Add confined host escalation transport

This commit is contained in:
ms2k 2026-07-15 19:48:42 +09:00
commit 407b75b1ff
2 changed files with 2888 additions and 21 deletions

View file

@ -1,18 +1,299 @@
use std::fs::{File, OpenOptions};
use std::io::{self, Seek, SeekFrom, Write};
use std::os::fd::AsFd;
use std::io::{self, Read, Seek, SeekFrom, Write};
use std::os::fd::{AsFd, OwnedFd};
use std::os::unix::ffi::{OsStrExt, OsStringExt};
use std::os::unix::fs::MetadataExt;
use std::path::Path;
use std::sync::atomic::{AtomicU64, Ordering};
use lazaret_core::identity::{DeviceId, InodeId, MountId, MountIdentity, NamespaceIdentity};
use rustix::fs::{
AtFlags, FlockOperation, MemfdFlags, SealFlags, StatxFlags, fcntl_add_seals, fcntl_get_seals,
flock, fstat, makedev, memfd_create, statx,
AtFlags, FlockOperation, MemfdFlags, Mode, OFlags, RawDir, RenameFlags, SealFlags, StatxFlags,
fchmod, fcntl_add_seals, fcntl_get_seals, flock, fstat, makedev, memfd_create, openat,
renameat_with, statat, statx, unlinkat,
};
static TEMPORARY_SEQUENCE: AtomicU64 = AtomicU64::new(0);
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct MailboxEntryIdentity {
device: u64,
inode: u64,
}
#[derive(Debug)]
pub struct RegularFileMailbox {
directory: File,
}
impl AsFd for RegularFileMailbox {
fn as_fd(&self) -> std::os::fd::BorrowedFd<'_> {
self.directory.as_fd()
}
}
impl RegularFileMailbox {
/// # Errors
/// Refuses entries that are not invoking-user-owned mode-0700 directories opened without links.
pub fn openat_private(
parent: impl AsFd,
name: &std::ffi::OsStr,
invoking_uid: u32,
) -> io::Result<Self> {
validate_mailbox_name(name)?;
let directory = File::from(openat(
parent.as_fd(),
name,
OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
Mode::empty(),
)?);
Self::from_private_directory(directory, invoking_uid)
}
/// # Errors
/// Refuses paths that are not invoking-user-owned mode-0700 directories opened without links.
pub fn open_private(path: &Path, invoking_uid: u32) -> io::Result<Self> {
let directory = File::from(openat(
rustix::fs::CWD,
path,
OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
Mode::empty(),
)?);
Self::from_private_directory(directory, invoking_uid)
}
fn from_private_directory(directory: File, invoking_uid: u32) -> io::Result<Self> {
let metadata = directory.metadata()?;
if !metadata.file_type().is_dir()
|| metadata.uid() != invoking_uid
|| metadata.mode() & 0o7777 != 0o700
{
return Err(io::Error::other(
"mailbox directory is not invoking-user-owned mode 0700",
));
}
Ok(Self { directory })
}
/// # Errors
/// Refuses a path that is not a directory opened without following its final component.
pub fn open(path: &Path) -> io::Result<Self> {
let directory = File::from(openat(
rustix::fs::CWD,
path,
OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
Mode::empty(),
)?);
Ok(Self { directory })
}
/// Publishes one complete regular-file record using a no-replace atomic rename.
///
/// # Errors
/// Refuses invalid names, existing destinations, or incomplete durable publication.
pub fn publish(
&self,
name: &std::ffi::OsStr,
bytes: &[u8],
) -> io::Result<MailboxEntryIdentity> {
validate_mailbox_name(name)?;
let temporary_name = loop {
let sequence = TEMPORARY_SEQUENCE.fetch_add(1, Ordering::Relaxed);
let candidate = std::ffi::OsString::from_vec(
format!(".lazaret-{}-{sequence}.tmp", std::process::id()).into_bytes(),
);
match openat(
self.directory.as_fd(),
&candidate,
OFlags::CREATE | OFlags::EXCL | OFlags::WRONLY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
Mode::RUSR | Mode::WUSR,
) {
Ok(fd) => break (candidate, fd),
Err(rustix::io::Errno::EXIST) => {}
Err(error) => return Err(error.into()),
}
};
let (temporary_name, fd) = temporary_name;
let mut file = File::from(fd);
let result = (|| {
fchmod(file.as_fd(), Mode::RUSR | Mode::WUSR)?;
file.write_all(bytes)?;
file.sync_all()?;
let metadata = file.metadata()?;
if !metadata.file_type().is_file()
|| metadata.nlink() != 1
|| usize::try_from(metadata.len()).ok() != Some(bytes.len())
{
return Err(io::Error::other("mailbox temporary record changed"));
}
renameat_with(
self.directory.as_fd(),
&temporary_name,
self.directory.as_fd(),
name,
RenameFlags::NOREPLACE,
)?;
self.directory.sync_all()?;
Ok(MailboxEntryIdentity {
device: metadata.dev(),
inode: metadata.ino(),
})
})();
if result.is_err() {
let _ = unlinkat(self.directory.as_fd(), &temporary_name, AtFlags::empty());
}
result
}
/// Reads and removes one exact regular-file entry after bounding it before allocation.
///
/// # Errors
/// Refuses links, non-regular files, oversize records, replacement, growth, or short reads.
pub fn consume(&self, name: &std::ffi::OsStr, maximum: usize) -> io::Result<Vec<u8>> {
let (bytes, identity) = self.read(name, maximum)?;
self.remove_if_matches(name, identity)?;
Ok(bytes)
}
/// Reads one exact regular-file entry after bounding it before allocation.
///
/// # Errors
/// Refuses links, non-regular files, oversize records, replacement, growth, or short reads.
pub fn read(
&self,
name: &std::ffi::OsStr,
maximum: usize,
) -> io::Result<(Vec<u8>, MailboxEntryIdentity)> {
validate_mailbox_name(name)?;
let mut file = File::from(openat(
self.directory.as_fd(),
name,
OFlags::RDONLY | OFlags::NONBLOCK | OFlags::NOFOLLOW | OFlags::CLOEXEC,
Mode::empty(),
)?);
let before = file.metadata()?;
let length = usize::try_from(before.len())
.map_err(|_| io::Error::other("mailbox record length is out of range"))?;
if !before.file_type().is_file() || before.nlink() != 1 {
return Err(io::Error::other(
"mailbox entry is linked or is not a regular file",
));
}
if length > maximum {
return Err(io::Error::other("mailbox record exceeds the size limit"));
}
let mut bytes = Vec::with_capacity(length.saturating_add(1));
Read::by_ref(&mut file)
.take(u64::try_from(maximum).unwrap_or(u64::MAX).saturating_add(1))
.read_to_end(&mut bytes)?;
let after = file.metadata()?;
if bytes.len() != length
|| after.dev() != before.dev()
|| after.ino() != before.ino()
|| after.len() != before.len()
|| after.nlink() != 1
|| !after.file_type().is_file()
{
return Err(io::Error::other("mailbox record changed while being read"));
}
Ok((
bytes,
MailboxEntryIdentity {
device: before.dev(),
inode: before.ino(),
},
))
}
/// Removes a host-published record only while the pathname retains its exact identity.
///
/// # Errors
/// Refuses a missing or replaced pathname.
