Files
codex/codex-rs/core/src/exec.rs
Michael Bolin 89ef4efdcf fix: overhaul how we spawn commands under seccomp/landlock on Linux (#1086)
Historically, we spawned the Seatbelt and Landlock sandboxes in
substantially different ways:

For **Seatbelt**, we would run `/usr/bin/sandbox-exec` with our policy
specified as an arg followed by the original command:


d1de7bb383/codex-rs/core/src/exec.rs (L147-L219)

For **Landlock/Seccomp**, we would do
`tokio::runtime::Builder::new_current_thread()`, _invoke
Landlock/Seccomp APIs to modify the permissions of that new thread_, and
then spawn the command:


d1de7bb383/codex-rs/core/src/exec_linux.rs (L28-L49)

While it is neat that Landlock/Seccomp supports applying a policy to
only one thread without having to apply it to the entire process, it
requires us to maintain two different codepaths and is a bit harder to
reason about. The tipping point was
https://github.com/openai/codex/pull/1061, in which we had to start
building up the `env` in an unexpected way for the existing
Landlock/Seccomp approach to continue to work.

This PR overhauls things so that we do similar things for Mac and Linux.
It turned out that we were already building our own "helper binary"
comparable to Mac's `sandbox-exec` as part of the `cli` crate:


d1de7bb383/codex-rs/cli/Cargo.toml (L10-L12)

We originally created this to build a small binary to include with the
Node.js version of the Codex CLI to provide support for Linux
sandboxing.

Though the sticky bit is that, at this point, we still want to deploy
the Rust version of Codex as a single, standalone binary rather than a
CLI and a supporting sandboxing binary. To satisfy this goal, we use
"the arg0 trick," in which we:

* use `std::env::current_exe()` to get the path to the CLI that is
currently running
* use the CLI as the `program` for the `Command`
* set `"codex-linux-sandbox"` as arg0 for the `Command`

A CLI that supports sandboxing should check arg0 at the start of the
program. If it is `"codex-linux-sandbox"`, it must invoke
`codex_linux_sandbox::run_main()`, which runs the CLI as if it were
`codex-linux-sandbox`. When acting as `codex-linux-sandbox`, we make the
appropriate Landlock/Seccomp API calls and then use `execvp(3)` to spawn
the original command, so do _replace_ the process rather than spawn a
subprocess. Incidentally, we do this before starting the Tokio runtime,
so the process should only have one thread when `execvp(3)` is called.

Because the `core` crate that needs to spawn the Linux sandboxing is not
a CLI in its own right, this means that every CLI that includes `core`
and relies on this behavior has to (1) implement it and (2) provide the
path to the sandboxing executable. While the path is almost always
`std::env::current_exe()`, we needed to make this configurable for
integration tests, so `Config` now has a `codex_linux_sandbox_exe:
Option<PathBuf>` property to facilitate threading this through,
introduced in https://github.com/openai/codex/pull/1089.

This common pattern is now captured in
`codex_linux_sandbox::run_with_sandbox()` and all of the `main.rs`
functions that should use it have been updated as part of this PR.

The `codex-linux-sandbox` crate added to the Cargo workspace as part of
this PR now has the bulk of the Landlock/Seccomp logic, which makes
`core` a bit simpler. Indeed, `core/src/exec_linux.rs` and
`core/src/landlock.rs` were removed/ported as part of this PR. I also
moved the unit tests for this code into an integration test,
`linux-sandbox/tests/landlock.rs`, in which I use
`env!("CARGO_BIN_EXE_codex-linux-sandbox")` as the value for
`codex_linux_sandbox_exe` since `std::env::current_exe()` is not
appropriate in that case.
2025-05-23 11:37:07 -07:00

