unix_events.py 33.7 KB
Newer Older
1
"""Selector event loop for Unix with signal handling."""
2 3 4 5 6 7 8 9

import errno
import os
import signal
import socket
import stat
import subprocess
import sys
10
import threading
11
import warnings
12 13


14
from . import base_events
15
from . import base_subprocess
16
from . import compat
17
from . import constants
18
from . import coroutines
19
from . import events
20
from . import futures
21
from . import selector_events
22
from . import selectors
23
from . import transports
24
from .coroutines import coroutine
25
from .log import logger
26 27


28
__all__ = ['SelectorEventLoop',
29 30 31
           'AbstractChildWatcher', 'SafeChildWatcher',
           'FastChildWatcher', 'DefaultEventLoopPolicy',
           ]
32 33 34 35 36

if sys.platform == 'win32':  # pragma: no cover
    raise ImportError('Signals are not really supported on Windows')


37 38 39 40 41
def _sighandler_noop(signum, frame):
    """Dummy signal handler."""
    pass


42
class _UnixSelectorEventLoop(selector_events.BaseSelectorEventLoop):
43
    """Unix event loop.
44

45
    Adds signal handling and UNIX Domain Socket support to SelectorEventLoop.
46 47 48 49 50 51 52 53 54
    """

    def __init__(self, selector=None):
        super().__init__(selector)
        self._signal_handlers = {}

    def _socketpair(self):
        return socket.socketpair()

55
    def close(self):
56
        super().close()
57 58 59
        for sig in list(self._signal_handlers):
            self.remove_signal_handler(sig)

60 61 62 63 64 65 66
    def _process_self_data(self, data):
        for signum in data:
            if not signum:
                # ignore null bytes written by _write_to_self()
                continue
            self._handle_signal(signum)

67 68 69 70 71 72
    def add_signal_handler(self, sig, callback, *args):
        """Add a handler for a signal.  UNIX only.

        Raise ValueError if the signal number is invalid or uncatchable.
        Raise RuntimeError if there is a problem setting up the handler.
        """
73 74
        if (coroutines.iscoroutine(callback)
        or coroutines.iscoroutinefunction(callback)):
75 76
            raise TypeError("coroutines cannot be used "
                            "with add_signal_handler()")
77
        self._check_signal(sig)
78
        self._check_closed()
79 80 81 82 83 84
        try:
            # set_wakeup_fd() raises ValueError if this is not the
            # main thread.  By calling it early we ensure that an
            # event loop running in another thread cannot add a signal
            # handler.
            signal.set_wakeup_fd(self._csock.fileno())
85
        except (ValueError, OSError) as exc:
86 87
            raise RuntimeError(str(exc))

88
        handle = events.Handle(callback, args, self)
89 90 91
        self._signal_handlers[sig] = handle

        try:
92 93 94 95 96
            # Register a dummy signal handler to ask Python to write the signal
            # number in the wakup file descriptor. _process_self_data() will
            # read signal numbers from this file descriptor to handle signals.
            signal.signal(sig, _sighandler_noop)

97 98
            # Set SA_RESTART to limit EINTR occurrences.
            signal.siginterrupt(sig, False)
99 100 101 102 103
        except OSError as exc:
            del self._signal_handlers[sig]
            if not self._signal_handlers:
                try:
                    signal.set_wakeup_fd(-1)
104
                except (ValueError, OSError) as nexc:
105
                    logger.info('set_wakeup_fd(-1) failed: %s', nexc)
106 107 108 109 110 111

            if exc.errno == errno.EINVAL:
                raise RuntimeError('sig {} cannot be caught'.format(sig))
            else:
                raise

112
    def _handle_signal(self, sig):
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148
        """Internal helper that is the actual signal handler."""
        handle = self._signal_handlers.get(sig)
        if handle is None:
            return  # Assume it's some race condition.
        if handle._cancelled:
            self.remove_signal_handler(sig)  # Remove it properly.
        else:
            self._add_callback_signalsafe(handle)

    def remove_signal_handler(self, sig):
        """Remove a handler for a signal.  UNIX only.

