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.
18 * N.B. You most likely do _not_ want to use RWSpinLock or any other
19 * kind of spinlock. Use SharedMutex instead.
21 * In short, spinlocks in preemptive multi-tasking operating systems
22 * have serious problems and fast mutexes like SharedMutex are almost
23 * certainly the better choice, because letting the OS scheduler put a
24 * thread to sleep is better for system responsiveness and throughput
25 * than wasting a timeslice repeatedly querying a lock held by a
26 * thread that's blocked, and you can't prevent userspace
29 * Spinlocks in an operating system kernel make much more sense than
30 * they do in userspace.
32 * -------------------------------------------------------------------
34 * Two Read-Write spin lock implementations.
36 * Ref: http://locklessinc.com/articles/locks
38 * Both locks here are faster than pthread_rwlock and have very low
39 * overhead (usually 20-30ns). They don't use any system mutexes and
40 * are very compact (4/8 bytes), so are suitable for per-instance
41 * based locking, particularly when contention is not expected.
43 * For a spinlock, RWSpinLock is a reasonable choice. (See the note
44 * about for why a spin lock is frequently a bad idea generally.)
45 * RWSpinLock has minimal overhead, and comparable contention
46 * performance when the number of competing threads is less than or
47 * equal to the number of logical CPUs. Even as the number of
48 * threads gets larger, RWSpinLock can still be very competitive in
49 * READ, although it is slower on WRITE, and also inherently unfair
52 * RWTicketSpinLock shows more balanced READ/WRITE performance. If
53 * your application really needs a lot more threads, and a
54 * higher-priority writer, prefer one of the RWTicketSpinLock locks.
58 * RWTicketSpinLock locks can only be used with GCC on x86/x86-64
61 * RWTicketSpinLock<32> only allows up to 2^8 - 1 concurrent
62 * readers and writers.
64 * RWTicketSpinLock<64> only allows up to 2^16 - 1 concurrent
65 * readers and writers.
67 * RWTicketSpinLock<..., true> (kFavorWriter = true, that is, strict
68 * writer priority) is NOT reentrant, even for lock_shared().
70 * The lock will not grant any new shared (read) accesses while a thread
71 * attempting to acquire the lock in write mode is blocked. (That is,
72 * if the lock is held in shared mode by N threads, and a thread attempts
73 * to acquire it in write mode, no one else can acquire it in shared mode
74 * until these N threads release the lock and then the blocked thread
75 * acquires and releases the exclusive lock.) This also applies for
76 * attempts to reacquire the lock in shared mode by threads that already
77 * hold it in shared mode, making the lock non-reentrant.
79 * RWSpinLock handles 2^30 - 1 concurrent readers.
81 * @author Xin Liu <xliux@fb.com>
87 ========================================================================
88 Benchmark on (Intel(R) Xeon(R) CPU L5630 @ 2.13GHz) 8 cores(16 HTs)
89 ========================================================================
91 ------------------------------------------------------------------------------
92 1. Single thread benchmark (read/write lock + unlock overhead)
93 Benchmark Iters Total t t/iter iter/sec
94 -------------------------------------------------------------------------------
