m/n/core: run dedicated PID 1 reaper
This introduces minit, a tiny init implementation, written in C, built
against musl. It does one thing: reap children. No support for TTY, no
configurability, just the bare minimum for a working system.
We also drive-by remove some dead code from main.go.
This solves https://github.com/monogon-dev/monogon/issues/15
Change-Id: I666ff2042f19639465ff918590a39b8e219ee7d6
Reviewed-on: https://review.monogon.dev/c/monogon/+/346
Reviewed-by: Lorenz Brun <lorenz@monogon.tech>
diff --git a/metropolis/node/core/main.go b/metropolis/node/core/main.go
index fa768c2..5d495fa 100644
--- a/metropolis/node/core/main.go
+++ b/metropolis/node/core/main.go
@@ -18,15 +18,9 @@
import (
"context"
- "crypto/ed25519"
- "crypto/rand"
- "crypto/x509"
"fmt"
- "log"
- "math/big"
"net"
"os"
- "os/signal"
"runtime/debug"
"time"
@@ -51,14 +45,16 @@
func main() {
defer func() {
if r := recover(); r != nil {
- fmt.Println("Init panicked:", r)
+ fmt.Fprintf(os.Stderr, "\n\n")
+ fmt.Fprintf(os.Stderr, " Metropolis encountered an uncorrectable error and this node must be restarted.\n")
+ fmt.Fprintf(os.Stderr, " Core panicked: %v\n\n", 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.
+ // TODO(lorenz): Switch this to Reboot when init panics are less likely.
if err := unix.Reboot(unix.LINUX_REBOOT_CMD_POWER_OFF); err != nil {
+ // Best effort, nothing we can do if this fails except printing the error to the
+ // console.
panic(fmt.Sprintf("failed to halt node: %v\n", err))
}
}()
@@ -93,9 +89,6 @@
logger.Info("Starting Metropolis node 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)
}
@@ -114,8 +107,7 @@
}
// 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.
+ // low-level, unrecoverable failure occured.
trapdoor := make(chan struct{})
// Make context for supervisor. We cancel it when we reach the trapdoor.
@@ -242,87 +234,7 @@
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("Metropolis encountered an uncorrectable error and this node must be")
- log.Printf("restarted.")
- log.Printf("")
- 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
+ <-trapdoor
+ logger.Infof("Trapdoor closed, exiting core.")
+ ctxC()
}
diff --git a/metropolis/node/core/minit/BUILD.bazel b/metropolis/node/core/minit/BUILD.bazel
new file mode 100644
index 0000000..573ca92
--- /dev/null
+++ b/metropolis/node/core/minit/BUILD.bazel
@@ -0,0 +1,9 @@
+cc_binary(
+ name = "minit",
+ srcs = [
+ "main.c",
+ ],
+ visibility = [
+ "//metropolis/node:__pkg__",
+ ],
+)
diff --git a/metropolis/node/core/minit/main.c b/metropolis/node/core/minit/main.c
new file mode 100644
index 0000000..f2611b4
--- /dev/null
+++ b/metropolis/node/core/minit/main.c
@@ -0,0 +1,132 @@
+// minit is a barebones Linux-compatible init (PID 1) process.
+//
+// Its goal is to run the Metropolis core executable and reap any children that
+// it stumbles upon. It does not support running under a TTY and is not
+// configurable in any way.
+//
+// The only reason this exists is because Go's child process reaping (when
+// using os/exec.Command) races any PID 1 process reaping, thereby preventing
+// running a complex Go binary as PID 1. In the future this might be rewritten
+// in a memory-safe language like Zig or Rust, but this implementation will do
+// for now, as long as it keeps having basically zero attack surface.
+//
+// This code has been vaguely inspired by github.com/Yelp/dumb-init and
+// github.com/krallin/tini, two already existing minimal init implementations.
+// These, however, attempt to handle being run in a TTY and some
+// configurability, as they're meant to be run in containers. We don't need any
+// of that, and we'd rather have as little C as possible.
+
+#include <errno.h>
+#include <linux/reboot.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/reboot.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+void handle_signal(pid_t child_pid, int signum);
+
+int main() {
+ // Block all signals. We'll unblock them in the child.
+ sigset_t all_signals;
+ sigfillset(&all_signals);
+ sigprocmask(SIG_BLOCK, &all_signals, NULL);
+
+ // Say hello.
+ fprintf(stderr,
+ "\n"
+ " Metropolis Cluster Operating System\n"
+ " Copyright 2020-2021 The Monogon Project Authors\n"
+ "\n"
+ );
+
+
+ pid_t pid = fork();
+ if (pid < 0) {
+ fprintf(stderr, "fork(): %s\n", strerror(errno));
+ return 1;
+ }
+
+ if (pid == 0) {
+ // In the child. Unblock all signals.
+ sigprocmask(SIG_UNBLOCK, &all_signals, NULL);
+ if (setsid() == -1) {
+ fprintf(stderr, "setsid: %s\n", strerror(errno));
+ return 1;
+ }
+
+ // Then, start the core executable.
+ char *argv[] = {
+ "/core",
+ NULL,
+ };
+ execvp(argv[0], argv);
+ fprintf(stderr, "execvpe(/core) failed: %s\n", strerror(errno));
+ return 1;
+ }
+
+ // In the parent. Wait for any signal, then handle it and any other pending
+ // ones.
+ for (;;) {
+ int signum;
+ sigwait(&all_signals, &signum);
+ handle_signal(pid, signum);
+ }
+}
+
+// handle_signal is called by the main reap loop for every signal received. It
+// reaps children if SIGCHLD is received, and otherwise dispatches the signal to
+// its direct child.
+void handle_signal(pid_t child_pid, int signum) {
+ // Anything other than SIGCHLD should just be forwarded to the child.
+ if (signum != SIGCHLD) {
+ kill(-child_pid, signum);
+ return;
+ }
+
+ // A SIGCHLD was received. Go through all children and reap them, checking
+ // if any of them is our direct child.
+
+ // exit_status will be set if the direct child process exited.
+ int exit_status = -1;
+
+ pid_t killed_pid;
+ int status;
+ while ((killed_pid = waitpid(-1, &status, WNOHANG)) > 0) {
+ if (killed_pid != child_pid) {
+ // Something else than our direct child died, just reap it.
+ continue;
+ }
+
+ // Our direct child exited. Translate its status into an exit code.
+ if (WIFEXITED(status)) {
+ // For processes which exited, just use the exit code directly.
+ exit_status = WEXITSTATUS(status);
+ } else if (WIFSIGNALED(status)) {
+ // Otherwise, emulate what sh/bash do and return 128 + the signal
+ // number that the child received.
+ exit_status = 128 + WTERMSIG(status);
+ } else {
+ // Something unexpected happened. Attempt to handle this gracefully,
+ // but complain.
+ fprintf(stderr, "child status not EXITED nor SIGNALED: %d\n", status);
+ exit_status = 1;
+ }
+ }
+
+ // Direct child exited, let's also exit.
+ if (exit_status >= 0) {
+ fprintf(stderr, "\n Metropolis core exited with status: %d\n", exit_status);
+ sync();
+ if (exit_status != 0) {
+ fprintf(stderr, " Disks synced, rebooting in 30 seconds...\n", exit_status);
+ sleep(30);
+ fprintf(stderr, " Rebooting...\n\n", exit_status);
+ } else {
+ fprintf(stderr, " Disks synced, rebooting...\n\n");
+ }
+ reboot(LINUX_REBOOT_CMD_RESTART);
+ }
+}