nil, errors.New("secure boot was enabled but no key exchange keys were known")
+	}
+	if len(out.PermittedKeys) == 0 && len(out.PermittedHashes) == 0 {
+		return nil, errors.New("secure boot was enabled but no keys or hashes were permitted")
+	}
+	return &out, nil
+}
diff --git a/metropolis/node/core/tpm/tpm.go b/metropolis/node/core/tpm/tpm.go
new file mode 100644
index 0000000..76f4f92
--- /dev/null
+++ b/metropolis/node/core/tpm/tpm.go
@@ -0,0 +1,561 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tpm
+
+import (
+	"bytes"
+	"crypto"
+	"crypto/rand"
+	"crypto/rsa"
+	"crypto/x509"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/gogo/protobuf/proto"
+	tpmpb "github.com/google/go-tpm-tools/proto"
+	"github.com/google/go-tpm-tools/tpm2tools"
+	"github.com/google/go-tpm/tpm2"
+	"github.com/google/go-tpm/tpmutil"
+	"github.com/pkg/errors"
+	"golang.org/x/sys/unix"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/sysfs"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+)
+
+var (
+	// SecureBootPCRs are all PCRs that measure the current Secure Boot configuration.
+	// This is what we want if we rely on secure boot to verify boot integrity. The firmware
+	// hashes the secure boot policy and custom keys into the PCR.
+	//
+	// This requires an extra step that provisions the custom keys.
+	//
+	// Some background: https://mjg59.dreamwidth.org/48897.html?thread=1847297
+	// (the initramfs issue mentioned in the article has been solved by integrating
+	// it into the kernel binary, and we don't have a shim bootloader)
+	//
+	// PCR7 alone is not sufficient - it needs to be combined with firmware measurements.
+	SecureBootPCRs = []int{7}
+
+	// FirmwarePCRs are alle PCRs that contain the firmware measurements
+	// See https://trustedcomputinggroup.org/wp-content/uploads/TCG_EFI_Platform_1_22_Final_-v15.pdf
+	FirmwarePCRs = []int{
+		0, // platform firmware
+		2, // option ROM code
+		3, // option ROM configuration and data
+	}
+
+	// FullSystemPCRs are all PCRs that contain any measurements up to the currently running EFI payload.
+	FullSystemPCRs = []int{
+		0, // platform firmware
+		1, // host platform configuration
+		2, // option ROM code
+		3, // option ROM configuration and data
+		4, // EFI payload
+	}
+
+	// Using FullSystemPCRs is the most secure, but also the most brittle option since updating the EFI
+	// binary, updating the platform firmware, changing platform settings or updating the binary
+	// would invalidate the sealed data. It's annoying (but possible) to predict values for PCR4,
+	// and even more annoying for the firmware PCR (comparison to known values on similar hardware
+	// is the only thing that comes to mind).
+	//
+	// See also: https://github.com/mxre/sealkey (generates PCR4 from EFI image, BSD license)
+	//
+	// Using only SecureBootPCRs is the easiest and still reasonably secure, if we assume that the
+	// platform knows how to take care of itself (i.e. Intel Boot Guard), and that secure boot
+	// is implemented properly. It is, however, a much larger amount of code we need to trust.
+	//
+	// We do not care about PCR 5 (GPT partition table) since modifying it is harmless. All of
+	// the boot options and cmdline are hardcoded in the kernel image, and we use no bootloader,
+	// so there's no PCR for bootloader configuration or kernel cmdline.
+)
+
+var (
+	numSRTMPCRs = 16
+	srtmPCRs    = tpm2.PCRSelection{Hash: tpm2.AlgSHA256, PCRs: []int{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}}
+	// TCG Trusted Platform Module Library Level 00 Revision 0.99 Table 6
+	tpmGeneratedValue = uint32(0xff544347)
+)
+
+var (
+	// ErrNotExists is returned when no TPMs are available in the system
+	ErrNotExists = errors.New("no TPMs found")
+	// ErrNotInitialized is returned when this package was not initialized successfully
+	ErrNotInitialized = errors.New("no TPM was initialized")
+)
+
+// Singleton since the TPM is too
+var tpm *TPM
+
+// We're serializing all TPM operations since it has a limited number of handles and recovering
+// if it runs out is difficult to implement correctly. Might also be marginally more secure.
+var lock sync.Mutex
+
+// TPM represents a high-level interface to a connected TPM 2.0
+type TPM struct {
+	logger logtree.LeveledLogger
+	device io.ReadWriteCloser
+
+	// We keep the AK loaded since it's used fairly often and deriving it is expensive
+	akHandleCache tpmutil.Handle
+	akPublicKey   crypto.PublicKey
+}
+
+// Initialize finds and opens the TPM (if any). If there is no TPM available it returns
+// ErrNotExists
+func Initialize(logger logtree.LeveledLogger) error {
+	lock.Lock()
+	defer lock.Unlock()
+	tpmDir, err := os.Open("/sys/class/tpm")
+	if err != nil {
+		return errors.Wrap(err, "failed to open sysfs TPM class")
+	}
+	defer tpmDir.Close()
+
+	tpms, err := tpmDir.Readdirnames(2)
+	if err != nil {
+		return errors.Wrap(err, "failed to read TPM device class")
+	}
+
+	if len(tpms) == 0 {
+		return ErrNotExists
+	}
+	if len(tpms) > 1 {
+		// If this is changed GetMeasurementLog() needs to be updated too
+		logger.Warningf("Found more than one TPM, using the first one")
+	}
+	tpmName := tpms[0]
+	ueventData, err := sysfs.ReadUevents(filepath.Join("/sys/class/tpm", tpmName, "uevent"))
+	majorDev, err := strconv.Atoi(ueventData["MAJOR"])
+	if err != nil {
+		return fmt.Errorf("failed to convert uevent: %w", err)
+	}
+	minorDev, err := strconv.Atoi(ueventData["MINOR"])
+	if err != nil {
+		return fmt.Errorf("failed to convert uevent: %w", err)
+	}
+	if err := unix.Mknod("/dev/tpm", 0600|unix.S_IFCHR, int(unix.Mkdev(uint32(majorDev), uint32(minorDev)))); err != nil {
+		return errors.Wrap(err, "failed to create TPM device node")
+	}
+	device, err := tpm2.OpenTPM("/dev/tpm")
+	if err != nil {
+		return errors.Wrap(err, "failed to open TPM")
+	}
+	tpm = &TPM{
+		device: device,
+		logger: logger,
+	}
+	return nil
+}
+
+// GenerateSafeKey uses two sources of randomness (Kernel & TPM) to generate the key
+func GenerateSafeKey(size uint16) ([]byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return []byte{}, ErrNotInitialized
+	}
+	encryptionKeyHost := make([]byte, size)
+	if _, err := io.ReadFull(rand.Reader, encryptionKeyHost); err != nil {
+		return []byte{}, errors.Wrap(err, "failed to generate host portion of new key")
+	}
+	var encryptionKeyTPM []byte
+	for i := 48; i > 0; i-- {
+		tpmKeyPart, err := tpm2.GetRandom(tpm.device, size-uint16(len(encryptionKeyTPM)))
+		if err != nil {
+			return []byte{}, errors.Wrap(err, "failed to generate TPM portion of new key")
+		}
+		encryptionKeyTPM = append(encryptionKeyTPM, tpmKeyPart...)
+		if len(encryptionKeyTPM) >= int(size) {
+			break
+		}
+	}
+
+	if len(encryptionKeyTPM) != int(size) {
+		return []byte{}, fmt.Errorf("got incorrect amount of TPM randomess: %v, requested %v", len(encryptionKeyTPM), size)
+	}
+
+	encryptionKey := make([]byte, size)
+	for i := uint16(0); i < size; i++ {
+		encryptionKey[i] = encryptionKeyHost[i] ^ encryptionKeyTPM[i]
+	}
+	return encryptionKey, nil
+}
+
+// Seal seals sensitive data and only allows access if the current platform configuration in
+// matches the one the data was sealed on.
+func Seal(data []byte, pcrs []int) ([]byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return []byte{}, ErrNotInitialized
+	}
+	srk, err := tpm2tools.StorageRootKeyRSA(tpm.device)
+	if err != nil {
+		return []byte{}, errors.Wrap(err, "failed to load TPM SRK")
+	}
+	defer srk.Close()
+	sealedKey, err := srk.Seal(pcrs, data)
+	sealedKeyRaw, err := proto.Marshal(sealedKey)
+	if err != nil {
+		return []byte{}, errors.Wrapf(err, "failed to marshal sealed data")
+	}
+	return sealedKeyRaw, nil
+}
+
+// Unseal unseals sensitive data if the current platform configuration allows and sealing constraints
+// allow it.
+func Unseal(data []byte) ([]byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return []byte{}, ErrNotInitialized
+	}
+	srk, err := tpm2tools.StorageRootKeyRSA(tpm.device)
+	if err != nil {
+		return []byte{}, errors.Wrap(err, "failed to load TPM SRK")
+	}
+	defer srk.Close()
+
+	var sealedKey tpmpb.SealedBytes
+	if err := proto.Unmarshal(data, &sealedKey); err != nil {
+		return []byte{}, errors.Wrap(err, "failed to decode sealed data")
+	}
+	// Logging this for auditing purposes
+	pcrList := []string{}
+	for _, pcr := range sealedKey.Pcrs {
+		pcrList = append(pcrList, string(pcr))
+	}
+	tpm.logger.Infof("Attempting to unseal data protected with PCRs %s", strings.Join(pcrList, ","))
+	unsealedData, err := srk.Unseal(&sealedKey)
+	if err != nil {
+		return []byte{}, errors.Wrap(err, "failed to unseal data")
+	}
+	return unsealedData, nil
+}
+
+// Standard AK template for RSA2048 non-duplicatable restricted signing for attestation
+var akTemplate = tpm2.Public{
+	Type:       tpm2.AlgRSA,
+	NameAlg:    tpm2.AlgSHA256,
+	Attributes: tpm2.FlagSignerDefault,
+	RSAParameters: &tpm2.RSAParams{
+		Sign: &tpm2.SigScheme{
+			Alg:  tpm2.AlgRSASSA,
+			Hash: tpm2.AlgSHA256,
+		},
+		KeyBits: 2048,
+	},
+}
+
+func loadAK() error {
+	var err error
+	// Rationale: The AK is an EK-equivalent key and used only for attestation. Using a non-primary
+	// key here would require us to store the wrapped version somewhere, which is inconvenient.
+	// This being a primary key in the Endorsement hierarchy means that it can always be recreated
+	// and can never be "destroyed". Under our security model this is of no concern since we identify
+	// a node by its IK (Identity Key) which we can destroy.
+	tpm.akHandleCache, tpm.akPublicKey, err = tpm2.CreatePrimary(tpm.device, tpm2.HandleEndorsement,
+		tpm2.PCRSelection{}, "", "", akTemplate)
+	return err
+}
+
+// Process documented in TCG EK Credential Profile 2.2.1
+func loadEK() (tpmutil.Handle, crypto.PublicKey, error) {
+	// The EK is a primary key which is supposed to be certified by the manufacturer of the TPM.
+	// Its public attributes are standardized in TCG EK Credential Profile 2.0 Table 1. These need
+	// to match exactly or we aren't getting the key the manufacturere signed. tpm2tools contains
+	// such a template already, so we're using that instead of redoing it ourselves.
+	// This ignores the more complicated ways EKs can be specified, the additional stuff you can do
+	// is just absolutely crazy (see 2.2.1.2 onward)
+	return tpm2.CreatePrimary(tpm.device, tpm2.HandleEndorsement,
+		tpm2.PCRSelection{}, "", "", tpm2tools.DefaultEKTemplateRSA())
+}
+
+// GetAKPublic gets the TPM2T_PUBLIC of the AK key
+func GetAKPublic() ([]byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return []byte{}, ErrNotInitialized
+	}
+	if tpm.akHandleCache == tpmutil.Handle(0) {
+		if err := loadAK(); err != nil {
+			return []byte{}, fmt.Errorf("failed to load AK primary key: %w", err)
+		}
+	}
+	public, _, _, err := tpm2.ReadPublic(tpm.device, tpm.akHandleCache)
+	if err != nil {
+		return []byte{}, err
+	}
+	return public.Encode()
+}
+
+// TCG TPM v2.0 Provisioning Guidance v1.0 7.8 Table 2 and
+// TCG EK Credential Profile v2.1 2.2.1.4 de-facto Standard for Windows
+// These are both non-normative and reference Windows 10 documentation that's no longer available :(
+// But in practice this is what people are using, so if it's normative or not doesn't really matter
+const ekCertHandle = 0x01c00002
+
+// GetEKPublic gets the public key and (if available) Certificate of the EK
+func GetEKPublic() ([]byte, []byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return []byte{}, []byte{}, ErrNotInitialized
+	}
+	ekHandle, publicRaw, err := loadEK()
+	if err != nil {
+		return []byte{}, []byte{}, fmt.Errorf("failed to load EK primary key: %w", err)
+	}
+	defer tpm2.FlushContext(tpm.device, ekHandle)
+	// Don't question the use of HandleOwner, that's the Standard™
+	ekCertRaw, err := tpm2.NVReadEx(tpm.device, ekCertHandle, tpm2.HandleOwner, "", 0)
+	if err != nil {
+		return []byte{}, []byte{}, err
+	}
+
+	publicKey, err := x509.MarshalPKIXPublicKey(publicRaw)
+	if err != nil {
+		return []byte{}, []byte{}, err
+	}
+
+	return publicKey, ekCertRaw, nil
+}
+
+// MakeAKChallenge generates a challenge for TPM residency and attributes of the AK
+func MakeAKChallenge(ekPubKey, akPub []byte, nonce []byte) ([]byte, []byte, error) {
+	ekPubKeyData, err := x509.ParsePKIXPublicKey(ekPubKey)
+	if err != nil {
+		return []byte{}, []byte{}, fmt.Errorf("failed to decode EK pubkey: %w", err)
+	}
+	akPubData, err := tpm2.DecodePublic(akPub)
+	if err != nil {
+		return []byte{}, []byte{}, fmt.Errorf("failed to decode AK public part: %w", err)
+	}
+	// Make sure we're attesting the right attributes (in particular Restricted)
+	if !akPubData.MatchesTemplate(akTemplate) {
+		return []byte{}, []byte{}, errors.New("the key being challenged is not a valid AK")
+	}
+	akName, err := akPubData.Name()
+	if err != nil {
+		return []byte{}, []byte{}, fmt.Errorf("failed to derive AK name: %w", err)
+	}
+	return generateRSA(akName.Digest, ekPubKeyData.(*rsa.PublicKey), 16, nonce, rand.Reader)
+}
+
+// SolveAKChallenge solves a challenge for TPM residency of the AK
+func SolveAKChallenge(credBlob, secretChallenge []byte) ([]byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return []byte{}, ErrNotInitialized
+	}
+	if tpm.akHandleCache == tpmutil.Handle(0) {
+		if err := loadAK(); err != nil {
+			return []byte{}, fmt.Errorf("failed to load AK primary key: %w", err)
+		}
+	}
+
+	ekHandle, _, err := loadEK()
+	if err != nil {
+		return []byte{}, fmt.Errorf("failed to load EK: %w", err)
+	}
+	defer tpm2.FlushContext(tpm.device, ekHandle)
+
+	// This is necessary since the EK requires an endorsement handle policy in its session
+	// For us this is stupid because we keep all hierarchies open anyways since a) we cannot safely
+	// store secrets on the OS side pre-global unlock and b) it makes no sense in this security model
+	// since an uncompromised host OS will not let an untrusted entity attest as itself and a
+	// compromised OS can either not pass PCR policy checks or the game's already over (you
+	// successfully runtime-exploited a production Smalltown Core)
+	endorsementSession, _, err := tpm2.StartAuthSession(
+		tpm.device,
+		tpm2.HandleNull,
+		tpm2.HandleNull,
+		make([]byte, 16),
+		nil,
+		tpm2.SessionPolicy,
+		tpm2.AlgNull,
+		tpm2.AlgSHA256)
+	if err != nil {
+		panic(err)
+	}
+	defer tpm2.FlushContext(tpm.device, endorsementSession)
+
+	_, err = tpm2.PolicySecret(tpm.device, tpm2.HandleEndorsement, tpm2.AuthCommand{Session: tpm2.HandlePasswordSession, Attributes: tpm2.AttrContinueSession}, endorsementSession, nil, nil, nil, 0)
+	if err != nil {
+		return []byte{}, fmt.Errorf("failed to make a policy secret session: %w", err)
+	}
+
+	for {
+		solution, err := tpm2.ActivateCredentialUsingAuth(tpm.device, []tpm2.AuthCommand{
+			{Session: tpm2.HandlePasswordSession, Attributes: tpm2.AttrContinueSession}, // Use standard no-password authentication
+			{Session: endorsementSession, Attributes: tpm2.AttrContinueSession},         // Use a full policy session for the EK
+		}, tpm.akHandleCache, ekHandle, credBlob, secretChallenge)
+		if warn, ok := err.(tpm2.Warning); ok && warn.Code == tpm2.RCRetry {
+			time.Sleep(100 * time.Millisecond)
+			continue
+		}
+		return solution, err
+	}
+}
+
+// FlushTransientHandles flushes all sessions and non-persistent handles
+func FlushTransientHandles() error {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return ErrNotInitialized
+	}
+	flushHandleTypes := []tpm2.HandleType{tpm2.HandleTypeTransient, tpm2.HandleTypeLoadedSession, tpm2.HandleTypeSavedSession}
+	for _, handleType := range flushHandleTypes {
+		handles, err := tpm2tools.Handles(tpm.device, handleType)
+		if err != nil {
+			return err
+		}
+		for _, handle := range handles {
+			if err := tpm2.FlushContext(tpm.device, handle); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// AttestPlatform performs a PCR quote using the AK and returns the quote and its signature
+func AttestPlatform(nonce []byte) ([]byte, []byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return []byte{}, []byte{}, ErrNotInitialized
+	}
+	if tpm.akHandleCache == tpmutil.Handle(0) {
+		if err := loadAK(); err != nil {
+			return []byte{}, []byte{}, fmt.Errorf("failed to load AK primary key: %w", err)
+		}
+	}
+	// We only care about SHA256 since SHA1 is weak. This is supported on at least GCE and
+	// Intel / AMD fTPM, which is good enough for now. Alg is null because that would just hash the
+	// nonce, which is dumb.
+	quote, signature, err := tpm2.Quote(tpm.device, tpm.akHandleCache, "", "", nonce, srtmPCRs,
+		tpm2.AlgNull)
+	if err != nil {
+		return []byte{}, []byte{}, fmt.Errorf("failed to quote PCRs: %w", err)
+	}
+	return quote, signature.RSA.Signature, err
+}
+
+// VerifyAttestPlatform verifies a given attestation. You can rely on all data coming back as being
+// from the TPM on which the AK is bound to.
+func VerifyAttestPlatform(nonce, akPub, quote, signature []byte) (*tpm2.AttestationData, error) {
+	hash := crypto.SHA256.New()
+	hash.Write(quote)
+
+	akPubData, err := tpm2.DecodePublic(akPub)
+	if err != nil {
+		return nil, fmt.Errorf("invalid AK: %w", err)
+	}
+	akPublicKey, err := akPubData.Key()
+	if err != nil {
+		return nil, fmt.Errorf("invalid AK: %w", err)
+	}
+	akRSAKey, ok := akPublicKey.(*rsa.PublicKey)
+	if !ok {
+		return nil, errors.New("invalid AK: invalid key type")
+	}
+
+	if err := rsa.VerifyPKCS1v15(akRSAKey, crypto.SHA256, hash.Sum(nil), signature); err != nil {
+		return nil, err
+	}
+
+	quoteData, err := tpm2.DecodeAttestationData(quote)
+	if err != nil {
+		return nil, err
+	}
+	// quoteData.Magic works together with the TPM's Restricted key attribute. If this attribute is set
+	// (which it needs to be for the AK to be considered valid) the TPM will not sign external data
+	// having this prefix with such a key. Only data that originates inside the TPM like quotes and
+	// key certifications can have this prefix and sill be signed by a restricted key. This check
+	// is thus vital, otherwise somebody can just feed the TPM an arbitrary attestation to sign with
+	// its AK and this function will happily accept the forged attestation.
+	if quoteData.Magic != tpmGeneratedValue {
+		return nil, errors.New("invalid TPM quote: data marker for internal data not set - forged attestation")
+	}
+	if quoteData.Type != tpm2.TagAttestQuote {
+		return nil, errors.New("invalid TPM qoute: not a TPM quote")
+	}
+	if !bytes.Equal(quoteData.ExtraData, nonce) {
+		return nil, errors.New("invalid TPM quote: wrong nonce")
+	}
+
+	return quoteData, nil
+}
+
+// GetPCRs returns all SRTM PCRs in-order
+func GetPCRs() ([][]byte, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	if tpm == nil {
+		return [][]byte{}, ErrNotInitialized
+	}
+	pcrs := make([][]byte, numSRTMPCRs)
+
+	// The TPM can (and most do) return partial results. Let's just retry as many times as we have
+	// PCRs since each read should return at least one PCR.
+readLoop:
+	for i := 0; i < numSRTMPCRs; i++ {
+		sel := tpm2.PCRSelection{Hash: tpm2.AlgSHA256}
+		for pcrN := 0; pcrN < numSRTMPCRs; pcrN++ {
+			if len(pcrs[pcrN]) == 0 {
+				sel.PCRs = append(sel.PCRs, pcrN)
+			}
+		}
+
+		readPCRs, err := tpm2.ReadPCRs(tpm.device, sel)
+		if err != nil {
+			return nil, fmt.Errorf("failed to read PCRs: %w", err)
+		}
+
+		for pcrN, pcr := range readPCRs {
+			pcrs[pcrN] = pcr
+		}
+		for _, pcr := range pcrs {
+			// If at least one PCR is still not read, continue
+			if len(pcr) == 0 {
+				continue readLoop
+			}
+		}
+		break
+	}
+
+	return pcrs, nil
+}
+
+// GetMeasurmentLog returns the binary log of all data hashed into PCRs. The result can be parsed by eventlog.
+// As this library currently doesn't support extending PCRs it just returns the log as supplied by the EFI interface.
+func GetMeasurementLog() ([]byte, error) {
+	return ioutil.ReadFile("/sys/kernel/security/tpm0/binary_bios_measurements")
+}
diff --git a/metropolis/node/kubernetes/BUILD.bazel b/metropolis/node/kubernetes/BUILD.bazel
new file mode 100644
index 0000000..f1fa849
--- /dev/null
+++ b/metropolis/node/kubernetes/BUILD.bazel
@@ -0,0 +1,54 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "apiserver.go",
+        "controller-manager.go",
+        "csi.go",
+        "kubelet.go",
+        "provisioner.go",
+        "scheduler.go",
+        "service.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes",
+    visibility = ["//metropolis/node:__subpackages__"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/node/common/fileargs:go_default_library",
+        "//metropolis/node/common/fsquota:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/localstorage:go_default_library",
+        "//metropolis/node/core/localstorage/declarative:go_default_library",
+        "//metropolis/node/core/logtree:go_default_library",
+        "//metropolis/node/core/network/dns:go_default_library",
+        "//metropolis/node/kubernetes/clusternet:go_default_library",
+        "//metropolis/node/kubernetes/nfproxy:go_default_library",
+        "//metropolis/node/kubernetes/pki:go_default_library",
+        "//metropolis/node/kubernetes/reconciler:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+        "@com_github_container_storage_interface_spec//lib/go/csi:go_default_library",
+        "@io_bazel_rules_go//proto/wkt:wrappers_go_proto",
+        "@io_k8s_api//core/v1:go_default_library",
+        "@io_k8s_api//storage/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/api/errors:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+        "@io_k8s_client_go//informers:go_default_library",
+        "@io_k8s_client_go//informers/core/v1:go_default_library",
+        "@io_k8s_client_go//informers/storage/v1:go_default_library",
+        "@io_k8s_client_go//kubernetes:go_default_library",
+        "@io_k8s_client_go//kubernetes/scheme:go_default_library",
+        "@io_k8s_client_go//kubernetes/typed/core/v1:go_default_library",
+        "@io_k8s_client_go//tools/cache:go_default_library",
+        "@io_k8s_client_go//tools/clientcmd:go_default_library",
+        "@io_k8s_client_go//tools/record:go_default_library",
+        "@io_k8s_client_go//tools/reference:go_default_library",
+        "@io_k8s_client_go//util/workqueue:go_default_library",
+        "@io_k8s_kubelet//config/v1beta1:go_default_library",
+        "@io_k8s_kubelet//pkg/apis/pluginregistration/v1:go_default_library",
+        "@org_golang_google_grpc//:go_default_library",
+        "@org_golang_google_grpc//codes:go_default_library",
+        "@org_golang_google_grpc//status:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/node/kubernetes/apiserver.go b/metropolis/node/kubernetes/apiserver.go
new file mode 100644
index 0000000..583e268
--- /dev/null
+++ b/metropolis/node/kubernetes/apiserver.go
@@ -0,0 +1,137 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kubernetes
+
+import (
+	"context"
+	"encoding/pem"
+	"fmt"
+	"io"
+	"net"
+	"os/exec"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fileargs"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/pki"
+)
+
+type apiserverService struct {
+	KPKI                        *pki.KubernetesPKI
+	AdvertiseAddress            net.IP
+	ServiceIPRange              net.IPNet
+	Output                      io.Writer
+	EphemeralConsensusDirectory *localstorage.EphemeralConsensusDirectory
+
+	// All PKI-related things are in DER
+	idCA                  []byte
+	kubeletClientCert     []byte
+	kubeletClientKey      []byte
+	aggregationCA         []byte
+	aggregationClientCert []byte
+	aggregationClientKey  []byte
+	serviceAccountPrivKey []byte // In PKIX form
+	serverCert            []byte
+	serverKey             []byte
+}
+
+func (s *apiserverService) loadPKI(ctx context.Context) error {
+	for _, el := range []struct {
+		targetCert *[]byte
+		targetKey  *[]byte
+		name       pki.KubeCertificateName
+	}{
+		{&s.idCA, nil, pki.IdCA},
+		{&s.kubeletClientCert, &s.kubeletClientKey, pki.KubeletClient},
+		{&s.aggregationCA, nil, pki.AggregationCA},
+		{&s.aggregationClientCert, &s.aggregationClientKey, pki.FrontProxyClient},
+		{&s.serverCert, &s.serverKey, pki.APIServer},
+	} {
+		cert, key, err := s.KPKI.Certificate(ctx, el.name)
+		if err != nil {
+			return fmt.Errorf("could not load certificate %q from PKI: %w", el.name, err)
+		}
+		if el.targetCert != nil {
+			*el.targetCert = cert
+		}
+		if el.targetKey != nil {
+			*el.targetKey = key
+		}
+	}
+
+	var err error
+	s.serviceAccountPrivKey, err = s.KPKI.ServiceAccountKey(ctx)
+	if err != nil {
+		return fmt.Errorf("could not load serviceaccount privkey: %w", err)
+	}
+	return nil
+}
+
+func (s *apiserverService) Run(ctx context.Context) error {
+	if err := s.loadPKI(ctx); err != nil {
+		return fmt.Errorf("loading PKI data failed: %w", err)
+	}
+	args, err := fileargs.New()
+	if err != nil {
+		panic(err) // If this fails, something is very wrong. Just crash.
+	}
+	defer args.Close()
+
+	cmd := exec.CommandContext(ctx, "/kubernetes/bin/kube", "kube-apiserver",
+		fmt.Sprintf("--advertise-address=%v", s.AdvertiseAddress.String()),
+		"--authorization-mode=Node,RBAC",
+		args.FileOpt("--client-ca-file", "client-ca.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: s.idCA})),
+		"--enable-admission-plugins=NodeRestriction,PodSecurityPolicy",
+		"--enable-aggregator-routing=true",
+		"--insecure-port=0",
+		fmt.Sprintf("--secure-port=%v", common.KubernetesAPIPort),
+		fmt.Sprintf("--etcd-servers=unix:///%s:0", s.EphemeralConsensusDirectory.ClientSocket.FullPath()),
+		args.FileOpt("--kubelet-client-certificate", "kubelet-client-cert.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: s.kubeletClientCert})),
+		args.FileOpt("--kubelet-client-key", "kubelet-client-key.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: s.kubeletClientKey})),
+		"--kubelet-preferred-address-types=InternalIP",
+		args.FileOpt("--proxy-client-cert-file", "aggregation-client-cert.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: s.aggregationClientCert})),
+		args.FileOpt("--proxy-client-key-file", "aggregation-client-key.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: s.aggregationClientKey})),
+		"--requestheader-allowed-names=front-proxy-client",
+		args.FileOpt("--requestheader-client-ca-file", "aggregation-ca.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: s.aggregationCA})),
+		"--requestheader-extra-headers-prefix=X-Remote-Extra-",
+		"--requestheader-group-headers=X-Remote-Group",
+		"--requestheader-username-headers=X-Remote-User",
+		args.FileOpt("--service-account-key-file", "service-account-pubkey.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: s.serviceAccountPrivKey})),
+		fmt.Sprintf("--service-cluster-ip-range=%v", s.ServiceIPRange.String()),
+		args.FileOpt("--tls-cert-file", "server-cert.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: s.serverCert})),
+		args.FileOpt("--tls-private-key-file", "server-key.pem",
+			pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: s.serverKey})),
+	)
+	if args.Error() != nil {
+		return err
+	}
+	cmd.Stdout = s.Output
+	cmd.Stderr = s.Output
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	err = cmd.Run()
+	fmt.Fprintf(s.Output, "apiserver stopped: %v\n", err)
+	return err
+}
diff --git a/metropolis/node/kubernetes/clusternet/BUILD.bazel b/metropolis/node/kubernetes/clusternet/BUILD.bazel
new file mode 100644
index 0000000..9e9cc01
--- /dev/null
+++ b/metropolis/node/kubernetes/clusternet/BUILD.bazel
@@ -0,0 +1,27 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "clusternet.go",
+        "netlink_compat.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/clusternet",
+    visibility = ["//metropolis/node/kubernetes:__subpackages__"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/node/common/jsonpatch:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/localstorage:go_default_library",
+        "//metropolis/node/core/logtree:go_default_library",
+        "@com_github_vishvananda_netlink//:go_default_library",
+        "@com_zx2c4_golang_wireguard_wgctrl//:go_default_library",
+        "@com_zx2c4_golang_wireguard_wgctrl//wgtypes:go_default_library",
+        "@io_k8s_api//core/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/types:go_default_library",
+        "@io_k8s_client_go//informers:go_default_library",
+        "@io_k8s_client_go//kubernetes:go_default_library",
+        "@io_k8s_client_go//tools/cache:go_default_library",
+    ],
+)
diff --git a/metropolis/node/kubernetes/clusternet/clusternet.go b/metropolis/node/kubernetes/clusternet/clusternet.go
new file mode 100644
index 0000000..d8dc7ad
--- /dev/null
+++ b/metropolis/node/kubernetes/clusternet/clusternet.go
@@ -0,0 +1,276 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package clusternet implements a WireGuard-based overlay network for Kubernetes. It relies on controller-manager's
+// IPAM to assign IP ranges to nodes and on Kubernetes' Node objects to distribute the Node IPs and public keys.
