running_process_platform_internal/
platform_linux.rs

1//! Linux implementation root for the process capability.
2
3#[cfg(feature = "ipc")]
4#[path = "platform_linux/ipc.rs"]
5mod ipc;
6#[cfg(feature = "ipc")]
7pub use ipc::{
8    current_user_id as ipc_current_user_id, Endpoint as IpcEndpoint, Listener as IpcListener,
9    ListenerNonblockingMode as IpcListenerNonblockingMode, PeerIdentity as IpcPeerIdentity,
10    Stream as IpcStream,
11};
12#[cfg(feature = "ipc-async")]
13pub use ipc::{AsyncListener as IpcAsyncListener, AsyncStream as IpcAsyncStream};
14
15#[cfg(feature = "session-relay")]
16#[path = "platform_linux_session_relay.rs"]
17mod session_relay;
18#[cfg(feature = "session-relay")]
19pub use session_relay::relay_local_socket_session;
20
21#[path = "platform_linux/terminal.rs"]
22pub mod terminal;
23pub use terminal::active_graphics_probe;
24pub use crate::platform::terminal_input;
25
26#[path = "platform_linux/window_icon.rs"]
27mod window_icon;
28pub use window_icon::{icon_support as window_icon_support_impl, set_icon as set_window_icon_impl};
29
30pub fn shell_command(command: &str) -> std::process::Command {
31    let mut shell = std::process::Command::new("/bin/sh");
32    shell.arg("-lc").arg(command);
33    shell
34}
35
36pub fn compat_shell_command(command: &str) -> std::process::Command {
37    let mut shell = std::process::Command::new("/bin/sh");
38    shell.arg("-lc").arg(command);
39    shell
40}
41
42pub fn canonical_environment_pairs(pairs: Vec<(String, String)>) -> Vec<(String, String)> {
43    pairs
44}
45
46pub fn monitor_console_windows(
47    _duration: std::time::Duration,
48) -> Vec<crate::platform::process::ConsoleWindowInfo> {
49    Vec::new()
50}
51
52use std::ffi::OsStr;
53use std::io;
54use std::io::Read;
55use std::os::fd::{AsRawFd, RawFd};
56use std::os::unix::net::UnixStream;
57use std::sync::Mutex;
58
59use tokio::process::{Child, Command};
60
61use crate::SpawnSpec;
62
63#[path = "platform_linux_descendants.rs"]
64mod descendants;
65pub use descendants::start_descendant_monitor;
66
67#[path = "platform_linux_trace.rs"]
68mod exact_trace;
69pub use exact_trace::{configure_exact_trace, start_exact_trace, TracedChild};
70
71pub fn exact_trace_capability() -> crate::platform::process::ExactTraceCapability {
72    crate::platform::process::ExactTraceCapability {
73        available: true,
74        backend: "linux-ptrace",
75        reason: "launch-time PTRACE_TRACEME with follow-fork/clone/exec/exit supervision",
76        non_invasive_backend: "proc-descendant-snapshot",
77        non_invasive_grade:
78            crate::platform::process::NonInvasiveObservationGrade::SnapshotInferred,
79    }
80}
81
82pub struct WindowsJobHandle;
83
84pub fn assign_child_to_windows_job(
85    _child: &std::process::Child,
86    _direct_pid: u32,
87    _address_space_limit_bytes: Option<u64>,
88    _emit: Option<Box<dyn Fn(crate::platform::process::DescendantEvent) + Send>>,
89) -> io::Result<WindowsJobHandle> {
90    Err(io::Error::new(
91        io::ErrorKind::Unsupported,
92        "Windows Job Objects are unavailable on Linux",
93    ))
94}
95
96#[derive(Default)]
97pub struct CaptureCancellation {
98    wakers: Mutex<CaptureWakers>,
99}
100
101#[derive(Default)]
102struct CaptureWakers {
103    stdout: Option<UnixStream>,
104    stderr: Option<UnixStream>,
105}
106
107struct CancelableCaptureReader<R> {
108    reader: R,
109    wake_reader: UnixStream,
110}
111
112impl<R: Read + AsRawFd> Read for CancelableCaptureReader<R> {
113    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
114        if buf.is_empty() { return Ok(0); }
115        loop {
116            let mut poll_fds = [
117                libc::pollfd { fd: self.reader.as_raw_fd(), events: libc::POLLIN | libc::POLLHUP | libc::POLLERR, revents: 0 },
118                libc::pollfd { fd: self.wake_reader.as_raw_fd(), events: libc::POLLIN | libc::POLLHUP | libc::POLLERR, revents: 0 },
119            ];
120            // SAFETY: both descriptors remain owned by this reader for the call.
