2 * Copyright 2014 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 <linux/futex.h>
27 #include <sys/syscall.h>
28 #include <type_traits>
31 #include <folly/Traits.h>
32 #include <folly/detail/CacheLocality.h>
33 #include <folly/detail/Futex.h>
39 template<typename T, template<typename> class Atom>
40 class SingleElementQueue;
42 template <typename T> class MPMCPipelineStageImpl;
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 /// The underlying implementation uses a ticket dispenser for the head and
53 /// the tail, spreading accesses across N single-element queues to produce
54 /// a queue with capacity N. The ticket dispensers use atomic increment,
55 /// which is more robust to contention than a CAS loop. Each of the
56 /// single-element queues uses its own CAS to serialize access, with an
57 /// adaptive spin cutoff. When spinning fails on a single-element queue
58 /// it uses futex()'s _BITSET operations to reduce unnecessary wakeups
59 /// even if multiple waiters are present on an individual queue (such as
60 /// when the MPMCQueue's capacity is smaller than the number of enqueuers
63 /// NOEXCEPT INTERACTION: Ticket-based queues separate the assignment
64 /// of In benchmarks (contained in tao/queues/ConcurrentQueueTests)
65 /// it handles 1 to 1, 1 to N, N to 1, and N to M thread counts better
66 /// than any of the alternatives present in fbcode, for both small (~10)
67 /// and large capacities. In these benchmarks it is also faster than
68 /// tbb::concurrent_bounded_queue for all configurations. When there are
69 /// many more threads than cores, MPMCQueue is _much_ faster than the tbb
70 /// queue because it uses futex() to block and unblock waiting threads,
71 /// rather than spinning with sched_yield.
73 /// queue positions from the actual construction of the in-queue elements,
74 /// which means that the T constructor used during enqueue must not throw
75 /// an exception. This is enforced at compile time using type traits,
76 /// which requires that T be adorned with accurate noexcept information.
77 /// If your type does not use noexcept, you will have to wrap it in
78 /// something that provides the guarantee. We provide an alternate
79 /// safe implementation for types that don't use noexcept but that are
80 /// marked folly::IsRelocatable and boost::has_nothrow_constructor,
81 /// which is common for folly types. In particular, if you can declare
82 /// FOLLY_ASSUME_FBVECTOR_COMPATIBLE then your type can be put in
85 template<typename> class Atom = std::atomic>
86 class MPMCQueue : boost::noncopyable {
88 static_assert(std::is_nothrow_constructible<T,T&&>::value ||
89 folly::IsRelocatable<T>::value,
90 "T must be relocatable or have a noexcept move constructor");
92 friend class detail::MPMCPipelineStageImpl<T>;
96 explicit MPMCQueue(size_t queueCapacity)
97 : capacity_(queueCapacity)
98 , slots_(new detail::SingleElementQueue<T,Atom>[queueCapacity +
100 , stride_(computeStride(queueCapacity))
106 // ideally this would be a static assert, but g++ doesn't allow it
107 assert(alignof(MPMCQueue<T,Atom>)
108 >= detail::CacheLocality::kFalseSharingRange);
109 assert(static_cast<uint8_t*>(static_cast<void*>(&popTicket_))
110 - static_cast<uint8_t*>(static_cast<void*>(&pushTicket_))
111 >= detail::CacheLocality::kFalseSharingRange);
114 /// A default-constructed queue is useful because a usable (non-zero
115 /// capacity) queue can be moved onto it or swapped with it
126 /// IMPORTANT: The move constructor is here to make it easier to perform
127 /// the initialization phase, it is not safe to use when there are any
128 /// concurrent accesses (this is not checked).
129 MPMCQueue(MPMCQueue<T,Atom>&& rhs) noexcept
130 : capacity_(rhs.capacity_)
132 , stride_(rhs.stride_)
133 , pushTicket_(rhs.pushTicket_.load(std::memory_order_relaxed))
134 , popTicket_(rhs.popTicket_.load(std::memory_order_relaxed))
135 , pushSpinCutoff_(rhs.pushSpinCutoff_.load(std::memory_order_relaxed))
136 , popSpinCutoff_(rhs.popSpinCutoff_.load(std::memory_order_relaxed))
138 // relaxed ops are okay for the previous reads, since rhs queue can't
139 // be in concurrent use
143 rhs.slots_ = nullptr;
145 rhs.pushTicket_.store(0, std::memory_order_relaxed);
146 rhs.popTicket_.store(0, std::memory_order_relaxed);
147 rhs.pushSpinCutoff_.store(0, std::memory_order_relaxed);
148 rhs.popSpinCutoff_.store(0, std::memory_order_relaxed);
151 /// IMPORTANT: The move operator is here to make it easier to perform
152 /// the initialization phase, it is not safe to use when there are any
153 /// concurrent accesses (this is not checked).
