Quoting Tvrtko Ursulin (2018-05-04 10:15:20) > On 03/05/2018 20:50, Chris Wilson wrote: > The second hunk is a bit more difficult. It creates a longer softirq-off > section, which is a slight negative, and I am unsure how much it > actually closes the race with tasklets in practice. So it may be the > only benefit is to reduce fiddles of both preempt count and > local_bh_disable, to fiddle just one. Remember that local_bh_disable() is just a fiddle with preempt-count and that the local tasklet doesn't run during the !preemptible section anyway. So it does remove one preemption point (in clearing the preempt-count) but that is of dubious merit since we would then need to kick the submission tasklet again immediately (from the preempt-count pov) afterwards. -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
