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.
17 // @author Nathan Bronson (ngbronson@fb.com)
25 #include <type_traits>
27 #include <folly/Likely.h>
28 #include <folly/concurrency/CacheLocality.h>
29 #include <folly/detail/Futex.h>
30 #include <folly/portability/Asm.h>
31 #include <folly/portability/SysResource.h>
33 // SharedMutex is a reader-writer lock. It is small, very fast, scalable
34 // on multi-core, and suitable for use when readers or writers may block.
35 // Unlike most other reader-writer locks, its throughput with concurrent
36 // readers scales linearly; it is able to acquire and release the lock
37 // in shared mode without cache line ping-ponging. It is suitable for
38 // a wide range of lock hold times because it starts with spinning,
39 // proceeds to using sched_yield with a preemption heuristic, and then
40 // waits using futex and precise wakeups.
42 // SharedMutex provides all of the methods of folly::RWSpinLock,
43 // boost::shared_mutex, boost::upgrade_mutex, and C++14's
44 // std::shared_timed_mutex. All operations that can block are available
45 // in try, try-for, and try-until (system_clock or steady_clock) versions.
47 // SharedMutexReadPriority gives priority to readers,
48 // SharedMutexWritePriority gives priority to writers. SharedMutex is an
49 // alias for SharedMutexWritePriority, because writer starvation is more
50 // likely than reader starvation for the read-heavy workloads targetted
53 // In my tests SharedMutex is as good or better than the other
54 // reader-writer locks in use at Facebook for almost all use cases,
55 // sometimes by a wide margin. (If it is rare that there are actually
56 // concurrent readers then RWSpinLock can be a few nanoseconds faster.)
57 // I compared it to folly::RWSpinLock, folly::RWTicketSpinLock64,
58 // boost::shared_mutex, pthread_rwlock_t, and a RWLock that internally uses
59 // spinlocks to guard state and pthread_mutex_t+pthread_cond_t to block.
60 // (Thrift's ReadWriteMutex is based underneath on pthread_rwlock_t.)
61 // It is generally as good or better than the rest when evaluating size,
62 // speed, scalability, or latency outliers. In the corner cases where
63 // it is not the fastest (such as single-threaded use or heavy write
64 // contention) it is never very much worse than the best. See the bottom
65 // of folly/test/SharedMutexTest.cpp for lots of microbenchmark results.
67 // Comparison to folly::RWSpinLock:
69 // * SharedMutex is faster than RWSpinLock when there are actually
70 // concurrent read accesses (sometimes much faster), and ~5 nanoseconds
71 // slower when there is not actually any contention. SharedMutex is
72 // faster in every (benchmarked) scenario where the shared mode of
73 // the lock is actually useful.
75 // * Concurrent shared access to SharedMutex scales linearly, while total
76 // RWSpinLock throughput drops as more threads try to access the lock
77 // in shared mode. Under very heavy read contention SharedMutex can
78 // be two orders of magnitude faster than RWSpinLock (or any reader
79 // writer lock that doesn't use striping or deferral).
81 // * SharedMutex can safely protect blocking calls, because after an
82 // initial period of spinning it waits using futex().
84 // * RWSpinLock prioritizes readers, SharedMutex has both reader- and
85 // writer-priority variants, but defaults to write priority.
87 // * RWSpinLock's upgradeable mode blocks new readers, while SharedMutex's
88 // doesn't. Both semantics are reasonable. The boost documentation
89 // doesn't explicitly talk about this behavior (except by omitting
90 // any statement that those lock modes conflict), but the boost
91 // implementations do allow new readers while the upgradeable mode
92 // is held. See https://github.com/boostorg/thread/blob/master/
93 // include/boost/thread/pthread/shared_mutex.hpp
95 // * RWSpinLock::UpgradedHolder maps to SharedMutex::UpgradeHolder
96 // (UpgradeableHolder would be even more pedantically correct).
97 // SharedMutex's holders have fewer methods (no reset) and are less
98 // tolerant (promotion and downgrade crash if the donor doesn't own
99 // the lock, and you must use the default constructor rather than
100 // passing a nullptr to the pointer constructor).
102 // Both SharedMutex and RWSpinLock provide "exclusive", "upgrade",
103 // and "shared" modes. At all times num_threads_holding_exclusive +
104 // num_threads_holding_upgrade <= 1, and num_threads_holding_exclusive ==
105 // 0 || num_threads_holding_shared == 0. RWSpinLock has the additional
106 // constraint that num_threads_holding_shared cannot increase while
107 // num_threads_holding_upgrade is non-zero.
109 // Comparison to the internal RWLock:
111 // * SharedMutex doesn't allow a maximum reader count to be configured,
112 // so it can't be used as a semaphore in the same way as RWLock.
114 // * SharedMutex is 4 bytes, RWLock is 256.
116 // * SharedMutex is as fast or faster than RWLock in all of my
117 // microbenchmarks, and has positive rather than negative scalability.
119 // * RWLock and SharedMutex are both writer priority locks.
121 // * SharedMutex avoids latency outliers as well as RWLock.
123 // * SharedMutex uses different names (t != 0 below):
125 // RWLock::lock(0) => SharedMutex::lock()
127 // RWLock::lock(t) => SharedMutex::try_lock_for(milliseconds(t))
129 // RWLock::tryLock() => SharedMutex::try_lock()
131 // RWLock::unlock() => SharedMutex::unlock()
133 // RWLock::enter(0) => SharedMutex::lock_shared()
135 // RWLock::enter(t) =>
136 // SharedMutex::try_lock_shared_for(milliseconds(t))
138 // RWLock::tryEnter() => SharedMutex::try_lock_shared()
140 // RWLock::leave() => SharedMutex::unlock_shared()
142 // * RWLock allows the reader count to be adjusted by a value other
143 // than 1 during enter() or leave(). SharedMutex doesn't currently
144 // implement this feature.
146 // * RWLock's methods are marked const, SharedMutex's aren't.
148 // Reader-writer locks have the potential to allow concurrent access
149 // to shared read-mostly data, but in practice they often provide no
150 // improvement over a mutex. The problem is the cache coherence protocol
151 // of modern CPUs. Coherence is provided by making sure that when a cache
152 // line is written it is present in only one core's cache. Since a memory
153 // write is required to acquire a reader-writer lock in shared mode, the
154 // cache line holding the lock is invalidated in all of the other caches.
155 // This leads to cache misses when another thread wants to acquire or
156 // release the lock concurrently. When the RWLock is colocated with the
157 // data it protects (common), cache misses can also continue occur when
158 // a thread that already holds the lock tries to read the protected data.
160 // Ideally, a reader-writer lock would allow multiple cores to acquire
161 // and release the lock in shared mode without incurring any cache misses.
