2 * Copyright 2017 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.
17 #include <folly/detail/Futex.h>
20 #include <condition_variable>
22 #include <boost/intrusive/list.hpp>
23 #include <folly/Hash.h>
24 #include <folly/ScopeGuard.h>
28 # include <linux/futex.h>
29 # include <sys/syscall.h>
32 using namespace std::chrono;
34 namespace folly { namespace detail {
38 ////////////////////////////////////////////////////
39 // native implementation using the futex() syscall
43 /// Certain toolchains (like Android's) don't include the full futex API in
44 /// their headers even though they support it. Make sure we have our constants
45 /// even if the headers don't have them.
46 #ifndef FUTEX_WAIT_BITSET
47 # define FUTEX_WAIT_BITSET 9
49 #ifndef FUTEX_WAKE_BITSET
50 # define FUTEX_WAKE_BITSET 10
52 #ifndef FUTEX_PRIVATE_FLAG
53 # define FUTEX_PRIVATE_FLAG 128
55 #ifndef FUTEX_CLOCK_REALTIME
56 # define FUTEX_CLOCK_REALTIME 256
59 int nativeFutexWake(void* addr, int count, uint32_t wakeMask) {
60 int rv = syscall(__NR_futex,
62 FUTEX_WAKE_BITSET | FUTEX_PRIVATE_FLAG, /* op */
64 nullptr, /* timeout */
68 /* NOTE: we ignore errors on wake for the case of a futex
69 guarding its own destruction, similar to this
70 glibc bug with sem_post/sem_wait:
71 https://sourceware.org/bugzilla/show_bug.cgi?id=12674 */
78 template <class Clock>
80 timeSpecFromTimePoint(time_point<Clock> absTime)
82 auto epoch = absTime.time_since_epoch();
83 if (epoch.count() < 0) {
84 // kernel timespec_valid requires non-negative seconds and nanos in [0,1G)
85 epoch = Clock::duration::zero();
88 // timespec-safe seconds and nanoseconds;
89 // chrono::{nano,}seconds are `long long int`
90 // whereas timespec uses smaller types
91 using time_t_seconds = duration<std::time_t, seconds::period>;
92 using long_nanos = duration<long int, nanoseconds::period>;
94 auto secs = duration_cast<time_t_seconds>(epoch);
95 auto nanos = duration_cast<long_nanos>(epoch - secs);
96 struct timespec result = { secs.count(), nanos.count() };
100 FutexResult nativeFutexWaitImpl(void* addr,
102 time_point<system_clock>* absSystemTime,
103 time_point<steady_clock>* absSteadyTime,
105 assert(absSystemTime == nullptr || absSteadyTime == nullptr);
107 int op = FUTEX_WAIT_BITSET | FUTEX_PRIVATE_FLAG;
109 struct timespec* timeout = nullptr;
111 if (absSystemTime != nullptr) {
112 op |= FUTEX_CLOCK_REALTIME;
113 ts = timeSpecFromTimePoint(*absSystemTime);
115 } else if (absSteadyTime != nullptr) {
116 ts = timeSpecFromTimePoint(*absSteadyTime);
120 // Unlike FUTEX_WAIT, FUTEX_WAIT_BITSET requires an absolute timeout
121 // value - http://locklessinc.com/articles/futex_cheat_sheet/
122 int rv = syscall(__NR_futex,
126 timeout, /* timeout */
128 waitMask); /* val3 */
131 return FutexResult::AWOKEN;
135 assert(timeout != nullptr);
136 return FutexResult::TIMEDOUT;
138 return FutexResult::INTERRUPTED;
140 return FutexResult::VALUE_CHANGED;
143 // EINVAL, EACCESS, or EFAULT. EINVAL means there was an invalid
144 // op (should be impossible) or an invalid timeout (should have
145 // been sanitized by timeSpecFromTimePoint). EACCESS or EFAULT
146 // means *addr points to invalid memory, which is unlikely because
147 // the caller should have segfaulted already. We can either
148 // crash, or return a value that lets the process continue for
149 // a bit. We choose the latter. VALUE_CHANGED probably turns the
150 // caller into a spin lock.
151 return FutexResult::VALUE_CHANGED;
158 ///////////////////////////////////////////////////////
159 // compatibility implementation using standard C++ API
161 // Our emulated futex uses 4096 lists of wait nodes. There are two levels
162 // of locking: a per-list mutex that controls access to the list and a
163 // per-node mutex, condvar, and bool that are used for the actual wakeups.