        Return True if a signal handler was removed, False if not.
        """
        self._check_signal(sig)
        try:
            del self._signal_handlers[sig]
        except KeyError:
            return False

        if sig == signal.SIGINT:
            handler = signal.default_int_handler
        else:
            handler = signal.SIG_DFL

        try:
            signal.signal(sig, handler)
        except OSError as exc:
            if exc.errno == errno.EINVAL:
                raise RuntimeError('sig {} cannot be caught'.format(sig))
            else:
                raise

        if not self._signal_handlers:
            try:
                signal.set_wakeup_fd(-1)
149
            except (ValueError, OSError) as exc:
150
                logger.info('set_wakeup_fd(-1) failed: %s', exc)
151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174

        return True

    def _check_signal(self, sig):
        """Internal helper to validate a signal.

        Raise ValueError if the signal number is invalid or uncatchable.
        Raise RuntimeError if there is a problem setting up the handler.
        """
        if not isinstance(sig, int):
            raise TypeError('sig must be an int, not {!r}'.format(sig))

        if not (1 <= sig < signal.NSIG):
            raise ValueError(
                'sig {} out of range(1, {})'.format(sig, signal.NSIG))

    def _make_read_pipe_transport(self, pipe, protocol, waiter=None,
                                  extra=None):
        return _UnixReadPipeTransport(self, pipe, protocol, waiter, extra)

    def _make_write_pipe_transport(self, pipe, protocol, waiter=None,
                                   extra=None):
        return _UnixWritePipeTransport(self, pipe, protocol, waiter, extra)

175
    @coroutine
176 177 178
    def _make_subprocess_transport(self, protocol, args, shell,
                                   stdin, stdout, stderr, bufsize,
                                   extra=None, **kwargs):
179
        with events.get_child_watcher() as watcher:
180
            waiter = futures.Future(loop=self)
181 182
            transp = _UnixSubprocessTransport(self, protocol, args, shell,
                                              stdin, stdout, stderr, bufsize,
183 184 185 186 187
                                              waiter=waiter, extra=extra,
                                              **kwargs)

            watcher.add_child_handler(transp.get_pid(),
                                      self._child_watcher_callback, transp)
188
            try:
189
                yield from waiter
190 191 192 193 194 195 196 197 198
            except Exception as exc:
                # Workaround CPython bug #23353: using yield/yield-from in an
                # except block of a generator doesn't clear properly
                # sys.exc_info()
                err = exc
            else:
                err = None

            if err is not None:
199
                transp.close()
200
                yield from transp._wait()
201
                raise err
202

203 204
        return transp

205 206
    def _child_watcher_callback(self, pid, returncode, transp):
        self.call_soon_threadsafe(transp._process_exited, returncode)
207

208
    @coroutine
209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
    def create_unix_connection(self, protocol_factory, path, *,
                               ssl=None, sock=None,
                               server_hostname=None):
        assert server_hostname is None or isinstance(server_hostname, str)
        if ssl:
            if server_hostname is None:
                raise ValueError(
                    'you have to pass server_hostname when using ssl')
        else:
            if server_hostname is not None:
                raise ValueError('server_hostname is only meaningful with ssl')

        if path is not None:
            if sock is not None:
                raise ValueError(
                    'path and sock can not be specified at the same time')

226
            sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM, 0)
227 228 229
            try:
                sock.setblocking(False)
                yield from self.sock_connect(sock, path)
230 231
            except:
                sock.close()
232 233 234 235 236 237 238 239 240 241 242
                raise

        else:
            if sock is None:
                raise ValueError('no path and sock were specified')
            sock.setblocking(False)

        transport, protocol = yield from self._create_connection_transport(
            sock, protocol_factory, ssl, server_hostname)
        return transport, protocol

243
    @coroutine
244 245 246 247 248 249
    def create_unix_server(self, protocol_factory, path=None, *,
                           sock=None, backlog=100, ssl=None):
        if isinstance(ssl, bool):
            raise TypeError('ssl argument must be an SSLContext or None')

        if path is not None:
250 251 252 253
            if sock is not None:
                raise ValueError(
                    'path and sock can not be specified at the same time')

