2 * Copyright 2016 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
21 #include <sys/types.h>
31 #include <folly/FileUtil.h>
32 #include <folly/io/async/EventBase.h>
33 #include <folly/io/async/EventHandler.h>
34 #include <folly/io/async/DelayedDestruction.h>
35 #include <folly/io/async/Request.h>
36 #include <folly/Likely.h>
37 #include <folly/ScopeGuard.h>
38 #include <folly/SpinLock.h>
40 #include <glog/logging.h>
42 #if __linux__ && !__ANDROID__
43 #define FOLLY_HAVE_EVENTFD
44 #include <folly/io/async/EventFDWrapper.h>
50 * A producer-consumer queue for passing messages between EventBase threads.
52 * Messages can be added to the queue from any thread. Multiple consumers may
53 * listen to the queue from multiple EventBase threads.
55 * A NotificationQueue may not be destroyed while there are still consumers
56 * registered to receive events from the queue. It is the user's
57 * responsibility to ensure that all consumers are unregistered before the
60 * MessageT should be MoveConstructible (i.e., must support either a move
61 * constructor or a copy constructor, or both). Ideally it's move constructor
62 * (or copy constructor if no move constructor is provided) should never throw
63 * exceptions. If the constructor may throw, the consumers could end up
64 * spinning trying to move a message off the queue and failing, and then
67 template<typename MessageT>
68 class NotificationQueue {
71 * A callback interface for consuming messages from the queue as they arrive.
73 class Consumer : public DelayedDestruction, private EventHandler {
75 enum : uint16_t { kDefaultMaxReadAtOnce = 10 };
79 destroyedFlagPtr_(nullptr),
80 maxReadAtOnce_(kDefaultMaxReadAtOnce) {}
82 // create a consumer in-place, without the need to build new class
83 template <typename TCallback>
84 static std::unique_ptr<Consumer, DelayedDestruction::Destructor> make(
85 TCallback&& callback);
88 * messageAvailable() will be invoked whenever a new
89 * message is available from the pipe.
91 virtual void messageAvailable(MessageT&& message) = 0;
94 * Begin consuming messages from the specified queue.
96 * messageAvailable() will be called whenever a message is available. This
97 * consumer will continue to consume messages until stopConsuming() is
100 * A Consumer may only consume messages from a single NotificationQueue at
101 * a time. startConsuming() should not be called if this consumer is
104 void startConsuming(EventBase* eventBase, NotificationQueue* queue) {
105 init(eventBase, queue);
106 registerHandler(READ | PERSIST);
110 * Same as above but registers this event handler as internal so that it
111 * doesn't count towards the pending reader count for the IOLoop.
113 void startConsumingInternal(
114 EventBase* eventBase, NotificationQueue* queue) {
115 init(eventBase, queue);
116 registerInternalHandler(READ | PERSIST);
120 * Stop consuming messages.
122 * startConsuming() may be called again to resume consumption of messages
123 * at a later point in time.
125 void stopConsuming();
128 * Consume messages off the queue until it is empty. No messages may be
129 * added to the queue while it is draining, so that the process is bounded.
130 * To that end, putMessage/tryPutMessage will throw an std::runtime_error,
131 * and tryPutMessageNoThrow will return false.
133 * @returns true if the queue was drained, false otherwise. In practice,
134 * this will only fail if someone else is already draining the queue.
136 bool consumeUntilDrained(size_t* numConsumed = nullptr) noexcept;
139 * Get the NotificationQueue that this consumer is currently consuming
140 * messages from. Returns nullptr if the consumer is not currently
141 * consuming events from any queue.
143 NotificationQueue* getCurrentQueue() const {
148 * Set a limit on how many messages this consumer will read each iteration
149 * around the event loop.
151 * This helps rate-limit how much work the Consumer will do each event loop
152 * iteration, to prevent it from starving other event handlers.
