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).
88 int NumLocalLists_ = 32,
89 int LocalListLimit_ = 200,
90 template<typename> class Atom = std::atomic,
91 bool EagerRecycleWhenTrivial = false,
92 bool EagerRecycleWhenNotTrivial = true>
93 struct IndexedMemPool : boost::noncopyable {
96 typedef std::unique_ptr<T, detail::IndexedMemPoolRecycler<IndexedMemPool>>
99 static_assert(LocalListLimit_ <= 255, "LocalListLimit must fit in 8 bits");
101 NumLocalLists = NumLocalLists_,
102 LocalListLimit = LocalListLimit_
106 static constexpr bool eagerRecycle() {
107 return std::is_trivial<T>::value
108 ? EagerRecycleWhenTrivial : EagerRecycleWhenNotTrivial;
111 // these are public because clients may need to reason about the number
112 // of bits required to hold indices from a pool, given its capacity
114 static constexpr uint32_t maxIndexForCapacity(uint32_t capacity) {
115 // index of std::numeric_limits<uint32_t>::max() is reserved for isAllocated
117 return uint32_t(std::min(
118 uint64_t(capacity) + (NumLocalLists - 1) * LocalListLimit,
119 uint64_t(std::numeric_limits<uint32_t>::max() - 1)));
122 static constexpr uint32_t capacityForMaxIndex(uint32_t maxIndex) {
123 return maxIndex - (NumLocalLists - 1) * LocalListLimit;
127 /// Constructs a pool that can allocate at least _capacity_ elements,
128 /// even if all the local lists are full
129 explicit IndexedMemPool(uint32_t capacity)
130 : actualCapacity_(maxIndexForCapacity(capacity))
132 , globalHead_(TaggedPtr{})
134 const size_t needed = sizeof(Slot) * (actualCapacity_ + 1);
135 size_t pagesize = size_t(sysconf(_SC_PAGESIZE));
136 mmapLength_ = ((needed - 1) & ~(pagesize - 1)) + pagesize;
137 assert(needed <= mmapLength_ && mmapLength_ < needed + pagesize);
138 assert((mmapLength_ % pagesize) == 0);
140 slots_ = static_cast<Slot*>(mmap(nullptr, mmapLength_,
141 PROT_READ | PROT_WRITE,
142 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0));
143 if (slots_ == MAP_FAILED) {
144 assert(errno == ENOMEM);
145 throw std::bad_alloc();
149 /// Destroys all of the contained elements
151 if (!eagerRecycle()) {
152 for (size_t i = size_; i > 0; --i) {
156 munmap(slots_, mmapLength_);
159 /// Returns a lower bound on the number of elements that may be
160 /// simultaneously allocated and not yet recycled. Because of the
161 /// local lists it is possible that more elements than this are returned
164 return capacityForMaxIndex(actualCapacity_);
167 /// Finds a slot with a non-zero index, emplaces a T there if we're
168 /// using the eager recycle lifecycle mode, and returns the index,
169 /// or returns 0 if no elements are available.