543 lines
17 KiB
Rust

#[cfg(unix)]
use std::os::unix::process::ExitStatusExt;
use std::collections::HashMap;
use std::io;
use std::path::Path;
use std::path::PathBuf;
use std::process::ExitStatus;
use std::process::Stdio;
use std::sync::Arc;
use std::time::Duration;
use std::time::Instant;
use tokio::io::AsyncRead;
use tokio::io::AsyncReadExt;
use tokio::io::BufReader;
use tokio::process::Child;
use tokio::process::Command;
use tokio::sync::Notify;
use crate::error::CodexErr;
use crate::error::Result;
use crate::error::SandboxErr;
use crate::protocol::SandboxPolicy;
// Maximum we send for each stream, which is either:
// - 10KiB OR
// - 256 lines
const MAX_STREAM_OUTPUT: usize = 10 * 1024;
const MAX_STREAM_OUTPUT_LINES: usize = 256;
const DEFAULT_TIMEOUT_MS: u64 = 10_000;
// Hardcode these since it does not seem worth including the libc crate just
// for these.
const SIGKILL_CODE: i32 = 9;
const TIMEOUT_CODE: i32 = 64;
const MACOS_SEATBELT_BASE_POLICY: &str = include_str!("seatbelt_base_policy.sbpl");
/// When working with `sandbox-exec`, only consider `sandbox-exec` in `/usr/bin`
/// to defend against an attacker trying to inject a malicious version on the
/// PATH. If /usr/bin/sandbox-exec has been tampered with, then the attacker
/// already has root access.
const MACOS_PATH_TO_SEATBELT_EXECUTABLE: &str = "/usr/bin/sandbox-exec";
/// Experimental environment variable that will be set to some non-empty value
/// if both of the following are true:
///
/// 1. The process was spawned by Codex as part of a shell tool call.
/// 2. SandboxPolicy.has_full_network_access() was false for the tool call.
///
/// We may try to have just one environment variable for all sandboxing
/// attributes, so this may change in the future.
pub const CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR: &str = "CODEX_SANDBOX_NETWORK_DISABLED";
#[derive(Debug, Clone)]
pub struct ExecParams {
pub command: Vec<String>,
pub cwd: PathBuf,
pub timeout_ms: Option<u64>,
pub env: HashMap<String, String>,
}
#[derive(Clone, Copy, Debug, PartialEq)]
pub enum SandboxType {
None,
/// Only available on macOS.
MacosSeatbelt,
/// Only available on Linux.
LinuxSeccomp,
}
pub async fn process_exec_tool_call(
params: ExecParams,
sandbox_type: SandboxType,
ctrl_c: Arc<Notify>,
sandbox_policy: &SandboxPolicy,
codex_linux_sandbox_exe: &Option<PathBuf>,
) -> Result<ExecToolCallOutput> {
let start = Instant::now();
let raw_output_result = match sandbox_type {
SandboxType::None => exec(params, sandbox_policy, ctrl_c).await,
SandboxType::MacosSeatbelt => {
let ExecParams {
command,
cwd,
timeout_ms,
env,
} = params;
let child = spawn_command_under_seatbelt(
command,
sandbox_policy,
cwd,
StdioPolicy::RedirectForShellTool,
env,
)
.await?;
consume_truncated_output(child, ctrl_c, timeout_ms).await
}
SandboxType::LinuxSeccomp => {
let ExecParams {
command,
cwd,
timeout_ms,
env,
} = params;
let codex_linux_sandbox_exe = codex_linux_sandbox_exe
.as_ref()
.ok_or(CodexErr::LandlockSandboxExecutableNotProvided)?;
let child = spawn_command_under_linux_sandbox(
codex_linux_sandbox_exe,
command,
sandbox_policy,
cwd,
StdioPolicy::RedirectForShellTool,
env,
)
.await?;
consume_truncated_output(child, ctrl_c, timeout_ms).await
}
};
let duration = start.elapsed();
match raw_output_result {
Ok(raw_output) => {
let stdout = String::from_utf8_lossy(&raw_output.stdout).to_string();
let stderr = String::from_utf8_lossy(&raw_output.stderr).to_string();
#[cfg(target_family = "unix")]
match raw_output.exit_status.signal() {
Some(TIMEOUT_CODE) => return Err(CodexErr::Sandbox(SandboxErr::Timeout)),
Some(signal) => {
return Err(CodexErr::Sandbox(SandboxErr::Signal(signal)));
}
None => {}
}
let exit_code = raw_output.exit_status.code().unwrap_or(-1);
// NOTE(ragona): This is much less restrictive than the previous check. If we exec
// a command, and it returns anything other than success, we assume that it may have
// been a sandboxing error and allow the user to retry. (The user of course may choose
// not to retry, or in a non-interactive mode, would automatically reject the approval.)
if exit_code != 0 && sandbox_type != SandboxType::None {
return Err(CodexErr::Sandbox(SandboxErr::Denied(
exit_code, stdout, stderr,
)));
}
Ok(ExecToolCallOutput {
exit_code,
stdout,
stderr,
duration,
})
}
Err(err) => {
tracing::error!("exec error: {err}");
Err(err)
}
}
}
pub async fn spawn_command_under_seatbelt(
command: Vec<String>,
sandbox_policy: &SandboxPolicy,
cwd: PathBuf,
stdio_policy: StdioPolicy,
env: HashMap<String, String>,
) -> std::io::Result<Child> {
let args = create_seatbelt_command_args(command, sandbox_policy, &cwd);
let arg0 = None;
spawn_child_async(
PathBuf::from(MACOS_PATH_TO_SEATBELT_EXECUTABLE),
args,
arg0,
cwd,
sandbox_policy,
stdio_policy,
env,
)
.await
}
/// Spawn a shell tool command under the Linux Landlock+seccomp sandbox helper
/// (codex-linux-sandbox).
///
/// Unlike macOS Seatbelt where we directly embed the policy text, the Linux
/// helper accepts a list of `--sandbox-permission`/`-s` flags mirroring the
/// public CLI. We convert the internal [`SandboxPolicy`] representation into
/// the equivalent CLI options.
pub async fn spawn_command_under_linux_sandbox<P>(
codex_linux_sandbox_exe: P,
command: Vec<String>,
sandbox_policy: &SandboxPolicy,
cwd: PathBuf,
stdio_policy: StdioPolicy,
env: HashMap<String, String>,
) -> std::io::Result<Child>
where
P: AsRef<Path>,
{
let args = create_linux_sandbox_command_args(command, sandbox_policy, &cwd);
let arg0 = Some("codex-linux-sandbox");
spawn_child_async(
codex_linux_sandbox_exe.as_ref().to_path_buf(),
args,
arg0,
cwd,
sandbox_policy,
stdio_policy,
env,
)
.await
}
/// Converts the sandbox policy into the CLI invocation for `codex-linux-sandbox`.
fn create_linux_sandbox_command_args(
command: Vec<String>,
sandbox_policy: &SandboxPolicy,
cwd: &Path,
) -> Vec<String> {
let mut linux_cmd: Vec<String> = vec![];
// Translate individual permissions.
// Use high-level helper methods to infer flags when we cannot see the
// exact permission list.
if sandbox_policy.has_full_disk_read_access() {
linux_cmd.extend(["-s", "disk-full-read-access"].map(String::from));
}
if sandbox_policy.has_full_disk_write_access() {
linux_cmd.extend(["-s", "disk-full-write-access"].map(String::from));
} else {
// Derive granular writable paths (includes cwd if `DiskWriteCwd` is
// present).
for root in sandbox_policy.get_writable_roots_with_cwd(cwd) {
// Check if this path corresponds exactly to cwd to map to
// `disk-write-cwd`, otherwise use the generic folder rule.
if root == cwd {
linux_cmd.extend(["-s", "disk-write-cwd"].map(String::from));
} else {
linux_cmd.extend([
"-s".to_string(),
format!("disk-write-folder={}", root.to_string_lossy()),
]);
}
}
}
if sandbox_policy.has_full_network_access() {
linux_cmd.extend(["-s", "network-full-access"].map(String::from));
}
// Separator so that command arguments starting with `-` are not parsed as
// options of the helper itself.
linux_cmd.push("--".to_string());
// Append the original tool command.
linux_cmd.extend(command);
linux_cmd
}
