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/Likely.h>
26 #include <folly/detail/Futex.h>
27 #include <folly/detail/MemoryIdler.h>
28 #include <folly/portability/Asm.h>
32 /// A Baton allows a thread to block once and be awoken. Captures a
33 /// single handoff, and during its lifecycle (from construction/reset
34 /// to destruction/reset) a baton must either be post()ed and wait()ed
35 /// exactly once each, or not at all.
37 /// Baton includes no internal padding, and is only 4 bytes in size.
38 /// Any alignment or padding to avoid false sharing is up to the user.
40 /// This is basically a stripped-down semaphore that supports only a
41 /// single call to sem_post and a single call to sem_wait.
43 /// The non-blocking version (Blocking == false) provides more speed
44 /// by using only load acquire and store release operations in the
45 /// critical path, at the cost of disallowing blocking and timing out.
47 /// The current posix semaphore sem_t isn't too bad, but this provides
48 /// more a bit more speed, inlining, smaller size, a guarantee that
49 /// the implementation won't change, and compatibility with
50 /// DeterministicSchedule. By having a much more restrictive
51 /// lifecycle we can also add a bunch of assertions that can help to
52 /// catch race conditions ahead of time.
54 template <typename> class Atom = std::atomic,
55 bool Blocking = true> // blocking vs spinning
57 constexpr Baton() noexcept : state_(INIT) {}
59 Baton(Baton const&) = delete;
60 Baton& operator=(Baton const&) = delete;
62 /// It is an error to destroy a Baton on which a thread is currently
63 /// wait()ing. In practice this means that the waiter usually takes
64 /// responsibility for destroying the Baton.
66 // The docblock for this function says that it can't be called when
67 // there is a concurrent waiter. We assume a strong version of this
68 // requirement in which the caller must _know_ that this is true, they
69 // are not allowed to be merely lucky. If two threads are involved,
70 // the destroying thread must actually have synchronized with the
71 // waiting thread after wait() returned. To convey causality the the
72 // waiting thread must have used release semantics and the destroying
73 // thread must have used acquire semantics for that communication,
74 // so we are guaranteed to see the post-wait() value of state_,
75 // which cannot be WAITING.
77 // Note that since we only care about a single memory location,
78 // the only two plausible memory orders here are relaxed and seq_cst.
79 assert(state_.load(std::memory_order_relaxed) != WAITING);
82 /// Equivalent to destroying the Baton and creating a new one. It is
83 /// a bug to call this while there is a waiting thread, so in practice
84 /// the waiter will be the one that resets the baton.
85 void reset() noexcept {
86 // See ~Baton for a discussion about why relaxed is okay here
87 assert(state_.load(std::memory_order_relaxed) != WAITING);
89 // We use a similar argument to justify the use of a relaxed store
90 // here. Since both wait() and post() are required to be called
91 // only once per lifetime, no thread can actually call those methods
92 // correctly after a reset() unless it synchronizes with the thread
93 // that performed the reset(). If a post() or wait() on another thread
94 // didn't synchronize, then regardless of what operation we performed
95 // here there would be a race on proper use of the Baton's spec
96 // (although not on any particular load and store). Put another way,
97 // we don't need to synchronize here because anybody that might rely
98 // on such synchronization is required by the baton rules to perform
99 // an additional synchronization that has the desired effect anyway.
101 // There is actually a similar argument to be made about the
102 // constructor, in which the fenceless constructor initialization
103 // of state_ is piggybacked on whatever synchronization mechanism
104 // distributes knowledge of the Baton's existence
105 state_.store(INIT, std::memory_order_relaxed);
108 /// Causes wait() to wake up. For each lifetime of a Baton (where a
109 /// lifetime starts at construction or reset() and ends at
110 /// destruction or reset()) there can be at most one call to post(),
111 /// in the single poster version. Any thread may call post().
112 void post() noexcept {
114 /// Non-blocking version
117 auto state = state_.load(std::memory_order_relaxed);
118 return (state == INIT || state == EARLY_DELIVERY);
120 state_.store(EARLY_DELIVERY, std::memory_order_release);
124 /// Blocking versions
126 uint32_t before = state_.load(std::memory_order_acquire);
128 assert(before == INIT || before == WAITING || before == TIMED_OUT);
130 if (before == INIT &&
131 state_.compare_exchange_strong(
134 std::memory_order_release,
135 std::memory_order_relaxed)) {
139 assert(before == WAITING || before == TIMED_OUT);
141 if (before == TIMED_OUT) {
145 assert(before == WAITING);
146 state_.store(LATE_DELIVERY, std::memory_order_release);
150 /// Waits until post() has been called in the current Baton lifetime.
151 /// May be called at most once during a Baton lifetime (construction
152 /// |reset until destruction|reset). If post is called before wait in
153 /// the current lifetime then this method returns immediately.
155 /// The restriction that there can be at most one wait() per lifetime
156 /// could be relaxed somewhat without any perf or size regressions,
157 /// but by making this condition very restrictive we can provide better
158 /// checking in debug builds.
159 FOLLY_ALWAYS_INLINE void wait() noexcept {
167 /// Similar to wait, but doesn't block the thread if it hasn't been posted.
169 /// try_wait has the following semantics:
170 /// - It is ok to call try_wait any number times on the same baton until
171 /// try_wait reports that the baton has been posted.
172 /// - It is ok to call timed_wait or wait on the same baton if try_wait
173 /// reports that baton hasn't been posted.
174 /// - If try_wait indicates that the baton has been posted, it is invalid to
175 /// call wait, try_wait or timed_wait on the same baton without resetting
177 /// @return true if baton has been posted, false othewise
178 FOLLY_ALWAYS_INLINE bool try_wait() const noexcept {
179 auto s = state_.load(std::memory_order_acquire);
180 assert(s == INIT || s == EARLY_DELIVERY);
181 return LIKELY(s == EARLY_DELIVERY);
184 /// Similar to wait, but with a timeout. The thread is unblocked if the
186 /// Note: Only a single call to wait/try_wait_for/try_wait_until is allowed
187 /// during a baton's life-cycle (from ctor/reset to dtor/reset). In other
188 /// words, after try_wait_for the caller can't invoke
189 /// wait/try_wait/try_wait_for/try_wait_until
190 /// again on the same baton without resetting it.
192 /// @param timeout Time until which the thread can block
193 /// @return true if the baton was posted to before timeout,
195 template <typename Rep, typename Period>
196 FOLLY_ALWAYS_INLINE bool try_wait_for(
197 const std::chrono::duration<Rep, Period>& timeout) noexcept {
199 Blocking, "Non-blocking Baton does not support try_wait_for.");
205 auto deadline = std::chrono::steady_clock::now() + timeout;
206 return tryWaitUntilSlow(deadline);
209 /// Similar to wait, but with a deadline. The thread is unblocked if the
210 /// deadline expires.
211 /// Note: Only a single call to wait/try_wait_for/try_wait_until is allowed
212 /// during a baton's life-cycle (from ctor/reset to dtor/reset). In other
213 /// words, after try_wait_until the caller can't invoke
214 /// wait/try_wait/try_wait_for/try_wait_until
215 /// again on the same baton without resetting it.
217 /// @param deadline Time until which the thread can block
218 /// @return true if the baton was posted to before deadline,
220 template <typename Clock, typename Duration>
221 FOLLY_ALWAYS_INLINE bool try_wait_until(
222 const std::chrono::time_point<Clock, Duration>& deadline) noexcept {
224 Blocking, "Non-blocking Baton does not support try_wait_until.");
230 return tryWaitUntilSlow(deadline);
233 /// Alias to try_wait_for. Deprecated.
234 template <typename Rep, typename Period>
235 FOLLY_ALWAYS_INLINE bool timed_wait(
236 const std::chrono::duration<Rep, Period>& timeout) noexcept {
237 return try_wait_for(timeout);
240 /// Alias to try_wait_until. Deprecated.
