On Wed, May 11, 2016 at 05:11:55PM +0300, Imre Deak wrote: > On Wed, 2016-04-20 at 20:00 +0100, Chris Wilson wrote: > > During hibernation, all objects will have had their page contents > > written to disk and then restored upon resume. This means that every > > page will be dirty and we need to treat all objects as being in the CPU > > domain and require their contents to be flushed before use. > > > > At present we only do so for those objects bound into the Global GTT, > > however we need to mark all allocated objects as being unclean. > > > > v2: Actually restrict the clflushing of object content to post-hibernate > > as we always call restore_gtt_mappings because we cannot trust the BIOS > > not to scribble over the GGTT (but we can be confident that they will > > not touch system pages i.e. normal shmemfs objects). > > Then it should be (also) done from the PM restore hook as that's the > one called during resuming from hibernation. The thaw hook is called > right after creating the hibernation image (before writing it to disk), > so you'll have (not-dirty) cached data in that case too, but processes > are freezed so nothing should use the corresponding objects. Ah. Right, s/thaw/restore indeed. > Perhaps the best would be to move the objects to the CPU domain already > in the freeze hook (called right before creating the hibernation image) > as that covers both the above restore and thaw cases and that would > also make resume a bit faster. Hmm. I think I have an idea that should make us both a bit happier... -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