154 MPMCQueue<T,Atom> const& operator= (MPMCQueue<T,Atom>&& rhs) {
157 new (this) MPMCQueue(std::move(rhs));
162 /// MPMCQueue can only be safely destroyed when there are no
163 /// pending enqueuers or dequeuers (this is not checked).
168 /// Returns the number of successful reads minus the number of successful
169 /// writes. Waiting blockingRead and blockingWrite calls are included,
170 /// so this value can be negative.
171 ssize_t size() const noexcept {
172 // since both pushes and pops increase monotonically, we can get a
173 // consistent snapshot either by bracketing a read of popTicket_ with
174 // two reads of pushTicket_ that return the same value, or the other
175 // way around. We maximize our chances by alternately attempting
177 uint64_t pushes = pushTicket_.load(std::memory_order_acquire); // A
178 uint64_t pops = popTicket_.load(std::memory_order_acquire); // B
180 uint64_t nextPushes = pushTicket_.load(std::memory_order_acquire); // C
181 if (pushes == nextPushes) {
182 // pushTicket_ didn't change from A (or the previous C) to C,
183 // so we can linearize at B (or D)
184 return pushes - pops;
187 uint64_t nextPops = popTicket_.load(std::memory_order_acquire); // D
188 if (pops == nextPops) {
189 // popTicket_ didn't chance from B (or the previous D), so we
190 // can linearize at C
191 return pushes - pops;
197 /// Returns true if there are no items available for dequeue
198 bool isEmpty() const noexcept {
202 /// Returns true if there is currently no empty space to enqueue
203 bool isFull() const noexcept {
204 // careful with signed -> unsigned promotion, since size can be negative
205 return size() >= static_cast<ssize_t>(capacity_);
208 /// Returns is a guess at size() for contexts that don't need a precise
209 /// value, such as stats.
210 ssize_t sizeGuess() const noexcept {
211 return writeCount() - readCount();
215 size_t capacity() const noexcept {
219 /// Returns the total number of calls to blockingWrite or successful
220 /// calls to write, including those blockingWrite calls that are
221 /// currently blocking
222 uint64_t writeCount() const noexcept {
223 return pushTicket_.load(std::memory_order_acquire);
226 /// Returns the total number of calls to blockingRead or successful
227 /// calls to read, including those blockingRead calls that are currently
229 uint64_t readCount() const noexcept {
230 return popTicket_.load(std::memory_order_acquire);
233 /// Enqueues a T constructed from args, blocking until space is
234 /// available. Note that this method signature allows enqueue via
235 /// move, if args is a T rvalue, via copy, if args is a T lvalue, or
236 /// via emplacement if args is an initializer list that can be passed
237 /// to a T constructor.
238 template <typename ...Args>
239 void blockingWrite(Args&&... args) noexcept {
240 enqueueWithTicket(pushTicket_++, std::forward<Args>(args)...);
243 /// If an item can be enqueued with no blocking, does so and returns
244 /// true, otherwise returns false. This method is similar to
245 /// writeIfNotFull, but if you don't have a specific need for that
246 /// method you should use this one.
248 /// One of the common usages of this method is to enqueue via the
249 /// move constructor, something like q.write(std::move(x)). If write
250 /// returns false because the queue is full then x has not actually been
251 /// consumed, which looks strange. To understand why it is actually okay
252 /// to use x afterward, remember that std::move is just a typecast that
253 /// provides an rvalue reference that enables use of a move constructor
254 /// or operator. std::move doesn't actually move anything. It could
255 /// more accurately be called std::rvalue_cast or std::move_permission.
256 template <typename ...Args>
257 bool write(Args&&... args) noexcept {
259 if (tryObtainReadyPushTicket(ticket)) {
260 // we have pre-validated that the ticket won't block
261 enqueueWithTicket(ticket, std::forward<Args>(args)...);
268 /// If the queue is not full, enqueues and returns true, otherwise
269 /// returns false. Unlike write this method can be blocked by another
270 /// thread, specifically a read that has linearized (been assigned
271 /// a ticket) but not yet completed. If you don't really need this
272 /// function you should probably use write.
274 /// MPMCQueue isn't lock-free, so just because a read operation has
275 /// linearized (and isFull is false) doesn't mean that space has been
276 /// made available for another write. In this situation write will
277 /// return false, but writeIfNotFull will wait for the dequeue to finish.
278 /// This method is required if you are composing queues and managing
279 /// your own wakeup, because it guarantees that after every successful
280 /// write a readIfNotFull will succeed.
281 template <typename ...Args>
282 bool writeIfNotFull(Args&&... args) noexcept {
284 if (tryObtainPromisedPushTicket(ticket)) {
285 // some other thread is already dequeuing the slot into which we
286 // are going to enqueue, but we might have to wait for them to finish
287 enqueueWithTicket(ticket, std::forward<Args>(args)...);
294 /// Moves a dequeued element onto elem, blocking until an element
296 void blockingRead(T& elem) noexcept {
297 dequeueWithTicket(popTicket_++, elem);
300 /// If an item can be dequeued with no blocking, does so and returns
301 /// true, otherwise returns false.
302 bool read(T& elem) noexcept {
304 if (tryObtainReadyPopTicket(ticket)) {
305 // the ticket has been pre-validated to not block
306 dequeueWithTicket(ticket, elem);
313 /// If the queue is not empty, dequeues and returns true, otherwise
314 /// returns false. If the matching write is still in progress then this
315 /// method may block waiting for it. If you don't rely on being able
316 /// to dequeue (such as by counting completed write) then you should
318 bool readIfNotEmpty(T& elem) noexcept {
320 if (tryObtainPromisedPopTicket(ticket)) {
321 // the matching enqueue already has a ticket, but might not be done
322 dequeueWithTicket(ticket, elem);
331 /// Once every kAdaptationFreq we will spin longer, to try to estimate
332 /// the proper spin backoff
333 kAdaptationFreq = 128,
335 /// To avoid false sharing in slots_ with neighboring memory
336 /// allocations, we pad it with this many SingleElementQueue-s at
338 kSlotPadding = (detail::CacheLocality::kFalseSharingRange - 1)
339 / sizeof(detail::SingleElementQueue<T,Atom>) + 1
342 /// The maximum number of items in the queue at once
343 size_t FOLLY_ALIGN_TO_AVOID_FALSE_SHARING capacity_;
345 /// An array of capacity_ SingleElementQueue-s, each of which holds
346 /// either 0 or 1 item. We over-allocate by 2 * kSlotPadding and don't
347 /// touch the slots at either end, to avoid false sharing
348 detail::SingleElementQueue<T,Atom>* slots_;
350 /// The number of slots_ indices that we advance for each ticket, to
351 /// avoid false sharing. Ideally slots_[i] and slots_[i + stride_]
352 /// aren't on the same cache line
355 /// Enqueuers get tickets from here
356 Atom<uint64_t> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING pushTicket_;
358 /// Dequeuers get tickets from here
359 Atom<uint64_t> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING popTicket_;
361 /// This is how many times we will spin before using FUTEX_WAIT when
362 /// the queue is full on enqueue, adaptively computed by occasionally
363 /// spinning for longer and smoothing with an exponential moving average
364 Atom<int> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING pushSpinCutoff_;
366 /// The adaptive spin cutoff when the queue is empty on dequeue
367 Atom<int> FOLLY_ALIGN_TO_AVOID_FALSE_SHARING popSpinCutoff_;
369 /// Alignment doesn't prevent false sharing at the end of the struct,
370 /// so fill out the last cache line
371 char padding_[detail::CacheLocality::kFalseSharingRange - sizeof(Atom<int>)];
374 /// We assign tickets in increasing order, but we don't want to
375 /// access neighboring elements of slots_ because that will lead to
376 /// false sharing (multiple cores accessing the same cache line even
377 /// though they aren't accessing the same bytes in that cache line).
378 /// To avoid this we advance by stride slots per ticket.
380 /// We need gcd(capacity, stride) to be 1 so that we will use all
381 /// of the slots. We ensure this by only considering prime strides,
382 /// which either have no common divisors with capacity or else have
383 /// a zero remainder after dividing by capacity. That is sufficient
384 /// to guarantee correctness, but we also want to actually spread the
385 /// accesses away from each other to avoid false sharing (consider a
386 /// stride of 7 with a capacity of 8). To that end we try a few taking
387 /// care to observe that advancing by -1 is as bad as advancing by 1
388 /// when in comes to false sharing.
390 /// The simple way to avoid false sharing would be to pad each
391 /// SingleElementQueue, but since we have capacity_ of them that could
392 /// waste a lot of space.
393 static int computeStride(size_t capacity) noexcept {
394 static const int smallPrimes[] = { 2, 3, 5, 7, 11, 13, 17, 19, 23 };
398 for (int stride : smallPrimes) {
399 if ((stride % capacity) == 0 || (capacity % stride) == 0) {
402 size_t sep = stride % capacity;
403 sep = std::min(sep, capacity - sep);
412 /// Returns the index into slots_ that should be used when enqueuing or
413 /// dequeuing with the specified ticket
414 size_t idx(uint64_t ticket) noexcept {
415 return ((ticket * stride_) % capacity_) + kSlotPadding;
418 /// Maps an enqueue or dequeue ticket to the turn should be used at the
419 /// corresponding SingleElementQueue
420 uint32_t turn(uint64_t ticket) noexcept {
421 return ticket / capacity_;
424 /// Tries to obtain a push ticket for which SingleElementQueue::enqueue
425 /// won't block. Returns true on immediate success, false on immediate
427 bool tryObtainReadyPushTicket(uint64_t& rv) noexcept {
428 auto ticket = pushTicket_.load(std::memory_order_acquire); // A
430 if (!slots_[idx(ticket)].mayEnqueue(turn(ticket))) {
431 // if we call enqueue(ticket, ...) on the SingleElementQueue
432 // right now it would block, but this might no longer be the next
433 // ticket. We can increase the chance of tryEnqueue success under
434 // contention (without blocking) by rechecking the ticket dispenser
436 ticket = pushTicket_.load(std::memory_order_acquire); // B
437 if (prev == ticket) {
438 // mayEnqueue was bracketed by two reads (A or prev B or prev
439 // failing CAS to B), so we are definitely unable to enqueue
443 // we will bracket the mayEnqueue check with a read (A or prev B
444 // or prev failing CAS) and the following CAS. If the CAS fails
445 // it will effect a load of pushTicket_
446 if (pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
454 /// Tries to obtain a push ticket which can be satisfied if all
455 /// in-progress pops complete. This function does not block, but
456 /// blocking may be required when using the returned ticket if some
457 /// other thread's pop is still in progress (ticket has been granted but
458 /// pop has not yet completed).
459 bool tryObtainPromisedPushTicket(uint64_t& rv) noexcept {
460 auto numPushes = pushTicket_.load(std::memory_order_acquire); // A
462 auto numPops = popTicket_.load(std::memory_order_acquire); // B
463 // n will be negative if pops are pending
464 int64_t n = numPushes - numPops;
465 if (n >= static_cast<ssize_t>(capacity_)) {
466 // Full, linearize at B. We don't need to recheck the read we
467 // performed at A, because if numPushes was stale at B then the
468 // real numPushes value is even worse
471 if (pushTicket_.compare_exchange_strong(numPushes, numPushes + 1)) {
478 /// Tries to obtain a pop ticket for which SingleElementQueue::dequeue
479 /// won't block. Returns true on immediate success, false on immediate
481 bool tryObtainReadyPopTicket(uint64_t& rv) noexcept {
482 auto ticket = popTicket_.load(std::memory_order_acquire);
484 if (!slots_[idx(ticket)].mayDequeue(turn(ticket))) {
486 ticket = popTicket_.load(std::memory_order_acquire);
487 if (prev == ticket) {
491 if (popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
499 /// Similar to tryObtainReadyPopTicket, but returns a pop ticket whose
500 /// corresponding push ticket has already been handed out, rather than
501 /// returning one whose corresponding push ticket has already been
502 /// completed. This means that there is a possibility that the caller
503 /// will block when using the ticket, but it allows the user to rely on
504 /// the fact that if enqueue has succeeded, tryObtainPromisedPopTicket
505 /// will return true. The "try" part of this is that we won't have
506 /// to block waiting for someone to call enqueue, although we might
507 /// have to block waiting for them to finish executing code inside the
508 /// MPMCQueue itself.
509 bool tryObtainPromisedPopTicket(uint64_t& rv) noexcept {
510 auto numPops = popTicket_.load(std::memory_order_acquire); // A
512 auto numPushes = pushTicket_.load(std::memory_order_acquire); // B
513 if (numPops >= numPushes) {
514 // Empty, or empty with pending pops. Linearize at B. We don't
515 // need to recheck the read we performed at A, because if numPops
516 // is stale then the fresh value is larger and the >= is still true
519 if (popTicket_.compare_exchange_strong(numPops, numPops + 1)) {
526 // Given a ticket, constructs an enqueued item using args
527 template <typename ...Args>
528 void enqueueWithTicket(uint64_t ticket, Args&&... args) noexcept {
529 slots_[idx(ticket)].enqueue(turn(ticket),
531 (ticket % kAdaptationFreq) == 0,
532 std::forward<Args>(args)...);
535 // Given a ticket, dequeues the corresponding element
536 void dequeueWithTicket(uint64_t ticket, T& elem) noexcept {
537 slots_[idx(ticket)].dequeue(turn(ticket),
539 (ticket % kAdaptationFreq) == 0,
547 /// A TurnSequencer allows threads to order their execution according to
548 /// a monotonically increasing (with wraparound) "turn" value. The two
549 /// operations provided are to wait for turn T, and to move to the next
550 /// turn. Every thread that is waiting for T must have arrived before
551 /// that turn is marked completed (for MPMCQueue only one thread waits
552 /// for any particular turn, so this is trivially true).
554 /// TurnSequencer's state_ holds 26 bits of the current turn (shifted
555 /// left by 6), along with a 6 bit saturating value that records the
556 /// maximum waiter minus the current turn. Wraparound of the turn space
557 /// is expected and handled. This allows us to atomically adjust the
558 /// number of outstanding waiters when we perform a FUTEX_WAKE operation.
559 /// Compare this strategy to sem_t's separate num_waiters field, which
560 /// isn't decremented until after the waiting thread gets scheduled,
561 /// during which time more enqueues might have occurred and made pointless
562 /// FUTEX_WAKE calls.
564 /// TurnSequencer uses futex() directly. It is optimized for the
565 /// case that the highest awaited turn is 32 or less higher than the
566 /// current turn. We use the FUTEX_WAIT_BITSET variant, which lets
567 /// us embed 32 separate wakeup channels in a single futex. See
568 /// http://locklessinc.com/articles/futex_cheat_sheet for a description.
570 /// We only need to keep exact track of the delta between the current
571 /// turn and the maximum waiter for the 32 turns that follow the current
572 /// one, because waiters at turn t+32 will be awoken at turn t. At that
573 /// point they can then adjust the delta using the higher base. Since we
574 /// need to encode waiter deltas of 0 to 32 inclusive, we use 6 bits.
575 /// We actually store waiter deltas up to 63, since that might reduce
576 /// the number of CAS operations a tiny bit.
578 /// To avoid some futex() calls entirely, TurnSequencer uses an adaptive
579 /// spin cutoff before waiting. The overheads (and convergence rate)
580 /// of separately tracking the spin cutoff for each TurnSequencer would
581 /// be prohibitive, so the actual storage is passed in as a parameter and
582 /// updated atomically. This also lets the caller use different adaptive
583 /// cutoffs for different operations (read versus write, for example).
584 /// To avoid contention, the spin cutoff is only updated when requested
586 template <template<typename> class Atom>
587 struct TurnSequencer {
588 explicit TurnSequencer(const uint32_t firstTurn = 0) noexcept
589 : state_(encode(firstTurn << kTurnShift, 0))
592 /// Returns true iff a call to waitForTurn(turn, ...) won't block
593 bool isTurn(const uint32_t turn) const noexcept {
594 auto state = state_.load(std::memory_order_acquire);
595 return decodeCurrentSturn(state) == (turn << kTurnShift);
598 // Internally we always work with shifted turn values, which makes the
599 // truncation and wraparound work correctly. This leaves us bits at
600 // the bottom to store the number of waiters. We call shifted turns
601 // "sturns" inside this class.