pub fn remove_if_matches(
&self,
name: &std::ffi::OsStr,
expected: MailboxEntryIdentity,
) -> io::Result<()> {
validate_mailbox_name(name)?;
let quarantine = loop {
let sequence = TEMPORARY_SEQUENCE.fetch_add(1, Ordering::Relaxed);
let candidate = std::ffi::OsString::from_vec(
format!(".lazaret-remove-{}-{sequence}.tmp", std::process::id()).into_bytes(),
);
match renameat_with(
self.directory.as_fd(),
name,
self.directory.as_fd(),
&candidate,
RenameFlags::NOREPLACE,
) {
Ok(()) => break candidate,
Err(rustix::io::Errno::EXIST) => {}
Err(error) => return Err(error.into()),
}
};
let current = statat(
self.directory.as_fd(),
&quarantine,
AtFlags::SYMLINK_NOFOLLOW,
)?;
if current.st_dev != expected.device || current.st_ino != expected.inode {
let restore = renameat_with(
self.directory.as_fd(),
&quarantine,
self.directory.as_fd(),
name,
RenameFlags::NOREPLACE,
);
return match restore {
Ok(()) => Err(io::Error::other("mailbox entry identity mismatch")),
Err(error) => Err(io::Error::other(format!(
"mailbox entry identity mismatch and restore failed: {error}"
))),
};
}
unlinkat(self.directory.as_fd(), &quarantine, AtFlags::empty())?;
self.directory.sync_all()
}
/// # Errors
/// Returns an error when the retained directory cannot be enumerated.
pub fn names(&self) -> io::Result<Vec<std::ffi::OsString>> {
let fd: OwnedFd = openat(
self.directory.as_fd(),
".",
OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
Mode::empty(),
)?;
let mut buffer = vec![std::mem::MaybeUninit::<u8>::uninit(); 8192];
let mut directory = RawDir::new(fd, &mut buffer);
let mut names = Vec::new();
while let Some(entry) = directory.next() {
let entry = entry?;
let name = entry.file_name().to_bytes();
if name != b"." && name != b".." {
names.push(std::ffi::OsString::from_vec(name.to_vec()));
}
}
Ok(names)
}
}
fn validate_mailbox_name(name: &std::ffi::OsStr) -> io::Result<()> {
let bytes = name.as_bytes();
if bytes.is_empty()
|| bytes.len() > 255
|| bytes == b"."
|| bytes == b".."
|| bytes.contains(&0)
|| bytes.contains(&b'/')
{
Err(io::Error::other("invalid mailbox entry name"))
} else {
Ok(())
}
}
const REQUIRED_SEALS: SealFlags = SealFlags::SHRINK
.union(SealFlags::GROW)
.union(SealFlags::WRITE)
@ -159,12 +440,13 @@ pub fn require_mount_identity(path: &Path, expected: MountIdentity) -> io::Resul
mod tests {
use std::fs::OpenOptions;
use std::io::{Read, Write};
use std::os::unix::fs::symlink;
use rustix::fs::{MemfdFlags, memfd_create};
use super::{
atomic_replace, current_mount_identity, namespace_identity, require_mount_identity,
require_write_sealed, sealed_memfd, try_exclusive_flock,
RegularFileMailbox, atomic_replace, current_mount_identity, namespace_identity,
require_mount_identity, require_write_sealed, sealed_memfd, try_exclusive_flock,
};
fn test_directory(name: &str) -> std::path::PathBuf {
@ -265,4 +547,69 @@ mod tests {
assert!(require_mount_identity(&old, expected).is_ok());
std::fs::remove_dir_all(directory).unwrap();
}
#[test]
fn mailbox_publication_is_atomic_and_no_replace() {
let directory = test_directory("mailbox-publish");
let mailbox = RegularFileMailbox::open(&directory).unwrap();
let name = std::ffi::OsStr::new("request.ready");
mailbox.publish(name, b"complete").unwrap();
