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.
17 #ifndef FOLLY_ATOMICHASHARRAY_H_
18 #error "This should only be included by AtomicHashArray.h"
21 #include <folly/Bits.h>
22 #include <folly/detail/AtomicHashUtils.h>
26 // AtomicHashArray private constructor --
27 template <class KeyT, class ValueT,
28 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
29 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
30 AtomicHashArray(size_t capacity, KeyT emptyKey, KeyT lockedKey,
31 KeyT erasedKey, double _maxLoadFactor, size_t cacheSize)
32 : capacity_(capacity),
33 maxEntries_(size_t(_maxLoadFactor * capacity_ + 0.5)),
34 kEmptyKey_(emptyKey), kLockedKey_(lockedKey), kErasedKey_(erasedKey),
35 kAnchorMask_(nextPowTwo(capacity_) - 1), numEntries_(0, cacheSize),
36 numPendingEntries_(0, cacheSize), isFull_(0), numErases_(0) {
42 * Sets ret.second to value found and ret.index to index
43 * of key and returns true, or if key does not exist returns false and
44 * ret.index is set to capacity_.
46 template <class KeyT, class ValueT,
47 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
48 typename AtomicHashArray<KeyT, ValueT,
49 HashFcn, EqualFcn, Allocator, ProbeFcn>::SimpleRetT
50 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
51 findInternal(const KeyT key_in) {
52 DCHECK_NE(key_in, kEmptyKey_);
53 DCHECK_NE(key_in, kLockedKey_);
54 DCHECK_NE(key_in, kErasedKey_);
55 for (size_t idx = keyToAnchorIdx(key_in), numProbes = 0;
57 idx = ProbeFcn()(idx, numProbes, capacity_)) {
58 const KeyT key = acquireLoadKey(cells_[idx]);
59 if (LIKELY(EqualFcn()(key, key_in))) {
60 return SimpleRetT(idx, true);
62 if (UNLIKELY(key == kEmptyKey_)) {
63 // if we hit an empty element, this key does not exist
64 return SimpleRetT(capacity_, false);
66 // NOTE: the way we count numProbes must be same in find(), insert(),
67 // and erase(). Otherwise it may break probing.
69 if (UNLIKELY(numProbes >= capacity_)) {
70 // probed every cell...fail
71 return SimpleRetT(capacity_, false);
79 * Returns false on failure due to key collision or full.
80 * Also sets ret.index to the index of the key. If the map is full, sets
81 * ret.index = capacity_. Also sets ret.second to cell value, thus if insert
82 * successful this will be what we just inserted, if there is a key collision
83 * this will be the previously inserted value, and if the map is full it is
86 template <class KeyT, class ValueT,
87 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
88 template <typename... ArgTs>
89 typename AtomicHashArray<KeyT, ValueT,
90 HashFcn, EqualFcn, Allocator, ProbeFcn>::SimpleRetT
91 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
92 insertInternal(KeyT key_in, ArgTs&&... vCtorArgs) {
93 const short NO_NEW_INSERTS = 1;
94 const short NO_PENDING_INSERTS = 2;
95 CHECK_NE(key_in, kEmptyKey_);
96 CHECK_NE(key_in, kLockedKey_);
97 CHECK_NE(key_in, kErasedKey_);
99 size_t idx = keyToAnchorIdx(key_in);
100 size_t numProbes = 0;
102 DCHECK_LT(idx, capacity_);
103 value_type* cell = &cells_[idx];
104 if (relaxedLoadKey(*cell) == kEmptyKey_) {
105 // NOTE: isFull_ is set based on numEntries_.readFast(), so it's
106 // possible to insert more than maxEntries_ entries. However, it's not
107 // possible to insert past capacity_.
108 ++numPendingEntries_;
109 if (isFull_.load(std::memory_order_acquire)) {
110 --numPendingEntries_;
112 // Before deciding whether this insert succeeded, this thread needs to
113 // wait until no other thread can add a new entry.
115 // Correctness assumes isFull_ is true at this point. If
116 // another thread now does ++numPendingEntries_, we expect it
117 // to pass the isFull_.load() test above. (It shouldn't insert
119 detail::atomic_hash_spin_wait([&] {
121 (isFull_.load(std::memory_order_acquire) != NO_PENDING_INSERTS) &&
122 (numPendingEntries_.readFull() != 0);
124 isFull_.store(NO_PENDING_INSERTS, std::memory_order_release);
126 if (relaxedLoadKey(*cell) == kEmptyKey_) {
127 // Don't insert past max load factor
128 return SimpleRetT(capacity_, false);
131 // An unallocated cell. Try once to lock it. If we succeed, insert here.
