2 * Copyright 2017 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.
22 #include <boost/noncopyable.hpp>
25 #include <type_traits>
27 #include <folly/Traits.h>
28 #include <folly/detail/CacheLocality.h>
29 #include <folly/detail/TurnSequencer.h>
30 #include <folly/portability/Unistd.h>
36 template<typename T, template<typename> class Atom>
37 struct SingleElementQueue;
39 template <typename T> class MPMCPipelineStageImpl;
41 /// MPMCQueue base CRTP template
42 template <typename> class MPMCQueueBase;
46 /// MPMCQueue<T> is a high-performance bounded concurrent queue that
47 /// supports multiple producers, multiple consumers, and optional blocking.
48 /// The queue has a fixed capacity, for which all memory will be allocated
49 /// up front. The bulk of the work of enqueuing and dequeuing can be
50 /// performed in parallel.
52 /// MPMCQueue is linearizable. That means that if a call to write(A)
53 /// returns before a call to write(B) begins, then A will definitely end up
54 /// in the queue before B, and if a call to read(X) returns before a call
55 /// to read(Y) is started, that X will be something from earlier in the
56 /// queue than Y. This also means that if a read call returns a value, you
57 /// can be sure that all previous elements of the queue have been assigned
58 /// a reader (that reader might not yet have returned, but it exists).
60 /// The underlying implementation uses a ticket dispenser for the head and
61 /// the tail, spreading accesses across N single-element queues to produce
62 /// a queue with capacity N. The ticket dispensers use atomic increment,
63 /// which is more robust to contention than a CAS loop. Each of the
64 /// single-element queues uses its own CAS to serialize access, with an
65 /// adaptive spin cutoff. When spinning fails on a single-element queue
66 /// it uses futex()'s _BITSET operations to reduce unnecessary wakeups
67 /// even if multiple waiters are present on an individual queue (such as
68 /// when the MPMCQueue's capacity is smaller than the number of enqueuers
71 /// In benchmarks (contained in tao/queues/ConcurrentQueueTests)
72 /// it handles 1 to 1, 1 to N, N to 1, and N to M thread counts better
73 /// than any of the alternatives present in fbcode, for both small (~10)
74 /// and large capacities. In these benchmarks it is also faster than
75 /// tbb::concurrent_bounded_queue for all configurations. When there are
76 /// many more threads than cores, MPMCQueue is _much_ faster than the tbb
77 /// queue because it uses futex() to block and unblock waiting threads,
78 /// rather than spinning with sched_yield.
80 /// NOEXCEPT INTERACTION: tl;dr; If it compiles you're fine. Ticket-based
81 /// queues separate the assignment of queue positions from the actual
82 /// construction of the in-queue elements, which means that the T
83 /// constructor used during enqueue must not throw an exception. This is
84 /// enforced at compile time using type traits, which requires that T be
85 /// adorned with accurate noexcept information. If your type does not
86 /// use noexcept, you will have to wrap it in something that provides
87 /// the guarantee. We provide an alternate safe implementation for types
88 /// that don't use noexcept but that are marked folly::IsRelocatable
89 /// and std::is_nothrow_constructible, which is common for folly types.
90 /// In particular, if you can declare FOLLY_ASSUME_FBVECTOR_COMPATIBLE
91 /// then your type can be put in MPMCQueue.
93 /// If you have a pool of N queue consumers that you want to shut down
94 /// after the queue has drained, one way is to enqueue N sentinel values
95 /// to the queue. If the producer doesn't know how many consumers there
96 /// are you can enqueue one sentinel and then have each consumer requeue
97 /// two sentinels after it receives it (by requeuing 2 the shutdown can
98 /// complete in O(log P) time instead of O(P)).
99 template<typename T, template<typename> class Atom = std::atomic,
100 bool Dynamic = false>
101 class MPMCQueue : public detail::MPMCQueueBase<MPMCQueue<T,Atom,Dynamic>> {
102 friend class detail::MPMCPipelineStageImpl<T>;
103 using Slot = detail::SingleElementQueue<T,Atom>;
106 explicit MPMCQueue(size_t queueCapacity)
107 : detail::MPMCQueueBase<MPMCQueue<T,Atom,Dynamic>>(queueCapacity)
109 this->stride_ = this->computeStride(queueCapacity);
110 this->slots_ = new Slot[queueCapacity + 2 * this->kSlotPadding];
113 MPMCQueue() noexcept { }
116 /// The dynamic version of MPMCQueue allows dynamic expansion of queue
117 /// capacity, such that a queue may start with a smaller capacity than
118 /// specified and expand only if needed. Users may optionally specify
119 /// the initial capacity and the expansion multiplier.
121 /// The design uses a seqlock to enforce mutual exclusion among
122 /// expansion attempts. Regular operations read up-to-date queue
123 /// information (slots array, capacity, stride) inside read-only
124 /// seqlock sections, which are unimpeded when no expansion is in
127 /// An expansion computes a new capacity, allocates a new slots array,
128 /// and updates stride. No information needs to be copied from the
129 /// current slots array to the new one. When this happens, new slots
130 /// will not have sequence numbers that match ticket numbers. The
131 /// expansion needs to compute a ticket offset such that operations
132 /// that use new arrays can adjust the calculations of slot indexes
133 /// and sequence numbers that take into account that the new slots
134 /// start with sequence numbers of zero. The current ticket offset is
135 /// packed with the seqlock in an atomic 64-bit integer. The initial
138 /// Lagging write and read operations with tickets lower than the
139 /// ticket offset of the current slots array (i.e., the minimum ticket
140 /// number that can be served by the current array) must use earlier
141 /// closed arrays instead of the current one. Information about closed
142 /// slots arrays (array address, capacity, stride, and offset) is
143 /// maintained in a logarithmic-sized structure. Each entry in that
144 /// structure never need to be changed once set. The number of closed
145 /// arrays is half the value of the seqlock (when unlocked).
147 /// The acquisition of the seqlock to perform an expansion does not
148 /// prevent the issuing of new push and pop tickets concurrently. The
149 /// expansion must set the new ticket offset to a value that couldn't
150 /// have been issued to an operation that has already gone through a
151 /// seqlock read-only section (and hence obtained information for
152 /// older closed arrays).
154 /// Note that the total queue capacity can temporarily exceed the
155 /// specified capacity when there are lagging consumers that haven't
156 /// yet consumed all the elements in closed arrays. Users should not
157 /// rely on the capacity of dynamic queues for synchronization, e.g.,
158 /// they should not expect that a thread will definitely block on a
159 /// call to blockingWrite() when the queue size is known to be equal
162 /// Note that some writeIfNotFull() and tryWriteUntil() operations may
163 /// fail even if the size of the queue is less than its maximum
164 /// capacity and despite the success of expansion, if the operation
165 /// happens to acquire a ticket that belongs to a closed array. This
166 /// is a transient condition. Typically, one or two ticket values may
167 /// be subject to such condition per expansion.
169 /// The dynamic version is a partial specialization of MPMCQueue with
171 template <typename T, template<typename> class Atom>
172 class MPMCQueue<T,Atom,true> :
173 public detail::MPMCQueueBase<MPMCQueue<T,Atom,true>> {
174 friend class detail::MPMCQueueBase<MPMCQueue<T,Atom,true>>;
175 using Slot = detail::SingleElementQueue<T,Atom>;
178 uint64_t offset_ {0};
179 Slot* slots_ {nullptr};
180 size_t capacity_ {0};
186 explicit MPMCQueue(size_t queueCapacity)
187 : detail::MPMCQueueBase<MPMCQueue<T,Atom,true>>(queueCapacity)
189 size_t cap = std::min<size_t>(kDefaultMinDynamicCapacity, queueCapacity);
190 initQueue(cap, kDefaultExpansionMultiplier);
193 explicit MPMCQueue(size_t queueCapacity,
195 size_t expansionMultiplier)
196 : detail::MPMCQueueBase<MPMCQueue<T,Atom,true>>(queueCapacity)
198 minCapacity = std::max<size_t>(1, minCapacity);
199 size_t cap = std::min<size_t>(minCapacity, queueCapacity);
200 expansionMultiplier = std::max<size_t>(2, expansionMultiplier);
201 initQueue(cap, expansionMultiplier);
204 MPMCQueue() noexcept {
209 MPMCQueue(MPMCQueue<T,Atom,true>&& rhs) noexcept {
210 this->capacity_ = rhs.capacity_;
211 this->slots_ = rhs.slots_;
212 this->stride_ = rhs.stride_;
213 this->dstate_.store(rhs.dstate_.load(std::memory_order_relaxed),
214 std::memory_order_relaxed);
215 this->dcapacity_.store(rhs.dcapacity_.load(std::memory_order_relaxed),
216 std::memory_order_relaxed);
217 this->pushTicket_.store(rhs.pushTicket_.load(std::memory_order_relaxed),
218 std::memory_order_relaxed);
219 this->popTicket_.store(rhs.popTicket_.load(std::memory_order_relaxed),
220 std::memory_order_relaxed);
221 this->pushSpinCutoff_.store(
222 rhs.pushSpinCutoff_.load(std::memory_order_relaxed),
223 std::memory_order_relaxed);
224 this->popSpinCutoff_.store(
225 rhs.popSpinCutoff_.load(std::memory_order_relaxed),
226 std::memory_order_relaxed);
228 closed_ = rhs.closed_;
231 rhs.slots_ = nullptr;
233 rhs.dstate_.store(0, std::memory_order_relaxed);
234 rhs.dcapacity_.store(0, std::memory_order_relaxed);
235 rhs.pushTicket_.store(0, std::memory_order_relaxed);
236 rhs.popTicket_.store(0, std::memory_order_relaxed);
237 rhs.pushSpinCutoff_.store(0, std::memory_order_relaxed);
238 rhs.popSpinCutoff_.store(0, std::memory_order_relaxed);
240 rhs.closed_ = nullptr;
243 MPMCQueue<T,Atom, true> const& operator= (MPMCQueue<T,Atom, true>&& rhs) {
246 new (this) MPMCQueue(std::move(rhs));
252 if (closed_ != nullptr) {
253 for (int i = getNumClosed(this->dstate_.load()) - 1; i >= 0; --i) {
254 delete[] closed_[i].slots_;
260 size_t allocatedCapacity() const noexcept {
261 return this->dcapacity_.load(std::memory_order_relaxed);
264 template <typename ...Args>
265 void blockingWrite(Args&&... args) noexcept {
266 uint64_t ticket = this->pushTicket_++;
273 if (!trySeqlockReadSection(state, slots, cap, stride)) {
274 asm_volatile_pause();
277 if (maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride)) {
278 // There was an expansion after this ticket was issued.
281 if (slots[this->idx((ticket-offset), cap, stride)]
282 .mayEnqueue(this->turn(ticket-offset, cap))) {
283 // A slot is ready. No need to expand.
285 } else if (this->popTicket_.load(std::memory_order_relaxed) + cap
287 // May block, but a pop is in progress. No need to expand.
288 // Get seqlock read section info again in case an expansion
289 // occurred with an equal or higher ticket.
292 // May block. See if we can expand.
293 if (tryExpand(state, cap)) {
294 // This or another thread started an expansion. Get updated info.
