Re: [PATCH v2 2/4] drm/xe/multi_queue: handle CGP_SYNC wait timeout during VF recovery

On Fri, Jul 31, 2026 at 04:26:25PM -0700, Niranjana Vishwanathapura wrote:
> The CGP_SYNC wait in xe_guc_exec_queue_group_cgp_sync() only woke on
> sync_pending clearing or GuC stop. During VF migration the CT wakes all
> waiters but sync_pending may still be set (the GuC lost the in-flight
> message), causing the timeout path to fire and ban the group.
> 
> Add vf_recovery() as a third wake condition and exclude it from the
> error path so a migration-woken wait returns success rather than
> triggering a GT reset.
> 
> Assisted-by: Github-Copilot:Claude-opus-4.8
> Signed-off-by: Niranjana Vishwanathapura <[email protected]>

Reviewed-by: Matthew Brost <[email protected]>

> ---
>  drivers/gpu/drm/xe/xe_guc_submit.c | 7 +++----
>  1 file changed, 3 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c
> index 314747336fb9..13d0ab8052e5 100644
> --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> @@ -815,13 +815,12 @@ static void xe_guc_exec_queue_group_cgp_sync(struct xe_guc *guc,
>  	 * Hence, no locking is required here.
>  	 * Wait for any pending CGP_SYNC_DONE response before updating the
>  	 * CGP page and sending CGP_SYNC message.
> -	 *
> -	 * FIXME: Support VF migration
>  	 */
>  	ret = wait_event_timeout(guc->ct.wq,
>  				 !READ_ONCE(group->sync_pending) ||
> -				 xe_guc_read_stopped(guc), HZ);
> -	if (!ret || xe_guc_read_stopped(guc)) {
> +				 xe_guc_read_stopped(guc) || vf_recovery(guc),
> +				 HZ);
> +	if ((!ret && !vf_recovery(guc)) || xe_guc_read_stopped(guc)) {
>  		/* CGP_SYNC failed. Reset gt, cleanup the group */
>  		xe_gt_warn(guc_to_gt(guc), "Wait for CGP_SYNC_DONE response failed!\n");
>  		set_exec_queue_group_banned(q);
> -- 
> 2.43.0
>