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.
25 #include <folly/detail/Futex.h>
26 #include <folly/detail/MemoryIdler.h>
27 #include <folly/portability/Asm.h>
31 /// A Baton allows a thread to block once and be awoken. The single
32 /// poster version (with SinglePoster == true) captures a single
33 /// handoff, and during its lifecycle (from construction/reset to
34 /// destruction/reset) a baton must either be post()ed and wait()ed
35 /// exactly once each, or not at all.
37 /// The multi-poster version (SinglePoster == false) allows multiple
38 /// concurrent handoff attempts, the first of which completes the
39 /// handoff and the rest if any are idempotent.
41 /// Baton includes no internal padding, and is only 4 bytes in size.
42 /// Any alignment or padding to avoid false sharing is up to the user.
44 /// This is basically a stripped-down semaphore that supports (only a
45 /// single call to sem_post, when SinglePoster == true) and a single
48 /// The non-blocking version (Blocking == false) provides more speed
49 /// by using only load acquire and store release operations in the
50 /// critical path, at the cost of disallowing blocking and timing out.
52 /// The current posix semaphore sem_t isn't too bad, but this provides
53 /// more a bit more speed, inlining, smaller size, a guarantee that
54 /// the implementation won't change, and compatibility with
55 /// DeterministicSchedule. By having a much more restrictive
56 /// lifecycle we can also add a bunch of assertions that can help to
57 /// catch race conditions ahead of time.
59 template <typename> class Atom = std::atomic,
60 bool SinglePoster = true, // single vs multiple posters
61 bool Blocking = true> // blocking vs spinning
63 constexpr Baton() : state_(INIT) {}
65 Baton(Baton const&) = delete;
66 Baton& operator=(Baton const&) = delete;
68 /// It is an error to destroy a Baton on which a thread is currently
69 /// wait()ing. In practice this means that the waiter usually takes
70 /// responsibility for destroying the Baton.
72 // The docblock for this function says that it can't be called when
73 // there is a concurrent waiter. We assume a strong version of this
74 // requirement in which the caller must _know_ that this is true, they
75 // are not allowed to be merely lucky. If two threads are involved,
76 // the destroying thread must actually have synchronized with the
77 // waiting thread after wait() returned. To convey causality the the
78 // waiting thread must have used release semantics and the destroying
79 // thread must have used acquire semantics for that communication,
80 // so we are guaranteed to see the post-wait() value of state_,
81 // which cannot be WAITING.
83 // Note that since we only care about a single memory location,
84 // the only two plausible memory orders here are relaxed and seq_cst.
85 assert(state_.load(std::memory_order_relaxed) != WAITING);
88 /// Equivalent to destroying the Baton and creating a new one. It is
89 /// a bug to call this while there is a waiting thread, so in practice
90 /// the waiter will be the one that resets the baton.
92 // See ~Baton for a discussion about why relaxed is okay here
93 assert(state_.load(std::memory_order_relaxed) != WAITING);
95 // We use a similar argument to justify the use of a relaxed store
96 // here. Since both wait() and post() are required to be called
97 // only once per lifetime, no thread can actually call those methods
98 // correctly after a reset() unless it synchronizes with the thread
99 // that performed the reset(). If a post() or wait() on another thread
100 // didn't synchronize, then regardless of what operation we performed
101 // here there would be a race on proper use of the Baton's spec
102 // (although not on any particular load and store). Put another way,
103 // we don't need to synchronize here because anybody that might rely
104 // on such synchronization is required by the baton rules to perform
105 // an additional synchronization that has the desired effect anyway.
107 // There is actually a similar argument to be made about the
108 // constructor, in which the fenceless constructor initialization
109 // of state_ is piggybacked on whatever synchronization mechanism
110 // distributes knowledge of the Baton's existence
111 state_.store(INIT, std::memory_order_relaxed);
114 /// Causes wait() to wake up. For each lifetime of a Baton (where a
115 /// lifetime starts at construction or reset() and ends at
116 /// destruction or reset()) there can be at most one call to post(),
117 /// in the single poster version. Any thread may call post().
120 /// Non-blocking version
123 auto state = state_.load(std::memory_order_relaxed);
124 return (state == INIT || state == EARLY_DELIVERY);
126 state_.store(EARLY_DELIVERY, std::memory_order_release);
130 /// Blocking versions
133 /// Single poster version
135 uint32_t before = state_.load(std::memory_order_acquire);
137 assert(before == INIT || before == WAITING || before == TIMED_OUT);
139 if (before == INIT &&
140 state_.compare_exchange_strong(before, EARLY_DELIVERY)) {
144 assert(before == WAITING || before == TIMED_OUT);
146 if (before == TIMED_OUT) {
150 assert(before == WAITING);
151 state_.store(LATE_DELIVERY, std::memory_order_release);
154 /// Multi-poster version
157 uint32_t before = state_.load(std::memory_order_acquire);
159 if (before == INIT &&
160 state_.compare_exchange_strong(before, EARLY_DELIVERY)) {
164 if (before == TIMED_OUT) {
168 if (before == EARLY_DELIVERY || before == LATE_DELIVERY) {
169 // The reason for not simply returning (without the following
170 // atomic operation) is to avoid the following case:
173 // local1.post(); local2.post(); global.wait();
174 // global.post(); global.post(); local1.try_wait() == true;
175 // local2.try_wait() == false;
177 if (state_.fetch_add(0) != before) {
183 assert(before == WAITING);
184 if (!state_.compare_exchange_weak(before, LATE_DELIVERY)) {
193 /// Waits until post() has been called in the current Baton lifetime.
194 /// May be called at most once during a Baton lifetime (construction
195 /// |reset until destruction|reset). If post is called before wait in
196 /// the current lifetime then this method returns immediately.
198 /// The restriction that there can be at most one wait() per lifetime
199 /// could be relaxed somewhat without any perf or size regressions,
200 /// but by making this condition very restrictive we can provide better
201 /// checking in debug builds.
203 if (spinWaitForEarlyDelivery()) {
204 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
209 while (!try_wait()) {
210 std::this_thread::yield();
215 // guess we have to block :(
216 uint32_t expected = INIT;
217 if (!state_.compare_exchange_strong(expected, WAITING)) {
218 // CAS failed, last minute reprieve
219 assert(expected == EARLY_DELIVERY);
224 detail::MemoryIdler::futexWait(state_, WAITING);
226 // state_ is the truth even if FUTEX_WAIT reported a matching
227 // FUTEX_WAKE, since we aren't using type-stable storage and we
228 // don't guarantee reuse. The scenario goes like this: thread
229 // A's last touch of a Baton is a call to wake(), which stores
230 // LATE_DELIVERY and gets an unlucky context switch before delivering
231 // the corresponding futexWake. Thread B sees LATE_DELIVERY
232 // without consuming a futex event, because it calls futexWait
233 // with an expected value of WAITING and hence doesn't go to sleep.
234 // B returns, so the Baton's memory is reused and becomes another
235 // Baton (or a reuse of this one). B calls futexWait on the new
236 // Baton lifetime, then A wakes up and delivers a spurious futexWake
237 // to the same memory location. B's futexWait will then report a
238 // consumed wake event even though state_ is still WAITING.
240 // It would be possible to add an extra state_ dance to communicate
241 // that the futexWake has been sent so that we can be sure to consume
242 // it before returning, but that would be a perf and complexity hit.
243 uint32_t s = state_.load(std::memory_order_acquire);
244 assert(s == WAITING || s == LATE_DELIVERY);
246 if (s == LATE_DELIVERY) {
253 /// Similar to wait, but with a timeout. The thread is unblocked if the
255 /// Note: Only a single call to timed_wait/wait is allowed during a baton's
256 /// life-cycle (from construction/reset to destruction/reset). In other
257 /// words, after timed_wait the caller can't invoke wait/timed_wait/try_wait
258 /// again on the same baton without resetting it.
