1 // Copyright 2014 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
14 use c_str::{CString, ToCStr};
17 use io::process::{ProcessExit, ExitStatus, ExitSignal};
18 use io::{mod, IoResult, IoError, EndOfFile};
19 use libc::{mod, pid_t, c_void, c_int};
22 use path::BytesContainer;
24 use sync::mpsc::{channel, Sender, Receiver};
25 use sys::fs::FileDesc;
26 use sys::{mod, retry, c, wouldblock, set_nonblocking, ms_to_timeval};
27 use sys_common::helper_thread::Helper;
28 use sys_common::{AsInner, mkerr_libc, timeout};
30 pub use sys_common::ProcessConfig;
32 helper_init! { static HELPER: Helper<Req> }
34 /// The unique id of the process (this should never be negative).
40 NewChild(libc::pid_t, Sender<ProcessExit>, u64),
43 const CLOEXEC_MSG_FOOTER: &'static [u8] = b"NOEX";
46 pub fn id(&self) -> pid_t {
50 pub unsafe fn kill(&self, signal: int) -> IoResult<()> {
51 Process::killpid(self.pid, signal)
54 pub unsafe fn killpid(pid: pid_t, signal: int) -> IoResult<()> {
55 let r = libc::funcs::posix88::signal::kill(pid, signal as c_int);
59 pub fn spawn<K, V, C, P>(cfg: &C, in_fd: Option<P>,
60 out_fd: Option<P>, err_fd: Option<P>)
62 where C: ProcessConfig<K, V>, P: AsInner<FileDesc>,
63 K: BytesContainer + Eq + Hash, V: BytesContainer
65 use libc::funcs::posix88::unistd::{fork, dup2, close, chdir, execvp};
66 use libc::funcs::bsd44::getdtablesize;
70 pub fn rust_unset_sigprocmask();
74 #[cfg(target_os = "macos")]
75 unsafe fn set_environ(envp: *const c_void) {
76 extern { fn _NSGetEnviron() -> *mut *const c_void; }
78 *_NSGetEnviron() = envp;
80 #[cfg(not(target_os = "macos"))]
81 unsafe fn set_environ(envp: *const c_void) {
82 extern { static mut environ: *const c_void; }
86 unsafe fn set_cloexec(fd: c_int) {
87 let ret = c::ioctl(fd, c::FIOCLEX);
91 let dirp = cfg.cwd().map(|c| c.as_ptr()).unwrap_or(ptr::null());
93 // temporary until unboxed closures land
95 mem::transmute::<&ProcessConfig<K,V>,&'static ProcessConfig<K,V>>(cfg)
98 with_envp(cfg.env(), move|: envp: *const c_void| {
99 with_argv(cfg.program(), cfg.args(), move|: argv: *const *const libc::c_char| unsafe {
100 let (input, mut output) = try!(sys::os::pipe());
102 // We may use this in the child, so perform allocations before the
104 let devnull = "/dev/null".to_c_str();
106 set_cloexec(output.fd());
110 return Err(super::last_error())
113 fn combine(arr: &[u8]) -> i32 {
114 let a = arr[0] as u32;
115 let b = arr[1] as u32;
116 let c = arr[2] as u32;
117 let d = arr[3] as u32;
119 ((a << 24) | (b << 16) | (c << 8) | (d << 0)) as i32
122 let p = Process{ pid: pid };
124 let mut bytes = [0; 8];
125 return match input.read(&mut bytes) {
127 assert!(combine(CLOEXEC_MSG_FOOTER) == combine(bytes.slice(4, 8)),
128 "Validation on the CLOEXEC pipe failed: {}", bytes);
129 let errno = combine(bytes.slice(0, 4));
130 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
131 Err(super::decode_error(errno))
133 Err(ref e) if e.kind == EndOfFile => Ok(p),
135 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
136 panic!("the CLOEXEC pipe failed: {}", e)
138 Ok(..) => { // pipe I/O up to PIPE_BUF bytes should be atomic
139 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
140 panic!("short read on the CLOEXEC pipe")
145 // And at this point we've reached a special time in the life of the
146 // child. The child must now be considered hamstrung and unable to
147 // do anything other than syscalls really. Consider the following
150 // 1. Thread A of process 1 grabs the malloc() mutex
151 // 2. Thread B of process 1 forks(), creating thread C
152 // 3. Thread C of process 2 then attempts to malloc()
153 // 4. The memory of process 2 is the same as the memory of
154 // process 1, so the mutex is locked.
