gecko/xpcom/threads/nsTimerImpl.cpp

832 lines
21 KiB
C++
Raw Normal View History

/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim: set ts=8 sts=2 et sw=2 tw=80: */
/* This Source Code Form is subject to the terms of the Mozilla Public
2012-05-21 04:12:37 -07:00
* 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/. */
#include "nsTimerImpl.h"
#include "TimerThread.h"
#include "nsAutoPtr.h"
#include "nsThreadManager.h"
#include "nsThreadUtils.h"
#include "plarena.h"
#include "pratom.h"
#include "GeckoProfiler.h"
#include "mozilla/Atomics.h"
using mozilla::Atomic;
using mozilla::TimeDuration;
using mozilla::TimeStamp;
static Atomic<int32_t> gGenerator;
static TimerThread* gThread = nullptr;
#ifdef DEBUG_TIMERS
PRLogModuleInfo*
GetTimerLog()
{
static PRLogModuleInfo* sLog;
if (!sLog) {
sLog = PR_NewLogModule("nsTimerImpl");
}
return sLog;
}
#include <math.h>
double nsTimerImpl::sDeltaSumSquared = 0;
double nsTimerImpl::sDeltaSum = 0;
double nsTimerImpl::sDeltaNum = 0;
static void
myNS_MeanAndStdDev(double n, double sumOfValues, double sumOfSquaredValues,
double* meanResult, double* stdDevResult)
{
double mean = 0.0, var = 0.0, stdDev = 0.0;
if (n > 0.0 && sumOfValues >= 0) {
mean = sumOfValues / n;
double temp = (n * sumOfSquaredValues) - (sumOfValues * sumOfValues);
if (temp < 0.0 || n <= 1) {
var = 0.0;
} else {
var = temp / (n * (n - 1));
}
// for some reason, Windows says sqrt(0.0) is "-1.#J" (?!) so do this:
stdDev = var != 0.0 ? sqrt(var) : 0.0;
}
*meanResult = mean;
*stdDevResult = stdDev;
}
#endif
namespace {
// TimerEventAllocator is a thread-safe allocator used only for nsTimerEvents.
// It's needed to avoid contention over the default allocator lock when
// firing timer events (see bug 733277). The thread-safety is required because
// nsTimerEvent objects are allocated on the timer thread, and freed on another
// thread. Because TimerEventAllocator has its own lock, contention over that
// lock is limited to the allocation and deallocation of nsTimerEvent objects.
//
// Because this allocator is layered over PLArenaPool, it never shrinks -- even
// "freed" nsTimerEvents aren't truly freed, they're just put onto a free-list
// for later recycling. So the amount of memory consumed will always be equal
// to the high-water mark consumption. But nsTimerEvents are small and it's
// unusual to have more than a few hundred of them, so this shouldn't be a
// problem in practice.
class TimerEventAllocator
{
private:
struct FreeEntry
{
FreeEntry* mNext;
};
PLArenaPool mPool;
FreeEntry* mFirstFree;
mozilla::Monitor mMonitor;
public:
TimerEventAllocator()
: mFirstFree(nullptr)
, mMonitor("TimerEventAllocator")
{
PL_InitArenaPool(&mPool, "TimerEventPool", 4096, /* align = */ 0);
}
~TimerEventAllocator()
{
PL_FinishArenaPool(&mPool);
}
void* Alloc(size_t aSize);
void Free(void* aPtr);
};
} // anonymous namespace
class nsTimerEvent : public nsRunnable
{
public:
NS_IMETHOD Run();
nsTimerEvent()
: mTimer()
, mGeneration(0)
{
MOZ_COUNT_CTOR(nsTimerEvent);
MOZ_ASSERT(gThread->IsOnTimerThread(),
"nsTimer must always be allocated on the timer thread");
sAllocatorUsers++;
}
#ifdef DEBUG_TIMERS
TimeStamp mInitTime;
#endif
static void Init();
static void Shutdown();
static void DeleteAllocatorIfNeeded();
static void* operator new(size_t aSize) CPP_THROW_NEW
{