260 /// @param deadline Time until which the thread can block
261 /// @return true if the baton was posted to before timeout,
263 template <typename Clock, typename Duration = typename Clock::duration>
264 bool timed_wait(const std::chrono::time_point<Clock,Duration>& deadline) {
265 static_assert(Blocking, "Non-blocking Baton does not support timed wait.");
267 if (spinWaitForEarlyDelivery()) {
268 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
272 // guess we have to block :(
273 uint32_t expected = INIT;
274 if (!state_.compare_exchange_strong(expected, WAITING)) {
275 // CAS failed, last minute reprieve
276 assert(expected == EARLY_DELIVERY);
281 auto rv = state_.futexWaitUntil(WAITING, deadline);
282 if (rv == folly::detail::FutexResult::TIMEDOUT) {
283 state_.store(TIMED_OUT, std::memory_order_release);
287 uint32_t s = state_.load(std::memory_order_acquire);
288 assert(s == WAITING || s == LATE_DELIVERY);
289 if (s == LATE_DELIVERY) {
295 /// Similar to timed_wait, but with a duration.
296 template <typename Clock = std::chrono::steady_clock, typename Duration>
297 bool timed_wait(const Duration& duration) {
298 auto deadline = Clock::now() + duration;
299 return timed_wait(deadline);
302 /// Similar to wait, but doesn't block the thread if it hasn't been posted.
304 /// try_wait has the following semantics:
305 /// - It is ok to call try_wait any number times on the same baton until
306 /// try_wait reports that the baton has been posted.
307 /// - It is ok to call timed_wait or wait on the same baton if try_wait
308 /// reports that baton hasn't been posted.
309 /// - If try_wait indicates that the baton has been posted, it is invalid to
310 /// call wait, try_wait or timed_wait on the same baton without resetting
312 /// @return true if baton has been posted, false othewise
313 bool try_wait() const {
314 auto s = state_.load(std::memory_order_acquire);
315 assert(s == INIT || s == EARLY_DELIVERY);
316 return s == EARLY_DELIVERY;
320 enum State : uint32_t {
329 // Must be positive. If multiple threads are actively using a
330 // higher-level data structure that uses batons internally, it is
331 // likely that the post() and wait() calls happen almost at the same
332 // time. In this state, we lose big 50% of the time if the wait goes
333 // to sleep immediately. On circa-2013 devbox hardware it costs about
334 // 7 usec to FUTEX_WAIT and then be awoken (half the t/iter as the
335 // posix_sem_pingpong test in BatonTests). We can improve our chances
336 // of EARLY_DELIVERY by spinning for a bit, although we have to balance
337 // this against the loss if we end up sleeping any way. Spins on this
338 // hw take about 7 nanos (all but 0.5 nanos is the pause instruction).
339 // We give ourself 300 spins, which is about 2 usec of waiting. As a
340 // partial consolation, since we are using the pause instruction we
341 // are giving a speed boost to the colocated hyperthread.
342 PreBlockAttempts = 300,
345 // Spin for "some time" (see discussion on PreBlockAttempts) waiting
348 // @return true if we received an early delivery during the wait,
349 // false otherwise. If the function returns true then
350 // state_ is guaranteed to be EARLY_DELIVERY
351 bool spinWaitForEarlyDelivery() {
353 static_assert(PreBlockAttempts > 0,
354 "isn't this assert clearer than an uninitialized variable warning?");
355 for (int i = 0; i < PreBlockAttempts; ++i) {
360 // The pause instruction is the polite way to spin, but it doesn't
361 // actually affect correctness to omit it if we don't have it.
362 // Pausing donates the full capabilities of the current core to
363 // its other hyperthreads for a dozen cycles or so
364 asm_volatile_pause();
370 detail::Futex<Atom> state_;