Implemented new fence syncs which replaces the old sync points.

content/common/gpu/client/gpu_channel_host.cc

Index: content/common/gpu/client/gpu_channel_host.cc
diff --git a/content/common/gpu/client/gpu_channel_host.cc b/content/common/gpu/client/gpu_channel_host.cc
index 6b1706cdbb614fb320611d359ac5ceb45b6eb5f8..8189b092da242ac285ef6960b7e38d33a2536280 100644
--- a/content/common/gpu/client/gpu_channel_host.cc
+++ b/content/common/gpu/client/gpu_channel_host.cc
@@ -35,10 +35,14 @@ base::StaticAtomicSequenceNumber g_next_transfer_buffer_id;
 }  // namespace
 
 GpuChannelHost::StreamFlushInfo::StreamFlushInfo()
-    : flush_pending(false),
+    : next_stream_flush_id(1),
+      flushed_stream_flush_id(0),
+      verified_stream_flush_id(0),
+      flush_pending(false),
       route_id(MSG_ROUTING_NONE),
       put_offset(0),
-      flush_count(0) {}
+      flush_count(0),
+      flush_id(0) {}
 
 GpuChannelHost::StreamFlushInfo::~StreamFlushInfo() {}
 
@@ -127,7 +131,7 @@ bool GpuChannelHost::Send(IPC::Message* msg) {
   return result;
 }
 
-void GpuChannelHost::OrderingBarrier(
+uint32_t GpuChannelHost::OrderingBarrier(
     int32 route_id,
     int32 stream_id,
     int32 put_offset,
@@ -138,30 +142,37 @@ void GpuChannelHost::OrderingBarrier(
   AutoLock lock(context_lock_);
   StreamFlushInfo& flush_info = stream_flush_info_[stream_id];
   if (flush_info.flush_pending && flush_info.route_id != route_id)
-    InternalFlush(stream_id);
+    InternalFlush(flush_info);
 
   if (put_offset_changed) {
+    const uint32_t flush_id = flush_info.next_stream_flush_id++;
     flush_info.flush_pending = true;
     flush_info.route_id = route_id;
     flush_info.put_offset = put_offset;
     flush_info.flush_count = flush_count;
+    flush_info.flush_id = flush_id;
     flush_info.latency_info.insert(flush_info.latency_info.end(),
                                    latency_info.begin(), latency_info.end());
 
     if (do_flush)
-      InternalFlush(stream_id);
+      InternalFlush(flush_info);
+
+    return flush_id;
   }
+  return 0;
 }
 
-void GpuChannelHost::InternalFlush(int32 stream_id) {
+void GpuChannelHost::InternalFlush(StreamFlushInfo& flush_info) {
   context_lock_.AssertAcquired();
-  StreamFlushInfo& flush_info = stream_flush_info_[stream_id];
   DCHECK(flush_info.flush_pending);
+  DCHECK(flush_info.flushed_stream_flush_id < flush_info.flush_id);
   Send(new GpuCommandBufferMsg_AsyncFlush(
       flush_info.route_id, flush_info.put_offset, flush_info.flush_count,
       flush_info.latency_info));
   flush_info.latency_info.clear();
   flush_info.flush_pending = false;
+
+  flush_info.flushed_stream_flush_id = flush_info.flush_id;
 }
 
 scoped_ptr<CommandBufferProxyImpl> GpuChannelHost::CreateViewCommandBuffer(
@@ -290,8 +301,9 @@ void GpuChannelHost::DestroyCommandBuffer(
   RemoveRoute(route_id);
 
   AutoLock lock(context_lock_);
-  if (stream_flush_info_[stream_id].route_id == route_id)
-    stream_flush_info_.erase(stream_id);
+  StreamFlushInfo& flush_info = stream_flush_info_[stream_id];
+  if (flush_info.flush_pending && flush_info.route_id == route_id)
+    flush_info.flush_pending = false;
 }
 
 void GpuChannelHost::DestroyChannel() {
@@ -390,6 +402,49 @@ int32 GpuChannelHost::GenerateStreamID() {
   return next_stream_id_.GetNext();
 }
 
+uint32_t GpuChannelHost::ValidateFlushIDReachedServer(int32 stream_id) {
+  // Store what flush ids we will be validating for all streams.
+  base::hash_map<int32, uint32_t> validate_flushes;
+  {
+    AutoLock lock(context_lock_);
+    for (const auto& iter : stream_flush_info_) {
+      const int32 iter_stream_id = iter.first;
+      const StreamFlushInfo& flush_info = iter.second;
+      if (flush_info.flushed_stream_flush_id >
+          flush_info.verified_stream_flush_id) {
+        validate_flushes[iter_stream_id] = flush_info.flushed_stream_flush_id;

[piman, 2015/09/26 00:09:51]

If there is a pending flush on that stream, you also need to InternalFlush it,
otherwise the flush id isn't an guarantee that it's visible on the GPU process.

[David Yen, 2015/09/26 00:24:56]

The pending flushes here are ordering barriers which have not been flushed yet.
Those shouldn't be counted as "flushed" right?

[piman, 2015/09/26 00:34:51]

Oh, right. flushed_stream_flush_id is only set after the real flush. Maybe it's
ok to require a real flush instead of implicitly flushing pending
OrderingBarriers. Either way is reasonable.

+      }
+    }
+  }
+
+  if (Send(new GpuChannelMsg_Nop())) {

[piman, 2015/09/26 00:09:51]

nit: you can skip if validate_flushes is empty.

[David Yen, 2015/09/28 17:38:18]

Even better, I made it skip the validation if the current stream has no flushes
to validate. Other streams can validate themselves later and hopefully be able
to validate even more flushes.

+    // Update verified flush id for all streams.
+    uint32_t highest_flush_id = 0;
+    AutoLock lock(context_lock_);
+    for (const auto& iter : validate_flushes) {
+      const int32_t validated_stream_id = iter.first;
+      const uint32_t validated_flush_id = iter.second;
+      StreamFlushInfo& flush_info = stream_flush_info_[validated_stream_id];
+      if (flush_info.verified_stream_flush_id < validated_flush_id) {
+        flush_info.verified_stream_flush_id = validated_flush_id;
+      }
+
+      if (validated_stream_id == stream_id)
+        highest_flush_id = flush_info.verified_stream_flush_id;
+    }
+
+    return highest_flush_id;
+  }
+
+  return 0;
+}
+
+uint32_t GpuChannelHost::GetHighestValidatedFlushID(int32 stream_id) {
+  AutoLock lock(context_lock_);
+  StreamFlushInfo& flush_info = stream_flush_info_[stream_id];
+  return flush_info.verified_stream_flush_id;
+}
+
 GpuChannelHost::~GpuChannelHost() {
 #if DCHECK_IS_ON()
   AutoLock lock(context_lock_);