assert!(mailbox.publish(name, b"replacement").is_err());
assert_eq!(std::fs::read(directory.join(name)).unwrap(), b"complete");
assert!(
mailbox
.names()
.unwrap()
.iter()
.all(|entry| !entry.as_encoded_bytes().ends_with(b".tmp"))
);
std::fs::remove_dir_all(directory).unwrap();
}
#[test]
fn mailbox_refuses_links_nonregular_and_oversize_records() {
let directory = test_directory("mailbox-hostile");
let mailbox = RegularFileMailbox::open(&directory).unwrap();
std::fs::write(directory.join("target"), b"target").unwrap();
symlink("target", directory.join("symlink.ready")).unwrap();
std::fs::hard_link(directory.join("target"), directory.join("hardlink.ready")).unwrap();
std::fs::create_dir(directory.join("directory.ready")).unwrap();
std::fs::write(directory.join("oversize.ready"), [0_u8; 17]).unwrap();
assert!(
mailbox
.consume(std::ffi::OsStr::new("symlink.ready"), 16)
.is_err()
);
assert!(
mailbox
.consume(std::ffi::OsStr::new("hardlink.ready"), 16)
.is_err()
);
assert!(
mailbox
.consume(std::ffi::OsStr::new("directory.ready"), 16)
.is_err()
);
assert!(
mailbox
.consume(std::ffi::OsStr::new("oversize.ready"), 16)
.is_err()
);
std::fs::remove_dir_all(directory).unwrap();
}
#[test]
fn mailbox_removes_only_the_identity_it_read() {
let directory = test_directory("mailbox-replacement");
let mailbox = RegularFileMailbox::open(&directory).unwrap();
let name = std::ffi::OsStr::new("result.ready");
let identity = mailbox.publish(name, b"first").unwrap();
std::fs::rename(directory.join(name), directory.join("old")).unwrap();
std::fs::write(directory.join(name), b"foreign").unwrap();
assert!(mailbox.remove_if_matches(name, identity).is_err());
assert_eq!(std::fs::read(directory.join(name)).unwrap(), b"foreign");
std::fs::remove_dir_all(directory).unwrap();
}
}

View file

@ -1,7 +1,7 @@
use std::ffi::{OsStr, OsString};
use std::fs::File;
use std::io::{self, Read, Write};
use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
use std::os::fd::{AsFd, AsRawFd, BorrowedFd, OwnedFd};
use std::process::Stdio;
use std::sync::{Mutex, OnceLock};
@ -13,7 +13,7 @@ use lazaret_core::records::{
HostProcessOwnershipSlot, ProcessOwnershipCommitAck, StagedProcessOwnership,
};
use rustix::event::{PollFd, PollFlags, Timespec, poll};
use rustix::fs::{fstat, fstatfs};
use rustix::fs::{OFlags, fcntl_getfl, fcntl_setfl, fstat, fstatfs};
use rustix::io::{FdFlags, fcntl_setfd};
use rustix::pipe::{PipeFlags, pipe_with};
use rustix::process::{Pid, PidfdFlags, Signal, WaitId, WaitIdOptions};
@ -25,11 +25,12 @@ const WRAPPER_ROLE: &str = "launch-wrapper";
const SIGCHLD_PROBE_ROLE: &str = "sigchld-probe";
static CHILD_BROKER: OnceLock<ChildBroker> = OnceLock::new();
#[derive(Clone, Debug, Eq, PartialEq)]
#[derive(Debug)]
pub struct TargetCommand {
program: OsString,
arguments: Vec<OsString>,
parent_edge: ParentEdge,
executable: Option<OwnedFd>,
}
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
@ -45,9 +46,27 @@ impl TargetCommand {
program: program.into(),
arguments: Vec::new(),
parent_edge: ParentEdge::ParentDeathKill,
executable: None,
}
}
/// Constructs an argv-only target bound to the exact inherited executable descriptor.
///
/// # Errors
/// Returns an error when the descriptor cannot be made inheritable by the launch wrapper.
