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.
19 #include <folly/Baton.h>
20 #include <folly/Function.h>
21 #include <folly/IndexedMemPool.h>
22 #include <folly/Portability.h>
23 #include <folly/detail/CacheLocality.h>
31 /// Flat combining (FC) was introduced in the SPAA 2010 paper Flat
32 /// Combining and the Synchronization-Parallelism Tradeoff, by Danny
33 /// Hendler, Itai Incze, Nir Shavit, and Moran Tzafrir.
34 /// http://mcg.cs.tau.ac.il/projects/projects/flat-combining
36 /// FC is an alternative to coarse-grained locking for making
37 /// sequential data structures thread-safe while minimizing the
38 /// synchroniation overheads and cache coherence traffic associated
41 /// Under FC, when a thread finds the lock contended, it can
42 /// request (using a request record) that the lock holder execute its
43 /// operation on the shared data structure. There can be a designated
44 /// combiner thread or any thread can act as the combiner when it
47 /// Potential advantages of FC include:
48 /// - Reduced cache coherence traffic
49 /// - Reduced synchronization overheads, as the overheads of releasing
50 /// and acquiring the lock are eliminated from the critical path of
51 /// operating on the data structure.
52 /// - Opportunities for smart combining, where executing multiple
53 /// operations together may take less time than executng the
54 /// operations separately, e.g., K delete_min operations on a
55 /// priority queue may be combined to take O(K + log N) time instead
58 /// This implementation of flat combining supports:
60 /// - A simple interface that requires minimal extra code by the
61 /// user. To use this interface efficiently the user-provided
62 /// functions must be copyable to folly::Functio without dynamic
63 /// allocation. If this is impossible or inconvenient, the user is
64 /// encouraged to use the custom interface described below.
65 /// - A custom interface that supports custom combinining and custom
66 /// request structure, either for the sake of smart combining or for
67 /// efficiently supporting operations that are not be copyable to
68 /// folly::Function without synamic allocation.
69 /// - Both synchronous and asynchronous operations.
70 /// - Request records with and without thread-caching.
71 /// - Combining with and without a dedicated combiner thread.
73 /// This implementation differs from the algorithm in the SPAA 2010 paper:
74 /// - It does not require thread caching of request records
75 /// - It supports a dedicated combiner
76 /// - It supports asynchronous operations
78 /// The generic FC class template supports generic data structures and
79 /// utilities with arbitrary operations. The template supports static
80 /// polymorphism for the combining function to enable custom smart
83 /// A simple example of using the FC template:
84 /// class ConcurrentFoo : public FlatCombining<ConcurrentFoo> {
85 /// Foo foo_; // sequential data structure
87 /// T bar(V& v) { // thread-safe execution of foo_.bar(v)
89 /// // Note: fn must be copyable to folly::Function without dynamic
90 /// // allocation. Otherwise, it is recommended to use the custom
91 /// // interface and manage the function arguments and results
92 /// // explicitly in a custom request structure.
93 /// auto fn = [&] { result = foo_.bar(v); };
94 /// this->requestFC(fn);
99 /// See test/FlatCombiningExamples.h for more examples. See the
100 /// comments for requestFC() below for a list of simple and custom
101 /// variants of that function.
104 typename T, // concurrent data structure using FC interface
105 typename Mutex = std::mutex,
106 template <typename> class Atom = std::atomic,
107 typename Req = /* default dummy type */ bool>
108 class FlatCombining {
109 using SavedFn = folly::Function<void()>;
112 /// Combining request record.
