On Fri, Apr 24, 2015 at 07:45:23AM +0100, Chris Wilson wrote: > v2: Akash worried that we were discarding the serialisation point with > outstanding GTT writes before doing the unbind, so restore the wmb() for > the writes. We only lose the write domain now, and not both - 1 one > sided barrier now, rather than 2 full barriers previously.
More thought says that worrying about outstanding kernel writes (which I was since we explicitly flush userspace mmaps) is pointless since we use iomap_atomic - i.e. we also explicitly flush writes upon unmap, inside pwrite, reloc etc. So we can go back to v1! :) -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