170 template <typename ...Args>
171 uint32_t allocIndex(Args&&... args) {
172 static_assert(sizeof...(Args) == 0 || eagerRecycle(),
173 "emplace-style allocation requires eager recycle, "
174 "which is defaulted only for non-trivial types");
175 auto idx = localPop(localHead());
176 if (idx != 0 && eagerRecycle()) {
177 T* ptr = &slot(idx).elem;
178 new (ptr) T(std::forward<Args>(args)...);
183 /// If an element is available, returns a std::unique_ptr to it that will
184 /// recycle the element to the pool when it is reclaimed, otherwise returns
185 /// a null (falsy) std::unique_ptr
186 template <typename ...Args>
187 UniquePtr allocElem(Args&&... args) {
188 auto idx = allocIndex(std::forward<Args>(args)...);
189 T* ptr = idx == 0 ? nullptr : &slot(idx).elem;
190 return UniquePtr(ptr, typename UniquePtr::deleter_type(this));
193 /// Gives up ownership previously granted by alloc()
194 void recycleIndex(uint32_t idx) {
195 assert(isAllocated(idx));
196 if (eagerRecycle()) {
199 localPush(localHead(), idx);
202 /// Provides access to the pooled element referenced by idx
203 T& operator[](uint32_t idx) {
204 return slot(idx).elem;
207 /// Provides access to the pooled element referenced by idx
208 const T& operator[](uint32_t idx) const {
209 return slot(idx).elem;
212 /// If elem == &pool[idx], then pool.locateElem(elem) == idx. Also,
213 /// pool.locateElem(nullptr) == 0
214 uint32_t locateElem(const T* elem) const {
219 static_assert(std::is_standard_layout<Slot>::value, "offsetof needs POD");
221 auto slot = reinterpret_cast<const Slot*>(
222 reinterpret_cast<const char*>(elem) - offsetof(Slot, elem));
223 auto rv = uint32_t(slot - slots_);
225 // this assert also tests that rv is in range
226 assert(elem == &(*this)[rv]);
230 /// Returns true iff idx has been alloc()ed and not recycleIndex()ed
231 bool isAllocated(uint32_t idx) const {
232 return slot(idx).localNext == uint32_t(-1);
244 Slot() : localNext{}, globalNext{} {}
250 // size is bottom 8 bits, tag in top 24. g++'s code generation for
251 // bitfields seems to depend on the phase of the moon, plus we can
252 // do better because we can rely on other checks to avoid masking
257 SizeMask = (1U << SizeBits) - 1,
258 TagIncr = 1U << SizeBits,
261 uint32_t size() const {
262 return tagAndSize & SizeMask;
265 TaggedPtr withSize(uint32_t repl) const {
266 assert(repl <= LocalListLimit);
267 return TaggedPtr{ idx, (tagAndSize & ~SizeMask) | repl };
270 TaggedPtr withSizeIncr() const {
271 assert(size() < LocalListLimit);
272 return TaggedPtr{ idx, tagAndSize + 1 };
275 TaggedPtr withSizeDecr() const {
277 return TaggedPtr{ idx, tagAndSize - 1 };
280 TaggedPtr withIdx(uint32_t repl) const {
281 return TaggedPtr{ repl, tagAndSize + TagIncr };
284 TaggedPtr withEmpty() const {
285 return withIdx(0).withSize(0);
289 struct FOLLY_ALIGN_TO_AVOID_FALSE_SHARING LocalList {
290 AtomicStruct<TaggedPtr,Atom> head;
292 LocalList() : head(TaggedPtr{}) {}
297 /// the actual number of slots that we will allocate, to guarantee
298 /// that we will satisfy the capacity requested at construction time.
299 /// They will be numbered 1..actualCapacity_ (note the 1-based counting),
300 /// and occupy slots_[1..actualCapacity_].
301 size_t actualCapacity_;
303 /// the number of bytes allocated from mmap, which is a multiple of
304 /// the page size of the machine
307 /// this records the number of slots that have actually been constructed.
308 /// To allow use of atomic ++ instead of CAS, we let this overflow.