302 this->enqueueWithTicketBase(ticket-offset, slots, cap, stride,
303 std::forward<Args>(args)...);
306 void blockingReadWithTicket(uint64_t& ticket, T& elem) noexcept {
307 ticket = this->popTicket_++;
313 while (!trySeqlockReadSection(state, slots, cap, stride)) {
314 asm_volatile_pause();
316 // If there was an expansion after the corresponding push ticket
317 // was issued, adjust accordingly
318 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
319 this->dequeueWithTicketBase(ticket-offset, slots, cap, stride, elem);
326 kDefaultMinDynamicCapacity = 10,
327 kDefaultExpansionMultiplier = 10,
332 // Info about closed slots arrays for use by lagging operations
333 ClosedArray* closed_;
335 void initQueue(const size_t cap, const size_t mult) {
336 this->stride_ = this->computeStride(cap);
337 this->slots_ = new Slot[cap + 2 * this->kSlotPadding];
338 this->dstate_.store(0);
339 this->dcapacity_.store(cap);
341 size_t maxClosed = 0;
342 for (size_t expanded = cap;
343 expanded < this->capacity_;
347 closed_ = (maxClosed > 0) ? new ClosedArray[maxClosed] : nullptr;
350 bool tryObtainReadyPushTicket(
351 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
355 ticket = this->pushTicket_.load(std::memory_order_acquire); // A
356 if (!trySeqlockReadSection(state, slots, cap, stride)) {
357 asm_volatile_pause();
361 // If there was an expansion with offset greater than this ticket,
362 // adjust accordingly
364 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
366 if (slots[this->idx((ticket-offset), cap, stride)]
367 .mayEnqueue(this->turn(ticket-offset, cap))) {
369 if (this->pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
377 if (ticket != this->pushTicket_.load(std::memory_order_relaxed)) { // B
378 // Try again. Ticket changed.
382 // Try to expand unless the ticket is for a closed array
383 if (offset == getOffset(state)) {
384 if (tryExpand(state, cap)) {
385 // This or another thread started an expansion. Get up-to-date info.
394 bool tryObtainPromisedPushTicket(
395 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
399 ticket = this->pushTicket_.load(std::memory_order_acquire);
400 auto numPops = this->popTicket_.load(std::memory_order_acquire);
401 if (!trySeqlockReadSection(state, slots, cap, stride)) {
402 asm_volatile_pause();
406 const auto curCap = cap;
407 // If there was an expansion with offset greater than this ticket,
408 // adjust accordingly
410 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
412 int64_t n = ticket - numPops;
414 if (n >= static_cast<ssize_t>(cap)) {
415 if ((cap == curCap) && tryExpand(state, cap)) {
416 // This or another thread started an expansion. Start over.
424 if (this->pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
432 bool tryObtainReadyPopTicket(
433 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
437 ticket = this->popTicket_.load(std::memory_order_relaxed);
438 if (!trySeqlockReadSection(state, slots, cap, stride)) {
439 asm_volatile_pause();
443 // If there was an expansion after the corresponding push ticket
444 // was issued, adjust accordingly
446 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
448 if (slots[this->idx((ticket-offset), cap, stride)]
449 .mayDequeue(this->turn(ticket-offset, cap))) {
450 if (this->popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
461 bool tryObtainPromisedPopTicket(
462 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
466 ticket = this->popTicket_.load(std::memory_order_acquire);
467 auto numPushes = this->pushTicket_.load(std::memory_order_acquire);
468 if (!trySeqlockReadSection(state, slots, cap, stride)) {
469 asm_volatile_pause();
474 // If there was an expansion after the corresponding push
475 // ticket was issued, adjust accordingly
476 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
478 if (ticket >= numPushes) {
482 if (this->popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
489 /// Enqueues an element with a specific ticket number
490 template <typename ...Args>
491 void enqueueWithTicket(const uint64_t ticket, Args&&... args) noexcept {
498 while (!trySeqlockReadSection(state, slots, cap, stride)) {}
500 // If there was an expansion after this ticket was issued, adjust
502 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
504 this->enqueueWithTicketBase(ticket-offset, slots, cap, stride,
505 std::forward<Args>(args)...);
508 uint64_t getOffset(const uint64_t state) const noexcept {
509 return state >> kSeqlockBits;
512 int getNumClosed(const uint64_t state) const noexcept {
513 return (state & ((1 << kSeqlockBits) - 1)) >> 1;
516 /// Try to expand the queue. Returns true if this expansion was
517 /// successful or a concurent expansion is in progress. Returns
518 /// false if the queue has reached its maximum capacity or
519 /// allocation has failed.
520 bool tryExpand(const uint64_t state, const size_t cap) noexcept {
521 if (cap == this->capacity_) {
525 uint64_t oldval = state;
526 assert((state & 1) == 0);
527 if (this->dstate_.compare_exchange_strong(oldval, state + 1)) {
528 assert(cap == this->dcapacity_.load());
529 uint64_t ticket = 1 + std::max(this->pushTicket_.load(),
530 this->popTicket_.load());
532 std::min(dmult_ * cap, this->capacity_);
534 new (std::nothrow) Slot[newCapacity + 2 * this->kSlotPadding];
535 if (newSlots == nullptr) {
536 // Expansion failed. Restore the seqlock
537 this->dstate_.store(state);
540 // Successful expansion
541 // calculate the current ticket offset
542 uint64_t offset = getOffset(state);
543 // calculate index in closed array
544 int index = getNumClosed(state);
545 assert((index << 1) < (1 << kSeqlockBits));
546 // fill the info for the closed slots array
547 closed_[index].offset_ = offset;
548 closed_[index].slots_ = this->dslots_.load();
549 closed_[index].capacity_ = cap;
550 closed_[index].stride_ = this->dstride_.load();
551 // update the new slots array info
552 this->dslots_.store(newSlots);
553 this->dcapacity_.store(newCapacity);
554 this->dstride_.store(this->computeStride(newCapacity));
555 // Release the seqlock and record the new ticket offset
556 this->dstate_.store((ticket << kSeqlockBits) + (2 * (index + 1)));
558 } else { // failed to acquire seqlock
559 // Someone acaquired the seqlock. Go back to the caller and get
565 /// Seqlock read-only section
566 bool trySeqlockReadSection(
567 uint64_t& state, Slot*& slots, size_t& cap, int& stride
569 state = this->dstate_.load(std::memory_order_acquire);
574 // Start read-only section.
575 slots = this->dslots_.load(std::memory_order_relaxed);
576 cap = this->dcapacity_.load(std::memory_order_relaxed);
577 stride = this->dstride_.load(std::memory_order_relaxed);
578 // End of read-only section. Validate seqlock.
