Re: [PATCH v4 27/33] drm/xe: Add BO flags required for SVM

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

On Fri, Feb 07, 2025 at 02:54:45PM +0100, Thomas Hellström wrote:
> On Wed, 2025-01-29 at 11:52 -0800, Matthew Brost wrote:
> > Add XE_BO_FLAG_CPU_ADDR_MIRROR to indicate BO is tied to SVM range.
> > While these BO's are kernel allocations, we need a VM reference in
> > this
> > case which this flag indicates. In addition, we do not support CCS on
> > these BO's either. The later can be revisited later.
> > 
> > v2:
> >  - Take VM ref for system allocator BOs
> > v3:
> >  - s/XE_BO_FLAG_SYSTEM_ALLOC/XE_BO_FLAG_CPU_ADDR_MIRROR (Thomas)
> >  - Better commit message (Thomas)
> >  - Drop XE_BO_FLAG_SKIP_CLEAR for now
> >  - Add comment about possibly supporting CCS (Thomas)
> > v4:
> >  - Fix alignment issue (Checkpatch)
> > 
> > Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> 
> I was wondering, since the bo might as well be an external bo and
> benefit from finer resv granularity on eviction, (multi-device actually
> uses this), can't we drop the bo->vm reference? And, assuming tile is
> not needed either (is it)? Can we skip the flag altogether?
> 

If we make these external BO's, then this patch could just be dropped.

I feel like I tried external BO's a while a back and for some reason it
did not work but falling recall why. If external BO's work, then sure we
can make that change drop or revert this patch.

Matt

> /Thomas
> 
> > ---
> >  drivers/gpu/drm/xe/xe_bo.c | 12 ++++++++----
> >  drivers/gpu/drm/xe/xe_bo.h |  1 +
> >  2 files changed, 9 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> > index e914a60b8afc..20c96709e267 100644
> > --- a/drivers/gpu/drm/xe/xe_bo.c
> > +++ b/drivers/gpu/drm/xe/xe_bo.c
> > @@ -1239,7 +1239,7 @@ static void xe_ttm_bo_destroy(struct
> > ttm_buffer_object *ttm_bo)
> >  		xe_drm_client_remove_bo(bo);
> >  #endif
> >  
> > -	if (bo->vm && xe_bo_is_user(bo))
> > +	if (bo->vm && (xe_bo_is_user(bo) || bo->flags &
> > XE_BO_FLAG_CPU_ADDR_MIRROR))
> >  		xe_vm_put(bo->vm);
> >  
> >  	mutex_lock(&xe->mem_access.vram_userfault.lock);
> > @@ -1435,7 +1435,8 @@ struct xe_bo *___xe_bo_create_locked(struct
> > xe_device *xe, struct xe_bo *bo,
> >  	int err;
> >  
> >  	/* Only kernel objects should set GT */
> > -	xe_assert(xe, !tile || type == ttm_bo_type_kernel);
> > +	xe_assert(xe, !tile || type == ttm_bo_type_kernel ||
> > +		  flags & XE_BO_FLAG_CPU_ADDR_MIRROR);
> >  
> >  	if (XE_WARN_ON(!size)) {
> >  		xe_bo_free(bo);
> > @@ -1631,7 +1632,7 @@ __xe_bo_create_locked(struct xe_device *xe,
> >  	 * by having all the vm's bo refereferences released at vm
> > close
> >  	 * time.
> >  	 */
> > -	if (vm && xe_bo_is_user(bo))
> > +	if (vm && (xe_bo_is_user(bo) || bo->flags &
> > XE_BO_FLAG_CPU_ADDR_MIRROR))
> >  		xe_vm_get(vm);
> >  	bo->vm = vm;
> >  
> > @@ -2503,8 +2504,11 @@ bool xe_bo_needs_ccs_pages(struct xe_bo *bo)
> >  	 * system memory (i.e., it allows XE_PL_TT placement),
> > FlatCCS
> >  	 * can't be used since there's no CCS storage associated
> > with
> >  	 * non-VRAM addresses.
> > +	 *
> > +	 * XXX: Can we support CCS with CPU address mirroring?
> >  	 */
> > -	if (IS_DGFX(xe) && (bo->flags & XE_BO_FLAG_SYSTEM))
> > +	if (IS_DGFX(xe) && ((bo->flags & XE_BO_FLAG_SYSTEM) ||
> > +			    (bo->flags &
> > XE_BO_FLAG_CPU_ADDR_MIRROR)))
> >  		return false;
> >  
> >  	return true;
> > diff --git a/drivers/gpu/drm/xe/xe_bo.h b/drivers/gpu/drm/xe/xe_bo.h
> > index ce55a2bb13f6..c01ed535a8c3 100644
> > --- a/drivers/gpu/drm/xe/xe_bo.h
> > +++ b/drivers/gpu/drm/xe/xe_bo.h
> > @@ -47,6 +47,7 @@
> >  					 XE_BO_FLAG_GGTT1 | \
> >  					 XE_BO_FLAG_GGTT2 | \
> >  					 XE_BO_FLAG_GGTT3)
> > +#define XE_BO_FLAG_CPU_ADDR_MIRROR	BIT(22)
> >  
> >  /* this one is trigger internally only */
> >  #define XE_BO_FLAG_INTERNAL_TEST	BIT(30)
>