121            let polled = unsafe { libc::poll(poll_fds.as_mut_ptr(), poll_fds.len() as _, -1) };
122            if polled < 0 {
123                let error = io::Error::last_os_error();
124                if error.kind() == io::ErrorKind::Interrupted { continue; }
125                return Err(error);
126            }
127            if poll_fds[1].revents != 0 {
128                return Err(io::Error::new(io::ErrorKind::Interrupted, "capture reader cancelled"));
129            }
130            if poll_fds[0].revents != 0 {
131                match self.reader.read(buf) {
132                    Err(error) if error.kind() == io::ErrorKind::WouldBlock => continue,
133                    result => return result,
134                }
135            }
136        }
137    }
138}
139
140pub fn prepare_capture_reader<R>(
141    reader: R,
142    cancellation: &CaptureCancellation,
143    stream: crate::platform::process::CaptureStream,
144) -> io::Result<Box<dyn Read + Send>>
145where R: Read + AsRawFd + Send + 'static {
146    set_nonblocking(reader.as_raw_fd())?;
147    let (wake_reader, wake_writer) = UnixStream::pair()?;
148    wake_writer.set_nonblocking(true)?;
149    let mut wakers = cancellation.wakers.lock().expect("capture wakers mutex poisoned");
150    match stream {
151        crate::platform::process::CaptureStream::Stdout => wakers.stdout = Some(wake_writer),
152        crate::platform::process::CaptureStream::Stderr => wakers.stderr = Some(wake_writer),
153    }
154    Ok(Box::new(CancelableCaptureReader { reader, wake_reader }))
155}
156
157pub fn capture_reader_done(cancellation: &CaptureCancellation, stream: crate::platform::process::CaptureStream) {
158    let mut wakers = cancellation.wakers.lock().expect("capture wakers mutex poisoned");
159    match stream {
160        crate::platform::process::CaptureStream::Stdout => wakers.stdout = None,
161        crate::platform::process::CaptureStream::Stderr => wakers.stderr = None,
162    }
163}
164
165pub fn cancel_capture_reader(cancellation: &CaptureCancellation) {
166    let wakers = cancellation.wakers.lock().expect("capture wakers mutex poisoned");
167    let byte = [1_u8; 1];
168    for writer in [&wakers.stdout, &wakers.stderr].into_iter().flatten() {
169        // SAFETY: the stored wake socket stays alive while the mutex is held.
170        let _ = unsafe { libc::write(writer.as_raw_fd(), byte.as_ptr().cast(), byte.len()) };
171    }
172}
173
174fn set_nonblocking(fd: RawFd) -> io::Result<()> {
175    // SAFETY: `fd` is borrowed from a live reader for both calls.
176    let flags = unsafe { libc::fcntl(fd, libc::F_GETFL) };
177    if flags < 0 { return Err(io::Error::last_os_error()); }
178    // SAFETY: `fd` is borrowed from a live reader for both calls.
