2 * Copyright 2013 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>
30 #include <folly/Traits.h>
31 #include <folly/detail/Futex.h>
37 template<typename T, template<typename> class Atom>
38 class SingleElementQueue;
40 template <typename T> class MPMCPipelineStageImpl;
44 /// MPMCQueue<T> is a high-performance bounded concurrent queue that
45 /// supports multiple producers, multiple consumers, and optional blocking.
46 /// The queue has a fixed capacity, for which all memory will be allocated
47 /// up front. The bulk of the work of enqueuing and dequeuing can be
48 /// performed in parallel.
50 /// The underlying implementation uses a ticket dispenser for the head and
51 /// the tail, spreading accesses across N single-element queues to produce
52 /// a queue with capacity N. The ticket dispensers use atomic increment,
53 /// which is more robust to contention than a CAS loop. Each of the
54 /// single-element queues uses its own CAS to serialize access, with an
55 /// adaptive spin cutoff. When spinning fails on a single-element queue
56 /// it uses futex()'s _BITSET operations to reduce unnecessary wakeups
57 /// even if multiple waiters are present on an individual queue (such as
58 /// when the MPMCQueue's capacity is smaller than the number of enqueuers
61 /// NOEXCEPT INTERACTION: Ticket-based queues separate the assignment
62 /// of In benchmarks (contained in tao/queues/ConcurrentQueueTests)
63 /// it handles 1 to 1, 1 to N, N to 1, and N to M thread counts better
64 /// than any of the alternatives present in fbcode, for both small (~10)
65 /// and large capacities. In these benchmarks it is also faster than
66 /// tbb::concurrent_bounded_queue for all configurations. When there are
67 /// many more threads than cores, MPMCQueue is _much_ faster than the tbb
68 /// queue because it uses futex() to block and unblock waiting threads,
69 /// rather than spinning with sched_yield.
71 /// queue positions from the actual construction of the in-queue elements,
72 /// which means that the T constructor used during enqueue must not throw
73 /// an exception. This is enforced at compile time using type traits,
74 /// which requires that T be adorned with accurate noexcept information.
75 /// If your type does not use noexcept, you will have to wrap it in
76 /// something that provides the guarantee. We provide an alternate
77 /// safe implementation for types that don't use noexcept but that are
78 /// marked folly::IsRelocatable and boost::has_nothrow_constructor,
79 /// which is common for folly types. In particular, if you can declare
80 /// FOLLY_ASSUME_FBVECTOR_COMPATIBLE then your type can be put in
83 template<typename> class Atom = std::atomic,
84 typename = typename std::enable_if<
85 std::is_nothrow_constructible<T,T&&>::value ||
86 folly::IsRelocatable<T>::value>::type>
87 class MPMCQueue : boost::noncopyable {
88 friend class detail::MPMCPipelineStageImpl<T>;
92 explicit MPMCQueue(size_t capacity)
94 , slots_(new detail::SingleElementQueue<T,Atom>[capacity +
96 , stride_(computeStride(capacity))
102 // ideally this would be a static assert, but g++ doesn't allow it
103 assert(alignof(MPMCQueue<T,Atom>) >= kFalseSharingRange);
106 /// A default-constructed queue is useful because a usable (non-zero
107 /// capacity) queue can be moved onto it or swapped with it
118 /// IMPORTANT: The move constructor is here to make it easier to perform
119 /// the initialization phase, it is not safe to use when there are any
120 /// concurrent accesses (this is not checked).
121 MPMCQueue(MPMCQueue<T,Atom>&& rhs) noexcept
122 : capacity_(rhs.capacity_)
124 , stride_(rhs.stride_)
125 , pushTicket_(rhs.pushTicket_.load(std::memory_order_relaxed))
126 , popTicket_(rhs.popTicket_.load(std::memory_order_relaxed))
127 , pushSpinCutoff_(rhs.pushSpinCutoff_.load(std::memory_order_relaxed))
128 , popSpinCutoff_(rhs.popSpinCutoff_.load(std::memory_order_relaxed))
130 // relaxed ops are okay for the previous reads, since rhs queue can't
131 // be in concurrent use
135 rhs.slots_ = nullptr;
137 rhs.pushTicket_.store(0, std::memory_order_relaxed);
138 rhs.popTicket_.store(0, std::memory_order_relaxed);
139 rhs.pushSpinCutoff_.store(0, std::memory_order_relaxed);
140 rhs.popSpinCutoff_.store(0, std::memory_order_relaxed);
143 /// IMPORTANT: The move operator is here to make it easier to perform
144 /// the initialization phase, it is not safe to use when there are any
145 /// concurrent accesses (this is not checked).
146 MPMCQueue<T,Atom> const& operator= (MPMCQueue<T,Atom>&& rhs) {
149 new (this) MPMCQueue(std::move(rhs));
154 /// MPMCQueue can only be safely destroyed when there are no
155 /// pending enqueuers or dequeuers (this is not checked).
160 /// Returns the number of successful reads minus the number of successful
161 /// writes. Waiting blockingRead and blockingWrite calls are included,
162 /// so this value can be negative.
163 ssize_t size() const noexcept {
164 // since both pushes and pops increase monotonically, we can get a
165 // consistent snapshot either by bracketing a read of popTicket_ with
166 // two reads of pushTicket_ that return the same value, or the other
167 // way around. We maximize our chances by alternately attempting
169 uint64_t pushes = pushTicket_.load(std::memory_order_acquire); // A
170 uint64_t pops = popTicket_.load(std::memory_order_acquire); // B
172 uint64_t nextPushes = pushTicket_.load(std::memory_order_acquire); // C
173 if (pushes == nextPushes) {
174 // pushTicket_ didn't change from A (or the previous C) to C,
175 // so we can linearize at B (or D)
176 return pushes - pops;
179 uint64_t nextPops = popTicket_.load(std::memory_order_acquire); // D
180 if (pops == nextPops) {
181 // popTicket_ didn't chance from B (or the previous D), so we
182 // can linearize at C
183 return pushes - pops;
189 /// Returns true if there are no items available for dequeue
190 bool isEmpty() const noexcept {
194 /// Returns true if there is currently no empty space to enqueue
195 bool isFull() const noexcept {
196 // careful with signed -> unsigned promotion, since size can be negative
197 return size() >= static_cast<ssize_t>(capacity_);
200 /// Returns is a guess at size() for contexts that don't need a precise
201 /// value, such as stats.
202 ssize_t sizeGuess() const noexcept {
203 return writeCount() - readCount();
207 size_t capacity() const noexcept {
211 /// Returns the total number of calls to blockingWrite or successful
212 /// calls to write, including those blockingWrite calls that are
213 /// currently blocking
214 uint64_t writeCount() const noexcept {
215 return pushTicket_.load(std::memory_order_acquire);
218 /// Returns the total number of calls to blockingRead or successful
219 /// calls to read, including those blockingRead calls that are currently
221 uint64_t readCount() const noexcept {
222 return popTicket_.load(std::memory_order_acquire);
225 /// Enqueues a T constructed from args, blocking until space is
226 /// available. Note that this method signature allows enqueue via
227 /// move, if args is a T rvalue, via copy, if args is a T lvalue, or
228 /// via emplacement if args is an initializer list that can be passed
229 /// to a T constructor.
230 template <typename ...Args>
231 void blockingWrite(Args&&... args) noexcept {
232 enqueueWithTicket(pushTicket_++, std::forward<Args>(args)...);
235 /// If an item can be enqueued with no blocking, does so and returns
236 /// true, otherwise returns false. This method is similar to
237 /// writeIfNotFull, but if you don't have a specific need for that
238 /// method you should use this one.
240 /// One of the common usages of this method is to enqueue via the
241 /// move constructor, something like q.write(std::move(x)). If write
242 /// returns false because the queue is full then x has not actually been
243 /// consumed, which looks strange. To understand why it is actually okay
244 /// to use x afterward, remember that std::move is just a typecast that
245 /// provides an rvalue reference that enables use of a move constructor
246 /// or operator. std::move doesn't actually move anything. It could
247 /// more accurately be called std::rvalue_cast or std::move_permission.
248 template <typename ...Args>
249 bool write(Args&&... args) noexcept {
251 if (tryObtainReadyPushTicket(ticket)) {
252 // we have pre-validated that the ticket won't block
253 enqueueWithTicket(ticket, std::forward<Args>(args)...);
260 /// If the queue is not full, enqueues and returns true, otherwise
261 /// returns false. Unlike write this method can be blocked by another
262 /// thread, specifically a read that has linearized (been assigned
263 /// a ticket) but not yet completed. If you don't really need this
264 /// function you should probably use write.
