2 * Copyright 2012 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 #if defined(__GNUC__) && (defined(__i386) || defined(__x86_64__) || \
111 #define RW_SPINLOCK_USE_X86_INTRINSIC_
112 #include <x86intrin.h>
114 #undef RW_SPINLOCK_USE_X86_INTRINSIC_
120 #include <boost/noncopyable.hpp>
123 #include <glog/logging.h>
125 #include "folly/Likely.h"
130 * A simple, small (4-bytes), but unfair rwlock. Use it when you want
131 * a nice writer and don't expect a lot of write/read contention, or
132 * when you need small rwlocks since you are creating a large number
135 * Note that the unfairness here is extreme: if the lock is
136 * continually accessed for read, writers will never get a chance. If
137 * the lock can be that highly contended this class is probably not an
138 * ideal choice anyway.
140 * It currently implements most of the Lockable, SharedLockable and
141 * UpgradeLockable concepts except the TimedLockable related locking/unlocking
144 class RWSpinLock : boost::noncopyable {
145 enum : int32_t { READER = 4, UPGRADED = 2, WRITER = 1 };
147 RWSpinLock() : bits_(0) {}
152 while (!LIKELY(try_lock())) {
153 if (++count > 1000) sched_yield();
157 // Writer is responsible for clearing up both the UPGRADED and WRITER bits.
159 static_assert(READER > WRITER + UPGRADED, "wrong bits!");
160 bits_.fetch_and(~(WRITER | UPGRADED), std::memory_order_release);
163 // SharedLockable Concept
166 while (!LIKELY(try_lock_shared())) {
167 if (++count > 1000) sched_yield();
171 void unlock_shared() {
172 bits_.fetch_add(-READER, std::memory_order_release);
175 // Downgrade the lock from writer status to reader status.
176 void unlock_and_lock_shared() {
177 bits_.fetch_add(READER, std::memory_order_acquire);
181 // UpgradeLockable Concept
182 void lock_upgrade() {
184 while (!try_lock_upgrade()) {
185 if (++count > 1000) sched_yield();
189 void unlock_upgrade() {
190 bits_.fetch_add(-UPGRADED, std::memory_order_acq_rel);
193 // unlock upgrade and try to acquire write lock
194 void unlock_upgrade_and_lock() {
196 while (!try_unlock_upgrade_and_lock()) {
197 if (++count > 1000) sched_yield();
201 // unlock upgrade and read lock atomically
202 void unlock_upgrade_and_lock_shared() {
203 bits_.fetch_add(READER - UPGRADED, std::memory_order_acq_rel);
206 // write unlock and upgrade lock atomically
207 void unlock_and_lock_upgrade() {
208 // need to do it in two steps here -- as the UPGRADED bit might be OR-ed at
209 // the same time when other threads are trying do try_lock_upgrade().
210 bits_.fetch_or(UPGRADED, std::memory_order_acquire);
211 bits_.fetch_add(-WRITER, std::memory_order_release);
215 // Attempt to acquire writer permission. Return false if we didn't get it.
218 return bits_.compare_exchange_strong(expect, WRITER,
219 std::memory_order_acq_rel);
222 // Try to get reader permission on the lock. This can fail if we
223 // find out someone is a writer or upgrader.
224 // Setting the UPGRADED bit would allow a writer-to-be to indicate
225 // its intention to write and block any new readers while waiting
226 // for existing readers to finish and release their read locks. This
227 // helps avoid starving writers (promoted from upgraders).
228 bool try_lock_shared() {
229 // fetch_add is considerably (100%) faster than compare_exchange,
230 // so here we are optimizing for the common (lock success) case.
231 int32_t value = bits_.fetch_add(READER, std::memory_order_acquire);
232 if (UNLIKELY(value & (WRITER|UPGRADED))) {
233 bits_.fetch_add(-READER, std::memory_order_release);
239 // try to unlock upgrade and write lock atomically
240 bool try_unlock_upgrade_and_lock() {
241 int32_t expect = UPGRADED;
242 return bits_.compare_exchange_strong(expect, WRITER,
243 std::memory_order_acq_rel);
246 // try to acquire an upgradable lock.
247 bool try_lock_upgrade() {
248 int32_t value = bits_.fetch_or(UPGRADED, std::memory_order_acquire);
250 // Note: when failed, we cannot flip the UPGRADED bit back,
251 // as in this case there is either another upgrade lock or a write lock.
252 // If it's a write lock, the bit will get cleared up when that lock's done
254 return ((value & (UPGRADED | WRITER)) == 0);
257 // mainly for debugging purposes.
258 int32_t bits() const { return bits_.load(std::memory_order_acquire); }
261 class UpgradedHolder;
266 explicit ReadHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
267 if (lock_) lock_->lock_shared();
270 explicit ReadHolder(RWSpinLock& lock) : lock_(&lock) {
271 lock_->lock_shared();
274 ReadHolder(ReadHolder&& other) : lock_(other.lock_) {
275 other.lock_ = nullptr;
279 explicit ReadHolder(UpgradedHolder&& upgraded) : lock_(upgraded.lock_) {
280 upgraded.lock_ = nullptr;
281 if (lock_) lock_->unlock_upgrade_and_lock_shared();
284 explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
285 writer.lock_ = nullptr;
286 if (lock_) lock_->unlock_and_lock_shared();
289 ReadHolder& operator=(ReadHolder&& other) {
291 swap(lock_, other.lock_);
295 ReadHolder(const ReadHolder& other) = delete;
296 ReadHolder& operator=(const ReadHolder& other) = delete;
298 ~ReadHolder() { if (lock_) lock_->unlock_shared(); }
300 void reset(RWSpinLock* lock = nullptr) {
301 if (lock == lock_) return;
302 if (lock_) lock_->unlock_shared();
304 if (lock_) lock_->lock_shared();
307 void swap(ReadHolder* other) {
308 std::swap(lock_, other->lock_);
312 friend class UpgradedHolder;
313 friend class WriteHolder;
317 class UpgradedHolder {
319 explicit UpgradedHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
320 if (lock_) lock_->lock_upgrade();
323 explicit UpgradedHolder(RWSpinLock& lock) : lock_(&lock) {
324 lock_->lock_upgrade();
327 explicit UpgradedHolder(WriteHolder&& writer) {
328 lock_ = writer.lock_;
329 writer.lock_ = nullptr;
330 if (lock_) lock_->unlock_and_lock_upgrade();
333 UpgradedHolder(UpgradedHolder&& other) : lock_(other.lock_) {
334 other.lock_ = nullptr;
337 UpgradedHolder& operator =(UpgradedHolder&& other) {
339 swap(lock_, other.lock_);
343 UpgradedHolder(const UpgradedHolder& other) = delete;
344 UpgradedHolder& operator =(const UpgradedHolder& other) = delete;
346 ~UpgradedHolder() { if (lock_) lock_->unlock_upgrade(); }
348 void reset(RWSpinLock* lock = nullptr) {
349 if (lock == lock_) return;
350 if (lock_) lock_->unlock_upgrade();
352 if (lock_) lock_->lock_upgrade();
355 void swap(UpgradedHolder* other) {
357 swap(lock_, other->lock_);
361 friend class WriteHolder;
362 friend class ReadHolder;
368 explicit WriteHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
369 if (lock_) lock_->lock();
372 explicit WriteHolder(RWSpinLock& lock) : lock_(&lock) {
376 // promoted from an upgrade lock holder
377 explicit WriteHolder(UpgradedHolder&& upgraded) {
378 lock_ = upgraded.lock_;
379 upgraded.lock_ = nullptr;
380 if (lock_) lock_->unlock_upgrade_and_lock();
383 WriteHolder(WriteHolder&& other) : lock_(other.lock_) {
384 other.lock_ = nullptr;
387 WriteHolder& operator =(WriteHolder&& other) {
389 swap(lock_, other.lock_);
393 WriteHolder(const WriteHolder& other) = delete;
394 WriteHolder& operator =(const WriteHolder& other) = delete;
396 ~WriteHolder () { if (lock_) lock_->unlock(); }
398 void reset(RWSpinLock* lock = nullptr) {
399 if (lock == lock_) return;
400 if (lock_) lock_->unlock();
402 if (lock_) lock_->lock();
405 void swap(WriteHolder* other) {
407 swap(lock_, other->lock_);
411 friend class ReadHolder;
412 friend class UpgradedHolder;
416 // Synchronized<> adaptors
417 friend void acquireRead(RWSpinLock& l) { return l.lock_shared(); }
418 friend void acquireReadWrite(RWSpinLock& l) { return l.lock(); }
419 friend void releaseRead(RWSpinLock& l) { return l.unlock_shared(); }
420 friend void releaseReadWrite(RWSpinLock& l) { return l.unlock(); }
423 std::atomic<int32_t> bits_;
427 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
428 // A more balanced Read-Write spin lock implemented based on GCC intrinsics.
431 template <size_t kBitWidth> struct RWTicketIntTrait {
432 static_assert(kBitWidth == 32 || kBitWidth == 64,
433 "bit width has to be either 32 or 64 ");
437 struct RWTicketIntTrait<64> {
438 typedef uint64_t FullInt;
439 typedef uint32_t HalfInt;
440 typedef uint16_t QuarterInt;
443 static __m128i make128(const uint16_t v[4]) {
444 return _mm_set_epi16(0, 0, 0, 0, v[3], v[2], v[1], v[0]);
446 static inline __m128i fromInteger(uint64_t from) {
447 return _mm_cvtsi64_si128(from);
449 static inline uint64_t toInteger(__m128i in) {
450 return _mm_cvtsi128_si64(in);
452 static inline uint64_t addParallel(__m128i in, __m128i kDelta) {
453 return toInteger(_mm_add_epi16(in, kDelta));
459 struct RWTicketIntTrait<32> {
460 typedef uint32_t FullInt;
461 typedef uint16_t HalfInt;
462 typedef uint8_t QuarterInt;
465 static __m128i make128(const uint8_t v[4]) {
466 return _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
467 0, 0, 0, 0, v[3], v[2], v[1], v[0]);
469 static inline __m128i fromInteger(uint32_t from) {
470 return _mm_cvtsi32_si128(from);
472 static inline uint32_t toInteger(__m128i in) {
473 return _mm_cvtsi128_si32(in);
475 static inline uint32_t addParallel(__m128i in, __m128i kDelta) {
476 return toInteger(_mm_add_epi8(in, kDelta));
483 template<size_t kBitWidth, bool kFavorWriter=false>
484 class RWTicketSpinLockT : boost::noncopyable {
485 typedef detail::RWTicketIntTrait<kBitWidth> IntTraitType;
486 typedef typename detail::RWTicketIntTrait<kBitWidth>::FullInt FullInt;
487 typedef typename detail::RWTicketIntTrait<kBitWidth>::HalfInt HalfInt;
488 typedef typename detail::RWTicketIntTrait<kBitWidth>::QuarterInt
494 __extension__ struct {
501 private: // Some x64-specific utilities for atomic access to ticket.
502 template<class T> static T load_acquire(T* addr) {
503 T t = *addr; // acquire barrier
504 asm volatile("" : : : "memory");
509 static void store_release(T* addr, T v) {
510 asm volatile("" : : : "memory");
511 *addr = v; // release barrier
516 RWTicketSpinLockT() {
517 store_release(&ticket.whole, FullInt(0));
522 writeLockAggressive();
529 * Both try_lock and try_lock_shared diverge in our implementation from the
530 * lock algorithm described in the link above.
532 * In the read case, it is undesirable that the readers could wait
533 * for another reader (before increasing ticket.read in the other
534 * implementation). Our approach gives up on
535 * first-come-first-serve, but our benchmarks showed improve
536 * performance for both readers and writers under heavily contended
537 * cases, particularly when the number of threads exceeds the number
540 * We have writeLockAggressive() using the original implementation
541 * for a writer, which gives some advantage to the writer over the
542 * readers---for that path it is guaranteed that the writer will
543 * acquire the lock after all the existing readers exit.
547 FullInt old = t.whole = load_acquire(&ticket.whole);
548 if (t.users != t.write) return false;
550 return __sync_bool_compare_and_swap(&ticket.whole, old, t.whole);
554 * Call this if you want to prioritize writer to avoid starvation.
555 * Unlike writeLockNice, immediately acquires the write lock when
556 * the existing readers (arriving before the writer) finish their
559 void writeLockAggressive() {
560 // sched_yield() is needed here to avoid a pathology if the number
561 // of threads attempting concurrent writes is >= the number of real
562 // cores allocated to this process. This is less likely than the
563 // corresponding situation in lock_shared(), but we still want to
566 QuarterInt val = __sync_fetch_and_add(&ticket.users, 1);
567 while (val != load_acquire(&ticket.write)) {
568 asm volatile("pause");
569 if (UNLIKELY(++count > 1000)) sched_yield();
573 // Call this when the writer should be nicer to the readers.
574 void writeLockNice() {
575 // Here it doesn't cpu-relax the writer.
577 // This is because usually we have many more readers than the
578 // writers, so the writer has less chance to get the lock when
579 // there are a lot of competing readers. The aggressive spinning
580 // can help to avoid starving writers.
582 // We don't worry about sched_yield() here because the caller
583 // has already explicitly abandoned fairness.
584 while (!try_lock()) {}
587 // Atomically unlock the write-lock from writer and acquire the read-lock.
588 void unlock_and_lock_shared() {
589 QuarterInt val = __sync_fetch_and_add(&ticket.read, 1);
592 // Release writer permission on the lock.
595 t.whole = load_acquire(&ticket.whole);
596 FullInt old = t.whole;
599 // SSE2 can reduce the lock and unlock overhead by 10%
600 static const QuarterInt kDeltaBuf[4] = { 1, 1, 0, 0 }; // write/read/user
601 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
602 __m128i m = IntTraitType::fromInteger(old);
603 t.whole = IntTraitType::addParallel(m, kDelta);
608 store_release(&ticket.readWrite, t.readWrite);
612 // sched_yield() is important here because we can't grab the
613 // shared lock if there is a pending writeLockAggressive, so we
614 // need to let threads that already have a shared lock complete
616 while (!LIKELY(try_lock_shared())) {
617 asm volatile("pause");
618 if (UNLIKELY((++count & 1023) == 0)) sched_yield();
622 bool try_lock_shared() {
624 old.whole = t.whole = load_acquire(&ticket.whole);
625 old.users = old.read;
627 // SSE2 may reduce the total lock and unlock overhead by 10%
628 static const QuarterInt kDeltaBuf[4] = { 0, 1, 1, 0 }; // write/read/user
629 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
630 __m128i m = IntTraitType::fromInteger(old.whole);
631 t.whole = IntTraitType::addParallel(m, kDelta);
636 return __sync_bool_compare_and_swap(&ticket.whole, old.whole, t.whole);
639 void unlock_shared() {
640 QuarterInt val = __sync_fetch_and_add(&ticket.write, 1);
645 typedef RWTicketSpinLockT<kBitWidth, kFavorWriter> RWSpinLock;
646 class ReadHolder : boost::noncopyable {
648 explicit ReadHolder(RWSpinLock *lock = nullptr) :
650 if (lock_) lock_->lock_shared();
653 explicit ReadHolder(RWSpinLock &lock) : lock_ (&lock) {
654 if (lock_) lock_->lock_shared();
657 // atomically unlock the write-lock from writer and acquire the read-lock
658 explicit ReadHolder(WriteHolder *writer) : lock_(nullptr) {
659 std::swap(this->lock_, writer->lock_);
661 lock_->unlock_and_lock_shared();
666 if (lock_) lock_->unlock_shared();
669 void reset(RWSpinLock *lock = nullptr) {
670 if (lock_) lock_->unlock_shared();
672 if (lock_) lock_->lock_shared();
675 void swap(ReadHolder *other) {
676 std::swap(this->lock_, other->lock_);
683 class WriteHolder : boost::noncopyable {
685 explicit WriteHolder(RWSpinLock *lock = nullptr) : lock_(lock) {
686 if (lock_) lock_->lock();
688 explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) {
689 if (lock_) lock_->lock();
693 if (lock_) lock_->unlock();
696 void reset(RWSpinLock *lock = nullptr) {
697 if (lock == lock_) return;
698 if (lock_) lock_->unlock();
700 if (lock_) lock_->lock();
703 void swap(WriteHolder *other) {
704 std::swap(this->lock_, other->lock_);
708 friend class ReadHolder;
712 // Synchronized<> adaptors.
713 friend void acquireRead(RWTicketSpinLockT& mutex) {
716 friend void acquireReadWrite(RWTicketSpinLockT& mutex) {
719 friend bool acquireReadWrite(RWTicketSpinLockT& mutex,
720 unsigned int milliseconds) {
724 friend void releaseRead(RWTicketSpinLockT& mutex) {
725 mutex.unlock_shared();
727 friend void releaseReadWrite(RWTicketSpinLockT& mutex) {
732 typedef RWTicketSpinLockT<32> RWTicketSpinLock32;
733 typedef RWTicketSpinLockT<64> RWTicketSpinLock64;
735 #endif // RW_SPINLOCK_USE_X86_INTRINSIC_
739 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
740 #undef RW_SPINLOCK_USE_X86_INTRINSIC_
743 #endif // FOLLY_RWSPINLOCK_H_