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.
24 #include <type_traits>
26 #include <boost/noncopyable.hpp>
28 #include <folly/Traits.h>
29 #include <folly/concurrency/CacheLocality.h>
30 #include <folly/detail/TurnSequencer.h>
31 #include <folly/portability/Unistd.h>
37 template<typename T, template<typename> class Atom>
38 struct SingleElementQueue;
40 template <typename T> class MPMCPipelineStageImpl;
42 /// MPMCQueue base CRTP template
43 template <typename> class MPMCQueueBase;
47 /// MPMCQueue<T> is a high-performance bounded concurrent queue that
48 /// supports multiple producers, multiple consumers, and optional blocking.
49 /// The queue has a fixed capacity, for which all memory will be allocated
50 /// up front. The bulk of the work of enqueuing and dequeuing can be
51 /// performed in parallel.
53 /// MPMCQueue is linearizable. That means that if a call to write(A)
54 /// returns before a call to write(B) begins, then A will definitely end up
55 /// in the queue before B, and if a call to read(X) returns before a call
56 /// to read(Y) is started, that X will be something from earlier in the
57 /// queue than Y. This also means that if a read call returns a value, you
58 /// can be sure that all previous elements of the queue have been assigned
59 /// a reader (that reader might not yet have returned, but it exists).
61 /// The underlying implementation uses a ticket dispenser for the head and
62 /// the tail, spreading accesses across N single-element queues to produce
63 /// a queue with capacity N. The ticket dispensers use atomic increment,
64 /// which is more robust to contention than a CAS loop. Each of the
65 /// single-element queues uses its own CAS to serialize access, with an
66 /// adaptive spin cutoff. When spinning fails on a single-element queue
67 /// it uses futex()'s _BITSET operations to reduce unnecessary wakeups
68 /// even if multiple waiters are present on an individual queue (such as
69 /// when the MPMCQueue's capacity is smaller than the number of enqueuers
72 /// In benchmarks (contained in tao/queues/ConcurrentQueueTests)
73 /// it handles 1 to 1, 1 to N, N to 1, and N to M thread counts better
74 /// than any of the alternatives present in fbcode, for both small (~10)
75 /// and large capacities. In these benchmarks it is also faster than
76 /// tbb::concurrent_bounded_queue for all configurations. When there are
77 /// many more threads than cores, MPMCQueue is _much_ faster than the tbb
78 /// queue because it uses futex() to block and unblock waiting threads,
79 /// rather than spinning with sched_yield.
81 /// NOEXCEPT INTERACTION: tl;dr; If it compiles you're fine. Ticket-based
82 /// queues separate the assignment of queue positions from the actual
83 /// construction of the in-queue elements, which means that the T
84 /// constructor used during enqueue must not throw an exception. This is
85 /// enforced at compile time using type traits, which requires that T be
86 /// adorned with accurate noexcept information. If your type does not
87 /// use noexcept, you will have to wrap it in something that provides
88 /// the guarantee. We provide an alternate safe implementation for types
89 /// that don't use noexcept but that are marked folly::IsRelocatable
90 /// and std::is_nothrow_constructible, which is common for folly types.
91 /// In particular, if you can declare FOLLY_ASSUME_FBVECTOR_COMPATIBLE
92 /// then your type can be put in MPMCQueue.
94 /// If you have a pool of N queue consumers that you want to shut down
95 /// after the queue has drained, one way is to enqueue N sentinel values
96 /// to the queue. If the producer doesn't know how many consumers there
97 /// are you can enqueue one sentinel and then have each consumer requeue
98 /// two sentinels after it receives it (by requeuing 2 the shutdown can
99 /// complete in O(log P) time instead of O(P)).
100 template<typename T, template<typename> class Atom = std::atomic,
101 bool Dynamic = false>
102 class MPMCQueue : public detail::MPMCQueueBase<MPMCQueue<T,Atom,Dynamic>> {
103 friend class detail::MPMCPipelineStageImpl<T>;
104 using Slot = detail::SingleElementQueue<T,Atom>;
107 explicit MPMCQueue(size_t queueCapacity)
108 : detail::MPMCQueueBase<MPMCQueue<T,Atom,Dynamic>>(queueCapacity)
110 this->stride_ = this->computeStride(queueCapacity);
111 this->slots_ = new Slot[queueCapacity + 2 * this->kSlotPadding];
114 MPMCQueue() noexcept { }
117 /// The dynamic version of MPMCQueue allows dynamic expansion of queue
118 /// capacity, such that a queue may start with a smaller capacity than
119 /// specified and expand only if needed. Users may optionally specify
120 /// the initial capacity and the expansion multiplier.
122 /// The design uses a seqlock to enforce mutual exclusion among
123 /// expansion attempts. Regular operations read up-to-date queue
124 /// information (slots array, capacity, stride) inside read-only
125 /// seqlock sections, which are unimpeded when no expansion is in
128 /// An expansion computes a new capacity, allocates a new slots array,
129 /// and updates stride. No information needs to be copied from the
130 /// current slots array to the new one. When this happens, new slots
131 /// will not have sequence numbers that match ticket numbers. The
132 /// expansion needs to compute a ticket offset such that operations
133 /// that use new arrays can adjust the calculations of slot indexes
134 /// and sequence numbers that take into account that the new slots
135 /// start with sequence numbers of zero. The current ticket offset is
136 /// packed with the seqlock in an atomic 64-bit integer. The initial
139 /// Lagging write and read operations with tickets lower than the
140 /// ticket offset of the current slots array (i.e., the minimum ticket
141 /// number that can be served by the current array) must use earlier
142 /// closed arrays instead of the current one. Information about closed
143 /// slots arrays (array address, capacity, stride, and offset) is
144 /// maintained in a logarithmic-sized structure. Each entry in that
145 /// structure never need to be changed once set. The number of closed
146 /// arrays is half the value of the seqlock (when unlocked).
