GLInProcessContext: support async flushes and dedicated GPU thread

gpu/command_buffer/service/in_process_command_buffer.cc

Index: gpu/command_buffer/service/in_process_command_buffer.cc
diff --git a/gpu/command_buffer/service/in_process_command_buffer.cc b/gpu/command_buffer/service/in_process_command_buffer.cc
new file mode 100644
index 0000000000000000000000000000000000000000..e78de08794110ec8a221ebcb025f39a7e085a532
--- /dev/null
+++ b/gpu/command_buffer/service/in_process_command_buffer.cc
@@ -0,0 +1,546 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "gpu/command_buffer/service/in_process_command_buffer.h"
+
+#include <queue>
+#include <utility>
+#include <vector>
+
+#include <GLES2/gl2.h>
+#ifndef GL_GLEXT_PROTOTYPES
+#define GL_GLEXT_PROTOTYPES 1
+#endif
+#include <GLES2/gl2ext.h>
+#include <GLES2/gl2extchromium.h>
+
+#include "base/bind.h"
+#include "base/bind_helpers.h"
+#include "base/lazy_instance.h"
+#include "base/logging.h"
+#include "base/memory/scoped_ptr.h"
+#include "base/memory/weak_ptr.h"
+#include "base/message_loop/message_loop.h"
+#include "base/message_loop/message_loop_proxy.h"
+#include "base/synchronization/waitable_event.h"
+#include "base/threading/non_thread_safe.h"
+#include "base/threading/thread.h"
+#include "gpu/command_buffer/common/id_allocator.h"
+#include "gpu/command_buffer/service/command_buffer_service.h"
+#include "gpu/command_buffer/service/context_group.h"
+#include "gpu/command_buffer/service/gl_context_virtual.h"
+#include "gpu/command_buffer/service/gpu_scheduler.h"
+#include "gpu/command_buffer/service/image_manager.h"
+#include "gpu/command_buffer/service/transfer_buffer_manager.h"
+#include "ui/gfx/size.h"
+#include "ui/gl/gl_context.h"
+#include "ui/gl/gl_image.h"
+#include "ui/gl/gl_share_group.h"
+#include "ui/gl/gl_surface.h"
+
+namespace gpu {
+
+namespace {
+
+static base::LazyInstance<std::set<InProcessCommandBuffer*> >
+    g_all_shared_contexts = LAZY_INSTANCE_INITIALIZER;
+
+static bool g_use_virtualized_gl_context = false;
+
+template <typename T>
+static void RunTaskWithResult(base::Callback<T(void)> task,
+                              T* result,
+                              base::WaitableEvent* completion) {
+  *result = task.Run();
+  completion->Signal();
+}
+
+class GpuCommandQueue {
+ public:
+  GpuCommandQueue();
+  virtual ~GpuCommandQueue();
+
+  void QueueTask(const base::Closure& task);
+  void RunTasks();
+
+ private:
+  base::Lock tasks_lock_;
+  std::queue<base::Closure> tasks_;
+
+  DISALLOW_COPY_AND_ASSIGN(GpuCommandQueue);
+};
+
+GpuCommandQueue::GpuCommandQueue() {
+}
+
+GpuCommandQueue::~GpuCommandQueue() {
+  base::AutoLock lock(tasks_lock_);
+  DCHECK(tasks_.empty());
+}
+
+void GpuCommandQueue::QueueTask(const base::Closure& task) {
+  base::AutoLock lock(tasks_lock_);
+  tasks_.push(task);
+}
+
+void GpuCommandQueue::RunTasks() {
+  size_t num_tasks;
+  {
+    base::AutoLock lock(tasks_lock_);
+    num_tasks = tasks_.size();
+  }
+
+  while (num_tasks) {
+    base::Closure task;
+    {
+      base::AutoLock lock(tasks_lock_);
+      task = tasks_.front();
+      tasks_.pop();
+      num_tasks = tasks_.size();
+    }
+
+    task.Run();
+  }
+}
+
+static base::LazyInstance<base::Closure> g_schedule_work_callback =
+    LAZY_INSTANCE_INITIALIZER;
+
+class GpuInProcessThread : public base::Thread {
+ public:
+  GpuInProcessThread();
+  virtual ~GpuInProcessThread();
+
+ private:
+  DISALLOW_COPY_AND_ASSIGN(GpuInProcessThread);
+};
+
+GpuInProcessThread::GpuInProcessThread() : base::Thread("GpuThread") {
+  Start();
+}
+
+GpuInProcessThread::~GpuInProcessThread() {}
+
+static base::LazyInstance<GpuInProcessThread> g_gpu_thread =
+    LAZY_INSTANCE_INITIALIZER;
+
+static base::LazyInstance<GpuCommandQueue> g_gpu_queue =
+    LAZY_INSTANCE_INITIALIZER;