254 255 256 257 258
            sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)

            try:
                sock.bind(path)
            except OSError as exc:
259
                sock.close()
260 261 262 263 264 265 266
                if exc.errno == errno.EADDRINUSE:
                    # Let's improve the error message by adding
                    # with what exact address it occurs.
                    msg = 'Address {!r} is already in use'.format(path)
                    raise OSError(errno.EADDRINUSE, msg) from None
                else:
                    raise
267 268 269
            except:
                sock.close()
                raise
270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
        else:
            if sock is None:
                raise ValueError(
                    'path was not specified, and no sock specified')

            if sock.family != socket.AF_UNIX:
                raise ValueError(
                    'A UNIX Domain Socket was expected, got {!r}'.format(sock))

        server = base_events.Server(self, [sock])
        sock.listen(backlog)
        sock.setblocking(False)
        self._start_serving(protocol_factory, sock, ssl, server)
        return server

285

286 287 288 289
if hasattr(os, 'set_blocking'):
    def _set_nonblocking(fd):
        os.set_blocking(fd, False)
else:
290 291
    import fcntl

292 293 294 295
    def _set_nonblocking(fd):
        flags = fcntl.fcntl(fd, fcntl.F_GETFL)
        flags = flags | os.O_NONBLOCK
        fcntl.fcntl(fd, fcntl.F_SETFL, flags)
296 297 298 299


class _UnixReadPipeTransport(transports.ReadTransport):

300
    max_size = 256 * 1024  # max bytes we read in one event loop iteration
301 302 303 304 305 306 307

    def __init__(self, loop, pipe, protocol, waiter=None, extra=None):
        super().__init__(extra)
        self._extra['pipe'] = pipe
        self._loop = loop
        self._pipe = pipe
        self._fileno = pipe.fileno()
308
        mode = os.fstat(self._fileno).st_mode
309 310 311
        if not (stat.S_ISFIFO(mode) or
                stat.S_ISSOCK(mode) or
                stat.S_ISCHR(mode)):
312
            raise ValueError("Pipe transport is for pipes/sockets only.")
313 314 315 316
        _set_nonblocking(self._fileno)
        self._protocol = protocol
        self._closing = False
        self._loop.call_soon(self._protocol.connection_made, self)
317 318 319
        # only start reading when connection_made() has been called
        self._loop.call_soon(self._loop.add_reader,
                             self._fileno, self._read_ready)
320
        if waiter is not None:
321
            # only wake up the waiter when connection_made() has been called
322 323
            self._loop.call_soon(futures._set_result_unless_cancelled,
                                 waiter, None)
324

325
    def __repr__(self):
326 327 328 329 330 331
        info = [self.__class__.__name__]
        if self._pipe is None:
            info.append('closed')
        elif self._closing:
            info.append('closing')
        info.append('fd=%s' % self._fileno)
332 333 334 335 336 337 338 339 340 341 342 343
        if self._pipe is not None:
            polling = selector_events._test_selector_event(
                          self._loop._selector,
                          self._fileno, selectors.EVENT_READ)
            if polling:
                info.append('polling')
            else:
                info.append('idle')
        else:
            info.append('closed')
        return '<%s>' % ' '.join(info)

344 345 346 347 348 349
    def _read_ready(self):
        try:
            data = os.read(self._fileno, self.max_size)
        except (BlockingIOError, InterruptedError):
            pass
        except OSError as exc:
350
            self._fatal_error(exc, 'Fatal read error on pipe transport')
351 352 353 354
        else:
            if data:
                self._protocol.data_received(data)
            else:
355 356
                if self._loop.get_debug():
                    logger.info("%r was closed by peer", self)
357 358 359 360 361
                self._closing = True
                self._loop.remove_reader(self._fileno)
                self._loop.call_soon(self._protocol.eof_received)
                self._loop.call_soon(self._call_connection_lost, None)

362
    def pause_reading(self):
363 364
        self._loop.remove_reader(self._fileno)

365
    def resume_reading(self):
366 367
        self._loop.add_reader(self._fileno, self._read_ready)

