Re: [PATCH 2/6] drm/xe/bb: drop per-batch prefetch padding

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Fri, 2026-10-30 at 11:49 +0000, Matthew Auld wrote:
> xe_bb_new() currently appends bb_prefetch() (512B or 1K) to every
> batch suballocation, inflating a 36B identity-mapped clear batch to
> ~1K and also causing xe_sa_bo_flush_write() to copy up to 1K of
> unused
> padding to VRAM on every batch.
> 
> However, both kernel_bb_pool and usm.bb_pool are initialised via
> xe_sa_bo_manager_init(), which already reserves a 4K guard region at
> the end of the backing BO to cover CS prefetch, and prefetching into
> adjacent suballocations within the mapped pool should be harmless.
> Drop
> bb_prefetch().
> 
> This should reduce pressure on the bb pool.
> 
> Assisted-by: LLM
> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/work_items/9644
> Signed-off-by: Matthew Auld <matthew.auld@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Cc: Matthew Brost <matthew.brost@intel.com>

Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>


> ---
>  drivers/gpu/drm/xe/xe_bb.c | 26 ++++++--------------------
>  1 file changed, 6 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_bb.c b/drivers/gpu/drm/xe/xe_bb.c
> index b896b6f6615c..b2603902c147 100644
> --- a/drivers/gpu/drm/xe/xe_bb.c
> +++ b/drivers/gpu/drm/xe/xe_bb.c
> @@ -14,20 +14,6 @@
>  #include "xe_sched_job.h"
>  #include "xe_vm_types.h"
>  
> -static int bb_prefetch(struct xe_gt *gt)
> -{
> -	struct xe_device *xe = gt_to_xe(gt);
> -
> -	if (GRAPHICS_VERx100(xe) >= 1250 && xe_gt_is_main_type(gt))
> -		/*
> -		 * RCS and CCS require 1K, although other engines
> would be
> -		 * okay with 512.
> -		 */
> -		return SZ_1K;
> -	else
> -		return SZ_512;
> -}
> -
>  struct xe_bb *xe_bb_new(struct xe_gt *gt, u32 dwords, bool usm)
>  {
>  	struct xe_tile *tile = gt_to_tile(gt);
> @@ -38,13 +24,13 @@ struct xe_bb *xe_bb_new(struct xe_gt *gt, u32
> dwords, bool usm)
>  		return ERR_PTR(-ENOMEM);
>  
>  	/*
> -	 * We need to allocate space for the requested number of
> dwords,
> -	 * one additional MI_BATCH_BUFFER_END dword, and additional
> buffer
> -	 * space to accommodate the platform-specific hardware
> prefetch
> -	 * requirements.
> +	 * We need to allocate space for the requested number of
> dwords and
> +	 * one additional MI_BATCH_BUFFER_END dword. Hardware CS
> prefetch
> +	 * at the end of the pool BO is covered by the SZ_4K guard
> region
> +	 * reserved by xe_sa_bo_manager_init().
>  	 */
>  	bb->bo = xe_sa_bo_new(!usm ? tile->mem.kernel_bb_pool : gt-
> >usm.bb_pool,
> -			      4 * (dwords + 1) + bb_prefetch(gt));
> +			      4 * (dwords + 1));
>  	if (IS_ERR(bb->bo)) {
>  		err = PTR_ERR(bb->bo);
>  		goto err;
> @@ -129,7 +115,7 @@ __xe_bb_create_job(struct xe_exec_queue *q,
> struct xe_bb *bb, u64 *addr)
>  	if (bb->len == 0 || bb->cs[bb->len - 1] !=
> MI_BATCH_BUFFER_END)
>  		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
>  
> -	xe_gt_assert(q->gt, bb->len * 4 + bb_prefetch(q->gt) <=
> size);
> +	xe_gt_assert(q->gt, bb->len * 4 <= size);
>  
>  	xe_sa_bo_flush_write(bb->bo);
>