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 <type_traits>
23 #include <boost/noncopyable.hpp>
24 #include <folly/AtomicStruct.h>
25 #include <folly/detail/CacheLocality.h>
26 #include <folly/portability/SysMman.h>
27 #include <folly/portability/Unistd.h>
29 // Ignore shadowing warnings within this file, so includers can use -Wshadow.
30 #pragma GCC diagnostic push
31 #pragma GCC diagnostic ignored "-Wshadow"
36 template <typename Pool>
37 struct IndexedMemPoolRecycler;
40 /// Instances of IndexedMemPool dynamically allocate and then pool their
41 /// element type (T), returning 4-byte integer indices that can be passed
42 /// to the pool's operator[] method to access or obtain pointers to the
43 /// actual elements. The memory backing items returned from the pool
44 /// will always be readable, even if items have been returned to the pool.
45 /// These two features are useful for lock-free algorithms. The indexing
46 /// behavior makes it easy to build tagged pointer-like-things, since
47 /// a large number of elements can be managed using fewer bits than a
48 /// full pointer. The access-after-free behavior makes it safe to read
49 /// from T-s even after they have been recycled, since it is guaranteed
50 /// that the memory won't have been returned to the OS and unmapped
51 /// (the algorithm must still use a mechanism to validate that the read
52 /// was correct, but it doesn't have to worry about page faults), and if
53 /// the elements use internal sequence numbers it can be guaranteed that
54 /// there won't be an ABA match due to the element being overwritten with
55 /// a different type that has the same bit pattern.
57 /// IndexedMemPool has two object lifecycle strategies. The first
58 /// is to construct objects when they are allocated from the pool and
59 /// destroy them when they are recycled. In this mode allocIndex and
60 /// allocElem have emplace-like semantics. In the second mode, objects
61 /// are default-constructed the first time they are removed from the pool,
62 /// and deleted when the pool itself is deleted. By default the first
63 /// mode is used for non-trivial T, and the second is used for trivial T.
65 /// IMPORTANT: Space for extra elements is allocated to account for those
66 /// that are inaccessible because they are in other local lists, so the
67 /// actual number of items that can be allocated ranges from capacity to
68 /// capacity + (NumLocalLists_-1)*LocalListLimit_. This is important if
69 /// you are trying to maximize the capacity of the pool while constraining
70 /// the bit size of the resulting pointers, because the pointers will
71 /// actually range up to the boosted capacity. See maxIndexForCapacity
72 /// and capacityForMaxIndex.
74 /// To avoid contention, NumLocalLists_ free lists of limited (less than
75 /// or equal to LocalListLimit_) size are maintained, and each thread
76 /// retrieves and returns entries from its associated local list. If the
77 /// local list becomes too large then elements are placed in bulk in a
78 /// global free list. This allows items to be efficiently recirculated
79 /// from consumers to producers. AccessSpreader is used to access the
80 /// local lists, so there is no performance advantage to having more
81 /// local lists than L1 caches.
83 /// The pool mmap-s the entire necessary address space when the pool is
84 /// constructed, but delays element construction. This means that only
85 /// elements that are actually returned to the caller get paged into the
86 /// process's resident set (RSS).