162 // This requires that each core records its shared access in a cache line
163 // that isn't read or written by other read-locking cores. (Writers will
164 // have to check all of the cache lines.) Typical server hardware when
165 // this comment was written has 16 L1 caches and cache lines of 64 bytes,
166 // so a lock striped over all L1 caches would occupy a prohibitive 1024
167 // bytes. Nothing says that we need a separate set of per-core memory
168 // locations for each lock, however. Each SharedMutex instance is only
169 // 4 bytes, but all locks together share a 2K area in which they make a
170 // core-local record of lock acquisitions.
172 // SharedMutex's strategy of using a shared set of core-local stripes has
173 // a potential downside, because it means that acquisition of any lock in
174 // write mode can conflict with acquisition of any lock in shared mode.
175 // If a lock instance doesn't actually experience concurrency then this
176 // downside will outweight the upside of improved scalability for readers.
177 // To avoid this problem we dynamically detect concurrent accesses to
178 // SharedMutex, and don't start using the deferred mode unless we actually
179 // observe concurrency. See kNumSharedToStartDeferring.
181 // It is explicitly allowed to call lock_unshared() from a different
182 // thread than lock_shared(), so long as they are properly paired.
183 // lock_unshared() needs to find the location at which lock_shared()
184 // recorded the lock, which might be in the lock itself or in any of
185 // the shared slots. If you can conveniently pass state from lock
186 // acquisition to release then the fastest mechanism is to std::move
187 // the SharedMutex::ReadHolder instance or an SharedMutex::Token (using
188 // lock_shared(Token&) and unlock_shared(Token&)). The guard or token
189 // will tell unlock_shared where in deferredReaders[] to look for the
190 // deferred lock. The Token-less version of unlock_shared() works in all
191 // cases, but is optimized for the common (no inter-thread handoff) case.
193 // In both read- and write-priority mode, a waiting lock() (exclusive mode)
194 // only blocks readers after it has waited for an active upgrade lock to be
195 // released; until the upgrade lock is released (or upgraded or downgraded)
196 // readers will still be able to enter. Preferences about lock acquisition
197 // are not guaranteed to be enforced perfectly (even if they were, there
198 // is theoretically the chance that a thread could be arbitrarily suspended
199 // between calling lock() and SharedMutex code actually getting executed).
201 // try_*_for methods always try at least once, even if the duration
202 // is zero or negative. The duration type must be compatible with
203 // std::chrono::steady_clock. try_*_until methods also always try at
204 // least once. std::chrono::system_clock and std::chrono::steady_clock
207 // If you have observed by profiling that your SharedMutex-s are getting
208 // cache misses on deferredReaders[] due to another SharedMutex user, then
209 // you can use the tag type to create your own instantiation of the type.
210 // The contention threshold (see kNumSharedToStartDeferring) should make
211 // this unnecessary in all but the most extreme cases. Make sure to check
212 // that the increased icache and dcache footprint of the tagged result is
215 // SharedMutex's use of thread local storage is as an optimization, so
216 // for the case where thread local storage is not supported, define it
218 #ifndef FOLLY_SHAREDMUTEX_TLS
220 #define FOLLY_SHAREDMUTEX_TLS FOLLY_TLS
222 #define FOLLY_SHAREDMUTEX_TLS
228 struct SharedMutexToken {
229 enum class Type : uint16_t {
239 template <bool ReaderPriority,
240 typename Tag_ = void,
241 template <typename> class Atom = std::atomic,
242 bool BlockImmediately = false>
243 class SharedMutexImpl {
245 static constexpr bool kReaderPriority = ReaderPriority;
248 typedef SharedMutexToken Token;
254 constexpr SharedMutexImpl() noexcept : state_(0) {}
256 SharedMutexImpl(const SharedMutexImpl&) = delete;
257 SharedMutexImpl(SharedMutexImpl&&) = delete;
258 SharedMutexImpl& operator = (const SharedMutexImpl&) = delete;
259 SharedMutexImpl& operator = (SharedMutexImpl&&) = delete;
261 // It is an error to destroy an SharedMutex that still has
262 // any outstanding locks. This is checked if NDEBUG isn't defined.
263 // SharedMutex's exclusive mode can be safely used to guard the lock's
264 // own destruction. If, for example, you acquire the lock in exclusive
265 // mode and then observe that the object containing the lock is no longer
266 // needed, you can unlock() and then immediately destroy the lock.
267 // See https://sourceware.org/bugzilla/show_bug.cgi?id=13690 for a
268 // description about why this property needs to be explicitly mentioned.
270 auto state = state_.load(std::memory_order_relaxed);
271 if (UNLIKELY((state & kHasS) != 0)) {
272 cleanupTokenlessSharedDeferred(state);
276 // if a futexWait fails to go to sleep because the value has been
277 // changed, we don't necessarily clean up the wait bits, so it is
278 // possible they will be set here in a correct system
279 assert((state & ~(kWaitingAny | kMayDefer)) == 0);
280 if ((state & kMayDefer) != 0) {
281 for (uint32_t slot = 0; slot < kMaxDeferredReaders; ++slot) {
282 auto slotValue = deferredReader(slot)->load(std::memory_order_relaxed);
283 assert(!slotValueIsThis(slotValue));
291 (void)lockExclusiveImpl(kHasSolo, ctx);
296 return lockExclusiveImpl(kHasSolo, ctx);
299 template <class Rep, class Period>
300 bool try_lock_for(const std::chrono::duration<Rep, Period>& duration) {
301 WaitForDuration<Rep, Period> ctx(duration);
302 return lockExclusiveImpl(kHasSolo, ctx);
305 template <class Clock, class Duration>
307 const std::chrono::time_point<Clock, Duration>& absDeadline) {
308 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
309 return lockExclusiveImpl(kHasSolo, ctx);
313 // It is possible that we have a left-over kWaitingNotS if the last
314 // unlock_shared() that let our matching lock() complete finished
315 // releasing before lock()'s futexWait went to sleep. Clean it up now
316 auto state = (state_ &= ~(kWaitingNotS | kPrevDefer | kHasE));
317 assert((state & ~kWaitingAny) == 0);
318 wakeRegisteredWaiters(state, kWaitingE | kWaitingU | kWaitingS);
321 // Managing the token yourself makes unlock_shared a bit faster
325 (void)lockSharedImpl(nullptr, ctx);
328 void lock_shared(Token& token) {
330 (void)lockSharedImpl(&token, ctx);
333 bool try_lock_shared() {
335 return lockSharedImpl(nullptr, ctx);
338 bool try_lock_shared(Token& token) {
340 return lockSharedImpl(&token, ctx);
343 template <class Rep, class Period>
344 bool try_lock_shared_for(const std::chrono::duration<Rep, Period>& duration) {
345 WaitForDuration<Rep, Period> ctx(duration);
346 return lockSharedImpl(nullptr, ctx);
349 template <class Rep, class Period>
350 bool try_lock_shared_for(const std::chrono::duration<Rep, Period>& duration,
352 WaitForDuration<Rep, Period> ctx(duration);
353 return lockSharedImpl(&token, ctx);
356 template <class Clock, class Duration>
357 bool try_lock_shared_until(
358 const std::chrono::time_point<Clock, Duration>& absDeadline) {
359 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
360 return lockSharedImpl(nullptr, ctx);
363 template <class Clock, class Duration>
364 bool try_lock_shared_until(
365 const std::chrono::time_point<Clock, Duration>& absDeadline,
367 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
368 return lockSharedImpl(&token, ctx);
371 void unlock_shared() {
372 auto state = state_.load(std::memory_order_acquire);
374 // kPrevDefer can only be set if HasE or BegunE is set
375 assert((state & (kPrevDefer | kHasE | kBegunE)) != kPrevDefer);
377 // lock() strips kMayDefer immediately, but then copies it to
378 // kPrevDefer so we can tell if the pre-lock() lock_shared() might
380 if ((state & (kMayDefer | kPrevDefer)) == 0 ||
381 !tryUnlockTokenlessSharedDeferred()) {
382 // Matching lock_shared() couldn't have deferred, or the deferred
383 // lock has already been inlined by applyDeferredReaders()
384 unlockSharedInline();
388 void unlock_shared(Token& token) {
389 assert(token.type_ == Token::Type::INLINE_SHARED ||
390 token.type_ == Token::Type::DEFERRED_SHARED);
392 if (token.type_ != Token::Type::DEFERRED_SHARED ||
393 !tryUnlockSharedDeferred(token.slot_)) {
394 unlockSharedInline();
397 token.type_ = Token::Type::INVALID;
401 void unlock_and_lock_shared() {
402 // We can't use state_ -=, because we need to clear 2 bits (1 of which
403 // has an uncertain initial state) and set 1 other. We might as well
404 // clear the relevant wake bits at the same time. Note that since S
405 // doesn't block the beginning of a transition to E (writer priority
406 // can cut off new S, reader priority grabs BegunE and blocks deferred
407 // S) we need to wake E as well.
