gecko-dev/dom/media/GraphDriver.h

577 строки
21 KiB
C
Исходник Обычный вид История

/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* 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 GRAPHDRIVER_H_
#define GRAPHDRIVER_H_
#include "nsAutoPtr.h"
#include "nsAutoRef.h"
#include "AudioBufferUtils.h"
#include "AudioMixer.h"
#include "AudioSegment.h"
#include "SelfRef.h"
#include "mozilla/Atomics.h"
#include "mozilla/SharedThreadPool.h"
#include "mozilla/StaticPtr.h"
struct cubeb_stream;
template <>
class nsAutoRefTraits<cubeb_stream> : public nsPointerRefTraits<cubeb_stream>
{
public:
static void Release(cubeb_stream* aStream) { cubeb_stream_destroy(aStream); }
};
namespace mozilla {
/**
* Assume we can run an iteration of the MediaStreamGraph loop in this much time
* or less.
* We try to run the control loop at this rate.
*/
static const int MEDIA_GRAPH_TARGET_PERIOD_MS = 10;
/**
* Assume that we might miss our scheduled wakeup of the MediaStreamGraph by
* this much.
*/
static const int SCHEDULE_SAFETY_MARGIN_MS = 10;
/**
* Try have this much audio buffered in streams and queued to the hardware.
* The maximum delay to the end of the next control loop
* is 2*MEDIA_GRAPH_TARGET_PERIOD_MS + SCHEDULE_SAFETY_MARGIN_MS.
* There is no point in buffering more audio than this in a stream at any
* given time (until we add processing).
* This is not optimal yet.
*/
static const int AUDIO_TARGET_MS = 2*MEDIA_GRAPH_TARGET_PERIOD_MS +
SCHEDULE_SAFETY_MARGIN_MS;
class MediaStreamGraphImpl;
class AudioCallbackDriver;
class OfflineClockDriver;
/**
* A driver is responsible for the scheduling of the processing, the thread
* management, and give the different clocks to a MediaStreamGraph. This is an
* abstract base class. A MediaStreamGraph can be driven by an
* OfflineClockDriver, if the graph is offline, or a SystemClockDriver, if the
* graph is real time.
* A MediaStreamGraph holds an owning reference to its driver.
*
* The lifetime of drivers is a complicated affair. Here are the different
* scenarii that can happen:
*
* Starting a MediaStreamGraph with an AudioCallbackDriver
* - A new thread T is created, from the main thread.
* - On this thread T, cubeb is initialized if needed, and a cubeb_stream is
* created and started
* - The thread T posts a message to the main thread to terminate itself.
* - The graph runs off the audio thread
*
* Starting a MediaStreamGraph with a SystemClockDriver:
* - A new thread T is created from the main thread.
* - The graph runs off this thread.
*
* Switching from a SystemClockDriver to an AudioCallbackDriver:
* - A new AudioCallabackDriver is created and initialized on the graph thread
* - At the end of the MSG iteration, the SystemClockDriver transfers its timing
* info and a reference to itself to the AudioCallbackDriver. It then starts
* the AudioCallbackDriver.
* - When the AudioCallbackDriver starts, it checks if it has been switched from
* a SystemClockDriver, and if that is the case, sends a message to the main
* thread to shut the SystemClockDriver thread down.
* - The graph now runs off an audio callback
*
* Switching from an AudioCallbackDriver to a SystemClockDriver:
* - A new SystemClockDriver is created, and set as mNextDriver.
* - At the end of the MSG iteration, the AudioCallbackDriver transfers its
* timing info and a reference to itself to the SystemClockDriver. A new
* SystemClockDriver is started from the current audio thread.
* - When starting, the SystemClockDriver checks if it has been switched from an
* AudioCallbackDriver. If yes, it creates a new temporary thread to release
* the cubeb_streams. This temporary thread closes the cubeb_stream, and
* then dispatches a message to the main thread to be terminated.
* - The graph now runs off a normal thread.
*
* Two drivers cannot run at the same time for the same graph. The thread safety
* of the different attributes of drivers, and they access pattern is documented
* next to the members themselves.
*
*/
class GraphDriver
{
public:
explicit GraphDriver(MediaStreamGraphImpl* aGraphImpl);
NS_INLINE_DECL_THREADSAFE_REFCOUNTING(GraphDriver);
/* For real-time graphs, this waits until it's time to process more data. For
* offline graphs, this is a no-op. */
virtual void WaitForNextIteration() = 0;
/* Wakes up the graph if it is waiting. */
virtual void WakeUp() = 0;
virtual void Destroy() {}
/* Start the graph, init the driver, start the thread. */
virtual void Start() = 0;
/* Stop the graph, shutting down the thread. */
virtual void Stop() = 0;
/* Resume after a stop */
virtual void Resume() = 0;
/* Revive this driver, as more messages just arrived. */
virtual void Revive() = 0;
/* Remove Mixer callbacks when switching */
virtual void RemoveCallback() = 0;
void Shutdown();
/* Rate at which the GraphDriver runs, in ms. This can either be user
* controlled (because we are using a {System,Offline}ClockDriver, and decide
* how often we want to wakeup/how much we want to process per iteration), or
* it can be indirectly set by the latency of the audio backend, and the
* number of buffers of this audio backend: say we have four buffers, and 40ms
* latency, we will get a callback approximately every 10ms. */
virtual uint32_t IterationDuration() = 0;
/* Return whether we are switching or not. */
bool Switching();
// Those are simply or setting the associated pointer, but assert that the
// lock is held.
GraphDriver* NextDriver();
GraphDriver* PreviousDriver();
void SetNextDriver(GraphDriver* aNextDriver);
void SetPreviousDriver(GraphDriver* aPreviousDriver);
/**
* If we are running a real time graph, get the current time stamp to schedule
* video frames. This has to be reimplemented by real time drivers.
*/
virtual TimeStamp GetCurrentTimeStamp() {
return mCurrentTimeStamp;
}
GraphTime IterationEnd() {
return mIterationEnd;
}
virtual AudioCallbackDriver* AsAudioCallbackDriver() {
return nullptr;
}
virtual OfflineClockDriver* AsOfflineClockDriver() {
return nullptr;
}
/**
* Tell the driver it has to stop and return the current time of the graph, so
* another driver can start from the right point in time.
*/
virtual void SwitchAtNextIteration(GraphDriver* aDriver);
/**
* Set the time for a graph, on a driver. This is used so a new driver just
* created can start at the right point in time.
*/
void SetGraphTime(GraphDriver* aPreviousDriver,
GraphTime aLastSwitchNextIterationStart,
GraphTime aLastSwitchNextIterationEnd);
/**
* Call this to indicate that another iteration of the control loop is
* required on its regular schedule. The monitor must not be held.
* This function has to be idempotent.
*/
void EnsureNextIteration();
/**
* Same thing, but not locked.
*/
void EnsureNextIterationLocked();
MediaStreamGraphImpl* GraphImpl() {
return mGraphImpl;
}
virtual bool OnThread() = 0;
// These are invoked on the MSG thread (or MainThread in shutdown)
virtual void SetInputListener(AudioDataListener *aListener) {
mAudioInput = aListener;
}
// XXX do we need the param? probably no
virtual void RemoveInputListener(AudioDataListener *aListener) {
mAudioInput = nullptr;
}
protected:
GraphTime StateComputedTime() const;
// Time of the start of this graph iteration. This must be accessed while
// having the monitor.
GraphTime mIterationStart;
// Time of the end of this graph iteration. This must be accessed while having
// the monitor.
GraphTime mIterationEnd;
// The MediaStreamGraphImpl that owns this driver. This has a lifetime longer
// than the driver, and will never be null. Hence, it can be accesed without
// monitor.
MediaStreamGraphImpl* mGraphImpl;
// This enum specifies the wait state of the driver.
enum WaitState {
// RunThread() is running normally
WAITSTATE_RUNNING,
// RunThread() is paused waiting for its next iteration, which will
// happen soon
WAITSTATE_WAITING_FOR_NEXT_ITERATION,
// RunThread() is paused indefinitely waiting for something to change
WAITSTATE_WAITING_INDEFINITELY,
// Something has signaled RunThread() to wake up immediately,
// but it hasn't done so yet
WAITSTATE_WAKING_UP
};
// This must be access with the monitor.
WaitState mWaitState;
// Callback for mic data, if any
AudioDataListener *mAudioInput;
// This is used on the main thread (during initialization), and the graph
// thread. No monitor needed because we know the graph thread does not run
// during the initialization.
TimeStamp mCurrentTimeStamp;
// This is non-null only when this driver has recently switched from an other
// driver, and has not cleaned it up yet (for example because the audio stream
// is currently calling the callback during initialization).
//
// This is written to when changing driver, from the previous driver's thread,
// or a thread created for the occasion. This is read each time we need to
// check whether we're changing driver (in Switching()), from the graph
// thread.
// This must be accessed using the {Set,Get}PreviousDriver methods.
Bug 1207245 - part 6 - rename nsRefPtr<T> to RefPtr<T>; r=ehsan; a=Tomcat The bulk of this commit was generated with a script, executed at the top level of a typical source code checkout. The only non-machine-generated part was modifying MFBT's moz.build to reflect the new naming. CLOSED TREE makes big refactorings like this a piece of cake. # The main substitution. find . -name '*.cpp' -o -name '*.cc' -o -name '*.h' -o -name '*.mm' -o -name '*.idl'| \ xargs perl -p -i -e ' s/nsRefPtr\.h/RefPtr\.h/g; # handle includes s/nsRefPtr ?</RefPtr</g; # handle declarations and variables ' # Handle a special friend declaration in gfx/layers/AtomicRefCountedWithFinalize.h. perl -p -i -e 's/::nsRefPtr;/::RefPtr;/' gfx/layers/AtomicRefCountedWithFinalize.h # Handle nsRefPtr.h itself, a couple places that define constructors # from nsRefPtr, and code generators specially. We do this here, rather # than indiscriminantly s/nsRefPtr/RefPtr/, because that would rename # things like nsRefPtrHashtable. perl -p -i -e 's/nsRefPtr/RefPtr/g' \ mfbt/nsRefPtr.h \ xpcom/glue/nsCOMPtr.h \ xpcom/base/OwningNonNull.h \ ipc/ipdl/ipdl/lower.py \ ipc/ipdl/ipdl/builtin.py \ dom/bindings/Codegen.py \ python/lldbutils/lldbutils/utils.py # In our indiscriminate substitution above, we renamed # nsRefPtrGetterAddRefs, the class behind getter_AddRefs. Fix that up. find . -name '*.cpp' -o -name '*.h' -o -name '*.idl' | \ xargs perl -p -i -e 's/nsRefPtrGetterAddRefs/RefPtrGetterAddRefs/g' if [ -d .git ]; then git mv mfbt/nsRefPtr.h mfbt/RefPtr.h else hg mv mfbt/nsRefPtr.h mfbt/RefPtr.h fi --HG-- rename : mfbt/nsRefPtr.h => mfbt/RefPtr.h
2015-10-18 08:24:48 +03:00
RefPtr<GraphDriver> mPreviousDriver;
// This is non-null only when this driver is going to switch to an other
// driver at the end of this iteration.
// This must be accessed using the {Set,Get}NextDriver methods.
