On Sat, Aug 09, 2014 at 05:37:23PM +0100, Chris Wilson wrote: > With ppgtt, it is no longer correct to mark an object as > map_and_fenceable if we simply unbind it from the global gtt. This has > consequences during execbuffer where we simply use > obj->map_and_fenceable in use_cpu_reloc() to decide which access method > to use for writing into the object. Now for a ppgtt object, > map_and_fenceable will be true when it is not bound into the ggtt but > only in a ppgtt, leading to an invalid access on a non-llc architecture. The last sentence confused me for some reason. I think the first two sentences were perfectly succinct in describing the problem. > > v2: Revamp and resend to ease future patches. > > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> > Cc: Ben Widawsky <benjamin.widawsky@xxxxxxxxx> > Cc: Daniel Vetter <daniel.vetter@xxxxxxxx> I didn't read the patch closely, but the problem is clear. It seems like the one hunk: > - if (use_cpu_reloc(obj)) > + if (use_cpu_reloc(obj) || !i915_gem_obj_ggtt_bound(obj)) would be sufficient to fix the problem. I had a local solution at one point (I'm too lazy to check what I've sent out these days) where I turned map_and_fenceable into a getter(). I thought that looked pretty nice as well. lgtm [snip] -- Ben Widawsky, Intel Open Source Technology Center _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/intel-gfx