579 std::atomic_thread_fence(std::memory_order_acquire);
580 return (state == this->dstate_.load(std::memory_order_relaxed));
583 /// If there was an expansion after ticket was issued, update local variables
584 /// of the lagging operation using the most recent closed array with
585 /// offset <= ticket and return true. Otherwise, return false;
586 bool maybeUpdateFromClosed(
587 const uint64_t state,
588 const uint64_t ticket,
592 int& stride) noexcept {
593 offset = getOffset(state);
594 if (ticket >= offset) {
597 for (int i = getNumClosed(state) - 1; i >= 0; --i) {
598 offset = closed_[i].offset_;
599 if (offset <= ticket) {
600 slots = closed_[i].slots_;
601 cap = closed_[i].capacity_;
602 stride = closed_[i].stride_;
606 // A closed array with offset <= ticket should have been found
614 /// CRTP specialization of MPMCQueueBase
617 typename T, template<typename> class Atom, bool Dynamic> class Derived,
618 typename T, template<typename> class Atom, bool Dynamic>
619 class MPMCQueueBase<Derived<T, Atom, Dynamic>> : boost::noncopyable {
621 // Note: Using CRTP static casts in several functions of this base
622 // template instead of making called functions virtual or duplicating
623 // the code of calling functions in the derived partially specialized
626 static_assert(std::is_nothrow_constructible<T,T&&>::value ||
627 folly::IsRelocatable<T>::value,
628 "T must be relocatable or have a noexcept move constructor");
631 typedef T value_type;
633 using Slot = detail::SingleElementQueue<T,Atom>;
635 explicit MPMCQueueBase(size_t queueCapacity)
636 : capacity_(queueCapacity)
642 if (queueCapacity == 0) {
643 throw std::invalid_argument(
644 "MPMCQueue with explicit capacity 0 is impossible"
645 // Stride computation in derived classes would sigfpe if capacity is 0
649 // ideally this would be a static assert, but g++ doesn't allow it
650 assert(alignof(MPMCQueue<T,Atom>)
651 >= detail::CacheLocality::kFalseSharingRange);
652 assert(static_cast<uint8_t*>(static_cast<void*>(&popTicket_))
653 - static_cast<uint8_t*>(static_cast<void*>(&pushTicket_))
654 >= detail::CacheLocality::kFalseSharingRange);
657 /// A default-constructed queue is useful because a usable (non-zero
658 /// capacity) queue can be moved onto it or swapped with it
659 MPMCQueueBase() noexcept
671 /// IMPORTANT: The move constructor is here to make it easier to perform
672 /// the initialization phase, it is not safe to use when there are any
673 /// concurrent accesses (this is not checked).
674 MPMCQueueBase(MPMCQueueBase<Derived<T,Atom,Dynamic>>&& rhs) noexcept
675 : capacity_(rhs.capacity_)
677 , stride_(rhs.stride_)
678 , dstate_(rhs.dstate_.load(std::memory_order_relaxed))
679 , dcapacity_(rhs.dcapacity_.load(std::memory_order_relaxed))
680 , pushTicket_(rhs.pushTicket_.load(std::memory_order_relaxed))
681 , popTicket_(rhs.popTicket_.load(std::memory_order_relaxed))
682 , pushSpinCutoff_(rhs.pushSpinCutoff_.load(std::memory_order_relaxed))
683 , popSpinCutoff_(rhs.popSpinCutoff_.load(std::memory_order_relaxed))
685 // relaxed ops are okay for the previous reads, since rhs queue can't
686 // be in concurrent use
690 rhs.slots_ = nullptr;
692 rhs.dstate_.store(0, std::memory_order_relaxed);
693 rhs.dcapacity_.store(0, std::memory_order_relaxed);
694 rhs.pushTicket_.store(0, std::memory_order_relaxed);
695 rhs.popTicket_.store(0, std::memory_order_relaxed);
696 rhs.pushSpinCutoff_.store(0, std::memory_order_relaxed);
697 rhs.popSpinCutoff_.store(0, std::memory_order_relaxed);
700 /// IMPORTANT: The move operator is here to make it easier to perform
701 /// the initialization phase, it is not safe to use when there are any
702 /// concurrent accesses (this is not checked).
703 MPMCQueueBase<Derived<T,Atom,Dynamic>> const& operator=
704 (MPMCQueueBase<Derived<T,Atom,Dynamic>>&& rhs) {
706 this->~MPMCQueueBase();
707 new (this) MPMCQueueBase(std::move(rhs));
712 /// MPMCQueue can only be safely destroyed when there are no
713 /// pending enqueuers or dequeuers (this is not checked).
718 /// Returns the number of writes (including threads that are blocked waiting
719 /// to write) minus the number of reads (including threads that are blocked
720 /// waiting to read). So effectively, it becomes:
721 /// elements in queue + pending(calls to write) - pending(calls to read).
722 /// If nothing is pending, then the method returns the actual number of
723 /// elements in the queue.
724 /// The returned value can be negative if there are no writers and the queue
725 /// is empty, but there is one reader that is blocked waiting to read (in
726 /// which case, the returned size will be -1).
727 ssize_t size() const noexcept {
728 // since both pushes and pops increase monotonically, we can get a
729 // consistent snapshot either by bracketing a read of popTicket_ with
730 // two reads of pushTicket_ that return the same value, or the other
731 // way around. We maximize our chances by alternately attempting
733 uint64_t pushes = pushTicket_.load(std::memory_order_acquire); // A
734 uint64_t pops = popTicket_.load(std::memory_order_acquire); // B
736 uint64_t nextPushes = pushTicket_.load(std::memory_order_acquire); // C
737 if (pushes == nextPushes) {
738 // pushTicket_ didn't change from A (or the previous C) to C,
739 // so we can linearize at B (or D)
740 return ssize_t(pushes - pops);
743 uint64_t nextPops = popTicket_.load(std::memory_order_acquire); // D
744 if (pops == nextPops) {
745 // popTicket_ didn't chance from B (or the previous D), so we
746 // can linearize at C
747 return ssize_t(pushes - pops);
753 /// Returns true if there are no items available for dequeue
754 bool isEmpty() const noexcept {
758 /// Returns true if there is currently no empty space to enqueue
759 bool isFull() const noexcept {
760 // careful with signed -> unsigned promotion, since size can be negative
761 return size() >= static_cast<ssize_t>(capacity_);
764 /// Returns is a guess at size() for contexts that don't need a precise
765 /// value, such as stats. More specifically, it returns the number of writes
766 /// minus the number of reads, but after reading the number of writes, more
767 /// writers could have came before the number of reads was sampled,
768 /// and this method doesn't protect against such case.
