4 files changed, 23 insertions, 13 deletions
diff --git a/content/common/gpu/gpu_command_buffer_stub.cc b/content/common/gpu/gpu_command_buffer_stub.cc
index d78f4bc..b9d8450 100644
--- a/content/common/gpu/gpu_command_buffer_stub.cc
+++ b/content/common/gpu/gpu_command_buffer_stub.cc
@@ -258,6 +258,7 @@ void GpuCommandBufferStub::OnResizeOffscreenFrameBuffer(const gfx::Size& size) {
 
 void GpuCommandBufferStub::OnSwapBuffers() {
   GPU_TRACE_EVENT0("gpu", "GpuCommandBufferStub::OnSwapBuffers");
+  ReportState();
   Send(new GpuCommandBufferMsg_SwapBuffers(route_id_));
 }
 
diff --git a/gpu/command_buffer/client/gles2_implementation.cc b/gpu/command_buffer/client/gles2_implementation.cc
index aa94fbb..6b85a55 100644
--- a/gpu/command_buffer/client/gles2_implementation.cc
+++ b/gpu/command_buffer/client/gles2_implementation.cc
@@ -684,15 +684,22 @@ void GLES2Implementation::Finish() {
 }
 
 void GLES2Implementation::SwapBuffers() {
-  // Wait if this would add too many swap buffers.
-  if (swap_buffers_tokens_.size() == kMaxSwapBuffers) {
+  // TODO(piman): Strictly speaking we'd want to insert the token after the
+  // swap, but the state update with the updated token might not have happened
+  // by the time the SwapBuffer callback gets called, forcing us to synchronize
+  // with the GPU process more than needed. So instead, make it happen before.
+  // All it means is that we could be slightly looser on the kMaxSwapBuffers
+  // semantics if the client doesn't use the callback mechanism, and by chance
+  // the scheduler yields between the InsertToken and the SwapBuffers.
+  swap_buffers_tokens_.push(helper_->InsertToken());
+  helper_->SwapBuffers();
+  helper_->YieldScheduler();
+  helper_->CommandBufferHelper::Flush();
+  // Wait if we added too many swap buffers.
+  if (swap_buffers_tokens_.size() > kMaxSwapBuffers) {
     helper_->WaitForToken(swap_buffers_tokens_.front());
     swap_buffers_tokens_.pop();
   }
-  helper_->SwapBuffers();
-  swap_buffers_tokens_.push(helper_->InsertToken());
-  helper_->YieldScheduler();
-  Flush();
 }
 
 void GLES2Implementation::CopyTextureToParentTextureCHROMIUM(
diff --git a/gpu/command_buffer/service/gpu_scheduler.cc b/gpu/command_buffer/service/gpu_scheduler.cc
index c1b7aaa..09a24fd 100644
--- a/gpu/command_buffer/service/gpu_scheduler.cc
+++ b/gpu/command_buffer/service/gpu_scheduler.cc
@@ -163,6 +163,11 @@ void GpuScheduler::ProcessCommands() {
          !parser_->IsEmpty()) {
     error = parser_->ProcessCommand();
 
+    // TODO(piman): various classes duplicate various pieces of state, leading
+    // to needlessly complex update logic. It should be possible to simply share
+    // the state across all of them.
+    command_buffer_->SetGetOffset(static_cast<int32>(parser_->get()));
+
     if (error == error::kWaiting || error == error::kYield) {
       break;
     } else if (error::IsError(error)) {
@@ -179,8 +184,6 @@ void GpuScheduler::ProcessCommands() {
     }
   }
 
-  command_buffer_->SetGetOffset(static_cast<int32>(parser_->get()));
-
   if (unscheduled_count_ == 0 &&
       error != error::kWaiting &&
       !parser_->IsEmpty()) {
diff --git a/gpu/command_buffer/service/gpu_scheduler_unittest.cc b/gpu/command_buffer/service/gpu_scheduler_unittest.cc
index a7f17b9..965353c 100644
--- a/gpu/command_buffer/service/gpu_scheduler_unittest.cc
+++ b/gpu/command_buffer/service/gpu_scheduler_unittest.cc
@@ -93,7 +93,6 @@ TEST_F(GpuSchedulerTest, SchedulerDoesNothingIfRingBufferIsEmpty) {
   state.put_offset = 0;
   EXPECT_CALL(*command_buffer_, GetState())
     .WillRepeatedly(Return(state));
-  EXPECT_CALL(*command_buffer_, SetGetOffset(0));
 
   EXPECT_CALL(*command_buffer_, SetParseError(_))
     .Times(0);
@@ -136,13 +135,14 @@ TEST_F(GpuSchedulerTest, ProcessesTwoCommands) {
   state.put_offset = 3;
   EXPECT_CALL(*command_buffer_, GetState())
     .WillRepeatedly(Return(state));
-  EXPECT_CALL(*command_buffer_, SetGetOffset(3));
 
   EXPECT_CALL(*async_api_, DoCommand(7, 1, &buffer_[0]))
     .WillOnce(Return(error::kNoError));
+  EXPECT_CALL(*command_buffer_, SetGetOffset(2));
 
   EXPECT_CALL(*async_api_, DoCommand(8, 0, &buffer_[2]))
     .WillOnce(Return(error::kNoError));
+  EXPECT_CALL(*command_buffer_, SetGetOffset(3));
 
   scheduler_->PutChanged(true);
 }
@@ -157,8 +157,6 @@ TEST_F(GpuSchedulerTest, SchedulerSetsTheGLContext) {
   EXPECT_CALL(*command_buffer_, GetState())
     .WillRepeatedly(Return(state));
 
-  EXPECT_CALL(*command_buffer_, SetGetOffset(0));
-
   scheduler_->PutChanged(true);
 }
 
@@ -180,10 +178,10 @@ TEST_F(GpuSchedulerTest, PostsTaskToFinishRemainingCommands) {
 
   EXPECT_CALL(*async_api_, DoCommand(7, 1, &buffer_[0]))
     .WillOnce(Return(error::kNoError));
+  EXPECT_CALL(*command_buffer_, SetGetOffset(2));
 
   EXPECT_CALL(*async_api_, DoCommand(8, 0, &buffer_[2]))
     .WillOnce(Return(error::kNoError));
-
   EXPECT_CALL(*command_buffer_, SetGetOffset(3));
 
   scheduler_->PutChanged(true);
@@ -216,6 +214,7 @@ TEST_F(GpuSchedulerTest, SetsErrorCodeOnCommandBuffer) {
   EXPECT_CALL(*async_api_, DoCommand(7, 0, &buffer_[0]))
     .WillOnce(Return(
         error::kUnknownCommand));
+  EXPECT_CALL(*command_buffer_, SetGetOffset(1));
 
   EXPECT_CALL(*command_buffer_,
       SetParseError(error::kUnknownCommand));