154 * A limit of 0 means no limit will be enforced. If unset, the limit
155 * defaults to kDefaultMaxReadAtOnce (defined to 10 above).
157 void setMaxReadAtOnce(uint32_t maxAtOnce) {
158 maxReadAtOnce_ = maxAtOnce;
160 uint32_t getMaxReadAtOnce() const {
161 return maxReadAtOnce_;
164 EventBase* getEventBase() {
168 void handlerReady(uint16_t events) noexcept override;
172 void destroy() override;
174 virtual ~Consumer() {}
178 * Consume messages off the the queue until
179 * - the queue is empty (1), or
180 * - until the consumer is destroyed, or
181 * - until the consumer is uninstalled, or
182 * - an exception is thrown in the course of dequeueing, or
183 * - unless isDrain is true, until the maxReadAtOnce_ limit is hit
185 * (1) Well, maybe. See logic/comments around "wasEmpty" in implementation.
187 void consumeMessages(bool isDrain, size_t* numConsumed = nullptr) noexcept;
189 void setActive(bool active, bool shouldLock = false) {
195 queue_->spinlock_.lock();
197 if (!active_ && active) {
198 ++queue_->numActiveConsumers_;
199 } else if (active_ && !active) {
200 --queue_->numActiveConsumers_;
204 queue_->spinlock_.unlock();
207 void init(EventBase* eventBase, NotificationQueue* queue);
209 NotificationQueue* queue_;
210 bool* destroyedFlagPtr_;
211 uint32_t maxReadAtOnce_;
218 #ifdef FOLLY_HAVE_EVENTFD
224 * Create a new NotificationQueue.
226 * If the maxSize parameter is specified, this sets the maximum queue size
227 * that will be enforced by tryPutMessage(). (This size is advisory, and may
228 * be exceeded if producers explicitly use putMessage() instead of
231 * The fdType parameter determines the type of file descriptor used
232 * internally to signal message availability. The default (eventfd) is
233 * preferable for performance and because it won't fail when the queue gets
234 * too long. It is not available on on older and non-linux kernels, however.
235 * In this case the code will fall back to using a pipe, the parameter is
236 * mostly for testing purposes.
238 explicit NotificationQueue(uint32_t maxSize = 0,
239 #ifdef FOLLY_HAVE_EVENTFD
240 FdType fdType = FdType::EVENTFD)
242 FdType fdType = FdType::PIPE)
246 advisoryMaxQueueSize_(maxSize),
247 pid_(pid_t(getpid())),
250 RequestContext::saveContext();
252 #ifdef FOLLY_HAVE_EVENTFD
253 if (fdType == FdType::EVENTFD) {
254 eventfd_ = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
255 if (eventfd_ == -1) {
256 if (errno == ENOSYS || errno == EINVAL) {
257 // eventfd not availalble
258 LOG(ERROR) << "failed to create eventfd for NotificationQueue: "
259 << errno << ", falling back to pipe mode (is your kernel "
261 fdType = FdType::PIPE;
264 folly::throwSystemError("Failed to create eventfd for "
265 "NotificationQueue", errno);
270 if (fdType == FdType::PIPE) {
271 if (pipe(pipeFds_)) {
272 folly::throwSystemError("Failed to create pipe for NotificationQueue",
276 // put both ends of the pipe into non-blocking mode
277 if (fcntl(pipeFds_[0], F_SETFL, O_RDONLY | O_NONBLOCK) != 0) {
278 folly::throwSystemError("failed to put NotificationQueue pipe read "
279 "endpoint into non-blocking mode", errno);
281 if (fcntl(pipeFds_[1], F_SETFL, O_WRONLY | O_NONBLOCK) != 0) {
282 folly::throwSystemError("failed to put NotificationQueue pipe write "
283 "endpoint into non-blocking mode", errno);
286 ::close(pipeFds_[0]);
287 ::close(pipeFds_[1]);
293 ~NotificationQueue() {
298 if (pipeFds_[0] >= 0) {
299 ::close(pipeFds_[0]);
302 if (pipeFds_[1] >= 0) {
303 ::close(pipeFds_[1]);
309 * Set the advisory maximum queue size.