266 /// MPMCQueue isn't lock-free, so just because a read operation has
267 /// linearized (and isFull is false) doesn't mean that space has been
268 /// made available for another write. In this situation write will
269 /// return false, but writeIfNotFull will wait for the dequeue to finish.
270 /// This method is required if you are composing queues and managing
271 /// your own wakeup, because it guarantees that after every successful
272 /// write a readIfNotFull will succeed.
273 template <typename ...Args>
274 bool writeIfNotFull(Args&&... args) noexcept {
276 if (tryObtainPromisedPushTicket(ticket)) {
277 // some other thread is already dequeuing the slot into which we
278 // are going to enqueue, but we might have to wait for them to finish
279 enqueueWithTicket(ticket, std::forward<Args>(args)...);
286 /// Moves a dequeued element onto elem, blocking until an element
288 void blockingRead(T& elem) noexcept {
289 dequeueWithTicket(popTicket_++, elem);
292 /// If an item can be dequeued with no blocking, does so and returns
293 /// true, otherwise returns false.
294 bool read(T& elem) noexcept {
296 if (tryObtainReadyPopTicket(ticket)) {
297 // the ticket has been pre-validated to not block
298 dequeueWithTicket(ticket, elem);
305 /// If the queue is not empty, dequeues and returns true, otherwise
306 /// returns false. If the matching write is still in progress then this
307 /// method may block waiting for it. If you don't rely on being able
308 /// to dequeue (such as by counting completed write) then you should
310 bool readIfNotEmpty(T& elem) noexcept {
312 if (tryObtainPromisedPopTicket(ticket)) {
313 // the matching enqueue already has a ticket, but might not be done
314 dequeueWithTicket(ticket, elem);
323 /// Once every kAdaptationFreq we will spin longer, to try to estimate
324 /// the proper spin backoff
325 kAdaptationFreq = 128,
327 /// Memory locations on the same cache line are subject to false
328 /// sharing, which is very bad for performance
329 kFalseSharingRange = 64,
331 /// To avoid false sharing in slots_ with neighboring memory
332 /// allocations, we pad it with this many SingleElementQueue-s at
335 (kFalseSharingRange - 1) / sizeof(detail::SingleElementQueue<T,Atom>)
338 #define FOLLY_ON_NEXT_CACHE_LINE __attribute__((aligned(kFalseSharingRange)))
340 /// The maximum number of items in the queue at once
341 size_t capacity_ FOLLY_ON_NEXT_CACHE_LINE;
343 /// An array of capacity_ SingleElementQueue-s, each of which holds
344 /// either 0 or 1 item. We over-allocate by 2 * kSlotPadding and don't
345 /// touch the slots at either end, to avoid false sharing
346 detail::SingleElementQueue<T,Atom>* slots_;
348 /// The number of slots_ indices that we advance for each ticket, to
349 /// avoid false sharing. Ideally slots_[i] and slots_[i + stride_]
350 /// aren't on the same cache line
353 /// Enqueuers get tickets from here
354 Atom<uint64_t> pushTicket_ FOLLY_ON_NEXT_CACHE_LINE;
356 /// Dequeuers get tickets from here
357 Atom<uint64_t> popTicket_ FOLLY_ON_NEXT_CACHE_LINE;
359 /// This is how many times we will spin before using FUTEX_WAIT when
360 /// the queue is full on enqueue, adaptively computed by occasionally
361 /// spinning for longer and smoothing with an exponential moving average
362 Atom<int> pushSpinCutoff_ FOLLY_ON_NEXT_CACHE_LINE;
364 /// The adaptive spin cutoff when the queue is empty on dequeue
365 Atom<int> popSpinCutoff_ FOLLY_ON_NEXT_CACHE_LINE;
367 /// Alignment doesn't avoid false sharing at the end of the struct,
368 /// so fill out the last cache line
369 char padding_[kFalseSharingRange - sizeof(Atom<int>)];
371 #undef FOLLY_ON_NEXT_CACHE_LINE
373 /// We assign tickets in increasing order, but we don't want to
374 /// access neighboring elements of slots_ because that will lead to
375 /// false sharing (multiple cores accessing the same cache line even
376 /// though they aren't accessing the same bytes in that cache line).
377 /// To avoid this we advance by stride slots per ticket.