132 // If we fail, fall through to comparison below; maybe the insert that
133 // just beat us was for this very key....
134 if (tryLockCell(cell)) {
135 // Write the value - done before unlocking
137 DCHECK(relaxedLoadKey(*cell) == kLockedKey_);
138 new (&cell->second) ValueT(std::forward<ArgTs>(vCtorArgs)...);
139 unlockCell(cell, key_in); // Sets the new key
141 // Transition back to empty key---requires handling
142 // locked->empty below.
143 unlockCell(cell, kEmptyKey_);
144 --numPendingEntries_;
147 // Direct comparison rather than EqualFcn ok here
148 // (we just inserted it)
149 DCHECK(relaxedLoadKey(*cell) == key_in);
150 --numPendingEntries_;
151 ++numEntries_; // This is a thread cached atomic increment :)
152 if (numEntries_.readFast() >= maxEntries_) {
153 isFull_.store(NO_NEW_INSERTS, std::memory_order_relaxed);
155 return SimpleRetT(idx, true);
157 --numPendingEntries_;
160 DCHECK(relaxedLoadKey(*cell) != kEmptyKey_);
161 if (kLockedKey_ == acquireLoadKey(*cell)) {
162 detail::atomic_hash_spin_wait([&] {
163 return kLockedKey_ == acquireLoadKey(*cell);
167 const KeyT thisKey = acquireLoadKey(*cell);
168 if (EqualFcn()(thisKey, key_in)) {
169 // Found an existing entry for our key, but we don't overwrite the
171 return SimpleRetT(idx, false);
172 } else if (thisKey == kEmptyKey_ || thisKey == kLockedKey_) {
173 // We need to try again (i.e., don't increment numProbes or
174 // advance idx): this case can happen if the constructor for
175 // ValueT threw for this very cell (the rethrow block above).
180 // NOTE: the way we count numProbes must be same in find(),
181 // insert(), and erase(). Otherwise it may break probing.
183 if (UNLIKELY(numProbes >= capacity_)) {
184 // probed every cell...fail
185 return SimpleRetT(capacity_, false);
188 idx = ProbeFcn()(idx, numProbes, capacity_);
196 * This will attempt to erase the given key key_in if the key is found. It
197 * returns 1 iff the key was located and marked as erased, and 0 otherwise.
199 * Memory is not freed or reclaimed by erase, i.e. the cell containing the
200 * erased key will never be reused. If there's an associated value, we won't
203 template <class KeyT, class ValueT,
204 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
205 size_t AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
207 CHECK_NE(key_in, kEmptyKey_);
208 CHECK_NE(key_in, kLockedKey_);
209 CHECK_NE(key_in, kErasedKey_);
211 for (size_t idx = keyToAnchorIdx(key_in), numProbes = 0;
213 idx = ProbeFcn()(idx, numProbes, capacity_)) {
214 DCHECK_LT(idx, capacity_);
215 value_type* cell = &cells_[idx];
216 KeyT currentKey = acquireLoadKey(*cell);
217 if (currentKey == kEmptyKey_ || currentKey == kLockedKey_) {
218 // If we hit an empty (or locked) element, this key does not exist. This
219 // is similar to how it's handled in find().
222 if (EqualFcn()(currentKey, key_in)) {
223 // Found an existing entry for our key, attempt to mark it erased.
224 // Some other thread may have erased our key, but this is ok.
225 KeyT expect = currentKey;
226 if (cellKeyPtr(*cell)->compare_exchange_strong(expect, kErasedKey_)) {
227 numErases_.fetch_add(1, std::memory_order_relaxed);
229 // Even if there's a value in the cell, we won't delete (or even
230 // default construct) it because some other thread may be accessing it.
231 // Locking it meanwhile won't work either since another thread may be
232 // holding a pointer to it.
234 // We found the key and successfully erased it.
237 // If another thread succeeds in erasing our key, we'll stop our search.
241 // NOTE: the way we count numProbes must be same in find(), insert(),
242 // and erase(). Otherwise it may break probing.