311 * This maximum queue size affects calls to tryPutMessage(). Message
312 * producers can still use the putMessage() call to unconditionally put a
313 * message on the queue, ignoring the configured maximum queue size. This
314 * can cause the queue size to exceed the configured maximum.
316 void setMaxQueueSize(uint32_t max) {
317 advisoryMaxQueueSize_ = max;
321 * Attempt to put a message on the queue if the queue is not already full.
323 * If the queue is full, a std::overflow_error will be thrown. The
324 * setMaxQueueSize() function controls the maximum queue size.
326 * If the queue is currently draining, an std::runtime_error will be thrown.
328 * This method may contend briefly on a spinlock if many threads are
329 * concurrently accessing the queue, but for all intents and purposes it will
330 * immediately place the message on the queue and return.
332 * tryPutMessage() may throw std::bad_alloc if memory allocation fails, and
333 * may throw any other exception thrown by the MessageT move/copy
336 void tryPutMessage(MessageT&& message) {
337 putMessageImpl(std::move(message), advisoryMaxQueueSize_);
339 void tryPutMessage(const MessageT& message) {
340 putMessageImpl(message, advisoryMaxQueueSize_);
344 * No-throw versions of the above. Instead returns true on success, false on
347 * Only std::overflow_error (the common exception case) and std::runtime_error
348 * (which indicates that the queue is being drained) are prevented from being
349 * thrown. User code must still catch std::bad_alloc errors.
351 bool tryPutMessageNoThrow(MessageT&& message) {
352 return putMessageImpl(std::move(message), advisoryMaxQueueSize_, false);
354 bool tryPutMessageNoThrow(const MessageT& message) {
355 return putMessageImpl(message, advisoryMaxQueueSize_, false);
359 * Unconditionally put a message on the queue.
361 * This method is like tryPutMessage(), but ignores the maximum queue size
362 * and always puts the message on the queue, even if the maximum queue size
365 * putMessage() may throw
366 * - std::bad_alloc if memory allocation fails, and may
367 * - std::runtime_error if the queue is currently draining
368 * - any other exception thrown by the MessageT move/copy constructor.
370 void putMessage(MessageT&& message) {
371 putMessageImpl(std::move(message), 0);
373 void putMessage(const MessageT& message) {
374 putMessageImpl(message, 0);
378 * Put several messages on the queue.
380 template<typename InputIteratorT>
381 void putMessages(InputIteratorT first, InputIteratorT last) {
382 typedef typename std::iterator_traits<InputIteratorT>::iterator_category
384 putMessagesImpl(first, last, IterCategory());
388 * Try to immediately pull a message off of the queue, without blocking.
390 * If a message is immediately available, the result parameter will be
391 * updated to contain the message contents and true will be returned.
393 * If no message is available, false will be returned and result will be left
396 bool tryConsume(MessageT& result) {
397 SCOPE_EXIT { syncSignalAndQueue(); };
401 folly::SpinLockGuard g(spinlock_);
403 if (UNLIKELY(queue_.empty())) {
407 auto data = std::move(queue_.front());
409 RequestContext::setContext(data.second);
416 size_t size() const {
417 folly::SpinLockGuard g(spinlock_);
418 return queue_.size();
422 * Check that the NotificationQueue is being used from the correct process.
424 * If you create a NotificationQueue in one process, then fork, and try to
425 * send messages to the queue from the child process, you're going to have a
426 * bad time. Unfortunately users have (accidentally) run into this.
428 * Because we use an eventfd/pipe, the child process can actually signal the
429 * parent process that an event is ready. However, it can't put anything on
430 * the parent's queue, so the parent wakes up and finds an empty queue. This
431 * check ensures that we catch the problem in the misbehaving child process
432 * code, and crash before signalling the parent process.
