X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=folly%2FRWSpinLock.h;h=cc4f480a4e29b123cafd16af71963bbd3b63f03e;hb=cd201c86411f960e6e63c8222a334cea5ed67768;hp=2e7bbba2a27508fc3bba72e5c4aea029e1082245;hpb=99cbb9cec777f20b2f6f5cad92709f1a032f2ad8;p=folly.git diff --git a/folly/RWSpinLock.h b/folly/RWSpinLock.h index 2e7bbba2..cc4f480a 100644 --- a/folly/RWSpinLock.h +++ b/folly/RWSpinLock.h @@ -1,5 +1,5 @@ /* - * Copyright 2015 Facebook, Inc. + * Copyright 2017 Facebook, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -15,6 +15,22 @@ */ /* + * N.B. You most likely do _not_ want to use RWSpinLock or any other + * kind of spinlock. Use SharedMutex instead. + * + * In short, spinlocks in preemptive multi-tasking operating systems + * have serious problems and fast mutexes like SharedMutex are almost + * certainly the better choice, because letting the OS scheduler put a + * thread to sleep is better for system responsiveness and throughput + * than wasting a timeslice repeatedly querying a lock held by a + * thread that's blocked, and you can't prevent userspace + * programs blocking. + * + * Spinlocks in an operating system kernel make much more sense than + * they do in userspace. + * + * ------------------------------------------------------------------- + * * Two Read-Write spin lock implementations. * * Ref: http://locklessinc.com/articles/locks @@ -24,12 +40,14 @@ * are very compact (4/8 bytes), so are suitable for per-instance * based locking, particularly when contention is not expected. * - * In most cases, RWSpinLock is a reasonable choice. It has minimal - * overhead, and comparable contention performance when the number of - * competing threads is less than or equal to the number of logical - * CPUs. Even as the number of threads gets larger, RWSpinLock can - * still be very competitive in READ, although it is slower on WRITE, - * and also inherently unfair to writers. + * For a spinlock, RWSpinLock is a reasonable choice. (See the note + * about for why a spin lock is frequently a bad idea generally.) + * RWSpinLock has minimal overhead, and comparable contention + * performance when the number of competing threads is less than or + * equal to the number of logical CPUs. Even as the number of + * threads gets larger, RWSpinLock can still be very competitive in + * READ, although it is slower on WRITE, and also inherently unfair + * to writers. * * RWTicketSpinLock shows more balanced READ/WRITE performance. If * your application really needs a lot more threads, and a @@ -63,8 +81,7 @@ * @author Xin Liu */ -#ifndef FOLLY_RWSPINLOCK_H_ -#define FOLLY_RWSPINLOCK_H_ +#pragma once /* ======================================================================== @@ -119,6 +136,7 @@ pthread_rwlock_t Read 728698 24us 101ns 7.28ms 194us */ #include +#include #if defined(__GNUC__) && \ (defined(__i386) || FOLLY_X64 || \ @@ -132,17 +150,17 @@ pthread_rwlock_t Read 728698 24us 101ns 7.28ms 194us #endif // iOS doesn't define _mm_cvtsi64_si128 and friends -#if (FOLLY_SSE >= 2) && !TARGET_OS_IPHONE +#if (FOLLY_SSE >= 2) && !FOLLY_MOBILE #define RW_SPINLOCK_USE_SSE_INSTRUCTIONS_ #else #undef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_ #endif +#include #include #include -#include +#include -#include #include #include @@ -176,7 +194,7 @@ class RWSpinLock { void lock() { int count = 0; while (!LIKELY(try_lock())) { - if (++count > 1000) sched_yield(); + if (++count > 1000) std::this_thread::yield(); } } @@ -190,7 +208,7 @@ class RWSpinLock { void lock_shared() { int count = 0; while (!LIKELY(try_lock_shared())) { - if (++count > 1000) sched_yield(); + if (++count > 1000) std::this_thread::yield(); } } @@ -208,7 +226,7 @@ class RWSpinLock { void lock_upgrade() { int count = 0; while (!try_lock_upgrade()) { - if (++count > 1000) sched_yield(); + if (++count > 1000) std::this_thread::yield(); } } @@ -220,7 +238,7 @@ class RWSpinLock { void unlock_upgrade_and_lock() { int64_t count = 0; while (!try_unlock_upgrade_and_lock()) { - if (++count > 1000) sched_yield(); + if (++count > 1000) std::this_thread::yield(); } } @@ -289,7 +307,7 @@ class RWSpinLock { class ReadHolder { public: - explicit ReadHolder(RWSpinLock* lock = nullptr) : lock_(lock) { + explicit ReadHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock_shared(); } @@ -342,7 +360,7 @@ class RWSpinLock { class UpgradedHolder { public: - explicit UpgradedHolder(RWSpinLock* lock = nullptr) : lock_(lock) { + explicit UpgradedHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock_upgrade(); } @@ -391,7 +409,7 @@ class RWSpinLock { class WriteHolder { public: - explicit WriteHolder(RWSpinLock* lock = nullptr) : lock_(lock) { + explicit WriteHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock(); } @@ -439,12 +457,6 @@ class RWSpinLock { RWSpinLock* lock_; }; - // Synchronized<> adaptors - friend void acquireRead(RWSpinLock& l) { return l.lock_shared(); } - friend void acquireReadWrite(RWSpinLock& l) { return l.lock(); } - friend void releaseRead(RWSpinLock& l) { return l.unlock_shared(); } - friend void releaseReadWrite(RWSpinLock& l) { return l.unlock(); } - private: std::atomic bits_; }; @@ -467,13 +479,14 @@ struct RWTicketIntTrait<64> { #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_ static __m128i make128(const uint16_t v[4]) { - return _mm_set_epi16(0, 0, 0, 0, v[3], v[2], v[1], v[0]); + return _mm_set_epi16(0, 0, 0, 0, + short(v[3]), short(v[2]), short(v[1]), short(v[0])); } static inline __m128i fromInteger(uint64_t from) { - return _mm_cvtsi64_si128(from); + return _mm_cvtsi64_si128(int64_t(from)); } static inline uint64_t toInteger(__m128i in) { - return _mm_cvtsi128_si64(in); + return uint64_t(_mm_cvtsi128_si64(in)); } static inline uint64_t addParallel(__m128i in, __m128i kDelta) { return toInteger(_mm_add_epi16(in, kDelta)); @@ -489,14 +502,17 @@ struct RWTicketIntTrait<32> { #ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_ static __m128i make128(const uint8_t v[4]) { - return _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, - 0, 0, 0, 0, v[3], v[2], v[1], v[0]); + return _mm_set_epi8( + 0, 0, 0, 0, + 0, 0, 0, 0, + 0, 0, 0, 0, + char(v[3]), char(v[2]), char(v[1]), char(v[0])); } static inline __m128i fromInteger(uint32_t from) { - return _mm_cvtsi32_si128(from); + return _mm_cvtsi32_si128(int32_t(from)); } static inline uint32_t toInteger(__m128i in) { - return _mm_cvtsi128_si32(in); + return uint32_t(_mm_cvtsi128_si32(in)); } static inline uint32_t addParallel(__m128i in, __m128i kDelta) { return toInteger(_mm_add_epi8(in, kDelta)); @@ -585,7 +601,7 @@ class RWTicketSpinLockT { * turns. */ void writeLockAggressive() { - // sched_yield() is needed here to avoid a pathology if the number + // std::this_thread::yield() is needed here to avoid a pathology if the number // of threads attempting concurrent writes is >= the number of real // cores allocated to this process. This is less likely than the // corresponding situation in lock_shared(), but we still want to @@ -594,7 +610,7 @@ class RWTicketSpinLockT { QuarterInt val = __sync_fetch_and_add(&ticket.users, 1); while (val != load_acquire(&ticket.write)) { asm_volatile_pause(); - if (UNLIKELY(++count > 1000)) sched_yield(); + if (UNLIKELY(++count > 1000)) std::this_thread::yield(); } } @@ -607,7 +623,7 @@ class RWTicketSpinLockT { // there are a lot of competing readers. The aggressive spinning // can help to avoid starving writers. // - // We don't worry about sched_yield() here because the caller + // We don't worry about std::this_thread::yield() here because the caller // has already explicitly abandoned fairness. while (!try_lock()) {} } @@ -637,13 +653,13 @@ class RWTicketSpinLockT { } void lock_shared() { - // sched_yield() is important here because we can't grab the + // std::this_thread::yield() is important here because we can't grab the // shared lock if there is a pending writeLockAggressive, so we // need to let threads that already have a shared lock complete int count = 0; while (!LIKELY(try_lock_shared())) { asm_volatile_pause(); - if (UNLIKELY((++count & 1023) == 0)) sched_yield(); + if (UNLIKELY((++count & 1023) == 0)) std::this_thread::yield(); } } @@ -676,8 +692,7 @@ class RWTicketSpinLockT { ReadHolder(ReadHolder const&) = delete; ReadHolder& operator=(ReadHolder const&) = delete; - explicit ReadHolder(RWSpinLock *lock = nullptr) : - lock_(lock) { + explicit ReadHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock_shared(); } @@ -716,7 +731,7 @@ class RWTicketSpinLockT { WriteHolder(WriteHolder const&) = delete; WriteHolder& operator=(WriteHolder const&) = delete; - explicit WriteHolder(RWSpinLock *lock = nullptr) : lock_(lock) { + explicit WriteHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock(); } explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) { @@ -742,20 +757,6 @@ class RWTicketSpinLockT { friend class ReadHolder; RWSpinLock *lock_; }; - - // Synchronized<> adaptors. - friend void acquireRead(RWTicketSpinLockT& mutex) { - mutex.lock_shared(); - } - friend void acquireReadWrite(RWTicketSpinLockT& mutex) { - mutex.lock(); - } - friend void releaseRead(RWTicketSpinLockT& mutex) { - mutex.unlock_shared(); - } - friend void releaseReadWrite(RWTicketSpinLockT& mutex) { - mutex.unlock(); - } }; typedef RWTicketSpinLockT<32> RWTicketSpinLock32; @@ -768,5 +769,3 @@ typedef RWTicketSpinLockT<64> RWTicketSpinLock64; #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_ #undef RW_SPINLOCK_USE_X86_INTRINSIC_ #endif - -#endif // FOLLY_RWSPINLOCK_H_