408 auto state = state_.load(std::memory_order_acquire);
410 assert((state & ~(kWaitingAny | kPrevDefer)) == kHasE);
411 } while (!state_.compare_exchange_strong(
412 state, (state & ~(kWaitingAny | kPrevDefer | kHasE)) + kIncrHasS));
413 if ((state & (kWaitingE | kWaitingU | kWaitingS)) != 0) {
414 futexWakeAll(kWaitingE | kWaitingU | kWaitingS);
418 void unlock_and_lock_shared(Token& token) {
419 unlock_and_lock_shared();
420 token.type_ = Token::Type::INLINE_SHARED;
423 void lock_upgrade() {
425 (void)lockUpgradeImpl(ctx);
428 bool try_lock_upgrade() {
430 return lockUpgradeImpl(ctx);
433 template <class Rep, class Period>
434 bool try_lock_upgrade_for(
435 const std::chrono::duration<Rep, Period>& duration) {
436 WaitForDuration<Rep, Period> ctx(duration);
437 return lockUpgradeImpl(ctx);
440 template <class Clock, class Duration>
441 bool try_lock_upgrade_until(
442 const std::chrono::time_point<Clock, Duration>& absDeadline) {
443 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
444 return lockUpgradeImpl(ctx);
447 void unlock_upgrade() {
448 auto state = (state_ -= kHasU);
449 assert((state & (kWaitingNotS | kHasSolo)) == 0);
450 wakeRegisteredWaiters(state, kWaitingE | kWaitingU);
453 void unlock_upgrade_and_lock() {
454 // no waiting necessary, so waitMask is empty
456 (void)lockExclusiveImpl(0, ctx);
459 void unlock_upgrade_and_lock_shared() {
460 auto state = (state_ -= kHasU - kIncrHasS);
461 assert((state & (kWaitingNotS | kHasSolo)) == 0);
462 wakeRegisteredWaiters(state, kWaitingE | kWaitingU);
465 void unlock_upgrade_and_lock_shared(Token& token) {
466 unlock_upgrade_and_lock_shared();
467 token.type_ = Token::Type::INLINE_SHARED;
470 void unlock_and_lock_upgrade() {
471 // We can't use state_ -=, because we need to clear 2 bits (1 of
472 // which has an uncertain initial state) and set 1 other. We might
473 // as well clear the relevant wake bits at the same time.
474 auto state = state_.load(std::memory_order_acquire);
476 assert((state & ~(kWaitingAny | kPrevDefer)) == kHasE);
478 (state & ~(kWaitingNotS | kWaitingS | kPrevDefer | kHasE)) + kHasU;
479 if (state_.compare_exchange_strong(state, after)) {
480 if ((state & kWaitingS) != 0) {
481 futexWakeAll(kWaitingS);
489 typedef typename folly::detail::Futex<Atom> Futex;
491 // Internally we use four kinds of wait contexts. These are structs
492 // that provide a doWait method that returns true if a futex wake
493 // was issued that intersects with the waitMask, false if there was a
494 // timeout and no more waiting should be performed. Spinning occurs
495 // before the wait context is invoked.
498 bool canBlock() { return true; }
499 bool canTimeOut() { return false; }
500 bool shouldTimeOut() { return false; }
502 bool doWait(Futex& futex, uint32_t expected, uint32_t waitMask) {
503 futex.futexWait(expected, waitMask);
509 bool canBlock() { return false; }
510 bool canTimeOut() { return true; }
511 bool shouldTimeOut() { return true; }
513 bool doWait(Futex& /* futex */,
514 uint32_t /* expected */,
515 uint32_t /* waitMask */) {
520 template <class Rep, class Period>
521 struct WaitForDuration {
522 std::chrono::duration<Rep, Period> duration_;
523 bool deadlineComputed_;
524 std::chrono::steady_clock::time_point deadline_;
526 explicit WaitForDuration(const std::chrono::duration<Rep, Period>& duration)
527 : duration_(duration), deadlineComputed_(false) {}
529 std::chrono::steady_clock::time_point deadline() {
530 if (!deadlineComputed_) {
531 deadline_ = std::chrono::steady_clock::now() + duration_;
532 deadlineComputed_ = true;
537 bool canBlock() { return duration_.count() > 0; }
538 bool canTimeOut() { return true; }
540 bool shouldTimeOut() {
541 return std::chrono::steady_clock::now() > deadline();
544 bool doWait(Futex& futex, uint32_t expected, uint32_t waitMask) {
545 auto result = futex.futexWaitUntil(expected, deadline(), waitMask);
546 return result != folly::detail::FutexResult::TIMEDOUT;
550 template <class Clock, class Duration>
551 struct WaitUntilDeadline {
552 std::chrono::time_point<Clock, Duration> absDeadline_;
554 bool canBlock() { return true; }
555 bool canTimeOut() { return true; }
556 bool shouldTimeOut() { return Clock::now() > absDeadline_; }
558 bool doWait(Futex& futex, uint32_t expected, uint32_t waitMask) {
559 auto result = futex.futexWaitUntil(expected, absDeadline_, waitMask);
560 return result != folly::detail::FutexResult::TIMEDOUT;
567 // S count needs to be on the end, because we explicitly allow it to
568 // underflow. This can occur while we are in the middle of applying
569 // deferred locks (we remove them from deferredReaders[] before
570 // inlining them), or during token-less unlock_shared() if a racing
571 // lock_shared();unlock_shared() moves the deferredReaders slot while
572 // the first unlock_shared() is scanning. The former case is cleaned
573 // up before we finish applying the locks. The latter case can persist
574 // until destruction, when it is cleaned up.
