This commit is contained in:
@@ -0,0 +1,8 @@
|
||||
[package]
|
||||
name = "slim-init"
|
||||
version = "0.1.0"
|
||||
edition = "2024"
|
||||
|
||||
[dependencies]
|
||||
base64 = "0.23"
|
||||
nix = { version = "0.31", features = ["mount", "fs", "process", "user", "reboot"] }
|
||||
@@ -0,0 +1,429 @@
|
||||
//! slim universal init.
|
||||
//!
|
||||
//! Injected by `slim build` at /slim/init and invoked via init=/slim/init.
|
||||
//! Sets up devices, filesystems, networking, and 9p shares, then execs the
|
||||
//! container's CMD/ENTRYPOINT (or a runtime override) with execvp so the
|
||||
//! command becomes PID 1 directly.
|
||||
|
||||
use base64::Engine;
|
||||
use nix::mount::{MsFlags, mount};
|
||||
use nix::sys::reboot::{RebootMode, reboot};
|
||||
use nix::sys::stat::{Mode, SFlag, makedev, mknod};
|
||||
use nix::sys::wait::waitpid;
|
||||
use nix::unistd::{ForkResult, Gid, Uid, User, execvp, fork, setgid, setuid};
|
||||
use std::env;
|
||||
use std::ffi::CString;
|
||||
use std::fs;
|
||||
use std::os::unix::fs::PermissionsExt;
|
||||
use std::path::Path;
|
||||
use std::process::Command;
|
||||
|
||||
fn warn(msg: &str) {
|
||||
eprintln!("slim-init: {msg}");
|
||||
}
|
||||
|
||||
/// Run a fallible operation, logging on error.
|
||||
fn try_io<F, T>(label: &str, f: F)
|
||||
where
|
||||
F: FnOnce() -> std::io::Result<T>,
|
||||
{
|
||||
if let Err(e) = f() {
|
||||
warn(&format!("{label}: {e}"));
|
||||
}
|
||||
}
|
||||
|
||||
/// Run a command, logging on error or non-zero exit.
|
||||
fn try_cmd(label: &str, cmd: &mut Command) {
|
||||
match cmd.status() {
|
||||
Ok(status) => {
|
||||
if !status.success() {
|
||||
warn(&format!("{label}: exited with {status}"));
|
||||
}
|
||||
}
|
||||
Err(e) => warn(&format!("{label}: {e}")),
|
||||
}
|
||||
}
|
||||
|
||||
/// The kernel passes a minimal environment to init. Set a default PATH
|
||||
/// so that `ip` and other tools can be found by name.
|
||||
fn ensure_path() {
|
||||
if env::var("PATH").is_err() {
|
||||
// SAFETY: we are single-threaded before fork/exec.
|
||||
unsafe {
|
||||
env::set_var(
|
||||
"PATH",
|
||||
"/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn find_in_path(prog: &str) -> Option<String> {
|
||||
let path = env::var("PATH").ok()?;
|
||||
for dir in path.split(':') {
|
||||
let candidate = format!("{dir}/{prog}");
|
||||
if Path::new(&candidate).exists() {
|
||||
return Some(candidate);
|
||||
}
|
||||
}
|
||||
None
|
||||
}
|
||||
|
||||
fn try_mount(
|
||||
source: Option<&str>,
|
||||
target: &str,
|
||||
fstype: Option<&str>,
|
||||
flags: MsFlags,
|
||||
data: Option<&str>,
|
||||
) {
|
||||
if let Err(e) = mount(source, target, fstype, flags, data) {
|
||||
warn(&format!("mount {target}: {e}"));
|
||||
}
|
||||
}
|
||||
|
||||
fn setup_filesystems() {
|
||||
if !Path::new("/dev/console").exists() {
|
||||
try_io("mknod /dev/console", || {
|
||||
mknod(
|
||||
"/dev/console",
|
||||
SFlag::S_IFCHR,
|
||||
Mode::S_IRUSR | Mode::S_IWUSR,
|
||||
makedev(5, 1),
|
||||
)
|
||||