368 369 370
    def is_closing(self):
        return self._closing

371 372 373 374
    def close(self):
        if not self._closing:
            self._close(None)

375 376 377
    # On Python 3.3 and older, objects with a destructor part of a reference
    # cycle are never destroyed. It's not more the case on Python 3.4 thanks
    # to the PEP 442.
378
    if compat.PY34:
379 380 381 382 383
        def __del__(self):
            if self._pipe is not None:
                warnings.warn("unclosed transport %r" % self, ResourceWarning)
                self._pipe.close()

384
    def _fatal_error(self, exc, message='Fatal error on pipe transport'):
385
        # should be called by exception handler only
386 387 388 389
        if (isinstance(exc, OSError) and exc.errno == errno.EIO):
            if self._loop.get_debug():
                logger.debug("%r: %s", self, message, exc_info=True)
        else:
390
            self._loop.call_exception_handler({
391
                'message': message,
392 393 394 395
                'exception': exc,
                'transport': self,
                'protocol': self._protocol,
            })
396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412
        self._close(exc)

    def _close(self, exc):
        self._closing = True
        self._loop.remove_reader(self._fileno)
        self._loop.call_soon(self._call_connection_lost, exc)

    def _call_connection_lost(self, exc):
        try:
            self._protocol.connection_lost(exc)
        finally:
            self._pipe.close()
            self._pipe = None
            self._protocol = None
            self._loop = None


413
class _UnixWritePipeTransport(transports._FlowControlMixin,
414
                              transports.WriteTransport):
415 416

    def __init__(self, loop, pipe, protocol, waiter=None, extra=None):
417
        super().__init__(extra, loop)
418 419 420
        self._extra['pipe'] = pipe
        self._pipe = pipe
        self._fileno = pipe.fileno()
421 422
        mode = os.fstat(self._fileno).st_mode
        is_socket = stat.S_ISSOCK(mode)
423 424 425 426 427
        if not (is_socket or
                stat.S_ISFIFO(mode) or
                stat.S_ISCHR(mode)):
            raise ValueError("Pipe transport is only for "
                             "pipes, sockets and character devices")
428 429 430 431 432
        _set_nonblocking(self._fileno)
        self._protocol = protocol
        self._buffer = []
        self._conn_lost = 0
        self._closing = False  # Set when close() or write_eof() called.
433

434 435 436 437 438
        self._loop.call_soon(self._protocol.connection_made, self)

        # On AIX, the reader trick (to be notified when the read end of the
        # socket is closed) only works for sockets. On other platforms it
        # works for pipes and sockets. (Exception: OS X 10.4?  Issue #19294.)
439
        if is_socket or not sys.platform.startswith("aix"):
440 441 442
            # only start reading when connection_made() has been called
            self._loop.call_soon(self._loop.add_reader,
                                 self._fileno, self._read_ready)
443 444

        if waiter is not None:
445
            # only wake up the waiter when connection_made() has been called
446 447
            self._loop.call_soon(futures._set_result_unless_cancelled,
                                 waiter, None)
448

449
    def __repr__(self):
450 451 452 453 454 455
        info = [self.__class__.__name__]
        if self._pipe is None:
            info.append('closed')
        elif self._closing:
            info.append('closing')
        info.append('fd=%s' % self._fileno)
456 457 458 459 460 461 462 463 464 465 466 467 468 469 470
        if self._pipe is not None:
            polling = selector_events._test_selector_event(
                          self._loop._selector,
                          self._fileno, selectors.EVENT_WRITE)
            if polling:
                info.append('polling')
            else:
                info.append('idle')

            bufsize = self.get_write_buffer_size()
            info.append('bufsize=%s' % bufsize)
        else:
            info.append('closed')
        return '<%s>' % ' '.join(info)

471 472 473
    def get_write_buffer_size(self):
        return sum(len(data) for data in self._buffer)

