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 FOLLY_ALWAYS_INLINE bool ready() const noexcept {
83 auto s = state_.load(std::memory_order_acquire);
84 assert(s == INIT || s == EARLY_DELIVERY);
85 return LIKELY(s == EARLY_DELIVERY);
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.
91 void reset() noexcept {
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().
118 void post() noexcept {
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
132 uint32_t before = state_.load(std::memory_order_acquire);
134 assert(before == INIT || before == WAITING || before == TIMED_OUT);
136 if (before == INIT &&
137 state_.compare_exchange_strong(
140 std::memory_order_release,
141 std::memory_order_relaxed)) {
145 assert(before == WAITING || before == TIMED_OUT);
147 if (before == TIMED_OUT) {
151 assert(before == WAITING);
152 state_.store(LATE_DELIVERY, std::memory_order_release);
156 /// Waits until post() has been called in the current Baton lifetime.
157 /// May be called at most once during a Baton lifetime (construction
158 /// |reset until destruction|reset). If post is called before wait in
159 /// the current lifetime then this method returns immediately.
161 /// The restriction that there can be at most one wait() per lifetime
162 /// could be relaxed somewhat without any perf or size regressions,
163 /// but by making this condition very restrictive we can provide better
164 /// checking in debug builds.
165 FOLLY_ALWAYS_INLINE void wait() noexcept {
173 /// Similar to wait, but doesn't block the thread if it hasn't been posted.
175 /// try_wait has the following semantics:
176 /// - It is ok to call try_wait any number times on the same baton until
177 /// try_wait reports that the baton has been posted.
178 /// - It is ok to call timed_wait or wait on the same baton if try_wait
179 /// reports that baton hasn't been posted.
180 /// - If try_wait indicates that the baton has been posted, it is invalid to
181 /// call wait, try_wait or timed_wait on the same baton without resetting
183 /// @return true if baton has been posted, false othewise
184 FOLLY_ALWAYS_INLINE bool try_wait() const noexcept {
188 /// Similar to wait, but with a timeout. The thread is unblocked if the
190 /// Note: Only a single call to wait/try_wait_for/try_wait_until is allowed
191 /// during a baton's life-cycle (from ctor/reset to dtor/reset). In other
192 /// words, after try_wait_for the caller can't invoke
193 /// wait/try_wait/try_wait_for/try_wait_until
194 /// again on the same baton without resetting it.
196 /// @param timeout Time until which the thread can block
197 /// @return true if the baton was posted to before timeout,
199 template <typename Rep, typename Period>
200 FOLLY_ALWAYS_INLINE bool try_wait_for(
201 const std::chrono::duration<Rep, Period>& timeout) noexcept {
203 Blocking, "Non-blocking Baton does not support try_wait_for.");
209 auto deadline = std::chrono::steady_clock::now() + timeout;
210 return tryWaitUntilSlow(deadline);
213 /// Similar to wait, but with a deadline. The thread is unblocked if the
214 /// deadline expires.
215 /// Note: Only a single call to wait/try_wait_for/try_wait_until is allowed
216 /// during a baton's life-cycle (from ctor/reset to dtor/reset). In other
217 /// words, after try_wait_until the caller can't invoke
218 /// wait/try_wait/try_wait_for/try_wait_until
219 /// again on the same baton without resetting it.
221 /// @param deadline Time until which the thread can block
222 /// @return true if the baton was posted to before deadline,
224 template <typename Clock, typename Duration>
225 FOLLY_ALWAYS_INLINE bool try_wait_until(
226 const std::chrono::time_point<Clock, Duration>& deadline) noexcept {
228 Blocking, "Non-blocking Baton does not support try_wait_until.");
234 return tryWaitUntilSlow(deadline);
237 /// Alias to try_wait_for. Deprecated.
238 template <typename Rep, typename Period>
239 FOLLY_ALWAYS_INLINE bool timed_wait(
240 const std::chrono::duration<Rep, Period>& timeout) noexcept {
241 return try_wait_for(timeout);
244 /// Alias to try_wait_until. Deprecated.