379 /// We need gcd(capacity, stride) to be 1 so that we will use all
380 /// of the slots. We ensure this by only considering prime strides,
381 /// which either have no common divisors with capacity or else have
382 /// a zero remainder after dividing by capacity. That is sufficient
383 /// to guarantee correctness, but we also want to actually spread the
384 /// accesses away from each other to avoid false sharing (consider a
385 /// stride of 7 with a capacity of 8). To that end we try a few taking
386 /// care to observe that advancing by -1 is as bad as advancing by 1
387 /// when in comes to false sharing.
389 /// The simple way to avoid false sharing would be to pad each
390 /// SingleElementQueue, but since we have capacity_ of them that could
391 /// waste a lot of space.
392 static int computeStride(size_t capacity) noexcept {
393 static const int smallPrimes[] = { 2, 3, 5, 7, 11, 13, 17, 19, 23 };
397 for (int stride : smallPrimes) {
398 if ((stride % capacity) == 0 || (capacity % stride) == 0) {
401 size_t sep = stride % capacity;
402 sep = std::min(sep, capacity - sep);
411 /// Returns the index into slots_ that should be used when enqueuing or
412 /// dequeuing with the specified ticket
413 size_t idx(uint64_t ticket) noexcept {
414 return ((ticket * stride_) % capacity_) + kSlotPadding;
417 /// Maps an enqueue or dequeue ticket to the turn should be used at the
418 /// corresponding SingleElementQueue
419 uint32_t turn(uint64_t ticket) noexcept {
420 return ticket / capacity_;
423 /// Tries to obtain a push ticket for which SingleElementQueue::enqueue
424 /// won't block. Returns true on immediate success, false on immediate
426 bool tryObtainReadyPushTicket(uint64_t& rv) noexcept {
427 auto ticket = pushTicket_.load(std::memory_order_acquire); // A
429 if (!slots_[idx(ticket)].mayEnqueue(turn(ticket))) {
430 // if we call enqueue(ticket, ...) on the SingleElementQueue
431 // right now it would block, but this might no longer be the next
432 // ticket. We can increase the chance of tryEnqueue success under
433 // contention (without blocking) by rechecking the ticket dispenser
435 ticket = pushTicket_.load(std::memory_order_acquire); // B
436 if (prev == ticket) {
437 // mayEnqueue was bracketed by two reads (A or prev B or prev
438 // failing CAS to B), so we are definitely unable to enqueue
442 // we will bracket the mayEnqueue check with a read (A or prev B
443 // or prev failing CAS) and the following CAS. If the CAS fails
444 // it will effect a load of pushTicket_
445 if (pushTicket_.compare_exchange_strong(ticket, ticket + 1)) {
453 /// Tries to obtain a push ticket which can be satisfied if all
454 /// in-progress pops complete. This function does not block, but
455 /// blocking may be required when using the returned ticket if some
456 /// other thread's pop is still in progress (ticket has been granted but
457 /// pop has not yet completed).
458 bool tryObtainPromisedPushTicket(uint64_t& rv) noexcept {
459 auto numPushes = pushTicket_.load(std::memory_order_acquire); // A
461 auto numPops = popTicket_.load(std::memory_order_acquire); // B
462 // n will be negative if pops are pending
463 int64_t n = numPushes - numPops;
464 if (n >= static_cast<ssize_t>(capacity_)) {
465 // Full, linearize at B. We don't need to recheck the read we
466 // performed at A, because if numPushes was stale at B then the
467 // real numPushes value is even worse
470 if (pushTicket_.compare_exchange_strong(numPushes, numPushes + 1)) {
477 /// Tries to obtain a pop ticket for which SingleElementQueue::dequeue
478 /// won't block. Returns true on immediate success, false on immediate
480 bool tryObtainReadyPopTicket(uint64_t& rv) noexcept {
481 auto ticket = popTicket_.load(std::memory_order_acquire);
483 if (!slots_[idx(ticket)].mayDequeue(turn(ticket))) {
485 ticket = popTicket_.load(std::memory_order_acquire);
486 if (prev == ticket) {
490 if (popTicket_.compare_exchange_strong(ticket, ticket + 1)) {
498 /// Similar to tryObtainReadyPopTicket, but returns a pop ticket whose
499 /// corresponding push ticket has already been handed out, rather than
500 /// returning one whose corresponding push ticket has already been
501 /// completed. This means that there is a possibility that the caller
502 /// will block when using the ticket, but it allows the user to rely on
503 /// the fact that if enqueue has succeeded, tryObtainPromisedPopTicket
504 /// will return true. The "try" part of this is that we won't have
505 /// to block waiting for someone to call enqueue, although we might
506 /// have to block waiting for them to finish executing code inside the
507 /// MPMCQueue itself.
