[Intel-gfx] [PATCH v4 1/7] drm/i915: Implement a framework for batch buffer pools

[Daniel Vetter]

On Fri, Nov 21, 2014 at 05:28:11PM -0800, Michael H. Nguyen wrote:
> Hi Daniel, Chris
> 
> On 11/12/2014 08:38 AM, Chris Wilson wrote:
> >On Wed, Nov 12, 2014 at 05:33:08PM +0100, Daniel Vetter wrote:
> >>On Wed, Nov 12, 2014 at 10:46 AM, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> >>>On Wed, Nov 12, 2014 at 09:44:34AM +0100, Daniel Vetter wrote:
> >>>>On Fri, Nov 07, 2014 at 02:22:01PM -0800, bradley.d.volkin at intel.com wrote:
> >>>>>+           if (obj && obj->madv == __I915_MADV_PURGED) {
> >>>>>+                   was_purged = true;
> >>>>>+                   list_del(&obj->batch_pool_list);
> >>>>>+                   drm_gem_object_unreference(&obj->base);
> >>>>>+                   obj = NULL;
> >>>>>+           }
> >>>>
> >>>>Minor issue: We should move the purged check into the loop so that purge
> >>>>buffer structs get released even when they're too small/big. Otherwise
> >>>>we'll have a good chance to hang onto gobloads of structs forever.
> >>>
> >>>I mentioned that we should do the purge of structs in our oom-notifier
> >>>as well to be safe.
> I understand Daniel's suggestion to move the purge check into the loop (will
> do that) but I'm not familiar w/ the oom-notifier at all and so don't know
> how to do what Chris is asking w/ out ramping up. Is it not critical, follow
> up type work or is it absolutely necessary to have before merge? It sounded
> like the first.

Imo the oom integration isn't critical and we can do that if we ever have
a workload where this matters. And my gut suspicion is that mostly we'll
get away with normal shrinking as already implemented.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch