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_DETAIL_THREADLOCALDETAIL_H_
18 #define FOLLY_DETAIL_THREADLOCALDETAIL_H_
27 #include <glog/logging.h>
29 #include <folly/Foreach.h>
30 #include <folly/Exception.h>
31 #include <folly/Malloc.h>
33 // In general, emutls cleanup is not guaranteed to play nice with the way
34 // StaticMeta mixes direct pthread calls and the use of __thread. This has
35 // caused problems on multiple platforms so don't use __thread there.
37 // XXX: Ideally we would instead determine if emutls is in use at runtime as it
38 // is possible to configure glibc on Linux to use emutls regardless.
39 #if !__APPLE__ && !__ANDROID__
40 #define FOLLY_TLD_USE_FOLLY_TLS 1
42 #undef FOLLY_TLD_USE_FOLLY_TLS
46 namespace threadlocal_detail {
49 * Base class for deleters.
53 virtual ~DeleterBase() { }
54 virtual void dispose(void* ptr, TLPDestructionMode mode) const = 0;
58 * Simple deleter class that calls delete on the passed-in pointer.
61 class SimpleDeleter : public DeleterBase {
63 virtual void dispose(void* ptr, TLPDestructionMode /*mode*/) const {
64 delete static_cast<Ptr>(ptr);
69 * Custom deleter that calls a given callable.
71 template <class Ptr, class Deleter>
72 class CustomDeleter : public DeleterBase {
74 explicit CustomDeleter(Deleter d) : deleter_(d) { }
75 virtual void dispose(void* ptr, TLPDestructionMode mode) const {
76 deleter_(static_cast<Ptr>(ptr), mode);
84 * POD wrapper around an element (a void*) and an associated deleter.
85 * This must be POD, as we memset() it to 0 and memcpy() it around.
87 struct ElementWrapper {
88 bool dispose(TLPDestructionMode mode) {
93 DCHECK(deleter != nullptr);
94 deleter->dispose(ptr, mode);
102 if (ptr != nullptr) {
111 DCHECK(ptr == nullptr);
112 DCHECK(deleter == nullptr);
115 // We leak a single object here but that is ok. If we used an
116 // object directly, there is a chance that the destructor will be
117 // called on that static object before any of the ElementWrappers
118 // are disposed and that isn't so nice.
119 static auto d = new SimpleDeleter<Ptr>();
126 template <class Ptr, class Deleter>
127 void set(Ptr p, Deleter d) {
128 DCHECK(ptr == nullptr);
129 DCHECK(deleter == nullptr);
132 deleter = new CustomDeleter<Ptr,Deleter>(d);
147 DeleterBase* deleter;
152 * Per-thread entry. Each thread using a StaticMeta object has one.
153 * This is written from the owning thread only (under the lock), read
154 * from the owning thread (no lock necessary), and read from other threads
158 ElementWrapper* elements;
159 size_t elementsCapacity;
164 constexpr uint32_t kEntryIDInvalid = std::numeric_limits<uint32_t>::max();
167 * We want to disable onThreadExit call at the end of shutdown, we don't care
168 * about leaking memory at that point.
170 * Otherwise if ThreadLocal is used in a shared library, onThreadExit may be
171 * called after dlclose().
173 class PthreadKeyUnregister {
175 ~PthreadKeyUnregister() {
176 std::lock_guard<std::mutex> lg(mutex_);
178 for (const auto& key: keys_) {
179 pthread_key_delete(key);
183 static void registerKey(pthread_key_t key) {
184 instance_.registerKeyImpl(key);
188 PthreadKeyUnregister() {}
190 void registerKeyImpl(pthread_key_t key) {
191 std::lock_guard<std::mutex> lg(mutex_);
193 keys_.push_back(key);
197 std::vector<pthread_key_t> keys_;
199 static PthreadKeyUnregister instance_;
203 // Held in a singleton to track our global instances.