pub fn from_executable(
executable: OwnedFd,
arguments: impl IntoIterator<Item = OsString>,
) -> io::Result<Self> {
fcntl_setfd(executable.as_fd(), FdFlags::empty())?;
Ok(Self {
program: OsString::from(format!("/proc/self/fd/{}", executable.as_raw_fd())),
arguments: arguments.into_iter().collect(),
parent_edge: ParentEdge::ParentDeathKill,
executable: Some(executable),
})
}
#[must_use]
pub fn argument(mut self, argument: impl Into<OsString>) -> Self {
self.arguments.push(argument.into());
@ -73,6 +92,7 @@ impl TargetCommand {
#[allow(clippy::disallowed_types)]
fn append_to(&self, command: &mut std::process::Command) {
let _ = &self.executable;
command
.arg("--parent-action")
.arg(match self.parent_edge {
@ -149,6 +169,35 @@ pub struct RetainedProcess {
identity: PidfsProcessIdentityV1,
}
#[derive(Debug)]
pub struct CapturedProcess {
process: RetainedProcess,
stdout: File,
stderr: File,
}
impl CapturedProcess {
#[must_use]
pub const fn process(&self) -> &RetainedProcess {
&self.process
}
pub fn streams(&mut self) -> (&mut File, &mut File) {
(&mut self.stdout, &mut self.stderr)
}
#[must_use]
pub fn into_process(self) -> RetainedProcess {
self.process
}
}
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct ProcessExit {
pub exit_code: Option<i32>,
pub terminating_signal: Option<i32>,
}
impl RetainedProcess {
#[must_use]
pub const fn identity(&self) -> PidfsProcessIdentityV1 {
@ -212,6 +261,16 @@ fn require_parent_edge_standard_io(
}
}
fn require_captured_parent_edge(parent_edge: ParentEdge) -> io::Result<()> {
if parent_edge == ParentEdge::ParentDeathKill {
Ok(())
} else {
Err(io::Error::other(
"captured output requires the parent-death edge",
))
}
}
impl ChildBroker {
/// Spawns the repository wrapper, stages exact ownership, accepts its acknowledgment,
/// and only then releases target execution.
@ -312,18 +371,144 @@ impl ChildBroker {
Ok(process)
}
/// Acquires an exact child with bounded-output pipe ownership retained by the caller.
///
/// # Errors
/// Returns an error when serialization or exact pidfd reaping fails.
#[allow(clippy::disallowed_methods)]
pub fn reap(&self, process: RetainedProcess) -> io::Result<()> {
/// Refuses and reaps the blocked wrapper on every failed acquisition or callback.
#[allow(clippy::disallowed_methods, clippy::disallowed_types)]
pub fn acquire_captured<Stage, Acknowledge>(
&self,
expected_parent: PidfsProcessIdentityV1,
ownership: ProcessOwnershipContext,
target: &TargetCommand,
stage: Stage,
acknowledge: Acknowledge,
) -> io::Result<CapturedProcess>
where
Stage: FnOnce(StagedProcessOwnership) -> io::Result<()>,
Acknowledge: FnOnce(StagedProcessOwnership) -> io::Result<ProcessOwnershipCommitAck>,
{
let _guard = self
.exclusion
.lock()
.map_err(|_| io::Error::other("child broker lock poisoned"))?;
rustix::process::waitid(WaitId::PidFd(process.pidfd.as_fd()), WaitIdOptions::EXITED)?;
drop(process);
Ok(())
require_captured_parent_edge(target.parent_edge())?;
require_default_sigchld()?;
require_reapable_sigchld()?;
let (release, wait) = launch_barrier()?;
let mut command = private_role_command(WRAPPER_ROLE);
append_parent_identity(&mut command, expected_parent);
target.append_to(&mut command);
command
.stdin(wait.into_stdio())
.stdout(Stdio::piped())
.stderr(Stdio::piped());
let mut child = command.spawn()?;
let host_pid = child.id();
let process = match acquire_spawned_child(host_pid) {
Ok(process) => process,
Err(error) => return fail_before_pidfd(release, &mut child, error),
};
let staged = StagedProcessOwnership::v1(
ownership.domain_id,
ownership.generation_id,
ownership.coordinator_id,
ownership.slot,
process.identity,
);
if let Err(error) = stage(staged) {
return fail_after_pidfd(release, process, error);
}
let acknowledgment = match acknowledge(staged) {
Ok(acknowledgment) => acknowledgment,
Err(error) => return fail_after_pidfd(release, process, error),
};
if !acknowledgment.acknowledges(&staged) {
return fail_after_pidfd(
release,
process,
io::Error::other(
"process ownership acknowledgment does not match staged ownership",
),
);
}
if let Err(error) = release.release() {
return fail_after_pidfd_closed(process, error);
}
let Some(stdout) = child.stdout.take() else {
return fail_running_captured(
process,
child,
io::Error::other("captured child stdout is unavailable"),
);
};
let Some(stderr) = child.stderr.take() else {
return fail_running_captured(
process,
child,
io::Error::other("captured child stderr is unavailable"),
);
};
let stdout = match child_pipe_file(stdout.into()) {
Ok(stdout) => stdout,
Err(error) => return fail_running_captured(process, child, error),
};
let stderr = match child_pipe_file(stderr.into()) {
Ok(stderr) => stderr,
Err(error) => return fail_running_captured(process, child, error),
};
drop(child);
Ok(CapturedProcess {
process,
stdout,
stderr,
})
}
/// # Errors
/// Returns an error when serialization or exact pidfd reaping fails.
