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 void unlock_shared_and_lock_upgrade() {
211 // write unlock and upgrade lock atomically
212 void unlock_and_lock_upgrade() {
213 // need to do it in two steps here -- as the UPGRADED bit might be OR-ed at
214 // the same time when other threads are trying do try_lock_upgrade().
215 bits_.fetch_or(UPGRADED, std::memory_order_acquire);
216 bits_.fetch_add(-WRITER, std::memory_order_release);
220 // Attempt to acquire writer permission. Return false if we didn't get it.
223 return bits_.compare_exchange_strong(expect, WRITER,
224 std::memory_order_acq_rel);
227 // Try to get reader permission on the lock. This can fail if we
228 // find out someone is a writer.
229 bool try_lock_shared() {
230 // fetch_add is considerably (100%) faster than compare_exchange,
231 // so here we are optimizing for the common (lock success) case.
232 int32_t value = bits_.fetch_add(READER, std::memory_order_acquire);
233 if (UNLIKELY(value & WRITER)) {
234 bits_.fetch_add(-READER, std::memory_order_release);
240 // try to unlock upgrade and write lock atomically
241 bool try_unlock_upgrade_and_lock() {
242 int32_t expect = UPGRADED;
243 return bits_.compare_exchange_strong(expect, WRITER,
244 std::memory_order_acq_rel);
247 // try to acquire an upgradable lock.
248 bool try_lock_upgrade() {
249 int32_t value = bits_.fetch_or(UPGRADED, std::memory_order_acquire);
251 // Note: when failed, we cannot flip the UPGRADED bit back,
252 // as in this case there is either another upgrade lock or a write lock.
253 // If it's a write lock, the bit will get cleared up when that lock's done
255 return ((value & (UPGRADED | WRITER)) == 0);
258 // mainly for debugging purposes.
259 int32_t bits() const { return bits_.load(std::memory_order_acquire); }
262 class UpgradedHolder;
267 explicit ReadHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
268 if (lock_) lock_->lock_shared();
271 explicit ReadHolder(RWSpinLock& lock) : lock_(&lock) {
272 lock_->lock_shared();
275 ReadHolder(ReadHolder&& other) : lock_(other.lock_) {
276 other.lock_ = nullptr;
280 explicit ReadHolder(UpgradedHolder&& upgraded) : lock_(upgraded.lock_) {
281 upgraded.lock_ = nullptr;
282 if (lock_) lock_->unlock_upgrade_and_lock_shared();
285 explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
286 writer.lock_ = nullptr;
287 if (lock_) lock_->unlock_and_lock_shared();
290 ReadHolder& operator=(ReadHolder&& other) {
292 swap(lock_, other.lock_);
296 ReadHolder(const ReadHolder& other) = delete;
297 ReadHolder& operator=(const ReadHolder& other) = delete;
299 ~ReadHolder() { if (lock_) lock_->unlock_shared(); }
301 void reset(RWSpinLock* lock = nullptr) {
302 if (lock == lock_) return;
303 if (lock_) lock_->unlock_shared();
305 if (lock_) lock_->lock_shared();
308 void swap(ReadHolder* other) {
309 std::swap(lock_, other->lock_);
313 friend class UpgradedHolder;
314 friend class WriteHolder;
318 class UpgradedHolder {
320 explicit UpgradedHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
321 if (lock_) lock_->lock_upgrade();
324 explicit UpgradedHolder(RWSpinLock& lock) : lock_(&lock) {
325 lock_->lock_upgrade();
328 explicit UpgradedHolder(ReadHolder&& reader) {
329 lock_ = reader.lock_;
330 reader.lock_ = nullptr;
331 if (lock_) lock_->unlock_shared_and_lock_upgrade();
334 explicit UpgradedHolder(WriteHolder&& writer) {
335 lock_ = writer.lock_;
336 writer.lock_ = nullptr;
337 if (lock_) lock_->unlock_and_lock_upgrade();
340 UpgradedHolder(UpgradedHolder&& other) : lock_(other.lock_) {
341 other.lock_ = nullptr;
344 UpgradedHolder& operator =(UpgradedHolder&& other) {
346 swap(lock_, other.lock_);
350 UpgradedHolder(const UpgradedHolder& other) = delete;
351 UpgradedHolder& operator =(const UpgradedHolder& other) = delete;
353 ~UpgradedHolder() { if (lock_) lock_->unlock_upgrade(); }
355 void reset(RWSpinLock* lock = nullptr) {
356 if (lock == lock_) return;
357 if (lock_) lock_->unlock_upgrade();
359 if (lock_) lock_->lock_upgrade();
362 void swap(UpgradedHolder* other) {
364 swap(lock_, other->lock_);
368 friend class WriteHolder;
369 friend class ReadHolder;
375 explicit WriteHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
376 if (lock_) lock_->lock();
379 explicit WriteHolder(RWSpinLock& lock) : lock_(&lock) {
383 // promoted from an upgrade lock holder
384 explicit WriteHolder(UpgradedHolder&& upgraded) {
385 lock_ = upgraded.lock_;
386 upgraded.lock_ = nullptr;
387 if (lock_) lock_->unlock_upgrade_and_lock();
390 WriteHolder(WriteHolder&& other) : lock_(other.lock_) {
391 other.lock_ = nullptr;
394 WriteHolder& operator =(WriteHolder&& other) {
396 swap(lock_, other.lock_);
400 WriteHolder(const WriteHolder& other) = delete;
401 WriteHolder& operator =(const WriteHolder& other) = delete;
403 ~WriteHolder () { if (lock_) lock_->unlock(); }
405 void reset(RWSpinLock* lock = nullptr) {
406 if (lock == lock_) return;
407 if (lock_) lock_->unlock();
409 if (lock_) lock_->lock();
412 void swap(WriteHolder* other) {
414 swap(lock_, other->lock_);
418 friend class ReadHolder;
419 friend class UpgradedHolder;
423 // Synchronized<> adaptors
424 friend void acquireRead(RWSpinLock& l) { return l.lock_shared(); }
425 friend void acquireReadWrite(RWSpinLock& l) { return l.lock(); }
426 friend void releaseRead(RWSpinLock& l) { return l.unlock_shared(); }
427 friend void releaseReadWrite(RWSpinLock& l) { return l.unlock(); }
430 std::atomic<int32_t> bits_;
434 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
435 // A more balanced Read-Write spin lock implemented based on GCC intrinsics.
438 template <size_t kBitWidth> struct RWTicketIntTrait {
439 static_assert(kBitWidth == 32 || kBitWidth == 64,
440 "bit width has to be either 32 or 64 ");
444 struct RWTicketIntTrait<64> {
445 typedef uint64_t FullInt;
446 typedef uint32_t HalfInt;
447 typedef uint16_t QuarterInt;
450 static __m128i make128(const uint16_t v[4]) {
451 return _mm_set_epi16(0, 0, 0, 0, v[3], v[2], v[1], v[0]);
453 static inline __m128i fromInteger(uint64_t from) {
454 return _mm_cvtsi64_si128(from);
456 static inline uint64_t toInteger(__m128i in) {
457 return _mm_cvtsi128_si64(in);
459 static inline uint64_t addParallel(__m128i in, __m128i kDelta) {
460 return toInteger(_mm_add_epi16(in, kDelta));
466 struct RWTicketIntTrait<32> {
467 typedef uint32_t FullInt;
468 typedef uint16_t HalfInt;
469 typedef uint8_t QuarterInt;
472 static __m128i make128(const uint8_t v[4]) {
473 return _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
474 0, 0, 0, 0, v[3], v[2], v[1], v[0]);
476 static inline __m128i fromInteger(uint32_t from) {
477 return _mm_cvtsi32_si128(from);
479 static inline uint32_t toInteger(__m128i in) {
480 return _mm_cvtsi128_si32(in);
482 static inline uint32_t addParallel(__m128i in, __m128i kDelta) {
483 return toInteger(_mm_add_epi8(in, kDelta));
490 template<size_t kBitWidth, bool kFavorWriter=false>
491 class RWTicketSpinLockT : boost::noncopyable {
492 typedef detail::RWTicketIntTrait<kBitWidth> IntTraitType;
493 typedef typename detail::RWTicketIntTrait<kBitWidth>::FullInt FullInt;
494 typedef typename detail::RWTicketIntTrait<kBitWidth>::HalfInt HalfInt;
495 typedef typename detail::RWTicketIntTrait<kBitWidth>::QuarterInt
501 __extension__ struct {
508 private: // Some x64-specific utilities for atomic access to ticket.
509 template<class T> static T load_acquire(T* addr) {
510 T t = *addr; // acquire barrier
511 asm volatile("" : : : "memory");
516 static void store_release(T* addr, T v) {
517 asm volatile("" : : : "memory");
518 *addr = v; // release barrier
523 RWTicketSpinLockT() {
524 store_release(&ticket.whole, FullInt(0));
529 writeLockAggressive();
536 * Both try_lock and try_lock_shared diverge in our implementation from the
537 * lock algorithm described in the link above.
539 * In the read case, it is undesirable that the readers could wait
540 * for another reader (before increasing ticket.read in the other
541 * implementation). Our approach gives up on
542 * first-come-first-serve, but our benchmarks showed improve
543 * performance for both readers and writers under heavily contended
544 * cases, particularly when the number of threads exceeds the number
547 * We have writeLockAggressive() using the original implementation
548 * for a writer, which gives some advantage to the writer over the
549 * readers---for that path it is guaranteed that the writer will
550 * acquire the lock after all the existing readers exit.
554 FullInt old = t.whole = load_acquire(&ticket.whole);
555 if (t.users != t.write) return false;
557 return __sync_bool_compare_and_swap(&ticket.whole, old, t.whole);
561 * Call this if you want to prioritize writer to avoid starvation.
562 * Unlike writeLockNice, immediately acquires the write lock when
563 * the existing readers (arriving before the writer) finish their
566 void writeLockAggressive() {
567 // sched_yield() is needed here to avoid a pathology if the number
568 // of threads attempting concurrent writes is >= the number of real
569 // cores allocated to this process. This is less likely than the
570 // corresponding situation in lock_shared(), but we still want to
573 QuarterInt val = __sync_fetch_and_add(&ticket.users, 1);
574 while (val != load_acquire(&ticket.write)) {
575 asm volatile("pause");
576 if (UNLIKELY(++count > 1000)) sched_yield();
580 // Call this when the writer should be nicer to the readers.
581 void writeLockNice() {
582 // Here it doesn't cpu-relax the writer.
584 // This is because usually we have many more readers than the
585 // writers, so the writer has less chance to get the lock when
586 // there are a lot of competing readers. The aggressive spinning
587 // can help to avoid starving writers.
589 // We don't worry about sched_yield() here because the caller
590 // has already explicitly abandoned fairness.
591 while (!try_lock()) {}
594 // Atomically unlock the write-lock from writer and acquire the read-lock.
595 void unlock_and_lock_shared() {
596 QuarterInt val = __sync_fetch_and_add(&ticket.read, 1);
599 // Release writer permission on the lock.
602 t.whole = load_acquire(&ticket.whole);
603 FullInt old = t.whole;
606 // SSE2 can reduce the lock and unlock overhead by 10%
607 static const QuarterInt kDeltaBuf[4] = { 1, 1, 0, 0 }; // write/read/user
608 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
609 __m128i m = IntTraitType::fromInteger(old);
610 t.whole = IntTraitType::addParallel(m, kDelta);
615 store_release(&ticket.readWrite, t.readWrite);
619 // sched_yield() is important here because we can't grab the
620 // shared lock if there is a pending writeLockAggressive, so we
621 // need to let threads that already have a shared lock complete
623 while (!LIKELY(try_lock_shared())) {
624 asm volatile("pause");
625 if (UNLIKELY((++count & 1023) == 0)) sched_yield();
629 bool try_lock_shared() {
631 old.whole = t.whole = load_acquire(&ticket.whole);
632 old.users = old.read;
634 // SSE2 may reduce the total lock and unlock overhead by 10%
635 static const QuarterInt kDeltaBuf[4] = { 0, 1, 1, 0 }; // write/read/user
636 static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
637 __m128i m = IntTraitType::fromInteger(old.whole);
638 t.whole = IntTraitType::addParallel(m, kDelta);
643 return __sync_bool_compare_and_swap(&ticket.whole, old.whole, t.whole);
646 void unlock_shared() {
647 QuarterInt val = __sync_fetch_and_add(&ticket.write, 1);
652 typedef RWTicketSpinLockT<kBitWidth, kFavorWriter> RWSpinLock;
653 class ReadHolder : boost::noncopyable {
655 explicit ReadHolder(RWSpinLock *lock = nullptr) :
657 if (lock_) lock_->lock_shared();
660 explicit ReadHolder(RWSpinLock &lock) : lock_ (&lock) {
661 if (lock_) lock_->lock_shared();
664 // atomically unlock the write-lock from writer and acquire the read-lock
665 explicit ReadHolder(WriteHolder *writer) : lock_(nullptr) {
666 std::swap(this->lock_, writer->lock_);
668 lock_->unlock_and_lock_shared();
673 if (lock_) lock_->unlock_shared();
676 void reset(RWSpinLock *lock = nullptr) {
677 if (lock_) lock_->unlock_shared();
679 if (lock_) lock_->lock_shared();
682 void swap(ReadHolder *other) {
683 std::swap(this->lock_, other->lock_);
690 class WriteHolder : boost::noncopyable {
692 explicit WriteHolder(RWSpinLock *lock = nullptr) : lock_(lock) {
693 if (lock_) lock_->lock();
695 explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) {
696 if (lock_) lock_->lock();
700 if (lock_) lock_->unlock();
703 void reset(RWSpinLock *lock = nullptr) {
704 if (lock == lock_) return;
705 if (lock_) lock_->unlock();
707 if (lock_) lock_->lock();
710 void swap(WriteHolder *other) {
711 std::swap(this->lock_, other->lock_);
715 friend class ReadHolder;
719 // Synchronized<> adaptors.
720 friend void acquireRead(RWTicketSpinLockT& mutex) {
723 friend void acquireReadWrite(RWTicketSpinLockT& mutex) {
726 friend bool acquireReadWrite(RWTicketSpinLockT& mutex,
727 unsigned int milliseconds) {
731 friend void releaseRead(RWTicketSpinLockT& mutex) {
732 mutex.unlock_shared();
734 friend void releaseReadWrite(RWTicketSpinLockT& mutex) {
739 typedef RWTicketSpinLockT<32> RWTicketSpinLock32;
740 typedef RWTicketSpinLockT<64> RWTicketSpinLock64;
742 #endif // RW_SPINLOCK_USE_X86_INTRINSIC_
746 #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
747 #undef RW_SPINLOCK_USE_X86_INTRINSIC_
750 #endif // FOLLY_RWSPINLOCK_H_