89 uint32_t NumLocalLists_ = 32,
90 uint32_t LocalListLimit_ = 200,
91 template <typename> class Atom = std::atomic,
92 bool EagerRecycleWhenTrivial = false,
93 bool EagerRecycleWhenNotTrivial = true>
94 struct IndexedMemPool : boost::noncopyable {
97 typedef std::unique_ptr<T, detail::IndexedMemPoolRecycler<IndexedMemPool>>
100 static_assert(LocalListLimit_ <= 255, "LocalListLimit must fit in 8 bits");
102 NumLocalLists = NumLocalLists_,
103 LocalListLimit = LocalListLimit_
107 static constexpr bool eagerRecycle() {
108 return std::is_trivial<T>::value
109 ? EagerRecycleWhenTrivial : EagerRecycleWhenNotTrivial;
112 // these are public because clients may need to reason about the number
113 // of bits required to hold indices from a pool, given its capacity
115 static constexpr uint32_t maxIndexForCapacity(uint32_t capacity) {
116 // index of std::numeric_limits<uint32_t>::max() is reserved for isAllocated
118 return uint32_t(std::min(
119 uint64_t(capacity) + (NumLocalLists - 1) * LocalListLimit,
120 uint64_t(std::numeric_limits<uint32_t>::max() - 1)));
123 static constexpr uint32_t capacityForMaxIndex(uint32_t maxIndex) {
124 return maxIndex - (NumLocalLists - 1) * LocalListLimit;
128 /// Constructs a pool that can allocate at least _capacity_ elements,
129 /// even if all the local lists are full
130 explicit IndexedMemPool(uint32_t capacity)
131 : actualCapacity_(maxIndexForCapacity(capacity))
133 , globalHead_(TaggedPtr{})
135 const size_t needed = sizeof(Slot) * (actualCapacity_ + 1);
136 size_t pagesize = size_t(sysconf(_SC_PAGESIZE));
137 mmapLength_ = ((needed - 1) & ~(pagesize - 1)) + pagesize;
138 assert(needed <= mmapLength_ && mmapLength_ < needed + pagesize);
139 assert((mmapLength_ % pagesize) == 0);
141 slots_ = static_cast<Slot*>(mmap(nullptr, mmapLength_,
142 PROT_READ | PROT_WRITE,
143 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0));
144 if (slots_ == MAP_FAILED) {
145 assert(errno == ENOMEM);
146 throw std::bad_alloc();
150 /// Destroys all of the contained elements
152 if (!eagerRecycle()) {
153 // Take the minimum since it is possible that size_ > actualCapacity_.
154 // This can happen if there are multiple concurrent requests
155 // when size_ == actualCapacity_ - 1.
156 uint32_t last = std::min(uint32_t(size_), uint32_t(actualCapacity_));
157 for (uint32_t i = last; i > 0; --i) {
161 munmap(slots_, mmapLength_);
164 /// Returns a lower bound on the number of elements that may be
165 /// simultaneously allocated and not yet recycled. Because of the
166 /// local lists it is possible that more elements than this are returned
168 uint32_t capacity() {
169 return capacityForMaxIndex(actualCapacity_);
172 /// Finds a slot with a non-zero index, emplaces a T there if we're
173 /// using the eager recycle lifecycle mode, and returns the index,
174 /// or returns 0 if no elements are available.
175 template <typename ...Args>
176 uint32_t allocIndex(Args&&... args) {
177 static_assert(sizeof...(Args) == 0 || eagerRecycle(),
178 "emplace-style allocation requires eager recycle, "
179 "which is defaulted only for non-trivial types");
180 auto idx = localPop(localHead());
181 if (idx != 0 && eagerRecycle()) {
182 T* ptr = &slot(idx).elem;