Bug 1207245 - part 6 - rename nsRefPtr<T> to RefPtr<T>; r=ehsan; a=Tomcat The bulk of this commit was generated with a script, executed at the top level of a typical source code checkout. The only non-machine-generated part was modifying MFBT's moz.build to reflect the new naming. CLOSED TREE makes big refactorings like this a piece of cake. # The main substitution. find . -name '*.cpp' -o -name '*.cc' -o -name '*.h' -o -name '*.mm' -o -name '*.idl'| \ xargs perl -p -i -e ' s/nsRefPtr\.h/RefPtr\.h/g; # handle includes s/nsRefPtr ?</RefPtr</g; # handle declarations and variables ' # Handle a special friend declaration in gfx/layers/AtomicRefCountedWithFinalize.h. perl -p -i -e 's/::nsRefPtr;/::RefPtr;/' gfx/layers/AtomicRefCountedWithFinalize.h # Handle nsRefPtr.h itself, a couple places that define constructors # from nsRefPtr, and code generators specially. We do this here, rather # than indiscriminantly s/nsRefPtr/RefPtr/, because that would rename # things like nsRefPtrHashtable. perl -p -i -e 's/nsRefPtr/RefPtr/g' \ mfbt/nsRefPtr.h \ xpcom/glue/nsCOMPtr.h \ xpcom/base/OwningNonNull.h \ ipc/ipdl/ipdl/lower.py \ ipc/ipdl/ipdl/builtin.py \ dom/bindings/Codegen.py \ python/lldbutils/lldbutils/utils.py # In our indiscriminate substitution above, we renamed # nsRefPtrGetterAddRefs, the class behind getter_AddRefs. Fix that up. find . -name '*.cpp' -o -name '*.h' -o -name '*.idl' | \ xargs perl -p -i -e 's/nsRefPtrGetterAddRefs/RefPtrGetterAddRefs/g' if [ -d .git ]; then git mv mfbt/nsRefPtr.h mfbt/RefPtr.h else hg mv mfbt/nsRefPtr.h mfbt/RefPtr.h fi --HG-- rename : mfbt/nsRefPtr.h => mfbt/RefPtr.h
2015-10-18 08:24:48 +03:00
RefPtr<GraphDriver> mNextDriver;
virtual ~GraphDriver()
{ }
};
class MediaStreamGraphInitThreadRunnable;
/**
* This class is a driver that manages its own thread.
*/
class ThreadedDriver : public GraphDriver
{
public:
explicit ThreadedDriver(MediaStreamGraphImpl* aGraphImpl);
virtual ~ThreadedDriver();
void Start() override;
void Stop() override;
void Resume() override;
void Revive() override;
void RemoveCallback() override;
/**
* Runs main control loop on the graph thread. Normally a single invocation
* of this runs for the entire lifetime of the graph thread.
*/
void RunThread();
friend class MediaStreamGraphInitThreadRunnable;
uint32_t IterationDuration() override {
return MEDIA_GRAPH_TARGET_PERIOD_MS;
}
bool OnThread() override { return !mThread || NS_GetCurrentThread() == mThread; }
/* When the graph wakes up to do an iteration, implementations return the
* range of time that will be processed. This is called only once per
* iteration; it may determine the interval from state in a previous
* call. */
virtual MediaTime GetIntervalForIteration() = 0;
protected:
nsCOMPtr<nsIThread> mThread;
};
/**
* A SystemClockDriver drives a MediaStreamGraph using a system clock, and waits
* using a monitor, between each iteration.
*/
class SystemClockDriver : public ThreadedDriver
{
public:
explicit SystemClockDriver(MediaStreamGraphImpl* aGraphImpl);
virtual ~SystemClockDriver();
MediaTime GetIntervalForIteration() override;
void WaitForNextIteration() override;
void WakeUp() override;
private:
// Those are only modified (after initialization) on the graph thread. The
// graph thread does not run during the initialization.
TimeStamp mInitialTimeStamp;
TimeStamp mLastTimeStamp;
};
/**
* An OfflineClockDriver runs the graph as fast as possible, without waiting
* between iteration.
*/
class OfflineClockDriver : public ThreadedDriver
{
public:
OfflineClockDriver(MediaStreamGraphImpl* aGraphImpl, GraphTime aSlice);
virtual ~OfflineClockDriver();
MediaTime GetIntervalForIteration() override;
void WaitForNextIteration() override;
void WakeUp() override;
TimeStamp GetCurrentTimeStamp() override;
OfflineClockDriver* AsOfflineClockDriver() override {
return this;
}
private:
// Time, in GraphTime, for each iteration
GraphTime mSlice;
};
Bug 1094764 - Implement AudioContext.suspend and friends. r=roc,ehsan - Relevant spec text: - http://webaudio.github.io/web-audio-api/#widl-AudioContext-suspend-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-resume-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-close-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-state - http://webaudio.github.io/web-audio-api/#widl-AudioContext-onstatechange - In a couple words, the behavior we want: - Closed context cannot have new nodes created, but can do decodeAudioData, and create buffers, and such. - OfflineAudioContexts don't support those methods, transitions happen at startRendering and at the end of processing. onstatechange is used to make this observable. - (regular) AudioContexts support those methods. The promises and onstatechange should be resolved/called when the operation has actually completed on the rendering thread. Once a context has been closed, it cannot transition back to "running". An AudioContext switches to "running" when the audio callback start running, this allow authors to know how long the audio stack takes to start running. - MediaStreams that feed in/go out of a suspended graph should respectively not buffer at the graph input, and output silence - suspended context should not be doing much on the CPU, and we should try to pause audio streams if we can (this behaviour is the main reason we need this in the first place, for saving battery on mobile, and CPU on all platforms) - Now, the implementation: - AudioNodeStreams are now tagged with a context id, to be able to operate on all the streams of a given AudioContext on the Graph thread without having to go and lock everytime to touch the AudioContext. This happens in the AudioNodeStream ctor. IDs are of course constant for the lifetime of the node. - When an AudioContext goes into suspended mode, streams for this AudioContext are moved out of the mStreams array to a second array, mSuspendedStreams. Streams in mSuspendedStream are not ordered, and are not processed. - The MSG will automatically switch to a SystemClockDriver when it finds that there are no more AudioNodeStream/Stream with an audio track. This is how pausing the audio subsystem and saving battery works. Subsequently, when the MSG finds that there are only streams in mSuspendedStreams, it will go to sleep (block on a monitor), so we save CPU, but it does not shut itself down. This is mostly not a new behaviour (this is what the MSG does since the refactoring), but is important to note. - Promises are gripped (addref-ed) on the main thread, and then shepherd down other threads and to the GraphDriver, if needed (sometimes we can resolve them right away). They move between threads as void* to prevent calling methods on them, as they are not thread safe. Then, the driver executes the operation, and when it's done (initializing and closing audio streams can take some time), we send the promise back to the main thread, and resolve it, casting back to Promise* after asserting we're back on the main thread. This way, we can send them back on the main thread once an operation has complete (suspending an audio stream, starting it again on resume(), etc.), without having to do bookkeeping between suspend calls and their result. Promises are not thread safe, so we can't move them around AddRef-ed. - The stream destruction logic now takes into account that a stream can be destroyed while not being in mStreams. - A graph can now switch GraphDriver twice or more per iteration, for example if an author goes suspend()/resume()/suspend() in the same script. - Some operation have to be done on suspended stream, so we now use double for-loop around mSuspendedStreams and mStreams in some places in MediaStreamGraph.cpp. - A tricky part was making sure everything worked at AudioContext boundaries. TrackUnionStream that have one of their input stream suspended append null ticks instead. - The graph ordering algorithm had to be altered to not include suspended streams. - There are some edge cases (adding a stream on a suspended graph, calling suspend/resume when a graph has just been close()d).
2015-02-27 20:22:05 +03:00
struct StreamAndPromiseForOperation
{
StreamAndPromiseForOperation(MediaStream* aStream,
void* aPromise,
dom::AudioContextOperation aOperation);
Bug 1207245 - part 6 - rename nsRefPtr<T> to RefPtr<T>; r=ehsan; a=Tomcat The bulk of this commit was generated with a script, executed at the top level of a typical source code checkout. The only non-machine-generated part was modifying MFBT's moz.build to reflect the new naming. CLOSED TREE makes big refactorings like this a piece of cake. # The main substitution. find . -name '*.cpp' -o -name '*.cc' -o -name '*.h' -o -name '*.mm' -o -name '*.idl'| \ xargs perl -p -i -e ' s/nsRefPtr\.h/RefPtr\.h/g; # handle includes s/nsRefPtr ?</RefPtr</g; # handle declarations and variables ' # Handle a special friend declaration in gfx/layers/AtomicRefCountedWithFinalize.h. perl -p -i -e 's/::nsRefPtr;/::RefPtr;/' gfx/layers/AtomicRefCountedWithFinalize.h # Handle nsRefPtr.h itself, a couple places that define constructors # from nsRefPtr, and code generators specially. We do this here, rather # than indiscriminantly s/nsRefPtr/RefPtr/, because that would rename # things like nsRefPtrHashtable. perl -p -i -e 's/nsRefPtr/RefPtr/g' \ mfbt/nsRefPtr.h \ xpcom/glue/nsCOMPtr.h \ xpcom/base/OwningNonNull.h \ ipc/ipdl/ipdl/lower.py \ ipc/ipdl/ipdl/builtin.py \ dom/bindings/Codegen.py \ python/lldbutils/lldbutils/utils.py # In our indiscriminate substitution above, we renamed # nsRefPtrGetterAddRefs, the class behind getter_AddRefs. Fix that up. find . -name '*.cpp' -o -name '*.h' -o -name '*.idl' | \ xargs perl -p -i -e 's/nsRefPtrGetterAddRefs/RefPtrGetterAddRefs/g' if [ -d .git ]; then git mv mfbt/nsRefPtr.h mfbt/RefPtr.h else hg mv mfbt/nsRefPtr.h mfbt/RefPtr.h fi --HG-- rename : mfbt/nsRefPtr.h => mfbt/RefPtr.h
2015-10-18 08:24:48 +03:00
RefPtr<MediaStream> mStream;
Bug 1094764 - Implement AudioContext.suspend and friends. r=roc,ehsan - Relevant spec text: - http://webaudio.github.io/web-audio-api/#widl-AudioContext-suspend-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-resume-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-close-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-state - http://webaudio.github.io/web-audio-api/#widl-AudioContext-onstatechange - In a couple words, the behavior we want: - Closed context cannot have new nodes created, but can do decodeAudioData, and create buffers, and such. - OfflineAudioContexts don't support those methods, transitions happen at startRendering and at the end of processing. onstatechange is used to make this observable. - (regular) AudioContexts support those methods. The promises and onstatechange should be resolved/called when the operation has actually completed on the rendering thread. Once a context has been closed, it cannot transition back to "running". An AudioContext switches to "running" when the audio callback start running, this allow authors to know how long the audio stack takes to start running. - MediaStreams that feed in/go out of a suspended graph should respectively not buffer at the graph input, and output silence - suspended context should not be doing much on the CPU, and we should try to pause audio streams if we can (this behaviour is the main reason we need this in the first place, for saving battery on mobile, and CPU on all platforms) - Now, the implementation: - AudioNodeStreams are now tagged with a context id, to be able to operate on all the streams of a given AudioContext on the Graph thread without having to go and lock everytime to touch the AudioContext. This happens in the AudioNodeStream ctor. IDs are of course constant for the lifetime of the node. - When an AudioContext goes into suspended mode, streams for this AudioContext are moved out of the mStreams array to a second array, mSuspendedStreams. Streams in mSuspendedStream are not ordered, and are not processed. - The MSG will automatically switch to a SystemClockDriver when it finds that there are no more AudioNodeStream/Stream with an audio track. This is how pausing the audio subsystem and saving battery works. Subsequently, when the MSG finds that there are only streams in mSuspendedStreams, it will go to sleep (block on a monitor), so we save CPU, but it does not shut itself down. This is mostly not a new behaviour (this is what the MSG does since the refactoring), but is important to note. - Promises are gripped (addref-ed) on the main thread, and then shepherd down other threads and to the GraphDriver, if needed (sometimes we can resolve them right away). They move between threads as void* to prevent calling methods on them, as they are not thread safe. Then, the driver executes the operation, and when it's done (initializing and closing audio streams can take some time), we send the promise back to the main thread, and resolve it, casting back to Promise* after asserting we're back on the main thread. This way, we can send them back on the main thread once an operation has complete (suspending an audio stream, starting it again on resume(), etc.), without having to do bookkeeping between suspend calls and their result. Promises are not thread safe, so we can't move them around AddRef-ed. - The stream destruction logic now takes into account that a stream can be destroyed while not being in mStreams. - A graph can now switch GraphDriver twice or more per iteration, for example if an author goes suspend()/resume()/suspend() in the same script. - Some operation have to be done on suspended stream, so we now use double for-loop around mSuspendedStreams and mStreams in some places in MediaStreamGraph.cpp. - A tricky part was making sure everything worked at AudioContext boundaries. TrackUnionStream that have one of their input stream suspended append null ticks instead. - The graph ordering algorithm had to be altered to not include suspended streams. - There are some edge cases (adding a stream on a suspended graph, calling suspend/resume when a graph has just been close()d).
