A very simple initrd implementation that mounts / and /usr

simple_initrd.c

/*
 * Author: Christian Seiler <[email protected]>
 * License: GPLv3+
 *
 * IMPORTANT NOTE:
 * This was just a proof of concept, there's a better version available from
 * https://github.com/chris-se/tiny-initrd
 *
 * Usage: musl-gcc -O2 -Wall -Wextra -static -o init simple_initrd.c
 *        mkdir initramfs initramfs/{dev,proc,target}
 *        cp init initramfs/
 *        strip initramfs/init
 *        cd initramfs ; find . | cpio --quiet -R 0:0 -o -H newc | gzip > ../initrd.img ; cd ..
 *
 * (You don't HAVE to use musl-gcc, but if you want your binary to be small
 * then you really should use something like musl, dietlibc, uclibc or
 * similar. In principle this could also work with dynamic linking if you
 * add the dynamlic linker to your initrd image + the libraries required,
 * but again, this would defeat the purpose of having a very small binary.)
 *
 * Specify root=/dev/... and x-mount.usr=/dev/... on kernel command line to have
 * this initrd mount / and /usr before the system is booted. x-mount.usr= is
 * optional.
 *
 * This relies on devtmpfs and kernel device names. root=UUID=... is not supported,
 * neither is root=/dev/disk/by-... or similar symlinks. Only native kernel device
 * names are supported.
 *
 * Everything required for mounting / and /usr needs to be compiled into the
 * kernel, no modules will be loaded.
 *
 * This is basically meant to replace the no-initrd situation, but also mount /usr.
 * It's not supposed to support fancy things. If you need that, use a proper initrd
 * and this code isn't for you.
 *
 * First draft, designed to be small in code size. Could probably be improved
 * quite a bit...
 *
 * compile-tested on amd64, but should be portable. Tell me if it isn't.
 */

#include <sys/types.h>
#include <sys/stat.h>
#include <sys/mount.h>
#include <sys/uio.h>
#include <sys/time.h>
#include <errno.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <stdlib.h>
#include <time.h>

#ifndef CMDLINE_FILE
#define CMDLINE_FILE "/proc/cmdline"
#endif /* !defined(CMDLINE_FILE) */

#ifndef CMDLINE_BUFSZ
#define CMDLINE_BUFSZ 4096
#endif /* !defined(CMDLINE_BUFSZ) */

#ifndef FSTYPES_FILE
#define FSTYPES_FILE "/proc/filesystems"
#endif /* !defined(FSTYPES_FILE) */

#ifndef FSTYPES_BUFSZ
#define FSTYPES_BUFSZ 16384
#endif /* !defined(FSTYPES_BUFSZ) */

#ifndef FS_BUFSZ
#define FS_BUFSZ 256
#endif /* !defined(FS_BUFSZ) */

#ifndef DEVTMPFS_MOUNTOPTS
#define DEVTMPFS_MOUNTOPTS "size=10240k,mode=0755"
#endif /* !defined(DEVTMPFS_MOUNTOPTS) */

#ifndef FSDEV_TIMEOUT
#define FSDEV_TIMEOUT 180
#endif /* !defined(FSDEV_TIMEOUT) */

#define N_MAX_SUPPORTED_FILESYSTEMS 256
#define N_MAX_FSTYPELEN 32

static void panic(int e, const char *message) __attribute__((noreturn));

void panic(int e, const char *message)
{
  struct iovec iov[4];
  const char *error = e != 0 ? strerror(e) : "";

  /* Don't use stdio functions, because we link statically
   * and they bloat the binary. */
  iov[0].iov_base = (char *)message;
  iov[0].iov_len = strlen(message);
  iov[1].iov_base = (char *)": ";
  iov[1].iov_len = 2;
  iov[2].iov_base = (char *)error;
  iov[2].iov_len = strlen(error);
  iov[3].iov_base = (char *)"\n";
  iov[3].iov_len = 1;

  writev(2, iov, 4);
  _exit(1);
}

void warn3(const char *p1, const char *p2, const char *p3)
{
  struct iovec iov[4];