769 /// The returned value can be negative.
770 ssize_t sizeGuess() const noexcept {
771 return writeCount() - readCount();
775 size_t capacity() const noexcept {
779 /// Doesn't change for non-dynamic
780 size_t allocatedCapacity() const noexcept {
784 /// Returns the total number of calls to blockingWrite or successful
785 /// calls to write, including those blockingWrite calls that are
786 /// currently blocking
787 uint64_t writeCount() const noexcept {
788 return pushTicket_.load(std::memory_order_acquire);
791 /// Returns the total number of calls to blockingRead or successful
792 /// calls to read, including those blockingRead calls that are currently
794 uint64_t readCount() const noexcept {
795 return popTicket_.load(std::memory_order_acquire);
798 /// Enqueues a T constructed from args, blocking until space is
799 /// available. Note that this method signature allows enqueue via
800 /// move, if args is a T rvalue, via copy, if args is a T lvalue, or
801 /// via emplacement if args is an initializer list that can be passed
802 /// to a T constructor.
803 template <typename ...Args>
804 void blockingWrite(Args&&... args) noexcept {
805 enqueueWithTicketBase(pushTicket_++, slots_, capacity_, stride_,
806 std::forward<Args>(args)...);
809 /// If an item can be enqueued with no blocking, does so and returns
810 /// true, otherwise returns false. This method is similar to
811 /// writeIfNotFull, but if you don't have a specific need for that
812 /// method you should use this one.
814 /// One of the common usages of this method is to enqueue via the
815 /// move constructor, something like q.write(std::move(x)). If write
816 /// returns false because the queue is full then x has not actually been
817 /// consumed, which looks strange. To understand why it is actually okay
818 /// to use x afterward, remember that std::move is just a typecast that
819 /// provides an rvalue reference that enables use of a move constructor
820 /// or operator. std::move doesn't actually move anything. It could
821 /// more accurately be called std::rvalue_cast or std::move_permission.
822 template <typename ...Args>
823 bool write(Args&&... args) noexcept {
828 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
829 tryObtainReadyPushTicket(ticket, slots, cap, stride)) {
830 // we have pre-validated that the ticket won't block
831 enqueueWithTicketBase(ticket, slots, cap, stride,
832 std::forward<Args>(args)...);
839 template <class Clock, typename... Args>
840 bool tryWriteUntil(const std::chrono::time_point<Clock>& when,
841 Args&&... args) noexcept {
846 if (tryObtainPromisedPushTicketUntil(ticket, slots, cap, stride, when)) {
847 // we have pre-validated that the ticket won't block, or rather that
848 // it won't block longer than it takes another thread to dequeue an
849 // element from the slot it identifies.
850 enqueueWithTicketBase(ticket, slots, cap, stride,
851 std::forward<Args>(args)...);
858 /// If the queue is not full, enqueues and returns true, otherwise
859 /// returns false. Unlike write this method can be blocked by another
860 /// thread, specifically a read that has linearized (been assigned
861 /// a ticket) but not yet completed. If you don't really need this
862 /// function you should probably use write.
864 /// MPMCQueue isn't lock-free, so just because a read operation has
865 /// linearized (and isFull is false) doesn't mean that space has been
866 /// made available for another write. In this situation write will
867 /// return false, but writeIfNotFull will wait for the dequeue to finish.
868 /// This method is required if you are composing queues and managing
869 /// your own wakeup, because it guarantees that after every successful
870 /// write a readIfNotEmpty will succeed.
871 template <typename ...Args>
872 bool writeIfNotFull(Args&&... args) noexcept {
877 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
878 tryObtainPromisedPushTicket(ticket, slots, cap, stride)) {
879 // some other thread is already dequeuing the slot into which we
880 // are going to enqueue, but we might have to wait for them to finish
881 enqueueWithTicketBase(ticket, slots, cap, stride,
882 std::forward<Args>(args)...);
889 /// Moves a dequeued element onto elem, blocking until an element
891 void blockingRead(T& elem) noexcept {
893 static_cast<Derived<T,Atom,Dynamic>*>(this)->
894 blockingReadWithTicket(ticket, elem);
897 /// Same as blockingRead() but also records the ticket nunmer
898 void blockingReadWithTicket(uint64_t& ticket, T& elem) noexcept {
899 assert(capacity_ != 0);
900 ticket = popTicket_++;
901 dequeueWithTicketBase(ticket, slots_, capacity_, stride_, elem);
904 /// If an item can be dequeued with no blocking, does so and returns
905 /// true, otherwise returns false.
906 bool read(T& elem) noexcept {
908 return readAndGetTicket(ticket, elem);
911 /// Same as read() but also records the ticket nunmer
912 bool readAndGetTicket(uint64_t& ticket, T& elem) noexcept {
916 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
917 tryObtainReadyPopTicket(ticket, slots, cap, stride)) {
918 // the ticket has been pre-validated to not block
919 dequeueWithTicketBase(ticket, slots, cap, stride, elem);
926 template <class Clock, typename... Args>
928 const std::chrono::time_point<Clock>& when,
934 if (tryObtainPromisedPopTicketUntil(ticket, slots, cap, stride, when)) {
935 // we have pre-validated that the ticket won't block, or rather that
936 // it won't block longer than it takes another thread to enqueue an
937 // element on the slot it identifies.
