def write(self, data):
"""
Writes the given data to the channel.
:param data: the data to be written to the channel.
:type data: string
:returns: An :class:`~kaa.InProgress` object which is finished when the
given data is fully written to the channel. The InProgress
is finished with the number of bytes sent in the last write
required to commit the given data to the channel. (This may
not be the actual number of bytes of the given data.)
If the channel closes unexpectedly before the data was
written, an IOError is thrown to the InProgress.
It is not required that the channel be open in order to write to it.
Written data is queued until the channel open and then flushed. As
writes are asynchronous, all written data is queued. It is the
caller's responsibility to ensure the internal write queue does not
exceed the desired size by waiting for past write() InProgress to
finish before writing more data.
If a write does not complete because the channel was closed
prematurely, an IOError is thrown to the InProgress.
"""
if not (self._mode & IO_WRITE):
raise IOError(9, 'Cannot write to a read-only channel')
if not self.writable:
raise IOError(9, 'Channel is not writable')
if self.write_queue_used + len(data) > self._queue_size:
raise ValueError('Data would exceed write queue limit')
if not isinstance(data, BYTES_TYPE):
raise ValueError('data must be bytes, not unicode')
ip = InProgress()
if data:
ip.signals['abort'].connect(self._abort_write_inprogress, data, ip)
self._write_queue.append((data, ip))
if self._channel and self._wmon and not self._wmon.active:
self._wmon.register(self.fileno, IO_WRITE)
else:
# We're writing the null string, nothing really to do. We're
# implicitly done.
ip.finish(0)
return ip
def __call__(self, *args, **kwargs):
in_progress = InProgress()
if CoreThreading.is_mainthread():
try:
result = super(MainThreadCallable, self).__call__(*args, **kwargs)
except BaseException, e:
# All exceptions, including SystemExit and KeyboardInterrupt,
# are caught and thrown to the InProgress, because it may be
# waiting in another thread. However SE and KI are reraised
# in here the main thread so they can be propagated back up
# the mainloop.
in_progress.throw(*sys.exc_info())
if isinstance(e, (KeyboardInterrupt, SystemExit)):
raise
else:
in_progress.finish(result)
return in_progress
def __call__(self, *args, **kwargs):
in_progress = InProgress()
if CoreThreading.is_mainthread():
try:
result = super(MainThreadCallable, self).__call__(*args, **kwargs)
except BaseException, e:
# All exceptions, including SystemExit and KeyboardInterrupt,
# are caught and thrown to the InProgress, because it may be
# waiting in another thread. However SE and KI are reraised
# in here the main thread so they can be propagated back up
# the mainloop.
in_progress.throw()
if isinstance(e, (KeyboardInterrupt, SystemExit)):
raise
else:
in_progress.finish(result)
return in_progress
class Process(Object):
STATE_STOPPED = 0 # Idle state, no child.
STATE_RUNNING = 1 # start() was called and child is running
STATE_STOPPING = 2 # stop() was called
STATE_DYING = 3 # in the midst of cleanup during child death
STATE_HUNG = 4 # a SIGKILL failed to stop the process
__kaasignals__ = {
'read':
'''
Emitted for each chunk of data read from either stdout or stderr
of the child process.
.. describe:: def callback(chunk, ...)
:param chunk: data read from the child's stdout or stderr.
:type chunk: str
When a callback is connected to the *read* signal, data is automatically
read from the child as soon as it becomes available, and the signal
is emitted.
It is allowed to have a callback connected to the *read* signal
and simultaneously use the :meth:`~kaa.Process.read` and
:meth:`~kaa.Process.readline` methods.
''',
'readline':
'''
Emitted for each line read from either stdout or stderr of the
child process.
.. describe:: def callback(line, ...)
:param line: line read from the child's stdout or stderr.
:type line: str
It is not allowed to have a callback connected to the *readline* signal
and simultaneously use the :meth:`~kaa.Process.readline` method.
Refer to :meth:`~kaa.Process.readline` for more details.
''',
'finished':
'''
Emitted when the child is dead and all data from stdout and stderr
has been consumed.
.. describe:: def callback(exitcode, ...)
:param exitcode: the exit code of the child
:type expected: int
Due to buffering, a child process may be terminated, but the pipes
to its stdout and stderr still open possibly containing buffered
data yet to be read. This signal emits only when the child has
exited and all data has been consumed (or stdout and stderr
explicitly closed).
After this signal emits, the :attr:`~kaa.Process.readable`
property will be False.
''',
'exited':
'''
Emitted when the child process has terminated.
.. describe:: def callback(exitcode, ...)
:param exitcode: the exit code of the child
:type expected: int
Unlike the :attr:`~kaa.Process.signals.finished` signal, this
signal emits when the child is dead (and has been reaped), however
the Process may or may not still be :attr:`~kaa.Process.readable`.
'''
}
def __init__(self, cmd, shell=False, dumpfile=None):
"""
Create a Process object. The subprocess is not started until
:meth:`~kaa.Process.start` is called.
:param cmd: the command to be executed.
:type cmd: string or list of strings
:param shell: True if the command should be executed through a shell.
This allows for shell-like syntax (redirection, pipes,
etc.), but in this case *cmd* must be a string.
:type shell: bool
:param dumpfile: File to which all child stdout and stderr will be
dumped, or None to disable output dumping.
:type dumpfile: None, string (path to filename), file object, IOChannel
Process objects passed to :func:`kaa.inprogress` return a
:class:`~kaa.InProgress` that corresponds to the
:attr:`~kaa.Process.signals.finished` signal (not the
:attr:`~kaa.Process.signals.exited` signal).
