blob: be13a9b0aa064289b1543d10741b9d823dc3d528 [file] [log] [blame]
// Copyright 2012 The Chromium Authors
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "gpu/ipc/service/gpu_channel_manager.h"
#include <algorithm>
#include <memory>
#include <utility>
#include <variant>
#include "base/command_line.h"
#include "base/debug/crash_logging.h"
#include "base/debug/dump_without_crashing.h"
#include "base/feature_list.h"
#include "base/functional/bind.h"
#include "base/location.h"
#include "base/metrics/histogram_macros.h"
#include "base/run_loop.h"
#include "base/strings/stringprintf.h"
#include "base/system/sys_info.h"
#include "base/task/bind_post_task.h"
#include "base/task/single_thread_task_runner.h"
#include "base/trace_event/trace_event.h"
#include "base/trace_event/traced_value.h"
#include "build/build_config.h"
#include "gpu/command_buffer/common/constants.h"
#include "gpu/command_buffer/common/context_creation_attribs.h"
#include "gpu/command_buffer/common/sync_token.h"
#include "gpu/command_buffer/service/feature_info.h"
#include "gpu/command_buffer/service/gl_utils.h"
#include "gpu/command_buffer/service/gpu_tracer.h"
#include "gpu/command_buffer/service/memory_program_cache.h"
#include "gpu/command_buffer/service/passthrough_program_cache.h"
#include "gpu/command_buffer/service/scheduler.h"
#include "gpu/command_buffer/service/sync_point_manager.h"
#include "gpu/config/gpu_crash_keys.h"
#include "gpu/config/gpu_finch_features.h"
#include "gpu/config/gpu_switches.h"
#include "gpu/ipc/common/gpu_client_ids.h"
#include "gpu/ipc/common/memory_stats.h"
#include "gpu/ipc/service/gpu_channel.h"
#include "gpu/ipc/service/gpu_channel_manager_delegate.h"
#include "gpu/ipc/service/gpu_watchdog_thread.h"
#include "third_party/perfetto/include/perfetto/tracing/track.h"
#include "third_party/skia/include/core/SkGraphics.h"
#include "third_party/skia/include/gpu/ganesh/GrDirectContext.h"
#include "third_party/skia/include/gpu/ganesh/GrTypes.h"
#include "ui/gl/gl_bindings.h"
#include "ui/gl/gl_enums.h"
#include "ui/gl/gl_features.h"
#include "ui/gl/gl_share_group.h"
#include "ui/gl/gl_surface_egl.h"
#include "ui/gl/gl_version_info.h"
#include "ui/gl/init/gl_factory.h"
#if BUILDFLAG(USE_DAWN)
#include "gpu/command_buffer/service/dawn_caching_interface.h"
#endif
#if BUILDFLAG(SKIA_USE_DAWN)
#include "gpu/command_buffer/service/dawn_context_provider.h"
#endif
#if BUILDFLAG(IS_WIN)
#include <dxgi1_3.h>
#include "ui/gl/gl_angle_util_win.h"
#endif
#if BUILDFLAG(ENABLE_VULKAN)
#include "gpu/vulkan/vulkan_device_queue.h"
#include "gpu/vulkan/vulkan_fence_helper.h"
#endif
namespace gpu {
namespace {
#if BUILDFLAG(IS_ANDROID)
// Amount of time we expect the GPU to stay powered up without being used.
const int kMaxGpuIdleTimeMs = 40;
// Maximum amount of time we keep pinging the GPU waiting for the client to
// draw.
const int kMaxKeepAliveTimeMs = 200;
#endif
#if BUILDFLAG(IS_WIN)
void TrimD3DResources(const scoped_refptr<SharedContextState>& context_state) {
// Graphics drivers periodically allocate internal memory buffers in
// order to speed up subsequent rendering requests. These memory allocations
// in general lead to increased memory usage by the overall system.
// Calling Trim discards internal memory buffers allocated for the app,
// reducing its memory footprint.
// Calling Trim method does not change the rendering state of the
// graphics device and has no effect on rendering operations.
// There is a brief performance hit when internal buffers are reallocated
// during the first rendering operations after the Trim call, therefore
// apps should only call Trim when going idle for a period of time or during
// low memory conditions.
Microsoft::WRL::ComPtr<ID3D11Device> d3d11_device;
if (context_state) {
d3d11_device = context_state->GetD3D11Device();
}
if (d3d11_device) {
Microsoft::WRL::ComPtr<IDXGIDevice3> dxgi_device;
if (SUCCEEDED(d3d11_device.As(&dxgi_device))) {
dxgi_device->Trim();
}
}
Microsoft::WRL::ComPtr<ID3D11Device> angle_d3d11_device =
gl::QueryD3D11DeviceObjectFromANGLE();
if (angle_d3d11_device && angle_d3d11_device != d3d11_device) {
Microsoft::WRL::ComPtr<IDXGIDevice3> dxgi_device;
if (SUCCEEDED(angle_d3d11_device.As(&dxgi_device))) {
dxgi_device->Trim();
}
}
}
#endif
void GL_APIENTRY CrashReportOnGLErrorDebugCallback(GLenum source,
GLenum type,
GLuint id,
GLenum severity,
GLsizei length,
const GLchar* message,
const GLvoid* user_param) {
if (type == GL_DEBUG_TYPE_ERROR && source == GL_DEBUG_SOURCE_API &&
user_param) {
// Note: log_message cannot contain any user data. The error strings
// generated from ANGLE are all static strings and do not contain user
// information such as shader source code. Be careful if updating the
// contents of this string.
std::string log_message = gl::GLEnums::GetStringEnum(id);
if (message && length > 0) {
log_message += ": " + std::string(message, length);
}
LOG(ERROR) << log_message;
crash_keys::gpu_gl_error_message.Set(log_message);
int* remaining_reports =
const_cast<int*>(static_cast<const int*>(user_param));
if (*remaining_reports > 0) {
base::debug::DumpWithoutCrashing();
(*remaining_reports)--;
}
}
}
void FormatAllocationSourcesForTracing(
base::trace_event::TracedValue* dict,
base::flat_map<GpuPeakMemoryAllocationSource, uint64_t>&
allocation_sources) {
dict->SetInteger("UNKNOWN",
allocation_sources[GpuPeakMemoryAllocationSource::UNKNOWN]);
dict->SetInteger(
"COMMAND_BUFFER",
allocation_sources[GpuPeakMemoryAllocationSource::COMMAND_BUFFER]);
dict->SetInteger(
"SHARED_CONTEXT_STATE",
allocation_sources[GpuPeakMemoryAllocationSource::SHARED_CONTEXT_STATE]);
dict->SetInteger(
"SHARED_IMAGE_STUB",
allocation_sources[GpuPeakMemoryAllocationSource::SHARED_IMAGE_STUB]);
dict->SetInteger("SKIA",
allocation_sources[GpuPeakMemoryAllocationSource::SKIA]);
}
void SetCrashKeyTimeDelta(base::debug::CrashKeyString* key,
base::TimeDelta time_delta) {
auto str = base::StringPrintf(
"%d hours, %d min, %lld sec, %lld ms", time_delta.InHours(),
time_delta.InMinutes() % 60, time_delta.InSeconds() % 60ll,
time_delta.InMilliseconds() % 1000ll);
base::debug::SetCrashKeyString(key, str);
}
} // namespace
GpuChannelManager::GpuPeakMemoryMonitor::GpuPeakMemoryMonitor() = default;
GpuChannelManager::GpuPeakMemoryMonitor::~GpuPeakMemoryMonitor() = default;
base::flat_map<GpuPeakMemoryAllocationSource, uint64_t>
GpuChannelManager::GpuPeakMemoryMonitor::GetPeakMemoryUsage(
uint32_t sequence_num,
uint64_t* out_peak_memory) {
base::AutoLock auto_lock(peak_mem_lock_);
auto sequence = sequence_trackers_.find(sequence_num);
base::flat_map<GpuPeakMemoryAllocationSource, uint64_t> allocation_per_source;
*out_peak_memory = 0u;
if (sequence != sequence_trackers_.end()) {
*out_peak_memory = sequence->second.total_memory_;
allocation_per_source = sequence->second.peak_memory_per_source_;
}
return allocation_per_source;
}
// Runs on GpuMain thread, called from GpuServiceImpl
void GpuChannelManager::GpuPeakMemoryMonitor::StartGpuMemoryTracking(
uint32_t sequence_num) {
base::AutoLock auto_lock(peak_mem_lock_);
sequence_trackers_.emplace(
sequence_num,
SequenceTracker(current_memory_, current_memory_per_source_));
TRACE_EVENT_BEGIN("gpu", "PeakMemoryTracking", perfetto::Track(sequence_num),
"start", current_memory_, "start_sources",
StartTrackingTracedValue());
}
// Runs on GpuMain thread, called from GpuServiceImpl
void GpuChannelManager::GpuPeakMemoryMonitor::StopGpuMemoryTracking(
uint32_t sequence_num) {
base::AutoLock auto_lock(peak_mem_lock_);
auto sequence = sequence_trackers_.find(sequence_num);
if (sequence != sequence_trackers_.end()) {
TRACE_EVENT_END("gpu", perfetto::Track(sequence_num), "peak",
sequence->second.total_memory_, "end_sources",
StopTrackingTracedValue(sequence->second));
sequence_trackers_.erase(sequence);
}
}
GpuChannelManager::GpuPeakMemoryMonitor::SequenceTracker::SequenceTracker(
uint64_t current_memory,
base::flat_map<GpuPeakMemoryAllocationSource, uint64_t>
current_memory_per_source)
: initial_memory_(current_memory),
total_memory_(current_memory),
initial_memory_per_source_(current_memory_per_source),
peak_memory_per_source_(std::move(current_memory_per_source)) {}
GpuChannelManager::GpuPeakMemoryMonitor::SequenceTracker::SequenceTracker(
const SequenceTracker& other) = default;
GpuChannelManager::GpuPeakMemoryMonitor::SequenceTracker::~SequenceTracker() =
default;
std::unique_ptr<base::trace_event::TracedValue>
GpuChannelManager::GpuPeakMemoryMonitor::StartTrackingTracedValue() {
peak_mem_lock_.AssertAcquired();
auto dict = std::make_unique<base::trace_event::TracedValue>();
FormatAllocationSourcesForTracing(dict.get(), current_memory_per_source_);
return dict;
}
std::unique_ptr<base::trace_event::TracedValue>
GpuChannelManager::GpuPeakMemoryMonitor::StopTrackingTracedValue(
SequenceTracker& sequence) {
peak_mem_lock_.AssertAcquired();
auto dict = std::make_unique<base::trace_event::TracedValue>();
dict->BeginDictionary("source_totals");
FormatAllocationSourcesForTracing(dict.get(),
sequence.peak_memory_per_source_);
dict->EndDictionary();
dict->BeginDictionary("difference");
int total_diff = sequence.total_memory_ - sequence.initial_memory_;
dict->SetInteger("TOTAL", total_diff);
dict->EndDictionary();
dict->BeginDictionary("source_difference");
for (auto it : sequence.peak_memory_per_source_) {
int diff = (it.second - sequence.initial_memory_per_source_[it.first]);