938 dequeueWithTicketBase(ticket, slots, cap, stride, elem);
945 /// If the queue is not empty, dequeues and returns true, otherwise
946 /// returns false. If the matching write is still in progress then this
947 /// method may block waiting for it. If you don't rely on being able
948 /// to dequeue (such as by counting completed write) then you should
950 bool readIfNotEmpty(T& elem) noexcept {
955 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
956 tryObtainPromisedPopTicket(ticket, slots, cap, stride)) {
957 // the matching enqueue already has a ticket, but might not be done
958 dequeueWithTicketBase(ticket, slots, cap, stride, elem);
967 /// Once every kAdaptationFreq we will spin longer, to try to estimate
968 /// the proper spin backoff
969 kAdaptationFreq = 128,
971 /// To avoid false sharing in slots_ with neighboring memory
972 /// allocations, we pad it with this many SingleElementQueue-s at
974 kSlotPadding = (detail::CacheLocality::kFalseSharingRange - 1)
978 /// The maximum number of items in the queue at once
979 size_t FOLLY_ALIGN_TO_AVOID_FALSE_SHARING capacity_;
981 /// Anonymous union for use when Dynamic = false and true, respectively
983 /// An array of capacity_ SingleElementQueue-s, each of which holds
984 /// either 0 or 1 item. We over-allocate by 2 * kSlotPadding and don't
985 /// touch the slots at either end, to avoid false sharing
987 /// Current dynamic slots array of dcapacity_ SingleElementQueue-s
991 /// Anonymous union for use when Dynamic = false and true, respectively
993 /// The number of slots_ indices that we advance for each ticket, to
994 /// avoid false sharing. Ideally slots_[i] and slots_[i + stride_]
995 /// aren't on the same cache line
1001 /// The following two memebers are used by dynamic MPMCQueue.
1002 /// Ideally they should be in MPMCQueue<T,Atom,true>, but we get
1003 /// better cache locality if they are in the same cache line as
1004 /// dslots_ and dstride_.
1006 /// Dynamic state. A packed seqlock and ticket offset
1007 Atom<uint64_t> dstate_;
1008 /// Dynamic capacity
1009 Atom<size_t> dcapacity_;
1011 /// Enqueuers get tickets from here
1012 Atom<uint64_t> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING pushTicket_;
1014 /// Dequeuers get tickets from here
1015 Atom<uint64_t> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING popTicket_;
1017 /// This is how many times we will spin before using FUTEX_WAIT when
1018 /// the queue is full on enqueue, adaptively computed by occasionally
1019 /// spinning for longer and smoothing with an exponential moving average
1020 Atom<uint32_t> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING pushSpinCutoff_;
1022 /// The adaptive spin cutoff when the queue is empty on dequeue
1023 Atom<uint32_t> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING popSpinCutoff_;
1025 /// Alignment doesn't prevent false sharing at the end of the struct,
1026 /// so fill out the last cache line
1027 char padding_[detail::CacheLocality::kFalseSharingRange -
1028 sizeof(Atom<uint32_t>)];
1030 /// We assign tickets in increasing order, but we don't want to
1031 /// access neighboring elements of slots_ because that will lead to
1032 /// false sharing (multiple cores accessing the same cache line even
1033 /// though they aren't accessing the same bytes in that cache line).
1034 /// To avoid this we advance by stride slots per ticket.
1036 /// We need gcd(capacity, stride) to be 1 so that we will use all
1037 /// of the slots. We ensure this by only considering prime strides,
1038 /// which either have no common divisors with capacity or else have
1039 /// a zero remainder after dividing by capacity. That is sufficient
1040 /// to guarantee correctness, but we also want to actually spread the
1041 /// accesses away from each other to avoid false sharing (consider a
1042 /// stride of 7 with a capacity of 8). To that end we try a few taking
1043 /// care to observe that advancing by -1 is as bad as advancing by 1
1044 /// when in comes to false sharing.
1046 /// The simple way to avoid false sharing would be to pad each
1047 /// SingleElementQueue, but since we have capacity_ of them that could
1048 /// waste a lot of space.
1049 static int computeStride(size_t capacity) noexcept {
1050 static const int smallPrimes[] = { 2, 3, 5, 7, 11, 13, 17, 19, 23 };
1054 for (int stride : smallPrimes) {
1055 if ((stride % capacity) == 0 || (capacity % stride) == 0) {
1058 size_t sep = stride % capacity;
1059 sep = std::min(sep, capacity - sep);
1060 if (sep > bestSep) {
1061 bestStride = stride;
1068 /// Returns the index into slots_ that should be used when enqueuing or
1069 /// dequeuing with the specified ticket
1070 size_t idx(uint64_t ticket, size_t cap, int stride) noexcept {
1071 return ((ticket * stride) % cap) + kSlotPadding;
1074 /// Maps an enqueue or dequeue ticket to the turn should be used at the
1075 /// corresponding SingleElementQueue
1076 uint32_t turn(uint64_t ticket, size_t cap) noexcept {
1078 return uint32_t(ticket / cap);
1081 /// Tries to obtain a push ticket for which SingleElementQueue::enqueue
1082 /// won't block. Returns true on immediate success, false on immediate
1084 bool tryObtainReadyPushTicket(
1085 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1087 ticket = pushTicket_.load(std::memory_order_acquire); // A
1092 if (!slots[idx(ticket, cap, stride)]
1093 .mayEnqueue(turn(ticket, cap))) {
1094 // if we call enqueue(ticket, ...) on the SingleElementQueue
1095 // right now it would block, but this might no longer be the next
1096 // ticket. We can increase the chance of tryEnqueue success under
1097 // contention (without blocking) by rechecking the ticket dispenser
1099 ticket = pushTicket_.load(std::memory_order_acquire); // B
1100 if (prev == ticket) {
1101 // mayEnqueue was bracketed by two reads (A or prev B or prev
1102 // failing CAS to B), so we are definitely unable to enqueue
1106 // we will bracket the mayEnqueue check with a read (A or prev B
1107 // or prev failing CAS) and the following CAS. If the CAS fails
1108 // it will effect a load of pushTicket_
1109 if (pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
1116 /// Tries until when to obtain a push ticket for which
1117 /// SingleElementQueue::enqueue won't block. Returns true on success, false
1119 /// ticket is filled on success AND failure.
