2 * Copyright 2014 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 * Two Read-Write spin lock implementations.
20 * Ref: http://locklessinc.com/articles/locks
22 * Both locks here are faster than pthread_rwlock and have very low
23 * overhead (usually 20-30ns). They don't use any system mutexes and
24 * are very compact (4/8 bytes), so are suitable for per-instance
25 * based locking, particularly when contention is not expected.
27 * In most cases, RWSpinLock is a reasonable choice. It has minimal
28 * overhead, and comparable contention performance when the number of
29 * competing threads is less than or equal to the number of logical
30 * CPUs. Even as the number of threads gets larger, RWSpinLock can
31 * still be very competitive in READ, although it is slower on WRITE,
32 * and also inherently unfair to writers.
34 * RWTicketSpinLock shows more balanced READ/WRITE performance. If
35 * your application really needs a lot more threads, and a
36 * higher-priority writer, prefer one of the RWTicketSpinLock locks.
40 * RWTicketSpinLock locks can only be used with GCC on x86/x86-64
43 * RWTicketSpinLock<32> only allows up to 2^8 - 1 concurrent
44 * readers and writers.
46 * RWTicketSpinLock<64> only allows up to 2^16 - 1 concurrent
47 * readers and writers.
49 * RWSpinLock handles 2^30 - 1 concurrent readers.
51 * @author Xin Liu <xliux@fb.com>
54 #ifndef FOLLY_RWSPINLOCK_H_
55 #define FOLLY_RWSPINLOCK_H_
58 ========================================================================
59 Benchmark on (Intel(R) Xeon(R) CPU L5630 @ 2.13GHz) 8 cores(16 HTs)
60 ========================================================================
62 ------------------------------------------------------------------------------
63 1. Single thread benchmark (read/write lock + unlock overhead)
64 Benchmark Iters Total t t/iter iter/sec
65 -------------------------------------------------------------------------------
66 * BM_RWSpinLockRead 100000 1.786 ms 17.86 ns 53.4M
67 +30.5% BM_RWSpinLockWrite 100000 2.331 ms 23.31 ns 40.91M
68 +85.7% BM_RWTicketSpinLock32Read 100000 3.317 ms 33.17 ns 28.75M
69 +96.0% BM_RWTicketSpinLock32Write 100000 3.5 ms 35 ns 27.25M
70 +85.6% BM_RWTicketSpinLock64Read 100000 3.315 ms 33.15 ns 28.77M
71 +96.0% BM_RWTicketSpinLock64Write 100000 3.5 ms 35 ns 27.25M
72 +85.7% BM_RWTicketSpinLock32FavorWriterRead 100000 3.317 ms 33.17 ns 28.75M
73 +29.7% BM_RWTicketSpinLock32FavorWriterWrite 100000 2.316 ms 23.16 ns 41.18M
74 +85.3% BM_RWTicketSpinLock64FavorWriterRead 100000 3.309 ms 33.09 ns 28.82M
75 +30.2% BM_RWTicketSpinLock64FavorWriterWrite 100000 2.325 ms 23.25 ns 41.02M
76 + 175% BM_PThreadRWMutexRead 100000 4.917 ms 49.17 ns 19.4M
77 + 166% BM_PThreadRWMutexWrite 100000 4.757 ms 47.57 ns 20.05M
79 ------------------------------------------------------------------------------
80 2. Contention Benchmark 90% read 10% write
81 Benchmark hits average min max sigma
82 ------------------------------------------------------------------------------
83 ---------- 8 threads ------------
84 RWSpinLock Write 142666 220ns 78ns 40.8us 269ns
85 RWSpinLock Read 1282297 222ns 80ns 37.7us 248ns
86 RWTicketSpinLock Write 85692 209ns 71ns 17.9us 252ns
87 RWTicketSpinLock Read 769571 215ns 78ns 33.4us 251ns
88 pthread_rwlock_t Write 84248 2.48us 99ns 269us 8.19us
89 pthread_rwlock_t Read 761646 933ns 101ns 374us 3.25us
91 ---------- 16 threads ------------
92 RWSpinLock Write 124236 237ns 78ns 261us 801ns
93 RWSpinLock Read 1115807 236ns 78ns 2.27ms 2.17us
94 RWTicketSpinLock Write 81781 231ns 71ns 31.4us 351ns
95 RWTicketSpinLock Read 734518 238ns 78ns 73.6us 379ns
96 pthread_rwlock_t Write 83363 7.12us 99ns 785us 28.1us
97 pthread_rwlock_t Read 754978 2.18us 101ns 1.02ms 14.3us
99 ---------- 50 threads ------------
100 RWSpinLock Write 131142 1.37us 82ns 7.53ms 68.2us
101 RWSpinLock Read 1181240 262ns 78ns 6.62ms 12.7us
102 RWTicketSpinLock Write 83045 397ns 73ns 7.01ms 31.5us
103 RWTicketSpinLock Read 744133 386ns 78ns 11ms 31.4us
104 pthread_rwlock_t Write 80849 112us 103ns 4.52ms 263us
105 pthread_rwlock_t Read 728698 24us 101ns 7.28ms 194us
109 #include "folly/Portability.h"
111 #if defined(__GNUC__) && !defined(__clang__) && \
112 (defined(__i386) || FOLLY_X64 || \
114 #define RW_SPINLOCK_USE_X86_INTRINSIC_
115 #include <x86intrin.h>
117 #undef RW_SPINLOCK_USE_X86_INTRINSIC_
123 #include <boost/noncopyable.hpp>
126 #include <glog/logging.h>
128 #include "folly/Likely.h"
133 * A simple, small (4-bytes), but unfair rwlock. Use it when you want
134 * a nice writer and don't expect a lot of write/read contention, or
135 * when you need small rwlocks since you are creating a large number
138 * Note that the unfairness here is extreme: if the lock is
139 * continually accessed for read, writers will never get a chance. If
140 * the lock can be that highly contended this class is probably not an
141 * ideal choice anyway.
143 * It currently implements most of the Lockable, SharedLockable and
144 * UpgradeLockable concepts except the TimedLockable related locking/unlocking
147 class RWSpinLock : boost::noncopyable {
148 enum : int32_t { READER = 4, UPGRADED = 2, WRITER = 1 };
150 RWSpinLock() : bits_(0) {}
155 while (!LIKELY(try_lock())) {
156 if (++count > 1000) sched_yield();
160 // Writer is responsible for clearing up both the UPGRADED and WRITER bits.
162 static_assert(READER > WRITER + UPGRADED, "wrong bits!");
163 bits_.fetch_and(~(WRITER | UPGRADED), std::memory_order_release);
166 // SharedLockable Concept
169 while (!LIKELY(try_lock_shared())) {
170 if (++count > 1000) sched_yield();
174 void unlock_shared() {
175 bits_.fetch_add(-READER, std::memory_order_release);
178 // Downgrade the lock from writer status to reader status.
179 void unlock_and_lock_shared() {
180 bits_.fetch_add(READER, std::memory_order_acquire);
184 // UpgradeLockable Concept
185 void lock_upgrade() {
187 while (!try_lock_upgrade()) {
188 if (++count > 1000) sched_yield();
192 void unlock_upgrade() {
193 bits_.fetch_add(-UPGRADED, std::memory_order_acq_rel);
196 // unlock upgrade and try to acquire write lock
197 void unlock_upgrade_and_lock() {
199 while (!try_unlock_upgrade_and_lock()) {
200 if (++count > 1000) sched_yield();
204 // unlock upgrade and read lock atomically
205 void unlock_upgrade_and_lock_shared() {
206 bits_.fetch_add(READER - UPGRADED, std::memory_order_acq_rel);
209 // write unlock and upgrade lock atomically
210 void unlock_and_lock_upgrade() {
211 // need to do it in two steps here -- as the UPGRADED bit might be OR-ed at
212 // the same time when other threads are trying do try_lock_upgrade().
213 bits_.fetch_or(UPGRADED, std::memory_order_acquire);
214 bits_.fetch_add(-WRITER, std::memory_order_release);
218 // Attempt to acquire writer permission. Return false if we didn't get it.
221 return bits_.compare_exchange_strong(expect, WRITER,
222 std::memory_order_acq_rel);
225 // Try to get reader permission on the lock. This can fail if we
226 // find out someone is a writer or upgrader.
227 // Setting the UPGRADED bit would allow a writer-to-be to indicate
228 // its intention to write and block any new readers while waiting
229 // for existing readers to finish and release their read locks. This
230 // helps avoid starving writers (promoted from upgraders).
231 bool try_lock_shared() {
232 // fetch_add is considerably (100%) faster than compare_exchange,
233 // so here we are optimizing for the common (lock success) case.
234 int32_t value = bits_.fetch_add(READER, std::memory_order_acquire);
235 if (UNLIKELY(value & (WRITER|UPGRADED))) {
236 bits_.fetch_add(-READER, std::memory_order_release);
242 // try to unlock upgrade and write lock atomically
243 bool try_unlock_upgrade_and_lock() {
244 int32_t expect = UPGRADED;
245 return bits_.compare_exchange_strong(expect, WRITER,
246 std::memory_order_acq_rel);
249 // try to acquire an upgradable lock.
250 bool try_lock_upgrade() {
251 int32_t value = bits_.fetch_or(UPGRADED, std::memory_order_acquire);
253 // Note: when failed, we cannot flip the UPGRADED bit back,
254 // as in this case there is either another upgrade lock or a write lock.
255 // If it's a write lock, the bit will get cleared up when that lock's done
257 return ((value & (UPGRADED | WRITER)) == 0);
260 // mainly for debugging purposes.
261 int32_t bits() const { return bits_.load(std::memory_order_acquire); }
264 class UpgradedHolder;
269 explicit ReadHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
270 if (lock_) lock_->lock_shared();
273 explicit ReadHolder(RWSpinLock& lock) : lock_(&lock) {
274 lock_->lock_shared();
277 ReadHolder(ReadHolder&& other) : lock_(other.lock_) {
278 other.lock_ = nullptr;
282 explicit ReadHolder(UpgradedHolder&& upgraded) : lock_(upgraded.lock_) {
283 upgraded.lock_ = nullptr;
284 if (lock_) lock_->unlock_upgrade_and_lock_shared();
287 explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
288 writer.lock_ = nullptr;
289 if (lock_) lock_->unlock_and_lock_shared();
292 ReadHolder& operator=(ReadHolder&& other) {
294 swap(lock_, other.lock_);
298 ReadHolder(const ReadHolder& other) = delete;
299 ReadHolder& operator=(const ReadHolder& other) = delete;
301 ~ReadHolder() { if (lock_) lock_->unlock_shared(); }
303 void reset(RWSpinLock* lock = nullptr) {
304 if (lock == lock_) return;
305 if (lock_) lock_->unlock_shared();
307 if (lock_) lock_->lock_shared();
310 void swap(ReadHolder* other) {
311 std::swap(lock_, other->lock_);
315 friend class UpgradedHolder;
316 friend class WriteHolder;
320 class UpgradedHolder {
322 explicit UpgradedHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
323 if (lock_) lock_->lock_upgrade();
326 explicit UpgradedHolder(RWSpinLock& lock) : lock_(&lock) {
327 lock_->lock_upgrade();
330 explicit UpgradedHolder(WriteHolder&& writer) {
331 lock_ = writer.lock_;
332 writer.lock_ = nullptr;
333 if (lock_) lock_->unlock_and_lock_upgrade();
336 UpgradedHolder(UpgradedHolder&& other) : lock_(other.lock_) {
337 other.lock_ = nullptr;
340 UpgradedHolder& operator =(UpgradedHolder&& other) {
342 swap(lock_, other.lock_);
346 UpgradedHolder(const UpgradedHolder& other) = delete;
347 UpgradedHolder& operator =(const UpgradedHolder& other) = delete;
349 ~UpgradedHolder() { if (lock_) lock_->unlock_upgrade(); }
351 void reset(RWSpinLock* lock = nullptr) {
352 if (lock == lock_) return;
353 if (lock_) lock_->unlock_upgrade();
355 if (lock_) lock_->lock_upgrade();
358 void swap(UpgradedHolder* other) {
360 swap(lock_, other->lock_);
364 friend class WriteHolder;
365 friend class ReadHolder;
371 explicit WriteHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
372 if (lock_) lock_->lock();
375 explicit WriteHolder(RWSpinLock& lock) : lock_(&lock) {
379 // promoted from an upgrade lock holder
380 explicit WriteHolder(UpgradedHolder&& upgraded) {
381 lock_ = upgraded.lock_;
382 upgraded.lock_ = nullptr;
383 if (lock_) lock_->unlock_upgrade_and_lock();
386 WriteHolder(WriteHolder&& other) : lock_(other.lock_) {
387 other.lock_ = nullptr;
390 WriteHolder& operator =(WriteHolder&& other) {
392 swap(lock_, other.lock_);
396 WriteHolder(const WriteHolder& other) = delete;
397 WriteHolder& operator =(const WriteHolder& other) = delete;
399 ~WriteHolder () { if (lock_) lock_->unlock(); }
401 void reset(RWSpinLock* lock = nullptr) {
402 if (lock == lock_) return;
403 if (lock_) lock_->unlock();
405 if (lock_) lock_->lock();
408 void swap(WriteHolder* other) {
410 swap(lock_, other->lock_);
414 friend class ReadHolder;
415 friend class UpgradedHolder;
419 // Synchronized<> adaptors
420 friend void acquireRead(RWSpinLock& l) { return l.lock_shared(); }
421 friend void acquireReadWrite(RWSpinLock& l) { return l.lock(); }
422 friend void releaseRead(RWSpinLock& l) { return l.unlock_shared(); }
423 friend void releaseReadWrite(RWSpinLock& l) { return l.unlock(); }
426 std::atomic<int32_t> bits_;
430 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
431 // A more balanced Read-Write spin lock implemented based on GCC intrinsics.
434 template <size_t kBitWidth> struct RWTicketIntTrait {
435 static_assert(kBitWidth == 32 || kBitWidth == 64,
436 "bit width has to be either 32 or 64 ");
440 struct RWTicketIntTrait<64> {
441 typedef uint64_t FullInt;
442 typedef uint32_t HalfInt;
443 typedef uint16_t QuarterInt;
446 static __m128i make128(const uint16_t v[4]) {
447 return _mm_set_epi16(0, 0, 0, 0, v[3], v[2], v[1], v[0]);
449 static inline __m128i fromInteger(uint64_t from) {
450 return _mm_cvtsi64_si128(from);
452 static inline uint64_t toInteger(__m128i in) {
453 return _mm_cvtsi128_si64(in);
455 static inline uint64_t addParallel(__m128i in, __m128i kDelta) {
456 return toInteger(_mm_add_epi16(in, kDelta));
462 struct RWTicketIntTrait<32> {
463 typedef uint32_t FullInt;
464 typedef uint16_t HalfInt;
465 typedef uint8_t QuarterInt;
468 static __m128i make128(const uint8_t v[4]) {
469 return _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
470 0, 0, 0, 0, v[3], v[2], v[1], v[0]);
472 static inline __m128i fromInteger(uint32_t from) {
473 return _mm_cvtsi32_si128(from);
475 static inline uint32_t toInteger(__m128i in) {
476 return _mm_cvtsi128_si32(in);
478 static inline uint32_t addParallel(__m128i in, __m128i kDelta) {
479 return toInteger(_mm_add_epi8(in, kDelta));
486 template<size_t kBitWidth, bool kFavorWriter=false>
487 class RWTicketSpinLockT : boost::noncopyable {
488 typedef detail::RWTicketIntTrait<kBitWidth> IntTraitType;
489 typedef typename detail::RWTicketIntTrait<kBitWidth>::FullInt FullInt;
490 typedef typename detail::RWTicketIntTrait<kBitWidth>::HalfInt HalfInt;
491 typedef typename detail::RWTicketIntTrait<kBitWidth>::QuarterInt
497 __extension__ struct {
504 private: // Some x64-specific utilities for atomic access to ticket.
505 template<class T> static T load_acquire(T* addr) {
506 T t = *addr; // acquire barrier
507 asm volatile("" : : : "memory");
512 static void store_release(T* addr, T v) {
513 asm volatile("" : : : "memory");
514 *addr = v; // release barrier
519 RWTicketSpinLockT() {
520 store_release(&ticket.whole, FullInt(0));
525 writeLockAggressive();
532 * Both try_lock and try_lock_shared diverge in our implementation from the
533 * lock algorithm described in the link above.
535 * In the read case, it is undesirable that the readers could wait
536 * for another reader (before increasing ticket.read in the other
537 * implementation). Our approach gives up on
538 * first-come-first-serve, but our benchmarks showed improve
539 * performance for both readers and writers under heavily contended
540 * cases, particularly when the number of threads exceeds the number
543 * We have writeLockAggressive() using the original implementation
544 * for a writer, which gives some advantage to the writer over the
545 * readers---for that path it is guaranteed that the writer will
546 * acquire the lock after all the existing readers exit.
550 FullInt old = t.whole = load_acquire(&ticket.whole);
551 if (t.users != t.write) return false;
553 return __sync_bool_compare_and_swap(&ticket.whole, old, t.whole);
557 * Call this if you want to prioritize writer to avoid starvation.
558 * Unlike writeLockNice, immediately acquires the write lock when
559 * the existing readers (arriving before the writer) finish their
562 void writeLockAggressive() {
563 // sched_yield() is needed here to avoid a pathology if the number
564 // of threads attempting concurrent writes is >= the number of real
565 // cores allocated to this process. This is less likely than the
566 // corresponding situation in lock_shared(), but we still want to
569 QuarterInt val = __sync_fetch_and_add(&ticket.users, 1);
570 while (val != load_acquire(&ticket.write)) {
571 asm volatile("pause");
572 if (UNLIKELY(++count > 1000)) sched_yield();
576 // Call this when the writer should be nicer to the readers.
577 void writeLockNice() {
578 // Here it doesn't cpu-relax the writer.
580 // This is because usually we have many more readers than the
581 // writers, so the writer has less chance to get the lock when
582 // there are a lot of competing readers. The aggressive spinning
583 // can help to avoid starving writers.
585 // We don't worry about sched_yield() here because the caller
586 // has already explicitly abandoned fairness.
587 while (!try_lock()) {}
590 // Atomically unlock the write-lock from writer and acquire the read-lock.
591 void unlock_and_lock_shared() {
592 QuarterInt val = __sync_fetch_and_add(&ticket.read, 1);
595 // Release writer permission on the lock.
598 t.whole = load_acquire(&ticket.whole);
599 FullInt old = t.whole;
602 // SSE2 can reduce the lock and unlock overhead by 10%
603 static const QuarterInt kDeltaBuf[4] = { 1, 1, 0, 0 }; // write/read/user
604 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
605 __m128i m = IntTraitType::fromInteger(old);
606 t.whole = IntTraitType::addParallel(m, kDelta);
611 store_release(&ticket.readWrite, t.readWrite);
615 // sched_yield() is important here because we can't grab the
616 // shared lock if there is a pending writeLockAggressive, so we
617 // need to let threads that already have a shared lock complete
619 while (!LIKELY(try_lock_shared())) {
620 asm volatile("pause");
621 if (UNLIKELY((++count & 1023) == 0)) sched_yield();
625 bool try_lock_shared() {
627 old.whole = t.whole = load_acquire(&ticket.whole);
628 old.users = old.read;
630 // SSE2 may reduce the total lock and unlock overhead by 10%
631 static const QuarterInt kDeltaBuf[4] = { 0, 1, 1, 0 }; // write/read/user
632 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
633 __m128i m = IntTraitType::fromInteger(old.whole);
634 t.whole = IntTraitType::addParallel(m, kDelta);
639 return __sync_bool_compare_and_swap(&ticket.whole, old.whole, t.whole);
642 void unlock_shared() {
643 QuarterInt val = __sync_fetch_and_add(&ticket.write, 1);
648 typedef RWTicketSpinLockT<kBitWidth, kFavorWriter> RWSpinLock;
649 class ReadHolder : boost::noncopyable {
651 explicit ReadHolder(RWSpinLock *lock = nullptr) :
653 if (lock_) lock_->lock_shared();
656 explicit ReadHolder(RWSpinLock &lock) : lock_ (&lock) {
657 if (lock_) lock_->lock_shared();
660 // atomically unlock the write-lock from writer and acquire the read-lock
661 explicit ReadHolder(WriteHolder *writer) : lock_(nullptr) {
662 std::swap(this->lock_, writer->lock_);
664 lock_->unlock_and_lock_shared();
669 if (lock_) lock_->unlock_shared();
672 void reset(RWSpinLock *lock = nullptr) {
673 if (lock_) lock_->unlock_shared();
675 if (lock_) lock_->lock_shared();
678 void swap(ReadHolder *other) {
679 std::swap(this->lock_, other->lock_);
686 class WriteHolder : boost::noncopyable {
688 explicit WriteHolder(RWSpinLock *lock = nullptr) : lock_(lock) {
689 if (lock_) lock_->lock();
691 explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) {
692 if (lock_) lock_->lock();
696 if (lock_) lock_->unlock();
699 void reset(RWSpinLock *lock = nullptr) {
700 if (lock == lock_) return;
701 if (lock_) lock_->unlock();
703 if (lock_) lock_->lock();
706 void swap(WriteHolder *other) {
707 std::swap(this->lock_, other->lock_);
711 friend class ReadHolder;
715 // Synchronized<> adaptors.
716 friend void acquireRead(RWTicketSpinLockT& mutex) {
719 friend void acquireReadWrite(RWTicketSpinLockT& mutex) {
722 friend bool acquireReadWrite(RWTicketSpinLockT& mutex,
723 unsigned int milliseconds) {
727 friend void releaseRead(RWTicketSpinLockT& mutex) {
728 mutex.unlock_shared();
730 friend void releaseReadWrite(RWTicketSpinLockT& mutex) {
735 typedef RWTicketSpinLockT<32> RWTicketSpinLock32;
736 typedef RWTicketSpinLockT<64> RWTicketSpinLock64;
738 #endif // RW_SPINLOCK_USE_X86_INTRINSIC_
742 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
743 #undef RW_SPINLOCK_USE_X86_INTRINSIC_
746 #endif // FOLLY_RWSPINLOCK_H_