  /* Don't use stdio functions, because we link statically
   * and they bloat the binary. */
  iov[0].iov_base = (char *)p1;
  iov[0].iov_len = strlen(p1);
  iov[1].iov_base = (char *)p2;
  iov[1].iov_len = strlen(p2);
  iov[2].iov_base = (char *)p3;
  iov[2].iov_len = strlen(p3);
  iov[3].iov_base = (char *)"\n";
  iov[3].iov_len = 1;

  writev(2, iov, 4);
}

static char rootfs[FS_BUFSZ];
static char usrfs[FS_BUFSZ];
static char init_binary[FS_BUFSZ];

static char supported_filesystems[N_MAX_SUPPORTED_FILESYSTEMS][N_MAX_FSTYPELEN];
static int n_supported_filesystems;

static int do_mount(const char *source, const char *target,
          const char *filesystemtype, unsigned long mountflags,
          const void *data)
{
  /* We need to loop through filesystem types as the kernel doesn't do
   * that for us if we call mount(). libmount does something similar,
   * but we don't want to link against it. */
  if (!filesystemtype && !(mountflags & (MS_MOVE | MS_REMOUNT | MS_BIND))) {
    int i, rc = -1;
    errno = EINVAL;
    for (i = 0; i < n_supported_filesystems; i++) {
      rc = mount(source, target, supported_filesystems[i], mountflags | MS_SILENT, data);
      if (rc == 0)
        return 0;
    }
    return rc;
  }
  return mount(source, target, filesystemtype, mountflags, data);
}

static void parse_cmdline()
{
  int fd;
  ssize_t r;
  char buf[CMDLINE_BUFSZ] = { 0 };
  char *saveptr;
  char *token;
  
  fd = open(CMDLINE_FILE, O_RDONLY);
  if (fd < 0)
    panic(errno, "Couldn't open " CMDLINE_FILE);

  r = read(fd, buf, CMDLINE_BUFSZ - 1);
  if (r < 0)
    panic(errno, "Couldn't read " CMDLINE_FILE);

  close(fd);

  for (token = strtok_r(buf, " \t\r\n", &saveptr); token != NULL; token = strtok_r(NULL, " \t\r\n", &saveptr)) {
    if (!strncmp(token, "root=", 5)) {
      token += 5;
      if (strlen(token) > FS_BUFSZ - 1)
        panic(0, "root= too long");
      strncpy(rootfs, token, FS_BUFSZ);
    } else if (!strncmp(token, "x-mount.usr=", 12)) {
      token += 12;
      if (strlen(token) > FS_BUFSZ - 1)
        panic(0, "x-mount.usr= too long");
      strncpy(usrfs, token, FS_BUFSZ);
    } else if (!strncmp(token, "init=", 5)) {
      token += 5;
      if (strlen(token) > FS_BUFSZ - 1)
        panic(0, "init= too long");
      strncpy(init_binary, token, FS_BUFSZ);
    }
  }
}

static void parse_filesystems()
{
  int fd;
  ssize_t r;
  char buf[FSTYPES_BUFSZ] = { 0 };
  char *saveptr;
  char *line;

  fd = open(FSTYPES_FILE, O_RDONLY);
  if (fd < 0)
    panic(errno, "Couldn't open " FSTYPES_FILE);

  r = read(fd, buf, FSTYPES_BUFSZ - 1);
  if (r < 0)
    panic(errno, "Couldn't read " FSTYPES_FILE);

  close(fd);

  for (line = strtok_r(buf, "\n", &saveptr); line != NULL; line = strtok_r(NULL, "\n", &saveptr)) {
    if (!strncmp(line, "nodev", 5))
      continue;
    while (line[0] == ' ' || line[0] == '\t')
      ++line;
    if (strlen(line) > N_MAX_FSTYPELEN - 1) {
      warn3("[initrd] filesystem type ", line, " is too long for us, ignoring.");
      continue;
    }
    if (n_supported_filesystems == N_MAX_SUPPORTED_FILESYSTEMS) {
      warn3("[initrd] kernel supports too many filesystem types, ignoring ", line, ", sorry.");
      continue;
    }
    strncpy(supported_filesystems[n_supported_filesystems++], line, N_MAX_FSTYPELEN - 1);
  }
}

static void try_exec(int orig_argc, char *const orig_argv[], const char *binary)
{
  char *argv[256];
  int i;

  if (orig_argc > 255)
    panic(0, "Too many arguments to init.");