179    if unsafe { libc::fcntl(fd, libc::F_SETFL, flags | libc::O_NONBLOCK) } < 0 {
180        return Err(io::Error::last_os_error());
181    }
182    Ok(())
183}
184
185#[path = "platform_linux_file_handles.rs"]
186mod file_handles;
187pub use file_handles::read_process_file_handles;
188#[path = "platform_linux_cmdline.rs"]
189mod cmdline;
190pub use cmdline::read_process_cmdline;
191
192#[path = "platform/process_tree.rs"]
193mod process_tree;
194
195pub fn kill_tree(pid: u32, timeout: std::time::Duration) -> io::Result<u32> {
196    process_tree::kill_tree(pid, timeout, |_pid, process| Ok(process.start_time()))
197}
198
199pub fn exit_code(status: std::process::ExitStatus) -> i32 {
200    use std::os::unix::process::ExitStatusExt;
201    status.code().unwrap_or_else(|| -status.signal().unwrap_or(1))
202}
203
204pub fn set_process_name(name: &str) {
205    let truncated: String = name.chars().take(15).collect();
206    let c_name = std::ffi::CString::new(truncated).unwrap_or_default();
207    unsafe { libc::prctl(libc::PR_SET_NAME, c_name.as_ptr() as libc::c_ulong, 0, 0, 0); }
208}
209
210pub fn configure_trampoline_command(_command: &mut std::process::Command) {}
211
212pub fn configure_process_command(
213    command: &mut std::process::Command,
214    config: crate::platform::process::ProcessCommandConfig,
215) -> io::Result<()> {
216    let create_process_group = config.create_process_group;
217    let nice = config.nice;
218    let address_space_limit_bytes = config.address_space_limit_bytes;
219    if !(create_process_group || nice.is_some() || address_space_limit_bytes.is_some()) {
220        return Ok(());
221    }
222    use std::os::unix::process::CommandExt;
223    unsafe {
224        command.pre_exec(move || {
225            if create_process_group && libc::setpgid(0, 0) == -1 {
226                return Err(io::Error::last_os_error());
227            }
228            if let Some(nice) = nice {
229                if libc::setpriority(libc::PRIO_PROCESS, 0, nice) == -1 {
230                    return Err(io::Error::last_os_error());
231                }
232            }
233            if let Some(limit) = address_space_limit_bytes {
234                let rlim = libc::rlimit { rlim_cur: limit, rlim_max: limit };
235                if libc::setrlimit(libc::RLIMIT_AS, &rlim) == -1 {
236                    return Err(io::Error::last_os_error());
237                }
238            }
239            Ok(())
240        });
241    }
242    Ok(())
243}
244
245pub fn trampoline_exit_code(status: std::process::ExitStatus) -> i32 {
246    use std::os::unix::process::ExitStatusExt;
247    status.signal().map_or_else(|| status.code().unwrap_or(1), |signal| 128 + signal)
248}
249
250/// Return the GNU build ID of the running executable without reading the
251/// executable from disk.
252///
253/// The dynamic loader has already mapped the main image's `PT_NOTE` segment,
254/// so callers that only need an image-generation identity do not need to hash
255/// a potentially large unoptimized binary. `None` preserves a clean fallback
256/// for binaries linked without a GNU build ID.
257pub fn current_executable_build_id() -> Option<Vec<u8>> {
258    unsafe extern "C" fn visit(
259        info: *mut libc::dl_phdr_info,
260        _size: libc::size_t,
261        output: *mut libc::c_void,
262    ) -> libc::c_int {
263        const MAX_NOTE_BYTES: usize = 1024 * 1024;
264
265        let info = unsafe { &*info };
266        let is_main_executable = info.dlpi_name.is_null()
267            || unsafe { std::ffi::CStr::from_ptr(info.dlpi_name) }
268                .to_bytes()
269                .is_empty();
270        if !is_main_executable || info.dlpi_phdr.is_null() || info.dlpi_phnum == 0 {
271            return 0;
272        }
273        let headers = unsafe {
274            std::slice::from_raw_parts(info.dlpi_phdr, usize::from(info.dlpi_phnum))
275        };
276        #[allow(clippy::unnecessary_cast)]
277        let load_bias = info.dlpi_addr as u64;
278        for header in headers {
279            if header.p_type != libc::PT_NOTE {
280                continue;
281            }
282            let Ok(length) = usize::try_from(header.p_memsz) else {
283                continue;
284            };
285            if length == 0 || length > MAX_NOTE_BYTES {
286                continue;
287            }
288            let Some(address) = load_bias.checked_add(header.p_vaddr) else {
289                continue;
290            };
291            let Some(note_end) = address.checked_add(length as u64) else {
292                continue;
293            };
294            let mapped_read_only = headers.iter().any(|load| {
