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 /// Unix-specific extensions to the Command builder
36 pub struct CommandExt {
41 /// The unique id of the process (this should never be negative).
47 NewChild(libc::pid_t, Sender<ProcessExit>, u64),
50 const CLOEXEC_MSG_FOOTER: &'static [u8] = b"NOEX";
53 pub fn id(&self) -> pid_t {
57 pub unsafe fn kill(&self, signal: int) -> IoResult<()> {
58 Process::killpid(self.pid, signal)
61 pub unsafe fn killpid(pid: pid_t, signal: int) -> IoResult<()> {
62 let r = libc::funcs::posix88::signal::kill(pid, signal as c_int);
66 pub fn spawn<K, V, C, P>(cfg: &C, in_fd: Option<P>,
67 out_fd: Option<P>, err_fd: Option<P>)
69 where C: ProcessConfig<K, V>, P: AsInner<FileDesc>,
70 K: BytesContainer + Eq + Hash<Hasher>, V: BytesContainer
72 use libc::funcs::posix88::unistd::{fork, dup2, close, chdir, execvp};
76 pub fn rust_unset_sigprocmask();
80 #[cfg(all(target_os = "android", target_arch = "aarch64"))]
81 unsafe fn getdtablesize() -> c_int {
82 libc::sysconf(libc::consts::os::sysconf::_SC_OPEN_MAX) as c_int
84 #[cfg(not(all(target_os = "android", target_arch = "aarch64")))]
85 unsafe fn getdtablesize() -> c_int {
86 libc::funcs::bsd44::getdtablesize()
89 unsafe fn set_cloexec(fd: c_int) {
90 let ret = c::ioctl(fd, c::FIOCLEX);
94 let dirp = cfg.cwd().map(|c| c.as_ptr()).unwrap_or(ptr::null());
96 // temporary until unboxed closures land
98 mem::transmute::<&ProcessConfig<K,V>,&'static ProcessConfig<K,V>>(cfg)
101 with_envp(cfg.env(), move|envp: *const c_void| {
102 with_argv(cfg.program(), cfg.args(), move|argv: *const *const libc::c_char| unsafe {
103 let (input, mut output) = try!(sys::os::pipe());
105 // We may use this in the child, so perform allocations before the
107 let devnull = b"/dev/null\0";
109 set_cloexec(output.fd());
113 return Err(super::last_error())
116 fn combine(arr: &[u8]) -> i32 {
117 let a = arr[0] as u32;
118 let b = arr[1] as u32;
119 let c = arr[2] as u32;
120 let d = arr[3] as u32;
122 ((a << 24) | (b << 16) | (c << 8) | (d << 0)) as i32
125 let p = Process{ pid: pid };
127 let mut bytes = [0; 8];
128 return match input.read(&mut bytes) {
130 assert!(combine(CLOEXEC_MSG_FOOTER) == combine(&bytes[4.. 8]),
131 "Validation on the CLOEXEC pipe failed: {:?}", bytes);
132 let errno = combine(&bytes[0.. 4]);
133 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
134 Err(super::decode_error(errno))
136 Err(ref e) if e.kind == EndOfFile => Ok(p),
138 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
139 panic!("the CLOEXEC pipe failed: {:?}", e)
141 Ok(..) => { // pipe I/O up to PIPE_BUF bytes should be atomic
142 assert!(p.wait(0).is_ok(), "wait(0) should either return Ok or panic");
143 panic!("short read on the CLOEXEC pipe")
148 // And at this point we've reached a special time in the life of the
149 // child. The child must now be considered hamstrung and unable to
150 // do anything other than syscalls really. Consider the following
153 // 1. Thread A of process 1 grabs the malloc() mutex
154 // 2. Thread B of process 1 forks(), creating thread C
155 // 3. Thread C of process 2 then attempts to malloc()
156 // 4. The memory of process 2 is the same as the memory of
157 // process 1, so the mutex is locked.
