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 // Held in a singleton to track our global instances.
167 // We have one of these per "Tag", by default one for the whole system
170 // Creating and destroying ThreadLocalPtr objects, as well as thread exit
171 // for threads that use ThreadLocalPtr objects collide on a lock inside
172 // StaticMeta; you can specify multiple Tag types to break that lock.
175 // Represents an ID of a thread local object. Initially set to the maximum
176 // uint. This representation allows us to avoid a branch in accessing TLS data
177 // (because if you test capacity > id if id = maxint then the test will always
178 // fail). It allows us to keep a constexpr constructor and avoid SIOF.
181 std::atomic<uint32_t> value;
183 constexpr EntryID() : value(kEntryIDInvalid) {
186 EntryID(EntryID&& other) noexcept : value(other.value.load()) {
187 other.value = kEntryIDInvalid;
190 EntryID& operator=(EntryID&& other) {
191 assert(this != &other);
192 value = other.value.load();
193 other.value = kEntryIDInvalid;
197 EntryID(const EntryID& other) = delete;
198 EntryID& operator=(const EntryID& other) = delete;
200 uint32_t getOrInvalid() {
201 // It's OK for this to be relaxed, even though we're effectively doing
202 // double checked locking in using this value. We only care about the
203 // uniqueness of IDs, getOrAllocate does not modify any other memory
204 // this thread will use.
205 return value.load(std::memory_order_relaxed);
208 uint32_t getOrAllocate() {
209 uint32_t id = getOrInvalid();
210 if (id != kEntryIDInvalid) {
213 // The lock inside allocate ensures that a single value is allocated
214 return instance().allocate(this);
218 static StaticMeta<Tag>& instance() {
219 // Leak it on exit, there's only one per process and we don't have to
220 // worry about synchronization with exiting threads.
221 static bool constructed = (inst_ = new StaticMeta<Tag>());
222 (void)constructed; // suppress unused warning
227 std::vector<uint32_t> freeIds_;
229 pthread_key_t pthreadKey_;
232 void push_back(ThreadEntry* t) {
234 t->prev = head_.prev;
235 head_.prev->next = t;
239 void erase(ThreadEntry* t) {
240 t->next->prev = t->prev;
241 t->prev->next = t->next;
242 t->next = t->prev = t;
245 #ifdef FOLLY_TLD_USE_FOLLY_TLS
246 static FOLLY_TLS ThreadEntry threadEntry_;
248 static StaticMeta<Tag>* inst_;
250 StaticMeta() : nextId_(1) {
251 head_.next = head_.prev = &head_;
252 int ret = pthread_key_create(&pthreadKey_, &onThreadExit);
253 checkPosixError(ret, "pthread_key_create failed");
255 #if FOLLY_HAVE_PTHREAD_ATFORK
256 ret = pthread_atfork(/*prepare*/ &StaticMeta::preFork,
257 /*parent*/ &StaticMeta::onForkParent,
258 /*child*/ &StaticMeta::onForkChild);
259 checkPosixError(ret, "pthread_atfork failed");
260 #elif !__ANDROID__ && !defined(_MSC_VER)
261 // pthread_atfork is not part of the Android NDK at least as of n9d. If
262 // something is trying to call native fork() directly at all with Android's
263 // process management model, this is probably the least of the problems.
265 // But otherwise, this is a problem.
266 #warning pthread_atfork unavailable
270 LOG(FATAL) << "StaticMeta lives forever!";
273 static ThreadEntry* getThreadEntry() {
274 #ifdef FOLLY_TLD_USE_FOLLY_TLS
275 return &threadEntry_;
277 auto key = instance().pthreadKey_;
278 ThreadEntry* threadEntry =
279 static_cast<ThreadEntry*>(pthread_getspecific(key));
281 threadEntry = new ThreadEntry();
282 int ret = pthread_setspecific(key, threadEntry);
283 checkPosixError(ret, "pthread_setspecific failed");
289 static void preFork(void) {
290 instance().lock_.lock(); // Make sure it's created
293 static void onForkParent(void) {
294 inst_->lock_.unlock();
297 static void onForkChild(void) {
298 // only the current thread survives
299 inst_->head_.next = inst_->head_.prev = &inst_->head_;
300 ThreadEntry* threadEntry = getThreadEntry();
301 // If this thread was in the list before the fork, add it back.