204 // We have one of these per "Tag", by default one for the whole system
207 // Creating and destroying ThreadLocalPtr objects, as well as thread exit
208 // for threads that use ThreadLocalPtr objects collide on a lock inside
209 // StaticMeta; you can specify multiple Tag types to break that lock.
212 // Represents an ID of a thread local object. Initially set to the maximum
213 // uint. This representation allows us to avoid a branch in accessing TLS data
214 // (because if you test capacity > id if id = maxint then the test will always
215 // fail). It allows us to keep a constexpr constructor and avoid SIOF.
218 std::atomic<uint32_t> value;
220 constexpr EntryID() : value(kEntryIDInvalid) {
223 EntryID(EntryID&& other) noexcept : value(other.value.load()) {
224 other.value = kEntryIDInvalid;
227 EntryID& operator=(EntryID&& other) {
228 assert(this != &other);
229 value = other.value.load();
230 other.value = kEntryIDInvalid;
234 EntryID(const EntryID& other) = delete;
235 EntryID& operator=(const EntryID& other) = delete;
237 uint32_t getOrInvalid() {
238 // It's OK for this to be relaxed, even though we're effectively doing
239 // double checked locking in using this value. We only care about the
240 // uniqueness of IDs, getOrAllocate does not modify any other memory
241 // this thread will use.
242 return value.load(std::memory_order_relaxed);
245 uint32_t getOrAllocate() {
246 uint32_t id = getOrInvalid();
247 if (id != kEntryIDInvalid) {
250 // The lock inside allocate ensures that a single value is allocated
251 return instance().allocate(this);
255 static StaticMeta<Tag>& instance() {
256 // Leak it on exit, there's only one per process and we don't have to
257 // worry about synchronization with exiting threads.
258 static bool constructed = (inst_ = new StaticMeta<Tag>());
259 (void)constructed; // suppress unused warning
264 std::vector<uint32_t> freeIds_;
266 pthread_key_t pthreadKey_;
269 void push_back(ThreadEntry* t) {
271 t->prev = head_.prev;
272 head_.prev->next = t;
276 void erase(ThreadEntry* t) {
277 t->next->prev = t->prev;
278 t->prev->next = t->next;
279 t->next = t->prev = t;
282 #ifdef FOLLY_TLD_USE_FOLLY_TLS
283 static FOLLY_TLS ThreadEntry threadEntry_;
285 static StaticMeta<Tag>* inst_;
287 StaticMeta() : nextId_(1) {
288 head_.next = head_.prev = &head_;
289 int ret = pthread_key_create(&pthreadKey_, &onThreadExit);
290 checkPosixError(ret, "pthread_key_create failed");
291 PthreadKeyUnregister::registerKey(pthreadKey_);
293 #if FOLLY_HAVE_PTHREAD_ATFORK
294 ret = pthread_atfork(/*prepare*/ &StaticMeta::preFork,
295 /*parent*/ &StaticMeta::onForkParent,
296 /*child*/ &StaticMeta::onForkChild);
297 checkPosixError(ret, "pthread_atfork failed");
298 #elif !__ANDROID__ && !defined(_MSC_VER)
299 // pthread_atfork is not part of the Android NDK at least as of n9d. If
300 // something is trying to call native fork() directly at all with Android's
301 // process management model, this is probably the least of the problems.
303 // But otherwise, this is a problem.
304 #warning pthread_atfork unavailable
308 LOG(FATAL) << "StaticMeta lives forever!";
311 static ThreadEntry* getThreadEntry() {
312 #ifdef FOLLY_TLD_USE_FOLLY_TLS
313 return &threadEntry_;
315 auto key = instance().pthreadKey_;
316 ThreadEntry* threadEntry =
317 static_cast<ThreadEntry*>(pthread_getspecific(key));
319 threadEntry = new ThreadEntry();
320 int ret = pthread_setspecific(key, threadEntry);
321 checkPosixError(ret, "pthread_setspecific failed");
327 static void preFork(void) {
328 instance().lock_.lock(); // Make sure it's created
331 static void onForkParent(void) {
332 inst_->lock_.unlock();
335 static void onForkChild(void) {
336 // only the current thread survives
337 inst_->head_.next = inst_->head_.prev = &inst_->head_;
338 ThreadEntry* threadEntry = getThreadEntry();
339 // If this thread was in the list before the fork, add it back.
340 if (threadEntry->elementsCapacity != 0) {
341 inst_->push_back(threadEntry);
343 inst_->lock_.unlock();
346 static void onThreadExit(void* ptr) {
347 auto& meta = instance();
348 #ifdef FOLLY_TLD_USE_FOLLY_TLS
349 ThreadEntry* threadEntry = getThreadEntry();
351 DCHECK_EQ(ptr, &meta);
352 DCHECK_GT(threadEntry->elementsCapacity, 0);
354 // pthread sets the thread-specific value corresponding
355 // to meta.pthreadKey_ to NULL before calling onThreadExit.
356 // We need to set it back to ptr to enable the correct behaviour
357 // of the subsequent calls of getThreadEntry
358 // (which may happen in user-provided custom deleters)
359 pthread_setspecific(meta.pthreadKey_, ptr);
360 ThreadEntry* threadEntry = static_cast<ThreadEntry*>(ptr);
363 std::lock_guard<std::mutex> g(meta.lock_);
364 meta.erase(threadEntry);
365 // No need to hold the lock any longer; the ThreadEntry is private to this
366 // thread now that it's been removed from meta.
368 // NOTE: User-provided deleter / object dtor itself may be using ThreadLocal
369 // with the same Tag, so dispose() calls below may (re)create some of the
370 // elements or even increase elementsCapacity, thus multiple cleanup rounds
372 for (bool shouldRun = true; shouldRun; ) {
374 FOR_EACH_RANGE(i, 0, threadEntry->elementsCapacity) {
375 if (threadEntry->elements[i].dispose(TLPDestructionMode::THIS_THREAD)) {
380 free(threadEntry->elements);
381 threadEntry->elements = nullptr;
382 pthread_setspecific(meta.pthreadKey_, nullptr);
384 #ifndef FOLLY_TLD_USE_FOLLY_TLS
385 // Allocated in getThreadEntry() when not using folly TLS; free it
390 static uint32_t allocate(EntryID* ent) {
392 auto & meta = instance();
393 std::lock_guard<std::mutex> g(meta.lock_);
395 id = ent->value.load();
396 if (id != kEntryIDInvalid) {
400 if (!meta.freeIds_.empty()) {
401 id = meta.freeIds_.back();
402 meta.freeIds_.pop_back();
407 uint32_t old_id = ent->value.exchange(id);
408 DCHECK_EQ(old_id, kEntryIDInvalid);
412 static void destroy(EntryID* ent) {
414 auto & meta = instance();
415 // Elements in other threads that use this id.
416 std::vector<ElementWrapper> elements;
418 std::lock_guard<std::mutex> g(meta.lock_);
419 uint32_t id = ent->value.exchange(kEntryIDInvalid);
420 if (id == kEntryIDInvalid) {
424 for (ThreadEntry* e = meta.head_.next; e != &meta.head_; e = e->next) {
425 if (id < e->elementsCapacity && e->elements[id].ptr) {
426 elements.push_back(e->elements[id]);
429 * Writing another thread's ThreadEntry from here is fine;
430 * the only other potential reader is the owning thread --
431 * from onThreadExit (which grabs the lock, so is properly
432 * synchronized with us) or from get(), which also grabs
433 * the lock if it needs to resize the elements vector.
435 * We can't conflict with reads for a get(id), because
436 * it's illegal to call get on a thread local that's
439 e->elements[id].ptr = nullptr;
440 e->elements[id].deleter = nullptr;
441 e->elements[id].ownsDeleter = false;
444 meta.freeIds_.push_back(id);
446 // Delete elements outside the lock
447 FOR_EACH(it, elements) {
448 it->dispose(TLPDestructionMode::ALL_THREADS);
450 } catch (...) { // Just in case we get a lock error or something anyway...
451 LOG(WARNING) << "Destructor discarding an exception that was thrown.";
456 * Reserve enough space in the ThreadEntry::elements for the item
459 static void reserve(EntryID* id) {
460 auto& meta = instance();
461 ThreadEntry* threadEntry = getThreadEntry();
462 size_t prevCapacity = threadEntry->elementsCapacity;
464 uint32_t idval = id->getOrAllocate();
465 if (prevCapacity > idval) {
468 // Growth factor < 2, see folly/docs/FBVector.md; + 5 to prevent
470 size_t newCapacity = static_cast<size_t>((idval + 5) * 1.7);
471 assert(newCapacity > prevCapacity);
472 ElementWrapper* reallocated = nullptr;
474 // Need to grow. Note that we can't call realloc, as elements is
475 // still linked in meta, so another thread might access invalid memory
476 // after realloc succeeds. We'll copy by hand and update our ThreadEntry
478 if (usingJEMalloc()) {
479 bool success = false;
480 size_t newByteSize = nallocx(newCapacity * sizeof(ElementWrapper), 0);
482 // Try to grow in place.
484 // Note that xallocx(MALLOCX_ZERO) will only zero newly allocated memory,
485 // even if a previous allocation allocated more than we requested.
486 // This is fine; we always use MALLOCX_ZERO with jemalloc and we
487 // always expand our allocation to the real size.
488 if (prevCapacity * sizeof(ElementWrapper) >=
489 jemallocMinInPlaceExpandable) {
490 success = (xallocx(threadEntry->elements, newByteSize, 0, MALLOCX_ZERO)
494 // In-place growth failed.
496 success = ((reallocated = static_cast<ElementWrapper*>(
497 mallocx(newByteSize, MALLOCX_ZERO))) != nullptr);
501 // Expand to real size
502 assert(newByteSize / sizeof(ElementWrapper) >= newCapacity);
503 newCapacity = newByteSize / sizeof(ElementWrapper);
505 throw std::bad_alloc();
507 } else { // no jemalloc
508 // calloc() is simpler than malloc() followed by memset(), and
509 // potentially faster when dealing with a lot of memory, as it can get
510 // already-zeroed pages from the kernel.
511 reallocated = static_cast<ElementWrapper*>(
512 calloc(newCapacity, sizeof(ElementWrapper)));
514 throw std::bad_alloc();
518 // Success, update the entry
520 std::lock_guard<std::mutex> g(meta.lock_);
522 if (prevCapacity == 0) {
523 meta.push_back(threadEntry);
528 * Note: we need to hold the meta lock when copying data out of
529 * the old vector, because some other thread might be
530 * destructing a ThreadLocal and writing to the elements vector
533 memcpy(reallocated, threadEntry->elements,
534 sizeof(ElementWrapper) * prevCapacity);
536 swap(reallocated, threadEntry->elements);
538 threadEntry->elementsCapacity = newCapacity;
543 #ifdef FOLLY_TLD_USE_FOLLY_TLS
544 if (prevCapacity == 0) {
545 pthread_setspecific(meta.pthreadKey_, &meta);
550 static ElementWrapper& get(EntryID* ent) {
551 ThreadEntry* threadEntry = getThreadEntry();
552 uint32_t id = ent->getOrInvalid();
553 // if id is invalid, it is equal to uint32_t's max value.
554 // x <= max value is always true
555 if (UNLIKELY(threadEntry->elementsCapacity <= id)) {
557 id = ent->getOrInvalid();
558 assert(threadEntry->elementsCapacity > id);
560 return threadEntry->elements[id];
564 #ifdef FOLLY_TLD_USE_FOLLY_TLS
566 FOLLY_TLS ThreadEntry StaticMeta<Tag>::threadEntry_ = {nullptr, 0,
569 template <class Tag> StaticMeta<Tag>* StaticMeta<Tag>::inst_ = nullptr;
571 } // namespace threadlocal_detail
574 #endif /* FOLLY_DETAIL_THREADLOCALDETAIL_H_ */