# -*- test-case-name: twisted.internet.test.test_core -*- # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. """ A reactor for integrating with U{CFRunLoop}, the CoreFoundation main loop used by macOS. This is useful for integrating Twisted with U{PyObjC} applications. """ from __future__ import annotations __all__ = ["install", "CFReactor"] import sys from zope.interface import implementer from CFNetwork import ( CFSocketCreateRunLoopSource, CFSocketCreateWithNative, CFSocketDisableCallBacks, CFSocketEnableCallBacks, CFSocketInvalidate, CFSocketSetSocketFlags, kCFSocketAutomaticallyReenableReadCallBack, kCFSocketAutomaticallyReenableWriteCallBack, kCFSocketConnectCallBack, kCFSocketReadCallBack, kCFSocketWriteCallBack, ) from CoreFoundation import ( CFAbsoluteTimeGetCurrent, CFRunLoopAddSource, CFRunLoopAddTimer, CFRunLoopGetCurrent, CFRunLoopRemoveSource, CFRunLoopRun, CFRunLoopStop, CFRunLoopTimerCreate, CFRunLoopTimerInvalidate, kCFAllocatorDefault, kCFRunLoopCommonModes, ) from twisted.internet.interfaces import IReactorFDSet from twisted.internet.posixbase import _NO_FILEDESC, PosixReactorBase from twisted.python import log # We know that we're going to run on macOS so we can just pick the # POSIX-appropriate waker. This also avoids having a dynamic base class and # so lets more things get type checked. from ._signals import _UnixWaker _READ = 0 _WRITE = 1 _preserveSOError = 1 << 6 class _WakerPlus(_UnixWaker): """ The normal Twisted waker will simply wake up the main loop, which causes an iteration to run, which in turn causes L{ReactorBase.runUntilCurrent} to get invoked. L{CFReactor} has a slightly different model of iteration, though: rather than have each iteration process the thread queue, then timed calls, then file descriptors, each callback is run as it is dispatched by the CFRunLoop observer which triggered it. So this waker needs to not only unblock the loop, but also make sure the work gets done; so, it reschedules the invocation of C{runUntilCurrent} to be immediate (0 seconds from now) even if there is no timed call work to do. """ def __init__(self, reactor): super().__init__() self.reactor = reactor def doRead(self): """ Wake up the loop and force C{runUntilCurrent} to run immediately in the next timed iteration. """ result = super().doRead() self.reactor._scheduleSimulate(True) return result @implementer(IReactorFDSet) class CFReactor(PosixReactorBase): """ The CoreFoundation reactor. You probably want to use this via the L{install} API. @ivar _fdmap: a dictionary, mapping an integer (a file descriptor) to a 4-tuple of: - source: a C{CFRunLoopSource}; the source associated with this socket. - socket: a C{CFSocket} wrapping the file descriptor. - descriptor: an L{IReadDescriptor} and/or L{IWriteDescriptor} provider. - read-write: a 2-C{list} of booleans: respectively, whether this descriptor is currently registered for reading or registered for writing. @ivar _idmap: a dictionary, mapping the id() of an L{IReadDescriptor} or L{IWriteDescriptor} to a C{fd} in L{_fdmap}. Implemented in this manner so that we don't have to rely (even more) on the hashability of L{IReadDescriptor} providers, and we know that they won't be collected since these are kept in sync with C{_fdmap}. Necessary because the .fileno() of a file descriptor may change at will, so we need to be able to look up what its file descriptor I{used} to be, so that we can look it up in C{_fdmap} @ivar _cfrunloop: the C{CFRunLoop} pyobjc object wrapped by this reactor. @ivar _inCFLoop: Is C{CFRunLoopRun} currently running? @type _inCFLoop: L{bool} @ivar _currentSimulator: if a CFTimer is currently scheduled with the CF run loop to run Twisted callLater calls, this is a reference to it. Otherwise, it is L{None} """ def __init__(self, runLoop=None, runner=None): self._fdmap = {} self._idmap = {} if runner is None: runner = CFRunLoopRun self._runner = runner if runLoop is None: runLoop = CFRunLoopGetCurrent() self._cfrunloop = runLoop PosixReactorBase.__init__(self) def _wakerFactory(self) -> _WakerPlus: return _WakerPlus(self) def _socketCallback( self, cfSocket, callbackType, ignoredAddress, ignoredData, context ): """ The socket callback issued by CFRunLoop. This will issue C{doRead} or C{doWrite} calls to the L{IReadDescriptor} and L{IWriteDescriptor} registered with the file descriptor that we are being notified of. @param cfSocket: The C{CFSocket} which has got some activity. @param callbackType: The type of activity that we are being notified of. Either C{kCFSocketReadCallBack} or C{kCFSocketWriteCallBack}. @param ignoredAddress: Unused, because this is not used for either of the callback types we register for. @param ignoredData: Unused, because this is not used for either of the callback types we register for. @param context: The data associated with this callback by C{CFSocketCreateWithNative} (in C{CFReactor._watchFD}). A 2-tuple of C{(int, CFRunLoopSource)}. """ (fd, smugglesrc) = context if fd not in self._fdmap: # Spurious notifications seem to be generated sometimes if you # CFSocketDisableCallBacks in the middle of an event. I don't know # about this FD, any more, so let's get rid of it. CFRunLoopRemoveSource(self._cfrunloop, smugglesrc, kCFRunLoopCommonModes) return src, skt, readWriteDescriptor, rw = self._fdmap[fd] def _drdw(): why = None isRead = False try: if readWriteDescriptor.fileno() == -1: why = _NO_FILEDESC else: isRead = callbackType == kCFSocketReadCallBack # CFSocket seems to deliver duplicate read/write # notifications sometimes, especially a duplicate # writability notification when first registering the # socket. This bears further investigation, since I may # have been mis-interpreting the behavior I was seeing. # (Running the full Twisted test suite, while thorough, is # not always entirely clear.) Until this has been more # thoroughly investigated , we consult our own # reading/writing state flags to determine whether we # should actually attempt a doRead/doWrite first. -glyph if isRead: if rw[_READ]: why = readWriteDescriptor.doRead() else: if rw[_WRITE]: why = readWriteDescriptor.doWrite() except BaseException: why = sys.exc_info()[1] log.err() if why: self._disconnectSelectable(readWriteDescriptor, why, isRead) log.callWithLogger(readWriteDescriptor, _drdw) def _watchFD(self, fd, descr, flag): """ Register a file descriptor with the C{CFRunLoop}, or modify its state so that it's listening for both notifications (read and write) rather than just one; used to implement C{addReader} and C{addWriter}. @param fd: The file descriptor. @type fd: L{int} @param descr: the L{IReadDescriptor} or L{IWriteDescriptor} @param flag: the flag to register for callbacks on, either C{kCFSocketReadCallBack} or C{kCFSocketWriteCallBack} """ if fd == -1: raise RuntimeError("Invalid file descriptor.") if fd in self._fdmap: src, cfs, gotdescr, rw = self._fdmap[fd] # do I need to verify that it's the same descr? else: ctx = [] ctx.append(fd) cfs = CFSocketCreateWithNative( kCFAllocatorDefault, fd, kCFSocketReadCallBack | kCFSocketWriteCallBack | kCFSocketConnectCallBack, self._socketCallback, ctx, ) CFSocketSetSocketFlags( cfs, kCFSocketAutomaticallyReenableReadCallBack | kCFSocketAutomaticallyReenableWriteCallBack | # This extra flag is to ensure that CF doesn't (destructively, # because destructively is the only way to do it) retrieve # SO_ERROR and thereby break twisted.internet.tcp.BaseClient, # which needs SO_ERROR to tell it whether or not it needs to # call connect_ex a second time. _preserveSOError, ) src = CFSocketCreateRunLoopSource(kCFAllocatorDefault, cfs, 0) ctx.append(src) CFRunLoopAddSource(self._cfrunloop, src, kCFRunLoopCommonModes) CFSocketDisableCallBacks( cfs, kCFSocketReadCallBack | kCFSocketWriteCallBack | kCFSocketConnectCallBack, ) rw = [False, False] self._idmap[id(descr)] = fd self._fdmap[fd] = src, cfs, descr, rw rw[self._flag2idx(flag)] = True CFSocketEnableCallBacks(cfs, flag) def _flag2idx(self, flag): """ Convert a C{kCFSocket...} constant to an index into the read/write state list (C{_READ} or C{_WRITE}) (the 4th element of the value of C{self._fdmap}). @param flag: C{kCFSocketReadCallBack} or C{kCFSocketWriteCallBack} @return: C{_READ} or C{_WRITE} """ return {kCFSocketReadCallBack: _READ, kCFSocketWriteCallBack: _WRITE}[flag] def _unwatchFD(self, fd, descr, flag): """ Unregister a file descriptor with the C{CFRunLoop}, or modify its state so that it's listening for only one notification (read or write) as opposed to both; used to implement C{removeReader} and C{removeWriter}. @param fd: a file descriptor @type fd: C{int} @param descr: an L{IReadDescriptor} or L{IWriteDescriptor} @param flag: C{kCFSocketWriteCallBack} C{kCFSocketReadCallBack} """ if id(descr) not in self._idmap: return if fd == -1: # need to deal with it in this case, I think. realfd = self._idmap[id(descr)] else: realfd = fd src, cfs, descr, rw = self._fdmap[realfd] CFSocketDisableCallBacks(cfs, flag) rw[self._flag2idx(flag)] = False if not rw[_READ] and not rw[_WRITE]: del self._idmap[id(descr)] del self._fdmap[realfd] CFRunLoopRemoveSource(self._cfrunloop, src, kCFRunLoopCommonModes) CFSocketInvalidate(cfs) def addReader(self, reader): """ Implement L{IReactorFDSet.addReader}. """ self._watchFD(reader.fileno(), reader, kCFSocketReadCallBack) def addWriter(self, writer): """ Implement L{IReactorFDSet.addWriter}. """ self._watchFD(writer.fileno(), writer, kCFSocketWriteCallBack) def removeReader(self, reader): """ Implement L{IReactorFDSet.removeReader}. """ self._unwatchFD(reader.fileno(), reader, kCFSocketReadCallBack) def removeWriter(self, writer): """ Implement L{IReactorFDSet.removeWriter}. """ self._unwatchFD(writer.fileno(), writer, kCFSocketWriteCallBack) def removeAll(self): """ Implement L{IReactorFDSet.removeAll}. """ allDesc = {descr for src, cfs, descr, rw in self._fdmap.values()} allDesc -= set(self._internalReaders) for desc in allDesc: self.removeReader(desc) self.removeWriter(desc) return list(allDesc) def getReaders(self): """ Implement L{IReactorFDSet.getReaders}. """ return [descr for src, cfs, descr, rw in self._fdmap.values() if rw[_READ]] def getWriters(self): """ Implement L{IReactorFDSet.getWriters}. """ return [descr for src, cfs, descr, rw in self._fdmap.values() if rw[_WRITE]] def _moveCallLaterSooner(self, tple): """ Override L{PosixReactorBase}'s implementation of L{IDelayedCall.reset} so that it will immediately reschedule. Normally C{_moveCallLaterSooner} depends on the fact that C{runUntilCurrent} is always run before the mainloop goes back to sleep, so this forces it to immediately recompute how long the loop needs to stay asleep. """ result = PosixReactorBase._moveCallLaterSooner(self, tple) self._scheduleSimulate() return result def startRunning(self, installSignalHandlers: bool = True) -> None: """ Start running the reactor, then kick off the timer that advances Twisted's clock to keep pace with CFRunLoop's. """ super().startRunning(installSignalHandlers) # Before 'startRunning' is called, the reactor is not attached to the # CFRunLoop[1]; specifically, the CFTimer that runs all of Twisted's # timers is not active and will not have been added to the loop by any # application code. Now that _running is probably[2] True, we need to # ensure that timed calls will actually run on the main loop. This # call needs to be here, rather than at the top of mainLoop, because # it's possible to use startRunning to *attach* a reactor to an # already-running CFRunLoop, i.e. within a plugin for an application # that doesn't otherwise use Twisted, rather than calling it via run(). self._scheduleSimulate(force=True) # [1]: readers & writers are still active in the loop, but arguably # they should not be. # [2]: application code within a 'startup' system event trigger *may* # have already crashed the reactor and thus set _started to False, # but that specific case is handled by mainLoop, since that case # is inherently irrelevant in an attach-to-application case and is # only necessary to handle mainLoop spuriously blocking. _inCFLoop = False def mainLoop(self) -> None: """ Run the runner (C{CFRunLoopRun} or something that calls it), which runs the run loop until C{crash()} is called. """ if not self._started: # If we arrive here, we were crashed by application code in a # 'startup' system event trigger, (or crashed manually before the # application calls 'mainLoop' directly for whatever reason; sigh, # this method should not be public). However, application code # doing obscure things will expect an invocation of this loop to # have at least *one* pass over ready readers, writers, and delayed # calls. iterate(), in particular, is emulated in exactly this way # in this reactor implementation. In order to ensure that we enter # the real implementation of the mainloop and do all of those # things, we need to set _started back to True so that callLater # actually schedules itself against the CFRunLoop, but immediately # crash once we are in the context of the loop where we've run # ready I/O and timers. def docrash() -> None: self.crash() self._started = True self.callLater(0, docrash) already = False try: while self._started: if already: # Sometimes CFRunLoopRun (or its equivalents) may exit # without CFRunLoopStop being called. # This is really only *supposed* to happen when it runs out # of sources & timers to process. However, in full Twisted # test-suite runs we have observed, extremely rarely (once # in every 3000 tests or so) CFRunLoopRun exiting in cases # where it seems as though there *is* still some work to # do. However, given the difficulty of reproducing the # race conditions necessary to make this happen, it's # possible that we have missed some nuance of when # CFRunLoop considers the list of work "empty" and