295                if load.p_type != libc::PT_LOAD || load.p_flags & libc::PF_R == 0 {
296                    return false;
297                }
298                let Some(start) = load_bias.checked_add(load.p_vaddr) else {
299                    return false;
300                };
301                let Some(end) = start.checked_add(load.p_memsz) else {
302                    return false;
303                };
304                address >= start && note_end <= end
305            });
306            if address == 0 || !mapped_read_only {
307                continue;
308            }
309            let notes = unsafe { std::slice::from_raw_parts(address as *const u8, length) };
310            if let Some(build_id) = gnu_build_id_from_notes(notes) {
311                let output = unsafe { &mut *output.cast::<Option<Vec<u8>>>() };
312                *output = Some(build_id.to_vec());
313                return 1;
314            }
315        }
316        0
317    }
318
319    let mut output = None;
320    unsafe {
321        libc::dl_iterate_phdr(
322            Some(visit),
323            (&mut output as *mut Option<Vec<u8>>).cast::<libc::c_void>(),
324        );
325    }
326    output
327}
328
329fn gnu_build_id_from_notes(mut notes: &[u8]) -> Option<&[u8]> {
330    fn aligned(value: usize) -> Option<usize> {
331        value.checked_add(3).map(|value| value & !3)
332    }
333
334    while notes.len() >= 12 {
335        let name_len = usize::try_from(u32::from_ne_bytes(notes[0..4].try_into().ok()?)).ok()?;
336        let desc_len = usize::try_from(u32::from_ne_bytes(notes[4..8].try_into().ok()?)).ok()?;
337        let kind = u32::from_ne_bytes(notes[8..12].try_into().ok()?);
338        let name_end = 12usize.checked_add(name_len)?;
339        let desc_start = 12usize.checked_add(aligned(name_len)?)?;
340        let desc_end = desc_start.checked_add(desc_len)?;
341        let next = desc_start.checked_add(aligned(desc_len)?)?;
342        if next > notes.len() || name_end > notes.len() || desc_end > notes.len() {
343            return None;
344        }
345        if kind == 3 && notes.get(12..name_end)?.starts_with(b"GNU") && desc_len > 0 {
346            return notes.get(desc_start..desc_end);
347        }
348        notes = &notes[next..];
349    }
350    None
351}
352
353/// Request a graceful shutdown for a child-owned POSIX process group.
354pub fn soft_terminate_process_group(pid: u32) -> io::Result<()> {
355    // SAFETY: `kill` receives only the numeric child-owned group id; no Rust
356    // references or borrowed state cross the OS boundary.
357    let result = unsafe { libc::kill(-(pid as i32), libc::SIGTERM) };
358    if result != 0 {
359        let error = io::Error::last_os_error();
360        if error.raw_os_error() != Some(libc::ESRCH) {
361            return Err(error);
362        }
363    }
364    Ok(())
365}
366
367pub fn process_snapshot() -> Vec<crate::platform::process::ProcessSnapshot> {
368    Vec::new()
369}
370
371pub fn process_snapshot_for_pid(_pid: u32) -> Option<crate::platform::process::ProcessSnapshot> {
372    None
373}
374
375/// Mark inherited descriptors close-on-exec without breaking std's exec-error pipe.
376///
377/// # Safety
378/// This must only be called from a post-fork `pre_exec` closure.
379pub unsafe fn unix_mark_extra_fds_close_on_exec() {
380    #[cfg(any(target_arch = "x86_64", target_arch = "aarch64", target_arch = "x86", target_arch = "arm", target_arch = "riscv64", target_arch = "powerpc64"))]
381    {
382        const SYS_CLOSE_RANGE: libc::c_long = 436;
383        const CLOSE_RANGE_CLOEXEC: libc::c_uint = 4;
384        if libc::syscall(SYS_CLOSE_RANGE, 3u32, libc::c_uint::MAX, CLOSE_RANGE_CLOEXEC) == 0 {
385            return;
386        }
387    }
388    mark_fds_from_directory_or_range();
389}
390
391pub fn configure_sync_daemon_command(command: &mut std::process::Command) -> io::Result<()> {
392    use std::os::unix::process::CommandExt;
393    unsafe {
394        command.pre_exec(|| {
395            let _ = libc::setsid();
396            unix_mark_extra_fds_close_on_exec();
397            Ok(())
398        });
399    }
400    Ok(())
401}
402
403pub fn configure_sync_contained_command(command: &mut std::process::Command) -> io::Result<()> {
404    use std::os::unix::process::CommandExt;
405    unsafe {
406        command.pre_exec(|| {
407            if libc::setpgid(0, 0) == -1 { return Err(io::Error::last_os_error()); }
408            if libc::prctl(libc::PR_SET_PDEATHSIG, libc::SIGKILL) == -1 {
409                return Err(io::Error::last_os_error());
410            }
411            if libc::getppid() == 1 { libc::_exit(1); }
