2 * Copyright 2016 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.
24 #include <folly/detail/Futex.h>
25 #include <folly/detail/MemoryIdler.h>
26 #include <folly/portability/Asm.h>
30 /// A Baton allows a thread to block once and be awoken: it captures
31 /// a single handoff. During its lifecycle (from construction/reset to
32 /// destruction/reset) a baton must either be post()ed and wait()ed exactly
33 /// once each, or not at all.
35 /// Baton includes no internal padding, and is only 4 bytes in size.
36 /// Any alignment or padding to avoid false sharing is up to the user.
38 /// This is basically a stripped-down semaphore that supports only a
39 /// single call to sem_post and a single call to sem_wait. The current
40 /// posix semaphore sem_t isn't too bad, but this provides more a bit more
41 /// speed, inlining, smaller size, a guarantee that the implementation
42 /// won't change, and compatibility with DeterministicSchedule. By having
43 /// a much more restrictive lifecycle we can also add a bunch of assertions
44 /// that can help to catch race conditions ahead of time.
45 template <template<typename> class Atom = std::atomic>
47 constexpr Baton() : state_(INIT) {}
49 Baton(Baton const&) = delete;
50 Baton& operator=(Baton const&) = delete;
52 /// It is an error to destroy a Baton on which a thread is currently
53 /// wait()ing. In practice this means that the waiter usually takes
54 /// responsibility for destroying the Baton.
56 // The docblock for this function says that it can't be called when
57 // there is a concurrent waiter. We assume a strong version of this
58 // requirement in which the caller must _know_ that this is true, they
59 // are not allowed to be merely lucky. If two threads are involved,
60 // the destroying thread must actually have synchronized with the
61 // waiting thread after wait() returned. To convey causality the the
62 // waiting thread must have used release semantics and the destroying
63 // thread must have used acquire semantics for that communication,
64 // so we are guaranteed to see the post-wait() value of state_,
65 // which cannot be WAITING.
67 // Note that since we only care about a single memory location,
68 // the only two plausible memory orders here are relaxed and seq_cst.
69 assert(state_.load(std::memory_order_relaxed) != WAITING);
72 /// Equivalent to destroying the Baton and creating a new one. It is
73 /// a bug to call this while there is a waiting thread, so in practice
74 /// the waiter will be the one that resets the baton.
76 // See ~Baton for a discussion about why relaxed is okay here
77 assert(state_.load(std::memory_order_relaxed) != WAITING);
79 // We use a similar argument to justify the use of a relaxed store
80 // here. Since both wait() and post() are required to be called
81 // only once per lifetime, no thread can actually call those methods
82 // correctly after a reset() unless it synchronizes with the thread
83 // that performed the reset(). If a post() or wait() on another thread
84 // didn't synchronize, then regardless of what operation we performed
85 // here there would be a race on proper use of the Baton's spec
86 // (although not on any particular load and store). Put another way,
87 // we don't need to synchronize here because anybody that might rely
88 // on such synchronization is required by the baton rules to perform
89 // an additional synchronization that has the desired effect anyway.
91 // There is actually a similar argument to be made about the
92 // constructor, in which the fenceless constructor initialization
93 // of state_ is piggybacked on whatever synchronization mechanism
94 // distributes knowledge of the Baton's existence
95 state_.store(INIT, std::memory_order_relaxed);
98 /// Causes wait() to wake up. For each lifetime of a Baton (where a
99 /// lifetime starts at construction or reset() and ends at destruction
100 /// or reset()) there can be at most one call to post(). Any thread
103 /// Although we could implement a more generic semaphore semantics
104 /// without any extra size or CPU overhead, the single-call limitation
105 /// allows us to have better assert-ions during debug builds.
107 uint32_t before = state_.load(std::memory_order_acquire);
109 assert(before == INIT || before == WAITING || before == TIMED_OUT);
111 if (before == INIT &&
112 state_.compare_exchange_strong(before, EARLY_DELIVERY)) {
116 assert(before == WAITING || before == TIMED_OUT);
118 if (before == TIMED_OUT) {
122 assert(before == WAITING);
123 state_.store(LATE_DELIVERY, std::memory_order_release);
127 /// Waits until post() has been called in the current Baton lifetime.
128 /// May be called at most once during a Baton lifetime (construction
129 /// |reset until destruction|reset). If post is called before wait in
130 /// the current lifetime then this method returns immediately.
132 /// The restriction that there can be at most one wait() per lifetime
133 /// could be relaxed somewhat without any perf or size regressions,
134 /// but by making this condition very restrictive we can provide better
135 /// checking in debug builds.
137 if (spinWaitForEarlyDelivery()) {
138 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
142 // guess we have to block :(
143 uint32_t expected = INIT;
144 if (!state_.compare_exchange_strong(expected, WAITING)) {
145 // CAS failed, last minute reprieve
146 assert(expected == EARLY_DELIVERY);
151 detail::MemoryIdler::futexWait(state_, WAITING);
153 // state_ is the truth even if FUTEX_WAIT reported a matching
154 // FUTEX_WAKE, since we aren't using type-stable storage and we
155 // don't guarantee reuse. The scenario goes like this: thread
156 // A's last touch of a Baton is a call to wake(), which stores
157 // LATE_DELIVERY and gets an unlucky context switch before delivering
158 // the corresponding futexWake. Thread B sees LATE_DELIVERY
159 // without consuming a futex event, because it calls futexWait
160 // with an expected value of WAITING and hence doesn't go to sleep.
