We've never tested our device reset code on TreeHerder, and this is a good time to start now that we will have a very simple way of creating failures: we can kill the GPU process. We don't just want to test that we can recover at the top of an event loop: we should be able to test the GPU process dying in the middle of an IPDL transaction, or while decoding video, or in the middle of a texture operation. Normal unit or integration tests would be a good place to start. Matt also mentioned a chaos mode on tree herder might help: We could randomly crash the GPU process during GPU entry-points and make sure we can recover to paint another frame.
Kevin, as I talked with david, this is another good practice to extend testing coverage of composition, especially with gpu process. Please have a look at it.
Assignee: nobody → kechen
You need to log in before you can comment on or make changes to this bug.