gecko-dev/netwerk/ipc/ChannelEventQueue.h
Shih-Chiang Chien e6a79fe88f Bug 1426751 - Prevent FlushQueue to be invoked on multiple threads simultaneously. r=mayhemer
This issue is triggered by off-main-thread ODA listener that return error cause from OnDataAvailable callback.
A CancelEvent will be prepend to event queue and trigger race condition between CompleteResume and EndForceEnqueueing.

The `mFlushing` is checked and set in separate critical sections, therefore two threads that executing MayFlushQueue
might both pass the `mFlushing` check and trying to call FlushQueue simultaneously.

The solution is to check and set `mFlushing` in single critical section, so we can guarantee that only one FlushQueue
can be executed at anytime.

In addition, resumption is postponed until no AutoEventEnqueuer is activated. Therefore, CompleteResume will only be
triggered while all the suspension requests and auto enqueue requests are finished.

MozReview-Commit-ID: HpxzgUqYm8C

--HG--
extra : rebase_source : 9be0d37bb4475a0817dafb6270585263160a9da1
2018-01-04 18:53:02 +08:00

358 lines
9.7 KiB
C++

/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*-
* vim: set sw=2 ts=8 et tw=80 :
*/
/* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
#ifndef mozilla_net_ChannelEventQueue_h
#define mozilla_net_ChannelEventQueue_h
#include "nsTArray.h"
#include "nsAutoPtr.h"
#include "nsIEventTarget.h"
#include "nsThreadUtils.h"
#include "nsXULAppAPI.h"
#include "mozilla/DebugOnly.h"
#include "mozilla/Mutex.h"
#include "mozilla/RecursiveMutex.h"
#include "mozilla/UniquePtr.h"
#include "mozilla/Unused.h"
class nsISupports;
namespace mozilla {
namespace net {
class ChannelEvent
{
public:
ChannelEvent() { MOZ_COUNT_CTOR(ChannelEvent); }
virtual ~ChannelEvent() { MOZ_COUNT_DTOR(ChannelEvent); }
virtual void Run() = 0;
virtual already_AddRefed<nsIEventTarget> GetEventTarget() = 0;
};
// Note that MainThreadChannelEvent should not be used in child process since
// GetEventTarget() directly returns an unlabeled event target.
class MainThreadChannelEvent : public ChannelEvent
{
public:
MainThreadChannelEvent() { MOZ_COUNT_CTOR(MainThreadChannelEvent); }
virtual ~MainThreadChannelEvent() { MOZ_COUNT_DTOR(MainThreadChannelEvent); }
already_AddRefed<nsIEventTarget>
GetEventTarget() override
{
MOZ_ASSERT(XRE_IsParentProcess());
return do_AddRef(GetMainThreadEventTarget());
}
};
// This event is designed to be only used for e10s child channels.
// The goal is to force the child channel to implement GetNeckoTarget()
// which should return a labeled main thread event target so that this
// channel event can be dispatched correctly.
template<typename T>
class NeckoTargetChannelEvent : public ChannelEvent
{
public:
explicit NeckoTargetChannelEvent(T *aChild)
: mChild(aChild)
{
MOZ_COUNT_CTOR(NeckoTargetChannelEvent);
}
virtual ~NeckoTargetChannelEvent()
{
MOZ_COUNT_DTOR(NeckoTargetChannelEvent);
}
already_AddRefed<nsIEventTarget>
GetEventTarget() override
{
MOZ_ASSERT(mChild);
return mChild->GetNeckoTarget();
}
protected:
T *mChild;
};
// Workaround for Necko re-entrancy dangers. We buffer IPDL messages in a
// queue if still dispatching previous one(s) to listeners/observers.
// Otherwise synchronous XMLHttpRequests and/or other code that spins the
// event loop (ex: IPDL rpc) could cause listener->OnDataAvailable (for