474
    def _read_ready(self):
475
        # Pipe was closed by peer.
476 477
        if self._loop.get_debug():
            logger.info("%r was closed by peer", self)
478 479 480 481
        if self._buffer:
            self._close(BrokenPipeError())
        else:
            self._close()
482 483

    def write(self, data):
484 485 486
        assert isinstance(data, (bytes, bytearray, memoryview)), repr(data)
        if isinstance(data, bytearray):
            data = memoryview(data)
487 488 489 490 491
        if not data:
            return

        if self._conn_lost or self._closing:
            if self._conn_lost >= constants.LOG_THRESHOLD_FOR_CONNLOST_WRITES:
492 493
                logger.warning('pipe closed by peer or '
                               'os.write(pipe, data) raised exception.')
494 495 496 497 498 499 500 501 502 503 504
            self._conn_lost += 1
            return

        if not self._buffer:
            # Attempt to send it right away first.
            try:
                n = os.write(self._fileno, data)
            except (BlockingIOError, InterruptedError):
                n = 0
            except Exception as exc:
                self._conn_lost += 1
505
                self._fatal_error(exc, 'Fatal write error on pipe transport')
506 507 508 509 510 511 512 513
                return
            if n == len(data):
                return
            elif n > 0:
                data = data[n:]
            self._loop.add_writer(self._fileno, self._write_ready)

        self._buffer.append(data)
514
        self._maybe_pause_protocol()
515 516 517 518 519 520 521 522 523 524 525 526 527 528 529

    def _write_ready(self):
        data = b''.join(self._buffer)
        assert data, 'Data should not be empty'

        self._buffer.clear()
        try:
            n = os.write(self._fileno, data)
        except (BlockingIOError, InterruptedError):
            self._buffer.append(data)
        except Exception as exc:
            self._conn_lost += 1
            # Remove writer here, _fatal_error() doesn't it
            # because _buffer is empty.
            self._loop.remove_writer(self._fileno)
530
            self._fatal_error(exc, 'Fatal write error on pipe transport')
531 532 533
        else:
            if n == len(data):
                self._loop.remove_writer(self._fileno)
534 535
                self._maybe_resume_protocol()  # May append to buffer.
                if not self._buffer and self._closing:
536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555
                    self._loop.remove_reader(self._fileno)
                    self._call_connection_lost(None)
                return
            elif n > 0:
                data = data[n:]

            self._buffer.append(data)  # Try again later.

    def can_write_eof(self):
        return True

    def write_eof(self):
        if self._closing:
            return
        assert self._pipe
        self._closing = True
        if not self._buffer:
            self._loop.remove_reader(self._fileno)
            self._loop.call_soon(self._call_connection_lost, None)

556 557 558
    def is_closing(self):
        return self._closing

559
    def close(self):
560
        if self._pipe is not None and not self._closing:
561 562 563
            # write_eof is all what we needed to close the write pipe
            self.write_eof()

564 565 566
    # On Python 3.3 and older, objects with a destructor part of a reference
    # cycle are never destroyed. It's not more the case on Python 3.4 thanks
    # to the PEP 442.
567
    if compat.PY34:
568 569 570 571 572
        def __del__(self):
            if self._pipe is not None:
                warnings.warn("unclosed transport %r" % self, ResourceWarning)
                self._pipe.close()

573 574 575
    def abort(self):
        self._close(None)

576
    def _fatal_error(self, exc, message='Fatal error on pipe transport'):
577
        # should be called by exception handler only
578 579 580 581
        if isinstance(exc, (BrokenPipeError, ConnectionResetError)):
            if self._loop.get_debug():
                logger.debug("%r: %s", self, message, exc_info=True)
        else:
582
            self._loop.call_exception_handler({
583
                'message': message,
584 585 586 587
                'exception': exc,
                'transport': self,
                'protocol': self._protocol,
            })
588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
        self._close(exc)

    def _close(self, exc=None):
        self._closing = True
        if self._buffer:
            self._loop.remove_writer(self._fileno)
        self._buffer.clear()
        self._loop.remove_reader(self._fileno)
        self._loop.call_soon(self._call_connection_lost, exc)

    def _call_connection_lost(self, exc):
        try:
            self._protocol.connection_lost(exc)
        finally:
            self._pipe.close()
            self._pipe = None
            self._protocol = None
            self._loop = None


608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623
if hasattr(os, 'set_inheritable'):
    # Python 3.4 and newer
    _set_inheritable = os.set_inheritable
else:
    import fcntl

    def _set_inheritable(fd, inheritable):
        cloexec_flag = getattr(fcntl, 'FD_CLOEXEC', 1)

        old = fcntl.fcntl(fd, fcntl.F_GETFD)
        if not inheritable:
            fcntl.fcntl(fd, fcntl.F_SETFD, old | cloexec_flag)
        else:
            fcntl.fcntl(fd, fcntl.F_SETFD, old & ~cloexec_flag)


624
class _UnixSubprocessTransport(base_subprocess.BaseSubprocessTransport):
625

626
    def _start(self, args, shell, stdin, stdout, stderr, bufsize, **kwargs):
627
        stdin_w = None
628
        if stdin == subprocess.PIPE:
629 630 631 632 633 634
            # Use a socket pair for stdin, since not all platforms
            # support selecting read events on the write end of a
            # socket (which we use in order to detect closing of the
            # other end).  Notably this is needed on AIX, and works
            # just fine on other platforms.
            stdin, stdin_w = self._loop._socketpair()
635 636 637 638 639 640

            # Mark the write end of the stdin pipe as non-inheritable,
            # needed by close_fds=False on Python 3.3 and older
            # (Python 3.4 implements the PEP 446, socketpair returns
            # non-inheritable sockets)
            _set_inheritable(stdin_w.fileno(), False)
641 642 643
        self._proc = subprocess.Popen(
            args, shell=shell, stdin=stdin, stdout=stdout, stderr=stderr,
            universal_newlines=False, bufsize=bufsize, **kwargs)
644 645
        if stdin_w is not None:
            stdin.close()
646
            self._proc.stdin = open(stdin_w.detach(), 'wb', buffering=bufsize)
647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678


class AbstractChildWatcher:
    """Abstract base class for monitoring child processes.

    Objects derived from this class monitor a collection of subprocesses and
    report their termination or interruption by a signal.

    New callbacks are registered with .add_child_handler(). Starting a new
    process must be done within a 'with' block to allow the watcher to suspend
    its activity until the new process if fully registered (this is needed to
    prevent a race condition in some implementations).

    Example:
        with watcher:
            proc = subprocess.Popen("sleep 1")
            watcher.add_child_handler(proc.pid, callback)

    Notes:
        Implementations of this class must be thread-safe.

        Since child watcher objects may catch the SIGCHLD signal and call
        waitpid(-1), there should be only one active object per process.
    """

    def add_child_handler(self, pid, callback, *args):
        """Register a new child handler.

        Arrange for callback(pid, returncode, *args) to be called when
        process 'pid' terminates. Specifying another callback for the same
        process replaces the previous handler.

679
        Note: callback() must be thread-safe.
680 681 682 683 684 685 686 687 688 689 690
        """
        raise NotImplementedError()

    def remove_child_handler(self, pid):
        """Removes the handler for process 'pid'.

        The function returns True if the handler was successfully removed,
        False if there was nothing to remove."""

        raise NotImplementedError()

691 692
    def attach_loop(self, loop):
        """Attach the watcher to an event loop.
693

694 695 696 697
        If the watcher was previously attached to an event loop, then it is
        first detached before attaching to the new loop.

        Note: loop may be None.
698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720
        """
        raise NotImplementedError()

    def close(self):
        """Close the watcher.

        This must be called to make sure that any underlying resource is freed.
        """
        raise NotImplementedError()

    def __enter__(self):
        """Enter the watcher's context and allow starting new processes

        This function must return self"""
        raise NotImplementedError()

    def __exit__(self, a, b, c):
        """Exit the watcher's context"""
        raise NotImplementedError()


class BaseChildWatcher(AbstractChildWatcher):

721
    def __init__(self):
722 723 724
        self._loop = None

    def close(self):
725
        self.attach_loop(None)
726 727 728 729 730 731 732

    def _do_waitpid(self, expected_pid):
        raise NotImplementedError()

    def _do_waitpid_all(self):
        raise NotImplementedError()