245 template <typename Clock, typename Duration>
246 FOLLY_ALWAYS_INLINE bool timed_wait(
247 const std::chrono::time_point<Clock, Duration>& deadline) noexcept {
248 return try_wait_until(deadline);
252 enum State : uint32_t {
261 // Must be positive. If multiple threads are actively using a
262 // higher-level data structure that uses batons internally, it is
263 // likely that the post() and wait() calls happen almost at the same
264 // time. In this state, we lose big 50% of the time if the wait goes
265 // to sleep immediately. On circa-2013 devbox hardware it costs about
266 // 7 usec to FUTEX_WAIT and then be awoken (half the t/iter as the
267 // posix_sem_pingpong test in BatonTests). We can improve our chances
268 // of EARLY_DELIVERY by spinning for a bit, although we have to balance
269 // this against the loss if we end up sleeping any way. Spins on this
270 // hw take about 7 nanos (all but 0.5 nanos is the pause instruction).
271 // We give ourself 300 spins, which is about 2 usec of waiting. As a
272 // partial consolation, since we are using the pause instruction we
273 // are giving a speed boost to the colocated hyperthread.
274 PreBlockAttempts = 300,
277 // Spin for "some time" (see discussion on PreBlockAttempts) waiting
280 // @return true if we received an early delivery during the wait,
281 // false otherwise. If the function returns true then
282 // state_ is guaranteed to be EARLY_DELIVERY
283 bool spinWaitForEarlyDelivery() noexcept {
285 PreBlockAttempts > 0,
286 "isn't this assert clearer than an uninitialized variable warning?");
287 for (int i = 0; i < PreBlockAttempts; ++i) {
292 // The pause instruction is the polite way to spin, but it doesn't
293 // actually affect correctness to omit it if we don't have it.
294 // Pausing donates the full capabilities of the current core to
295 // its other hyperthreads for a dozen cycles or so
296 asm_volatile_pause();
302 FOLLY_NOINLINE void waitSlow() noexcept {
303 if (spinWaitForEarlyDelivery()) {
304 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
309 while (!try_wait()) {
310 std::this_thread::yield();
315 // guess we have to block :(
316 uint32_t expected = INIT;
317 if (!state_.compare_exchange_strong(expected, WAITING)) {
318 // CAS failed, last minute reprieve
319 assert(expected == EARLY_DELIVERY);
324 detail::MemoryIdler::futexWait(state_, WAITING);
326 // state_ is the truth even if FUTEX_WAIT reported a matching
327 // FUTEX_WAKE, since we aren't using type-stable storage and we
328 // don't guarantee reuse. The scenario goes like this: thread
329 // A's last touch of a Baton is a call to wake(), which stores
330 // LATE_DELIVERY and gets an unlucky context switch before delivering
331 // the corresponding futexWake. Thread B sees LATE_DELIVERY
332 // without consuming a futex event, because it calls futexWait
333 // with an expected value of WAITING and hence doesn't go to sleep.
334 // B returns, so the Baton's memory is reused and becomes another
335 // Baton (or a reuse of this one). B calls futexWait on the new
336 // Baton lifetime, then A wakes up and delivers a spurious futexWake
337 // to the same memory location. B's futexWait will then report a
338 // consumed wake event even though state_ is still WAITING.
340 // It would be possible to add an extra state_ dance to communicate
341 // that the futexWake has been sent so that we can be sure to consume
342 // it before returning, but that would be a perf and complexity hit.
343 uint32_t s = state_.load(std::memory_order_acquire);
344 assert(s == WAITING || s == LATE_DELIVERY);
346 if (s == LATE_DELIVERY) {
353 template <typename Clock, typename Duration>
354 FOLLY_NOINLINE bool tryWaitUntilSlow(
355 const std::chrono::time_point<Clock, Duration>& deadline) noexcept {
356 if (spinWaitForEarlyDelivery()) {
357 assert(state_.load(std::memory_order_acquire) == EARLY_DELIVERY);
361 // guess we have to block :(
362 uint32_t expected = INIT;
363 if (!state_.compare_exchange_strong(expected, WAITING)) {
364 // CAS failed, last minute reprieve
365 assert(expected == EARLY_DELIVERY);
370 auto rv = state_.futexWaitUntil(WAITING, deadline);
371 if (rv == folly::detail::FutexResult::TIMEDOUT) {
372 state_.store(TIMED_OUT, std::memory_order_release);
376 uint32_t s = state_.load(std::memory_order_acquire);
377 assert(s == WAITING || s == LATE_DELIVERY);
378 if (s == LATE_DELIVERY) {
384 detail::Futex<Atom> state_;