"""
super(Process, self).__init__()
self._cmd = cmd
self._shell = shell
self._stop_command = None
# The subprocess.Popen object.
self._child = None
# Weakref of self used to invoke Process._cleanup callback on finalization.
self._cleanup_weakref = None
# The exit code returned by the child once it completes.
self._exitcode = None
if dumpfile:
# Dumpfile specified, create IOChannel which we'll later pass to
# IOSubChannels. dumpfile can be a string (path to file), or
# anything else you can pass to IOChannel (fd, file-like object,
# another IOChannel, etc.)
if isinstance(dumpfile, basestring):
try:
dumpfile = open(dumpfile, 'w')
log.info('Logging process activity to %s' % dumpfile.name)
except IOError:
log.warning('Unable to open %s for logging' % dumpfile)
logger = IOChannel(dumpfile, mode=IO_WRITE)
else:
logger = None
# Create the IOChannels for the child's stdin, stdout, and stderr.
self._stdin = IOChannel()
self._stdout = IOSubChannel(self, logger)
self._stderr = IOSubChannel(self, logger)
self._weak_closed_cbs = []
for fd in self._stdout, self._stderr:
fd.signals['read'].connect_weak(self.signals['read'].emit)
fd.signals['readline'].connect_weak(self.signals['readline'].emit)
# We need to keep track of the WeakCallbacks for _cleanup()
cb = fd.signals['closed'].connect_weak(self._check_dead)
self._weak_closed_cbs.append(cb)
self._stdin.signals['closed'].connect_weak(self._check_dead)
# The Process read and readline signals (aka "global" read/readline signals)
# encapsulate both stdout and stderr. When a new callback is connected
# to these signals, we invoke _update_read_monitor() on the IOSubChannel
# object which will register the fd with the mainloop if necessary.
# (If we didn't do this, the fd would not get registered and therefore
# data never read and therefore the callbacks connected to the global
# read/readline signals never invoked.)
cb = WeakCallback(self._update_read_monitor)
self.signals['read'].changed_cb = cb
self.signals['readline'].changed_cb = cb
self._state = Process.STATE_STOPPED
# InProgress for the whole process. Is created in start(), and finished
# when the process is terminated.
self._in_progress = None
def _update_read_monitor(self, signal=None, change=None):
"""
See IOChannel._update_read_monitor for docstring.
"""
self._stdout._update_read_monitor(signal, change)
self._stderr._update_read_monitor(signal, change)
def __inprogress__(self):
return self._in_progress
@property
def stdin(self):
"""
:class:`~kaa.IOChannel` of child process's stdin.
This object is valid even when the child is not running.
"""
return self._stdin
@property
def stdout(self):
"""
:class:`~kaa.IOChannel` of child process's stdout.
This object is valid even when the child is not running, although it is
obviously not readable until the child is started.
"""
return self._stdout
@property
def stderr(self):
"""
:class:`~kaa.IOChannel` of child process's stderr.
This object is valid even when the child is not running, although it is
obviously not readable until the child is started.
"""
return self._stderr
@property
def pid(self):
"""
The child's pid when it is running (or stopping), or None when it is not.
"""
if self._child:
return self._child.pid
@property
def exitcode(self):
"""
The child's exit code once it has terminated.
If the child is still running or it has not yet been started, this
value will be None.
"""
return self._exitcode
@property
def running(self):
"""
True if the child process is running.
A child that is currently stopping is still considered running. When
the ``running`` property is False, it means :meth:`~kaa.Process.start`
may safely be called.
To test whether :meth:`~kaa.Process.read` or :meth:`~kaa.Process.write`
may be called, use the :attr:`~kaa.Process.readable` and
:attr:`~kaa.Process.writable` properties respectively.
"""
return bool(self._child and self._state not in (Process.STATE_STOPPED, Process.STATE_HUNG))
@property
def readable(self):
"""
True if it is possible to read data from the child.
The child is readable if either the child's stdout or stderr channels
are still open, or if they are both closed but a read call would
succeed anyway due to data remaining in the read queue.
This doesn't necessarily mean the child is still running: a terminated
child may still be read from (there may be data buffered in its stdout
or stderr channels). Use the :attr:`~kaa.Process.running` property if
you want to see if the child is still running.
"""
return self._stdout.readable or self._stderr.readable
@property
def writable(self):
"""
True if it is possible to write data to the child.
If the child process is writable, :meth:`~kaa.Process.write` may
safely be called. A child that is in the process of stopping is not
writable.
"""
return bool(self._child and self._state == Process.STATE_RUNNING)
@property
def stop_command(self):
"""
Stop command for this process.
The command can be either a callable or a string. The command is
invoked (if it is a callback) or the command is written to the child's
stdin (if cmd is a string or unicode) when the process is being
terminated with a call to stop().
Shutdown handlers for the process should be set with this property.
"""
return self._stop_command
@stop_command.setter
def stop_command(self, cmd):
assert(callable(cmd) or type(cmd) in (str, unicode) or cmd == None)
self._stop_command = cmd
@property
def delimiter(self):
"""
String used to split data for use with :meth:`~kaa.Process.readline`.
"""
# stdout and stderr are the same.
return self._stdout.delimiter
@delimiter.setter
def delimiter(self, value):
self._stdout.delimiter = value
self._stderr.delimiter = value
def _normalize_cmd(self, cmd):
"""
Returns a list of arguments based on the given cmd. If cmd is a list,
empty strings and other zero values are removed and the list is
returned. If cmd is a string, it is converted to a list based on shell
semantics.
e.g. program -a "bar baz" \"blah -> ['program', '-a', 'bar baz', '"blah']
"""
if cmd and isinstance(cmd, basestring):
return shlex.split(cmd)
elif isinstance(cmd, (tuple, list)):
return [ x for x in cmd if cmd ]
elif not cmd:
return []
#@threaded() <-- don't
def start(self, args=''):
"""
Starts the process with the given arguments.
:param args: additional arguments when invoking the child, appended
to any arguments specified to the initializer.
:type args: string or list of strings
:return: An :class:`~kaa.InProgress` object, finished with the exitcode
when the child process terminates.
The Process is registered with a global supervisor which holds a strong
reference to the Process object while the child process remains
active.
.. warning::
If timeout() is called on the returned InProgress and the timeout
occurs, the InProgress returned by ``start()`` will be finished with
a TimeoutException even though the child process isn't actually
dead. You can always test the :attr:`~kaa.Process.running`
property, or use the :attr:`~kaa.Process.signals.finished` signal,
which doesn't emit until the child process is genuinely dead.
"""
if self._child and self._state != Process.STATE_HUNG:
raise IOError(errno.EEXIST, 'Child process has already been started')
if not self._shell:
cmd = self._normalize_cmd(self._cmd) + self._normalize_cmd(args)
else:
# If passing through the shell, user must provide cmd and args
# as strings.
if not isinstance(self._cmd, basestring) or not isinstance(args, basestring):
raise ValueError('Command and arguments must be strings when shell=True')
cmd = self._cmd + ' ' + args
supervisor.register(self)
log.debug("Spawning: %s", cmd)
self._child = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE,
stderr=subprocess.PIPE, close_fds=True, shell=self._shell)
self._stdin.wrap(self._child.stdin, IO_WRITE)
self._stdout.wrap(self._child.stdout, IO_READ)
self._stderr.wrap(self._child.stderr, IO_READ)
self._in_progress = InProgress()
self._in_progress.signals['abort'].connect_weak(lambda exc: self.stop())
self._exitcode = None
self._state = Process.STATE_RUNNING
return self._in_progress
@coroutine(policy=POLICY_SINGLETON)
def stop(self, cmd=None, wait=3.0):
"""
Stops the child process.
:param cmd: stop command used to attempt to terminate the child
gracefully; overrides the *stop_command* property if
specified.
:type cmd: string or callable; see :attr:`~kaa.Process.stop_command`
:param wait: number of seconds to wait between termination steps
(see below).
:type wait: float
:returns: A :class:`~kaa.InProgress`, finished (with None) when the
child terminates. If the child refuses to terminate (even
with a SIGKILL) an SystemError exception is thrown to the
InProgress.
The child process is terminated using the following steps:
1. The stop command is written (or invoked) if one is specified,
and up to *wait* seconds is given for the child to terminate.
2. A SIGTERM is issued to the child process, and, again, we wait up
to *wait* seconds for the child to terminate.
3. A SIGKILL is issued to the child process, and this time
we wait up to *wait*\*2 seconds.
If after step 3 the child is still not dead, a SystemError exception
is thrown to the InProgress, as well to the InProgress returned by
:meth:`~kaa.Process.start` and the *finished* signal will be emitted
with the value None.
"""
if self._state != Process.STATE_RUNNING:
# Process is either stopping or dying (or hung).
yield
self._state = Process.STATE_STOPPING
cmd = cmd or self._stop_command
pid = self.pid # See below why we save self.pid
if cmd:
log.debug('Stop command specified: %s, stdin=%d', cmd, self._stdin.alive)
if callable(cmd):
# XXX: should we allow coroutines for cmd and yield them?
cmd()
elif self._stdin.alive:
# This does get buffered. We could bypass the write queue
# by calling self.stdin._write() directly, but maybe the
# IOChannel isn't writable at this moment.
self._stdin.write(cmd)
yield InProgressAny(self._in_progress, delay(wait))
# If we're here, child is either dead or we timed out.
self._stdin.close(immediate=True)
# Either no stop command specified or our stop attempt timed out.
# Try a relatively polite SIGTERM, then SIGKILL.
for sig, pause in ((15, wait), (9, wait * 2)):
if self._state == Process.STATE_STOPPED:
# And we're done.
yield
try:
os.kill(pid, sig)
# Here we yield on the 'exited' signal instead of
# self._in_progress, because the InProgress could in fact be
# finished due to a timeout, not because the process is
# legitimately stopped, whereas the 'exited' signals truly is
# only emitted when the child is dead. (Also, because we
# don't care about losing stdout/err data, otherwise we would
# use 'finished')
yield InProgressAny(self.signals['exited'], delay(pause))
except OSError:
# Process is dead after all.
self._check_dead()
except:
log.exception("Some other error")
# If state isn't STOPPED, make sure pid hasn't changed. Because we yield
# on the 'exited' signal above, it's possible for the user to connect a
# callback to 'exited' that restarts the child, which gets executed before
# this coroutine resumes. If our pid has changed, we know we died. If the
# pid is the same, we have a hung process.
if self._state != Process.STATE_STOPPED and pid == self.pid:
# Child refuses to die even after SIGKILL. :(
self._state = Process.STATE_HUNG
exc = SystemError('Child process (pid=%d) refuses to die even after SIGKILL' % pid)
self._in_progress.throw(SystemError, exc, None)
raise exc
def _async_read(self, stdout_read, stderr_read):
"""
Common implementation for read() and readline().
"""
if not self._stdout.readable and not self._stderr.readable:
return InProgress().finish(None)
# TODO: if child is dead, attach handler to this IP and if len data <
# chunk size, can close the channel. (What makes this more complicated
# is knowing which channel to close, given pass_index=False.)
return InProgressAny(stdout_read(), stderr_read(), pass_index=False,
filter=lambda val: val in (None, ''))
def read(self):
"""
Reads a chunk of data from either stdout or stderr of the process.
There is no way to determine from which (stdout or stderr) the data
was read; if you require this, use the stdout or stderr attributes
directly (however see warning below).
:returns: A :class:`~kaa.InProgress`, finished with the data read.
If it is finished the empty string, it means the child's
stdout and stderr were both closed (which is almost certainly
because the process exited) and no data was available.
No exception is raised if the child is not readable. Like
Socket.read(), it is therefore possible to busy-loop by reading on a
dead child::
while True:
data = yield process.read()
# Or: data = process.read().wait()
So the return value of read() should be tested for non-None.