2015-02-27 20:22:05 +03:00
void* mPromise;
dom::AudioContextOperation mOperation;
};
enum AsyncCubebOperation {
INIT,
SHUTDOWN
};
/**
* This is a graph driver that is based on callback functions called by the
* audio api. This ensures minimal audio latency, because it means there is no
* buffering happening: the audio is generated inside the callback.
*
* This design is less flexible than running our own thread:
* - We have no control over the thread:
* - It cannot block, and it has to run for a shorter amount of time than the
* buffer it is going to fill, or an under-run is going to occur (short burst
* of silence in the final audio output).
* - We can't know for sure when the callback function is going to be called
* (although we compute an estimation so we can schedule video frames)
* - Creating and shutting the thread down is a blocking operation, that can
* take _seconds_ in some cases (because IPC has to be set up, and
* sometimes hardware components are involved and need to be warmed up)
* - We have no control on how much audio we generate, we have to return exactly
* the number of frames asked for by the callback. Since for the Web Audio
* API, we have to do block processing at 128 frames per block, we need to
* keep a little spill buffer to store the extra frames.
*/
class AudioCallbackDriver : public GraphDriver,
public MixerCallbackReceiver
{
public:
explicit AudioCallbackDriver(MediaStreamGraphImpl* aGraphImpl);
virtual ~AudioCallbackDriver();
void Destroy() override;
void Start() override;
void Stop() override;
void Resume() override;
void Revive() override;
void RemoveCallback() override;
void WaitForNextIteration() override;
void WakeUp() override;
/* Static wrapper function cubeb calls back. */
static long DataCallback_s(cubeb_stream * aStream,
void * aUser, void * aInputBuffer, void * aOutputBuffer,
long aFrames);
static void StateCallback_s(cubeb_stream* aStream, void * aUser,
cubeb_state aState);
static void DeviceChangedCallback_s(void * aUser);
/* This function is called by the underlying audio backend when a refill is
* needed. This is what drives the whole graph when it is used to output
* audio. If the return value is exactly aFrames, this function will get
* called again. If it is less than aFrames, the stream will go in draining
* mode, and this function will not be called again. */
long DataCallback(AudioDataValue* aInputBuffer, AudioDataValue* aOutputBuffer, long aFrames);
/* This function is called by the underlying audio backend, but is only used
* for informational purposes at the moment. */
void StateCallback(cubeb_state aState);
/* This is an approximation of the number of millisecond there are between two
* iterations of the graph. */
uint32_t IterationDuration() override;
/* This function gets called when the graph has produced the audio frames for
* this iteration. */
void MixerCallback(AudioDataValue* aMixedBuffer,
AudioSampleFormat aFormat,
uint32_t aChannels,
uint32_t aFrames,
uint32_t aSampleRate) override;
AudioCallbackDriver* AsAudioCallbackDriver() override {
return this;
}
Bug 1094764 - Implement AudioContext.suspend and friends. r=roc,ehsan - Relevant spec text: - http://webaudio.github.io/web-audio-api/#widl-AudioContext-suspend-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-resume-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-close-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-state - http://webaudio.github.io/web-audio-api/#widl-AudioContext-onstatechange - In a couple words, the behavior we want: - Closed context cannot have new nodes created, but can do decodeAudioData, and create buffers, and such. - OfflineAudioContexts don't support those methods, transitions happen at startRendering and at the end of processing. onstatechange is used to make this observable. - (regular) AudioContexts support those methods. The promises and onstatechange should be resolved/called when the operation has actually completed on the rendering thread. Once a context has been closed, it cannot transition back to "running". An AudioContext switches to "running" when the audio callback start running, this allow authors to know how long the audio stack takes to start running. - MediaStreams that feed in/go out of a suspended graph should respectively not buffer at the graph input, and output silence - suspended context should not be doing much on the CPU, and we should try to pause audio streams if we can (this behaviour is the main reason we need this in the first place, for saving battery on mobile, and CPU on all platforms) - Now, the implementation: - AudioNodeStreams are now tagged with a context id, to be able to operate on all the streams of a given AudioContext on the Graph thread without having to go and lock everytime to touch the AudioContext. This happens in the AudioNodeStream ctor. IDs are of course constant for the lifetime of the node. - When an AudioContext goes into suspended mode, streams for this AudioContext are moved out of the mStreams array to a second array, mSuspendedStreams. Streams in mSuspendedStream are not ordered, and are not processed. - The MSG will automatically switch to a SystemClockDriver when it finds that there are no more AudioNodeStream/Stream with an audio track. This is how pausing the audio subsystem and saving battery works. Subsequently, when the MSG finds that there are only streams in mSuspendedStreams, it will go to sleep (block on a monitor), so we save CPU, but it does not shut itself down. This is mostly not a new behaviour (this is what the MSG does since the refactoring), but is important to note. - Promises are gripped (addref-ed) on the main thread, and then shepherd down other threads and to the GraphDriver, if needed (sometimes we can resolve them right away). They move between threads as void* to prevent calling methods on them, as they are not thread safe. Then, the driver executes the operation, and when it's done (initializing and closing audio streams can take some time), we send the promise back to the main thread, and resolve it, casting back to Promise* after asserting we're back on the main thread. This way, we can send them back on the main thread once an operation has complete (suspending an audio stream, starting it again on resume(), etc.), without having to do bookkeeping between suspend calls and their result. Promises are not thread safe, so we can't move them around AddRef-ed. - The stream destruction logic now takes into account that a stream can be destroyed while not being in mStreams. - A graph can now switch GraphDriver twice or more per iteration, for example if an author goes suspend()/resume()/suspend() in the same script. - Some operation have to be done on suspended stream, so we now use double for-loop around mSuspendedStreams and mStreams in some places in MediaStreamGraph.cpp. - A tricky part was making sure everything worked at AudioContext boundaries. TrackUnionStream that have one of their input stream suspended append null ticks instead. - The graph ordering algorithm had to be altered to not include suspended streams. - There are some edge cases (adding a stream on a suspended graph, calling suspend/resume when a graph has just been close()d).
2015-02-27 20:22:05 +03:00
/* Enqueue a promise that is going to be resolved when a specific operation
* occurs on the cubeb stream. */
void EnqueueStreamAndPromiseForOperation(MediaStream* aStream,
void* aPromise,
dom::AudioContextOperation aOperation);
bool IsSwitchingDevice() {
#ifdef XP_MACOSX
return mSelfReference;
#else
return false;
#endif
}
/**
* Whether the audio callback is processing. This is for asserting only.
*/
bool InCallback();
bool OnThread() override { return !mStarted || InCallback(); }
/* Whether the underlying cubeb stream has been started. See comment for
* mStarted for details. */
bool IsStarted();
/* Tell the driver whether this process is using a microphone or not. This is
* thread safe. */
void SetMicrophoneActive(bool aActive);
Bug 1094764 - Implement AudioContext.suspend and friends. r=roc,ehsan - Relevant spec text: - http://webaudio.github.io/web-audio-api/#widl-AudioContext-suspend-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-resume-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-close-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-state - http://webaudio.github.io/web-audio-api/#widl-AudioContext-onstatechange - In a couple words, the behavior we want: - Closed context cannot have new nodes created, but can do decodeAudioData, and create buffers, and such. - OfflineAudioContexts don't support those methods, transitions happen at startRendering and at the end of processing. onstatechange is used to make this observable. - (regular) AudioContexts support those methods. The promises and onstatechange should be resolved/called when the operation has actually completed on the rendering thread. Once a context has been closed, it cannot transition back to "running". An AudioContext switches to "running" when the audio callback start running, this allow authors to know how long the audio stack takes to start running. - MediaStreams that feed in/go out of a suspended graph should respectively not buffer at the graph input, and output silence - suspended context should not be doing much on the CPU, and we should try to pause audio streams if we can (this behaviour is the main reason we need this in the first place, for saving battery on mobile, and CPU on all platforms) - Now, the implementation: - AudioNodeStreams are now tagged with a context id, to be able to operate on all the streams of a given AudioContext on the Graph thread without having to go and lock everytime to touch the AudioContext. This happens in the AudioNodeStream ctor. IDs are of course constant for the lifetime of the node. - When an AudioContext goes into suspended mode, streams for this AudioContext are moved out of the mStreams array to a second array, mSuspendedStreams. Streams in mSuspendedStream are not ordered, and are not processed. - The MSG will automatically switch to a SystemClockDriver when it finds that there are no more AudioNodeStream/Stream with an audio track. This is how pausing the audio subsystem and saving battery works. Subsequently, when the MSG finds that there are only streams in mSuspendedStreams, it will go to sleep (block on a monitor), so we save CPU, but it does not shut itself down. This is mostly not a new behaviour (this is what the MSG does since the refactoring), but is important to note. - Promises are gripped (addref-ed) on the main thread, and then shepherd down other threads and to the GraphDriver, if needed (sometimes we can resolve them right away). They move between threads as void* to prevent calling methods on them, as they are not thread safe. Then, the driver executes the operation, and when it's done (initializing and closing audio streams can take some time), we send the promise back to the main thread, and resolve it, casting back to Promise* after asserting we're back on the main thread. This way, we can send them back on the main thread once an operation has complete (suspending an audio stream, starting it again on resume(), etc.), without having to do bookkeeping between suspend calls and their result. Promises are not thread safe, so we can't move them around AddRef-ed. - The stream destruction logic now takes into account that a stream can be destroyed while not being in mStreams. - A graph can now switch GraphDriver twice or more per iteration, for example if an author goes suspend()/resume()/suspend() in the same script. - Some operation have to be done on suspended stream, so we now use double for-loop around mSuspendedStreams and mStreams in some places in MediaStreamGraph.cpp. - A tricky part was making sure everything worked at AudioContext boundaries. TrackUnionStream that have one of their input stream suspended append null ticks instead. - The graph ordering algorithm had to be altered to not include suspended streams. - There are some edge cases (adding a stream on a suspended graph, calling suspend/resume when a graph has just been close()d).
2015-02-27 20:22:05 +03:00
void CompleteAudioContextOperations(AsyncCubebOperation aOperation);
private:
/**
* On certain MacBookPro, the microphone is located near the left speaker.
* We need to pan the sound output to the right speaker if we are using the
* mic and the built-in speaker, or we will have terrible echo. */
void PanOutputIfNeeded(bool aMicrophoneActive);
/**
* This is called when the output device used by the cubeb stream changes. */
void DeviceChangedCallback();
/* Start the cubeb stream */
void StartStream();
friend class AsyncCubebTask;
void Init();
/* MediaStreamGraphs are always down/up mixed to stereo for now. */
static const uint32_t ChannelCount = 2;
/* The size of this buffer comes from the fact that some audio backends can
* call back with a number of frames lower than one block (128 frames), so we
* need to keep at most two block in the SpillBuffer, because we always round
* up to block boundaries during an iteration.
* This is only ever accessed on the audio callback thread. */
SpillBuffer<AudioDataValue, WEBAUDIO_BLOCK_SIZE * 2, ChannelCount> mScratchBuffer;
/* Wrapper to ensure we write exactly the number of frames we need in the
* audio buffer cubeb passes us. This is only ever accessed on the audio
* callback thread. */
AudioCallbackBufferWrapper<AudioDataValue, ChannelCount> mBuffer;
/* cubeb stream for this graph. This is guaranteed to be non-null after Init()
* has been called, and is synchronized internaly. */
nsAutoRef<cubeb_stream> mAudioStream;
/* The sample rate for the aforementionned cubeb stream. This is set on
* initialization and can be read safely afterwards. */
uint32_t mSampleRate;
/* Approximation of the time between two callbacks. This is used to schedule
* video frames. This is in milliseconds. Only even used (after
* inizatialization) on the audio callback thread. */
uint32_t mIterationDurationMS;
/* cubeb_stream_init calls the audio callback to prefill the buffers. The
* previous driver has to be kept alive until the audio stream has been
* started, because it is responsible to call cubeb_stream_start, so we delay
* the cleanup of the previous driver until it has started the audio stream.
* Otherwise, there is a race where we kill the previous driver thread
* between cubeb_stream_init and cubeb_stream_start,
* and callbacks after the prefill never get called.
* This is written on the previous driver's thread (if switching) or main
* thread (if this driver is the first one).
* This is read on previous driver's thread (during callbacks from
* cubeb_stream_init) and the audio thread (when switching away from this
* driver back to a SystemClockDriver).
* This is synchronized by the Graph's monitor.
* */
bool mStarted;
/* Listener for mic input, if any. */
RefPtr<AudioDataListener> mAudioInput;
struct AutoInCallback
{
explicit AutoInCallback(AudioCallbackDriver* aDriver);
~AutoInCallback();
AudioCallbackDriver* mDriver;
};
/* Thread for off-main-thread initialization and
* shutdown of the audio stream. */
nsCOMPtr<nsIThread> mInitShutdownThread;
/* This must be accessed with the graph monitor held. */
Bug 1094764 - Implement AudioContext.suspend and friends. r=roc,ehsan - Relevant spec text: - http://webaudio.github.io/web-audio-api/#widl-AudioContext-suspend-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-resume-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-close-Promise - http://webaudio.github.io/web-audio-api/#widl-AudioContext-state - http://webaudio.github.io/web-audio-api/#widl-AudioContext-onstatechange - In a couple words, the behavior we want: - Closed context cannot have new nodes created, but can do decodeAudioData, and create buffers, and such. - OfflineAudioContexts don't support those methods, transitions happen at startRendering and at the end of processing. onstatechange is used to make this observable. - (regular) AudioContexts support those methods. The promises and onstatechange should be resolved/called when the operation has actually completed on the rendering thread. Once a context has been closed, it cannot transition back to "running". An AudioContext switches to "running" when the audio callback start running, this allow authors to know how long the audio stack takes to start running. - MediaStreams that feed in/go out of a suspended graph should respectively not buffer at the graph input, and output silence - suspended context should not be doing much on the CPU, and we should try to pause audio streams if we can (this behaviour is the main reason we need this in the first place, for saving battery on mobile, and CPU on all platforms) - Now, the implementation: - AudioNodeStreams are now tagged with a context id, to be able to operate on all the streams of a given AudioContext on the Graph thread without having to go and lock everytime to touch the AudioContext. This happens in the AudioNodeStream ctor. IDs are of course constant for the lifetime of the node. - When an AudioContext goes into suspended mode, streams for this AudioContext are moved out of the mStreams array to a second array, mSuspendedStreams. Streams in mSuspendedStream are not ordered, and are not processed. - The MSG will automatically switch to a SystemClockDriver when it finds that there are no more AudioNodeStream/Stream with an audio track. This is how pausing the audio subsystem and saving battery works. Subsequently, when the MSG finds that there are only streams in mSuspendedStreams, it will go to sleep (block on a monitor), so we save CPU, but it does not shut itself down. This is mostly not a new behaviour (this is what the MSG does since the refactoring), but is important to note. - Promises are gripped (addref-ed) on the main thread, and then shepherd down other threads and to the GraphDriver, if needed (sometimes we can resolve them right away). They move between threads as void* to prevent calling methods on them, as they are not thread safe. Then, the driver executes the operation, and when it's done (initializing and closing audio streams can take some time), we send the promise back to the main thread, and resolve it, casting back to Promise* after asserting we're back on the main thread. This way, we can send them back on the main thread once an operation has complete (suspending an audio stream, starting it again on resume(), etc.), without having to do bookkeeping between suspend calls and their result. Promises are not thread safe, so we can't move them around AddRef-ed. - The stream destruction logic now takes into account that a stream can be destroyed while not being in mStreams. - A graph can now switch GraphDriver twice or more per iteration, for example if an author goes suspend()/resume()/suspend() in the same script. - Some operation have to be done on suspended stream, so we now use double for-loop around mSuspendedStreams and mStreams in some places in MediaStreamGraph.cpp. - A tricky part was making sure everything worked at AudioContext boundaries. TrackUnionStream that have one of their input stream suspended append null ticks instead. - The graph ordering algorithm had to be altered to not include suspended streams. - There are some edge cases (adding a stream on a suspended graph, calling suspend/resume when a graph has just been close()d).
2015-02-27 20:22:05 +03:00
nsAutoTArray<StreamAndPromiseForOperation, 1> mPromisesForOperation;
/* This is set during initialization, and can be read safely afterwards. */
dom::AudioChannel mAudioChannel;
/* Used to queue us to add the mixer callback on first run. */
bool mAddedMixer;
/* This is atomic and is set by the audio callback thread. It can be read by
* any thread safely. */
Atomic<bool> mInCallback;
/**
* True if microphone is being used by this process. This is synchronized by
* the graph's monitor. */
bool mMicrophoneActive;
#ifdef XP_MACOSX
/* Implements the workaround for the osx audio stack when changing output
* devices. See comments in .cpp */
bool OSXDeviceSwitchingWorkaround();
/* Self-reference that keep this driver alive when switching output audio
* device and making the graph running temporarily off a SystemClockDriver. */
SelfReference<AudioCallbackDriver> mSelfReference;
/* While switching devices, we keep track of the number of callbacks received,
* since OSX seems to still call us _sometimes_. */
uint32_t mCallbackReceivedWhileSwitching;
#endif
};
class AsyncCubebTask : public nsRunnable
{
public:
AsyncCubebTask(AudioCallbackDriver* aDriver, AsyncCubebOperation aOperation);
nsresult Dispatch()
{
nsresult rv = EnsureThread();
if (!NS_FAILED(rv)) {
rv = sThreadPool->Dispatch(this, NS_DISPATCH_NORMAL);
}
return rv;
}
protected:
virtual ~AsyncCubebTask();
private:
static nsresult EnsureThread();
NS_IMETHOD Run() override final;
static StaticRefPtr<nsIThreadPool> sThreadPool;
Bug 1207245 - part 6 - rename nsRefPtr<T> to RefPtr<T>; r=ehsan; a=Tomcat The bulk of this commit was generated with a script, executed at the top level of a typical source code checkout. The only non-machine-generated part was modifying MFBT's moz.build to reflect the new naming. CLOSED TREE makes big refactorings like this a piece of cake. # The main substitution. find . -name '*.cpp' -o -name '*.cc' -o -name '*.h' -o -name '*.mm' -o -name '*.idl'| \ xargs perl -p -i -e ' s/nsRefPtr\.h/RefPtr\.h/g; # handle includes s/nsRefPtr ?</RefPtr</g; # handle declarations and variables ' # Handle a special friend declaration in gfx/layers/AtomicRefCountedWithFinalize.h. perl -p -i -e 's/::nsRefPtr;/::RefPtr;/' gfx/layers/AtomicRefCountedWithFinalize.h # Handle nsRefPtr.h itself, a couple places that define constructors # from nsRefPtr, and code generators specially. We do this here, rather # than indiscriminantly s/nsRefPtr/RefPtr/, because that would rename # things like nsRefPtrHashtable. perl -p -i -e 's/nsRefPtr/RefPtr/g' \ mfbt/nsRefPtr.h \ xpcom/glue/nsCOMPtr.h \ xpcom/base/OwningNonNull.h \ ipc/ipdl/ipdl/lower.py \ ipc/ipdl/ipdl/builtin.py \ dom/bindings/Codegen.py \ python/lldbutils/lldbutils/utils.py # In our indiscriminate substitution above, we renamed # nsRefPtrGetterAddRefs, the class behind getter_AddRefs. Fix that up. find . -name '*.cpp' -o -name '*.h' -o -name '*.idl' | \ xargs perl -p -i -e 's/nsRefPtrGetterAddRefs/RefPtrGetterAddRefs/g' if [ -d .git ]; then git mv mfbt/nsRefPtr.h mfbt/RefPtr.h else hg mv mfbt/nsRefPtr.h mfbt/RefPtr.h fi --HG-- rename : mfbt/nsRefPtr.h => mfbt/RefPtr.h
2015-10-18 08:24:48 +03:00
RefPtr<AudioCallbackDriver> mDriver;
AsyncCubebOperation mOperation;
Bug 1207245 - part 6 - rename nsRefPtr<T> to RefPtr<T>; r=ehsan; a=Tomcat The bulk of this commit was generated with a script, executed at the top level of a typical source code checkout. The only non-machine-generated part was modifying MFBT's moz.build to reflect the new naming. CLOSED TREE makes big refactorings like this a piece of cake. # The main substitution. find . -name '*.cpp' -o -name '*.cc' -o -name '*.h' -o -name '*.mm' -o -name '*.idl'| \ xargs perl -p -i -e ' s/nsRefPtr\.h/RefPtr\.h/g; # handle includes s/nsRefPtr ?</RefPtr</g; # handle declarations and variables ' # Handle a special friend declaration in gfx/layers/AtomicRefCountedWithFinalize.h. perl -p -i -e 's/::nsRefPtr;/::RefPtr;/' gfx/layers/AtomicRefCountedWithFinalize.h # Handle nsRefPtr.h itself, a couple places that define constructors # from nsRefPtr, and code generators specially. We do this here, rather # than indiscriminantly s/nsRefPtr/RefPtr/, because that would rename # things like nsRefPtrHashtable. perl -p -i -e 's/nsRefPtr/RefPtr/g' \ mfbt/nsRefPtr.h \ xpcom/glue/nsCOMPtr.h \ xpcom/base/OwningNonNull.h \ ipc/ipdl/ipdl/lower.py \ ipc/ipdl/ipdl/builtin.py \ dom/bindings/Codegen.py \ python/lldbutils/lldbutils/utils.py # In our indiscriminate substitution above, we renamed # nsRefPtrGetterAddRefs, the class behind getter_AddRefs. Fix that up. find . -name '*.cpp' -o -name '*.h' -o -name '*.idl' | \ xargs perl -p -i -e 's/nsRefPtrGetterAddRefs/RefPtrGetterAddRefs/g' if [ -d .git ]; then git mv mfbt/nsRefPtr.h mfbt/RefPtr.h else hg mv mfbt/nsRefPtr.h mfbt/RefPtr.h fi --HG-- rename : mfbt/nsRefPtr.h => mfbt/RefPtr.h
2015-10-18 08:24:48 +03:00
RefPtr<MediaStreamGraphImpl> mShutdownGrip;
};
} // namespace mozilla
#endif // GRAPHDRIVER_H_