[Intel-xe] [PATCH] drm/xe/bo: handle PL_TT -> PL_TT

[Thomas Hellström]

Hi,

On 6/15/23 20:18, Matthew Auld wrote:
> When moving between PL_VRAM <-> PL_SYSTEM we have to have use PL_TT in
> the middle as a temporary resource for the actual copy. In some GL
> workloads it can be seen that once the resource has been moved to the
> PL_TT we might have to bail out of the ttm_bo_validate(), before
> finishing the final hop. If this happens the resource is left as
> TTM_PL_FLAG_TEMPORARY, and when the ttm_bo_validate() is restarted the
> current placement is always seen as incompatible, requiring us to
> complete the move.  However if the BO allows PL_TT as a possible
> placement we can end up attempting a PL_TT -> PL_TT move (like when
> running out of VRAM) which leads to explosions in xe_bo_move(), like
> triggering the XE_BUG_ON(!tile).
>
> Going from TTM_PL_FLAG_TEMPORARY with PL_TT -> PL_VRAM should already
> work as-is, so it looks like we only need to worry about PL_TT -> PL_TT
> and it looks like we can just treat it as a dummy move, since no real
> move is needed.
>
> Reported-by: José Roberto de Souza <jose.souza at intel.com>
> Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>

Could perhaps be merged with the SYSTEM-to-TT test above so we get 
any-ttm-backed to TT, but perhaps that will become hairy.

Either way

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


> ---
>   drivers/gpu/drm/xe/xe_bo.c | 10 ++++++++++
>   1 file changed, 10 insertions(+)
>
> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> index b94a80a32d86..5aed626cce80 100644
> --- a/drivers/gpu/drm/xe/xe_bo.c
> +++ b/drivers/gpu/drm/xe/xe_bo.c
> @@ -603,6 +603,16 @@ static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict,
>   		goto out;
>   	}
>   
> +	/*
> +	 * Failed multi-hop where the old_mem is still marked as
> +	 * TTM_PL_FLAG_TEMPORARY, should just be a dummy move.
> +	 */
> +	if (old_mem->mem_type == XE_PL_TT &&
> +	    new_mem->mem_type == XE_PL_TT) {
> +		ttm_bo_move_null(ttm_bo, new_mem);
> +		goto out;
> +	}
> +
>   	if (!move_lacks_source && !xe_bo_is_pinned(bo)) {
>   		ret = xe_bo_move_notify(bo, ctx);
>   		if (ret)