508 bool tryObtainPromisedPopTicket(uint64_t& rv) noexcept {
509 auto numPops = popTicket_.load(std::memory_order_acquire); // A
511 auto numPushes = pushTicket_.load(std::memory_order_acquire); // B
512 if (numPops >= numPushes) {
513 // Empty, or empty with pending pops. Linearize at B. We don't
514 // need to recheck the read we performed at A, because if numPops
515 // is stale then the fresh value is larger and the >= is still true
518 if (popTicket_.compare_exchange_strong(numPops, numPops + 1)) {
525 // Given a ticket, constructs an enqueued item using args
526 template <typename ...Args>
527 void enqueueWithTicket(uint64_t ticket, Args&&... args) noexcept {
528 slots_[idx(ticket)].enqueue(turn(ticket),
530 (ticket % kAdaptationFreq) == 0,
531 std::forward<Args>(args)...);
534 // Given a ticket, dequeues the corresponding element
535 void dequeueWithTicket(uint64_t ticket, T& elem) noexcept {
536 slots_[idx(ticket)].dequeue(turn(ticket),
538 (ticket % kAdaptationFreq) == 0,
546 /// A TurnSequencer allows threads to order their execution according to
547 /// a monotonically increasing (with wraparound) "turn" value. The two
548 /// operations provided are to wait for turn T, and to move to the next
549 /// turn. Every thread that is waiting for T must have arrived before
550 /// that turn is marked completed (for MPMCQueue only one thread waits
551 /// for any particular turn, so this is trivially true).
553 /// TurnSequencer's state_ holds 26 bits of the current turn (shifted
554 /// left by 6), along with a 6 bit saturating value that records the
555 /// maximum waiter minus the current turn. Wraparound of the turn space
556 /// is expected and handled. This allows us to atomically adjust the
557 /// number of outstanding waiters when we perform a FUTEX_WAKE operation.
558 /// Compare this strategy to sem_t's separate num_waiters field, which
559 /// isn't decremented until after the waiting thread gets scheduled,
560 /// during which time more enqueues might have occurred and made pointless
561 /// FUTEX_WAKE calls.
563 /// TurnSequencer uses futex() directly. It is optimized for the
564 /// case that the highest awaited turn is 32 or less higher than the
565 /// current turn. We use the FUTEX_WAIT_BITSET variant, which lets
566 /// us embed 32 separate wakeup channels in a single futex. See
567 /// http://locklessinc.com/articles/futex_cheat_sheet for a description.
569 /// We only need to keep exact track of the delta between the current
570 /// turn and the maximum waiter for the 32 turns that follow the current
571 /// one, because waiters at turn t+32 will be awoken at turn t. At that
572 /// point they can then adjust the delta using the higher base. Since we
573 /// need to encode waiter deltas of 0 to 32 inclusive, we use 6 bits.
574 /// We actually store waiter deltas up to 63, since that might reduce
575 /// the number of CAS operations a tiny bit.
577 /// To avoid some futex() calls entirely, TurnSequencer uses an adaptive
578 /// spin cutoff before waiting. The overheads (and convergence rate)
579 /// of separately tracking the spin cutoff for each TurnSequencer would
580 /// be prohibitive, so the actual storage is passed in as a parameter and
581 /// updated atomically. This also lets the caller use different adaptive
582 /// cutoffs for different operations (read versus write, for example).
583 /// To avoid contention, the spin cutoff is only updated when requested
585 template <template<typename> class Atom>
586 struct TurnSequencer {
587 explicit TurnSequencer(const uint32_t firstTurn = 0) noexcept
588 : state_(encode(firstTurn << kTurnShift, 0))
591 /// Returns true iff a call to waitForTurn(turn, ...) won't block
592 bool isTurn(const uint32_t turn) const noexcept {
593 auto state = state_.load(std::memory_order_acquire);
594 return decodeCurrentSturn(state) == (turn << kTurnShift);
597 // Internally we always work with shifted turn values, which makes the
598 // truncation and wraparound work correctly. This leaves us bits at
599 // the bottom to store the number of waiters. We call shifted turns
600 // "sturns" inside this class.
602 /// Blocks the current thread until turn has arrived. If
603 /// updateSpinCutoff is true then this will spin for up to kMaxSpins tries
604 /// before blocking and will adjust spinCutoff based on the results,
605 /// otherwise it will spin for at most spinCutoff spins.
606 void waitForTurn(const uint32_t turn,
607 Atom<int>& spinCutoff,
608 const bool updateSpinCutoff) noexcept {
609 int prevThresh = spinCutoff.load(std::memory_order_relaxed);
610 const int effectiveSpinCutoff =
611 updateSpinCutoff || prevThresh == 0 ? kMaxSpins : prevThresh;
614 const uint32_t sturn = turn << kTurnShift;
615 for (tries = 0; ; ++tries) {
616 uint32_t state = state_.load(std::memory_order_acquire);
617 uint32_t current_sturn = decodeCurrentSturn(state);
618 if (current_sturn == sturn) {
622 // wrap-safe version of assert(current_sturn < sturn)
623 assert(sturn - current_sturn < std::numeric_limits<uint32_t>::max() / 2);
625 // the first effectSpinCutoff tries are spins, after that we will
626 // record ourself as a waiter and block with futexWait
627 if (tries < effectiveSpinCutoff) {
628 asm volatile ("pause");
632 uint32_t current_max_waiter_delta = decodeMaxWaitersDelta(state);
633 uint32_t our_waiter_delta = (sturn - current_sturn) >> kTurnShift;
635 if (our_waiter_delta <= current_max_waiter_delta) {
636 // state already records us as waiters, probably because this
637 // isn't our first time around this loop
640 new_state = encode(current_sturn, our_waiter_delta);
641 if (state != new_state &&
642 !state_.compare_exchange_strong(state, new_state)) {
646 state_.futexWait(new_state, futexChannel(turn));
649 if (updateSpinCutoff || prevThresh == 0) {
650 // if we hit kMaxSpins then spinning was pointless, so the right
651 // spinCutoff is kMinSpins
653 if (tries >= kMaxSpins) {
656 // to account for variations, we allow ourself to spin 2*N when
657 // we think that N is actually required in order to succeed
658 target = std::min(int{kMaxSpins}, std::max(int{kMinSpins}, tries * 2));
661 if (prevThresh == 0) {
665 // try once, keep moving if CAS fails. Exponential moving average
667 spinCutoff.compare_exchange_weak(
668 prevThresh, prevThresh + (target - prevThresh) / 8);
673 /// Unblocks a thread running waitForTurn(turn + 1)
674 void completeTurn(const uint32_t turn) noexcept {
675 uint32_t state = state_.load(std::memory_order_acquire);
677 assert(state == encode(turn << kTurnShift, decodeMaxWaitersDelta(state)));
678 uint32_t max_waiter_delta = decodeMaxWaitersDelta(state);
679 uint32_t new_state = encode(
680 (turn + 1) << kTurnShift,
681 max_waiter_delta == 0 ? 0 : max_waiter_delta - 1);
682 if (state_.compare_exchange_strong(state, new_state)) {
683 if (max_waiter_delta != 0) {
684 state_.futexWake(std::numeric_limits<int>::max(),
685 futexChannel(turn + 1));
689 // failing compare_exchange_strong updates first arg to the value
690 // that caused the failure, so no need to reread state_
694 /// Returns the least-most significant byte of the current uncompleted
695 /// turn. The full 32 bit turn cannot be recovered.
696 uint8_t uncompletedTurnLSB() const noexcept {
697 return state_.load(std::memory_order_acquire) >> kTurnShift;
702 /// kTurnShift counts the bits that are stolen to record the delta
703 /// between the current turn and the maximum waiter. It needs to be big
704 /// enough to record wait deltas of 0 to 32 inclusive. Waiters more
705 /// than 32 in the future will be woken up 32*n turns early (since
706 /// their BITSET will hit) and will adjust the waiter count again.
707 /// We go a bit beyond and let the waiter count go up to 63, which
708 /// is free and might save us a few CAS
710 kWaitersMask = (1 << kTurnShift) - 1,
712 /// The minimum spin count that we will adaptively select
715 /// The maximum spin count that we will adaptively select, and the
716 /// spin count that will be used when probing to get a new data point
717 /// for the adaptation
721 /// This holds both the current turn, and the highest waiting turn,
722 /// stored as (current_turn << 6) | min(63, max(waited_turn - current_turn))
725 /// Returns the bitmask to pass futexWait or futexWake when communicating
726 /// about the specified turn
727 int futexChannel(uint32_t turn) const noexcept {
728 return 1 << (turn & 31);
731 uint32_t decodeCurrentSturn(uint32_t state) const noexcept {
732 return state & ~kWaitersMask;
735 uint32_t decodeMaxWaitersDelta(uint32_t state) const noexcept {
736 return state & kWaitersMask;
739 uint32_t encode(uint32_t currentSturn, uint32_t maxWaiterD) const noexcept {
740 return currentSturn | std::min(uint32_t{ kWaitersMask }, maxWaiterD);
745 /// SingleElementQueue implements a blocking queue that holds at most one
746 /// item, and that requires its users to assign incrementing identifiers
747 /// (turns) to each enqueue and dequeue operation. Note that the turns
748 /// used by SingleElementQueue are doubled inside the TurnSequencer
749 template <typename T, template <typename> class Atom>
750 struct SingleElementQueue {
752 ~SingleElementQueue() noexcept {
753 if ((sequencer_.uncompletedTurnLSB() & 1) == 1) {
754 // we are pending a dequeue, so we have a constructed item
759 /// enqueue using in-place noexcept construction
760 template <typename ...Args,
761 typename = typename std::enable_if<
762 std::is_nothrow_constructible<T,Args...>::value>::type>
763 void enqueue(const uint32_t turn,
764 Atom<int>& spinCutoff,
765 const bool updateSpinCutoff,
766 Args&&... args) noexcept {
767 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
768 new (contents_) T(std::forward<Args>(args)...);
769 sequencer_.completeTurn(turn * 2);
772 /// enqueue using move construction, either real (if
773 /// is_nothrow_move_constructible) or simulated using relocation and
774 /// default construction (if IsRelocatable and has_nothrow_constructor)
775 template <typename = typename std::enable_if<
776 (folly::IsRelocatable<T>::value &&
777 boost::has_nothrow_constructor<T>::value) ||
778 std::is_nothrow_constructible<T,T&&>::value>::type>
779 void enqueue(const uint32_t turn,
780 Atom<int>& spinCutoff,
781 const bool updateSpinCutoff,
782 T&& goner) noexcept {
783 if (std::is_nothrow_constructible<T,T&&>::value) {
785 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
786 new (contents_) T(std::move(goner));
787 sequencer_.completeTurn(turn * 2);
789 // simulate nothrow move with relocation, followed by default
790 // construction to fill the gap we created
791 sequencer_.waitForTurn(turn * 2, spinCutoff, updateSpinCutoff);
792 memcpy(contents_, &goner, sizeof(T));
793 sequencer_.completeTurn(turn * 2);
798 bool mayEnqueue(const uint32_t turn) const noexcept {
799 return sequencer_.isTurn(turn * 2);
802 void dequeue(uint32_t turn,
803 Atom<int>& spinCutoff,
804 const bool updateSpinCutoff,
806 if (folly::IsRelocatable<T>::value) {
807 // this version is preferred, because we do as much work as possible
812 // unlikely, but if we don't complete our turn the queue will die
814 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
815 memcpy(&elem, contents_, sizeof(T));
816 sequencer_.completeTurn(turn * 2 + 1);
818 // use nothrow move assignment
819 sequencer_.waitForTurn(turn * 2 + 1, spinCutoff, updateSpinCutoff);
820 elem = std::move(*ptr());
822 sequencer_.completeTurn(turn * 2 + 1);
826 bool mayDequeue(const uint32_t turn) const noexcept {
827 return sequencer_.isTurn(turn * 2 + 1);
831 /// Storage for a T constructed with placement new
832 char contents_[sizeof(T)] __attribute__((aligned(alignof(T))));
834 /// Even turns are pushes, odd turns are pops
835 TurnSequencer<Atom> sequencer_;
838 return static_cast<T*>(static_cast<void*>(contents_));
841 void destroyContents() noexcept {
845 // g++ doesn't seem to have std::is_nothrow_destructible yet
848 memset(contents_, 'Q', sizeof(T));
853 } // namespace detail