RE: [PATCH] drm/xe/guc: skip CLEANUP message when device is hot-unplugged

Hi Matt,

> -----Original Message-----
> From: Brost, Matthew <[email protected]>
> Sent: Wednesday, July 29, 2026 1:51 AM
> To: Gote, Nitin R <[email protected]>
> Cc: [email protected]; Auld, Matthew <[email protected]>
> Subject: Re: [PATCH] drm/xe/guc: skip CLEANUP message when device is hot-
> unplugged
> 
> On Tue, Jul 28, 2026 at 08:54:53PM +0530, Nitin Gote wrote:
> > When a device is hot-unplugged while a DRM fd with active exec queues
> > is still open, the fd close path runs during process exit and calls
> > guc_exec_queue_destroy(). This queues a CLEANUP message via
> > guc_exec_queue_add_msg(), which internally calls
> > xe_pm_runtime_get_noresume().
> >
> > After hot-unplug, pm_runtime_disable() has already been called, so
> > xe_pm_runtime_get_noresume() warns with "Missing outer runtime PM
> > protection".
> >
> > GuC communication is no longer valid after unplug, so skip the CLEANUP
> > message and destroy the exec queue locally via __guc_exec_queue_destroy().
> >
> > Observed with new IGT core_hotunplug subtests:
> >
> >   igt@core_hotunplug@hotreplug-with-load
> >   igt@core_hotunplug@hotunplug-rescan-with-load
> >
> > Link: https://patchwork.freedesktop.org/series/166744/
> > Cc: Matthew Auld <[email protected]>
> > Assisted-by: GitHub-Copilot:claude-sonnet-4.6
> > Signed-off-by: Nitin Gote <[email protected]>
> > ---
> >  drivers/gpu/drm/xe/xe_guc_submit.c | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c
> > b/drivers/gpu/drm/xe/xe_guc_submit.c
> > index 8aaed4fd13ea..b5035c19d7b0 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> > @@ -2131,12 +2131,15 @@ static bool guc_exec_queue_try_add_msg(struct
> xe_exec_queue *q,
> >  #define STATIC_MSG_RESUME	2
> >  static void guc_exec_queue_destroy(struct xe_exec_queue *q)  {
> > +	struct xe_guc *guc = exec_queue_to_guc(q);
> >  	struct xe_sched_msg *msg = q->guc->static_msgs +
> STATIC_MSG_CLEANUP;
> >
> > -	if (!(q->flags & EXEC_QUEUE_FLAG_PERMANENT) &&
> !exec_queue_wedged(q))
> > +	/* If the device has been hot-unplugged the GuC is already gone. */
> > +	if (!(q->flags & EXEC_QUEUE_FLAG_PERMANENT) &&
> !exec_queue_wedged(q) &&
> > +	    !drm_dev_is_unplugged(&guc_to_xe(guc)->drm))
> 
> Isn't there a TOCTOU here - immedaitely the value of drm_dev_is_unplugged can
> change after the CLEANUP message is added?

Agreed, the drm_dev_is_unplugged() check in guc_exec_queue_destroy() is racy since the device can be unplugged after the check.

For v2, I will drop the drm_dev_is_unplugged() check from guc_exec_queue_destroy() and instead gate 
the "Missing outer runtime PM protection" warning on !drm_dev_is_unplugged() in xe_pm_runtime_get_noresume().
After hot-unplug, runtime PM has already been disabled, so pm_runtime_get_if_in_use() failing is expected and does not indicate a
missing outer runtime PM guard.

> 
> >  		guc_exec_queue_add_msg(q, msg, CLEANUP);
> >  	else
> > -		__guc_exec_queue_destroy(exec_queue_to_guc(q), q);
> > +		__guc_exec_queue_destroy(guc, q);
> 
> I think after merging [1] we should also aim to get rid of this bypass teardown
> path and always add a CLEANUP message and within message processing either
> send the H2G to initiatate teardown of the queue or kick the worker which tears
> down the queue (i.e., the async worker is the only place where final queue down
> can occur).
>

Regarding removing the wedged bypass and always routing through CLEANUP message processing,
It seems the right direction to me. I'll look into that as a follow-up once this fix lands.

Thank you,
Nitin

> Matt
> 
> [1] https://patchwork.freedesktop.org/patch/740542/?series=168424&rev=10
> 
> >  }
> >
> >  static int guc_exec_queue_set_priority(struct xe_exec_queue *q,
> > --
> > 2.50.1
> >