1 // Copyright 2014-2015 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 collections::HashMap;
15 use collections::hash_map::Hasher;
18 use old_io::process::{ProcessExit, ExitStatus, ExitSignal};
19 use old_io::{self, IoResult, IoError, EndOfFile};
20 use libc::{self, pid_t, c_void, c_int};
23 use old_path::BytesContainer;
25 use sync::mpsc::{channel, Sender, Receiver};
26 use sys::fs::FileDesc;
27 use sys::{self, retry, c, wouldblock, set_nonblocking, ms_to_timeval};
28 use sys_common::helper_thread::Helper;
29 use sys_common::{AsInner, mkerr_libc, timeout};
31 pub use sys_common::ProcessConfig;
33 helper_init! { static HELPER: Helper<Req> }
35 /// The unique id of the process (this should never be negative).
41 NewChild(libc::pid_t, Sender<ProcessExit>, u64),
44 const CLOEXEC_MSG_FOOTER: &'static [u8] = b"NOEX";
47 pub fn id(&self) -> pid_t {
51 pub unsafe fn kill(&self, signal: int) -> IoResult<()> {
52 Process::killpid(self.pid, signal)
55 pub unsafe fn killpid(pid: pid_t, signal: int) -> IoResult<()> {
56 let r = libc::funcs::posix88::signal::kill(pid, signal as c_int);
60 pub fn spawn<K, V, C, P>(cfg: &C, in_fd: Option<P>,
61 out_fd: Option<P>, err_fd: Option<P>)
63 where C: ProcessConfig<K, V>, P: AsInner<FileDesc>,
64 K: BytesContainer + Eq + Hash<Hasher>, V: BytesContainer
66 use libc::funcs::posix88::unistd::{fork, dup2, close, chdir, execvp};
67 use libc::funcs::bsd44::getdtablesize;
71 pub fn rust_unset_sigprocmask();
75 unsafe fn set_cloexec(fd: c_int) {
76 let ret = c::ioctl(fd, c::FIOCLEX);
80 let dirp = cfg.cwd().map(|c| c.as_ptr()).unwrap_or(ptr::null());
82 // temporary until unboxed closures land
84 mem::transmute::<&ProcessConfig<K,V>,&'static ProcessConfig<K,V>>(cfg)
87 with_envp(cfg.env(), move|envp: *const c_void| {
88 with_argv(cfg.program(), cfg.args(), move|argv: *const *const libc::c_char| unsafe {
89 let (input, mut output) = try!(sys::os::pipe());
91 // We may use this in the child, so perform allocations before the
93 let devnull = b"/dev/null\0";
95 set_cloexec(output.fd());
99 return Err(super::last_error())
102 fn combine(arr: &[u8]) -> i32 {
103 let a = arr[0] as u32;
104 let b = arr[1] as u32;
105 let c = arr[2] as u32;
106 let d = arr[3] as u32;
108 ((a << 24) | (b << 16) | (c << 8) | (d << 0)) as i32
111 let p = Process{ pid: pid };
113 let mut bytes = [0; 8];
114 return match input.read(&mut bytes) {
116 assert!(combine(CLOEXEC_MSG_FOOTER) == combine(&bytes[4.. 8]),
117 "Validation on the CLOEXEC pipe failed: {:?}", bytes);
118 let errno = combine(&bytes[0.. 4]);
119 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
120 Err(super::decode_error(errno))
122 Err(ref e) if e.kind == EndOfFile => Ok(p),
124 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
125 panic!("the CLOEXEC pipe failed: {:?}", e)
127 Ok(..) => { // pipe I/O up to PIPE_BUF bytes should be atomic
128 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
129 panic!("short read on the CLOEXEC pipe")
134 // And at this point we've reached a special time in the life of the
135 // child. The child must now be considered hamstrung and unable to
136 // do anything other than syscalls really. Consider the following
139 // 1. Thread A of process 1 grabs the malloc() mutex
140 // 2. Thread B of process 1 forks(), creating thread C
141 // 3. Thread C of process 2 then attempts to malloc()
142 // 4. The memory of process 2 is the same as the memory of
143 // process 1, so the mutex is locked.