302 if (threadEntry->elementsCapacity != 0) {
303 inst_->push_back(threadEntry);
305 inst_->lock_.unlock();
308 static void onThreadExit(void* ptr) {
309 auto& meta = instance();
310 #ifdef FOLLY_TLD_USE_FOLLY_TLS
311 ThreadEntry* threadEntry = getThreadEntry();
313 DCHECK_EQ(ptr, &meta);
314 DCHECK_GT(threadEntry->elementsCapacity, 0);
316 // pthread sets the thread-specific value corresponding
317 // to meta.pthreadKey_ to NULL before calling onThreadExit.
318 // We need to set it back to ptr to enable the correct behaviour
319 // of the subsequent calls of getThreadEntry
320 // (which may happen in user-provided custom deleters)
321 pthread_setspecific(meta.pthreadKey_, ptr);
322 ThreadEntry* threadEntry = static_cast<ThreadEntry*>(ptr);
325 std::lock_guard<std::mutex> g(meta.lock_);
326 meta.erase(threadEntry);
327 // No need to hold the lock any longer; the ThreadEntry is private to this
328 // thread now that it's been removed from meta.
330 // NOTE: User-provided deleter / object dtor itself may be using ThreadLocal
331 // with the same Tag, so dispose() calls below may (re)create some of the
332 // elements or even increase elementsCapacity, thus multiple cleanup rounds
334 for (bool shouldRun = true; shouldRun; ) {
336 FOR_EACH_RANGE(i, 0, threadEntry->elementsCapacity) {
337 if (threadEntry->elements[i].dispose(TLPDestructionMode::THIS_THREAD)) {
342 free(threadEntry->elements);
343 threadEntry->elements = nullptr;
344 pthread_setspecific(meta.pthreadKey_, nullptr);
346 #ifndef FOLLY_TLD_USE_FOLLY_TLS
347 // Allocated in getThreadEntry() when not using folly TLS; free it
352 static uint32_t allocate(EntryID* ent) {
354 auto & meta = instance();
355 std::lock_guard<std::mutex> g(meta.lock_);
357 id = ent->value.load();
358 if (id != kEntryIDInvalid) {
362 if (!meta.freeIds_.empty()) {
363 id = meta.freeIds_.back();
364 meta.freeIds_.pop_back();
369 uint32_t old_id = ent->value.exchange(id);
370 DCHECK_EQ(old_id, kEntryIDInvalid);
374 static void destroy(EntryID* ent) {
376 auto & meta = instance();
377 // Elements in other threads that use this id.
378 std::vector<ElementWrapper> elements;
380 std::lock_guard<std::mutex> g(meta.lock_);
381 uint32_t id = ent->value.exchange(kEntryIDInvalid);
382 if (id == kEntryIDInvalid) {
386 for (ThreadEntry* e = meta.head_.next; e != &meta.head_; e = e->next) {
387 if (id < e->elementsCapacity && e->elements[id].ptr) {
388 elements.push_back(e->elements[id]);
391 * Writing another thread's ThreadEntry from here is fine;
392 * the only other potential reader is the owning thread --
393 * from onThreadExit (which grabs the lock, so is properly
394 * synchronized with us) or from get(), which also grabs
395 * the lock if it needs to resize the elements vector.
397 * We can't conflict with reads for a get(id), because
398 * it's illegal to call get on a thread local that's
401 e->elements[id].ptr = nullptr;
402 e->elements[id].deleter = nullptr;
403 e->elements[id].ownsDeleter = false;
406 meta.freeIds_.push_back(id);
408 // Delete elements outside the lock
409 FOR_EACH(it, elements) {
410 it->dispose(TLPDestructionMode::ALL_THREADS);
412 } catch (...) { // Just in case we get a lock error or something anyway...
413 LOG(WARNING) << "Destructor discarding an exception that was thrown.";
418 * Reserve enough space in the ThreadEntry::elements for the item
421 static void reserve(EntryID* id) {
422 auto& meta = instance();
423 ThreadEntry* threadEntry = getThreadEntry();
424 size_t prevCapacity = threadEntry->elementsCapacity;
426 uint32_t idval = id->getOrAllocate();
427 if (prevCapacity > idval) {
430 // Growth factor < 2, see folly/docs/FBVector.md; + 5 to prevent
432 size_t newCapacity = static_cast<size_t>((idval + 5) * 1.7);
433 assert(newCapacity > prevCapacity);
434 ElementWrapper* reallocated = nullptr;
436 // Need to grow. Note that we can't call realloc, as elements is
437 // still linked in meta, so another thread might access invalid memory
438 // after realloc succeeds. We'll copy by hand and update our ThreadEntry
440 if (usingJEMalloc()) {
441 bool success = false;
442 size_t newByteSize = nallocx(newCapacity * sizeof(ElementWrapper), 0);
444 // Try to grow in place.
446 // Note that xallocx(MALLOCX_ZERO) will only zero newly allocated memory,
447 // even if a previous allocation allocated more than we requested.
448 // This is fine; we always use MALLOCX_ZERO with jemalloc and we
449 // always expand our allocation to the real size.
450 if (prevCapacity * sizeof(ElementWrapper) >=
451 jemallocMinInPlaceExpandable) {
452 success = (xallocx(threadEntry->elements, newByteSize, 0, MALLOCX_ZERO)
456 // In-place growth failed.
458 success = ((reallocated = static_cast<ElementWrapper*>(
459 mallocx(newByteSize, MALLOCX_ZERO))) != nullptr);
463 // Expand to real size
464 assert(newByteSize / sizeof(ElementWrapper) >= newCapacity);
465 newCapacity = newByteSize / sizeof(ElementWrapper);
467 throw std::bad_alloc();
469 } else { // no jemalloc
470 // calloc() is simpler than malloc() followed by memset(), and
471 // potentially faster when dealing with a lot of memory, as it can get
472 // already-zeroed pages from the kernel.
473 reallocated = static_cast<ElementWrapper*>(
474 calloc(newCapacity, sizeof(ElementWrapper)));
476 throw std::bad_alloc();
480 // Success, update the entry
482 std::lock_guard<std::mutex> g(meta.lock_);
484 if (prevCapacity == 0) {
485 meta.push_back(threadEntry);
490 * Note: we need to hold the meta lock when copying data out of
491 * the old vector, because some other thread might be
492 * destructing a ThreadLocal and writing to the elements vector
495 memcpy(reallocated, threadEntry->elements,
496 sizeof(ElementWrapper) * prevCapacity);
498 swap(reallocated, threadEntry->elements);
500 threadEntry->elementsCapacity = newCapacity;
505 #ifdef FOLLY_TLD_USE_FOLLY_TLS
506 if (prevCapacity == 0) {
507 pthread_setspecific(meta.pthreadKey_, &meta);
512 static ElementWrapper& get(EntryID* ent) {
513 ThreadEntry* threadEntry = getThreadEntry();
514 uint32_t id = ent->getOrInvalid();
515 // if id is invalid, it is equal to uint32_t's max value.
516 // x <= max value is always true
517 if (UNLIKELY(threadEntry->elementsCapacity <= id)) {
519 id = ent->getOrInvalid();
520 assert(threadEntry->elementsCapacity > id);
522 return threadEntry->elements[id];
526 #ifdef FOLLY_TLD_USE_FOLLY_TLS
528 FOLLY_TLS ThreadEntry StaticMeta<Tag>::threadEntry_ = {nullptr, 0,
531 template <class Tag> StaticMeta<Tag>* StaticMeta<Tag>::inst_ = nullptr;
533 } // namespace threadlocal_detail
536 #endif /* FOLLY_DETAIL_THREADLOCALDETAIL_H_ */