2 * Copyright 2011-present 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 * N.B. You most likely do _not_ want to use RWSpinLock or any other
18 * kind of spinlock. Use SharedMutex instead.
20 * In short, spinlocks in preemptive multi-tasking operating systems
21 * have serious problems and fast mutexes like SharedMutex are almost
22 * certainly the better choice, because letting the OS scheduler put a
23 * thread to sleep is better for system responsiveness and throughput
24 * than wasting a timeslice repeatedly querying a lock held by a
25 * thread that's blocked, and you can't prevent userspace
28 * Spinlocks in an operating system kernel make much more sense than
29 * they do in userspace.
31 * -------------------------------------------------------------------
33 * Two Read-Write spin lock implementations.
35 * Ref: http://locklessinc.com/articles/locks
37 * Both locks here are faster than pthread_rwlock and have very low
38 * overhead (usually 20-30ns). They don't use any system mutexes and
39 * are very compact (4/8 bytes), so are suitable for per-instance
40 * based locking, particularly when contention is not expected.
42 * For a spinlock, RWSpinLock is a reasonable choice. (See the note
43 * about for why a spin lock is frequently a bad idea generally.)
44 * RWSpinLock has minimal overhead, and comparable contention
45 * performance when the number of competing threads is less than or
46 * equal to the number of logical CPUs. Even as the number of
47 * threads gets larger, RWSpinLock can still be very competitive in
48 * READ, although it is slower on WRITE, and also inherently unfair
51 * RWTicketSpinLock shows more balanced READ/WRITE performance. If
52 * your application really needs a lot more threads, and a
53 * higher-priority writer, prefer one of the RWTicketSpinLock locks.
57 * RWTicketSpinLock locks can only be used with GCC on x86/x86-64
60 * RWTicketSpinLock<32> only allows up to 2^8 - 1 concurrent
61 * readers and writers.
63 * RWTicketSpinLock<64> only allows up to 2^16 - 1 concurrent
64 * readers and writers.
66 * RWTicketSpinLock<..., true> (kFavorWriter = true, that is, strict
67 * writer priority) is NOT reentrant, even for lock_shared().
69 * The lock will not grant any new shared (read) accesses while a thread
70 * attempting to acquire the lock in write mode is blocked. (That is,
71 * if the lock is held in shared mode by N threads, and a thread attempts
72 * to acquire it in write mode, no one else can acquire it in shared mode
73 * until these N threads release the lock and then the blocked thread
74 * acquires and releases the exclusive lock.) This also applies for
75 * attempts to reacquire the lock in shared mode by threads that already
76 * hold it in shared mode, making the lock non-reentrant.
78 * RWSpinLock handles 2^30 - 1 concurrent readers.
80 * @author Xin Liu <xliux@fb.com>
86 ========================================================================
87 Benchmark on (Intel(R) Xeon(R) CPU L5630 @ 2.13GHz) 8 cores(16 HTs)
88 ========================================================================
90 ------------------------------------------------------------------------------
91 1. Single thread benchmark (read/write lock + unlock overhead)
92 Benchmark Iters Total t t/iter iter/sec
93 -------------------------------------------------------------------------------