Alternatively, the readable property could be tested::
while process.readable:
data = yield process.read()
.. warning::
You can read directly from stdout or stderr. However, beware of this
code, which is wrong::
while process.readable:
data = yield process.stdout.read()
In the above incorrect example, process.readable may be True even
though process.stdout is closed (because process.stderr may not be
closed). In this case, process.stdout.read() will finish immediately
with None, resulting in a busy loop. The solution is to test the
process.stdout.readable property instead::
while process.stdout.readable:
data = yield process.stdout.read()
"""
return self._async_read(self._stdout.read, self._stderr.read)
def readline(self):
"""
Reads a line from either stdout or stderr, whichever is available
first.
If finished with None or the empty string, it means that no data was
read and the process exited.
:returns: A :class:`~kaa.InProgress`, finished with the data read.
If it is finished the empty string, it means the child's
stdout and stderr were both closed (which is almost certainly
because the process exited) and no data was available.
Like :meth:`~kaa.Process.read`, it is possible busy-loop with this
method, so you should test its output or test the
:attr:`~kaa.Process.readable` property calling.
"""
return self._async_read(self._stdout.readline, self._stderr.readline)
def write(self, data):
"""
Write data to child's stdin.
Returns an InProgress, which is finished when the data has actually
been written to the child's stdin.
:param data: the data to be written to the channel.
:type data: string
:returns: An :class:`~kaa.InProgress` object, which is finished when the
data has actually been written to the child's stdin.
If the channel closes unexpectedly before the data was
written, an IOError is thrown to the InProgress.
This is a convenience function, as the caller could do
``process.stdin.write()``.
"""
if not self._stdin.alive:
raise IOError(9, 'Cannot write to closed child stdin')
return self._stdin.write(data)
def _emit_finished(self):
if not self._in_progress.finished:
self._in_progress.finish(self._exitcode)
self.signals['finished'].emit(self._exitcode)
def _check_dead(self, expected=None):
"""
Checks to see if the child process has died.
This method is called in the following circumstances:
1. Child is still running and Supervisor invokes it upon SIGCHLD.
(XX) 2. Child is finished and read() or readline() was called.
3. One of stdin/stdout/stderr closes on us.
"""
log.debug('Checking child dead child=%s, state=%d, weakref=%s, stdin=%s, stdout=%s, stderr=%s', self._child,
self._state, self._cleanup_weakref, self._stdin.alive, self._stdout.alive, self._stderr.alive)
if not self._child or self._state in (Process.STATE_STOPPED, Process.STATE_DYING):
# We're already dead or dying.
if not self._stdout.alive and not self._stderr.alive and self._cleanup_weakref:
# Child is dead and all IOChannels are closed. We no longer need
# our weakref cleanup crutch.
self._cleanup_weakref = None
# With child exited and both stdout/stderr closed, the child is
# considered finished.
self._emit_finished()
return
if self._child.poll() is not None:
self._handle_dead()
@classmethod
def _cleanup(cls, weakref, channels, callbacks):
"""
Called when the Process object is destroyed (similar to __del__ but
uses a weakref finalization callback instead to avoid the problems
associated with __del__).
The child process may be finished and reaped, but the stdout/stderr
IOChannels still alive so the caller can retrieve any buffered data
the child left behind. However, when the Process object is dead,
we can close the IOMonitors (which breaks a ref cycle internal to
IOMonitor).
We don't worry about stdin because if stdin is alive it means the
child is still running, and the Supervisor has a reference to us
and therefore it is impossible for this function to get called.
"""
log.debug('Process cleanup: stdout=%s stderr=%s', channels[0].fileno, channels[1].fileno)
for channel, callback in zip(channels, callbacks):
# We previously attached a weak callback for self._check_dead to
# the 'closed' signals of stdout and stderr. Since we're in
# _cleanup(), it means the Process object is destroyed, and all its
# weak references are now dead. When we call close() below, it
# will emit the 'closed' signal, so our weak _check_dead will be
# invoked in Signal.emit(), and fail because the weakref is dead.
#
# Now normally this isn't a problem, because when weak callbacks
# die, they automatically get disconnected. The weak _check_dead
# callbacks are on the chopping block already, and would get
# disconnected automatically after this function. (In other words,
# _cleanup() is invoked before Signal._weakref_destroyed.) So
# we need to disconnect the WeakCallback before invoking close()
# in order to avoid a CallbackError that Signal.emit() reraises.
channel.signals['closed'].disconnect(callback)
channel.close(immediate=True)
def _handle_dead(self):
# This state prevents reentry into this method via 'closed' signal
# of stdin IOChannel.
self._state = Process.STATE_DYING
# Should be safe to wait() to collect zombies. We shouldn't be here
# unless the process actually is done.
self._exitcode = self._child.wait()
log.debug('Child terminated, process=%s exitcode=%d', self, self._exitcode)
# We can close stdin since the child is dead. But stdout and stderr
# need to remain open, in case there is data buffered in them that the
# user may yet retrieve.
self._stdin.close(immediate=True)
self._child = None
# We no longer need help from the supervisor. Any future SIGCHLDs
# are not caused by us.
supervisor.unregister(self)
self._state = Process.STATE_STOPPED
# We don't emit 'finished' here, because stdout/stderr is still open
# and there may be data yet to be read. But we do emit 'exited'
self.signals['exited'].emit(self._exitcode)
if self._stdout.alive or self._stderr.alive:
# Use weakref finializer callback kludge to invoke Process._cleanup
# when Process object goes away in order to close stdout
# and stderr IOChannels.
cb = Callback(self.__class__._cleanup, (self._stdout, self._stderr), self._weak_closed_cbs)
self._cleanup_weakref = weakref.ref(self, cb)
else:
# Child exit and stdout/stderr closed. We're finished.
self._emit_finished()
class Process(Object):
STATE_STOPPED = 0 # Idle state, no child.