// instance) to be dispatched and called before mListener->OnStartRequest has
// completed.
class ChannelEventQueue final
{
NS_INLINE_DECL_THREADSAFE_REFCOUNTING(ChannelEventQueue)
public:
explicit ChannelEventQueue(nsISupports *owner)
: mSuspendCount(0)
, mSuspended(false)
, mForcedCount(0)
, mFlushing(false)
, mOwner(owner)
, mMutex("ChannelEventQueue::mMutex")
, mRunningMutex("ChannelEventQueue::mRunningMutex")
{}
// Puts IPDL-generated channel event into queue, to be run later
// automatically when EndForcedQueueing and/or Resume is called.
//
// @param aCallback - the ChannelEvent
// @param aAssertionWhenNotQueued - this optional param will be used in an
// assertion when the event is executed directly.
inline void RunOrEnqueue(ChannelEvent* aCallback,
bool aAssertionWhenNotQueued = false);
// Append ChannelEvent in front of the event queue.
inline nsresult PrependEvent(UniquePtr<ChannelEvent>& aEvent);
inline nsresult PrependEvents(nsTArray<UniquePtr<ChannelEvent>>& aEvents);
// After StartForcedQueueing is called, RunOrEnqueue() will start enqueuing
// events that will be run/flushed when EndForcedQueueing is called.
// - Note: queueing may still be required after EndForcedQueueing() (if the
// queue is suspended, etc): always call RunOrEnqueue() to avoid race
// conditions.
inline void StartForcedQueueing();
inline void EndForcedQueueing();
// Suspend/resume event queue. RunOrEnqueue() will start enqueuing
// events and they will be run/flushed when resume is called. These should be
// called when the channel owning the event queue is suspended/resumed.
void Suspend();
// Resume flushes the queue asynchronously, i.e. items in queue will be
// dispatched in a new event on the current thread.
void Resume();
private:
// Private destructor, to discourage deletion outside of Release():
~ChannelEventQueue()
{
}
void SuspendInternal();
void ResumeInternal();
inline void MaybeFlushQueue();
void FlushQueue();
inline void CompleteResume();
ChannelEvent* TakeEvent();
nsTArray<UniquePtr<ChannelEvent>> mEventQueue;
uint32_t mSuspendCount;
bool mSuspended;
uint32_t mForcedCount; // Support ForcedQueueing on multiple thread.
bool mFlushing;
// Keep ptr to avoid refcount cycle: only grab ref during flushing.
nsISupports *mOwner;
// For atomic mEventQueue operation and state update
Mutex mMutex;
// To guarantee event execution order among threads
RecursiveMutex mRunningMutex;
friend class AutoEventEnqueuer;
};
inline void
ChannelEventQueue::RunOrEnqueue(ChannelEvent* aCallback,
bool aAssertionWhenNotQueued)
{
MOZ_ASSERT(aCallback);
// Events execution could be a destruction of the channel (and our own
// destructor) unless we make sure its refcount doesn't drop to 0 while this
// method is running.
nsCOMPtr<nsISupports> kungFuDeathGrip(mOwner);
Unused << kungFuDeathGrip; // Not used in this function
// To avoid leaks.
UniquePtr<ChannelEvent> event(aCallback);
// To guarantee that the running event and all the events generated within
// it will be finished before events on other threads.
RecursiveMutexAutoLock lock(mRunningMutex);
{
MutexAutoLock lock(mMutex);
bool enqueue = !!mForcedCount || mSuspended || mFlushing || !mEventQueue.IsEmpty();
if (enqueue) {
mEventQueue.AppendElement(Move(event));
return;
}
nsCOMPtr<nsIEventTarget> target = event->GetEventTarget();
MOZ_ASSERT(target);
bool isCurrentThread = false;
DebugOnly<nsresult> rv = target->IsOnCurrentThread(&isCurrentThread);
MOZ_ASSERT(NS_SUCCEEDED(rv));
if (!isCurrentThread) {
// Leverage Suspend/Resume mechanism to trigger flush procedure without
// creating a new one.
SuspendInternal();
mEventQueue.AppendElement(Move(event));
ResumeInternal();
return;
}
}
MOZ_RELEASE_ASSERT(!aAssertionWhenNotQueued);
event->Run();
}
inline void
ChannelEventQueue::StartForcedQueueing()
{
MutexAutoLock lock(mMutex);
++mForcedCount;
}
inline void
ChannelEventQueue::EndForcedQueueing()
{
bool tryFlush = false;
{
MutexAutoLock lock(mMutex);
MOZ_ASSERT(mForcedCount > 0);
if(!--mForcedCount) {
tryFlush = true;
}
}
if (tryFlush) {
MaybeFlushQueue();
}
}
inline nsresult
ChannelEventQueue::PrependEvent(UniquePtr<ChannelEvent>& aEvent)
{
MutexAutoLock lock(mMutex);
// Prepending event while no queue flush foreseen might cause the following
// channel events not run. This assertion here guarantee there must be a
// queue flush, either triggered by Resume or EndForcedQueueing, to execute
// the added event.
MOZ_ASSERT(mSuspended || !!mForcedCount);
UniquePtr<ChannelEvent>* newEvent =
mEventQueue.InsertElementAt(0, Move(aEvent));
if (!newEvent) {
return NS_ERROR_OUT_OF_MEMORY;
}
return NS_OK;
}
inline nsresult
ChannelEventQueue::PrependEvents(nsTArray<UniquePtr<ChannelEvent>>& aEvents)
{
MutexAutoLock lock(mMutex);
// Prepending event while no queue flush foreseen might cause the following
// channel events not run. This assertion here guarantee there must be a
// queue flush, either triggered by Resume or EndForcedQueueing, to execute
// the added events.
MOZ_ASSERT(mSuspended || !!mForcedCount);
UniquePtr<ChannelEvent>* newEvents =
mEventQueue.InsertElementsAt(0, aEvents.Length());
if (!newEvents) {
return NS_ERROR_OUT_OF_MEMORY;
}
for (uint32_t i = 0; i < aEvents.Length(); i++) {
newEvents[i] = Move(aEvents[i]);
}
return NS_OK;
}
inline void
ChannelEventQueue::CompleteResume()
{
bool tryFlush = false;
{
MutexAutoLock lock(mMutex);
// channel may have been suspended again since Resume fired event to call
// this.
if (!mSuspendCount) {
// we need to remain logically suspended (for purposes of queuing incoming
// messages) until this point, else new incoming messages could run before
// queued ones.
mSuspended = false;
tryFlush = true;
}
}
if (tryFlush) {
MaybeFlushQueue();
}
}
inline void
ChannelEventQueue::MaybeFlushQueue()
{
// Don't flush if forced queuing on, we're already being flushed, or
// suspended, or there's nothing to flush
bool flushQueue = false;
{
MutexAutoLock lock(mMutex);
flushQueue = !mForcedCount && !mFlushing && !mSuspended &&
!mEventQueue.IsEmpty();
// Only one thread is allowed to run FlushQueue at a time.
if (flushQueue) {
mFlushing = true;
}
}
if (flushQueue) {
FlushQueue();
}
}
// Ensures that RunOrEnqueue() will be collecting events during its lifetime
// (letting caller know incoming IPDL msgs should be queued). Flushes the queue
// when it goes out of scope.
class MOZ_STACK_CLASS AutoEventEnqueuer
{
public:
explicit AutoEventEnqueuer(ChannelEventQueue *queue)
: mEventQueue(queue)
, mOwner(queue->mOwner)
{
mEventQueue->StartForcedQueueing();
}
~AutoEventEnqueuer() {
mEventQueue->EndForcedQueueing();
}
private:
RefPtr<ChannelEventQueue> mEventQueue;
// Ensure channel object lives longer than ChannelEventQueue.
nsCOMPtr<nsISupports> mOwner;
};
} // namespace net
} // namespace mozilla
#endif