Re: [PATCH v4] drm/xe: Invalidate media_gt TLBs

[Matthew Brost <matthew.brost@intel.com>]

On Tue, Aug 20, 2024 at 10:22:54AM -0700, Matt Roper wrote:
> On Tue, Aug 20, 2024 at 09:01:29AM -0700, Matthew Brost wrote:
> > Testing on LNL has shown media TLBs need to be invalidated via the GuC,
> > update xe_vm_invalidate_vma appropriately.
> > 
> > v2: Fix 2 tile case
> > v3: Include missing local changm
> > 
> > Fixes: 3330361543fc ("drm/xe/lnl: Add LNL platform definition")
> 
> Is this the right Fixes line?  This may have been discovered in testing

I was struggling with the fixes tag. Just picked a LNL one as yes I
discovered this bug on LNL working on SVM.

> on LNL, but presumably the fix applies to MTL as well since we also have
> a standalone media GT there too, and we've had MTL support in the driver
> since before LNL was added.
> 

Forgot about MTL, yes this will apply to MTL too.

> Of course both platforms are under force_probe at the moment (and MTL
> will be forever since it's officially supported by i915 instead of Xe),
> so it probably doesn't really matter too much for backporting purposes
> and such.

Yea this patch will be pain to backport as it is built on top a recent
change to TLB invalidations. Maybe we just drop it as these are under
force probe so not a huge deal if it is not fixed.

Matt

> 
> 
> Matt
> 
> > Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> > Reviewed-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_vm.c | 37 ++++++++++++++++++++++++-------------
> >  1 file changed, 24 insertions(+), 13 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> > index d1bfd0b6e955..4cc13eddb6b3 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.c
> > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > @@ -3158,9 +3158,10 @@ int xe_vm_invalidate_vma(struct xe_vma *vma)
> >  {
> >  	struct xe_device *xe = xe_vma_vm(vma)->xe;
> >  	struct xe_tile *tile;
> > -	struct xe_gt_tlb_invalidation_fence fence[XE_MAX_TILES_PER_DEVICE];
> > -	u32 tile_needs_invalidate = 0;
> > +	struct xe_gt_tlb_invalidation_fence
> > +		fence[XE_MAX_TILES_PER_DEVICE * XE_MAX_GT_PER_TILE];
> >  	u8 id;
> > +	u32 fence_id = 0;
> >  	int ret = 0;
> >  
> >  	xe_assert(xe, !xe_vma_is_null(vma));
> > @@ -3188,27 +3189,37 @@ int xe_vm_invalidate_vma(struct xe_vma *vma)
> >  		if (xe_pt_zap_ptes(tile, vma)) {
> >  			xe_device_wmb(xe);
> >  			xe_gt_tlb_invalidation_fence_init(tile->primary_gt,
> > -							  &fence[id], true);
> > +							  &fence[fence_id],
> > +							  true);
> >  
> > -			/*
> > -			 * FIXME: We potentially need to invalidate multiple
> > -			 * GTs within the tile
> > -			 */
> >  			ret = xe_gt_tlb_invalidation_vma(tile->primary_gt,
> > -							 &fence[id], vma);
> > +							 &fence[fence_id], vma);
> >  			if (ret < 0) {
> > -				xe_gt_tlb_invalidation_fence_fini(&fence[id]);
> > +				xe_gt_tlb_invalidation_fence_fini(&fence[fence_id]);
> >  				goto wait;
> >  			}
> > +			++fence_id;
> >  
> > -			tile_needs_invalidate |= BIT(id);
> > +			if (!tile->media_gt)
> > +				continue;
> > +
> > +			xe_gt_tlb_invalidation_fence_init(tile->media_gt,
> > +							  &fence[fence_id],
> > +							  true);
> > +
> > +			ret = xe_gt_tlb_invalidation_vma(tile->media_gt,
> > +							 &fence[fence_id], vma);
> > +			if (ret < 0) {
> > +				xe_gt_tlb_invalidation_fence_fini(&fence[fence_id]);
> > +				goto wait;
> > +			}
> > +			++fence_id;
> >  		}
> >  	}
> >  
> >  wait:
> > -	for_each_tile(tile, xe, id)
> > -		if (tile_needs_invalidate & BIT(id))
> > -			xe_gt_tlb_invalidation_fence_wait(&fence[id]);
> > +	for (id = 0; id < fence_id; ++id)
> > +		xe_gt_tlb_invalidation_fence_wait(&fence[id]);
> >  
> >  	vma->tile_invalidated = vma->tile_mask;
> >  
> > -- 
> > 2.34.1
> > 
> 
> -- 
> Matt Roper
> Graphics Software Engineer
> Linux GPU Platform Enablement
> Intel Corporation