[Intel-gfx] [PATCH 11/11] drm/i915: Allow userspace to request an object at a specific offset

Daniel Vetter daniel at ffwll.ch
Wed Jan 16 17:55:52 CET 2013


On Tue, Jan 08, 2013 at 10:53:19AM +0000, Chris Wilson wrote:
> Certain workarounds and workloads require objects at specific or at
> least known offsets. Privileged users could pin an object into the GTT,
> but that has obvious limitations for the general case. Instead, the user
> can construct a batch assuming a particular layout for an object and
> request that the kernel try its utmost to provide the object at that
> location. This has the advantage that not only can it fail, but also
> such allocations are transitory - although contention should be rare and
> the object persist at the same location between batches. The benefit for
> userspace is that it can then avoid all relocations referencing this
> object as it resides at a known space - this becomes even more useful
> with per-process GTT spaces where there will be virtually no contention
> between applications.
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>

I'm unsure about api implications of this one here, and slightly afraid of
userspace managing to put itself into an ugly corner. Also, I'm not sure
whether this is the interface we want for real ppgtt, a simple (real) pin
interface which reserves the ppgtt address (but doesn't necessarily force
the backing storage to be pinned) feels saner.

So until I see more torches&pitchforks from my appartment, I'd like to
defer this until we have real ppgtt available ...
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch



More information about the Intel-gfx mailing list