gecko-dev/gfx/ipc/GPUProcessManager.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1740 lines
58 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
* 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 "GPUProcessManager.h"
#include "gfxConfig.h"
#include "gfxPlatform.h"
#include "GPUProcessHost.h"
#include "GPUProcessListener.h"
#include "mozilla/AppShutdown.h"
#include "mozilla/MemoryReportingProcess.h"
#include "mozilla/Preferences.h"
#include "mozilla/Sprintf.h"
#include "mozilla/StaticPtr.h"
#include "mozilla/StaticPrefs_gfx.h"
#include "mozilla/StaticPrefs_layers.h"
#include "mozilla/StaticPrefs_media.h"
#include "mozilla/RemoteDecoderManagerChild.h"
#include "mozilla/RemoteDecoderManagerParent.h"
#include "mozilla/Telemetry.h"
#include "mozilla/dom/ContentParent.h"
#include "mozilla/gfx/gfxVars.h"
#include "mozilla/gfx/GPUChild.h"
#include "mozilla/gfx/GPUParent.h"
#include "mozilla/glean/GleanMetrics.h"
#include "mozilla/ipc/Endpoint.h"
#include "mozilla/ipc/ProcessChild.h"
#include "mozilla/layers/APZCTreeManagerChild.h"
#include "mozilla/layers/APZInputBridgeChild.h"
#include "mozilla/layers/CompositorBridgeChild.h"
#include "mozilla/layers/CompositorBridgeParent.h"
#include "mozilla/layers/CompositorManagerChild.h"
#include "mozilla/layers/CompositorManagerParent.h"
#include "mozilla/layers/CompositorOptions.h"
#include "mozilla/layers/ImageBridgeChild.h"
#include "mozilla/layers/ImageBridgeParent.h"
#include "mozilla/layers/InProcessCompositorSession.h"
#include "mozilla/layers/LayerTreeOwnerTracker.h"
#include "mozilla/layers/RemoteCompositorSession.h"
#include "mozilla/webrender/RenderThread.h"
#include "mozilla/widget/PlatformWidgetTypes.h"
#include "nsAppRunner.h"
#include "mozilla/widget/CompositorWidget.h"
#ifdef MOZ_WIDGET_SUPPORTS_OOP_COMPOSITING
# include "mozilla/widget/CompositorWidgetChild.h"
#endif
#include "nsBaseWidget.h"
#include "nsContentUtils.h"
#include "VRManagerChild.h"
#include "VRManagerParent.h"
#include "VsyncBridgeChild.h"
#include "VsyncIOThreadHolder.h"
#include "VsyncSource.h"
#include "nsExceptionHandler.h"
#include "nsPrintfCString.h"
#if defined(MOZ_WIDGET_ANDROID)
# include "mozilla/java/SurfaceControlManagerWrappers.h"
# include "mozilla/widget/AndroidUiThread.h"
# include "mozilla/layers/UiCompositorControllerChild.h"
#endif // defined(MOZ_WIDGET_ANDROID)
#if defined(XP_WIN)
# include "gfxWindowsPlatform.h"
#endif
namespace mozilla {
namespace gfx {
using namespace mozilla::layers;
static StaticAutoPtr<GPUProcessManager> sSingleton;
GPUProcessManager* GPUProcessManager::Get() { return sSingleton; }
void GPUProcessManager::Initialize() {
MOZ_ASSERT(XRE_IsParentProcess());
sSingleton = new GPUProcessManager();
}
void GPUProcessManager::Shutdown() { sSingleton = nullptr; }
GPUProcessManager::GPUProcessManager()
: mTaskFactory(this),
mNextNamespace(0),
mIdNamespace(0),
mResourceId(0),
mUnstableProcessAttempts(0),
mTotalProcessAttempts(0),
mDeviceResetCount(0),
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
mAppInForeground(true),
mProcess(nullptr),
mProcessToken(0),
mProcessStable(true),
mGPUChild(nullptr) {
MOZ_COUNT_CTOR(GPUProcessManager);
mIdNamespace = AllocateNamespace();
mDeviceResetLastTime = TimeStamp::Now();
LayerTreeOwnerTracker::Initialize();
CompositorBridgeParent::InitializeStatics();
}
GPUProcessManager::~GPUProcessManager() {
MOZ_COUNT_DTOR(GPUProcessManager);
LayerTreeOwnerTracker::Shutdown();
// The GPU process should have already been shut down.
MOZ_ASSERT(!mProcess && !mGPUChild);
// We should have already removed observers.
MOZ_ASSERT(!mObserver);
}
NS_IMPL_ISUPPORTS(GPUProcessManager::Observer, nsIObserver);
GPUProcessManager::Observer::Observer(GPUProcessManager* aManager)
: mManager(aManager) {}
NS_IMETHODIMP
GPUProcessManager::Observer::Observe(nsISupports* aSubject, const char* aTopic,
const char16_t* aData) {
if (!strcmp(aTopic, NS_XPCOM_SHUTDOWN_OBSERVER_ID)) {
mManager->OnXPCOMShutdown();
} else if (!strcmp(aTopic, "nsPref:changed")) {
mManager->OnPreferenceChange(aData);
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
} else if (!strcmp(aTopic, "application-foreground")) {
mManager->mAppInForeground = true;
if (!mManager->mProcess && gfxConfig::IsEnabled(Feature::GPU_PROCESS)) {
Unused << mManager->LaunchGPUProcess();
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
}
} else if (!strcmp(aTopic, "application-background")) {
mManager->mAppInForeground = false;
} else if (!strcmp(aTopic, "screen-information-changed")) {
mManager->ScreenInformationChanged();
}
return NS_OK;
}
GPUProcessManager::BatteryObserver::BatteryObserver(GPUProcessManager* aManager)
: mManager(aManager) {
hal::RegisterBatteryObserver(this);
}
void GPUProcessManager::BatteryObserver::Notify(
const hal::BatteryInformation& aBatteryInfo) {
mManager->NotifyBatteryInfo(aBatteryInfo);
}
void GPUProcessManager::BatteryObserver::ShutDown() {
hal::UnregisterBatteryObserver(this);
}
GPUProcessManager::BatteryObserver::~BatteryObserver() {}
void GPUProcessManager::OnXPCOMShutdown() {
if (mObserver) {
nsContentUtils::UnregisterShutdownObserver(mObserver);
Preferences::RemoveObserver(mObserver, "");
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
nsCOMPtr<nsIObserverService> obsServ = services::GetObserverService();
if (obsServ) {
obsServ->RemoveObserver(mObserver, "application-foreground");
obsServ->RemoveObserver(mObserver, "application-background");
obsServ->RemoveObserver(mObserver, "screen-information-changed");
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
}
mObserver = nullptr;
}
CleanShutdown();
}
void GPUProcessManager::OnPreferenceChange(const char16_t* aData) {
// We know prefs are ASCII here.
NS_LossyConvertUTF16toASCII strData(aData);
mozilla::dom::Pref pref(strData, /* isLocked */ false,
/* isSanitized */ false, Nothing(), Nothing());
Preferences::GetPreference(&pref, GeckoProcessType_GPU,
/* remoteType */ ""_ns);
if (!!mGPUChild) {
MOZ_ASSERT(mQueuedPrefs.IsEmpty());
mGPUChild->SendPreferenceUpdate(pref);
} else if (IsGPUProcessLaunching()) {
mQueuedPrefs.AppendElement(pref);
}
}
void GPUProcessManager::ScreenInformationChanged() {
#if defined(XP_WIN)
if (!!mGPUChild) {
mGPUChild->SendScreenInformationChanged();
}
#endif
}
void GPUProcessManager::NotifyBatteryInfo(
const hal::BatteryInformation& aBatteryInfo) {
if (mGPUChild) {
mGPUChild->SendNotifyBatteryInfo(aBatteryInfo);
}
}
void GPUProcessManager::ResetProcessStable() {
mTotalProcessAttempts++;
mProcessStable = false;
mProcessAttemptLastTime = TimeStamp::Now();
}
bool GPUProcessManager::IsProcessStable(const TimeStamp& aNow) {
if (mTotalProcessAttempts > 0) {
auto delta = (int32_t)(aNow - mProcessAttemptLastTime).ToMilliseconds();
if (delta < StaticPrefs::layers_gpu_process_stable_min_uptime_ms()) {
return false;
}
}
return mProcessStable;
}
bool GPUProcessManager::LaunchGPUProcess() {
if (mProcess) {
return true;
}
if (AppShutdown::IsInOrBeyond(ShutdownPhase::XPCOMShutdown)) {
return false;
}
// Start listening for pref changes so we can
// forward them to the process once it is running.
if (!mObserver) {
mObserver = new Observer(this);
nsContentUtils::RegisterShutdownObserver(mObserver);
Preferences::AddStrongObserver(mObserver, "");
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
nsCOMPtr<nsIObserverService> obsServ = services::GetObserverService();
if (obsServ) {
obsServ->AddObserver(mObserver, "application-foreground", false);
obsServ->AddObserver(mObserver, "application-background", false);
obsServ->AddObserver(mObserver, "screen-information-changed", false);
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
}
}
// Start the Vsync I/O thread so can use it as soon as the process launches.
EnsureVsyncIOThread();
// If the previous process didn't live long enough, increment our unstable
// attempts counter so that we don't end up in a restart loop. If the process
// did live long enough, reset the counter so that we don't disable the
// process too eagerly.
auto newTime = TimeStamp::Now();
if (IsProcessStable(newTime)) {
mUnstableProcessAttempts = 0;
} else {
mUnstableProcessAttempts++;
mozilla::glean::gpu_process::unstable_launch_attempts.Set(
mUnstableProcessAttempts);
}
mTotalProcessAttempts++;
mozilla::glean::gpu_process::total_launch_attempts.Set(mTotalProcessAttempts);
mProcessAttemptLastTime = newTime;
mProcessStable = false;
// If the process is launched whilst we're in the background it may never get
// a chance to be declared stable before it is killed again. We don't want
// this happening repeatedly to result in the GPU process being disabled, so
// we assume that processes launched whilst in the background are stable.
if (!mAppInForeground) {
gfxCriticalNote
<< "GPU process is being launched whilst app is in background";
mProcessStable = true;
}
std::vector<std::string> extraArgs;
ipc::ProcessChild::AddPlatformBuildID(extraArgs);
// The subprocess is launched asynchronously, so we wait for a callback to
// acquire the IPDL actor.
mProcess = new GPUProcessHost(this);
if (!mProcess->Launch(extraArgs)) {
DisableGPUProcess("Failed to launch GPU process");
}
return true;
}
bool GPUProcessManager::IsGPUProcessLaunching() {
MOZ_ASSERT(NS_IsMainThread());
return !!mProcess && !mGPUChild;
}
void GPUProcessManager::DisableGPUProcess(const char* aMessage) {
MaybeDisableGPUProcess(aMessage, /* aAllowRestart */ false);
}
bool GPUProcessManager::MaybeDisableGPUProcess(const char* aMessage,
bool aAllowRestart) {
if (!gfxConfig::IsEnabled(Feature::GPU_PROCESS)) {
return true;
}
if (!aAllowRestart) {
gfxConfig::SetFailed(Feature::GPU_PROCESS, FeatureStatus::Failed, aMessage);
}
bool wantRestart;
if (mLastError) {
wantRestart =
FallbackFromAcceleration(mLastError.value(), mLastErrorMsg.ref());
mLastError.reset();
mLastErrorMsg.reset();
} else {
wantRestart = gfxPlatform::FallbackFromAcceleration(
FeatureStatus::Unavailable, aMessage,
"FEATURE_FAILURE_GPU_PROCESS_ERROR"_ns);
}
if (aAllowRestart && wantRestart) {
// The fallback method can make use of the GPU process.
return false;
}
if (aAllowRestart) {
gfxConfig::SetFailed(Feature::GPU_PROCESS, FeatureStatus::Failed, aMessage);
}
MOZ_ASSERT(!gfxConfig::IsEnabled(Feature::GPU_PROCESS));
gfxCriticalNote << aMessage;
gfxPlatform::DisableGPUProcess();
mozilla::glean::gpu_process::feature_status.Set(
gfxConfig::GetFeature(Feature::GPU_PROCESS)
.GetStatusAndFailureIdString());
mozilla::glean::gpu_process::crash_fallbacks.Get("disabled"_ns).Add(1);
DestroyProcess();
ShutdownVsyncIOThread();
// Now the stability state is based upon the in process compositor session.
ResetProcessStable();
// We may have been in the middle of guaranteeing our various services are
// available when one failed. Some callers may fallback to using the same
// process equivalent, and we need to make sure those services are setup
// correctly. We cannot re-enter DisableGPUProcess from this call because we
// know that it is disabled in the config above.
DebugOnly<bool> ready = EnsureProtocolsReady();
MOZ_ASSERT_IF(!ready,
AppShutdown::IsInOrBeyond(ShutdownPhase::XPCOMShutdown));
// If we disable the GPU process during reinitialization after a previous
// crash, then we need to tell the content processes again, because they
// need to rebind to the UI process.
HandleProcessLost();
return true;
}
nsresult GPUProcessManager::EnsureGPUReady(
bool aRetryAfterFallback /* = true */) {
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
MOZ_ASSERT(NS_IsMainThread());
// We only wait to fail with NS_ERROR_ILLEGAL_DURING_SHUTDOWN if we would
// cause a state change or if we are in the middle of relaunching the GPU
// process.
bool inShutdown = AppShutdown::IsInOrBeyond(ShutdownPhase::XPCOMShutdown);
while (true) {
// Launch the GPU process if it is enabled but hasn't been (re-)launched
// yet.
if (!mProcess && gfxConfig::IsEnabled(Feature::GPU_PROCESS)) {
if (NS_WARN_IF(inShutdown)) {
return NS_ERROR_ILLEGAL_DURING_SHUTDOWN;
}
if (!LaunchGPUProcess()) {
return NS_ERROR_FAILURE;
}
}
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
if (mProcess && !mProcess->IsConnected()) {
if (NS_WARN_IF(inShutdown)) {
return NS_ERROR_ILLEGAL_DURING_SHUTDOWN;
}
if (!mProcess->WaitForLaunch()) {
// If this fails, we should have fired OnProcessLaunchComplete and
// removed the process.
MOZ_ASSERT(!mProcess && !mGPUChild);
return NS_ERROR_FAILURE;
}
}
// The only scenario this should be possible is if we raced with the
// initialization, which failed, and has already decided to disable the GPU
// process.
if (!mGPUChild) {
MOZ_DIAGNOSTIC_ASSERT(!gfxConfig::IsEnabled(Feature::GPU_PROCESS));
break;
}
if (mGPUChild->EnsureGPUReady()) {
return NS_OK;
}
// If the initialization above fails, we likely have a GPU process teardown
// waiting in our message queue (or will soon). If the fallback wants us to
// give up on the GPU process, we will exit the loop.
if (MaybeDisableGPUProcess("Failed to initialize GPU process",
/* aAllowRestart */ true)) {
MOZ_DIAGNOSTIC_ASSERT(!gfxConfig::IsEnabled(Feature::GPU_PROCESS));
break;
}
// Otherwise HandleProcessLost will explicitly teardown the process and
// prevent any pending events from triggering our fallback logic again, and
// we will retry with a different configuration.
MOZ_DIAGNOSTIC_ASSERT(gfxConfig::IsEnabled(Feature::GPU_PROCESS));
OnBlockingProcessUnexpectedShutdown();
// Some callers may need to reconfigure if we fellback.
if (!aRetryAfterFallback) {
return NS_ERROR_NOT_AVAILABLE;
}
}
// This is the first time we are trying to use the in-process compositor.
if (mTotalProcessAttempts == 0) {
if (NS_WARN_IF(inShutdown)) {
return NS_ERROR_ILLEGAL_DURING_SHUTDOWN;
}
ResetProcessStable();
}
return NS_ERROR_FAILURE;
}
bool GPUProcessManager::EnsureProtocolsReady() {
return EnsureCompositorManagerChild() && EnsureImageBridgeChild() &&
EnsureVRManager();
}
bool GPUProcessManager::EnsureCompositorManagerChild() {
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return false;
}
if (CompositorManagerChild::IsInitialized(mProcessToken)) {
return true;
}
if (NS_FAILED(rv)) {
CompositorManagerChild::InitSameProcess(AllocateNamespace(), mProcessToken);
return true;
}
ipc::Endpoint<PCompositorManagerParent> parentPipe;
ipc::Endpoint<PCompositorManagerChild> childPipe;
rv = PCompositorManager::CreateEndpoints(
mGPUChild->OtherPid(), base::GetCurrentProcId(), &parentPipe, &childPipe);
if (NS_FAILED(rv)) {
DisableGPUProcess("Failed to create PCompositorManager endpoints");
return true;
}
uint32_t cmNamespace = AllocateNamespace();
mGPUChild->SendInitCompositorManager(std::move(parentPipe), cmNamespace);
CompositorManagerChild::Init(std::move(childPipe), cmNamespace,
mProcessToken);
return true;
}
bool GPUProcessManager::EnsureImageBridgeChild() {
if (ImageBridgeChild::GetSingleton()) {
return true;
}
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return false;
}
if (NS_FAILED(rv)) {
ImageBridgeChild::InitSameProcess(AllocateNamespace());
return true;
}
ipc::Endpoint<PImageBridgeParent> parentPipe;
ipc::Endpoint<PImageBridgeChild> childPipe;
rv = PImageBridge::CreateEndpoints(
mGPUChild->OtherPid(), base::GetCurrentProcId(), &parentPipe, &childPipe);
if (NS_FAILED(rv)) {
DisableGPUProcess("Failed to create PImageBridge endpoints");
return true;
}
mGPUChild->SendInitImageBridge(std::move(parentPipe));
ImageBridgeChild::InitWithGPUProcess(std::move(childPipe),
AllocateNamespace());
return true;
}
bool GPUProcessManager::EnsureVRManager() {
if (VRManagerChild::IsCreated()) {
return true;
}
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return false;
}
if (NS_FAILED(rv)) {
VRManagerChild::InitSameProcess();
return true;
}
ipc::Endpoint<PVRManagerParent> parentPipe;
ipc::Endpoint<PVRManagerChild> childPipe;
rv = PVRManager::CreateEndpoints(
mGPUChild->OtherPid(), base::GetCurrentProcId(), &parentPipe, &childPipe);
if (NS_FAILED(rv)) {
DisableGPUProcess("Failed to create PVRManager endpoints");
return true;
}
mGPUChild->SendInitVRManager(std::move(parentPipe));
VRManagerChild::InitWithGPUProcess(std::move(childPipe));
return true;
}
#if defined(MOZ_WIDGET_ANDROID)
already_AddRefed<UiCompositorControllerChild>
GPUProcessManager::CreateUiCompositorController(nsBaseWidget* aWidget,
const LayersId aId) {
RefPtr<UiCompositorControllerChild> result;
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return nullptr;
}
if (NS_FAILED(rv)) {
Bug 1756700 - Delay compositor creation on Android to allow time for GPU process to launch. r=gfx-reviewers,geckoview-reviewers,aosmond,calu We noticed a cold_view_nav_start regression on Fenix from enabling the GPU process, and profiles showed time spent synchronously waiting for the GPU process to launch. This occured because the compositor was being created in nsWindow::Create, and as it requires the GPU process to be running it had to block until launch completed. The process is launched when the gfxPlatform is first initialized, but that was only occuring immediately prior to creating the compositor, which did not give it enough time to complete asynchronously. This patch makes it so that we initialize the gfxPlatform slightly earlier, and importantly delay creating the compositor until it is actually required. This gives the process enough time to launch asynchronously meaning we do not have to block. We started deliberately creating the compositor early on Android because of bug 1453501, to avoid a race condition where the compositor didn't exist when RemoteLayerTreeOwner::Initialize was called, causing us to use a basic layer manager. However, since bug 1741156 landed we now create the compositor on-demand, meaning this is no longer a possibility. Delaying compositor creation can, however, uncover another race condition. If the UICompositorControllerChild is opened on the UI thread before the main thread is able to set its pointer to the widget, then the java GeckoSession will never be notified that the compositor has been opened, and composition will never be resumed. This patch fixes this issue by setting the UiCompositorControllerChild's widget pointer in its constructor rather than immediately afterwards. Differential Revision: https://phabricator.services.mozilla.com/D139842
2022-03-02 16:56:28 +00:00
result = UiCompositorControllerChild::CreateForSameProcess(aId, aWidget);
} else {
ipc::Endpoint<PUiCompositorControllerParent> parentPipe;
ipc::Endpoint<PUiCompositorControllerChild> childPipe;
rv = PUiCompositorController::CreateEndpoints(mGPUChild->OtherPid(),
base::GetCurrentProcId(),
