Fixing zero timeout handling for fence functions

Christian König deathsimple at vodafone.de
Tue Oct 25 12:25:07 UTC 2016


Hi Sumit,

sending this once more with all the patches in once set, cause the last one
turned out to be a bit chaotic because I send from the wrong branch.

The following patch set fixes the handling in the fence and reservation object
wait function when the timeout is zero.

An AMD developer introduced this a while ago to work around some issues in TTM
and our amdgpu driver, but essentially this effort was a bit flawed because
even with a zero timeout enable_signalling() should be called.

Otherwise someone busy waiting for the fence might never be signaled when you
have hardware with faulty interrupts for example.

Please review and/or comment,
Christian.



More information about the dri-devel mailing list