[piman, 2013/07/23 01:57:14]

I'm fairly uncomfortable with all these globals being destructed in random
order, with pretty important destruction side-effects (e.g. joining a thread).

Could you move all these objects into a single class (i.e. a single singleton
that has all the other objects as fields), so that at least we control the
destruction order and can reason about it?

E.g. if g_gpu_queue is destroyed before g_gpu_thread, then the PostTask in
QueueTask below is unsafe.

[no sievers, 2013/07/25 00:41:23]

Done.

+
+static void QueueTask(const base::Closure& task) {
+  g_gpu_queue.Get().QueueTask(task);
+
+  if (!g_schedule_work_callback.Get().is_null()) {
+    g_schedule_work_callback.Get().Run();
+    return;
+  }
+  g_gpu_thread.Get().message_loop()
+      ->PostTask(FROM_HERE,
+                 base::Bind(&GpuCommandQueue::RunTasks,
+                            base::Unretained(g_gpu_queue.Pointer())));
+}
+
+}  // anonyous namespace
+
+InProcessCommandBuffer::InProcessCommandBuffer()
+    : context_lost_(false), last_put_offset_(-1) {}
+
+InProcessCommandBuffer::~InProcessCommandBuffer() {
+  Destroy();
+}
+
+bool InProcessCommandBuffer::IsContextLost() {
+  if (context_lost_ || !command_buffer_) {
+    return true;
+  }
+  CommandBuffer::State state = GetState();
+  return error::IsError(state.error);
+}
+
+void InProcessCommandBuffer::OnResizeView(gfx::Size size, float scale_factor) {
+  DCHECK(!surface_->IsOffscreen());
+  surface_->Resize(size);
+}
+
+bool InProcessCommandBuffer::MakeCurrent() {
+  if (decoder_->MakeCurrent())
+    return true;
+  DLOG(ERROR) << "Context lost because MakeCurrent failed.";
+  command_buffer_->SetContextLostReason(decoder_->GetContextLostReason());
+  command_buffer_->SetParseError(gpu::error::kLostContext);
+  return false;
+}
+
+void InProcessCommandBuffer::PumpCommands() {
+  base::AutoLock lock(service_lock_);

[piman, 2013/07/23 01:57:14]

I think it would be safer if the lock was taken before we execute the
command_buffer_->Flush() (which ends up calling this), otherwise I fear
command_buffer is not properly protected.
This could become a service_lock_->AssertAcquired() then (also, in other
functions that are meant to be called with the lock held, e.g. MakeCurrent or
GetBufferChanged).

[no sievers, 2013/07/25 00:41:23]

Oops, my bad. Done.