STATE_RUNNING = 1 # start() was called and child is running
STATE_STOPPING = 2 # stop() was called
STATE_DYING = 3 # in the midst of cleanup during child death
STATE_HUNG = 4 # a SIGKILL failed to stop the process
__kaasignals__ = {
'read':
"""
Emitted for each chunk of data read from either stdout or stderr
of the child process.
.. describe:: def callback(chunk, ...)
:param chunk: data read from the child's stdout or stderr.
:type chunk: str
When a callback is connected to the *read* signal, data is
automatically read from the child as soon as it becomes available,
and the signal is emitted.
It is allowed to have a callback connected to the *read* signal and
simultaneously use the :meth:`read` and :meth:`readline` methods.
""",
'readline':
"""
Emitted for each line read from either stdout or stderr of the
child process.
.. describe:: def callback(line, ...)
:param line: line read from the child's stdout or stderr.
:type line: str
It is not allowed to have a callback connected to the *readline* signal
and simultaneously use the :meth:`readline` method.
Refer to :meth:`readline` for more details.
""",
'finished':
"""
Emitted when the child is dead and all data from stdout and stderr
has been consumed.
.. describe:: def callback(exitcode, ...)
:param exitcode: the exit code of the child
:type expected: int
Due to buffering, a child process may be terminated, but the pipes
to its stdout and stderr still open possibly containing buffered
data yet to be read. This signal emits only when the child has
exited and all data has been consumed (or stdout and stderr
explicitly closed).
After this signal emits, the :attr:`readable` property will be
False.
""",
'exited':
"""
Emitted when the child process has terminated.
.. describe:: def callback(exitcode, ...)
:param exitcode: the exit code of the child
:type expected: int
Unlike the :attr:`~ksignals.finished` signal, this signal emits
when the child is dead (and has been reaped), however the Process
may or may not still be :attr:`readable`.
"""
}
def __init__(self, cmd, shell=False, dumpfile=None):
"""
Create a Process object. The subprocess is not started until
:meth:`start` is called.
:param cmd: the command to be executed.
:type cmd: string or list of strings
:param shell: True if the command should be executed through a shell.
This allows for shell-like syntax (redirection, pipes,
etc.), but in this case *cmd* must be a string.
:type shell: bool
:param dumpfile: File to which all child stdout and stderr will be
dumped, or None to disable output dumping.
:type dumpfile: None, string (path to filename), file object, IOChannel
Process objects passed to :func:`kaa.inprogress` return a
:class:`~kaa.InProgress` that corresponds to the
:attr:`~kaa.Process.signals.exited` signal (not the
:attr:`~kaa.Process.signals.finished` signal).
"""
super(Process, self).__init__()
self._cmd = cmd
self._shell = shell
self._stop_command = None
# The subprocess.Popen object.
self._child = None
# Weakref of self used to invoke Process._cleanup callback on finalization.
self._cleanup_weakref = None
# The exit code returned by the child once it completes.
self._exitcode = None
if dumpfile:
# Dumpfile specified, create IOChannel which we'll later pass to
# IOSubChannels. dumpfile can be a string (path to file), or
# anything else you can pass to IOChannel (fd, file-like object,
# another IOChannel, etc.)
if isinstance(dumpfile, basestring):
try:
dumpfile = open(dumpfile, 'w')
log.info('Logging process activity to %s' % dumpfile.name)
except IOError:
log.warning('Unable to open %s for logging' % dumpfile)
logger = IOChannel(dumpfile, mode=IO_WRITE)
else:
logger = None
# Create the IOChannels for the child's stdin, stdout, and stderr.
self._stdin = IOChannel()
self._stdout = IOSubChannel(self, logger)
self._stderr = IOSubChannel(self, logger)
self._weak_closed_cbs = []
for fd in self._stdout, self._stderr:
fd.signals['read'].connect_weak(self.signals['read'].emit)
fd.signals['readline'].connect_weak(self.signals['readline'].emit)
# We need to keep track of the WeakCallables for _cleanup()
cb = fd.signals['closed'].connect_weak(self._check_dead)
self._weak_closed_cbs.append(cb)
self._stdin.signals['closed'].connect_weak(self._check_dead)
# The Process read and readline signals (aka "global" read/readline signals)
# encapsulate both stdout and stderr. When a new callback is connected
# to these signals, we invoke _update_read_monitor() on the IOSubChannel
# object which will register the fd with the mainloop if necessary.
# (If we didn't do this, the fd would not get registered and therefore
# data never read and therefore the callbacks connected to the global
# read/readline signals never invoked.)
cb = WeakCallable(self._update_read_monitor)
self.signals['read'].changed_cb = cb
self.signals['readline'].changed_cb = cb
self._state = Process.STATE_STOPPED
# InProgress for the whole process. Is recreated in start() for
# multiple invocations, and finished when the process is terminated.
self._in_progress = InProgress()
def _update_read_monitor(self, signal=None, change=None):
"""
See IOChannel._update_read_monitor for docstring.
"""
self._stdout._update_read_monitor(signal, change)
self._stderr._update_read_monitor(signal, change)
def __inprogress__(self):
return self._in_progress
@property
def stdin(self):
"""
:class:`~kaa.IOChannel` of child process's stdin.
This object is valid even when the child is not running.
"""
return self._stdin
@property
def stdout(self):
"""
:class:`~kaa.IOChannel` of child process's stdout.
This object is valid even when the child is not running, although it is
obviously not readable until the child is started.
"""
return self._stdout
@property
def stderr(self):
"""
:class:`~kaa.IOChannel` of child process's stderr.
This object is valid even when the child is not running, although it is
obviously not readable until the child is started.
"""
return self._stderr
@property
def pid(self):
"""
The child's pid when it is running (or stopping), or None when it is not.
"""
if self._child:
return self._child.pid
@property
def exitcode(self):
"""
The child's exit code once it has terminated.
If the child is still running or it has not yet been started, this
value will be None.
"""
return self._exitcode
@property
def running(self):
"""
True if the child process is running.
A child that is currently stopping is still considered running. When
the :attr:`running` property is False, it means :meth:`start`
may safely be called.
To test whether :meth:`read` or :meth:`write` may be called, use the
:attr:`readable` and :attr:`writable` properties respectively.
"""
return bool(self._child and self._state not in (Process.STATE_STOPPED, Process.STATE_HUNG))
@property
def stopping(self):
"""
True if the child process is currently being shut down.
True when :meth:`stop` was called and the process is not stopped yet.
"""
return bool(self._child and self._state == Process.STATE_STOPPING)
@property
def readable(self):
"""
True if it is possible to read data from the child.
The child is readable if either the child's :attr:`stdout` or
:attr:`stderr` channels are still open, or if they are both closed but
a read call would succeed anyway due to data remaining in the read
queue.
This doesn't necessarily mean the child is still running: a terminated
child may still be read from (there may be data buffered in its
:attr:`stdout` or :attr:`stderr` channels). Use the :attr:`running`
property if you want to see if the child is still running.
"""
return self._stdout.readable or self._stderr.readable
@property
def writable(self):
"""
True if it is possible to write data to the child.
If the child process is writable, :meth:`write` may safely be called.
A child that is in the process of :attr:`stopping` is not writable.
"""
return bool(self._child and self._state == Process.STATE_RUNNING)
@property
def stop_command(self):
"""
Stop command for this process.
The command can be either a callable or a string. The command is
invoked (if it is a callable) or the command is written to the child's
stdin (if cmd is a string or unicode) when the process is being
terminated with a call to stop().
Shutdown handlers for the process should be set with this property.
"""
return self._stop_command
@stop_command.setter
def stop_command(self, cmd):
assert(callable(cmd) or type(cmd) in (str, unicode) or cmd == None)
self._stop_command = cmd
@property
def delimiter(self):
"""
String used to split data for use with :meth:`readline`.
"""
# stdout and stderr are the same.
return self._stdout.delimiter
@delimiter.setter
def delimiter(self, value):
self._stdout.delimiter = value
self._stderr.delimiter = value
def _normalize_cmd(self, cmd):
"""
Returns a list of arguments based on the given cmd. If cmd is a list,
empty strings and other zero values are removed and the list is
returned. If cmd is a string, it is converted to a list based on shell
semantics.
e.g. program -a "bar baz" \"blah -> ['program', '-a', 'bar baz', '"blah']
"""
if cmd and isinstance(cmd, basestring):
return shlex.split(cmd)
elif isinstance(cmd, (tuple, list)):
return [ x for x in cmd if cmd ]
elif not cmd:
return []
def _child_preexec(self):
"""
Callback function for Popen object that gets invoked in the child after
forking but prior to execing.
"""
# Children will inherit any ignored signals, so before we exec, reset
# any ignored signals to their defaults.
for sig in range(1, signal.NSIG):
if signal.getsignal(sig) == signal.SIG_IGN:
signal.signal(sig, signal.SIG_DFL)
#@threaded() <-- don't
def start(self, args=''):
"""
Starts the process with the given arguments.
:param args: additional arguments when invoking the child, appended
to any arguments specified to the initializer.
:type args: string or list of strings
:return: An :class:`~kaa.InProgress` object, finished with the exitcode
when the child process terminates (when the
:attr:`~signals.exited` signal is emitted).
The Process is registered with a global supervisor which holds a strong
reference to the Process object while the child process remains
active.
.. warning::
If :meth:`~kaa.InProgress.timeout` is called on the returned
InProgress and the timeout occurs, the InProgress returned by
:meth:`start` will be finished with a :class:`~kaa.TimeoutException`
even though the child process isn't actually dead. You can always
test the :attr:`running` property, or use the
:attr:`~signals.finished` signal, which doesn't emit until the child
process is genuinely dead.
"""
if self._child and self._state != Process.STATE_HUNG:
raise IOError(errno.EEXIST, 'Child process has already been started')
if not self._shell:
cmd = self._normalize_cmd(self._cmd) + self._normalize_cmd(args)
else:
# If passing through the shell, user must provide cmd and args
# as strings.
if not isinstance(self._cmd, basestring) or not isinstance(args, basestring):
raise ValueError('Command and arguments must be strings when shell=True')
cmd = self._cmd + ' ' + args
if self._in_progress.finished:
self._in_progress = InProgress()
self._in_progress.signals['abort'].connect_weak(lambda exc: self.stop())
self._exitcode = None
supervisor.register(self)
log.debug("Spawning: %s", cmd)
self._child = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE,
stderr=subprocess.PIPE, preexec_fn=self._child_preexec,
close_fds=True, shell=self._shell)
self._stdin.wrap(self._child.stdin, IO_WRITE)
self._stdout.wrap(self._child.stdout, IO_READ)
self._stderr.wrap(self._child.stderr, IO_READ)
self._state = Process.STATE_RUNNING
return self._in_progress
@coroutine(policy=POLICY_SINGLETON)
def stop(self, cmd=None, wait=3.0):
"""
Stops the child process.
:param cmd: stop command used to attempt to terminate the child
gracefully; overrides the :attr:`stop_command` property if
specified.
:type cmd: string or callable
:param wait: number of seconds to wait between termination steps
(see below).
:type wait: float
:returns: A :class:`~kaa.InProgress`, finished (with None) when the
child terminates. If the child refuses to terminate (even
with a SIGKILL) an SystemError exception is thrown to the
InProgress.
The child process is terminated using the following steps:
1. The stop command is written (or invoked) if one is specified,
and up to *wait* seconds is given for the child to terminate.
2. A SIGTERM is issued to the child process, and, again, we wait up
to *wait* seconds for the child to terminate.
3. A SIGKILL is issued to the child process, and this time
we wait up to *wait*\*2 seconds.