161 // B returns, so the Baton's memory is reused and becomes another
162 // Baton (or a reuse of this one). B calls futexWait on the new
163 // Baton lifetime, then A wakes up and delivers a spurious futexWake
164 // to the same memory location. B's futexWait will then report a
165 // consumed wake event even though state_ is still WAITING.
167 // It would be possible to add an extra state_ dance to communicate
168 // that the futexWake has been sent so that we can be sure to consume
169 // it before returning, but that would be a perf and complexity hit.
170 uint32_t s = state_.load(std::memory_order_acquire);
171 assert(s == WAITING || s == LATE_DELIVERY);
173 if (s == LATE_DELIVERY) {
180 /// Similar to wait, but with a timeout. The thread is unblocked if the
182 /// Note: Only a single call to timed_wait/wait is allowed during a baton's
183 /// life-cycle (from construction/reset to destruction/reset). In other
184 /// words, after timed_wait the caller can't invoke wait/timed_wait/try_wait
185 /// again on the same baton without resetting it.
187 /// @param deadline Time until which the thread can block
188 /// @return true if the baton was posted to before timeout,
190 template <typename Clock, typename Duration = typename Clock::duration>
191 bool timed_wait(const std::chrono::time_point<Clock,Duration>& deadline) {
192 if (spinWaitForEarlyDelivery()) {
193 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
197 // guess we have to block :(
198 uint32_t expected = INIT;
199 if (!state_.compare_exchange_strong(expected, WAITING)) {
200 // CAS failed, last minute reprieve
201 assert(expected == EARLY_DELIVERY);
206 auto rv = state_.futexWaitUntil(WAITING, deadline);
207 if (rv == folly::detail::FutexResult::TIMEDOUT) {
208 state_.store(TIMED_OUT, std::memory_order_release);
212 uint32_t s = state_.load(std::memory_order_acquire);
213 assert(s == WAITING || s == LATE_DELIVERY);
214 if (s == LATE_DELIVERY) {
220 /// Similar to timed_wait, but with a duration.
221 template <typename Clock = std::chrono::steady_clock, typename Duration>
222 bool timed_wait(const Duration& duration) {
223 auto deadline = Clock::now() + duration;
224 return timed_wait(deadline);
227 /// Similar to wait, but doesn't block the thread if it hasn't been posted.
229 /// try_wait has the following semantics:
230 /// - It is ok to call try_wait any number times on the same baton until
231 /// try_wait reports that the baton has been posted.
232 /// - It is ok to call timed_wait or wait on the same baton if try_wait
233 /// reports that baton hasn't been posted.
234 /// - If try_wait indicates that the baton has been posted, it is invalid to
235 /// call wait, try_wait or timed_wait on the same baton without resetting
237 /// @return true if baton has been posted, false othewise
239 auto s = state_.load(std::memory_order_acquire);
240 assert(s == INIT || s == EARLY_DELIVERY);
241 return s == EARLY_DELIVERY;
245 enum State : uint32_t {
254 // Must be positive. If multiple threads are actively using a
255 // higher-level data structure that uses batons internally, it is
256 // likely that the post() and wait() calls happen almost at the same
257 // time. In this state, we lose big 50% of the time if the wait goes
258 // to sleep immediately. On circa-2013 devbox hardware it costs about
259 // 7 usec to FUTEX_WAIT and then be awoken (half the t/iter as the
260 // posix_sem_pingpong test in BatonTests). We can improve our chances
261 // of EARLY_DELIVERY by spinning for a bit, although we have to balance
262 // this against the loss if we end up sleeping any way. Spins on this
263 // hw take about 7 nanos (all but 0.5 nanos is the pause instruction).
264 // We give ourself 300 spins, which is about 2 usec of waiting. As a
265 // partial consolation, since we are using the pause instruction we
266 // are giving a speed boost to the colocated hyperthread.
267 PreBlockAttempts = 300,
270 // Spin for "some time" (see discussion on PreBlockAttempts) waiting
273 // @return true if we received an early delivery during the wait,
274 // false otherwise. If the function returns true then
275 // state_ is guaranteed to be EARLY_DELIVERY
276 bool spinWaitForEarlyDelivery() {
278 static_assert(PreBlockAttempts > 0,
279 "isn't this assert clearer than an uninitialized variable warning?");
280 for (int i = 0; i < PreBlockAttempts; ++i) {
285 // The pause instruction is the polite way to spin, but it doesn't
286 // actually affect correctness to omit it if we don't have it.
287 // Pausing donates the full capabilities of the current core to
288 // its other hyperthreads for a dozen cycles or so
289 asm_volatile_pause();
295 detail::Futex<Atom> state_;