blob: 2b296dafd75905d0e6c3e6281c7c6cec1f3f3ad1 [file] [log] [blame]
// Copyright (c) 2012 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "base/synchronization/waitable_event_watcher.h"
#include <utility>
#include "base/bind.h"
#include "base/logging.h"
#include "base/synchronization/lock.h"
#include "base/threading/sequenced_task_runner_handle.h"
namespace base {
// -----------------------------------------------------------------------------
// WaitableEventWatcher (async waits).
// The basic design is that we add an AsyncWaiter to the wait-list of the event.
// That AsyncWaiter has a pointer to SequencedTaskRunner, and a Task to be
// posted to it. The task ends up calling the callback when it runs on the
// sequence.
// Since the wait can be canceled, we have a thread-safe Flag object which is
// set when the wait has been canceled. At each stage in the above, we check the
// flag before going onto the next stage. Since the wait may only be canceled in
// the sequence which runs the Task, we are assured that the callback cannot be
// called after canceling...
// -----------------------------------------------------------------------------
// A thread-safe, reference-counted, write-once flag.
// -----------------------------------------------------------------------------
class Flag : public RefCountedThreadSafe<Flag> {
Flag() { flag_ = false; }
void Set() {
AutoLock locked(lock_);
flag_ = true;
bool value() const {
AutoLock locked(lock_);
return flag_;
friend class RefCountedThreadSafe<Flag>;
~Flag() = default;
mutable Lock lock_;
bool flag_;
// -----------------------------------------------------------------------------
// This is an asynchronous waiter which posts a task to a SequencedTaskRunner
// when fired. An AsyncWaiter may only be in a single wait-list.
// -----------------------------------------------------------------------------
class AsyncWaiter : public WaitableEvent::Waiter {
AsyncWaiter(scoped_refptr<SequencedTaskRunner> task_runner,
base::OnceClosure callback,
Flag* flag)
: task_runner_(std::move(task_runner)),
flag_(flag) {}
bool Fire(WaitableEvent* event) override {
// Post the callback if we haven't been cancelled.
if (!flag_->value())
task_runner_->PostTask(FROM_HERE, std::move(callback_));
// We are removed from the wait-list by the WaitableEvent itself. It only
// remains to delete ourselves.
delete this;
// We can always return true because an AsyncWaiter is never in two
// different wait-lists at the same time.
return true;
// See StopWatching for discussion
bool Compare(void* tag) override { return tag == flag_.get(); }
const scoped_refptr<SequencedTaskRunner> task_runner_;
base::OnceClosure callback_;
const scoped_refptr<Flag> flag_;
// -----------------------------------------------------------------------------
// For async waits we need to run a callback on a sequence. We do this by
// posting an AsyncCallbackHelper task, which calls the callback and keeps track
// of when the event is canceled.
// -----------------------------------------------------------------------------
void AsyncCallbackHelper(Flag* flag,
WaitableEventWatcher::EventCallback callback,
WaitableEvent* event) {
// Runs on the sequence that called StartWatching().
if (!flag->value()) {
// This is to let the WaitableEventWatcher know that the event has occured.
WaitableEventWatcher::WaitableEventWatcher() {
WaitableEventWatcher::~WaitableEventWatcher() {
// The destructor may be called from a different sequence than StartWatching()
// when there is no active watch. To avoid triggering a DCHECK in
// StopWatching(), do not call it when there is no active watch.
if (cancel_flag_ && !cancel_flag_->value())
// -----------------------------------------------------------------------------
// The Handle is how the user cancels a wait. After deleting the Handle we
// insure that the delegate cannot be called.
// -----------------------------------------------------------------------------
bool WaitableEventWatcher::StartWatching(
WaitableEvent* event,
EventCallback callback,
scoped_refptr<SequencedTaskRunner> task_runner) {
// A user may call StartWatching from within the callback function. In this
// case, we won't know that we have finished watching, expect that the Flag
// will have been set in AsyncCallbackHelper().
if (cancel_flag_.get() && cancel_flag_->value())
cancel_flag_ = nullptr;
DCHECK(!cancel_flag_) << "StartWatching called while still watching";
cancel_flag_ = new Flag;
OnceClosure internal_callback =
base::BindOnce(&AsyncCallbackHelper, base::RetainedRef(cancel_flag_),
std::move(callback), event);
WaitableEvent::WaitableEventKernel* kernel = event->kernel_.get();
AutoLock locked(kernel->lock_);
if (kernel->signaled_) {
if (!kernel->manual_reset_)
kernel->signaled_ = false;
// No hairpinning - we can't call the delegate directly here. We have to
// post a task to |task_runner| as usual.
task_runner->PostTask(FROM_HERE, std::move(internal_callback));
return true;
kernel_ = kernel;
waiter_ = new AsyncWaiter(std::move(task_runner),
std::move(internal_callback), cancel_flag_.get());
return true;
void WaitableEventWatcher::StopWatching() {
if (!cancel_flag_.get()) // if not currently watching...
if (cancel_flag_->value()) {
// In this case, the event has fired, but we haven't figured that out yet.
// The WaitableEvent may have been deleted too.
cancel_flag_ = nullptr;
if (!kernel_.get()) {
// We have no kernel. This means that we never enqueued a Waiter on an
// event because the event was already signaled when StartWatching was
// called.
// In this case, a task was enqueued on the MessageLoop and will run.
// We set the flag in case the task hasn't yet run. The flag will stop the
// delegate getting called. If the task has run then we have the last
// reference to the flag and it will be deleted immedately after.
cancel_flag_ = nullptr;
AutoLock locked(kernel_->lock_);
// We have a lock on the kernel. No one else can signal the event while we
// have it.
// We have a possible ABA issue here. If Dequeue was to compare only the
// pointer values then it's possible that the AsyncWaiter could have been
// fired, freed and the memory reused for a different Waiter which was
// enqueued in the same wait-list. We would think that that waiter was our
// AsyncWaiter and remove it.
// To stop this, Dequeue also takes a tag argument which is passed to the
// virtual Compare function before the two are considered a match. So we need
// a tag which is good for the lifetime of this handle: the Flag. Since we
// have a reference to the Flag, its memory cannot be reused while this object
// still exists. So if we find a waiter with the correct pointer value, and
// which shares a Flag pointer, we have a real match.
if (kernel_->Dequeue(waiter_, cancel_flag_.get())) {
// Case 2: the waiter hasn't been signaled yet; it was still on the wait
// list. We've removed it, thus we can delete it and the task (which cannot
// have been enqueued with the MessageLoop because the waiter was never
// signaled)
delete waiter_;
cancel_flag_ = nullptr;
// Case 3: the waiter isn't on the wait-list, thus it was signaled. It may not
// have run yet, so we set the flag to tell it not to bother enqueuing the
// task on the SequencedTaskRunner, but to delete it instead. The Waiter
// deletes itself once run.
cancel_flag_ = nullptr;
// If the waiter has already run then the task has been enqueued. If the Task
// hasn't yet run, the flag will stop the delegate from getting called. (This
// is thread safe because one may only delete a Handle from the sequence that
// called StartWatching()).
// If the delegate has already been called then we have nothing to do. The
// task has been deleted by the MessageLoop.
} // namespace base