733
    def attach_loop(self, loop):
734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749
        assert loop is None or isinstance(loop, events.AbstractEventLoop)

        if self._loop is not None:
            self._loop.remove_signal_handler(signal.SIGCHLD)

        self._loop = loop
        if loop is not None:
            loop.add_signal_handler(signal.SIGCHLD, self._sig_chld)

            # Prevent a race condition in case a child terminated
            # during the switch.
            self._do_waitpid_all()

    def _sig_chld(self):
        try:
            self._do_waitpid_all()
750 751 752 753 754 755 756 757
        except Exception as exc:
            # self._loop should always be available here
            # as '_sig_chld' is added as a signal handler
            # in 'attach_loop'
            self._loop.call_exception_handler({
                'message': 'Unknown exception in SIGCHLD handler',
                'exception': exc,
            })
758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783

    def _compute_returncode(self, status):
        if os.WIFSIGNALED(status):
            # The child process died because of a signal.
            return -os.WTERMSIG(status)
        elif os.WIFEXITED(status):
            # The child process exited (e.g sys.exit()).
            return os.WEXITSTATUS(status)
        else:
            # The child exited, but we don't understand its status.
            # This shouldn't happen, but if it does, let's just
            # return that status; perhaps that helps debug it.
            return status


class SafeChildWatcher(BaseChildWatcher):
    """'Safe' child watcher implementation.

    This implementation avoids disrupting other code spawning processes by
    polling explicitly each process in the SIGCHLD handler instead of calling
    os.waitpid(-1).

    This is a safe solution but it has a significant overhead when handling a
    big number of children (O(n) each time SIGCHLD is raised)
    """

784 785 786 787 788 789 790 791
    def __init__(self):
        super().__init__()
        self._callbacks = {}

    def close(self):
        self._callbacks.clear()
        super().close()

792 793 794 795 796 797 798
    def __enter__(self):
        return self

    def __exit__(self, a, b, c):
        pass

    def add_child_handler(self, pid, callback, *args):
799
        self._callbacks[pid] = (callback, args)
800 801 802 803

        # Prevent a race condition in case the child is already terminated.
        self._do_waitpid(pid)

804 805 806 807 808 809 810
    def remove_child_handler(self, pid):
        try:
            del self._callbacks[pid]
            return True
        except KeyError:
            return False

811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834
    def _do_waitpid_all(self):

        for pid in list(self._callbacks):
            self._do_waitpid(pid)

    def _do_waitpid(self, expected_pid):
        assert expected_pid > 0

        try:
            pid, status = os.waitpid(expected_pid, os.WNOHANG)
        except ChildProcessError:
            # The child process is already reaped
            # (may happen if waitpid() is called elsewhere).
            pid = expected_pid
            returncode = 255
            logger.warning(
                "Unknown child process pid %d, will report returncode 255",
                pid)
        else:
            if pid == 0:
                # The child process is still alive.
                return

            returncode = self._compute_returncode(status)
835 836 837
            if self._loop.get_debug():
                logger.debug('process %s exited with returncode %s',
                             expected_pid, returncode)
838 839 840 841 842 843

        try:
            callback, args = self._callbacks.pop(pid)
        except KeyError:  # pragma: no cover
            # May happen if .remove_child_handler() is called
            # after os.waitpid() returns.
844 845 846
            if self._loop.get_debug():
                logger.warning("Child watcher got an unexpected pid: %r",
                               pid, exc_info=True)
847 848 849 850 851 852 853 854 855 856 857 858 859 860
        else:
            callback(pid, returncode, *args)


class FastChildWatcher(BaseChildWatcher):
    """'Fast' child watcher implementation.

    This implementation reaps every terminated processes by calling
    os.waitpid(-1) directly, possibly breaking other code spawning processes
    and waiting for their termination.

