Revert of Re-land: cc: Use worker context for one-copy tile initialization. (patchset #22 id:420001 of https://codereview.chromium.org/1230203007/ )

Reason for revert:
I'm sorry, but this patch definitely causes failures of one of the context_lost tests on Windows on multiple bots as described in https://code.google.com/p/chromium/issues/detail?id=490295#c15 . I'm sorry the CQ isn't catching these, and I'm not sure why it isn't -- but it's important to fix these failures rather than leaving them on the tree.

Original issue's description:
> Re-land: cc: Use worker context for one-copy tile initialization.
>
> This moves management of staging resources to
> OneCopyTileTaskWorkerPool class. This makes it possible
> to use a worker context to issue and detect when copy
> operations complete.
>
> BUG=490295
> CQ_INCLUDE_TRYBOTS=tryserver.blink:linux_blink_rel
>
> Committed: https://crrev.com/a6b998600660f34bcce9fbc2c75bbc0a3f7f03cc
> Cr-Commit-Position: refs/heads/master@{#342567}

TBR=piman@chromium.org,primiano@chromium.org,vmiura@chromium.org,vmpstr@chromium.org,reveman@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=490295

Review URL: https://codereview.chromium.org/1282183002

Cr-Commit-Position: refs/heads/master@{#342568}

5 files changed, 384 insertions, 530 deletions

diff --git a/cc/raster/gpu_tile_task_worker_pool.cc b/cc/raster/gpu_tile_task_worker_pool.cc
index 2380455..3567308 100644
--- a/cc/raster/gpu_tile_task_worker_pool.cc
+++ b/cc/raster/gpu_tile_task_worker_pool.cc
@@ -46,9 +46,13 @@ class RasterBufferImpl : public RasterBuffer {
     ContextProvider* context_provider = rasterizer_->resource_provider()
                                             ->output_surface()
                                             ->worker_context_provider();
-    DCHECK(context_provider);
 
-    ContextProvider::ScopedContextLock scoped_context(context_provider);
+    // The context lock must be held while accessing the context on a
+    // worker thread.
+    base::AutoLock context_lock(*context_provider->GetLock());
+
+    // Allow this worker thread to bind to context_provider.
+    context_provider->DetachFromThread();
 
     gfx::Rect playback_rect = raster_full_rect;
     if (resource_has_previous_content_) {
@@ -63,7 +67,10 @@ class RasterBufferImpl : public RasterBuffer {
                                  playback_rect, scale);
 
     // Barrier to sync worker context output to cc context.
-    scoped_context.ContextGL()->OrderingBarrierCHROMIUM();
+    context_provider->ContextGL()->OrderingBarrierCHROMIUM();
+
+    // Allow compositor thread to bind to context_provider.
+    context_provider->DetachFromThread();
   }
 
  private:
diff --git a/cc/raster/one_copy_tile_task_worker_pool.cc b/cc/raster/one_copy_tile_task_worker_pool.cc
index 51e4510..64da37e 100644
--- a/cc/raster/one_copy_tile_task_worker_pool.cc
+++ b/cc/raster/one_copy_tile_task_worker_pool.cc
@@ -8,20 +8,16 @@
 #include <limits>
 
 #include "base/strings/stringprintf.h"
-#include "base/thread_task_runner_handle.h"
-#include "base/trace_event/memory_dump_manager.h"
 #include "base/trace_event/trace_event.h"
 #include "base/trace_event/trace_event_argument.h"
 #include "cc/base/math_util.h"
 #include "cc/debug/traced_value.h"
 #include "cc/raster/raster_buffer.h"
 #include "cc/resources/platform_color.h"
-#include "cc/resources/resource_format.h"
-#include "cc/resources/resource_util.h"
+#include "cc/resources/resource_pool.h"
 #include "cc/resources/scoped_resource.h"
-#include "gpu/GLES2/gl2extchromium.h"
 #include "gpu/command_buffer/client/gles2_interface.h"
-#include "gpu/command_buffer/client/gpu_memory_buffer_manager.h"
+#include "ui/gfx/gpu_memory_buffer.h"
 
 namespace cc {
 namespace {
@@ -30,15 +26,49 @@ class RasterBufferImpl : public RasterBuffer {
  public:
   RasterBufferImpl(OneCopyTileTaskWorkerPool* worker_pool,
                    ResourceProvider* resource_provider,
+                   ResourcePool* resource_pool,
                    ResourceFormat resource_format,
-                   const Resource* resource,
+                   const Resource* output_resource,
                    uint64_t previous_content_id)
       : worker_pool_(worker_pool),
-        resource_(resource),
-        lock_(resource_provider, resource->id()),
-        previous_content_id_(previous_content_id) {}
+        resource_provider_(resource_provider),
+        resource_pool_(resource_pool),
+        output_resource_(output_resource),
+        raster_content_id_(0),
+        sequence_(0) {
+    if (worker_pool->have_persistent_gpu_memory_buffers() &&
+        previous_content_id) {
+      raster_resource_ =
+          resource_pool->TryAcquireResourceWithContentId(previous_content_id);
+    }
+    if (raster_resource_) {
+      raster_content_id_ = previous_content_id;
+      DCHECK_EQ(resource_format, raster_resource_->format());
+      DCHECK_EQ(output_resource->size().ToString(),
+                raster_resource_->size().ToString());
+    } else {
+      raster_resource_ = resource_pool->AcquireResource(output_resource->size(),
+                                                        resource_format);
+    }
+
+    lock_.reset(new ResourceProvider::ScopedWriteLockGpuMemoryBuffer(
+        resource_provider_, raster_resource_->id()));
+  }
 
-  ~RasterBufferImpl() override {}
+  ~RasterBufferImpl() override {
+    // Release write lock in case a copy was never scheduled.
+    lock_.reset();
+
+    // Make sure any scheduled copy operations are issued before we release the
+    // raster resource.
+    if (sequence_)
+      worker_pool_->AdvanceLastIssuedCopyTo(sequence_);
+
+    // Return resources to pool so they can be used by another RasterBuffer
+    // instance.
+    resource_pool_->ReleaseResource(raster_resource_.Pass(),
+                                    raster_content_id_);
+  }
 
   // Overridden from RasterBuffer:
   void Playback(const RasterSource* raster_source,
@@ -47,117 +77,55 @@ class RasterBufferImpl : public RasterBuffer {
                 uint64_t new_content_id,
                 float scale,
                 bool include_images) override {
-    worker_pool_->PlaybackAndCopyOnWorkerThread(
-        resource_, &lock_, raster_source, raster_full_rect, raster_dirty_rect,
-        scale, include_images, previous_content_id_, new_content_id);
+    // If there's a raster_content_id_, we are reusing a resource with that
+    // content id.
+    bool reusing_raster_resource = raster_content_id_ != 0;
+    sequence_ = worker_pool_->PlaybackAndScheduleCopyOnWorkerThread(
+        reusing_raster_resource, lock_.Pass(), raster_resource_.get(),
+        output_resource_, raster_source, raster_full_rect, raster_dirty_rect,
+        scale, include_images);
+    // Store the content id of the resource to return to the pool.
+    raster_content_id_ = new_content_id;
   }
 
  private:
   OneCopyTileTaskWorkerPool* worker_pool_;
-  const Resource* resource_;
-  ResourceProvider::ScopedWriteLockGL lock_;
-  uint64_t previous_content_id_;
+  ResourceProvider* resource_provider_;
+  ResourcePool* resource_pool_;
+  const Resource* output_resource_;
+  uint64_t raster_content_id_;
+  scoped_ptr<ScopedResource> raster_resource_;
+  scoped_ptr<ResourceProvider::ScopedWriteLockGpuMemoryBuffer> lock_;
+  CopySequenceNumber sequence_;
 
