abort() when uv_signal_start() after fork()

[ibc]

libuv master. This crash happens in OSX (in Linux it gets frozen). My code does the following:

• 1. Run the following code:

uv_loop_t loop;
int err;

err = uv_loop_init(&loop);
if (err)
    exit(1);

err = uv_run(&loop, UV_RUN_NOWAIT);
if (err)
    exit(1);

err = uv_loop_close(&loop);
if (err)
    exit(1);

• 2. The typical double fork() to become a daemon.
• 3. Init a new UV loop, init a uv_signal handler and call uv_signal_start().

It crashes as shown below. It works perfectly if I don't fork.

lldb output in OSX:

(lldb) bt all
* thread #1: tid = 0x0000, 0x00007fff894fc866 libsystem_kernel.dylib`__pthread_kill + 10, stop reason = signal SIGSTOP
  * frame #0: 0x00007fff894fc866 libsystem_kernel.dylib`__pthread_kill + 10
    frame #1: 0x00007fff8687535c libsystem_pthread.dylib`pthread_kill + 92
    frame #2: 0x00007fff854a3b1a libsystem_c.dylib`abort + 125
    frame #3: 0x0000000101b52716 libuv.11.dylib`uv_signal_start [inlined] uv__signal_block_and_lock(saved_sigmask=0xffffffff00000000) + 182 at signal.c:106
    frame #4: 0x0000000101b526b8 libuv.11.dylib`uv_signal_start(handle=0x00007ffa43404e10, signal_cb=0x0000000101ac01b0, signum=30) + 88 at signal.c:315

The UV code calling abort() is in src/unix/signal.c line 106:

static void uv__signal_block_and_lock(sigset_t* saved_sigmask) {
  sigset_t new_mask;

  if (sigfillset(&new_mask))
    abort();

  if (pthread_sigmask(SIG_SETMASK, &new_mask, saved_sigmask))
    abort();

  if (uv__signal_lock())
    abort();           // <----- HERE
}

So uv__signal_lock() returns != 0:

static int uv__signal_lock(void) {
  int r;
  char data;

  do {
    r = read(uv__signal_lock_pipefd[0], &data, sizeof data);
  } while (r < 0 && errno == EINTR);

  return (r < 0) ? -1 : 0;
}

Can not I run (and close) a UV loop before forking and then run other after forking??

[saghul]

There is some global state associated with signal handling, so I wouldn't do that. We don't currently have a good story for forking scenarios, I'm afraid.

[ibc]

Opps, and let me know: such a "global state" (which I assume is the static Unix sockets for firing signal events to each loop) is created upon initiation of the first UV loop, right?

[saghul]

Exactly: https://github.com/joyent/libuv/blob/master/src/unix/loop.c#L105

[ibc]

So no imminent fix, right?

[saghul]

Nope, I'm afraid.

[ibc]

Afraid of what?

[saghul]

It's not happening any time soon, I mean.

[mmalecki]

Could we possibly have a uv_fork function which would properly set up uv's global state after forking?

[saghul]

Ideally it would be nice to make it automagic. Maybe we can use pthread_atfork for resetting global state in the child.

I don't think we can support all cases, so I'd say that creating an eventloop in the parent and then trying to use in the child would be unsupported. This should make things easier, though I haven't really sat down to think this all the way through.

You are welcome to work on this and come up with a design which we can discuss :-)

[ibc]

2014-08-10 21:28 GMT+02:00 Saúl Ibarra Corretgé [email protected]:

I don't think we can support all cases, so I'd say that creating an
eventloop in the parent and then trying to use in the child would be
unsupported. This should make things easier, though I haven't really sat
down to think this all the way through.

BTW my use-case was not su hard:

• Create UV loop in the parent.
• Run it and properly stop and close it.
• Fork.
• Create another loop.

I was never trying to reuse the same loop.

Iñaki Baz Castillo
[email protected]

[saghul]

@ibc yeah, your use case would be supported. But we would need to reinitialize some stuff like the signal handlers thing and probably something in the threadpool as well.