434 void checkPid() const { CHECK_EQ(pid_, pid_t(getpid())); }
437 // Forbidden copy constructor and assignment operator
438 NotificationQueue(NotificationQueue const &) = delete;
439 NotificationQueue& operator=(NotificationQueue const &) = delete;
441 inline bool checkQueueSize(size_t maxSize, bool throws=true) const {
442 DCHECK(0 == spinlock_.trylock());
443 if (maxSize > 0 && queue_.size() >= maxSize) {
445 throw std::overflow_error("unable to add message to NotificationQueue: "
453 inline bool checkDraining(bool throws=true) {
454 if (UNLIKELY(draining_ && throws)) {
455 throw std::runtime_error("queue is draining, cannot add message");
461 // TODO 10860938 Remove after figuring out crash
462 mutable std::atomic<int> eventBytes_{0};
463 mutable std::atomic<int> maxEventBytes_{0};
466 void ensureSignalLocked() const {
467 // semantics: empty fd == empty queue <=> !signal_
472 ssize_t bytes_written = 0;
473 ssize_t bytes_expected = 0;
477 // eventfd(2) dictates that we must write a 64-bit integer
479 bytes_expected = static_cast<ssize_t>(sizeof(signal));
480 bytes_written = ::write(eventfd_, &signal, bytes_expected);
483 bytes_expected = static_cast<ssize_t>(sizeof(signal));
484 bytes_written = ::write(pipeFds_[1], &signal, bytes_expected);
486 } while (bytes_written == -1 && errno == EINTR);
489 if (bytes_written > 0) {
490 eventBytes_ += bytes_written;
491 maxEventBytes_ = std::max((int)maxEventBytes_, (int)eventBytes_);
495 if (bytes_written == bytes_expected) {
499 LOG(ERROR) << "NotificationQueue Write Error=" << errno
500 << " bytesInPipe=" << eventBytes_
501 << " maxInPipe=" << maxEventBytes_ << " queue=" << size();
503 folly::throwSystemError("failed to signal NotificationQueue after "
508 void drainSignalsLocked() {
509 ssize_t bytes_read = 0;
512 bytes_read = readNoInt(eventfd_, &message, sizeof(message));
513 CHECK(bytes_read != -1 || errno == EAGAIN);
515 // There should only be one byte in the pipe. To avoid potential leaks we still drain.
518 while ((result = readNoInt(pipeFds_[0], &message, sizeof(message))) != -1) {
519 bytes_read += result;
521 CHECK(result == -1 && errno == EAGAIN);
522 LOG_IF(ERROR, bytes_read > 1)
523 << "[NotificationQueue] Unexpected state while draining pipe: bytes_read="
524 << bytes_read << " bytes, expected <= 1";
526 LOG_IF(ERROR, (signal_ && bytes_read == 0) || (!signal_ && bytes_read > 0))
527 << "[NotificationQueue] Unexpected state while draining signals: signal_="
528 << signal_ << " bytes_read=" << bytes_read;
533 if (bytes_read > 0) {
534 eventBytes_ -= bytes_read;
539 void ensureSignal() const {
540 folly::SpinLockGuard g(spinlock_);
541 ensureSignalLocked();
544 void syncSignalAndQueue() {
545 folly::SpinLockGuard g(spinlock_);
547 if (queue_.empty()) {
548 drainSignalsLocked();
550 ensureSignalLocked();
554 bool putMessageImpl(MessageT&& message, size_t maxSize, bool throws=true) {
558 folly::SpinLockGuard g(spinlock_);
559 if (checkDraining(throws) || !checkQueueSize(maxSize, throws)) {
562 // We only need to signal an event if not all consumers are
564 if (numActiveConsumers_ < numConsumers_) {
567 queue_.emplace_back(std::move(message), RequestContext::saveContext());
569 ensureSignalLocked();
576 const MessageT& message, size_t maxSize, bool throws=true) {
580 folly::SpinLockGuard g(spinlock_);
581 if (checkDraining(throws) || !checkQueueSize(maxSize, throws)) {