164 // The per-node mutex allows us to do precise wakeups without thundering
167 struct EmulatedFutexWaitNode : public boost::intrusive::list_base_hook<> {
169 const uint32_t waitMask_;
171 // tricky: hold both bucket and node mutex to write, either to read
174 std::condition_variable cond_;
176 EmulatedFutexWaitNode(void* addr, uint32_t waitMask)
178 , waitMask_(waitMask)
184 struct EmulatedFutexBucket {
186 boost::intrusive::list<EmulatedFutexWaitNode> waiters_;
188 static const size_t kNumBuckets = 4096;
190 static EmulatedFutexBucket& bucketFor(void* addr) {
191 static auto gBuckets = new EmulatedFutexBucket[kNumBuckets];
192 uint64_t mixedBits = folly::hash::twang_mix64(
193 reinterpret_cast<uintptr_t>(addr));
194 return gBuckets[mixedBits % kNumBuckets];
198 int emulatedFutexWake(void* addr, int count, uint32_t waitMask) {
199 auto& bucket = EmulatedFutexBucket::bucketFor(addr);
200 std::unique_lock<std::mutex> bucketLock(bucket.mutex_);
203 for (auto iter = bucket.waiters_.begin();
204 numAwoken < count && iter != bucket.waiters_.end(); ) {
206 auto& node = *iter++;
207 if (node.addr_ == addr && (node.waitMask_ & waitMask) != 0) {
210 // we unlink, but waiter destroys the node
211 bucket.waiters_.erase(current);
213 std::unique_lock<std::mutex> nodeLock(node.mutex_);
214 node.signaled_ = true;
215 node.cond_.notify_one();
221 template <typename F>
222 FutexResult emulatedFutexWaitImpl(
225 time_point<system_clock>* absSystemTime,
226 time_point<steady_clock>* absSteadyTime,
229 std::is_same<F, Futex<std::atomic>>::value ||
230 std::is_same<F, Futex<EmulatedFutexAtomic>>::value,
231 "Type F must be either Futex<std::atomic> or Futex<EmulatedFutexAtomic>");
232 void* addr = static_cast<void*>(futex);
233 auto& bucket = EmulatedFutexBucket::bucketFor(addr);
234 EmulatedFutexWaitNode node(addr, waitMask);
237 std::unique_lock<std::mutex> bucketLock(bucket.mutex_);
239 if (futex->load(std::memory_order_relaxed) != expected) {
240 return FutexResult::VALUE_CHANGED;
243 bucket.waiters_.push_back(node);
244 } // bucketLock scope
246 std::cv_status status = std::cv_status::no_timeout;
248 std::unique_lock<std::mutex> nodeLock(node.mutex_);
249 while (!node.signaled_ && status != std::cv_status::timeout) {
250 if (absSystemTime != nullptr) {
251 status = node.cond_.wait_until(nodeLock, *absSystemTime);
252 } else if (absSteadyTime != nullptr) {
253 status = node.cond_.wait_until(nodeLock, *absSteadyTime);
255 node.cond_.wait(nodeLock);
260 if (status == std::cv_status::timeout) {
261 // it's not really a timeout until we unlink the unsignaled node
262 std::unique_lock<std::mutex> bucketLock(bucket.mutex_);
263 if (!node.signaled_) {
264 bucket.waiters_.erase(bucket.waiters_.iterator_to(node));
265 return FutexResult::TIMEDOUT;
268 return FutexResult::AWOKEN;
274 /////////////////////////////////
275 // Futex<> specializations
279 Futex<std::atomic>::futexWake(int count, uint32_t wakeMask) {
281 return nativeFutexWake(this, count, wakeMask);
283 return emulatedFutexWake(this, count, wakeMask);
289 Futex<EmulatedFutexAtomic>::futexWake(int count, uint32_t wakeMask) {
290 return emulatedFutexWake(this, count, wakeMask);
295 Futex<std::atomic>::futexWaitImpl(uint32_t expected,
296 time_point<system_clock>* absSystemTime,
297 time_point<steady_clock>* absSteadyTime,
300 return nativeFutexWaitImpl(
301 this, expected, absSystemTime, absSteadyTime, waitMask);
303 return emulatedFutexWaitImpl(
304 this, expected, absSystemTime, absSteadyTime, waitMask);
310 Futex<EmulatedFutexAtomic>::futexWaitImpl(
312 time_point<system_clock>* absSystemTime,
313 time_point<steady_clock>* absSteadyTime,
315 return emulatedFutexWaitImpl(
316 this, expected, absSystemTime, absSteadyTime, waitMask);
319 }} // namespace folly::detail