.map_err(std::io::Error::from)
|
||||
});
|
||||
}
|
||||
|
||||
let dirs = [
|
||||
"/proc",
|
||||
"/sys",
|
||||
"/dev/pts",
|
||||
"/dev/shm",
|
||||
"/run",
|
||||
"/tmp",
|
||||
"/sys/fs/cgroup",
|
||||
];
|
||||
for d in &dirs {
|
||||
try_io(&format!("mkdir {d}"), || fs::create_dir_all(d));
|
||||
}
|
||||
|
||||
try_mount(Some("proc"), "/proc", Some("proc"), MsFlags::empty(), None);
|
||||
try_mount(Some("sysfs"), "/sys", Some("sysfs"), MsFlags::empty(), None);
|
||||
try_mount(
|
||||
Some("devpts"),
|
||||
"/dev/pts",
|
||||
Some("devpts"),
|
||||
MsFlags::empty(),
|
||||
None,
|
||||
);
|
||||
try_mount(
|
||||
Some("tmpfs"),
|
||||
"/dev/shm",
|
||||
Some("tmpfs"),
|
||||
MsFlags::empty(),
|
||||
None,
|
||||
);
|
||||
try_mount(
|
||||
Some("tmpfs"),
|
||||
"/run",
|
||||
Some("tmpfs"),
|
||||
MsFlags::empty(),
|
||||
Some("mode=755"),
|
||||
);
|
||||
try_mount(
|
||||
Some("tmpfs"),
|
||||
"/tmp",
|
||||
Some("tmpfs"),
|
||||
MsFlags::empty(),
|
||||
Some("mode=1777"),
|
||||
);
|
||||
|
||||
// cgroup2
|
||||
try_mount(
|
||||
Some("cgroup2"),
|
||||
"/sys/fs/cgroup",
|
||||
Some("cgroup2"),
|
||||
MsFlags::empty(),
|
||||
None,
|
||||
);
|
||||
if let Ok(controllers) = fs::read_to_string("/sys/fs/cgroup/cgroup.controllers") {
|
||||
for c in controllers.split_whitespace() {
|
||||
try_io("write cgroup.subtree_control", || {
|
||||
fs::write("/sys/fs/cgroup/cgroup.subtree_control", format!("+{c}"))
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn setup_rootless_prereqs() {
|
||||
// make-rshared /
|
||||
if let Err(e) =
|
||||
mount::<str, str, str, str>(None, "/", None, MsFlags::MS_REC | MsFlags::MS_SHARED, None)
|
||||
{
|
||||
warn(&format!("make-rshared /: {e}"));
|
||||
}
|
||||
|
||||
// chmod u+s newuidmap newgidmap
|
||||
for bin in &["/usr/bin/newuidmap", "/usr/bin/newgidmap"] {
|
||||
match fs::metadata(bin) {
|
||||
Ok(meta) => {
|
||||
let mut perms = meta.permissions();
|
||||
perms.set_mode(perms.mode() | 0o4000);
|
||||
try_io(&format!("chmod u+s {bin}"), || {
|
||||
fs::set_permissions(bin, perms)
|
||||
});
|
||||
}
|
||||
Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
|
||||
Err(e) => warn(&format!("stat {bin}: {e}")),
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn setup_timezone() {
|
||||
try_io("symlink timezone", || {
|
||||
std::os::unix::fs::symlink("/usr/share/zoneinfo/Europe/Stockholm", "/etc/localtime")
|
||||
});
|
||||
}
|
||||
|
||||
fn setup_networking() {
|
||||
let ip = find_in_path("ip").unwrap_or_else(|| "/sbin/ip".to_string());
|
||||
try_cmd(
|
||||
"ip link set lo up",
|
||||
Command::new(&ip).args(["link", "set", "lo", "up"]),
|
||||
);
|
||||
try_cmd(
|
||||
"ip link set eth0 up",
|
||||
Command::new(&ip).args(["link", "set", "eth0", "up"]),
|
||||
);
|
||||
try_cmd(
|
||||
"ip addr add 10.0.2.15/24 dev eth0",
|
||||
Command::new(&ip).args(["addr", "add", "10.0.2.15/24", "dev", "eth0"]),
|
||||
);
|
||||
try_cmd(
|
||||
"ip route add default via 10.0.2.2",
|
||||
Command::new(&ip).args(["route", "add", "default", "via", "10.0.2.2"]),