156 // This situation looks a lot like deadlock, right? It turns out
157 // that this is what pthread_atfork() takes care of, which is
158 // presumably implemented across platforms. The first thing that
159 // threads to *before* forking is to do things like grab the malloc
160 // mutex, and then after the fork they unlock it.
162 // Despite this information, libnative's spawn has been witnessed to
163 // deadlock on both OSX and FreeBSD. I'm not entirely sure why, but
164 // all collected backtraces point at malloc/free traffic in the
165 // child spawned process.
167 // For this reason, the block of code below should contain 0
168 // invocations of either malloc of free (or their related friends).
170 // As an example of not having malloc/free traffic, we don't close
171 // this file descriptor by dropping the FileDesc (which contains an
172 // allocation). Instead we just close it manually. This will never
173 // have the drop glue anyway because this code never returns (the
174 // child will either exec() or invoke libc::exit)
175 let _ = libc::close(input.fd());
177 fn fail(output: &mut FileDesc) -> ! {
178 let errno = sys::os::errno() as u32;
184 CLOEXEC_MSG_FOOTER[0], CLOEXEC_MSG_FOOTER[1],
185 CLOEXEC_MSG_FOOTER[2], CLOEXEC_MSG_FOOTER[3]
187 // pipe I/O up to PIPE_BUF bytes should be atomic
188 assert!(output.write(&bytes).is_ok());
189 unsafe { libc::_exit(1) }
192 rustrt::rust_unset_sigprocmask();
194 // If a stdio file descriptor is set to be ignored (via a -1 file
195 // descriptor), then we don't actually close it, but rather open
196 // up /dev/null into that file descriptor. Otherwise, the first file
197 // descriptor opened up in the child would be numbered as one of the
198 // stdio file descriptors, which is likely to wreak havoc.
199 let setup = |&: src: Option<P>, dst: c_int| {
200 let src = match src {
202 let flags = if dst == libc::STDIN_FILENO {
207 libc::open(devnull.as_ptr(), flags, 0)
210 let fd = obj.as_inner().fd();
211 // Leak the memory and the file descriptor. We're in the
212 // child now an all our resources are going to be
213 // cleaned up very soon
218 src != -1 && retry(|| dup2(src, dst)) != -1
221 if !setup(in_fd, libc::STDIN_FILENO) { fail(&mut output) }
222 if !setup(out_fd, libc::STDOUT_FILENO) { fail(&mut output) }
223 if !setup(err_fd, libc::STDERR_FILENO) { fail(&mut output) }
225 // close all other fds
226 for fd in range(3, getdtablesize()).rev() {
227 if fd != output.fd() {
228 let _ = close(fd as c_int);
234 if libc::setgid(u as libc::gid_t) != 0 {
242 // When dropping privileges from root, the `setgroups` call
243 // will remove any extraneous groups. If we don't call this,
244 // then even though our uid has dropped, we may still have
245 // groups that enable us to do super-user things. This will
246 // fail if we aren't root, so don't bother checking the
247 // return value, this is just done as an optimistic
248 // privilege dropping function.
250 fn setgroups(ngroups: libc::c_int,
251 ptr: *const libc::c_void) -> libc::c_int;
253 let _ = setgroups(0, 0 as *const libc::c_void);
255 if libc::setuid(u as libc::uid_t) != 0 {
262 // Don't check the error of setsid because it fails if we're the
263 // process leader already. We just forked so it shouldn't return
264 // error, but ignore it anyway.