+
+  if (!MakeCurrent())
+    return;
+
+  gpu_scheduler_->PutChanged();
+  CommandBuffer::State state = command_buffer_->GetState();
+  DCHECK((!error::IsError(state.error) && !context_lost_) ||
+         (error::IsError(state.error) && context_lost_));
+}
+
+bool InProcessCommandBuffer::GetBufferChanged(int32 transfer_buffer_id) {
+  command_buffer_->SetGetBuffer(transfer_buffer_id);
+  return true;
+}
+
+bool InProcessCommandBuffer::Initialize(
+    bool is_offscreen,
+    bool share_resources,
+    gfx::AcceleratedWidget window,
+    const gfx::Size& size,
+    const char* allowed_extensions,
+    const std::vector<int32>& attribs,
+    gfx::GpuPreference gpu_preference,
+    const base::Closure& context_lost_callback) {
+
+  share_resources_ = share_resources;
+  context_lost_callback_ = WrapCallback(context_lost_callback);
+
+  base::WaitableEvent completion(true, false);
+  bool result;
+  base::Callback<bool(void)> init_task =
+      base::Bind(&InProcessCommandBuffer::InitializeOnGpuThread,
+                 base::Unretained(this),

[piman, 2013/07/23 01:57:14]

What ensures |this| outlives the thread? If the destruction of |this| relies on
global destruction, the order is undefined and there's no guarantee.

[no sievers, 2013/07/25 00:41:23]

Was this comment intended for the Destroy() further down?
Is it terrible that I made the GpuCommandQueue '::Leaky'?

+                 is_offscreen,
+                 window,
+                 size,
+                 allowed_extensions,
+                 attribs,
+                 gpu_preference);
+  QueueTask(
+      base::Bind(&RunTaskWithResult<bool>, init_task, &result, &completion));
+  completion.Wait();
+  return result;
+}
+
+bool InProcessCommandBuffer::InitializeOnGpuThread(
+    bool is_offscreen,
+    gfx::AcceleratedWidget window,
+    const gfx::Size& size,
+    const char* allowed_extensions,
+    const std::vector<int32>& attribs,
+    gfx::GpuPreference gpu_preference) {
+  // Use one share group for all contexts.
+  CR_DEFINE_STATIC_LOCAL(scoped_refptr<gfx::GLShareGroup>, share_group,
+                         (new gfx::GLShareGroup));
+
+  DCHECK(size.width() >= 0 && size.height() >= 0);
+
+  TransferBufferManager* manager = new TransferBufferManager();
+  transfer_buffer_manager_.reset(manager);
+  manager->Initialize();
+
+  scoped_ptr<CommandBufferService> command_buffer(
+      new CommandBufferService(transfer_buffer_manager_.get()));
+  command_buffer->SetPutOffsetChangeCallback(base::Bind(
+      &InProcessCommandBuffer::PumpCommands, base::Unretained(this)));
+  command_buffer->SetParseErrorCallback(base::Bind(
+      &InProcessCommandBuffer::OnContextLost, base::Unretained(this)));
+
+  if (!command_buffer->Initialize()) {
+    LOG(ERROR) << "Could not initialize command buffer.";
+    Destroy();
+    return false;
+  }
+
+  InProcessCommandBuffer* context_group = NULL;
+
+  if (share_resources_ && !g_all_shared_contexts.Get().empty()) {
+    for (std::set<InProcessCommandBuffer*>::iterator it =
+             g_all_shared_contexts.Get().begin();
+         it != g_all_shared_contexts.Get().end();
+         ++it) {
+      if (!(*it)->IsContextLost()) {
+        context_group = *it;
+        break;
+      }
+    }
+    if (!context_group)
+      share_group = new gfx::GLShareGroup;
+  }
+
+  // TODO(gman): This needs to be true if this is Pepper.
+  bool bind_generates_resource = false;
+  decoder_.reset(gles2::GLES2Decoder::Create(
+      context_group ? context_group->decoder_->GetContextGroup()
+                    : new gles2::ContextGroup(
+                          NULL, NULL, NULL, NULL, bind_generates_resource)));
+
+  gpu_scheduler_.reset(
+      new GpuScheduler(command_buffer.get(), decoder_.get(), decoder_.get()));
+  command_buffer->SetGetBufferChangeCallback(base::Bind(
+      &GpuScheduler::SetGetBuffer, base::Unretained(gpu_scheduler_.get())));
+  command_buffer_ = command_buffer.Pass();
+
+  decoder_->set_engine(gpu_scheduler_.get());
+
+  if (is_offscreen)
+    surface_ = gfx::GLSurface::CreateOffscreenGLSurface(size);
+  else
+    surface_ = gfx::GLSurface::CreateViewGLSurface(window);
+
+  if (!surface_.get()) {
+    LOG(ERROR) << "Could not create GLSurface.";
+    Destroy();
+    return false;
+  }
+
+  if (g_use_virtualized_gl_context) {
+    context_ = share_group->GetSharedContext();
+    if (!context_.get()) {
+      context_ = gfx::GLContext::CreateGLContext(
+          share_group.get(), surface_.get(), gpu_preference);
+      share_group->SetSharedContext(context_.get());
+    }
+
+    context_ = new GLContextVirtual(
+        share_group.get(), context_.get(), decoder_->AsWeakPtr());
+    if (context_->Initialize(surface_.get(), gpu_preference)) {
+      VLOG(1) << "Created virtual GL context.";
+    } else {
+      context_ = NULL;
+    }
+  } else {
+    context_ = gfx::GLContext::CreateGLContext(
+        share_group.get(), surface_.get(), gpu_preference);
+  }
+
+  if (!context_.get()) {
+    LOG(ERROR) << "Could not create GLContext.";
+    Destroy();
+    return false;
+  }
+
+  if (!context_->MakeCurrent(surface_.get())) {
+    LOG(ERROR) << "Could not make context current.";
+    Destroy();
+    return false;
+  }
+
+  gles2::DisallowedFeatures disallowed_features;
+  disallowed_features.swap_buffer_complete_callback = true;
+  disallowed_features.gpu_memory_manager = true;
+  if (!decoder_->Initialize(surface_,
+                            context_,
+                            is_offscreen,
+                            size,
+                            disallowed_features,
+                            allowed_extensions,
+                            attribs)) {
+    LOG(ERROR) << "Could not initialize decoder.";
+    Destroy();
+    return false;
+  }
+
+  if (!is_offscreen) {
+    decoder_->SetResizeCallback(base::Bind(
+        &InProcessCommandBuffer::OnResizeView, base::Unretained(this)));
+  }
+
+  if (share_resources_) {
+    g_all_shared_contexts.Pointer()->insert(this);
+  }
+
+  return true;
+}
+
+void InProcessCommandBuffer::Destroy() {
+  base::WaitableEvent completion(true, false);
+  bool result;
+  base::Callback<bool(void)> destroy_task = base::Bind(
+      &InProcessCommandBuffer::DestroyOnGpuThread, base::Unretained(this));
+  QueueTask(
+      base::Bind(&RunTaskWithResult<bool>, destroy_task, &result, &completion));
+  completion.Wait();
+}
+
+bool InProcessCommandBuffer::DestroyOnGpuThread() {
+  if (decoder_) {
+    decoder_->Destroy(!IsContextLost());
+    decoder_.reset();
+  }
+
+  g_all_shared_contexts.Pointer()->erase(this);
+  return true;
+}
+
+unsigned int InProcessCommandBuffer::AddImage(
+    scoped_refptr<gfx::GLImage> image) {
+  base::AutoLock lock(service_lock_);
+  gles2::ContextGroup* group = decoder_->GetContextGroup();

[no sievers, 2013/07/25 00:41:23]

Oops, neither GLImage nor ImageManager are RefCountedThreadSafe.
I have moved this entirely over to the service thread.

+  unsigned int image_id =
+      group->GetIdAllocator(gles2::id_namespaces::kImages)->AllocateID();
+  base::Closure image_manager_task = base::Bind(
+      &gles2::ImageManager::AddImage, group->image_manager(), image, image_id);
+  QueueTask(image_manager_task);
+  return image_id;
+}
+
+void InProcessCommandBuffer::RemoveImage(unsigned int image_id) {
+  base::AutoLock lock(service_lock_);
+  gles2::ContextGroup* group = decoder_->GetContextGroup();
+  group->GetIdAllocator(gles2::id_namespaces::kImages)->FreeID(image_id);
+  base::Closure image_manager_task = base::Bind(
+      &gles2::ImageManager::RemoveImage, group->image_manager(), image_id);
+  QueueTask(image_manager_task);
+}
+
+void InProcessCommandBuffer::OnContextLost() {
+  if (!context_lost_callback_.is_null())
+    context_lost_callback_.Run();
+
+  context_lost_ = true;
+  if (share_resources_) {
+      for (std::set<InProcessCommandBuffer*>::iterator it =
+               g_all_shared_contexts.Get().begin();
+           it != g_all_shared_contexts.Get().end();
+           ++it)
+        (*it)->context_lost_ = true;
+  }
+}
+
+CommandBuffer::State InProcessCommandBuffer::GetStateFast() {
+  base::AutoLock lock(service_lock_);
+  return last_state_ = command_buffer_->GetState();
+}
+
+CommandBuffer::State InProcessCommandBuffer::GetState() {
+  while (last_put_offset_ != last_state_.get_offset &&
+         last_state_.error == gpu::error::kNoError) {
+    GetStateFast();
+    base::PlatformThread::YieldCurrentThread();
+  }

[piman, 2013/07/23 01:57:14]

I'm not sure what you mean to do here. It looks like it's spinning waiting for
the command buffer to be done with the commands, i.e. a Finish.

GetState was never originally meant to do a Finish, so doing the same as
GetStateFast should be ok.

[no sievers, 2013/07/25 00:41:23]

Done.

+
+  return last_state_;
+}
+
+CommandBuffer::State InProcessCommandBuffer::GetLastState() {
+  return last_state_;
+}
+
+int32 InProcessCommandBuffer::GetLastToken() { return last_state_.token; }
+
+void InProcessCommandBuffer::Flush(int32 put_offset) {
+  if (last_state_.error != gpu::error::kNoError)
+    return;
+
+  if (last_put_offset_ == put_offset)
+    return;
+
+  last_put_offset_ = put_offset;
+  base::Closure task = base::Bind(&CommandBuffer::Flush,
+                                  base::Unretained(command_buffer_.get()),
+                                  put_offset);
+
+  QueueTask(task);
+}
+
+CommandBuffer::State InProcessCommandBuffer::FlushSync(int32 put_offset,
+                                                       int32 last_known_get) {
+  Flush(put_offset);
+  if (last_known_get == last_state_.get_offset) {
+    GetStateFast();
+    base::PlatformThread::YieldCurrentThread();
+  }

[piman, 2013/07/23 01:57:14]

The semantics of FlushSync is that it only returns if some progress has been
made since last_known_get - i.e. upon return, last_state_.get_offset !=
last_known_get (or the command buffer is lost). It's essentially the way to wait
on the command buffer execution.
The current implementation doesn't guarantee that. I think what you'd want is
something like a ConditionVariable that the execution thread would trigger every
time it processes a Flush, that this could wait on in a poll loop, like:

base::AutoLock lock(service_lock_);
while (last_known_get == !last_state_.get_offset) {
  cond_var_.Wait();
  last_state_ = command_buffer->GetState();
}


and in the PumpCommands, do a cond_var_.Signal() while the lock is held.

[no sievers, 2013/07/25 00:41:23]

Hmm, I don't see CommandBufferProxyImpl guarantee that 'last_known_get !=
last_state_.get_offset' after returning because it uses GetStateFast which is
inserted to the front of the queue. It does seem to eventually guarantee
progress on the gpu thread by disallowing GetStateFast messages to be handled
back to back.

I have put a mutex though (mutex per context rather than cond. var with lock). I
hope I got it right.