various # callbacks and timers to be "invalidated". Therefore we # are not fully confident that this is a platform bug, but # it is nevertheless unexpected behavior from our reading # of the documentation. # To accommodate this rare and slightly ambiguous stress # case, we make extra sure that our scheduled timer is # re-created on the loop as a CFRunLoopTimer, which # reliably gives the loop some work to do and 'fixes' it if # it exited due to having no active sources or timers. self._scheduleSimulate() # At this point, there may be a little more code that we # would need to put here for full correctness for a very # peculiar type of application: if you're writing a # command-line tool using CFReactor, adding *nothing* to # the reactor itself, disabling even the internal Waker # file descriptors, then there's a possibility that # CFRunLoopRun will exit early, and if we have no timers, # we might busy-loop here. Because we cannot seem to force # this to happen under normal circumstances, we're leaving # that code out. already = True self._inCFLoop = True try: self._runner() finally: self._inCFLoop = False finally: self._stopSimulating() _currentSimulator: object | None = None def _stopSimulating(self) -> None: """ If we have a CFRunLoopTimer registered with the CFRunLoop, invalidate it and set it to None. """ if self._currentSimulator is None: return CFRunLoopTimerInvalidate(self._currentSimulator) self._currentSimulator = None def _scheduleSimulate(self, force: bool = False) -> None: """ Schedule a call to C{self.runUntilCurrent}. This will cancel the currently scheduled call if it is already scheduled. @param force: Even if there are no timed calls, make sure that C{runUntilCurrent} runs immediately (in a 0-seconds-from-now C{CFRunLoopTimer}). This is necessary for calls which need to trigger behavior of C{runUntilCurrent} other than running timed calls, such as draining the thread call queue or calling C{crash()} when the appropriate flags are set. @type force: C{bool} """ self._stopSimulating() if not self._started: # If the reactor is not running (e.g. we are scheduling callLater # calls before starting the reactor) we should not be scheduling # CFRunLoopTimers against the global CFRunLoop. return timeout = 0.0 if force else self.timeout() if timeout is None: return fireDate = CFAbsoluteTimeGetCurrent() + timeout def simulate(cftimer, extra): self._currentSimulator = None self.runUntilCurrent() self._scheduleSimulate() c = self._currentSimulator = CFRunLoopTimerCreate( kCFAllocatorDefault, fireDate, 0, 0, 0, simulate, None ) CFRunLoopAddTimer(self._cfrunloop, c, kCFRunLoopCommonModes) def callLater(self, _seconds, _f, *args, **kw): """ Implement L{IReactorTime.callLater}. """ delayedCall = PosixReactorBase.callLater(self, _seconds, _f, *args, **kw) self._scheduleSimulate() return delayedCall def stop(self): """ Implement L{IReactorCore.stop}. """ PosixReactorBase.stop(self) self._scheduleSimulate(True) def crash(self): """ Implement L{IReactorCore.crash} """ PosixReactorBase.crash(self) if not self._inCFLoop: return CFRunLoopStop(self._cfrunloop) def iterate(self, delay=0): """ Emulate the behavior of C{iterate()} for things that want to call it, by letting the loop run for a little while and then scheduling a timed call to exit it. """ self._started = True # Since the CoreFoundation loop doesn't have the concept of "iterate" # we can't ask it to do this. Instead we will make arrangements to # crash it *very* soon and then make it run. This is a rough # approximation of "an iteration". Using crash and mainLoop here # means that it's safe (as safe as anything using "iterate" can be) to # do this repeatedly. self.callLater(0, self.crash) self.mainLoop() def install(runLoop=None, runner=None): """ Configure the twisted mainloop to be run inside CFRunLoop. @param runLoop: the run loop to use. @param runner: the function to call in order to actually invoke the main loop. This will default to C{CFRunLoopRun} if not specified. However, this is not an appropriate choice for GUI applications, as you need to run NSApplicationMain (or something like it). For example, to run the Twisted mainloop in a PyObjC application, your C{main.py} should look something like this:: from PyObjCTools import AppHelper from twisted.internet.cfreactor import install install(runner=AppHelper.runEventLoop) # initialize your application reactor.run() @return: The installed reactor. @rtype: C{CFReactor} """ reactor = CFReactor(runLoop=runLoop, runner=runner) from twisted.internet.main import installReactor installReactor(reactor) return reactor