fn create_seatbelt_command_args(
command: Vec<String>,
sandbox_policy: &SandboxPolicy,
cwd: &Path,
) -> Vec<String> {
let (file_write_policy, extra_cli_args) = {
if sandbox_policy.has_full_disk_write_access() {
// Allegedly, this is more permissive than `(allow file-write*)`.
(
r#"(allow file-write* (regex #"^/"))"#.to_string(),
Vec::<String>::new(),
)
} else {
let writable_roots = sandbox_policy.get_writable_roots_with_cwd(cwd);
let (writable_folder_policies, cli_args): (Vec<String>, Vec<String>) = writable_roots
.iter()
.enumerate()
.map(|(index, root)| {
let param_name = format!("WRITABLE_ROOT_{index}");
let policy: String = format!("(subpath (param \"{param_name}\"))");
let cli_arg = format!("-D{param_name}={}", root.to_string_lossy());
(policy, cli_arg)
})
.unzip();
if writable_folder_policies.is_empty() {
("".to_string(), Vec::<String>::new())
} else {
let file_write_policy = format!(
"(allow file-write*\n{}\n)",
writable_folder_policies.join(" ")
);
(file_write_policy, cli_args)
}
}
};
let file_read_policy = if sandbox_policy.has_full_disk_read_access() {
"; allow read-only file operations\n(allow file-read*)"
} else {
""
};
// TODO(mbolin): apply_patch calls must also honor the SandboxPolicy.
let network_policy = if sandbox_policy.has_full_network_access() {
"(allow network-outbound)\n(allow network-inbound)\n(allow system-socket)"
} else {
""
};
let full_policy = format!(
"{MACOS_SEATBELT_BASE_POLICY}\n{file_read_policy}\n{file_write_policy}\n{network_policy}"
);
let mut seatbelt_args: Vec<String> = vec!["-p".to_string(), full_policy];
seatbelt_args.extend(extra_cli_args);
seatbelt_args.push("--".to_string());
seatbelt_args.extend(command);
seatbelt_args
}
#[derive(Debug)]
pub struct RawExecToolCallOutput {
pub exit_status: ExitStatus,
pub stdout: Vec<u8>,
pub stderr: Vec<u8>,
}
#[derive(Debug)]
pub struct ExecToolCallOutput {
pub exit_code: i32,
pub stdout: String,
pub stderr: String,
pub duration: Duration,
}
async fn exec(
ExecParams {
command,
cwd,
timeout_ms,
env,
}: ExecParams,
sandbox_policy: &SandboxPolicy,
ctrl_c: Arc<Notify>,
) -> Result<RawExecToolCallOutput> {
let (program, args) = command.split_first().ok_or_else(|| {
CodexErr::Io(io::Error::new(
io::ErrorKind::InvalidInput,
"command args are empty",
))
})?;
let arg0 = None;
let child = spawn_child_async(
PathBuf::from(program),
args.into(),
arg0,
cwd,
sandbox_policy,
StdioPolicy::RedirectForShellTool,
env,
)
.await?;
consume_truncated_output(child, ctrl_c, timeout_ms).await
}
#[derive(Debug, Clone, Copy)]
pub enum StdioPolicy {
RedirectForShellTool,
Inherit,
}
/// Spawns the appropriate child process for the ExecParams and SandboxPolicy,
/// ensuring the args and environment variables used to create the `Command`
/// (and `Child`) honor the configuration.
///
/// For now, we take `SandboxPolicy` as a parameter to spawn_child() because
/// we need to determine whether to set the
/// `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR` environment variable.
async fn spawn_child_async(
program: PathBuf,
args: Vec<String>,
#[cfg_attr(not(unix), allow(unused_variables))] arg0: Option<&str>,
cwd: PathBuf,
sandbox_policy: &SandboxPolicy,
stdio_policy: StdioPolicy,
env: HashMap<String, String>,
) -> std::io::Result<Child> {
let mut cmd = Command::new(&program);
#[cfg(unix)]
cmd.arg0(arg0.map_or_else(|| program.to_string_lossy().to_string(), String::from));
cmd.args(args);
cmd.current_dir(cwd);
cmd.env_clear();
cmd.envs(env);
if !sandbox_policy.has_full_network_access() {
cmd.env(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR, "1");