   DISALLOW_COPY_AND_ASSIGN(RasterBufferImpl);
 };
 
-// Delay between checking for query result to be available.
-const int kCheckForQueryResultAvailableTickRateMs = 1;
+// Number of in-flight copy operations to allow.
+const int kMaxCopyOperations = 32;
+
+// Delay been checking for copy operations to complete.
+const int kCheckForCompletedCopyOperationsTickRateMs = 1;
 
-// Number of attempts to allow before we perform a check that will wait for
-// query to complete.
-const int kMaxCheckForQueryResultAvailableAttempts = 256;
+// Number of failed attempts to allow before we perform a check that will
+// wait for copy operations to complete if needed.
+const int kFailedAttemptsBeforeWaitIfNeeded = 256;
 
 // 4MiB is the size of 4 512x512 tiles, which has proven to be a good
 // default batch size for copy operations.
 const int kMaxBytesPerCopyOperation = 1024 * 1024 * 4;
 
-// Delay before a staging buffer might be released.
-const int kStagingBufferExpirationDelayMs = 1000;
-
-bool CheckForQueryResult(gpu::gles2::GLES2Interface* gl, unsigned query_id) {
-  unsigned complete = 1;
-  gl->GetQueryObjectuivEXT(query_id, GL_QUERY_RESULT_AVAILABLE_EXT, &complete);
-  return !!complete;
-}
-
-void WaitForQueryResult(gpu::gles2::GLES2Interface* gl, unsigned query_id) {
-  TRACE_EVENT0("cc", "WaitForQueryResult");
-
-  int attempts_left = kMaxCheckForQueryResultAvailableAttempts;
-  while (attempts_left--) {
-    if (CheckForQueryResult(gl, query_id))
-      break;
-
-    base::PlatformThread::Sleep(base::TimeDelta::FromMilliseconds(
-        kCheckForQueryResultAvailableTickRateMs));
-  }
-
-  unsigned result = 0;
-  gl->GetQueryObjectuivEXT(query_id, GL_QUERY_RESULT_EXT, &result);
-}
-
 }  // namespace
 
-OneCopyTileTaskWorkerPool::StagingBuffer::StagingBuffer(const gfx::Size& size)
-    : size(size), texture_id(0), image_id(0), query_id(0), content_id(0) {}
-
-OneCopyTileTaskWorkerPool::StagingBuffer::~StagingBuffer() {
-  DCHECK_EQ(texture_id, 0u);
-  DCHECK_EQ(image_id, 0u);
-  DCHECK_EQ(query_id, 0u);
+OneCopyTileTaskWorkerPool::CopyOperation::CopyOperation(
+    scoped_ptr<ResourceProvider::ScopedWriteLockGpuMemoryBuffer> src_write_lock,
+    const Resource* src,
+    const Resource* dst,
+    const gfx::Rect& rect)
+    : src_write_lock(src_write_lock.Pass()), src(src), dst(dst), rect(rect) {
 }
 
-void OneCopyTileTaskWorkerPool::StagingBuffer::DestroyGLResources(
-    gpu::gles2::GLES2Interface* gl) {
-  if (query_id) {
-    gl->DeleteQueriesEXT(1, &query_id);
-    query_id = 0;
-  }
-  if (image_id) {
-    gl->DestroyImageCHROMIUM(image_id);
-    image_id = 0;
-  }
-  if (texture_id) {
-    gl->DeleteTextures(1, &texture_id);
-    texture_id = 0;
-  }
-}
-
-void OneCopyTileTaskWorkerPool::StagingBuffer::OnMemoryDump(
-    base::trace_event::ProcessMemoryDump* pmd,
-    ResourceFormat format,
-    bool in_free_list) const {
-  if (!gpu_memory_buffer)
-    return;
-
-  gfx::GpuMemoryBufferId buffer_id = gpu_memory_buffer->GetId();
-  std::string buffer_dump_name =
-      base::StringPrintf("cc/one_copy/staging_memory/buffer_%d", buffer_id);
-  base::trace_event::MemoryAllocatorDump* buffer_dump =
-      pmd->CreateAllocatorDump(buffer_dump_name);
-
-  uint64_t buffer_size_in_bytes =
-      ResourceUtil::UncheckedSizeInBytes<uint64_t>(size, format);
-  buffer_dump->AddScalar(base::trace_event::MemoryAllocatorDump::kNameSize,
-                         base::trace_event::MemoryAllocatorDump::kUnitsBytes,
-                         buffer_size_in_bytes);
-  buffer_dump->AddScalar("free_size",
-                         base::trace_event::MemoryAllocatorDump::kUnitsBytes,
-                         in_free_list ? buffer_size_in_bytes : 0);
-
-  // Emit an ownership edge towards a global allocator dump node.
-  const uint64 tracing_process_id =
-      base::trace_event::MemoryDumpManager::GetInstance()
-          ->GetTracingProcessId();
-  base::trace_event::MemoryAllocatorDumpGuid shared_buffer_guid =
-      gfx::GetGpuMemoryBufferGUIDForTracing(tracing_process_id, buffer_id);
-  pmd->CreateSharedGlobalAllocatorDump(shared_buffer_guid);
-
-  // By creating an edge with a higher |importance| (w.r.t. browser-side dumps)
-  // the tracing UI will account the effective size of the buffer to the child.
-  const int kImportance = 2;
-  pmd->AddOwnershipEdge(buffer_dump->guid(), shared_buffer_guid, kImportance);
+OneCopyTileTaskWorkerPool::CopyOperation::~CopyOperation() {
 }
 
 // static
@@ -166,49 +134,51 @@ scoped_ptr<TileTaskWorkerPool> OneCopyTileTaskWorkerPool::Create(
     TaskGraphRunner* task_graph_runner,
     ContextProvider* context_provider,
     ResourceProvider* resource_provider,
+    ResourcePool* resource_pool,
     int max_copy_texture_chromium_size,
-    bool use_persistent_gpu_memory_buffers,
-    int max_staging_buffers) {
+    bool have_persistent_gpu_memory_buffers) {
   return make_scoped_ptr<TileTaskWorkerPool>(new OneCopyTileTaskWorkerPool(
-      task_runner, task_graph_runner, resource_provider,
-      max_copy_texture_chromium_size, use_persistent_gpu_memory_buffers,
-      max_staging_buffers));
+      task_runner, task_graph_runner, context_provider, resource_provider,
+      resource_pool, max_copy_texture_chromium_size,
+      have_persistent_gpu_memory_buffers));
 }
 
 OneCopyTileTaskWorkerPool::OneCopyTileTaskWorkerPool(
     base::SequencedTaskRunner* task_runner,
     TaskGraphRunner* task_graph_runner,
+    ContextProvider* context_provider,
     ResourceProvider* resource_provider,
+    ResourcePool* resource_pool,
     int max_copy_texture_chromium_size,
-    bool use_persistent_gpu_memory_buffers,
-    int max_staging_buffers)
+    bool have_persistent_gpu_memory_buffers)
     : task_runner_(task_runner),
       task_graph_runner_(task_graph_runner),
       namespace_token_(task_graph_runner->GetNamespaceToken()),
+      context_provider_(context_provider),
       resource_provider_(resource_provider),
+      resource_pool_(resource_pool),
       max_bytes_per_copy_operation_(
           max_copy_texture_chromium_size
               ? std::min(kMaxBytesPerCopyOperation,
                          max_copy_texture_chromium_size)
               : kMaxBytesPerCopyOperation),
-      use_persistent_gpu_memory_buffers_(use_persistent_gpu_memory_buffers),
+      have_persistent_gpu_memory_buffers_(have_persistent_gpu_memory_buffers),
+      last_issued_copy_operation_(0),
+      last_flushed_copy_operation_(0),
+      lock_(),
+      copy_operation_count_cv_(&lock_),
       bytes_scheduled_since_last_flush_(0),
-      max_staging_buffers_(max_staging_buffers),
-      staging_buffer_expiration_delay_(
-          base::TimeDelta::FromMilliseconds(kStagingBufferExpirationDelayMs)),
-      reduce_memory_usage_pending_(false),
+      issued_copy_operation_count_(0),
+      next_copy_operation_sequence_(1),
+      check_for_completed_copy_operations_pending_(false),
+      shutdown_(false),
       weak_ptr_factory_(this),
       task_set_finished_weak_ptr_factory_(this) {
-  base::trace_event::MemoryDumpManager::GetInstance()->RegisterDumpProvider(
-      this, base::ThreadTaskRunnerHandle::Get());
-  reduce_memory_usage_callback_ =
-      base::Bind(&OneCopyTileTaskWorkerPool::ReduceMemoryUsage,
-                 weak_ptr_factory_.GetWeakPtr());
+  DCHECK(context_provider_);
 }
 
 OneCopyTileTaskWorkerPool::~OneCopyTileTaskWorkerPool() {
-  base::trace_event::MemoryDumpManager::GetInstance()->UnregisterDumpProvider(
-      this);
+  DCHECK_EQ(pending_copy_operations_.size(), 0u);
 }
 
 TileTaskRunner* OneCopyTileTaskWorkerPool::AsTileTaskRunner() {
@@ -222,21 +192,28 @@ void OneCopyTileTaskWorkerPool::SetClient(TileTaskRunnerClient* client) {
 void OneCopyTileTaskWorkerPool::Shutdown() {
   TRACE_EVENT0("cc", "OneCopyTileTaskWorkerPool::Shutdown");
 
+  {
+    base::AutoLock lock(lock_);
+
+    shutdown_ = true;
+    copy_operation_count_cv_.Signal();
+  }
+
   TaskGraph empty;
   task_graph_runner_->ScheduleTasks(namespace_token_, &empty);
   task_graph_runner_->WaitForTasksToFinishRunning(namespace_token_);
-
-  base::AutoLock lock(lock_);
-
-  if (buffers_.empty())
-    return;
-
-  ReleaseBuffersNotUsedSince(base::TimeTicks() + base::TimeDelta::Max());
 }
 
 void OneCopyTileTaskWorkerPool::ScheduleTasks(TileTaskQueue* queue) {
   TRACE_EVENT0("cc", "OneCopyTileTaskWorkerPool::ScheduleTasks");
 
+#if DCHECK_IS_ON()
+  {
+    base::AutoLock lock(lock_);
+    DCHECK(!shutdown_);
+  }
+#endif
+
   if (tasks_pending_.none())
     TRACE_EVENT_ASYNC_BEGIN0("cc", "ScheduledTasks", this);
 
@@ -261,6 +238,8 @@ void OneCopyTileTaskWorkerPool::ScheduleTasks(TileTaskQueue* queue) {
                    task_set_finished_weak_ptr_factory_.GetWeakPtr(), task_set));
   }
 
+  resource_pool_->CheckBusyResources(false);
+
   for (TileTaskQueue::Item::Vector::const_iterator it = queue->items.begin();
        it != queue->items.end(); ++it) {
     const TileTaskQueue::Item& item = *it;
@@ -287,19 +266,14 @@ void OneCopyTileTaskWorkerPool::ScheduleTasks(TileTaskQueue* queue) {
   }
 
   ScheduleTasksOnOriginThread(this, &graph_);
-
-  // Barrier to sync any new resources to the worker context.
-  resource_provider_->output_surface()
-      ->context_provider()
-      ->ContextGL()
-      ->OrderingBarrierCHROMIUM();
-
   task_graph_runner_->ScheduleTasks(namespace_token_, &graph_);
 
   std::copy(new_task_set_finished_tasks,
             new_task_set_finished_tasks + kNumberOfTaskSets,
             task_set_finished_tasks_);
 
+  resource_pool_->ReduceResourceUsage();
+
   TRACE_EVENT_ASYNC_STEP_INTO1("cc", "ScheduledTasks", this, "running", "state",
                                StateAsValue());
 }
@@ -338,9 +312,10 @@ scoped_ptr<RasterBuffer> OneCopyTileTaskWorkerPool::AcquireBufferForRaster(
   // TODO(danakj): If resource_content_id != 0, we only need to copy/upload
   // the dirty rect.
   DCHECK_EQ(resource->format(), resource_provider_->best_texture_format());
-  return make_scoped_ptr<RasterBuffer>(new RasterBufferImpl(
-      this, resource_provider_, resource_provider_->best_texture_format(),
-      resource, previous_content_id));
+  return make_scoped_ptr<RasterBuffer>(
+      new RasterBufferImpl(this, resource_provider_, resource_pool_,
+                           resource_provider_->best_texture_format(), resource,
+                           previous_content_id));
 }
 
 void OneCopyTileTaskWorkerPool::ReleaseBufferForRaster(
@@ -348,372 +323,236 @@ void OneCopyTileTaskWorkerPool::ReleaseBufferForRaster(
   // Nothing to do here. RasterBufferImpl destructor cleans up after itself.
 }
 
-void OneCopyTileTaskWorkerPool::PlaybackAndCopyOnWorkerThread(
-    const Resource* resource,
-    const ResourceProvider::ScopedWriteLockGL* resource_lock,
+CopySequenceNumber
+OneCopyTileTaskWorkerPool::PlaybackAndScheduleCopyOnWorkerThread(
+    bool reusing_raster_resource,
+    scoped_ptr<ResourceProvider::ScopedWriteLockGpuMemoryBuffer>
+        raster_resource_write_lock,
+    const Resource* raster_resource,
+    const Resource* output_resource,
     const RasterSource* raster_source,
     const gfx::Rect& raster_full_rect,
     const gfx::Rect& raster_dirty_rect,
     float scale,
-    bool include_images,
-    uint64_t previous_content_id,
-    uint64_t new_content_id) {
-  base::AutoLock lock(lock_);
-
-  scoped_ptr<StagingBuffer> staging_buffer =
-      AcquireStagingBuffer(resource, previous_content_id);
-  DCHECK(staging_buffer);
-
-  {
-    base::AutoUnlock unlock(lock_);
-
-    // Allocate GpuMemoryBuffer if necessary.
-    if (!staging_buffer->gpu_memory_buffer) {
-      staging_buffer->gpu_memory_buffer =
-          resource_provider_->gpu_memory_buffer_manager()
-              ->AllocateGpuMemoryBuffer(
-                  staging_buffer->size,
-                  BufferFormat(resource_provider_->best_texture_format()),
-                  use_persistent_gpu_memory_buffers_
-                      ? gfx::BufferUsage::PERSISTENT_MAP
-                      : gfx::BufferUsage::MAP);
-    }
+    bool include_images) {
+  gfx::GpuMemoryBuffer* gpu_memory_buffer =
+      raster_resource_write_lock->GetGpuMemoryBuffer();
+  if (gpu_memory_buffer) {
+    void* data = NULL;
+    bool rv = gpu_memory_buffer->Map(&data);
+    DCHECK(rv);
+    int stride;
+    gpu_memory_buffer->GetStride(&stride);
+    // TileTaskWorkerPool::PlaybackToMemory only supports unsigned strides.
+    DCHECK_GE(stride, 0);
 
     gfx::Rect playback_rect = raster_full_rect;
-    if (use_persistent_gpu_memory_buffers_ && previous_content_id) {
-      // Reduce playback rect to dirty region if the content id of the staging
-      // buffer matches the prevous content id.
-      if (previous_content_id == staging_buffer->content_id)
-        playback_rect.Intersect(raster_dirty_rect);
-    }
-
-    if (staging_buffer->gpu_memory_buffer) {
-      void* data = nullptr;
-      bool rv = staging_buffer->gpu_memory_buffer->Map(&data);
-      DCHECK(rv);
-      int stride;
-      staging_buffer->gpu_memory_buffer->GetStride(&stride);
-      // TileTaskWorkerPool::PlaybackToMemory only supports unsigned strides.
-      DCHECK_GE(stride, 0);
-
-      DCHECK(!playback_rect.IsEmpty())
-          << "Why are we rastering a tile that's not dirty?";
-      TileTaskWorkerPool::PlaybackToMemory(
-          data, resource_provider_->best_texture_format(), staging_buffer->size,
-          static_cast<size_t>(stride), raster_source, raster_full_rect,
-          playback_rect, scale, include_images);
-      staging_buffer->gpu_memory_buffer->Unmap();
-      staging_buffer->content_id = new_content_id;
+    if (reusing_raster_resource) {
+      playback_rect.Intersect(raster_dirty_rect);
     }
+    DCHECK(!playback_rect.IsEmpty())
+        << "Why are we rastering a tile that's not dirty?";
+    TileTaskWorkerPool::PlaybackToMemory(
+        data, raster_resource->format(), raster_resource->size(),
+        static_cast<size_t>(stride), raster_source, raster_full_rect,
+        playback_rect, scale, include_images);
+    gpu_memory_buffer->Unmap();
   }
 
-  ContextProvider* context_provider =
-      resource_provider_->output_surface()->worker_context_provider();
-  DCHECK(context_provider);
+  base::AutoLock lock(lock_);
 
-  {
-    ContextProvider::ScopedContextLock scoped_context(context_provider);
-
-    gpu::gles2::GLES2Interface* gl = scoped_context.ContextGL();
-    DCHECK(gl);
-
-    unsigned image_target = resource_provider_->GetImageTextureTarget(
-        resource_provider_->best_texture_format());
-
-    // Create and bind staging texture.
-    if (!staging_buffer->texture_id) {
-      gl->GenTextures(1, &staging_buffer->texture_id);
-      gl->BindTexture(image_target, staging_buffer->texture_id);
-      gl->TexParameteri(image_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-      gl->TexParameteri(image_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-      gl->TexParameteri(image_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-      gl->TexParameteri(image_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
-    } else {
-      gl->BindTexture(image_target, staging_buffer->texture_id);
-    }
+  CopySequenceNumber sequence = 0;
+  int bytes_per_row = (BitsPerPixel(raster_resource->format()) *
+                       raster_resource->size().width()) /
+                      8;
+  int chunk_size_in_rows =
+      std::max(1, max_bytes_per_copy_operation_ / bytes_per_row);
+  // Align chunk size to 4. Required to support compressed texture formats.
+  chunk_size_in_rows = MathUtil::UncheckedRoundUp(chunk_size_in_rows, 4);
+  int y = 0;
+  int height = raster_resource->size().height();
+  while (y < height) {
+    int failed_attempts = 0;
+    while ((pending_copy_operations_.size() + issued_copy_operation_count_) >=
+           kMaxCopyOperations) {
+      // Ignore limit when shutdown is set.
+      if (shutdown_)
+        break;
 
-    // Create and bind image.
-    if (!staging_buffer->image_id) {
-      if (staging_buffer->gpu_memory_buffer) {
-        staging_buffer->image_id = gl->CreateImageCHROMIUM(
-            staging_buffer->gpu_memory_buffer->AsClientBuffer(),
-            staging_buffer->size.width(), staging_buffer->size.height(),
-            GLInternalFormat(resource_provider_->best_texture_format()));
-        gl->BindTexImage2DCHROMIUM(image_target, staging_buffer->image_id);
-      }
-    } else {
-      gl->ReleaseTexImage2DCHROMIUM(image_target, staging_buffer->image_id);
-      gl->BindTexImage2DCHROMIUM(image_target, staging_buffer->image_id);
-    }
+      ++failed_attempts;
 
-    // Unbind staging texture.
-    gl->BindTexture(image_target, 0);
-
-    if (resource_provider_->use_sync_query()) {
-      if (!staging_buffer->query_id)
-        gl->GenQueriesEXT(1, &staging_buffer->query_id);
-
-#if defined(OS_CHROMEOS)
-      // TODO(reveman): This avoids a performance problem on some ChromeOS
-      // devices. This needs to be removed to support native GpuMemoryBuffer
-      // implementations. crbug.com/436314
-      gl->BeginQueryEXT(GL_COMMANDS_ISSUED_CHROMIUM, staging_buffer->query_id);
-#else
-      gl->BeginQueryEXT(GL_COMMANDS_COMPLETED_CHROMIUM,
-                        staging_buffer->query_id);
-#endif
-    }
+      // Schedule a check that will also wait for operations to complete
+      // after too many failed attempts.
+      bool wait_if_needed = failed_attempts > kFailedAttemptsBeforeWaitIfNeeded;
+
+      // Schedule a check for completed copy operations if too many operations
+      // are currently in-flight.
+      ScheduleCheckForCompletedCopyOperationsWithLockAcquired(wait_if_needed);
 
-    int bytes_per_row =
-        (BitsPerPixel(resource_provider_->best_texture_format()) *
-         resource->size().width()) /
-        8;
-    int chunk_size_in_rows =
-        std::max(1, max_bytes_per_copy_operation_ / bytes_per_row);
-    // Align chunk size to 4. Required to support compressed texture formats.
-    chunk_size_in_rows = MathUtil::UncheckedRoundUp(chunk_size_in_rows, 4);
-    int y = 0;
-    int height = resource->size().height();
-    while (y < height) {
-      // Copy at most |chunk_size_in_rows|.
-      int rows_to_copy = std::min(chunk_size_in_rows, height - y);
-      DCHECK_GT(rows_to_copy, 0);
-
-      gl->CopySubTextureCHROMIUM(GL_TEXTURE_2D, staging_buffer->texture_id,
-                                 resource_lock->texture_id(), 0, y, 0, y,
-                                 resource->size().width(), rows_to_copy, false,
-                                 false, false);
-      y += rows_to_copy;
-
-      // Increment |bytes_scheduled_since_last_flush_| by the amount of memory
-      // used for this copy operation.
-      bytes_scheduled_since_last_flush_ += rows_to_copy * bytes_per_row;
-
-      if (bytes_scheduled_since_last_flush_ >= max_bytes_per_copy_operation_) {
-        gl->ShallowFlushCHROMIUM();
-        bytes_scheduled_since_last_flush_ = 0;
+      {
+        TRACE_EVENT0("cc", "WaitingForCopyOperationsToComplete");
+
+        // Wait for in-flight copy operations to drop below limit.
+        copy_operation_count_cv_.Wait();
       }
     }
 
-    if (resource_provider_->use_sync_query()) {
-#if defined(OS_CHROMEOS)
-      gl->EndQueryEXT(GL_COMMANDS_ISSUED_CHROMIUM);
-#else
-      gl->EndQueryEXT(GL_COMMANDS_COMPLETED_CHROMIUM);
-#endif
+    // There may be more work available, so wake up another worker thread.
+    copy_operation_count_cv_.Signal();
+
+    // Copy at most |chunk_size_in_rows|.
+    int rows_to_copy = std::min(chunk_size_in_rows, height - y);
+    DCHECK_GT(rows_to_copy, 0);
+
+    // |raster_resource_write_lock| is passed to the first copy operation as it
+    // needs to be released before we can issue a copy.
+    pending_copy_operations_.push_back(make_scoped_ptr(new CopyOperation(
+        raster_resource_write_lock.Pass(), raster_resource, output_resource,
+        gfx::Rect(0, y, raster_resource->size().width(), rows_to_copy))));
+    y += rows_to_copy;
+
+    // Acquire a sequence number for this copy operation.
+    sequence = next_copy_operation_sequence_++;
+
+    // Increment |bytes_scheduled_since_last_flush_| by the amount of memory
+    // used for this copy operation.
+    bytes_scheduled_since_last_flush_ += rows_to_copy * bytes_per_row;
+
+    // Post task that will advance last flushed copy operation to |sequence|
+    // when |bytes_scheduled_since_last_flush_| has reached
+    // |max_bytes_per_copy_operation_|.
+    if (bytes_scheduled_since_last_flush_ >= max_bytes_per_copy_operation_) {
+      task_runner_->PostTask(
+          FROM_HERE,
+          base::Bind(&OneCopyTileTaskWorkerPool::AdvanceLastFlushedCopyTo,
+                     weak_ptr_factory_.GetWeakPtr(), sequence));
+      bytes_scheduled_since_last_flush_ = 0;
     }
-
-    // Barrier to sync worker context output to cc context.
-    gl->OrderingBarrierCHROMIUM();
   }
 
-  staging_buffer->last_usage = base::TimeTicks::Now();
-  busy_buffers_.push_back(staging_buffer.Pass());
-
-  ScheduleReduceMemoryUsage();
+  return sequence;
 }
 
-bool OneCopyTileTaskWorkerPool::OnMemoryDump(
-    const base::trace_event::MemoryDumpArgs& args,
-    base::trace_event::ProcessMemoryDump* pmd) {
-  base::AutoLock lock(lock_);
-
-  for (const auto& buffer : buffers_) {
-    buffer->OnMemoryDump(pmd, resource_provider_->best_texture_format(),
-                         std::find(free_buffers_.begin(), free_buffers_.end(),
-                                   buffer) != free_buffers_.end());
-  }
+void OneCopyTileTaskWorkerPool::AdvanceLastIssuedCopyTo(
+    CopySequenceNumber sequence) {
+  if (last_issued_copy_operation_ >= sequence)
+    return;
 
-  return true;
+  IssueCopyOperations(sequence - last_issued_copy_operation_);
+  last_issued_copy_operation_ = sequence;
 }
 
-scoped_ptr<OneCopyTileTaskWorkerPool::StagingBuffer>
-OneCopyTileTaskWorkerPool::AcquireStagingBuffer(const Resource* resource,
-                                                uint64_t previous_content_id) {
-  lock_.AssertAcquired();
+void OneCopyTileTaskWorkerPool::AdvanceLastFlushedCopyTo(
+    CopySequenceNumber sequence) {
+  if (last_flushed_copy_operation_ >= sequence)
+    return;
 
-  scoped_ptr<StagingBuffer> staging_buffer;
+  AdvanceLastIssuedCopyTo(sequence);
 
-  ContextProvider* context_provider =
-      resource_provider_->output_surface()->worker_context_provider();
-  DCHECK(context_provider);
+  // Flush all issued copy operations.
+  context_provider_->ContextGL()->ShallowFlushCHROMIUM();
+  last_flushed_copy_operation_ = last_issued_copy_operation_;
+}
 
-  ContextProvider::ScopedContextLock scoped_context(context_provider);
+void OneCopyTileTaskWorkerPool::OnTaskSetFinished(TaskSet task_set) {
+  TRACE_EVENT1("cc", "OneCopyTileTaskWorkerPool::OnTaskSetFinished", "task_set",
+               task_set);
 
-  gpu::gles2::GLES2Interface* gl = scoped_context.ContextGL();
-  DCHECK(gl);
+  DCHECK(tasks_pending_[task_set]);
+  tasks_pending_[task_set] = false;
+  if (tasks_pending_.any()) {
+    TRACE_EVENT_ASYNC_STEP_INTO1("cc", "ScheduledTasks", this, "running",
+                                 "state", StateAsValue());
+  } else {
+    TRACE_EVENT_ASYNC_END0("cc", "ScheduledTasks", this);
+  }
+  client_->DidFinishRunningTileTasks(task_set);
+}
 
-  // Check if any busy buffers have become available.
-  if (resource_provider_->use_sync_query()) {
-    while (!busy_buffers_.empty()) {
-      if (!CheckForQueryResult(gl, busy_buffers_.front()->query_id))
-        break;
+void OneCopyTileTaskWorkerPool::IssueCopyOperations(int64 count) {
+  TRACE_EVENT1("cc", "OneCopyTileTaskWorkerPool::IssueCopyOperations", "count",
+               count);
 
-      free_buffers_.push_back(busy_buffers_.take_front());
-    }
-  }
+  CopyOperation::Deque copy_operations;
 
-  // Wait for number of non-free buffers to become less than the limit.
-  while ((buffers_.size() - free_buffers_.size()) >= max_staging_buffers_) {
-    // Stop when there are no more busy buffers to wait for.
-    if (busy_buffers_.empty())
-      break;
+  {
+    base::AutoLock lock(lock_);
 
-    if (resource_provider_->use_sync_query()) {
-      WaitForQueryResult(gl, busy_buffers_.front()->query_id);
-      free_buffers_.push_back(busy_buffers_.take_front());
-    } else {
-      // Fall-back to glFinish if CHROMIUM_sync_query is not available.
-      gl->Finish();
-      while (!busy_buffers_.empty())
-        free_buffers_.push_back(busy_buffers_.take_front());
+    for (int64 i = 0; i < count; ++i) {
+      DCHECK(!pending_copy_operations_.empty());
+      copy_operations.push_back(pending_copy_operations_.take_front());
     }
-  }
 