183 new (ptr) T(std::forward<Args>(args)...);
188 /// If an element is available, returns a std::unique_ptr to it that will
189 /// recycle the element to the pool when it is reclaimed, otherwise returns
190 /// a null (falsy) std::unique_ptr
191 template <typename ...Args>
192 UniquePtr allocElem(Args&&... args) {
193 auto idx = allocIndex(std::forward<Args>(args)...);
194 T* ptr = idx == 0 ? nullptr : &slot(idx).elem;
195 return UniquePtr(ptr, typename UniquePtr::deleter_type(this));
198 /// Gives up ownership previously granted by alloc()
199 void recycleIndex(uint32_t idx) {
200 assert(isAllocated(idx));
201 if (eagerRecycle()) {
204 localPush(localHead(), idx);
207 /// Provides access to the pooled element referenced by idx
208 T& operator[](uint32_t idx) {
209 return slot(idx).elem;
212 /// Provides access to the pooled element referenced by idx
213 const T& operator[](uint32_t idx) const {
214 return slot(idx).elem;
217 /// If elem == &pool[idx], then pool.locateElem(elem) == idx. Also,
218 /// pool.locateElem(nullptr) == 0
219 uint32_t locateElem(const T* elem) const {
224 static_assert(std::is_standard_layout<Slot>::value, "offsetof needs POD");
226 auto slot = reinterpret_cast<const Slot*>(
227 reinterpret_cast<const char*>(elem) - offsetof(Slot, elem));
228 auto rv = uint32_t(slot - slots_);
230 // this assert also tests that rv is in range
231 assert(elem == &(*this)[rv]);
235 /// Returns true iff idx has been alloc()ed and not recycleIndex()ed
236 bool isAllocated(uint32_t idx) const {
237 return slot(idx).localNext.load(std::memory_order_relaxed) == uint32_t(-1);
246 Atom<uint32_t> localNext;
247 Atom<uint32_t> globalNext;
249 Slot() : localNext{}, globalNext{} {}
255 // size is bottom 8 bits, tag in top 24. g++'s code generation for
256 // bitfields seems to depend on the phase of the moon, plus we can
257 // do better because we can rely on other checks to avoid masking
262 SizeMask = (1U << SizeBits) - 1,
263 TagIncr = 1U << SizeBits,
266 uint32_t size() const {
267 return tagAndSize & SizeMask;
270 TaggedPtr withSize(uint32_t repl) const {
271 assert(repl <= LocalListLimit);
272 return TaggedPtr{ idx, (tagAndSize & ~SizeMask) | repl };
275 TaggedPtr withSizeIncr() const {
276 assert(size() < LocalListLimit);
277 return TaggedPtr{ idx, tagAndSize + 1 };
280 TaggedPtr withSizeDecr() const {
282 return TaggedPtr{ idx, tagAndSize - 1 };
285 TaggedPtr withIdx(uint32_t repl) const {
286 return TaggedPtr{ repl, tagAndSize + TagIncr };
289 TaggedPtr withEmpty() const {
290 return withIdx(0).withSize(0);
294 struct FOLLY_ALIGN_TO_AVOID_FALSE_SHARING LocalList {
295 AtomicStruct<TaggedPtr,Atom> head;
297 LocalList() : head(TaggedPtr{}) {}
302 /// the number of bytes allocated from mmap, which is a multiple of
303 /// the page size of the machine
306 /// the actual number of slots that we will allocate, to guarantee
307 /// that we will satisfy the capacity requested at construction time.
308 /// They will be numbered 1..actualCapacity_ (note the 1-based counting),
309 /// and occupy slots_[1..actualCapacity_].
310 uint32_t actualCapacity_;
312 /// this records the number of slots that have actually been constructed.
313 /// To allow use of atomic ++ instead of CAS, we let this overflow.