265 let _ = libc::setsid();
267 if !dirp.is_null() && chdir(dirp) == -1 {
273 let _ = execvp(*argv, argv as *mut _);
279 pub fn wait(&self, deadline: u64) -> IoResult<ProcessExit> {
281 use sync::mpsc::TryRecvError;
283 static mut WRITE_FD: libc::c_int = 0;
285 let mut status = 0 as c_int;
287 return match retry(|| unsafe { c::waitpid(self.pid, &mut status, 0) }) {
288 -1 => panic!("unknown waitpid error: {}", super::last_error()),
289 _ => Ok(translate_status(status)),
293 // On unix, wait() and its friends have no timeout parameters, so there is
294 // no way to time out a thread in wait(). From some googling and some
295 // thinking, it appears that there are a few ways to handle timeouts in
296 // wait(), but the only real reasonable one for a multi-threaded program is
297 // to listen for SIGCHLD.
299 // With this in mind, the waiting mechanism with a timeout barely uses
300 // waitpid() at all. There are a few times that waitpid() is invoked with
301 // WNOHANG, but otherwise all the necessary blocking is done by waiting for
302 // a SIGCHLD to arrive (and that blocking has a timeout). Note, however,
303 // that waitpid() is still used to actually reap the child.
305 // Signal handling is super tricky in general, and this is no exception. Due
306 // to the async nature of SIGCHLD, we use the self-pipe trick to transmit
307 // data out of the signal handler to the rest of the application. The first
308 // idea would be to have each thread waiting with a timeout to read this
309 // output file descriptor, but a write() is akin to a signal(), not a
310 // broadcast(), so it would only wake up one thread, and possibly the wrong
311 // thread. Hence a helper thread is used.
313 // The helper thread here is responsible for farming requests for a
314 // waitpid() with a timeout, and then processing all of the wait requests.
315 // By guaranteeing that only this helper thread is reading half of the
316 // self-pipe, we're sure that we'll never lose a SIGCHLD. This helper thread
317 // is also responsible for select() to wait for incoming messages or
318 // incoming SIGCHLD messages, along with passing an appropriate timeout to
319 // select() to wake things up as necessary.
321 // The ordering of the following statements is also very purposeful. First,
322 // we must be guaranteed that the helper thread is booted and available to
323 // receive SIGCHLD signals, and then we must also ensure that we do a
324 // nonblocking waitpid() at least once before we go ask the sigchld helper.
325 // This prevents the race where the child exits, we boot the helper, and
326 // then we ask for the child's exit status (never seeing a sigchld).
328 // The actual communication between the helper thread and this thread is
329 // quite simple, just a channel moving data around.
331 unsafe { HELPER.boot(register_sigchld, waitpid_helper) }
333 match self.try_wait() {
334 Some(ret) => return Ok(ret),
338 let (tx, rx) = channel();
339 unsafe { HELPER.send(NewChild(self.pid, tx, deadline)); }
340 return match rx.recv() {
342 Err(..) => Err(timeout("wait timed out")),
345 // Register a new SIGCHLD handler, returning the reading half of the
346 // self-pipe plus the old handler registered (return value of sigaction).
348 // Be sure to set up the self-pipe first because as soon as we register a
349 // handler we're going to start receiving signals.
350 fn register_sigchld() -> (libc::c_int, c::sigaction) {
352 let mut pipes = [0; 2];
353 assert_eq!(libc::pipe(pipes.as_mut_ptr()), 0);
354 set_nonblocking(pipes[0], true).ok().unwrap();
355 set_nonblocking(pipes[1], true).ok().unwrap();
358 let mut old: c::sigaction = mem::zeroed();
359 let mut new: c::sigaction = mem::zeroed();
360 new.sa_handler = sigchld_handler;
361 new.sa_flags = c::SA_NOCLDSTOP;
362 assert_eq!(c::sigaction(c::SIGCHLD, &new, &mut old), 0);
367 // Helper thread for processing SIGCHLD messages
368 fn waitpid_helper(input: libc::c_int,
369 messages: Receiver<Req>,
370 (read_fd, old): (libc::c_int, c::sigaction)) {
371 set_nonblocking(input, true).ok().unwrap();
372 let mut set: c::fd_set = unsafe { mem::zeroed() };
373 let mut tv: libc::timeval;
374 let mut active = Vec::<(libc::pid_t, Sender<ProcessExit>, u64)>::new();
375 let max = cmp::max(input, read_fd) + 1;
378 // Figure out the timeout of our syscall-to-happen. If we're waiting
379 // for some processes, then they'll have a timeout, otherwise we
380 // wait indefinitely for a message to arrive.
