[Intel-gfx] [PATCH 1/5] drm/i915: Stop engines around GPU reset preparations

Chris Wilson chris at chris-wilson.co.uk
Sat Mar 3 08:46:03 UTC 2018

Quoting Chris Wilson (2018-03-02 14:32:42)
> As we make preparations to reset the GPU state, we assume that the GPU
> is hung and will not advance. Make this assumption more explicit by
> setting the STOP_RING bit on the engines as part of our early reset
> preparations.

I had to drop this one as it causes a CS arbitration event (i.e. context
switch) when the CS parser hits the STOP_RING. As we were doing the stop
prior to disabling the tasklet, this meant we were handling a spurious
interrupt. Note this means that we are generating spurious interrupts
prior to the reset currently, just we are effectively masking them :)

As for the problem it was trying to solve? I am thinking along the lines
of more set-wedge races that used to be prevented by the stop_machine().

More information about the Intel-gfx mailing list