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.
16 #include <folly/ThreadLocal.h>
21 namespace folly { namespace threadlocal_detail {
23 StaticMetaBase::StaticMetaBase(ThreadEntry* (*threadEntry)())
24 : nextId_(1), threadEntry_(threadEntry) {
25 head_.next = head_.prev = &head_;
26 int ret = pthread_key_create(&pthreadKey_, &onThreadExit);
27 checkPosixError(ret, "pthread_key_create failed");
28 PthreadKeyUnregister::registerKey(pthreadKey_);
31 void StaticMetaBase::onThreadExit(void* ptr) {
32 #ifdef FOLLY_TLD_USE_FOLLY_TLS
33 auto threadEntry = static_cast<ThreadEntry*>(ptr);
35 std::unique_ptr<ThreadEntry> threadEntry(static_cast<ThreadEntry*>(ptr));
37 DCHECK_GT(threadEntry->elementsCapacity, 0);
38 auto& meta = *threadEntry->meta;
40 // Make sure this ThreadEntry is available if ThreadLocal A is accessed in
41 // ThreadLocal B destructor.
42 pthread_setspecific(meta.pthreadKey_, &(*threadEntry));
44 pthread_setspecific(meta.pthreadKey_, nullptr);
48 std::lock_guard<std::mutex> g(meta.lock_);
49 meta.erase(&(*threadEntry));
50 // No need to hold the lock any longer; the ThreadEntry is private to this
51 // thread now that it's been removed from meta.
53 // NOTE: User-provided deleter / object dtor itself may be using ThreadLocal
54 // with the same Tag, so dispose() calls below may (re)create some of the
55 // elements or even increase elementsCapacity, thus multiple cleanup rounds
57 for (bool shouldRun = true; shouldRun;) {
59 FOR_EACH_RANGE (i, 0, threadEntry->elementsCapacity) {
60 if (threadEntry->elements[i].dispose(TLPDestructionMode::THIS_THREAD)) {
65 free(threadEntry->elements);
66 threadEntry->elements = nullptr;
67 threadEntry->meta = nullptr;
70 uint32_t StaticMetaBase::allocate(EntryID* ent) {
73 std::lock_guard<std::mutex> g(meta.lock_);
75 id = ent->value.load();
76 if (id != kEntryIDInvalid) {
80 if (!meta.freeIds_.empty()) {
81 id = meta.freeIds_.back();
82 meta.freeIds_.pop_back();
87 uint32_t old_id = ent->value.exchange(id);
88 DCHECK_EQ(old_id, kEntryIDInvalid);
92 void StaticMetaBase::destroy(EntryID* ent) {
95 // Elements in other threads that use this id.
96 std::vector<ElementWrapper> elements;
98 std::lock_guard<std::mutex> g(meta.lock_);
99 uint32_t id = ent->value.exchange(kEntryIDInvalid);
100 if (id == kEntryIDInvalid) {
104 for (ThreadEntry* e = meta.head_.next; e != &meta.head_; e = e->next) {
105 if (id < e->elementsCapacity && e->elements[id].ptr) {
106 elements.push_back(e->elements[id]);
109 * Writing another thread's ThreadEntry from here is fine;
110 * the only other potential reader is the owning thread --
111 * from onThreadExit (which grabs the lock, so is properly
112 * synchronized with us) or from get(), which also grabs
113 * the lock if it needs to resize the elements vector.
115 * We can't conflict with reads for a get(id), because
116 * it's illegal to call get on a thread local that's
119 e->elements[id].ptr = nullptr;
120 e->elements[id].deleter1 = nullptr;
121 e->elements[id].ownsDeleter = false;
124 meta.freeIds_.push_back(id);
126 // Delete elements outside the lock
127 for (ElementWrapper& elem : elements) {
128 elem.dispose(TLPDestructionMode::ALL_THREADS);
130 } catch (...) { // Just in case we get a lock error or something anyway...
131 LOG(WARNING) << "Destructor discarding an exception that was thrown.";
136 * Reserve enough space in the ThreadEntry::elements for the item
139 void StaticMetaBase::reserve(EntryID* id) {
141 ThreadEntry* threadEntry = (*threadEntry_)();
142 size_t prevCapacity = threadEntry->elementsCapacity;
144 uint32_t idval = id->getOrAllocate(meta);
145 if (prevCapacity > idval) {
148 // Growth factor < 2, see folly/docs/FBVector.md; + 5 to prevent
150 size_t newCapacity = static_cast<size_t>((idval + 5) * 1.7);
151 assert(newCapacity > prevCapacity);
152 ElementWrapper* reallocated = nullptr;
154 // Need to grow. Note that we can't call realloc, as elements is
155 // still linked in meta, so another thread might access invalid memory
156 // after realloc succeeds. We'll copy by hand and update our ThreadEntry
158 if (usingJEMalloc()) {
159 bool success = false;
160 size_t newByteSize = nallocx(newCapacity * sizeof(ElementWrapper), 0);
162 // Try to grow in place.
164 // Note that xallocx(MALLOCX_ZERO) will only zero newly allocated memory,
165 // even if a previous allocation allocated more than we requested.
166 // This is fine; we always use MALLOCX_ZERO with jemalloc and we
167 // always expand our allocation to the real size.
168 if (prevCapacity * sizeof(ElementWrapper) >= jemallocMinInPlaceExpandable) {
170 (xallocx(threadEntry->elements, newByteSize, 0, MALLOCX_ZERO) ==
174 // In-place growth failed.
177 ((reallocated = static_cast<ElementWrapper*>(
178 mallocx(newByteSize, MALLOCX_ZERO))) != nullptr);
182 // Expand to real size
183 assert(newByteSize / sizeof(ElementWrapper) >= newCapacity);
184 newCapacity = newByteSize / sizeof(ElementWrapper);
186 throw std::bad_alloc();
188 } else { // no jemalloc
189 // calloc() is simpler than malloc() followed by memset(), and
190 // potentially faster when dealing with a lot of memory, as it can get
191 // already-zeroed pages from the kernel.
192 reallocated = static_cast<ElementWrapper*>(
193 calloc(newCapacity, sizeof(ElementWrapper)));
195 throw std::bad_alloc();
199 // Success, update the entry
201 std::lock_guard<std::mutex> g(meta.lock_);
203 if (prevCapacity == 0) {
204 meta.push_back(threadEntry);
209 * Note: we need to hold the meta lock when copying data out of
210 * the old vector, because some other thread might be
211 * destructing a ThreadLocal and writing to the elements vector
214 if (prevCapacity != 0) {
217 threadEntry->elements,
218 sizeof(*reallocated) * prevCapacity);
220 std::swap(reallocated, threadEntry->elements);
222 threadEntry->elementsCapacity = newCapacity;
231 folly::Function<void()> prepare;
232 folly::Function<void()> parent;
233 folly::Function<void()> child;
238 static AtForkList& instance() {
239 static auto instance = new AtForkList();
243 static void prepare() noexcept {
244 instance().tasksLock.lock();
245 auto& tasks = instance().tasks;
246 for (auto task = tasks.rbegin(); task != tasks.rend(); ++task) {
251 static void parent() noexcept {
252 auto& tasks = instance().tasks;
253 for (auto& task : tasks) {
256 instance().tasksLock.unlock();
259 static void child() noexcept {
260 auto& tasks = instance().tasks;
261 for (auto& task : tasks) {
264 instance().tasksLock.unlock();
267 std::mutex tasksLock;
268 std::list<AtForkTask> tasks;
272 #if FOLLY_HAVE_PTHREAD_ATFORK
273 int ret = pthread_atfork(
274 &AtForkList::prepare, &AtForkList::parent, &AtForkList::child);
275 checkPosixError(ret, "pthread_atfork failed");
276 #elif !__ANDROID__ && !defined(_MSC_VER)
277 // pthread_atfork is not part of the Android NDK at least as of n9d. If
278 // something is trying to call native fork() directly at all with Android's
279 // process management model, this is probably the least of the problems.
281 // But otherwise, this is a problem.
282 #warning pthread_atfork unavailable
288 void StaticMetaBase::initAtFork() {
289 AtForkList::instance();
292 void StaticMetaBase::registerAtFork(
293 folly::Function<void()> prepare,
294 folly::Function<void()> parent,
295 folly::Function<void()> child) {
296 std::lock_guard<std::mutex> lg(AtForkList::instance().tasksLock);
297 AtForkList::instance().tasks.push_back(
298 {std::move(prepare), std::move(parent), std::move(child)});
301 FOLLY_STATIC_CTOR_PRIORITY_MAX
302 PthreadKeyUnregister PthreadKeyUnregister::instance_;