|
||||
);
|
||||
try_io("write /etc/resolv.conf", || {
|
||||
fs::write("/etc/resolv.conf", "nameserver 10.0.2.3\n")
|
||||
});
|
||||
}
|
||||
|
||||
fn setup_9p_shares(workdir: &str) {
|
||||
let cmdline = match fs::read_to_string("/proc/cmdline") {
|
||||
Ok(c) => c,
|
||||
Err(e) => {
|
||||
warn(&format!("read /proc/cmdline: {e}"));
|
||||
return;
|
||||
}
|
||||
};
|
||||
let b64 = base64::engine::general_purpose::STANDARD;
|
||||
|
||||
for tok in cmdline.split_whitespace() {
|
||||
if let Some(rest) = tok.strip_prefix("slim.mount=") {
|
||||
let Some((tag, dest_b64)) = rest.split_once(':') else {
|
||||
continue;
|
||||
};
|
||||
let dest = match b64.decode(dest_b64) {
|
||||
Ok(bytes) => match String::from_utf8(bytes) {
|
||||
Ok(s) => s,
|
||||
Err(e) => {
|
||||
warn(&format!("9p: invalid UTF-8 in guest path: {e}"));
|
||||
continue;
|
||||
}
|
||||
},
|
||||
Err(e) => {
|
||||
warn(&format!("9p: base64 decode failed: {e}"));
|
||||
continue;
|
||||
}
|
||||
};
|
||||
|
||||
// Resolve relative guest paths against WORKINGDIR
|
||||
let dest = if dest.starts_with('/') {
|
||||
dest
|
||||
} else {
|
||||
format!("{}/{}", workdir, dest)
|
||||
};
|
||||
|
||||
try_io(&format!("mkdir {dest}"), || fs::create_dir_all(&dest));
|
||||
try_mount(
|
||||
Some(tag),
|
||||
&dest,
|
||||
Some("9p"),
|
||||
MsFlags::empty(),
|
||||
Some("trans=virtio,version=9p2000.L"),
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn read_file_opt(path: &str) -> Option<String> {
|
||||
fs::read_to_string(path)
|
||||
.ok()
|
||||
.filter(|s| !s.trim().is_empty())
|
||||
}
|
||||
|
||||
fn split_null_delimited(data: &[u8]) -> Vec<String> {
|
||||
data.split(|&b| b == 0)
|
||||
.filter(|s| !s.is_empty())
|
||||
.map(|s| String::from_utf8_lossy(s).into_owned())
|
||||
.collect()
|
||||
}
|
||||
|
||||
fn read_argv() -> Vec<String> {
|
||||
// Check for runtime override: slim.cmd=<base64(null-delimited argv)>
|
||||
let b64 = base64::engine::general_purpose::STANDARD;
|
||||
if let Ok(cmdline) = fs::read_to_string("/proc/cmdline") {
|
||||
for tok in cmdline.split_whitespace() {
|
||||
if let Some(rest) = tok.strip_prefix("slim.cmd=")
|
||||
&& let Ok(decoded) = b64.decode(rest)
|
||||
{
|
||||
let argv = split_null_delimited(&decoded);
|
||||
if !argv.is_empty() {
|
||||
return argv;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Fall back to build-time argv from /slim/exec
|
||||
match fs::read("/slim/exec") {
|
||||
Ok(data) => {
|
||||
let argv = split_null_delimited(&data);
|
||||
if !argv.is_empty() {
|
||||
return argv;
|
||||
}
|
||||
}
|
||||
Err(e) => warn(&format!("read /slim/exec: {e}")),
|
||||
}
|
||||
|
||||
warn("no command found, falling back to /bin/sh");
|
||||
vec!["/bin/sh".to_string()]
|
||||
}
|
||||
|
||||
fn apply_env() {
|
||||
if let Ok(data) = fs::read("/slim/env") {
|
||||
for entry in split_null_delimited(&data) {
|
||||
if let Some((key, val)) = entry.split_once('=') {
|
||||
// SAFETY: we are single-threaded before fork/exec.