148 /// The acquisition of the seqlock to perform an expansion does not
149 /// prevent the issuing of new push and pop tickets concurrently. The
150 /// expansion must set the new ticket offset to a value that couldn't
151 /// have been issued to an operation that has already gone through a
152 /// seqlock read-only section (and hence obtained information for
153 /// older closed arrays).
155 /// Note that the total queue capacity can temporarily exceed the
156 /// specified capacity when there are lagging consumers that haven't
157 /// yet consumed all the elements in closed arrays. Users should not
158 /// rely on the capacity of dynamic queues for synchronization, e.g.,
159 /// they should not expect that a thread will definitely block on a
160 /// call to blockingWrite() when the queue size is known to be equal
163 /// Note that some writeIfNotFull() and tryWriteUntil() operations may
164 /// fail even if the size of the queue is less than its maximum
165 /// capacity and despite the success of expansion, if the operation
166 /// happens to acquire a ticket that belongs to a closed array. This
167 /// is a transient condition. Typically, one or two ticket values may
168 /// be subject to such condition per expansion.
170 /// The dynamic version is a partial specialization of MPMCQueue with
172 template <typename T, template<typename> class Atom>
173 class MPMCQueue<T,Atom,true> :
174 public detail::MPMCQueueBase<MPMCQueue<T,Atom,true>> {
175 friend class detail::MPMCQueueBase<MPMCQueue<T,Atom,true>>;
176 using Slot = detail::SingleElementQueue<T,Atom>;
179 uint64_t offset_ {0};
180 Slot* slots_ {nullptr};
181 size_t capacity_ {0};
187 explicit MPMCQueue(size_t queueCapacity)
188 : detail::MPMCQueueBase<MPMCQueue<T,Atom,true>>(queueCapacity)
190 size_t cap = std::min<size_t>(kDefaultMinDynamicCapacity, queueCapacity);
191 initQueue(cap, kDefaultExpansionMultiplier);
194 explicit MPMCQueue(size_t queueCapacity,
196 size_t expansionMultiplier)
197 : detail::MPMCQueueBase<MPMCQueue<T,Atom,true>>(queueCapacity)
199 minCapacity = std::max<size_t>(1, minCapacity);
200 size_t cap = std::min<size_t>(minCapacity, queueCapacity);
201 expansionMultiplier = std::max<size_t>(2, expansionMultiplier);
202 initQueue(cap, expansionMultiplier);
205 MPMCQueue() noexcept {
210 MPMCQueue(MPMCQueue<T,Atom,true>&& rhs) noexcept {
211 this->capacity_ = rhs.capacity_;
212 this->slots_ = rhs.slots_;
213 this->stride_ = rhs.stride_;
214 this->dstate_.store(rhs.dstate_.load(std::memory_order_relaxed),
215 std::memory_order_relaxed);
216 this->dcapacity_.store(rhs.dcapacity_.load(std::memory_order_relaxed),
217 std::memory_order_relaxed);
218 this->pushTicket_.store(rhs.pushTicket_.load(std::memory_order_relaxed),
219 std::memory_order_relaxed);
220 this->popTicket_.store(rhs.popTicket_.load(std::memory_order_relaxed),
221 std::memory_order_relaxed);
222 this->pushSpinCutoff_.store(
223 rhs.pushSpinCutoff_.load(std::memory_order_relaxed),
224 std::memory_order_relaxed);
225 this->popSpinCutoff_.store(
226 rhs.popSpinCutoff_.load(std::memory_order_relaxed),
227 std::memory_order_relaxed);
229 closed_ = rhs.closed_;
232 rhs.slots_ = nullptr;
234 rhs.dstate_.store(0, std::memory_order_relaxed);
235 rhs.dcapacity_.store(0, std::memory_order_relaxed);
236 rhs.pushTicket_.store(0, std::memory_order_relaxed);
237 rhs.popTicket_.store(0, std::memory_order_relaxed);
238 rhs.pushSpinCutoff_.store(0, std::memory_order_relaxed);
239 rhs.popSpinCutoff_.store(0, std::memory_order_relaxed);
241 rhs.closed_ = nullptr;
244 MPMCQueue<T,Atom, true> const& operator= (MPMCQueue<T,Atom, true>&& rhs) {
247 new (this) MPMCQueue(std::move(rhs));
253 if (closed_ != nullptr) {
254 for (int i = getNumClosed(this->dstate_.load()) - 1; i >= 0; --i) {
255 delete[] closed_[i].slots_;
261 size_t allocatedCapacity() const noexcept {
262 return this->dcapacity_.load(std::memory_order_relaxed);
265 template <typename ...Args>
266 void blockingWrite(Args&&... args) noexcept {
267 uint64_t ticket = this->pushTicket_++;
274 if (!trySeqlockReadSection(state, slots, cap, stride)) {
275 asm_volatile_pause();
278 if (maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride)) {
279 // There was an expansion after this ticket was issued.
282 if (slots[this->idx((ticket-offset), cap, stride)]
283 .mayEnqueue(this->turn(ticket-offset, cap))) {
284 // A slot is ready. No need to expand.
286 } else if (this->popTicket_.load(std::memory_order_relaxed) + cap
288 // May block, but a pop is in progress. No need to expand.
289 // Get seqlock read section info again in case an expansion
290 // occurred with an equal or higher ticket.
293 // May block. See if we can expand.
294 if (tryExpand(state, cap)) {
295 // This or another thread started an expansion. Get updated info.