114 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
115 folly::Baton<Atom, true, false> valid_;
116 folly::Baton<Atom, true, false> done_;
117 folly::Baton<Atom, true, false> disconnected_;
138 bool isValid() const {
139 return valid_.try_wait();
150 bool isDone() const {
151 return done_.try_wait();
158 void setDisconnected() {
159 disconnected_.post();
162 void clearDisconnected() {
163 disconnected_.reset();
166 bool isDisconnected() const {
167 return disconnected_.try_wait();
170 void setIndex(const size_t index) {
174 size_t getIndex() const {
178 void setNext(const size_t next) {
182 size_t getNext() const {
186 void setLast(const uint64_t pass) {
190 uint64_t getLast() const {
198 template <typename Func>
199 void setFn(Func&& fn) {
201 std::is_nothrow_constructible<
202 folly::Function<void()>,
203 _t<std::decay<Func>>>::value,
204 "Try using a smaller function object that can fit in folly::Function "
205 "without allocation, or use the custom interface of requestFC() to "
206 "manage the requested function's arguments and results explicitly "
207 "in a custom request structure without allocation.");
208 fn_ = std::forward<Func>(fn);
228 using Pool = folly::IndexedMemPool<Rec, 32, 4, Atom, false, false>;
231 /// The constructor takes three optional arguments:
232 /// - Optional dedicated combiner thread (default true)
233 /// - Number of records (if 0, then kDefaultNumRecs)
234 /// - A hint for the max. number of combined operations per
235 /// combining session that is checked at the beginning of each pass
236 /// on the request records (if 0, then kDefaultMaxops)
237 explicit FlatCombining(
238 const bool dedicated = true,
239 const uint32_t numRecs = 0, // number of combining records
240 const uint32_t maxOps = 0 // hint of max ops per combining session
242 : numRecs_(numRecs == 0 ? kDefaultNumRecs : numRecs),
243 maxOps_(maxOps == 0 ? kDefaultMaxOps : maxOps),
245 dedicated_(dedicated),
246 recsPool_(numRecs_) {
248 // dedicated combiner thread
249 combiner_ = std::thread([this] { dedicatedCombining(); });
253 /// Destructor: If there is a dedicated combiner, the destructor
254 /// flags it to shutdown. Otherwise, the destructor waits for all
255 /// pending asynchronous requests to be completed.
265 // Wait for all pending operations to complete. Useful primarily
266 // when there are asynchronous operations without a dedicated
269 for (size_t i = getRecsHead(); i != NULL_INDEX; i = nextIndex(i)) {
270 Rec& rec = recsPool_[i];
275 // Give the caller exclusive access.
276 void acquireExclusive() {
280 // Try to give the caller exclusive access. Returns true iff successful.
281 bool tryExclusive() {
282 return m_.try_lock();
285 // Release exclusive access. The caller must have exclusive access.
286 void releaseExclusive() {
290 // Give the lock holder ownership of the mutex and exclusive access.
291 // No need for explicit release.
292 template <typename LockHolder>
293 void holdLock(LockHolder& l) {
297 // Give the caller's lock holder ownership of the mutex but without
298 // exclusive access. The caller can later use the lock holder to try
299 // to acquire exclusive access.
300 template <typename LockHolder>
301 void holdLock(LockHolder& l, std::defer_lock_t) {
302 l = LockHolder(m_, std::defer_lock);
305 // Execute an operation without combining
306 template <typename OpFunc>
307 void requestNoFC(OpFunc& opFn) {
308 std::lock_guard<Mutex> guard(m_);
312 // This function first tries to execute the operation without
313 // combining. If unuccessful, it allocates a combining record if
314 // needed. If there are no available records, it waits for exclusive
315 // access and executes the operation. If a record is available and
316 // ready for use, it fills the record and indicates that the request
317 // is valid for combining. If the request is synchronous (by default
318 // or necessity), it waits for the operation to be completed by a
319 // combiner and optionally extracts the result, if any.
321 // This function can be called in several forms:
322 // Simple forms that do not require the user to define a Req structure
323 // or to override any request processing member functions:
325 // requestFC(opFn, rec) // provides its own pre-allocated record
326 // requestFC(opFn, rec, syncop) // asynchronous if syncop == false
327 // Custom forms that require the user to define a Req structure and to
328 // override some request processing member functions:
329 // requestFC(opFn, fillFn)
330 // requestFC(opFn, fillFn, rec)
331 // requestFC(opFn, fillFn, rec, syncop)
332 // requestFC(opFn, fillFn, resFn)
333 // requestFC(opFn, fillFn, resFn, rec)
334 template <typename OpFunc>
335 void requestFC(OpFunc&& opFn, Rec* rec = nullptr, bool syncop = true) {
336 auto dummy = [](Req&) {};
338 std::forward<OpFunc>(opFn),
345 template <typename OpFunc, typename FillFunc>
348 const FillFunc& fillFn,
350 bool syncop = true) {
351 auto dummy = [](Req&) {};
353 std::forward<OpFunc>(opFn),
360 template <typename OpFunc, typename FillFunc, typename ResFn>
363 const FillFunc& fillFn,
365 Rec* rec = nullptr) {
366 // must wait for result to execute resFn -- so it must be synchronous
368 std::forward<OpFunc>(opFn),
376 // Allocate a record.
378 auto idx = recsPool_.allocIndex();
379 if (idx == NULL_INDEX) {
382 Rec& rec = recsPool_[idx];
388 void freeRec(Rec* rec) {
389 if (rec == nullptr) {
392 auto idx = rec->getIndex();
393 recsPool_.recycleIndex(idx);
396 // Returns the number of uncombined operations so far.