+//
+// It sets up a single WireGuard network interface and routes the entire ClusterCIDR into that network interface,
+// relying on WireGuard's AllowedIPs mechanism to look up the correct peer node to send the traffic to. This means
+// that the routing table doesn't change and doesn't have to be separately managed. When clusternet is started
+// it annotates its WireGuard public key onto its node object.
+// For each node object that's created or updated on the K8s apiserver it checks if a public key annotation is set and
+// if yes a peer with that public key, its InternalIP as endpoint and the CIDR for that node as AllowedIPs is created.
+package clusternet
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net"
+	"os"
+
+	"github.com/vishvananda/netlink"
+	"golang.zx2c4.com/wireguard/wgctrl"
+	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
+	corev1 "k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/types"
+	"k8s.io/client-go/informers"
+	"k8s.io/client-go/kubernetes"
+	"k8s.io/client-go/tools/cache"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/jsonpatch"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+)
+
+const (
+	clusterNetDeviceName = "clusternet"
+	publicKeyAnnotation  = "node.smalltown.nexantic.com/wg-pubkey"
+)
+
+type Service struct {
+	NodeName        string
+	Kubernetes      kubernetes.Interface
+	ClusterNet      net.IPNet
+	InformerFactory informers.SharedInformerFactory
+	DataDirectory   *localstorage.DataKubernetesClusterNetworkingDirectory
+
+	wgClient *wgctrl.Client
+	privKey  wgtypes.Key
+	logger   logtree.LeveledLogger
+}
+
+// ensureNode creates/updates the corresponding WireGuard peer entry for the given node objet
+func (s *Service) ensureNode(newNode *corev1.Node) error {
+	if newNode.Name == s.NodeName {
+		// Node doesn't need to connect to itself
+		return nil
+	}
+	pubKeyRaw := newNode.Annotations[publicKeyAnnotation]
+	if pubKeyRaw == "" {
+		return nil
+	}
+	pubKey, err := wgtypes.ParseKey(pubKeyRaw)
+	if err != nil {
+		return fmt.Errorf("failed to parse public-key annotation: %w", err)
+	}
+	var internalIP net.IP
+	for _, addr := range newNode.Status.Addresses {
+		if addr.Type == corev1.NodeInternalIP {
+			if internalIP != nil {
+				s.logger.Warningf("More than one NodeInternalIP specified, using the first one")
+				break
+			}
+			internalIP = net.ParseIP(addr.Address)
+			if internalIP == nil {
+				s.logger.Warningf("Failed to parse Internal IP %s", addr.Address)
+			}
+		}
+	}
+	if internalIP == nil {
+		return errors.New("node has no Internal IP")
+	}
+	var allowedIPs []net.IPNet
+	for _, podNetStr := range newNode.Spec.PodCIDRs {
+		_, podNet, err := net.ParseCIDR(podNetStr)
+		if err != nil {
+			s.logger.Warningf("Node %s PodCIDR failed to parse, ignored: %v", newNode.Name, err)
+			continue
+		}
+		allowedIPs = append(allowedIPs, *podNet)
+	}
+	allowedIPs = append(allowedIPs, net.IPNet{IP: internalIP, Mask: net.CIDRMask(32, 32)})
+	s.logger.V(1).Infof("Adding/Updating WireGuard peer node %s, endpoint %s, allowedIPs %+v", newNode.Name, internalIP.String(), allowedIPs)
+	// WireGuard's kernel side has create/update semantics on peers by default. So we can just add the peer multiple
+	// times to update it.
+	err = s.wgClient.ConfigureDevice(clusterNetDeviceName, wgtypes.Config{
+		Peers: []wgtypes.PeerConfig{{
+			PublicKey:         pubKey,
+			Endpoint:          &net.UDPAddr{Port: common.WireGuardPort, IP: internalIP},
+			ReplaceAllowedIPs: true,
+			AllowedIPs:        allowedIPs,
+		}},
+	})
+	if err != nil {
+		return fmt.Errorf("failed to add WireGuard peer node: %w", err)
+	}
+	return nil
+}
+
+// removeNode removes the corresponding WireGuard peer entry for the given node object
+func (s *Service) removeNode(oldNode *corev1.Node) error {
+	if oldNode.Name == s.NodeName {
+		// Node doesn't need to connect to itself
+		return nil
+	}
+	pubKeyRaw := oldNode.Annotations[publicKeyAnnotation]
+	if pubKeyRaw == "" {
+		return nil
+	}
+	pubKey, err := wgtypes.ParseKey(pubKeyRaw)
+	if err != nil {
+		return fmt.Errorf("node public-key annotation not decodable: %w", err)
+	}
+	err = s.wgClient.ConfigureDevice(clusterNetDeviceName, wgtypes.Config{
+		Peers: []wgtypes.PeerConfig{{
+			PublicKey: pubKey,
+			Remove:    true,
+		}},
+	})
+	if err != nil {
+		return fmt.Errorf("failed to remove WireGuard peer node: %w", err)
+	}
+	return nil
+}
+
+// ensureOnDiskKey loads the private key from disk or (if none exists) generates one and persists it.
+func (s *Service) ensureOnDiskKey() error {
+	keyRaw, err := s.DataDirectory.Key.Read()
+	if os.IsNotExist(err) {
+		key, err := wgtypes.GeneratePrivateKey()
+		if err != nil {
+			return fmt.Errorf("failed to generate private key: %w", err)
+		}
+		if err := s.DataDirectory.Key.Write([]byte(key.String()), 0600); err != nil {
+			return fmt.Errorf("failed to store newly generated key: %w", err)
+		}
+
+		s.privKey = key
+		return nil
+	} else if err != nil {
+		return fmt.Errorf("failed to load on-disk key: %w", err)
+	}
+
+	key, err := wgtypes.ParseKey(string(keyRaw))
+	if err != nil {
+		return fmt.Errorf("invalid private key in file: %w", err)
+	}
+	s.privKey = key
+	return nil
+}
+
+// annotateThisNode annotates the node (as defined by NodeName) with the wireguard public key of this node.
+func (s *Service) annotateThisNode(ctx context.Context) error {
+	patch := []jsonpatch.JsonPatchOp{{
+		Operation: "add",
+		Path:      "/metadata/annotations/" + jsonpatch.EncodeJSONRefToken(publicKeyAnnotation),
+		Value:     s.privKey.PublicKey().String(),
+	}}
+
+	patchRaw, err := json.Marshal(patch)
+	if err != nil {
+		return fmt.Errorf("failed to encode JSONPatch: %w", err)
+	}
+
+	if _, err := s.Kubernetes.CoreV1().Nodes().Patch(ctx, s.NodeName, types.JSONPatchType, patchRaw, metav1.PatchOptions{}); err != nil {
+		return fmt.Errorf("failed to patch resource: %w", err)
+	}
+
+	return nil
+}
+
+// Run runs the ClusterNet service. See package description for what it does.
+func (s *Service) Run(ctx context.Context) error {
+	logger := supervisor.Logger(ctx)
+	s.logger = logger
+
+	wgClient, err := wgctrl.New()
+	if err != nil {
+		return fmt.Errorf("failed to connect to netlink's WireGuard config endpoint: %w", err)
+	}
+	s.wgClient = wgClient
+
+	if err := s.ensureOnDiskKey(); err != nil {
+		return fmt.Errorf("failed to ensure on-disk key: %w", err)
+	}
+
+	wgInterface := &Wireguard{LinkAttrs: netlink.LinkAttrs{Name: clusterNetDeviceName, Flags: net.FlagUp}}
+	if err := netlink.LinkAdd(wgInterface); err != nil {
+		return fmt.Errorf("failed to add WireGuard network interfacee: %w", err)
+	}
+	defer netlink.LinkDel(wgInterface)
+
+	listenPort := common.WireGuardPort
+	if err := wgClient.ConfigureDevice(clusterNetDeviceName, wgtypes.Config{
+		PrivateKey: &s.privKey,
+		ListenPort: &listenPort,
+	}); err != nil {
+		return fmt.Errorf("failed to set up WireGuard interface: %w", err)
+	}
+
+	if err := netlink.RouteAdd(&netlink.Route{
+		Dst:       &s.ClusterNet,
+		LinkIndex: wgInterface.Index,
+	}); err != nil && !os.IsExist(err) {
+		return fmt.Errorf("failed to add cluster net route to Wireguard interface: %w", err)
+	}
+
+	if err := s.annotateThisNode(ctx); err != nil {
+		return fmt.Errorf("when annotating this node with public key: %w", err)
+	}
+
+	nodeInformer := s.InformerFactory.Core().V1().Nodes()
+	nodeInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
+		AddFunc: func(new interface{}) {
+			newNode, ok := new.(*corev1.Node)
+			if !ok {
+				logger.Errorf("Received non-node item %+v in node event handler", new)
+				return
+			}
+			if err := s.ensureNode(newNode); err != nil {
+				logger.Warningf("Failed to sync node: %v", err)
+			}
+		},
+		UpdateFunc: func(old, new interface{}) {
+			newNode, ok := new.(*corev1.Node)
+			if !ok {
+				logger.Errorf("Received non-node item %+v in node event handler", new)
+				return
+			}
+			if err := s.ensureNode(newNode); err != nil {
+				logger.Warningf("Failed to sync node: %v", err)
+			}
+		},
+		DeleteFunc: func(old interface{}) {
+			oldNode, ok := old.(*corev1.Node)
+			if !ok {
+				logger.Errorf("Received non-node item %+v in node event handler", oldNode)
+				return
+			}
+			if err := s.removeNode(oldNode); err != nil {
+				logger.Warningf("Failed to sync node: %v", err)
+			}
+		},
+	})
+
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	nodeInformer.Informer().Run(ctx.Done())
+	return ctx.Err()
+}
diff --git a/metropolis/node/kubernetes/clusternet/netlink_compat.go b/metropolis/node/kubernetes/clusternet/netlink_compat.go
new file mode 100644
index 0000000..a90cc47
--- /dev/null
+++ b/metropolis/node/kubernetes/clusternet/netlink_compat.go
@@ -0,0 +1,33 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Can be removed once https://github.com/vishvananda/netlink/pull/464 lands
+package clusternet
+
+import "github.com/vishvananda/netlink"
+
+// Wireguard represent links of type "wireguard", see https://www.wireguard.com/
+type Wireguard struct {
+	netlink.LinkAttrs
+}
+
+func (wg *Wireguard) Attrs() *netlink.LinkAttrs {
+	return &wg.LinkAttrs
+}
+
+func (wg *Wireguard) Type() string {
+	return "wireguard"
+}
diff --git a/metropolis/node/kubernetes/containerd/BUILD.bazel b/metropolis/node/kubernetes/containerd/BUILD.bazel
new file mode 100644
index 0000000..9e42595
--- /dev/null
+++ b/metropolis/node/kubernetes/containerd/BUILD.bazel
@@ -0,0 +1,20 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/containerd",
+    visibility = ["//metropolis/node/core:__subpackages__"],
+    deps = [
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/localstorage:go_default_library",
+        "@com_github_containerd_containerd//:go_default_library",
+        "@com_github_containerd_containerd//namespaces:go_default_library",
+    ],
+)
+
+exports_files([
+    "config.toml",
+    "runsc.toml",
+    "cnispec.gojson",
+])
diff --git a/metropolis/node/kubernetes/containerd/cnispec.gojson b/metropolis/node/kubernetes/containerd/cnispec.gojson
new file mode 100644
index 0000000..0057036
--- /dev/null
+++ b/metropolis/node/kubernetes/containerd/cnispec.gojson
@@ -0,0 +1,29 @@
+{{- /*gotype: github.com/containerd/cri/pkg/server.cniConfigTemplate*/ -}}
+{
+    "name": "k8s-pod-network",
+    "cniVersion": "0.3.1",
+    "plugins": [
+        {
+            "type": "ptp",
+            "mtu": 1420,
+            "ipam": {
+                "type": "host-local",
+                "dataDir": "/containerd/run/ipam",
+                "ranges": [
+                    {{range $i, $range := .PodCIDRRanges}}{{if $i}},
+            {{end}}[
+            {
+                "subnet": "{{$range}}"
+            }
+        ]
+        {{end}}
+    ],
+    "routes": [
+        {{range $i, $route := .Routes}}{{if $i}},
+    {{end}}{
+    "dst": "{{$route}}"
+}{{end}}]
+}
+}
+]
+}
\ No newline at end of file
diff --git a/metropolis/node/kubernetes/containerd/config.toml b/metropolis/node/kubernetes/containerd/config.toml
new file mode 100644
index 0000000..f8c7fb1
--- /dev/null
+++ b/metropolis/node/kubernetes/containerd/config.toml
@@ -0,0 +1,125 @@
+version = 2
+root = "/data/containerd"
+state = "/ephemeral/containerd"
+plugin_dir = ""
+disabled_plugins = []
+required_plugins = []
+oom_score = 0
+
+[grpc]
+  address = "/ephemeral/containerd/client.sock"
+  tcp_address = ""
+  tcp_tls_cert = ""
+  tcp_tls_key = ""
+  uid = 0
+  gid = 0
+  max_recv_message_size = 16777216
+  max_send_message_size = 16777216
+
+[ttrpc]
+  address = ""
+  uid = 0
+  gid = 0
+
+[debug]
+  address = ""
+  uid = 0
+  gid = 0
+  level = ""
+
+[metrics]
+  address = ""
+  grpc_histogram = false
+
+[cgroup]
+  path = ""
+
+[timeouts]
+  "io.containerd.timeout.shim.cleanup" = "5s"
+  "io.containerd.timeout.shim.load" = "5s"
+  "io.containerd.timeout.shim.shutdown" = "3s"
+  "io.containerd.timeout.task.state" = "2s"
+
+[plugins]
+  [plugins."io.containerd.gc.v1.scheduler"]
+    pause_threshold = 0.02
+    deletion_threshold = 0
+    mutation_threshold = 100
+    schedule_delay = "0s"
+    startup_delay = "100ms"
+  [plugins."io.containerd.grpc.v1.cri"]
+    disable_tcp_service = true
+    stream_server_address = "127.0.0.1"
+    stream_server_port = "0"
+    stream_idle_timeout = "4h0m0s"
+    enable_selinux = false
+    sandbox_image = "k8s.gcr.io/pause:3.1"
+    stats_collect_period = 10
+    systemd_cgroup = false
+    enable_tls_streaming = false
+    ignore_image_defined_volumes = true
+    max_container_log_line_size = 16384
+    disable_cgroup = false
+    disable_apparmor = true
+    restrict_oom_score_adj = false
+    max_concurrent_downloads = 3
+    disable_proc_mount = false
+    [plugins."io.containerd.grpc.v1.cri".containerd]
+      snapshotter = "overlayfs"
+      default_runtime_name = "runsc"
+      no_pivot = false
+      [plugins."io.containerd.grpc.v1.cri".containerd.default_runtime]
+        runtime_type = ""
+        runtime_engine = ""
+        runtime_root = ""
+        privileged_without_host_devices = false
+      [plugins."io.containerd.grpc.v1.cri".containerd.untrusted_workload_runtime]
+        runtime_type = ""
+        runtime_engine = ""
+        runtime_root = ""
+        privileged_without_host_devices = false
+      [plugins."io.containerd.grpc.v1.cri".containerd.runtimes]
+        [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runsc]
+          runtime_type = "io.containerd.runsc.v1"
+          runtime_engine = ""
+          runtime_root = ""
+          privileged_without_host_devices = false
+          [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runsc.options]
+            TypeUrl = "io.containerd.runsc.v1.options"
+            ConfigPath = "/containerd/conf/runsc.toml"
+        [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runc]
+          runtime_type = "io.containerd.runc.v2"
+          runtime_engine = ""
+          runtime_root = ""
+          privileged_without_host_devices = false
+          base_runtime_spec = ""
+    [plugins."io.containerd.grpc.v1.cri".cni]
+      bin_dir = "/containerd/bin/cni"
+      conf_dir = "/containerd/conf/cni"
+      max_conf_num = 0
+      conf_template = "/containerd/conf/cnispec.gojson"
+    [plugins."io.containerd.grpc.v1.cri".registry]
+      [plugins."io.containerd.grpc.v1.cri".registry.mirrors]
+        [plugins."io.containerd.grpc.v1.cri".registry.mirrors."docker.io"]
+          endpoint = ["https://registry-1.docker.io"]
+    [plugins."io.containerd.grpc.v1.cri".x509_key_pair_streaming]
+      tls_cert_file = ""
+      tls_key_file = ""
+  [plugins."io.containerd.internal.v1.opt"]
+    path = "/containerd/bin"
+  [plugins."io.containerd.internal.v1.restart"]
+    interval = "10s"
+  [plugins."io.containerd.metadata.v1.bolt"]
+    content_sharing_policy = "shared"
+  [plugins."io.containerd.monitor.v1.cgroups"]
+    no_prometheus = false
+  [plugins."io.containerd.runtime.v1.linux"]
+    shim = "containerd-shim"
+    runtime = "noop"
+    runtime_root = ""
+    no_shim = false
+    shim_debug = false
+  [plugins."io.containerd.runtime.v2.task"]
+    platforms = ["linux/amd64"]
+  [plugins."io.containerd.service.v1.diff-service"]
+    default = ["walking"]
\ No newline at end of file
diff --git a/metropolis/node/kubernetes/containerd/main.go b/metropolis/node/kubernetes/containerd/main.go
new file mode 100644
index 0000000..366f902
--- /dev/null
+++ b/metropolis/node/kubernetes/containerd/main.go
@@ -0,0 +1,146 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package containerd
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"time"
+
+	ctr "github.com/containerd/containerd"
+	"github.com/containerd/containerd/namespaces"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+)
+
+const (
+	preseedNamespacesDir = "/containerd/preseed/"
+)
+
+type Service struct {
+	EphemeralVolume *localstorage.EphemeralContainerdDirectory
+}
+
+func (s *Service) Run(ctx context.Context) error {
+	cmd := exec.CommandContext(ctx, "/containerd/bin/containerd", "--config", "/containerd/conf/config.toml")
+	cmd.Env = []string{"PATH=/containerd/bin", "TMPDIR=" + s.EphemeralVolume.Tmp.FullPath()}
+
+	runscFifo, err := os.OpenFile(s.EphemeralVolume.RunSCLogsFIFO.FullPath(), os.O_CREATE|os.O_RDONLY, os.ModeNamedPipe|0777)
+	if err != nil {
+		return err
+	}
+
+	if err := supervisor.Run(ctx, "runsc", s.logPump(runscFifo)); err != nil {
+		return fmt.Errorf("failed to start runsc log pump: %w", err)
+	}
+
+	if err := supervisor.Run(ctx, "preseed", s.runPreseed); err != nil {
+		return fmt.Errorf("failed to start preseed runnable: %w", err)
+	}
+	return supervisor.RunCommand(ctx, cmd)
+}
+
+// logPump returns a runnable that pipes data from a file/FIFO into its raw logger.
+// TODO(q3k): refactor this out to a generic function in supervisor or logtree.
+func (s *Service) logPump(fifo *os.File) supervisor.Runnable {
+	return func(ctx context.Context) error {
+		supervisor.Signal(ctx, supervisor.SignalHealthy)
+		for {
+			// Quit if requested.
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			default:
+			}
+
+			n, err := io.Copy(supervisor.RawLogger(ctx), fifo)
+			if n == 0 && err == nil {
+				// Hack because pipes/FIFOs can return zero reads when nobody is writing. To avoid busy-looping,
+				// sleep a bit before retrying. This does not loose data since the FIFO internal buffer will
+				// stall writes when it becomes full. 10ms maximum stall in a non-latency critical process (reading
+				// debug logs) is not an issue for us.
+				time.Sleep(10 * time.Millisecond)
+			} else if err != nil {
+				return fmt.Errorf("log pump failed: %v", err)
+			}
+		}
+	}
+}
+
+// runPreseed loads OCI bundles in tar form from preseedNamespacesDir into containerd at startup.
+// This can be run multiple times, containerd will automatically dedup the layers.
+// containerd uses namespaces to keep images (and everything else) separate so to define where the images will be loaded
+// to they need to be in a folder named after the namespace they should be loaded into.
+// containerd's CRI plugin (which is built as part of containerd) uses a hardcoded namespace ("k8s.io") for everything
+// accessed through CRI, so if an image should be available on K8s it needs to be in that namespace.
+// As an example if image helloworld should be loaded for use with Kubernetes, the OCI bundle needs to be at
+// <preseedNamespacesDir>/k8s.io/helloworld.tar. No tagging beyond what's in the bundle is performed.
+func (s *Service) runPreseed(ctx context.Context) error {
+	client, err := ctr.New(s.EphemeralVolume.ClientSocket.FullPath())
+	if err != nil {
+		return fmt.Errorf("failed to connect to containerd: %w", err)
+	}
+	logger := supervisor.Logger(ctx)
+	preseedNamespaceDirs, err := ioutil.ReadDir(preseedNamespacesDir)
+	if err != nil {
+		return fmt.Errorf("failed to open preseed dir: %w", err)
+	}
+	for _, dir := range preseedNamespaceDirs {
+		if !dir.IsDir() {
+			logger.Warningf("Non-Directory %q found in preseed folder, ignoring", dir.Name())
+			continue
+		}
+		namespace := dir.Name()
+		images, err := ioutil.ReadDir(filepath.Join(preseedNamespacesDir, namespace))
+		if err != nil {
+			return fmt.Errorf("failed to list namespace preseed directory for ns \"%v\": %w", namespace, err)
+		}
+		ctxWithNS := namespaces.WithNamespace(ctx, namespace)
+		for _, image := range images {
+			if image.IsDir() {
+				logger.Warningf("Directory %q found in preseed namespaced folder, ignoring", image.Name())
+				continue
+			}
+			imageFile, err := os.Open(filepath.Join(preseedNamespacesDir, namespace, image.Name()))
+			if err != nil {
+				return fmt.Errorf("failed to open preseed image \"%v\": %w", image.Name(), err)
+			}
+			// defer in this loop is fine since we're never going to preseed more than ~1M images which is where our
+			// file descriptor limit is.
+			defer imageFile.Close()
+			importedImages, err := client.Import(ctxWithNS, imageFile)
+			if err != nil {
+				return fmt.Errorf("failed to import preseed image: %w", err)
+			}
+			var importedImageNames []string
+			for _, img := range importedImages {
+				importedImageNames = append(importedImageNames, img.Name)
+			}
+			logger.Infof("Successfully imported preseeded bundle %s/%s into containerd", namespace, strings.Join(importedImageNames, ","))
+		}
+	}
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	supervisor.Signal(ctx, supervisor.SignalDone)
+	return nil
+}
diff --git a/metropolis/node/kubernetes/containerd/runsc.toml b/metropolis/node/kubernetes/containerd/runsc.toml
new file mode 100644
index 0000000..4fe0751
--- /dev/null
+++ b/metropolis/node/kubernetes/containerd/runsc.toml
@@ -0,0 +1,6 @@
+root = "/ephemeral/containerd/runsc"
+[runsc_config]
+debug = "true"
+debug-log = "/ephemeral/containerd/runsc-logs.fifo"
+panic-log = "/ephemeral/containerd/runsc-logs.fifo"
+log = "/ephemeral/containerd/runsc-logs.fifo"
diff --git a/metropolis/node/kubernetes/controller-manager.go b/metropolis/node/kubernetes/controller-manager.go
new file mode 100644
index 0000000..487511f
--- /dev/null
+++ b/metropolis/node/kubernetes/controller-manager.go
@@ -0,0 +1,93 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kubernetes
+
+import (
+	"context"
+	"encoding/pem"
+	"fmt"
+	"net"
+	"os/exec"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fileargs"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/pki"
+)
+
+type controllerManagerConfig struct {
+	clusterNet net.IPNet
+	// All PKI-related things are in DER
+	kubeConfig            []byte
+	rootCA                []byte
+	serviceAccountPrivKey []byte // In PKCS#8 form
+	serverCert            []byte
+	serverKey             []byte
+}
+
+func getPKIControllerManagerConfig(ctx context.Context, kpki *pki.KubernetesPKI) (*controllerManagerConfig, error) {
+	var config controllerManagerConfig
+	var err error
+	config.rootCA, _, err = kpki.Certificate(ctx, pki.IdCA)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get ID root CA: %w", err)
+	}
+	config.serverCert, config.serverKey, err = kpki.Certificate(ctx, pki.ControllerManager)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get controller-manager serving certificate: %w", err)
+	}
+	config.serviceAccountPrivKey, err = kpki.ServiceAccountKey(ctx)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get serviceaccount privkey: %w", err)
+	}
+	config.kubeConfig, err = kpki.Kubeconfig(ctx, pki.ControllerManagerClient)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get controller-manager kubeconfig: %w", err)
+	}
+	return &config, nil
+}
+
+func runControllerManager(config controllerManagerConfig) supervisor.Runnable {
+	return func(ctx context.Context) error {
+		args, err := fileargs.New()
+		if err != nil {
+			panic(err) // If this fails, something is very wrong. Just crash.
+		}
+		defer args.Close()
+
+		cmd := exec.CommandContext(ctx, "/kubernetes/bin/kube", "kube-controller-manager",
+			args.FileOpt("--kubeconfig", "kubeconfig", config.kubeConfig),
+			args.FileOpt("--service-account-private-key-file", "service-account-privkey.pem",
+				pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: config.serviceAccountPrivKey})),
+			args.FileOpt("--root-ca-file", "root-ca.pem",
+				pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: config.rootCA})),
+			"--port=0",                               // Kill insecure serving
+			"--use-service-account-credentials=true", // Enables things like PSP enforcement
+			fmt.Sprintf("--cluster-cidr=%v", config.clusterNet.String()),
+			args.FileOpt("--tls-cert-file", "server-cert.pem",
+				pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: config.serverCert})),
+			args.FileOpt("--tls-private-key-file", "server-key.pem",
+				pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: config.serverKey})),
+			"--allocate-node-cidrs",
+			"--cluster-cidr="+config.clusterNet.String(),
+		)
+
+		if args.Error() != nil {
+			return fmt.Errorf("failed to use fileargs: %w", err)
+		}
+		return supervisor.RunCommand(ctx, cmd)
+	}
+}
diff --git a/metropolis/node/kubernetes/csi.go b/metropolis/node/kubernetes/csi.go
new file mode 100644
index 0000000..4b44a1a
--- /dev/null
+++ b/metropolis/node/kubernetes/csi.go
@@ -0,0 +1,246 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kubernetes
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"os"
+	"path/filepath"
+	"regexp"
+
+	"github.com/container-storage-interface/spec/lib/go/csi"
+	"github.com/golang/protobuf/ptypes/wrappers"
+	"golang.org/x/sys/unix"
+	"google.golang.org/grpc"
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/status"
+	pluginregistration "k8s.io/kubelet/pkg/apis/pluginregistration/v1"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fsquota"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+)
+
+// Derived from K8s spec for acceptable names, but shortened to 130 characters to avoid issues with
+// maximum path length. We don't provision longer names so this applies only if you manually create
+// a volume with a name of more than 130 characters.
+var acceptableNames = regexp.MustCompile("^[a-z][a-bz0-9-.]{0,128}[a-z0-9]$")
+
+const volumeDir = "volumes"
+
+type csiPluginServer struct {
+	KubeletDirectory *localstorage.DataKubernetesKubeletDirectory
+	VolumesDirectory *localstorage.DataVolumesDirectory
+
+	logger logtree.LeveledLogger
+}
+
+func (s *csiPluginServer) Run(ctx context.Context) error {
+	s.logger = supervisor.Logger(ctx)
+
+	pluginListener, err := net.ListenUnix("unix", &net.UnixAddr{Name: s.KubeletDirectory.Plugins.VFS.FullPath(), Net: "unix"})
+	if err != nil {
+		return fmt.Errorf("failed to listen on CSI socket: %w", err)
+	}
+	pluginListener.SetUnlinkOnClose(true)
+
+	pluginServer := grpc.NewServer()
+	csi.RegisterIdentityServer(pluginServer, s)
+	csi.RegisterNodeServer(pluginServer, s)
+	// Enable graceful shutdown since we don't have long-running RPCs and most of them shouldn't and can't be
+	// cancelled anyways.
+	if err := supervisor.Run(ctx, "csi-node", supervisor.GRPCServer(pluginServer, pluginListener, true)); err != nil {
+		return err
+	}
+
+	registrationListener, err := net.ListenUnix("unix", &net.UnixAddr{Name: s.KubeletDirectory.PluginsRegistry.VFSReg.FullPath(), Net: "unix"})
+	if err != nil {
+		return fmt.Errorf("failed to listen on CSI registration socket: %w", err)
+	}
+	registrationListener.SetUnlinkOnClose(true)
+
+	registrationServer := grpc.NewServer()
+	pluginregistration.RegisterRegistrationServer(registrationServer, s)
+	if err := supervisor.Run(ctx, "registration", supervisor.GRPCServer(registrationServer, registrationListener, true)); err != nil {
+		return err
+	}
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	supervisor.Signal(ctx, supervisor.SignalDone)
+	return nil
+}
+
+func (*csiPluginServer) NodeStageVolume(ctx context.Context, req *csi.NodeStageVolumeRequest) (*csi.NodeStageVolumeResponse, error) {
+	return nil, status.Errorf(codes.Unimplemented, "method NodeStageVolume not supported")
+}
+
+func (*csiPluginServer) NodeUnstageVolume(ctx context.Context, req *csi.NodeUnstageVolumeRequest) (*csi.NodeUnstageVolumeResponse, error) {
+	return nil, status.Errorf(codes.Unimplemented, "method NodeUnstageVolume not supported")
+}
+
+func (s *csiPluginServer) NodePublishVolume(ctx context.Context, req *csi.NodePublishVolumeRequest) (*csi.NodePublishVolumeResponse, error) {
+	if !acceptableNames.MatchString(req.VolumeId) {
+		return nil, status.Error(codes.InvalidArgument, "invalid characters in volume id")
+	}
+
+	// TODO(q3k): move this logic to localstorage?
+	volumePath := filepath.Join(s.VolumesDirectory.FullPath(), req.VolumeId)
+
+	switch req.VolumeCapability.AccessMode.Mode {
+	case csi.VolumeCapability_AccessMode_SINGLE_NODE_WRITER:
+	case csi.VolumeCapability_AccessMode_SINGLE_NODE_READER_ONLY:
+	default:
+		return nil, status.Error(codes.InvalidArgument, "unsupported access mode")
+	}
+	switch req.VolumeCapability.AccessType.(type) {
+	case *csi.VolumeCapability_Mount:
+	default:
+		return nil, status.Error(codes.InvalidArgument, "unsupported access type")
+	}
+
+	err := unix.Mount(volumePath, req.TargetPath, "", unix.MS_BIND, "")
+	switch {
+	case err == unix.ENOENT:
+		return nil, status.Error(codes.NotFound, "volume not found")
+	case err != nil:
+		return nil, status.Errorf(codes.Unavailable, "failed to bind-mount volume: %v", err)
+	}
+
+	if req.Readonly {
+		err := unix.Mount(volumePath, req.TargetPath, "", unix.MS_BIND|unix.MS_REMOUNT|unix.MS_RDONLY, "")
+		if err != nil {
+			_ = unix.Unmount(req.TargetPath, 0) // Best-effort
+			return nil, status.Errorf(codes.Unavailable, "failed to remount volume: %v", err)
+		}
+	}
+	return &csi.NodePublishVolumeResponse{}, nil
+}
+
+func (*csiPluginServer) NodeUnpublishVolume(ctx context.Context, req *csi.NodeUnpublishVolumeRequest) (*csi.NodeUnpublishVolumeResponse, error) {
+	if err := unix.Unmount(req.TargetPath, 0); err != nil {
+		return nil, status.Errorf(codes.Unavailable, "failed to unmount volume: %v", err)
+	}
+	return &csi.NodeUnpublishVolumeResponse{}, nil
+}
+
+func (*csiPluginServer) NodeGetVolumeStats(ctx context.Context, req *csi.NodeGetVolumeStatsRequest) (*csi.NodeGetVolumeStatsResponse, error) {
+	quota, err := fsquota.GetQuota(req.VolumePath)
+	if os.IsNotExist(err) {
+		return nil, status.Error(codes.NotFound, "volume does not exist at this path")
+	} else if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "failed to get quota: %v", err)
+	}
+
+	return &csi.NodeGetVolumeStatsResponse{
+		Usage: []*csi.VolumeUsage{
+			{
+				Total:     int64(quota.Bytes),
+				Unit:      csi.VolumeUsage_BYTES,
+				Used:      int64(quota.BytesUsed),
+				Available: int64(quota.Bytes - quota.BytesUsed),
+			},
+			{
+				Total:     int64(quota.Inodes),
+				Unit:      csi.VolumeUsage_INODES,
+				Used:      int64(quota.InodesUsed),
+				Available: int64(quota.Inodes - quota.InodesUsed),
+			},
+		},
+	}, nil
+}
+
+func (*csiPluginServer) NodeExpandVolume(ctx context.Context, req *csi.NodeExpandVolumeRequest) (*csi.NodeExpandVolumeResponse, error) {
+	if req.CapacityRange.LimitBytes <= 0 {
+		return nil, status.Error(codes.InvalidArgument, "invalid expanded volume size: at or below zero bytes")
+	}
+	if err := fsquota.SetQuota(req.VolumePath, uint64(req.CapacityRange.LimitBytes), 0); err != nil {
+		return nil, status.Errorf(codes.Unavailable, "failed to update quota: %v", err)
+	}
+	return &csi.NodeExpandVolumeResponse{CapacityBytes: req.CapacityRange.LimitBytes}, nil
+}
+
+func rpcCapability(cap csi.NodeServiceCapability_RPC_Type) *csi.NodeServiceCapability {
+	return &csi.NodeServiceCapability{
+		Type: &csi.NodeServiceCapability_Rpc{
+			Rpc: &csi.NodeServiceCapability_RPC{Type: cap},
+		},
+	}
+}
+
+func (*csiPluginServer) NodeGetCapabilities(ctx context.Context, req *csi.NodeGetCapabilitiesRequest) (*csi.NodeGetCapabilitiesResponse, error) {
+	return &csi.NodeGetCapabilitiesResponse{
+		Capabilities: []*csi.NodeServiceCapability{
+			rpcCapability(csi.NodeServiceCapability_RPC_EXPAND_VOLUME),
+			rpcCapability(csi.NodeServiceCapability_RPC_GET_VOLUME_STATS),
+		},
+	}, nil
+}
+
+func (*csiPluginServer) NodeGetInfo(ctx context.Context, req *csi.NodeGetInfoRequest) (*csi.NodeGetInfoResponse, error) {
+	hostname, err := os.Hostname()
+	if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "failed to get node identity: %v", err)
+	}
+	return &csi.NodeGetInfoResponse{
+		NodeId: hostname,
+	}, nil
+}
+
+// CSI Identity endpoints
+func (*csiPluginServer) GetPluginInfo(ctx context.Context, req *csi.GetPluginInfoRequest) (*csi.GetPluginInfoResponse, error) {
+	return &csi.GetPluginInfoResponse{
+		Name:          "com.smalltown.vfs",
+		VendorVersion: "0.0.1", // TODO(lorenz): Maybe stamp?
+	}, nil
+}
+
+func (*csiPluginServer) GetPluginCapabilities(ctx context.Context, req *csi.GetPluginCapabilitiesRequest) (*csi.GetPluginCapabilitiesResponse, error) {
+	return &csi.GetPluginCapabilitiesResponse{
+		Capabilities: []*csi.PluginCapability{
+			{
+				Type: &csi.PluginCapability_VolumeExpansion_{
+					VolumeExpansion: &csi.PluginCapability_VolumeExpansion{
+						Type: csi.PluginCapability_VolumeExpansion_ONLINE,
+					},
+				},
+			},
+		},
+	}, nil
+}
+
+func (s *csiPluginServer) Probe(ctx context.Context, req *csi.ProbeRequest) (*csi.ProbeResponse, error) {
+	return &csi.ProbeResponse{Ready: &wrappers.BoolValue{Value: true}}, nil
+}
+
+// Registration endpoints
+func (s *csiPluginServer) GetInfo(ctx context.Context, req *pluginregistration.InfoRequest) (*pluginregistration.PluginInfo, error) {
+	return &pluginregistration.PluginInfo{
+		Type:              "CSIPlugin",
+		Name:              "com.smalltown.vfs",
+		Endpoint:          s.KubeletDirectory.Plugins.VFS.FullPath(),
+		SupportedVersions: []string{"1.2"}, // Keep in sync with container-storage-interface/spec package version
+	}, nil
+}
+
+func (s *csiPluginServer) NotifyRegistrationStatus(ctx context.Context, req *pluginregistration.RegistrationStatus) (*pluginregistration.RegistrationStatusResponse, error) {
+	if req.Error != "" {
+		s.logger.Warningf("Kubelet failed registering CSI plugin: %v", req.Error)
+	}
+	return &pluginregistration.RegistrationStatusResponse{}, nil
+}
diff --git a/metropolis/node/kubernetes/hyperkube/BUILD b/metropolis/node/kubernetes/hyperkube/BUILD
new file mode 100644
index 0000000..dced1c7
--- /dev/null
+++ b/metropolis/node/kubernetes/hyperkube/BUILD
@@ -0,0 +1,29 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+load("@//third_party/go:kubernetes_version_def.bzl", "version_x_defs")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/hyperkube",
+    visibility = ["//visibility:private"],
+    deps = [
+        "@com_github_spf13_cobra//:go_default_library",
+        "@com_github_spf13_pflag//:go_default_library",
+        "@io_k8s_component_base//cli/flag:go_default_library",
+        "@io_k8s_component_base//logs:go_default_library",
+        "@io_k8s_component_base//metrics/prometheus/restclient:go_default_library",
+        "@io_k8s_component_base//metrics/prometheus/version:go_default_library",
+        "@io_k8s_kubernetes//cmd/kube-apiserver/app:go_default_library",
+        "@io_k8s_kubernetes//cmd/kube-controller-manager/app:go_default_library",
+        "@io_k8s_kubernetes//cmd/kube-scheduler/app:go_default_library",
+        "@io_k8s_kubernetes//cmd/kubelet/app:go_default_library",
+    ],
+)
+
+go_binary(
+    name = "hyperkube",
+    embed = [":go_default_library"],
+    pure = "on",
+    visibility = ["//visibility:public"],
+    x_defs = version_x_defs(),
+)
diff --git a/metropolis/node/kubernetes/hyperkube/main.go b/metropolis/node/kubernetes/hyperkube/main.go
new file mode 100644
index 0000000..3b4ac08
--- /dev/null
+++ b/metropolis/node/kubernetes/hyperkube/main.go
@@ -0,0 +1,122 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+/*
+Copyright 2014 The Kubernetes Authors.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+// Adapted from https://github.com/dims/hyperkube
+
+package main
+
+import (
+	goflag "flag"
+	"math/rand"
+	"os"
+	"path/filepath"
+	"time"
+
+	"github.com/spf13/cobra"
+	"github.com/spf13/pflag"
+
+	cliflag "k8s.io/component-base/cli/flag"
+	"k8s.io/component-base/logs"
+	_ "k8s.io/component-base/metrics/prometheus/restclient" // for client metric registration
+	_ "k8s.io/component-base/metrics/prometheus/version"    // for version metric registration
+	kubeapiserver "k8s.io/kubernetes/cmd/kube-apiserver/app"
+	kubecontrollermanager "k8s.io/kubernetes/cmd/kube-controller-manager/app"
+	kubescheduler "k8s.io/kubernetes/cmd/kube-scheduler/app"
+	kubelet "k8s.io/kubernetes/cmd/kubelet/app"
+)
+
+func main() {
+	rand.Seed(time.Now().UnixNano())
+
+	hyperkubeCommand, allCommandFns := NewHyperKubeCommand()
+
+	// TODO: once we switch everything over to Cobra commands, we can go back to calling
+	// cliflag.InitFlags() (by removing its pflag.Parse() call). For now, we have to set the
+	// normalize func and add the go flag set by hand.
+	pflag.CommandLine.SetNormalizeFunc(cliflag.WordSepNormalizeFunc)
+	pflag.CommandLine.AddGoFlagSet(goflag.CommandLine)
+	// cliflag.InitFlags()
+	logs.InitLogs()
+	defer logs.FlushLogs()
+
+	basename := filepath.Base(os.Args[0])
+	if err := commandFor(basename, hyperkubeCommand, allCommandFns).Execute(); err != nil {
+		os.Exit(1)
+	}
+}
+
+func commandFor(basename string, defaultCommand *cobra.Command, commands []func() *cobra.Command) *cobra.Command {
+	for _, commandFn := range commands {
+		command := commandFn()
+		if command.Name() == basename {
+			return command
+		}
+		for _, alias := range command.Aliases {
+			if alias == basename {
+				return command
+			}
+		}
+	}
+
+	return defaultCommand
+}
+
+// NewHyperKubeCommand is the entry point for hyperkube
+func NewHyperKubeCommand() (*cobra.Command, []func() *cobra.Command) {
+	// these have to be functions since the command is polymorphic. Cobra wants you to be top level
+	// command to get executed
+	apiserver := func() *cobra.Command { return kubeapiserver.NewAPIServerCommand() }
+	controller := func() *cobra.Command { return kubecontrollermanager.NewControllerManagerCommand() }
+	scheduler := func() *cobra.Command { return kubescheduler.NewSchedulerCommand() }
+	kubelet := func() *cobra.Command { return kubelet.NewKubeletCommand() }
+
+	commandFns := []func() *cobra.Command{
+		apiserver,
+		controller,
+		scheduler,
+		kubelet,
+	}
+
+	cmd := &cobra.Command{
+		Use:   "kube",
+		Short: "Combines all Kubernetes components in a single binary",
+		Run: func(cmd *cobra.Command, args []string) {
+			if len(args) != 0 {
+				cmd.Help()
+				os.Exit(1)
+			}
+		},
+	}
+
+	for i := range commandFns {
+		cmd.AddCommand(commandFns[i]())
+	}
+
+	return cmd, commandFns
+}
diff --git a/metropolis/node/kubernetes/kubelet.go b/metropolis/node/kubernetes/kubelet.go
new file mode 100644
index 0000000..e9c6ce5
--- /dev/null
+++ b/metropolis/node/kubernetes/kubelet.go
@@ -0,0 +1,145 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kubernetes
+
+import (
+	"context"
+	"encoding/json"
+	"encoding/pem"
+	"fmt"
+	"io"
+	"net"
+	"os/exec"
+
+	v1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	kubeletconfig "k8s.io/kubelet/config/v1beta1"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fileargs"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage/declarative"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/pki"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/reconciler"
+)
+
+type kubeletService struct {
+	NodeName           string
+	ClusterDNS         []net.IP
+	KubeletDirectory   *localstorage.DataKubernetesKubeletDirectory
+	EphemeralDirectory *localstorage.EphemeralDirectory
+	Output             io.Writer
+	KPKI               *pki.KubernetesPKI
+}
+
+func (s *kubeletService) createCertificates(ctx context.Context) error {
+	identity := fmt.Sprintf("system:node:%s", s.NodeName)
+
+	ca := s.KPKI.Certificates[pki.IdCA]
+	cacert, _, err := ca.Ensure(ctx, s.KPKI.KV)
+	if err != nil {
+		return fmt.Errorf("could not ensure ca certificate: %w", err)
+	}
+
+	kubeconfig, err := pki.New(ca, "", pki.Client(identity, []string{"system:nodes"})).Kubeconfig(ctx, s.KPKI.KV)
+	if err != nil {
+		return fmt.Errorf("could not create volatile kubelet client cert: %w", err)
+	}
+
+	cert, key, err := pki.New(ca, "", pki.Server([]string{s.NodeName}, nil)).Ensure(ctx, s.KPKI.KV)
+	if err != nil {
+		return fmt.Errorf("could not create volatile kubelet server cert: %w", err)
+	}
+
+	// TODO(q3k): this should probably become its own function //metropolis/node/kubernetes/pki.
+	for _, el := range []struct {
+		target declarative.FilePlacement
+		data   []byte
+	}{
+		{s.KubeletDirectory.Kubeconfig, kubeconfig},
+		{s.KubeletDirectory.PKI.CACertificate, pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: cacert})},
+		{s.KubeletDirectory.PKI.Certificate, pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: cert})},
+		{s.KubeletDirectory.PKI.Key, pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: key})},
+	} {
+		if err := el.target.Write(el.data, 0400); err != nil {
+			return fmt.Errorf("could not write %v: %w", el.target, err)
+		}
+	}
+
+	return nil
+}
+
+func (s *kubeletService) configure() *kubeletconfig.KubeletConfiguration {
+	var clusterDNS []string
+	for _, dnsIP := range s.ClusterDNS {
+		clusterDNS = append(clusterDNS, dnsIP.String())
+	}
+
+	return &kubeletconfig.KubeletConfiguration{
+		TypeMeta: v1.TypeMeta{
+			Kind:       "KubeletConfiguration",
+			APIVersion: kubeletconfig.GroupName + "/v1beta1",
+		},
+		TLSCertFile:       s.KubeletDirectory.PKI.Certificate.FullPath(),
+		TLSPrivateKeyFile: s.KubeletDirectory.PKI.Key.FullPath(),
+		TLSMinVersion:     "VersionTLS13",
+		ClusterDNS:        clusterDNS,
+		Authentication: kubeletconfig.KubeletAuthentication{
+			X509: kubeletconfig.KubeletX509Authentication{
+				ClientCAFile: s.KubeletDirectory.PKI.CACertificate.FullPath(),
+			},
+		},
+		// TODO(q3k): move reconciler.False to a generic package, fix the following references.
+		ClusterDomain:                "cluster.local", // cluster.local is hardcoded in the certificate too currently
+		EnableControllerAttachDetach: reconciler.False(),
+		HairpinMode:                  "none",
+		MakeIPTablesUtilChains:       reconciler.False(), // We don't have iptables
+		FailSwapOn:                   reconciler.False(), // Our kernel doesn't have swap enabled which breaks Kubelet's detection
+		KubeReserved: map[string]string{
+			"cpu":    "200m",
+			"memory": "300Mi",
+		},
+
+		// We're not going to use this, but let's make it point to a known-empty directory in case anybody manages to
+		// trigger it.
+		VolumePluginDir: s.EphemeralDirectory.FlexvolumePlugins.FullPath(),
+	}
+}
+
+func (s *kubeletService) Run(ctx context.Context) error {
+	if err := s.createCertificates(ctx); err != nil {
+		return fmt.Errorf("when creating certificates: %w", err)
+	}
+
+	configRaw, err := json.Marshal(s.configure())
+	if err != nil {
+		return fmt.Errorf("when marshaling kubelet configuration: %w", err)
+	}
+
+	fargs, err := fileargs.New()
+	if err != nil {
+		return err
+	}
+	cmd := exec.CommandContext(ctx, "/kubernetes/bin/kube", "kubelet",
+		fargs.FileOpt("--config", "config.json", configRaw),
+		"--container-runtime=remote",
+		fmt.Sprintf("--container-runtime-endpoint=unix://%s", s.EphemeralDirectory.Containerd.ClientSocket.FullPath()),
+		fmt.Sprintf("--kubeconfig=%s", s.KubeletDirectory.Kubeconfig.FullPath()),
+		fmt.Sprintf("--root-dir=%s", s.KubeletDirectory.FullPath()),
+	)
+	cmd.Env = []string{"PATH=/kubernetes/bin"}
+	return supervisor.RunCommand(ctx, cmd)
+}
diff --git a/metropolis/node/kubernetes/nfproxy/BUILD.bazel b/metropolis/node/kubernetes/nfproxy/BUILD.bazel
new file mode 100644
index 0000000..29124a6
--- /dev/null
+++ b/metropolis/node/kubernetes/nfproxy/BUILD.bazel
@@ -0,0 +1,22 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["nfproxy.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/nfproxy",
+    visibility = ["//metropolis/node/kubernetes:__subpackages__"],
+    deps = [
+        "//metropolis/node/common/supervisor:go_default_library",
+        "@com_github_sbezverk_nfproxy//pkg/controller:go_default_library",
+        "@com_github_sbezverk_nfproxy//pkg/nftables:go_default_library",
+        "@com_github_sbezverk_nfproxy//pkg/proxy:go_default_library",
+        "@io_k8s_api//core/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/labels:go_default_library",
+        "@io_k8s_apimachinery//pkg/selection:go_default_library",
+        "@io_k8s_client_go//informers:go_default_library",
+        "@io_k8s_client_go//kubernetes:go_default_library",
+        "@io_k8s_client_go//kubernetes/scheme:go_default_library",
+        "@io_k8s_client_go//tools/record:go_default_library",
+    ],
+)
diff --git a/metropolis/node/kubernetes/nfproxy/nfproxy.go b/metropolis/node/kubernetes/nfproxy/nfproxy.go
new file mode 100644
index 0000000..5fc9a11
--- /dev/null
+++ b/metropolis/node/kubernetes/nfproxy/nfproxy.go
@@ -0,0 +1,104 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package nfproxy is a Kubernetes Service IP proxy based exclusively on the Linux nftables interface.
+// It uses netfilter's NAT capabilities to accept traffic on service IPs and DNAT it to the respective endpoint.
+package nfproxy
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+	"os"
+	"time"
+
+	"github.com/sbezverk/nfproxy/pkg/controller"
+	"github.com/sbezverk/nfproxy/pkg/nftables"
+	"github.com/sbezverk/nfproxy/pkg/proxy"
+	v1 "k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/labels"
+	"k8s.io/apimachinery/pkg/selection"
+	kubeinformers "k8s.io/client-go/informers"
+	"k8s.io/client-go/kubernetes"
+	"k8s.io/client-go/kubernetes/scheme"
+	"k8s.io/client-go/tools/record"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+)
+
+type Service struct {
+	// Traffic in ClusterCIDR is assumed to be originated inside the cluster and will not be SNATed
+	ClusterCIDR net.IPNet
+	// A Kubernetes ClientSet with read access to endpoints and services
+	ClientSet kubernetes.Interface
+}
+
+func (s *Service) Run(ctx context.Context) error {
+	var ipv4ClusterCIDR string
+	var ipv6ClusterCIDR string
+	if s.ClusterCIDR.IP.To4() == nil && s.ClusterCIDR.IP.To16() != nil {
+		ipv6ClusterCIDR = s.ClusterCIDR.String()
+	} else if s.ClusterCIDR.IP.To4() != nil {
+		ipv4ClusterCIDR = s.ClusterCIDR.String()
+	} else {
+		return errors.New("invalid ClusterCIDR")
+	}
+	nfti, err := nftables.InitNFTables(ipv4ClusterCIDR, ipv6ClusterCIDR)
+	if err != nil {
+		return fmt.Errorf("failed to initialize nftables with error: %w", err)
+	}
+
+	// Create event recorder to report events into K8s
+	hostname, err := os.Hostname()
+	if err != nil {
+		return fmt.Errorf("failed to get local host name with error: %w", err)
+	}
+	eventBroadcaster := record.NewBroadcaster()
+	recorder := eventBroadcaster.NewRecorder(scheme.Scheme, v1.EventSource{Component: "nfproxy", Host: hostname})
+
+	// Create new proxy controller with endpoint slices enabled
+	// https://kubernetes.io/docs/concepts/services-networking/endpoint-slices/
+	nfproxy := proxy.NewProxy(nfti, hostname, recorder, true)
+
+	// Create special informer which doesn't track headless services
+	noHeadlessEndpoints, err := labels.NewRequirement(v1.IsHeadlessService, selection.DoesNotExist, nil)
+	if err != nil {
+		return fmt.Errorf("failed to create Requirement for noHeadlessEndpoints: %w", err)
+	}
+	labelSelector := labels.NewSelector()
+	labelSelector = labelSelector.Add(*noHeadlessEndpoints)
+
+	kubeInformerFactory := kubeinformers.NewSharedInformerFactoryWithOptions(s.ClientSet, time.Minute*5,
+		kubeinformers.WithTweakListOptions(func(options *metav1.ListOptions) {
+			options.LabelSelector = labelSelector.String()
+		}))
+
+	svcController := controller.NewServiceController(nfproxy, s.ClientSet, kubeInformerFactory.Core().V1().Services())
+	ep := controller.NewEndpointSliceController(nfproxy, s.ClientSet, kubeInformerFactory.Discovery().V1beta1().EndpointSlices())
+	kubeInformerFactory.Start(ctx.Done())
+
+	if err = svcController.Start(ctx.Done()); err != nil {
+		return fmt.Errorf("error running Service controller: %w", err)
+	}
+	if err = ep.Start(ctx.Done()); err != nil {
+		return fmt.Errorf("error running endpoint controller: %w", err)
+	}
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	supervisor.Signal(ctx, supervisor.SignalDone)
+	return nil
+}
diff --git a/metropolis/node/kubernetes/pki/BUILD.bazel b/metropolis/node/kubernetes/pki/BUILD.bazel
new file mode 100644
index 0000000..f82603d
--- /dev/null
+++ b/metropolis/node/kubernetes/pki/BUILD.bazel
@@ -0,0 +1,19 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "ca.go",
+        "certificate.go",
+        "kubernetes.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/pki",
+    visibility = ["//metropolis/node:__subpackages__"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/node/core/logtree:go_default_library",
+        "@io_etcd_go_etcd//clientv3:go_default_library",
+        "@io_k8s_client_go//tools/clientcmd:go_default_library",
+        "@io_k8s_client_go//tools/clientcmd/api:go_default_library",
+    ],
+)
diff --git a/metropolis/node/kubernetes/pki/ca.go b/metropolis/node/kubernetes/pki/ca.go
new file mode 100644
index 0000000..64453cd
--- /dev/null
+++ b/metropolis/node/kubernetes/pki/ca.go
@@ -0,0 +1,151 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pki
+
+import (
+	"context"
+	"crypto"
+	"crypto/ed25519"
+	"crypto/rand"
+	"crypto/sha1"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/asn1"
+	"fmt"
+	"math/big"
+	"time"
+
+	"go.etcd.io/etcd/clientv3"
+)
+
+// Issuer is a CA that can issue certificates. Two issuers are currently implemented:
+//  - SelfSigned, which will generated a certificate signed by its corresponding private key.
+//  - Certificate, which will use another existing Certificate as a CA.
+type Issuer interface {
+	// CACertificate returns the DER-encoded x509 certificate of the CA that will sign certificates when Issue is
+	// called, or nil if this is self-signing issuer.
+	CACertificate(ctx context.Context, kv clientv3.KV) ([]byte, error)
+	// Issue will generate a key and certificate signed by the Issuer. The returned certificate is x509 DER-encoded,
+	// while the key is a bare ed25519 key.
+	Issue(ctx context.Context, template x509.Certificate, kv clientv3.KV) (cert, key []byte, err error)
+}
+
+var (
+	// From RFC 5280 Section 4.1.2.5
+	unknownNotAfter = time.Unix(253402300799, 0)
+)
+
+// Workaround for https://github.com/golang/go/issues/26676 in Go's crypto/x509. Specifically Go
+// violates Section 4.2.1.2 of RFC 5280 without this.
+// Fixed for 1.15 in https://go-review.googlesource.com/c/go/+/227098/.
+//
+// Taken from https://github.com/FiloSottile/mkcert/blob/master/cert.go#L295 written by one of Go's
+// crypto engineers
+func calculateSKID(pubKey crypto.PublicKey) ([]byte, error) {
+	spkiASN1, err := x509.MarshalPKIXPublicKey(pubKey)
+	if err != nil {
+		return nil, err
+	}
+
+	var spki struct {
+		Algorithm        pkix.AlgorithmIdentifier
+		SubjectPublicKey asn1.BitString
+	}
+	_, err = asn1.Unmarshal(spkiASN1, &spki)
+	if err != nil {
+		return nil, err
+	}
+	skid := sha1.Sum(spki.SubjectPublicKey.Bytes)
+	return skid[:], nil
+}
+
+// issueCertificate is a generic low level certificate-and-key issuance function. If ca or cakey is null, the
+// certificate will be self-signed. The returned certificate is DER-encoded, while the returned key is internal.
+func issueCertificate(template x509.Certificate, ca *x509.Certificate, caKey interface{}) (cert, key []byte, err error) {
+	pubKey, privKey, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		panic(err)
+	}
+
+	serialNumberLimit := new(big.Int).Lsh(big.NewInt(1), 127)
+	serialNumber, err := rand.Int(rand.Reader, serialNumberLimit)
+	if err != nil {
+		err = fmt.Errorf("failed to generate serial number: %w", err)
+		return
+	}
+
+	skid, err := calculateSKID(pubKey)
+	if err != nil {
+		return []byte{}, privKey, err
+	}
+
+	template.SerialNumber = serialNumber
+	template.NotBefore = time.Now()
+	template.NotAfter = unknownNotAfter
+	template.BasicConstraintsValid = true
+	template.SubjectKeyId = skid
+
+	// Set the AuthorityKeyID to the SKID of the signing certificate (or self, if self-signing).
+	if ca != nil && caKey != nil {
+		template.AuthorityKeyId = ca.AuthorityKeyId
+	} else {
+		template.AuthorityKeyId = template.SubjectKeyId
+	}
+
+	if ca == nil || caKey == nil {
+		ca = &template
+		caKey = privKey
+	}
+
+	caCertRaw, err := x509.CreateCertificate(rand.Reader, &template, ca, pubKey, caKey)
+	return caCertRaw, privKey, err
+}
+
+type selfSigned struct{}
+
+func (s *selfSigned) Issue(ctx context.Context, template x509.Certificate, kv clientv3.KV) (cert, key []byte, err error) {
+	return issueCertificate(template, nil, nil)
+}
+
+func (s *selfSigned) CACertificate(ctx context.Context, kv clientv3.KV) ([]byte, error) {
+	return nil, nil
+}
+
+var (
+	// SelfSigned is an Issuer that generates self-signed certificates.
+	SelfSigned = &selfSigned{}
+)
+
+func (c *Certificate) Issue(ctx context.Context, template x509.Certificate, kv clientv3.KV) (cert, key []byte, err error) {
+	caCert, caKey, err := c.ensure(ctx, kv)
+	if err != nil {
+		return nil, nil, fmt.Errorf("could not ensure CA certificate %q exists: %w", c.name, err)
+	}
+
+	ca, err := x509.ParseCertificate(caCert)
+	if err != nil {
+		return nil, nil, fmt.Errorf("could not parse CA certificate: %w", err)
+	}
+	// Ensure only one level of CAs exist, and that they are created explicitly.
+	template.IsCA = false
+	return issueCertificate(template, ca, ed25519.PrivateKey(caKey))
+}
+
+func (c *Certificate) CACertificate(ctx context.Context, kv clientv3.KV) ([]byte, error) {
+	cert, _, err := c.ensure(ctx, kv)
+	return cert, err
+}
diff --git a/metropolis/node/kubernetes/pki/certificate.go b/metropolis/node/kubernetes/pki/certificate.go
new file mode 100644
index 0000000..6bd50f9
--- /dev/null
+++ b/metropolis/node/kubernetes/pki/certificate.go
@@ -0,0 +1,192 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pki
+
+import (
+	"context"
+	"crypto/ed25519"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"fmt"
+	"net"
+
+	"go.etcd.io/etcd/clientv3"
+)
+
+// Certificate is the promise of a Certificate being available to the caller. In this case, Certificate refers to a
+// pair of x509 certificate and corresponding private key.
+// Certificates can be stored in etcd, and their issuers might also be store on etcd. As such, this type's methods
+// contain references to an etcd KV client.
+// This Certificate type is agnostic to usage, but mostly geared towards Kubernetes certificates.
+type Certificate struct {
+	// issuer is the Issuer that will generate this certificate if one doesn't yet exist or etcd, or the requested
+	// certificate is volatile (not to be stored on etcd).
+	issuer Issuer
+	// name is a unique key for storing the certificate in etcd. If empty, certificate is 'volatile', will not be stored
+	// on etcd, and every .Ensure() call will generate a new pair.
+	name string
+	// template is an x509 certificate definition that will be used to generate the certificate when issuing it.
+	template x509.Certificate
+}
+
+const (
+	// etcdPrefix is where all the PKI data is stored in etcd.
+	etcdPrefix = "/kube-pki/"
+)
+
+func etcdPath(f string, args ...interface{}) string {
+	return etcdPrefix + fmt.Sprintf(f, args...)
+}
+
+// New creates a new Certificate, or to be more precise, a promise that a certificate will exist once Ensure is called.
+// Issuer must be a valid certificate issuer (SelfSigned or another Certificate). Name must be unique among all
+// certificates, or empty (which will cause the certificate to be volatile, ie. not stored in etcd).
+func New(issuer Issuer, name string, template x509.Certificate) *Certificate {
+	return &Certificate{
+		issuer:   issuer,
+		name:     name,
+		template: template,
+	}
+}
+
+// Client makes a Kubernetes PKI-compatible client certificate template.
+// Directly derived from Kubernetes PKI requirements documented at
+// https://kubernetes.io/docs/setup/best-practices/certificates/#configure-certificates-manually
+func Client(identity string, groups []string) x509.Certificate {
+	return x509.Certificate{
+		Subject: pkix.Name{
+			CommonName:   identity,
+			Organization: groups,
+		},
+		KeyUsage:    x509.KeyUsageDigitalSignature | x509.KeyUsageKeyEncipherment,
+		ExtKeyUsage: []x509.ExtKeyUsage{x509.ExtKeyUsageClientAuth},
+	}
+}
+
+// Server makes a Kubernetes PKI-compatible server certificate template.
+func Server(dnsNames []string, ips []net.IP) x509.Certificate {
+	return x509.Certificate{
+		Subject:     pkix.Name{},
+		KeyUsage:    x509.KeyUsageDigitalSignature | x509.KeyUsageKeyEncipherment,
+		ExtKeyUsage: []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
+		DNSNames:    dnsNames,
+		IPAddresses: ips,
+	}
+}
+
+// CA makes a Certificate that can sign other certificates.
+func CA(cn string) x509.Certificate {
+	return x509.Certificate{
+		Subject: pkix.Name{
+			CommonName: cn,
+		},
+		IsCA:        true,
+		KeyUsage:    x509.KeyUsageCertSign | x509.KeyUsageCRLSign | x509.KeyUsageDigitalSignature,
+		ExtKeyUsage: []x509.ExtKeyUsage{x509.ExtKeyUsageClientAuth, x509.ExtKeyUsageServerAuth, x509.ExtKeyUsageOCSPSigning},
+	}
+}
+
+func (c *Certificate) etcdPaths() (cert, key string) {
+	return etcdPath("%s-cert.der", c.name), etcdPath("%s-key.der", c.name)
+}
+
+// ensure returns a DER-encoded x509 certificate and internally encoded bare ed25519 key for a given Certificate,
+// in memory (if volatile), loading it from etcd, or creating and saving it on etcd if needed.
+// This function is safe to call in parallel from multiple etcd clients (including across machines), but it will error
+// in case a concurrent certificate generation happens. These errors are, however, safe to retry - as long as all the
+// certificate creators (ie., Smalltown nodes) run the same version of this code.
+// TODO(q3k): in the future, this should be handled better - especially as we introduce new certificates, or worse,
+// change the issuance chain. As a stopgap measure, an explicit per-certificate or even global lock can be implemented.
+// And, even before that, we can handle concurrency errors in a smarter way.
+func (c *Certificate) ensure(ctx context.Context, kv clientv3.KV) (cert, key []byte, err error) {
+	if c.name == "" {
+		// Volatile certificate - generate.
+		// TODO(q3k): cache internally?
+		cert, key, err = c.issuer.Issue(ctx, c.template, kv)
+		if err != nil {
+			err = fmt.Errorf("failed to issue: %w", err)
+			return
+		}
+		return
+	}
+
+	certPath, keyPath := c.etcdPaths()
+
+	// Try loading certificate and key from etcd.
+	certRes, err := kv.Get(ctx, certPath)
+	if err != nil {
+		err = fmt.Errorf("failed to get certificate from etcd: %w", err)
+		return
+	}
+	keyRes, err := kv.Get(ctx, keyPath)
+	if err != nil {
+		err = fmt.Errorf("failed to get key from etcd: %w", err)
+		return
+	}
+
+	if len(certRes.Kvs) == 1 && len(keyRes.Kvs) == 1 {
+		// Certificate and key exists in etcd, return that.
+		cert = certRes.Kvs[0].Value
+		key = keyRes.Kvs[0].Value
+
+		err = nil
+		// TODO(q3k): check for expiration
+		return
+	}
+
+	// No certificate found - issue one.
+	cert, key, err = c.issuer.Issue(ctx, c.template, kv)
+	if err != nil {
+		err = fmt.Errorf("failed to issue: %w", err)
+		return
+	}
+
+	// Save to etcd in transaction. This ensures that no partial writes happen, and that we haven't been raced to the
+	// save.
+	res, err := kv.Txn(ctx).
+		If(
+			clientv3.Compare(clientv3.CreateRevision(certPath), "=", 0),
+			clientv3.Compare(clientv3.CreateRevision(keyPath), "=", 0),
+		).
+		Then(
+			clientv3.OpPut(certPath, string(cert)),
+			clientv3.OpPut(keyPath, string(key)),
+		).Commit()
+	if err != nil {
+		err = fmt.Errorf("failed to write newly issued certificate: %w", err)
+	} else if !res.Succeeded {
+		err = fmt.Errorf("certificate issuance transaction failed: concurrent write")
+	}
+
+	return
+}
+
+// Ensure returns an x509 DER-encoded (but not PEM-encoded) certificate and key for a given Certificate.
+// If the certificate is volatile, each call to Ensure will cause a new certificate to be generated.
+// Otherwise, it will be retrieved from etcd, or generated and stored there if needed.
+func (c *Certificate) Ensure(ctx context.Context, kv clientv3.KV) (cert, key []byte, err error) {
+	cert, key, err = c.ensure(ctx, kv)
+	if err != nil {
+		return nil, nil, err
+	}
+	key, err = x509.MarshalPKCS8PrivateKey(ed25519.PrivateKey(key))
+	if err != nil {
+		err = fmt.Errorf("could not marshal private key (data corruption?): %w", err)
+		return
+	}
+	return cert, key, err
+}
diff --git a/metropolis/node/kubernetes/pki/kubernetes.go b/metropolis/node/kubernetes/pki/kubernetes.go
new file mode 100644
index 0000000..c4827a9
--- /dev/null
+++ b/metropolis/node/kubernetes/pki/kubernetes.go
@@ -0,0 +1,228 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pki
+
+import (
+	"context"
+	"crypto/rand"
+	"crypto/rsa"
+	"crypto/x509"
+	"encoding/pem"
+	"fmt"
+	"net"
+
+	"go.etcd.io/etcd/clientv3"
+	"k8s.io/client-go/tools/clientcmd"
+	configapi "k8s.io/client-go/tools/clientcmd/api"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+)
+
+// KubeCertificateName is an enum-like unique name of a static Kubernetes certificate. The value of the name is used
+// as the unique part of an etcd path where the certificate and key are stored.
+type KubeCertificateName string
+
+const (
+	// The main Kubernetes CA, used to authenticate API consumers, and servers.
+	IdCA KubeCertificateName = "id-ca"
+
+	// Kubernetes apiserver server certificate.
+	APIServer KubeCertificateName = "apiserver"
+
+	// Kubelet client certificate, used to authenticate to the apiserver.
+	KubeletClient KubeCertificateName = "kubelet-client"
+
+	// Kubernetes Controller manager client certificate, used to authenticate to the apiserver.
+	ControllerManagerClient KubeCertificateName = "controller-manager-client"
+	// Kubernetes Controller manager server certificate, used to run its HTTP server.
+	ControllerManager KubeCertificateName = "controller-manager"
+
+	// Kubernetes Scheduler client certificate, used to authenticate to the apiserver.
+	SchedulerClient KubeCertificateName = "scheduler-client"
+	// Kubernetes scheduler server certificate, used to run its HTTP server.
+	Scheduler KubeCertificateName = "scheduler"
+
+	// Root-on-kube (system:masters) client certificate. Used to control the apiserver (and resources) by Smalltown
+	// internally.
+	Master KubeCertificateName = "master"
+
+	// OpenAPI Kubernetes Aggregation CA.
+	// See: https://kubernetes.io/docs/tasks/extend-kubernetes/configure-aggregation-layer/#ca-reusage-and-conflicts
+	AggregationCA    KubeCertificateName = "aggregation-ca"
+	FrontProxyClient KubeCertificateName = "front-proxy-client"
+)
+
+const (
+	// serviceAccountKeyName is the etcd path part that is used to store the ServiceAccount authentication secret.
+	// This is not a certificate, just an RSA key.
+	serviceAccountKeyName = "service-account-privkey"
+)
+
+// KubernetesPKI manages all PKI resources required to run Kubernetes on Smalltown. It contains all static certificates,
+// which can be retrieved, or be used to generate Kubeconfigs from.
+type KubernetesPKI struct {
+	logger       logtree.LeveledLogger
+	KV           clientv3.KV
+	Certificates map[KubeCertificateName]*Certificate
+}
+
+func NewKubernetes(l logtree.LeveledLogger, kv clientv3.KV) *KubernetesPKI {
+	pki := KubernetesPKI{
+		logger:       l,
+		KV:           kv,
+		Certificates: make(map[KubeCertificateName]*Certificate),
+	}
+
+	make := func(i, name KubeCertificateName, template x509.Certificate) {
+		pki.Certificates[name] = New(pki.Certificates[i], string(name), template)
+	}
+
+	pki.Certificates[IdCA] = New(SelfSigned, string(IdCA), CA("Smalltown Kubernetes ID CA"))
+	make(IdCA, APIServer, Server(
+		[]string{
+			"kubernetes",
+			"kubernetes.default",
+			"kubernetes.default.svc",
+			"kubernetes.default.svc.cluster",
+			"kubernetes.default.svc.cluster.local",
+			"localhost",
+		},
+		[]net.IP{{10, 0, 255, 1}, {127, 0, 0, 1}}, // TODO(q3k): add service network internal apiserver address
+	))
+	make(IdCA, KubeletClient, Client("smalltown:apiserver-kubelet-client", nil))
+	make(IdCA, ControllerManagerClient, Client("system:kube-controller-manager", nil))
+	make(IdCA, ControllerManager, Server([]string{"kube-controller-manager.local"}, nil))
+	make(IdCA, SchedulerClient, Client("system:kube-scheduler", nil))
+	make(IdCA, Scheduler, Server([]string{"kube-scheduler.local"}, nil))
+	make(IdCA, Master, Client("smalltown:master", []string{"system:masters"}))
+
+	pki.Certificates[AggregationCA] = New(SelfSigned, string(AggregationCA), CA("Smalltown OpenAPI Aggregation CA"))
+	make(AggregationCA, FrontProxyClient, Client("front-proxy-client", nil))
+
+	return &pki
+}
+
+// EnsureAll ensures that all static certificates (and the serviceaccount key) are present on etcd.
+func (k *KubernetesPKI) EnsureAll(ctx context.Context) error {
+	for n, v := range k.Certificates {
+		k.logger.Infof("Ensuring %s exists", string(n))
+		_, _, err := v.Ensure(ctx, k.KV)
+		if err != nil {
+			return fmt.Errorf("could not ensure certificate %q exists: %w", n, err)
+		}
+	}
+	_, err := k.ServiceAccountKey(ctx)
+	if err != nil {
+		return fmt.Errorf("could not ensure service account key exists: %w", err)
+	}
+	return nil
+}
+
+// Kubeconfig generates a kubeconfig blob for a given certificate name. The same lifetime semantics as in .Certificate
+// apply.
+func (k *KubernetesPKI) Kubeconfig(ctx context.Context, name KubeCertificateName) ([]byte, error) {
+	c, ok := k.Certificates[name]
+	if !ok {
+		return nil, fmt.Errorf("no certificate %q", name)
+	}
+	return c.Kubeconfig(ctx, k.KV)
+}
+
+// Certificate retrieves an x509 DER-encoded (but not PEM-wrapped) key and certificate for a given certificate name.
+// If the requested certificate is volatile, it will be created on demand. Otherwise it will be created on etcd (if not
+// present), and retrieved from there.
+func (k *KubernetesPKI) Certificate(ctx context.Context, name KubeCertificateName) (cert, key []byte, err error) {
+	c, ok := k.Certificates[name]
+	if !ok {
+		return nil, nil, fmt.Errorf("no certificate %q", name)
+	}
+	return c.Ensure(ctx, k.KV)
+}
+
+// Kubeconfig generates a kubeconfig blob for this certificate. The same lifetime semantics as in .Ensure apply.
+func (c *Certificate) Kubeconfig(ctx context.Context, kv clientv3.KV) ([]byte, error) {
+
+	cert, key, err := c.Ensure(ctx, kv)
+	if err != nil {
+		return nil, fmt.Errorf("could not ensure certificate exists: %w", err)
+	}
+
+	kubeconfig := configapi.NewConfig()
+
+	cluster := configapi.NewCluster()
+	cluster.Server = fmt.Sprintf("https://127.0.0.1:%v", common.KubernetesAPIPort)
+
+	ca, err := c.issuer.CACertificate(ctx, kv)
+	if err != nil {
+		return nil, fmt.Errorf("could not get CA certificate: %w", err)
+	}
+	if ca != nil {
+		cluster.CertificateAuthorityData = pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: ca})
+	}
+	kubeconfig.Clusters["default"] = cluster
+
+	authInfo := configapi.NewAuthInfo()
+	authInfo.ClientCertificateData = pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: cert})
+	authInfo.ClientKeyData = pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: key})
+	kubeconfig.AuthInfos["default"] = authInfo
+
+	ct := configapi.NewContext()
+	ct.Cluster = "default"
+	ct.AuthInfo = "default"
+	kubeconfig.Contexts["default"] = ct
+
+	kubeconfig.CurrentContext = "default"
+	return clientcmd.Write(*kubeconfig)
+}
+
+// ServiceAccountKey retrieves (and possible generates and stores on etcd) the Kubernetes service account key. The
+// returned data is ready to be used by Kubernetes components (in PKIX form).
+func (k *KubernetesPKI) ServiceAccountKey(ctx context.Context) ([]byte, error) {
+	// TODO(q3k): this should be abstracted away once we abstract away etcd access into a library with try-or-create
+	// semantics.
+
+	path := etcdPath("%s.der", serviceAccountKeyName)
+
+	// Try loading  key from etcd.
+	keyRes, err := k.KV.Get(ctx, path)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get key from etcd: %w", err)
+	}
+
+	if len(keyRes.Kvs) == 1 {
+		// Certificate and key exists in etcd, return that.
+		return keyRes.Kvs[0].Value, nil
+	}
+
+	// No key found - generate one.
+	keyRaw, err := rsa.GenerateKey(rand.Reader, 2048)
+	if err != nil {
+		panic(err)
+	}
+	key, err := x509.MarshalPKCS8PrivateKey(keyRaw)
+	if err != nil {
+		panic(err) // Always a programmer error
+	}
+
+	// Save to etcd.
+	_, err = k.KV.Put(ctx, path, string(key))
+	if err != nil {
+		err = fmt.Errorf("failed to write newly generated key: %w", err)
+	}
+	return key, nil
+}
diff --git a/metropolis/node/kubernetes/provisioner.go b/metropolis/node/kubernetes/provisioner.go
new file mode 100644
index 0000000..b671125
--- /dev/null
+++ b/metropolis/node/kubernetes/provisioner.go
@@ -0,0 +1,368 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kubernetes
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+
+	v1 "k8s.io/api/core/v1"
+	storagev1 "k8s.io/api/storage/v1"
+	apierrs "k8s.io/apimachinery/pkg/api/errors"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/informers"
+	coreinformers "k8s.io/client-go/informers/core/v1"
+	storageinformers "k8s.io/client-go/informers/storage/v1"
+	"k8s.io/client-go/kubernetes"
+	"k8s.io/client-go/kubernetes/scheme"
+	typedcorev1 "k8s.io/client-go/kubernetes/typed/core/v1"
+	"k8s.io/client-go/tools/cache"
+	"k8s.io/client-go/tools/record"
+	ref "k8s.io/client-go/tools/reference"
+	"k8s.io/client-go/util/workqueue"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fsquota"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/logtree"
+)
+
+// ONCHANGE(//metropolis/node/kubernetes/reconciler:resources_csi.go): needs to match csiProvisionerServerName declared.
+const csiProvisionerServerName = "com.nexantic.smalltown.vfs"
+
+// csiProvisionerServer is responsible for the provisioning and deprovisioning of CSI-based container volumes. It runs on all
+// nodes and watches PVCs for ones assigned to the node it's running on and fulfills the provisioning request by
+// creating a directory, applying a quota and creating the corresponding PV. When the PV is released and its retention
+// policy is Delete, the directory and the PV resource are deleted.
+type csiProvisionerServer struct {
+	NodeName         string
+	Kubernetes       kubernetes.Interface
+	InformerFactory  informers.SharedInformerFactory
+	VolumesDirectory *localstorage.DataVolumesDirectory
+
+	claimQueue           workqueue.RateLimitingInterface
+	pvQueue              workqueue.RateLimitingInterface
+	recorder             record.EventRecorder
+	pvcInformer          coreinformers.PersistentVolumeClaimInformer
+	pvInformer           coreinformers.PersistentVolumeInformer
+	storageClassInformer storageinformers.StorageClassInformer
+	logger               logtree.LeveledLogger
+}
+
+// runCSIProvisioner runs the main provisioning machinery. It consists of a bunch of informers which keep track of
+// the events happening on the Kubernetes control plane and informs us when something happens. If anything happens to
+// PVCs or PVs, we enqueue the identifier of that resource in a work queue. Queues are being worked on by only one
+// worker to limit load and avoid complicated locking infrastructure. Failed items are requeued.
+func (p *csiProvisionerServer) Run(ctx context.Context) error {
+	// The recorder is used to log Kubernetes events for successful or failed volume provisions. These events then
+	// show up in `kubectl describe pvc` and can be used by admins to debug issues with this provisioner.
+	eventBroadcaster := record.NewBroadcaster()
+	eventBroadcaster.StartRecordingToSink(&typedcorev1.EventSinkImpl{Interface: p.Kubernetes.CoreV1().Events("")})
+	p.recorder = eventBroadcaster.NewRecorder(scheme.Scheme, v1.EventSource{Component: csiProvisionerServerName, Host: p.NodeName})
+
+	p.pvInformer = p.InformerFactory.Core().V1().PersistentVolumes()
+	p.pvcInformer = p.InformerFactory.Core().V1().PersistentVolumeClaims()
+	p.storageClassInformer = p.InformerFactory.Storage().V1().StorageClasses()
+
+	p.claimQueue = workqueue.NewRateLimitingQueue(workqueue.DefaultControllerRateLimiter())
+	p.pvQueue = workqueue.NewRateLimitingQueue(workqueue.DefaultControllerRateLimiter())
+
+	p.pvcInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
+		AddFunc: p.enqueueClaim,
+		UpdateFunc: func(old, new interface{}) {
+			p.enqueueClaim(new)
+		},
+	})
+	p.pvInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
+		AddFunc: p.enqueuePV,
+		UpdateFunc: func(old, new interface{}) {
+			p.enqueuePV(new)
+		},
+	})
+	p.logger = supervisor.Logger(ctx)
+
+	go p.pvcInformer.Informer().Run(ctx.Done())
+	go p.pvInformer.Informer().Run(ctx.Done())
+	go p.storageClassInformer.Informer().Run(ctx.Done())
+
+	// These will self-terminate once the queues are shut down
+	go p.processQueueItems(p.claimQueue, func(key string) error {
+		return p.processPVC(key)
+	})
+	go p.processQueueItems(p.pvQueue, func(key string) error {
+		return p.processPV(key)
+	})
+
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	<-ctx.Done()
+	p.claimQueue.ShutDown()
+	p.pvQueue.ShutDown()
+	return nil
+}
+
+// isOurPVC checks if the given PVC is is to be provisioned by this provisioner and has been scheduled onto this node
+func (p *csiProvisionerServer) isOurPVC(pvc *v1.PersistentVolumeClaim) bool {
+	if pvc.ObjectMeta.Annotations["volume.beta.kubernetes.io/storage-provisioner"] != csiProvisionerServerName {
+		return false
+	}
+	if pvc.ObjectMeta.Annotations["volume.kubernetes.io/selected-node"] != p.NodeName {
+		return false
+	}
+	return true
+}
+
+// isOurPV checks if the given PV has been provisioned by this provisioner and has been scheduled onto this node
+func (p *csiProvisionerServer) isOurPV(pv *v1.PersistentVolume) bool {
+	if pv.ObjectMeta.Annotations["pv.kubernetes.io/provisioned-by"] != csiProvisionerServerName {
+		return false
+	}
+	if pv.Spec.NodeAffinity.Required.NodeSelectorTerms[0].MatchExpressions[0].Values[0] != p.NodeName {
+		return false
+	}
+	return true
+}
+
+// enqueueClaim adds an added/changed PVC to the work queue
+func (p *csiProvisionerServer) enqueueClaim(obj interface{}) {
+	key, err := cache.MetaNamespaceKeyFunc(obj)
+	if err != nil {
+		p.logger.Errorf("Not queuing PVC because key could not be derived: %v", err)
+		return
+	}
+	p.claimQueue.Add(key)
+}
+
+// enqueuePV adds an added/changed PV to the work queue
+func (p *csiProvisionerServer) enqueuePV(obj interface{}) {
+	key, err := cache.MetaNamespaceKeyFunc(obj)
+	if err != nil {
+		p.logger.Errorf("Not queuing PV because key could not be derived: %v", err)
+		return
+	}
+	p.pvQueue.Add(key)
+}
+
+// processQueueItems gets items from the given work queue and calls the process function for each of them. It self-
+// terminates once the queue is shut down.
+func (p *csiProvisionerServer) processQueueItems(queue workqueue.RateLimitingInterface, process func(key string) error) {
+	for {
+		obj, shutdown := queue.Get()
+		if shutdown {
+			return
+		}
+
+		func(obj interface{}) {
+			defer queue.Done(obj)
+			key, ok := obj.(string)
+			if !ok {
+				queue.Forget(obj)
+				p.logger.Errorf("Expected string in workqueue, got %+v", obj)
+				return
+			}
+
+			if err := process(key); err != nil {
+				p.logger.Warningf("Failed processing item %q, requeueing (numrequeues: %d): %v", key, queue.NumRequeues(obj), err)
+				queue.AddRateLimited(obj)
+			}
+
+			queue.Forget(obj)
+		}(obj)
+	}
+}
+
+// volumePath gets the path where the volume is stored.
+func (p *csiProvisionerServer) volumePath(volumeID string) string {
+	return filepath.Join(p.VolumesDirectory.FullPath(), volumeID)
+}
+
+// processPVC looks at a single PVC item from the queue, determines if it needs to be provisioned and logs the
+// provisioning result to the recorder
+func (p *csiProvisionerServer) processPVC(key string) error {
+	namespace, name, err := cache.SplitMetaNamespaceKey(key)
+	if err != nil {
+		return fmt.Errorf("invalid resource key: %s", key)
+	}
+	pvc, err := p.pvcInformer.Lister().PersistentVolumeClaims(namespace).Get(name)
+	if apierrs.IsNotFound(err) {
+		return nil // nothing to do, no error
+	} else if err != nil {
+		return fmt.Errorf("failed to get PVC for processing: %w", err)
+	}
+
+	if !p.isOurPVC(pvc) {
+		return nil
+	}
+
+	if pvc.Status.Phase != "Pending" {
+		// If the PVC is not pending, we don't need to provision anything
+		return nil
+	}
+
+	storageClass, err := p.storageClassInformer.Lister().Get(*pvc.Spec.StorageClassName)
+	if err != nil {
+		return fmt.Errorf("")
+	}
+
+	if storageClass.Provisioner != csiProvisionerServerName {
+		// We're not responsible for this PVC. Can only happen if controller-manager makes a mistake
+		// setting the annotations, but we're bailing here anyways for safety.
+		return nil
+	}
+
+	err = p.provisionPVC(pvc, storageClass)
+
+	if err != nil {
+		p.recorder.Eventf(pvc, v1.EventTypeWarning, "ProvisioningFailed", "Failed to provision PV: %v", err)
+		return err
+	}
+	p.recorder.Eventf(pvc, v1.EventTypeNormal, "Provisioned", "Successfully provisioned PV")
+
+	return nil
+}
+
+// provisionPVC creates the directory where the volume lives, sets a quota for the requested amount of storage and
+// creates the PV object representing this new volume
+func (p *csiProvisionerServer) provisionPVC(pvc *v1.PersistentVolumeClaim, storageClass *storagev1.StorageClass) error {
+	claimRef, err := ref.GetReference(scheme.Scheme, pvc)
+	if err != nil {
+		return fmt.Errorf("failed to get reference to PVC: %w", err)
+	}
+
+	storageReq := pvc.Spec.Resources.Requests[v1.ResourceStorage]
+	if storageReq.IsZero() {
+		return fmt.Errorf("PVC is not requesting any storage, this is not supported")
+	}
+	capacity, ok := storageReq.AsInt64()
+	if !ok {
+		return fmt.Errorf("PVC requesting more than 2^63 bytes of storage, this is not supported")
+	}
+
+	if *pvc.Spec.VolumeMode == v1.PersistentVolumeBlock {
+		return fmt.Errorf("Block PVCs are not supported by Smalltown")
+	}
+
+	volumeID := "pvc-" + string(pvc.ObjectMeta.UID)
+	volumePath := p.volumePath(volumeID)
+
+	p.logger.Infof("Creating local PV %s", volumeID)
+	if err := os.Mkdir(volumePath, 0644); err != nil && !os.IsExist(err) {
+		return fmt.Errorf("failed to create volume directory: %w", err)
+	}
+	files, err := ioutil.ReadDir(volumePath)
+	if err != nil {
+		return fmt.Errorf("failed to list files in newly-created volume: %w", err)
+	}
+	if len(files) > 0 {
+		return errors.New("newly-created volume already contains data, bailing")
+	}
+	if err := fsquota.SetQuota(volumePath, uint64(capacity), 100000); err != nil {
+		return fmt.Errorf("failed to update quota: %v", err)
+	}
+
+	vol := &v1.PersistentVolume{
+		ObjectMeta: metav1.ObjectMeta{
+			Name: volumeID,
+			Annotations: map[string]string{
+				"pv.kubernetes.io/provisioned-by": csiProvisionerServerName},
+		},
+		Spec: v1.PersistentVolumeSpec{
+			AccessModes: []v1.PersistentVolumeAccessMode{v1.ReadWriteOnce},
+			Capacity: v1.ResourceList{
+				v1.ResourceStorage: storageReq, // We're always giving the exact amount
+			},
+			PersistentVolumeSource: v1.PersistentVolumeSource{
+				CSI: &v1.CSIPersistentVolumeSource{
+					Driver:       csiProvisionerServerName,
+					VolumeHandle: volumeID,
+				},
+			},
+			ClaimRef: claimRef,
+			NodeAffinity: &v1.VolumeNodeAffinity{
+				Required: &v1.NodeSelector{
+					NodeSelectorTerms: []v1.NodeSelectorTerm{
+						{
+							MatchExpressions: []v1.NodeSelectorRequirement{
+								{
+									Key:      "kubernetes.io/hostname",
+									Operator: v1.NodeSelectorOpIn,
+									Values:   []string{p.NodeName},
+								},
+							},
+						},
+					},
+				},
+			},
+			StorageClassName:              *pvc.Spec.StorageClassName,
+			PersistentVolumeReclaimPolicy: *storageClass.ReclaimPolicy,
+		},
+	}
+
+	_, err = p.Kubernetes.CoreV1().PersistentVolumes().Create(context.Background(), vol, metav1.CreateOptions{})
+	if err != nil && !apierrs.IsAlreadyExists(err) {
+		return fmt.Errorf("failed to create PV object: %w", err)
+	}
+	return nil
+}
+
+// processPV looks at a single PV item from the queue and checks if it has been released and needs to be deleted. If yes
+// it deletes the associated quota, directory and the PV object and logs the result to the recorder.
+func (p *csiProvisionerServer) processPV(key string) error {
+	_, name, err := cache.SplitMetaNamespaceKey(key)
+	if err != nil {
+		return fmt.Errorf("invalid resource key: %s", key)
+	}
+	pv, err := p.pvInformer.Lister().Get(name)
+	if apierrs.IsNotFound(err) {
+		return nil // nothing to do, no error
+	} else if err != nil {
+		return fmt.Errorf("failed to get PV for processing: %w", err)
+	}
+
+	if !p.isOurPV(pv) {
+		return nil
+	}
+	if pv.Spec.PersistentVolumeReclaimPolicy != v1.PersistentVolumeReclaimDelete || pv.Status.Phase != "Released" {
+		return nil
+	}
+	volumePath := p.volumePath(pv.Spec.CSI.VolumeHandle)
+
+	// Log deletes for auditing purposes
+	p.logger.Infof("Deleting persistent volume %s", pv.Spec.CSI.VolumeHandle)
+	if err := fsquota.SetQuota(volumePath, 0, 0); err != nil {
+		// We record these here manually since a successful deletion removes the PV we'd be attaching them to
+		p.recorder.Eventf(pv, v1.EventTypeWarning, "DeprovisioningFailed", "Failed to remove quota: %v", err)
+		return fmt.Errorf("failed to remove quota: %w", err)
+	}
+	err = os.RemoveAll(volumePath)
+	if os.IsNotExist(err) {
+		return nil
+	} else if err != nil {
+		p.recorder.Eventf(pv, v1.EventTypeWarning, "DeprovisioningFailed", "Failed to delete volume: %v", err)
+		return fmt.Errorf("failed to delete volume: %w", err)
+	}
+
+	err = p.Kubernetes.CoreV1().PersistentVolumes().Delete(context.Background(), pv.Name, metav1.DeleteOptions{})
+	if err != nil && !apierrs.IsNotFound(err) {
+		p.recorder.Eventf(pv, v1.EventTypeWarning, "DeprovisioningFailed", "Failed to delete PV object from K8s API: %v", err)
+		return fmt.Errorf("failed to delete PV object: %w", err)
+	}
+	return nil
+}
diff --git a/metropolis/node/kubernetes/reconciler/BUILD.bazel b/metropolis/node/kubernetes/reconciler/BUILD.bazel
new file mode 100644
index 0000000..d8f2db6
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/BUILD.bazel
@@ -0,0 +1,38 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "reconciler.go",
+        "resources_csi.go",
+        "resources_podsecuritypolicy.go",
+        "resources_rbac.go",
+        "resources_runtimeclass.go",
+        "resources_storageclass.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/reconciler",
+    visibility = ["//metropolis/node:__subpackages__"],
+    deps = [
+        "//metropolis/node/common/supervisor:go_default_library",
+        "@io_k8s_api//core/v1:go_default_library",
+        "@io_k8s_api//node/v1beta1:go_default_library",
+        "@io_k8s_api//policy/v1beta1:go_default_library",
+        "@io_k8s_api//rbac/v1:go_default_library",
+        "@io_k8s_api//storage/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+        "@io_k8s_client_go//kubernetes:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    srcs = ["reconciler_test.go"],
+    embed = [":go_default_library"],
+    deps = [
+        "@io_k8s_api//node/v1beta1:go_default_library",
+        "@io_k8s_api//policy/v1beta1:go_default_library",
+        "@io_k8s_api//rbac/v1:go_default_library",
+        "@io_k8s_api//storage/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+    ],
+)
diff --git a/metropolis/node/kubernetes/reconciler/reconciler.go b/metropolis/node/kubernetes/reconciler/reconciler.go
new file mode 100644
index 0000000..9c5ba4e
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/reconciler.go
@@ -0,0 +1,163 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// The reconciler ensures that a base set of K8s resources is always available in the cluster. These are necessary to
+// ensure correct out-of-the-box functionality. All resources containing the smalltown.com/builtin=true label are assumed
+// to be managed by the reconciler.
+// It currently does not revert modifications made by admins, it is  planned to create an admission plugin prohibiting
+// such modifications to resources with the smalltown.com/builtin label to deal with that problem. This would also solve a
+// potential issue where you could delete resources just by adding the smalltown.com/builtin=true label.
+package reconciler
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	meta "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+)
+
+// Sad workaround for all the pointer booleans in K8s specs
+func True() *bool {
+	val := true
+	return &val
+}
+func False() *bool {
+	val := false
+	return &val
+}
+
+const (
+	// BuiltinLabelKey is used as a k8s label to mark built-in objects (ie., managed by the reconciler)
+	BuiltinLabelKey = "smalltown.com/builtin"
+	// BuiltinLabelValue is used as a k8s label value, under the BuiltinLabelKey key.
+	BuiltinLabelValue = "true"
+	// BuiltinRBACPrefix is used to prefix all built-in objects that are part of the rbac/v1 API (eg.
+	// {Cluster,}Role{Binding,} objects). This corresponds to the colon-separated 'namespaces' notation used by
+	// Kubernetes system (system:) objects.
+	BuiltinRBACPrefix = "smalltown:"
+)
+
+// builtinLabels makes a kubernetes-compatible label dictionary (key->value) that is used to mark objects that are
+// built-in into Smalltown (ie., managed by the reconciler). These are then subsequently retrieved by listBuiltins.
+// The extra argument specifies what other labels are to be merged into the the labels dictionary, for convenience. If
+// nil or empty, no extra labels will be applied.
+func builtinLabels(extra map[string]string) map[string]string {
+	l := map[string]string{
+		BuiltinLabelKey: BuiltinLabelValue,
+	}
+	if extra != nil {
+		for k, v := range extra {
+			l[k] = v
+		}
+	}
+	return l
+}
+
+// listBuiltins returns a k8s client ListOptions structure that allows to retrieve all objects that are built-in into
+// Smalltown currently present in the API server (ie., ones that are to be managed by the reconciler). These are created
+// by applying builtinLabels to their metadata labels.