584 if (numActiveConsumers_ < numConsumers_) {
587 queue_.emplace_back(message, RequestContext::saveContext());
589 ensureSignalLocked();
595 template<typename InputIteratorT>
596 void putMessagesImpl(InputIteratorT first, InputIteratorT last,
597 std::input_iterator_tag) {
602 folly::SpinLockGuard g(spinlock_);
604 while (first != last) {
605 queue_.emplace_back(*first, RequestContext::saveContext());
609 if (numActiveConsumers_ < numConsumers_) {
613 ensureSignalLocked();
618 mutable folly::SpinLock spinlock_;
619 mutable bool signal_{false};
621 int pipeFds_[2]; // to fallback to on older/non-linux systems
622 uint32_t advisoryMaxQueueSize_;
624 std::deque<std::pair<MessageT, std::shared_ptr<RequestContext>>> queue_;
625 int numConsumers_{0};
626 std::atomic<int> numActiveConsumers_{0};
627 bool draining_{false};
630 template<typename MessageT>
631 void NotificationQueue<MessageT>::Consumer::destroy() {
632 // If we are in the middle of a call to handlerReady(), destroyedFlagPtr_
633 // will be non-nullptr. Mark the value that it points to, so that
634 // handlerReady() will know the callback is destroyed, and that it cannot
635 // access any member variables anymore.
636 if (destroyedFlagPtr_) {
637 *destroyedFlagPtr_ = true;
640 DelayedDestruction::destroy();
643 template<typename MessageT>
644 void NotificationQueue<MessageT>::Consumer::handlerReady(uint16_t /*events*/)
646 consumeMessages(false);
649 template<typename MessageT>
650 void NotificationQueue<MessageT>::Consumer::consumeMessages(
651 bool isDrain, size_t* numConsumed) noexcept {
652 DestructorGuard dg(this);
653 uint32_t numProcessed = 0;
657 queue_->syncSignalAndQueue();
660 SCOPE_EXIT { setActive(false, /* shouldLock = */ true); };
662 if (numConsumed != nullptr) {
663 *numConsumed = numProcessed;
667 // Now pop the message off of the queue.
669 // We have to manually acquire and release the spinlock here, rather than
670 // using SpinLockHolder since the MessageT has to be constructed while
671 // holding the spinlock and available after we release it. SpinLockHolder
672 // unfortunately doesn't provide a release() method. (We can't construct
673 // MessageT first since we have no guarantee that MessageT has a default
675 queue_->spinlock_.lock();
679 if (UNLIKELY(queue_->queue_.empty())) {
680 // If there is no message, we've reached the end of the queue, return.
682 queue_->spinlock_.unlock();
686 // Pull a message off the queue.
687 auto& data = queue_->queue_.front();
689 MessageT msg(std::move(data.first));
691 RequestContext::setContext(data.second);
692 queue_->queue_.pop_front();
694 // Check to see if the queue is empty now.
695 // We use this as an optimization to see if we should bother trying to
696 // loop again and read another message after invoking this callback.
697 bool wasEmpty = queue_->queue_.empty();
702 // Now unlock the spinlock before we invoke the callback.
703 queue_->spinlock_.unlock();
707 bool callbackDestroyed = false;
708 CHECK(destroyedFlagPtr_ == nullptr);
709 destroyedFlagPtr_ = &callbackDestroyed;
710 messageAvailable(std::move(msg));
711 destroyedFlagPtr_ = nullptr;
713 RequestContext::setContext(old_ctx);
715 // If the callback was destroyed before it returned, we are done
716 if (callbackDestroyed) {
720 // If the callback is no longer installed, we are done.