return sAllocator->Alloc(aSize);
}
void operator delete(void* aPtr)
{
sAllocator->Free(aPtr);
DeleteAllocatorIfNeeded();
}
already_AddRefed<nsTimerImpl> ForgetTimer()
{
return mTimer.forget();
}
void SetTimer(already_AddRefed<nsTimerImpl> aTimer)
{
mTimer = aTimer;
mGeneration = mTimer->GetGeneration();
}
private:
~nsTimerEvent()
{
MOZ_COUNT_DTOR(nsTimerEvent);
MOZ_ASSERT(!sCanDeleteAllocator || sAllocatorUsers > 0,
"This will result in us attempting to deallocate the nsTimerEvent allocator twice");
sAllocatorUsers--;
}
nsRefPtr<nsTimerImpl> mTimer;
int32_t mGeneration;
static TimerEventAllocator* sAllocator;
static Atomic<int32_t> sAllocatorUsers;
static bool sCanDeleteAllocator;
};
TimerEventAllocator* nsTimerEvent::sAllocator = nullptr;
Atomic<int32_t> nsTimerEvent::sAllocatorUsers;
bool nsTimerEvent::sCanDeleteAllocator = false;
namespace {
void*
TimerEventAllocator::Alloc(size_t aSize)
{
MOZ_ASSERT(aSize == sizeof(nsTimerEvent));
mozilla::MonitorAutoLock lock(mMonitor);
void* p;
if (mFirstFree) {
p = mFirstFree;
mFirstFree = mFirstFree->mNext;
} else {
PL_ARENA_ALLOCATE(p, &mPool, aSize);
if (!p) {
return nullptr;
}
}
return p;
}
void
TimerEventAllocator::Free(void* aPtr)
{
mozilla::MonitorAutoLock lock(mMonitor);
FreeEntry* entry = reinterpret_cast<FreeEntry*>(aPtr);
entry->mNext = mFirstFree;
mFirstFree = entry;
}
} // anonymous namespace
NS_IMPL_QUERY_INTERFACE(nsTimerImpl, nsITimer)
NS_IMPL_ADDREF(nsTimerImpl)
NS_IMETHODIMP_(MozExternalRefCountType)
nsTimerImpl::Release(void)
{
nsrefcnt count;
MOZ_ASSERT(int32_t(mRefCnt) > 0, "dup release");
count = --mRefCnt;
NS_LOG_RELEASE(this, count, "nsTimerImpl");
if (count == 0) {
mRefCnt = 1; /* stabilize */
/* enable this to find non-threadsafe destructors: */
/* NS_ASSERT_OWNINGTHREAD(nsTimerImpl); */
delete this;
return 0;
}
// If only one reference remains, and mArmed is set, then the ref must be
// from the TimerThread::mTimers array, so we Cancel this timer to remove
// the mTimers element, and return 0 if Cancel in fact disarmed the timer.
//
// We use an inlined version of nsTimerImpl::Cancel here to check for the
// NS_ERROR_NOT_AVAILABLE code returned by gThread->RemoveTimer when this
// timer is not found in the mTimers array -- i.e., when the timer was not
// in fact armed once we acquired TimerThread::mLock, in spite of mArmed
// being true here. That can happen if the armed timer is being fired by
// TimerThread::Run as we race and test mArmed just before it is cleared by
// the timer thread. If the RemoveTimer call below doesn't find this timer
// in the mTimers array, then the last ref to this timer is held manually
// and temporarily by the TimerThread, so we should fall through to the
// final return and return 1, not 0.
//
// The original version of this thread-based timer code kept weak refs from
// TimerThread::mTimers, removing this timer's weak ref in the destructor,
// but that leads to double-destructions in the race described above, and
// adding mArmed doesn't help, because destructors can't be deferred, once
// begun. But by combining reference-counting and a specialized Release
// method with "is this timer still in the mTimers array once we acquire
// the TimerThread's lock" testing, we defer destruction until we're sure
// that only one thread has its hot little hands on this timer.