244 if (UNLIKELY(numProbes >= capacity_)) {
245 // probed every cell...fail
251 template <class KeyT, class ValueT,
252 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
253 typename AtomicHashArray<KeyT, ValueT,
254 HashFcn, EqualFcn, Allocator, ProbeFcn>::SmartPtr
255 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
256 create(size_t maxSize, const Config& c) {
257 CHECK_LE(c.maxLoadFactor, 1.0);
258 CHECK_GT(c.maxLoadFactor, 0.0);
259 CHECK_NE(c.emptyKey, c.lockedKey);
260 size_t capacity = size_t(maxSize / c.maxLoadFactor);
261 size_t sz = sizeof(AtomicHashArray) + sizeof(value_type) * capacity;
263 auto const mem = Allocator().allocate(sz);
265 new (mem) AtomicHashArray(capacity, c.emptyKey, c.lockedKey, c.erasedKey,
266 c.maxLoadFactor, c.entryCountThreadCacheSize);
268 Allocator().deallocate(mem, sz);
272 SmartPtr map(static_cast<AtomicHashArray*>((void *)mem));
275 * Mark all cells as empty.
277 * Note: we're bending the rules a little here accessing the key
278 * element in our cells even though the cell object has not been
279 * constructed, and casting them to atomic objects (see cellKeyPtr).
280 * (Also, in fact we never actually invoke the value_type
281 * constructor.) This is in order to avoid needing to default
282 * construct a bunch of value_type when we first start up: if you
283 * have an expensive default constructor for the value type this can
284 * noticeably speed construction time for an AHA.
286 FOR_EACH_RANGE(i, 0, map->capacity_) {
287 cellKeyPtr(map->cells_[i])->store(map->kEmptyKey_,
288 std::memory_order_relaxed);
293 template <class KeyT, class ValueT,
294 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
295 void AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
296 destroy(AtomicHashArray* p) {
299 size_t sz = sizeof(AtomicHashArray) + sizeof(value_type) * p->capacity_;
301 FOR_EACH_RANGE(i, 0, p->capacity_) {
302 if (p->cells_[i].first != p->kEmptyKey_) {
303 p->cells_[i].~value_type();
306 p->~AtomicHashArray();
308 Allocator().deallocate((char *)p, sz);
311 // clear -- clears all keys and values in the map and resets all counters
312 template <class KeyT, class ValueT,
313 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
314 void AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
316 FOR_EACH_RANGE(i, 0, capacity_) {
317 if (cells_[i].first != kEmptyKey_) {
318 cells_[i].~value_type();
319 *const_cast<KeyT*>(&cells_[i].first) = kEmptyKey_;
321 CHECK(cells_[i].first == kEmptyKey_);
324 numPendingEntries_.set(0);
325 isFull_.store(0, std::memory_order_relaxed);
326 numErases_.store(0, std::memory_order_relaxed);
330 // Iterator implementation
332 template <class KeyT, class ValueT,
333 class HashFcn, class EqualFcn, class Allocator, class ProbeFcn>
334 template <class ContT, class IterVal>
335 struct AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator, ProbeFcn>::
337 : boost::iterator_facade<aha_iterator<ContT,IterVal>,
339 boost::forward_traversal_tag>
341 explicit aha_iterator() : aha_(0) {}
343 // Conversion ctor for interoperability between const_iterator and
344 // iterator. The enable_if<> magic keeps us well-behaved for
345 // is_convertible<> (v. the iterator_facade documentation).
346 template<class OtherContT, class OtherVal>
347 aha_iterator(const aha_iterator<OtherContT,OtherVal>& o,
348 typename std::enable_if<
349 std::is_convertible<OtherVal*,IterVal*>::value >::type* = 0)
354 explicit aha_iterator(ContT* array, size_t offset)
359 // Returns unique index that can be used with findAt().
360 // WARNING: The following function will fail silently for hashtable
361 // with capacity > 2^32
362 uint32_t getIndex() const { return offset_; }
364 void advancePastEmpty() {
365 while (offset_ < aha_->capacity_ && !isValid()) {
371 friend class AtomicHashArray;
372 friend class boost::iterator_core_access;
379 bool equal(const aha_iterator& o) const {
380 return aha_ == o.aha_ && offset_ == o.offset_;
383 IterVal& dereference() const {
384 return aha_->cells_[offset_];
387 bool isValid() const {
388 KeyT key = acquireLoadKey(aha_->cells_[offset_]);
389 return key != aha_->kEmptyKey_ &&
390 key != aha_->kLockedKey_ &&
391 key != aha_->kErasedKey_;