Re: [PATCH 6/7] drm/xe: Kick start GPU scheduler on teardown

[Matthew Brost <matthew.brost@intel.com>]

On Mon, Oct 13, 2025 at 04:32:57PM -0600, Summers, Stuart wrote:
> On Mon, 2025-10-13 at 22:31 +0000, Stuart Summers wrote:
> > Ensure all remaining GPU scheduler messages are processed
> > on teardown. We already have a wait here, but if the
> > scheduler was stopped while we have outstanding transactions,
> > those queues will never be fully drained.
> > 
> > Signed-off-by: Stuart Summers <stuart.summers@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_guc_submit.c | 6 ++++++
> >  1 file changed, 6 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c
> > b/drivers/gpu/drm/xe/xe_guc_submit.c
> > index 5ec1e4a83d68..c923f13a13ef 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> > @@ -276,6 +276,12 @@ static void guc_submit_fini(struct drm_device
> > *drm, void *arg)
> >         struct xe_gt *gt = guc_to_gt(guc);
> >         int ret;
> >  
> > +       /*
> > +        * Kick start the scheduler here to ensure any remaining
> > +        * messages are processed
> > +        */
> > +       xe_guc_submit_start(guc);
> 
> Let me know what you think here Matt. We can also extend the below
> timeout. I'm hoping this covers both the case I have and your VF
> submission drain case.
> 

I think you will need some code similar to guc_exec_queue_stop somewhere
in the wedge path which cleans up lost H2G. Off the top of my head, I'm
unsure exactly where to stick that part.

Matt

> Thanks,
> Stuart
> 
> > +
> >         ret = wait_event_timeout(guc->submission_state.fini_wq,
> >                                  xa_empty(&guc-
> > >submission_state.exec_queue_lookup),
> >                                  HZ * 5);
>