309 /// The actual number of constructed elements is min(actualCapacity_,
311 Atom<uint32_t> size_;
313 /// raw storage, only 1..min(size_,actualCapacity_) (inclusive) are
314 /// actually constructed. Note that slots_[0] is not constructed or used
315 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING Slot* slots_;
317 /// use AccessSpreader to find your list. We use stripes instead of
318 /// thread-local to avoid the need to grow or shrink on thread start
319 /// or join. These are heads of lists chained with localNext
320 LocalList local_[NumLocalLists];
322 /// this is the head of a list of node chained by globalNext, that are
323 /// themselves each the head of a list chained by localNext
324 FOLLY_ALIGN_TO_AVOID_FALSE_SHARING AtomicStruct<TaggedPtr,Atom> globalHead_;
326 ///////////// private methods
328 size_t slotIndex(uint32_t idx) const {
330 idx <= actualCapacity_ &&
331 idx <= size_.load(std::memory_order_acquire));
335 Slot& slot(uint32_t idx) {
336 return slots_[slotIndex(idx)];
339 const Slot& slot(uint32_t idx) const {
340 return slots_[slotIndex(idx)];
343 // localHead references a full list chained by localNext. s should
344 // reference slot(localHead), it is passed as a micro-optimization
345 void globalPush(Slot& s, uint32_t localHead) {
347 TaggedPtr gh = globalHead_.load(std::memory_order_acquire);
348 s.globalNext = gh.idx;
349 if (globalHead_.compare_exchange_strong(gh, gh.withIdx(localHead))) {
356 // idx references a single node
357 void localPush(AtomicStruct<TaggedPtr,Atom>& head, uint32_t idx) {
359 TaggedPtr h = head.load(std::memory_order_acquire);
363 if (h.size() == LocalListLimit) {
364 // push will overflow local list, steal it instead
365 if (head.compare_exchange_strong(h, h.withEmpty())) {
366 // steal was successful, put everything in the global list
371 // local list has space
372 if (head.compare_exchange_strong(h, h.withIdx(idx).withSizeIncr())) {
377 // h was updated by failing CAS
381 // returns 0 if empty
382 uint32_t globalPop() {
384 TaggedPtr gh = globalHead_.load(std::memory_order_acquire);
385 if (gh.idx == 0 || globalHead_.compare_exchange_strong(
386 gh, gh.withIdx(slot(gh.idx).globalNext))) {
387 // global list is empty, or pop was successful
393 // returns 0 if allocation failed
394 uint32_t localPop(AtomicStruct<TaggedPtr,Atom>& head) {
396 TaggedPtr h = head.load(std::memory_order_acquire);
398 // local list is non-empty, try to pop
399 Slot& s = slot(h.idx);
400 if (head.compare_exchange_strong(
401 h, h.withIdx(s.localNext).withSizeDecr())) {
403 s.localNext = uint32_t(-1);
409 uint32_t idx = globalPop();
411 // global list is empty, allocate and construct new slot
412 if (size_.load(std::memory_order_relaxed) >= actualCapacity_ ||
413 (idx = ++size_) > actualCapacity_) {
417 // default-construct it now if we aren't going to construct and
418 // destroy on each allocation
419 if (!eagerRecycle()) {
420 T* ptr = &slot(idx).elem;
423 slot(idx).localNext = uint32_t(-1);
428 if (head.compare_exchange_strong(
429 h, h.withIdx(s.localNext).withSize(LocalListLimit))) {
430 // global list moved to local list, keep head for us
431 s.localNext = uint32_t(-1);
434 // local bulk push failed, return idx to the global list and try again
439 AtomicStruct<TaggedPtr,Atom>& localHead() {
440 auto stripe = detail::AccessSpreader<Atom>::current(NumLocalLists);
441 return local_[stripe].head;
447 /// This is a stateful Deleter functor, which allows std::unique_ptr
448 /// to track elements allocated from an IndexedMemPool by tracking the
449 /// associated pool. See IndexedMemPool::allocElem.
450 template <typename Pool>
451 struct IndexedMemPoolRecycler {
454 explicit IndexedMemPoolRecycler(Pool* pool) : pool(pool) {}
456 IndexedMemPoolRecycler(const IndexedMemPoolRecycler<Pool>& rhs)
458 IndexedMemPoolRecycler& operator= (const IndexedMemPoolRecycler<Pool>& rhs)
461 void operator()(typename Pool::value_type* elem) const {
462 pool->recycleIndex(pool->locateElem(elem));
470 # pragma GCC diagnostic pop