+var listBuiltins = meta.ListOptions{
+	LabelSelector: fmt.Sprintf("%s=%s", BuiltinLabelKey, BuiltinLabelValue),
+}
+
+// builtinRBACName returns a name that is compatible with colon-delimited 'namespaced' objects, a la system:*.
+// These names are to be used by all builtins created as part of the rbac/v1 Kubernetes API.
+func builtinRBACName(name string) string {
+	return BuiltinRBACPrefix + name
+}
+
+// resource is a type of resource to be managed by the reconciler. All builti-ins/reconciled objects must implement
+// this interface to be managed correctly by the reconciler.
+type resource interface {
+	// List returns a list of names of objects current present on the target (ie. k8s API server).
+	List(ctx context.Context) ([]string, error)
+	// Create creates an object on the target. The el interface{} argument is the black box object returned by the
+	// Expected() call.
+	Create(ctx context.Context, el interface{}) error
+	// Delete delete an object, by name, from the target.
+	Delete(ctx context.Context, name string) error
+	// Expected returns a map of all objects expected to be present on the target. The keys are names (which must
+	// correspond to the names returned by List() and used by Delete(), and the values are blackboxes that will then
+	// be passed to the Create() call if their corresponding key (name) does not exist on the target.
+	Expected() map[string]interface{}
+}
+
+func allResources(clientSet kubernetes.Interface) map[string]resource {
+	return map[string]resource{
+		"psps":                resourcePodSecurityPolicies{clientSet},
+		"clusterroles":        resourceClusterRoles{clientSet},
+		"clusterrolebindings": resourceClusterRoleBindings{clientSet},
+		"storageclasses":      resourceStorageClasses{clientSet},
+		"csidrivers":          resourceCSIDrivers{clientSet},
+		"runtimeclasses":      resourceRuntimeClasses{clientSet},
+	}
+}
+
+func Run(clientSet kubernetes.Interface) supervisor.Runnable {
+	return func(ctx context.Context) error {
+		log := supervisor.Logger(ctx)
+		resources := allResources(clientSet)
+		t := time.NewTicker(10 * time.Second)
+		reconcileAll := func() {
+			for name, resource := range resources {
+				if err := reconcile(ctx, resource); err != nil {
+					log.Warningf("Failed to reconcile built-in resources %s: %v", name, err)
+				}
+			}
+		}
+		supervisor.Signal(ctx, supervisor.SignalHealthy)
+		reconcileAll()
+		for {
+			select {
+			case <-t.C:
+				reconcileAll()
+			case <-ctx.Done():
+				return nil
+			}
+		}
+	}
+}
+
+func reconcile(ctx context.Context, r resource) error {
+	present, err := r.List(ctx)
+	if err != nil {
+		return err
+	}
+	presentSet := make(map[string]bool)
+	for _, el := range present {
+		presentSet[el] = true
+	}
+	expectedMap := r.Expected()
+	for name, el := range expectedMap {
+		if !presentSet[name] {
+			if err := r.Create(ctx, el); err != nil {
+				return err
+			}
+		}
+	}
+	for name, _ := range presentSet {
+		if _, ok := expectedMap[name]; !ok {
+			if err := r.Delete(ctx, name); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
diff --git a/metropolis/node/kubernetes/reconciler/reconciler_test.go b/metropolis/node/kubernetes/reconciler/reconciler_test.go
new file mode 100644
index 0000000..b58d4af
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/reconciler_test.go
@@ -0,0 +1,184 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package reconciler
+
+import (
+	"context"
+	"fmt"
+	"testing"
+
+	node "k8s.io/api/node/v1beta1"
+	policy "k8s.io/api/policy/v1beta1"
+	rbac "k8s.io/api/rbac/v1"
+	storage "k8s.io/api/storage/v1"
+	meta "k8s.io/apimachinery/pkg/apis/meta/v1"
+)
+
+// kubernetesMeta unwraps an interface{} that might contain a Kubernetes resource of type that is managed by the
+// reconciler. Any time a new Kubernetes type is managed by the reconciler, the following switch should be extended
+// to cover that type.
+func kubernetesMeta(v interface{}) *meta.ObjectMeta {
+	switch v2 := v.(type) {
+	case *rbac.ClusterRole:
+		return &v2.ObjectMeta
+	case *rbac.ClusterRoleBinding:
+		return &v2.ObjectMeta
+	case *storage.CSIDriver:
+		return &v2.ObjectMeta
+	case *storage.StorageClass:
+		return &v2.ObjectMeta
+	case *policy.PodSecurityPolicy:
+		return &v2.ObjectMeta
+	case *node.RuntimeClass:
+		return &v2.ObjectMeta
+	}
+	return nil
+}
+
+// TestExpectedNamedCorrectly ensures that all the Expected objects of all resource types have a correspondence between
+// their returned key and inner name. This contract must be met in order for the reconciler to not create runaway
+// resources. This assumes all managed resources are Kubernetes resources.
+func TestExpectedNamedCorrectly(t *testing.T) {
+	for reconciler, r := range allResources(nil) {
+		for outer, v := range r.Expected() {
+			meta := kubernetesMeta(v)
+			if meta == nil {
+				t.Errorf("reconciler %q, object %q: could not decode kubernetes metadata", reconciler, outer)
+				continue
+			}
+			if inner := meta.Name; outer != inner {
+				t.Errorf("reconciler %q, object %q: inner name mismatch (%q)", reconciler, outer, inner)
+				continue
+			}
+		}
+	}
+}
+
+// TestExpectedLabeledCorrectly ensures that all the Expected objects of all resource types have a Kubernetes metadata
+// label that signifies it's a builtin object, to be retrieved afterwards. This contract must be met in order for the
+// reconciler to not keep overwriting objects (and possibly failing), when a newly created object is not then
+// retrievable using a selector corresponding to this label. This assumes all managed resources are Kubernetes objects.
+func TestExpectedLabeledCorrectly(t *testing.T) {
+	for reconciler, r := range allResources(nil) {
+		for outer, v := range r.Expected() {
+			meta := kubernetesMeta(v)
+			if meta == nil {
+				t.Errorf("reconciler %q, object %q: could not decode kubernetes metadata", reconciler, outer)
+				continue
+			}
+			if data := meta.Labels[BuiltinLabelKey]; data != BuiltinLabelValue {
+				t.Errorf("reconciler %q, object %q: %q=%q, wanted =%q", reconciler, outer, BuiltinLabelKey, data, BuiltinLabelValue)
+				continue
+			}
+		}
+	}
+}
+
+// testResource is a resource type used for testing. The inner type is a string that is equal to its name (key).
+// It simulates a target (ie. k8s apiserver mock) that always acts nominally (all resources are created, deleted as
+// requested, and the state is consistent with requests).
+type testResource struct {
+	// current is the simulated state of resources in the target.
+	current map[string]string
+	// expected is what this type will report as the Expected() resources.
+	expected map[string]string
+}
+
+func (r *testResource) List(ctx context.Context) ([]string, error) {
+	var keys []string
+	for k, _ := range r.current {
+		keys = append(keys, k)
+	}
+	return keys, nil
+}
+
+func (r *testResource) Create(ctx context.Context, el interface{}) error {
+	r.current[el.(string)] = el.(string)
+	return nil
+}
+
+func (r *testResource) Delete(ctx context.Context, name string) error {
+	delete(r.current, name)
+	return nil
+}
+
+func (r *testResource) Expected() map[string]interface{} {
+	exp := make(map[string]interface{})
+	for k, v := range r.expected {
+		exp[k] = v
+	}
+	return exp
+}
+
+// newTestResource creates a test resource with a list of expected resource strings.
+func newTestResource(want ...string) *testResource {
+	expected := make(map[string]string)
+	for _, w := range want {
+		expected[w] = w
+	}
+	return &testResource{
+		current:  make(map[string]string),
+		expected: expected,
+	}
+}
+
+// currentDiff returns a human-readable string showing the different between the current state and the given resource
+// strings. If no difference is present, the returned string is empty.
+func (r *testResource) currentDiff(want ...string) string {
+	expected := make(map[string]string)
+	for _, w := range want {
+		if _, ok := r.current[w]; !ok {
+			return fmt.Sprintf("%q missing in current", w)
+		}
+		expected[w] = w
+	}
+	for _, g := range r.current {
+		if _, ok := expected[g]; !ok {
+			return fmt.Sprintf("%q spurious in current", g)
+		}
+	}
+	return ""
+}
+
+// TestBasicReconciliation ensures that the reconcile function does manipulate a target state based on a set of
+// expected resources.
+func TestBasicReconciliation(t *testing.T) {
+	ctx := context.Background()
+	r := newTestResource("foo", "bar", "baz")
+
+	// nothing should have happened yet (testing the test)
+	if diff := r.currentDiff(); diff != "" {
+		t.Fatalf("wrong state after creation: %s", diff)
+	}
+
+	if err := reconcile(ctx, r); err != nil {
+		t.Fatalf("reconcile: %v", err)
+	}
+	// everything requested should have been created
+	if diff := r.currentDiff("foo", "bar", "baz"); diff != "" {
+		t.Fatalf("wrong state after reconciliation: %s", diff)
+	}
+
+	delete(r.expected, "foo")
+	if err := reconcile(ctx, r); err != nil {
+		t.Fatalf("reconcile: %v", err)
+	}
+	// foo should not be missing
+	if diff := r.currentDiff("bar", "baz"); diff != "" {
+		t.Fatalf("wrong state after deleting foo: %s", diff)
+	}
+}
diff --git a/metropolis/node/kubernetes/reconciler/resources_csi.go b/metropolis/node/kubernetes/reconciler/resources_csi.go
new file mode 100644
index 0000000..ecbcb4b
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/resources_csi.go
@@ -0,0 +1,71 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package reconciler
+
+import (
+	"context"
+
+	storage "k8s.io/api/storage/v1"
+	meta "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
+)
+
+// TODO(q3k): this is duplicated with //metropolis/node/kubernetes:provisioner.go; integrate this once provisioner.go
+// gets moved into a subpackage.
+// ONCHANGE(//metropolis/node/kubernetes:provisioner.go): needs to match csiProvisionerName declared.
+const csiProvisionerName = "com.nexantic.smalltown.vfs"
+
+type resourceCSIDrivers struct {
+	kubernetes.Interface
+}
+
+func (r resourceCSIDrivers) List(ctx context.Context) ([]string, error) {
+	res, err := r.StorageV1().CSIDrivers().List(ctx, listBuiltins)
+	if err != nil {
+		return nil, err
+	}
+	objs := make([]string, len(res.Items))
+	for i, el := range res.Items {
+		objs[i] = el.ObjectMeta.Name
+	}
+	return objs, nil
+}
+
+func (r resourceCSIDrivers) Create(ctx context.Context, el interface{}) error {
+	_, err := r.StorageV1().CSIDrivers().Create(ctx, el.(*storage.CSIDriver), meta.CreateOptions{})
+	return err
+}
+
+func (r resourceCSIDrivers) Delete(ctx context.Context, name string) error {
+	return r.StorageV1().CSIDrivers().Delete(ctx, name, meta.DeleteOptions{})
+}
+
+func (r resourceCSIDrivers) Expected() map[string]interface{} {
+	return map[string]interface{}{
+		csiProvisionerName: &storage.CSIDriver{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   csiProvisionerName,
+				Labels: builtinLabels(nil),
+			},
+			Spec: storage.CSIDriverSpec{
+				AttachRequired:       False(),
+				PodInfoOnMount:       False(),
+				VolumeLifecycleModes: []storage.VolumeLifecycleMode{storage.VolumeLifecyclePersistent},
+			},
+		},
+	}
+}
diff --git a/metropolis/node/kubernetes/reconciler/resources_podsecuritypolicy.go b/metropolis/node/kubernetes/reconciler/resources_podsecuritypolicy.go
new file mode 100644
index 0000000..507089f
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/resources_podsecuritypolicy.go
@@ -0,0 +1,108 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package reconciler
+
+import (
+	"context"
+
+	core "k8s.io/api/core/v1"
+	policy "k8s.io/api/policy/v1beta1"
+	meta "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
+)
+
+type resourcePodSecurityPolicies struct {
+	kubernetes.Interface
+}
+
+func (r resourcePodSecurityPolicies) List(ctx context.Context) ([]string, error) {
+	res, err := r.PolicyV1beta1().PodSecurityPolicies().List(ctx, listBuiltins)
+	if err != nil {
+		return nil, err
+	}
+	objs := make([]string, len(res.Items))
+	for i, el := range res.Items {
+		objs[i] = el.ObjectMeta.Name
+	}
+	return objs, nil
+}
+
+func (r resourcePodSecurityPolicies) Create(ctx context.Context, el interface{}) error {
+	_, err := r.PolicyV1beta1().PodSecurityPolicies().Create(ctx, el.(*policy.PodSecurityPolicy), meta.CreateOptions{})
+	return err
+}
+
+func (r resourcePodSecurityPolicies) Delete(ctx context.Context, name string) error {
+	return r.PolicyV1beta1().PodSecurityPolicies().Delete(ctx, name, meta.DeleteOptions{})
+}
+
+func (r resourcePodSecurityPolicies) Expected() map[string]interface{} {
+	return map[string]interface{}{
+		"default": &policy.PodSecurityPolicy{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   "default",
+				Labels: builtinLabels(nil),
+				Annotations: map[string]string{
+					"kubernetes.io/description": "This default PSP allows the creation of pods using features that are" +
+						" generally considered safe against any sort of escape.",
+				},
+			},
+			Spec: policy.PodSecurityPolicySpec{
+				AllowPrivilegeEscalation: True(),
+				AllowedCapabilities: []core.Capability{ // runc's default list of allowed capabilities
+					"SETPCAP",
+					"MKNOD",
+					"AUDIT_WRITE",
+					"CHOWN",
+					"NET_RAW",
+					"DAC_OVERRIDE",
+					"FOWNER",
+					"FSETID",
+					"KILL",
+					"SETGID",
+					"SETUID",
+					"NET_BIND_SERVICE",
+					"SYS_CHROOT",
+					"SETFCAP",
+				},
+				HostNetwork: false,
+				HostIPC:     false,
+				HostPID:     false,
+				FSGroup: policy.FSGroupStrategyOptions{
+					Rule: policy.FSGroupStrategyRunAsAny,
+				},
+				RunAsUser: policy.RunAsUserStrategyOptions{
+					Rule: policy.RunAsUserStrategyRunAsAny,
+				},
+				SELinux: policy.SELinuxStrategyOptions{
+					Rule: policy.SELinuxStrategyRunAsAny,
+				},
+				SupplementalGroups: policy.SupplementalGroupsStrategyOptions{
+					Rule: policy.SupplementalGroupsStrategyRunAsAny,
+				},
+				Volumes: []policy.FSType{ // Volumes considered safe to use
+					policy.ConfigMap,
+					policy.EmptyDir,
+					policy.Projected,
+					policy.Secret,
+					policy.DownwardAPI,
+					policy.PersistentVolumeClaim,
+				},
+			},
+		},
+	}
+}
diff --git a/metropolis/node/kubernetes/reconciler/resources_rbac.go b/metropolis/node/kubernetes/reconciler/resources_rbac.go
new file mode 100644
index 0000000..40ca879
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/resources_rbac.go
@@ -0,0 +1,154 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package reconciler
+
+import (
+	"context"
+
+	rbac "k8s.io/api/rbac/v1"
+	meta "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
+)
+
+var (
+	clusterRolePSPDefault                    = builtinRBACName("psp-default")
+	clusterRoleBindingDefaultPSP             = builtinRBACName("default-psp-for-sa")
+	clusterRoleBindingAPIServerKubeletClient = builtinRBACName("apiserver-kubelet-client")
+)
+
+type resourceClusterRoles struct {
+	kubernetes.Interface
+}
+
+func (r resourceClusterRoles) List(ctx context.Context) ([]string, error) {
+	res, err := r.RbacV1().ClusterRoles().List(ctx, listBuiltins)
+	if err != nil {
+		return nil, err
+	}
+	objs := make([]string, len(res.Items))
+	for i, el := range res.Items {
+		objs[i] = el.ObjectMeta.Name
+	}
+	return objs, nil
+}
+
+func (r resourceClusterRoles) Create(ctx context.Context, el interface{}) error {
+	_, err := r.RbacV1().ClusterRoles().Create(ctx, el.(*rbac.ClusterRole), meta.CreateOptions{})
+	return err
+}
+
+func (r resourceClusterRoles) Delete(ctx context.Context, name string) error {
+	return r.RbacV1().ClusterRoles().Delete(ctx, name, meta.DeleteOptions{})
+}
+
+func (r resourceClusterRoles) Expected() map[string]interface{} {
+	return map[string]interface{}{
+		clusterRolePSPDefault: &rbac.ClusterRole{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   clusterRolePSPDefault,
+				Labels: builtinLabels(nil),
+				Annotations: map[string]string{
+					"kubernetes.io/description": "This role grants access to the \"default\" PSP.",
+				},
+			},
+			Rules: []rbac.PolicyRule{
+				{
+					APIGroups:     []string{"policy"},
+					Resources:     []string{"podsecuritypolicies"},
+					ResourceNames: []string{"default"},
+					Verbs:         []string{"use"},
+				},
+			},
+		},
+	}
+}
+
+type resourceClusterRoleBindings struct {
+	kubernetes.Interface
+}
+
+func (r resourceClusterRoleBindings) List(ctx context.Context) ([]string, error) {
+	res, err := r.RbacV1().ClusterRoleBindings().List(ctx, listBuiltins)
+	if err != nil {
+		return nil, err
+	}
+	objs := make([]string, len(res.Items))
+	for i, el := range res.Items {
+		objs[i] = el.ObjectMeta.Name
+	}
+	return objs, nil
+}
+
+func (r resourceClusterRoleBindings) Create(ctx context.Context, el interface{}) error {
+	_, err := r.RbacV1().ClusterRoleBindings().Create(ctx, el.(*rbac.ClusterRoleBinding), meta.CreateOptions{})
+	return err
+}
+
+func (r resourceClusterRoleBindings) Delete(ctx context.Context, name string) error {
+	return r.RbacV1().ClusterRoleBindings().Delete(ctx, name, meta.DeleteOptions{})
+}
+
+func (r resourceClusterRoleBindings) Expected() map[string]interface{} {
+	return map[string]interface{}{
+		clusterRoleBindingDefaultPSP: &rbac.ClusterRoleBinding{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   clusterRoleBindingDefaultPSP,
+				Labels: builtinLabels(nil),
+				Annotations: map[string]string{
+					"kubernetes.io/description": "This binding grants every service account access to the \"default\" PSP. " +
+						"Creation of Pods is still restricted by other RBAC roles. Otherwise no pods (unprivileged or not) " +
+						"can be created.",
+				},
+			},
+			RoleRef: rbac.RoleRef{
+				APIGroup: rbac.GroupName,
+				Kind:     "ClusterRole",
+				Name:     clusterRolePSPDefault,
+			},
+			Subjects: []rbac.Subject{
+				{
+					APIGroup: rbac.GroupName,
+					Kind:     "Group",
+					Name:     "system:serviceaccounts",
+				},
+			},
+		},
+		clusterRoleBindingAPIServerKubeletClient: &rbac.ClusterRoleBinding{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   clusterRoleBindingAPIServerKubeletClient,
+				Labels: builtinLabels(nil),
+				Annotations: map[string]string{
+					"kubernetes.io/description": "This binding grants the apiserver access to the kubelets. This enables " +
+						"lots of built-in functionality like reading logs or forwarding ports via the API.",
+				},
+			},
+			RoleRef: rbac.RoleRef{
+				APIGroup: rbac.GroupName,
+				Kind:     "ClusterRole",
+				Name:     "system:kubelet-api-admin",
+			},
+			Subjects: []rbac.Subject{
+				{
+					APIGroup: rbac.GroupName,
+					Kind:     "User",
+					// TODO(q3k): describe this name's contract, or unify with whatever creates this.
+					Name: "smalltown:apiserver-kubelet-client",
+				},
+			},
+		},
+	}
+}
diff --git a/metropolis/node/kubernetes/reconciler/resources_runtimeclass.go b/metropolis/node/kubernetes/reconciler/resources_runtimeclass.go
new file mode 100644
index 0000000..c202c0e
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/resources_runtimeclass.go
@@ -0,0 +1,69 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package reconciler
+
+import (
+	"context"
+
+	node "k8s.io/api/node/v1beta1"
+	meta "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
+)
+
+type resourceRuntimeClasses struct {
+	kubernetes.Interface
+}
+
+func (r resourceRuntimeClasses) List(ctx context.Context) ([]string, error) {
+	res, err := r.NodeV1beta1().RuntimeClasses().List(ctx, listBuiltins)
+	if err != nil {
+		return nil, err
+	}
+	objs := make([]string, len(res.Items))
+	for i, el := range res.Items {
+		objs[i] = el.ObjectMeta.Name
+	}
+	return objs, nil
+}
+
+func (r resourceRuntimeClasses) Create(ctx context.Context, el interface{}) error {
+	_, err := r.NodeV1beta1().RuntimeClasses().Create(ctx, el.(*node.RuntimeClass), meta.CreateOptions{})
+	return err
+}
+
+func (r resourceRuntimeClasses) Delete(ctx context.Context, name string) error {
+	return r.NodeV1beta1().RuntimeClasses().Delete(ctx, name, meta.DeleteOptions{})
+}
+
+func (r resourceRuntimeClasses) Expected() map[string]interface{} {
+	return map[string]interface{}{
+		"gvisor": &node.RuntimeClass{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   "gvisor",
+				Labels: builtinLabels(nil),
+			},
+			Handler: "runsc",
+		},
+		"runc": &node.RuntimeClass{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   "runc",
+				Labels: builtinLabels(nil),
+			},
+			Handler: "runc",
+		},
+	}
+}
diff --git a/metropolis/node/kubernetes/reconciler/resources_storageclass.go b/metropolis/node/kubernetes/reconciler/resources_storageclass.go
new file mode 100644
index 0000000..72476ec
--- /dev/null
+++ b/metropolis/node/kubernetes/reconciler/resources_storageclass.go
@@ -0,0 +1,72 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package reconciler
+
+import (
+	"context"
+
+	core "k8s.io/api/core/v1"
+	storage "k8s.io/api/storage/v1"
+	meta "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
+)
+
+var reclaimPolicyDelete = core.PersistentVolumeReclaimDelete
+var waitForConsumerBinding = storage.VolumeBindingWaitForFirstConsumer
+
+type resourceStorageClasses struct {
+	kubernetes.Interface
+}
+
+func (r resourceStorageClasses) List(ctx context.Context) ([]string, error) {
+	res, err := r.StorageV1().StorageClasses().List(ctx, listBuiltins)
+	if err != nil {
+		return nil, err
+	}
+	objs := make([]string, len(res.Items))
+	for i, el := range res.Items {
+		objs[i] = el.ObjectMeta.Name
+	}
+	return objs, nil
+}
+
+func (r resourceStorageClasses) Create(ctx context.Context, el interface{}) error {
+	_, err := r.StorageV1().StorageClasses().Create(ctx, el.(*storage.StorageClass), meta.CreateOptions{})
+	return err
+}
+
+func (r resourceStorageClasses) Delete(ctx context.Context, name string) error {
+	return r.StorageV1().StorageClasses().Delete(ctx, name, meta.DeleteOptions{})
+}
+
+func (r resourceStorageClasses) Expected() map[string]interface{} {
+	return map[string]interface{}{
+		"local": &storage.StorageClass{
+			ObjectMeta: meta.ObjectMeta{
+				Name:   "local",
+				Labels: builtinLabels(nil),
+				Annotations: map[string]string{
+					"storageclass.kubernetes.io/is-default-class": "true",
+				},
+			},
+			AllowVolumeExpansion: True(),
+			Provisioner:          csiProvisionerName,
+			ReclaimPolicy:        &reclaimPolicyDelete,
+			VolumeBindingMode:    &waitForConsumerBinding,
+		},
+	}
+}
diff --git a/metropolis/node/kubernetes/scheduler.go b/metropolis/node/kubernetes/scheduler.go
new file mode 100644
index 0000000..21e6663
--- /dev/null
+++ b/metropolis/node/kubernetes/scheduler.go
@@ -0,0 +1,70 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kubernetes
+
+import (
+	"context"
+	"encoding/pem"
+	"fmt"
+	"os/exec"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/fileargs"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/pki"
+)
+
+type schedulerConfig struct {
+	kubeConfig []byte
+	serverCert []byte
+	serverKey  []byte
+}
+
+func getPKISchedulerConfig(ctx context.Context, kpki *pki.KubernetesPKI) (*schedulerConfig, error) {
+	var config schedulerConfig
+	var err error
+	config.serverCert, config.serverKey, err = kpki.Certificate(ctx, pki.Scheduler)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get scheduler serving certificate: %w", err)
+	}
+	config.kubeConfig, err = kpki.Kubeconfig(ctx, pki.SchedulerClient)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get scheduler kubeconfig: %w", err)
+	}
+	return &config, nil
+}
+
+func runScheduler(config schedulerConfig) supervisor.Runnable {
+	return func(ctx context.Context) error {
+		args, err := fileargs.New()
+		if err != nil {
+			panic(err) // If this fails, something is very wrong. Just crash.
+		}
+		defer args.Close()
+		cmd := exec.CommandContext(ctx, "/kubernetes/bin/kube", "kube-scheduler",
+			args.FileOpt("--kubeconfig", "kubeconfig", config.kubeConfig),
+			"--port=0", // Kill insecure serving
+			args.FileOpt("--tls-cert-file", "server-cert.pem",
+				pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: config.serverCert})),
+			args.FileOpt("--tls-private-key-file", "server-key.pem",
+				pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: config.serverKey})),
+		)
+		if args.Error() != nil {
+			return fmt.Errorf("failed to use fileargs: %w", err)
+		}
+		return supervisor.RunCommand(ctx, cmd)
+	}
+}
diff --git a/metropolis/node/kubernetes/service.go b/metropolis/node/kubernetes/service.go
new file mode 100644
index 0000000..2917bfc
--- /dev/null
+++ b/metropolis/node/kubernetes/service.go
@@ -0,0 +1,177 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kubernetes
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"os"
+	"time"
+
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/status"
+	"k8s.io/client-go/informers"
+	"k8s.io/client-go/kubernetes"
+	"k8s.io/client-go/tools/clientcmd"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/localstorage"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dns"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/clusternet"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/nfproxy"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/pki"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/kubernetes/reconciler"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+type Config struct {
+	AdvertiseAddress net.IP
+	ServiceIPRange   net.IPNet
+	ClusterNet       net.IPNet
+
+	KPKI                    *pki.KubernetesPKI
+	Root                    *localstorage.Root
+	CorednsRegistrationChan chan *dns.ExtraDirective
+}
+
+type Service struct {
+	c Config
+}
+
+func New(c Config) *Service {
+	s := &Service{
+		c: c,
+	}
+	return s
+}
+
+func (s *Service) Run(ctx context.Context) error {
+	controllerManagerConfig, err := getPKIControllerManagerConfig(ctx, s.c.KPKI)
+	if err != nil {
+		return fmt.Errorf("could not generate controller manager pki config: %w", err)
+	}
+	controllerManagerConfig.clusterNet = s.c.ClusterNet
+	schedulerConfig, err := getPKISchedulerConfig(ctx, s.c.KPKI)
+	if err != nil {
+		return fmt.Errorf("could not generate scheduler pki config: %w", err)
+	}
+
+	masterKubeconfig, err := s.c.KPKI.Kubeconfig(ctx, pki.Master)
+	if err != nil {
+		return fmt.Errorf("could not generate master kubeconfig: %w", err)
+	}
+
+	rawClientConfig, err := clientcmd.NewClientConfigFromBytes(masterKubeconfig)
+	if err != nil {
+		return fmt.Errorf("could not generate kubernetes client config: %w", err)
+	}
+
+	clientConfig, err := rawClientConfig.ClientConfig()
+	clientSet, err := kubernetes.NewForConfig(clientConfig)
+	if err != nil {
+		return fmt.Errorf("could not generate kubernetes client: %w", err)
+	}
+
+	informerFactory := informers.NewSharedInformerFactory(clientSet, 5*time.Minute)
+
+	hostname, err := os.Hostname()
+	if err != nil {
+		return fmt.Errorf("failed to get hostname: %w", err)
+	}
+
+	dnsHostIP := s.c.AdvertiseAddress // TODO: Which IP to use
+
+	apiserver := &apiserverService{
+		KPKI:                        s.c.KPKI,
+		AdvertiseAddress:            s.c.AdvertiseAddress,
+		ServiceIPRange:              s.c.ServiceIPRange,
+		EphemeralConsensusDirectory: &s.c.Root.Ephemeral.Consensus,
+	}
+
+	kubelet := kubeletService{
+		NodeName:           hostname,
+		ClusterDNS:         []net.IP{dnsHostIP},
+		KubeletDirectory:   &s.c.Root.Data.Kubernetes.Kubelet,
+		EphemeralDirectory: &s.c.Root.Ephemeral,
+		KPKI:               s.c.KPKI,
+	}
+
+	csiPlugin := csiPluginServer{
+		KubeletDirectory: &s.c.Root.Data.Kubernetes.Kubelet,
+		VolumesDirectory: &s.c.Root.Data.Volumes,
+	}
+
+	csiProvisioner := csiProvisionerServer{
+		NodeName:         hostname,
+		Kubernetes:       clientSet,
+		InformerFactory:  informerFactory,
+		VolumesDirectory: &s.c.Root.Data.Volumes,
+	}
+
+	clusternet := clusternet.Service{
+		NodeName:        hostname,
+		Kubernetes:      clientSet,
+		ClusterNet:      s.c.ClusterNet,
+		InformerFactory: informerFactory,
+		DataDirectory:   &s.c.Root.Data.Kubernetes.ClusterNetworking,
+	}
+
+	nfproxy := nfproxy.Service{
+		ClusterCIDR: s.c.ClusterNet,
+		ClientSet:   clientSet,
+	}
+
+	for _, sub := range []struct {
+		name     string
+		runnable supervisor.Runnable
+	}{
+		{"apiserver", apiserver.Run},
+		{"controller-manager", runControllerManager(*controllerManagerConfig)},
+		{"scheduler", runScheduler(*schedulerConfig)},
+		{"kubelet", kubelet.Run},
+		{"reconciler", reconciler.Run(clientSet)},
+		{"csi-plugin", csiPlugin.Run},
+		{"csi-provisioner", csiProvisioner.Run},
+		{"clusternet", clusternet.Run},
+		{"nfproxy", nfproxy.Run},
+	} {
+		err := supervisor.Run(ctx, sub.name, sub.runnable)
+		if err != nil {
+			return fmt.Errorf("could not run sub-service %q: %w", sub.name, err)
+		}
+	}
+
+	supervisor.Logger(ctx).Info("Registering K8s CoreDNS")
+	clusterDNSDirective := dns.NewKubernetesDirective("cluster.local", masterKubeconfig)
+	s.c.CorednsRegistrationChan <- clusterDNSDirective
+
+	supervisor.Signal(ctx, supervisor.SignalHealthy)
+	<-ctx.Done()
+	s.c.CorednsRegistrationChan <- dns.CancelDirective(clusterDNSDirective)
+	return nil
+}
+
+// GetDebugKubeconfig issues a kubeconfig for an arbitrary given identity. Useful for debugging and testing.
+func (s *Service) GetDebugKubeconfig(ctx context.Context, request *apb.GetDebugKubeconfigRequest) (*apb.GetDebugKubeconfigResponse, error) {
+	ca := s.c.KPKI.Certificates[pki.IdCA]
+	debugKubeconfig, err := pki.New(ca, "", pki.Client(request.Id, request.Groups)).Kubeconfig(ctx, s.c.KPKI.KV)
+	if err != nil {
+		return nil, status.Errorf(codes.Unavailable, "Failed to generate kubeconfig: %v", err)
+	}
+	return &apb.GetDebugKubeconfigResponse{DebugKubeconfig: string(debugKubeconfig)}, nil
+}
diff --git a/metropolis/node/ports.go b/metropolis/node/ports.go
new file mode 100644
index 0000000..c63ec38
--- /dev/null
+++ b/metropolis/node/ports.go
@@ -0,0 +1,28 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package node
+
+const (
+	NodeServicePort     = 7835
+	ConsensusPort       = 7834
+	MasterServicePort   = 7833
+	ExternalServicePort = 7836
+	DebugServicePort    = 7837
+	WireGuardPort       = 7838
+	KubernetesAPIPort   = 6443
+	DebuggerPort        = 2345
+)
diff --git a/metropolis/proto/api/BUILD.bazel b/metropolis/proto/api/BUILD.bazel
new file mode 100644
index 0000000..ec40562
--- /dev/null
+++ b/metropolis/proto/api/BUILD.bazel
@@ -0,0 +1,27 @@
+load("@rules_proto//proto:defs.bzl", "proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+
+proto_library(
+    name = "api_proto",
+    srcs = [
+        "debug.proto",
+        "enrolment.proto",
+    ],
+    visibility = ["//visibility:public"],
+)
+
+go_proto_library(
+    name = "api_go_proto",
+    compilers = ["@io_bazel_rules_go//proto:go_grpc"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/proto/api",
+    proto = ":api_proto",
+    visibility = ["//visibility:public"],
+)
+
+go_library(
+    name = "go_default_library",
+    embed = [":api_go_proto"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/proto/api",
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/proto/api/debug.proto b/metropolis/proto/api/debug.proto
new file mode 100644
index 0000000..2483360
--- /dev/null
+++ b/metropolis/proto/api/debug.proto
@@ -0,0 +1,164 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+package smalltown.core.proto.api;
+option go_package = "git.monogon.dev/source/nexantic.git/metropolis/proto/api";
+
+import "metropolis/proto/api/enrolment.proto";
+
+// NodeDebugService exposes debug and testing endpoints that allow introspection into a running Smalltown instance.
+// It is not authenticated and will be disabled in production. It is currently consumed by metropolis/cli/dbg and
+// by tests.
+service NodeDebugService {
+    // GetDebugKubeconfig issues kubeconfigs with arbitrary identities and groups for debugging
+    rpc GetDebugKubeconfig(GetDebugKubeconfigRequest) returns (GetDebugKubeconfigResponse);
+
+    // GetLogs Returns historical and/or streaming logs for a given DN with given filters from the system global
+    // LogTree.
+    //
+    // For more information about this API, see //metropolis/node/core/logtree. But, in summary:
+    //   - All logging is performed to a DN (distinguished name), which is a dot-delimited string like foo.bar.baz.
+    //   - Log entries can be either raw (coming from unstructured logging from an external service, like a running
+    //     process) or leveled (emitted by Smalltown code with a source line, timestamp, and severity).
+    //   - The DNs form a tree of logging nodes - and when requesting logs, a given subtree of DNs can be requested,
+    //     instead of just a given DN.
+    //   - All supervised processes live at `root.<supervisor DN>`. For more example paths, see the console logs of
+    //     a running Smalltown instance, or request all logs (at DN "").
+    //
+    // TODO(q3k): move method and its related messages to the non-debug node endpoint once we have one.
+    rpc GetLogs(GetLogsRequest) returns (stream GetLogsResponse);
+
+    // GetGoldenTicket requests a 'golden ticket' which can be used to enroll any node into the cluster.
+    // This bypasses integrity checks.
+    rpc GetGoldenTicket(GetGoldenTicketRequest) returns (GetGoldenTicketResponse);
+}
+
+
+message GetDebugKubeconfigRequest {
+    string id = 1; // Kubernetes identity (user)
+    repeated string groups = 2; // Kubernetes groups
+}
+
+message GetDebugKubeconfigResponse {
+    string debug_kubeconfig = 1;
+}
+
+// Severity level corresponding to //metropolis/node/core/logtree.Severity.
+enum LeveledLogSeverity {
+    INVALID = 0;
+    INFO = 1;
+    WARNING = 2;
+    ERROR = 3;
+    FATAL = 4;
+}
+
+// Filter set when requesting logs for a given DN. This message is equivalent to the following GADT enum:
+// data LogFilter = WithChildren
+//                | OnlyRaw
+//                | OnlyLeveled
+//                | LeveledWithMinimumSeverity(Severity)
+//
+// Multiple LogFilters can be chained/combined when requesting logs, as long as they do not conflict.
+message LogFilter {
+    // Entries will be returned not only for the given DN, but all child DNs as well. For instance, if the
+    // requested DN is foo, entries logged to foo, foo.bar and foo.bar.baz will all be returned.
+    message WithChildren {
+    }
+    // Only raw logging entries will be returned. Conflicts with OnlyLeveled filters.
+    message OnlyRaw {
+    }
+    // Only leveled logging entries will be returned. Conflicts with OnlyRaw filters.
+    message OnlyLeveled {
+    }
+    // If leveled logs are returned, all entries at severity lower than `minimum` will be discarded.
+    message LeveledWithMinimumSeverity {
+        LeveledLogSeverity minimum = 1;
+    }
+    oneof filter {
+        WithChildren with_children = 1;
+        OnlyRaw only_raw = 3;
+        OnlyLeveled only_leveled = 4;
+        LeveledWithMinimumSeverity leveled_with_minimum_severity = 5;
+    }
+}
+
+message GetLogsRequest {
+    // DN from which to request logs. All supervised runnables live at `root.`, the init code lives at `init.`.
+    string dn = 1;
+    // Filters to apply to returned data.
+    repeated LogFilter filters = 2;
+
+    enum BacklogMode {
+        BACKLOG_INVALID = 0;
+        // No historic data will be returned.
+        BACKLOG_DISABLE = 1;
+        // All available historic data will be returned.
+        BACKLOG_ALL = 2;
+        // At most backlog_count entries will be returned, if available.
+        BACKLOG_COUNT = 3;
+    }
+    BacklogMode backlog_mode = 3;
+    int64 backlog_count = 4;
+
+    enum StreamMode {
+        STREAM_INVALID = 0;
+        // No streaming entries, gRPC stream will be closed as soon as all backlog data is served.
+        STREAM_DISABLE = 1;
+        // Entries will be streamed as early as available right after all backlog data is served.
+        STREAM_UNBUFFERED = 2;
+    }
+    StreamMode stream_mode = 5;
+}
+
+message GetLogsResponse {
+    // Entries from the requested historical entries (via WithBackLog). They will all be served before the first
+    // stream_entries are served (if any).
+    repeated LogEntry backlog_entries = 1;
+    // Entries streamed as they arrive. Currently no server-side buffering is enabled, instead every line is served
+    // as early as it arrives. However, this might change in the future, so this behaviour cannot be depended
+    // upon.
+    repeated LogEntry stream_entries = 2;
+}
+
+message LogEntry {
+    message Leveled {
+        repeated string lines = 1;
+        int64 timestamp = 2;
+        LeveledLogSeverity severity = 3;
+        string location = 4;
+    }
+    message Raw {
+        string data = 1;
+        int64 original_length = 2;
+    }
+
+    string dn = 1;
+    oneof kind {
+        Leveled leveled = 2;
+        Raw raw = 3;
+    }
+}
+
+message GetGoldenTicketRequest {
+    // IP address at which the new node will run.
+    string external_ip = 1;
+}
+
+message GetGoldenTicketResponse {
+    // Ticket to use in the new node's EnrolmentConfig.
+    GoldenTicket ticket = 1;
+}
diff --git a/metropolis/proto/api/enrolment.proto b/metropolis/proto/api/enrolment.proto
new file mode 100644
index 0000000..c07e4ea
--- /dev/null
+++ b/metropolis/proto/api/enrolment.proto
@@ -0,0 +1,63 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+package smalltown.core.proto.api;
+option go_package = "git.monogon.dev/source/nexantic.git/metropolis/proto/api";
+
+// EnrolmentConfig is the single boot configuration file contained in the Smalltown ESP. It configures
+// the way the node will start up (what cluster it will join/enroll into/create).
+message EnrolmentConfig {
+    // Debug/temporary cluster enrolment method. If set, the node will attempt to enroll into the
+    // cluster that this ticket was generated for. Otherwise, a new cluster will be created.
+    GoldenTicket golden_ticket = 1;
+
+    // Filled in by node after it is enrolled
+    string node_id = 2;
+}
+
+// GoldenTicket is a ticket that allows any node to enroll into a cluster, bypassing any integrity
+// checks.
+//
+// Currently, enrolling into a cluster does not use a TPM-based workflow, and instead
+// bases on a simplified workflow of joining consensus by being started with a
+// TLS client certificate. This is a short-circuit fix to allow multi-node
+// clusters for testing before we design the final cluster node lifecycle system.
+message GoldenTicket {
+    // Etcd peer CA certificate.
+    bytes etcd_ca_cert = 1;
+    // Etcd peer client certificate.
+    bytes etcd_client_cert = 2;
+    // Etcd peer client key.
+    bytes etcd_client_key = 3;
+    // Initial etcd peer CRL.
+    bytes etcd_crl = 4;
+
+    message EtcdPeer {
+        string name = 1;
+        string address = 2;
+    }
+    // All other current etcd peers in the cluster.
+    repeated EtcdPeer peers = 5;
+    // The peer that this node should start running.
+    EtcdPeer this = 6;
+
+    // Node configuration. Currently unused (in the future, this will be used to run a node
+    // management service separate from etcd clustering).
+    string node_id = 7;
+    bytes node_cert = 8;
+    bytes node_key = 9;
+}
diff --git a/metropolis/proto/common/BUILD.bazel b/metropolis/proto/common/BUILD.bazel
new file mode 100644
index 0000000..882bb61
--- /dev/null
+++ b/metropolis/proto/common/BUILD.bazel
@@ -0,0 +1,23 @@
+load("@rules_proto//proto:defs.bzl", "proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+
+proto_library(
+    name = "common_proto",
+    srcs = ["common.proto"],
+    visibility = ["//visibility:public"],
+)
+
+go_proto_library(
+    name = "common_go_proto",
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/proto/common",
+    proto = ":common_proto",
+    visibility = ["//visibility:public"],
+)
+
+go_library(
+    name = "go_default_library",
+    embed = [":common_go_proto"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/proto/common",
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/proto/common/common.proto b/metropolis/proto/common/common.proto
new file mode 100644
index 0000000..def91e8
--- /dev/null
+++ b/metropolis/proto/common/common.proto
@@ -0,0 +1,24 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+package smalltown.core.proto.common;
+option go_package = "git.monogon.dev/source/nexantic.git/metropolis/proto/common";
+
+enum TrustBackend {
+    DUMMY = 0;
+    TPM = 1;
+}
diff --git a/metropolis/proto/internal/BUILD.bazel b/metropolis/proto/internal/BUILD.bazel
new file mode 100644
index 0000000..2951a08
--- /dev/null
+++ b/metropolis/proto/internal/BUILD.bazel
@@ -0,0 +1,23 @@
+load("@rules_proto//proto:defs.bzl", "proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+
+proto_library(
+    name = "internal_proto",
+    srcs = ["internal.proto"],
+    visibility = ["//metropolis:__subpackages__"],
+)
+
+go_proto_library(
+    name = "internal_go_proto",
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/proto/internal",
+    proto = ":internal_proto",
+    visibility = ["//metropolis:__subpackages__"],
+)
+
+go_library(
+    name = "go_default_library",
+    embed = [":internal_go_proto"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/proto/internal",
+    visibility = ["//metropolis:__subpackages__"],
+)
diff --git a/metropolis/proto/internal/internal.proto b/metropolis/proto/internal/internal.proto
new file mode 100644
index 0000000..c5aa259
--- /dev/null
+++ b/metropolis/proto/internal/internal.proto
@@ -0,0 +1,43 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+option go_package = "git.monogon.dev/source/nexantic.git/metropolis/proto/internal";
+package smalltown.core.proto.internal;
+
+// Node describes a single node's state in etcd
+message Node {
+    // Individual node service certificate.
+    bytes certificate = 1;
+    // Node's individual cluster part of the disk encryption key.
+    bytes cluster_unlock_key = 2;
+    // Node address. This is currently an IPv4 address because that is what the
+    // network service uses, but would also support IPv6. See EnrolmentConfig for a
+    // discussion on why we use IP addresses rather than some kind of resolver.
+    string address = 3;
+
+    message Roles {
+        message ConsensusMember {
+            string etcd_member_name = 1;
+        }
+        ConsensusMember consensus_member = 1;
+        message KubernetesWorker {
+            string node_name = 1;
+        }
+        KubernetesWorker kubernetes_worker = 2;
+    }
+    Roles roles = 4;
+}
diff --git a/metropolis/test/e2e/BUILD.bazel b/metropolis/test/e2e/BUILD.bazel
new file mode 100644
index 0000000..b9bb6f8
--- /dev/null
+++ b/metropolis/test/e2e/BUILD.bazel
@@ -0,0 +1,43 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "go_default_library",
+    srcs = [
+        "kubernetes_helpers.go",
+        "utils.go",
+    ],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/e2e",
+    visibility = ["//metropolis/test:__subpackages__"],
+    deps = [
+        "//metropolis/proto/api:go_default_library",
+        "@io_k8s_api//apps/v1:go_default_library",
+        "@io_k8s_api//core/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/api/resource:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/util/intstr:go_default_library",
+        "@io_k8s_client_go//kubernetes:go_default_library",
+        "@io_k8s_client_go//tools/clientcmd:go_default_library",
+    ],
+)
+
+go_test(
+    name = "go_default_test",
+    size = "large",
+    srcs = ["main_test.go"],
+    data = [
+        "//metropolis/node:image",
+        "//metropolis/node:swtpm_data",
+        "//third_party/edk2:firmware",
+    ],
+    embed = [":go_default_library"],
+    rundir = ".",
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+        "//metropolis/test/launch:go_default_library",
+        "@io_k8s_api//core/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+        "@io_k8s_kubernetes//pkg/api/v1/pod:go_default_library",
+        "@org_golang_google_grpc//:go_default_library",
+    ],
+)
diff --git a/metropolis/test/e2e/k8s_cts/BUILD.bazel b/metropolis/test/e2e/k8s_cts/BUILD.bazel
new file mode 100644
index 0000000..0e43c24
--- /dev/null
+++ b/metropolis/test/e2e/k8s_cts/BUILD.bazel
@@ -0,0 +1,55 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+load("@io_bazel_rules_docker//go:image.bzl", "go_image")
+load("@io_bazel_rules_docker//container:container.bzl", "container_image")
+
+go_image(
+    name = "kubectl",
+    binary = "@io_k8s_kubernetes//cmd/kubectl",
+    pure = "on",
+)
+
+container_image(
+    name = "kubectl_in_path",
+    base = ":kubectl",
+    env = {
+        # Don't include FHS paths since they aren't available anyways
+        "PATH": "/app/cmd/kubectl",
+    },
+)
+
+go_image(
+    name = "k8s_cts_image",
+    base = ":kubectl_in_path",
+    binary = "@io_k8s_kubernetes//test/e2e:_go_default_test-pure",
+    pure = "on",
+    visibility = ["//visibility:public"],
+)
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/e2e/k8s_cts",
+    visibility = ["//visibility:private"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/test/e2e:go_default_library",
+        "//metropolis/test/launch:go_default_library",
+        "@io_k8s_api//core/v1:go_default_library",
+        "@io_k8s_api//rbac/v1:go_default_library",
+        "@io_k8s_apimachinery//pkg/apis/meta/v1:go_default_library",
+    ],
+)
+
+go_binary(
+    name = "k8s_cts",
+    data = [
+        "//metropolis/node:image",
+        "//metropolis/node:swtpm_data",
+        "//metropolis/test/nanoswitch:initramfs",
+        "//metropolis/test/ktest:linux-testing",
+        "//third_party/edk2:firmware",
+        "@com_github_bonzini_qboot//:qboot-bin",
+    ],
+    embed = [":go_default_library"],
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/test/e2e/k8s_cts/main.go b/metropolis/test/e2e/k8s_cts/main.go
new file mode 100644
index 0000000..728a890
--- /dev/null
+++ b/metropolis/test/e2e/k8s_cts/main.go
@@ -0,0 +1,175 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// This package launches a Smalltown Cluster with two nodes and spawns in the CTS container. Then it streams its output
+// to the console. When the CTS has finished it exits with the appropriate error code.
+package main
+
+import (
+	"context"
+	"io"
+	"log"
+	"os"
+	"os/signal"
+	"strings"
+	"syscall"
+	"time"
+
+	corev1 "k8s.io/api/core/v1"
+	rbacv1 "k8s.io/api/rbac/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/test/e2e"
+	"git.monogon.dev/source/nexantic.git/metropolis/test/launch"
+)
+
+// makeCTSPodSpec generates a spec for a standalone pod running the Kubernetes CTS. It also sets the test configuration
+// for the Kubernetes E2E test suite to only run CTS tests and excludes known-broken ones.
+func makeCTSPodSpec(name string, saName string) *corev1.Pod {
+	skipRegexes := []string{
+		// hostNetworking cannot be supported since we run different network stacks for the host and containers
+		"should function for node-pod communication",
+		// gVisor misreports statfs() syscalls: https://github.com/google/gvisor/issues/3339
+		`should support \((non-)?root,`,
+		"volume on tmpfs should have the correct mode",
+		"volume on default medium should have the correct mode",
+		// gVisor doesn't support the full Linux privilege machinery including SUID and NewPrivs
+		// https://github.com/google/gvisor/issues/189#issuecomment-481064000
+		"should run the container as unprivileged when false",
+	}
+	return &corev1.Pod{
+		ObjectMeta: metav1.ObjectMeta{
+			Name: name,
+			Labels: map[string]string{
+				"name": name,
+			},
+		},
+		Spec: corev1.PodSpec{
+			Containers: []corev1.Container{
+				{
+					Name:  "cts",
+					Image: "bazel/metropolis/test/e2e/k8s_cts:k8s_cts_image",
+					Args: []string{
+						"-cluster-ip-range=10.0.0.0/17",
+						"-dump-systemd-journal=false",
+						"-ginkgo.focus=\\[Conformance\\]",
+						"-ginkgo.skip=" + strings.Join(skipRegexes, "|"),
+						"-test.parallel=8",
+					},
+					ImagePullPolicy: corev1.PullNever,
+				},
+			},
+			Tolerations: []corev1.Toleration{{ // Tolerate all taints, otherwise the CTS likes to self-evict
+				Operator: "Exists",
+			}},
+			PriorityClassName:  "system-cluster-critical", // Don't evict the CTS pod
+			RestartPolicy:      corev1.RestartPolicyNever,
+			ServiceAccountName: saName,
+		},
+	}
+}
+
+func main() {
+	sigs := make(chan os.Signal, 1)
+	signal.Notify(sigs, syscall.SIGINT, syscall.SIGTERM)
+	ctx, cancel := context.WithCancel(context.Background())
+	go func() {
+		<-sigs
+		cancel()
+	}()
+
+	debugClient, portMap, err := launch.LaunchCluster(ctx, launch.ClusterOptions{NumNodes: 2})
+	if err != nil {
+		log.Fatalf("Failed to launch cluster: %v", err)
+	}
+	log.Println("Cluster initialized")
+
+	clientSet, err := e2e.GetKubeClientSet(ctx, debugClient, portMap[common.KubernetesAPIPort])
+	if err != nil {
+		log.Fatalf("Failed to get clientSet: %v", err)
+	}
+	log.Println("Credentials available")
+
+	saName := "cts"
+	ctsSA := &corev1.ServiceAccount{ObjectMeta: metav1.ObjectMeta{Name: saName}}
+	for {
+		if _, err := clientSet.CoreV1().ServiceAccounts("default").Create(ctx, ctsSA, metav1.CreateOptions{}); err != nil {
+			log.Printf("Failed to create ServiceAccount: %v", err)
+			time.Sleep(1 * time.Second)
+			continue
+		}
+		break
+	}
+	ctsRoleBinding := &rbacv1.ClusterRoleBinding{
+		ObjectMeta: metav1.ObjectMeta{
+			Name: saName,
+		},
+		Subjects: []rbacv1.Subject{
+			{
+				Namespace: "default",
+				Name:      saName,
+				Kind:      rbacv1.ServiceAccountKind,
+			},
+		},
+		RoleRef: rbacv1.RoleRef{
+			Kind: "ClusterRole",
+			Name: "cluster-admin",
+		},
+	}
+	podName := "cts"
+	if _, err := clientSet.RbacV1().ClusterRoleBindings().Create(ctx, ctsRoleBinding, metav1.CreateOptions{}); err != nil {
+		log.Fatalf("Failed to create ClusterRoleBinding: %v", err)
+	}
+	for {
+		if _, err := clientSet.CoreV1().Pods("default").Create(ctx, makeCTSPodSpec(podName, saName), metav1.CreateOptions{}); err != nil {
+			log.Printf("Failed to create Pod: %v", err)
+			time.Sleep(1 * time.Second)
+			continue
+		}
+		break
+	}
+	var logs io.ReadCloser
+	go func() {
+		// This loops the whole .Stream()/io.Copy process because the API sometimes returns streams that immediately return EOF
+		for {
+			logs, err = clientSet.CoreV1().Pods("default").GetLogs(podName, &corev1.PodLogOptions{Follow: true}).Stream(ctx)
+			if err == nil {
+				if _, err := io.Copy(os.Stdout, logs); err != nil {
+					log.Printf("Log pump error: %v", err)
+				}
+				logs.Close()
+			} else {
+				log.Printf("Pod logs not ready yet: %v", err)
+			}
+			time.Sleep(1 * time.Second)
+		}
+	}()
+	for {
+		time.Sleep(1 * time.Second)
+		pod, err := clientSet.CoreV1().Pods("default").Get(ctx, podName, metav1.GetOptions{})
+		if err != nil {
+			log.Printf("Failed to get CTS pod: %v", err)
+			continue
+		}
+		if pod.Status.Phase == corev1.PodSucceeded {
+			return
+		}
+		if pod.Status.Phase == corev1.PodFailed {
+			log.Fatalf("CTS failed")
+		}
+	}
+}
diff --git a/metropolis/test/e2e/kubernetes_helpers.go b/metropolis/test/e2e/kubernetes_helpers.go
new file mode 100644
index 0000000..fffbd1b
--- /dev/null
+++ b/metropolis/test/e2e/kubernetes_helpers.go
@@ -0,0 +1,146 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package e2e
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"time"
+
+	appsv1 "k8s.io/api/apps/v1"
+	corev1 "k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/api/resource"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/intstr"
+	"k8s.io/client-go/kubernetes"
+	"k8s.io/client-go/tools/clientcmd"
+
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+// GetKubeClientSet gets a Kubeconfig from the debug API and creates a K8s ClientSet using it. The identity used has
+// the system:masters group and thus has RBAC access to everything.
+func GetKubeClientSet(ctx context.Context, client apb.NodeDebugServiceClient, port uint16) (kubernetes.Interface, error) {
+	var lastErr = errors.New("context canceled before any operation completed")
+	for {
+		reqT, cancel := context.WithTimeout(ctx, 5*time.Second)
+		defer cancel()
+		res, err := client.GetDebugKubeconfig(reqT, &apb.GetDebugKubeconfigRequest{Id: "debug-user", Groups: []string{"system:masters"}})
+		if err == nil {
+			rawClientConfig, err := clientcmd.NewClientConfigFromBytes([]byte(res.DebugKubeconfig))
+			if err != nil {
+				return nil, err // Invalid Kubeconfigs are immediately fatal
+			}
+
+			clientConfig, err := rawClientConfig.ClientConfig()
+			clientConfig.Host = fmt.Sprintf("localhost:%v", port)
+			clientSet, err := kubernetes.NewForConfig(clientConfig)
+			if err != nil {
+				return nil, err
+			}
+			return clientSet, nil
+		}
+		if err != nil && err == ctx.Err() {
+			return nil, lastErr
+		}
+		lastErr = err
+		select {
+		case <-ctx.Done():
+			return nil, lastErr
+		case <-time.After(1 * time.Second):
+		}
+	}
+}
+
+// makeTestDeploymentSpec generates a Deployment spec for a single pod running NGINX with a readiness probe. This allows
+// verifying that the control plane is capable of scheduling simple pods and that kubelet works, its runtime is set up
+// well enough to run a simple container and the network to the pod can pass readiness probe traffic.
+func makeTestDeploymentSpec(name string) *appsv1.Deployment {
+	return &appsv1.Deployment{
+		ObjectMeta: metav1.ObjectMeta{Name: name},
+		Spec: appsv1.DeploymentSpec{
+			Selector: &metav1.LabelSelector{MatchLabels: map[string]string{
+				"name": name,
+			}},
+			Template: corev1.PodTemplateSpec{
+				ObjectMeta: metav1.ObjectMeta{
+					Labels: map[string]string{
+						"name": name,
+					},
+				},
+				Spec: corev1.PodSpec{
+					Containers: []corev1.Container{
+						{
+							Name: "test",
+							// TODO(phab/T793): Build and preseed our own container images
+							Image: "nginx:alpine",
+							ReadinessProbe: &corev1.Probe{
+								Handler: corev1.Handler{
+									HTTPGet: &corev1.HTTPGetAction{Port: intstr.FromInt(80)},
+								},
+							},
+						},
+					},
+				},
+			},
+		},
+	}
+}
+
+// makeTestStatefulSet generates a StatefulSet spec
+func makeTestStatefulSet(name string) *appsv1.StatefulSet {
+	return &appsv1.StatefulSet{
+		ObjectMeta: metav1.ObjectMeta{Name: name},
+		Spec: appsv1.StatefulSetSpec{
+			Selector: &metav1.LabelSelector{MatchLabels: map[string]string{
+				"name": name,
+			}},
+			VolumeClaimTemplates: []corev1.PersistentVolumeClaim{
+				{
+					ObjectMeta: metav1.ObjectMeta{Name: "www"},
+					Spec: corev1.PersistentVolumeClaimSpec{
+						AccessModes: []corev1.PersistentVolumeAccessMode{corev1.ReadWriteOnce},
+						Resources: corev1.ResourceRequirements{
+							Requests: map[corev1.ResourceName]resource.Quantity{corev1.ResourceStorage: resource.MustParse("50Mi")},
+						},
+					},
+				},
+			},
+			Template: corev1.PodTemplateSpec{
+				ObjectMeta: metav1.ObjectMeta{
+					Labels: map[string]string{
+						"name": name,
+					},
+				},
+				Spec: corev1.PodSpec{
+					Containers: []corev1.Container{
+						{
+							Name:  "test",
+							Image: "nginx:alpine",
+							ReadinessProbe: &corev1.Probe{
+								Handler: corev1.Handler{
+									HTTPGet: &corev1.HTTPGetAction{Port: intstr.FromInt(80)},
+								},
+							},
+						},
+					},
+				},
+			},
+		},
+	}
+}
diff --git a/metropolis/test/e2e/main_test.go b/metropolis/test/e2e/main_test.go
new file mode 100644
index 0000000..46e862c
--- /dev/null
+++ b/metropolis/test/e2e/main_test.go
@@ -0,0 +1,244 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package e2e
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log"
+	"net"
+	"net/http"
+	_ "net/http"
+	_ "net/http/pprof"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	"google.golang.org/grpc"
+	corev1 "k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	podv1 "k8s.io/kubernetes/pkg/api/v1/pod"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+	"git.monogon.dev/source/nexantic.git/metropolis/test/launch"
+)
+
+const (
+	// Timeout for the global test context.
+	//
+	// Bazel would eventually time out the test after 900s ("large") if, for some reason,
+	// the context cancellation fails to abort it.
+	globalTestTimeout = 600 * time.Second
+
+	// Timeouts for individual end-to-end tests of different sizes.
+	smallTestTimeout = 60 * time.Second
+	largeTestTimeout = 120 * time.Second
+)
+
+// TestE2E is the main E2E test entrypoint for single-node freshly-bootstrapped E2E tests. It starts a full Smalltown node
+// in bootstrap mode and then runs tests against it. The actual tests it performs are located in the RunGroup subtest.
+func TestE2E(t *testing.T) {
+	// Run pprof server for debugging
+	go func() {
+		addr, err := net.ResolveTCPAddr("tcp", "localhost:0")
+		if err != nil {
+			panic(err)
+		}
+
+		l, err := net.ListenTCP("tcp", addr)
+		if err != nil {
+			log.Fatalf("Failed to listen on pprof port: %s", l.Addr())
+		}
+		defer l.Close()
+
+		log.Printf("pprof server listening on %s", l.Addr())
+		log.Printf("pprof server returned an error: %v", http.Serve(l, nil))
+	}()
+
+	// Set a global timeout to make sure this terminates
+	ctx, cancel := context.WithTimeout(context.Background(), globalTestTimeout)
+	portMap, err := launch.ConflictFreePortMap(launch.NodePorts)
+	if err != nil {
+		t.Fatalf("Failed to acquire ports for e2e test: %v", err)
+	}
+
+	procExit := make(chan struct{})
+
+	go func() {
+		if err := launch.Launch(ctx, launch.Options{Ports: portMap, SerialPort: os.Stdout}); err != nil {
+			panic(err)
+		}
+		close(procExit)
+	}()
+	grpcClient, err := portMap.DialGRPC(common.DebugServicePort, grpc.WithInsecure())
+	if err != nil {
+		fmt.Printf("Failed to dial debug service (is it running): %v\n", err)
+	}
+	debugClient := apb.NewNodeDebugServiceClient(grpcClient)
+
+	// This exists to keep the parent around while all the children race
+	// It currently tests both a set of OS-level conditions and Kubernetes Deployments and StatefulSets
+	t.Run("RunGroup", func(t *testing.T) {
+		t.Run("Get Kubernetes Debug Kubeconfig", func(t *testing.T) {
+			t.Parallel()
+			selfCtx, cancel := context.WithTimeout(ctx, largeTestTimeout)
+			defer cancel()
+			clientSet, err := GetKubeClientSet(selfCtx, debugClient, portMap[common.KubernetesAPIPort])
+			if err != nil {
+				t.Fatal(err)
+			}
+			testEventual(t, "Node is registered and ready", ctx, largeTestTimeout, func(ctx context.Context) error {
+				nodes, err := clientSet.CoreV1().Nodes().List(ctx, metav1.ListOptions{})
+				if err != nil {
+					return err
+				}
+				if len(nodes.Items) < 1 {
+					return errors.New("node not registered")
+				}
+				if len(nodes.Items) > 1 {
+					return errors.New("more than one node registered (but there is only one)")
+				}
+				node := nodes.Items[0]
+				for _, cond := range node.Status.Conditions {
+					if cond.Type != corev1.NodeReady {
+						continue
+					}
+					if cond.Status != corev1.ConditionTrue {
+						return fmt.Errorf("node not ready: %v", cond.Message)
+					}
+				}
+				return nil
+			})
+			testEventual(t, "Simple deployment", ctx, largeTestTimeout, func(ctx context.Context) error {
+				_, err := clientSet.AppsV1().Deployments("default").Create(ctx, makeTestDeploymentSpec("test-deploy-1"), metav1.CreateOptions{})
+				return err
+			})
+			testEventual(t, "Simple deployment is running", ctx, largeTestTimeout, func(ctx context.Context) error {
+				res, err := clientSet.CoreV1().Pods("default").List(ctx, metav1.ListOptions{LabelSelector: "name=test-deploy-1"})
+				if err != nil {
+					return err
+				}
+				if len(res.Items) == 0 {
+					return errors.New("pod didn't get created")
+				}
+				pod := res.Items[0]
+				if podv1.IsPodAvailable(&pod, 1, metav1.NewTime(time.Now())) {
+					return nil
+				}
+				events, err := clientSet.CoreV1().Events("default").List(ctx, metav1.ListOptions{FieldSelector: fmt.Sprintf("involvedObject.name=%s,involvedObject.namespace=default", pod.Name)})
+				if err != nil || len(events.Items) == 0 {
+					return fmt.Errorf("pod is not ready: %v", pod.Status.Phase)
+				} else {
+					return fmt.Errorf("pod is not ready: %v", events.Items[0].Message)
+				}
+			})
+			testEventual(t, "Simple deployment with runc", ctx, largeTestTimeout, func(ctx context.Context) error {
+				deployment := makeTestDeploymentSpec("test-deploy-2")
+				var runcStr = "runc"
+				deployment.Spec.Template.Spec.RuntimeClassName = &runcStr
+				_, err := clientSet.AppsV1().Deployments("default").Create(ctx, deployment, metav1.CreateOptions{})
+				return err
+			})
+			testEventual(t, "Simple deployment is running on runc", ctx, largeTestTimeout, func(ctx context.Context) error {
+				res, err := clientSet.CoreV1().Pods("default").List(ctx, metav1.ListOptions{LabelSelector: "name=test-deploy-2"})
+				if err != nil {
+					return err
+				}
+				if len(res.Items) == 0 {
+					return errors.New("pod didn't get created")
+				}
+				pod := res.Items[0]
+				if podv1.IsPodAvailable(&pod, 1, metav1.NewTime(time.Now())) {
+					return nil
+				}
+				events, err := clientSet.CoreV1().Events("default").List(ctx, metav1.ListOptions{FieldSelector: fmt.Sprintf("involvedObject.name=%s,involvedObject.namespace=default", pod.Name)})
+				if err != nil || len(events.Items) == 0 {
+					return fmt.Errorf("pod is not ready: %v", pod.Status.Phase)
+				} else {
+					var errorMsg strings.Builder
+					for _, msg := range events.Items {
+						errorMsg.WriteString(" | ")
+						errorMsg.WriteString(msg.Message)
+					}
+					return fmt.Errorf("pod is not ready: %v", errorMsg.String())
+				}
+			})
+			testEventual(t, "Simple StatefulSet with PVC", ctx, largeTestTimeout, func(ctx context.Context) error {
+				_, err := clientSet.AppsV1().StatefulSets("default").Create(ctx, makeTestStatefulSet("test-statefulset-1"), metav1.CreateOptions{})
+				return err
+			})
+			testEventual(t, "Simple StatefulSet with PVC is running", ctx, largeTestTimeout, func(ctx context.Context) error {
+				res, err := clientSet.CoreV1().Pods("default").List(ctx, metav1.ListOptions{LabelSelector: "name=test-statefulset-1"})
+				if err != nil {
+					return err
+				}
+				if len(res.Items) == 0 {
+					return errors.New("pod didn't get created")
+				}
+				pod := res.Items[0]
+				if podv1.IsPodAvailable(&pod, 1, metav1.NewTime(time.Now())) {
+					return nil
+				}
+				events, err := clientSet.CoreV1().Events("default").List(ctx, metav1.ListOptions{FieldSelector: fmt.Sprintf("involvedObject.name=%s,involvedObject.namespace=default", pod.Name)})
+				if err != nil || len(events.Items) == 0 {
+					return fmt.Errorf("pod is not ready: %v", pod.Status.Phase)
+				} else {
+					return fmt.Errorf("pod is not ready: %v", events.Items[0].Message)
+				}
+			})
+			testEventual(t, "Pod with preseeded image", ctx, smallTestTimeout, func(ctx context.Context) error {
+				_, err := clientSet.CoreV1().Pods("default").Create(ctx, &corev1.Pod{
+					ObjectMeta: metav1.ObjectMeta{
+						Name: "preseed-test-1",
+					},
+					Spec: corev1.PodSpec{
+						Containers: []corev1.Container{{
+							Name:            "preseed-test-1",
+							ImagePullPolicy: corev1.PullNever,
+							Image:           "bazel/metropolis/test/e2e/preseedtest:preseedtest",
+						}},
+						RestartPolicy: corev1.RestartPolicyNever,
+					},
+				}, metav1.CreateOptions{})
+				return err
+			})
+			testEventual(t, "Pod with preseeded image is completed", ctx, largeTestTimeout, func(ctx context.Context) error {
+				pod, err := clientSet.CoreV1().Pods("default").Get(ctx, "preseed-test-1", metav1.GetOptions{})
+				if err != nil {
+					return fmt.Errorf("failed to get pod: %w", err)
+				}
+				if pod.Status.Phase == corev1.PodSucceeded {
+					return nil
+				}
+				events, err := clientSet.CoreV1().Events("default").List(ctx, metav1.ListOptions{FieldSelector: fmt.Sprintf("involvedObject.name=%s,involvedObject.namespace=default", pod.Name)})
+				if err != nil || len(events.Items) == 0 {
+					return fmt.Errorf("pod is not ready: %v", pod.Status.Phase)
+				} else {
+					return fmt.Errorf("pod is not ready: %v", events.Items[len(events.Items)-1].Message)
+				}
+			})
+		})
+	})
+
+	// Cancel the main context and wait for our subprocesses to exit
+	// to avoid leaking them and blocking the parent.
+	cancel()
+	<-procExit
+}
diff --git a/metropolis/test/e2e/preseedtest/BUILD.bazel b/metropolis/test/e2e/preseedtest/BUILD.bazel
new file mode 100644
index 0000000..41b32e3
--- /dev/null
+++ b/metropolis/test/e2e/preseedtest/BUILD.bazel
@@ -0,0 +1,16 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("@io_bazel_rules_docker//go:image.bzl", "go_image")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/e2e/preseedtest",
+    visibility = ["//visibility:private"],
+)
+
+go_image(
+    name = "preseedtest",
+    embed = [":go_default_library"],
+    pure = "on",
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/test/e2e/preseedtest/main.go b/metropolis/test/e2e/preseedtest/main.go
new file mode 100644
index 0000000..ceb3898
--- /dev/null
+++ b/metropolis/test/e2e/preseedtest/main.go
@@ -0,0 +1,23 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import "fmt"
+
+func main() {
+	fmt.Println("Hello world from preseeded image")
+}
diff --git a/metropolis/test/e2e/utils.go b/metropolis/test/e2e/utils.go
new file mode 100644
index 0000000..f888189
--- /dev/null
+++ b/metropolis/test/e2e/utils.go
@@ -0,0 +1,51 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package e2e
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+)
+
+// testEventual creates a new subtest looping the given function until it either doesn't return an error anymore or
+// the timeout is exceeded. The last returned non-context-related error is being used as the test error.
+func testEventual(t *testing.T, name string, ctx context.Context, timeout time.Duration, f func(context.Context) error) {
+	ctx, cancel := context.WithTimeout(ctx, timeout)
+	t.Helper()
+	t.Run(name, func(t *testing.T) {
+		defer cancel()
+		var lastErr = errors.New("test didn't run to completion at least once")
+		t.Parallel()
+		for {
+			err := f(ctx)
+			if err == nil {
+				return
+			}
+			if err == ctx.Err() {
+				t.Fatal(lastErr)
+			}
+			lastErr = err
+			select {
+			case <-ctx.Done():
+				t.Fatal(lastErr)
+			case <-time.After(1 * time.Second):
+			}
+		}
+	})
+}
diff --git a/metropolis/test/ktest/BUILD b/metropolis/test/ktest/BUILD
new file mode 100644
index 0000000..d94831c
--- /dev/null
+++ b/metropolis/test/ktest/BUILD
@@ -0,0 +1,63 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+load("//metropolis/node/build/kconfig-patcher:kconfig-patcher.bzl", "kconfig_patch")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/ktest",
+    visibility = ["//visibility:private"],
+    deps = ["//metropolis/test/launch:go_default_library"],
+)
+
+go_binary(
+    name = "ktest",
+    embed = [":go_default_library"],
+    pure = "on",
+    visibility = ["//visibility:public"],
+)
+
+kconfig_patch(
+    name = "testing-config",
+    src = "//third_party/linux:kernel-config",
+    out = "testing.config",
+    override_configs = {
+        # Unlock command line
+        "CONFIG_CMDLINE_OVERRIDE": "n",
+        "CONFIG_CMDLINE_BOOL": "n",
+        # Shave off 1 second from boot time
+        "CONFIG_SERIO_I8042": "",
+        "CONFIG_KEYBOARD_ATKBD": "",
+        "CONFIG_RTC_DRV_CMOS": "",
+        # Shave off an additional 18ms (half of the boot time)
+        "CONFIG_DEBUG_WX": "",
+    },
+)
+
+genrule(
+    name = "linux-testing",
+    srcs = [
+        "@linux//:all",
+        ":testing-config",
+    ],
+    outs = [
+        "linux-testing.elf",
+    ],
+    cmd = """
+    DIR=external/linux
+
+    mkdir $$DIR/.bin
+
+    cp $(location :testing-config) $$DIR/.config
+
+    (cd $$DIR && make -j $$(nproc) vmlinux >/dev/null)
+
+    cp $$DIR/vmlinux $@
+    """,
+    visibility = ["//visibility:public"],
+)
+
+filegroup(
+    name = "test-script",
+    srcs = ["run_ktest.sh"],
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/test/ktest/init/BUILD.bazel b/metropolis/test/ktest/init/BUILD.bazel
new file mode 100644
index 0000000..4161146
--- /dev/null
+++ b/metropolis/test/ktest/init/BUILD.bazel
@@ -0,0 +1,16 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/ktest/init",
+    visibility = ["//visibility:private"],
+    deps = ["@org_golang_x_sys//unix:go_default_library"],
+)
+
+go_binary(
+    name = "init",
+    embed = [":go_default_library"],
+    pure = "on",
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/test/ktest/init/main.go b/metropolis/test/ktest/init/main.go
new file mode 100644
index 0000000..f6049db
--- /dev/null
+++ b/metropolis/test/ktest/init/main.go
@@ -0,0 +1,83 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// ktestinit is an init designed to run inside a lightweight VM for running tests in there.
+// It performs basic platform initialization like mounting kernel filesystems and launches the
+// test executable at /tester, passes the exit code back out over the control socket to ktest and
+// then terminates the VM kernel.
+package main
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+
+	"golang.org/x/sys/unix"
+)
+
+func mountInit() error {
+	for _, el := range []struct {
+		dir   string
+		fs    string
+		flags uintptr
+	}{
+		{"/sys", "sysfs", unix.MS_NOEXEC | unix.MS_NOSUID | unix.MS_NODEV},
+		{"/proc", "proc", unix.MS_NOEXEC | unix.MS_NOSUID | unix.MS_NODEV},
+		{"/dev", "devtmpfs", unix.MS_NOEXEC | unix.MS_NOSUID},
+		{"/dev/pts", "devpts", unix.MS_NOEXEC | unix.MS_NOSUID},
+	} {
+		if err := os.Mkdir(el.dir, 0755); err != nil && !os.IsExist(err) {
+			return fmt.Errorf("could not make %s: %w", el.dir, err)
+		}
+		if err := unix.Mount(el.fs, el.dir, el.fs, el.flags, ""); err != nil {
+			return fmt.Errorf("could not mount %s on %s: %w", el.fs, el.dir, err)
+		}
+	}
+	return nil
+}
+
+func main() {
+	if err := mountInit(); err != nil {
+		panic(err)
+	}
+
+	// First virtual serial is always stdout, second is control
+	ioConn, err := os.OpenFile("/dev/vport1p1", os.O_RDWR, 0)
+	if err != nil {
+		fmt.Printf("Failed to open communication device: %v\n", err)
+		return
+	}
+	cmd := exec.Command("/tester", "-test.v")
+	cmd.Stderr = os.Stderr
+	cmd.Stdout = os.Stdout
+	cmd.Env = append(cmd.Env, "IN_KTEST=true")
+	if err := cmd.Run(); err != nil {
+		var exerr *exec.ExitError
+		if errors.As(err, &exerr) {
+			if _, err := ioConn.Write([]byte{uint8(exerr.ExitCode())}); err != nil {
+				panic(err)
+			}
+		} else if err != nil {
+			fmt.Printf("Failed to execute tests (tests didn't run): %v", err)
+		}
+	} else {
+		ioConn.Write([]byte{0})
+	}
+	ioConn.Close()
+
+	unix.Reboot(unix.LINUX_REBOOT_CMD_RESTART)
+}
diff --git a/metropolis/test/ktest/ktest.bzl b/metropolis/test/ktest/ktest.bzl
new file mode 100644
index 0000000..fdbff20
--- /dev/null
+++ b/metropolis/test/ktest/ktest.bzl
@@ -0,0 +1,62 @@
+#  Copyright 2020 The Monogon Project Authors.
+#
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+
+"""
+Ktest provides a simple macro to run tests inside the normal Smalltown kernel
+"""
+
+def ktest(deps, tester, initramfs_extra, cmdline):
+    native.genrule(
+        name = "test_initramfs",
+        srcs = [
+            "//metropolis/test/ktest/init",
+        ] + deps + [tester],
+        outs = [
+            "initramfs.cpio.lz4",
+        ],
+        testonly = True,
+        cmd = """
+        $(location @linux//:gen_init_cpio) - <<- 'EOF' | lz4 -l > \"$@\" 
+dir /dev 0755 0 0
+nod /dev/console 0600 0 0 c 5 1
+nod /dev/null 0644 0 0 c 1 3
+file /init $(location //metropolis/test/ktest/init) 0755 0 0
+file /tester $(location """ + tester + """) 0755 0 0
+""" + initramfs_extra + """
+EOF
+        """,
+        tools = [
+            "@linux//:gen_init_cpio",
+        ],
+    )
+
+    native.sh_test(
+        name = "ktest",
+        args = [
+            "$(location //metropolis/test/ktest)",
+            "$(location :test_initramfs)",
+            "$(location //metropolis/test/ktest:linux-testing)",
+            cmdline,
+        ],
+        size = "small",
+        srcs = ["//metropolis/test/ktest:test-script"],
+        data = [
+            "//metropolis/test/ktest",
+            ":test_initramfs",
+            "//metropolis/test/ktest:linux-testing",
+            "@com_github_bonzini_qboot//:qboot-bin",
+        ],
+    )
diff --git a/metropolis/test/ktest/main.go b/metropolis/test/ktest/main.go
new file mode 100644
index 0000000..7f750b8
--- /dev/null
+++ b/metropolis/test/ktest/main.go
@@ -0,0 +1,75 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// ktest is a test launcher for running tests inside a custom kernel and passes the results
+// back out.
+package main
+
+import (
+	"context"
+	"flag"
+	"io"
+	"log"
+	"os"
+	"time"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/test/launch"
+)
+
+var (
+	kernelPath = flag.String("kernel-path", "", "Path of the Kernel ELF file")
+	initrdPath = flag.String("initrd-path", "", "Path of the initrd image")
+	cmdline    = flag.String("cmdline", "", "Additional kernel command line options")
+)
+
+func main() {
+	flag.Parse()
+
+	hostFeedbackConn, vmFeedbackConn, err := launch.NewSocketPair()
+	if err != nil {
+		log.Fatalf("Failed to create socket pair: %v", err)
+	}
+
+	exitCodeChan := make(chan uint8, 1)
+
+	go func() {
+		defer hostFeedbackConn.Close()
+
+		returnCode := make([]byte, 1)
+		if _, err := io.ReadFull(hostFeedbackConn, returnCode); err != nil {
+			log.Fatalf("Failed to read socket: %v", err)
+		}
+		exitCodeChan <- returnCode[0]
+	}()
+
+	if err := launch.RunMicroVM(context.Background(), &launch.MicroVMOptions{
+		KernelPath:                  *kernelPath,
+		InitramfsPath:               *initrdPath,
+		Cmdline:                     *cmdline,
+		SerialPort:                  os.Stdout,
+		ExtraChardevs:               []*os.File{vmFeedbackConn},
+		DisableHostNetworkInterface: true,
+	}); err != nil {
+		log.Fatalf("Failed to run ktest VM: %v", err)
+	}
+
+	select {
+	case exitCode := <-exitCodeChan:
+		os.Exit(int(exitCode))
+	case <-time.After(1 * time.Second):
+		log.Fatal("Failed to get an error code back (test runtime probably crashed)")
+	}
+}
diff --git a/metropolis/test/ktest/run_ktest.sh b/metropolis/test/ktest/run_ktest.sh
new file mode 100755
index 0000000..02920a1
--- /dev/null
+++ b/metropolis/test/ktest/run_ktest.sh
@@ -0,0 +1,2 @@
+#!/bin/bash
+exec "$1" -initrd-path "$2" -kernel-path "$3" -cmdline "$4"
\ No newline at end of file
diff --git a/metropolis/test/launch/BUILD.bazel b/metropolis/test/launch/BUILD.bazel
new file mode 100644
index 0000000..b6245e1
--- /dev/null
+++ b/metropolis/test/launch/BUILD.bazel
@@ -0,0 +1,17 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["launch.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/launch",
+    visibility = ["//metropolis:__subpackages__"],
+    deps = [
+        "//golibs/common:go_default_library",
+        "//metropolis/node:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+        "@com_github_golang_protobuf//proto:go_default_library",
+        "@com_github_grpc_ecosystem_go_grpc_middleware//retry:go_default_library",
+        "@org_golang_google_grpc//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/metropolis/test/launch/cli/launch-multi2/BUILD.bazel b/metropolis/test/launch/cli/launch-multi2/BUILD.bazel
new file mode 100644
index 0000000..9f27860
--- /dev/null
+++ b/metropolis/test/launch/cli/launch-multi2/BUILD.bazel
@@ -0,0 +1,29 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/launch/cli/launch-multi2",
+    visibility = ["//visibility:private"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/proto/api:go_default_library",
+        "//metropolis/test/launch:go_default_library",
+        "@com_github_grpc_ecosystem_go_grpc_middleware//retry:go_default_library",
+        "@org_golang_google_grpc//:go_default_library",
+    ],
+)
+
+go_binary(
+    name = "launch-multi2",
+    data = [
+        "//metropolis/node:image",
+        "//metropolis/node:swtpm_data",
+        "//metropolis/test/nanoswitch:initramfs",
+        "//metropolis/test/ktest:linux-testing",
+        "//third_party/edk2:firmware",
+        "@com_github_bonzini_qboot//:qboot-bin",
+    ],
+    embed = [":go_default_library"],
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/test/launch/cli/launch-multi2/main.go b/metropolis/test/launch/cli/launch-multi2/main.go
new file mode 100644
index 0000000..265d6a0
--- /dev/null
+++ b/metropolis/test/launch/cli/launch-multi2/main.go
@@ -0,0 +1,102 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"context"
+	"log"
+	"os"
+	"os/signal"
+	"syscall"
+	"time"
+
+	grpcretry "github.com/grpc-ecosystem/go-grpc-middleware/retry"
+	"google.golang.org/grpc"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+	"git.monogon.dev/source/nexantic.git/metropolis/test/launch"
+)
+
+func main() {
+	sigs := make(chan os.Signal, 1)
+	signal.Notify(sigs, syscall.SIGINT, syscall.SIGTERM)
+	ctx, cancel := context.WithCancel(context.Background())
+	go func() {
+		<-sigs
+		cancel()
+	}()
+	sw0, vm0, err := launch.NewSocketPair()
+	if err != nil {
+		log.Fatalf("Failed to create network pipe: %v\n", err)
+	}
+	sw1, vm1, err := launch.NewSocketPair()
+	if err != nil {
+		log.Fatalf("Failed to create network pipe: %v\n", err)
+	}
+
+	go func() {
+		if err := launch.Launch(ctx, launch.Options{ConnectToSocket: vm0, SerialPort: os.Stdout}); err != nil {
+			log.Fatalf("Failed to launch vm0: %v", err)
+		}
+	}()
+	nanoswitchPortMap := make(launch.PortMap)
+	identityPorts := []uint16{
+		common.ExternalServicePort,
+		common.DebugServicePort,
+		common.KubernetesAPIPort,
+	}
+	for _, port := range identityPorts {
+		nanoswitchPortMap[port] = port
+	}
+	go func() {
+		opts := []grpcretry.CallOption{
+			grpcretry.WithBackoff(grpcretry.BackoffExponential(100 * time.Millisecond)),
+		}
+		conn, err := nanoswitchPortMap.DialGRPC(common.DebugServicePort, grpc.WithInsecure(),
+			grpc.WithUnaryInterceptor(grpcretry.UnaryClientInterceptor(opts...)))
+		if err != nil {
+			panic(err)
+		}
+		defer conn.Close()
+		debug := apb.NewNodeDebugServiceClient(conn)
+		res, err := debug.GetGoldenTicket(ctx, &apb.GetGoldenTicketRequest{
+			// HACK: this is assigned by DHCP, and we assume that everything goes well.
+			ExternalIp: "10.1.0.3",
+		}, grpcretry.WithMax(10))
+		if err != nil {
+			log.Fatalf("Failed to get golden ticket: %v", err)
+		}
+
+		ec := &apb.EnrolmentConfig{
+			GoldenTicket: res.Ticket,
+		}
+
+		if err := launch.Launch(ctx, launch.Options{ConnectToSocket: vm1, EnrolmentConfig: ec, SerialPort: os.Stdout}); err != nil {
+			log.Fatalf("Failed to launch vm1: %v", err)
+		}
+	}()
+	if err := launch.RunMicroVM(ctx, &launch.MicroVMOptions{
+		SerialPort:             os.Stdout,
+		KernelPath:             "metropolis/test/ktest/linux-testing.elf",
+		InitramfsPath:          "metropolis/test/nanoswitch/initramfs.lz4",
+		ExtraNetworkInterfaces: []*os.File{sw0, sw1},
+		PortMap:                nanoswitchPortMap,
+	}); err != nil {
+		log.Fatalf("Failed to launch nanoswitch: %v", err)
+	}
+}
diff --git a/metropolis/test/launch/cli/launch/BUILD.bazel b/metropolis/test/launch/cli/launch/BUILD.bazel
new file mode 100644
index 0000000..6b1461d
--- /dev/null
+++ b/metropolis/test/launch/cli/launch/BUILD.bazel
@@ -0,0 +1,20 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["main.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/launch/cli/launch",
+    visibility = ["//visibility:private"],
+    deps = ["//metropolis/test/launch:go_default_library"],
+)
+
+go_binary(
+    name = "launch",
+    data = [
+        "//metropolis/node:image",
+        "//metropolis/node:swtpm_data",
+        "//third_party/edk2:firmware",
+    ],
+    embed = [":go_default_library"],
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/test/launch/cli/launch/main.go b/metropolis/test/launch/cli/launch/main.go
new file mode 100644
index 0000000..852c8e1
--- /dev/null
+++ b/metropolis/test/launch/cli/launch/main.go
@@ -0,0 +1,43 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package main
+
+import (
+	"context"
+	"log"
+	"os"
+	"os/signal"
+	"syscall"
+
+	"git.monogon.dev/source/nexantic.git/metropolis/test/launch"
+)
+
+func main() {
+	sigs := make(chan os.Signal, 1)
+	signal.Notify(sigs, syscall.SIGINT, syscall.SIGTERM)
+	ctx, cancel := context.WithCancel(context.Background())
+	go func() {
+		<-sigs
+		cancel()
+	}()
+	if err := launch.Launch(ctx, launch.Options{Ports: launch.IdentityPortMap(launch.NodePorts), SerialPort: os.Stdout}); err != nil {
+		if err == ctx.Err() {
+			return
+		}
+		log.Fatalf("Failed to execute: %v\n", err)
+	}
+}
diff --git a/metropolis/test/launch/launch.go b/metropolis/test/launch/launch.go
new file mode 100644
index 0000000..2d495e0
--- /dev/null
+++ b/metropolis/test/launch/launch.go
@@ -0,0 +1,557 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package launch
+
+import (
+	"bytes"
+	"context"
+	"crypto/rand"
+	"errors"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"log"
+	"net"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"syscall"
+	"time"
+
+	"github.com/golang/protobuf/proto"
+	grpcretry "github.com/grpc-ecosystem/go-grpc-middleware/retry"
+	"golang.org/x/sys/unix"
+	"google.golang.org/grpc"
+
+	freeport "git.monogon.dev/source/nexantic.git/golibs/common"
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	apb "git.monogon.dev/source/nexantic.git/metropolis/proto/api"
+)
+
+type qemuValue map[string][]string
+
+// toOption encodes structured data into a QEMU option.
+// Example: "test", {"key1": {"val1"}, "key2": {"val2", "val3"}} returns "test,key1=val1,key2=val2,key2=val3"
+func (value qemuValue) toOption(name string) string {
+	var optionValues []string
+	if name != "" {
+		optionValues = append(optionValues, name)
+	}
+	for name, values := range value {
+		if len(values) == 0 {
+			optionValues = append(optionValues, name)
+		}
+		for _, val := range values {
+			optionValues = append(optionValues, fmt.Sprintf("%v=%v", name, val))
+		}
+	}
+	return strings.Join(optionValues, ",")
+}
+
+func copyFile(src, dst string) error {
+	in, err := os.Open(src)
+	if err != nil {
+		return err
+	}
+	defer in.Close()
+
+	out, err := os.Create(dst)
+	if err != nil {
+		return err
+	}
+	defer out.Close()
+
+	_, err = io.Copy(out, in)
+	if err != nil {
+		return err
+	}
+	return out.Close()
+}
+
+// PortMap represents where VM ports are mapped to on the host. It maps from the VM port number to the host port number.
+type PortMap map[uint16]uint16
+
+// toQemuForwards generates QEMU hostfwd values (https://qemu.weilnetz.de/doc/qemu-doc.html#:~:text=hostfwd=) for all
+// mapped ports.
+func (p PortMap) toQemuForwards() []string {
+	var hostfwdOptions []string
+	for vmPort, hostPort := range p {
+		hostfwdOptions = append(hostfwdOptions, fmt.Sprintf("tcp::%v-:%v", hostPort, vmPort))
+	}
+	return hostfwdOptions
+}
+
+// DialGRPC creates a gRPC client for a VM port that's forwarded/mapped to the host. The given port is automatically
+// resolved to the host-mapped port.
+func (p PortMap) DialGRPC(port uint16, opts ...grpc.DialOption) (*grpc.ClientConn, error) {
+	mappedPort, ok := p[port]
+	if !ok {
+		return nil, fmt.Errorf("cannot dial port: port %v is not mapped/forwarded", port)
+	}
+	grpcClient, err := grpc.Dial(fmt.Sprintf("localhost:%v", mappedPort), opts...)
+	if err != nil {
+		return nil, fmt.Errorf("failed to dial port %v: %w", port, err)
+	}
+	return grpcClient, nil
+}
+
+// Options contains all options that can be passed to Launch()
+type Options struct {
+	// Ports contains the port mapping where to expose the internal ports of the VM to the host. See IdentityPortMap()
+	// and ConflictFreePortMap(). Ignored when ConnectToSocket is set.
+	Ports PortMap
+
+	// If set to true, reboots are honored. Otherwise all reboots exit the Launch() command. Smalltown generally restarts
+	// on almost all errors, so unless you want to test reboot behavior this should be false.
+	AllowReboot bool
+
+	// By default the Smalltown VM is connected to the Host via SLIRP. If ConnectToSocket is set, it is instead connected
+	// to the given file descriptor/socket. If this is set, all port maps from the Ports option are ignored.
+	// Intended for networking this instance together with others for running  more complex network configurations.
+	ConnectToSocket *os.File
+
+	// SerialPort is a File(descriptor) over which you can communicate with the serial port of the machine
+	// It can be set to an existing file descriptor (like os.Stdout/os.Stderr) or you can use NewSocketPair() to get one
+	// end to talk to from Go.
+	SerialPort *os.File
+
+	// EnrolmentConfig is passed into the VM and subsequently used for bootstrapping if no enrolment config is built-in
+	EnrolmentConfig *apb.EnrolmentConfig
+}
+
+// NodePorts is the list of ports a fully operational Smalltown node listens on
+var NodePorts = []uint16{common.ConsensusPort, common.NodeServicePort, common.MasterServicePort,
+	common.ExternalServicePort, common.DebugServicePort, common.KubernetesAPIPort, common.DebuggerPort}
+
+// IdentityPortMap returns a port map where each given port is mapped onto itself on the host. This is mainly useful
+// for development against Smalltown. The dbg command requires this mapping.
+func IdentityPortMap(ports []uint16) PortMap {
+	portMap := make(PortMap)
+	for _, port := range ports {
+		portMap[port] = port
+	}
+	return portMap
+}
+
+// ConflictFreePortMap returns a port map where each given port is mapped onto a random free port on the host. This is
+// intended for automated testing where multiple instances of Smalltown might be running. Please call this function for
+// each Launch command separately and as close to it as possible since it cannot guarantee that the ports will remain
+// free.
+func ConflictFreePortMap(ports []uint16) (PortMap, error) {
+	portMap := make(PortMap)
+	for _, port := range ports {
+		mappedPort, listenCloser, err := freeport.AllocateTCPPort()
+		if err != nil {
+			return portMap, fmt.Errorf("failed to get free host port: %w", err)
+		}
+		// Defer closing of the listening port until the function is done and all ports are allocated
+		defer listenCloser.Close()
+		portMap[port] = mappedPort
+	}
+	return portMap, nil
+}
+
+// Gets a random EUI-48 Ethernet MAC address
+func generateRandomEthernetMAC() (*net.HardwareAddr, error) {
+	macBuf := make([]byte, 6)
+	_, err := rand.Read(macBuf)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read randomness for MAC: %v", err)
+	}
+
+	// Set U/L bit and clear I/G bit (locally administered individual MAC)
+	// Ref IEEE 802-2014 Section 8.2.2
+	macBuf[0] = (macBuf[0] | 2) & 0xfe
+	mac := net.HardwareAddr(macBuf)
+	return &mac, nil
+}
+
+// Launch launches a Smalltown instance with the given options. The instance runs mostly paravirtualized but with some
+// emulated hardware similar to how a cloud provider might set up its VMs. The disk is fully writable but is run
+// in snapshot mode meaning that changes are not kept beyond a single invocation.
+func Launch(ctx context.Context, options Options) error {
+	// Pin temp directory to /tmp until we can use abstract socket namespace in QEMU (next release after 5.0,
+	// https://github.com/qemu/qemu/commit/776b97d3605ed0fc94443048fdf988c7725e38a9). swtpm accepts already-open FDs
+	// so we can pass in an abstract socket namespace FD that we open and pass the name of it to QEMU. Not pinning this
+	// crashes both swtpm and qemu because we run into UNIX socket length limitations (for legacy reasons 108 chars).