95 * BM_RWSpinLockRead 100000 1.786 ms 17.86 ns 53.4M
96 +30.5% BM_RWSpinLockWrite 100000 2.331 ms 23.31 ns 40.91M
97 +85.7% BM_RWTicketSpinLock32Read 100000 3.317 ms 33.17 ns 28.75M
98 +96.0% BM_RWTicketSpinLock32Write 100000 3.5 ms 35 ns 27.25M
99 +85.6% BM_RWTicketSpinLock64Read 100000 3.315 ms 33.15 ns 28.77M
100 +96.0% BM_RWTicketSpinLock64Write 100000 3.5 ms 35 ns 27.25M
101 +85.7% BM_RWTicketSpinLock32FavorWriterRead 100000 3.317 ms 33.17 ns 28.75M
102 +29.7% BM_RWTicketSpinLock32FavorWriterWrite 100000 2.316 ms 23.16 ns 41.18M
103 +85.3% BM_RWTicketSpinLock64FavorWriterRead 100000 3.309 ms 33.09 ns 28.82M
104 +30.2% BM_RWTicketSpinLock64FavorWriterWrite 100000 2.325 ms 23.25 ns 41.02M
105 + 175% BM_PThreadRWMutexRead 100000 4.917 ms 49.17 ns 19.4M
106 + 166% BM_PThreadRWMutexWrite 100000 4.757 ms 47.57 ns 20.05M
108 ------------------------------------------------------------------------------
109 2. Contention Benchmark 90% read 10% write
110 Benchmark hits average min max sigma
111 ------------------------------------------------------------------------------
112 ---------- 8 threads ------------
113 RWSpinLock Write 142666 220ns 78ns 40.8us 269ns
114 RWSpinLock Read 1282297 222ns 80ns 37.7us 248ns
115 RWTicketSpinLock Write 85692 209ns 71ns 17.9us 252ns
116 RWTicketSpinLock Read 769571 215ns 78ns 33.4us 251ns
117 pthread_rwlock_t Write 84248 2.48us 99ns 269us 8.19us
118 pthread_rwlock_t Read 761646 933ns 101ns 374us 3.25us
120 ---------- 16 threads ------------
121 RWSpinLock Write 124236 237ns 78ns 261us 801ns
122 RWSpinLock Read 1115807 236ns 78ns 2.27ms 2.17us
123 RWTicketSpinLock Write 81781 231ns 71ns 31.4us 351ns
124 RWTicketSpinLock Read 734518 238ns 78ns 73.6us 379ns
125 pthread_rwlock_t Write 83363 7.12us 99ns 785us 28.1us
126 pthread_rwlock_t Read 754978 2.18us 101ns 1.02ms 14.3us
128 ---------- 50 threads ------------
129 RWSpinLock Write 131142 1.37us 82ns 7.53ms 68.2us
130 RWSpinLock Read 1181240 262ns 78ns 6.62ms 12.7us
131 RWTicketSpinLock Write 83045 397ns 73ns 7.01ms 31.5us
132 RWTicketSpinLock Read 744133 386ns 78ns 11ms 31.4us
133 pthread_rwlock_t Write 80849 112us 103ns 4.52ms 263us
134 pthread_rwlock_t Read 728698 24us 101ns 7.28ms 194us
138 #include <folly/Portability.h>
139 #include <folly/portability/Asm.h>
141 #if defined(__GNUC__) && (defined(__i386) || FOLLY_X64 || defined(ARCH_K8))
142 #define RW_SPINLOCK_USE_X86_INTRINSIC_
143 #include <x86intrin.h>
144 #elif defined(_MSC_VER) && defined(FOLLY_X64)
145 #define RW_SPINLOCK_USE_X86_INTRINSIC_
147 #undef RW_SPINLOCK_USE_X86_INTRINSIC_
150 // iOS doesn't define _mm_cvtsi64_si128 and friends
151 #if (FOLLY_SSE >= 2) && !FOLLY_MOBILE && FOLLY_X64
152 #define RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
154 #undef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
162 #include <glog/logging.h>
164 #include <folly/Likely.h>
169 * A simple, small (4-bytes), but unfair rwlock. Use it when you want
170 * a nice writer and don't expect a lot of write/read contention, or
171 * when you need small rwlocks since you are creating a large number
174 * Note that the unfairness here is extreme: if the lock is
175 * continually accessed for read, writers will never get a chance. If
176 * the lock can be that highly contended this class is probably not an
177 * ideal choice anyway.
