2 * Copyright 2017-present 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.
16 #include <folly/detail/Futex.h>
17 #include <folly/ScopeGuard.h>
18 #include <folly/hash/Hash.h>
19 #include <folly/portability/SysSyscall.h>
25 #include <folly/synchronization/ParkingLot.h>
28 #include <linux/futex.h>
31 using namespace std::chrono;
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 using Lot = ParkingLot<uint32_t>;
164 int emulatedFutexWake(void* addr, int count, uint32_t waitMask) {
166 parkingLot.unpark(addr, [&](const uint32_t& mask) {
167 if ((mask & waitMask) == 0) {
168 return UnparkControl::RetainContinue;
173 return count > 0 ? UnparkControl::RemoveContinue
174 : UnparkControl::RemoveBreak;
179 template <typename F>
180 FutexResult emulatedFutexWaitImpl(
183 time_point<system_clock>* absSystemTime,
184 time_point<steady_clock>* absSteadyTime,
187 std::is_same<F, Futex<std::atomic>>::value ||
188 std::is_same<F, Futex<EmulatedFutexAtomic>>::value,
189 "Type F must be either Futex<std::atomic> or Futex<EmulatedFutexAtomic>");
192 res = parkingLot.park_until(
195 [&] { return *futex == expected; },
198 } else if (absSteadyTime) {
199 res = parkingLot.park_until(
202 [&] { return *futex == expected; },
206 res = parkingLot.park(
207 futex, waitMask, [&] { return *futex == expected; }, [] {});
210 case ParkResult::Skip:
211 return FutexResult::VALUE_CHANGED;
212 case ParkResult::Unpark:
213 return FutexResult::AWOKEN;
214 case ParkResult::Timeout:
215 return FutexResult::TIMEDOUT;
218 return FutexResult::INTERRUPTED;
223 /////////////////////////////////
224 // Futex<> specializations
228 Futex<std::atomic>::futexWake(int count, uint32_t wakeMask) {
230 return nativeFutexWake(this, count, wakeMask);
232 return emulatedFutexWake(this, count, wakeMask);
238 Futex<EmulatedFutexAtomic>::futexWake(int count, uint32_t wakeMask) {
239 return emulatedFutexWake(this, count, wakeMask);
244 Futex<std::atomic>::futexWaitImpl(uint32_t expected,
245 time_point<system_clock>* absSystemTime,
246 time_point<steady_clock>* absSteadyTime,
249 return nativeFutexWaitImpl(
250 this, expected, absSystemTime, absSteadyTime, waitMask);
252 return emulatedFutexWaitImpl(
253 this, expected, absSystemTime, absSteadyTime, waitMask);
259 Futex<EmulatedFutexAtomic>::futexWaitImpl(
261 time_point<system_clock>* absSystemTime,
262 time_point<steady_clock>* absSteadyTime,
264 return emulatedFutexWaitImpl(
265 this, expected, absSystemTime, absSteadyTime, waitMask);
268 } // namespace detail