cc: Mark async texture uploads as completed from the upload thread.

This reduces the latency between when an upload completes and when the
client is notified.

BUG=173802


Review URL: https://chromiumcodereview.appspot.com/12213073

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@181883 0039d316-1c4b-4281-b951-d872f2087c98

gpu/command_buffer/service/async_pixel_transfer_delegate_mock.h

@@ -32,8 +32,9 @@ class MockAsyncPixelTransferDelegate : public gfx::AsyncPixelTransferDelegate {
   // Implement AsyncPixelTransferDelegate.
   MOCK_METHOD1(CreateRawPixelTransferState,
       gfx::AsyncPixelTransferState*(GLuint service_id));
-  MOCK_METHOD1(AsyncNotifyCompletion,
-      void(const base::Closure& task));
+  MOCK_METHOD2(AsyncNotifyCompletion,
+      void(const AsyncMemoryParams& mem_params,
+           const CompletionCallback& callback));
   MOCK_METHOD3(AsyncTexImage2D,
       void(gfx::AsyncPixelTransferState*,
           const AsyncTexImage2DParams& tex_params,

gpu/command_buffer/service/gles2_cmd_decoder.cc

@@ -2719,6 +2719,9 @@ bool GLES2DecoderImpl::MakeCurrent() {
     return false;
   }
 
+  if (query_manager_.get())
+    query_manager_->ProcessPendingTransferQueries();
+
   // TODO(epenner): Is there a better place to do this? Transfers
   // can complete any time we yield the main thread. So we *must*
   // process transfers after any such yield, before resuming.
@@ -9116,6 +9119,8 @@ error::Error GLES2DecoderImpl::HandleEndQueryEXT(
     return error::kOutOfBounds;
   }
 
+  query_manager_->ProcessPendingTransferQueries();
+
   state_.current_query = NULL;
   return error::kNoError;
 }

gpu/command_buffer/service/query_manager.cc

@@ -3,13 +3,15 @@
 // found in the LICENSE file.
 
 #include "gpu/command_buffer/service/query_manager.h"
+
 #include "base/atomicops.h"
 #include "base/bind.h"
 #include "base/logging.h"
+#include "base/shared_memory.h"
 #include "base/time.h"
 #include "gpu/command_buffer/common/gles2_cmd_format.h"
-#include "gpu/command_buffer/service/gles2_cmd_decoder.h"
 #include "gpu/command_buffer/service/feature_info.h"
+#include "gpu/command_buffer/service/gles2_cmd_decoder.h"
 #include "ui/gl/async_pixel_transfer_delegate.h"
 
 namespace gpu {
@@ -176,10 +178,20 @@ class AsyncPixelTransfersCompletedQuery
   virtual bool Process() OVERRIDE;
   virtual void Destroy(bool have_context) OVERRIDE;
 
-  void MarkAsCompletedCallback() { MarkAsCompleted(1); }
-
  protected:
   virtual ~AsyncPixelTransfersCompletedQuery();
+
+  static void MarkAsCompletedThreadSafe(
+      uint32 submit_count, const gfx::AsyncMemoryParams& mem_params) {
+    DCHECK(mem_params.shared_memory);
+    DCHECK(mem_params.shared_memory->memory());
+    void *data = static_cast<int8*>(mem_params.shared_memory->memory()) +
+        mem_params.shm_data_offset;
+    QuerySync* sync = static_cast<QuerySync*>(data);
+
+    // No need for a MemoryBarrier here as sync->result is not written.
+    sync->process_count = submit_count;
+  }
 };
 
 AsyncPixelTransfersCompletedQuery::AsyncPixelTransfersCompletedQuery(
@@ -192,28 +204,41 @@ bool AsyncPixelTransfersCompletedQuery::Begin() {
 }
 
 bool AsyncPixelTransfersCompletedQuery::End(uint32 submit_count) {
-  MarkAsPending(submit_count);
-
-  // This will call MarkAsCompleted(1) as a reply to a task on
-  // the async upload thread, such that it occurs after all previous
-  // async transfers have completed.
+  gfx::AsyncMemoryParams mem_params;
+  // Get the real shared memory since it might need to be duped to prevent
+  // use-after-free of the memory.
+  Buffer buffer = manager()->decoder()->GetSharedMemoryBuffer(shm_id());
+  if (!buffer.shared_memory)
+    return false;
+  mem_params.shared_memory = buffer.shared_memory;
+  mem_params.shm_size = buffer.size;
+  mem_params.shm_data_offset = shm_offset();
+  mem_params.shm_data_size = sizeof(QuerySync);
+
+  // Ask AsyncPixelTransferDelegate to run completion callback after all
+  // previous async transfers are done. No guarantee that callback is run
+  // on the current thread.
   manager()->decoder()->GetAsyncPixelTransferDelegate()->AsyncNotifyCompletion(
-      base::Bind(
-          &AsyncPixelTransfersCompletedQuery::MarkAsCompletedCallback,
-          AsWeakPtr()));
-
-  // TODO(epenner): The async task occurs outside the normal
-  // flow, via a callback on this thread. Is there anything
-  // missing or wrong with that?
+      mem_params,
+      base::Bind(AsyncPixelTransfersCompletedQuery::MarkAsCompletedThreadSafe,
+                 submit_count));
 
-  // TODO(epenner): Could we possibly trigger the completion on
-  // the upload thread by writing to the query shared memory
-  // directly?
-  return true;
+  return AddToPendingTransferQueue(submit_count);
 }
 
 bool AsyncPixelTransfersCompletedQuery::Process() {
-  NOTREACHED();
+  QuerySync* sync = manager()->decoder()->GetSharedMemoryAs<QuerySync*>(
+      shm_id(), shm_offset(), sizeof(*sync));
+  if (!sync)
+    return false;
+
+  // Check if completion callback has been run. sync->process_count atomicity
+  // is guaranteed as this is already used to notify client of a completed
+  // query.
+  if (sync->process_count != submit_count())
+    return true;
+
+  UnmarkAsPending();
   return true;
 }
 
@@ -435,7 +460,7 @@ bool QueryManager::ProcessPendingQueries() {
       return false;
     }
     if (query->pending()) {
-      return true;
+      break;
     }
     pending_queries_.pop_front();
   }
@@ -447,6 +472,25 @@ bool QueryManager::HavePendingQueries() {
   return !pending_queries_.empty();
 }
 
+bool QueryManager::ProcessPendingTransferQueries() {
+  while (!pending_transfer_queries_.empty()) {
+    Query* query = pending_transfer_queries_.front().get();
+    if (!query->Process()) {
+      return false;
+    }
+    if (query->pending()) {
+      break;
+    }
+    pending_transfer_queries_.pop_front();
+  }
+
+  return true;
+}
+
+bool QueryManager::HavePendingTransferQueries() {
+  return !pending_transfer_queries_.empty();
+}
+
 bool QueryManager::AddPendingQuery(Query* query, uint32 submit_count) {
   DCHECK(query);
   DCHECK(!query->IsDeleted());
@@ -458,6 +502,17 @@ bool QueryManager::AddPendingQuery(Query* query, uint32 submit_count) {
   return true;
 }
 
+bool QueryManager::AddPendingTransferQuery(Query* query, uint32 submit_count) {
+  DCHECK(query);
+  DCHECK(!query->IsDeleted());
+  if (!RemovePendingQuery(query)) {
+    return false;
+  }
+  query->MarkAsPending(submit_count);
+  pending_transfer_queries_.push_back(query);
+  return true;
+}
+
 bool QueryManager::RemovePendingQuery(Query* query) {
   DCHECK(query);
   if (query->pending()) {
@@ -471,6 +526,13 @@ bool QueryManager::RemovePendingQuery(Query* query) {
         break;
       }
     }
+    for (QueryQueue::iterator it = pending_transfer_queries_.begin();
+         it != pending_transfer_queries_.end(); ++it) {
+      if (it->get() == query) {
+        pending_transfer_queries_.erase(it);
+        break;
+      }
+    }
     if (!query->MarkAsCompleted(0)) {
       return false;
     }

gpu/command_buffer/service/query_manager.h

@@ -89,11 +89,21 @@ class GPU_EXPORT QueryManager {
       submit_count_ = submit_count;
     }
 
+    void UnmarkAsPending() {
+      DCHECK(pending_);
+      pending_ = false;
+    }
+
     // Returns false if shared memory for sync is invalid.
     bool AddToPendingQueue(uint32 submit_count) {
       return manager_->AddPendingQuery(this, submit_count);
     }
 