94 * BM_RWSpinLockRead 100000 1.786 ms 17.86 ns 53.4M
95 +30.5% BM_RWSpinLockWrite 100000 2.331 ms 23.31 ns 40.91M
96 +85.7% BM_RWTicketSpinLock32Read 100000 3.317 ms 33.17 ns 28.75M
97 +96.0% BM_RWTicketSpinLock32Write 100000 3.5 ms 35 ns 27.25M
98 +85.6% BM_RWTicketSpinLock64Read 100000 3.315 ms 33.15 ns 28.77M
99 +96.0% BM_RWTicketSpinLock64Write 100000 3.5 ms 35 ns 27.25M
100 +85.7% BM_RWTicketSpinLock32FavorWriterRead 100000 3.317 ms 33.17 ns 28.75M
101 +29.7% BM_RWTicketSpinLock32FavorWriterWrite 100000 2.316 ms 23.16 ns 41.18M
102 +85.3% BM_RWTicketSpinLock64FavorWriterRead 100000 3.309 ms 33.09 ns 28.82M
103 +30.2% BM_RWTicketSpinLock64FavorWriterWrite 100000 2.325 ms 23.25 ns 41.02M
104 + 175% BM_PThreadRWMutexRead 100000 4.917 ms 49.17 ns 19.4M
105 + 166% BM_PThreadRWMutexWrite 100000 4.757 ms 47.57 ns 20.05M
107 ------------------------------------------------------------------------------
108 2. Contention Benchmark 90% read 10% write
109 Benchmark hits average min max sigma
110 ------------------------------------------------------------------------------
111 ---------- 8 threads ------------
112 RWSpinLock Write 142666 220ns 78ns 40.8us 269ns
113 RWSpinLock Read 1282297 222ns 80ns 37.7us 248ns
114 RWTicketSpinLock Write 85692 209ns 71ns 17.9us 252ns
115 RWTicketSpinLock Read 769571 215ns 78ns 33.4us 251ns
116 pthread_rwlock_t Write 84248 2.48us 99ns 269us 8.19us
117 pthread_rwlock_t Read 761646 933ns 101ns 374us 3.25us
119 ---------- 16 threads ------------
120 RWSpinLock Write 124236 237ns 78ns 261us 801ns
121 RWSpinLock Read 1115807 236ns 78ns 2.27ms 2.17us
122 RWTicketSpinLock Write 81781 231ns 71ns 31.4us 351ns
123 RWTicketSpinLock Read 734518 238ns 78ns 73.6us 379ns
124 pthread_rwlock_t Write 83363 7.12us 99ns 785us 28.1us
125 pthread_rwlock_t Read 754978 2.18us 101ns 1.02ms 14.3us
127 ---------- 50 threads ------------
128 RWSpinLock Write 131142 1.37us 82ns 7.53ms 68.2us
129 RWSpinLock Read 1181240 262ns 78ns 6.62ms 12.7us
130 RWTicketSpinLock Write 83045 397ns 73ns 7.01ms 31.5us
131 RWTicketSpinLock Read 744133 386ns 78ns 11ms 31.4us
132 pthread_rwlock_t Write 80849 112us 103ns 4.52ms 263us
133 pthread_rwlock_t Read 728698 24us 101ns 7.28ms 194us
137 #include <folly/Portability.h>
138 #include <folly/portability/Asm.h>
140 #if defined(__GNUC__) && (defined(__i386) || FOLLY_X64 || defined(ARCH_K8))
141 #define RW_SPINLOCK_USE_X86_INTRINSIC_
142 #include <x86intrin.h>
143 #elif defined(_MSC_VER) && defined(FOLLY_X64)
144 #define RW_SPINLOCK_USE_X86_INTRINSIC_
146 #define RW_SPINLOCK_USE_X86_INTRINSIC_
148 #undef RW_SPINLOCK_USE_X86_INTRINSIC_
151 // iOS doesn't define _mm_cvtsi64_si128 and friends
152 #if (FOLLY_SSE >= 2) && !FOLLY_MOBILE && FOLLY_X64
153 #define RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
155 #undef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
162 #include <folly/Likely.h>
167 * A simple, small (4-bytes), but unfair rwlock. Use it when you want
168 * a nice writer and don't expect a lot of write/read contention, or
169 * when you need small rwlocks since you are creating a large number
172 * Note that the unfairness here is extreme: if the lock is
173 * continually accessed for read, writers will never get a chance. If
174 * the lock can be that highly contended this class is probably not an
175 * ideal choice anyway.
177 * It currently implements most of the Lockable, SharedLockable and
178 * UpgradeLockable concepts except the TimedLockable related locking/unlocking
182 enum : int32_t { READER = 4, UPGRADED = 2, WRITER = 1 };
184 constexpr RWSpinLock() : bits_(0) {}
186 RWSpinLock(RWSpinLock const&) = delete;
187 RWSpinLock& operator=(RWSpinLock const&) = delete;
191 uint_fast32_t count = 0;
192 while (!LIKELY(try_lock())) {
193 if (++count > 1000) {
194 std::this_thread::yield();
199 // Writer is responsible for clearing up both the UPGRADED and WRITER bits.
201 static_assert(READER > WRITER + UPGRADED, "wrong bits!");
202 bits_.fetch_and(~(WRITER | UPGRADED), std::memory_order_release);
205 // SharedLockable Concept
207 uint_fast32_t count = 0;
208 while (!LIKELY(try_lock_shared())) {
209 if (++count > 1000) {
210 std::this_thread::yield();
215 void unlock_shared() {
216 bits_.fetch_add(-READER, std::memory_order_release);
219 // Downgrade the lock from writer status to reader status.
220 void unlock_and_lock_shared() {
221 bits_.fetch_add(READER, std::memory_order_acquire);
225 // UpgradeLockable Concept
226 void lock_upgrade() {
227 uint_fast32_t count = 0;
228 while (!try_lock_upgrade()) {
229 if (++count > 1000) {
230 std::this_thread::yield();
235 void unlock_upgrade() {
236 bits_.fetch_add(-UPGRADED, std::memory_order_acq_rel);
239 // unlock upgrade and try to acquire write lock
240 void unlock_upgrade_and_lock() {
242 while (!try_unlock_upgrade_and_lock()) {
243 if (++count > 1000) {
244 std::this_thread::yield();
249 // unlock upgrade and read lock atomically
250 void unlock_upgrade_and_lock_shared() {
251 bits_.fetch_add(READER - UPGRADED, std::memory_order_acq_rel);
254 // write unlock and upgrade lock atomically
255 void unlock_and_lock_upgrade() {
256 // need to do it in two steps here -- as the UPGRADED bit might be OR-ed at
257 // the same time when other threads are trying do try_lock_upgrade().
258 bits_.fetch_or(UPGRADED, std::memory_order_acquire);
259 bits_.fetch_add(-WRITER, std::memory_order_release);
263 // Attempt to acquire writer permission. Return false if we didn't get it.
266 return bits_.compare_exchange_strong(expect, WRITER,
267 std::memory_order_acq_rel);
270 // Try to get reader permission on the lock. This can fail if we
271 // find out someone is a writer or upgrader.
272 // Setting the UPGRADED bit would allow a writer-to-be to indicate
273 // its intention to write and block any new readers while waiting
274 // for existing readers to finish and release their read locks. This
275 // helps avoid starving writers (promoted from upgraders).
276 bool try_lock_shared() {
277 // fetch_add is considerably (100%) faster than compare_exchange,
278 // so here we are optimizing for the common (lock success) case.
279 int32_t value = bits_.fetch_add(READER, std::memory_order_acquire);
280 if (UNLIKELY(value & (WRITER|UPGRADED))) {
281 bits_.fetch_add(-READER, std::memory_order_release);
287 // try to unlock upgrade and write lock atomically
288 bool try_unlock_upgrade_and_lock() {
289 int32_t expect = UPGRADED;
290 return bits_.compare_exchange_strong(expect, WRITER,
291 std::memory_order_acq_rel);
294 // try to acquire an upgradable lock.
295 bool try_lock_upgrade() {
296 int32_t value = bits_.fetch_or(UPGRADED, std::memory_order_acquire);
298 // Note: when failed, we cannot flip the UPGRADED bit back,
299 // as in this case there is either another upgrade lock or a write lock.
300 // If it's a write lock, the bit will get cleared up when that lock's done
302 return ((value & (UPGRADED | WRITER)) == 0);
305 // mainly for debugging purposes.
306 int32_t bits() const { return bits_.load(std::memory_order_acquire); }
309 class UpgradedHolder;
314 explicit ReadHolder(RWSpinLock* lock) : lock_(lock) {
316 lock_->lock_shared();
320 explicit ReadHolder(RWSpinLock& lock) : lock_(&lock) {
321 lock_->lock_shared();
324 ReadHolder(ReadHolder&& other) noexcept : lock_(other.lock_) {
325 other.lock_ = nullptr;
329 explicit ReadHolder(UpgradedHolder&& upgraded) : lock_(upgraded.lock_) {
330 upgraded.lock_ = nullptr;
332 lock_->unlock_upgrade_and_lock_shared();
336 explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
337 writer.lock_ = nullptr;
339 lock_->unlock_and_lock_shared();
343 ReadHolder& operator=(ReadHolder&& other) {
345 swap(lock_, other.lock_);
349 ReadHolder(const ReadHolder& other) = delete;
350 ReadHolder& operator=(const ReadHolder& other) = delete;
354 lock_->unlock_shared();
358 void reset(RWSpinLock* lock = nullptr) {
363 lock_->unlock_shared();
367 lock_->lock_shared();
371 void swap(ReadHolder* other) {
372 std::swap(lock_, other->lock_);
376 friend class UpgradedHolder;
377 friend class WriteHolder;
381 class UpgradedHolder {
383 explicit UpgradedHolder(RWSpinLock* lock) : lock_(lock) {
385 lock_->lock_upgrade();
389 explicit UpgradedHolder(RWSpinLock& lock) : lock_(&lock) {
390 lock_->lock_upgrade();
393 explicit UpgradedHolder(WriteHolder&& writer) {
394 lock_ = writer.lock_;
395 writer.lock_ = nullptr;
397 lock_->unlock_and_lock_upgrade();
401 UpgradedHolder(UpgradedHolder&& other) noexcept : lock_(other.lock_) {
402 other.lock_ = nullptr;
405 UpgradedHolder& operator =(UpgradedHolder&& other) {
407 swap(lock_, other.lock_);
411 UpgradedHolder(const UpgradedHolder& other) = delete;
412 UpgradedHolder& operator =(const UpgradedHolder& other) = delete;
416 lock_->unlock_upgrade();
420 void reset(RWSpinLock* lock = nullptr) {
425 lock_->unlock_upgrade();
429 lock_->lock_upgrade();
433 void swap(UpgradedHolder* other) {
435 swap(lock_, other->lock_);
439 friend class WriteHolder;
440 friend class ReadHolder;
446 explicit WriteHolder(RWSpinLock* lock) : lock_(lock) {
452 explicit WriteHolder(RWSpinLock& lock) : lock_(&lock) {
456 // promoted from an upgrade lock holder
457 explicit WriteHolder(UpgradedHolder&& upgraded) {
458 lock_ = upgraded.lock_;
459 upgraded.lock_ = nullptr;
461 lock_->unlock_upgrade_and_lock();
465 WriteHolder(WriteHolder&& other) noexcept : lock_(other.lock_) {
466 other.lock_ = nullptr;
469 WriteHolder& operator =(WriteHolder&& other) {
471 swap(lock_, other.lock_);
475 WriteHolder(const WriteHolder& other) = delete;
476 WriteHolder& operator =(const WriteHolder& other) = delete;
484 void reset(RWSpinLock* lock = nullptr) {
497 void swap(WriteHolder* other) {
499 swap(lock_, other->lock_);
503 friend class ReadHolder;
504 friend class UpgradedHolder;
509 std::atomic<int32_t> bits_;
513 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
514 // A more balanced Read-Write spin lock implemented based on GCC intrinsics.
517 template <size_t kBitWidth> struct RWTicketIntTrait {
518 static_assert(kBitWidth == 32 || kBitWidth == 64,
519 "bit width has to be either 32 or 64 ");
523 struct RWTicketIntTrait<64> {
524 typedef uint64_t FullInt;
525 typedef uint32_t HalfInt;
526 typedef uint16_t QuarterInt;
528 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
529 static __m128i make128(const uint16_t v[4]) {
530 return _mm_set_epi16(0, 0, 0, 0,
531 short(v[3]), short(v[2]), short(v[1]), short(v[0]));
533 static inline __m128i fromInteger(uint64_t from) {
534 return _mm_cvtsi64_si128(int64_t(from));
536 static inline uint64_t toInteger(__m128i in) {
537 return uint64_t(_mm_cvtsi128_si64(in));
539 static inline uint64_t addParallel(__m128i in, __m128i kDelta) {
540 return toInteger(_mm_add_epi16(in, kDelta));
546 struct RWTicketIntTrait<32> {
547 typedef uint32_t FullInt;
548 typedef uint16_t HalfInt;
549 typedef uint8_t QuarterInt;
551 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
552 static __m128i make128(const uint8_t v[4]) {
557 char(v[3]), char(v[2]), char(v[1]), char(v[0]));
559 static inline __m128i fromInteger(uint32_t from) {
560 return _mm_cvtsi32_si128(int32_t(from));
562 static inline uint32_t toInteger(__m128i in) {
563 return uint32_t(_mm_cvtsi128_si32(in));
565 static inline uint32_t addParallel(__m128i in, __m128i kDelta) {
566 return toInteger(_mm_add_epi8(in, kDelta));
570 } // namespace detail
572 template <size_t kBitWidth, bool kFavorWriter = false>
573 class RWTicketSpinLockT {
574 typedef detail::RWTicketIntTrait<kBitWidth> IntTraitType;
575 typedef typename detail::RWTicketIntTrait<kBitWidth>::FullInt FullInt;
576 typedef typename detail::RWTicketIntTrait<kBitWidth>::HalfInt HalfInt;
577 typedef typename detail::RWTicketIntTrait<kBitWidth>::QuarterInt
581 constexpr RWTicket() : whole(0) {}
584 __extension__ struct {
591 private: // Some x64-specific utilities for atomic access to ticket.
592 template <class T> static T load_acquire(T* addr) {
593 T t = *addr; // acquire barrier
594 asm_volatile_memory();
599 static void store_release(T* addr, T v) {
600 asm_volatile_memory();
601 *addr = v; // release barrier
606 constexpr RWTicketSpinLockT() {}
608 RWTicketSpinLockT(RWTicketSpinLockT const&) = delete;
609 RWTicketSpinLockT& operator=(RWTicketSpinLockT const&) = delete;
613 writeLockAggressive();
620 * Both try_lock and try_lock_shared diverge in our implementation from the
621 * lock algorithm described in the link above.
623 * In the read case, it is undesirable that the readers could wait
624 * for another reader (before increasing ticket.read in the other
625 * implementation). Our approach gives up on
626 * first-come-first-serve, but our benchmarks showed improve
627 * performance for both readers and writers under heavily contended
628 * cases, particularly when the number of threads exceeds the number
631 * We have writeLockAggressive() using the original implementation
632 * for a writer, which gives some advantage to the writer over the
633 * readers---for that path it is guaranteed that the writer will
634 * acquire the lock after all the existing readers exit.
638 FullInt old = t.whole = load_acquire(&ticket.whole);
639 if (t.users != t.write) {
643 return __sync_bool_compare_and_swap(&ticket.whole, old, t.whole);
647 * Call this if you want to prioritize writer to avoid starvation.
648 * Unlike writeLockNice, immediately acquires the write lock when
649 * the existing readers (arriving before the writer) finish their
652 void writeLockAggressive() {
653 // std::this_thread::yield() is needed here to avoid a pathology if the number
654 // of threads attempting concurrent writes is >= the number of real
655 // cores allocated to this process. This is less likely than the
656 // corresponding situation in lock_shared(), but we still want to
658 uint_fast32_t count = 0;
659 QuarterInt val = __sync_fetch_and_add(&ticket.users, 1);
660 while (val != load_acquire(&ticket.write)) {
661 asm_volatile_pause();
662 if (UNLIKELY(++count > 1000)) {
663 std::this_thread::yield();
668 // Call this when the writer should be nicer to the readers.
669 void writeLockNice() {
670 // Here it doesn't cpu-relax the writer.
672 // This is because usually we have many more readers than the
673 // writers, so the writer has less chance to get the lock when
674 // there are a lot of competing readers. The aggressive spinning
675 // can help to avoid starving writers.
677 // We don't worry about std::this_thread::yield() here because the caller
678 // has already explicitly abandoned fairness.
679 while (!try_lock()) {}
682 // Atomically unlock the write-lock from writer and acquire the read-lock.
683 void unlock_and_lock_shared() {
684 QuarterInt val = __sync_fetch_and_add(&ticket.read, 1);
687 // Release writer permission on the lock.
690 t.whole = load_acquire(&ticket.whole);
692 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
693 FullInt old = t.whole;
694 // SSE2 can reduce the lock and unlock overhead by 10%
695 static const QuarterInt kDeltaBuf[4] = { 1, 1, 0, 0 }; // write/read/user
696 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
697 __m128i m = IntTraitType::fromInteger(old);
698 t.whole = IntTraitType::addParallel(m, kDelta);
703 store_release(&ticket.readWrite, t.readWrite);
707 // std::this_thread::yield() is important here because we can't grab the
708 // shared lock if there is a pending writeLockAggressive, so we
709 // need to let threads that already have a shared lock complete
710 uint_fast32_t count = 0;
711 while (!LIKELY(try_lock_shared())) {
712 asm_volatile_pause();
713 if (UNLIKELY((++count & 1023) == 0)) {
714 std::this_thread::yield();
719 bool try_lock_shared() {
721 old.whole = t.whole = load_acquire(&ticket.whole);
722 old.users = old.read;
723 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
724 // SSE2 may reduce the total lock and unlock overhead by 10%
725 static const QuarterInt kDeltaBuf[4] = { 0, 1, 1, 0 }; // write/read/user
726 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
727 __m128i m = IntTraitType::fromInteger(old.whole);
728 t.whole = IntTraitType::addParallel(m, kDelta);
733 return __sync_bool_compare_and_swap(&ticket.whole, old.whole, t.whole);
736 void unlock_shared() {
737 __sync_fetch_and_add(&ticket.write, 1);
742 typedef RWTicketSpinLockT<kBitWidth, kFavorWriter> RWSpinLock;
745 ReadHolder(ReadHolder const&) = delete;
746 ReadHolder& operator=(ReadHolder const&) = delete;
748 explicit ReadHolder(RWSpinLock* lock) : lock_(lock) {
750 lock_->lock_shared();
754 explicit ReadHolder(RWSpinLock &lock) : lock_ (&lock) {
756 lock_->lock_shared();
760 // atomically unlock the write-lock from writer and acquire the read-lock
761 explicit ReadHolder(WriteHolder *writer) : lock_(nullptr) {
762 std::swap(this->lock_, writer->lock_);
764 lock_->unlock_and_lock_shared();
770 lock_->unlock_shared();
774 void reset(RWSpinLock *lock = nullptr) {
776 lock_->unlock_shared();
780 lock_->lock_shared();
784 void swap(ReadHolder *other) {
785 std::swap(this->lock_, other->lock_);
794 WriteHolder(WriteHolder const&) = delete;
795 WriteHolder& operator=(WriteHolder const&) = delete;
797 explicit WriteHolder(RWSpinLock* lock) : lock_(lock) {
802 explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) {
814 void reset(RWSpinLock *lock = nullptr) {
827 void swap(WriteHolder *other) {
828 std::swap(this->lock_, other->lock_);
832 friend class ReadHolder;
837 typedef RWTicketSpinLockT<32> RWTicketSpinLock32;
838 typedef RWTicketSpinLockT<64> RWTicketSpinLock64;
840 #endif // RW_SPINLOCK_USE_X86_INTRINSIC_
844 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
845 #undef RW_SPINLOCK_USE_X86_INTRINSIC_