switch (it.first) {
case GpuPeakMemoryAllocationSource::UNKNOWN:
dict->SetInteger("UNKNOWN", diff);
break;
case GpuPeakMemoryAllocationSource::COMMAND_BUFFER:
dict->SetInteger("COMMAND_BUFFER", diff);
break;
case GpuPeakMemoryAllocationSource::SHARED_CONTEXT_STATE:
dict->SetInteger("SHARED_CONTEXT_STATE", diff);
break;
case GpuPeakMemoryAllocationSource::SHARED_IMAGE_STUB:
dict->SetInteger("SHARED_IMAGE_STUB", diff);
break;
case GpuPeakMemoryAllocationSource::SKIA:
dict->SetInteger("SKIA", diff);
break;
}
}
dict->EndDictionary();
return dict;
}
void GpuChannelManager::GpuPeakMemoryMonitor::OnMemoryAllocatedChange(
CommandBufferId id,
uint64_t old_size,
uint64_t new_size,
GpuPeakMemoryAllocationSource source) {
base::AutoLock auto_lock(peak_mem_lock_);
uint64_t diff = new_size - old_size;
current_memory_ += diff;
current_memory_per_source_[source] += diff;
if (old_size < new_size) {
// When memory has increased, iterate over the sequences to update their
// peak.
// TODO(jonross): This should be fine if we typically have 1-2 sequences.
// However if that grows we may end up iterating many times are memory
// approaches peak. If that is the case we should track a
// |peak_since_last_sequence_update_| on the the memory changes. Then only
// update the sequences with a new one is added, or the peak is requested.
for (auto& seq : sequence_trackers_) {
if (current_memory_ > seq.second.total_memory_) {
seq.second.total_memory_ = current_memory_;
for (auto& sequence : sequence_trackers_) {
TRACE_EVENT_ASYNC_STEP_INTO1("gpu", "PeakMemoryTracking",
sequence.first, "Peak", "peak",
current_memory_);
}
for (auto& memory_per_source : current_memory_per_source_) {
seq.second.peak_memory_per_source_[memory_per_source.first] =
memory_per_source.second;
}
}
}
}
}
GpuChannelManager::GpuChannelManager(
const GpuPreferences& gpu_preferences,
GpuChannelManagerDelegate* delegate,
GpuWatchdogThread* watchdog,
scoped_refptr<base::SingleThreadTaskRunner> task_runner,
scoped_refptr<base::SingleThreadTaskRunner> io_task_runner,
Scheduler* scheduler,
SyncPointManager* sync_point_manager,
SharedImageManager* shared_image_manager,
const GpuFeatureInfo& gpu_feature_info,
GpuProcessShmCount* use_shader_cache_shm_count,
scoped_refptr<gl::GLSurface> default_offscreen_surface,
ImageDecodeAcceleratorWorker* image_decode_accelerator_worker,
viz::VulkanContextProvider* vulkan_context_provider,
viz::MetalContextProvider* metal_context_provider,
DawnContextProvider* dawn_context_provider,
webgpu::DawnCachingInterfaceFactory* dawn_caching_interface_factory,
const SharedContextState::GrContextOptionsProvider*
gr_context_options_provider)
: task_runner_(task_runner),
io_task_runner_(io_task_runner),
gpu_preferences_(gpu_preferences),
gpu_driver_bug_workarounds_(
gpu_feature_info.enabled_gpu_driver_bug_workarounds),
delegate_(delegate),
watchdog_(watchdog),
share_group_(new gl::GLShareGroup()),
scheduler_(scheduler),
sync_point_manager_(sync_point_manager),
shared_image_manager_(shared_image_manager),
shader_translator_cache_(gpu_preferences_),
default_offscreen_surface_(std::move(default_offscreen_surface)),
gpu_feature_info_(gpu_feature_info),
discardable_manager_(gpu_preferences_),
passthrough_discardable_manager_(gpu_preferences_),
image_decode_accelerator_worker_(image_decode_accelerator_worker),
use_shader_cache_shm_count_(use_shader_cache_shm_count),
memory_pressure_listener_(
FROM_HERE,
base::MemoryPressureListenerTag::kGpuChannelManager,
base::BindRepeating(&GpuChannelManager::HandleMemoryPressure,
base::Unretained(this))),
dawn_caching_interface_factory_(dawn_caching_interface_factory),
vulkan_context_provider_(vulkan_context_provider),
metal_context_provider_(metal_context_provider),
dawn_context_provider_(dawn_context_provider),
peak_memory_monitor_(base::MakeRefCounted<GpuPeakMemoryMonitor>()),
gr_context_options_provider_(gr_context_options_provider) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
DCHECK(task_runner->BelongsToCurrentThread());
DCHECK(io_task_runner);
DCHECK(scheduler);
const bool enable_gr_shader_cache =
(gpu_feature_info_
.status_values[GPU_FEATURE_TYPE_GPU_TILE_RASTERIZATION] ==
gpu::kGpuFeatureStatusEnabled) &&
!gpu_preferences_.disable_gpu_shader_disk_cache;
UMA_HISTOGRAM_BOOLEAN("Gpu.GrShaderCacheEnabled", enable_gr_shader_cache);
if (enable_gr_shader_cache) {
size_t gr_shader_cache_size = gpu_preferences.gpu_program_cache_size;
if (base::FeatureList::IsEnabled(features::kANGLEPerContextBlobCache)) {
// When ANGLE shares the shader cache with Skia, double the size of the
// cache so that there is room for both APIs to cache together.
gr_shader_cache_size *= 2;
}
gr_shader_cache_.emplace(gr_shader_cache_size, this);
gr_shader_cache_->CacheClientIdOnDisk(gpu::kDisplayCompositorClientId);
}
}
GpuChannelManager::~GpuChannelManager() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
// Clear |gpu_channels_| first to prevent reentrancy problems from GpuChannel
// destructor.
auto gpu_channels = std::move(gpu_channels_);
gpu_channels_.clear();
gpu_channels.clear();
if (default_offscreen_surface_.get()) {
default_offscreen_surface_->Destroy();
default_offscreen_surface_ = nullptr;
}
// Try to make the context current so that GPU resources can be destroyed
// correctly.
if (shared_context_state_)
shared_context_state_->MakeCurrent(nullptr);
}
gles2::Outputter* GpuChannelManager::outputter() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
if (!outputter_) {
outputter_ =
std::make_unique<gles2::TraceOutputter>("GpuChannelManager Trace");
}
return outputter_.get();
}
gles2::ProgramCache* GpuChannelManager::program_cache() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
if (!program_cache_.get()) {
const GpuDriverBugWorkarounds& workarounds = gpu_driver_bug_workarounds_;
bool disable_disk_cache =
gpu_preferences_.disable_gpu_shader_disk_cache ||
workarounds.disable_program_disk_cache;
// Use the EGL blob cache extension for the passthrough decoder.
if (use_passthrough_cmd_decoder()) {
program_cache_ = std::make_unique<gles2::PassthroughProgramCache>(
gpu_preferences_.gpu_program_cache_size, disable_disk_cache);
} else {
program_cache_ = std::make_unique<gles2::MemoryProgramCache>(
gpu_preferences_.gpu_program_cache_size, disable_disk_cache,
workarounds.disable_program_caching_for_transform_feedback,
use_shader_cache_shm_count_);
}
}
return program_cache_.get();
}
void GpuChannelManager::RemoveChannel(int client_id) {
// Using sequence enforcement to avoid further wrong-thread accesses
// in production.
CHECK(task_runner_->RunsTasksInCurrentSequence());
auto it = gpu_channels_.find(client_id);
if (it == gpu_channels_.end())
return;
delegate_->DidDestroyChannel(client_id);
// Erase the |gpu_channels_| entry before destroying the GpuChannel object to
// avoid reentrancy problems from the GpuChannel destructor.
std::unique_ptr<GpuChannel> channel = std::move(it->second);
gpu_channels_.erase(it);
channel.reset();
if (gpu_channels_.empty()) {
delegate_->DidDestroyAllChannels();
}
}
GpuChannel* GpuChannelManager::LookupChannel(int32_t client_id) const {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
const auto& it = gpu_channels_.find(client_id);
return it != gpu_channels_.end() ? it->second.get() : nullptr;
}
GpuChannel* GpuChannelManager::EstablishChannel(
const base::UnguessableToken& channel_token,
int client_id,
uint64_t client_tracing_id,
bool is_gpu_host,
const gfx::GpuExtraInfo& gpu_extra_info) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
// Remove existing GPU channel with same client id before creating
// new GPU channel. if not, new SyncPointClientState in SyncPointManager
// will be destroyed when existing GPU channel is destroyed.
// We can't call RemoveChannel() because it will clear GpuDiskCache
// with the client id.
auto it = gpu_channels_.find(client_id);
if (it != gpu_channels_.end()) {
std::unique_ptr<GpuChannel> channel = std::move(it->second);
gpu_channels_.erase(it);
channel.reset();
}
std::unique_ptr<GpuChannel> gpu_channel = GpuChannel::Create(
this, channel_token, scheduler_, sync_point_manager_, share_group_,
task_runner_, io_task_runner_, client_id, client_tracing_id, is_gpu_host,
image_decode_accelerator_worker_, gpu_extra_info);
if (!gpu_channel)
return nullptr;
GpuChannel* gpu_channel_ptr = gpu_channel.get();
gpu_channels_[client_id] = std::move(gpu_channel);
return gpu_channel_ptr;
}
void GpuChannelManager::SetChannelClientPid(int client_id,
base::ProcessId client_pid) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
GpuChannel* gpu_channel = LookupChannel(client_id);
if (gpu_channel) {
// TODO(rockot): It's possible to receive different PIDs for the same
// GpuChannel because some clients may reuse a client ID. For example, if a
// Content renderer crashes and restarts, the new process will use the same
// GPU client ID that the crashed process used. In such cases, this
// SetChannelClientPid (which comes from the GPU host, not the client
// process) may arrive late with the crashed process PID, followed shortly
// thereafter by the current PID of the client.
//
// For a short window of time this means a GpuChannel may have a stale PID
// value. It's not a serious issue since the PID is only informational and
// not required for security or application correctness, but we should still
// address it. One option is to introduce a separate host-controlled
// interface that is paired with the GpuChannel during Establish, which the
// host can then use to asynchronously push down a PID for the specific
// channel instance.
gpu_channel->set_client_pid(client_pid);
}
}
void GpuChannelManager::SetChannelDiskCacheHandle(
int client_id,
const gpu::GpuDiskCacheHandle& handle) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
GpuChannel* gpu_channel = LookupChannel(client_id);
if (gpu_channel) {
gpu_channel->RegisterCacheHandle(handle);
}
// Record the client id for the shader specific cache.
if (gr_shader_cache_ &&
gpu::GetHandleType(handle) == gpu::GpuDiskCacheType::kGlShaders) {
gr_shader_cache_->CacheClientIdOnDisk(client_id);
}
}
void GpuChannelManager::OnDiskCacheHandleDestoyed(
const gpu::GpuDiskCacheHandle& handle) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
switch (gpu::GetHandleType(handle)) {
case gpu::GpuDiskCacheType::kGlShaders: {
// Currently there isn't any handling necessary for when the disk cache is
// destroyed for the shader cache because it consists of just 2 massive
// caches that are long-living and shared across all channels (i.e.
// unfortunately there is currently no access partitioning for it w.r.t
// different handles).
break;
}
case gpu::GpuDiskCacheType::kDawnWebGPU:
case gpu::GpuDiskCacheType::kDawnGraphite: {
#if BUILDFLAG(USE_DAWN)
dawn_caching_interface_factory()->ReleaseHandle(handle);
#endif
break;
}
}
}
void GpuChannelManager::PopulateCache(const gpu::GpuDiskCacheHandle& handle,
const std::string& key,
const std::string& data) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
switch (gpu::GetHandleType(handle)) {
case gpu::GpuDiskCacheType::kGlShaders: {
auto gl_shader_handle = std::get<gpu::GpuDiskCacheGlShaderHandle>(handle);
if (gl_shader_handle == kGrShaderGpuDiskCacheHandle) {
if (gr_shader_cache_)
gr_shader_cache_->PopulateCache(key, data);
return;
}
if (program_cache())
program_cache()->LoadProgram(key, data);
break;
}
case gpu::GpuDiskCacheType::kDawnWebGPU:
case gpu::GpuDiskCacheType::kDawnGraphite: {
#if BUILDFLAG(USE_DAWN) || BUILDFLAG(SKIA_USE_DAWN)
TRACE_EVENT1("gpu", "GpuChannelManager::PopulateCacheDawn", "handle_type",
gpu::GetHandleType(handle));
std::unique_ptr<gpu::webgpu::DawnCachingInterface>
dawn_caching_interface =
dawn_caching_interface_factory()->CreateInstance(handle);
if (!dawn_caching_interface) {
return;
}
dawn_caching_interface->StoreData(key.data(), key.size(), data.data(),
data.size());
#endif
break;
}
}
}
void GpuChannelManager::LoseAllContexts() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
discardable_manager_.OnContextLost();
passthrough_discardable_manager_.OnContextLost();
share_group_ = base::MakeRefCounted<gl::GLShareGroup>();
for (auto& kv : gpu_channels_) {
kv.second->MarkAllContextsLost();
}
task_runner_->PostTask(FROM_HERE,
base::BindOnce(&GpuChannelManager::DestroyAllChannels,
weak_factory_.GetWeakPtr()));
if (shared_context_state_) {
shared_context_state_->MarkContextLost();
shared_context_state_.reset();
}
}
SharedContextState::ContextLostCallback
GpuChannelManager::GetContextLostCallback() {
return base::BindPostTask(
task_runner_,
base::BindOnce(&GpuChannelManager::OnContextLost,
weak_factory_.GetWeakPtr(), context_lost_count_ + 1));
}
void GpuChannelManager::DestroyAllChannels() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
// Clear |gpu_channels_| first to prevent reentrancy problems from GpuChannel
// destructor.
auto gpu_channels = std::move(gpu_channels_);
gpu_channels_.clear();
gpu_channels.clear();
}
void GpuChannelManager::GetVideoMemoryUsageStats(
VideoMemoryUsageStats* video_memory_usage_stats) const {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
// For each context group, assign its memory usage to its PID
video_memory_usage_stats->process_map.clear();
uint64_t total_size = 0;
for (const auto& entry : gpu_channels_) {
const GpuChannel* channel = entry.second.get();
if (channel->client_pid() == base::kNullProcessId)
continue;
uint64_t size = channel->GetMemoryUsage();
total_size += size;
video_memory_usage_stats->process_map[channel->client_pid()].video_memory +=
size;
}
// Add the SharedContextState memory from the CrGpuMain thread to the total.