575 static constexpr uint32_t kIncrHasS = 1 << 10;
576 static constexpr uint32_t kHasS = ~(kIncrHasS - 1);
578 // If false, then there are definitely no deferred read locks for this
579 // instance. Cleared after initialization and when exclusively locked.
580 static constexpr uint32_t kMayDefer = 1 << 9;
582 // lock() cleared kMayDefer as soon as it starts draining readers (so
583 // that it doesn't have to do a second CAS once drain completes), but
584 // unlock_shared() still needs to know whether to scan deferredReaders[]
585 // or not. We copy kMayDefer to kPrevDefer when setting kHasE or
586 // kBegunE, and clear it when clearing those bits.
587 static constexpr uint32_t kPrevDefer = 1 << 8;
589 // Exclusive-locked blocks all read locks and write locks. This bit
590 // may be set before all readers have finished, but in that case the
591 // thread that sets it won't return to the caller until all read locks
592 // have been released.
593 static constexpr uint32_t kHasE = 1 << 7;
595 // Exclusive-draining means that lock() is waiting for existing readers
596 // to leave, but that new readers may still acquire shared access.
597 // This is only used in reader priority mode. New readers during
598 // drain must be inline. The difference between this and kHasU is that
599 // kBegunE prevents kMayDefer from being set.
600 static constexpr uint32_t kBegunE = 1 << 6;
602 // At most one thread may have either exclusive or upgrade lock
603 // ownership. Unlike exclusive mode, ownership of the lock in upgrade
604 // mode doesn't preclude other threads holding the lock in shared mode.
605 // boost's concept for this doesn't explicitly say whether new shared
606 // locks can be acquired one lock_upgrade has succeeded, but doesn't
607 // list that as disallowed. RWSpinLock disallows new read locks after
608 // lock_upgrade has been acquired, but the boost implementation doesn't.
609 // We choose the latter.
610 static constexpr uint32_t kHasU = 1 << 5;
612 // There are three states that we consider to be "solo", in that they
613 // cannot coexist with other solo states. These are kHasE, kBegunE,
614 // and kHasU. Note that S doesn't conflict with any of these, because
615 // setting the kHasE is only one of the two steps needed to actually
616 // acquire the lock in exclusive mode (the other is draining the existing
618 static constexpr uint32_t kHasSolo = kHasE | kBegunE | kHasU;
620 // Once a thread sets kHasE it needs to wait for the current readers
621 // to exit the lock. We give this a separate wait identity from the
622 // waiting to set kHasE so that we can perform partial wakeups (wake
623 // one instead of wake all).
624 static constexpr uint32_t kWaitingNotS = 1 << 4;
626 // When waking writers we can either wake them all, in which case we
627 // can clear kWaitingE, or we can call futexWake(1). futexWake tells
628 // us if anybody woke up, but even if we detect that nobody woke up we
629 // can't clear the bit after the fact without issuing another wakeup.
630 // To avoid thundering herds when there are lots of pending lock()
631 // without needing to call futexWake twice when there is only one
632 // waiter, kWaitingE actually encodes if we have observed multiple
633 // concurrent waiters. Tricky: ABA issues on futexWait mean that when
634 // we see kWaitingESingle we can't assume that there is only one.
635 static constexpr uint32_t kWaitingESingle = 1 << 2;
636 static constexpr uint32_t kWaitingEMultiple = 1 << 3;
637 static constexpr uint32_t kWaitingE = kWaitingESingle | kWaitingEMultiple;
639 // kWaitingU is essentially a 1 bit saturating counter. It always
640 // requires a wakeAll.
641 static constexpr uint32_t kWaitingU = 1 << 1;
643 // All blocked lock_shared() should be awoken, so it is correct (not
644 // suboptimal) to wakeAll if there are any shared readers.
645 static constexpr uint32_t kWaitingS = 1 << 0;
647 // kWaitingAny is a mask of all of the bits that record the state of
648 // threads, rather than the state of the lock. It is convenient to be
649 // able to mask them off during asserts.
650 static constexpr uint32_t kWaitingAny =
651 kWaitingNotS | kWaitingE | kWaitingU | kWaitingS;
653 // The reader count at which a reader will attempt to use the lock
654 // in deferred mode. If this value is 2, then the second concurrent
655 // reader will set kMayDefer and use deferredReaders[]. kMayDefer is
656 // cleared during exclusive access, so this threshold must be reached
657 // each time a lock is held in exclusive mode.
658 static constexpr uint32_t kNumSharedToStartDeferring = 2;
660 // The typical number of spins that a thread will wait for a state
661 // transition. There is no bound on the number of threads that can wait
662 // for a writer, so we are pretty conservative here to limit the chance
663 // that we are starving the writer of CPU. Each spin is 6 or 7 nanos,
664 // almost all of which is in the pause instruction.
665 static constexpr uint32_t kMaxSpinCount = !BlockImmediately ? 1000 : 2;
667 // The maximum number of soft yields before falling back to futex.
668 // If the preemption heuristic is activated we will fall back before
669 // this. A soft yield takes ~900 nanos (two sched_yield plus a call
670 // to getrusage, with checks of the goal at each step). Soft yields
671 // aren't compatible with deterministic execution under test (unlike
672 // futexWaitUntil, which has a capricious but deterministic back end).
673 static constexpr uint32_t kMaxSoftYieldCount = !BlockImmediately ? 1000 : 0;
675 // If AccessSpreader assigns indexes from 0..k*n-1 on a system where some
676 // level of the memory hierarchy is symmetrically divided into k pieces
677 // (NUMA nodes, last-level caches, L1 caches, ...), then slot indexes
678 // that are the same after integer division by k share that resource.
679 // Our strategy for deferred readers is to probe up to numSlots/4 slots,
680 // using the full granularity of AccessSpreader for the start slot
681 // and then search outward. We can use AccessSpreader::current(n)
682 // without managing our own spreader if kMaxDeferredReaders <=
683 // AccessSpreader::kMaxCpus, which is currently 128.
685 // Our 2-socket E5-2660 machines have 8 L1 caches on each chip,
686 // with 64 byte cache lines. That means we need 64*16 bytes of
687 // deferredReaders[] to give each L1 its own playground. On x86_64
688 // each DeferredReaderSlot is 8 bytes, so we need kMaxDeferredReaders
689 // * kDeferredSeparationFactor >= 64 * 16 / 8 == 128. If
690 // kDeferredSearchDistance * kDeferredSeparationFactor <=
691 // 64 / 8 then we will search only within a single cache line, which
692 // guarantees we won't have inter-L1 contention. We give ourselves
693 // a factor of 2 on the core count, which should hold us for a couple
694 // processor generations. deferredReaders[] is 2048 bytes currently.
696 static constexpr uint32_t kMaxDeferredReaders = 64;
697 static constexpr uint32_t kDeferredSearchDistance = 2;
698 static constexpr uint32_t kDeferredSeparationFactor = 4;
702 static_assert(!(kMaxDeferredReaders & (kMaxDeferredReaders - 1)),
703 "kMaxDeferredReaders must be a power of 2");
704 static_assert(!(kDeferredSearchDistance & (kDeferredSearchDistance - 1)),
705 "kDeferredSearchDistance must be a power of 2");
707 // The number of deferred locks that can be simultaneously acquired
708 // by a thread via the token-less methods without performing any heap
709 // allocations. Each of these costs 3 pointers (24 bytes, probably)
710 // per thread. There's not much point in making this larger than
711 // kDeferredSearchDistance.
712 static constexpr uint32_t kTokenStackTLSCapacity = 2;
714 // We need to make sure that if there is a lock_shared()
715 // and lock_shared(token) followed by unlock_shared() and
716 // unlock_shared(token), the token-less unlock doesn't null
717 // out deferredReaders[token.slot_]. If we allowed that, then
718 // unlock_shared(token) wouldn't be able to assume that its lock
719 // had been inlined by applyDeferredReaders when it finds that
720 // deferredReaders[token.slot_] no longer points to this. We accomplish
721 // this by stealing bit 0 from the pointer to record that the slot's
722 // element has no token, hence our use of uintptr_t in deferredReaders[].
723 static constexpr uintptr_t kTokenless = 0x1;
725 // This is the starting location for Token-less unlock_shared().
726 static FOLLY_SHAREDMUTEX_TLS uint32_t tls_lastTokenlessSlot;
728 // Last deferred reader slot used.
729 static FOLLY_SHAREDMUTEX_TLS uint32_t tls_lastDeferredReaderSlot;
732 // Only indexes divisible by kDeferredSeparationFactor are used.
733 // If any of those elements points to a SharedMutexImpl, then it
734 // should be considered that there is a shared lock on that instance.
737 typedef Atom<uintptr_t> DeferredReaderSlot;
740 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING static DeferredReaderSlot deferredReaders
741 [kMaxDeferredReaders *
742 kDeferredSeparationFactor];
744 // Performs an exclusive lock, waiting for state_ & waitMask to be
746 template <class WaitContext>
747 bool lockExclusiveImpl(uint32_t preconditionGoalMask, WaitContext& ctx) {
748 uint32_t state = state_.load(std::memory_order_acquire);
750 (state & (preconditionGoalMask | kMayDefer | kHasS)) == 0 &&
751 state_.compare_exchange_strong(state, (state | kHasE) & ~kHasU))) {
754 return lockExclusiveImpl(state, preconditionGoalMask, ctx);
758 template <class WaitContext>
759 bool lockExclusiveImpl(uint32_t& state,
760 uint32_t preconditionGoalMask,
763 if (UNLIKELY((state & preconditionGoalMask) != 0) &&
764 !waitForZeroBits(state, preconditionGoalMask, kWaitingE, ctx) &&
769 uint32_t after = (state & kMayDefer) == 0 ? 0 : kPrevDefer;
770 if (!kReaderPriority || (state & (kMayDefer | kHasS)) == 0) {
771 // Block readers immediately, either because we are in write
772 // priority mode or because we can acquire the lock in one
773 // step. Note that if state has kHasU, then we are doing an
774 // unlock_upgrade_and_lock() and we should clear it (reader
775 // priority branch also does this).
776 after |= (state | kHasE) & ~(kHasU | kMayDefer);
778 after |= (state | kBegunE) & ~(kHasU | kMayDefer);
780 if (state_.compare_exchange_strong(state, after)) {
784 // If we set kHasE (writer priority) then no new readers can
785 // arrive. If we set kBegunE then they can still enter, but
786 // they must be inline. Either way we need to either spin on
787 // deferredReaders[] slots, or inline them so that we can wait on
788 // kHasS to zero itself. deferredReaders[] is pointers, which on
789 // x86_64 are bigger than futex() can handle, so we inline the
790 // deferred locks instead of trying to futexWait on each slot.
791 // Readers are responsible for rechecking state_ after recording
792 // a deferred read to avoid atomicity problems between the state_
793 // CAS and applyDeferredReader's reads of deferredReaders[].
794 if (UNLIKELY((before & kMayDefer) != 0)) {
795 applyDeferredReaders(state, ctx);
798 assert((state & (kHasE | kBegunE)) != 0 && (state & kHasU) == 0);
799 if (UNLIKELY((state & kHasS) != 0) &&
800 !waitForZeroBits(state, kHasS, kWaitingNotS, ctx) &&
802 // Ugh. We blocked new readers and other writers for a while,
803 // but were unable to complete. Move on. On the plus side
804 // we can clear kWaitingNotS because nobody else can piggyback
806 state = (state_ &= ~(kPrevDefer | kHasE | kBegunE | kWaitingNotS));
807 wakeRegisteredWaiters(state, kWaitingE | kWaitingU | kWaitingS);
811 if (kReaderPriority && (state & kHasE) == 0) {
812 assert((state & kBegunE) != 0);
813 if (!state_.compare_exchange_strong(state,
814 (state & ~kBegunE) | kHasE)) {
825 template <class WaitContext>
826 bool waitForZeroBits(uint32_t& state,
830 uint32_t spinCount = 0;
832 state = state_.load(std::memory_order_acquire);
833 if ((state & goal) == 0) {
836 asm_volatile_pause();
838 if (UNLIKELY(spinCount >= kMaxSpinCount)) {
839 return ctx.canBlock() &&
840 yieldWaitForZeroBits(state, goal, waitMask, ctx);
845 template <class WaitContext>
846 bool yieldWaitForZeroBits(uint32_t& state,
854 for (uint32_t yieldCount = 0; yieldCount < kMaxSoftYieldCount;
856 for (int softState = 0; softState < 3; ++softState) {
858 std::this_thread::yield();
861 getrusage(RUSAGE_THREAD, &usage);
864 if (((state = state_.load(std::memory_order_acquire)) & goal) == 0) {
867 if (ctx.shouldTimeOut()) {
872 if (before >= 0 && usage.ru_nivcsw >= before + 2) {
873 // One involuntary csw might just be occasional background work,
874 // but if we get two in a row then we guess that there is someone
875 // else who can profitably use this CPU. Fall back to futex
878 before = usage.ru_nivcsw;
881 return futexWaitForZeroBits(state, goal, waitMask, ctx);
884 template <class WaitContext>
885 bool futexWaitForZeroBits(uint32_t& state,
889 assert(waitMask == kWaitingNotS || waitMask == kWaitingE ||
890 waitMask == kWaitingU || waitMask == kWaitingS);
893 state = state_.load(std::memory_order_acquire);
894 if ((state & goal) == 0) {
899 if (waitMask == kWaitingE) {
900 if ((state & kWaitingESingle) != 0) {
901 after |= kWaitingEMultiple;
903 after |= kWaitingESingle;
909 // CAS is better than atomic |= here, because it lets us avoid
910 // setting the wait flag when the goal is concurrently achieved
911 if (after != state && !state_.compare_exchange_strong(state, after)) {
915 if (!ctx.doWait(state_, after, waitMask)) {
922 // Wakes up waiters registered in state_ as appropriate, clearing the
923 // awaiting bits for anybody that was awoken. Tries to perform direct
924 // single wakeup of an exclusive waiter if appropriate
925 void wakeRegisteredWaiters(uint32_t& state, uint32_t wakeMask) {
926 if (UNLIKELY((state & wakeMask) != 0)) {
927 wakeRegisteredWaitersImpl(state, wakeMask);
931 void wakeRegisteredWaitersImpl(uint32_t& state, uint32_t wakeMask) {
932 // If there are multiple lock() pending only one of them will actually
933 // get to wake up, so issuing futexWakeAll will make a thundering herd.