+    // Returns false if shared memory for sync is invalid.
+    bool AddToPendingTransferQueue(uint32 submit_count) {
+      return manager_->AddPendingTransferQuery(this, submit_count);
+    }
+
     void BeginQueryHelper(GLenum target, GLuint id) {
       manager_->BeginQueryHelper(target, id);
     }
@@ -102,15 +112,15 @@ class GPU_EXPORT QueryManager {
       manager_->EndQueryHelper(target);
     }
 
+    uint32 submit_count() const {
+      return submit_count_;
+    }
+
    private:
     friend class QueryManager;
     friend class QueryManagerTest;
     friend class base::RefCounted<Query>;
 
-    uint32 submit_count() const {
-      return submit_count_;
-    }
-
     // The manager that owns this Query.
     QueryManager* manager_;
 
@@ -162,6 +172,13 @@ class GPU_EXPORT QueryManager {
   // True if there are pending queries.
   bool HavePendingQueries();
 
+  // Processes pending transfer queries. Returns false if any queries are
+  // pointing to invalid shared memory.
+  bool ProcessPendingTransferQueries();
+
+  // True if there are pending transfer queries.
+  bool HavePendingTransferQueries();
+
   GLES2Decoder* decoder() const {
     return decoder_;
   }
@@ -179,6 +196,10 @@ class GPU_EXPORT QueryManager {
   // Returns false if any query is pointing to invalid shared memory.
   bool AddPendingQuery(Query* query, uint32 submit_count);
 
+  // Adds to queue of transfer queries waiting for completion.
+  // Returns false if any query is pointing to invalid shared memory.
+  bool AddPendingTransferQuery(Query* query, uint32 submit_count);
+
   // Removes a query from the queue of pending queries.
   // Returns false if any query is pointing to invalid shared memory.
   bool RemovePendingQuery(Query* query);
@@ -205,6 +226,9 @@ class GPU_EXPORT QueryManager {
   typedef std::deque<Query::Ref> QueryQueue;
   QueryQueue pending_queries_;
 
+  // Async pixel transfer queries waiting for completion.
+  QueryQueue pending_transfer_queries_;
+
   DISALLOW_COPY_AND_ASSIGN(QueryManager);
 };
 

ui/gl/async_pixel_transfer_delegate.h

@@ -78,6 +78,8 @@ class GL_EXPORT AsyncPixelTransferState :
 
 class GL_EXPORT AsyncPixelTransferDelegate {
  public:
+  typedef base::Callback<void(const AsyncMemoryParams&)> CompletionCallback;
+
   static scoped_ptr<AsyncPixelTransferDelegate>
       Create(gfx::GLContext* context);
   virtual ~AsyncPixelTransferDelegate() {}
@@ -89,8 +91,10 @@ class GL_EXPORT AsyncPixelTransferDelegate {
     return make_scoped_ptr(CreateRawPixelTransferState(texture_id));
   }
 
+  // There's no guarantee that callback will run on the caller thread.
   virtual void AsyncNotifyCompletion(
-      const base::Closure& notify_task) = 0;
+      const AsyncMemoryParams& mem_params,
+      const CompletionCallback& callback) = 0;
 
   virtual void AsyncTexImage2D(
       AsyncPixelTransferState* state,

ui/gl/async_pixel_transfer_delegate_android.cc

@@ -159,8 +159,7 @@ base::MessageLoopProxy* transfer_message_loop_proxy() {
 // Class which holds async pixel transfers state (EGLImage).
 // The EGLImage is accessed by either thread, but everything
 // else accessed only on the main thread.
-class TransferStateInternal
-    : public base::RefCountedThreadSafe<TransferStateInternal> {
+class TransferStateInternal : public base::RefCounted<TransferStateInternal> {
  public:
   explicit TransferStateInternal(GLuint texture_id,
                                  bool wait_for_uploads,
@@ -251,7 +250,7 @@ class TransferStateInternal
   }
 
  protected:
-  friend class base::RefCountedThreadSafe<TransferStateInternal>;
+  friend class base::RefCounted<TransferStateInternal>;
   friend class AsyncPixelTransferDelegateAndroid;
 
   static void DeleteTexture(GLuint id) {
@@ -330,7 +329,8 @@ class AsyncPixelTransferDelegateAndroid
 