303 this->enqueueWithTicketBase(ticket-offset, slots, cap, stride,
304 std::forward<Args>(args)...);
307 void blockingReadWithTicket(uint64_t& ticket, T& elem) noexcept {
308 ticket = this->popTicket_++;
314 while (!trySeqlockReadSection(state, slots, cap, stride)) {
315 asm_volatile_pause();
317 // If there was an expansion after the corresponding push ticket
318 // was issued, adjust accordingly
319 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
320 this->dequeueWithTicketBase(ticket-offset, slots, cap, stride, elem);
327 kDefaultMinDynamicCapacity = 10,
328 kDefaultExpansionMultiplier = 10,
333 // Info about closed slots arrays for use by lagging operations
334 ClosedArray* closed_;
336 void initQueue(const size_t cap, const size_t mult) {
337 this->stride_ = this->computeStride(cap);
338 this->slots_ = new Slot[cap + 2 * this->kSlotPadding];
339 this->dstate_.store(0);
340 this->dcapacity_.store(cap);
342 size_t maxClosed = 0;
343 for (size_t expanded = cap;
344 expanded < this->capacity_;
348 closed_ = (maxClosed > 0) ? new ClosedArray[maxClosed] : nullptr;
351 bool tryObtainReadyPushTicket(
352 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
356 ticket = this->pushTicket_.load(std::memory_order_acquire); // A
357 if (!trySeqlockReadSection(state, slots, cap, stride)) {
358 asm_volatile_pause();
362 // If there was an expansion with offset greater than this ticket,
363 // adjust accordingly
365 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
367 if (slots[this->idx((ticket-offset), cap, stride)]
368 .mayEnqueue(this->turn(ticket-offset, cap))) {
370 if (this->pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
378 if (ticket != this->pushTicket_.load(std::memory_order_relaxed)) { // B
379 // Try again. Ticket changed.
383 // Try to expand unless the ticket is for a closed array
384 if (offset == getOffset(state)) {
385 if (tryExpand(state, cap)) {
386 // This or another thread started an expansion. Get up-to-date info.
395 bool tryObtainPromisedPushTicket(
396 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
400 ticket = this->pushTicket_.load(std::memory_order_acquire);
401 auto numPops = this->popTicket_.load(std::memory_order_acquire);
402 if (!trySeqlockReadSection(state, slots, cap, stride)) {
403 asm_volatile_pause();
407 const auto curCap = cap;
408 // If there was an expansion with offset greater than this ticket,
409 // adjust accordingly
411 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
413 int64_t n = ticket - numPops;
415 if (n >= static_cast<ssize_t>(cap)) {
416 if ((cap == curCap) && tryExpand(state, cap)) {
417 // This or another thread started an expansion. Start over.
425 if (this->pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
433 bool tryObtainReadyPopTicket(
434 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
438 ticket = this->popTicket_.load(std::memory_order_relaxed);
439 if (!trySeqlockReadSection(state, slots, cap, stride)) {
440 asm_volatile_pause();
444 // If there was an expansion after the corresponding push ticket
445 // was issued, adjust accordingly
447 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
449 if (slots[this->idx((ticket-offset), cap, stride)]
450 .mayDequeue(this->turn(ticket-offset, cap))) {
451 if (this->popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
462 bool tryObtainPromisedPopTicket(
463 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
467 ticket = this->popTicket_.load(std::memory_order_acquire);
468 auto numPushes = this->pushTicket_.load(std::memory_order_acquire);
469 if (!trySeqlockReadSection(state, slots, cap, stride)) {
470 asm_volatile_pause();
475 // If there was an expansion after the corresponding push
476 // ticket was issued, adjust accordingly
477 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
479 if (ticket >= numPushes) {
483 if (this->popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
490 /// Enqueues an element with a specific ticket number
491 template <typename ...Args>
492 void enqueueWithTicket(const uint64_t ticket, Args&&... args) noexcept {
499 while (!trySeqlockReadSection(state, slots, cap, stride)) {}
501 // If there was an expansion after this ticket was issued, adjust
503 maybeUpdateFromClosed(state, ticket, offset, slots, cap, stride);
505 this->enqueueWithTicketBase(ticket-offset, slots, cap, stride,
506 std::forward<Args>(args)...);
509 uint64_t getOffset(const uint64_t state) const noexcept {
510 return state >> kSeqlockBits;
513 int getNumClosed(const uint64_t state) const noexcept {
514 return (state & ((1 << kSeqlockBits) - 1)) >> 1;
517 /// Try to expand the queue. Returns true if this expansion was
518 /// successful or a concurent expansion is in progress. Returns
519 /// false if the queue has reached its maximum capacity or
520 /// allocation has failed.
521 bool tryExpand(const uint64_t state, const size_t cap) noexcept {
522 if (cap == this->capacity_) {
526 uint64_t oldval = state;
527 assert((state & 1) == 0);
528 if (this->dstate_.compare_exchange_strong(oldval, state + 1)) {
529 assert(cap == this->dcapacity_.load());
530 uint64_t ticket = 1 + std::max(this->pushTicket_.load(),
531 this->popTicket_.load());
533 std::min(dmult_ * cap, this->capacity_);
535 new (std::nothrow) Slot[newCapacity + 2 * this->kSlotPadding];
536 if (newSlots == nullptr) {
537 // Expansion failed. Restore the seqlock
538 this->dstate_.store(state);
541 // Successful expansion
542 // calculate the current ticket offset
543 uint64_t offset = getOffset(state);
544 // calculate index in closed array
545 int index = getNumClosed(state);
546 assert((index << 1) < (1 << kSeqlockBits));
547 // fill the info for the closed slots array
548 closed_[index].offset_ = offset;
549 closed_[index].slots_ = this->dslots_.load();
550 closed_[index].capacity_ = cap;
551 closed_[index].stride_ = this->dstride_.load();
552 // update the new slots array info
553 this->dslots_.store(newSlots);
554 this->dcapacity_.store(newCapacity);
555 this->dstride_.store(this->computeStride(newCapacity));
556 // Release the seqlock and record the new ticket offset
557 this->dstate_.store((ticket << kSeqlockBits) + (2 * (index + 1)));
559 } else { // failed to acquire seqlock
560 // Someone acaquired the seqlock. Go back to the caller and get
566 /// Seqlock read-only section
567 bool trySeqlockReadSection(
568 uint64_t& state, Slot*& slots, size_t& cap, int& stride
570 state = this->dstate_.load(std::memory_order_acquire);
575 // Start read-only section.
