2 * Copyright 2015 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)
24 #include <type_traits>
25 #include <folly/Likely.h>
26 #include <folly/detail/CacheLocality.h>
27 #include <folly/detail/Futex.h>
28 #include <sys/resource.h>
30 // SharedMutex is a reader-writer lock. It is small, very fast, scalable
31 // on multi-core, and suitable for use when readers or writers may block.
32 // Unlike most other reader-writer locks, its throughput with concurrent
33 // readers scales linearly; it is able to acquire and release the lock
34 // in shared mode without cache line ping-ponging. It is suitable for
35 // a wide range of lock hold times because it starts with spinning,
36 // proceeds to using sched_yield with a preemption heuristic, and then
37 // waits using futex and precise wakeups.
39 // SharedMutex provides all of the methods of folly::RWSpinLock,
40 // boost::shared_mutex, boost::upgrade_mutex, and C++14's
41 // std::shared_timed_mutex. All operations that can block are available
42 // in try, try-for, and try-until (system_clock or steady_clock) versions.
44 // SharedMutexReadPriority gives priority to readers,
45 // SharedMutexWritePriority gives priority to writers. SharedMutex is an
46 // alias for SharedMutexWritePriority, because writer starvation is more
47 // likely than reader starvation for the read-heavy workloads targetted
50 // In my tests SharedMutex is as good or better than the other
51 // reader-writer locks in use at Facebook for almost all use cases,
52 // sometimes by a wide margin. (If it is rare that there are actually
53 // concurrent readers then RWSpinLock can be a few nanoseconds faster.)
54 // I compared it to folly::RWSpinLock, folly::RWTicketSpinLock64,
55 // boost::shared_mutex, pthread_rwlock_t, and a RWLock that internally uses
56 // spinlocks to guard state and pthread_mutex_t+pthread_cond_t to block.
57 // (Thrift's ReadWriteMutex is based underneath on pthread_rwlock_t.)
58 // It is generally as good or better than the rest when evaluating size,
59 // speed, scalability, or latency outliers. In the corner cases where
60 // it is not the fastest (such as single-threaded use or heavy write
61 // contention) it is never very much worse than the best. See the bottom
62 // of folly/test/SharedMutexTest.cpp for lots of microbenchmark results.
64 // Comparison to folly::RWSpinLock:
66 // * SharedMutex is faster than RWSpinLock when there are actually
67 // concurrent read accesses (sometimes much faster), and ~5 nanoseconds
68 // slower when there is not actually any contention. SharedMutex is
69 // faster in every (benchmarked) scenario where the shared mode of
70 // the lock is actually useful.
72 // * Concurrent shared access to SharedMutex scales linearly, while total
73 // RWSpinLock throughput drops as more threads try to access the lock
74 // in shared mode. Under very heavy read contention SharedMutex can
75 // be two orders of magnitude faster than RWSpinLock (or any reader
76 // writer lock that doesn't use striping or deferral).
78 // * SharedMutex can safely protect blocking calls, because after an
79 // initial period of spinning it waits using futex().
81 // * RWSpinLock prioritizes readers, SharedMutex has both reader- and
82 // writer-priority variants, but defaults to write priority.
84 // * RWSpinLock's upgradeable mode blocks new readers, while SharedMutex's
85 // doesn't. Both semantics are reasonable. The boost documentation
86 // doesn't explicitly talk about this behavior (except by omitting
87 // any statement that those lock modes conflict), but the boost
88 // implementations do allow new readers while the upgradeable mode
89 // is held. See https://github.com/boostorg/thread/blob/master/
90 // include/boost/thread/pthread/shared_mutex.hpp
92 // * RWSpinLock::UpgradedHolder maps to SharedMutex::UpgradeHolder
93 // (UpgradeableHolder would be even more pedantically correct).
94 // SharedMutex's holders have fewer methods (no reset) and are less
95 // tolerant (promotion and downgrade crash if the donor doesn't own
96 // the lock, and you must use the default constructor rather than
97 // passing a nullptr to the pointer constructor).
99 // Both SharedMutex and RWSpinLock provide "exclusive", "upgrade",
100 // and "shared" modes. At all times num_threads_holding_exclusive +
101 // num_threads_holding_upgrade <= 1, and num_threads_holding_exclusive ==
102 // 0 || num_threads_holding_shared == 0. RWSpinLock has the additional
103 // constraint that num_threads_holding_shared cannot increase while
104 // num_threads_holding_upgrade is non-zero.
106 // Comparison to the internal RWLock:
108 // * SharedMutex doesn't allow a maximum reader count to be configured,
109 // so it can't be used as a semaphore in the same way as RWLock.
111 // * SharedMutex is 4 bytes, RWLock is 256.
113 // * SharedMutex is as fast or faster than RWLock in all of my
114 // microbenchmarks, and has positive rather than negative scalability.
116 // * RWLock and SharedMutex are both writer priority locks.
118 // * SharedMutex avoids latency outliers as well as RWLock.
120 // * SharedMutex uses different names (t != 0 below):
122 // RWLock::lock(0) => SharedMutex::lock()
124 // RWLock::lock(t) => SharedMutex::try_lock_for(milliseconds(t))
126 // RWLock::tryLock() => SharedMutex::try_lock()
128 // RWLock::unlock() => SharedMutex::unlock()
130 // RWLock::enter(0) => SharedMutex::lock_shared()
132 // RWLock::enter(t) =>
133 // SharedMutex::try_lock_shared_for(milliseconds(t))
135 // RWLock::tryEnter() => SharedMutex::try_lock_shared()
137 // RWLock::leave() => SharedMutex::unlock_shared()
139 // * RWLock allows the reader count to be adjusted by a value other
140 // than 1 during enter() or leave(). SharedMutex doesn't currently
141 // implement this feature.
143 // * RWLock's methods are marked const, SharedMutex's aren't.
145 // Reader-writer locks have the potential to allow concurrent access
146 // to shared read-mostly data, but in practice they often provide no
147 // improvement over a mutex. The problem is the cache coherence protocol
148 // of modern CPUs. Coherence is provided by making sure that when a cache
149 // line is written it is present in only one core's cache. Since a memory
150 // write is required to acquire a reader-writer lock in shared mode, the
151 // cache line holding the lock is invalidated in all of the other caches.
152 // This leads to cache misses when another thread wants to acquire or
153 // release the lock concurrently. When the RWLock is colocated with the
154 // data it protects (common), cache misses can also continue occur when
155 // a thread that already holds the lock tries to read the protected data.
157 // Ideally, a reader-writer lock would allow multiple cores to acquire
158 // and release the lock in shared mode without incurring any cache misses.
159 // This requires that each core records its shared access in a cache line
160 // that isn't read or written by other read-locking cores. (Writers will
161 // have to check all of the cache lines.) Typical server hardware when
162 // this comment was written has 16 L1 caches and cache lines of 64 bytes,
163 // so a lock striped over all L1 caches would occupy a prohibitive 1024
164 // bytes. Nothing says that we need a separate set of per-core memory
165 // locations for each lock, however. Each SharedMutex instance is only
166 // 4 bytes, but all locks together share a 2K area in which they make a
167 // core-local record of lock acquisitions.
169 // SharedMutex's strategy of using a shared set of core-local stripes has
170 // a potential downside, because it means that acquisition of any lock in
171 // write mode can conflict with acquisition of any lock in shared mode.
172 // If a lock instance doesn't actually experience concurrency then this
173 // downside will outweight the upside of improved scalability for readers.
174 // To avoid this problem we dynamically detect concurrent accesses to
175 // SharedMutex, and don't start using the deferred mode unless we actually
176 // observe concurrency. See kNumSharedToStartDeferring.
178 // It is explicitly allowed to call lock_unshared() from a different
179 // thread than lock_shared(), so long as they are properly paired.
180 // lock_unshared() needs to find the location at which lock_shared()
181 // recorded the lock, which might be in the lock itself or in any of
182 // the shared slots. If you can conveniently pass state from lock
183 // acquisition to release then the fastest mechanism is to std::move
184 // the SharedMutex::ReadHolder instance or an SharedMutex::Token (using
185 // lock_shared(Token&) and unlock_sahred(Token&)). The guard or token
186 // will tell unlock_shared where in deferredReaders[] to look for the
187 // deferred lock. The Token-less version of unlock_shared() works in all
188 // cases, but is optimized for the common (no inter-thread handoff) case.
190 // In both read- and write-priority mode, a waiting lock() (exclusive mode)
191 // only blocks readers after it has waited for an active upgrade lock to be
192 // released; until the upgrade lock is released (or upgraded or downgraded)
193 // readers will still be able to enter. Preferences about lock acquisition
194 // are not guaranteed to be enforced perfectly (even if they were, there
195 // is theoretically the chance that a thread could be arbitrarily suspended
196 // between calling lock() and SharedMutex code actually getting executed).
198 // try_*_for methods always try at least once, even if the duration
199 // is zero or negative. The duration type must be compatible with
200 // std::chrono::steady_clock. try_*_until methods also always try at
201 // least once. std::chrono::system_clock and std::chrono::steady_clock
204 // If you have observed by profiling that your SharedMutex-s are getting
205 // cache misses on deferredReaders[] due to another SharedMutex user, then
206 // you can use the tag type plus the RWDEFERREDLOCK_DECLARE_STATIC_STORAGE
207 // macro to create your own instantiation of the type. The contention
208 // threshold (see kNumSharedToStartDeferring) should make this unnecessary
209 // in all but the most extreme cases. Make sure to check that the
210 // increased icache and dcache footprint of the tagged result is worth it.
214 struct SharedMutexToken {
215 enum class Type : uint16_t {
225 template <bool ReaderPriority,
226 typename Tag_ = void,
227 template <typename> class Atom = std::atomic,
228 bool BlockImmediately = false>
229 class SharedMutexImpl {
231 static constexpr bool kReaderPriority = ReaderPriority;
234 typedef SharedMutexToken Token;
240 SharedMutexImpl() : state_(0) {}
242 SharedMutexImpl(const SharedMutexImpl&) = delete;
243 SharedMutexImpl(SharedMutexImpl&&) = delete;
244 SharedMutexImpl& operator = (const SharedMutexImpl&) = delete;
245 SharedMutexImpl& operator = (SharedMutexImpl&&) = delete;
247 // It is an error to destroy an SharedMutex that still has
248 // any outstanding locks. This is checked if NDEBUG isn't defined.
249 // SharedMutex's exclusive mode can be safely used to guard the lock's
250 // own destruction. If, for example, you acquire the lock in exclusive
251 // mode and then observe that the object containing the lock is no longer
252 // needed, you can unlock() and then immediately destroy the lock.
253 // See https://sourceware.org/bugzilla/show_bug.cgi?id=13690 for a
254 // description about why this property needs to be explicitly mentioned.
257 auto state = state_.load(std::memory_order_acquire);
259 // if a futexWait fails to go to sleep because the value has been
260 // changed, we don't necessarily clean up the wait bits, so it is
261 // possible they will be set here in a correct system
262 assert((state & ~(kWaitingAny | kMayDefer)) == 0);
263 if ((state & kMayDefer) != 0) {
264 for (uint32_t slot = 0; slot < kMaxDeferredReaders; ++slot) {
265 auto slotValue = deferredReader(slot)->load(std::memory_order_acquire);
266 assert(!slotValueIsThis(slotValue));
274 (void)lockExclusiveImpl(kHasSolo, ctx);
279 return lockExclusiveImpl(kHasSolo, ctx);
282 template <class Rep, class Period>
283 bool try_lock_for(const std::chrono::duration<Rep, Period>& duration) {
284 WaitForDuration<Rep, Period> ctx(duration);
285 return lockExclusiveImpl(kHasSolo, ctx);
288 template <class Clock, class Duration>
290 const std::chrono::time_point<Clock, Duration>& absDeadline) {
291 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
292 return lockExclusiveImpl(kHasSolo, ctx);
296 // It is possible that we have a left-over kWaitingNotS if the last
297 // unlock_shared() that let our matching lock() complete finished
298 // releasing before lock()'s futexWait went to sleep. Clean it up now
299 auto state = (state_ &= ~(kWaitingNotS | kPrevDefer | kHasE));
300 assert((state & ~kWaitingAny) == 0);
301 wakeRegisteredWaiters(state, kWaitingE | kWaitingU | kWaitingS);
304 // Managing the token yourself makes unlock_shared a bit faster
308 (void)lockSharedImpl(nullptr, ctx);
311 void lock_shared(Token& token) {
313 (void)lockSharedImpl(&token, ctx);
316 bool try_lock_shared() {
318 return lockSharedImpl(nullptr, ctx);
321 bool try_lock_shared(Token& token) {
323 return lockSharedImpl(&token, ctx);
326 template <class Rep, class Period>
327 bool try_lock_shared_for(const std::chrono::duration<Rep, Period>& duration) {
328 WaitForDuration<Rep, Period> ctx(duration);
329 return lockSharedImpl(nullptr, ctx);
332 template <class Rep, class Period>
333 bool try_lock_shared_for(const std::chrono::duration<Rep, Period>& duration,
335 WaitForDuration<Rep, Period> ctx(duration);
336 return lockSharedImpl(&token, ctx);
339 template <class Clock, class Duration>
340 bool try_lock_shared_until(
341 const std::chrono::time_point<Clock, Duration>& absDeadline) {
342 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
343 return lockSharedImpl(nullptr, ctx);
346 template <class Clock, class Duration>
347 bool try_lock_shared_until(
348 const std::chrono::time_point<Clock, Duration>& absDeadline,
350 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
351 return lockSharedImpl(&token, ctx);
354 void unlock_shared() {
355 auto state = state_.load(std::memory_order_acquire);
357 // kPrevDefer can only be set if HasE or BegunE is set
358 assert((state & (kPrevDefer | kHasE | kBegunE)) != kPrevDefer);
360 // lock() strips kMayDefer immediately, but then copies it to
361 // kPrevDefer so we can tell if the pre-lock() lock_shared() might
363 if ((state & (kMayDefer | kPrevDefer)) == 0 ||
364 !tryUnlockAnySharedDeferred()) {
365 // Matching lock_shared() couldn't have deferred, or the deferred
366 // lock has already been inlined by applyDeferredReaders()
367 unlockSharedInline();
371 void unlock_shared(Token& token) {
372 assert(token.type_ == Token::Type::INLINE_SHARED ||
373 token.type_ == Token::Type::DEFERRED_SHARED);
375 if (token.type_ != Token::Type::DEFERRED_SHARED ||
376 !tryUnlockSharedDeferred(token.slot_)) {
377 unlockSharedInline();
380 token.type_ = Token::Type::INVALID;
384 void unlock_and_lock_shared() {
385 // We can't use state_ -=, because we need to clear 2 bits (1 of which
386 // has an uncertain initial state) and set 1 other. We might as well
387 // clear the relevant wake bits at the same time. Note that since S
388 // doesn't block the beginning of a transition to E (writer priority
389 // can cut off new S, reader priority grabs BegunE and blocks deferred
390 // S) we need to wake E as well.
391 auto state = state_.load(std::memory_order_acquire);
393 assert((state & ~(kWaitingAny | kPrevDefer)) == kHasE);
394 } while (!state_.compare_exchange_strong(
395 state, (state & ~(kWaitingAny | kPrevDefer | kHasE)) + kIncrHasS));
396 if ((state & (kWaitingE | kWaitingU | kWaitingS)) != 0) {
397 futexWakeAll(kWaitingE | kWaitingU | kWaitingS);
401 void unlock_and_lock_shared(Token& token) {
402 unlock_and_lock_shared();
403 token.type_ = Token::Type::INLINE_SHARED;
406 void lock_upgrade() {
408 (void)lockUpgradeImpl(ctx);
411 bool try_lock_upgrade() {
413 return lockUpgradeImpl(ctx);
416 template <class Rep, class Period>
417 bool try_lock_upgrade_for(
418 const std::chrono::duration<Rep, Period>& duration) {
419 WaitForDuration<Rep, Period> ctx(duration);
420 return lockUpgradeImpl(ctx);
423 template <class Clock, class Duration>
424 bool try_lock_upgrade_until(
425 const std::chrono::time_point<Clock, Duration>& absDeadline) {
426 WaitUntilDeadline<Clock, Duration> ctx{absDeadline};
427 return lockUpgradeImpl(ctx);
430 void unlock_upgrade() {
431 auto state = (state_ -= kHasU);
432 assert((state & (kWaitingNotS | kHasSolo)) == 0);
433 wakeRegisteredWaiters(state, kWaitingE | kWaitingU);
436 void unlock_upgrade_and_lock() {
437 // no waiting necessary, so waitMask is empty
439 (void)lockExclusiveImpl(0, ctx);
442 void unlock_upgrade_and_lock_shared() {
443 auto state = (state_ -= kHasU - kIncrHasS);
444 assert((state & (kWaitingNotS | kHasSolo)) == 0 && (state & kHasS) != 0);
445 wakeRegisteredWaiters(state, kWaitingE | kWaitingU);
448 void unlock_upgrade_and_lock_shared(Token& token) {
449 unlock_upgrade_and_lock_shared();
450 token.type_ = Token::Type::INLINE_SHARED;
453 void unlock_and_lock_upgrade() {
454 // We can't use state_ -=, because we need to clear 2 bits (1 of
455 // which has an uncertain initial state) and set 1 other. We might
456 // as well clear the relevant wake bits at the same time.
457 auto state = state_.load(std::memory_order_acquire);
459 assert((state & ~(kWaitingAny | kPrevDefer)) == kHasE);
461 (state & ~(kWaitingNotS | kWaitingS | kPrevDefer | kHasE)) + kHasU;
462 if (state_.compare_exchange_strong(state, after)) {
463 if ((state & kWaitingS) != 0) {
464 futexWakeAll(kWaitingS);
472 typedef typename folly::detail::Futex<Atom> Futex;
474 // Internally we use four kinds of wait contexts. These are structs
475 // that provide a doWait method that returns true if a futex wake
476 // was issued that intersects with the waitMask, false if there was a
477 // timeout and no more waiting should be performed. Spinning occurs
478 // before the wait context is invoked.
481 bool canBlock() { return true; }
482 bool canTimeOut() { return false; }
483 bool shouldTimeOut() { return false; }
485 bool doWait(Futex& futex, uint32_t expected, uint32_t waitMask) {
486 futex.futexWait(expected, waitMask);
492 bool canBlock() { return false; }
493 bool canTimeOut() { return true; }
494 bool shouldTimeOut() { return true; }
496 bool doWait(Futex& futex, uint32_t expected, uint32_t waitMask) {
501 template <class Rep, class Period>
502 struct WaitForDuration {
503 std::chrono::duration<Rep, Period> duration_;
504 bool deadlineComputed_;
505 std::chrono::steady_clock::time_point deadline_;
507 explicit WaitForDuration(const std::chrono::duration<Rep, Period>& duration)
508 : duration_(duration), deadlineComputed_(false) {}
510 std::chrono::steady_clock::time_point deadline() {
511 if (!deadlineComputed_) {
512 deadline_ = std::chrono::steady_clock::now() + duration_;
513 deadlineComputed_ = true;
518 bool canBlock() { return duration_.count() > 0; }
519 bool canTimeOut() { return true; }
521 bool shouldTimeOut() {
522 return std::chrono::steady_clock::now() > deadline();
525 bool doWait(Futex& futex, uint32_t expected, uint32_t waitMask) {
526 auto result = futex.futexWaitUntil(expected, deadline(), waitMask);
527 return result != folly::detail::FutexResult::TIMEDOUT;
531 template <class Clock, class Duration>
532 struct WaitUntilDeadline {
533 std::chrono::time_point<Clock, Duration> absDeadline_;
535 bool canBlock() { return true; }
536 bool canTimeOut() { return true; }
537 bool shouldTimeOut() { return Clock::now() > absDeadline_; }
539 bool doWait(Futex& futex, uint32_t expected, uint32_t waitMask) {
540 auto result = futex.futexWaitUntil(expected, absDeadline_, waitMask);
541 return result != folly::detail::FutexResult::TIMEDOUT;
548 static constexpr uint32_t kIncrHasS = 1 << 10;
549 static constexpr uint32_t kHasS = ~(kIncrHasS - 1);
551 // If false, then there are definitely no deferred read locks for this
552 // instance. Cleared after initialization and when exclusively locked.
553 static constexpr uint32_t kMayDefer = 1 << 9;
555 // lock() cleared kMayDefer as soon as it starts draining readers (so
556 // that it doesn't have to do a second CAS once drain completes), but
557 // unlock_shared() still needs to know whether to scan deferredReaders[]
558 // or not. We copy kMayDefer to kPrevDefer when setting kHasE or
559 // kBegunE, and clear it when clearing those bits.
560 static constexpr uint32_t kPrevDefer = 1 << 8;
562 // Exclusive-locked blocks all read locks and write locks. This bit
563 // may be set before all readers have finished, but in that case the
564 // thread that sets it won't return to the caller until all read locks
565 // have been released.
566 static constexpr uint32_t kHasE = 1 << 7;
568 // Exclusive-draining means that lock() is waiting for existing readers
569 // to leave, but that new readers may still acquire shared access.
570 // This is only used in reader priority mode. New readers during
571 // drain must be inline. The difference between this and kHasU is that
572 // kBegunE prevents kMayDefer from being set.
573 static constexpr uint32_t kBegunE = 1 << 6;
575 // At most one thread may have either exclusive or upgrade lock
576 // ownership. Unlike exclusive mode, ownership of the lock in upgrade
577 // mode doesn't preclude other threads holding the lock in shared mode.
578 // boost's concept for this doesn't explicitly say whether new shared
579 // locks can be acquired one lock_upgrade has succeeded, but doesn't
580 // list that as disallowed. RWSpinLock disallows new read locks after
581 // lock_upgrade has been acquired, but the boost implementation doesn't.
582 // We choose the latter.
583 static constexpr uint32_t kHasU = 1 << 5;
585 // There are three states that we consider to be "solo", in that they
586 // cannot coexist with other solo states. These are kHasE, kBegunE,
587 // and kHasU. Note that S doesn't conflict with any of these, because
588 // setting the kHasE is only one of the two steps needed to actually
589 // acquire the lock in exclusive mode (the other is draining the existing
591 static constexpr uint32_t kHasSolo = kHasE | kBegunE | kHasU;
593 // Once a thread sets kHasE it needs to wait for the current readers
594 // to exit the lock. We give this a separate wait identity from the
595 // waiting to set kHasE so that we can perform partial wakeups (wake
596 // one instead of wake all).
597 static constexpr uint32_t kWaitingNotS = 1 << 4;
599 // When waking writers we can either wake them all, in which case we
600 // can clear kWaitingE, or we can call futexWake(1). futexWake tells
601 // us if anybody woke up, but even if we detect that nobody woke up we
602 // can't clear the bit after the fact without issuing another wakeup.
603 // To avoid thundering herds when there are lots of pending lock()
604 // without needing to call futexWake twice when there is only one
605 // waiter, kWaitingE actually encodes if we have observed multiple
606 // concurrent waiters. Tricky: ABA issues on futexWait mean that when
607 // we see kWaitingESingle we can't assume that there is only one.
608 static constexpr uint32_t kWaitingESingle = 1 << 2;
609 static constexpr uint32_t kWaitingEMultiple = 1 << 3;
610 static constexpr uint32_t kWaitingE = kWaitingESingle | kWaitingEMultiple;
612 // kWaitingU is essentially a 1 bit saturating counter. It always
613 // requires a wakeAll.
614 static constexpr uint32_t kWaitingU = 1 << 1;
616 // All blocked lock_shared() should be awoken, so it is correct (not
617 // suboptimal) to wakeAll if there are any shared readers.
618 static constexpr uint32_t kWaitingS = 1 << 0;
620 // kWaitingAny is a mask of all of the bits that record the state of
621 // threads, rather than the state of the lock. It is convenient to be
622 // able to mask them off during asserts.
623 static constexpr uint32_t kWaitingAny =
624 kWaitingNotS | kWaitingE | kWaitingU | kWaitingS;
626 // The reader count at which a reader will attempt to use the lock
627 // in deferred mode. If this value is 2, then the second concurrent
628 // reader will set kMayDefer and use deferredReaders[]. kMayDefer is
629 // cleared during exclusive access, so this threshold must be reached
630 // each time a lock is held in exclusive mode.
631 static constexpr uint32_t kNumSharedToStartDeferring = 2;
633 // The typical number of spins that a thread will wait for a state
634 // transition. There is no bound on the number of threads that can wait
635 // for a writer, so we are pretty conservative here to limit the chance
636 // that we are starving the writer of CPU. Each spin is 6 or 7 nanos,
637 // almost all of which is in the pause instruction.
638 static constexpr uint32_t kMaxSpinCount = !BlockImmediately ? 1000 : 2;
640 // The maximum number of soft yields before falling back to futex.
641 // If the preemption heuristic is activated we will fall back before
642 // this. A soft yield takes ~900 nanos (two sched_yield plus a call
643 // to getrusage, with checks of the goal at each step). Soft yields
644 // aren't compatible with deterministic execution under test (unlike
645 // futexWaitUntil, which has a capricious but deterministic back end).
646 static constexpr uint32_t kMaxSoftYieldCount = !BlockImmediately ? 1000 : 0;
648 // If AccessSpreader assigns indexes from 0..k*n-1 on a system where some
649 // level of the memory hierarchy is symmetrically divided into k pieces
650 // (NUMA nodes, last-level caches, L1 caches, ...), then slot indexes
651 // that are the same after integer division by k share that resource.
652 // Our strategy for deferred readers is to probe up to numSlots/4 slots,
653 // using the full granularity of AccessSpreader for the start slot
654 // and then search outward. We can use AccessSpreader::current(n)
655 // without managing our own spreader if kMaxDeferredReaders <=
656 // AccessSpreader::kMaxCpus, which is currently 128.
658 // Our 2-socket E5-2660 machines have 8 L1 caches on each chip,
659 // with 64 byte cache lines. That means we need 64*16 bytes of
660 // deferredReaders[] to give each L1 its own playground. On x86_64
661 // each DeferredReaderSlot is 8 bytes, so we need kMaxDeferredReaders
662 // * kDeferredSeparationFactor >= 64 * 16 / 8 == 128. If
663 // kDeferredSearchDistance * kDeferredSeparationFactor <=
664 // 64 / 8 then we will search only within a single cache line, which
665 // guarantees we won't have inter-L1 contention. We give ourselves
666 // a factor of 2 on the core count, which should hold us for a couple
667 // processor generations. deferredReaders[] is 2048 bytes currently.
668 static constexpr uint32_t kMaxDeferredReaders = 64;
669 static constexpr uint32_t kDeferredSearchDistance = 2;
670 static constexpr uint32_t kDeferredSeparationFactor = 4;
672 static_assert(!(kMaxDeferredReaders & (kMaxDeferredReaders - 1)),
673 "kMaxDeferredReaders must be a power of 2");
674 static_assert(!(kDeferredSearchDistance & (kDeferredSearchDistance - 1)),
675 "kDeferredSearchDistance must be a power of 2");
677 // The number of deferred locks that can be simultaneously acquired
678 // by a thread via the token-less methods without performing any heap
679 // allocations. Each of these costs 3 pointers (24 bytes, probably)
680 // per thread. There's not much point in making this larger than
681 // kDeferredSearchDistance.
682 static constexpr uint32_t kTokenStackTLSCapacity = 2;
684 // We need to make sure that if there is a lock_shared()
685 // and lock_shared(token) followed by unlock_shared() and
686 // unlock_shared(token), the token-less unlock doesn't null
687 // out deferredReaders[token.slot_]. If we allowed that, then
688 // unlock_shared(token) wouldn't be able to assume that its lock
689 // had been inlined by applyDeferredReaders when it finds that
690 // deferredReaders[token.slot_] no longer points to this. We accomplish
691 // this by stealing bit 0 from the pointer to record that the slot's
692 // element has no token, hence our use of uintptr_t in deferredReaders[].
693 static constexpr uintptr_t kTokenless = 0x1;
695 // This is the starting location for Token-less unlock_shared().
696 static FOLLY_TLS uint32_t tls_lastTokenlessSlot;
698 // Only indexes divisible by kDeferredSeparationFactor are used.
699 // If any of those elements points to a SharedMutexImpl, then it
700 // should be considered that there is a shared lock on that instance.
702 typedef Atom<uintptr_t> DeferredReaderSlot;
703 static DeferredReaderSlot deferredReaders
704 [kMaxDeferredReaders *
705 kDeferredSeparationFactor] FOLLY_ALIGN_TO_AVOID_FALSE_SHARING;
707 // Performs an exclusive lock, waiting for state_ & waitMask to be
709 template <class WaitContext>
710 bool lockExclusiveImpl(uint32_t preconditionGoalMask, WaitContext& ctx) {
711 uint32_t state = state_.load(std::memory_order_acquire);
713 (state & (preconditionGoalMask | kMayDefer | kHasS)) == 0 &&
714 state_.compare_exchange_strong(state, (state | kHasE) & ~kHasU))) {
717 return lockExclusiveImpl(state, preconditionGoalMask, ctx);
721 template <class WaitContext>
722 bool lockExclusiveImpl(uint32_t& state,
723 uint32_t preconditionGoalMask,
726 if (UNLIKELY((state & preconditionGoalMask) != 0) &&
727 !waitForZeroBits(state, preconditionGoalMask, kWaitingE, ctx) &&
732 uint32_t after = (state & kMayDefer) == 0 ? 0 : kPrevDefer;
733 if (!ReaderPriority || (state & (kMayDefer | kHasS)) == 0) {
734 // Block readers immediately, either because we are in write
735 // priority mode or because we can acquire the lock in one
736 // step. Note that if state has kHasU, then we are doing an
737 // unlock_upgrade_and_lock() and we should clear it (reader
738 // priority branch also does this).
739 after |= (state | kHasE) & ~(kHasU | kMayDefer);
741 after |= (state | kBegunE) & ~(kHasU | kMayDefer);
743 if (state_.compare_exchange_strong(state, after)) {
747 // If we set kHasE (writer priority) then no new readers can
748 // arrive. If we set kBegunE then they can still enter, but
749 // they must be inline. Either way we need to either spin on
750 // deferredReaders[] slots, or inline them so that we can wait on
751 // kHasS to zero itself. deferredReaders[] is pointers, which on
752 // x86_64 are bigger than futex() can handle, so we inline the
753 // deferred locks instead of trying to futexWait on each slot.
754 // Readers are responsible for rechecking state_ after recording
755 // a deferred read to avoid atomicity problems between the state_
756 // CAS and applyDeferredReader's reads of deferredReaders[].
757 if (UNLIKELY((before & kMayDefer) != 0)) {
758 applyDeferredReaders(state, ctx);
761 assert((state & (kHasE | kBegunE)) != 0 && (state & kHasU) == 0);
762 if (UNLIKELY((state & kHasS) != 0) &&
763 !waitForZeroBits(state, kHasS, kWaitingNotS, ctx) &&
765 // Ugh. We blocked new readers and other writers for a while,
766 // but were unable to complete. Move on. On the plus side
767 // we can clear kWaitingNotS because nobody else can piggyback
769 state = (state_ &= ~(kPrevDefer | kHasE | kBegunE | kWaitingNotS));
770 wakeRegisteredWaiters(state, kWaitingE | kWaitingU | kWaitingS);
774 if (ReaderPriority && (state & kHasE) == 0) {
775 assert((state & kBegunE) != 0);
776 if (!state_.compare_exchange_strong(state,
777 (state & ~kBegunE) | kHasE)) {
788 template <class WaitContext>
789 bool waitForZeroBits(uint32_t& state,
793 uint32_t spinCount = 0;
795 state = state_.load(std::memory_order_acquire);
796 if ((state & goal) == 0) {
799 asm_volatile_pause();
801 if (UNLIKELY(spinCount >= kMaxSpinCount)) {
802 return ctx.canBlock() &&
803 yieldWaitForZeroBits(state, goal, waitMask, ctx);
808 template <class WaitContext>
809 bool yieldWaitForZeroBits(uint32_t& state,
817 for (uint32_t yieldCount = 0; yieldCount < kMaxSoftYieldCount;
819 for (int softState = 0; softState < 3; ++softState) {
821 std::this_thread::yield();
824 getrusage(RUSAGE_THREAD, &usage);
827 if (((state = state_.load(std::memory_order_acquire)) & goal) == 0) {
830 if (ctx.shouldTimeOut()) {
835 if (before >= 0 && usage.ru_nivcsw >= before + 2) {
836 // One involuntary csw might just be occasional background work,
837 // but if we get two in a row then we guess that there is someone
838 // else who can profitably use this CPU. Fall back to futex
841 before = usage.ru_nivcsw;
844 return futexWaitForZeroBits(state, goal, waitMask, ctx);
847 template <class WaitContext>
848 bool futexWaitForZeroBits(uint32_t& state,
852 assert(waitMask == kWaitingNotS || waitMask == kWaitingE ||
853 waitMask == kWaitingU || waitMask == kWaitingS);
856 state = state_.load(std::memory_order_acquire);
857 if ((state & goal) == 0) {
862 if (waitMask == kWaitingE) {
863 if ((state & kWaitingESingle) != 0) {
864 after |= kWaitingEMultiple;
866 after |= kWaitingESingle;
872 // CAS is better than atomic |= here, because it lets us avoid
873 // setting the wait flag when the goal is concurrently achieved
874 if (after != state && !state_.compare_exchange_strong(state, after)) {
878 if (!ctx.doWait(state_, after, waitMask)) {
885 // Wakes up waiters registered in state_ as appropriate, clearing the
886 // awaiting bits for anybody that was awoken. Tries to perform direct
887 // single wakeup of an exclusive waiter if appropriate
888 void wakeRegisteredWaiters(uint32_t& state, uint32_t wakeMask) {
889 if (UNLIKELY((state & wakeMask) != 0)) {
890 wakeRegisteredWaitersImpl(state, wakeMask);
894 void wakeRegisteredWaitersImpl(uint32_t& state, uint32_t wakeMask) {
895 // If there are multiple lock() pending only one of them will actually
896 // get to wake up, so issuing futexWakeAll will make a thundering herd.
897 // There's nothing stopping us from issuing futexWake(1) instead,
898 // so long as the wait bits are still an accurate reflection of
899 // the waiters. If we notice (via futexWake's return value) that
900 // nobody woke up then we can try again with the normal wake-all path.
901 // Note that we can't just clear the bits at that point; we need to
902 // clear the bits and then issue another wakeup.
904 // It is possible that we wake an E waiter but an outside S grabs the
905 // lock instead, at which point we should wake pending U and S waiters.
906 // Rather than tracking state to make the failing E regenerate the
907 // wakeup, we just disable the optimization in the case that there
908 // are waiting U or S that we are eligible to wake.
909 if ((wakeMask & kWaitingE) == kWaitingE &&
910 (state & wakeMask) == kWaitingE &&
911 state_.futexWake(1, kWaitingE) > 0) {
912 // somebody woke up, so leave state_ as is and clear it later
916 if ((state & wakeMask) != 0) {
917 auto prev = state_.fetch_and(~wakeMask);
918 if ((prev & wakeMask) != 0) {
919 futexWakeAll(wakeMask);
921 state = prev & ~wakeMask;
925 void futexWakeAll(uint32_t wakeMask) {
926 state_.futexWake(std::numeric_limits<int>::max(), wakeMask);
929 DeferredReaderSlot* deferredReader(uint32_t slot) {
930 return &deferredReaders[slot * kDeferredSeparationFactor];
933 uintptr_t tokenfulSlotValue() { return reinterpret_cast<uintptr_t>(this); }
935 uintptr_t tokenlessSlotValue() { return tokenfulSlotValue() | kTokenless; }
937 bool slotValueIsThis(uintptr_t slotValue) {
938 return (slotValue & ~kTokenless) == tokenfulSlotValue();
941 // Clears any deferredReaders[] that point to this, adjusting the inline
942 // shared lock count to compensate. Does some spinning and yielding
943 // to avoid the work. Always finishes the application, even if ctx
945 template <class WaitContext>
946 void applyDeferredReaders(uint32_t& state, WaitContext& ctx) {
949 uint32_t spinCount = 0;
951 while (!slotValueIsThis(
952 deferredReader(slot)->load(std::memory_order_acquire))) {
953 if (++slot == kMaxDeferredReaders) {
958 if (UNLIKELY(++spinCount >= kMaxSpinCount)) {
959 applyDeferredReaders(state, ctx, slot);
965 template <class WaitContext>
966 void applyDeferredReaders(uint32_t& state, WaitContext& ctx, uint32_t slot) {
972 for (uint32_t yieldCount = 0; yieldCount < kMaxSoftYieldCount;
974 for (int softState = 0; softState < 3; ++softState) {
976 std::this_thread::yield();
979 getrusage(RUSAGE_THREAD, &usage);
982 while (!slotValueIsThis(
983 deferredReader(slot)->load(std::memory_order_acquire))) {
984 if (++slot == kMaxDeferredReaders) {
988 if (ctx.shouldTimeOut()) {
989 // finish applying immediately on timeout
994 if (before >= 0 && usage.ru_nivcsw >= before + 2) {
995 // heuristic says run queue is not empty
998 before = usage.ru_nivcsw;
1002 uint32_t movedSlotCount = 0;
1003 for (; slot < kMaxDeferredReaders; ++slot) {
1004 auto slotPtr = deferredReader(slot);
1005 auto slotValue = slotPtr->load(std::memory_order_acquire);
1006 if (slotValueIsThis(slotValue) &&
1007 slotPtr->compare_exchange_strong(slotValue, 0)) {
1012 if (movedSlotCount > 0) {
1013 state = (state_ += movedSlotCount * kIncrHasS);
1015 assert((state & (kHasE | kBegunE)) != 0);
1017 // if state + kIncrHasS overflows (off the end of state) then either
1018 // we have 2^(32-9) readers (almost certainly an application bug)
1019 // or we had an underflow (also a bug)
1020 assert(state < state + kIncrHasS);
1023 // It is straightfoward to make a token-less lock_shared() and
1024 // unlock_shared() either by making the token-less version always use
1025 // INLINE_SHARED mode or by removing the token version. Supporting
1026 // deferred operation for both types is trickier than it appears, because
1027 // the purpose of the token it so that unlock_shared doesn't have to
1028 // look in other slots for its deferred lock. Token-less unlock_shared
1029 // might place a deferred lock in one place and then release a different
1030 // slot that was originally used by the token-ful version. If this was
1031 // important we could solve the problem by differentiating the deferred
1032 // locks so that cross-variety release wouldn't occur. The best way
1033 // is probably to steal a bit from the pointer, making deferredLocks[]
1034 // an array of Atom<uintptr_t>.
1036 template <class WaitContext>
1037 bool lockSharedImpl(Token* token, WaitContext& ctx) {
1038 uint32_t state = state_.load(std::memory_order_relaxed);
1039 if ((state & (kHasS | kMayDefer | kHasE)) == 0 &&
1040 state_.compare_exchange_strong(state, state + kIncrHasS)) {
1041 if (token != nullptr) {
1042 token->type_ = Token::Type::INLINE_SHARED;
1046 return lockSharedImpl(state, token, ctx);
1049 template <class WaitContext>
1050 bool lockSharedImpl(uint32_t& state, Token* token, WaitContext& ctx) {
1052 if (UNLIKELY((state & kHasE) != 0) &&
1053 !waitForZeroBits(state, kHasE, kWaitingS, ctx) && ctx.canTimeOut()) {
1058 uintptr_t slotValue = 1; // any non-zero value will do
1060 bool canAlreadyDefer = (state & kMayDefer) != 0;
1061 bool aboveDeferThreshold =
1062 (state & kHasS) >= (kNumSharedToStartDeferring - 1) * kIncrHasS;
1063 bool drainInProgress = ReaderPriority && (state & kBegunE) != 0;
1064 if (canAlreadyDefer || (aboveDeferThreshold && !drainInProgress)) {
1065 // starting point for our empty-slot search, can change after
1066 // calling waitForZeroBits
1068 (uint32_t)folly::detail::AccessSpreader<Atom>::current(
1069 kMaxDeferredReaders);
1071 // deferred readers are already enabled, or it is time to
1072 // enable them if we can find a slot
1073 for (uint32_t i = 0; i < kDeferredSearchDistance; ++i) {
1074 slot = bestSlot ^ i;
1075 assert(slot < kMaxDeferredReaders);
1076 slotValue = deferredReader(slot)->load(std::memory_order_relaxed);
1077 if (slotValue == 0) {
1084 if (slotValue != 0) {
1085 // not yet deferred, or no empty slots
1086 if (state_.compare_exchange_strong(state, state + kIncrHasS)) {
1087 // successfully recorded the read lock inline
1088 if (token != nullptr) {
1089 token->type_ = Token::Type::INLINE_SHARED;
1093 // state is updated, try again
1097 // record that deferred readers might be in use if necessary
1098 if ((state & kMayDefer) == 0) {
1099 if (!state_.compare_exchange_strong(state, state | kMayDefer)) {
1100 // keep going if CAS failed because somebody else set the bit
1102 if ((state & (kHasE | kMayDefer)) != kMayDefer) {
1106 // state = state | kMayDefer;
1109 // try to use the slot
1110 bool gotSlot = deferredReader(slot)->compare_exchange_strong(
1112 token == nullptr ? tokenlessSlotValue() : tokenfulSlotValue());
1114 // If we got the slot, we need to verify that an exclusive lock
1115 // didn't happen since we last checked. If we didn't get the slot we
1116 // need to recheck state_ anyway to make sure we don't waste too much
1117 // work. It is also possible that since we checked state_ someone
1118 // has acquired and released the write lock, clearing kMayDefer.
1119 // Both cases are covered by looking for the readers-possible bit,
1120 // because it is off when the exclusive lock bit is set.
1121 state = state_.load(std::memory_order_acquire);
1127 if (token == nullptr) {
1128 tls_lastTokenlessSlot = slot;
1131 if ((state & kMayDefer) != 0) {
1132 assert((state & kHasE) == 0);
1134 if (token != nullptr) {
1135 token->type_ = Token::Type::DEFERRED_SHARED;
1136 token->slot_ = (uint16_t)slot;
1141 // release the slot before retrying
1142 if (token == nullptr) {
1143 // We can't rely on slot. Token-less slot values can be freed by
1144 // any unlock_shared(), so we need to do the full deferredReader
1145 // search during unlock. Unlike unlock_shared(), we can't trust
1146 // kPrevDefer here. This deferred lock isn't visible to lock()
1147 // (that's the whole reason we're undoing it) so there might have
1148 // subsequently been an unlock() and lock() with no intervening
1149 // transition to deferred mode.
1150 if (!tryUnlockAnySharedDeferred()) {
1151 unlockSharedInline();
1154 if (!tryUnlockSharedDeferred(slot)) {
1155 unlockSharedInline();
1159 // We got here not because the lock was unavailable, but because
1160 // we lost a compare-and-swap. Try-lock is typically allowed to
1161 // have spurious failures, but there is no lock efficiency gain
1162 // from exploiting that freedom here.
1166 bool tryUnlockAnySharedDeferred() {
1167 auto bestSlot = tls_lastTokenlessSlot;
1168 for (uint32_t i = 0; i < kMaxDeferredReaders; ++i) {
1169 auto slotPtr = deferredReader(bestSlot ^ i);
1170 auto slotValue = slotPtr->load(std::memory_order_relaxed);
1171 if (slotValue == tokenlessSlotValue() &&
1172 slotPtr->compare_exchange_strong(slotValue, 0)) {
1173 tls_lastTokenlessSlot = bestSlot ^ i;
1180 bool tryUnlockSharedDeferred(uint32_t slot) {
1181 assert(slot < kMaxDeferredReaders);
1182 auto slotValue = tokenfulSlotValue();
1183 return deferredReader(slot)->compare_exchange_strong(slotValue, 0);
1186 uint32_t unlockSharedInline() {
1187 uint32_t state = (state_ -= kIncrHasS);
1188 assert((state & (kHasE | kBegunE)) != 0 || state < state + kIncrHasS);
1189 if ((state & kHasS) == 0) {
1190 // Only the second half of lock() can be blocked by a non-zero
1191 // reader count, so that's the only thing we need to wake
1192 wakeRegisteredWaiters(state, kWaitingNotS);
1197 template <class WaitContext>
1198 bool lockUpgradeImpl(WaitContext& ctx) {
1201 if (!waitForZeroBits(state, kHasSolo, kWaitingU, ctx)) {
1204 } while (!state_.compare_exchange_strong(state, state | kHasU));
1211 ReadHolder() : lock_(nullptr) {}
1213 explicit ReadHolder(const SharedMutexImpl* lock) : ReadHolder(*lock) {}
1215 explicit ReadHolder(const SharedMutexImpl& lock)
1216 : lock_(const_cast<SharedMutexImpl*>(&lock)) {
1217 lock_->lock_shared(token_);
1220 ReadHolder(ReadHolder&& rhs) noexcept : lock_(rhs.lock_),
1221 token_(rhs.token_) {
1222 rhs.lock_ = nullptr;
1225 // Downgrade from upgrade mode
1226 explicit ReadHolder(UpgradeHolder&& upgraded) : lock_(upgraded.lock_) {
1227 assert(upgraded.lock_ != nullptr);
1228 upgraded.lock_ = nullptr;
1229 lock_->unlock_upgrade_and_lock_shared(token_);
1232 // Downgrade from exclusive mode
1233 explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
1234 assert(writer.lock_ != nullptr);
1235 writer.lock_ = nullptr;
1236 lock_->unlock_and_lock_shared(token_);
1239 ReadHolder& operator=(ReadHolder&& rhs) noexcept {
1240 std::swap(lock_, rhs.lock_);
1241 std::swap(token_, rhs.token_);
1245 ReadHolder(const ReadHolder& rhs) = delete;
1246 ReadHolder& operator=(const ReadHolder& rhs) = delete;
1250 lock_->unlock_shared(token_);
1255 friend class UpgradeHolder;
1256 friend class WriteHolder;
1257 SharedMutexImpl* lock_;
1258 SharedMutexToken token_;
1261 class UpgradeHolder {
1263 UpgradeHolder() : lock_(nullptr) {}
1265 explicit UpgradeHolder(SharedMutexImpl* lock) : UpgradeHolder(*lock) {}
1267 explicit UpgradeHolder(SharedMutexImpl& lock) : lock_(&lock) {
1268 lock_->lock_upgrade();
1271 // Downgrade from exclusive mode
1272 explicit UpgradeHolder(WriteHolder&& writer) : lock_(writer.lock_) {
1273 assert(writer.lock_ != nullptr);
1274 writer.lock_ = nullptr;
1275 lock_->unlock_and_lock_upgrade();
1278 UpgradeHolder(UpgradeHolder&& rhs) noexcept : lock_(rhs.lock_) {
1279 rhs.lock_ = nullptr;
1282 UpgradeHolder& operator=(UpgradeHolder&& rhs) noexcept {
1283 std::swap(lock_, rhs.lock_);
1287 UpgradeHolder(const UpgradeHolder& rhs) = delete;
1288 UpgradeHolder& operator=(const UpgradeHolder& rhs) = delete;
1292 lock_->unlock_upgrade();
1297 friend class WriteHolder;
1298 friend class ReadHolder;
1299 SharedMutexImpl* lock_;
1304 WriteHolder() : lock_(nullptr) {}
1306 explicit WriteHolder(SharedMutexImpl* lock) : WriteHolder(*lock) {}
1308 explicit WriteHolder(SharedMutexImpl& lock) : lock_(&lock) {
1312 // Promotion from upgrade mode
1313 explicit WriteHolder(UpgradeHolder&& upgrade) : lock_(upgrade.lock_) {
1314 assert(upgrade.lock_ != nullptr);
1315 upgrade.lock_ = nullptr;
1316 lock_->unlock_upgrade_and_lock();
1319 WriteHolder(WriteHolder&& rhs) noexcept : lock_(rhs.lock_) {
1320 rhs.lock_ = nullptr;
1323 WriteHolder& operator=(WriteHolder&& rhs) noexcept {
1324 std::swap(lock_, rhs.lock_);
1328 WriteHolder(const WriteHolder& rhs) = delete;
1329 WriteHolder& operator=(const WriteHolder& rhs) = delete;
1338 friend class ReadHolder;
1339 friend class UpgradeHolder;
1340 SharedMutexImpl* lock_;
1343 // Adapters for Synchronized<>
1344 friend void acquireRead(SharedMutexImpl& lock) { lock.lock_shared(); }
1345 friend void acquireReadWrite(SharedMutexImpl& lock) { lock.lock(); }
1346 friend void releaseRead(SharedMutexImpl& lock) { lock.unlock_shared(); }
1347 friend void releaseReadWrite(SharedMutexImpl& lock) { lock.unlock(); }
1350 #define COMMON_CONCURRENCY_SHARED_MUTEX_DECLARE_STATIC_STORAGE(type) \
1352 type::DeferredReaderSlot \
1353 type::deferredReaders[type::kMaxDeferredReaders * \
1354 type::kDeferredSeparationFactor] = {}; \
1356 FOLLY_TLS uint32_t type::tls_lastTokenlessSlot = 0;
1358 typedef SharedMutexImpl<true> SharedMutexReadPriority;
1359 typedef SharedMutexImpl<false> SharedMutexWritePriority;
1360 typedef SharedMutexWritePriority SharedMutex;
1362 } // namespace folly