   // implement AsyncPixelTransferDelegate:
   virtual void AsyncNotifyCompletion(
-      const base::Closure& task) OVERRIDE;
+      const AsyncMemoryParams& mem_params,
+      const CompletionCallback& callback) OVERRIDE;
   virtual void AsyncTexImage2D(
       AsyncPixelTransferState* state,
       const AsyncTexImage2DParams& tex_params,
@@ -360,6 +360,12 @@ class AsyncPixelTransferDelegateAndroid
       AsyncTexSubImage2DParams tex_params,
       base::SharedMemory* shared_memory,
       uint32 shared_memory_data_offset);
+  static void PerformNotifyCompletion(
+      base::SharedMemory* shared_memory,
+      uint32 shared_memory_size,
+      uint32 shared_memory_data_offset,
+      uint32 shared_memory_data_size,
+      const CompletionCallback& callback);
 
   // Returns true if a work-around was used.
   bool WorkAroundAsyncTexImage2D(
@@ -444,19 +450,23 @@ AsyncPixelTransferState*
                                     use_image_preserved));
 }
 
-namespace {
-// Dummy function to measure completion on
-// the upload thread.
-void NoOp() {}
-} // namespace
-
 void AsyncPixelTransferDelegateAndroid::AsyncNotifyCompletion(
-      const base::Closure& task) {
-  // Post a no-op task to the upload thread followed
-  // by a reply to the callback. The reply will then occur after
-  // all async transfers are complete.
-  transfer_message_loop_proxy()->PostTaskAndReply(FROM_HERE,
-      base::Bind(&NoOp), task);
+    const AsyncMemoryParams& mem_params,
+    const CompletionCallback& callback) {
+  DCHECK(mem_params.shared_memory);
+  DCHECK_LE(mem_params.shm_data_offset + mem_params.shm_data_size,
+            mem_params.shm_size);
+  // Post a PerformNotifyCompletion task to the upload thread. This task
+  // will run after all async transfers are complete.
+  transfer_message_loop_proxy()->PostTask(
+      FROM_HERE,
+      base::Bind(&AsyncPixelTransferDelegateAndroid::PerformNotifyCompletion,
+                 base::Owned(DuplicateSharedMemory(mem_params.shared_memory,
+                                                   mem_params.shm_size)),
+                 mem_params.shm_size,
+                 mem_params.shm_data_offset,
+                 mem_params.shm_data_size,
+                 callback));
 }
 
 void AsyncPixelTransferDelegateAndroid::AsyncTexImage2D(
@@ -660,6 +670,20 @@ void AsyncPixelTransferDelegateAndroid::PerformAsyncTexSubImage2D(
   state->last_transfer_time_ = base::TimeTicks::HighResNow() - begin_time;
 }
 
+void AsyncPixelTransferDelegateAndroid::PerformNotifyCompletion(
+    base::SharedMemory* shared_memory,
+    uint32 shared_memory_size,
+    uint32 shared_memory_data_offset,
+    uint32 shared_memory_data_size,
+    const CompletionCallback& callback) {
+  TRACE_EVENT0("gpu", "PerformNotifyCompletion");
+  gfx::AsyncMemoryParams mem_params;
+  mem_params.shared_memory = shared_memory;
+  mem_params.shm_size = shared_memory_size;
+  mem_params.shm_data_offset = shared_memory_data_offset;
+  mem_params.shm_data_size = shared_memory_data_size;
+  callback.Run(mem_params);
+}
 
 namespace {
 bool IsPowerOfTwo (unsigned int x) {

ui/gl/async_pixel_transfer_delegate_stub.cc

@@ -77,8 +77,9 @@ AsyncPixelTransferState*
 }
 
 void AsyncPixelTransferDelegateStub::AsyncNotifyCompletion(
-    const base::Closure& task) {
-  task.Run();
+    const AsyncMemoryParams& mem_params,
+    const CompletionCallback& callback) {
+  callback.Run(mem_params);
 }
 
 void AsyncPixelTransferDelegateStub::AsyncTexImage2D(

ui/gl/async_pixel_transfer_delegate_stub.h

@@ -36,7 +36,9 @@ class AsyncPixelTransferDelegateStub : public AsyncPixelTransferDelegate {
   virtual ~AsyncPixelTransferDelegateStub();
 
   // implement AsyncPixelTransferDelegate:
-  virtual void AsyncNotifyCompletion(const base::Closure& task) OVERRIDE;
+  virtual void AsyncNotifyCompletion(
+      const AsyncMemoryParams& mem_params,
+      const CompletionCallback& callback) OVERRIDE;
   virtual void AsyncTexImage2D(
       AsyncPixelTransferState* transfer_state,
       const AsyncTexImage2DParams& tex_params,