934 // There's nothing stopping us from issuing futexWake(1) instead,
935 // so long as the wait bits are still an accurate reflection of
936 // the waiters. If we notice (via futexWake's return value) that
937 // nobody woke up then we can try again with the normal wake-all path.
938 // Note that we can't just clear the bits at that point; we need to
939 // clear the bits and then issue another wakeup.
941 // It is possible that we wake an E waiter but an outside S grabs the
942 // lock instead, at which point we should wake pending U and S waiters.
943 // Rather than tracking state to make the failing E regenerate the
944 // wakeup, we just disable the optimization in the case that there
945 // are waiting U or S that we are eligible to wake.
946 if ((wakeMask & kWaitingE) == kWaitingE &&
947 (state & wakeMask) == kWaitingE &&
948 state_.futexWake(1, kWaitingE) > 0) {
949 // somebody woke up, so leave state_ as is and clear it later
953 if ((state & wakeMask) != 0) {
954 auto prev = state_.fetch_and(~wakeMask);
955 if ((prev & wakeMask) != 0) {
956 futexWakeAll(wakeMask);
958 state = prev & ~wakeMask;
962 void futexWakeAll(uint32_t wakeMask) {
963 state_.futexWake(std::numeric_limits<int>::max(), wakeMask);
966 DeferredReaderSlot* deferredReader(uint32_t slot) {
967 return &deferredReaders[slot * kDeferredSeparationFactor];
970 uintptr_t tokenfulSlotValue() { return reinterpret_cast<uintptr_t>(this); }
972 uintptr_t tokenlessSlotValue() { return tokenfulSlotValue() | kTokenless; }
974 bool slotValueIsThis(uintptr_t slotValue) {
975 return (slotValue & ~kTokenless) == tokenfulSlotValue();
978 // Clears any deferredReaders[] that point to this, adjusting the inline
979 // shared lock count to compensate. Does some spinning and yielding
980 // to avoid the work. Always finishes the application, even if ctx
982 template <class WaitContext>
983 void applyDeferredReaders(uint32_t& state, WaitContext& ctx) {
986 uint32_t spinCount = 0;
988 while (!slotValueIsThis(
989 deferredReader(slot)->load(std::memory_order_acquire))) {
990 if (++slot == kMaxDeferredReaders) {
994 asm_volatile_pause();
995 if (UNLIKELY(++spinCount >= kMaxSpinCount)) {
996 applyDeferredReaders(state, ctx, slot);
1002 template <class WaitContext>
1003 void applyDeferredReaders(uint32_t& state, WaitContext& ctx, uint32_t slot) {
1005 #ifdef RUSAGE_THREAD
1006 struct rusage usage;
1009 for (uint32_t yieldCount = 0; yieldCount < kMaxSoftYieldCount;
1011 for (int softState = 0; softState < 3; ++softState) {
1012 if (softState < 2) {
1013 std::this_thread::yield();
1015 #ifdef RUSAGE_THREAD
1016 getrusage(RUSAGE_THREAD, &usage);
1019 while (!slotValueIsThis(
1020 deferredReader(slot)->load(std::memory_order_acquire))) {
1021 if (++slot == kMaxDeferredReaders) {
1025 if (ctx.shouldTimeOut()) {
1026 // finish applying immediately on timeout
1030 #ifdef RUSAGE_THREAD
1031 if (before >= 0 && usage.ru_nivcsw >= before + 2) {
1032 // heuristic says run queue is not empty
1035 before = usage.ru_nivcsw;
1039 uint32_t movedSlotCount = 0;
1040 for (; slot < kMaxDeferredReaders; ++slot) {
1041 auto slotPtr = deferredReader(slot);
1042 auto slotValue = slotPtr->load(std::memory_order_acquire);
1043 if (slotValueIsThis(slotValue) &&
1044 slotPtr->compare_exchange_strong(slotValue, 0)) {
1049 if (movedSlotCount > 0) {
1050 state = (state_ += movedSlotCount * kIncrHasS);
1052 assert((state & (kHasE | kBegunE)) != 0);
1054 // if state + kIncrHasS overflows (off the end of state) then either
1055 // we have 2^(32-9) readers (almost certainly an application bug)
1056 // or we had an underflow (also a bug)
1057 assert(state < state + kIncrHasS);
1060 // It is straightfoward to make a token-less lock_shared() and
1061 // unlock_shared() either by making the token-less version always use
1062 // INLINE_SHARED mode or by removing the token version. Supporting
1063 // deferred operation for both types is trickier than it appears, because
1064 // the purpose of the token it so that unlock_shared doesn't have to
1065 // look in other slots for its deferred lock. Token-less unlock_shared
1066 // might place a deferred lock in one place and then release a different
1067 // slot that was originally used by the token-ful version. If this was
1068 // important we could solve the problem by differentiating the deferred
1069 // locks so that cross-variety release wouldn't occur. The best way
1070 // is probably to steal a bit from the pointer, making deferredLocks[]
1071 // an array of Atom<uintptr_t>.