397 uint64_t getNumUncombined() const {
401 // Returns the number of combined operations so far.
402 uint64_t getNumCombined() const {
406 // Returns the number of combining passes so far.
407 uint64_t getNumPasses() const {
411 // Returns the number of combining sessions so far.
412 uint64_t getNumSessions() const {
417 const size_t NULL_INDEX = 0;
418 const uint32_t kDefaultMaxOps = 100;
419 const uint64_t kDefaultNumRecs = 64;
420 const uint64_t kIdleThreshold = 10;
422 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
425 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
426 folly::Baton<Atom, false, true> pending_;
427 Atom<bool> shutdown_{false};
429 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
434 std::thread combiner_;
437 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
438 uint64_t uncombined_ = 0;
439 uint64_t combined_ = 0;
440 uint64_t passes_ = 0;
441 uint64_t sessions_ = 0;
443 template <typename OpFunc, typename FillFunc, typename ResFn>
446 const FillFunc& fillFn,
451 std::unique_lock<Mutex> l(this->m_, std::defer_lock);
461 bool tc = (rec != nullptr);
463 // if an async op doesn't have a thread-cached record then turn
464 // it into a synchronous op.
468 if (rec == nullptr) {
469 // Can't use FC - Must acquire lock
478 // Wait if record is in use
483 // Fill the request (custom)
484 Req& req = rec->getReq();
488 rec->setFn(std::forward<OpFunc>(opFn));
490 // Indicate that record is valid
491 assert(!rec->isValid());
493 // end of combining critical path
495 // store-load order setValid before isDisconnected
496 std::atomic_thread_fence(std::memory_order_seq_cst);
497 if (rec->isDisconnected()) {
498 rec->clearDisconnected();
499 pushRec(rec->getIndex());
502 // If synchronous wait for the request to be completed
506 Req& req = rec->getReq();
507 resFn(req); // Extract the result (custom)
510 freeRec(rec); // Free the temporary record.
515 void pushRec(size_t idx) {
516 Rec& rec = recsPool_[idx];
518 auto head = recs_.load(std::memory_order_acquire);
519 rec.setNext(head); // there shouldn't be a data race here
520 if (recs_.compare_exchange_weak(head, idx)) {
526 size_t getRecsHead() {
527 return recs_.load(std::memory_order_acquire);
530 size_t nextIndex(size_t idx) {
531 return recsPool_[idx].getNext();
534 void clearPending() {
542 bool isPending() const {
543 return pending_.try_wait();
546 void awaitPending() {
550 uint64_t combiningSession() {
551 uint64_t combined = 0;
553 uint64_t count = static_cast<T*>(this)->combiningPass();
559 } while (combined < this->maxOps_);
563 void tryCombining() {
565 while (isPending()) {
568 combined_ += combiningSession();
573 void dedicatedCombining() {
577 if (shutdown_.load()) {
584 std::lock_guard<Mutex> guard(m_);
585 count = combiningSession();
588 if (count < maxOps_) {
595 void awaitDone(Rec& rec) {
599 awaitDoneTryLock(rec);
603 /// Waits for the request to be done and occasionally tries to
604 /// acquire the lock and to do combining. Used only in the absence
605 /// of a dedicated combiner.
606 void awaitDoneTryLock(Rec& rec) {
609 while (!rec.isDone()) {
611 std::unique_lock<Mutex> l(m_, std::defer_lock);
617 folly::asm_volatile_pause();
618 if (++count == 1000) {
626 shutdown_.store(true);
630 /// The following member functions may be overridden for customization
632 void combinedOp(Req&) {
633 throw std::runtime_error(
634 "FlatCombining::combinedOp(Req&) must be overridden in the derived"
635 " class if called.");
638 void processReq(Rec& rec) {
639 SavedFn& opFn = rec.getFn();
645 Req& req = rec.getReq();
646 static_cast<T*>(this)->combinedOp(req); // defined in derived class
648 rec.setLast(passes_);
652 uint64_t combiningPass() {
654 auto idx = getRecsHead();
656 while (idx != NULL_INDEX) {
657 Rec& rec = recsPool_[idx];
658 auto next = rec.getNext();
659 bool valid = rec.isValid();
660 if (!valid && (passes_ - rec.getLast() > kIdleThreshold) &&
664 rec.setDisconnected();
665 // store-load order setDisconnected before isValid
666 std::atomic_thread_fence(std::memory_order_seq_cst);
667 valid = rec.isValid();
681 } // namespace folly {