2 * Copyright 2016 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 * Improved thread local storage for non-trivial types (similar speed as
19 * pthread_getspecific but only consumes a single pthread_key_t, and 4x faster
20 * than boost::thread_specific_ptr).
22 * Also includes an accessor interface to walk all the thread local child
23 * objects of a parent. accessAllThreads() initializes an accessor which holds
24 * a global lock *that blocks all creation and destruction of ThreadLocal
25 * objects with the same Tag* and can be used as an iterable container.
27 * Intended use is for frequent write, infrequent read data access patterns such
30 * There are two classes here - ThreadLocal and ThreadLocalPtr. ThreadLocalPtr
31 * has semantics similar to boost::thread_specific_ptr. ThreadLocal is a thin
32 * wrapper around ThreadLocalPtr that manages allocation automatically.
34 * @author Spencer Ahrens (sahrens)
39 #include <folly/Likely.h>
40 #include <folly/Portability.h>
41 #include <folly/ScopeGuard.h>
42 #include <boost/iterator/iterator_facade.hpp>
43 #include <type_traits>
47 enum class TLPDestructionMode {
53 #include <folly/detail/ThreadLocalDetail.h>
57 template<class T, class Tag> class ThreadLocalPtr;
59 template<class T, class Tag=void>
62 constexpr ThreadLocal() : constructor_([]() {
66 explicit ThreadLocal(std::function<T*()> constructor) :
67 constructor_(constructor) {
72 if (LIKELY(ptr != nullptr)) {
76 // separated new item creation out to speed up the fast path.
80 T* operator->() const {
84 T& operator*() const {
88 void reset(T* newPtr = nullptr) {
92 typedef typename ThreadLocalPtr<T,Tag>::Accessor Accessor;
93 Accessor accessAllThreads() const {
94 return tlp_.accessAllThreads();
98 ThreadLocal(ThreadLocal&&) = default;
99 ThreadLocal& operator=(ThreadLocal&&) = default;
103 ThreadLocal(const ThreadLocal&) = delete;
104 ThreadLocal& operator=(const ThreadLocal&) = delete;
107 auto ptr = constructor_();
112 mutable ThreadLocalPtr<T,Tag> tlp_;
113 std::function<T*()> constructor_;
117 * The idea here is that __thread is faster than pthread_getspecific, so we
118 * keep a __thread array of pointers to objects (ThreadEntry::elements) where
119 * each array has an index for each unique instance of the ThreadLocalPtr
120 * object. Each ThreadLocalPtr object has a unique id that is an index into
121 * these arrays so we can fetch the correct object from thread local storage
124 * In order to prevent unbounded growth of the id space and thus huge
125 * ThreadEntry::elements, arrays, for example due to continuous creation and
126 * destruction of ThreadLocalPtr objects, we keep a set of all active
127 * instances. When an instance is destroyed we remove it from the active
128 * set and insert the id into freeIds_ for reuse. These operations require a
129 * global mutex, but only happen at construction and destruction time.
131 * We use a single global pthread_key_t per Tag to manage object destruction and
132 * memory cleanup upon thread exit because there is a finite number of
133 * pthread_key_t's available per machine.
135 * NOTE: Apple platforms don't support the same semantics for __thread that
136 * Linux does (and it's only supported at all on i386). For these, use
137 * pthread_setspecific()/pthread_getspecific() for the per-thread
138 * storage. Windows (MSVC and GCC) does support the same semantics
139 * with __declspec(thread)
142 template<class T, class Tag=void>
143 class ThreadLocalPtr {
145 typedef threadlocal_detail::StaticMeta<Tag> StaticMeta;
147 constexpr ThreadLocalPtr() : id_() {}
149 ThreadLocalPtr(ThreadLocalPtr&& other) noexcept :
150 id_(std::move(other.id_)) {
153 ThreadLocalPtr& operator=(ThreadLocalPtr&& other) {
154 assert(this != &other);
156 id_ = std::move(other.id_);
165 threadlocal_detail::ElementWrapper& w = StaticMeta::instance().get(&id_);
166 return static_cast<T*>(w.ptr);
169 T* operator->() const {
173 T& operator*() const {
178 threadlocal_detail::ElementWrapper& w = StaticMeta::instance().get(&id_);
180 return static_cast<T*>(w.release());
183 void reset(T* newPtr = nullptr) {
184 auto guard = makeGuard([&] { delete newPtr; });
185 threadlocal_detail::ElementWrapper& w = StaticMeta::instance().get(&id_);
187 w.dispose(TLPDestructionMode::THIS_THREAD);
192 explicit operator bool() const {
193 return get() != nullptr;
197 * reset() that transfers ownership from a smart pointer
202 typename = typename std::enable_if<
203 std::is_convertible<SourceT*, T*>::value>::type>
204 void reset(std::unique_ptr<SourceT, Deleter> source) {
205 auto deleter = [delegate = source.get_deleter()](
206 T * ptr, TLPDestructionMode) {
209 reset(source.release(), deleter);
213 * reset() that transfers ownership from a smart pointer with the default
218 typename = typename std::enable_if<
219 std::is_convertible<SourceT*, T*>::value>::type>
220 void reset(std::unique_ptr<SourceT> source) {
221 reset(source.release());
225 * reset() with a custom deleter:
226 * deleter(T* ptr, TLPDestructionMode mode)
227 * "mode" is ALL_THREADS if we're destructing this ThreadLocalPtr (and thus
228 * deleting pointers for all threads), and THIS_THREAD if we're only deleting
229 * the member for one thread (because of thread exit or reset()).
230 * Invoking the deleter must not throw.
232 template <class Deleter>
233 void reset(T* newPtr, const Deleter& deleter) {
234 auto guard = makeGuard([&] {
236 deleter(newPtr, TLPDestructionMode::THIS_THREAD);
239 threadlocal_detail::ElementWrapper& w = StaticMeta::instance().get(&id_);
240 w.dispose(TLPDestructionMode::THIS_THREAD);
242 w.set(newPtr, deleter);
245 // Holds a global lock for iteration through all thread local child objects.
246 // Can be used as an iterable container.
247 // Use accessAllThreads() to obtain one.
249 friend class ThreadLocalPtr<T,Tag>;
251 threadlocal_detail::StaticMetaBase& meta_;
257 friend class Iterator;
259 // The iterators obtained from Accessor are bidirectional iterators.
260 class Iterator : public boost::iterator_facade<
263 boost::bidirectional_traversal_tag> { // traversal
264 friend class Accessor;
265 friend class boost::iterator_core_access;
266 const Accessor* accessor_;
267 threadlocal_detail::ThreadEntry* e_;
279 T& dereference() const {
280 return *static_cast<T*>(e_->elements[accessor_->id_].ptr);
283 bool equal(const Iterator& other) const {
284 return (accessor_->id_ == other.accessor_->id_ &&
288 explicit Iterator(const Accessor* accessor)
289 : accessor_(accessor),
290 e_(&accessor_->meta_.head_) {
294 return (e_->elements &&
295 accessor_->id_ < e_->elementsCapacity &&
296 e_->elements[accessor_->id_].ptr);
299 void incrementToValid() {
300 for (; e_ != &accessor_->meta_.head_ && !valid(); e_ = e_->next) { }
303 void decrementToValid() {
304 for (; e_ != &accessor_->meta_.head_ && !valid(); e_ = e_->prev) { }
312 Iterator begin() const {
313 return ++Iterator(this);
316 Iterator end() const {
317 return Iterator(this);
320 Accessor(const Accessor&) = delete;
321 Accessor& operator=(const Accessor&) = delete;
323 Accessor(Accessor&& other) noexcept
324 : meta_(other.meta_),
328 other.lock_ = nullptr;
331 Accessor& operator=(Accessor&& other) noexcept {
332 // Each Tag has its own unique meta, and accessors with different Tags
333 // have different types. So either *this is empty, or this and other
334 // have the same tag. But if they have the same tag, they have the same
335 // meta (and lock), so they'd both hold the lock at the same time,
336 // which is impossible, which leaves only one possible scenario --
337 // *this is empty. Assert it.
338 assert(&meta_ == &other.meta_);
339 assert(lock_ == nullptr);
341 swap(lock_, other.lock_);
342 swap(id_, other.id_);
346 : meta_(threadlocal_detail::StaticMeta<Tag>::instance()),
352 explicit Accessor(uint32_t id)
353 : meta_(threadlocal_detail::StaticMeta<Tag>::instance()),
354 lock_(&meta_.lock_) {
368 // accessor allows a client to iterate through all thread local child
369 // elements of this ThreadLocal instance. Holds a global lock for each <Tag>
370 Accessor accessAllThreads() const {
371 static_assert(!std::is_same<Tag, void>::value,
372 "Must use a unique Tag to use the accessAllThreads feature");
373 return Accessor(id_.getOrAllocate(StaticMeta::instance()));
378 StaticMeta::instance().destroy(&id_);
382 ThreadLocalPtr(const ThreadLocalPtr&) = delete;
383 ThreadLocalPtr& operator=(const ThreadLocalPtr&) = delete;
385 mutable typename StaticMeta::EntryID id_;