314 /// The actual number of constructed elements is min(actualCapacity_,
316 Atom<uint32_t> size_;
318 /// raw storage, only 1..min(size_,actualCapacity_) (inclusive) are
319 /// actually constructed. Note that slots_[0] is not constructed or used
320 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING Slot* slots_;
322 /// use AccessSpreader to find your list. We use stripes instead of
323 /// thread-local to avoid the need to grow or shrink on thread start
324 /// or join. These are heads of lists chained with localNext
325 LocalList local_[NumLocalLists];
327 /// this is the head of a list of node chained by globalNext, that are
328 /// themselves each the head of a list chained by localNext
329 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING AtomicStruct<TaggedPtr,Atom> globalHead_;
331 ///////////// private methods
333 uint32_t slotIndex(uint32_t idx) const {
335 idx <= actualCapacity_ &&
336 idx <= size_.load(std::memory_order_acquire));
340 Slot& slot(uint32_t idx) {
341 return slots_[slotIndex(idx)];
344 const Slot& slot(uint32_t idx) const {
345 return slots_[slotIndex(idx)];
348 // localHead references a full list chained by localNext. s should
349 // reference slot(localHead), it is passed as a micro-optimization
350 void globalPush(Slot& s, uint32_t localHead) {
352 TaggedPtr gh = globalHead_.load(std::memory_order_acquire);
353 s.globalNext.store(gh.idx, std::memory_order_relaxed);
354 if (globalHead_.compare_exchange_strong(gh, gh.withIdx(localHead))) {
361 // idx references a single node
362 void localPush(AtomicStruct<TaggedPtr,Atom>& head, uint32_t idx) {
364 TaggedPtr h = head.load(std::memory_order_acquire);
366 s.localNext.store(h.idx, std::memory_order_relaxed);
368 if (h.size() == LocalListLimit) {
369 // push will overflow local list, steal it instead
370 if (head.compare_exchange_strong(h, h.withEmpty())) {
371 // steal was successful, put everything in the global list
376 // local list has space
377 if (head.compare_exchange_strong(h, h.withIdx(idx).withSizeIncr())) {
382 // h was updated by failing CAS
386 // returns 0 if empty
387 uint32_t globalPop() {
389 TaggedPtr gh = globalHead_.load(std::memory_order_acquire);
391 globalHead_.compare_exchange_strong(
394 slot(gh.idx).globalNext.load(std::memory_order_relaxed)))) {
395 // global list is empty, or pop was successful
401 // returns 0 if allocation failed
402 uint32_t localPop(AtomicStruct<TaggedPtr,Atom>& head) {
404 TaggedPtr h = head.load(std::memory_order_acquire);
406 // local list is non-empty, try to pop
407 Slot& s = slot(h.idx);
408 auto next = s.localNext.load(std::memory_order_relaxed);
409 if (head.compare_exchange_strong(h, h.withIdx(next).withSizeDecr())) {
411 s.localNext.store(uint32_t(-1), std::memory_order_relaxed);
417 uint32_t idx = globalPop();
419 // global list is empty, allocate and construct new slot
420 if (size_.load(std::memory_order_relaxed) >= actualCapacity_ ||
421 (idx = ++size_) > actualCapacity_) {
425 // default-construct it now if we aren't going to construct and
426 // destroy on each allocation
427 if (!eagerRecycle()) {
428 T* ptr = &slot(idx).elem;
431 slot(idx).localNext.store(uint32_t(-1), std::memory_order_relaxed);
436 auto next = s.localNext.load(std::memory_order_relaxed);
437 if (head.compare_exchange_strong(
438 h, h.withIdx(next).withSize(LocalListLimit))) {
439 // global list moved to local list, keep head for us
440 s.localNext.store(uint32_t(-1), std::memory_order_relaxed);
443 // local bulk push failed, return idx to the global list and try again
448 AtomicStruct<TaggedPtr,Atom>& localHead() {
449 auto stripe = detail::AccessSpreader<Atom>::current(NumLocalLists);
450 return local_[stripe].head;
456 /// This is a stateful Deleter functor, which allows std::unique_ptr
457 /// to track elements allocated from an IndexedMemPool by tracking the
458 /// associated pool. See IndexedMemPool::allocElem.
459 template <typename Pool>
460 struct IndexedMemPoolRecycler {
463 explicit IndexedMemPoolRecycler(Pool* pool) : pool(pool) {}
465 IndexedMemPoolRecycler(const IndexedMemPoolRecycler<Pool>& rhs)
467 IndexedMemPoolRecycler& operator= (const IndexedMemPoolRecycler<Pool>& rhs)
470 void operator()(typename Pool::value_type* elem) const {
471 pool->recycleIndex(pool->locateElem(elem));
479 # pragma GCC diagnostic pop