}
match stdio_policy {
StdioPolicy::RedirectForShellTool => {
// Do not create a file descriptor for stdin because otherwise some
// commands may hang forever waiting for input. For example, ripgrep has
// a heuristic where it may try to read from stdin as explained here:
// https://github.com/BurntSushi/ripgrep/blob/e2362d4d5185d02fa857bf381e7bd52e66fafc73/crates/core/flags/hiargs.rs#L1101-L1103
cmd.stdin(Stdio::null());
cmd.stdout(Stdio::piped()).stderr(Stdio::piped());
}
StdioPolicy::Inherit => {
// Inherit stdin, stdout, and stderr from the parent process.
cmd.stdin(Stdio::inherit())
.stdout(Stdio::inherit())
.stderr(Stdio::inherit());
}
}
cmd.kill_on_drop(true).spawn()
}
/// Consumes the output of a child process, truncating it so it is suitable for
/// use as the output of a `shell` tool call. Also enforces specified timeout.
pub(crate) async fn consume_truncated_output(
mut child: Child,
ctrl_c: Arc<Notify>,
timeout_ms: Option<u64>,
) -> Result<RawExecToolCallOutput> {
// Both stdout and stderr were configured with `Stdio::piped()`
// above, therefore `take()` should normally return `Some`. If it doesn't
// we treat it as an exceptional I/O error
let stdout_reader = child.stdout.take().ok_or_else(|| {
CodexErr::Io(io::Error::other(
"stdout pipe was unexpectedly not available",
))
})?;
let stderr_reader = child.stderr.take().ok_or_else(|| {
CodexErr::Io(io::Error::other(
"stderr pipe was unexpectedly not available",
))
})?;
let stdout_handle = tokio::spawn(read_capped(
BufReader::new(stdout_reader),
MAX_STREAM_OUTPUT,
MAX_STREAM_OUTPUT_LINES,
));
let stderr_handle = tokio::spawn(read_capped(
BufReader::new(stderr_reader),
MAX_STREAM_OUTPUT,
MAX_STREAM_OUTPUT_LINES,
));
let interrupted = ctrl_c.notified();
let timeout = Duration::from_millis(timeout_ms.unwrap_or(DEFAULT_TIMEOUT_MS));
let exit_status = tokio::select! {
result = tokio::time::timeout(timeout, child.wait()) => {
match result {
Ok(Ok(exit_status)) => exit_status,
Ok(e) => e?,
Err(_) => {
// timeout
child.start_kill()?;
// Debatable whether `child.wait().await` should be called here.
synthetic_exit_status(128 + TIMEOUT_CODE)
}
}
}
_ = interrupted => {
child.start_kill()?;
synthetic_exit_status(128 + SIGKILL_CODE)
}
};
let stdout = stdout_handle.await??;
let stderr = stderr_handle.await??;
Ok(RawExecToolCallOutput {
exit_status,
stdout,
stderr,
})
}
async fn read_capped<R: AsyncRead + Unpin>(
mut reader: R,
max_output: usize,
max_lines: usize,
) -> io::Result<Vec<u8>> {
let mut buf = Vec::with_capacity(max_output.min(8 * 1024));
let mut tmp = [0u8; 8192];
let mut remaining_bytes = max_output;
let mut remaining_lines = max_lines;
loop {
let n = reader.read(&mut tmp).await?;
if n == 0 {
break;
}
// Copy into the buffer only while we still have byte and line budget.
if remaining_bytes > 0 && remaining_lines > 0 {
let mut copy_len = 0;
for &b in &tmp[..n] {
if remaining_bytes == 0 || remaining_lines == 0 {
break;
}
copy_len += 1;
remaining_bytes -= 1;
if b == b'\n' {
remaining_lines -= 1;
}
}
buf.extend_from_slice(&tmp[..copy_len]);
}
// Continue reading to EOF to avoid back-pressure, but discard once caps are hit.
}
Ok(buf)
}
#[cfg(unix)]
fn synthetic_exit_status(code: i32) -> ExitStatus {
use std::os::unix::process::ExitStatusExt;
std::process::ExitStatus::from_raw(code)
}
#[cfg(windows)]
fn synthetic_exit_status(code: i32) -> ExitStatus {
use std::os::windows::process::ExitStatusExt;
#[expect(clippy::unwrap_used)]
std::process::ExitStatus::from_raw(code.try_into().unwrap())
}