145 // This situation looks a lot like deadlock, right? It turns out
146 // that this is what pthread_atfork() takes care of, which is
147 // presumably implemented across platforms. The first thing that
148 // threads to *before* forking is to do things like grab the malloc
149 // mutex, and then after the fork they unlock it.
151 // Despite this information, libnative's spawn has been witnessed to
152 // deadlock on both OSX and FreeBSD. I'm not entirely sure why, but
153 // all collected backtraces point at malloc/free traffic in the
154 // child spawned process.
156 // For this reason, the block of code below should contain 0
157 // invocations of either malloc of free (or their related friends).
159 // As an example of not having malloc/free traffic, we don't close
160 // this file descriptor by dropping the FileDesc (which contains an
161 // allocation). Instead we just close it manually. This will never
162 // have the drop glue anyway because this code never returns (the
163 // child will either exec() or invoke libc::exit)
164 let _ = libc::close(input.fd());
166 fn fail(output: &mut FileDesc) -> ! {
167 let errno = sys::os::errno() as u32;
173 CLOEXEC_MSG_FOOTER[0], CLOEXEC_MSG_FOOTER[1],
174 CLOEXEC_MSG_FOOTER[2], CLOEXEC_MSG_FOOTER[3]
176 // pipe I/O up to PIPE_BUF bytes should be atomic
177 assert!(output.write(&bytes).is_ok());
178 unsafe { libc::_exit(1) }
181 rustrt::rust_unset_sigprocmask();
183 // If a stdio file descriptor is set to be ignored (via a -1 file
184 // descriptor), then we don't actually close it, but rather open
185 // up /dev/null into that file descriptor. Otherwise, the first file
186 // descriptor opened up in the child would be numbered as one of the
187 // stdio file descriptors, which is likely to wreak havoc.
188 let setup = |src: Option<P>, dst: c_int| {
189 let src = match src {
191 let flags = if dst == libc::STDIN_FILENO {
196 libc::open(devnull.as_ptr() as *const _, flags, 0)
199 let fd = obj.as_inner().fd();
200 // Leak the memory and the file descriptor. We're in the
201 // child now an all our resources are going to be
202 // cleaned up very soon
207 src != -1 && retry(|| dup2(src, dst)) != -1
210 if !setup(in_fd, libc::STDIN_FILENO) { fail(&mut output) }
211 if !setup(out_fd, libc::STDOUT_FILENO) { fail(&mut output) }
212 if !setup(err_fd, libc::STDERR_FILENO) { fail(&mut output) }
214 // close all other fds
215 for fd in (3..getdtablesize()).rev() {
216 if fd != output.fd() {
217 let _ = close(fd as c_int);
223 if libc::setgid(u as libc::gid_t) != 0 {
231 // When dropping privileges from root, the `setgroups` call
232 // will remove any extraneous groups. If we don't call this,
233 // then even though our uid has dropped, we may still have
234 // groups that enable us to do super-user things. This will
235 // fail if we aren't root, so don't bother checking the
236 // return value, this is just done as an optimistic
237 // privilege dropping function.
239 fn setgroups(ngroups: libc::c_int,
240 ptr: *const libc::c_void) -> libc::c_int;
242 let _ = setgroups(0, ptr::null());
244 if libc::setuid(u as libc::uid_t) != 0 {
251 // Don't check the error of setsid because it fails if we're the
252 // process leader already. We just forked so it shouldn't return
253 // error, but ignore it anyway.
