> >>>> + /** > >>>> + * @DRM_ASAHI_BIND_SINGLE_PAGE: Map a single page of the BO repeatedly > >>>> + * across the VA range. > >>>> + * > >>>> + * This is useful to fill a VA range with scratch pages or zero pages. > >>>> + * It is intended as a mechanism to accelerate sparse. > >>>> + */ > >>>> + DRM_ASAHI_BIND_SINGLE_PAGE = (1L << 2), > >> > >> Does this require the BO to be a single page? If so, does it require offset==0? Or does it just take whatever page is at the specified offset? > > > > I believe the intention is that it takes whatever page is at the > > specified offset and just maps that a bunch of times. HK doesn't use > > this yet though it probably should (this was added to help reduce > > overhead when emulating sparse with scratch/zero pages, which is still > > very new functionality in hk). > > > > Accelerating this properly involves GPUVM patches - although even without > > that, moving the loop into the kernel so it's only a single ioctl > > (user-kernel roundtrip) seems worth keeping the flag for. > > FYI: I will be posting a patch for the GPUVM abstraction soon. Great to hear :) Although in this case, I meant that accelerating DRM_ASAHI_BIND_SINGLE_PAGE requires patches to extend the actual C implementation of drm/gpuvm, not just the Rust abstraction. Which is a bit annoying for non-essential functionality with regards to upstreaming things...