179 * It currently implements most of the Lockable, SharedLockable and
180 * UpgradeLockable concepts except the TimedLockable related locking/unlocking
184 enum : int32_t { READER = 4, UPGRADED = 2, WRITER = 1 };
186 constexpr RWSpinLock() : bits_(0) {}
188 RWSpinLock(RWSpinLock const&) = delete;
189 RWSpinLock& operator=(RWSpinLock const&) = delete;
193 uint_fast32_t count = 0;
194 while (!LIKELY(try_lock())) {
195 if (++count > 1000) {
196 std::this_thread::yield();
201 // Writer is responsible for clearing up both the UPGRADED and WRITER bits.
203 static_assert(READER > WRITER + UPGRADED, "wrong bits!");
204 bits_.fetch_and(~(WRITER | UPGRADED), std::memory_order_release);
207 // SharedLockable Concept
209 uint_fast32_t count = 0;
210 while (!LIKELY(try_lock_shared())) {
211 if (++count > 1000) {
212 std::this_thread::yield();
217 void unlock_shared() {
218 bits_.fetch_add(-READER, std::memory_order_release);
221 // Downgrade the lock from writer status to reader status.
222 void unlock_and_lock_shared() {
223 bits_.fetch_add(READER, std::memory_order_acquire);
227 // UpgradeLockable Concept
228 void lock_upgrade() {
229 uint_fast32_t count = 0;
230 while (!try_lock_upgrade()) {
231 if (++count > 1000) {
232 std::this_thread::yield();
237 void unlock_upgrade() {
238 bits_.fetch_add(-UPGRADED, std::memory_order_acq_rel);
241 // unlock upgrade and try to acquire write lock
242 void unlock_upgrade_and_lock() {
244 while (!try_unlock_upgrade_and_lock()) {
245 if (++count > 1000) {
246 std::this_thread::yield();
251 // unlock upgrade and read lock atomically
252 void unlock_upgrade_and_lock_shared() {
253 bits_.fetch_add(READER - UPGRADED, std::memory_order_acq_rel);
256 // write unlock and upgrade lock atomically
257 void unlock_and_lock_upgrade() {
258 // need to do it in two steps here -- as the UPGRADED bit might be OR-ed at
259 // the same time when other threads are trying do try_lock_upgrade().
260 bits_.fetch_or(UPGRADED, std::memory_order_acquire);
261 bits_.fetch_add(-WRITER, std::memory_order_release);
265 // Attempt to acquire writer permission. Return false if we didn't get it.
268 return bits_.compare_exchange_strong(expect, WRITER,
269 std::memory_order_acq_rel);
272 // Try to get reader permission on the lock. This can fail if we
273 // find out someone is a writer or upgrader.
274 // Setting the UPGRADED bit would allow a writer-to-be to indicate
275 // its intention to write and block any new readers while waiting
276 // for existing readers to finish and release their read locks. This
277 // helps avoid starving writers (promoted from upgraders).
278 bool try_lock_shared() {
279 // fetch_add is considerably (100%) faster than compare_exchange,
280 // so here we are optimizing for the common (lock success) case.
281 int32_t value = bits_.fetch_add(READER, std::memory_order_acquire);
282 if (UNLIKELY(value & (WRITER|UPGRADED))) {
283 bits_.fetch_add(-READER, std::memory_order_release);
289 // try to unlock upgrade and write lock atomically
290 bool try_unlock_upgrade_and_lock() {
291 int32_t expect = UPGRADED;
292 return bits_.compare_exchange_strong(expect, WRITER,
293 std::memory_order_acq_rel);
296 // try to acquire an upgradable lock.
297 bool try_lock_upgrade() {
298 int32_t value = bits_.fetch_or(UPGRADED, std::memory_order_acquire);
300 // Note: when failed, we cannot flip the UPGRADED bit back,
301 // as in this case there is either another upgrade lock or a write lock.
302 // If it's a write lock, the bit will get cleared up when that lock's done
304 return ((value & (UPGRADED | WRITER)) == 0);
307 // mainly for debugging purposes.
308 int32_t bits() const { return bits_.load(std::memory_order_acquire); }
311 class UpgradedHolder;
316 explicit ReadHolder(RWSpinLock* lock) : lock_(lock) {
318 lock_->lock_shared();
322 explicit ReadHolder(RWSpinLock& lock) : lock_(&lock) {
323 lock_->lock_shared();
326 ReadHolder(ReadHolder&& other) noexcept : lock_(other.lock_) {
327 other.lock_ = nullptr;
331 explicit ReadHolder(UpgradedHolder&& upgraded) : lock_(upgraded.lock_) {
332 upgraded.lock_ = nullptr;
334 lock_->unlock_upgrade_and_lock_shared();
338 explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
339 writer.lock_ = nullptr;
341 lock_->unlock_and_lock_shared();
345 ReadHolder& operator=(ReadHolder&& other) {
347 swap(lock_, other.lock_);
351 ReadHolder(const ReadHolder& other) = delete;
352 ReadHolder& operator=(const ReadHolder& other) = delete;
356 lock_->unlock_shared();
360 void reset(RWSpinLock* lock = nullptr) {
365 lock_->unlock_shared();
369 lock_->lock_shared();
373 void swap(ReadHolder* other) {
374 std::swap(lock_, other->lock_);
378 friend class UpgradedHolder;
379 friend class WriteHolder;
383 class UpgradedHolder {
385 explicit UpgradedHolder(RWSpinLock* lock) : lock_(lock) {
387 lock_->lock_upgrade();
391 explicit UpgradedHolder(RWSpinLock& lock) : lock_(&lock) {
392 lock_->lock_upgrade();
395 explicit UpgradedHolder(WriteHolder&& writer) {
396 lock_ = writer.lock_;
397 writer.lock_ = nullptr;
399 lock_->unlock_and_lock_upgrade();
403 UpgradedHolder(UpgradedHolder&& other) noexcept : lock_(other.lock_) {
404 other.lock_ = nullptr;
407 UpgradedHolder& operator =(UpgradedHolder&& other) {
409 swap(lock_, other.lock_);
413 UpgradedHolder(const UpgradedHolder& other) = delete;
414 UpgradedHolder& operator =(const UpgradedHolder& other) = delete;
418 lock_->unlock_upgrade();
422 void reset(RWSpinLock* lock = nullptr) {
427 lock_->unlock_upgrade();
431 lock_->lock_upgrade();
435 void swap(UpgradedHolder* other) {
437 swap(lock_, other->lock_);
441 friend class WriteHolder;
442 friend class ReadHolder;
448 explicit WriteHolder(RWSpinLock* lock) : lock_(lock) {
454 explicit WriteHolder(RWSpinLock& lock) : lock_(&lock) {
458 // promoted from an upgrade lock holder
459 explicit WriteHolder(UpgradedHolder&& upgraded) {
460 lock_ = upgraded.lock_;
461 upgraded.lock_ = nullptr;
463 lock_->unlock_upgrade_and_lock();
467 WriteHolder(WriteHolder&& other) noexcept : lock_(other.lock_) {
468 other.lock_ = nullptr;
471 WriteHolder& operator =(WriteHolder&& other) {
473 swap(lock_, other.lock_);
477 WriteHolder(const WriteHolder& other) = delete;
478 WriteHolder& operator =(const WriteHolder& other) = delete;
486 void reset(RWSpinLock* lock = nullptr) {
499 void swap(WriteHolder* other) {
501 swap(lock_, other->lock_);
505 friend class ReadHolder;
506 friend class UpgradedHolder;
511 std::atomic<int32_t> bits_;
515 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
516 // A more balanced Read-Write spin lock implemented based on GCC intrinsics.
519 template <size_t kBitWidth> struct RWTicketIntTrait {
520 static_assert(kBitWidth == 32 || kBitWidth == 64,
521 "bit width has to be either 32 or 64 ");
525 struct RWTicketIntTrait<64> {
526 typedef uint64_t FullInt;
527 typedef uint32_t HalfInt;
528 typedef uint16_t QuarterInt;
530 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
531 static __m128i make128(const uint16_t v[4]) {
532 return _mm_set_epi16(0, 0, 0, 0,
533 short(v[3]), short(v[2]), short(v[1]), short(v[0]));
535 static inline __m128i fromInteger(uint64_t from) {
536 return _mm_cvtsi64_si128(int64_t(from));
538 static inline uint64_t toInteger(__m128i in) {
539 return uint64_t(_mm_cvtsi128_si64(in));
541 static inline uint64_t addParallel(__m128i in, __m128i kDelta) {
542 return toInteger(_mm_add_epi16(in, kDelta));
548 struct RWTicketIntTrait<32> {
549 typedef uint32_t FullInt;
550 typedef uint16_t HalfInt;
551 typedef uint8_t QuarterInt;
553 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
554 static __m128i make128(const uint8_t v[4]) {
559 char(v[3]), char(v[2]), char(v[1]), char(v[0]));
561 static inline __m128i fromInteger(uint32_t from) {
562 return _mm_cvtsi32_si128(int32_t(from));
564 static inline uint32_t toInteger(__m128i in) {
565 return uint32_t(_mm_cvtsi128_si32(in));
567 static inline uint32_t addParallel(__m128i in, __m128i kDelta) {
568 return toInteger(_mm_add_epi8(in, kDelta));
572 } // namespace detail
574 template <size_t kBitWidth, bool kFavorWriter = false>
575 class RWTicketSpinLockT {
576 typedef detail::RWTicketIntTrait<kBitWidth> IntTraitType;
577 typedef typename detail::RWTicketIntTrait<kBitWidth>::FullInt FullInt;
578 typedef typename detail::RWTicketIntTrait<kBitWidth>::HalfInt HalfInt;
579 typedef typename detail::RWTicketIntTrait<kBitWidth>::QuarterInt
583 constexpr RWTicket() : whole(0) {}
586 __extension__ struct {
593 private: // Some x64-specific utilities for atomic access to ticket.
594 template <class T> static T load_acquire(T* addr) {
595 T t = *addr; // acquire barrier
596 asm_volatile_memory();
601 static void store_release(T* addr, T v) {
602 asm_volatile_memory();
603 *addr = v; // release barrier
608 constexpr RWTicketSpinLockT() {}
610 RWTicketSpinLockT(RWTicketSpinLockT const&) = delete;
611 RWTicketSpinLockT& operator=(RWTicketSpinLockT const&) = delete;
615 writeLockAggressive();
622 * Both try_lock and try_lock_shared diverge in our implementation from the
623 * lock algorithm described in the link above.
625 * In the read case, it is undesirable that the readers could wait
626 * for another reader (before increasing ticket.read in the other
627 * implementation). Our approach gives up on
628 * first-come-first-serve, but our benchmarks showed improve
629 * performance for both readers and writers under heavily contended
630 * cases, particularly when the number of threads exceeds the number
633 * We have writeLockAggressive() using the original implementation
634 * for a writer, which gives some advantage to the writer over the
635 * readers---for that path it is guaranteed that the writer will
636 * acquire the lock after all the existing readers exit.
640 FullInt old = t.whole = load_acquire(&ticket.whole);
641 if (t.users != t.write) {
645 return __sync_bool_compare_and_swap(&ticket.whole, old, t.whole);
649 * Call this if you want to prioritize writer to avoid starvation.
650 * Unlike writeLockNice, immediately acquires the write lock when
651 * the existing readers (arriving before the writer) finish their
654 void writeLockAggressive() {
655 // std::this_thread::yield() is needed here to avoid a pathology if the number
656 // of threads attempting concurrent writes is >= the number of real
657 // cores allocated to this process. This is less likely than the
658 // corresponding situation in lock_shared(), but we still want to
660 uint_fast32_t count = 0;
661 QuarterInt val = __sync_fetch_and_add(&ticket.users, 1);
662 while (val != load_acquire(&ticket.write)) {
663 asm_volatile_pause();
664 if (UNLIKELY(++count > 1000)) {
665 std::this_thread::yield();
670 // Call this when the writer should be nicer to the readers.
671 void writeLockNice() {
672 // Here it doesn't cpu-relax the writer.
674 // This is because usually we have many more readers than the
675 // writers, so the writer has less chance to get the lock when
676 // there are a lot of competing readers. The aggressive spinning
677 // can help to avoid starving writers.
679 // We don't worry about std::this_thread::yield() here because the caller
680 // has already explicitly abandoned fairness.
681 while (!try_lock()) {}
684 // Atomically unlock the write-lock from writer and acquire the read-lock.
685 void unlock_and_lock_shared() {
686 QuarterInt val = __sync_fetch_and_add(&ticket.read, 1);
689 // Release writer permission on the lock.
692 t.whole = load_acquire(&ticket.whole);
693 FullInt old = t.whole;
695 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
696 // SSE2 can reduce the lock and unlock overhead by 10%
697 static const QuarterInt kDeltaBuf[4] = { 1, 1, 0, 0 }; // write/read/user
698 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
699 __m128i m = IntTraitType::fromInteger(old);
700 t.whole = IntTraitType::addParallel(m, kDelta);
705 store_release(&ticket.readWrite, t.readWrite);
709 // std::this_thread::yield() is important here because we can't grab the
710 // shared lock if there is a pending writeLockAggressive, so we
711 // need to let threads that already have a shared lock complete
712 uint_fast32_t count = 0;
713 while (!LIKELY(try_lock_shared())) {
714 asm_volatile_pause();
715 if (UNLIKELY((++count & 1023) == 0)) {
716 std::this_thread::yield();
721 bool try_lock_shared() {
723 old.whole = t.whole = load_acquire(&ticket.whole);
724 old.users = old.read;
725 #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_
726 // SSE2 may reduce the total lock and unlock overhead by 10%
727 static const QuarterInt kDeltaBuf[4] = { 0, 1, 1, 0 }; // write/read/user
728 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
729 __m128i m = IntTraitType::fromInteger(old.whole);
730 t.whole = IntTraitType::addParallel(m, kDelta);
735 return __sync_bool_compare_and_swap(&ticket.whole, old.whole, t.whole);
738 void unlock_shared() {
739 __sync_fetch_and_add(&ticket.write, 1);
744 typedef RWTicketSpinLockT<kBitWidth, kFavorWriter> RWSpinLock;
747 ReadHolder(ReadHolder const&) = delete;
748 ReadHolder& operator=(ReadHolder const&) = delete;
750 explicit ReadHolder(RWSpinLock* lock) : lock_(lock) {
752 lock_->lock_shared();
756 explicit ReadHolder(RWSpinLock &lock) : lock_ (&lock) {
758 lock_->lock_shared();
762 // atomically unlock the write-lock from writer and acquire the read-lock
763 explicit ReadHolder(WriteHolder *writer) : lock_(nullptr) {
764 std::swap(this->lock_, writer->lock_);
766 lock_->unlock_and_lock_shared();
772 lock_->unlock_shared();
776 void reset(RWSpinLock *lock = nullptr) {
778 lock_->unlock_shared();
782 lock_->lock_shared();
786 void swap(ReadHolder *other) {
787 std::swap(this->lock_, other->lock_);
796 WriteHolder(WriteHolder const&) = delete;
797 WriteHolder& operator=(WriteHolder const&) = delete;
799 explicit WriteHolder(RWSpinLock* lock) : lock_(lock) {
804 explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) {
816 void reset(RWSpinLock *lock = nullptr) {
829 void swap(WriteHolder *other) {
830 std::swap(this->lock_, other->lock_);
834 friend class ReadHolder;
839 typedef RWTicketSpinLockT<32> RWTicketSpinLock32;
840 typedef RWTicketSpinLockT<64> RWTicketSpinLock64;
842 #endif // RW_SPINLOCK_USE_X86_INTRINSIC_
846 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
847 #undef RW_SPINLOCK_USE_X86_INTRINSIC_