[Intel-xe] [PATCH] drm/xe/migrate: fix MI_ARB_ON_OFF usage

[Matt Roper]

On Fri, Oct 20, 2023 at 07:24:38PM +0100, Matthew Auld wrote:
> Spec says: "This is a privileged command; it will not be effective (will
> be converted to a no-op) if executed from within a non-privileged batch
> buffer." However here it looks like we are just emitting it inside some
> bb which was jumped to via the ppGTT, which should be considered
> a non-privileged address space.
> 
> It looks like we just need some way of preventing things like the
> emit_pte() and later copy/clear being preempted in-between so rather
> just emit directly in the ring for migration jobs.
> 
> Bspec: 45716
> Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> Cc: Matthew Brost <matthew.brost at intel.com>

Reviewed-by: Matt Roper <matthew.d.roper at intel.com>

> ---
>  drivers/gpu/drm/xe/xe_migrate.c  | 16 ----------------
>  drivers/gpu/drm/xe/xe_ring_ops.c |  2 ++
>  2 files changed, 2 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c
> index b4baecde60e6..009d728af762 100644
> --- a/drivers/gpu/drm/xe/xe_migrate.c
> +++ b/drivers/gpu/drm/xe/xe_migrate.c
> @@ -406,12 +406,6 @@ struct xe_migrate *xe_migrate_init(struct xe_tile *tile)
>  	return m;
>  }
>  
> -static void emit_arb_clear(struct xe_bb *bb)
> -{
> -	/* 1 dword */
> -	bb->cs[bb->len++] = MI_ARB_ON_OFF | MI_ARB_DISABLE;
> -}
> -
>  static u64 xe_migrate_res_sizes(struct xe_res_cursor *cur)
>  {
>  	/*
> @@ -745,10 +739,6 @@ struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
>  			goto err_sync;
>  		}
>  
> -		/* Preemption is enabled again by the ring ops. */
> -		if (!src_is_vram || !dst_is_vram)
> -			emit_arb_clear(bb);
> -
>  		if (!src_is_vram)
>  			emit_pte(m, bb, src_L0_pt, src_is_vram, &src_it, src_L0,
>  				 src_bo);
> @@ -994,7 +984,6 @@ struct dma_fence *xe_migrate_clear(struct xe_migrate *m,
>  
>  		/* Preemption is enabled again by the ring ops. */
>  		if (!clear_vram) {
> -			emit_arb_clear(bb);
>  			emit_pte(m, bb, clear_L0_pt, clear_vram, &src_it, clear_L0,
>  				 bo);
>  		} else {
> @@ -1285,9 +1274,6 @@ xe_migrate_update_pgtables(struct xe_migrate *m,
>  				VM_SA_UPDATE_UNIT_SIZE;
>  		}
>  
> -		/* Preemption is enabled again by the ring ops. */
> -		emit_arb_clear(bb);
> -
>  		/* Map our PT's to gtt */
>  		bb->cs[bb->len++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(num_updates);
>  		bb->cs[bb->len++] = ppgtt_ofs * XE_PAGE_SIZE + page_ofs;
> @@ -1316,8 +1302,6 @@ xe_migrate_update_pgtables(struct xe_migrate *m,
>  		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
>  		update_idx = bb->len;
>  
> -		/* Preemption is enabled again by the ring ops. */
> -		emit_arb_clear(bb);
>  		for (i = 0; i < num_updates; i++)
>  			write_pgtable(tile, bb, 0, &updates[i], pt_update);
>  	}
> diff --git a/drivers/gpu/drm/xe/xe_ring_ops.c b/drivers/gpu/drm/xe/xe_ring_ops.c
> index 58676f4b989f..59e0aa2d6a4c 100644
> --- a/drivers/gpu/drm/xe/xe_ring_ops.c
> +++ b/drivers/gpu/drm/xe/xe_ring_ops.c
> @@ -355,6 +355,8 @@ static void emit_migration_job_gen12(struct xe_sched_job *job,
>  	i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
>  				seqno, dw, i);
>  
> +	dw[i++] = MI_ARB_ON_OFF | MI_ARB_DISABLE; /* Enabled again below */
> +
>  	i = emit_bb_start(job->batch_addr[0], BIT(8), dw, i);
>  
>  	/* XXX: Do we need this? Leaving for now. */
> -- 
> 2.41.0
> 

-- 
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation