GpuScheduler can unschedule a command buffer until the GPU has made progress up to a fence.

gpu/command_buffer/service/gpu_scheduler.cc

 // Copyright (c) 2011 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/gpu_scheduler.h"
 
+#include "base/bind.h"
 #include "base/callback.h"
 #include "base/command_line.h"
 #include "base/compiler_specific.h"
 #include "base/debug/trace_event.h"
 #include "base/message_loop.h"
 #include "base/time.h"
 #include "ui/gfx/gl/gl_context.h"
 #include "ui/gfx/gl/gl_bindings.h"
 #include "ui/gfx/gl/gl_surface.h"
 #include "ui/gfx/gl/gl_switches.h"
 
 using ::base::SharedMemory;
 
 namespace gpu {
+namespace {
+const uint64 kPollFencePeriod = 1;
+}
 
 GpuScheduler::GpuScheduler(CommandBuffer* command_buffer,
                            gles2::GLES2Decoder* decoder,
                            CommandParser* parser)
     : command_buffer_(command_buffer),
       decoder_(decoder),
       parser_(parser),
       unscheduled_count_(0) {
   // Map the ring buffer and create the parser.
   if (!parser) {
@@ skipping 17 matching lines @@
 
 void GpuScheduler::PutChanged() {
   TRACE_EVENT1("gpu", "GpuScheduler:PutChanged", "this", this);
 
   DCHECK(IsScheduled());
 
   CommandBuffer::State state = command_buffer_->GetState();
   parser_->set_put(state.put_offset);
   if (state.error != error::kNoError)
     return;
+glGenFencesNV = NULL;

[apatrick_chromium, 2011/10/25 00:44:24]

glFinish code path tested scientifically. Poor performance confirmed. I deleted
this line already.

+  // Check that the GPU has passed all fences.
+  if (!unschedule_fences_.empty()) {
+    if (glGenFencesNV) {
+      while (!unschedule_fences_.empty()) {
+        if (glTestFenceNV(unschedule_fences_.front().fence)) {
+          glDeleteFencesNV(1, &unschedule_fences_.front().fence);
+          unschedule_fences_.front().task.Run();
+          unschedule_fences_.pop();
+        } else {
+          SetScheduled(false);
+          MessageLoop::current()->PostDelayedTask(
+              FROM_HERE,
+              base::Bind(&GpuScheduler::SetScheduled, AsWeakPtr(), true),
+              kPollFencePeriod);
+          return;
+        }
+      }
+    } else {
+      // Hopefully no recent drivers don't support GL_NV_fence and this will
+      // not happen in practice.
+      glFinish();
+
+      while (!unschedule_fences_.empty()) {
+        unschedule_fences_.front().task.Run();
+        unschedule_fences_.pop();
+      }
+    }
+  }
+
+  // One of the unschedule fence tasks might have unscheduled us.
+  if (!IsScheduled())
+    return;
 
   error::Error error = error::kNoError;
   while (!parser_->IsEmpty()) {
+    DCHECK(IsScheduled());
+    DCHECK(unschedule_fences_.empty());
+
     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::IsError(error)) {
       command_buffer_->SetContextLostReason(decoder_->GetContextLostReason());
       command_buffer_->SetParseError(error);
@@ skipping 49 matching lines @@
 
 int32 GpuScheduler::GetGetOffset() {
   return parser_->get();
 }
 
 void GpuScheduler::SetCommandProcessedCallback(
     Callback0::Type* callback) {
   command_processed_callback_.reset(callback);
 }
 
+void GpuScheduler::DeferToFence(base::Closure task) {
+  UnscheduleFence fence;
+
+  // What if either of these GL calls fails? TestFenceNV will return true and
+  // PutChanged will treat the fence as having been crossed and thereby not
+  // poll indefinately. See spec:
+  // http://www.opengl.org/registry/specs/NV/fence.txt
+  //
+  // What should happen if TestFenceNV is called for a name before SetFenceNV
+  // is called?
+  //     We generate an INVALID_OPERATION error, and return TRUE.
+  //     This follows the semantics for texture object names before
+  //     they are bound, in that they acquire their state upon binding.
+  //     We will arbitrarily return TRUE for consistency.
+  if (glGenFencesNV) {
+    glGenFencesNV(1, &fence.fence);
+    glSetFenceNV(fence.fence, GL_ALL_COMPLETED_NV);
+    glFlush();
+  }
+
+  fence.task = task;
+
+  unschedule_fences_.push(fence);
+}
+
+GpuScheduler::UnscheduleFence::UnscheduleFence() : fence(0) {
+}
+
+GpuScheduler::UnscheduleFence::~UnscheduleFence() {
+}
+
 }  // namespace gpu