&parentPipe, &childPipe);
if (NS_FAILED(rv)) {
DisableGPUProcess("Failed to create PUiCompositorController endpoints");
return nullptr;
}
mGPUChild->SendInitUiCompositorController(aId, std::move(parentPipe));
result = UiCompositorControllerChild::CreateForGPUProcess(
Bug 1756700 - Delay compositor creation on Android to allow time for GPU process to launch. r=gfx-reviewers,geckoview-reviewers,aosmond,calu We noticed a cold_view_nav_start regression on Fenix from enabling the GPU process, and profiles showed time spent synchronously waiting for the GPU process to launch. This occured because the compositor was being created in nsWindow::Create, and as it requires the GPU process to be running it had to block until launch completed. The process is launched when the gfxPlatform is first initialized, but that was only occuring immediately prior to creating the compositor, which did not give it enough time to complete asynchronously. This patch makes it so that we initialize the gfxPlatform slightly earlier, and importantly delay creating the compositor until it is actually required. This gives the process enough time to launch asynchronously meaning we do not have to block. We started deliberately creating the compositor early on Android because of bug 1453501, to avoid a race condition where the compositor didn't exist when RemoteLayerTreeOwner::Initialize was called, causing us to use a basic layer manager. However, since bug 1741156 landed we now create the compositor on-demand, meaning this is no longer a possibility. Delaying compositor creation can, however, uncover another race condition. If the UICompositorControllerChild is opened on the UI thread before the main thread is able to set its pointer to the widget, then the java GeckoSession will never be notified that the compositor has been opened, and composition will never be resumed. This patch fixes this issue by setting the UiCompositorControllerChild's widget pointer in its constructor rather than immediately afterwards. Differential Revision: https://phabricator.services.mozilla.com/D139842
2022-03-02 16:56:28 +00:00
mProcessToken, std::move(childPipe), aWidget);
Bug 1741156 - Initial GPU process implementation on Android. r=aosmond,agi Declare a GPU process and corresponding Service in the AndroidManifest. This is of a new class GeckoServiceGpuProcess which inherits from GeckoServiceChildProcess, and provides a binder interface ICompositorSurfaceManager which allows the parent process to set the compositor Surface for a given widget ID, and the compositor in the GPU process to look up the Surface for a widget ID. The ICompositorSurfaceManager interface is exposed to the parent process through a new method getCompositorSurfaceManager() in the IChildProcess interface. Add a new connection type for GPU processes to GeckoProcessManager, along with a function to look up the GPU process connection and fetch the ICompositorSurfaceManager binder. When the GPU process is launched we store the returned binder in the GPUProcessHost, and when each widget's compositor is created we store a reference to the binder in the UiCompositorControllerChild. Each nsWindow is given a unique ID, and whenever the Surface changes due to an Android resume event, it sends the new surface for that ID to the GPU process (if enabled) by calling ICompositorSurfaceManager.onSurfaceChanged(). Stop inheriting AndroidCompositorWidget from InProcessCompositorWidget and instead inherit from CompositorWidget directly. This class holds a reference to the Surface that will be rendered in to. The CompositorBridgeParent notifies the CompositorWidget whenever it has been resumed, allowing it to fetch the new Surface. For the cross-process CompositorWidgetParent implementation it fetches that Surface from the CompositorSurfaceManagerService, whereas the InProcessAndroidCompositorWidget can read it directly from the real widget. AndroidCompositorWidget::GetClientSize() can now calculate its size from the Surface, rather than racily reading the value from the nsWindow. This means RenderCompositorEGL and RenderCompositorOGLSWGL can now use GetClientSize() again rather than querying their own size from the Surface. With this patch, setting layers.gpu-process.enabled to true will cause us to launch a GPU process and render from it. We do not yet gracefully recover from a GPU process crash, nor can we render anything using SurfaceTextures (eg video or webgl). Those will come in future patches. Differential Revision: https://phabricator.services.mozilla.com/D131231
2021-11-29 20:52:31 +00:00
if (result) {
result->SetCompositorSurfaceManager(
mProcess->GetCompositorSurfaceManager());
}
}
return result.forget();
}
#endif // defined(MOZ_WIDGET_ANDROID)
void GPUProcessManager::OnProcessLaunchComplete(GPUProcessHost* aHost) {
MOZ_ASSERT(mProcess && mProcess == aHost);
if (!mProcess->IsConnected()) {
DisableGPUProcess("Failed to connect GPU process");
return;
}
mGPUChild = mProcess->GetActor();
mProcessToken = mProcess->GetProcessToken();
#if defined(XP_WIN)
if (mAppInForeground) {
SetProcessIsForeground();
}
#endif
ipc::Endpoint<PVsyncBridgeParent> vsyncParent;
ipc::Endpoint<PVsyncBridgeChild> vsyncChild;
nsresult rv = PVsyncBridge::CreateEndpoints(mGPUChild->OtherPid(),
base::GetCurrentProcId(),
&vsyncParent, &vsyncChild);
if (NS_FAILED(rv)) {
DisableGPUProcess("Failed to create PVsyncBridge endpoints");
return;
}
mVsyncBridge = VsyncBridgeChild::Create(mVsyncIOThread, mProcessToken,
std::move(vsyncChild));
mGPUChild->SendInitVsyncBridge(std::move(vsyncParent));
MOZ_ASSERT(!mBatteryObserver);
mBatteryObserver = new BatteryObserver(this);
// Flush any pref updates that happened during launch and weren't
// included in the blobs set up in LaunchGPUProcess.
for (const mozilla::dom::Pref& pref : mQueuedPrefs) {
Unused << NS_WARN_IF(!mGPUChild->SendPreferenceUpdate(pref));
}
mQueuedPrefs.Clear();
Bug 1831092 - Use the new pull-based API for all crash annotations and remove the global annotations table r=jgilbert,necko-reviewers,media-playback-reviewers,profiler-reviewers,win-reviewers,padenot,handyman,afranchuk,valentin,alwu,sotaro This changes comes with several different refactorings all rolled into one, unfotunately I couldn't find a way to pull them apart: - First of all annotations now can either recorded (that is, we copy the value and have the crash reporting code own the copy) or registered. Several annotations are changed to use this functionality so that we don't need to update them as their value change. - The code in the exception handler is modified to read the annotations from the mozannotation_client crate. This has the unfortunate side-effect that we need three different bits of code to serialize them: one for annotations read from a child process, one for reading annotations from the main process outside of the exception handler and one for reading annotations from the main process within the exception handler. As we move to fully out-of-process crash reporting the last two methods will go away. - The mozannotation_client crate now doesn't record annotation types anymore. I realized as I was working on this that storing types at runtime has two issues: the first one is that buggy code might change the type of an annotation (that is record it under two different types at two different moments), the second issue is that types might become corrupt during a crash, so better enforce them at annotation-writing time. The end result is that the mozannotation_* crates now only store byte buffers, track the format the data is stored in (null-terminated string, fixed size buffer, etc...) but not the type of data each annotation is supposed to contain. - Which brings us to the next change: concrete types for annotations are now enforced when they're written out. If an annotation doesn't match the expected type it's skipped. Storing an annotation with the wrong type will also trigger an assertion in debug builds. Differential Revision: https://phabricator.services.mozilla.com/D195248
2024-03-04 10:24:43 +00:00
CrashReporter::RecordAnnotationCString(
CrashReporter::Annotation::GPUProcessStatus, "Running");
Bug 1831092 - Use the new pull-based API for all crash annotations and remove the global annotations table r=jgilbert,necko-reviewers,media-playback-reviewers,profiler-reviewers,win-reviewers,padenot,handyman,afranchuk,valentin,alwu,sotaro This changes comes with several different refactorings all rolled into one, unfotunately I couldn't find a way to pull them apart: - First of all annotations now can either recorded (that is, we copy the value and have the crash reporting code own the copy) or registered. Several annotations are changed to use this functionality so that we don't need to update them as their value change. - The code in the exception handler is modified to read the annotations from the mozannotation_client crate. This has the unfortunate side-effect that we need three different bits of code to serialize them: one for annotations read from a child process, one for reading annotations from the main process outside of the exception handler and one for reading annotations from the main process within the exception handler. As we move to fully out-of-process crash reporting the last two methods will go away. - The mozannotation_client crate now doesn't record annotation types anymore. I realized as I was working on this that storing types at runtime has two issues: the first one is that buggy code might change the type of an annotation (that is record it under two different types at two different moments), the second issue is that types might become corrupt during a crash, so better enforce them at annotation-writing time. The end result is that the mozannotation_* crates now only store byte buffers, track the format the data is stored in (null-terminated string, fixed size buffer, etc...) but not the type of data each annotation is supposed to contain. - Which brings us to the next change: concrete types for annotations are now enforced when they're written out. If an annotation doesn't match the expected type it's skipped. Storing an annotation with the wrong type will also trigger an assertion in debug builds. Differential Revision: https://phabricator.services.mozilla.com/D195248
2024-03-04 10:24:43 +00:00
CrashReporter::RecordAnnotationU32(
CrashReporter::Annotation::GPUProcessLaunchCount, mTotalProcessAttempts);
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
ReinitializeRendering();
}
void GPUProcessManager::OnProcessDeclaredStable() { mProcessStable = true; }
static bool ShouldLimitDeviceResets(uint32_t count, int32_t deltaMilliseconds) {
// We decide to limit by comparing the amount of resets that have happened
// and time since the last reset to two prefs.