254 let _ = libc::setsid();
256 if !dirp.is_null() && chdir(dirp) == -1 {
260 *sys::os::environ() = envp as *const _;
262 let _ = execvp(*argv, argv as *mut _);
268 pub fn wait(&self, deadline: u64) -> IoResult<ProcessExit> {
270 use sync::mpsc::TryRecvError;
272 static mut WRITE_FD: libc::c_int = 0;
274 let mut status = 0 as c_int;
276 return match retry(|| unsafe { c::waitpid(self.pid, &mut status, 0) }) {
277 -1 => panic!("unknown waitpid error: {:?}", super::last_error()),
278 _ => Ok(translate_status(status)),
282 // On unix, wait() and its friends have no timeout parameters, so there is
283 // no way to time out a thread in wait(). From some googling and some
284 // thinking, it appears that there are a few ways to handle timeouts in
285 // wait(), but the only real reasonable one for a multi-threaded program is
286 // to listen for SIGCHLD.
288 // With this in mind, the waiting mechanism with a timeout barely uses
289 // waitpid() at all. There are a few times that waitpid() is invoked with
290 // WNOHANG, but otherwise all the necessary blocking is done by waiting for
291 // a SIGCHLD to arrive (and that blocking has a timeout). Note, however,
292 // that waitpid() is still used to actually reap the child.
294 // Signal handling is super tricky in general, and this is no exception. Due
295 // to the async nature of SIGCHLD, we use the self-pipe trick to transmit
296 // data out of the signal handler to the rest of the application. The first
297 // idea would be to have each thread waiting with a timeout to read this
298 // output file descriptor, but a write() is akin to a signal(), not a
299 // broadcast(), so it would only wake up one thread, and possibly the wrong
300 // thread. Hence a helper thread is used.
302 // The helper thread here is responsible for farming requests for a
303 // waitpid() with a timeout, and then processing all of the wait requests.
304 // By guaranteeing that only this helper thread is reading half of the
305 // self-pipe, we're sure that we'll never lose a SIGCHLD. This helper thread
306 // is also responsible for select() to wait for incoming messages or
307 // incoming SIGCHLD messages, along with passing an appropriate timeout to
308 // select() to wake things up as necessary.
310 // The ordering of the following statements is also very purposeful. First,
311 // we must be guaranteed that the helper thread is booted and available to
312 // receive SIGCHLD signals, and then we must also ensure that we do a
313 // nonblocking waitpid() at least once before we go ask the sigchld helper.
314 // This prevents the race where the child exits, we boot the helper, and
315 // then we ask for the child's exit status (never seeing a sigchld).
317 // The actual communication between the helper thread and this thread is
318 // quite simple, just a channel moving data around.
320 unsafe { HELPER.boot(register_sigchld, waitpid_helper) }
322 match self.try_wait() {
323 Some(ret) => return Ok(ret),
327 let (tx, rx) = channel();
328 unsafe { HELPER.send(NewChild(self.pid, tx, deadline)); }
329 return match rx.recv() {
331 Err(..) => Err(timeout("wait timed out")),
334 // Register a new SIGCHLD handler, returning the reading half of the
335 // self-pipe plus the old handler registered (return value of sigaction).
337 // Be sure to set up the self-pipe first because as soon as we register a
338 // handler we're going to start receiving signals.
339 fn register_sigchld() -> (libc::c_int, c::sigaction) {
341 let mut pipes = [0; 2];
342 assert_eq!(libc::pipe(pipes.as_mut_ptr()), 0);
343 set_nonblocking(pipes[0], true).ok().unwrap();
344 set_nonblocking(pipes[1], true).ok().unwrap();
347 let mut old: c::sigaction = mem::zeroed();
348 let mut new: c::sigaction = mem::zeroed();
349 new.sa_handler = sigchld_handler;
350 new.sa_flags = c::SA_NOCLDSTOP;
351 assert_eq!(c::sigaction(c::SIGCHLD, &new, &mut old), 0);
356 // Helper thread for processing SIGCHLD messages
357 fn waitpid_helper(input: libc::c_int,
358 messages: Receiver<Req>,
359 (read_fd, old): (libc::c_int, c::sigaction)) {
360 set_nonblocking(input, true).ok().unwrap();
361 let mut set: c::fd_set = unsafe { mem::zeroed() };
362 let mut tv: libc::timeval;
363 let mut active = Vec::<(libc::pid_t, Sender<ProcessExit>, u64)>::new();
364 let max = cmp::max(input, read_fd) + 1;
367 // Figure out the timeout of our syscall-to-happen. If we're waiting
368 // for some processes, then they'll have a timeout, otherwise we
369 // wait indefinitely for a message to arrive.
371 // FIXME: sure would be nice to not have to scan the entire array
372 let min = active.iter().map(|a| a.2).enumerate().min_by(|p| {
375 let (p, idx) = match min {
376 Some((idx, deadline)) => {
377 let now = sys::timer::now();
378 let ms = if now < deadline {deadline - now} else {0};
379 tv = ms_to_timeval(ms);
380 (&mut tv as *mut _, idx)
382 None => (ptr::null_mut(), -1),
385 // Wait for something to happen
386 c::fd_set(&mut set, input);
387 c::fd_set(&mut set, read_fd);
388 match unsafe { c::select(max, &mut set, ptr::null_mut(),
389 ptr::null_mut(), p) } {
390 // interrupted, retry
391 -1 if os::errno() == libc::EINTR as uint => continue,
393 // We read something, break out and process
396 // Timeout, the pending request is removed
398 drop(active.remove(idx));
402 n => panic!("error in select {:?} ({:?})", os::errno(), n),
405 // Process any pending messages
408 match messages.try_recv() {
409 Ok(NewChild(pid, tx, deadline)) => {
410 active.push((pid, tx, deadline));
412 Err(TryRecvError::Disconnected) => {
413 assert!(active.len() == 0);
416 Err(TryRecvError::Empty) => break,
421 // If a child exited (somehow received SIGCHLD), then poll all
422 // children to see if any of them exited.
424 // We also attempt to be responsible netizens when dealing with
425 // SIGCHLD by invoking any previous SIGCHLD handler instead of just
426 // ignoring any previous SIGCHLD handler. Note that we don't provide
427 // a 1:1 mapping of our handler invocations to the previous handler
428 // invocations because we drain the `read_fd` entirely. This is
429 // probably OK because the kernel is already allowed to coalesce
430 // simultaneous signals, we're just doing some extra coalescing.
432 // Another point of note is that this likely runs the signal handler
433 // on a different thread than the one that received the signal. I
434 // *think* this is ok at this time.
436 // The main reason for doing this is to allow stdtest to run native
437 // tests as well. Both libgreen and libnative are running around
438 // with process timeouts, but libgreen should get there first
439 // (currently libuv doesn't handle old signal handlers).
441 let i: uint = unsafe { mem::transmute(old.sa_handler) };
443 assert!(old.sa_flags & c::SA_SIGINFO == 0);
444 (old.sa_handler)(c::SIGCHLD);
447 // FIXME: sure would be nice to not have to scan the entire
449 active.retain(|&(pid, ref tx, _)| {
450 let pr = Process { pid: pid };
451 match pr.try_wait() {
452 Some(msg) => { tx.send(msg).unwrap(); false }
459 // Once this helper thread is done, we re-register the old sigchld
460 // handler and close our intermediate file descriptors.
462 assert_eq!(c::sigaction(c::SIGCHLD, &old, ptr::null_mut()), 0);
463 let _ = libc::close(read_fd);
464 let _ = libc::close(WRITE_FD);
469 // Drain all pending data from the file descriptor, returning if any data
470 // could be drained. This requires that the file descriptor is in
472 fn drain(fd: libc::c_int) -> bool {
475 let mut buf = [0u8; 1];
477 libc::read(fd, buf.as_mut_ptr() as *mut libc::c_void,
478 buf.len() as libc::size_t)
480 n if n > 0 => { ret = true; }
482 -1 if wouldblock() => return ret,
483 n => panic!("bad read {:?} ({:?})", os::last_os_error(), n),
488 // Signal handler for SIGCHLD signals, must be async-signal-safe!
