[Intel-gfx] [PATCH 3/3] drm/vgem: Attach sw fences to exported vGEM dma-buf (ioctl)

[Daniel Vetter]

On Tue, Jul 12, 2016 at 12:04:03PM +0100, Chris Wilson wrote:
> On Tue, Jul 12, 2016 at 12:44:17PM +0200, Daniel Vetter wrote:
> > On Mon, Jul 11, 2016 at 04:24:45PM +0100, Chris Wilson wrote:
> > > That doesn't fit the out-of-order unbound nature of the interface. The
> > > interface is just a collection of fences that userspace associates with
> > > the buffer that it may signal at any time. (Having no strict timeline is
> > > an advantage!)
> > 
> > Fences on the same timeline are supposed to be signalled in-order. If you
> > want full out-of-order fences then you need to grab a new timeline number
> > for each one. Drivers can and do merge fences on the same timeline and
> > just carry the one with the largest seqno around.
> 
> Ugh. Timelines simply don't mean anything everywhere - a very leaky
> abstration.
> 
> Nevertheless, a fence_context per vgem_fence would do the trick.

Yeah it's a bit meh, but allocating plenty of them is how we currently
cope with it. I suggested that we have a special FENCE_TIMELINE_UNORDERED
flag (we need it for fence_array too), but that wasn't popular. I still
expect it to happen eventually ;-)
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch