X-Git-Url: http://plrg.eecs.uci.edu/git/?p=folly.git;a=blobdiff_plain;f=folly%2FRWSpinLock.h;h=cc4f480a4e29b123cafd16af71963bbd3b63f03e;hp=99f018e45c9d8d1fc531186938dbb1558c13a9fb;hb=9883bc3793fc71b8d4c8d1adf6c6fe796e9131e5;hpb=1805d922175a0cb620461f8873f6a7061a683413 diff --git a/folly/RWSpinLock.h b/folly/RWSpinLock.h index 99f018e4..cc4f480a 100644 --- a/folly/RWSpinLock.h +++ b/folly/RWSpinLock.h @@ -1,5 +1,5 @@ /* - * Copyright 2012 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 @@ -46,13 +64,24 @@ * RWTicketSpinLock<64> only allows up to 2^16 - 1 concurrent * readers and writers. * + * RWTicketSpinLock<..., true> (kFavorWriter = true, that is, strict + * writer priority) is NOT reentrant, even for lock_shared(). + * + * The lock will not grant any new shared (read) accesses while a thread + * attempting to acquire the lock in write mode is blocked. (That is, + * if the lock is held in shared mode by N threads, and a thread attempts + * to acquire it in write mode, no one else can acquire it in shared mode + * until these N threads release the lock and then the blocked thread + * acquires and releases the exclusive lock.) This also applies for + * attempts to reacquire the lock in shared mode by threads that already + * hold it in shared mode, making the lock non-reentrant. + * * RWSpinLock handles 2^30 - 1 concurrent readers. * * @author Xin Liu */ -#ifndef FOLLY_RWSPINLOCK_H_ -#define FOLLY_RWSPINLOCK_H_ +#pragma once /* ======================================================================== @@ -106,23 +135,35 @@ pthread_rwlock_t Read 728698 24us 101ns 7.28ms 194us */ -#if defined(__GNUC__) && (defined(__i386) || defined(__x86_64__) || \ - defined(ARCH_K8)) -#define RW_SPINLOCK_USE_X86_INTRINSIC_ -#include +#include +#include + +#if defined(__GNUC__) && \ + (defined(__i386) || FOLLY_X64 || \ + defined(ARCH_K8)) +# define RW_SPINLOCK_USE_X86_INTRINSIC_ +# include +#elif defined(_MSC_VER) && defined(FOLLY_X64) +# define RW_SPINLOCK_USE_X86_INTRINSIC_ #else -#undef RW_SPINLOCK_USE_X86_INTRINSIC_ +# undef RW_SPINLOCK_USE_X86_INTRINSIC_ #endif +// iOS doesn't define _mm_cvtsi64_si128 and friends +#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 -#include "folly/Likely.h" +#include namespace folly { @@ -141,16 +182,19 @@ namespace folly { * UpgradeLockable concepts except the TimedLockable related locking/unlocking * interfaces. */ -class RWSpinLock : boost::noncopyable { +class RWSpinLock { enum : int32_t { READER = 4, UPGRADED = 2, WRITER = 1 }; public: - RWSpinLock() : bits_(0) {} + constexpr RWSpinLock() : bits_(0) {} + + RWSpinLock(RWSpinLock const&) = delete; + RWSpinLock& operator=(RWSpinLock const&) = delete; // Lockable Concept void lock() { int count = 0; while (!LIKELY(try_lock())) { - if (++count > 1000) sched_yield(); + if (++count > 1000) std::this_thread::yield(); } } @@ -164,7 +208,7 @@ class RWSpinLock : boost::noncopyable { void lock_shared() { int count = 0; while (!LIKELY(try_lock_shared())) { - if (++count > 1000) sched_yield(); + if (++count > 1000) std::this_thread::yield(); } } @@ -182,7 +226,7 @@ class RWSpinLock : boost::noncopyable { void lock_upgrade() { int count = 0; while (!try_lock_upgrade()) { - if (++count > 1000) sched_yield(); + if (++count > 1000) std::this_thread::yield(); } } @@ -194,7 +238,7 @@ class RWSpinLock : boost::noncopyable { 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(); } } @@ -203,11 +247,6 @@ class RWSpinLock : boost::noncopyable { bits_.fetch_add(READER - UPGRADED, std::memory_order_acq_rel); } - void unlock_shared_and_lock_upgrade() { - lock_upgrade(); - unlock_shared(); - } - // write unlock and upgrade lock atomically void unlock_and_lock_upgrade() { // need to do it in two steps here -- as the UPGRADED bit might be OR-ed at @@ -225,12 +264,16 @@ class RWSpinLock : boost::noncopyable { } // Try to get reader permission on the lock. This can fail if we - // find out someone is a writer. + // find out someone is a writer or upgrader. + // Setting the UPGRADED bit would allow a writer-to-be to indicate + // its intention to write and block any new readers while waiting + // for existing readers to finish and release their read locks. This + // helps avoid starving writers (promoted from upgraders). bool try_lock_shared() { // fetch_add is considerably (100%) faster than compare_exchange, // so here we are optimizing for the common (lock success) case. int32_t value = bits_.fetch_add(READER, std::memory_order_acquire); - if (UNLIKELY(value & WRITER)) { + if (UNLIKELY(value & (WRITER|UPGRADED))) { bits_.fetch_add(-READER, std::memory_order_release); return false; } @@ -264,7 +307,7 @@ class RWSpinLock : boost::noncopyable { class ReadHolder { public: - explicit ReadHolder(RWSpinLock* lock = nullptr) : lock_(lock) { + explicit ReadHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock_shared(); } @@ -272,7 +315,7 @@ class RWSpinLock : boost::noncopyable { lock_->lock_shared(); } - ReadHolder(ReadHolder&& other) : lock_(other.lock_) { + ReadHolder(ReadHolder&& other) noexcept : lock_(other.lock_) { other.lock_ = nullptr; } @@ -317,7 +360,7 @@ class RWSpinLock : boost::noncopyable { class UpgradedHolder { public: - explicit UpgradedHolder(RWSpinLock* lock = nullptr) : lock_(lock) { + explicit UpgradedHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock_upgrade(); } @@ -325,19 +368,13 @@ class RWSpinLock : boost::noncopyable { lock_->lock_upgrade(); } - explicit UpgradedHolder(ReadHolder&& reader) { - lock_ = reader.lock_; - reader.lock_ = nullptr; - if (lock_) lock_->unlock_shared_and_lock_upgrade(); - } - explicit UpgradedHolder(WriteHolder&& writer) { lock_ = writer.lock_; writer.lock_ = nullptr; if (lock_) lock_->unlock_and_lock_upgrade(); } - UpgradedHolder(UpgradedHolder&& other) : lock_(other.lock_) { + UpgradedHolder(UpgradedHolder&& other) noexcept : lock_(other.lock_) { other.lock_ = nullptr; } @@ -372,7 +409,7 @@ class RWSpinLock : boost::noncopyable { class WriteHolder { public: - explicit WriteHolder(RWSpinLock* lock = nullptr) : lock_(lock) { + explicit WriteHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock(); } @@ -387,7 +424,7 @@ class RWSpinLock : boost::noncopyable { if (lock_) lock_->unlock_upgrade_and_lock(); } - WriteHolder(WriteHolder&& other) : lock_(other.lock_) { + WriteHolder(WriteHolder&& other) noexcept : lock_(other.lock_) { other.lock_ = nullptr; } @@ -420,12 +457,6 @@ class RWSpinLock : boost::noncopyable { 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_; }; @@ -446,15 +477,16 @@ struct RWTicketIntTrait<64> { typedef uint32_t HalfInt; typedef uint16_t QuarterInt; -#ifdef __SSE2__ +#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)); @@ -468,16 +500,19 @@ struct RWTicketIntTrait<32> { typedef uint16_t HalfInt; typedef uint8_t QuarterInt; -#ifdef __SSE2__ +#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)); @@ -488,7 +523,7 @@ struct RWTicketIntTrait<32> { template -class RWTicketSpinLockT : boost::noncopyable { +class RWTicketSpinLockT { typedef detail::RWTicketIntTrait IntTraitType; typedef typename detail::RWTicketIntTrait::FullInt FullInt; typedef typename detail::RWTicketIntTrait::HalfInt HalfInt; @@ -496,6 +531,7 @@ class RWTicketSpinLockT : boost::noncopyable { QuarterInt; union RWTicket { + constexpr RWTicket() : whole(0) {} FullInt whole; HalfInt readWrite; __extension__ struct { @@ -508,21 +544,22 @@ class RWTicketSpinLockT : boost::noncopyable { private: // Some x64-specific utilities for atomic access to ticket. template static T load_acquire(T* addr) { T t = *addr; // acquire barrier - asm volatile("" : : : "memory"); + asm_volatile_memory(); return t; } template static void store_release(T* addr, T v) { - asm volatile("" : : : "memory"); + asm_volatile_memory(); *addr = v; // release barrier } public: - RWTicketSpinLockT() { - store_release(&ticket.whole, FullInt(0)); - } + constexpr RWTicketSpinLockT() {} + + RWTicketSpinLockT(RWTicketSpinLockT const&) = delete; + RWTicketSpinLockT& operator=(RWTicketSpinLockT const&) = delete; void lock() { if (kFavorWriter) { @@ -564,7 +601,7 @@ class RWTicketSpinLockT : boost::noncopyable { * 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 @@ -572,8 +609,8 @@ class RWTicketSpinLockT : boost::noncopyable { int count = 0; QuarterInt val = __sync_fetch_and_add(&ticket.users, 1); while (val != load_acquire(&ticket.write)) { - asm volatile("pause"); - if (UNLIKELY(++count > 1000)) sched_yield(); + asm_volatile_pause(); + if (UNLIKELY(++count > 1000)) std::this_thread::yield(); } } @@ -586,7 +623,7 @@ class RWTicketSpinLockT : boost::noncopyable { // 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()) {} } @@ -602,7 +639,7 @@ class RWTicketSpinLockT : boost::noncopyable { t.whole = load_acquire(&ticket.whole); FullInt old = t.whole; -#ifdef __SSE2__ +#ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_ // SSE2 can reduce the lock and unlock overhead by 10% static const QuarterInt kDeltaBuf[4] = { 1, 1, 0, 0 }; // write/read/user static const __m128i kDelta = IntTraitType::make128(kDeltaBuf); @@ -616,13 +653,13 @@ class RWTicketSpinLockT : boost::noncopyable { } 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(); + asm_volatile_pause(); + if (UNLIKELY((++count & 1023) == 0)) std::this_thread::yield(); } } @@ -630,7 +667,7 @@ class RWTicketSpinLockT : boost::noncopyable { RWTicket t, old; old.whole = t.whole = load_acquire(&ticket.whole); old.users = old.read; -#ifdef __SSE2__ +#ifdef RW_SPINLOCK_USE_SSE_INSTRUCTIONS_ // SSE2 may reduce the total lock and unlock overhead by 10% static const QuarterInt kDeltaBuf[4] = { 0, 1, 1, 0 }; // write/read/user static const __m128i kDelta = IntTraitType::make128(kDeltaBuf); @@ -650,10 +687,12 @@ class RWTicketSpinLockT : boost::noncopyable { class WriteHolder; typedef RWTicketSpinLockT RWSpinLock; - class ReadHolder : boost::noncopyable { + class ReadHolder { public: - explicit ReadHolder(RWSpinLock *lock = nullptr) : - lock_(lock) { + ReadHolder(ReadHolder const&) = delete; + ReadHolder& operator=(ReadHolder const&) = delete; + + explicit ReadHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock_shared(); } @@ -687,9 +726,12 @@ class RWTicketSpinLockT : boost::noncopyable { RWSpinLock *lock_; }; - class WriteHolder : boost::noncopyable { + class WriteHolder { public: - explicit WriteHolder(RWSpinLock *lock = nullptr) : lock_(lock) { + WriteHolder(WriteHolder const&) = delete; + WriteHolder& operator=(WriteHolder const&) = delete; + + explicit WriteHolder(RWSpinLock* lock) : lock_(lock) { if (lock_) lock_->lock(); } explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) { @@ -715,25 +757,6 @@ class RWTicketSpinLockT : boost::noncopyable { friend class ReadHolder; RWSpinLock *lock_; }; - - // Synchronized<> adaptors. - friend void acquireRead(RWTicketSpinLockT& mutex) { - mutex.lock_shared(); - } - friend void acquireReadWrite(RWTicketSpinLockT& mutex) { - mutex.lock(); - } - friend bool acquireReadWrite(RWTicketSpinLockT& mutex, - unsigned int milliseconds) { - mutex.lock(); - return true; - } - friend void releaseRead(RWTicketSpinLockT& mutex) { - mutex.unlock_shared(); - } - friend void releaseReadWrite(RWTicketSpinLockT& mutex) { - mutex.unlock(); - } }; typedef RWTicketSpinLockT<32> RWTicketSpinLock32; @@ -746,5 +769,3 @@ typedef RWTicketSpinLockT<64> RWTicketSpinLock64; #ifdef RW_SPINLOCK_USE_X86_INTRINSIC_ #undef RW_SPINLOCK_USE_X86_INTRINSIC_ #endif - -#endif // FOLLY_RWSPINLOCK_H_