[Intel-gfx] [PATCH 13/13] drm/i915: Support explicit fencing for execbuf

[Chris Wilson]

On Fri, Aug 26, 2016 at 04:29:57PM +0100, John Harrison wrote:
> On 26/08/2016 16:08, John Harrison wrote:
> >On 25/08/2016 10:08, Chris Wilson wrote:
> >>Now that the user can opt-out of implicit fencing, we need to give them
> >>back control over the fencing. We employ sync_file to wrap our
> >>drm_i915_gem_request and provide an fd that userspace can merge with
> >>other sync_file fds and pass back to the kernel to wait upon before
> >>future execution.
> It is worth mentioning in the description that this isn't just
> useful for pushing the synchronisation down to user land. It also
> allows synchronisation with non-rendering operations.

But we already cover that in dma-buf. This is only useful for explicit
synchronisation outside of the kernel, since we are trying to get
everybody inside talking dma-buf.

Software controlled fences are a no-go due to resitance from devs not
wanting userspace being able to arbitrary hang the kernel. From that
perspective, the only source of a fence is from the kernel. And the only
desire to use explicit fencing is because implicit may be too coarse, or
too accurate depending on your pov ;-)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre