1 // Copyright 2009 The Go Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style 3 // license that can be found in the LICENSE file. 4 5 // Package sync provides basic synchronization primitives such as mutual 6 // exclusion locks. Other than the Once and WaitGroup types, most are intended 7 // for use by low-level library routines. Higher-level synchronization is 8 // better done via channels and communication. 9 // 10 // Values containing the types defined in this package should not be copied. 11 package sync 12 13 import ( 14 "internal/race" 15 "sync/atomic" 16 "unsafe" 17 ) 18 19 func throw(string) // provided by runtime 20 21 // A Mutex is a mutual exclusion lock. 22 // The zero value for a Mutex is an unlocked mutex. 23 // 24 // A Mutex must not be copied after first use. 25 type Mutex struct { 26 state int32 27 sema uint32 28 } 29 30 // A Locker represents an object that can be locked and unlocked. 31 type Locker interface { 32 Lock() 33 Unlock() 34 } 35 36 const ( 37 mutexLocked = 1 << iota // mutex is locked 38 mutexWoken 39 mutexStarving 40 mutexWaiterShift = iota 41 42 // Mutex fairness. 43 // 44 // Mutex can be in 2 modes of operations: normal and starvation. 45 // In normal mode waiters are queued in FIFO order, but a woken up waiter 46 // does not own the mutex and competes with new arriving goroutines over 47 // the ownership. New arriving goroutines have an advantage -- they are 48 // already running on CPU and there can be lots of them, so a woken up 49 // waiter has good chances of losing. In such case it is queued at front 50 // of the wait queue. If a waiter fails to acquire the mutex for more than 1ms, 51 // it switches mutex to the starvation mode. 52 // 53 // In starvation mode ownership of the mutex is directly handed off from 54 // the unlocking goroutine to the waiter at the front of the queue. 55 // New arriving goroutines don't try to acquire the mutex even if it appears 56 // to be unlocked, and don't try to spin. Instead they queue themselves at 57 // the tail of the wait queue. 58 // 59 // If a waiter receives ownership of the mutex and sees that either 60 // (1) it is the last waiter in the queue, or (2) it waited for less than 1 ms, 61 // it switches mutex back to normal operation mode. 62 // 63 // Normal mode has considerably better performance as a goroutine can acquire 64 // a mutex several times in a row even if there are blocked waiters. 65 // Starvation mode is important to prevent pathological cases of tail latency. 66 starvationThresholdNs = 1e6 67 ) 68 69 // Lock locks m. 70 // If the lock is already in use, the calling goroutine 71 // blocks until the mutex is available. 72 func (m *Mutex) Lock() { 73 // Fast path: grab unlocked mutex. 74 if atomic.CompareAndSwapInt32(&m.state, 0, mutexLocked) { 75 if race.Enabled { 76 race.Acquire(unsafe.Pointer(m)) 77 } 78 return 79 } 80 // Slow path (outlined so that the fast path can be inlined) 81 m.lockSlow() 82 } 83 84 func (m *Mutex) lockSlow() { 85 var waitStartTime int64 86 starving := false 87 awoke := false 88 iter := 0 89 old := m.state 90 for { 91 // Don't spin in starvation mode, ownership is handed off to waiters 92 // so we won't be able to acquire the mutex anyway. 93 if old&(mutexLocked|mutexStarving) == mutexLocked && runtime_canSpin(iter) { 94 // Active spinning makes sense. 95 // Try to set mutexWoken flag to inform Unlock 96 // to not wake other blocked goroutines. 97 if !awoke && old&mutexWoken == 0 && old>>mutexWaiterShift != 0 && 98 atomic.CompareAndSwapInt32(&m.state, old, old|mutexWoken) { 99 awoke = true 100 } 101 runtime_doSpin() 102 iter++ 103 old = m.state 104 continue 105 } 106 new := old 107 // Don't try to acquire starving mutex, new arriving goroutines must queue. 108 if old&mutexStarving == 0 { 109 new |= mutexLocked 110 } 111 if old&(mutexLocked|mutexStarving) != 0 { 112 new += 1 << mutexWaiterShift 113 } 114 // The current goroutine switches mutex to starvation mode. 115 // But if the mutex is currently unlocked, don't do the switch. 116 // Unlock expects that starving mutex has waiters, which will not 117 // be true in this case. 118 if starving && old&mutexLocked != 0 { 119 new |= mutexStarving 120 } 121 if awoke { 122 // The goroutine has been woken from sleep, 123 // so we need to reset the flag in either case. 124 if new&mutexWoken == 0 { 125 throw("sync: inconsistent mutex state") 126 } 127 new &^= mutexWoken 128 } 129 if atomic.CompareAndSwapInt32(&m.state, old, new) { 130 if old&(mutexLocked|mutexStarving) == 0 { 131 break // locked the mutex with CAS 132 } 133 // If we were already waiting before, queue at the front of the queue. 134 queueLifo := waitStartTime != 0 135 if waitStartTime == 0 { 136 waitStartTime = runtime_nanotime() 137 } 138 runtime_SemacquireMutex(&m.sema, queueLifo, 1) 139 starving = starving || runtime_nanotime()-waitStartTime > starvationThresholdNs 140 old = m.state 141 if old&mutexStarving != 0 { 142 // If this goroutine was woken and mutex is in starvation mode, 143 // ownership was handed off to us but mutex is in somewhat 144 // inconsistent state: mutexLocked is not set and we are still 145 // accounted as waiter. Fix that. 146 if old&(mutexLocked|mutexWoken) != 0 || old>>mutexWaiterShift == 0 { 147 throw("sync: inconsistent mutex state") 148 } 149 delta := int32(mutexLocked - 1<<mutexWaiterShift) 150 if !starving || old>>mutexWaiterShift == 1 { 151 // Exit starvation mode. 152 // Critical to do it here and consider wait time. 153 // Starvation mode is so inefficient, that two goroutines 154 // can go lock-step infinitely once they switch mutex 155 // to starvation mode. 156 delta -= mutexStarving 157 } 158 atomic.AddInt32(&m.state, delta) 159 break 160 } 161 awoke = true 162 iter = 0 163 } else { 164 old = m.state 165 } 166 } 167 168 if race.Enabled { 169 race.Acquire(unsafe.Pointer(m)) 170 } 171 } 172 173 // Unlock unlocks m. 174 // It is a run-time error if m is not locked on entry to Unlock. 175 // 176 // A locked Mutex is not associated with a particular goroutine. 177 // It is allowed for one goroutine to lock a Mutex and then 178 // arrange for another goroutine to unlock it. 179 func (m *Mutex) Unlock() { 180 if race.Enabled { 181 _ = m.state 182 race.Release(unsafe.Pointer(m)) 183 } 184 185 // Fast path: drop lock bit. 186 new := atomic.AddInt32(&m.state, -mutexLocked) 187 if new != 0 { 188 // Outlined slow path to allow inlining the fast path. 189 // To hide unlockSlow during tracing we skip one extra frame when tracing GoUnblock. 190 m.unlockSlow(new) 191 } 192 } 193 194 func (m *Mutex) unlockSlow(new int32) { 195 if (new+mutexLocked)&mutexLocked == 0 { 196 throw("sync: unlock of unlocked mutex") 197 } 198 if new&mutexStarving == 0 { 199 old := new 200 for { 201 // If there are no waiters or a goroutine has already 202 // been woken or grabbed the lock, no need to wake anyone. 203 // In starvation mode ownership is directly handed off from unlocking 204 // goroutine to the next waiter. We are not part of this chain, 205 // since we did not observe mutexStarving when we unlocked the mutex above. 206 // So get off the way. 207 if old>>mutexWaiterShift == 0 || old&(mutexLocked|mutexWoken|mutexStarving) != 0 { 208 return 209 } 210 // Grab the right to wake someone. 211 new = (old - 1<<mutexWaiterShift) | mutexWoken 212 if atomic.CompareAndSwapInt32(&m.state, old, new) { 213 runtime_Semrelease(&m.sema, false, 1) 214 return 215 } 216 old = m.state 217 } 218 } else { 219 // Starving mode: handoff mutex ownership to the next waiter, and yield 220 // our time slice so that the next waiter can start to run immediately. 221 // Note: mutexLocked is not set, the waiter will set it after wakeup. 222 // But mutex is still considered locked if mutexStarving is set, 223 // so new coming goroutines won't acquire it. 224 runtime_Semrelease(&m.sema, true, 1) 225 } 226 } 227