[Intel-gfx] [PATCH 2/4] drm/i915: Improve dynamic management/eviction of lrc backing objects

[Chris Wilson]

On Wed, Oct 07, 2015 at 06:05:46PM +0200, Daniel Vetter wrote:
> On Tue, Oct 06, 2015 at 03:52:02PM +0100, Nick Hoath wrote:
> > Shovel all context related objects through the active queue and obj
> > management.
> > 
> > - Added callback in vma_(un)bind to add CPU (un)mapping at same time
> >   if desired
> > - Inserted LRC hw context & ringbuf to vma active list
> > 
> > Issue: VIZ-4277
> > Signed-off-by: Nick Hoath <nicholas.hoath at intel.com>
> > ---
> >  drivers/gpu/drm/i915/i915_drv.h         |  4 ++
> >  drivers/gpu/drm/i915/i915_gem.c         |  3 ++
> >  drivers/gpu/drm/i915/i915_gem_gtt.c     |  8 ++++
> >  drivers/gpu/drm/i915/intel_lrc.c        | 28 +++++++++++--
> >  drivers/gpu/drm/i915/intel_ringbuffer.c | 71 ++++++++++++++++++---------------
> >  drivers/gpu/drm/i915/intel_ringbuffer.h |  3 --
> >  6 files changed, 79 insertions(+), 38 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > index 3d217f9..d660ee3 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.h
> > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > @@ -2169,6 +2169,10 @@ struct drm_i915_gem_object {
> >  			struct work_struct *work;
> >  		} userptr;
> >  	};
> > +
> > +	/** Support for automatic CPU side mapping of object */
> > +	int (*mmap)(struct drm_i915_gem_object *obj, bool unmap);
> 
> I don't think we need a map hook, that can still be done (if not done so
> already) by the callers. Also it's better to rename this to vma_unbind
> (and it should be at the vma level I think) since there's other potential
> users. So explicit maping, lazy unmapping for the kmaps we need. That's
> the same design we're using for binding objects into gpu address spaces.
> 
> Also Chris Wilson has something similar, please align with him on the
> precise design of this callback.

We need the unbind hook because of the movement in the first patch (it
is a separate issue, the code should work without it albeit having to
remap the ring/context state more often). The changelog in this patch
simply explains the i915_vma_move_to_active() additions. But to get the
shrink accurate we do need the context unpin on retirement and to do the
pin_count check in i915_vma_unbind() after waiting (rather than before,
as we currently do). However, the eviction code will not inspect the
active contexts objects yet (as it will continue to skip over the
ggtt->pin_count on them). The way I allowed ctx objects to be evicted was
to only keep the ctx->state pinned for the duration of the request
construction.

Note that I think it should be a vma->unbind hook not an object level
one (it is i915_vma_unbind, without only a modicum of object level state
being modified in that function).
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre