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.
18 * AtomicHashArray is the building block for AtomicHashMap. It provides the
19 * core lock-free functionality, but is limitted by the fact that it cannot
20 * grow past it's initialization size and is a little more awkward (no public
21 * constructor, for example). If you're confident that you won't run out of
22 * space, don't mind the awkardness, and really need bare-metal performance,
23 * feel free to use AHA directly.
25 * Check out AtomicHashMap.h for more thorough documentation on perf and
26 * general pros and cons relative to other hash maps.
28 * @author Spencer Ahrens <sahrens@fb.com>
29 * @author Jordan DeLong <delong.j@fb.com>
32 #ifndef FOLLY_ATOMICHASHARRAY_H_
33 #define FOLLY_ATOMICHASHARRAY_H_
37 #include <boost/iterator/iterator_facade.hpp>
38 #include <boost/noncopyable.hpp>
40 #include <folly/Hash.h>
41 #include <folly/ThreadCachedInt.h>
45 struct AtomicHashArrayLinearProbeFcn
47 inline size_t operator()(size_t idx, size_t numProbes, size_t capacity) const{
48 idx += 1; // linear probing
50 // Avoid modulus because it's slow
51 return LIKELY(idx < capacity) ? idx : (idx - capacity);
55 struct AtomicHashArrayQuadraticProbeFcn
57 inline size_t operator()(size_t idx, size_t numProbes, size_t capacity) const{
58 idx += numProbes; // quadratic probing
60 // Avoid modulus because it's slow
61 return LIKELY(idx < capacity) ? idx : (idx - capacity);
65 template <class KeyT, class ValueT,
66 class HashFcn = std::hash<KeyT>,
67 class EqualFcn = std::equal_to<KeyT>,
68 class Allocator = std::allocator<char>,
69 class ProbeFcn = AtomicHashArrayLinearProbeFcn>
72 template <class KeyT, class ValueT,
73 class HashFcn = std::hash<KeyT>,
74 class EqualFcn = std::equal_to<KeyT>,
75 class Allocator = std::allocator<char>,
76 class ProbeFcn = AtomicHashArrayLinearProbeFcn>
77 class AtomicHashArray : boost::noncopyable {
78 static_assert((std::is_convertible<KeyT,int32_t>::value ||
79 std::is_convertible<KeyT,int64_t>::value ||
80 std::is_convertible<KeyT,const void*>::value),
81 "You are trying to use AtomicHashArray with disallowed key "
82 "types. You must use atomically compare-and-swappable integer "
83 "keys, or a different container class.");
85 typedef KeyT key_type;
86 typedef ValueT mapped_type;
87 typedef std::pair<const KeyT, ValueT> value_type;
88 typedef std::size_t size_type;
89 typedef std::ptrdiff_t difference_type;
90 typedef value_type& reference;
91 typedef const value_type& const_reference;
92 typedef value_type* pointer;
93 typedef const value_type* const_pointer;
95 const size_t capacity_;
96 const size_t maxEntries_;
97 const KeyT kEmptyKey_;
98 const KeyT kLockedKey_;
99 const KeyT kErasedKey_;
101 template<class ContT, class IterVal>
104 typedef aha_iterator<const AtomicHashArray,const value_type> const_iterator;
105 typedef aha_iterator<AtomicHashArray,value_type> iterator;
107 // You really shouldn't need this if you use the SmartPtr provided by create,
108 // but if you really want to do something crazy like stick the released
109 // pointer into a DescriminatedPtr or something, you'll need this to clean up
111 static void destroy(AtomicHashArray*);
114 const size_t kAnchorMask_;
117 void operator()(AtomicHashArray* ptr) {
118 AtomicHashArray::destroy(ptr);
123 typedef std::unique_ptr<AtomicHashArray, Deleter> SmartPtr;
128 * Creates AtomicHashArray objects. Use instead of constructor/destructor.
130 * We do things this way in order to avoid the perf penalty of a second
131 * pointer indirection when composing these into AtomicHashMap, which needs
132 * to store an array of pointers so that it can perform atomic operations on
135 * Instead of a mess of arguments, we take a max size and a Config struct to
136 * simulate named ctor parameters. The Config struct has sensible defaults
137 * for everything, but is overloaded - if you specify a positive capacity,
138 * that will be used directly instead of computing it based on
141 * Create returns an AHA::SmartPtr which is a unique_ptr with a custom
142 * deleter to make sure everything is cleaned up properly.
148 double maxLoadFactor;
150 int entryCountThreadCacheSize;
151 size_t capacity; // if positive, overrides maxLoadFactor
154 // Cannot have constexpr ctor because some compilers rightly complain.
155 Config() : emptyKey((KeyT)-1),
160 entryCountThreadCacheSize(1000),
164 // Cannot have pre-instantiated const Config instance because of SIOF.
165 static SmartPtr create(size_t maxSize, const Config& c = Config());
167 iterator find(KeyT k) {
168 return iterator(this, findInternal(k).idx);
170 const_iterator find(KeyT k) const {
171 return const_cast<AtomicHashArray*>(this)->find(k);
177 * Returns a pair with iterator to the element at r.first and bool success.
178 * Retrieve the index with ret.first.getIndex().
180 * Fails on key collision (does not overwrite) or if map becomes
181 * full, at which point no element is inserted, iterator is set to end(),
182 * and success is set false. On collisions, success is set false, but the
183 * iterator is set to the existing entry.
185 std::pair<iterator,bool> insert(const value_type& r) {
186 return emplace(r.first, r.second);
188 std::pair<iterator,bool> insert(value_type&& r) {
189 return emplace(r.first, std::move(r.second));
195 * Same contract as insert(), but performs in-place construction
196 * of the value type using the specified arguments.
198 template <typename... ArgTs>
199 std::pair<iterator,bool> emplace(KeyT key_in, ArgTs&&... vCtorArgs) {
200 SimpleRetT ret = insertInternal(key_in, std::forward<ArgTs>(vCtorArgs)...);
201 return std::make_pair(iterator(this, ret.idx), ret.success);
204 // returns the number of elements erased - should never exceed 1
205 size_t erase(KeyT k);
207 // clears all keys and values in the map and resets all counters. Not thread
211 // Exact number of elements in the map - note that readFull() acquires a
212 // mutex. See folly/ThreadCachedInt.h for more details.
213 size_t size() const {
214 return numEntries_.readFull() -
215 numErases_.load(std::memory_order_relaxed);
218 bool empty() const { return size() == 0; }
221 iterator it(this, 0);
222 it.advancePastEmpty();
225 const_iterator begin() const {
226 const_iterator it(this, 0);
227 it.advancePastEmpty();
231 iterator end() { return iterator(this, capacity_); }
232 const_iterator end() const { return const_iterator(this, capacity_); }
234 // See AtomicHashMap::findAt - access elements directly
235 // WARNING: The following 2 functions will fail silently for hashtable
236 // with capacity > 2^32
237 iterator findAt(uint32_t idx) {
238 DCHECK_LT(idx, capacity_);
239 return iterator(this, idx);
241 const_iterator findAt(uint32_t idx) const {
242 return const_cast<AtomicHashArray*>(this)->findAt(idx);
245 iterator makeIter(size_t idx) { return iterator(this, idx); }
246 const_iterator makeIter(size_t idx) const {
247 return const_iterator(this, idx);
250 // The max load factor allowed for this map
251 double maxLoadFactor() const { return ((double) maxEntries_) / capacity_; }
253 void setEntryCountThreadCacheSize(uint32_t newSize) {
254 numEntries_.setCacheSize(newSize);
255 numPendingEntries_.setCacheSize(newSize);
258 int getEntryCountThreadCacheSize() const {
259 return numEntries_.getCacheSize();
262 /* Private data and helper functions... */
265 friend class AtomicHashMap<KeyT,
272 struct SimpleRetT { size_t idx; bool success;
273 SimpleRetT(size_t i, bool s) : idx(i), success(s) {}
274 SimpleRetT() = default;
279 template <typename... ArgTs>
280 SimpleRetT insertInternal(KeyT key, ArgTs&&... vCtorArgs);
282 SimpleRetT findInternal(const KeyT key);
284 static std::atomic<KeyT>* cellKeyPtr(const value_type& r) {
285 // We need some illegal casting here in order to actually store
286 // our value_type as a std::pair<const,>. But a little bit of
287 // undefined behavior never hurt anyone ...
288 static_assert(sizeof(std::atomic<KeyT>) == sizeof(KeyT),
289 "std::atomic is implemented in an unexpected way for AHM");
291 const_cast<std::atomic<KeyT>*>(
292 reinterpret_cast<std::atomic<KeyT> const*>(&r.first));
295 static KeyT relaxedLoadKey(const value_type& r) {
296 return cellKeyPtr(r)->load(std::memory_order_relaxed);
299 static KeyT acquireLoadKey(const value_type& r) {
300 return cellKeyPtr(r)->load(std::memory_order_acquire);
303 // Fun with thread local storage - atomic increment is expensive
304 // (relatively), so we accumulate in the thread cache and periodically
305 // flush to the actual variable, and walk through the unflushed counts when
306 // reading the value, so be careful of calling size() too frequently. This
307 // increases insertion throughput several times over while keeping the count
309 ThreadCachedInt<uint64_t> numEntries_; // Successful key inserts
310 ThreadCachedInt<uint64_t> numPendingEntries_; // Used by insertInternal
311 std::atomic<int64_t> isFull_; // Used by insertInternal
312 std::atomic<int64_t> numErases_; // Successful key erases
314 value_type cells_[0]; // This must be the last field of this class
316 // Force constructor/destructor private since create/destroy should be
317 // used externally instead
318 AtomicHashArray(size_t capacity, KeyT emptyKey, KeyT lockedKey,
319 KeyT erasedKey, double maxLoadFactor, size_t cacheSize);
321 ~AtomicHashArray() = default;
323 inline void unlockCell(value_type* const cell, KeyT newKey) {
324 cellKeyPtr(*cell)->store(newKey, std::memory_order_release);
327 inline bool tryLockCell(value_type* const cell) {
328 KeyT expect = kEmptyKey_;
329 return cellKeyPtr(*cell)->compare_exchange_strong(expect, kLockedKey_,
330 std::memory_order_acq_rel);
333 inline size_t keyToAnchorIdx(const KeyT k) const {
334 const size_t hashVal = HashFcn()(k);
335 const size_t probe = hashVal & kAnchorMask_;
336 return LIKELY(probe < capacity_) ? probe : hashVal % capacity_;
340 }; // AtomicHashArray
344 #include <folly/AtomicHashArray-inl.h>
346 #endif // FOLLY_ATOMICHASHARRAY_H_