int32_t timeLimit = StaticPrefs::gfx_device_reset_threshold_ms_AtStartup();
int32_t countLimit = StaticPrefs::gfx_device_reset_limit_AtStartup();
bool hasTimeLimit = timeLimit >= 0;
bool hasCountLimit = countLimit >= 0;
bool triggeredTime = deltaMilliseconds < timeLimit;
bool triggeredCount = count > (uint32_t)countLimit;
// If we have both prefs set then it needs to trigger both limits,
// otherwise we only test the pref that is set or none
if (hasTimeLimit && hasCountLimit) {
return triggeredTime && triggeredCount;
} else if (hasTimeLimit) {
return triggeredTime;
} else if (hasCountLimit) {
return triggeredCount;
}
return false;
}
void GPUProcessManager::ResetCompositors() {
// Note: this will recreate devices in addition to recreating compositors.
// This isn't optimal, but this is only used on linux where acceleration
// isn't enabled by default, and this way we don't need a new code path.
SimulateDeviceReset();
}
void GPUProcessManager::SimulateDeviceReset() {
// Make sure we rebuild environment and configuration for accelerated
// features.
gfxPlatform::GetPlatform()->CompositorUpdated();
if (mProcess) {
if (mGPUChild) {
mGPUChild->SendSimulateDeviceReset();
}
} else {
wr::RenderThread::Get()->SimulateDeviceReset();
}
}
bool GPUProcessManager::FallbackFromAcceleration(wr::WebRenderError aError,
const nsCString& aMsg) {
if (aError == wr::WebRenderError::INITIALIZE) {
return gfxPlatform::FallbackFromAcceleration(
gfx::FeatureStatus::Unavailable, "WebRender initialization failed",
aMsg);
} else if (aError == wr::WebRenderError::MAKE_CURRENT) {
return gfxPlatform::FallbackFromAcceleration(
gfx::FeatureStatus::Unavailable,
"Failed to make render context current",
"FEATURE_FAILURE_WEBRENDER_MAKE_CURRENT"_ns);
} else if (aError == wr::WebRenderError::RENDER) {
return gfxPlatform::FallbackFromAcceleration(
gfx::FeatureStatus::Unavailable, "Failed to render WebRender",
"FEATURE_FAILURE_WEBRENDER_RENDER"_ns);
} else if (aError == wr::WebRenderError::NEW_SURFACE) {
// If we cannot create a new Surface even in the final fallback
// configuration then force a crash.
return gfxPlatform::FallbackFromAcceleration(
gfx::FeatureStatus::Unavailable, "Failed to create new surface",
"FEATURE_FAILURE_WEBRENDER_NEW_SURFACE"_ns,
/* aCrashAfterFinalFallback */ true);
} else if (aError == wr::WebRenderError::BEGIN_DRAW) {
return gfxPlatform::FallbackFromAcceleration(
gfx::FeatureStatus::Unavailable, "BeginDraw() failed",
"FEATURE_FAILURE_WEBRENDER_BEGIN_DRAW"_ns);
} else if (aError == wr::WebRenderError::EXCESSIVE_RESETS) {
return gfxPlatform::FallbackFromAcceleration(
gfx::FeatureStatus::Unavailable, "Device resets exceeded threshold",
"FEATURE_FAILURE_WEBRENDER_EXCESSIVE_RESETS"_ns);
} else {
MOZ_ASSERT_UNREACHABLE("Invalid value");
return gfxPlatform::FallbackFromAcceleration(
gfx::FeatureStatus::Unavailable, "Unhandled failure reason",
"FEATURE_FAILURE_WEBRENDER_UNHANDLED"_ns);
}
}
bool GPUProcessManager::DisableWebRenderConfig(wr::WebRenderError aError,
const nsCString& aMsg) {
// If we have a stable compositor process, this may just be due to an OOM or
// bad driver state. In that case, we should consider restarting the GPU
// process, or simulating a device reset to teardown the compositors to
// hopefully alleviate the situation.
if (IsProcessStable(TimeStamp::Now())) {
if (mProcess) {
mProcess->KillProcess(/* aGenerateMinidump */ false);
} else {
SimulateDeviceReset();
}
mLastError = Some(aError);
mLastErrorMsg = Some(aMsg);
return false;
}
mLastError.reset();
mLastErrorMsg.reset();
// Disable WebRender
bool wantRestart = FallbackFromAcceleration(aError, aMsg);
gfxVars::SetUseWebRenderDCompVideoHwOverlayWin(false);
gfxVars::SetUseWebRenderDCompVideoSwOverlayWin(false);
// If we still have the GPU process, and we fallback to a new configuration
// that prefers to have the GPU process, reset the counter. Because we
// updated the gfxVars, we want to flag the GPUChild to wait for the update
// to be processed before creating new compositor sessions, otherwise we risk
// them being out of sync with the content/parent processes.
if (wantRestart && mProcess && mGPUChild) {
mUnstableProcessAttempts = 1;
mGPUChild->MarkWaitForVarUpdate();
}
return true;
}
void GPUProcessManager::DisableWebRender(wr::WebRenderError aError,
const nsCString& aMsg) {
if (DisableWebRenderConfig(aError, aMsg)) {
if (mProcess) {
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
DestroyRemoteCompositorSessions();
} else {
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
DestroyInProcessCompositorSessions();
}
NotifyListenersOnCompositeDeviceReset();
}
}
void GPUProcessManager::NotifyWebRenderError(wr::WebRenderError aError) {
gfxCriticalNote << "Handling webrender error " << (unsigned int)aError;
#ifdef XP_WIN
if (aError == wr::WebRenderError::VIDEO_OVERLAY) {
gfxVars::SetUseWebRenderDCompVideoHwOverlayWin(false);
gfxVars::SetUseWebRenderDCompVideoSwOverlayWin(false);
return;
}
if (aError == wr::WebRenderError::VIDEO_HW_OVERLAY) {
gfxVars::SetUseWebRenderDCompVideoHwOverlayWin(false);
return;
}
if (aError == wr::WebRenderError::VIDEO_SW_OVERLAY) {
gfxVars::SetUseWebRenderDCompVideoSwOverlayWin(false);
return;
}
#else
if (aError == wr::WebRenderError::VIDEO_OVERLAY ||
aError == wr::WebRenderError::VIDEO_HW_OVERLAY ||
aError == wr::WebRenderError::VIDEO_SW_OVERLAY) {
MOZ_ASSERT_UNREACHABLE("unexpected to be called");
return;
}
#endif
DisableWebRender(aError, nsCString());
}
/* static */
void GPUProcessManager::RecordDeviceReset(DeviceResetReason aReason) {
if (aReason != DeviceResetReason::FORCED_RESET) {
Telemetry::Accumulate(Telemetry::DEVICE_RESET_REASON, uint32_t(aReason));
}
Bug 1831092 - Use the new pull-based API for all crash annotations and remove the global annotations table r=jgilbert,necko-reviewers,media-playback-reviewers,profiler-reviewers,win-reviewers,padenot,handyman,afranchuk,valentin,alwu,sotaro This changes comes with several different refactorings all rolled into one, unfotunately I couldn't find a way to pull them apart: - First of all annotations now can either recorded (that is, we copy the value and have the crash reporting code own the copy) or registered. Several annotations are changed to use this functionality so that we don't need to update them as their value change. - The code in the exception handler is modified to read the annotations from the mozannotation_client crate. This has the unfortunate side-effect that we need three different bits of code to serialize them: one for annotations read from a child process, one for reading annotations from the main process outside of the exception handler and one for reading annotations from the main process within the exception handler. As we move to fully out-of-process crash reporting the last two methods will go away. - The mozannotation_client crate now doesn't record annotation types anymore. I realized as I was working on this that storing types at runtime has two issues: the first one is that buggy code might change the type of an annotation (that is record it under two different types at two different moments), the second issue is that types might become corrupt during a crash, so better enforce them at annotation-writing time. The end result is that the mozannotation_* crates now only store byte buffers, track the format the data is stored in (null-terminated string, fixed size buffer, etc...) but not the type of data each annotation is supposed to contain. - Which brings us to the next change: concrete types for annotations are now enforced when they're written out. If an annotation doesn't match the expected type it's skipped. Storing an annotation with the wrong type will also trigger an assertion in debug builds. Differential Revision: https://phabricator.services.mozilla.com/D195248
2024-03-04 10:24:43 +00:00
CrashReporter::RecordAnnotationU32(
CrashReporter::Annotation::DeviceResetReason,
static_cast<uint32_t>(aReason));
}
/* static */
void GPUProcessManager::NotifyDeviceReset(DeviceResetReason aReason,
DeviceResetDetectPlace aPlace) {
if (XRE_IsGPUProcess()) {
if (!GPUParent::GetSingleton()) {
MOZ_ASSERT_UNREACHABLE("unexpected to be called");
return;
}
// End up to GPUProcessManager::OnRemoteProcessDeviceReset()
GPUParent::GetSingleton()->NotifyDeviceReset(aReason, aPlace);
} else {
if (!GPUProcessManager::Get()) {
MOZ_ASSERT_UNREACHABLE("unexpected to be called");
return;
}
if (NS_IsMainThread()) {
GPUProcessManager::Get()->OnInProcessDeviceReset(aReason, aPlace);
} else {
NS_DispatchToMainThread(NS_NewRunnableFunction(
"gfx::GPUProcessManager::OnInProcessDeviceReset",
[aReason, aPlace]() -> void {
gfx::GPUProcessManager::Get()->OnInProcessDeviceReset(aReason,
aPlace);
}));
}
}
}
bool GPUProcessManager::OnDeviceReset(bool aTrackThreshold) {
// Ignore resets for thresholding if requested.