If after step 3 the child is still not dead, a SystemError exception is
thrown to the InProgress, as well to the InProgress returned by
:meth:`start` and the :attr:`~signals.finished` signal will be emitted
with the value None.
"""
if self._state != Process.STATE_RUNNING:
# Process is either stopping or dying (or hung).
yield
self._state = Process.STATE_STOPPING
cmd = cmd or self._stop_command
pid = self.pid # See below why we save self.pid
if cmd:
log.debug('Stop command specified: %s, stdin=%d', cmd, self._stdin.alive)
if callable(cmd):
# XXX: should we allow coroutines for cmd and yield them?
cmd()
elif self._stdin.alive:
# This does get buffered. We could bypass the write queue
# by calling self.stdin._write() directly, but maybe the
# IOChannel isn't writable at this moment.
self._stdin.write(cmd)
yield InProgressAny(self._in_progress, delay(wait))
# If we're here, child is either dead or we timed out.
self._stdin.close(immediate=True)
# Either no stop command specified or our stop attempt timed out.
# Try a relatively polite SIGTERM, then SIGKILL.
for sig, pause in ((15, wait), (9, wait * 2)):
if self._state == Process.STATE_STOPPED:
# And we're done.
yield
try:
os.kill(pid, sig)
# Here we yield on the 'exited' signal instead of
# self._in_progress, because the InProgress could in fact be
# finished due to a timeout, not because the process is
# legitimately stopped, whereas the 'exited' signals truly is
# only emitted when the child is dead. (Also, because we
# don't care about losing stdout/err data, otherwise we would
# use 'finished')
yield InProgressAny(self.signals['exited'], delay(pause))
except OSError:
# Process is dead after all.
self._check_dead()
except:
log.exception("Some other error")
# If state isn't STOPPED, make sure pid hasn't changed. Because we yield
# on the 'exited' signal above, it's possible for the user to connect a
# callback to 'exited' that restarts the child, which gets executed before
# this coroutine resumes. If our pid has changed, we know we died. If the
# pid is the same, we have a hung process.
if self._state != Process.STATE_STOPPED and pid == self.pid:
# Child refuses to die even after SIGKILL. :(
self._state = Process.STATE_HUNG
exc = SystemError('Child process (pid=%d) refuses to die even after SIGKILL' % pid)
self._in_progress.throw(SystemError, exc, None)
raise exc
def _async_read(self, stdout_read, stderr_read):
"""
Common implementation for read() and readline().
"""
if not self._stdout.readable and not self._stderr.readable:
return InProgress().finish(None)
# TODO: if child is dead, attach handler to this IP and if len data <
# chunk size, can close the channel. (What makes this more complicated
# is knowing which channel to close, given pass_index=False.)
return InProgressAny(stdout_read(), stderr_read(), pass_index=False,
filter=lambda val: val in (None, ''))
def read(self):
"""
Reads a chunk of data from either stdout or stderr of the process.
There is no way to determine from which (stdout or stderr) the data
was read; if you require this, use the :attr:`stdout` or :attr:`stderr`
attributes directly (however see warning below).
:returns: A :class:`~kaa.InProgress`, finished with the data read.
If it is finished the empty string, it means the child's
stdout and stderr were both closed (which is almost certainly
because the process exited) and no data was available.
No exception is raised if the child is not readable. Like
:meth:`Socket.read`, it is therefore possible to busy-loop by reading
on a dead child::
while True:
data = yield process.read()
# Or: data = process.read().wait()
So the return value of read() should be tested for non-None.
Alternatively, the :attr:`readable` property could be tested::
while process.readable:
data = yield process.read()
.. warning::
You can read directly from stdout or stderr. However, beware of this
code, which is wrong::
while process.readable:
data = yield process.stdout.read()
In the above incorrect example, process.readable may be True even
though process.stdout is closed (because process.stderr may not be
closed). In this case, process.stdout.read() will finish immediately
with None, resulting in a busy loop. The solution is to test the
process.stdout.readable property instead::
while process.stdout.readable:
data = yield process.stdout.read()
"""
return self._async_read(self._stdout.read, self._stderr.read)
def readline(self):
"""
Reads a line from either stdout or stderr, whichever is available
first.
If finished with None or the empty string, it means that no data was
read and the process exited.
:returns: A :class:`~kaa.InProgress`, finished with the data read.
If it is finished the empty string, it means the child's
stdout and stderr were both closed (which is almost certainly
because the process exited) and no data was available.
Like :meth:`read`, it is possible busy-loop with this method, so you
should test its output or test the :attr:`readable` property calling.
"""
return self._async_read(self._stdout.readline, self._stderr.readline)
def write(self, data):
"""
Write data to child's stdin.
Returns an InProgress, which is finished when the data has actually
been written to the child's stdin.
:param data: the data to be written to the channel.
:type data: string
:returns: An :class:`~kaa.InProgress` object, which is finished when the
data has actually been written to the child's stdin.
If the channel closes unexpectedly before the data was
written, an IOError is thrown to the InProgress.
This is a convenience function, as the caller could do
``process.stdin.write()``.
"""
if not self._stdin.alive:
raise IOError(9, 'Cannot write to closed child stdin')
return self._stdin.write(data)
@coroutine()
def communicate(self, input=None):
"""
One-time interaction with the process, sending the given input, and
receiving all output from the child.
:param input: the data to send to the child's stdin
:type input: str
:return: an :class:`~kaa.InProgress`, which will be finished with a
2-tuple (stdoutdata, stderrdata)
If the process has not yet been started, :meth:`start` will be
called implicitly.
Any data previously written to the child with :meth:`write` will be
flushed and the pipe to the child's stdin will be closed. All
subsequent data from the child's stdout and stderr will be read until
EOF. The child will be terminated before returning.
This method is modeled after Python's standard library call
``subprocess.Popen.communicate()``.
