2 * Copyright 2012 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, class HashFcn>
28 AtomicHashArray<KeyT, ValueT, HashFcn>::
29 AtomicHashArray(size_t capacity, KeyT emptyKey, KeyT lockedKey,
30 KeyT erasedKey, double maxLoadFactor, size_t cacheSize)
31 : capacity_(capacity), maxEntries_(size_t(maxLoadFactor * capacity_ + 0.5)),
32 kEmptyKey_(emptyKey), kLockedKey_(lockedKey), kErasedKey_(erasedKey),
33 kAnchorMask_(nextPowTwo(capacity_) - 1), numEntries_(0, cacheSize),
34 numPendingEntries_(0, cacheSize), isFull_(0), numErases_(0) {
40 * Sets ret.second to value found and ret.index to index
41 * of key and returns true, or if key does not exist returns false and
42 * ret.index is set to capacity_.
44 template <class KeyT, class ValueT, class HashFcn>
45 typename AtomicHashArray<KeyT, ValueT, HashFcn>::SimpleRetT
46 AtomicHashArray<KeyT, ValueT, HashFcn>::
47 findInternal(const KeyT key_in) {
48 DCHECK_NE(key_in, kEmptyKey_);
49 DCHECK_NE(key_in, kLockedKey_);
50 DCHECK_NE(key_in, kErasedKey_);
51 for (size_t idx = keyToAnchorIdx(key_in), numProbes = 0;
53 idx = probeNext(idx, numProbes)) {
54 const KeyT key = acquireLoadKey(cells_[idx]);
55 if (LIKELY(key == key_in)) {
56 return SimpleRetT(idx, true);
58 if (UNLIKELY(key == kEmptyKey_)) {
59 // if we hit an empty element, this key does not exist
60 return SimpleRetT(capacity_, false);
63 if (UNLIKELY(numProbes >= capacity_)) {
64 // probed every cell...fail
65 return SimpleRetT(capacity_, false);
73 * Returns false on failure due to key collision or full.
74 * Also sets ret.index to the index of the key. If the map is full, sets
75 * ret.index = capacity_. Also sets ret.second to cell value, thus if insert
76 * successful this will be what we just inserted, if there is a key collision
77 * this will be the previously inserted value, and if the map is full it is
80 template <class KeyT, class ValueT, class HashFcn>
82 typename AtomicHashArray<KeyT, ValueT, HashFcn>::SimpleRetT
83 AtomicHashArray<KeyT, ValueT, HashFcn>::
84 insertInternal(KeyT key_in, T&& value) {
85 const short NO_NEW_INSERTS = 1;
86 const short NO_PENDING_INSERTS = 2;
87 CHECK_NE(key_in, kEmptyKey_);
88 CHECK_NE(key_in, kLockedKey_);
89 CHECK_NE(key_in, kErasedKey_);
91 size_t idx = keyToAnchorIdx(key_in);
94 DCHECK_LT(idx, capacity_);
95 value_type* cell = &cells_[idx];
96 if (relaxedLoadKey(*cell) == kEmptyKey_) {
97 // NOTE: isFull_ is set based on numEntries_.readFast(), so it's
98 // possible to insert more than maxEntries_ entries. However, it's not
99 // possible to insert past capacity_.
100 ++numPendingEntries_;
101 if (isFull_.load(std::memory_order_acquire)) {
102 --numPendingEntries_;
104 // Before deciding whether this insert succeeded, this thread needs to
105 // wait until no other thread can add a new entry.
107 // Correctness assumes isFull_ is true at this point. If
108 // another thread now does ++numPendingEntries_, we expect it
109 // to pass the isFull_.load() test above. (It shouldn't insert
112 isFull_.load(std::memory_order_acquire) != NO_PENDING_INSERTS
113 && numPendingEntries_.readFull() != 0
115 isFull_.store(NO_PENDING_INSERTS, std::memory_order_release);
117 if (relaxedLoadKey(*cell) == kEmptyKey_) {
118 // Don't insert past max load factor
119 return SimpleRetT(capacity_, false);
122 // An unallocated cell. Try once to lock it. If we succeed, insert here.
123 // If we fail, fall through to comparison below; maybe the insert that
124 // just beat us was for this very key....
125 if (tryLockCell(cell)) {
126 // Write the value - done before unlocking
128 DCHECK(relaxedLoadKey(*cell) == kLockedKey_);
130 * This happens using the copy constructor because we won't have
131 * constructed a lhs to use an assignment operator on when
132 * values are being set.
134 new (&cell->second) ValueT(std::forward<T>(value));
135 unlockCell(cell, key_in); // Sets the new key
137 // Transition back to empty key---requires handling
138 // locked->empty below.
139 unlockCell(cell, kEmptyKey_);
140 --numPendingEntries_;
143 DCHECK(relaxedLoadKey(*cell) == key_in);
144 --numPendingEntries_;
145 ++numEntries_; // This is a thread cached atomic increment :)
146 if (numEntries_.readFast() >= maxEntries_) {
147 isFull_.store(NO_NEW_INSERTS, std::memory_order_relaxed);
149 return SimpleRetT(idx, true);
151 --numPendingEntries_;
154 DCHECK(relaxedLoadKey(*cell) != kEmptyKey_);
155 if (kLockedKey_ == acquireLoadKey(*cell)) {
157 kLockedKey_ == acquireLoadKey(*cell)
161 const KeyT thisKey = acquireLoadKey(*cell);
162 if (thisKey == key_in) {
163 // Found an existing entry for our key, but we don't overwrite the
165 return SimpleRetT(idx, false);
166 } else if (thisKey == kEmptyKey_ || thisKey == kLockedKey_) {
167 // We need to try again (i.e., don't increment numProbes or
168 // advance idx): this case can happen if the constructor for
169 // ValueT threw for this very cell (the rethrow block above).
174 if (UNLIKELY(numProbes >= capacity_)) {
175 // probed every cell...fail
176 return SimpleRetT(capacity_, false);
179 idx = probeNext(idx, numProbes);
187 * This will attempt to erase the given key key_in if the key is found. It
188 * returns 1 iff the key was located and marked as erased, and 0 otherwise.
190 * Memory is not freed or reclaimed by erase, i.e. the cell containing the
191 * erased key will never be reused. If there's an associated value, we won't
194 template <class KeyT, class ValueT, class HashFcn>
195 size_t AtomicHashArray<KeyT, ValueT, HashFcn>::
197 CHECK_NE(key_in, kEmptyKey_);
198 CHECK_NE(key_in, kLockedKey_);
199 CHECK_NE(key_in, kErasedKey_);
200 for (size_t idx = keyToAnchorIdx(key_in), numProbes = 0;
202 idx = probeNext(idx, numProbes)) {
203 DCHECK_LT(idx, capacity_);
204 value_type* cell = &cells_[idx];
205 KeyT currentKey = acquireLoadKey(*cell);
206 if (currentKey == kEmptyKey_ || currentKey == kLockedKey_) {
207 // If we hit an empty (or locked) element, this key does not exist. This
208 // is similar to how it's handled in find().
211 if (key_in == currentKey) {
212 // Found an existing entry for our key, attempt to mark it erased.
213 // Some other thread may have erased our key, but this is ok.
214 KeyT expect = key_in;
215 if (cellKeyPtr(*cell)->compare_exchange_strong(expect, kErasedKey_)) {
216 numErases_.fetch_add(1, std::memory_order_relaxed);
218 // Even if there's a value in the cell, we won't delete (or even
219 // default construct) it because some other thread may be accessing it.
220 // Locking it meanwhile won't work either since another thread may be
221 // holding a pointer to it.
223 // We found the key and successfully erased it.
226 // If another thread succeeds in erasing our key, we'll stop our search.
230 if (UNLIKELY(numProbes >= capacity_)) {
231 // probed every cell...fail
237 template <class KeyT, class ValueT, class HashFcn>
238 const typename AtomicHashArray<KeyT, ValueT, HashFcn>::Config
239 AtomicHashArray<KeyT, ValueT, HashFcn>::defaultConfig;
241 template <class KeyT, class ValueT, class HashFcn>
242 typename AtomicHashArray<KeyT, ValueT, HashFcn>::SmartPtr
243 AtomicHashArray<KeyT, ValueT, HashFcn>::
244 create(size_t maxSize, const Config& c) {
245 CHECK_LE(c.maxLoadFactor, 1.0);
246 CHECK_GT(c.maxLoadFactor, 0.0);
247 CHECK_NE(c.emptyKey, c.lockedKey);
248 size_t capacity = size_t(maxSize / c.maxLoadFactor);
249 size_t sz = sizeof(AtomicHashArray) + sizeof(value_type) * capacity;
251 std::unique_ptr<void, void(*)(void*)> mem(malloc(sz), free);
252 new(mem.get()) AtomicHashArray(capacity, c.emptyKey, c.lockedKey, c.erasedKey,
253 c.maxLoadFactor, c.entryCountThreadCacheSize);
254 SmartPtr map(static_cast<AtomicHashArray*>(mem.release()));
257 * Mark all cells as empty.
259 * Note: we're bending the rules a little here accessing the key
260 * element in our cells even though the cell object has not been
261 * constructed, and casting them to atomic objects (see cellKeyPtr).
262 * (Also, in fact we never actually invoke the value_type
263 * constructor.) This is in order to avoid needing to default
264 * construct a bunch of value_type when we first start up: if you
265 * have an expensive default constructor for the value type this can
266 * noticeably speed construction time for an AHA.
268 FOR_EACH_RANGE(i, 0, map->capacity_) {
269 cellKeyPtr(map->cells_[i])->store(map->kEmptyKey_,
270 std::memory_order_relaxed);
275 template <class KeyT, class ValueT, class HashFcn>
276 void AtomicHashArray<KeyT, ValueT, HashFcn>::
277 destroy(AtomicHashArray* p) {
279 FOR_EACH_RANGE(i, 0, p->capacity_) {
280 if (p->cells_[i].first != p->kEmptyKey_) {
281 p->cells_[i].~value_type();
284 p->~AtomicHashArray();
288 // clear -- clears all keys and values in the map and resets all counters
289 template <class KeyT, class ValueT, class HashFcn>
290 void AtomicHashArray<KeyT, ValueT, HashFcn>::
292 FOR_EACH_RANGE(i, 0, capacity_) {
293 if (cells_[i].first != kEmptyKey_) {
294 cells_[i].~value_type();
295 *const_cast<KeyT*>(&cells_[i].first) = kEmptyKey_;
297 CHECK(cells_[i].first == kEmptyKey_);
300 numPendingEntries_.set(0);
301 isFull_.store(0, std::memory_order_relaxed);
302 numErases_.store(0, std::memory_order_relaxed);
306 // Iterator implementation
308 template <class KeyT, class ValueT, class HashFcn>
309 template <class ContT, class IterVal>
310 struct AtomicHashArray<KeyT, ValueT, HashFcn>::aha_iterator
311 : boost::iterator_facade<aha_iterator<ContT,IterVal>,
313 boost::forward_traversal_tag>
315 explicit aha_iterator() : aha_(0) {}
317 // Conversion ctor for interoperability between const_iterator and
318 // iterator. The enable_if<> magic keeps us well-behaved for
319 // is_convertible<> (v. the iterator_facade documentation).
320 template<class OtherContT, class OtherVal>
321 aha_iterator(const aha_iterator<OtherContT,OtherVal>& o,
322 typename std::enable_if<
323 std::is_convertible<OtherVal*,IterVal*>::value >::type* = 0)
328 explicit aha_iterator(ContT* array, size_t offset)
335 // Returns unique index that can be used with findAt().
336 // WARNING: The following function will fail silently for hashtable
337 // with capacity > 2^32
338 uint32_t getIndex() const { return offset_; }
341 friend class AtomicHashArray;
342 friend class boost::iterator_core_access;
349 bool equal(const aha_iterator& o) const {
350 return aha_ == o.aha_ && offset_ == o.offset_;
353 IterVal& dereference() const {
354 return aha_->cells_[offset_];
357 void advancePastEmpty() {
358 while (offset_ < aha_->capacity_ && !isValid()) {
363 bool isValid() const {
364 KeyT key = acquireLoadKey(aha_->cells_[offset_]);
365 return key != aha_->kEmptyKey_ &&
366 key != aha_->kLockedKey_ &&
367 key != aha_->kErasedKey_;
377 #undef FOLLY_SPIN_WAIT