+	tempDir, err := ioutil.TempDir("/tmp", "launch*")
+	if err != nil {
+		return fmt.Errorf("failed to create temporary directory: %w", err)
+	}
+	defer os.RemoveAll(tempDir)
+
+	// Copy TPM state into a temporary directory since it's being modified by the emulator
+	tpmTargetDir := filepath.Join(tempDir, "tpm")
+	tpmSrcDir := "metropolis/node/tpm"
+	if err := os.Mkdir(tpmTargetDir, 0644); err != nil {
+		return fmt.Errorf("failed to create TPM state directory: %w", err)
+	}
+	tpmFiles, err := ioutil.ReadDir(tpmSrcDir)
+	if err != nil {
+		return fmt.Errorf("failed to read TPM directory: %w", err)
+	}
+	for _, file := range tpmFiles {
+		name := file.Name()
+		if err := copyFile(filepath.Join(tpmSrcDir, name), filepath.Join(tpmTargetDir, name)); err != nil {
+			return fmt.Errorf("failed to copy TPM directory: %w", err)
+		}
+	}
+
+	var qemuNetType string
+	var qemuNetConfig qemuValue
+	if options.ConnectToSocket != nil {
+		qemuNetType = "socket"
+		qemuNetConfig = qemuValue{
+			"id": {"net0"},
+			"fd": {"3"},
+		}
+	} else {
+		qemuNetType = "user"
+		qemuNetConfig = qemuValue{
+			"id":        {"net0"},
+			"net":       {"10.42.0.0/24"},
+			"dhcpstart": {"10.42.0.10"},
+			"hostfwd":   options.Ports.toQemuForwards(),
+		}
+	}
+
+	tpmSocketPath := filepath.Join(tempDir, "tpm-socket")
+
+	mac, err := generateRandomEthernetMAC()
+	if err != nil {
+		return err
+	}
+
+	qemuArgs := []string{"-machine", "q35", "-accel", "kvm", "-nographic", "-nodefaults", "-m", "4096",
+		"-cpu", "host", "-smp", "sockets=1,cpus=1,cores=2,threads=2,maxcpus=4",
+		"-drive", "if=pflash,format=raw,readonly,file=external/edk2/OVMF_CODE.fd",
+		"-drive", "if=pflash,format=raw,snapshot=on,file=external/edk2/OVMF_VARS.fd",
+		"-drive", "if=virtio,format=raw,snapshot=on,cache=unsafe,file=metropolis/node/smalltown.img",
+		"-netdev", qemuNetConfig.toOption(qemuNetType),
+		"-device", "virtio-net-pci,netdev=net0,mac=" + mac.String(),
+		"-chardev", "socket,id=chrtpm,path=" + tpmSocketPath,
+		"-tpmdev", "emulator,id=tpm0,chardev=chrtpm",
+		"-device", "tpm-tis,tpmdev=tpm0",
+		"-device", "virtio-rng-pci",
+		"-serial", "stdio"}
+
+	if !options.AllowReboot {
+		qemuArgs = append(qemuArgs, "-no-reboot")
+	}
+
+	if options.EnrolmentConfig != nil {
+		enrolmentConfigPath := filepath.Join(tempDir, "enrolment.pb")
+		enrolmentConfigRaw, err := proto.Marshal(options.EnrolmentConfig)
+		if err != nil {
+			return fmt.Errorf("failed to encode enrolment config: %w", err)
+		}
+		if err := ioutil.WriteFile(enrolmentConfigPath, enrolmentConfigRaw, 0644); err != nil {
+			return fmt.Errorf("failed to write enrolment config: %w", err)
+		}
+		qemuArgs = append(qemuArgs, "-fw_cfg", "name=com.nexantic.smalltown/enrolment.pb,file="+enrolmentConfigPath)
+	}
+
+	// Start TPM emulator as a subprocess
+	tpmCtx, tpmCancel := context.WithCancel(ctx)
+	defer tpmCancel()
+
+	tpmEmuCmd := exec.CommandContext(tpmCtx, "swtpm", "socket", "--tpm2", "--tpmstate", "dir="+tpmTargetDir, "--ctrl", "type=unixio,path="+tpmSocketPath)
+	tpmEmuCmd.Stderr = os.Stderr
+	tpmEmuCmd.Stdout = os.Stdout
+
+	err = tpmEmuCmd.Start()
+	if err != nil {
+		return fmt.Errorf("failed to start TPM emulator: %w", err)
+	}
+
+	// Start the main qemu binary
+	systemCmd := exec.CommandContext(ctx, "qemu-system-x86_64", qemuArgs...)
+	if options.ConnectToSocket != nil {
+		systemCmd.ExtraFiles = []*os.File{options.ConnectToSocket}
+	}
+
+	var stdErrBuf bytes.Buffer
+	systemCmd.Stderr = &stdErrBuf
+	systemCmd.Stdout = options.SerialPort
+
+	err = systemCmd.Run()
+
+	// Stop TPM emulator and wait for it to exit to properly reap the child process
+	tpmCancel()
+	log.Print("Waiting for TPM emulator to exit")
+	// Wait returns a SIGKILL error because we just cancelled its context.
+	// We still need to call it to avoid creating zombies.
+	_ = tpmEmuCmd.Wait()
+
+	var exerr *exec.ExitError
+	if err != nil && errors.As(err, &exerr) {
+		status := exerr.ProcessState.Sys().(syscall.WaitStatus)
+		if status.Signaled() && status.Signal() == syscall.SIGKILL {
+			// Process was killed externally (most likely by our context being canceled).
+			// This is a normal exit for us, so return nil
+			return nil
+		}
+		exerr.Stderr = stdErrBuf.Bytes()
+		newErr := QEMUError(*exerr)
+		return &newErr
+	}
+	return err
+}
+
+// NewSocketPair creates a new socket pair. By connecting both ends to different instances you can connect them
+// with a virtual "network cable". The ends can be passed into the ConnectToSocket option.
+func NewSocketPair() (*os.File, *os.File, error) {
+	fds, err := unix.Socketpair(unix.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		return nil, nil, fmt.Errorf("failed to call socketpair: %w", err)
+	}
+
+	fd1 := os.NewFile(uintptr(fds[0]), "network0")
+	fd2 := os.NewFile(uintptr(fds[1]), "network1")
+	return fd1, fd2, nil
+}
+
+// HostInterfaceMAC is the MAC address the host SLIRP network interface has if it is not disabled (see
+// DisableHostNetworkInterface in MicroVMOptions)
+var HostInterfaceMAC = net.HardwareAddr{0x02, 0x72, 0x82, 0xbf, 0xc3, 0x56}
+
+// MicroVMOptions contains all options to start a MicroVM
+type MicroVMOptions struct {
+	// Path to the ELF kernel binary
+	KernelPath string
+
+	// Path to the Initramfs
+	InitramfsPath string
+
+	// Cmdline contains additional kernel commandline options
+	Cmdline string
+
+	// SerialPort is a File(descriptor) over which you can communicate with the serial port of the machine
+	// It can be set to an existing file descriptor (like os.Stdout/os.Stderr) or you can use NewSocketPair() to get one
+	// end to talk to from Go.
+	SerialPort *os.File
+
+	// ExtraChardevs can be used similar to SerialPort, but can contain an arbitrary number of additional serial ports
+	ExtraChardevs []*os.File
+
+	// ExtraNetworkInterfaces can contain an arbitrary number of file descriptors which are mapped into the VM as virtio
+	// network interfaces. The first interface is always a SLIRP-backed interface for communicating with the host.
+	ExtraNetworkInterfaces []*os.File
+
+	// PortMap contains ports that are mapped to the host through the built-in SLIRP network interface.
+	PortMap PortMap
+
+	// DisableHostNetworkInterface disables the SLIRP-backed host network interface that is normally the first network
+	// interface. If this is set PortMap is ignored. Mostly useful for speeding up QEMU's startup time for tests.
+	DisableHostNetworkInterface bool
+}
+
+// RunMicroVM launches a tiny VM mostly intended for testing. Very quick to boot (<40ms).
+func RunMicroVM(ctx context.Context, opts *MicroVMOptions) error {
+	// Generate options for all the file descriptors we'll be passing as virtio "serial ports"
+	var extraArgs []string
+	for idx, _ := range opts.ExtraChardevs {
+		idxStr := strconv.Itoa(idx)
+		id := "extra" + idxStr
+		// That this works is pretty much a hack, but upstream QEMU doesn't have a bidirectional chardev backend not
+		// based around files/sockets on the disk which are a giant pain to work with.
+		// We're using QEMU's fdset functionality to make FDs available as pseudo-files and then "ab"using the pipe
+		// backend's fallback functionality to get a single bidirectional chardev backend backed by a passed-down
+		// RDWR fd.
+		// Ref https://lists.gnu.org/archive/html/qemu-devel/2015-12/msg01256.html
+		addFdConf := qemuValue{
+			"set": {idxStr},
+			"fd":  {strconv.Itoa(idx + 3)},
+		}
+		chardevConf := qemuValue{
+			"id":   {id},
+			"path": {"/dev/fdset/" + idxStr},
+		}
+		deviceConf := qemuValue{
+			"chardev": {id},
+		}
+		extraArgs = append(extraArgs, "-add-fd", addFdConf.toOption(""),
+			"-chardev", chardevConf.toOption("pipe"), "-device", deviceConf.toOption("virtserialport"))
+	}
+
+	for idx, _ := range opts.ExtraNetworkInterfaces {
+		id := fmt.Sprintf("net%v", idx)
+		netdevConf := qemuValue{
+			"id": {id},
+			"fd": {strconv.Itoa(idx + 3 + len(opts.ExtraChardevs))},
+		}
+		extraArgs = append(extraArgs, "-netdev", netdevConf.toOption("socket"), "-device", "virtio-net-device,netdev="+id)
+	}
+
+	// This sets up a minimum viable environment for our Linux kernel.
+	// It clears all standard QEMU configuration and sets up a MicroVM machine
+	// (https://github.com/qemu/qemu/blob/master/docs/microvm.rst) with all legacy emulation turned off. This means
+	// the only "hardware" the Linux kernel inside can communicate with is a single virtio-mmio region. Over that MMIO
+	// interface we run a paravirtualized RNG (since the kernel in there has nothing to gather that from and it
+	// delays booting), a single paravirtualized console and an arbitrary number of extra serial ports for talking to
+	// various things that might run inside. The kernel, initramfs and command line are mapped into VM memory at boot
+	// time and not loaded from any sort of disk. Booting and shutting off one of these VMs takes <100ms.
+	baseArgs := []string{"-nodefaults", "-no-user-config", "-nographic", "-no-reboot",
+		"-accel", "kvm", "-cpu", "host",
+		// Needed until QEMU updates their bundled qboot version (needs https://github.com/bonzini/qboot/pull/28)
+		"-bios", "external/com_github_bonzini_qboot/bios.bin",
+		"-M", "microvm,x-option-roms=off,pic=off,pit=off,rtc=off,isa-serial=off",
+		"-kernel", opts.KernelPath,
+		// We force using a triple-fault reboot strategy since otherwise the kernel first tries others (like ACPI) which
+		// are not available in this very restricted environment. Similarly we need to override the boot console since
+		// there's nothing on the ISA bus that the kernel could talk to. We also force quiet for performance reasons.
+		"-append", "reboot=t console=hvc0 quiet " + opts.Cmdline,
+		"-initrd", opts.InitramfsPath,
+		"-device", "virtio-rng-device,max-bytes=1024,period=1000",
+		"-device", "virtio-serial-device,max_ports=16",
+		"-chardev", "stdio,id=con0", "-device", "virtconsole,chardev=con0",
+	}
+
+	if !opts.DisableHostNetworkInterface {
+		qemuNetType := "user"
+		qemuNetConfig := qemuValue{
+			"id":        {"usernet0"},
+			"net":       {"10.42.0.0/24"},
+			"dhcpstart": {"10.42.0.10"},
+		}
+		if opts.PortMap != nil {
+			qemuNetConfig["hostfwd"] = opts.PortMap.toQemuForwards()
+		}
+
+		baseArgs = append(baseArgs, "-netdev", qemuNetConfig.toOption(qemuNetType),
+			"-device", "virtio-net-device,netdev=usernet0,mac="+HostInterfaceMAC.String())
+	}
+
+	var stdErrBuf bytes.Buffer
+	cmd := exec.CommandContext(ctx, "qemu-system-x86_64", append(baseArgs, extraArgs...)...)
+	cmd.Stdout = opts.SerialPort
+	cmd.Stderr = &stdErrBuf
+
+	cmd.ExtraFiles = append(cmd.ExtraFiles, opts.ExtraChardevs...)
+	cmd.ExtraFiles = append(cmd.ExtraFiles, opts.ExtraNetworkInterfaces...)
+
+	err := cmd.Run()
+	var exerr *exec.ExitError
+	if err != nil && errors.As(err, &exerr) {
+		exerr.Stderr = stdErrBuf.Bytes()
+		newErr := QEMUError(*exerr)
+		return &newErr
+	}
+	return err
+}
+
+// QEMUError is a special type of ExitError used when QEMU fails. In addition to normal ExitError features it
+// prints stderr for debugging.
+type QEMUError exec.ExitError
+
+func (e *QEMUError) Error() string {
+	return fmt.Sprintf("%v: %v", e.String(), string(e.Stderr))
+}
+
+// NanoswitchPorts contains all ports forwarded by Nanoswitch to the first VM
+var NanoswitchPorts = []uint16{
+	common.ExternalServicePort,
+	common.DebugServicePort,
+	common.KubernetesAPIPort,
+}
+
+// ClusterOptions contains all options for launching a Smalltown cluster
+type ClusterOptions struct {
+	// The number of nodes this cluster should be started with initially
+	NumNodes int
+}
+
+// LaunchCluster launches a cluster of Smalltown VMs together with a Nanoswitch instance to network them all together.
+func LaunchCluster(ctx context.Context, opts ClusterOptions) (apb.NodeDebugServiceClient, PortMap, error) {
+	var switchPorts []*os.File
+	var vmPorts []*os.File
+	for i := 0; i < opts.NumNodes; i++ {
+		switchPort, vmPort, err := NewSocketPair()
+		if err != nil {
+			return nil, nil, fmt.Errorf("failed to get socketpair: %w", err)
+		}
+		switchPorts = append(switchPorts, switchPort)
+		vmPorts = append(vmPorts, vmPort)
+	}
+
+	if opts.NumNodes == 0 {
+		return nil, nil, errors.New("refusing to start cluster with zero nodes")
+	}
+
+	if opts.NumNodes > 2 {
+		return nil, nil, errors.New("launching more than 2 nodes is unsupported pending replacement of golden tickets")
+	}
+
+	go func() {
+		if err := Launch(ctx, Options{ConnectToSocket: vmPorts[0]}); err != nil {
+			// Launch() only terminates when QEMU has terminated. At that point our function probably doesn't run anymore
+			// so we have no way of communicating the error back up, so let's just log it. Also a failure in launching
+			// VMs should be very visible by the unavailability of the clients we return.
+			log.Printf("Failed to launch vm0: %v", err)
+		}
+	}()
+
+	portMap, err := ConflictFreePortMap(NanoswitchPorts)
+	if err != nil {
+		return nil, nil, fmt.Errorf("failed to allocate ephemeral ports: %w", err)
+	}
+
+	go func() {
+		if err := RunMicroVM(ctx, &MicroVMOptions{
+			KernelPath:             "metropolis/test/ktest/linux-testing.elf",
+			InitramfsPath:          "metropolis/test/nanoswitch/initramfs.lz4",
+			ExtraNetworkInterfaces: switchPorts,
+			PortMap:                portMap,
+		}); err != nil {
+			log.Printf("Failed to launch nanoswitch: %v", err)
+		}
+	}()
+	copts := []grpcretry.CallOption{
+		grpcretry.WithBackoff(grpcretry.BackoffExponential(100 * time.Millisecond)),
+	}
+	conn, err := portMap.DialGRPC(common.DebugServicePort, grpc.WithInsecure(),
+		grpc.WithUnaryInterceptor(grpcretry.UnaryClientInterceptor(copts...)))
+	if err != nil {
+		return nil, nil, fmt.Errorf("failed to dial debug service: %w", err)
+	}
+	defer conn.Close()
+	debug := apb.NewNodeDebugServiceClient(conn)
+
+	if opts.NumNodes == 2 {
+		res, err := debug.GetGoldenTicket(ctx, &apb.GetGoldenTicketRequest{
+			// HACK: this is assigned by DHCP, and we assume that everything goes well.
+			ExternalIp: "10.1.0.3",
+		}, grpcretry.WithMax(10))
+		if err != nil {
+			return nil, nil, fmt.Errorf("failed to get golden ticket: %w", err)
+		}
+
+		ec := &apb.EnrolmentConfig{
+			GoldenTicket: res.Ticket,
+		}
+
+		go func() {
+			if err := Launch(ctx, Options{ConnectToSocket: vmPorts[1], EnrolmentConfig: ec}); err != nil {
+				log.Printf("Failed to launch vm1: %v", err)
+			}
+		}()
+	}
+
+	return debug, portMap, nil
+}
diff --git a/metropolis/test/nanoswitch/BUILD b/metropolis/test/nanoswitch/BUILD
new file mode 100644
index 0000000..fc4f932
--- /dev/null
+++ b/metropolis/test/nanoswitch/BUILD
@@ -0,0 +1,40 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library")
+load("//metropolis/node/build:def.bzl", "smalltown_initramfs")
+
+go_library(
+    name = "go_default_library",
+    srcs = ["nanoswitch.go"],
+    importpath = "git.monogon.dev/source/nexantic.git/metropolis/test/nanoswitch",
+    visibility = ["//visibility:private"],
+    deps = [
+        "//metropolis/node:go_default_library",
+        "//metropolis/node/common/supervisor:go_default_library",
+        "//metropolis/node/core/network/dhcp4c:go_default_library",
+        "//metropolis/node/core/network/dhcp4c/callback:go_default_library",
+        "//metropolis/test/launch:go_default_library",
+        "@com_github_google_nftables//:go_default_library",
+        "@com_github_google_nftables//expr:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4:go_default_library",
+        "@com_github_insomniacslk_dhcp//dhcpv4/server4:go_default_library",
+        "@com_github_vishvananda_netlink//:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+go_binary(
+    name = "nanoswitch",
+    embed = [":go_default_library"],
+    pure = "on",
+    visibility = ["//visibility:public"],
+)
+
+smalltown_initramfs(
+    name = "initramfs",
+    files = {
+        ":nanoswitch": "/init",
+
+        # CA Certificate bundle
+        "@cacerts//file": "/etc/ssl/cert.pem",
+    },
+    visibility = ["//visibility:public"],
+)
diff --git a/metropolis/test/nanoswitch/nanoswitch.go b/metropolis/test/nanoswitch/nanoswitch.go
new file mode 100644
index 0000000..1fe6740
--- /dev/null
+++ b/metropolis/test/nanoswitch/nanoswitch.go
@@ -0,0 +1,301 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// nanoswitch is a virtualized switch/router combo intended for testing.
+// It uses the first interface as an external interface to connect to the host and pass traffic in and out. All other
+// interfaces are switched together and served by a built-in DHCP server. Traffic from that network to the
+// SLIRP/external network is SNATed as the host-side SLIRP ignores routed packets.
+// It also has built-in userspace proxying support for debugging.
+package main
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"net"
+	"os"
+	"time"
+
+	"github.com/google/nftables"
+	"github.com/google/nftables/expr"
+	"github.com/insomniacslk/dhcp/dhcpv4"
+	"github.com/insomniacslk/dhcp/dhcpv4/server4"
+	"github.com/vishvananda/netlink"
+	"golang.org/x/sys/unix"
+
+	common "git.monogon.dev/source/nexantic.git/metropolis/node"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/common/supervisor"
+	"git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c"
+	dhcpcb "git.monogon.dev/source/nexantic.git/metropolis/node/core/network/dhcp4c/callback"
+	"git.monogon.dev/source/nexantic.git/metropolis/test/launch"
+)
+
+var switchIP = net.IP{10, 1, 0, 1}
+var switchSubnetMask = net.CIDRMask(24, 32)
+
+// defaultLeaseOptions sets the lease options needed to properly configure connectivity to nanoswitch
+func defaultLeaseOptions(reply *dhcpv4.DHCPv4) {
+	reply.GatewayIPAddr = switchIP
+	reply.UpdateOption(dhcpv4.OptDNS(net.IPv4(10, 42, 0, 3))) // SLIRP fake DNS server
+	reply.UpdateOption(dhcpv4.OptRouter(switchIP))
+	reply.UpdateOption(dhcpv4.OptIPAddressLeaseTime(30 * time.Second)) // Make sure we exercise our DHCP client in E2E tests
+	reply.UpdateOption(dhcpv4.OptSubnetMask(switchSubnetMask))
+}
+
+// runDHCPServer runs an extremely minimal DHCP server with most options hardcoded, a wrapping bump allocator for the
+// IPs, 12h Lease timeout and no support for DHCP collision detection.
+func runDHCPServer(link netlink.Link) supervisor.Runnable {
+	currentIP := net.IP{10, 1, 0, 1}
+
+	return func(ctx context.Context) error {
+		laddr := net.UDPAddr{
+			IP:   net.IPv4(0, 0, 0, 0),
+			Port: 67,
+		}
+		server, err := server4.NewServer(link.Attrs().Name, &laddr, func(conn net.PacketConn, peer net.Addr, m *dhcpv4.DHCPv4) {
+			if m == nil {
+				return
+			}
+			reply, err := dhcpv4.NewReplyFromRequest(m)
+			if err != nil {
+				supervisor.Logger(ctx).Warningf("Failed to generate DHCP reply: %v", err)
+				return
+			}
+			reply.UpdateOption(dhcpv4.OptServerIdentifier(switchIP))
+			reply.ServerIPAddr = switchIP
+
+			switch m.MessageType() {
+			case dhcpv4.MessageTypeDiscover:
+				reply.UpdateOption(dhcpv4.OptMessageType(dhcpv4.MessageTypeOffer))
+				defaultLeaseOptions(reply)
+				currentIP[3]++ // Works only because it's a /24
+				reply.YourIPAddr = currentIP
+				supervisor.Logger(ctx).Infof("Replying with DHCP IP %s", reply.YourIPAddr.String())
+			case dhcpv4.MessageTypeRequest:
+				reply.UpdateOption(dhcpv4.OptMessageType(dhcpv4.MessageTypeAck))
+				defaultLeaseOptions(reply)
+				if m.RequestedIPAddress() != nil {
+					reply.YourIPAddr = m.RequestedIPAddress()
+				} else {
+					reply.YourIPAddr = m.ClientIPAddr
+				}
+			case dhcpv4.MessageTypeRelease, dhcpv4.MessageTypeDecline:
+				supervisor.Logger(ctx).Info("Ignoring Release/Decline")
+			}
+			if _, err := conn.WriteTo(reply.ToBytes(), peer); err != nil {
+				supervisor.Logger(ctx).Warningf("Cannot reply to client: %v", err)
+			}
+		})
+		if err != nil {
+			return err
+		}
+		supervisor.Signal(ctx, supervisor.SignalHealthy)
+		go func() {
+			<-ctx.Done()
+			server.Close()
+		}()
+		return server.Serve()
+	}
+}
+
+// userspaceProxy listens on port and proxies all TCP connections to the same port on targetIP
+func userspaceProxy(targetIP net.IP, port uint16) supervisor.Runnable {
+	return func(ctx context.Context) error {
+		logger := supervisor.Logger(ctx)
+		tcpListener, err := net.ListenTCP("tcp", &net.TCPAddr{IP: net.IPv4(0, 0, 0, 0), Port: int(port)})
+		if err != nil {
+			return err
+		}
+		supervisor.Signal(ctx, supervisor.SignalHealthy)
+		go func() {
+			<-ctx.Done()
+			tcpListener.Close()
+		}()
+		for {
+			conn, err := tcpListener.AcceptTCP()
+			if err != nil {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				return err
+			}
+			go func(conn *net.TCPConn) {
+				defer conn.Close()
+				upstreamConn, err := net.DialTCP("tcp", nil, &net.TCPAddr{IP: targetIP, Port: int(port)})
+				if err != nil {
+					logger.Infof("Userspace proxy failed to connect to upstream: %v", err)
+					return
+				}
+				defer upstreamConn.Close()
+				go io.Copy(upstreamConn, conn)
+				io.Copy(conn, upstreamConn)
+			}(conn)
+		}
+
+	}
+}
+
+// addNetworkRoutes sets up routing from DHCP
+func addNetworkRoutes(link netlink.Link, addr net.IPNet, gw net.IP) error {
+	if err := netlink.AddrReplace(link, &netlink.Addr{IPNet: &addr}); err != nil {
+		return fmt.Errorf("failed to add DHCP address to network interface \"%v\": %w", link.Attrs().Name, err)
+	}
+
+	if gw.IsUnspecified() {
+		return nil
+	}
+
+	route := &netlink.Route{
+		Dst:   &net.IPNet{IP: net.IPv4(0, 0, 0, 0), Mask: net.IPv4Mask(0, 0, 0, 0)},
+		Gw:    gw,
+		Scope: netlink.SCOPE_UNIVERSE,
+	}
+	if err := netlink.RouteAdd(route); err != nil {
+		return fmt.Errorf("could not add default route: netlink.RouteAdd(%+v): %v", route, err)
+	}
+	return nil
+}
+
+// nfifname converts an interface name into 16 bytes padded with zeroes (for nftables)
+func nfifname(n string) []byte {
+	b := make([]byte, 16)
+	copy(b, []byte(n+"\x00"))
+	return b
+}
+
+func main() {
+	supervisor.New(context.Background(), func(ctx context.Context) error {
+		logger := supervisor.Logger(ctx)
+		logger.Info("Starting NanoSwitch, a tiny TOR switch emulator")
+
+		// Set up target filesystems.
+		for _, el := range []struct {
+			dir   string
+			fs    string
+			flags uintptr
+		}{
+			{"/sys", "sysfs", unix.MS_NOEXEC | unix.MS_NOSUID | unix.MS_NODEV},
+			{"/proc", "proc", unix.MS_NOEXEC | unix.MS_NOSUID | unix.MS_NODEV},
+			{"/dev", "devtmpfs", unix.MS_NOEXEC | unix.MS_NOSUID},
+			{"/dev/pts", "devpts", unix.MS_NOEXEC | unix.MS_NOSUID},
+		} {
+			if err := os.Mkdir(el.dir, 0755); err != nil && !os.IsExist(err) {
+				return fmt.Errorf("could not make %s: %w", el.dir, err)
+			}
+			if err := unix.Mount(el.fs, el.dir, el.fs, el.flags, ""); err != nil {
+				return fmt.Errorf("could not mount %s on %s: %w", el.fs, el.dir, err)
+			}
+		}
+
+		c := &nftables.Conn{}
+
+		links, err := netlink.LinkList()
+		if err != nil {
+			logger.Fatalf("Failed to list links: %v", err)
+		}
+		var externalLink netlink.Link
+		var vmLinks []netlink.Link
+		for _, link := range links {
+			attrs := link.Attrs()
+			if link.Type() == "device" && len(attrs.HardwareAddr) > 0 {
+				if attrs.Flags&net.FlagUp != net.FlagUp {
+					netlink.LinkSetUp(link) // Attempt to take up all ethernet links
+				}
+				if bytes.Equal(attrs.HardwareAddr, launch.HostInterfaceMAC) {
+					externalLink = link
+				} else {
+					vmLinks = append(vmLinks, link)
+				}
+			}
+		}
+		vmBridgeLink := &netlink.Bridge{LinkAttrs: netlink.LinkAttrs{Name: "vmbridge", Flags: net.FlagUp}}
+		if err := netlink.LinkAdd(vmBridgeLink); err != nil {
+			logger.Fatalf("Failed to create vmbridge: %v", err)
+		}
+		for _, link := range vmLinks {
+			if err := netlink.LinkSetMaster(link, vmBridgeLink); err != nil {
+				logger.Fatalf("Failed to add VM interface to bridge: %v", err)
+			}
+			logger.Infof("Assigned interface %s to bridge", link.Attrs().Name)
+		}
+		if err := netlink.AddrReplace(vmBridgeLink, &netlink.Addr{IPNet: &net.IPNet{IP: switchIP, Mask: switchSubnetMask}}); err != nil {
+			logger.Fatalf("Failed to assign static IP to vmbridge: %v", err)
+		}
+		if externalLink != nil {
+			nat := c.AddTable(&nftables.Table{
+				Family: nftables.TableFamilyIPv4,
+				Name:   "nat",
+			})
+
+			postrouting := c.AddChain(&nftables.Chain{
+				Name:     "postrouting",
+				Hooknum:  nftables.ChainHookPostrouting,
+				Priority: nftables.ChainPriorityNATSource,
+				Table:    nat,
+				Type:     nftables.ChainTypeNAT,
+			})
+
+			// Masquerade/SNAT all traffic going out of the external interface
+			c.AddRule(&nftables.Rule{
+				Table: nat,
+				Chain: postrouting,
+				Exprs: []expr.Any{
+					&expr.Meta{Key: expr.MetaKeyOIFNAME, Register: 1},
+					&expr.Cmp{
+						Op:       expr.CmpOpEq,
+						Register: 1,
+						Data:     nfifname(externalLink.Attrs().Name),
+					},
+					&expr.Masq{},
+				},
+			})
+
+			if err := c.Flush(); err != nil {
+				panic(err)
+			}
+
+			netIface := &net.Interface{
+				Name:         externalLink.Attrs().Name,
+				MTU:          externalLink.Attrs().MTU,
+				Index:        externalLink.Attrs().Index,
+				Flags:        externalLink.Attrs().Flags,
+				HardwareAddr: externalLink.Attrs().HardwareAddr,
+			}
+			dhcpClient, err := dhcp4c.NewClient(netIface)
+			if err != nil {
+				logger.Fatalf("Failed to create DHCP client: %v", err)
+			}
+			dhcpClient.RequestedOptions = []dhcpv4.OptionCode{dhcpv4.OptionRouter}
+			dhcpClient.LeaseCallback = dhcpcb.Compose(dhcpcb.ManageIP(externalLink), dhcpcb.ManageDefaultRoute(externalLink))
+			supervisor.Run(ctx, "dhcp-client", dhcpClient.Run)
+			if err := ioutil.WriteFile("/proc/sys/net/ipv4/ip_forward", []byte("1\n"), 0644); err != nil {
+				logger.Fatalf("Failed to write ip forwards: %v", err)
+			}
+		} else {
+			logger.Info("No upstream interface detected")
+		}
+		supervisor.Run(ctx, "dhcp-server", runDHCPServer(vmBridgeLink))
+		supervisor.Run(ctx, "proxy-ext1", userspaceProxy(net.IPv4(10, 1, 0, 2), common.ExternalServicePort))
+		supervisor.Run(ctx, "proxy-dbg1", userspaceProxy(net.IPv4(10, 1, 0, 2), common.DebugServicePort))
+		supervisor.Run(ctx, "proxy-k8s-api1", userspaceProxy(net.IPv4(10, 1, 0, 2), common.KubernetesAPIPort))
+		supervisor.Signal(ctx, supervisor.SignalHealthy)
+		supervisor.Signal(ctx, supervisor.SignalDone)
+		return nil
+	})
+	select {}
+}