Fixed flakiness of context_lost tests on GPU bots.

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 afc0b5993c7e548060235b95bf1b865bf92bbe66..f503051fb162ee42d3f8ab442de42edc47265fdb 100644
--- a/content/common/gpu/client/gpu_channel_host.cc
+++ b/content/common/gpu/client/gpu_channel_host.cc
@@ -144,6 +144,19 @@ CommandBufferProxyImpl* GpuChannelHost::CreateViewCommandBuffer(
   int32 route_id = GenerateRouteID();
   if (!factory_->CreateViewCommandBuffer(surface_id, init_params, route_id)) {
     LOG(ERROR) << "GpuChannelHost::CreateViewCommandBuffer failed.";
+
+    // The most likely reason CreateViewCommandBuffer will fail is
+    // that the GPU process crashed. In this case the GPU channel
+    // needs to be considered lost. The caller will then set up a new
+    // connection, and the GPU channel and any view command buffers
+    // will all be associated with the same GPU process.
+    DCHECK(MessageLoopProxy::current().get());
+
+    scoped_refptr<base::MessageLoopProxy> io_loop = factory_->GetIOLoopProxy();
+    io_loop->PostTask(FROM_HERE,
+                      base::Bind(&GpuChannelHost::MessageFilter::OnChannelError,
+                                 channel_filter_.get()));
+
     return NULL;
   }