GitLab

Texture::Buffer is using the deprecated ContextFactoryObserver interface
to find out about main thread context loss. The ContextFactory adds
itself as a ContextLostObserver, then on context loss forwards that
message along to all the ContextFactoryObservers.

There are multiple message pipes between the browser and GPU process
that have indeterminate ordering for when they find out about connection
errors. This leads to the following ordering problem:

1. The GPU process crashes. The old shared main thread
   RasterContextProvider is lost and any message pipes between the
   browser and GPU process will encounter connection errors.
2. A new shared main thread RasterContextProvider is created.
3. Exo allocates a new Buffer::Texture using the new shared main thread
   RasterContextProvider.
4. The new Buffer::Texture adds itself as a ContextFactoryObserver.
5. All ContextFactoryObservers get notified about OnLostSharedContext()
   for the old shared main thread RasterContextProvider loss. For the
   new Buffer::Texture this is wrong, the RasterContextProvider is still
   valid.
6. The new Buffer::Texture destroys all of it's resources incorrectly.
   This triggers a flood of GL errors from the GPU process.

Make Buffer::Texture a ContextLossObserver instead. This way the
Buffer::Texture gets notified about context loss for only the
RasterContextProvider it's using.

Bug: 954151, 944597
Test: BufferTest.OnLostResources
Change-Id: Ic68dbaab8864b835b2bbb83b3b43f03495902aef
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/1597390
Commit-Queue: kylechar <kylechar@chromium.org>
Reviewed-by: Daniele Castagna <dcastagna@chromium.org>
Cr-Commit-Position: refs/heads/master@{#657432}