-----Original Message----- From: Sripada, Radhakrishna <[email protected]> Sent: Friday, November 3, 2023 1:02 PM To: Cavitt, Jonathan <[email protected]>; [email protected] Cc: Gupta, saurabhg <[email protected]>; Cavitt, Jonathan <[email protected]>; [email protected] Subject: RE: [Intel-gfx] [PATCH] drm/i915/gt: Temporarily disable CPU caching into DMA for MTL > > Hi Jonathan, > > > -----Original Message----- > > From: Intel-gfx <[email protected]> On Behalf Of > > Jonathan > > Cavitt > > Sent: Thursday, November 2, 2023 10:59 AM > > To: [email protected] > > Cc: Gupta, saurabhg <[email protected]>; Cavitt, Jonathan > > <[email protected]>; [email protected] > > Subject: [Intel-gfx] [PATCH] drm/i915/gt: Temporarily disable CPU caching > > into > > DMA for MTL > > > > FIXME: It is suspected that some Address Translation Service (ATS) > > issue on IOMMU is causing CAT errors to occur on some MTL workloads. > > Applying a write barrier to the ppgtt set entry functions appeared > > to have no effect, so we must temporarily use I915_MAP_WC in the > > map_pt_dma class of functions on MTL until a proper ATS solution is > > found. > > > What is the performance impact here? Are we disabling caching only > for the pte changes/scratch pages or does it extend beyond?
I don't actually know what map_pt_dma is used for, but if the name is indicative of its purpose, it should only impact mappings into the dma page table. As for the performance impact, I don't imagine it'll be much. Maybe a single-digit percentage slowdown? It might actually improve performance if we're avoiding enough cache misses, but the true performance impact would have to be measured empirically. -Jonathan Cavitt > > Regards, > Radhakrishna(RK) Sripada > > Signed-off-by: Jonathan Cavitt <[email protected]> > > CC: Chris Wilson <[email protected]> > > --- > > drivers/gpu/drm/i915/gt/intel_gtt.c | 20 ++++++++++++++++++++ > > 1 file changed, 20 insertions(+) > > > > diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c > > b/drivers/gpu/drm/i915/gt/intel_gtt.c > > index 4fbed27ef0ecc..21719563a602a 100644 > > --- a/drivers/gpu/drm/i915/gt/intel_gtt.c > > +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c > > @@ -95,6 +95,16 @@ int map_pt_dma(struct i915_address_space *vm, struct > > drm_i915_gem_object *obj) > > void *vaddr; > > > > type = intel_gt_coherent_map_type(vm->gt, obj, true); > > + /* > > + * FIXME: It is suspected that some Address Translation Service (ATS) > > + * issue on IOMMU is causing CAT errors to occur on some MTL > > workloads. > > + * Applying a write barrier to the ppgtt set entry functions appeared > > + * to have no effect, so we must temporarily use I915_MAP_WC here on > > + * MTL until a proper ATS solution is found. > > + */ > > + if (IS_METEORLAKE(vm->i915)) > > + type = I915_MAP_WC; > > + > > vaddr = i915_gem_object_pin_map_unlocked(obj, type); > > if (IS_ERR(vaddr)) > > return PTR_ERR(vaddr); > > @@ -109,6 +119,16 @@ int map_pt_dma_locked(struct i915_address_space > > *vm, struct drm_i915_gem_object > > void *vaddr; > > > > type = intel_gt_coherent_map_type(vm->gt, obj, true); > > + /* > > + * FIXME: It is suspected that some Address Translation Service (ATS) > > + * issue on IOMMU is causing CAT errors to occur on some MTL > > workloads. > > + * Applying a write barrier to the ppgtt set entry functions appeared > > + * to have no effect, so we must temporarily use I915_MAP_WC here on > > + * MTL until a proper ATS solution is found. > > + */ > > + if (IS_METEORLAKE(vm->i915)) > > + type = I915_MAP_WC; > > + > > vaddr = i915_gem_object_pin_map(obj, type); > > if (IS_ERR(vaddr)) > > return PTR_ERR(vaddr); > > -- > > 2.25.1 > >