// CompositorGpuThread::AddVideoMemoryUsageStatsOnCompositorGpu() adds the
// SharedContextState memory from CompositorGpuMain if DrDC is enabled.
if (shared_context_state_ && !shared_context_state_->context_lost()) {
total_size += shared_context_state_->GetMemoryUsage();
}
// Assign the total across all processes in the GPU process
video_memory_usage_stats->process_map[base::GetCurrentProcId()].video_memory =
total_size;
video_memory_usage_stats->process_map[base::GetCurrentProcId()]
.has_duplicates = true;
video_memory_usage_stats->bytes_allocated = total_size;
}
void GpuChannelManager::StartPeakMemoryMonitor(uint32_t sequence_num) {
peak_memory_monitor_->StartGpuMemoryTracking(sequence_num);
}
base::flat_map<GpuPeakMemoryAllocationSource, uint64_t>
GpuChannelManager::GetPeakMemoryUsage(uint32_t sequence_num,
uint64_t* out_peak_memory) {
auto allocation_per_source =
peak_memory_monitor_->GetPeakMemoryUsage(sequence_num, out_peak_memory);
peak_memory_monitor_->StopGpuMemoryTracking(sequence_num);
return allocation_per_source;
}
#if BUILDFLAG(IS_ANDROID)
void GpuChannelManager::DidAccessGpu() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
last_gpu_access_time_ = base::TimeTicks::Now();
}
void GpuChannelManager::WakeUpGpu() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
begin_wake_up_time_ = base::TimeTicks::Now();
ScheduleWakeUpGpu();
}
void GpuChannelManager::ScheduleWakeUpGpu() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
base::TimeTicks now = base::TimeTicks::Now();
TRACE_EVENT2("gpu", "GpuChannelManager::ScheduleWakeUp", "idle_time",
(now - last_gpu_access_time_).InMilliseconds(),
"keep_awake_time", (now - begin_wake_up_time_).InMilliseconds());
if (now - last_gpu_access_time_ < base::Milliseconds(kMaxGpuIdleTimeMs))
return;
if (now - begin_wake_up_time_ > base::Milliseconds(kMaxKeepAliveTimeMs))
return;
DoWakeUpGpu();
base::SingleThreadTaskRunner::GetCurrentDefault()->PostDelayedTask(
FROM_HERE,
base::BindOnce(&GpuChannelManager::ScheduleWakeUpGpu,
weak_factory_.GetWeakPtr()),
base::Milliseconds(kMaxGpuIdleTimeMs));
}
void GpuChannelManager::DoWakeUpGpu() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
const CommandBufferStub* stub = nullptr;
for (const auto& kv : gpu_channels_) {
const GpuChannel* channel = kv.second.get();
const CommandBufferStub* stub_candidate = channel->GetOneStub();
if (stub_candidate) {
DCHECK(stub_candidate->decoder_context());
// With Vulkan, Dawn, etc, RasterDecoders don't use GL.
if (stub_candidate->decoder_context()->GetGLContext()) {
stub = stub_candidate;
break;
}
}
}
if (!stub || !stub->decoder_context()->MakeCurrent())
return;
glFinish();
DidAccessGpu();
}
void GpuChannelManager::OnBackgroundCleanup() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
// Delete all the GL contexts when the channel does not use WebGL and Chrome
// goes to background on low-end devices.
std::vector<int> channels_to_clear;
for (auto& kv : gpu_channels_) {
// Stateful contexts (e.g. WebGL and WebGPU) support context lost
// notifications, but for now, skip those.
if (kv.second->HasActiveStatefulContext()) {
continue;
}
channels_to_clear.push_back(kv.first);
kv.second->MarkAllContextsLost();
}
for (int channel : channels_to_clear)
RemoveChannel(channel);
if (program_cache_)
program_cache_->Trim(0u);
if (shared_context_state_) {
shared_context_state_->MarkContextLost();
shared_context_state_.reset();
}
SkGraphics::PurgeAllCaches();
}
#endif // BUILDFLAG(IS_ANDROID)
void GpuChannelManager::OnApplicationBackgrounded() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
if (shared_context_state_) {
shared_context_state_->PurgeMemory(
base::MemoryPressureListener::MemoryPressureLevel::
MEMORY_PRESSURE_LEVEL_CRITICAL);
}
// Release all skia caching when the application is backgrounded.
SkGraphics::PurgeAllCaches();
// At that point, no frames are going to be produced. Make sure that
// e.g. pending SharedImage deletions happens promptly.
PerformImmediateCleanup();
application_backgrounded_ = true;
}
void GpuChannelManager::OnApplicationForegounded() {
application_backgrounded_ = false;
}
void GpuChannelManager::PerformImmediateCleanup() {
TRACE_EVENT0("viz", __PRETTY_FUNCTION__);
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
if (!shared_context_state_) {
return;
}
#if BUILDFLAG(ENABLE_VULKAN)
if (shared_context_state_->GrContextIsVulkan()) {
// TODO(lizeb): Also perform this on GL devices.
if (auto* context = shared_context_state_->gr_context()) {
context->flushAndSubmit(GrSyncCpu::kYes);
}
DCHECK(vulkan_context_provider_);
auto* fence_helper =
vulkan_context_provider_->GetDeviceQueue()->GetFenceHelper();
// PerformImmediateCleanup will ensure that all GPU work that was submitted
// before is finished before releasing resoucres, but skia might have
// recorded and not yet submitted work that reference them, so this must be
// called after GrContext::submit (or flushAndSubmit).
fence_helper->PerformImmediateCleanup();
}
#endif
}
void GpuChannelManager::HandleMemoryPressure(
base::MemoryPressureListener::MemoryPressureLevel memory_pressure_level) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
if (program_cache_)
program_cache_->HandleMemoryPressure(memory_pressure_level);
// These caches require a current context for cleanup.
if (shared_context_state_ &&
shared_context_state_->MakeCurrent(nullptr, true /* needs_gl */)) {
discardable_manager_.HandleMemoryPressure(memory_pressure_level);
passthrough_discardable_manager_.HandleMemoryPressure(
memory_pressure_level);
shared_context_state_->PurgeMemory(memory_pressure_level);
}
if (gr_shader_cache_) {
gr_shader_cache_->PurgeMemory(memory_pressure_level);
}
#if BUILDFLAG(USE_DAWN) || BUILDFLAG(SKIA_USE_DAWN)
if (dawn_caching_interface_factory()) {
dawn_caching_interface_factory()->PurgeMemory(memory_pressure_level);
}
#endif // BUILDFLAG(USE_DAWN) || BUILDFLAG(SKIA_USE_DAWN)
#if BUILDFLAG(IS_WIN)
TrimD3DResources(shared_context_state_);
#endif // BUILDFLAG(IS_WIN)
}
scoped_refptr<SharedContextState> GpuChannelManager::GetSharedContextState(
ContextResult* result) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
if (shared_context_state_ && !shared_context_state_->context_lost()) {
*result = ContextResult::kSuccess;
return shared_context_state_;
}
// Temporarily check the ANGLE metal experiment early in GPU process
// initialization. This will help us determine why users sometimes do not
// check the feature on subsequent runs of Chrome. crbug.com/1423439
[[maybe_unused]] bool default_angle_metal =
base::FeatureList::IsEnabled(features::kDefaultANGLEMetal);
scoped_refptr<gl::GLSurface> surface = default_offscreen_surface();
bool use_virtualized_gl_contexts = false;
#if BUILDFLAG(IS_MAC)
// Virtualize GpuPreference::kLowPower contexts by default on OS X to prevent
// performance regressions when enabling FCM.
// http://crbug.com/180463
use_virtualized_gl_contexts = true;
#endif
use_virtualized_gl_contexts |=
gpu_driver_bug_workarounds_.use_virtualized_gl_contexts;
bool enable_angle_validation = features::IsANGLEValidationEnabled();
scoped_refptr<gl::GLShareGroup> share_group;
bool use_passthrough_decoder = use_passthrough_cmd_decoder();
if (use_passthrough_decoder) {
share_group = new gl::GLShareGroup();
// Virtualized contexts don't work with passthrough command decoder.
// See https://crbug.com/914976
use_virtualized_gl_contexts = false;
} else {
share_group = share_group_;
}
scoped_refptr<gl::GLContext> context =
use_virtualized_gl_contexts ? share_group->shared_context() : nullptr;
if (context && (!context->MakeCurrent(surface.get()) ||
context->CheckStickyGraphicsResetStatus() != GL_NO_ERROR)) {
context = nullptr;
}
if (!context) {
gl::GLContextAttribs attribs =
gles2::GenerateGLContextAttribsForCompositor(use_passthrough_decoder);
// Disable robust resource initialization for raster decoder and compositor.
// TODO(crbug.com/40174948): disable robust_resource_initialization for
// SwANGLE.
if (gl::GLSurfaceEGL::GetGLDisplayEGL()->GetDisplayType() !=
gl::ANGLE_SWIFTSHADER) {
attribs.robust_resource_initialization = false;
}
attribs.can_skip_validation = !enable_angle_validation;
context =
gl::init::CreateGLContext(share_group.get(), surface.get(), attribs);
if (!context) {
// TODO(piman): This might not be fatal, we could recurse into
// CreateGLContext to get more info, tho it should be exceedingly
// rare and may not be recoverable anyway.
LOG(ERROR) << "ContextResult::kFatalFailure: "
"Failed to create shared context for virtualization.";
*result = ContextResult::kFatalFailure;
return nullptr;
}
// Ensure that context creation did not lose track of the intended share
// group.
DCHECK(context->share_group() == share_group.get());
gpu_feature_info_.ApplyToGLContext(context.get());
if (use_virtualized_gl_contexts)
share_group->SetSharedContext(context.get());
}
// This should be either:
// (1) a non-virtual GL context, or
// (2) a mock/stub context.