if (!aTrackThreshold) {
return false;
}
// Detect whether the device is resetting too quickly or too much
// indicating that we should give up and use software
mDeviceResetCount++;
auto newTime = TimeStamp::Now();
auto delta = (int32_t)(newTime - mDeviceResetLastTime).ToMilliseconds();
mDeviceResetLastTime = newTime;
// Returns true if we should disable acceleration due to the reset.
return ShouldLimitDeviceResets(mDeviceResetCount, delta);
}
void GPUProcessManager::OnInProcessDeviceReset(DeviceResetReason aReason,
DeviceResetDetectPlace aPlace) {
gfxCriticalNote << "Detect DeviceReset " << aReason << " " << aPlace
<< " in Parent process";
bool guilty;
switch (aReason) {
case DeviceResetReason::HUNG:
case DeviceResetReason::RESET:
case DeviceResetReason::INVALID_CALL:
guilty = true;
break;
default:
guilty = false;
break;
}
if (OnDeviceReset(guilty)) {
gfxCriticalNoteOnce << "In-process device reset threshold exceeded";
#ifdef MOZ_WIDGET_GTK
// FIXME(aosmond): Should we disable WebRender on other platforms?
DisableWebRenderConfig(wr::WebRenderError::EXCESSIVE_RESETS, nsCString());
#endif
}
#ifdef XP_WIN
// Ensure device reset handling before re-creating in process sessions.
// Normally nsWindow::OnPaint() already handled it.
gfxWindowsPlatform::GetPlatform()->HandleDeviceReset();
#endif
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
DestroyInProcessCompositorSessions();
NotifyListenersOnCompositeDeviceReset();
}
void GPUProcessManager::OnRemoteProcessDeviceReset(
GPUProcessHost* aHost, const DeviceResetReason& aReason,
const DeviceResetDetectPlace& aPlace) {
gfxCriticalNote << "Detect DeviceReset " << aReason << " " << aPlace
<< " in GPU process";
if (OnDeviceReset(/* aTrackThreshold */ true) &&
!DisableWebRenderConfig(wr::WebRenderError::EXCESSIVE_RESETS,
nsCString())) {
return;
}
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
DestroyRemoteCompositorSessions();
NotifyListenersOnCompositeDeviceReset();
}
void GPUProcessManager::NotifyListenersOnCompositeDeviceReset() {
for (const auto& listener : mListeners) {
listener->OnCompositorDeviceReset();
}
}
void GPUProcessManager::OnBlockingProcessUnexpectedShutdown() {
if (mProcess) {
CompositorManagerChild::OnGPUProcessLost(mProcess->GetProcessToken());
}
DestroyProcess(/* aUnexpectedShutdown */ true);
mUnstableProcessAttempts = 0;
HandleProcessLost();
}
void GPUProcessManager::OnProcessUnexpectedShutdown(GPUProcessHost* aHost) {
MOZ_ASSERT(mProcess && mProcess == aHost);
if (StaticPrefs::layers_gpu_process_crash_also_crashes_browser()) {
MOZ_CRASH("GPU process crashed and pref is set to crash the browser.");
}
CompositorManagerChild::OnGPUProcessLost(aHost->GetProcessToken());
DestroyProcess(/* aUnexpectedShutdown */ true);
if (mUnstableProcessAttempts >
uint32_t(StaticPrefs::layers_gpu_process_max_restarts())) {
char disableMessage[64];
SprintfLiteral(disableMessage, "GPU process disabled after %d attempts",
mTotalProcessAttempts);
if (!MaybeDisableGPUProcess(disableMessage, /* aAllowRestart */ true)) {
// Fallback wants the GPU process. Reset our counter.
MOZ_DIAGNOSTIC_ASSERT(gfxConfig::IsEnabled(Feature::GPU_PROCESS));
mUnstableProcessAttempts = 0;
HandleProcessLost();
} else {
MOZ_DIAGNOSTIC_ASSERT(!gfxConfig::IsEnabled(Feature::GPU_PROCESS));
}
} else if (mUnstableProcessAttempts >
uint32_t(StaticPrefs::
layers_gpu_process_max_restarts_with_decoder()) &&
mDecodeVideoOnGpuProcess) {
mDecodeVideoOnGpuProcess = false;
mozilla::glean::gpu_process::crash_fallbacks.Get("decoding_disabled"_ns)
.Add(1);
HandleProcessLost();
} else {
mozilla::glean::gpu_process::crash_fallbacks.Get("none"_ns).Add(1);
HandleProcessLost();
}
}
void GPUProcessManager::HandleProcessLost() {
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
MOZ_ASSERT(NS_IsMainThread());
// The shutdown and restart sequence for the GPU process is as follows:
//
// (1) The GPU process dies. IPDL will enqueue an ActorDestroy message on
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// each channel owning a bridge to the GPU process, on the thread owning
// that channel.
//
// (2) The first channel to process its ActorDestroy message will post a
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// message to the main thread to call NotifyRemoteActorDestroyed on the
// GPUProcessManager, which calls OnProcessUnexpectedShutdown if it has
// not handled shutdown for this process yet. OnProcessUnexpectedShutdown
// is responsible for tearing down the old process and deciding whether
// or not to disable the GPU process. It then calls this function,
// HandleProcessLost.
//
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// (3) We then notify each widget that its session with the compositor is now
// invalid. The widget is responsible for destroying its layer manager
// and CompositorBridgeChild. Note that at this stage, not all actors may
// have received ActorDestroy yet. CompositorBridgeChild may attempt to
// send messages, and if this happens, it will probably report a
// MsgDropped error. This is okay.
//
// (4) At this point, the UI process has a clean slate: no layers should
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// exist for the old compositor. We may make a decision on whether or not
// to re-launch the GPU process. Or, on Android if the app is in the
// background we may decide to wait until it comes to the foreground
// before re-launching.
//
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// (5) When we do decide to re-launch, or continue without a GPU process, we
// notify each ContentParent of the lost connection. It will request new
// endpoints from the GPUProcessManager and forward them to its
// ContentChild. The parent-side of these endpoints may come from the
// compositor thread of the UI process, or the compositor thread of the
// GPU process. However, no actual compositors should exist yet.
//
// (6) Each ContentChild will receive new endpoints. It will destroy its
// Compositor/ImageBridgeChild singletons and recreate them, as well
// as invalidate all retained layers.
//
// (7) In addition, each ContentChild will ask each of its BrowserChildren
// to re-request association with the compositor for the window
// owning the tab. The sequence of calls looks like:
// (a) [CONTENT] ContentChild::RecvReinitRendering
// (b) [CONTENT] BrowserChild::ReinitRendering
// (c) [CONTENT] BrowserChild::SendEnsureLayersConnected
// (d) [UI] BrowserParent::RecvEnsureLayersConnected
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// (e) [UI] RemoteLayerTreeOwner::EnsureLayersConnected
// (f) [UI] CompositorBridgeChild::SendNotifyChildRecreated
//
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// Note that at step (e), RemoteLayerTreeOwner will call
// GetWindowRenderer on the nsIWidget owning the tab. This step ensures
// that a compositor exists for the window. If we decided to launch a new
// GPU Process, at this point we block until the process has launched and
// we're able to create a new window compositor. Otherwise, if
// compositing is now in-process, this will simply create a new
// CompositorBridgeParent in the UI process. If there are multiple tabs
// in the same window, additional tabs will simply return the already-
// established compositor.
//
// Finally, this step serves one other crucial function: tabs must be
// associated with a window compositor or else they can't forward
// layer transactions. So this step both ensures that a compositor
// exists, and that the tab can forward layers.
//
// (8) Last, if the window had no remote tabs, step (7) will not have
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// applied, and the window will not have a new compositor just yet. The
// next refresh tick and paint will ensure that one exists, again via
// nsIWidget::GetWindowRenderer. On Android, we called
// nsIWidgetListener::RequestRepaint back in step (3) to ensure this
// tick occurs, but on other platforms this is not necessary.
DestroyRemoteCompositorSessions();
#ifdef MOZ_WIDGET_ANDROID
java::SurfaceControlManager::GetInstance()->OnGpuProcessLoss();
#endif
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// Re-launch the process if immediately if the GPU process is still enabled.
// Except on Android if the app is in the background, where we want to wait
// until the app is in the foreground again.
if (gfxConfig::IsEnabled(Feature::GPU_PROCESS)) {
#ifdef MOZ_WIDGET_ANDROID
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
if (mAppInForeground) {
#else
{
#endif
Unused << LaunchGPUProcess();
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
}
} else {
// If the GPU process is disabled we can reinitialize rendering immediately.
// This will be handled in OnProcessLaunchComplete() if the GPU process is
// enabled.
ReinitializeRendering();
}
}
void GPUProcessManager::ReinitializeRendering() {
// Notify content. This will ensure that each content process re-establishes
// a connection to the compositor thread (whether it's in-process or in a
// newly launched GPU process).
for (const auto& listener : mListeners) {
listener->OnCompositorUnexpectedShutdown();
}
// Notify any observers that the compositor has been reinitialized,
// eg the ZoomConstraintsClients for parent process documents.
nsCOMPtr<nsIObserverService> observerService = services::GetObserverService();
if (observerService) {
observerService->NotifyObservers(nullptr, "compositor-reinitialized",
nullptr);
}
}
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
void GPUProcessManager::DestroyRemoteCompositorSessions() {
// Build a list of sessions to notify, since notification might delete
// entries from the list.
nsTArray<RefPtr<RemoteCompositorSession>> sessions;
for (auto& session : mRemoteSessions) {
sessions.AppendElement(session);
}
// Notify each widget that we have lost the GPU process. This will ensure
// that each widget destroys its layer manager and CompositorBridgeChild.
for (const auto& session : sessions) {
session->NotifySessionLost();
}
}
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
void GPUProcessManager::DestroyInProcessCompositorSessions() {
// Build a list of sessions to notify, since notification might delete
// entries from the list.
nsTArray<RefPtr<InProcessCompositorSession>> sessions;
for (auto& session : mInProcessSessions) {
sessions.AppendElement(session);
}
// Notify each widget that we have lost the GPU process. This will ensure
// that each widget destroys its layer manager and CompositorBridgeChild.
for (const auto& session : sessions) {
session->NotifySessionLost();
}
// Ensure our stablility state is reset so that we don't necessarily crash
// right away on some WebRender errors.
CompositorBridgeParent::ResetStable();
ResetProcessStable();
}
void GPUProcessManager::NotifyRemoteActorDestroyed(
const uint64_t& aProcessToken) {
if (!NS_IsMainThread()) {
RefPtr<Runnable> task = mTaskFactory.NewRunnableMethod(
&GPUProcessManager::NotifyRemoteActorDestroyed, aProcessToken);
NS_DispatchToMainThread(task.forget());
return;
}
if (mProcessToken != aProcessToken) {
// This token is for an older process; we can safely ignore it.
return;
}
// One of the bridged top-level actors for the GPU process has been
// prematurely terminated, and we're receiving a notification. This
// can happen if the ActorDestroy for a bridged protocol fires
// before the ActorDestroy for PGPUChild.
OnProcessUnexpectedShutdown(mProcess);
}
void GPUProcessManager::CleanShutdown() {
DestroyProcess();
mVsyncIOThread = nullptr;
}
void GPUProcessManager::KillProcess(bool aGenerateMinidump) {
if (!mProcess) {
return;
}
mProcess->KillProcess(aGenerateMinidump);
}
void GPUProcessManager::CrashProcess() {
if (!mProcess) {
return;
}
mProcess->CrashProcess();
}
void GPUProcessManager::DestroyProcess(bool aUnexpectedShutdown) {
if (!mProcess) {
return;
}
mProcess->Shutdown(aUnexpectedShutdown);
mProcessToken = 0;
mProcess = nullptr;
mGPUChild = nullptr;
mQueuedPrefs.Clear();
if (mVsyncBridge) {
mVsyncBridge->Close();
mVsyncBridge = nullptr;
}
if (mBatteryObserver) {
mBatteryObserver->ShutDown();
mBatteryObserver = nullptr;
}
Bug 1831092 - Use the new pull-based API for all crash annotations and remove the global annotations table r=jgilbert,necko-reviewers,media-playback-reviewers,profiler-reviewers,win-reviewers,padenot,handyman,afranchuk,valentin,alwu,sotaro This changes comes with several different refactorings all rolled into one, unfotunately I couldn't find a way to pull them apart: - First of all annotations now can either recorded (that is, we copy the value and have the crash reporting code own the copy) or registered. Several annotations are changed to use this functionality so that we don't need to update them as their value change. - The code in the exception handler is modified to read the annotations from the mozannotation_client crate. This has the unfortunate side-effect that we need three different bits of code to serialize them: one for annotations read from a child process, one for reading annotations from the main process outside of the exception handler and one for reading annotations from the main process within the exception handler. As we move to fully out-of-process crash reporting the last two methods will go away. - The mozannotation_client crate now doesn't record annotation types anymore. I realized as I was working on this that storing types at runtime has two issues: the first one is that buggy code might change the type of an annotation (that is record it under two different types at two different moments), the second issue is that types might become corrupt during a crash, so better enforce them at annotation-writing time. The end result is that the mozannotation_* crates now only store byte buffers, track the format the data is stored in (null-terminated string, fixed size buffer, etc...) but not the type of data each annotation is supposed to contain. - Which brings us to the next change: concrete types for annotations are now enforced when they're written out. If an annotation doesn't match the expected type it's skipped. Storing an annotation with the wrong type will also trigger an assertion in debug builds. Differential Revision: https://phabricator.services.mozilla.com/D195248
2024-03-04 10:24:43 +00:00
CrashReporter::RecordAnnotationCString(
CrashReporter::Annotation::GPUProcessStatus, "Destroyed");
}
already_AddRefed<CompositorSession> GPUProcessManager::CreateTopLevelCompositor(
nsBaseWidget* aWidget, WebRenderLayerManager* aLayerManager,
CSSToLayoutDeviceScale aScale, const CompositorOptions& aOptions,
bool aUseExternalSurfaceSize, const gfx::IntSize& aSurfaceSize,
uint64_t aInnerWindowId, bool* aRetryOut) {
MOZ_ASSERT(aRetryOut);
LayersId layerTreeId = AllocateLayerTreeId();
RefPtr<CompositorSession> session;
nsresult rv = EnsureGPUReady(/* aRetryAfterFallback */ false);
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
*aRetryOut = false;
return nullptr;
}
// If we used fallback, then retry creating the compositor sessions because
// our configuration may have changed.
if (rv == NS_ERROR_NOT_AVAILABLE) {
*aRetryOut = true;
return nullptr;
}
if (!EnsureProtocolsReady()) {
*aRetryOut = false;
return nullptr;
}
if (NS_SUCCEEDED(rv)) {
session = CreateRemoteSession(aWidget, aLayerManager, layerTreeId, aScale,
aOptions, aUseExternalSurfaceSize,
aSurfaceSize, aInnerWindowId);
if (NS_WARN_IF(!session)) {
if (!MaybeDisableGPUProcess("Failed to create remote compositor",
/* aAllowRestart */ true)) {
// Fallback wants the GPU process. Reset our counter.
MOZ_DIAGNOSTIC_ASSERT(gfxConfig::IsEnabled(Feature::GPU_PROCESS));
OnBlockingProcessUnexpectedShutdown();
} else {
MOZ_DIAGNOSTIC_ASSERT(!gfxConfig::IsEnabled(Feature::GPU_PROCESS));
}
*aRetryOut = true;
return nullptr;
}
} else {
session = InProcessCompositorSession::Create(
aWidget, aLayerManager, layerTreeId, aScale, aOptions,
aUseExternalSurfaceSize, aSurfaceSize, AllocateNamespace(),
aInnerWindowId);
}
#if defined(MOZ_WIDGET_ANDROID)
if (session) {
// Nothing to do if controller gets a nullptr
RefPtr<UiCompositorControllerChild> controller =
CreateUiCompositorController(aWidget, session->RootLayerTreeId());
MOZ_ASSERT(controller);
session->SetUiCompositorControllerChild(controller);
}
#endif // defined(MOZ_WIDGET_ANDROID)
*aRetryOut = false;
return session.forget();
}
RefPtr<CompositorSession> GPUProcessManager::CreateRemoteSession(
nsBaseWidget* aWidget, WebRenderLayerManager* aLayerManager,
const LayersId& aRootLayerTreeId, CSSToLayoutDeviceScale aScale,
const CompositorOptions& aOptions, bool aUseExternalSurfaceSize,
const gfx::IntSize& aSurfaceSize, uint64_t aInnerWindowId) {
#ifdef MOZ_WIDGET_SUPPORTS_OOP_COMPOSITING
widget::CompositorWidgetInitData initData;
aWidget->GetCompositorWidgetInitData(&initData);
RefPtr<CompositorBridgeChild> child =
CompositorManagerChild::CreateWidgetCompositorBridge(
mProcessToken, aLayerManager, AllocateNamespace(), aScale, aOptions,
aUseExternalSurfaceSize, aSurfaceSize, aInnerWindowId);
if (!child) {
gfxCriticalNote << "Failed to create CompositorBridgeChild";
return nullptr;
}
RefPtr<CompositorVsyncDispatcher> dispatcher =
aWidget->GetCompositorVsyncDispatcher();
RefPtr<widget::CompositorWidgetVsyncObserver> observer =
new widget::CompositorWidgetVsyncObserver(mVsyncBridge, aRootLayerTreeId);
widget::CompositorWidgetChild* widget =
new widget::CompositorWidgetChild(dispatcher, observer, initData);
if (!child->SendPCompositorWidgetConstructor(widget, initData)) {
return nullptr;
}
if (!widget->Initialize()) {
return nullptr;
}
if (!child->SendInitialize(aRootLayerTreeId)) {
return nullptr;
}
RefPtr<APZCTreeManagerChild> apz = nullptr;
if (aOptions.UseAPZ()) {
PAPZCTreeManagerChild* papz =
child->SendPAPZCTreeManagerConstructor(LayersId{0});
if (!papz) {
return nullptr;
}
apz = static_cast<APZCTreeManagerChild*>(papz);
Bug 1340301 - Ensure APZ functions are called from correct threads on Android with GPU process. r=botond On Android the APZ controller thread is the android UI thread, rather than the Gecko main thread as on other platforms. There some places where the main thread requires to call IAPZCTreeManager functions that must run on the controller thread. Currently we use the function DispatchToControllerThread() prior to calling various IAPZCTreeManager APIs to achieve this. This works just fine for now, as there is no GPU process on Android. However, once we do add a GPU process we will encounter issues: Firstly, there will now be a cross-process APZInputBridge rather than using an in-process APZCTreeManager. The PAPZInputBridge protocol is managed by PGPU, and therefore must run on the main thread in the parent process. The input we require to send over the bridge, however, originates from the UI thread. To solve this we can convert PAPZInputBridge to a top-level protocol, and bind it to the UI thread on Android. We can then send input directly from the UI thread without issues. Secondly, the PAPZCTreeManager protocol must also run from the main thread in the parent process, as it is managed by PCompositorBridge. Unlike PAPZInputBridge we cannot convert PAPZCTreeManager in to a top level protocol, as it relies on the ordering guarantees with PCompositorBridge. We must therefore ensure that we only dispatch IAPZCTreeManager calls to the controller thread when using an in-process APZCTreeManager. Out-of-process calls, on the other hand, must be dispatched to the main thread where we can send IPDL commands from. To do this, we move the dispatch logic away from the callsites of IAPZCTreeManager APIs, and in to the APZCTreeManager and APZCTreeManagerChild implementations themselves. Differential Revision: https://phabricator.services.mozilla.com/D131120
2021-11-20 09:49:14 +00:00
ipc::Endpoint<PAPZInputBridgeParent> parentPipe;
ipc::Endpoint<PAPZInputBridgeChild> childPipe;
nsresult rv = PAPZInputBridge::CreateEndpoints(mGPUChild->OtherPid(),
base::GetCurrentProcId(),
&parentPipe, &childPipe);
if (NS_FAILED(rv)) {
return nullptr;
}
Bug 1340301 - Ensure APZ functions are called from correct threads on Android with GPU process. r=botond On Android the APZ controller thread is the android UI thread, rather than the Gecko main thread as on other platforms. There some places where the main thread requires to call IAPZCTreeManager functions that must run on the controller thread. Currently we use the function DispatchToControllerThread() prior to calling various IAPZCTreeManager APIs to achieve this. This works just fine for now, as there is no GPU process on Android. However, once we do add a GPU process we will encounter issues: Firstly, there will now be a cross-process APZInputBridge rather than using an in-process APZCTreeManager. The PAPZInputBridge protocol is managed by PGPU, and therefore must run on the main thread in the parent process. The input we require to send over the bridge, however, originates from the UI thread. To solve this we can convert PAPZInputBridge to a top-level protocol, and bind it to the UI thread on Android. We can then send input directly from the UI thread without issues. Secondly, the PAPZCTreeManager protocol must also run from the main thread in the parent process, as it is managed by PCompositorBridge. Unlike PAPZInputBridge we cannot convert PAPZCTreeManager in to a top level protocol, as it relies on the ordering guarantees with PCompositorBridge. We must therefore ensure that we only dispatch IAPZCTreeManager calls to the controller thread when using an in-process APZCTreeManager. Out-of-process calls, on the other hand, must be dispatched to the main thread where we can send IPDL commands from. To do this, we move the dispatch logic away from the callsites of IAPZCTreeManager APIs, and in to the APZCTreeManager and APZCTreeManagerChild implementations themselves. Differential Revision: https://phabricator.services.mozilla.com/D131120
2021-11-20 09:49:14 +00:00
mGPUChild->SendInitAPZInputBridge(aRootLayerTreeId, std::move(parentPipe));
RefPtr<APZInputBridgeChild> inputBridge =
APZInputBridgeChild::Create(mProcessToken, std::move(childPipe));
if (!inputBridge) {
return nullptr;
}
apz->SetInputBridge(inputBridge);
}
return new RemoteCompositorSession(aWidget, child, widget, apz,
aRootLayerTreeId);
#else
gfxCriticalNote << "Platform does not support out-of-process compositing";
return nullptr;
#endif
}
bool GPUProcessManager::CreateContentBridges(
base::ProcessId aOtherProcess,
ipc::Endpoint<PCompositorManagerChild>* aOutCompositor,
ipc::Endpoint<PImageBridgeChild>* aOutImageBridge,
ipc::Endpoint<PVRManagerChild>* aOutVRBridge,
ipc::Endpoint<PRemoteDecoderManagerChild>* aOutVideoManager,
dom::ContentParentId aChildId, nsTArray<uint32_t>* aNamespaces) {
const uint32_t cmNamespace = AllocateNamespace();
if (!CreateContentCompositorManager(aOtherProcess, aChildId, cmNamespace,
aOutCompositor) ||
!CreateContentImageBridge(aOtherProcess, aChildId, aOutImageBridge) ||
!CreateContentVRManager(aOtherProcess, aChildId, aOutVRBridge)) {
return false;
}
// VideoDeocderManager is only supported in the GPU process, so we allow this
// to be fallible.
CreateContentRemoteDecoderManager(aOtherProcess, aChildId, aOutVideoManager);
// Allocates 3 namespaces(for CompositorManagerChild, CompositorBridgeChild
// and ImageBridgeChild)
aNamespaces->AppendElement(cmNamespace);
aNamespaces->AppendElement(AllocateNamespace());
aNamespaces->AppendElement(AllocateNamespace());
return true;
}
bool GPUProcessManager::CreateContentCompositorManager(
base::ProcessId aOtherProcess, dom::ContentParentId aChildId,
uint32_t aNamespace, ipc::Endpoint<PCompositorManagerChild>* aOutEndpoint) {
ipc::Endpoint<PCompositorManagerParent> parentPipe;
ipc::Endpoint<PCompositorManagerChild> childPipe;
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return false;
}
base::ProcessId parentPid =
NS_SUCCEEDED(rv) ? mGPUChild->OtherPid() : base::GetCurrentProcId();
rv = PCompositorManager::CreateEndpoints(parentPid, aOtherProcess,
&parentPipe, &childPipe);
if (NS_FAILED(rv)) {
gfxCriticalNote << "Could not create content compositor manager: "
<< hexa(int(rv));
return false;
}
if (mGPUChild) {
mGPUChild->SendNewContentCompositorManager(std::move(parentPipe), aChildId,
aNamespace);
} else if (!CompositorManagerParent::Create(std::move(parentPipe), aChildId,
aNamespace,
/* aIsRoot */ false)) {
return false;
}
*aOutEndpoint = std::move(childPipe);
return true;
}
bool GPUProcessManager::CreateContentImageBridge(
base::ProcessId aOtherProcess, dom::ContentParentId aChildId,
ipc::Endpoint<PImageBridgeChild>* aOutEndpoint) {
if (!EnsureImageBridgeChild()) {
return false;
}
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return false;
}
base::ProcessId parentPid =
NS_SUCCEEDED(rv) ? mGPUChild->OtherPid() : base::GetCurrentProcId();
ipc::Endpoint<PImageBridgeParent> parentPipe;
ipc::Endpoint<PImageBridgeChild> childPipe;
rv = PImageBridge::CreateEndpoints(parentPid, aOtherProcess, &parentPipe,
&childPipe);
if (NS_FAILED(rv)) {
gfxCriticalNote << "Could not create content compositor bridge: "
<< hexa(int(rv));
return false;
}
if (mGPUChild) {
mGPUChild->SendNewContentImageBridge(std::move(parentPipe), aChildId);
} else {
if (!ImageBridgeParent::CreateForContent(std::move(parentPipe), aChildId)) {
return false;
}
}
*aOutEndpoint = std::move(childPipe);
return true;
}
base::ProcessId GPUProcessManager::GPUProcessPid() {
base::ProcessId gpuPid =
mGPUChild ? mGPUChild->OtherPid() : base::kInvalidProcessId;
return gpuPid;
}
bool GPUProcessManager::CreateContentVRManager(
base::ProcessId aOtherProcess, dom::ContentParentId aChildId,
ipc::Endpoint<PVRManagerChild>* aOutEndpoint) {
if (NS_WARN_IF(!EnsureVRManager())) {
return false;
}
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return false;
}
base::ProcessId parentPid =
NS_SUCCEEDED(rv) ? mGPUChild->OtherPid() : base::GetCurrentProcId();
ipc::Endpoint<PVRManagerParent> parentPipe;
ipc::Endpoint<PVRManagerChild> childPipe;
rv = PVRManager::CreateEndpoints(parentPid, aOtherProcess, &parentPipe,
&childPipe);
if (NS_FAILED(rv)) {
gfxCriticalNote << "Could not create content compositor bridge: "
<< hexa(int(rv));
return false;
}
if (mGPUChild) {
mGPUChild->SendNewContentVRManager(std::move(parentPipe), aChildId);
} else {
if (!VRManagerParent::CreateForContent(std::move(parentPipe), aChildId)) {
return false;
}
}
*aOutEndpoint = std::move(childPipe);
return true;
}
void GPUProcessManager::CreateContentRemoteDecoderManager(
base::ProcessId aOtherProcess, dom::ContentParentId aChildId,
ipc::Endpoint<PRemoteDecoderManagerChild>* aOutEndpoint) {
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return;
}
if (NS_FAILED(rv) || !StaticPrefs::media_gpu_process_decoder() ||
!mDecodeVideoOnGpuProcess) {
return;
}
ipc::Endpoint<PRemoteDecoderManagerParent> parentPipe;
ipc::Endpoint<PRemoteDecoderManagerChild> childPipe;
rv = PRemoteDecoderManager::CreateEndpoints(
mGPUChild->OtherPid(), aOtherProcess, &parentPipe, &childPipe);
if (NS_FAILED(rv)) {
gfxCriticalNote << "Could not create content video decoder: "
<< hexa(int(rv));
return;
}
mGPUChild->SendNewContentRemoteDecoderManager(std::move(parentPipe),
aChildId);
*aOutEndpoint = std::move(childPipe);
}
void GPUProcessManager::InitVideoBridge(
ipc::Endpoint<PVideoBridgeParent>&& aVideoBridge,
layers::VideoBridgeSource aSource) {
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return;
}
if (NS_SUCCEEDED(rv)) {
mGPUChild->SendInitVideoBridge(std::move(aVideoBridge), aSource);
}
}
void GPUProcessManager::MapLayerTreeId(LayersId aLayersId,
base::ProcessId aOwningId) {
nsresult rv = EnsureGPUReady();
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return;
}
if (NS_SUCCEEDED(rv)) {
mGPUChild->SendAddLayerTreeIdMapping(
LayerTreeIdMapping(aLayersId, aOwningId));
}
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// Must do this *after* the call to EnsureGPUReady, so that if the
// process is launched as a result then it is initialized without this
// LayersId, meaning it can be successfully mapped.
LayerTreeOwnerTracker::Get()->Map(aLayersId, aOwningId);
}
void GPUProcessManager::UnmapLayerTreeId(LayersId aLayersId,
base::ProcessId aOwningId) {
// Only call EnsureGPUReady() if we have already launched the process, to
// avoid launching a new process unnecesarily. (eg if we are backgrounded)
nsresult rv = mProcess ? EnsureGPUReady() : NS_ERROR_NOT_AVAILABLE;
if (NS_WARN_IF(rv == NS_ERROR_ILLEGAL_DURING_SHUTDOWN)) {
return;
}
if (NS_SUCCEEDED(rv)) {
mGPUChild->SendRemoveLayerTreeIdMapping(
LayerTreeIdMapping(aLayersId, aOwningId));
} else if (!mProcess) {
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
CompositorBridgeParent::DeallocateLayerTreeId(aLayersId);
}
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// Must do this *after* the call to EnsureGPUReady, so that if the
// process is launched as a result then it is initialized with this
// LayersId, meaning it can be successfully unmapped.
LayerTreeOwnerTracker::Get()->Unmap(aLayersId, aOwningId);
}
bool GPUProcessManager::IsLayerTreeIdMapped(LayersId aLayersId,
base::ProcessId aRequestingId) {
return LayerTreeOwnerTracker::Get()->IsMapped(aLayersId, aRequestingId);
}
LayersId GPUProcessManager::AllocateLayerTreeId() {
// Allocate tree id by using id namespace.
// By it, tree id does not conflict with external image id and
// async image pipeline id.
MOZ_ASSERT(NS_IsMainThread());
++mResourceId;
if (mResourceId == UINT32_MAX) {
// Move to next id namespace.
mIdNamespace = AllocateNamespace();
mResourceId = 1;
}
uint64_t layerTreeId = mIdNamespace;
layerTreeId = (layerTreeId << 32) | mResourceId;
return LayersId{layerTreeId};
}
uint32_t GPUProcessManager::AllocateNamespace() {
MOZ_ASSERT(NS_IsMainThread());
return ++mNextNamespace;
}
bool GPUProcessManager::AllocateAndConnectLayerTreeId(
PCompositorBridgeChild* aCompositorBridge, base::ProcessId aOtherPid,
LayersId* aOutLayersId, CompositorOptions* aOutCompositorOptions) {
LayersId layersId = AllocateLayerTreeId();
*aOutLayersId = layersId;
if (!mGPUChild || !aCompositorBridge) {
// If we're not remoting to another process, or there is no compositor,
// then we'll send at most one message. In this case we can just keep
// the old behavior of making sure the mapping occurs, and maybe sending
// a creation notification.
MapLayerTreeId(layersId, aOtherPid);
if (!aCompositorBridge) {
return false;
}
return aCompositorBridge->SendNotifyChildCreated(layersId,
aOutCompositorOptions);
}
// Use the combined message path.
LayerTreeOwnerTracker::Get()->Map(layersId, aOtherPid);
return aCompositorBridge->SendMapAndNotifyChildCreated(layersId, aOtherPid,
aOutCompositorOptions);
}
void GPUProcessManager::EnsureVsyncIOThread() {
if (mVsyncIOThread) {
return;
}
mVsyncIOThread = new VsyncIOThreadHolder();
MOZ_RELEASE_ASSERT(mVsyncIOThread->Start());
}
void GPUProcessManager::ShutdownVsyncIOThread() { mVsyncIOThread = nullptr; }
void GPUProcessManager::RegisterRemoteProcessSession(
RemoteCompositorSession* aSession) {
mRemoteSessions.AppendElement(aSession);
}
void GPUProcessManager::UnregisterRemoteProcessSession(
RemoteCompositorSession* aSession) {
mRemoteSessions.RemoveElement(aSession);
}
void GPUProcessManager::RegisterInProcessSession(
InProcessCompositorSession* aSession) {
mInProcessSessions.AppendElement(aSession);
}
void GPUProcessManager::UnregisterInProcessSession(
InProcessCompositorSession* aSession) {
mInProcessSessions.RemoveElement(aSession);
}
void GPUProcessManager::AddListener(GPUProcessListener* aListener) {
mListeners.AppendElement(aListener);
}
void GPUProcessManager::RemoveListener(GPUProcessListener* aListener) {
mListeners.RemoveElement(aListener);
}
bool GPUProcessManager::NotifyGpuObservers(const char* aTopic) {
if (NS_FAILED(EnsureGPUReady())) {
return false;
}
nsCString topic(aTopic);
mGPUChild->SendNotifyGpuObservers(topic);
return true;
}
class GPUMemoryReporter : public MemoryReportingProcess {
public:
NS_INLINE_DECL_THREADSAFE_REFCOUNTING(GPUMemoryReporter, override)
bool IsAlive() const override {
if (GPUProcessManager* gpm = GPUProcessManager::Get()) {
return !!gpm->GetGPUChild();
}
return false;
}
bool SendRequestMemoryReport(
const uint32_t& aGeneration, const bool& aAnonymize,
const bool& aMinimizeMemoryUsage,
const Maybe<ipc::FileDescriptor>& aDMDFile) override {
GPUChild* child = GetChild();
if (!child) {
return false;
}
return child->SendRequestMemoryReport(aGeneration, aAnonymize,
aMinimizeMemoryUsage, aDMDFile);
}
int32_t Pid() const override {
if (GPUChild* child = GetChild()) {
return (int32_t)child->OtherPid();
}
return 0;
}
private:
GPUChild* GetChild() const {
if (GPUProcessManager* gpm = GPUProcessManager::Get()) {
if (GPUChild* child = gpm->GetGPUChild()) {
return child;
}
}
return nullptr;
}
protected:
~GPUMemoryReporter() = default;
};
RefPtr<MemoryReportingProcess> GPUProcessManager::GetProcessMemoryReporter() {
Bug 1755381 - Avoid relaunching GPU process immediately on Android if app is in background. r=agi,aosmond If the android system kills the GPU process to free memory while the app is in the background, then we want to avoid immediately restarting the GPU process. To achieve this, we make GPUProcessManager keep track of whether it is in the foreground or background. If HandleProcessLost() gets called while in the background then we destroy the existing compositor sessions as before, but return early instead of immediately relaunching the process. If the process has not been launched when the app later gets foregrounded then we do so then. The final part of HandleProcessLost(), which reinitializes the content bridges and emits the "compositor-reinitialized" signal, has been moved to a new function ReinitializeRendering(). If the GPU process has been disabled, this gets called as-before at the end of HandleProcessLost(). When the GPU process is enabled, however, we now call it from OnProcessLaunchComplete(), so that it gets called regardless of whether the process is launched immediately or after a delay. While we're here, rename the functions RebuildRemoteSessions() and RebuildInProcessSessions() to DestroyRemoteCompositorSessions() and DestroyInProcessCompositorSessions(), to better reflect what they actually do: the "rebuilding" part occurs later on. Also update the mega-comment documenting the restart sequence, as it was somewhat outdated. In case a caller of EnsureGPUReady() gets called before the foreground signal arrives (eg in nsBaseWidget::CreateCompositorSession() due to a refresh tick paint), make EnsureGPUReady() launch the GPU process itself if the GPU process is enabled but not yet launched. As a consequence, to avoid launching the GPU process unnecessarily, change a couple callers of EnsureGPUReady() to simply check whether the process is enabled instead. Additionally, guard against a null pointer deref if the compositor has been destroyed when the widget receives a memory pressure event. This is now more likely to occur as there may be a gap between the compositor being destroyed and recreated. Differential Revision: https://phabricator.services.mozilla.com/D139042
2022-02-22 15:59:13 +00:00
// Ensure mProcess is non-null before calling EnsureGPUReady, to avoid
// launching the process if it has not already been launched.
if (!mProcess || NS_FAILED(EnsureGPUReady())) {
return nullptr;
}
return new GPUMemoryReporter();
}
void GPUProcessManager::SetAppInForeground(bool aInForeground) {
if (mAppInForeground == aInForeground) {
return;
}
mAppInForeground = aInForeground;
#if defined(XP_WIN)
SetProcessIsForeground();
#endif
}
#if defined(XP_WIN)
void GPUProcessManager::SetProcessIsForeground() {
NTSTATUS WINAPI NtSetInformationProcess(
IN HANDLE process_handle, IN ULONG info_class,
IN PVOID process_information, IN ULONG information_length);
constexpr unsigned int NtProcessInformationForeground = 25;
static bool alreadyInitialized = false;
static decltype(NtSetInformationProcess)* setInformationProcess = nullptr;
if (!alreadyInitialized) {
alreadyInitialized = true;
nsModuleHandle module(LoadLibrary(L"ntdll.dll"));
if (module) {
setInformationProcess =
(decltype(NtSetInformationProcess)*)GetProcAddress(
module, "NtSetInformationProcess");
}
}
if (MOZ_UNLIKELY(!setInformationProcess)) {
return;
}
unsigned pid = GPUProcessPid();
if (pid <= 0) {
return;
}
// Using the handle from mProcess->GetChildProcessHandle() fails;
// the PROCESS_SET_INFORMATION permission is probably missing.
nsAutoHandle processHandle(
::OpenProcess(PROCESS_SET_INFORMATION, FALSE, pid));
if (!processHandle) {
return;
}
BOOLEAN foreground = mAppInForeground;
setInformationProcess(processHandle, NtProcessInformationForeground,
(PVOID)&foreground, sizeof(foreground));
}
#endif
RefPtr<PGPUChild::TestTriggerMetricsPromise>
GPUProcessManager::TestTriggerMetrics() {
if (!NS_WARN_IF(!mGPUChild)) {
return mGPUChild->SendTestTriggerMetrics();
}
return PGPUChild::TestTriggerMetricsPromise::CreateAndReject(
ipc::ResponseRejectReason::SendError, __func__);
}
} // namespace gfx
} // namespace mozilla