  argv[0] = (char *)init_binary;
  for (i = 1; i < orig_argc; i++)
    argv[i] = orig_argv[i];
  argv[i] = NULL;

  execv(binary, argv);
}

static void wait_for_filesystems()
{
  /* We don't have udev running, but there is devtmpfs, so we just
   * do a very simple and stupid polling loop to wait until the
   * requested devices are present. This could be improved a bit,
   * but for now it's good enough. */
  int have_devices = 0;
  time_t start;
  struct timeval tv;
  int r;

  r = gettimeofday(&tv, NULL);
  if (r < 0)
    panic(errno, "Couldn't determine current time for timeout.");
  start = tv.tv_sec;

  have_devices = (access(rootfs, F_OK) == 0 && (strlen(usrfs) == 0 || access(usrfs, F_OK) == 0));

  while (!have_devices) {
    r = gettimeofday(&tv, NULL);
    if (r < 0)
      panic(errno, "Couldn't determine current time for timeout");
    if (tv.tv_sec - start > FSDEV_TIMEOUT)
      panic(0, "Timeout while waiting for devices for / and /usr filesystems to appear "
               "(did you specify the correct ones?)");
    /* Sleep for 50 milliseconds, then poll again. */
    struct timespec req = { 0, 50 * 1000 * 1000 };
    struct timespec rem;
    (void)nanosleep(&req, &rem);
    have_devices = (access(rootfs, F_OK) == 0 && (strlen(usrfs) == 0 || access(usrfs, F_OK) == 0));
  }
}

int main(int argc, char **argv)
{
  int r;
  
  r = do_mount("proc", "/proc", "proc", MS_NODEV | MS_NOEXEC | MS_NOSUID, NULL);
  if (r < 0)
    panic(errno, "Couldn't mount /proc filesystem");

  r = do_mount("udev", "/dev", "devtmpfs", 0, DEVTMPFS_MOUNTOPTS);
  if (r < 0)
    panic(errno, "Couldn't mount devtmpfs-/dev filesystem");

  parse_filesystems();
  parse_cmdline();

  if (!strlen(rootfs))
    panic(0, "No root filesystem (root=) specified");

  if (rootfs[0] != '/')
    panic(0, "root filesystem (root=) must be a (non-symlink) kernel device path");
  if (strlen(usrfs) && usrfs[0] != '/')
    panic(0, "/usr filesystem (x-mount.usr=) must be a (non-symlink) kernel device path");

  wait_for_filesystems();

  r = do_mount(rootfs, "/target", NULL, MS_RDONLY, NULL);
  if (r < 0)
    panic(errno, "Couldn't mount / filesystem");

  if (strlen(usrfs)) {
    r = do_mount(usrfs, "/target/usr", NULL, MS_RDONLY, NULL);
    if (r < 0)
      panic(errno, "Couldn't mount /usr filesystem");
  }

  if (access("/target/dev", F_OK) != 0)
    panic(errno, "/dev doesn't exist on rootfs");

  if (access("/target/proc", F_OK) != 0)
    panic(errno, "/proc doesn't exist on rootfs");

  r = do_mount("/dev", "/target/dev", NULL, MS_MOVE, NULL);
  if (r < 0)
    panic(errno, "Couldn't move /dev from initrd to root filesystem");

  r = do_mount("/proc", "/target/proc", NULL, MS_MOVE, NULL);
  if (r < 0)
    panic(errno, "Couldn't move /proc from initrd to root filesystem");

  r = chdir("/target");
  if (r < 0)
    panic(errno, "Couldn't change directory into new root filesystem");

  r = do_mount("/target", "/", NULL, MS_MOVE, NULL);
  if (r < 0)
    panic(errno, "Couldn't move new root filesystem to /");

  r = chroot(".");
  if (r < 0)
    panic(errno, "Couldn't switch rootfs");

  if (strlen(init_binary)) {
    try_exec(argc, argv, init_binary);
  } else {
    try_exec(argc, argv, "/sbin/init");
    try_exec(argc, argv, "/etc/init");
    try_exec(argc, argv, "/bin/init");
    try_exec(argc, argv, "/bin/sh");
  }

  /* Message stolen from Linux's init/main.c */
  panic(0, "No working init found. Try passing init= option to kernel. "
           "See Linux Documentation/init.txt for guidance.");
  return 1;
}