//
// Note that both approaches preclude a timer creator, and everyone else
// except the TimerThread who might have a strong ref, from dropping all
// their strong refs without implicitly canceling the timer. Timers need
// non-mTimers-element strong refs to stay alive.
if (count == 1 && mArmed) {
mCanceled = true;
MOZ_ASSERT(gThread, "Armed timer exists after the thread timer stopped.");
if (NS_SUCCEEDED(gThread->RemoveTimer(this))) {
return 0;
}
}
return count;
}
nsTimerImpl::nsTimerImpl() :
mClosure(nullptr),
mCallbackType(CALLBACK_TYPE_UNKNOWN),
mFiring(false),
mArmed(false),
mCanceled(false),
mGeneration(0),
mDelay(0)
{
// XXXbsmedberg: shouldn't this be in Init()?
mEventTarget = static_cast<nsIEventTarget*>(NS_GetCurrentThread());
mCallback.c = nullptr;
}
nsTimerImpl::~nsTimerImpl()
{
ReleaseCallback();
}
//static
nsresult
nsTimerImpl::Startup()
{
nsresult rv;
nsTimerEvent::Init();
gThread = new TimerThread();
if (!gThread) {
return NS_ERROR_OUT_OF_MEMORY;
}
NS_ADDREF(gThread);
rv = gThread->InitLocks();
if (NS_FAILED(rv)) {
NS_RELEASE(gThread);
}
return rv;
}
void
nsTimerImpl::Shutdown()
{
#ifdef DEBUG_TIMERS
if (PR_LOG_TEST(GetTimerLog(), PR_LOG_DEBUG)) {
double mean = 0, stddev = 0;
myNS_MeanAndStdDev(sDeltaNum, sDeltaSum, sDeltaSumSquared, &mean, &stddev);
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("sDeltaNum = %f, sDeltaSum = %f, sDeltaSumSquared = %f\n",
sDeltaNum, sDeltaSum, sDeltaSumSquared));
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("mean: %fms, stddev: %fms\n", mean, stddev));
}
#endif
if (!gThread) {
return;
}
gThread->Shutdown();
NS_RELEASE(gThread);
nsTimerEvent::Shutdown();
}
nsresult
nsTimerImpl::InitCommon(uint32_t aType, uint32_t aDelay)
{
nsresult rv;
if (NS_WARN_IF(!gThread)) {
return NS_ERROR_NOT_INITIALIZED;
}
if (!mEventTarget) {
NS_ERROR("mEventTarget is NULL");
return NS_ERROR_NOT_INITIALIZED;
}
rv = gThread->Init();
if (NS_WARN_IF(NS_FAILED(rv))) {
return rv;
}
/**
* In case of re-Init, both with and without a preceding Cancel, clear the
* mCanceled flag and assign a new mGeneration. But first, remove any armed
* timer from the timer thread's list.
*
* If we are racing with the timer thread to remove this timer and we lose,
* the RemoveTimer call made here will fail to find this timer in the timer
* thread's list, and will return false harmlessly. We test mArmed here to
* avoid the small overhead in RemoveTimer of locking the timer thread and
* checking its list for this timer. It's safe to test mArmed even though
* it might be cleared on another thread in the next cycle (or even already
* be cleared by another CPU whose store hasn't reached our CPU's cache),
* because RemoveTimer is idempotent.
*/
if (mArmed) {
gThread->RemoveTimer(this);
}
mCanceled = false;
mTimeout = TimeStamp();
mGeneration = gGenerator++;
mType = (uint8_t)aType;
SetDelayInternal(aDelay);
return gThread->AddTimer(this);
}
NS_IMETHODIMP
nsTimerImpl::InitWithFuncCallback(nsTimerCallbackFunc aFunc,
void* aClosure,
uint32_t aDelay,
uint32_t aType)
{
if (NS_WARN_IF(!aFunc)) {
return NS_ERROR_INVALID_ARG;
}
ReleaseCallback();
mCallbackType = CALLBACK_TYPE_FUNC;
mCallback.c = aFunc;
mClosure = aClosure;
return InitCommon(aType, aDelay);
}
NS_IMETHODIMP
nsTimerImpl::InitWithCallback(nsITimerCallback* aCallback,
uint32_t aDelay,
uint32_t aType)
{
if (NS_WARN_IF(!aCallback)) {
return NS_ERROR_INVALID_ARG;
}
ReleaseCallback();
mCallbackType = CALLBACK_TYPE_INTERFACE;
mCallback.i = aCallback;
NS_ADDREF(mCallback.i);
return InitCommon(aType, aDelay);
}
NS_IMETHODIMP
nsTimerImpl::Init(nsIObserver* aObserver, uint32_t aDelay, uint32_t aType)
{
if (NS_WARN_IF(!aObserver)) {
return NS_ERROR_INVALID_ARG;
}
ReleaseCallback();
mCallbackType = CALLBACK_TYPE_OBSERVER;
mCallback.o = aObserver;
NS_ADDREF(mCallback.o);
return InitCommon(aType, aDelay);
}
NS_IMETHODIMP
nsTimerImpl::Cancel()
{
mCanceled = true;
if (gThread) {
gThread->RemoveTimer(this);
}
ReleaseCallback();
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::SetDelay(uint32_t aDelay)
{
if (mCallbackType == CALLBACK_TYPE_UNKNOWN && mType == TYPE_ONE_SHOT) {
// This may happen if someone tries to re-use a one-shot timer
// by re-setting delay instead of reinitializing the timer.
NS_ERROR("nsITimer->SetDelay() called when the "
"one-shot timer is not set up.");
return NS_ERROR_NOT_INITIALIZED;
}
// If we're already repeating precisely, update mTimeout now so that the
// new delay takes effect in the future.
if (!mTimeout.IsNull() && mType == TYPE_REPEATING_PRECISE) {
mTimeout = TimeStamp::Now();
}
SetDelayInternal(aDelay);
if (!mFiring && gThread) {
gThread->TimerDelayChanged(this);
}
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::GetDelay(uint32_t* aDelay)
{
*aDelay = mDelay;
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::SetType(uint32_t aType)
{
mType = (uint8_t)aType;
// XXX if this is called, we should change the actual type.. this could effect
// repeating timers. we need to ensure in Fire() that if mType has changed
// during the callback that we don't end up with the timer in the queue twice.
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::GetType(uint32_t* aType)
{
*aType = mType;
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::GetClosure(void** aClosure)
{
*aClosure = mClosure;
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::GetCallback(nsITimerCallback** aCallback)
{
if (mCallbackType == CALLBACK_TYPE_INTERFACE) {
NS_IF_ADDREF(*aCallback = mCallback.i);
} else if (mTimerCallbackWhileFiring) {
NS_ADDREF(*aCallback = mTimerCallbackWhileFiring);
} else {
*aCallback = nullptr;
}
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::GetTarget(nsIEventTarget** aTarget)
{
NS_IF_ADDREF(*aTarget = mEventTarget);
return NS_OK;
}
NS_IMETHODIMP
nsTimerImpl::SetTarget(nsIEventTarget* aTarget)
{
if (NS_WARN_IF(mCallbackType != CALLBACK_TYPE_UNKNOWN)) {
return NS_ERROR_ALREADY_INITIALIZED;
}
if (aTarget) {
mEventTarget = aTarget;
} else {
mEventTarget = static_cast<nsIEventTarget*>(NS_GetCurrentThread());
}
return NS_OK;
}
void
nsTimerImpl::Fire()
{
if (mCanceled) {
return;
}
PROFILER_LABEL("Timer", "Fire",
js::ProfileEntry::Category::OTHER);
#ifdef MOZ_TASK_TRACER
mozilla::tasktracer::AutoRunFakeTracedTask runTracedTask(mTracedTask);
#endif
#ifdef DEBUG_TIMERS
TimeStamp now = TimeStamp::Now();
if (PR_LOG_TEST(GetTimerLog(), PR_LOG_DEBUG)) {
TimeDuration a = now - mStart; // actual delay in intervals
TimeDuration b = TimeDuration::FromMilliseconds(mDelay); // expected delay in intervals
TimeDuration delta = (a > b) ? a - b : b - a;
uint32_t d = delta.ToMilliseconds(); // delta in ms
sDeltaSum += d;
sDeltaSumSquared += double(d) * double(d);
sDeltaNum++;
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("[this=%p] expected delay time %4ums\n", this, mDelay));
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("[this=%p] actual delay time %fms\n", this,
a.ToMilliseconds()));
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("[this=%p] (mType is %d) -------\n", this, mType));
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("[this=%p] delta %4dms\n",
this, (a > b) ? (int32_t)d : -(int32_t)d));
mStart = mStart2;
mStart2 = TimeStamp();
}
#endif
TimeStamp timeout = mTimeout;
Bug 650379. Add a new XPCOM timer type that is like TYPE_REPEATING_PRECISE but does not swamp the event queue if the callback takes longer than the timer interval to run. r=cjones, sr=brendan This implements proposal 3 from bug 650379 comment 13. The main difference between TYPE_REPEATING_PRECISE and TYPE_REPEATING_PRECISE_CAN_SKIP is to not AddTimer the REPEATING_PRECISE_CAN_SKIP timer until after the callback has run; this guarantees that no more timer events will be posted until after the callback finishes executing. A secondary change is to make REPEATING_PRECISE_CAN_SKIP timers advance their firing time to mDelay from when PostTimerEvent is called, not mDelay from the old mTimeout. While this arguably makes them less precise, the alternative is that if a timer is significantly delayed for some reason (e.g. because the user puts the computer to sleep for a while) it will then fire a whole bunch of times to "catch up" to where it's supposed to be, advancing its firing time by mDelay at a time. That seems undesirable. An alternate approach would have been to readd the timer from inside PostTimerEvent, but only if we're not in the middle of firing the timer. That would allow more precise timers in the case when the callback is not taking too long, but still handle gracefully the case when the callback is slow. Unfortunately this falls down if something _else_ is hogging the main thread event loop (e.g. some other timer has a slow callback, or whatever); in that case we would post multiple events for the one precise timer while the event-loop-hogging operation is running. So I don't think we should do that.
2011-04-28 16:33:52 -07:00
if (IsRepeatingPrecisely()) {
// Precise repeating timers advance mTimeout by mDelay without fail before
// calling Fire().
timeout -= TimeDuration::FromMilliseconds(mDelay);
}
if (mCallbackType == CALLBACK_TYPE_INTERFACE) {
mTimerCallbackWhileFiring = mCallback.i;
}
mFiring = true;
// Handle callbacks that re-init the timer, but avoid leaking.
// See bug 330128.
CallbackUnion callback = mCallback;
unsigned callbackType = mCallbackType;
if (callbackType == CALLBACK_TYPE_INTERFACE) {
NS_ADDREF(callback.i);
} else if (callbackType == CALLBACK_TYPE_OBSERVER) {
NS_ADDREF(callback.o);
}
ReleaseCallback();
switch (callbackType) {
case CALLBACK_TYPE_FUNC:
callback.c(this, mClosure);
break;
case CALLBACK_TYPE_INTERFACE:
callback.i->Notify(this);
break;
case CALLBACK_TYPE_OBSERVER:
callback.o->Observe(static_cast<nsITimer*>(this),
NS_TIMER_CALLBACK_TOPIC,
nullptr);
break;
default:
;
}
// If the callback didn't re-init the timer, and it's not a one-shot timer,
// restore the callback state.
if (mCallbackType == CALLBACK_TYPE_UNKNOWN &&
mType != TYPE_ONE_SHOT && !mCanceled) {
mCallback = callback;
mCallbackType = callbackType;
} else {
// The timer was a one-shot, or the callback was reinitialized.
if (callbackType == CALLBACK_TYPE_INTERFACE) {
NS_RELEASE(callback.i);
} else if (callbackType == CALLBACK_TYPE_OBSERVER) {
NS_RELEASE(callback.o);
}
}
mFiring = false;
mTimerCallbackWhileFiring = nullptr;
#ifdef DEBUG_TIMERS
if (PR_LOG_TEST(GetTimerLog(), PR_LOG_DEBUG)) {
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("[this=%p] Took %fms to fire timer callback\n",
this, (TimeStamp::Now() - now).ToMilliseconds()));
}
#endif
Bug 650379. Add a new XPCOM timer type that is like TYPE_REPEATING_PRECISE but does not swamp the event queue if the callback takes longer than the timer interval to run. r=cjones, sr=brendan This implements proposal 3 from bug 650379 comment 13. The main difference between TYPE_REPEATING_PRECISE and TYPE_REPEATING_PRECISE_CAN_SKIP is to not AddTimer the REPEATING_PRECISE_CAN_SKIP timer until after the callback has run; this guarantees that no more timer events will be posted until after the callback finishes executing. A secondary change is to make REPEATING_PRECISE_CAN_SKIP timers advance their firing time to mDelay from when PostTimerEvent is called, not mDelay from the old mTimeout. While this arguably makes them less precise, the alternative is that if a timer is significantly delayed for some reason (e.g. because the user puts the computer to sleep for a while) it will then fire a whole bunch of times to "catch up" to where it's supposed to be, advancing its firing time by mDelay at a time. That seems undesirable. An alternate approach would have been to readd the timer from inside PostTimerEvent, but only if we're not in the middle of firing the timer. That would allow more precise timers in the case when the callback is not taking too long, but still handle gracefully the case when the callback is slow. Unfortunately this falls down if something _else_ is hogging the main thread event loop (e.g. some other timer has a slow callback, or whatever); in that case we would post multiple events for the one precise timer while the event-loop-hogging operation is running. So I don't think we should do that.
2011-04-28 16:33:52 -07:00
// Reschedule repeating timers, except REPEATING_PRECISE which already did
// that in PostTimerEvent, but make sure that we aren't armed already (which
// can happen if the callback reinitialized the timer).
if (IsRepeating() && mType != TYPE_REPEATING_PRECISE && !mArmed) {
if (mType == TYPE_REPEATING_SLACK) {
SetDelayInternal(mDelay); // force mTimeout to be recomputed. For
}
// REPEATING_PRECISE_CAN_SKIP timers this has
// already happened.
if (gThread) {
gThread->AddTimer(this);
}
}
}
void
nsTimerEvent::Init()
{
sAllocator = new TimerEventAllocator();
}
void
nsTimerEvent::Shutdown()
{
sCanDeleteAllocator = true;
DeleteAllocatorIfNeeded();
}
void
nsTimerEvent::DeleteAllocatorIfNeeded()
{
if (sCanDeleteAllocator && sAllocatorUsers == 0) {
delete sAllocator;
sAllocator = nullptr;
}
}
NS_IMETHODIMP
nsTimerEvent::Run()
{
if (mGeneration != mTimer->GetGeneration()) {
return NS_OK;
}
#ifdef DEBUG_TIMERS
if (PR_LOG_TEST(GetTimerLog(), PR_LOG_DEBUG)) {
TimeStamp now = TimeStamp::Now();
PR_LOG(GetTimerLog(), PR_LOG_DEBUG,
("[this=%p] time between PostTimerEvent() and Fire(): %fms\n",
this, (now - mInitTime).ToMilliseconds()));
}
#endif
mTimer->Fire();
// Since nsTimerImpl is not thread-safe, we should release |mTimer|
// here in the target thread to avoid race condition. Otherwise,
// ~nsTimerEvent() which calls nsTimerImpl::Release() could run in the
// timer thread and result in race condition.
mTimer = nullptr;
return NS_OK;
}
already_AddRefed<nsTimerImpl>
nsTimerImpl::PostTimerEvent(already_AddRefed<nsTimerImpl> aTimerRef)
{
nsRefPtr<nsTimerImpl> timer(aTimerRef);
if (!timer->mEventTarget) {
NS_ERROR("Attempt to post timer event to NULL event target");
return timer.forget();
}
// XXX we may want to reuse this nsTimerEvent in the case of repeating timers.
// Since TimerThread addref'd 'timer' for us, we don't need to addref here.
// We will release either in ~nsTimerEvent(), or pass the reference back to
// the caller. We need to copy the generation number from this timer into the
// event, so we can avoid firing a timer that was re-initialized after being
// canceled.
