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>
29 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::
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>
48 typename AtomicHashArray<KeyT, ValueT,
49 HashFcn, EqualFcn, Allocator>::SimpleRetT
50 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::
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 = probeNext(idx, numProbes)) {
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);
67 if (UNLIKELY(numProbes >= capacity_)) {
68 // probed every cell...fail
69 return SimpleRetT(capacity_, false);
77 * Returns false on failure due to key collision or full.
78 * Also sets ret.index to the index of the key. If the map is full, sets
79 * ret.index = capacity_. Also sets ret.second to cell value, thus if insert
80 * successful this will be what we just inserted, if there is a key collision
81 * this will be the previously inserted value, and if the map is full it is
84 template <class KeyT, class ValueT,
85 class HashFcn, class EqualFcn, class Allocator>
87 typename AtomicHashArray<KeyT, ValueT,
88 HashFcn, EqualFcn, Allocator>::SimpleRetT
89 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::
90 insertInternal(KeyT key_in, T&& value) {
91 const short NO_NEW_INSERTS = 1;
92 const short NO_PENDING_INSERTS = 2;
93 CHECK_NE(key_in, kEmptyKey_);
94 CHECK_NE(key_in, kLockedKey_);
95 CHECK_NE(key_in, kErasedKey_);
97 size_t idx = keyToAnchorIdx(key_in);
100 DCHECK_LT(idx, capacity_);
101 value_type* cell = &cells_[idx];
102 if (relaxedLoadKey(*cell) == kEmptyKey_) {
103 // NOTE: isFull_ is set based on numEntries_.readFast(), so it's
104 // possible to insert more than maxEntries_ entries. However, it's not
105 // possible to insert past capacity_.
106 ++numPendingEntries_;
107 if (isFull_.load(std::memory_order_acquire)) {
108 --numPendingEntries_;
110 // Before deciding whether this insert succeeded, this thread needs to
111 // wait until no other thread can add a new entry.
113 // Correctness assumes isFull_ is true at this point. If
114 // another thread now does ++numPendingEntries_, we expect it
115 // to pass the isFull_.load() test above. (It shouldn't insert
117 detail::atomic_hash_spin_wait([&] {
119 (isFull_.load(std::memory_order_acquire) != NO_PENDING_INSERTS) &&
120 (numPendingEntries_.readFull() != 0);
122 isFull_.store(NO_PENDING_INSERTS, std::memory_order_release);
124 if (relaxedLoadKey(*cell) == kEmptyKey_) {
125 // Don't insert past max load factor
126 return SimpleRetT(capacity_, false);
129 // An unallocated cell. Try once to lock it. If we succeed, insert here.
130 // If we fail, fall through to comparison below; maybe the insert that
131 // just beat us was for this very key....
132 if (tryLockCell(cell)) {
133 // Write the value - done before unlocking
135 DCHECK(relaxedLoadKey(*cell) == kLockedKey_);
137 * This happens using the copy constructor because we won't have
138 * constructed a lhs to use an assignment operator on when
139 * values are being set.
141 new (&cell->second) ValueT(std::forward<T>(value));
142 unlockCell(cell, key_in); // Sets the new key
144 // Transition back to empty key---requires handling
145 // locked->empty below.
146 unlockCell(cell, kEmptyKey_);
147 --numPendingEntries_;
150 // Direct comparison rather than EqualFcn ok here
151 // (we just inserted it)
152 DCHECK(relaxedLoadKey(*cell) == key_in);
153 --numPendingEntries_;
154 ++numEntries_; // This is a thread cached atomic increment :)
155 if (numEntries_.readFast() >= maxEntries_) {
156 isFull_.store(NO_NEW_INSERTS, std::memory_order_relaxed);
158 return SimpleRetT(idx, true);
160 --numPendingEntries_;
163 DCHECK(relaxedLoadKey(*cell) != kEmptyKey_);
164 if (kLockedKey_ == acquireLoadKey(*cell)) {
165 detail::atomic_hash_spin_wait([&] {
166 return kLockedKey_ == acquireLoadKey(*cell);
170 const KeyT thisKey = acquireLoadKey(*cell);
171 if (EqualFcn()(thisKey, key_in)) {
172 // Found an existing entry for our key, but we don't overwrite the
174 return SimpleRetT(idx, false);
175 } else if (thisKey == kEmptyKey_ || thisKey == kLockedKey_) {
176 // We need to try again (i.e., don't increment numProbes or
177 // advance idx): this case can happen if the constructor for
178 // ValueT threw for this very cell (the rethrow block above).
183 if (UNLIKELY(numProbes >= capacity_)) {
184 // probed every cell...fail
185 return SimpleRetT(capacity_, false);
188 idx = probeNext(idx, numProbes);
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>
205 size_t AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::
207 CHECK_NE(key_in, kEmptyKey_);
208 CHECK_NE(key_in, kLockedKey_);
209 CHECK_NE(key_in, kErasedKey_);
210 for (size_t idx = keyToAnchorIdx(key_in), numProbes = 0;
212 idx = probeNext(idx, numProbes)) {
213 DCHECK_LT(idx, capacity_);
214 value_type* cell = &cells_[idx];
215 KeyT currentKey = acquireLoadKey(*cell);
216 if (currentKey == kEmptyKey_ || currentKey == kLockedKey_) {
217 // If we hit an empty (or locked) element, this key does not exist. This
218 // is similar to how it's handled in find().
221 if (EqualFcn()(currentKey, key_in)) {
222 // Found an existing entry for our key, attempt to mark it erased.
223 // Some other thread may have erased our key, but this is ok.
224 KeyT expect = currentKey;
225 if (cellKeyPtr(*cell)->compare_exchange_strong(expect, kErasedKey_)) {
226 numErases_.fetch_add(1, std::memory_order_relaxed);
228 // Even if there's a value in the cell, we won't delete (or even
229 // default construct) it because some other thread may be accessing it.
230 // Locking it meanwhile won't work either since another thread may be
231 // holding a pointer to it.
233 // We found the key and successfully erased it.
236 // If another thread succeeds in erasing our key, we'll stop our search.
240 if (UNLIKELY(numProbes >= capacity_)) {
241 // probed every cell...fail
247 template <class KeyT, class ValueT,
248 class HashFcn, class EqualFcn, class Allocator>
249 typename AtomicHashArray<KeyT, ValueT,
250 HashFcn, EqualFcn, Allocator>::SmartPtr
251 AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::
252 create(size_t maxSize, const Config& c) {
253 CHECK_LE(c.maxLoadFactor, 1.0);
254 CHECK_GT(c.maxLoadFactor, 0.0);
255 CHECK_NE(c.emptyKey, c.lockedKey);
256 size_t capacity = size_t(maxSize / c.maxLoadFactor);
257 size_t sz = sizeof(AtomicHashArray) + sizeof(value_type) * capacity;
259 auto const mem = Allocator().allocate(sz);
261 new (mem) AtomicHashArray(capacity, c.emptyKey, c.lockedKey, c.erasedKey,
262 c.maxLoadFactor, c.entryCountThreadCacheSize);
264 Allocator().deallocate(mem, sz);
268 SmartPtr map(static_cast<AtomicHashArray*>((void *)mem));
271 * Mark all cells as empty.
273 * Note: we're bending the rules a little here accessing the key
274 * element in our cells even though the cell object has not been
275 * constructed, and casting them to atomic objects (see cellKeyPtr).
276 * (Also, in fact we never actually invoke the value_type
277 * constructor.) This is in order to avoid needing to default
278 * construct a bunch of value_type when we first start up: if you
279 * have an expensive default constructor for the value type this can
280 * noticeably speed construction time for an AHA.
282 FOR_EACH_RANGE(i, 0, map->capacity_) {
283 cellKeyPtr(map->cells_[i])->store(map->kEmptyKey_,
284 std::memory_order_relaxed);
289 template <class KeyT, class ValueT,
290 class HashFcn, class EqualFcn, class Allocator>
291 void AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::
292 destroy(AtomicHashArray* p) {
295 size_t sz = sizeof(AtomicHashArray) + sizeof(value_type) * p->capacity_;
297 FOR_EACH_RANGE(i, 0, p->capacity_) {
298 if (p->cells_[i].first != p->kEmptyKey_) {
299 p->cells_[i].~value_type();
302 p->~AtomicHashArray();
304 Allocator().deallocate((char *)p, sz);
307 // clear -- clears all keys and values in the map and resets all counters
308 template <class KeyT, class ValueT,
309 class HashFcn, class EqualFcn, class Allocator>
310 void AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::
312 FOR_EACH_RANGE(i, 0, capacity_) {
313 if (cells_[i].first != kEmptyKey_) {
314 cells_[i].~value_type();
315 *const_cast<KeyT*>(&cells_[i].first) = kEmptyKey_;
317 CHECK(cells_[i].first == kEmptyKey_);
320 numPendingEntries_.set(0);
321 isFull_.store(0, std::memory_order_relaxed);
322 numErases_.store(0, std::memory_order_relaxed);
326 // Iterator implementation
328 template <class KeyT, class ValueT,
329 class HashFcn, class EqualFcn, class Allocator>
330 template <class ContT, class IterVal>
331 struct AtomicHashArray<KeyT, ValueT, HashFcn, EqualFcn, Allocator>::aha_iterator
332 : boost::iterator_facade<aha_iterator<ContT,IterVal>,
334 boost::forward_traversal_tag>
336 explicit aha_iterator() : aha_(0) {}
338 // Conversion ctor for interoperability between const_iterator and
339 // iterator. The enable_if<> magic keeps us well-behaved for
340 // is_convertible<> (v. the iterator_facade documentation).
341 template<class OtherContT, class OtherVal>
342 aha_iterator(const aha_iterator<OtherContT,OtherVal>& o,
343 typename std::enable_if<
344 std::is_convertible<OtherVal*,IterVal*>::value >::type* = 0)
349 explicit aha_iterator(ContT* array, size_t offset)
354 // Returns unique index that can be used with findAt().
355 // WARNING: The following function will fail silently for hashtable
356 // with capacity > 2^32
357 uint32_t getIndex() const { return offset_; }
359 void advancePastEmpty() {
360 while (offset_ < aha_->capacity_ && !isValid()) {
366 friend class AtomicHashArray;
367 friend class boost::iterator_core_access;
374 bool equal(const aha_iterator& o) const {
375 return aha_ == o.aha_ && offset_ == o.offset_;
378 IterVal& dereference() const {
379 return aha_->cells_[offset_];
382 bool isValid() const {
383 KeyT key = acquireLoadKey(aha_->cells_[offset_]);
384 return key != aha_->kEmptyKey_ &&
385 key != aha_->kLockedKey_ &&
386 key != aha_->kErasedKey_;