241 template <typename Clock, typename Duration>
242 FOLLY_ALWAYS_INLINE bool timed_wait(
243 const std::chrono::time_point<Clock, Duration>& deadline) noexcept {
244 return try_wait_until(deadline);
248 enum State : uint32_t {
257 // Must be positive. If multiple threads are actively using a
258 // higher-level data structure that uses batons internally, it is
259 // likely that the post() and wait() calls happen almost at the same
260 // time. In this state, we lose big 50% of the time if the wait goes
261 // to sleep immediately. On circa-2013 devbox hardware it costs about
262 // 7 usec to FUTEX_WAIT and then be awoken (half the t/iter as the
263 // posix_sem_pingpong test in BatonTests). We can improve our chances
264 // of EARLY_DELIVERY by spinning for a bit, although we have to balance
265 // this against the loss if we end up sleeping any way. Spins on this
266 // hw take about 7 nanos (all but 0.5 nanos is the pause instruction).
267 // We give ourself 300 spins, which is about 2 usec of waiting. As a
268 // partial consolation, since we are using the pause instruction we
269 // are giving a speed boost to the colocated hyperthread.
270 PreBlockAttempts = 300,
273 // Spin for "some time" (see discussion on PreBlockAttempts) waiting
276 // @return true if we received an early delivery during the wait,
277 // false otherwise. If the function returns true then
278 // state_ is guaranteed to be EARLY_DELIVERY
279 bool spinWaitForEarlyDelivery() noexcept {
281 PreBlockAttempts > 0,
282 "isn't this assert clearer than an uninitialized variable warning?");
283 for (int i = 0; i < PreBlockAttempts; ++i) {
288 // The pause instruction is the polite way to spin, but it doesn't
289 // actually affect correctness to omit it if we don't have it.
290 // Pausing donates the full capabilities of the current core to
291 // its other hyperthreads for a dozen cycles or so
292 asm_volatile_pause();
298 FOLLY_NOINLINE void waitSlow() noexcept {
299 if (spinWaitForEarlyDelivery()) {
300 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
305 while (!try_wait()) {
306 std::this_thread::yield();
311 // guess we have to block :(
312 uint32_t expected = INIT;
313 if (!state_.compare_exchange_strong(expected, WAITING)) {
314 // CAS failed, last minute reprieve
315 assert(expected == EARLY_DELIVERY);
320 detail::MemoryIdler::futexWait(state_, WAITING);
322 // state_ is the truth even if FUTEX_WAIT reported a matching
323 // FUTEX_WAKE, since we aren't using type-stable storage and we
324 // don't guarantee reuse. The scenario goes like this: thread
325 // A's last touch of a Baton is a call to wake(), which stores
326 // LATE_DELIVERY and gets an unlucky context switch before delivering
327 // the corresponding futexWake. Thread B sees LATE_DELIVERY
328 // without consuming a futex event, because it calls futexWait
329 // with an expected value of WAITING and hence doesn't go to sleep.
330 // B returns, so the Baton's memory is reused and becomes another
331 // Baton (or a reuse of this one). B calls futexWait on the new
332 // Baton lifetime, then A wakes up and delivers a spurious futexWake
333 // to the same memory location. B's futexWait will then report a
334 // consumed wake event even though state_ is still WAITING.
336 // It would be possible to add an extra state_ dance to communicate
337 // that the futexWake has been sent so that we can be sure to consume
338 // it before returning, but that would be a perf and complexity hit.
339 uint32_t s = state_.load(std::memory_order_acquire);
340 assert(s == WAITING || s == LATE_DELIVERY);
342 if (s == LATE_DELIVERY) {
349 template <typename Clock, typename Duration>
350 FOLLY_NOINLINE bool tryWaitUntilSlow(
351 const std::chrono::time_point<Clock, Duration>& deadline) noexcept {
352 if (spinWaitForEarlyDelivery()) {
353 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
357 // guess we have to block :(
358 uint32_t expected = INIT;
359 if (!state_.compare_exchange_strong(expected, WAITING)) {
360 // CAS failed, last minute reprieve
361 assert(expected == EARLY_DELIVERY);
366 auto rv = state_.futexWaitUntil(WAITING, deadline);
367 if (rv == folly::detail::FutexResult::TIMEDOUT) {
368 state_.store(TIMED_OUT, std::memory_order_release);
372 uint32_t s = state_.load(std::memory_order_acquire);
373 assert(s == WAITING || s == LATE_DELIVERY);
374 if (s == LATE_DELIVERY) {
380 detail::Futex<Atom> state_;