159 // This situation looks a lot like deadlock, right? It turns out
160 // that this is what pthread_atfork() takes care of, which is
161 // presumably implemented across platforms. The first thing that
162 // threads to *before* forking is to do things like grab the malloc
163 // mutex, and then after the fork they unlock it.
165 // Despite this information, libnative's spawn has been witnessed to
166 // deadlock on both OSX and FreeBSD. I'm not entirely sure why, but
167 // all collected backtraces point at malloc/free traffic in the
168 // child spawned process.
170 // For this reason, the block of code below should contain 0
171 // invocations of either malloc of free (or their related friends).
173 // As an example of not having malloc/free traffic, we don't close
174 // this file descriptor by dropping the FileDesc (which contains an
175 // allocation). Instead we just close it manually. This will never
176 // have the drop glue anyway because this code never returns (the
177 // child will either exec() or invoke libc::exit)
178 let _ = libc::close(input.fd());
180 fn fail(output: &mut FileDesc) -> ! {
181 let errno = sys::os::errno() as u32;
187 CLOEXEC_MSG_FOOTER[0], CLOEXEC_MSG_FOOTER[1],
188 CLOEXEC_MSG_FOOTER[2], CLOEXEC_MSG_FOOTER[3]
190 // pipe I/O up to PIPE_BUF bytes should be atomic
191 assert!(output.write(&bytes).is_ok());
192 unsafe { libc::_exit(1) }
195 rustrt::rust_unset_sigprocmask();
197 // If a stdio file descriptor is set to be ignored (via a -1 file
198 // descriptor), then we don't actually close it, but rather open
199 // up /dev/null into that file descriptor. Otherwise, the first file
200 // descriptor opened up in the child would be numbered as one of the
201 // stdio file descriptors, which is likely to wreak havoc.
202 let setup = |src: Option<P>, dst: c_int| {
203 let src = match src {
205 let flags = if dst == libc::STDIN_FILENO {
210 libc::open(devnull.as_ptr() as *const _, flags, 0)
213 let fd = obj.as_inner().fd();
214 // Leak the memory and the file descriptor. We're in the
215 // child now an all our resources are going to be
216 // cleaned up very soon
221 src != -1 && retry(|| dup2(src, dst)) != -1
224 if !setup(in_fd, libc::STDIN_FILENO) { fail(&mut output) }
225 if !setup(out_fd, libc::STDOUT_FILENO) { fail(&mut output) }
226 if !setup(err_fd, libc::STDERR_FILENO) { fail(&mut output) }
228 // close all other fds
229 for fd in (3..getdtablesize()).rev() {
230 if fd != output.fd() {
231 let _ = close(fd as c_int);
237 if libc::setgid(u as libc::gid_t) != 0 {
245 // When dropping privileges from root, the `setgroups` call
246 // will remove any extraneous groups. If we don't call this,
247 // then even though our uid has dropped, we may still have
248 // groups that enable us to do super-user things. This will
249 // fail if we aren't root, so don't bother checking the
250 // return value, this is just done as an optimistic
251 // privilege dropping function.
253 fn setgroups(ngroups: libc::c_int,
254 ptr: *const libc::c_void) -> libc::c_int;
256 let _ = setgroups(0, ptr::null());
258 if libc::setuid(u as libc::uid_t) != 0 {
265 // Don't check the error of setsid because it fails if we're the
266 // process leader already. We just forked so it shouldn't return
267 // error, but ignore it anyway.
