Now with i915_reset_engine() marking the stalled request as guilty,
preemption timeout doesn't lead into a GPU hang death spiral; at the
loss of potentially resetting a context with no harm (in practice that
didn't work out!).
A bit ambivalent on the flip forcing reset, both the stutter and glitch
can be seen under load. I'm not sold on the UX; it's bad either way.
If we don't break the deadlock, the user can't interact with the system;
if we do, they may see a glitch in one app.
Intel-gfx mailing list