[PATCH 3/3] drm/xe/pf: Invalidate LMTT after completing changes

[Matthew Brost]

On Thu, Jul 03, 2025 at 12:30:41AM +0200, Michal Wajdeczko wrote:
> Once we finish populating all leaf pages in the VF's LMTT we should
> make sure that hardware will not access any stale data. Explicitly
> force LMTT invalidation (as it was already planned in the past).
> 
> Signed-off-by: Michal Wajdeczko <michal.wajdeczko at intel.com>
> Cc: Michał Winiarski <michal.winiarski at intel.com>
> Cc: Piotr Piórkowski <piotr.piorkowski at intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c | 12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> index a75ef4007a1e..a1b916e80fdd 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> @@ -1327,7 +1327,17 @@ static int pf_distribute_config_lmem(struct xe_gt *gt, unsigned int vfid, u64 si
>  
>  static void pf_force_lmtt_invalidate(struct xe_device *xe)
>  {
> -	/* TODO */
> +	struct xe_lmtt *lmtt;
> +	struct xe_tile *tile;
> +	unsigned int tid;
> +
> +	xe_assert(xe, xe_device_has_lmtt(xe));
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +
> +	for_each_tile(tile, xe, tid) {
> +		lmtt = &tile->sriov.pf.lmtt;
> +		xe_lmtt_invalidate_hw(lmtt);

If this is a critical path, you may want to consider issuing all TLB
invalidations across all GTs/tiles, then waiting on all fences to
pipeline these - we do this the VM code which is certainly a critical
path. If this is not-so-critical here, maybe this is ok.

Matt

> +	}
>  }
>  
>  static void pf_reset_vf_lmtt(struct xe_device *xe, unsigned int vfid)
> -- 
> 2.47.1
>