1073 template <class WaitContext>
1074 bool lockSharedImpl(Token* token, WaitContext& ctx) {
1075 uint32_t state = state_.load(std::memory_order_relaxed);
1076 if ((state & (kHasS | kMayDefer | kHasE)) == 0 &&
1077 state_.compare_exchange_strong(state, state + kIncrHasS)) {
1078 if (token != nullptr) {
1079 token->type_ = Token::Type::INLINE_SHARED;
1083 return lockSharedImpl(state, token, ctx);
1086 template <class WaitContext>
1087 bool lockSharedImpl(uint32_t& state, Token* token, WaitContext& ctx);
1089 // Updates the state in/out argument as if the locks were made inline,
1090 // but does not update state_
1091 void cleanupTokenlessSharedDeferred(uint32_t& state) {
1092 for (uint32_t i = 0; i < kMaxDeferredReaders; ++i) {
1093 auto slotPtr = deferredReader(i);
1094 auto slotValue = slotPtr->load(std::memory_order_relaxed);
1095 if (slotValue == tokenlessSlotValue()) {
1096 slotPtr->store(0, std::memory_order_relaxed);
1098 if ((state & kHasS) == 0) {
1105 bool tryUnlockTokenlessSharedDeferred();
1107 bool tryUnlockSharedDeferred(uint32_t slot) {
1108 assert(slot < kMaxDeferredReaders);
1109 auto slotValue = tokenfulSlotValue();
1110 return deferredReader(slot)->compare_exchange_strong(slotValue, 0);
1113 uint32_t unlockSharedInline() {
1114 uint32_t state = (state_ -= kIncrHasS);
1115 assert((state & (kHasE | kBegunE | kMayDefer)) != 0 ||
1116 state < state + kIncrHasS);
1117 if ((state & kHasS) == 0) {
1118 // Only the second half of lock() can be blocked by a non-zero
1119 // reader count, so that's the only thing we need to wake
1120 wakeRegisteredWaiters(state, kWaitingNotS);
1125 template <class WaitContext>
1126 bool lockUpgradeImpl(WaitContext& ctx) {
1129 if (!waitForZeroBits(state, kHasSolo, kWaitingU, ctx)) {
1132 } while (!state_.compare_exchange_strong(state, state | kHasU));
1138 ReadHolder() : lock_(nullptr) {}
1141 explicit ReadHolder(const SharedMutexImpl* lock)
1142 : lock_(const_cast<SharedMutexImpl*>(lock)) {
1144 lock_->lock_shared(token_);
1148 explicit ReadHolder(const SharedMutexImpl& lock)
1149 : lock_(const_cast<SharedMutexImpl*>(&lock)) {
1150 lock_->lock_shared(token_);
1153 ReadHolder(ReadHolder&& rhs) noexcept : lock_(rhs.lock_),
1154 token_(rhs.token_) {
1155 rhs.lock_ = nullptr;
1158 // Downgrade from upgrade mode
1159 explicit ReadHolder(UpgradeHolder&& upgraded) : lock_(upgraded.lock_) {
1160 assert(upgraded.lock_ != nullptr);
1161 upgraded.lock_ = nullptr;
1162 lock_->unlock_upgrade_and_lock_shared(token_);
1165 // Downgrade from exclusive mode
1166 explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
1167 assert(writer.lock_ != nullptr);
1168 writer.lock_ = nullptr;
1169 lock_->unlock_and_lock_shared(token_);
1172 ReadHolder& operator=(ReadHolder&& rhs) noexcept {
1173 std::swap(lock_, rhs.lock_);
1174 std::swap(token_, rhs.token_);
1178 ReadHolder(const ReadHolder& rhs) = delete;
1179 ReadHolder& operator=(const ReadHolder& rhs) = delete;
1187 lock_->unlock_shared(token_);
1193 friend class UpgradeHolder;
1194 friend class WriteHolder;
1195 SharedMutexImpl* lock_;
1196 SharedMutexToken token_;
1199 class UpgradeHolder {
1200 UpgradeHolder() : lock_(nullptr) {}
1203 explicit UpgradeHolder(SharedMutexImpl* lock) : lock_(lock) {
1205 lock_->lock_upgrade();
1209 explicit UpgradeHolder(SharedMutexImpl& lock) : lock_(&lock) {
1210 lock_->lock_upgrade();
1213 // Downgrade from exclusive mode
1214 explicit UpgradeHolder(WriteHolder&& writer) : lock_(writer.lock_) {
1215 assert(writer.lock_ != nullptr);
1216 writer.lock_ = nullptr;
1217 lock_->unlock_and_lock_upgrade();
1220 UpgradeHolder(UpgradeHolder&& rhs) noexcept : lock_(rhs.lock_) {
1221 rhs.lock_ = nullptr;
1224 UpgradeHolder& operator=(UpgradeHolder&& rhs) noexcept {
1225 std::swap(lock_, rhs.lock_);
1229 UpgradeHolder(const UpgradeHolder& rhs) = delete;
1230 UpgradeHolder& operator=(const UpgradeHolder& rhs) = delete;
1238 lock_->unlock_upgrade();
1244 friend class WriteHolder;
1245 friend class ReadHolder;
1246 SharedMutexImpl* lock_;
1250 WriteHolder() : lock_(nullptr) {}
1253 explicit WriteHolder(SharedMutexImpl* lock) : lock_(lock) {
1259 explicit WriteHolder(SharedMutexImpl& lock) : lock_(&lock) {
1263 // Promotion from upgrade mode
1264 explicit WriteHolder(UpgradeHolder&& upgrade) : lock_(upgrade.lock_) {
1265 assert(upgrade.lock_ != nullptr);
1266 upgrade.lock_ = nullptr;
1267 lock_->unlock_upgrade_and_lock();
1272 // It is intended that WriteHolder(ReadHolder&& rhs) do not exist.
1274 // Shared locks (read) can not safely upgrade to unique locks (write).
1275 // That upgrade path is a well-known recipe for deadlock, so we explicitly
1278 // If you need to do a conditional mutation, you have a few options:
1279 // 1. Check the condition under a shared lock and release it.
1280 // Then maybe check the condition again under a unique lock and maybe do
1282 // 2. Check the condition once under an upgradeable lock.
1283 // Then maybe upgrade the lock to a unique lock and do the mutation.
1284 // 3. Check the condition and maybe perform the mutation under a unique
1287 // Relevant upgradeable lock notes:
1288 // * At most one upgradeable lock can be held at a time for a given shared
1289 // mutex, just like a unique lock.
1290 // * An upgradeable lock may be held concurrently with any number of shared
1292 // * An upgradeable lock may be upgraded atomically to a unique lock.
1294 WriteHolder(WriteHolder&& rhs) noexcept : lock_(rhs.lock_) {
1295 rhs.lock_ = nullptr;
1298 WriteHolder& operator=(WriteHolder&& rhs) noexcept {
1299 std::swap(lock_, rhs.lock_);
1303 WriteHolder(const WriteHolder& rhs) = delete;
1304 WriteHolder& operator=(const WriteHolder& rhs) = delete;
1318 friend class ReadHolder;
1319 friend class UpgradeHolder;
1320 SharedMutexImpl* lock_;
1323 // Adapters for Synchronized<>
1324 friend void acquireRead(SharedMutexImpl& lock) { lock.lock_shared(); }
1325 friend void acquireReadWrite(SharedMutexImpl& lock) { lock.lock(); }
1326 friend void releaseRead(SharedMutexImpl& lock) { lock.unlock_shared(); }
1327 friend void releaseReadWrite(SharedMutexImpl& lock) { lock.unlock(); }
1328 friend bool acquireRead(SharedMutexImpl& lock, unsigned int ms) {
1329 return lock.try_lock_shared_for(std::chrono::milliseconds(ms));
1331 friend bool acquireReadWrite(SharedMutexImpl& lock, unsigned int ms) {
1332 return lock.try_lock_for(std::chrono::milliseconds(ms));
1336 typedef SharedMutexImpl<true> SharedMutexReadPriority;
1337 typedef SharedMutexImpl<false> SharedMutexWritePriority;
1338 typedef SharedMutexWritePriority SharedMutex;
1340 // Prevent the compiler from instantiating these in other translation units.
