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();
166 struct PthreadKeyUnregisterTester;
169 * We want to disable onThreadExit call at the end of shutdown, we don't care
170 * about leaking memory at that point.
172 * Otherwise if ThreadLocal is used in a shared library, onThreadExit may be
173 * called after dlclose().
175 * This class has one single static instance; however since it's so widely used,
176 * directly or indirectly, by so many classes, we need to take care to avoid
177 * problems stemming from the Static Initialization/Destruction Order Fiascos.
178 * Therefore this class needs to be constexpr-constructible, so as to avoid
179 * the need for this to participate in init/destruction order.
181 class PthreadKeyUnregister {
183 static constexpr size_t kMaxKeys = 1UL << 16;
185 ~PthreadKeyUnregister() {
186 std::lock_guard<std::mutex> lg(mutex_);
188 pthread_key_delete(keys_[--size_]);
192 static void registerKey(pthread_key_t key) {
193 instance_.registerKeyImpl(key);
198 * Only one global instance should exist, hence this is private.
199 * See also the important note at the top of this class about `constexpr`
202 constexpr PthreadKeyUnregister() : mutex_(), size_(0), keys_() { }
203 friend class folly::threadlocal_detail::PthreadKeyUnregisterTester;
205 void registerKeyImpl(pthread_key_t key) {
206 std::lock_guard<std::mutex> lg(mutex_);
207 CHECK_LT(size_, kMaxKeys);
208 keys_[size_++] = key;
213 pthread_key_t keys_[kMaxKeys];
215 static PthreadKeyUnregister instance_;
218 // Held in a singleton to track our global instances.
219 // We have one of these per "Tag", by default one for the whole system
222 // Creating and destroying ThreadLocalPtr objects, as well as thread exit
223 // for threads that use ThreadLocalPtr objects collide on a lock inside
224 // StaticMeta; you can specify multiple Tag types to break that lock.
227 // Represents an ID of a thread local object. Initially set to the maximum
228 // uint. This representation allows us to avoid a branch in accessing TLS data
229 // (because if you test capacity > id if id = maxint then the test will always
230 // fail). It allows us to keep a constexpr constructor and avoid SIOF.
233 std::atomic<uint32_t> value;
235 constexpr EntryID() : value(kEntryIDInvalid) {
238 EntryID(EntryID&& other) noexcept : value(other.value.load()) {
239 other.value = kEntryIDInvalid;
242 EntryID& operator=(EntryID&& other) {
243 assert(this != &other);
244 value = other.value.load();
245 other.value = kEntryIDInvalid;
249 EntryID(const EntryID& other) = delete;
250 EntryID& operator=(const EntryID& other) = delete;
252 uint32_t getOrInvalid() {
253 // It's OK for this to be relaxed, even though we're effectively doing
254 // double checked locking in using this value. We only care about the
255 // uniqueness of IDs, getOrAllocate does not modify any other memory
256 // this thread will use.
257 return value.load(std::memory_order_relaxed);
260 uint32_t getOrAllocate() {
261 uint32_t id = getOrInvalid();
262 if (id != kEntryIDInvalid) {
265 // The lock inside allocate ensures that a single value is allocated
266 return instance().allocate(this);
270 static StaticMeta<Tag>& instance() {
271 // Leak it on exit, there's only one per process and we don't have to
272 // worry about synchronization with exiting threads.
273 static bool constructed = (inst_ = new StaticMeta<Tag>());
274 (void)constructed; // suppress unused warning
279 std::vector<uint32_t> freeIds_;
281 pthread_key_t pthreadKey_;
284 void push_back(ThreadEntry* t) {
286 t->prev = head_.prev;
287 head_.prev->next = t;
291 void erase(ThreadEntry* t) {
292 t->next->prev = t->prev;
293 t->prev->next = t->next;
294 t->next = t->prev = t;
297 #ifdef FOLLY_TLD_USE_FOLLY_TLS
298 static FOLLY_TLS ThreadEntry threadEntry_;
300 static StaticMeta<Tag>* inst_;
302 StaticMeta() : nextId_(1) {
303 head_.next = head_.prev = &head_;
304 int ret = pthread_key_create(&pthreadKey_, &onThreadExit);
305 checkPosixError(ret, "pthread_key_create failed");
306 PthreadKeyUnregister::registerKey(pthreadKey_);
308 #if FOLLY_HAVE_PTHREAD_ATFORK
309 ret = pthread_atfork(/*prepare*/ &StaticMeta::preFork,
310 /*parent*/ &StaticMeta::onForkParent,
311 /*child*/ &StaticMeta::onForkChild);
312 checkPosixError(ret, "pthread_atfork failed");
313 #elif !__ANDROID__ && !defined(_MSC_VER)
314 // pthread_atfork is not part of the Android NDK at least as of n9d. If
315 // something is trying to call native fork() directly at all with Android's
316 // process management model, this is probably the least of the problems.
318 // But otherwise, this is a problem.
319 #warning pthread_atfork unavailable
323 LOG(FATAL) << "StaticMeta lives forever!";
326 static ThreadEntry* getThreadEntry() {
327 #ifdef FOLLY_TLD_USE_FOLLY_TLS
328 return &threadEntry_;
330 auto key = instance().pthreadKey_;
331 ThreadEntry* threadEntry =
332 static_cast<ThreadEntry*>(pthread_getspecific(key));
334 threadEntry = new ThreadEntry();
335 int ret = pthread_setspecific(key, threadEntry);
336 checkPosixError(ret, "pthread_setspecific failed");
342 static void preFork(void) {
343 instance().lock_.lock(); // Make sure it's created
346 static void onForkParent(void) {
347 inst_->lock_.unlock();
350 static void onForkChild(void) {
351 // only the current thread survives
352 inst_->head_.next = inst_->head_.prev = &inst_->head_;
353 ThreadEntry* threadEntry = getThreadEntry();
354 // If this thread was in the list before the fork, add it back.
355 if (threadEntry->elementsCapacity != 0) {
356 inst_->push_back(threadEntry);
358 inst_->lock_.unlock();
361 static void onThreadExit(void* ptr) {
362 auto& meta = instance();
363 #ifdef FOLLY_TLD_USE_FOLLY_TLS
364 ThreadEntry* threadEntry = getThreadEntry();
366 DCHECK_EQ(ptr, &meta);
367 DCHECK_GT(threadEntry->elementsCapacity, 0);
369 // pthread sets the thread-specific value corresponding
370 // to meta.pthreadKey_ to NULL before calling onThreadExit.
371 // We need to set it back to ptr to enable the correct behaviour
372 // of the subsequent calls of getThreadEntry
373 // (which may happen in user-provided custom deleters)
374 pthread_setspecific(meta.pthreadKey_, ptr);
375 ThreadEntry* threadEntry = static_cast<ThreadEntry*>(ptr);
378 std::lock_guard<std::mutex> g(meta.lock_);
379 meta.erase(threadEntry);
380 // No need to hold the lock any longer; the ThreadEntry is private to this
381 // thread now that it's been removed from meta.
383 // NOTE: User-provided deleter / object dtor itself may be using ThreadLocal
384 // with the same Tag, so dispose() calls below may (re)create some of the
385 // elements or even increase elementsCapacity, thus multiple cleanup rounds
387 for (bool shouldRun = true; shouldRun; ) {
389 FOR_EACH_RANGE(i, 0, threadEntry->elementsCapacity) {
390 if (threadEntry->elements[i].dispose(TLPDestructionMode::THIS_THREAD)) {
395 free(threadEntry->elements);
396 threadEntry->elements = nullptr;
397 pthread_setspecific(meta.pthreadKey_, nullptr);
399 #ifndef FOLLY_TLD_USE_FOLLY_TLS
400 // Allocated in getThreadEntry() when not using folly TLS; free it
405 static uint32_t allocate(EntryID* ent) {
407 auto & meta = instance();
408 std::lock_guard<std::mutex> g(meta.lock_);
410 id = ent->value.load();
411 if (id != kEntryIDInvalid) {
415 if (!meta.freeIds_.empty()) {
416 id = meta.freeIds_.back();
417 meta.freeIds_.pop_back();
422 uint32_t old_id = ent->value.exchange(id);
423 DCHECK_EQ(old_id, kEntryIDInvalid);
427 static void destroy(EntryID* ent) {
429 auto & meta = instance();
430 // Elements in other threads that use this id.
431 std::vector<ElementWrapper> elements;
433 std::lock_guard<std::mutex> g(meta.lock_);
434 uint32_t id = ent->value.exchange(kEntryIDInvalid);
435 if (id == kEntryIDInvalid) {
439 for (ThreadEntry* e = meta.head_.next; e != &meta.head_; e = e->next) {
440 if (id < e->elementsCapacity && e->elements[id].ptr) {
441 elements.push_back(e->elements[id]);
444 * Writing another thread's ThreadEntry from here is fine;
445 * the only other potential reader is the owning thread --
446 * from onThreadExit (which grabs the lock, so is properly
447 * synchronized with us) or from get(), which also grabs
448 * the lock if it needs to resize the elements vector.
450 * We can't conflict with reads for a get(id), because
451 * it's illegal to call get on a thread local that's
454 e->elements[id].ptr = nullptr;
455 e->elements[id].deleter = nullptr;
456 e->elements[id].ownsDeleter = false;
459 meta.freeIds_.push_back(id);
461 // Delete elements outside the lock
462 FOR_EACH(it, elements) {
463 it->dispose(TLPDestructionMode::ALL_THREADS);
465 } catch (...) { // Just in case we get a lock error or something anyway...
466 LOG(WARNING) << "Destructor discarding an exception that was thrown.";
471 * Reserve enough space in the ThreadEntry::elements for the item
474 static void reserve(EntryID* id) {
475 auto& meta = instance();
476 ThreadEntry* threadEntry = getThreadEntry();
477 size_t prevCapacity = threadEntry->elementsCapacity;
479 uint32_t idval = id->getOrAllocate();
480 if (prevCapacity > idval) {
483 // Growth factor < 2, see folly/docs/FBVector.md; + 5 to prevent
485 size_t newCapacity = static_cast<size_t>((idval + 5) * 1.7);
486 assert(newCapacity > prevCapacity);
487 ElementWrapper* reallocated = nullptr;
489 // Need to grow. Note that we can't call realloc, as elements is
490 // still linked in meta, so another thread might access invalid memory
491 // after realloc succeeds. We'll copy by hand and update our ThreadEntry
493 if (usingJEMalloc()) {
494 bool success = false;
495 size_t newByteSize = nallocx(newCapacity * sizeof(ElementWrapper), 0);
497 // Try to grow in place.
499 // Note that xallocx(MALLOCX_ZERO) will only zero newly allocated memory,
500 // even if a previous allocation allocated more than we requested.
501 // This is fine; we always use MALLOCX_ZERO with jemalloc and we
502 // always expand our allocation to the real size.
503 if (prevCapacity * sizeof(ElementWrapper) >=
504 jemallocMinInPlaceExpandable) {
505 success = (xallocx(threadEntry->elements, newByteSize, 0, MALLOCX_ZERO)
509 // In-place growth failed.
511 success = ((reallocated = static_cast<ElementWrapper*>(
512 mallocx(newByteSize, MALLOCX_ZERO))) != nullptr);
516 // Expand to real size
517 assert(newByteSize / sizeof(ElementWrapper) >= newCapacity);
518 newCapacity = newByteSize / sizeof(ElementWrapper);
520 throw std::bad_alloc();
522 } else { // no jemalloc
523 // calloc() is simpler than malloc() followed by memset(), and
524 // potentially faster when dealing with a lot of memory, as it can get
525 // already-zeroed pages from the kernel.
526 reallocated = static_cast<ElementWrapper*>(
527 calloc(newCapacity, sizeof(ElementWrapper)));
529 throw std::bad_alloc();
533 // Success, update the entry
535 std::lock_guard<std::mutex> g(meta.lock_);
537 if (prevCapacity == 0) {
538 meta.push_back(threadEntry);
543 * Note: we need to hold the meta lock when copying data out of
544 * the old vector, because some other thread might be
545 * destructing a ThreadLocal and writing to the elements vector
548 if (prevCapacity != 0) {
549 memcpy(reallocated, threadEntry->elements,
550 sizeof(*reallocated) * prevCapacity);
552 std::swap(reallocated, threadEntry->elements);
554 threadEntry->elementsCapacity = newCapacity;
559 #ifdef FOLLY_TLD_USE_FOLLY_TLS
560 if (prevCapacity == 0) {
561 pthread_setspecific(meta.pthreadKey_, &meta);
566 static ElementWrapper& get(EntryID* ent) {
567 ThreadEntry* threadEntry = getThreadEntry();
568 uint32_t id = ent->getOrInvalid();
569 // if id is invalid, it is equal to uint32_t's max value.
570 // x <= max value is always true
571 if (UNLIKELY(threadEntry->elementsCapacity <= id)) {
573 id = ent->getOrInvalid();
574 assert(threadEntry->elementsCapacity > id);
576 return threadEntry->elements[id];
580 #ifdef FOLLY_TLD_USE_FOLLY_TLS
582 FOLLY_TLS ThreadEntry StaticMeta<Tag>::threadEntry_ = {nullptr, 0,
585 template <class Tag> StaticMeta<Tag>* StaticMeta<Tag>::inst_ = nullptr;
587 } // namespace threadlocal_detail
590 #endif /* FOLLY_DETAIL_THREADLOCALDETAIL_H_ */