On Wed, Feb 08, 2012 at 06:17:00PM +0000, Chris Wilson wrote: > On Wed, 8 Feb 2012 19:02:47 +0100, Daniel Vetter <daniel at ffwll.ch> wrote: > > The issue is that the first one introduces a pretty decent perf > > regression, iirc Chris mentions something much large than 10x slowdown > > on certain cairo traces on sna. So we can only merge the first one > > together with the second one. > > As it turns out, the cairo-traces were successfully exploiting the > ringbuffer overflow to avoid work. The reason why it eventually broke is > that when I experimented with using the BLT and semaphores, I introduced > a dependency on accurate seqno processing, which was obviously broken by > the overwriting of commands in the ringbuffer. And so is consistent with > the hangs reported by Eugeni and Ben, but we have yet to see many more > examples in the wild. Otoh, I don't think this explains the > all-generation use-after-free bug. > > Net result, the 0.9s firefox-paintball result was a fantasy and the perf > gains provided by the second patch are only around 20% for the same > trace. Ergo -next material only if it passes scrutiny. Well, that explains a few things. In that case I'm fine with just merging the first patch for -fixes, I think there're quite a few bugs this does fix. -Daniel -- Daniel Vetter Mail: daniel at ffwll.ch Mobile: +41 (0)79 365 57 48