382 // FIXME: sure would be nice to not have to scan the entire array
383 let min = active.iter().map(|a| a.2).enumerate().min_by(|p| {
386 let (p, idx) = match min {
387 Some((idx, deadline)) => {
388 let now = sys::timer::now();
389 let ms = if now < deadline {deadline - now} else {0};
390 tv = ms_to_timeval(ms);
391 (&mut tv as *mut _, idx)
393 None => (ptr::null_mut(), -1),
396 // Wait for something to happen
397 c::fd_set(&mut set, input);
398 c::fd_set(&mut set, read_fd);
399 match unsafe { c::select(max, &mut set, ptr::null_mut(),
400 ptr::null_mut(), p) } {
401 // interrupted, retry
402 -1 if os::errno() == libc::EINTR as uint => continue,
404 // We read something, break out and process
407 // Timeout, the pending request is removed
409 drop(active.remove(idx));
413 n => panic!("error in select {} ({})", os::errno(), n),
416 // Process any pending messages
419 match messages.try_recv() {
420 Ok(NewChild(pid, tx, deadline)) => {
421 active.push((pid, tx, deadline));
423 Err(TryRecvError::Disconnected) => {
424 assert!(active.len() == 0);
427 Err(TryRecvError::Empty) => break,
432 // If a child exited (somehow received SIGCHLD), then poll all
433 // children to see if any of them exited.
435 // We also attempt to be responsible netizens when dealing with
436 // SIGCHLD by invoking any previous SIGCHLD handler instead of just
437 // ignoring any previous SIGCHLD handler. Note that we don't provide
438 // a 1:1 mapping of our handler invocations to the previous handler
439 // invocations because we drain the `read_fd` entirely. This is
440 // probably OK because the kernel is already allowed to coalesce
441 // simultaneous signals, we're just doing some extra coalescing.
443 // Another point of note is that this likely runs the signal handler
444 // on a different thread than the one that received the signal. I
445 // *think* this is ok at this time.
447 // The main reason for doing this is to allow stdtest to run native
448 // tests as well. Both libgreen and libnative are running around
449 // with process timeouts, but libgreen should get there first
450 // (currently libuv doesn't handle old signal handlers).
452 let i: uint = unsafe { mem::transmute(old.sa_handler) };
454 assert!(old.sa_flags & c::SA_SIGINFO == 0);
455 (old.sa_handler)(c::SIGCHLD);
458 // FIXME: sure would be nice to not have to scan the entire
460 active.retain(|&(pid, ref tx, _)| {
461 let pr = Process { pid: pid };
462 match pr.try_wait() {
463 Some(msg) => { tx.send(msg).unwrap(); false }
470 // Once this helper thread is done, we re-register the old sigchld
471 // handler and close our intermediate file descriptors.
473 assert_eq!(c::sigaction(c::SIGCHLD, &old, ptr::null_mut()), 0);
474 let _ = libc::close(read_fd);
475 let _ = libc::close(WRITE_FD);
480 // Drain all pending data from the file descriptor, returning if any data
481 // could be drained. This requires that the file descriptor is in
483 fn drain(fd: libc::c_int) -> bool {
486 let mut buf = [0u8; 1];
488 libc::read(fd, buf.as_mut_ptr() as *mut libc::c_void,
489 buf.len() as libc::size_t)
491 n if n > 0 => { ret = true; }
493 -1 if wouldblock() => return ret,
494 n => panic!("bad read {} ({})", os::last_os_error(), n),
499 // Signal handler for SIGCHLD signals, must be async-signal-safe!
501 // This function will write to the writing half of the "self pipe" to wake
502 // up the helper thread if it's waiting. Note that this write must be
503 // nonblocking because if it blocks and the reader is the thread we
504 // interrupted, then we'll deadlock.
506 // When writing, if the write returns EWOULDBLOCK then we choose to ignore
507 // it. At that point we're guaranteed that there's something in the pipe
508 // which will wake up the other end at some point, so we just allow this
509 // signal to be coalesced with the pending signals on the pipe.
510 extern fn sigchld_handler(_signum: libc::c_int) {
513 libc::write(WRITE_FD, &msg as *const _ as *const libc::c_void, 1)
516 -1 if wouldblock() => {} // see above comments
517 n => panic!("bad error on write fd: {} {}", n, os::errno()),
522 pub fn try_wait(&self) -> Option<ProcessExit> {
523 let mut status = 0 as c_int;
524 match retry(|| unsafe {
525 c::waitpid(self.pid, &mut status, c::WNOHANG)
527 n if n == self.pid => Some(translate_status(status)),
529 n => panic!("unknown waitpid error `{}`: {}", n,
530 super::last_error()),
535 fn with_argv<T,F>(prog: &CString, args: &[CString],
538 where F : FnOnce(*const *const libc::c_char) -> T
540 let mut ptrs: Vec<*const libc::c_char> = Vec::with_capacity(args.len()+1);
542 // Convert the CStrings into an array of pointers. Note: the
543 // lifetime of the various CStrings involved is guaranteed to be
544 // larger than the lifetime of our invocation of cb, but this is
545 // technically unsafe as the callback could leak these pointers
547 ptrs.push(prog.as_ptr());
548 ptrs.extend(args.iter().map(|tmp| tmp.as_ptr()));
550 // Add a terminating null pointer (required by libc).
551 ptrs.push(ptr::null());
556 fn with_envp<K,V,T,F>(env: Option<&collections::HashMap<K, V>>,
559 where F : FnOnce(*const c_void) -> T,
560 K : BytesContainer + Eq + Hash,
563 // On posixy systems we can pass a char** for envp, which is a
564 // null-terminated array of "k=v\0" strings. Since we must create
565 // these strings locally, yet expose a raw pointer to them, we
566 // create a temporary vector to own the CStrings that outlives the
570 let mut tmps = Vec::with_capacity(env.len());
572 for pair in env.iter() {
573 let mut kv = Vec::new();
574 kv.push_all(pair.0.container_as_bytes());
576 kv.push_all(pair.1.container_as_bytes());
577 kv.push(0); // terminating null
581 // As with `with_argv`, this is unsafe, since cb could leak the pointers.
582 let mut ptrs: Vec<*const libc::c_char> =
584 .map(|tmp| tmp.as_ptr() as *const libc::c_char)
586 ptrs.push(ptr::null());
588 cb(ptrs.as_ptr() as *const c_void)
594 fn translate_status(status: c_int) -> ProcessExit {
595 #![allow(non_snake_case)]
596 #[cfg(any(target_os = "linux", target_os = "android"))]
598 pub fn WIFEXITED(status: i32) -> bool { (status & 0xff) == 0 }
599 pub fn WEXITSTATUS(status: i32) -> i32 { (status >> 8) & 0xff }
600 pub fn WTERMSIG(status: i32) -> i32 { status & 0x7f }
603 #[cfg(any(target_os = "macos",
605 target_os = "freebsd",
606 target_os = "dragonfly"))]
608 pub fn WIFEXITED(status: i32) -> bool { (status & 0x7f) == 0 }
609 pub fn WEXITSTATUS(status: i32) -> i32 { status >> 8 }
610 pub fn WTERMSIG(status: i32) -> i32 { status & 0o177 }
613 if imp::WIFEXITED(status) {
614 ExitStatus(imp::WEXITSTATUS(status) as int)
616 ExitSignal(imp::WTERMSIG(status) as int)