412            unix_mark_extra_fds_close_on_exec();
413            Ok(())
414        });
415    }
416    Ok(())
417}
418
419pub fn parent_has_console() -> bool { false }
420
421pub fn sync_child_native_handle(_child: &std::process::Child) -> usize { 0 }
422
423unsafe fn mark_fds_from_directory_or_range() {
424    let dir = libc::opendir(c"/dev/fd".as_ptr());
425    if !dir.is_null() {
426        let dir_fd = libc::dirfd(dir);
427        loop {
428            let entry = libc::readdir(dir);
429            if entry.is_null() { break; }
430            let mut fd: libc::c_int = 0;
431            let mut cursor = (*entry).d_name.as_ptr();
432            let mut numeric = false;
433            while *cursor != 0 {
434                let byte = *cursor as u8;
435                if !byte.is_ascii_digit() { numeric = false; break; }
436                fd = fd * 10 + (byte - b'0') as libc::c_int;
437                cursor = cursor.add(1);
438                numeric = true;
439            }
440            if numeric && fd > 2 && fd != dir_fd { set_cloexec(fd); }
441        }
442        libc::closedir(dir);
443        return;
444    }
445    let maximum = libc::sysconf(libc::_SC_OPEN_MAX);
446    for fd in 3..if maximum < 0 { 4096 } else { maximum as libc::c_int } { set_cloexec(fd); }
447}
448
449unsafe fn set_cloexec(fd: libc::c_int) {
450    let flags = libc::fcntl(fd, libc::F_GETFD);
451    if flags != -1 { libc::fcntl(fd, libc::F_SETFD, flags | libc::FD_CLOEXEC); }
452}
453pub fn observer_backend(scope: crate::platform::process::ObserverScope, category: crate::platform::process::ObserverCategory) -> crate::platform::process::ObserverBackend {
454    use crate::platform::process::{ObserverBackend as B, ObserverCategory as C, ObserverScope as S, ObserverSupport as P};
455    match (scope, category) {
456        (S::SystemWide, C::File) => B { support:P::Unavailable, backend:"seccomp-user-notify", reason:"Phase 3: Linux seccomp user-notify file backend not yet implemented" },
457        (S::SystemWide, C::Network) => B { support:P::Unavailable, backend:"ebpf", reason:"Phase 3: Linux eBPF network backend not yet implemented" },
458        (S::SystemWide, C::Process) => B { support:P::Unavailable, backend:"seccomp-user-notify", reason:"Phase 3: Linux seccomp user-notify process backend not yet implemented" },
459        (S::LaunchedProcessTree, C::File) => B { support:P::Partial, backend:"proc-fd-snapshot", reason:"Linux /proc/<pid>/fd/* snapshot via read_process_file_handles (#539 slice 6 follow-up; no streaming file events)" },
460        (S::LaunchedProcessTree, C::Network) => B { support:P::Unavailable, backend:"none", reason:"#539: no-admin per-child network backend deferred to a follow-up issue" },
461        (S::LaunchedProcessTree, C::Process) => B { support:P::Supported, backend:"subreaper-proc-poll", reason:"Linux PR_SET_CHILD_SUBREAPER + /proc descendant polling (#539 slice 5)" },
462    }
463}
464
465pub fn unix_set_priority(pid: u32, nice: i32) -> io::Result<()> {
466    if unsafe { libc::setpriority(libc::PRIO_PROCESS, pid, nice) } == -1 { Err(io::Error::last_os_error()) } else { Ok(()) }
467}
468pub fn unix_signal_process(pid: u32, signal: crate::platform::process::UnixSignalKind) -> io::Result<()> {
469    if unsafe { libc::kill(pid as i32, unix_signal_raw(signal)) } == -1 { Err(io::Error::last_os_error()) } else { Ok(()) }
470}
471pub fn unix_signal_process_group(pid: i32, signal: crate::platform::process::UnixSignalKind) -> io::Result<()> {
472    if unsafe { libc::killpg(pid, unix_signal_raw(signal)) } == -1 { Err(io::Error::last_os_error()) } else { Ok(()) }
473}
474pub fn unix_signal_raw(signal: crate::platform::process::UnixSignalKind) -> i32 {
475    match signal { crate::platform::process::UnixSignalKind::Interrupt => libc::SIGINT, crate::platform::process::UnixSignalKind::Terminate => libc::SIGTERM, crate::platform::process::UnixSignalKind::Kill => libc::SIGKILL }
476}
477
478pub fn configure_compat_tokio_command(
479    command: &mut Command,
480    _show_console: bool,
481    kill_when_owner_dies: bool,
482) -> io::Result<()> {
483    configure_command(command, false, kill_when_owner_dies)
484}
485
486pub fn after_compat_tokio_spawn(_child: &Child, _kill_when_owner_dies: bool) {}
487
488pub(crate) fn configure_command(
489    command: &mut Command,
490    create_process_group: bool,
491    kill_when_owner_dies: bool,
492) -> io::Result<()> {
493    if create_process_group {
494        command.process_group(0);
495    }
496    if kill_when_owner_dies {
497        let owner_pid = unsafe { libc::getpid() };
498        // SAFETY: the closure invokes only async-signal-safe libc calls.
