Re: [PATCH v5 1/2] drm/xe/guc: Make xe_guc_submit_pause() available for non-VF cases

[Raag Jadav <raag.jadav@intel.com>]

On Wed, Oct 15, 2025 at 09:04:18AM +0200, Raag Jadav wrote:
> On Tue, Oct 14, 2025 at 11:09:19AM -0700, Matthew Brost wrote:
> > On Tue, Oct 14, 2025 at 01:00:35PM +0530, Raag Jadav wrote:
> > > Drop xe_gt_assert() meant for VF migration and make xe_guc_submit_pause()
> > > available for non-VF cases.
> > > 
> > > Signed-off-by: Raag Jadav <raag.jadav@intel.com>
> > > ---
> > >  drivers/gpu/drm/xe/xe_guc_submit.c | 2 --
> > >  1 file changed, 2 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c
> > > index 0ef67d3523a7..3cc428d45b4a 100644
> > > --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> > > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> > > @@ -2170,8 +2170,6 @@ void xe_guc_submit_pause(struct xe_guc *guc)
> > >  	struct xe_exec_queue *q;
> > >  	unsigned long index;
> > >  
> > > -	xe_gt_assert(guc_to_gt(guc), vf_recovery(guc));
> > > -
> > 
> > This function does a lot of things specific to VF recovery, it should
> > not be called in runtime PM flows. Same goes for xe_guc_submit_unpause.
> 
> Can we branch them out inside pause/unpause()? Or is there a better
> alternative?
> 
> Considering we'd still like to achieve the original pause/unpause()
> behaviour here (stop/start submission on a scheduler without loosing
> GuC state), do you suspect any side-effects?

I came up with something like this, but not sure if this is correct way
to do this. Let me know what you think.

diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c
index 0ef67d3523a7..8ee6069abe99 100644
--- a/drivers/gpu/drm/xe/xe_guc_submit.c
+++ b/drivers/gpu/drm/xe/xe_guc_submit.c
@@ -2127,6 +2127,11 @@ static void guc_exec_queue_pause(struct xe_guc *guc, struct xe_exec_queue *q)
 
 	/* Stop scheduling + flush any DRM scheduler operations */
 	xe_sched_submission_stop(sched);
+
+	/* Non-VF cases don't need migration */
+	if (!IS_SRIOV_VF(guc_to_xe(guc)))
+		return;
+
 	if (xe_exec_queue_is_lr(q))
 		cancel_work_sync(&q->guc->lr_tdr);
 	else
@@ -2170,8 +2175,6 @@ void xe_guc_submit_pause(struct xe_guc *guc)
 	struct xe_exec_queue *q;
 	unsigned long index;
 
-	xe_gt_assert(guc_to_gt(guc), vf_recovery(guc));
-
 	mutex_lock(&guc->submission_state.lock);
 	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
 		/* Prevent redundant attempts to stop parallel queues */
@@ -2325,6 +2328,12 @@ static void guc_exec_queue_unpause(struct xe_guc *guc, struct xe_exec_queue *q)
 
 	lockdep_assert_held(&guc->submission_state.lock);
 
+	/* Non-VF cases don't need migration */
+	if (!IS_SRIOV_VF(guc_to_xe(guc))) {
+		xe_sched_submission_start(sched);
+		return;
+	}
+
 	xe_sched_resubmit_jobs(sched);
 	guc_exec_queue_replay_pending_state_change(q);
 	xe_sched_submission_start(sched);