603 /// Blocks the current thread until turn has arrived. If
604 /// updateSpinCutoff is true then this will spin for up to kMaxSpins tries
605 /// before blocking and will adjust spinCutoff based on the results,
606 /// otherwise it will spin for at most spinCutoff spins.
607 void waitForTurn(const uint32_t turn,
608 Atom<int>& spinCutoff,
609 const bool updateSpinCutoff) noexcept {
610 int prevThresh = spinCutoff.load(std::memory_order_relaxed);
611 const int effectiveSpinCutoff =
612 updateSpinCutoff || prevThresh == 0 ? kMaxSpins : prevThresh;
615 const uint32_t sturn = turn << kTurnShift;
616 for (tries = 0; ; ++tries) {
617 uint32_t state = state_.load(std::memory_order_acquire);
618 uint32_t current_sturn = decodeCurrentSturn(state);
619 if (current_sturn == sturn) {
623 // wrap-safe version of assert(current_sturn < sturn)
624 assert(sturn - current_sturn < std::numeric_limits<uint32_t>::max() / 2);
626 // the first effectSpinCutoff tries are spins, after that we will
627 // record ourself as a waiter and block with futexWait
628 if (tries < effectiveSpinCutoff) {
629 asm volatile ("pause");
633 uint32_t current_max_waiter_delta = decodeMaxWaitersDelta(state);
634 uint32_t our_waiter_delta = (sturn - current_sturn) >> kTurnShift;
636 if (our_waiter_delta <= current_max_waiter_delta) {
637 // state already records us as waiters, probably because this
638 // isn't our first time around this loop
641 new_state = encode(current_sturn, our_waiter_delta);
642 if (state != new_state &&
643 !state_.compare_exchange_strong(state, new_state)) {
647 state_.futexWait(new_state, futexChannel(turn));
650 if (updateSpinCutoff || prevThresh == 0) {
651 // if we hit kMaxSpins then spinning was pointless, so the right
652 // spinCutoff is kMinSpins
654 if (tries >= kMaxSpins) {
657 // to account for variations, we allow ourself to spin 2*N when
658 // we think that N is actually required in order to succeed
659 target = std::min(int{kMaxSpins}, std::max(int{kMinSpins}, tries * 2));
662 if (prevThresh == 0) {
666 // try once, keep moving if CAS fails. Exponential moving average
668 spinCutoff.compare_exchange_weak(
669 prevThresh, prevThresh + (target - prevThresh) / 8);
674 /// Unblocks a thread running waitForTurn(turn + 1)
675 void completeTurn(const uint32_t turn) noexcept {
676 uint32_t state = state_.load(std::memory_order_acquire);
678 assert(state == encode(turn << kTurnShift, decodeMaxWaitersDelta(state)));
679 uint32_t max_waiter_delta = decodeMaxWaitersDelta(state);
680 uint32_t new_state = encode(
681 (turn + 1) << kTurnShift,
682 max_waiter_delta == 0 ? 0 : max_waiter_delta - 1);
683 if (state_.compare_exchange_strong(state, new_state)) {
684 if (max_waiter_delta != 0) {
685 state_.futexWake(std::numeric_limits<int>::max(),
686 futexChannel(turn + 1));
690 // failing compare_exchange_strong updates first arg to the value
691 // that caused the failure, so no need to reread state_
695 /// Returns the least-most significant byte of the current uncompleted
696 /// turn. The full 32 bit turn cannot be recovered.
697 uint8_t uncompletedTurnLSB() const noexcept {
698 return state_.load(std::memory_order_acquire) >> kTurnShift;
703 /// kTurnShift counts the bits that are stolen to record the delta
704 /// between the current turn and the maximum waiter. It needs to be big
705 /// enough to record wait deltas of 0 to 32 inclusive. Waiters more
706 /// than 32 in the future will be woken up 32*n turns early (since
707 /// their BITSET will hit) and will adjust the waiter count again.
708 /// We go a bit beyond and let the waiter count go up to 63, which
709 /// is free and might save us a few CAS
711 kWaitersMask = (1 << kTurnShift) - 1,
713 /// The minimum spin count that we will adaptively select
716 /// The maximum spin count that we will adaptively select, and the
717 /// spin count that will be used when probing to get a new data point
718 /// for the adaptation
722 /// This holds both the current turn, and the highest waiting turn,
723 /// stored as (current_turn << 6) | min(63, max(waited_turn - current_turn))
726 /// Returns the bitmask to pass futexWait or futexWake when communicating
727 /// about the specified turn
728 int futexChannel(uint32_t turn) const noexcept {
729 return 1 << (turn & 31);
732 uint32_t decodeCurrentSturn(uint32_t state) const noexcept {
733 return state & ~kWaitersMask;
736 uint32_t decodeMaxWaitersDelta(uint32_t state) const noexcept {
737 return state & kWaitersMask;
740 uint32_t encode(uint32_t currentSturn, uint32_t maxWaiterD) const noexcept {
741 return currentSturn | std::min(uint32_t{ kWaitersMask }, maxWaiterD);
746 /// SingleElementQueue implements a blocking queue that holds at most one
747 /// item, and that requires its users to assign incrementing identifiers
748 /// (turns) to each enqueue and dequeue operation. Note that the turns
749 /// used by SingleElementQueue are doubled inside the TurnSequencer
750 template <typename T, template <typename> class Atom>
751 struct SingleElementQueue {
753 ~SingleElementQueue() noexcept {
754 if ((sequencer_.uncompletedTurnLSB() & 1) == 1) {
755 // we are pending a dequeue, so we have a constructed item
760 /// enqueue using in-place noexcept construction
761 template <typename ...Args,
762 typename = typename std::enable_if<
763 std::is_nothrow_constructible<T,Args...>::value>::type>
764 void enqueue(const uint32_t turn,
765 Atom<int>& spinCutoff,
766 const bool updateSpinCutoff,
767 Args&&... args) noexcept {
768 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
769 new (&contents_) T(std::forward<Args>(args)...);
770 sequencer_.completeTurn(turn * 2);
773 /// enqueue using move construction, either real (if
774 /// is_nothrow_move_constructible) or simulated using relocation and
775 /// default construction (if IsRelocatable and has_nothrow_constructor)
776 template <typename = typename std::enable_if<
777 (folly::IsRelocatable<T>::value &&
778 boost::has_nothrow_constructor<T>::value) ||
779 std::is_nothrow_constructible<T,T&&>::value>::type>
780 void enqueue(const uint32_t turn,
781 Atom<int>& spinCutoff,
782 const bool updateSpinCutoff,
783 T&& goner) noexcept {
784 if (std::is_nothrow_constructible<T,T&&>::value) {
786 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
787 new (&contents_) T(std::move(goner));
788 sequencer_.completeTurn(turn * 2);
790 // simulate nothrow move with relocation, followed by default
791 // construction to fill the gap we created
792 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
793 memcpy(&contents_, &goner, sizeof(T));
794 sequencer_.completeTurn(turn * 2);
799 bool mayEnqueue(const uint32_t turn) const noexcept {
800 return sequencer_.isTurn(turn * 2);
803 void dequeue(uint32_t turn,
804 Atom<int>& spinCutoff,
805 const bool updateSpinCutoff,
807 if (folly::IsRelocatable<T>::value) {
808 // this version is preferred, because we do as much work as possible
813 // unlikely, but if we don't complete our turn the queue will die
815 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
816 memcpy(&elem, &contents_, sizeof(T));
817 sequencer_.completeTurn(turn * 2 + 1);
819 // use nothrow move assignment
820 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
821 elem = std::move(*ptr());
823 sequencer_.completeTurn(turn * 2 + 1);
827 bool mayDequeue(const uint32_t turn) const noexcept {
828 return sequencer_.isTurn(turn * 2 + 1);
832 /// Storage for a T constructed with placement new
833 typename std::aligned_storage<sizeof(T),alignof(T)>::type contents_;
835 /// Even turns are pushes, odd turns are pops
836 TurnSequencer<Atom> sequencer_;
839 return static_cast<T*>(static_cast<void*>(&contents_));
842 void destroyContents() noexcept {
846 // g++ doesn't seem to have std::is_nothrow_destructible yet
849 memset(&contents_, 'Q', sizeof(T));
854 } // namespace detail