1120 template <class Clock>
1121 bool tryObtainPromisedPushTicketUntil(
1122 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride,
1123 const std::chrono::time_point<Clock>& when
1125 bool deadlineReached = false;
1126 while (!deadlineReached) {
1127 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
1128 tryObtainPromisedPushTicket(ticket, slots, cap, stride)) {
1131 // ticket is a blocking ticket until the preceding ticket has been
1132 // processed: wait until this ticket's turn arrives. We have not reserved
1133 // this ticket so we will have to re-attempt to get a non-blocking ticket
1134 // if we wake up before we time-out.
1135 deadlineReached = !slots[idx(ticket, cap, stride)]
1136 .tryWaitForEnqueueTurnUntil(turn(ticket, cap), pushSpinCutoff_,
1137 (ticket % kAdaptationFreq) == 0, when);
1142 /// Tries to obtain a push ticket which can be satisfied if all
1143 /// in-progress pops complete. This function does not block, but
1144 /// blocking may be required when using the returned ticket if some
1145 /// other thread's pop is still in progress (ticket has been granted but
1146 /// pop has not yet completed).
1147 bool tryObtainPromisedPushTicket(
1148 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1150 auto numPushes = pushTicket_.load(std::memory_order_acquire); // A
1156 const auto numPops = popTicket_.load(std::memory_order_acquire); // B
1157 // n will be negative if pops are pending
1158 const int64_t n = int64_t(numPushes - numPops);
1159 if (n >= static_cast<ssize_t>(capacity_)) {
1160 // Full, linearize at B. We don't need to recheck the read we
1161 // performed at A, because if numPushes was stale at B then the
1162 // real numPushes value is even worse
1165 if (pushTicket_.compare_exchange_strong(numPushes, numPushes + 1)) {
1171 /// Tries to obtain a pop ticket for which SingleElementQueue::dequeue
1172 /// won't block. Returns true on immediate success, false on immediate
1174 bool tryObtainReadyPopTicket(
1175 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1177 ticket = popTicket_.load(std::memory_order_acquire);
1182 if (!slots[idx(ticket, cap, stride)]
1183 .mayDequeue(turn(ticket, cap))) {
1185 ticket = popTicket_.load(std::memory_order_acquire);
1186 if (prev == ticket) {
1190 if (popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
1197 /// Tries until when to obtain a pop ticket for which
1198 /// SingleElementQueue::dequeue won't block. Returns true on success, false
1200 /// ticket is filled on success AND failure.
1201 template <class Clock>
1202 bool tryObtainPromisedPopTicketUntil(
1207 const std::chrono::time_point<Clock>& when) noexcept {
1208 bool deadlineReached = false;
1209 while (!deadlineReached) {
1210 if (static_cast<Derived<T, Atom, Dynamic>*>(this)
1211 ->tryObtainPromisedPopTicket(ticket, slots, cap, stride)) {
1214 // ticket is a blocking ticket until the preceding ticket has been
1215 // processed: wait until this ticket's turn arrives. We have not reserved
1216 // this ticket so we will have to re-attempt to get a non-blocking ticket
1217 // if we wake up before we time-out.
1219 !slots[idx(ticket, cap, stride)].tryWaitForDequeueTurnUntil(
1222 (ticket % kAdaptationFreq) == 0,
1228 /// Similar to tryObtainReadyPopTicket, but returns a pop ticket whose
1229 /// corresponding push ticket has already been handed out, rather than
1230 /// returning one whose corresponding push ticket has already been
1231 /// completed. This means that there is a possibility that the caller
1232 /// will block when using the ticket, but it allows the user to rely on
1233 /// the fact that if enqueue has succeeded, tryObtainPromisedPopTicket
1234 /// will return true. The "try" part of this is that we won't have
1235 /// to block waiting for someone to call enqueue, although we might
1236 /// have to block waiting for them to finish executing code inside the
1237 /// MPMCQueue itself.
1238 bool tryObtainPromisedPopTicket(
1239 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1241 auto numPops = popTicket_.load(std::memory_order_acquire); // A
1247 const auto numPushes = pushTicket_.load(std::memory_order_acquire); // B
1248 if (numPops >= numPushes) {
1249 // Empty, or empty with pending pops. Linearize at B. We don't
1250 // need to recheck the read we performed at A, because if numPops
1251 // is stale then the fresh value is larger and the >= is still true
1254 if (popTicket_.compare_exchange_strong(numPops, numPops + 1)) {
1260 // Given a ticket, constructs an enqueued item using args
1261 template <typename ...Args>
1262 void enqueueWithTicketBase(
1263 uint64_t ticket, Slot* slots, size_t cap, int stride, Args&&... args
1265 slots[idx(ticket, cap, stride)]
1266 .enqueue(turn(ticket, cap),
1268 (ticket % kAdaptationFreq) == 0,
1269 std::forward<Args>(args)...);
1272 // To support tracking ticket numbers in MPMCPipelineStageImpl
1273 template <typename ...Args>
1274 void enqueueWithTicket(uint64_t ticket, Args&&... args) noexcept {
1275 enqueueWithTicketBase(ticket, slots_, capacity_, stride_,
1276 std::forward<Args>(args)...);
1279 // Given a ticket, dequeues the corresponding element
1280 void dequeueWithTicketBase(
1281 uint64_t ticket, Slot* slots, size_t cap, int stride, T& elem
1284 slots[idx(ticket, cap, stride)]
1285 .dequeue(turn(ticket, cap),
1287 (ticket % kAdaptationFreq) == 0,
1292 /// SingleElementQueue implements a blocking queue that holds at most one
1293 /// item, and that requires its users to assign incrementing identifiers
1294 /// (turns) to each enqueue and dequeue operation. Note that the turns
1295 /// used by SingleElementQueue are doubled inside the TurnSequencer
1296 template <typename T, template <typename> class Atom>
1297 struct SingleElementQueue {
1299 ~SingleElementQueue() noexcept {
1300 if ((sequencer_.uncompletedTurnLSB() & 1) == 1) {
1301 // we are pending a dequeue, so we have a constructed item
1306 /// enqueue using in-place noexcept construction
1307 template <typename ...Args,
1308 typename = typename std::enable_if<
1309 std::is_nothrow_constructible<T,Args...>::value>::type>
1310 void enqueue(const uint32_t turn,
1311 Atom<uint32_t>& spinCutoff,
1312 const bool updateSpinCutoff,
1313 Args&&... args) noexcept {
1314 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
1315 new (&contents_) T(std::forward<Args>(args)...);
1316 sequencer_.completeTurn(turn * 2);
1319 /// enqueue using move construction, either real (if
1320 /// is_nothrow_move_constructible) or simulated using relocation and
1321 /// default construction (if IsRelocatable and is_nothrow_constructible)
1323 typename = typename std::enable_if<
1324 (folly::IsRelocatable<T>::value &&
1325 std::is_nothrow_constructible<T>::value) ||
1326 std::is_nothrow_constructible<T, T&&>::value>::type>
1328 const uint32_t turn,
1329 Atom<uint32_t>& spinCutoff,
1330 const bool updateSpinCutoff,
1331 T&& goner) noexcept {
1337 typename std::conditional<std::is_nothrow_constructible<T,T&&>::value,
1338 ImplByMove, ImplByRelocation>::type());
1341 /// Waits until either:
1342 /// 1: the dequeue turn preceding the given enqueue turn has arrived
1343 /// 2: the given deadline has arrived
1344 /// Case 1 returns true, case 2 returns false.
1345 template <class Clock>
1346 bool tryWaitForEnqueueTurnUntil(
1347 const uint32_t turn,
1348 Atom<uint32_t>& spinCutoff,
1349 const bool updateSpinCutoff,
1350 const std::chrono::time_point<Clock>& when) noexcept {
1351 return sequencer_.tryWaitForTurn(
1352 turn * 2, spinCutoff, updateSpinCutoff, &when) !=
1353 TurnSequencer<Atom>::TryWaitResult::TIMEDOUT;
1356 bool mayEnqueue(const uint32_t turn) const noexcept {
1357 return sequencer_.isTurn(turn * 2);
1360 void dequeue(uint32_t turn,
1361 Atom<uint32_t>& spinCutoff,
1362 const bool updateSpinCutoff,
1368 typename std::conditional<folly::IsRelocatable<T>::value,
1370 ImplByMove>::type());
1373 /// Waits until either:
1374 /// 1: the enqueue turn preceding the given dequeue turn has arrived
1375 /// 2: the given deadline has arrived
1376 /// Case 1 returns true, case 2 returns false.
1377 template <class Clock>
1378 bool tryWaitForDequeueTurnUntil(
1379 const uint32_t turn,
1380 Atom<uint32_t>& spinCutoff,
1381 const bool updateSpinCutoff,
1382 const std::chrono::time_point<Clock>& when) noexcept {
1383 return sequencer_.tryWaitForTurn(
1384 turn * 2 + 1, spinCutoff, updateSpinCutoff, &when) !=
1385 TurnSequencer<Atom>::TryWaitResult::TIMEDOUT;
1388 bool mayDequeue(const uint32_t turn) const noexcept {
1389 return sequencer_.isTurn(turn * 2 + 1);
1393 /// Storage for a T constructed with placement new
1394 typename std::aligned_storage<sizeof(T),alignof(T)>::type contents_;
1396 /// Even turns are pushes, odd turns are pops
1397 TurnSequencer<Atom> sequencer_;
1400 return static_cast<T*>(static_cast<void*>(&contents_));
1403 void destroyContents() noexcept {
1407 // g++ doesn't seem to have std::is_nothrow_destructible yet
1410 memset(&contents_, 'Q', sizeof(T));
1414 /// Tag classes for dispatching to enqueue/dequeue implementation.
1415 struct ImplByRelocation {};
1416 struct ImplByMove {};
1418 /// enqueue using nothrow move construction.
1419 void enqueueImpl(const uint32_t turn,
1420 Atom<uint32_t>& spinCutoff,
1421 const bool updateSpinCutoff,
1423 ImplByMove) noexcept {
1424 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
1425 new (&contents_) T(std::move(goner));
1426 sequencer_.completeTurn(turn * 2);
1429 /// enqueue by simulating nothrow move with relocation, followed by
1430 /// default construction to a noexcept relocation.
1431 void enqueueImpl(const uint32_t turn,
1432 Atom<uint32_t>& spinCutoff,
1433 const bool updateSpinCutoff,
1435 ImplByRelocation) noexcept {
1436 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
1437 memcpy(&contents_, &goner, sizeof(T));
1438 sequencer_.completeTurn(turn * 2);
1442 /// dequeue by destructing followed by relocation. This version is preferred,
1443 /// because as much work as possible can be done before waiting.
1444 void dequeueImpl(uint32_t turn,
1445 Atom<uint32_t>& spinCutoff,
1446 const bool updateSpinCutoff,
1448 ImplByRelocation) noexcept {
1452 // unlikely, but if we don't complete our turn the queue will die
1454 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
1455 memcpy(&elem, &contents_, sizeof(T));
1456 sequencer_.completeTurn(turn * 2 + 1);
1459 /// dequeue by nothrow move assignment.
1460 void dequeueImpl(uint32_t turn,
1461 Atom<uint32_t>& spinCutoff,
1462 const bool updateSpinCutoff,
1464 ImplByMove) noexcept {
1465 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
1466 elem = std::move(*ptr());
1468 sequencer_.completeTurn(turn * 2 + 1);
1472 } // namespace detail
1474 } // namespace folly