576 slots = this->dslots_.load(std::memory_order_relaxed);
577 cap = this->dcapacity_.load(std::memory_order_relaxed);
578 stride = this->dstride_.load(std::memory_order_relaxed);
579 // End of read-only section. Validate seqlock.
580 std::atomic_thread_fence(std::memory_order_acquire);
581 return (state == this->dstate_.load(std::memory_order_relaxed));
584 /// If there was an expansion after ticket was issued, update local variables
585 /// of the lagging operation using the most recent closed array with
586 /// offset <= ticket and return true. Otherwise, return false;
587 bool maybeUpdateFromClosed(
588 const uint64_t state,
589 const uint64_t ticket,
593 int& stride) noexcept {
594 offset = getOffset(state);
595 if (ticket >= offset) {
598 for (int i = getNumClosed(state) - 1; i >= 0; --i) {
599 offset = closed_[i].offset_;
600 if (offset <= ticket) {
601 slots = closed_[i].slots_;
602 cap = closed_[i].capacity_;
603 stride = closed_[i].stride_;
607 // A closed array with offset <= ticket should have been found
615 /// CRTP specialization of MPMCQueueBase
618 typename T, template<typename> class Atom, bool Dynamic> class Derived,
619 typename T, template<typename> class Atom, bool Dynamic>
620 class MPMCQueueBase<Derived<T, Atom, Dynamic>> : boost::noncopyable {
622 // Note: Using CRTP static casts in several functions of this base
623 // template instead of making called functions virtual or duplicating
624 // the code of calling functions in the derived partially specialized
627 static_assert(std::is_nothrow_constructible<T,T&&>::value ||
628 folly::IsRelocatable<T>::value,
629 "T must be relocatable or have a noexcept move constructor");
632 typedef T value_type;
634 using Slot = detail::SingleElementQueue<T,Atom>;
636 explicit MPMCQueueBase(size_t queueCapacity)
637 : capacity_(queueCapacity)
643 if (queueCapacity == 0) {
644 throw std::invalid_argument(
645 "MPMCQueue with explicit capacity 0 is impossible"
646 // Stride computation in derived classes would sigfpe if capacity is 0
650 // ideally this would be a static assert, but g++ doesn't allow it
651 assert(alignof(MPMCQueue<T, Atom>) >= CacheLocality::kFalseSharingRange);
653 static_cast<uint8_t*>(static_cast<void*>(&popTicket_)) -
654 static_cast<uint8_t*>(static_cast<void*>(&pushTicket_)) >=
655 CacheLocality::kFalseSharingRange);
658 /// A default-constructed queue is useful because a usable (non-zero
659 /// capacity) queue can be moved onto it or swapped with it
660 MPMCQueueBase() noexcept
672 /// IMPORTANT: The move constructor is here to make it easier to perform
673 /// the initialization phase, it is not safe to use when there are any
674 /// concurrent accesses (this is not checked).
675 MPMCQueueBase(MPMCQueueBase<Derived<T,Atom,Dynamic>>&& rhs) noexcept
676 : capacity_(rhs.capacity_)
678 , stride_(rhs.stride_)
679 , dstate_(rhs.dstate_.load(std::memory_order_relaxed))
680 , dcapacity_(rhs.dcapacity_.load(std::memory_order_relaxed))
681 , pushTicket_(rhs.pushTicket_.load(std::memory_order_relaxed))
682 , popTicket_(rhs.popTicket_.load(std::memory_order_relaxed))
683 , pushSpinCutoff_(rhs.pushSpinCutoff_.load(std::memory_order_relaxed))
684 , popSpinCutoff_(rhs.popSpinCutoff_.load(std::memory_order_relaxed))
686 // relaxed ops are okay for the previous reads, since rhs queue can't
687 // be in concurrent use
691 rhs.slots_ = nullptr;
693 rhs.dstate_.store(0, std::memory_order_relaxed);
694 rhs.dcapacity_.store(0, std::memory_order_relaxed);
695 rhs.pushTicket_.store(0, std::memory_order_relaxed);
696 rhs.popTicket_.store(0, std::memory_order_relaxed);
697 rhs.pushSpinCutoff_.store(0, std::memory_order_relaxed);
698 rhs.popSpinCutoff_.store(0, std::memory_order_relaxed);
701 /// IMPORTANT: The move operator is here to make it easier to perform
702 /// the initialization phase, it is not safe to use when there are any
703 /// concurrent accesses (this is not checked).
704 MPMCQueueBase<Derived<T,Atom,Dynamic>> const& operator=
705 (MPMCQueueBase<Derived<T,Atom,Dynamic>>&& rhs) {
707 this->~MPMCQueueBase();
708 new (this) MPMCQueueBase(std::move(rhs));
713 /// MPMCQueue can only be safely destroyed when there are no
714 /// pending enqueuers or dequeuers (this is not checked).
719 /// Returns the number of writes (including threads that are blocked waiting
720 /// to write) minus the number of reads (including threads that are blocked
721 /// waiting to read). So effectively, it becomes:
722 /// elements in queue + pending(calls to write) - pending(calls to read).
723 /// If nothing is pending, then the method returns the actual number of
724 /// elements in the queue.
725 /// The returned value can be negative if there are no writers and the queue
726 /// is empty, but there is one reader that is blocked waiting to read (in
727 /// which case, the returned size will be -1).
