2 * Copyright 2015 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>
30 /// A TurnSequencer allows threads to order their execution according to
31 /// a monotonically increasing (with wraparound) "turn" value. The two
32 /// operations provided are to wait for turn T, and to move to the next
33 /// turn. Every thread that is waiting for T must have arrived before
34 /// that turn is marked completed (for MPMCQueue only one thread waits
35 /// for any particular turn, so this is trivially true).
37 /// TurnSequencer's state_ holds 26 bits of the current turn (shifted
38 /// left by 6), along with a 6 bit saturating value that records the
39 /// maximum waiter minus the current turn. Wraparound of the turn space
40 /// is expected and handled. This allows us to atomically adjust the
41 /// number of outstanding waiters when we perform a FUTEX_WAKE operation.
42 /// Compare this strategy to sem_t's separate num_waiters field, which
43 /// isn't decremented until after the waiting thread gets scheduled,
44 /// during which time more enqueues might have occurred and made pointless
47 /// TurnSequencer uses futex() directly. It is optimized for the
48 /// case that the highest awaited turn is 32 or less higher than the
49 /// current turn. We use the FUTEX_WAIT_BITSET variant, which lets
50 /// us embed 32 separate wakeup channels in a single futex. See
51 /// http://locklessinc.com/articles/futex_cheat_sheet for a description.
53 /// We only need to keep exact track of the delta between the current
54 /// turn and the maximum waiter for the 32 turns that follow the current
55 /// one, because waiters at turn t+32 will be awoken at turn t. At that
56 /// point they can then adjust the delta using the higher base. Since we
57 /// need to encode waiter deltas of 0 to 32 inclusive, we use 6 bits.
58 /// We actually store waiter deltas up to 63, since that might reduce
59 /// the number of CAS operations a tiny bit.
61 /// To avoid some futex() calls entirely, TurnSequencer uses an adaptive
62 /// spin cutoff before waiting. The overheads (and convergence rate)
63 /// of separately tracking the spin cutoff for each TurnSequencer would
64 /// be prohibitive, so the actual storage is passed in as a parameter and
65 /// updated atomically. This also lets the caller use different adaptive
66 /// cutoffs for different operations (read versus write, for example).
67 /// To avoid contention, the spin cutoff is only updated when requested
69 template <template<typename> class Atom>
70 struct TurnSequencer {
71 explicit TurnSequencer(const uint32_t firstTurn = 0) noexcept
72 : state_(encode(firstTurn << kTurnShift, 0))
75 /// Returns true iff a call to waitForTurn(turn, ...) won't block
76 bool isTurn(const uint32_t turn) const noexcept {
77 auto state = state_.load(std::memory_order_acquire);
78 return decodeCurrentSturn(state) == (turn << kTurnShift);
81 // Internally we always work with shifted turn values, which makes the
82 // truncation and wraparound work correctly. This leaves us bits at
83 // the bottom to store the number of waiters. We call shifted turns
84 // "sturns" inside this class.
86 /// Blocks the current thread until turn has arrived. If
87 /// updateSpinCutoff is true then this will spin for up to kMaxSpins tries
88 /// before blocking and will adjust spinCutoff based on the results,
89 /// otherwise it will spin for at most spinCutoff spins.