499        unsafe {
500            command.pre_exec(move || {
501                if libc::prctl(
502                    libc::PR_SET_PDEATHSIG,
503                    libc::SIGTERM as libc::c_ulong,
504                    0,
505                    0,
506                    0,
507                ) == -1
508                {
509                    return Err(io::Error::last_os_error());
510                }
511                if libc::getppid() != owner_pid {
512                    libc::kill(libc::getpid(), libc::SIGTERM);
513                }
514                Ok(())
515            });
516        }
517    }
518    Ok(())
519}
520
521pub(crate) fn after_spawn(_child: &Child, _kill_when_owner_dies: bool) {}
522
523pub(crate) fn signal_process(pid: u32) -> io::Result<()> {
524    unix_kill(pid as i32, libc::SIGKILL)
525}
526
527pub(crate) fn signal_process_group(pid: u32) -> io::Result<()> {
528    unix_kill(-(pid as i32), libc::SIGTERM)
529}
530
531fn unix_kill(target: i32, signal: i32) -> io::Result<()> {
532    let result = unsafe { libc::kill(target, signal) };
533    if result == 0 {
534        return Ok(());
535    }
536    let error = io::Error::last_os_error();
537    if error.raw_os_error() == Some(libc::ESRCH) {
538        Ok(())
539    } else {
540        Err(error)
541    }
542}
543
544pub(crate) fn shell_spec(command: &OsStr) -> SpawnSpec {
545    SpawnSpec::new("/bin/sh").arg("-c").arg(command)
546}
547
548#[cfg(test)]
549mod tests {
550    #[test]
551    fn shell_command_preserves_login_shell_contract_and_ignores_child_path() {
552        use std::ffi::OsStr;
553
554        let command_text = "printf '%s' 'alpha beta;\"gamma\"'";
555        let mut command = super::shell_command(command_text);
556        assert_eq!(command.get_program(), OsStr::new("/bin/sh"));
557        assert_eq!(
558            command.get_args().collect::<Vec<_>>(),
559            [OsStr::new("-lc"), OsStr::new(command_text)]
560        );
561        command
562            .env_clear()
563            .env("PATH", "/caller-supplied-path-override");
564        let output = command
565            .output()
566            .expect("absolute shell command should execute independently of child PATH");
567        assert!(output.status.success());
568        assert_eq!(output.stdout, b"alpha beta;\"gamma\"");
569    }
570
571    #[test]
572    #[cfg(not(target_env = "musl"))]
573    fn current_executable_exposes_a_gnu_build_id() {
574        let build_id = super::current_executable_build_id()
575            .expect("Linux test executable should carry a GNU build ID");
576        assert!(!build_id.is_empty());
577    }
578}
579#[cfg(test)]
580#[path = "tests/platform_linux_coverage.rs"]
581mod coverage_tests;
582#[path = "sync_spawn_group.rs"]
583mod sync_spawn;
584pub use sync_spawn::{spawn_sync, spawn_sync_daemon};
running_process_platform_internal/platform_linux.rs

running_process_platform_internal/
platform_linux.rs