"""
if self._state == Process.STATE_STOPPED:
self.start()
if input:
yield self.write(input)
self.stdin.close()
buf_out = BytesIO()
while self.stdout.readable:
buf_out.write((yield self.stdout.read()))
buf_err = BytesIO()
while self.stderr.readable:
buf_err.write((yield self.stderr.read()))
yield self.stop()
yield (buf_out.getvalue(), buf_err.getvalue())
def _check_dead(self, expected=None):
"""
Checks to see if the child process has died.
This method is called in the following circumstances:
1. Child is still running and Supervisor invokes it upon SIGCHLD.
(XX) 2. Child is finished and read() or readline() was called.
3. One of stdin/stdout/stderr closes on us.
"""
log.debug('Checking child dead child=%s, state=%d, weakref=%s, stdin=%s, stdout=%s, stderr=%s', self._child,
self._state, self._cleanup_weakref, self._stdin.alive, self._stdout.alive, self._stderr.alive)
if not self._child or self._state in (Process.STATE_STOPPED, Process.STATE_DYING):
# We're already dead or dying.
if not self._stdout.alive and not self._stderr.alive and self._cleanup_weakref:
# Child is dead and all IOChannels are closed. We no longer need
# our weakref cleanup crutch.
self._cleanup_weakref = None
# With child exited and both stdout/stderr closed, the child is
# considered finished.
self.signals['finished'].emit(self._exitcode)
return
if self._child.poll() is not None:
self._handle_dead()
@classmethod
def _cleanup(cls, weakref, channels, callbacks):
"""
Called when the Process object is destroyed (similar to __del__ but
uses a weakref finalization callback instead to avoid the problems
associated with __del__).
The child process may be finished and reaped, but the stdout/stderr
IOChannels still alive so the caller can retrieve any buffered data
the child left behind. However, when the Process object is dead,
we can close the IOMonitors (which breaks a ref cycle internal to
IOMonitor).
We don't worry about stdin because if stdin is alive it means the
child is still running, and the Supervisor has a reference to us
and therefore it is impossible for this function to get called.
"""
log.debug('Process cleanup: stdout=%s stderr=%s', channels[0].fileno, channels[1].fileno)
for channel, callback in zip(channels, callbacks):
# We previously attached a weak callback for self._check_dead to
# the 'closed' signals of stdout and stderr. Since we're in
# _cleanup(), it means the Process object is destroyed, and all its
# weak references are now dead. When we call close() below, it
# will emit the 'closed' signal, so our weak _check_dead will be
# invoked in Signal.emit(), and fail because the weakref is dead.
#
# Now normally this isn't a problem, because when weak callbacks
# die, they automatically get disconnected. The weak _check_dead
# callbacks are on the chopping block already, and would get
# disconnected automatically after this function. (In other words,
# _cleanup() is invoked before Signal._weakref_destroyed.) So
# we need to disconnect the WeakCallable before invoking close()
# in order to avoid a CallableError that Signal.emit() reraises.
channel.signals['closed'].disconnect(callback)
channel.close(immediate=True)
def _handle_dead(self):
# This state prevents reentry into this method via 'closed' signal
# of stdin IOChannel.
self._state = Process.STATE_DYING
# Should be safe to wait() to collect zombies. We shouldn't be here
# unless the process actually is done.
self._exitcode = self._child.wait()
log.debug('Child terminated, process=%s exitcode=%d', self, self._exitcode)
# We can close stdin since the child is dead. But stdout and stderr
# need to remain open, in case there is data buffered in them that the
# user may yet retrieve.
self._stdin.close(immediate=True)
self._child = None
# We no longer need help from the supervisor. Any future SIGCHLDs
# are not caused by us.
supervisor.unregister(self)
self._state = Process.STATE_STOPPED
# We don't emit 'finished' here, because stdout/stderr is still open
# and there may be data yet to be read. But we do emit 'exited' and
# finish the InProgress object.
self.signals['exited'].emit(self._exitcode)
if not self._in_progress.finished:
self._in_progress.finish(self._exitcode)
if self._stdout.alive or self._stderr.alive:
# Use weakref finializer callback kludge to invoke Process._cleanup
# when Process object goes away in order to close stdout
# and stderr IOChannels.
cb = Callable(self.__class__._cleanup, (self._stdout, self._stderr), self._weak_closed_cbs)
self._cleanup_weakref = weakref.ref(self, cb)
else:
# Child exit and stdout/stderr closed. We're finished.
self.signals['finished'].emit(self._exitcode)
class Generator(object):
"""
Generator for InProgress objects
"""
def __init__(self):
self._waiting = None
self._finished = []
self._generator_exit = False
self._populate = InProgress()
def __inprogress__(self):
"""
Wait until at least one item is produced or the generator
finished before that happens.
"""
return self._populate
@threaded(MAINTHREAD)
def send(self, result, exception=False):
"""
Send a new value (producer)
"""
delayed = [ InProgress(), result, False, exception ]
if result is GeneratorExit:
self._generator_exit = True
delayed = None
if not self._populate.finished:
# First item
self._delay = result
self._waiting = delayed
self._populate.finish(self)
return
ip, result, handled, exception = self._waiting
if not handled:
# InProgress not returned in __iter__, add it to the
# finished objects
self._finished.append(ip)
self._waiting = delayed
if exception:
ip.throw(*result)
else:
ip.finish(result)
def throw(self, type, value, tb):
"""
Throw an error, this will stop the generator
"""
self.send((type, value, tb), exception=True)
self.send(GeneratorExit)
return False
def finish(self, result):
"""
Finish the generator, the result will be ignored
"""
self.send(GeneratorExit)
def __iter__(self):
"""
Iterate over the values (consumer)
"""
while not self._generator_exit or self._waiting or self._finished:
if not self._finished:
# no finished items yet, return the waiting InProgress
self._waiting[2] = True
yield self._waiting[0]
else:
yield self._finished.pop(0)