|
||||
unsafe { env::set_var(key, val) };
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn apply_workdir() {
|
||||
if let Some(dir) = read_file_opt("/slim/workdir") {
|
||||
try_io("set working directory", || env::set_current_dir(&dir));
|
||||
}
|
||||
}
|
||||
|
||||
struct UserSpec {
|
||||
uid: Uid,
|
||||
gid: Gid,
|
||||
}
|
||||
|
||||
fn resolve_user(spec: &str) -> Option<UserSpec> {
|
||||
let (user_part, gid_part) = match spec.split_once(':') {
|
||||
Some((u, g)) => (u, Some(g)),
|
||||
None => (spec, None),
|
||||
};
|
||||
|
||||
let (uid, default_gid) = if let Ok(numeric) = user_part.parse::<u32>() {
|
||||
let user = User::from_uid(Uid::from_raw(numeric)).ok().flatten();
|
||||
let gid = user
|
||||
.as_ref()
|
||||
.map(|u| u.gid)
|
||||
.unwrap_or(Gid::from_raw(numeric));
|
||||
(Uid::from_raw(numeric), gid)
|
||||
} else {
|
||||
let user = User::from_name(user_part).ok().flatten()?;
|
||||
(user.uid, user.gid)
|
||||
};
|
||||
|
||||
let gid = match gid_part {
|
||||
Some(g) => match g.parse::<u32>() {
|
||||
Ok(n) => Gid::from_raw(n),
|
||||
Err(_) => default_gid,
|
||||
},
|
||||
None => default_gid,
|
||||
};
|
||||
|
||||
Some(UserSpec { uid, gid })
|
||||
}
|
||||
|
||||
fn exec_command(argv: &[String], user: Option<UserSpec>) -> ! {
|
||||
let cstrings: Vec<CString> = argv
|
||||
.iter()
|
||||
.map(|s| CString::new(s.as_str()).unwrap_or_default())
|
||||
.collect();
|
||||
let c_prog = cstrings[0].clone();
|
||||
let c_refs: Vec<&CString> = cstrings.iter().collect();
|
||||
|
||||
match user {
|
||||
None => {
|
||||
// No privilege drop: exec directly, PID 1 becomes the command.
|
||||
let Err(e) = execvp(&c_prog, &c_refs);
|
||||
warn(&format!("execvp {}: {e}", argv[0]));
|
||||
let _ = reboot(RebootMode::RB_POWER_OFF);
|
||||
std::process::exit(1);
|
||||
}
|
||||
Some(user) => {
|
||||
// Drop privileges in a child so PID 1 (root) can poweroff afterwards.
|
||||
match unsafe { fork() } {
|
||||
Ok(ForkResult::Child) => {
|
||||
// setgid before setuid to avoid losing privileges
|
||||
if let Err(e) = setgid(user.gid) {
|
||||
warn(&format!("setgid: {e}"));
|
||||
}
|
||||
if let Err(e) = setuid(user.uid) {
|
||||
warn(&format!("setuid: {e}"));
|
||||
}
|
||||
let Err(e) = execvp(&c_prog, &c_refs);
|
||||
warn(&format!("execvp {}: {e}", argv[0]));
|
||||
std::process::exit(127);
|
||||
}
|
||||
Ok(ForkResult::Parent { child }) => {
|
||||
let _ = waitpid(child, None);
|
||||
let _ = reboot(RebootMode::RB_POWER_OFF);
|
||||
std::process::exit(0);
|
||||
}
|
||||
Err(e) => {
|
||||
warn(&format!("fork: {e}"));
|
||||
let _ = reboot(RebootMode::RB_POWER_OFF);
|
||||
std::process::exit(1);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn main() {
|
||||
ensure_path();
|
||||
|
||||
// === Devices & special filesystems ===
|
||||
setup_filesystems();
|
||||
|
||||
// === Rootless container prerequisites (best-effort) ===
|
||||
setup_rootless_prereqs();
|
||||
|
||||
// === Timezone ===
|
||||
setup_timezone();
|
||||
|
||||
// === Networking ===
|
||||
setup_networking();
|
||||
|
||||
// === 9p shares ===
|
||||
let workdir = read_file_opt("/slim/workdir").unwrap_or_default();
|
||||
setup_9p_shares(&workdir);
|
||||
|
||||
// === Environment & working directory ===
|
||||
apply_env();
|
||||
apply_workdir();
|
||||
|
||||
// === Resolve user for privilege drop ===
|
||||
let user = read_file_opt("/slim/user").and_then(|spec| resolve_user(&spec));
|
||||
|
||||
// === Execute the configured command ===
|
||||
let argv = read_argv();
|
||||
exec_command(&argv, user);
|
||||
}
|
||||
Reference in New Issue
Block a user