1341 // They are instantiated once in SharedMutex.cpp
1342 extern template class SharedMutexImpl<true>;
1343 extern template class SharedMutexImpl<false>;
1346 bool ReaderPriority,
1348 template <typename> class Atom,
1349 bool BlockImmediately>
1350 typename SharedMutexImpl<ReaderPriority, Tag_, Atom, BlockImmediately>::
1352 SharedMutexImpl<ReaderPriority, Tag_, Atom, BlockImmediately>::
1353 deferredReaders[kMaxDeferredReaders * kDeferredSeparationFactor] =
1357 bool ReaderPriority,
1359 template <typename> class Atom,
1360 bool BlockImmediately>
1361 FOLLY_SHAREDMUTEX_TLS uint32_t
1362 SharedMutexImpl<ReaderPriority, Tag_, Atom, BlockImmediately>::
1363 tls_lastTokenlessSlot = 0;
1366 bool ReaderPriority,
1368 template <typename> class Atom,
1369 bool BlockImmediately>
1370 FOLLY_SHAREDMUTEX_TLS uint32_t
1371 SharedMutexImpl<ReaderPriority, Tag_, Atom, BlockImmediately>::
1372 tls_lastDeferredReaderSlot = 0;
1375 bool ReaderPriority,
1377 template <typename> class Atom,
1378 bool BlockImmediately>
1379 bool SharedMutexImpl<ReaderPriority, Tag_, Atom, BlockImmediately>::
1380 tryUnlockTokenlessSharedDeferred() {
1381 auto bestSlot = tls_lastTokenlessSlot;
1382 for (uint32_t i = 0; i < kMaxDeferredReaders; ++i) {
1383 auto slotPtr = deferredReader(bestSlot ^ i);
1384 auto slotValue = slotPtr->load(std::memory_order_relaxed);
1385 if (slotValue == tokenlessSlotValue() &&
1386 slotPtr->compare_exchange_strong(slotValue, 0)) {
1387 tls_lastTokenlessSlot = bestSlot ^ i;
1395 bool ReaderPriority,
1397 template <typename> class Atom,
1398 bool BlockImmediately>
1399 template <class WaitContext>
1400 bool SharedMutexImpl<ReaderPriority, Tag_, Atom, BlockImmediately>::
1401 lockSharedImpl(uint32_t& state, Token* token, WaitContext& ctx) {
1403 if (UNLIKELY((state & kHasE) != 0) &&
1404 !waitForZeroBits(state, kHasE, kWaitingS, ctx) && ctx.canTimeOut()) {
1408 uint32_t slot = tls_lastDeferredReaderSlot;
1409 uintptr_t slotValue = 1; // any non-zero value will do
1411 bool canAlreadyDefer = (state & kMayDefer) != 0;
1412 bool aboveDeferThreshold =
1413 (state & kHasS) >= (kNumSharedToStartDeferring - 1) * kIncrHasS;
1414 bool drainInProgress = ReaderPriority && (state & kBegunE) != 0;
1415 if (canAlreadyDefer || (aboveDeferThreshold && !drainInProgress)) {
1416 /* Try using the most recent slot first. */
1417 slotValue = deferredReader(slot)->load(std::memory_order_relaxed);
1418 if (slotValue != 0) {
1419 // starting point for our empty-slot search, can change after
1420 // calling waitForZeroBits
1422 (uint32_t)folly::AccessSpreader<Atom>::current(kMaxDeferredReaders);
1424 // deferred readers are already enabled, or it is time to
1425 // enable them if we can find a slot
1426 for (uint32_t i = 0; i < kDeferredSearchDistance; ++i) {
1427 slot = bestSlot ^ i;
1428 assert(slot < kMaxDeferredReaders);
1429 slotValue = deferredReader(slot)->load(std::memory_order_relaxed);
1430 if (slotValue == 0) {
1432 tls_lastDeferredReaderSlot = slot;
1439 if (slotValue != 0) {
1440 // not yet deferred, or no empty slots
1441 if (state_.compare_exchange_strong(state, state + kIncrHasS)) {
1442 // successfully recorded the read lock inline
1443 if (token != nullptr) {
1444 token->type_ = Token::Type::INLINE_SHARED;
1448 // state is updated, try again
1452 // record that deferred readers might be in use if necessary
1453 if ((state & kMayDefer) == 0) {
1454 if (!state_.compare_exchange_strong(state, state | kMayDefer)) {
1455 // keep going if CAS failed because somebody else set the bit
1457 if ((state & (kHasE | kMayDefer)) != kMayDefer) {
1461 // state = state | kMayDefer;
1464 // try to use the slot
1465 bool gotSlot = deferredReader(slot)->compare_exchange_strong(
1467 token == nullptr ? tokenlessSlotValue() : tokenfulSlotValue());
1469 // If we got the slot, we need to verify that an exclusive lock
1470 // didn't happen since we last checked. If we didn't get the slot we
1471 // need to recheck state_ anyway to make sure we don't waste too much
1472 // work. It is also possible that since we checked state_ someone
1473 // has acquired and released the write lock, clearing kMayDefer.
1474 // Both cases are covered by looking for the readers-possible bit,
1475 // because it is off when the exclusive lock bit is set.
1476 state = state_.load(std::memory_order_acquire);
1482 if (token == nullptr) {
1483 tls_lastTokenlessSlot = slot;
1486 if ((state & kMayDefer) != 0) {
1487 assert((state & kHasE) == 0);
1489 if (token != nullptr) {
1490 token->type_ = Token::Type::DEFERRED_SHARED;
1491 token->slot_ = (uint16_t)slot;
1496 // release the slot before retrying
1497 if (token == nullptr) {
1498 // We can't rely on slot. Token-less slot values can be freed by
1499 // any unlock_shared(), so we need to do the full deferredReader
1500 // search during unlock. Unlike unlock_shared(), we can't trust
1501 // kPrevDefer here. This deferred lock isn't visible to lock()
1502 // (that's the whole reason we're undoing it) so there might have
1503 // subsequently been an unlock() and lock() with no intervening
1504 // transition to deferred mode.
1505 if (!tryUnlockTokenlessSharedDeferred()) {
1506 unlockSharedInline();
1509 if (!tryUnlockSharedDeferred(slot)) {
1510 unlockSharedInline();
1514 // We got here not because the lock was unavailable, but because
1515 // we lost a compare-and-swap. Try-lock is typically allowed to
1516 // have spurious failures, but there is no lock efficiency gain
1517 // from exploiting that freedom here.
1521 } // namespace folly