721 if (queue_ == nullptr) {
725 // If we have hit maxReadAtOnce_, we are done.
727 if (!isDrain && maxReadAtOnce_ > 0 &&
728 numProcessed >= maxReadAtOnce_) {
732 // If the queue was empty before we invoked the callback, it's probable
733 // that it is still empty now. Just go ahead and return, rather than
734 // looping again and trying to re-read from the eventfd. (If a new
735 // message had in fact arrived while we were invoking the callback, we
736 // will simply be woken up the next time around the event loop and will
737 // process the message then.)
741 } catch (const std::exception& ex) {
742 // This catch block is really just to handle the case where the MessageT
743 // constructor throws. The messageAvailable() callback itself is
744 // declared as noexcept and should never throw.
746 // If the MessageT constructor does throw we try to handle it as best as
747 // we can, but we can't work miracles. We will just ignore the error for
748 // now and return. The next time around the event loop we will end up
749 // trying to read the message again. If MessageT continues to throw we
750 // will never make forward progress and will keep trying each time around
753 // Unlock the spinlock.
754 queue_->spinlock_.unlock();
762 template<typename MessageT>
763 void NotificationQueue<MessageT>::Consumer::init(
764 EventBase* eventBase,
765 NotificationQueue* queue) {
766 assert(eventBase->isInEventBaseThread());
767 assert(queue_ == nullptr);
768 assert(!isHandlerRegistered());
776 folly::SpinLockGuard g(queue_->spinlock_);
777 queue_->numConsumers_++;
779 queue_->ensureSignal();
781 if (queue_->eventfd_ >= 0) {
782 initHandler(eventBase, queue_->eventfd_);
784 initHandler(eventBase, queue_->pipeFds_[0]);
788 template<typename MessageT>
789 void NotificationQueue<MessageT>::Consumer::stopConsuming() {
790 if (queue_ == nullptr) {
791 assert(!isHandlerRegistered());
796 folly::SpinLockGuard g(queue_->spinlock_);
797 queue_->numConsumers_--;
801 assert(isHandlerRegistered());
807 template<typename MessageT>
808 bool NotificationQueue<MessageT>::Consumer::consumeUntilDrained(
809 size_t* numConsumed) noexcept {
810 DestructorGuard dg(this);
812 folly::SpinLockGuard g(queue_->spinlock_);
813 if (queue_->draining_) {
816 queue_->draining_ = true;
818 consumeMessages(true, numConsumed);
820 folly::SpinLockGuard g(queue_->spinlock_);
821 queue_->draining_ = false;
827 * Creates a NotificationQueue::Consumer wrapping a function object
828 * Modeled after AsyncTimeout::make
834 template <typename MessageT, typename TCallback>
835 struct notification_queue_consumer_wrapper
836 : public NotificationQueue<MessageT>::Consumer {
838 template <typename UCallback>
839 explicit notification_queue_consumer_wrapper(UCallback&& callback)
840 : callback_(std::forward<UCallback>(callback)) {}
842 // we are being stricter here and requiring noexcept for callback
843 void messageAvailable(MessageT&& message) override {
845 noexcept(std::declval<TCallback>()(std::forward<MessageT>(message))),
846 "callback must be declared noexcept, e.g.: `[]() noexcept {}`"
849 callback_(std::forward<MessageT>(message));
856 } // namespace detail
858 template <typename MessageT>
859 template <typename TCallback>
860 std::unique_ptr<typename NotificationQueue<MessageT>::Consumer,
861 DelayedDestruction::Destructor>
862 NotificationQueue<MessageT>::Consumer::make(TCallback&& callback) {
863 return std::unique_ptr<NotificationQueue<MessageT>::Consumer,
864 DelayedDestruction::Destructor>(
865 new detail::notification_queue_consumer_wrapper<
867 typename std::decay<TCallback>::type>(
868 std::forward<TCallback>(callback)));