490 // This function will write to the writing half of the "self pipe" to wake
491 // up the helper thread if it's waiting. Note that this write must be
492 // nonblocking because if it blocks and the reader is the thread we
493 // interrupted, then we'll deadlock.
495 // When writing, if the write returns EWOULDBLOCK then we choose to ignore
496 // it. At that point we're guaranteed that there's something in the pipe
497 // which will wake up the other end at some point, so we just allow this
498 // signal to be coalesced with the pending signals on the pipe.
499 extern fn sigchld_handler(_signum: libc::c_int) {
502 libc::write(WRITE_FD, &msg as *const _ as *const libc::c_void, 1)
505 -1 if wouldblock() => {} // see above comments
506 n => panic!("bad error on write fd: {:?} {:?}", n, os::errno()),
511 pub fn try_wait(&self) -> Option<ProcessExit> {
512 let mut status = 0 as c_int;
513 match retry(|| unsafe {
514 c::waitpid(self.pid, &mut status, c::WNOHANG)
516 n if n == self.pid => Some(translate_status(status)),
518 n => panic!("unknown waitpid error `{:?}`: {:?}", n,
519 super::last_error()),
524 fn with_argv<T,F>(prog: &CString, args: &[CString],
527 where F : FnOnce(*const *const libc::c_char) -> T
529 let mut ptrs: Vec<*const libc::c_char> = Vec::with_capacity(args.len()+1);
531 // Convert the CStrings into an array of pointers. Note: the
532 // lifetime of the various CStrings involved is guaranteed to be
533 // larger than the lifetime of our invocation of cb, but this is
534 // technically unsafe as the callback could leak these pointers
536 ptrs.push(prog.as_ptr());
537 ptrs.extend(args.iter().map(|tmp| tmp.as_ptr()));
539 // Add a terminating null pointer (required by libc).
540 ptrs.push(ptr::null());
545 fn with_envp<K,V,T,F>(env: Option<&HashMap<K, V>>,
548 where F : FnOnce(*const c_void) -> T,
549 K : BytesContainer + Eq + Hash<Hasher>,
552 // On posixy systems we can pass a char** for envp, which is a
553 // null-terminated array of "k=v\0" strings. Since we must create
554 // these strings locally, yet expose a raw pointer to them, we
555 // create a temporary vector to own the CStrings that outlives the
559 let mut tmps = Vec::with_capacity(env.len());
562 let mut kv = Vec::new();
563 kv.push_all(pair.0.container_as_bytes());
565 kv.push_all(pair.1.container_as_bytes());
566 kv.push(0); // terminating null
570 // As with `with_argv`, this is unsafe, since cb could leak the pointers.
571 let mut ptrs: Vec<*const libc::c_char> =
573 .map(|tmp| tmp.as_ptr() as *const libc::c_char)
575 ptrs.push(ptr::null());
577 cb(ptrs.as_ptr() as *const c_void)
583 fn translate_status(status: c_int) -> ProcessExit {
584 #![allow(non_snake_case)]
585 #[cfg(any(target_os = "linux", target_os = "android"))]
587 pub fn WIFEXITED(status: i32) -> bool { (status & 0xff) == 0 }
588 pub fn WEXITSTATUS(status: i32) -> i32 { (status >> 8) & 0xff }
589 pub fn WTERMSIG(status: i32) -> i32 { status & 0x7f }
592 #[cfg(any(target_os = "macos",
594 target_os = "freebsd",
595 target_os = "dragonfly",
596 target_os = "openbsd"))]
598 pub fn WIFEXITED(status: i32) -> bool { (status & 0x7f) == 0 }
599 pub fn WEXITSTATUS(status: i32) -> i32 { status >> 8 }
600 pub fn WTERMSIG(status: i32) -> i32 { status & 0o177 }
603 if imp::WIFEXITED(status) {
604 ExitStatus(imp::WEXITSTATUS(status) as int)
606 ExitSignal(imp::WTERMSIG(status) as int)