DCHECK(context->GetHandle() ||
gl::GetGLImplementation() == gl::kGLImplementationMockGL ||
gl::GetGLImplementation() == gl::kGLImplementationStubGL);
if (!context->MakeCurrent(surface.get())) {
LOG(ERROR)
<< "ContextResult::kTransientFailure, failed to make context current";
*result = ContextResult::kTransientFailure;
return nullptr;
}
// TODO(penghuang): https://crbug.com/899735 Handle device lost for Vulkan.
auto shared_context_state = base::MakeRefCounted<SharedContextState>(
std::move(share_group), std::move(surface), std::move(context),
use_virtualized_gl_contexts,
base::BindOnce(&GpuChannelManager::OnContextLost, base::Unretained(this),
context_lost_count_ + 1),
gpu_preferences_.gr_context_type, vulkan_context_provider_,
metal_context_provider_, dawn_context_provider_, peak_memory_monitor_,
/*direct_rendering_display_compositor_enabled=*/
features::IsDrDcEnabled(gpu_feature_info_),
/*created_on_compositor_gpu_thread=*/false, gr_context_options_provider_);
// Initialize GL context, so Vulkan and GL interop can work properly.
auto feature_info = base::MakeRefCounted<gles2::FeatureInfo>(
gpu_driver_bug_workarounds(), gpu_feature_info());
if (!shared_context_state->InitializeGL(gpu_preferences_,
feature_info.get())) {
LOG(ERROR) << "ContextResult::kFatalFailure: Failed to Initialize GL for "
" SharedContextState";
*result = ContextResult::kFatalFailure;
return nullptr;
}
// Log crash reports when GL errors are generated.
if (gl::GetGLImplementation() == gl::kGLImplementationEGLANGLE &&
enable_angle_validation && feature_info->feature_flags().khr_debug) {
// Limit the total number of gl error crash reports to 1 per GPU
// process.
static int remaining_gl_error_reports = 1;
gles2::InitializeGLDebugLogging(false, CrashReportOnGLErrorDebugCallback,
&remaining_gl_error_reports);
}
if (!shared_context_state->InitializeSkia(
gpu_preferences_, gpu_driver_bug_workarounds_, gr_shader_cache(),
use_shader_cache_shm_count_, watchdog_)) {
LOG(ERROR) << "ContextResult::kFatalFailure: Failed to initialize Skia for "
"SharedContextState";
*result = ContextResult::kFatalFailure;
return nullptr;
}
shared_context_state_ = std::move(shared_context_state);
*result = ContextResult::kSuccess;
return shared_context_state_;
}
void GpuChannelManager::OnContextLost(
int context_lost_count,
bool synthetic_loss,
error::ContextLostReason context_lost_reason) {
if (context_lost_count < 0)
context_lost_count = context_lost_count_ + 1;
// Because of the DrDC, we may receive context loss from the GPU main and
// thee DrDC thread. If a context loss happens on the GPU main thread first,
// a task will be post to the DrDC thread to trigger the context loss on
// the DrDC thread, and then the DrDC will report context loss to the GPU main
// thread again. So we use the |context_lost_count| to help us to ignore
// context loss which has been handled.
if (context_lost_count <= context_lost_count_)
return;
DCHECK_EQ(context_lost_count, context_lost_count_ + 1);
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
// ANGLE doesn't support recovering from context lost very well.
bool force_restart = use_passthrough_cmd_decoder();
// Add crash keys for context lost count and time.
static auto* const lost_count_crash_key = base::debug::AllocateCrashKeyString(
"context-lost-count", base::debug::CrashKeySize::Size32);
// The context lost time since creation of |GpuChannelManager|.
static auto* const lost_time_crash_key = base::debug::AllocateCrashKeyString(
"context-lost-time", base::debug::CrashKeySize::Size64);
// The context lost interval since last context lost event.
static auto* const lost_interval_crash_key =
base::debug::AllocateCrashKeyString("context-lost-interval",
base::debug::CrashKeySize::Size64);
base::debug::SetCrashKeyString(
lost_count_crash_key, base::StringPrintf("%d", ++context_lost_count_));
auto lost_time = base::TimeTicks::Now() - creation_time_;
SetCrashKeyTimeDelta(lost_time_crash_key, lost_time);
// If context lost 5 times, restart the GPU process.
force_restart |= context_lost_count_ >= 5;
if (!context_lost_time_.is_zero()) {
auto interval = lost_time - context_lost_time_;
SetCrashKeyTimeDelta(lost_interval_crash_key, interval);
// If context lost again in 5 seconds, restart the GPU process.
force_restart |= (interval <= base::Seconds(5));
}
context_lost_time_ = lost_time;
bool is_gl = gpu_preferences_.gr_context_type == GrContextType::kGL;
if (!force_restart && synthetic_loss && is_gl)
return;
// Lose all other contexts.
if (gl::GLContext::LosesAllContextsOnContextLost() ||
(shared_context_state_ &&
shared_context_state_->use_virtualized_gl_contexts())) {
delegate_->LoseAllContexts();
}
// Work around issues with recovery by allowing a new GPU process to launch.
if (force_restart || gpu_driver_bug_workarounds_.exit_on_context_lost ||
(shared_context_state_ && !shared_context_state_->GrContextIsGL())) {
delegate_->MaybeExitOnContextLost(context_lost_reason);
}
}
void GpuChannelManager::ScheduleGrContextCleanup() {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
if (shared_context_state_) {
shared_context_state_->ScheduleSkiaCleanup();
}
}
void GpuChannelManager::StoreShader(const std::string& key,
const std::string& shader) {
delegate_->StoreBlobToDisk(kGrShaderGpuDiskCacheHandle, key, shader);
}
void GpuChannelManager::SetImageDecodeAcceleratorWorkerForTesting(
ImageDecodeAcceleratorWorker* worker) {
DCHECK_CALLED_ON_VALID_THREAD(thread_checker_);
DCHECK(gpu_channels_.empty());
image_decode_accelerator_worker_ = worker;
}
} // namespace gpu