728 ssize_t size() const noexcept {
729 // since both pushes and pops increase monotonically, we can get a
730 // consistent snapshot either by bracketing a read of popTicket_ with
731 // two reads of pushTicket_ that return the same value, or the other
732 // way around. We maximize our chances by alternately attempting
734 uint64_t pushes = pushTicket_.load(std::memory_order_acquire); // A
735 uint64_t pops = popTicket_.load(std::memory_order_acquire); // B
737 uint64_t nextPushes = pushTicket_.load(std::memory_order_acquire); // C
738 if (pushes == nextPushes) {
739 // pushTicket_ didn't change from A (or the previous C) to C,
740 // so we can linearize at B (or D)
741 return ssize_t(pushes - pops);
744 uint64_t nextPops = popTicket_.load(std::memory_order_acquire); // D
745 if (pops == nextPops) {
746 // popTicket_ didn't chance from B (or the previous D), so we
747 // can linearize at C
748 return ssize_t(pushes - pops);
754 /// Returns true if there are no items available for dequeue
755 bool isEmpty() const noexcept {
759 /// Returns true if there is currently no empty space to enqueue
760 bool isFull() const noexcept {
761 // careful with signed -> unsigned promotion, since size can be negative
762 return size() >= static_cast<ssize_t>(capacity_);
765 /// Returns is a guess at size() for contexts that don't need a precise
766 /// value, such as stats. More specifically, it returns the number of writes
767 /// minus the number of reads, but after reading the number of writes, more
768 /// writers could have came before the number of reads was sampled,
769 /// and this method doesn't protect against such case.
770 /// The returned value can be negative.
771 ssize_t sizeGuess() const noexcept {
772 return writeCount() - readCount();
776 size_t capacity() const noexcept {
780 /// Doesn't change for non-dynamic
781 size_t allocatedCapacity() const noexcept {
785 /// Returns the total number of calls to blockingWrite or successful
786 /// calls to write, including those blockingWrite calls that are
787 /// currently blocking
788 uint64_t writeCount() const noexcept {
789 return pushTicket_.load(std::memory_order_acquire);
792 /// Returns the total number of calls to blockingRead or successful
793 /// calls to read, including those blockingRead calls that are currently
795 uint64_t readCount() const noexcept {
796 return popTicket_.load(std::memory_order_acquire);
799 /// Enqueues a T constructed from args, blocking until space is
800 /// available. Note that this method signature allows enqueue via
801 /// move, if args is a T rvalue, via copy, if args is a T lvalue, or
802 /// via emplacement if args is an initializer list that can be passed
803 /// to a T constructor.
804 template <typename ...Args>
805 void blockingWrite(Args&&... args) noexcept {
806 enqueueWithTicketBase(pushTicket_++, slots_, capacity_, stride_,
807 std::forward<Args>(args)...);
810 /// If an item can be enqueued with no blocking, does so and returns
811 /// true, otherwise returns false. This method is similar to
812 /// writeIfNotFull, but if you don't have a specific need for that
813 /// method you should use this one.
815 /// One of the common usages of this method is to enqueue via the
816 /// move constructor, something like q.write(std::move(x)). If write
817 /// returns false because the queue is full then x has not actually been
818 /// consumed, which looks strange. To understand why it is actually okay
819 /// to use x afterward, remember that std::move is just a typecast that
820 /// provides an rvalue reference that enables use of a move constructor
821 /// or operator. std::move doesn't actually move anything. It could
822 /// more accurately be called std::rvalue_cast or std::move_permission.
823 template <typename ...Args>
824 bool write(Args&&... args) noexcept {
829 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
830 tryObtainReadyPushTicket(ticket, slots, cap, stride)) {
831 // we have pre-validated that the ticket won't block
832 enqueueWithTicketBase(ticket, slots, cap, stride,
833 std::forward<Args>(args)...);
840 template <class Clock, typename... Args>
841 bool tryWriteUntil(const std::chrono::time_point<Clock>& when,
842 Args&&... args) noexcept {
847 if (tryObtainPromisedPushTicketUntil(ticket, slots, cap, stride, when)) {
848 // we have pre-validated that the ticket won't block, or rather that
849 // it won't block longer than it takes another thread to dequeue an
850 // element from the slot it identifies.
851 enqueueWithTicketBase(ticket, slots, cap, stride,
852 std::forward<Args>(args)...);
859 /// If the queue is not full, enqueues and returns true, otherwise
860 /// returns false. Unlike write this method can be blocked by another
861 /// thread, specifically a read that has linearized (been assigned
862 /// a ticket) but not yet completed. If you don't really need this
863 /// function you should probably use write.
865 /// MPMCQueue isn't lock-free, so just because a read operation has
866 /// linearized (and isFull is false) doesn't mean that space has been
867 /// made available for another write. In this situation write will
868 /// return false, but writeIfNotFull will wait for the dequeue to finish.
869 /// This method is required if you are composing queues and managing
870 /// your own wakeup, because it guarantees that after every successful
871 /// write a readIfNotEmpty will succeed.
872 template <typename ...Args>
873 bool writeIfNotFull(Args&&... args) noexcept {
878 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
879 tryObtainPromisedPushTicket(ticket, slots, cap, stride)) {
880 // some other thread is already dequeuing the slot into which we
881 // are going to enqueue, but we might have to wait for them to finish
882 enqueueWithTicketBase(ticket, slots, cap, stride,
883 std::forward<Args>(args)...);
890 /// Moves a dequeued element onto elem, blocking until an element
892 void blockingRead(T& elem) noexcept {
894 static_cast<Derived<T,Atom,Dynamic>*>(this)->
895 blockingReadWithTicket(ticket, elem);
898 /// Same as blockingRead() but also records the ticket nunmer
899 void blockingReadWithTicket(uint64_t& ticket, T& elem) noexcept {
900 assert(capacity_ != 0);
901 ticket = popTicket_++;
902 dequeueWithTicketBase(ticket, slots_, capacity_, stride_, elem);
905 /// If an item can be dequeued with no blocking, does so and returns
906 /// true, otherwise returns false.
907 bool read(T& elem) noexcept {
909 return readAndGetTicket(ticket, elem);
912 /// Same as read() but also records the ticket nunmer
913 bool readAndGetTicket(uint64_t& ticket, T& elem) noexcept {
917 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
918 tryObtainReadyPopTicket(ticket, slots, cap, stride)) {
919 // the ticket has been pre-validated to not block
920 dequeueWithTicketBase(ticket, slots, cap, stride, elem);
927 template <class Clock, typename... Args>
929 const std::chrono::time_point<Clock>& when,
935 if (tryObtainPromisedPopTicketUntil(ticket, slots, cap, stride, when)) {
936 // we have pre-validated that the ticket won't block, or rather that
937 // it won't block longer than it takes another thread to enqueue an
938 // element on the slot it identifies.
939 dequeueWithTicketBase(ticket, slots, cap, stride, elem);
946 /// If the queue is not empty, dequeues and returns true, otherwise
947 /// returns false. If the matching write is still in progress then this
948 /// method may block waiting for it. If you don't rely on being able
949 /// to dequeue (such as by counting completed write) then you should
951 bool readIfNotEmpty(T& elem) noexcept {
956 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
957 tryObtainPromisedPopTicket(ticket, slots, cap, stride)) {
958 // the matching enqueue already has a ticket, but might not be done
959 dequeueWithTicketBase(ticket, slots, cap, stride, elem);
968 /// Once every kAdaptationFreq we will spin longer, to try to estimate
969 /// the proper spin backoff
970 kAdaptationFreq = 128,
972 /// To avoid false sharing in slots_ with neighboring memory
973 /// allocations, we pad it with this many SingleElementQueue-s at
975 kSlotPadding = (CacheLocality::kFalseSharingRange - 1) / sizeof(Slot) + 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_[CacheLocality::kFalseSharingRange - sizeof(Atom<uint32_t>)];
1029 /// We assign tickets in increasing order, but we don't want to
1030 /// access neighboring elements of slots_ because that will lead to
1031 /// false sharing (multiple cores accessing the same cache line even
1032 /// though they aren't accessing the same bytes in that cache line).
1033 /// To avoid this we advance by stride slots per ticket.
1035 /// We need gcd(capacity, stride) to be 1 so that we will use all
1036 /// of the slots. We ensure this by only considering prime strides,
1037 /// which either have no common divisors with capacity or else have
1038 /// a zero remainder after dividing by capacity. That is sufficient
1039 /// to guarantee correctness, but we also want to actually spread the
1040 /// accesses away from each other to avoid false sharing (consider a
1041 /// stride of 7 with a capacity of 8). To that end we try a few taking
1042 /// care to observe that advancing by -1 is as bad as advancing by 1
1043 /// when in comes to false sharing.
1045 /// The simple way to avoid false sharing would be to pad each
1046 /// SingleElementQueue, but since we have capacity_ of them that could
1047 /// waste a lot of space.
1048 static int computeStride(size_t capacity) noexcept {
1049 static const int smallPrimes[] = { 2, 3, 5, 7, 11, 13, 17, 19, 23 };
1053 for (int stride : smallPrimes) {
1054 if ((stride % capacity) == 0 || (capacity % stride) == 0) {
1057 size_t sep = stride % capacity;
1058 sep = std::min(sep, capacity - sep);
1059 if (sep > bestSep) {
1060 bestStride = stride;
1067 /// Returns the index into slots_ that should be used when enqueuing or
1068 /// dequeuing with the specified ticket
1069 size_t idx(uint64_t ticket, size_t cap, int stride) noexcept {
1070 return ((ticket * stride) % cap) + kSlotPadding;
1073 /// Maps an enqueue or dequeue ticket to the turn should be used at the
1074 /// corresponding SingleElementQueue
1075 uint32_t turn(uint64_t ticket, size_t cap) noexcept {
1077 return uint32_t(ticket / cap);
1080 /// Tries to obtain a push ticket for which SingleElementQueue::enqueue
1081 /// won't block. Returns true on immediate success, false on immediate
1083 bool tryObtainReadyPushTicket(
1084 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1086 ticket = pushTicket_.load(std::memory_order_acquire); // A
1091 if (!slots[idx(ticket, cap, stride)]
1092 .mayEnqueue(turn(ticket, cap))) {
1093 // if we call enqueue(ticket, ...) on the SingleElementQueue
1094 // right now it would block, but this might no longer be the next
1095 // ticket. We can increase the chance of tryEnqueue success under
1096 // contention (without blocking) by rechecking the ticket dispenser
1098 ticket = pushTicket_.load(std::memory_order_acquire); // B
1099 if (prev == ticket) {
1100 // mayEnqueue was bracketed by two reads (A or prev B or prev
1101 // failing CAS to B), so we are definitely unable to enqueue
1105 // we will bracket the mayEnqueue check with a read (A or prev B
1106 // or prev failing CAS) and the following CAS. If the CAS fails
1107 // it will effect a load of pushTicket_
1108 if (pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
1115 /// Tries until when to obtain a push ticket for which
1116 /// SingleElementQueue::enqueue won't block. Returns true on success, false
1118 /// ticket is filled on success AND failure.
1119 template <class Clock>
1120 bool tryObtainPromisedPushTicketUntil(
1121 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride,
1122 const std::chrono::time_point<Clock>& when
1124 bool deadlineReached = false;
1125 while (!deadlineReached) {
1126 if (static_cast<Derived<T,Atom,Dynamic>*>(this)->
1127 tryObtainPromisedPushTicket(ticket, slots, cap, stride)) {
1130 // ticket is a blocking ticket until the preceding ticket has been
1131 // processed: wait until this ticket's turn arrives. We have not reserved
1132 // this ticket so we will have to re-attempt to get a non-blocking ticket
1133 // if we wake up before we time-out.
1134 deadlineReached = !slots[idx(ticket, cap, stride)]
1135 .tryWaitForEnqueueTurnUntil(turn(ticket, cap), pushSpinCutoff_,
1136 (ticket % kAdaptationFreq) == 0, when);
1141 /// Tries to obtain a push ticket which can be satisfied if all
1142 /// in-progress pops complete. This function does not block, but
1143 /// blocking may be required when using the returned ticket if some
1144 /// other thread's pop is still in progress (ticket has been granted but
1145 /// pop has not yet completed).
1146 bool tryObtainPromisedPushTicket(
1147 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1149 auto numPushes = pushTicket_.load(std::memory_order_acquire); // A
1155 const auto numPops = popTicket_.load(std::memory_order_acquire); // B
1156 // n will be negative if pops are pending
1157 const int64_t n = int64_t(numPushes - numPops);
1158 if (n >= static_cast<ssize_t>(capacity_)) {
1159 // Full, linearize at B. We don't need to recheck the read we
1160 // performed at A, because if numPushes was stale at B then the
1161 // real numPushes value is even worse
1164 if (pushTicket_.compare_exchange_strong(numPushes, numPushes + 1)) {
1170 /// Tries to obtain a pop ticket for which SingleElementQueue::dequeue
1171 /// won't block. Returns true on immediate success, false on immediate
1173 bool tryObtainReadyPopTicket(
1174 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1176 ticket = popTicket_.load(std::memory_order_acquire);
1181 if (!slots[idx(ticket, cap, stride)]
1182 .mayDequeue(turn(ticket, cap))) {
1184 ticket = popTicket_.load(std::memory_order_acquire);
1185 if (prev == ticket) {
1189 if (popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
1196 /// Tries until when to obtain a pop ticket for which
1197 /// SingleElementQueue::dequeue won't block. Returns true on success, false
1199 /// ticket is filled on success AND failure.
1200 template <class Clock>
1201 bool tryObtainPromisedPopTicketUntil(
1206 const std::chrono::time_point<Clock>& when) noexcept {
1207 bool deadlineReached = false;
1208 while (!deadlineReached) {
1209 if (static_cast<Derived<T, Atom, Dynamic>*>(this)
1210 ->tryObtainPromisedPopTicket(ticket, slots, cap, stride)) {
1213 // ticket is a blocking ticket until the preceding ticket has been
1214 // processed: wait until this ticket's turn arrives. We have not reserved
1215 // this ticket so we will have to re-attempt to get a non-blocking ticket
1216 // if we wake up before we time-out.
1218 !slots[idx(ticket, cap, stride)].tryWaitForDequeueTurnUntil(
1221 (ticket % kAdaptationFreq) == 0,
1227 /// Similar to tryObtainReadyPopTicket, but returns a pop ticket whose
1228 /// corresponding push ticket has already been handed out, rather than
1229 /// returning one whose corresponding push ticket has already been
1230 /// completed. This means that there is a possibility that the caller
1231 /// will block when using the ticket, but it allows the user to rely on
1232 /// the fact that if enqueue has succeeded, tryObtainPromisedPopTicket
1233 /// will return true. The "try" part of this is that we won't have
1234 /// to block waiting for someone to call enqueue, although we might
1235 /// have to block waiting for them to finish executing code inside the
1236 /// MPMCQueue itself.
1237 bool tryObtainPromisedPopTicket(
1238 uint64_t& ticket, Slot*& slots, size_t& cap, int& stride
1240 auto numPops = popTicket_.load(std::memory_order_acquire); // A
1246 const auto numPushes = pushTicket_.load(std::memory_order_acquire); // B
1247 if (numPops >= numPushes) {
1248 // Empty, or empty with pending pops. Linearize at B. We don't
1249 // need to recheck the read we performed at A, because if numPops
1250 // is stale then the fresh value is larger and the >= is still true
1253 if (popTicket_.compare_exchange_strong(numPops, numPops + 1)) {
1259 // Given a ticket, constructs an enqueued item using args
1260 template <typename ...Args>
1261 void enqueueWithTicketBase(
1262 uint64_t ticket, Slot* slots, size_t cap, int stride, Args&&... args
1264 slots[idx(ticket, cap, stride)]
1265 .enqueue(turn(ticket, cap),
1267 (ticket % kAdaptationFreq) == 0,
1268 std::forward<Args>(args)...);
1271 // To support tracking ticket numbers in MPMCPipelineStageImpl
1272 template <typename ...Args>
1273 void enqueueWithTicket(uint64_t ticket, Args&&... args) noexcept {
1274 enqueueWithTicketBase(ticket, slots_, capacity_, stride_,
1275 std::forward<Args>(args)...);
1278 // Given a ticket, dequeues the corresponding element
1279 void dequeueWithTicketBase(
1280 uint64_t ticket, Slot* slots, size_t cap, int stride, T& elem
1283 slots[idx(ticket, cap, stride)]
1284 .dequeue(turn(ticket, cap),
1286 (ticket % kAdaptationFreq) == 0,
1291 /// SingleElementQueue implements a blocking queue that holds at most one
1292 /// item, and that requires its users to assign incrementing identifiers
1293 /// (turns) to each enqueue and dequeue operation. Note that the turns
1294 /// used by SingleElementQueue are doubled inside the TurnSequencer
1295 template <typename T, template <typename> class Atom>
1296 struct SingleElementQueue {
1298 ~SingleElementQueue() noexcept {
1299 if ((sequencer_.uncompletedTurnLSB() & 1) == 1) {
1300 // we are pending a dequeue, so we have a constructed item
1305 /// enqueue using in-place noexcept construction
1306 template <typename ...Args,
1307 typename = typename std::enable_if<
1308 std::is_nothrow_constructible<T,Args...>::value>::type>
1309 void enqueue(const uint32_t turn,
1310 Atom<uint32_t>& spinCutoff,
1311 const bool updateSpinCutoff,
1312 Args&&... args) noexcept {
1313 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
1314 new (&contents_) T(std::forward<Args>(args)...);
1315 sequencer_.completeTurn(turn * 2);
1318 /// enqueue using move construction, either real (if
1319 /// is_nothrow_move_constructible) or simulated using relocation and
1320 /// default construction (if IsRelocatable and is_nothrow_constructible)
1322 typename = typename std::enable_if<
1323 (folly::IsRelocatable<T>::value &&
1324 std::is_nothrow_constructible<T>::value) ||
1325 std::is_nothrow_constructible<T, T&&>::value>::type>
1327 const uint32_t turn,
1328 Atom<uint32_t>& spinCutoff,
1329 const bool updateSpinCutoff,
1330 T&& goner) noexcept {
1336 typename std::conditional<std::is_nothrow_constructible<T,T&&>::value,
1337 ImplByMove, ImplByRelocation>::type());
1340 /// Waits until either:
1341 /// 1: the dequeue turn preceding the given enqueue turn has arrived
1342 /// 2: the given deadline has arrived
1343 /// Case 1 returns true, case 2 returns false.