#[allow(clippy::disallowed_methods)]
pub fn reap(&self, process: RetainedProcess) -> io::Result<()> {
self.reap_status(process).map(|_| ())
}
/// # Errors
/// Returns an error when serialization or exact pidfd reaping fails.
#[allow(clippy::disallowed_methods)]
pub fn reap_status(&self, process: RetainedProcess) -> io::Result<ProcessExit> {
let _guard = self
.exclusion
.lock()
.map_err(|_| io::Error::other("child broker lock poisoned"))?;
let status =
rustix::process::waitid(WaitId::PidFd(process.pidfd.as_fd()), WaitIdOptions::EXITED)?
.ok_or_else(|| io::Error::other("exact child exited without wait status"))?;
let exit = ProcessExit {
exit_code: status.exit_status(),
terminating_signal: status.terminating_signal(),
};
drop(process);
Ok(exit)
}
}
fn child_pipe_file(fd: OwnedFd) -> io::Result<File> {
let flags = fcntl_getfl(fd.as_fd())?;
fcntl_setfl(fd.as_fd(), flags | OFlags::NONBLOCK)?;
Ok(File::from(fd))
}
#[allow(clippy::disallowed_methods, clippy::disallowed_types)]
fn fail_running_captured<T>(
process: RetainedProcess,
child: std::process::Child,
error: io::Error,
) -> io::Result<T> {
let _ = process.send_signal(Signal::KILL);
rustix::process::waitid(WaitId::PidFd(process.pidfd.as_fd()), WaitIdOptions::EXITED)?;
drop(child);
drop(process);
Err(error)
}
#[allow(clippy::disallowed_types)]
@ -429,30 +614,27 @@ pub fn verify_direct_child_pidfd(pidfd: impl AsFd) -> io::Result<()> {
}
#[allow(clippy::disallowed_methods, clippy::disallowed_types)]
fn fail_before_pidfd(
fn fail_before_pidfd<T>(
release: BarrierRelease,
child: &mut std::process::Child,
error: io::Error,
) -> io::Result<RetainedProcess> {
) -> io::Result<T> {
drop(release);
child.wait()?;
Err(error)
}
fn fail_after_pidfd(
fn fail_after_pidfd<T>(
release: BarrierRelease,
process: RetainedProcess,
error: io::Error,
) -> io::Result<RetainedProcess> {
) -> io::Result<T> {
drop(release);
fail_after_pidfd_closed(process, error)
}
#[allow(clippy::disallowed_methods)]
fn fail_after_pidfd_closed(
process: RetainedProcess,
error: io::Error,
) -> io::Result<RetainedProcess> {
fn fail_after_pidfd_closed<T>(process: RetainedProcess, error: io::Error) -> io::Result<T> {
rustix::process::waitid(WaitId::PidFd(process.pidfd.as_fd()), WaitIdOptions::EXITED)?;
drop(process);
Err(error)
@ -848,6 +1030,7 @@ pub fn parse_wrapper_arguments(
program,
arguments: arguments.collect(),
parent_edge,
executable: None,
};
let boot = parse_boot_hex(&boot)?;
Ok((