[PATCH 3/3] drm/i915: Use CPU mapping for userspace dma-buf mmap()

Chris Wilson chris at chris-wilson.co.uk
Thu Aug 6 02:13:02 PDT 2015

On Thu, Aug 06, 2015 at 10:09:37AM +0100, Chris Wilson wrote:
> On Wed, Aug 05, 2015 at 07:13:11PM -0300, Tiago Vignatti wrote:
> > Userspace is the one in charge of flush CPU by wrapping mmap with
> > begin{,end}_cpu_access.
> > 
> > v2: Remove LLC check cause we have dma-buf sync providers now. Also, fix return
> > before transferring ownership when mmap fails.
> i915 doesn't implement end_cpu_access() as required for your sync ioctl.

An important point here is that performance on !llc (and for certain
objects with llc) will be absymal, utterly absymal. So what's the
rationale for this compared to doing the opposite and mapping the client
memory (a normal mmap) into the GPU with llc/snooping.

Chris Wilson, Intel Open Source Technology Centre

More information about the dri-devel mailing list