// Note: We override operator new for this class, and the override is
// fallible!
nsRefPtr<nsTimerEvent> event = new nsTimerEvent;
if (!event) {
return timer.forget();
}
#ifdef DEBUG_TIMERS
if (PR_LOG_TEST(GetTimerLog(), PR_LOG_DEBUG)) {
event->mInitTime = TimeStamp::Now();
}
#endif
// If this is a repeating precise timer, we need to calculate the time for
// the next timer to fire before we make the callback.
if (timer->IsRepeatingPrecisely()) {
timer->SetDelayInternal(timer->mDelay);
Bug 650379. Add a new XPCOM timer type that is like TYPE_REPEATING_PRECISE but does not swamp the event queue if the callback takes longer than the timer interval to run. r=cjones, sr=brendan This implements proposal 3 from bug 650379 comment 13. The main difference between TYPE_REPEATING_PRECISE and TYPE_REPEATING_PRECISE_CAN_SKIP is to not AddTimer the REPEATING_PRECISE_CAN_SKIP timer until after the callback has run; this guarantees that no more timer events will be posted until after the callback finishes executing. A secondary change is to make REPEATING_PRECISE_CAN_SKIP timers advance their firing time to mDelay from when PostTimerEvent is called, not mDelay from the old mTimeout. While this arguably makes them less precise, the alternative is that if a timer is significantly delayed for some reason (e.g. because the user puts the computer to sleep for a while) it will then fire a whole bunch of times to "catch up" to where it's supposed to be, advancing its firing time by mDelay at a time. That seems undesirable. An alternate approach would have been to readd the timer from inside PostTimerEvent, but only if we're not in the middle of firing the timer. That would allow more precise timers in the case when the callback is not taking too long, but still handle gracefully the case when the callback is slow. Unfortunately this falls down if something _else_ is hogging the main thread event loop (e.g. some other timer has a slow callback, or whatever); in that case we would post multiple events for the one precise timer while the event-loop-hogging operation is running. So I don't think we should do that.
2011-04-28 16:33:52 -07:00
// But only re-arm REPEATING_PRECISE timers.
if (gThread && timer->mType == TYPE_REPEATING_PRECISE) {
nsresult rv = gThread->AddTimer(timer);
if (NS_FAILED(rv)) {
return timer.forget();
}
}
}
nsIEventTarget* target = timer->mEventTarget;
event->SetTimer(timer.forget());
nsresult rv = target->Dispatch(event, NS_DISPATCH_NORMAL);
if (NS_FAILED(rv)) {
timer = event->ForgetTimer();
if (gThread) {
gThread->RemoveTimer(timer);
}
return timer.forget();
}
return nullptr;
}
void
nsTimerImpl::SetDelayInternal(uint32_t aDelay)
{
TimeDuration delayInterval = TimeDuration::FromMilliseconds(aDelay);
mDelay = aDelay;
TimeStamp now = TimeStamp::Now();
if (mTimeout.IsNull() || mType != TYPE_REPEATING_PRECISE) {
mTimeout = now;
}
mTimeout += delayInterval;
#ifdef DEBUG_TIMERS
if (PR_LOG_TEST(GetTimerLog(), PR_LOG_DEBUG)) {
if (mStart.IsNull()) {
mStart = now;
} else {
mStart2 = now;
}
}
#endif
}
size_t
nsTimerImpl::SizeOfIncludingThis(mozilla::MallocSizeOf aMallocSizeOf) const
{
return aMallocSizeOf(this);
}
// NOT FOR PUBLIC CONSUMPTION!
nsresult
NS_NewTimer(nsITimer** aResult, nsTimerCallbackFunc aCallback, void* aClosure,
uint32_t aDelay, uint32_t aType)
{
nsTimerImpl* timer = new nsTimerImpl();
if (!timer) {
return NS_ERROR_OUT_OF_MEMORY;
}
NS_ADDREF(timer);
nsresult rv = timer->InitWithFuncCallback(aCallback, aClosure,
aDelay, aType);
if (NS_FAILED(rv)) {
NS_RELEASE(timer);
return rv;
}
*aResult = timer;
return NS_OK;
}