90 void waitForTurn(const uint32_t turn,
91 Atom<uint32_t>& spinCutoff,
92 const bool updateSpinCutoff) noexcept {
93 uint32_t prevThresh = spinCutoff.load(std::memory_order_relaxed);
94 const uint32_t effectiveSpinCutoff =
95 updateSpinCutoff || prevThresh == 0 ? kMaxSpins : prevThresh;
98 const uint32_t sturn = turn << kTurnShift;
99 for (tries = 0; ; ++tries) {
100 uint32_t state = state_.load(std::memory_order_acquire);
101 uint32_t current_sturn = decodeCurrentSturn(state);
102 if (current_sturn == sturn) {
106 // wrap-safe version of assert(current_sturn < sturn)
107 assert(sturn - current_sturn < std::numeric_limits<uint32_t>::max() / 2);
109 // the first effectSpinCutoff tries are spins, after that we will
110 // record ourself as a waiter and block with futexWait
111 if (tries < effectiveSpinCutoff) {
112 asm volatile ("pause");
116 uint32_t current_max_waiter_delta = decodeMaxWaitersDelta(state);
117 uint32_t our_waiter_delta = (sturn - current_sturn) >> kTurnShift;
119 if (our_waiter_delta <= current_max_waiter_delta) {
120 // state already records us as waiters, probably because this
121 // isn't our first time around this loop
124 new_state = encode(current_sturn, our_waiter_delta);
125 if (state != new_state &&
126 !state_.compare_exchange_strong(state, new_state)) {
130 state_.futexWait(new_state, futexChannel(turn));
133 if (updateSpinCutoff || prevThresh == 0) {
134 // if we hit kMaxSpins then spinning was pointless, so the right
135 // spinCutoff is kMinSpins
137 if (tries >= kMaxSpins) {
140 // to account for variations, we allow ourself to spin 2*N when
141 // we think that N is actually required in order to succeed
142 target = std::min<uint32_t>(kMaxSpins,
143 std::max<uint32_t>(kMinSpins, tries * 2));
146 if (prevThresh == 0) {
148 spinCutoff.store(target);
150 // try once, keep moving if CAS fails. Exponential moving average
152 // Be careful that the quantity we add to prevThresh is signed.
153 spinCutoff.compare_exchange_weak(
154 prevThresh, prevThresh + int(target - prevThresh) / 8);
159 /// Unblocks a thread running waitForTurn(turn + 1)
160 void completeTurn(const uint32_t turn) noexcept {
161 uint32_t state = state_.load(std::memory_order_acquire);
163 assert(state == encode(turn << kTurnShift, decodeMaxWaitersDelta(state)));
164 uint32_t max_waiter_delta = decodeMaxWaitersDelta(state);
165 uint32_t new_state = encode(
166 (turn + 1) << kTurnShift,
167 max_waiter_delta == 0 ? 0 : max_waiter_delta - 1);
168 if (state_.compare_exchange_strong(state, new_state)) {
169 if (max_waiter_delta != 0) {
170 state_.futexWake(std::numeric_limits<int>::max(),
171 futexChannel(turn + 1));
175 // failing compare_exchange_strong updates first arg to the value
176 // that caused the failure, so no need to reread state_
180 /// Returns the least-most significant byte of the current uncompleted
181 /// turn. The full 32 bit turn cannot be recovered.
182 uint8_t uncompletedTurnLSB() const noexcept {
183 return state_.load(std::memory_order_acquire) >> kTurnShift;
188 /// kTurnShift counts the bits that are stolen to record the delta
189 /// between the current turn and the maximum waiter. It needs to be big
190 /// enough to record wait deltas of 0 to 32 inclusive. Waiters more
191 /// than 32 in the future will be woken up 32*n turns early (since
192 /// their BITSET will hit) and will adjust the waiter count again.
193 /// We go a bit beyond and let the waiter count go up to 63, which
194 /// is free and might save us a few CAS
196 kWaitersMask = (1 << kTurnShift) - 1,
198 /// The minimum spin count that we will adaptively select
201 /// The maximum spin count that we will adaptively select, and the
202 /// spin count that will be used when probing to get a new data point
203 /// for the adaptation
207 /// This holds both the current turn, and the highest waiting turn,
208 /// stored as (current_turn << 6) | min(63, max(waited_turn - current_turn))
211 /// Returns the bitmask to pass futexWait or futexWake when communicating
212 /// about the specified turn
213 int futexChannel(uint32_t turn) const noexcept {
214 return 1 << (turn & 31);
217 uint32_t decodeCurrentSturn(uint32_t state) const noexcept {
218 return state & ~kWaitersMask;
221 uint32_t decodeMaxWaitersDelta(uint32_t state) const noexcept {
222 return state & kWaitersMask;
225 uint32_t encode(uint32_t currentSturn, uint32_t maxWaiterD) const noexcept {
226 return currentSturn | std::min(uint32_t{ kWaitersMask }, maxWaiterD);
230 } // namespace detail