Delay GpuCommandBufferStub::AddDestructionObserver until we need it.

In certain failure cases at initialization time (caused by a race) we used to call
AddDestructionObserver before we would make the surface current, causing us to
destroy the surface before OnWillDestroyStub had a chance to be called, so we
wouldn't RemoveDestructionObserver and we'd call into a dead pointer.

This fixes it. By the time MakeCurrent is called, we know we have a ref to the
surface and it won't destroy it until GpuCommandBufferStub::Destroy, after
OnWillDestroyStub is called.

BUG=140502
TEST=attach debugger to renderer, add a breakpoint to CommandBufferProxyImpl::Initialize, right-click and "view page source" (breakpoint should hit), kill new tab, and only then continue breakpoint. Observe no GPU process crash.

Review URL: https://chromiumcodereview.appspot.com/10831169

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@150227 0039d316-1c4b-4281-b951-d872f2087c98

1 files changed, 1 insertions, 0 deletions

diff --git a/content/common/gpu/image_transport_surface.h b/content/common/gpu/image_transport_surface.h
index 4cbad57..348b5f6 100644
--- a/content/common/gpu/image_transport_surface.h
+++ b/content/common/gpu/image_transport_surface.h
@@ -131,6 +131,7 @@ class ImageTransportHelper : public IPC::Listener {
   void Suspend();
 
   GpuChannelManager* manager() const { return manager_; }
+  GpuCommandBufferStub* stub() const { return stub_.get(); }
 
  private:
   gpu::GpuScheduler* Scheduler();