-  // Find a staging buffer that allows us to perform partial raster when
-  // using persistent GpuMemoryBuffers.
-  if (use_persistent_gpu_memory_buffers_ && previous_content_id) {
-    StagingBufferDeque::iterator it =
-        std::find_if(free_buffers_.begin(), free_buffers_.end(),
-                     [previous_content_id](const StagingBuffer* buffer) {
-                       return buffer->content_id == previous_content_id;
-                     });
-    if (it != free_buffers_.end())
-      staging_buffer = free_buffers_.take(it);
+    // Increment |issued_copy_operation_count_| to reflect the transition of
+    // copy operations from "pending" to "issued" state.
+    issued_copy_operation_count_ += copy_operations.size();
   }
 
-  // Find staging buffer of correct size.
-  if (!staging_buffer) {
-    StagingBufferDeque::iterator it =
-        std::find_if(free_buffers_.begin(), free_buffers_.end(),
-                     [resource](const StagingBuffer* buffer) {
-                       return buffer->size == resource->size();
-                     });
-    if (it != free_buffers_.end())
-      staging_buffer = free_buffers_.take(it);
-  }
+  while (!copy_operations.empty()) {
+    scoped_ptr<CopyOperation> copy_operation = copy_operations.take_front();
 
-  // Create new staging buffer if necessary.
-  if (!staging_buffer) {
-    staging_buffer = make_scoped_ptr(new StagingBuffer(resource->size()));
-    buffers_.insert(staging_buffer.get());
-  }
+    // Remove the write lock.
+    copy_operation->src_write_lock.reset();
 
-  // Release enough free buffers to stay within the limit.
-  while (buffers_.size() > max_staging_buffers_) {
-    if (free_buffers_.empty())
-      break;
-
-    free_buffers_.front()->DestroyGLResources(gl);
-    buffers_.erase(free_buffers_.front());
-    free_buffers_.take_front();
+    // Copy contents of source resource to destination resource.
+    resource_provider_->CopyResource(copy_operation->src->id(),
+                                     copy_operation->dst->id(),
+                                     copy_operation->rect);
   }
-
-  return staging_buffer.Pass();
-}
-
-base::TimeTicks OneCopyTileTaskWorkerPool::GetUsageTimeForLRUBuffer() {
-  lock_.AssertAcquired();
-
-  if (!free_buffers_.empty())
-    return free_buffers_.front()->last_usage;
-
-  if (!busy_buffers_.empty())
-    return busy_buffers_.front()->last_usage;
-
-  return base::TimeTicks();
 }
 
-void OneCopyTileTaskWorkerPool::ScheduleReduceMemoryUsage() {
+void OneCopyTileTaskWorkerPool::
+    ScheduleCheckForCompletedCopyOperationsWithLockAcquired(
+        bool wait_if_needed) {
   lock_.AssertAcquired();
 
-  if (reduce_memory_usage_pending_)
+  if (check_for_completed_copy_operations_pending_)
     return;
 
-  reduce_memory_usage_pending_ = true;
+  base::TimeTicks now = base::TimeTicks::Now();
 
-  // Schedule a call to ReduceMemoryUsage at the time when the LRU buffer
-  // should be released.
-  base::TimeTicks reduce_memory_usage_time =
-      GetUsageTimeForLRUBuffer() + staging_buffer_expiration_delay_;
-  task_runner_->PostDelayedTask(
-      FROM_HERE, reduce_memory_usage_callback_,
-      reduce_memory_usage_time - base::TimeTicks::Now());
-}
-
-void OneCopyTileTaskWorkerPool::ReduceMemoryUsage() {
-  base::AutoLock lock(lock_);
-
-  reduce_memory_usage_pending_ = false;
+  // Schedule a check for completed copy operations as soon as possible but
+  // don't allow two consecutive checks to be scheduled to run less than the
+  // tick rate apart.
+  base::TimeTicks next_check_for_completed_copy_operations_time =
+      std::max(last_check_for_completed_copy_operations_time_ +
+                   base::TimeDelta::FromMilliseconds(
+                       kCheckForCompletedCopyOperationsTickRateMs),
+               now);
 
-  if (free_buffers_.empty() && busy_buffers_.empty())
-    return;
-
-  base::TimeTicks current_time = base::TimeTicks::Now();
-  ReleaseBuffersNotUsedSince(current_time - staging_buffer_expiration_delay_);
-
-  if (free_buffers_.empty() && busy_buffers_.empty())
-    return;
-
-  reduce_memory_usage_pending_ = true;
-
-  // Schedule another call to ReduceMemoryUsage at the time when the next
-  // buffer should be released.
-  base::TimeTicks reduce_memory_usage_time =
-      GetUsageTimeForLRUBuffer() + staging_buffer_expiration_delay_;
-  task_runner_->PostDelayedTask(FROM_HERE, reduce_memory_usage_callback_,
-                                reduce_memory_usage_time - current_time);
+  task_runner_->PostDelayedTask(
+      FROM_HERE,
+      base::Bind(&OneCopyTileTaskWorkerPool::CheckForCompletedCopyOperations,
+                 weak_ptr_factory_.GetWeakPtr(), wait_if_needed),
+      next_check_for_completed_copy_operations_time - now);
+
+  last_check_for_completed_copy_operations_time_ =
+      next_check_for_completed_copy_operations_time;
+  check_for_completed_copy_operations_pending_ = true;
 }
 
-void OneCopyTileTaskWorkerPool::ReleaseBuffersNotUsedSince(
-    base::TimeTicks time) {
-  lock_.AssertAcquired();
+void OneCopyTileTaskWorkerPool::CheckForCompletedCopyOperations(
+    bool wait_if_needed) {
+  TRACE_EVENT1("cc",
+               "OneCopyTileTaskWorkerPool::CheckForCompletedCopyOperations",
+               "wait_if_needed", wait_if_needed);
 
-  ContextProvider* context_provider =
-      resource_provider_->output_surface()->worker_context_provider();
-  DCHECK(context_provider);
+  resource_pool_->CheckBusyResources(wait_if_needed);
 
   {
-    ContextProvider::ScopedContextLock scoped_context(context_provider);
-
-    gpu::gles2::GLES2Interface* gl = scoped_context.ContextGL();
-    DCHECK(gl);
-
-    // Note: Front buffer is guaranteed to be LRU so we can stop releasing
-    // buffers as soon as we find a buffer that has been used since |time|.
-    while (!free_buffers_.empty()) {
-      if (free_buffers_.front()->last_usage > time)
-        return;
-
-      free_buffers_.front()->DestroyGLResources(gl);
-      buffers_.erase(free_buffers_.front());
-      free_buffers_.take_front();
-    }
+    base::AutoLock lock(lock_);
 
-    while (!busy_buffers_.empty()) {
-      if (busy_buffers_.front()->last_usage > time)
-        return;
+    DCHECK(check_for_completed_copy_operations_pending_);
+    check_for_completed_copy_operations_pending_ = false;
 
-      busy_buffers_.front()->DestroyGLResources(gl);
-      buffers_.erase(busy_buffers_.front());
-      busy_buffers_.take_front();
-    }
-  }
-}
+    // The number of busy resources in the pool reflects the number of issued
+    // copy operations that have not yet completed.
+    issued_copy_operation_count_ = resource_pool_->busy_resource_count();
 
-void OneCopyTileTaskWorkerPool::OnTaskSetFinished(TaskSet task_set) {
-  TRACE_EVENT1("cc", "OneCopyTileTaskWorkerPool::OnTaskSetFinished", "task_set",
-               task_set);
-
-  DCHECK(tasks_pending_[task_set]);
-  tasks_pending_[task_set] = false;
-  if (tasks_pending_.any()) {
-    TRACE_EVENT_ASYNC_STEP_INTO1("cc", "ScheduledTasks", this, "running",
-                                 "state", StateAsValue());
-  } else {
-    TRACE_EVENT_ASYNC_END0("cc", "ScheduledTasks", this);
+    // There may be work blocked on too many in-flight copy operations, so wake
+    // up a worker thread.
+    copy_operation_count_cv_.Signal();
   }
-  client_->DidFinishRunningTileTasks(task_set);
 }
 
 scoped_refptr<base::trace_event::ConvertableToTraceFormat>
@@ -734,14 +573,20 @@ OneCopyTileTaskWorkerPool::StateAsValue() const {
 
 void OneCopyTileTaskWorkerPool::StagingStateAsValueInto(
     base::trace_event::TracedValue* staging_state) const {
-  base::AutoLock lock(lock_);
-
-  staging_state->SetInteger("staging_buffer_count",
-                            static_cast<int>(buffers_.size()));
-  staging_state->SetInteger("busy_count",
-                            static_cast<int>(busy_buffers_.size()));
-  staging_state->SetInteger("free_count",
-                            static_cast<int>(free_buffers_.size()));
+  staging_state->SetInteger(
+      "staging_resource_count",
+      static_cast<int>(resource_pool_->total_resource_count()));
+  staging_state->SetInteger(
+      "bytes_used_for_staging_resources",
+      static_cast<int>(resource_pool_->total_memory_usage_bytes()));
+  staging_state->SetInteger(
+      "pending_copy_count",
+      static_cast<int>(resource_pool_->total_resource_count() -
+                       resource_pool_->acquired_resource_count()));
+  staging_state->SetInteger(
+      "bytes_pending_copy",
+      static_cast<int>(resource_pool_->total_memory_usage_bytes() -
+                       resource_pool_->acquired_memory_usage_bytes()));
 }
 
 }  // namespace cc
diff --git a/cc/raster/one_copy_tile_task_worker_pool.h b/cc/raster/one_copy_tile_task_worker_pool.h
index 851f1b4..4e050b4 100644
--- a/cc/raster/one_copy_tile_task_worker_pool.h
+++ b/cc/raster/one_copy_tile_task_worker_pool.h
@@ -5,12 +5,8 @@
 #ifndef CC_RASTER_ONE_COPY_TILE_TASK_WORKER_POOL_H_
 #define CC_RASTER_ONE_COPY_TILE_TASK_WORKER_POOL_H_
 
-#include <set>
-
 #include "base/memory/weak_ptr.h"
 #include "base/synchronization/lock.h"
-#include "base/time/time.h"
-#include "base/trace_event/memory_dump_provider.h"
 #include "base/values.h"
 #include "cc/base/scoped_ptr_deque.h"
 #include "cc/output/context_provider.h"
@@ -25,20 +21,15 @@ class TracedValue;
 }
 }
 
-namespace gpu {
-namespace gles2 {
-class GLES2Interface;
-}
-}
-
 namespace cc {
 class ResourcePool;
+class ScopedResource;
+
+typedef int64 CopySequenceNumber;
 
-class CC_EXPORT OneCopyTileTaskWorkerPool
-    : public TileTaskWorkerPool,
-      public TileTaskRunner,
-      public TileTaskClient,
-      public base::trace_event::MemoryDumpProvider {
+class CC_EXPORT OneCopyTileTaskWorkerPool : public TileTaskWorkerPool,
+                                            public TileTaskRunner,
+                                            public TileTaskClient {
  public:
   ~OneCopyTileTaskWorkerPool() override;
 
@@ -47,9 +38,9 @@ class CC_EXPORT OneCopyTileTaskWorkerPool
       TaskGraphRunner* task_graph_runner,
       ContextProvider* context_provider,
       ResourceProvider* resource_provider,
+      ResourcePool* resource_pool,
       int max_copy_texture_chromium_size,
-      bool use_persistent_gpu_memory_buffers,
-      int max_staging_buffers);
+      bool have_persistent_gpu_memory_buffers);
 
   // Overridden from TileTaskWorkerPool:
   TileTaskRunner* AsTileTaskRunner() override;
@@ -69,57 +60,61 @@ class CC_EXPORT OneCopyTileTaskWorkerPool
       uint64_t previous_content_id) override;
   void ReleaseBufferForRaster(scoped_ptr<RasterBuffer> buffer) override;
 
-  // Overridden from base::trace_event::MemoryDumpProvider:
-  bool OnMemoryDump(const base::trace_event::MemoryDumpArgs& args,
-                    base::trace_event::ProcessMemoryDump* pmd) override;
-
-  // Playback raster source and copy result into |resource|.
-  void PlaybackAndCopyOnWorkerThread(
-      const Resource* resource,
-      const ResourceProvider::ScopedWriteLockGL* resource_lock,
+  // Playback raster source and schedule copy of |raster_resource| resource to
+  // |output_resource|. Returns a non-zero sequence number for this copy
+  // operation.
+  CopySequenceNumber PlaybackAndScheduleCopyOnWorkerThread(
+      bool reusing_raster_resource,
+      scoped_ptr<ResourceProvider::ScopedWriteLockGpuMemoryBuffer>
+          raster_resource_write_lock,
+      const Resource* raster_resource,
+      const Resource* output_resource,
       const RasterSource* raster_source,
       const gfx::Rect& raster_full_rect,
       const gfx::Rect& raster_dirty_rect,
       float scale,
-      bool include_images,
-      uint64_t resource_content_id,
-      uint64_t previous_content_id);
+      bool include_images);
+
+  // Issues copy operations until |sequence| has been processed. This will
+  // return immediately if |sequence| has already been processed.
+  void AdvanceLastIssuedCopyTo(CopySequenceNumber sequence);
+
+  bool have_persistent_gpu_memory_buffers() const {
+    return have_persistent_gpu_memory_buffers_;
+  }
 
  protected:
   OneCopyTileTaskWorkerPool(base::SequencedTaskRunner* task_runner,
                             TaskGraphRunner* task_graph_runner,
+                            ContextProvider* context_provider,
                             ResourceProvider* resource_provider,
+                            ResourcePool* resource_pool,
                             int max_copy_texture_chromium_size,
-                            bool use_persistent_gpu_memory_buffers,
-                            int max_staging_buffers);
+                            bool have_persistent_gpu_memory_buffers);
 
  private:
-  struct StagingBuffer {
-    explicit StagingBuffer(const gfx::Size& size);
-    ~StagingBuffer();
-
-    void DestroyGLResources(gpu::gles2::GLES2Interface* gl);
-    void OnMemoryDump(base::trace_event::ProcessMemoryDump* pmd,
-                      ResourceFormat format,
-                      bool is_free) const;
-
-    const gfx::Size size;
-    scoped_ptr<gfx::GpuMemoryBuffer> gpu_memory_buffer;
-    base::TimeTicks last_usage;
-    unsigned texture_id;
-    unsigned image_id;
-    unsigned query_id;
-    uint64_t content_id;
+  struct CopyOperation {
+    typedef ScopedPtrDeque<CopyOperation> Deque;
+
+    CopyOperation(scoped_ptr<ResourceProvider::ScopedWriteLockGpuMemoryBuffer>
+                      src_write_lock,
+                  const Resource* src,
+                  const Resource* dst,
+                  const gfx::Rect& rect);
+    ~CopyOperation();
+
+    scoped_ptr<ResourceProvider::ScopedWriteLockGpuMemoryBuffer> src_write_lock;
+    const Resource* src;
+    const Resource* dst;
+    const gfx::Rect rect;
   };
 
-  scoped_ptr<StagingBuffer> AcquireStagingBuffer(const Resource* resource,
-                                                 uint64_t previous_content_id);
-  base::TimeTicks GetUsageTimeForLRUBuffer();
-  void ScheduleReduceMemoryUsage();
-  void ReduceMemoryUsage();
-  void ReleaseBuffersNotUsedSince(base::TimeTicks time);
-
   void OnTaskSetFinished(TaskSet task_set);
+  void AdvanceLastFlushedCopyTo(CopySequenceNumber sequence);
+  void IssueCopyOperations(int64 count);
+  void ScheduleCheckForCompletedCopyOperationsWithLockAcquired(
+      bool wait_if_needed);
+  void CheckForCompletedCopyOperations(bool wait_if_needed);
   scoped_refptr<base::trace_event::ConvertableToTraceFormat> StateAsValue()
       const;
   void StagingStateAsValueInto(
@@ -129,29 +124,31 @@ class CC_EXPORT OneCopyTileTaskWorkerPool
   TaskGraphRunner* task_graph_runner_;
   const NamespaceToken namespace_token_;
   TileTaskRunnerClient* client_;
-  ResourceProvider* const resource_provider_;
+  ContextProvider* context_provider_;
+  ResourceProvider* resource_provider_;
+  ResourcePool* resource_pool_;
   const int max_bytes_per_copy_operation_;
-  const bool use_persistent_gpu_memory_buffers_;
+  const bool have_persistent_gpu_memory_buffers_;
   TaskSetCollection tasks_pending_;
   scoped_refptr<TileTask> task_set_finished_tasks_[kNumberOfTaskSets];
+  CopySequenceNumber last_issued_copy_operation_;
+  CopySequenceNumber last_flushed_copy_operation_;
 
   // Task graph used when scheduling tasks and vector used to gather
   // completed tasks.
   TaskGraph graph_;
   Task::Vector completed_tasks_;
 
-  mutable base::Lock lock_;
+  base::Lock lock_;
   // |lock_| must be acquired when accessing the following members.
-  using StagingBufferSet = std::set<StagingBuffer*>;
-  StagingBufferSet buffers_;
-  using StagingBufferDeque = ScopedPtrDeque<StagingBuffer>;
-  StagingBufferDeque free_buffers_;
-  StagingBufferDeque busy_buffers_;
+  base::ConditionVariable copy_operation_count_cv_;
   int bytes_scheduled_since_last_flush_;
-  size_t max_staging_buffers_;
-  const base::TimeDelta staging_buffer_expiration_delay_;
-  bool reduce_memory_usage_pending_;
-  base::Closure reduce_memory_usage_callback_;
+  size_t issued_copy_operation_count_;
+  CopyOperation::Deque pending_copy_operations_;
+  CopySequenceNumber next_copy_operation_sequence_;
+  bool check_for_completed_copy_operations_pending_;
+  base::TimeTicks last_check_for_completed_copy_operations_time_;
+  bool shutdown_;
 
   base::WeakPtrFactory<OneCopyTileTaskWorkerPool> weak_ptr_factory_;
   // "raster finished" tasks need their own factory as they need to be
diff --git a/cc/raster/tile_task_worker_pool_perftest.cc b/cc/raster/tile_task_worker_pool_perftest.cc
index 1262ace..b49c498 100644
--- a/cc/raster/tile_task_worker_pool_perftest.cc
+++ b/cc/raster/tile_task_worker_pool_perftest.cc
@@ -265,11 +265,13 @@ class TileTaskWorkerPoolPerfTest
         break;
       case TILE_TASK_WORKER_POOL_TYPE_ONE_COPY:
         Create3dOutputSurfaceAndResourceProvider();
+        staging_resource_pool_ = ResourcePool::Create(resource_provider_.get(),
+                                                      GL_TEXTURE_2D);
         tile_task_worker_pool_ = OneCopyTileTaskWorkerPool::Create(
             task_runner_.get(), task_graph_runner_.get(),
             context_provider_.get(), resource_provider_.get(),
-            std::numeric_limits<int>::max(), false,
-            std::numeric_limits<int>::max());
+            staging_resource_pool_.get(), std::numeric_limits<int>::max(),
+            false);
         break;
       case TILE_TASK_WORKER_POOL_TYPE_GPU:
         Create3dOutputSurfaceAndResourceProvider();
@@ -429,6 +431,7 @@ class TileTaskWorkerPoolPerfTest
     return std::string();
   }
 
+  scoped_ptr<ResourcePool> staging_resource_pool_;
   scoped_ptr<TileTaskWorkerPool> tile_task_worker_pool_;
   TestGpuMemoryBufferManager gpu_memory_buffer_manager_;
   TestSharedBitmapManager shared_bitmap_manager_;
diff --git a/cc/raster/tile_task_worker_pool_unittest.cc b/cc/raster/tile_task_worker_pool_unittest.cc
index bf2a82a..6038b00 100644
--- a/cc/raster/tile_task_worker_pool_unittest.cc
+++ b/cc/raster/tile_task_worker_pool_unittest.cc
@@ -41,7 +41,6 @@ namespace {
 
 const size_t kMaxTransferBufferUsageBytes = 10000U;
 const size_t kMaxBytesPerCopyOperation = 1000U;
-const size_t kMaxStagingBuffers = 32U;
 
 // A resource of this dimension^2 * 4 must be greater than the above transfer
 // buffer constant.
@@ -166,10 +165,12 @@ class TileTaskWorkerPoolTest
         break;
       case TILE_TASK_WORKER_POOL_TYPE_ONE_COPY:
         Create3dOutputSurfaceAndResourceProvider();
+        staging_resource_pool_ = ResourcePool::Create(resource_provider_.get(),
+                                                      GL_TEXTURE_2D);
         tile_task_worker_pool_ = OneCopyTileTaskWorkerPool::Create(
             base::ThreadTaskRunnerHandle::Get().get(), &task_graph_runner_,
             context_provider_.get(), resource_provider_.get(),
-            kMaxBytesPerCopyOperation, false, kMaxStagingBuffers);
+            staging_resource_pool_.get(), kMaxBytesPerCopyOperation, false);
         break;
       case TILE_TASK_WORKER_POOL_TYPE_GPU:
         Create3dOutputSurfaceAndResourceProvider();
@@ -331,6 +332,7 @@ class TileTaskWorkerPoolTest
   FakeOutputSurfaceClient output_surface_client_;
   scoped_ptr<FakeOutputSurface> output_surface_;
   scoped_ptr<ResourceProvider> resource_provider_;
+  scoped_ptr<ResourcePool> staging_resource_pool_;
   scoped_ptr<TileTaskWorkerPool> tile_task_worker_pool_;
   TestGpuMemoryBufferManager gpu_memory_buffer_manager_;
   TestSharedBitmapManager shared_bitmap_manager_;