    There is no noticeable overhead when handling a big number of children
    (O(1) each time a child terminates).
    """
861 862 863
    def __init__(self):
        super().__init__()
        self._callbacks = {}
864 865 866 867 868
        self._lock = threading.Lock()
        self._zombies = {}
        self._forks = 0

    def close(self):
869
        self._callbacks.clear()
870
        self._zombies.clear()
871
        super().close()
872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894

    def __enter__(self):
        with self._lock:
            self._forks += 1

            return self

    def __exit__(self, a, b, c):
        with self._lock:
            self._forks -= 1

            if self._forks or not self._zombies:
                return

            collateral_victims = str(self._zombies)
            self._zombies.clear()

        logger.warning(
            "Caught subprocesses termination from unknown pids: %s",
            collateral_victims)

    def add_child_handler(self, pid, callback, *args):
        assert self._forks, "Must use the context manager"
895 896 897 898 899 900 901
        with self._lock:
            try:
                returncode = self._zombies.pop(pid)
            except KeyError:
                # The child is running.
                self._callbacks[pid] = callback, args
                return
902

903 904
        # The child is dead already. We can fire the callback.
        callback(pid, returncode, *args)
905

906 907 908 909 910 911 912
    def remove_child_handler(self, pid):
        try:
            del self._callbacks[pid]
            return True
        except KeyError:
            return False

913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928
    def _do_waitpid_all(self):
        # Because of signal coalescing, we must keep calling waitpid() as
        # long as we're able to reap a child.
        while True:
            try:
                pid, status = os.waitpid(-1, os.WNOHANG)
            except ChildProcessError:
                # No more child processes exist.
                return
            else:
                if pid == 0:
                    # A child process is still alive.
                    return

                returncode = self._compute_returncode(status)

929 930 931 932 933
            with self._lock:
                try:
                    callback, args = self._callbacks.pop(pid)
                except KeyError:
                    # unknown child
934 935 936
                    if self._forks:
                        # It may not be registered yet.
                        self._zombies[pid] = returncode
937 938 939 940
                        if self._loop.get_debug():
                            logger.debug('unknown process %s exited '
                                         'with returncode %s',
                                         pid, returncode)
941
                        continue
942
                    callback = None
943 944 945 946
                else:
                    if self._loop.get_debug():
                        logger.debug('process %s exited with returncode %s',
                                     pid, returncode)
947

948
            if callback is None:
949 950 951 952 953 954 955 956
                logger.warning(
                    "Caught subprocess termination from unknown pid: "
                    "%d -> %d", pid, returncode)
            else:
                callback(pid, returncode, *args)


class _UnixDefaultEventLoopPolicy(events.BaseDefaultEventLoopPolicy):
957
    """UNIX event loop policy with a watcher for child processes."""
958 959 960 961 962 963 964 965 966
    _loop_factory = _UnixSelectorEventLoop

    def __init__(self):
        super().__init__()
        self._watcher = None

    def _init_watcher(self):
        with events._lock:
            if self._watcher is None:  # pragma: no branch
967
                self._watcher = SafeChildWatcher()
968 969
                if isinstance(threading.current_thread(),
                              threading._MainThread):
970
                    self._watcher.attach_loop(self._local._loop)
971 972 973 974 975

    def set_event_loop(self, loop):
        """Set the event loop.

        As a side effect, if a child watcher was set before, then calling
976 977
        .set_event_loop() from the main thread will call .attach_loop(loop) on
        the child watcher.
978 979 980 981 982 983
        """

        super().set_event_loop(loop)

        if self._watcher is not None and \
            isinstance(threading.current_thread(), threading._MainThread):
984
            self._watcher.attach_loop(loop)
985 986

    def get_child_watcher(self):
987
        """Get the watcher for child processes.
988 989 990 991 992 993 994 995 996

        If not yet set, a SafeChildWatcher object is automatically created.
        """
        if self._watcher is None:
            self._init_watcher()

        return self._watcher

    def set_child_watcher(self, watcher):
997
        """Set the watcher for child processes."""
998 999 1000 1001 1002 1003 1004 1005 1006 1007

        assert watcher is None or isinstance(watcher, AbstractChildWatcher)

        if self._watcher is not None:
            self._watcher.close()

        self._watcher = watcher

SelectorEventLoop = _UnixSelectorEventLoop
DefaultEventLoopPolicy = _UnixDefaultEventLoopPolicy