1344 template <class Clock>
1345 bool tryWaitForEnqueueTurnUntil(
1346 const uint32_t turn,
1347 Atom<uint32_t>& spinCutoff,
1348 const bool updateSpinCutoff,
1349 const std::chrono::time_point<Clock>& when) noexcept {
1350 return sequencer_.tryWaitForTurn(
1351 turn * 2, spinCutoff, updateSpinCutoff, &when) !=
1352 TurnSequencer<Atom>::TryWaitResult::TIMEDOUT;
1355 bool mayEnqueue(const uint32_t turn) const noexcept {
1356 return sequencer_.isTurn(turn * 2);
1359 void dequeue(uint32_t turn,
1360 Atom<uint32_t>& spinCutoff,
1361 const bool updateSpinCutoff,
1367 typename std::conditional<folly::IsRelocatable<T>::value,
1369 ImplByMove>::type());
1372 /// Waits until either:
1373 /// 1: the enqueue turn preceding the given dequeue turn has arrived
1374 /// 2: the given deadline has arrived
1375 /// Case 1 returns true, case 2 returns false.
1376 template <class Clock>
1377 bool tryWaitForDequeueTurnUntil(
1378 const uint32_t turn,
1379 Atom<uint32_t>& spinCutoff,
1380 const bool updateSpinCutoff,
1381 const std::chrono::time_point<Clock>& when) noexcept {
1382 return sequencer_.tryWaitForTurn(
1383 turn * 2 + 1, spinCutoff, updateSpinCutoff, &when) !=
1384 TurnSequencer<Atom>::TryWaitResult::TIMEDOUT;
1387 bool mayDequeue(const uint32_t turn) const noexcept {
1388 return sequencer_.isTurn(turn * 2 + 1);
1392 /// Storage for a T constructed with placement new
1393 typename std::aligned_storage<sizeof(T),alignof(T)>::type contents_;
1395 /// Even turns are pushes, odd turns are pops
1396 TurnSequencer<Atom> sequencer_;
1399 return static_cast<T*>(static_cast<void*>(&contents_));
1402 void destroyContents() noexcept {
1406 // g++ doesn't seem to have std::is_nothrow_destructible yet
1409 memset(&contents_, 'Q', sizeof(T));
1413 /// Tag classes for dispatching to enqueue/dequeue implementation.
1414 struct ImplByRelocation {};
1415 struct ImplByMove {};
1417 /// enqueue using nothrow move construction.
1418 void enqueueImpl(const uint32_t turn,
1419 Atom<uint32_t>& spinCutoff,
1420 const bool updateSpinCutoff,
1422 ImplByMove) noexcept {
1423 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
1424 new (&contents_) T(std::move(goner));
1425 sequencer_.completeTurn(turn * 2);
1428 /// enqueue by simulating nothrow move with relocation, followed by
1429 /// default construction to a noexcept relocation.
1430 void enqueueImpl(const uint32_t turn,
1431 Atom<uint32_t>& spinCutoff,
1432 const bool updateSpinCutoff,
1434 ImplByRelocation) noexcept {
1435 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
1436 memcpy(&contents_, &goner, sizeof(T));
1437 sequencer_.completeTurn(turn * 2);
1441 /// dequeue by destructing followed by relocation. This version is preferred,
1442 /// because as much work as possible can be done before waiting.
1443 void dequeueImpl(uint32_t turn,
1444 Atom<uint32_t>& spinCutoff,
1445 const bool updateSpinCutoff,
1447 ImplByRelocation) noexcept {
1451 // unlikely, but if we don't complete our turn the queue will die
1453 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
1454 memcpy(&elem, &contents_, sizeof(T));
1455 sequencer_.completeTurn(turn * 2 + 1);
1458 /// dequeue by nothrow move assignment.
1459 void dequeueImpl(uint32_t turn,
1460 Atom<uint32_t>& spinCutoff,
1461 const bool updateSpinCutoff,
1463 ImplByMove) noexcept {
1464 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
1465 elem = std::move(*ptr());
1467 sequencer_.completeTurn(turn * 2 + 1);
1471 } // namespace detail
1473 } // namespace folly