+
+  return last_state_;
+}
+
+void InProcessCommandBuffer::SetGetBuffer(int32 shm_id) {
+  if (last_state_.error != gpu::error::kNoError)
+    return;
+
+  {
+    base::AutoLock lock(service_lock_);
+    command_buffer_->SetGetBuffer(shm_id);
+    last_put_offset_ = 0;
+  }
+  GetStateFast();
+}
+
+gpu::Buffer InProcessCommandBuffer::CreateTransferBuffer(size_t size,
+                                                         int32* id) {
+  base::AutoLock lock(service_lock_);
+  return command_buffer_->CreateTransferBuffer(size, id);
+}
+
+void InProcessCommandBuffer::DestroyTransferBuffer(int32 id) {
+  base::AutoLock lock(service_lock_);
+  command_buffer_->DestroyTransferBuffer(id);

[piman, 2013/07/23 01:57:14]

mmh, this would execute out-of-order wrt the commands that may still access the
transfer buffer that are in the queue. I think you want to QueueTask this.
I /think/ you can get away without QueueTask'ing CreateTransferBuffer though. If
not, then you'll also need to QueueTask the SetGetBuffer.

[no sievers, 2013/07/25 00:41:23]

It looked like the client only frees pending an explicitly inserted token (and
when we allocate a new shared memory segment it frees the unused ones for which
the token has passed), or with an explicit finish (FreeEverything() and
~GLESImplementation, for example). Otherwise, it would also have to guarantee a
flush is inserted.

But I guess it doesn't hurt to queue this, since it doesn't need to block. Done.

+}
+
+gpu::Buffer InProcessCommandBuffer::GetTransferBuffer(int32 id) {
+  NOTREACHED();
+  return gpu::Buffer();
+}
+
+uint32 InProcessCommandBuffer::InsertSyncPoint() {
+  NOTREACHED();
+  return 0;
+}
+void InProcessCommandBuffer::SignalSyncPoint(unsigned sync_point,
+                                             const base::Closure& callback) {
+  QueueTask(WrapCallback(callback));
+}
+
+gpu::error::Error InProcessCommandBuffer::GetLastError() {
+  return last_state_.error;
+}
+
+bool InProcessCommandBuffer::Initialize() {
+  NOTREACHED();
+  return false;
+}
+
+void InProcessCommandBuffer::SetGetOffset(int32 get_offset) { NOTREACHED(); }
+
+void InProcessCommandBuffer::SetToken(int32 token) { NOTREACHED(); }
+
+void InProcessCommandBuffer::SetParseError(gpu::error::Error error) {
+  NOTREACHED();
+}
+
+void InProcessCommandBuffer::SetContextLostReason(
+    gpu::error::ContextLostReason reason) {
+  NOTREACHED();
+}
+
+static void PostCallback(const scoped_refptr<base::MessageLoopProxy>& loop,
+                         const base::Closure& callback) {
+  if (loop != base::MessageLoopProxy::current())
+    loop->PostTask(FROM_HERE, callback);
+  else
+    callback.Run();
+}
+
+base::Closure InProcessCommandBuffer::WrapCallback(
+    const base::Closure& callback) {
+  base::Closure wrapped_callback =
+      base::Bind(&PostCallback, base::MessageLoopProxy::current(), callback);
+  return wrapped_callback;
+}
+
+// static
+void InProcessCommandBuffer::EnableVirtualizedContext() {
+  g_use_virtualized_gl_context = true;
+}
+
+// static
+void InProcessCommandBuffer::SetScheduleCallback(
+    const base::Closure& callback) {
+  DCHECK(g_schedule_work_callback.Get().is_null());
+  g_schedule_work_callback.Get() = callback;
+}
+
+// static
+void InProcessCommandBuffer::ProcessGpuWorkOnCurrentThread() {
+  g_gpu_queue.Get().RunTasks();
+}
+
+}  // namespace gpu