268 let _ = libc::setsid();
270 if !dirp.is_null() && chdir(dirp) == -1 {
274 *sys::os::environ() = envp as *const _;
276 let _ = execvp(*argv, argv as *mut _);
282 pub fn wait(&self, deadline: u64) -> IoResult<ProcessExit> {
284 use sync::mpsc::TryRecvError;
286 static mut WRITE_FD: libc::c_int = 0;
288 let mut status = 0 as c_int;
290 return match retry(|| unsafe { c::waitpid(self.pid, &mut status, 0) }) {
291 -1 => panic!("unknown waitpid error: {:?}", super::last_error()),
292 _ => Ok(translate_status(status)),
296 // On unix, wait() and its friends have no timeout parameters, so there is
297 // no way to time out a thread in wait(). From some googling and some
298 // thinking, it appears that there are a few ways to handle timeouts in
299 // wait(), but the only real reasonable one for a multi-threaded program is
300 // to listen for SIGCHLD.
302 // With this in mind, the waiting mechanism with a timeout barely uses
303 // waitpid() at all. There are a few times that waitpid() is invoked with
304 // WNOHANG, but otherwise all the necessary blocking is done by waiting for
305 // a SIGCHLD to arrive (and that blocking has a timeout). Note, however,
306 // that waitpid() is still used to actually reap the child.
308 // Signal handling is super tricky in general, and this is no exception. Due
309 // to the async nature of SIGCHLD, we use the self-pipe trick to transmit
310 // data out of the signal handler to the rest of the application. The first
311 // idea would be to have each thread waiting with a timeout to read this
312 // output file descriptor, but a write() is akin to a signal(), not a
313 // broadcast(), so it would only wake up one thread, and possibly the wrong
314 // thread. Hence a helper thread is used.
316 // The helper thread here is responsible for farming requests for a
317 // waitpid() with a timeout, and then processing all of the wait requests.
318 // By guaranteeing that only this helper thread is reading half of the
319 // self-pipe, we're sure that we'll never lose a SIGCHLD. This helper thread
320 // is also responsible for select() to wait for incoming messages or
321 // incoming SIGCHLD messages, along with passing an appropriate timeout to
322 // select() to wake things up as necessary.
324 // The ordering of the following statements is also very purposeful. First,
325 // we must be guaranteed that the helper thread is booted and available to
326 // receive SIGCHLD signals, and then we must also ensure that we do a
327 // nonblocking waitpid() at least once before we go ask the sigchld helper.
328 // This prevents the race where the child exits, we boot the helper, and
329 // then we ask for the child's exit status (never seeing a sigchld).
331 // The actual communication between the helper thread and this thread is
332 // quite simple, just a channel moving data around.
334 unsafe { HELPER.boot(register_sigchld, waitpid_helper) }
336 match self.try_wait() {
337 Some(ret) => return Ok(ret),
341 let (tx, rx) = channel();
342 unsafe { HELPER.send(NewChild(self.pid, tx, deadline)); }
343 return match rx.recv() {
345 Err(..) => Err(timeout("wait timed out")),
348 // Register a new SIGCHLD handler, returning the reading half of the
349 // self-pipe plus the old handler registered (return value of sigaction).
351 // Be sure to set up the self-pipe first because as soon as we register a
352 // handler we're going to start receiving signals.
353 fn register_sigchld() -> (libc::c_int, c::sigaction) {
355 let mut pipes = [0; 2];
356 assert_eq!(libc::pipe(pipes.as_mut_ptr()), 0);
357 set_nonblocking(pipes[0], true).ok().unwrap();
358 set_nonblocking(pipes[1], true).ok().unwrap();
361 let mut old: c::sigaction = mem::zeroed();
362 let mut new: c::sigaction = mem::zeroed();
363 new.sa_handler = sigchld_handler;
364 new.sa_flags = c::SA_NOCLDSTOP;
365 assert_eq!(c::sigaction(c::SIGCHLD, &new, &mut old), 0);
370 // Helper thread for processing SIGCHLD messages
371 fn waitpid_helper(input: libc::c_int,
372 messages: Receiver<Req>,
373 (read_fd, old): (libc::c_int, c::sigaction)) {
374 set_nonblocking(input, true).ok().unwrap();
375 let mut set: c::fd_set = unsafe { mem::zeroed() };
376 let mut tv: libc::timeval;
377 let mut active = Vec::<(libc::pid_t, Sender<ProcessExit>, u64)>::new();
378 let max = cmp::max(input, read_fd) + 1;
381 // Figure out the timeout of our syscall-to-happen. If we're waiting
382 // for some processes, then they'll have a timeout, otherwise we
383 // wait indefinitely for a message to arrive.
385 // FIXME: sure would be nice to not have to scan the entire array
386 let min = active.iter().map(|a| a.2).enumerate().min_by(|p| {
389 let (p, idx) = match min {
390 Some((idx, deadline)) => {
391 let now = sys::timer::now();
392 let ms = if now < deadline {deadline - now} else {0};
393 tv = ms_to_timeval(ms);
394 (&mut tv as *mut _, idx)
396 None => (ptr::null_mut(), -1),
399 // Wait for something to happen
400 c::fd_set(&mut set, input);
401 c::fd_set(&mut set, read_fd);
402 match unsafe { c::select(max, &mut set, ptr::null_mut(),
403 ptr::null_mut(), p) } {
404 // interrupted, retry
405 -1 if os::errno() == libc::EINTR as i32 => continue,
407 // We read something, break out and process
410 // Timeout, the pending request is removed
412 drop(active.remove(idx));
416 n => panic!("error in select {:?} ({:?})", os::errno(), n),
419 // Process any pending messages
422 match messages.try_recv() {
423 Ok(NewChild(pid, tx, deadline)) => {
424 active.push((pid, tx, deadline));
426 Err(TryRecvError::Disconnected) => {
427 assert!(active.len() == 0);
430 Err(TryRecvError::Empty) => break,
435 // If a child exited (somehow received SIGCHLD), then poll all
436 // children to see if any of them exited.
438 // We also attempt to be responsible netizens when dealing with
439 // SIGCHLD by invoking any previous SIGCHLD handler instead of just
440 // ignoring any previous SIGCHLD handler. Note that we don't provide
441 // a 1:1 mapping of our handler invocations to the previous handler
442 // invocations because we drain the `read_fd` entirely. This is
443 // probably OK because the kernel is already allowed to coalesce
444 // simultaneous signals, we're just doing some extra coalescing.
446 // Another point of note is that this likely runs the signal handler
447 // on a different thread than the one that received the signal. I
448 // *think* this is ok at this time.
450 // The main reason for doing this is to allow stdtest to run native
451 // tests as well. Both libgreen and libnative are running around
452 // with process timeouts, but libgreen should get there first
453 // (currently libuv doesn't handle old signal handlers).
455 let i: uint = unsafe { mem::transmute(old.sa_handler) };
457 assert!(old.sa_flags & c::SA_SIGINFO == 0);
458 (old.sa_handler)(c::SIGCHLD);
461 // FIXME: sure would be nice to not have to scan the entire
463 active.retain(|&(pid, ref tx, _)| {
464 let pr = Process { pid: pid };
465 match pr.try_wait() {
466 Some(msg) => { tx.send(msg).unwrap(); false }
473 // Once this helper thread is done, we re-register the old sigchld
474 // handler and close our intermediate file descriptors.
476 assert_eq!(c::sigaction(c::SIGCHLD, &old, ptr::null_mut()), 0);
477 let _ = libc::close(read_fd);
478 let _ = libc::close(WRITE_FD);
483 // Drain all pending data from the file descriptor, returning if any data
484 // could be drained. This requires that the file descriptor is in
486 fn drain(fd: libc::c_int) -> bool {
489 let mut buf = [0u8; 1];
491 libc::read(fd, buf.as_mut_ptr() as *mut libc::c_void,
492 buf.len() as libc::size_t)
494 n if n > 0 => { ret = true; }
496 -1 if wouldblock() => return ret,
497 n => panic!("bad read {:?} ({:?})", os::last_os_error(), n),
502 // Signal handler for SIGCHLD signals, must be async-signal-safe!
504 // This function will write to the writing half of the "self pipe" to wake
505 // up the helper thread if it's waiting. Note that this write must be
506 // nonblocking because if it blocks and the reader is the thread we
507 // interrupted, then we'll deadlock.
509 // When writing, if the write returns EWOULDBLOCK then we choose to ignore
510 // it. At that point we're guaranteed that there's something in the pipe
511 // which will wake up the other end at some point, so we just allow this
512 // signal to be coalesced with the pending signals on the pipe.
513 extern fn sigchld_handler(_signum: libc::c_int) {
516 libc::write(WRITE_FD, &msg as *const _ as *const libc::c_void, 1)
519 -1 if wouldblock() => {} // see above comments
520 n => panic!("bad error on write fd: {:?} {:?}", n, os::errno()),
525 pub fn try_wait(&self) -> Option<ProcessExit> {
526 let mut status = 0 as c_int;
527 match retry(|| unsafe {
528 c::waitpid(self.pid, &mut status, c::WNOHANG)
530 n if n == self.pid => Some(translate_status(status)),
532 n => panic!("unknown waitpid error `{:?}`: {:?}", n,
533 super::last_error()),
538 fn with_argv<T,F>(prog: &CString, args: &[CString],
541 where F : FnOnce(*const *const libc::c_char) -> T
543 let mut ptrs: Vec<*const libc::c_char> = Vec::with_capacity(args.len()+1);
545 // Convert the CStrings into an array of pointers. Note: the
546 // lifetime of the various CStrings involved is guaranteed to be
547 // larger than the lifetime of our invocation of cb, but this is
548 // technically unsafe as the callback could leak these pointers
550 ptrs.push(prog.as_ptr());
551 ptrs.extend(args.iter().map(|tmp| tmp.as_ptr()));
553 // Add a terminating null pointer (required by libc).
554 ptrs.push(ptr::null());
559 fn with_envp<K,V,T,F>(env: Option<&HashMap<K, V>>,
562 where F : FnOnce(*const c_void) -> T,
563 K : BytesContainer + Eq + Hash<Hasher>,
566 // On posixy systems we can pass a char** for envp, which is a
567 // null-terminated array of "k=v\0" strings. Since we must create
568 // these strings locally, yet expose a raw pointer to them, we
569 // create a temporary vector to own the CStrings that outlives the
573 let mut tmps = Vec::with_capacity(env.len());
576 let mut kv = Vec::new();
577 kv.push_all(pair.0.container_as_bytes());
579 kv.push_all(pair.1.container_as_bytes());
580 kv.push(0); // terminating null
584 // As with `with_argv`, this is unsafe, since cb could leak the pointers.
585 let mut ptrs: Vec<*const libc::c_char> =
587 .map(|tmp| tmp.as_ptr() as *const libc::c_char)
589 ptrs.push(ptr::null());
591 cb(ptrs.as_ptr() as *const c_void)
597 fn translate_status(status: c_int) -> ProcessExit {
598 #![allow(non_snake_case)]
599 #[cfg(any(target_os = "linux", target_os = "android"))]
601 pub fn WIFEXITED(status: i32) -> bool { (status & 0xff) == 0 }
602 pub fn WEXITSTATUS(status: i32) -> i32 { (status >> 8) & 0xff }
603 pub fn WTERMSIG(status: i32) -> i32 { status & 0x7f }
606 #[cfg(any(target_os = "macos",
608 target_os = "freebsd",
609 target_os = "dragonfly",
610 target_os = "openbsd"))]
612 pub fn WIFEXITED(status: i32) -> bool { (status & 0x7f) == 0 }
613 pub fn WEXITSTATUS(status: i32) -> i32 { status >> 8 }
614 pub fn WTERMSIG(status: i32) -> i32 { status & 0o177 }
617 if imp::WIFEXITED(status) {
618 ExitStatus(imp::WEXITSTATUS(status) as int)
620 ExitSignal(imp::WTERMSIG(status) as int)