Re: [PATCH v2 26/34] drm/xe/vf: Replay GuC submission state on pause / unpause

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

On Sat, Sep 27, 2025 at 03:33:43PM +0200, Lis, Tomasz wrote:
> 
> On 9/24/2025 3:15 AM, Matthew Brost wrote:
> > Fixup GuC submission pause / unpause functions to properly replay any
> > possible state lost during VF post migration recovery.
> > 
> > Signed-off-by: Matthew Brost<matthew.brost@intel.com>
> > ---
> >   drivers/gpu/drm/xe/xe_gpu_scheduler.c        |  14 ++
> >   drivers/gpu/drm/xe/xe_gpu_scheduler.h        |   2 +
> >   drivers/gpu/drm/xe/xe_gt_sriov_vf.c          |   1 +
> >   drivers/gpu/drm/xe/xe_guc_exec_queue_types.h |  15 ++
> >   drivers/gpu/drm/xe/xe_guc_submit.c           | 225 +++++++++++++++++--
> >   drivers/gpu/drm/xe/xe_guc_submit.h           |   1 +
> >   drivers/gpu/drm/xe/xe_sched_job_types.h      |   4 +
> >   7 files changed, 247 insertions(+), 15 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gpu_scheduler.c b/drivers/gpu/drm/xe/xe_gpu_scheduler.c
> > index 455ccaf17314..af300adc7e1a 100644
> > --- a/drivers/gpu/drm/xe/xe_gpu_scheduler.c
> > +++ b/drivers/gpu/drm/xe/xe_gpu_scheduler.c
> > @@ -135,3 +135,17 @@ void xe_sched_add_msg_locked(struct xe_gpu_scheduler *sched,
> >   	list_add_tail(&msg->link, &sched->msgs);
> >   	xe_sched_process_msg_queue(sched);
> >   }
> > +
> > +/**
> > + * xe_sched_add_msg_head() - Xe GPU scheduler add message to head of list
> > + * @sched: Xe GPU scheduler
> > + * @msg: Message to add
> > + */
> > +void xe_sched_add_msg_head(struct xe_gpu_scheduler *sched,
> > +			   struct xe_sched_msg *msg)
> > +{
> > +	lockdep_assert_held(&sched->base.job_list_lock);
> > +
> > +	list_add(&msg->link, &sched->msgs);
> > +	xe_sched_process_msg_queue(sched);
> > +}
> > diff --git a/drivers/gpu/drm/xe/xe_gpu_scheduler.h b/drivers/gpu/drm/xe/xe_gpu_scheduler.h
> > index e548b2aed95a..010003a6103a 100644
> > --- a/drivers/gpu/drm/xe/xe_gpu_scheduler.h
> > +++ b/drivers/gpu/drm/xe/xe_gpu_scheduler.h
> > @@ -29,6 +29,8 @@ void xe_sched_add_msg(struct xe_gpu_scheduler *sched,
> >   		      struct xe_sched_msg *msg);
> >   void xe_sched_add_msg_locked(struct xe_gpu_scheduler *sched,
> >   			     struct xe_sched_msg *msg);
> > +void xe_sched_add_msg_head(struct xe_gpu_scheduler *sched,
> > +			   struct xe_sched_msg *msg);
> >   static inline void xe_sched_msg_lock(struct xe_gpu_scheduler *sched)
> >   {
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
> > index a987560de2c7..91e7dbe80ab2 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
> > @@ -1217,6 +1217,7 @@ static int vf_post_migration_fixups(struct xe_gt *gt)
> >   static void vf_post_migration_rearm(struct xe_gt *gt)
> >   {
> >   	xe_guc_ct_restart(&gt->uc.guc.ct);
> > +	xe_guc_submit_unpause_prepare(&gt->uc.guc);
> >   }
> >   static void vf_post_migration_kickstart(struct xe_gt *gt)
> > diff --git a/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h b/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h
> > index c30c0e3ccbbb..a3b034e4b205 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h
> > +++ b/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h
> > @@ -51,6 +51,21 @@ struct xe_guc_exec_queue {
> >   	wait_queue_head_t suspend_wait;
> >   	/** @suspend_pending: a suspend of the exec_queue is pending */
> >   	bool suspend_pending;
> > +	/**
> > +	 * @needs_cleanup: Needs a cleanup message during VF post migration
> > +	 * recovery.
> > +	 */
> > +	bool needs_cleanup;
> > +	/**
> > +	 * @needs_suspend: Needs a suspend message during VF post migration
> > +	 * recovery.
> > +	 */
> > +	bool needs_suspend;
> > +	/**
> > +	 * @needs_resume: Needs a resume message during VF post migration
> > +	 * recovery.
> > +	 */
> > +	bool needs_resume;
> >   };
> >   #endif
> > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c
> > index 8bee65dd9ca6..b112a4a91a5b 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> > @@ -425,6 +425,11 @@ static void set_exec_queue_destroyed(struct xe_exec_queue *q)
> >   	atomic_or(EXEC_QUEUE_STATE_DESTROYED, &q->guc->state);
> >   }
> > +static void clear_exec_queue_destroyed(struct xe_exec_queue *q)
> > +{
> > +	atomic_and(~EXEC_QUEUE_STATE_DESTROYED, &q->guc->state);
> > +}
> > +
> >   static bool exec_queue_banned(struct xe_exec_queue *q)
> >   {
> >   	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_BANNED;
> > @@ -505,7 +510,12 @@ static void set_exec_queue_extra_ref(struct xe_exec_queue *q)
> >   	atomic_or(EXEC_QUEUE_STATE_EXTRA_REF, &q->guc->state);
> >   }
> > -static bool __maybe_unused exec_queue_pending_resume(struct xe_exec_queue *q)
> > +static void clear_exec_queue_extra_ref(struct xe_exec_queue *q)
> > +{
> > +	atomic_and(~EXEC_QUEUE_STATE_EXTRA_REF, &q->guc->state);
> > +}
> > +
> > +static bool exec_queue_pending_resume(struct xe_exec_queue *q)
> >   {
> >   	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_PENDING_RESUME;
> >   }
> > @@ -520,7 +530,7 @@ static void clear_exec_queue_pending_resume(struct xe_exec_queue *q)
> >   	atomic_and(~EXEC_QUEUE_STATE_PENDING_RESUME, &q->guc->state);
> >   }
> > -static bool __maybe_unused exec_queue_pending_tdr_exit(struct xe_exec_queue *q)
> > +static bool exec_queue_pending_tdr_exit(struct xe_exec_queue *q)
> >   {
> >   	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_PENDING_TDR_EXIT;
> >   }
> > @@ -1080,7 +1090,7 @@ static void wq_item_append(struct xe_exec_queue *q)
> >   }
> >   #define RESUME_PENDING	~0x0ull
> > -static void submit_exec_queue(struct xe_exec_queue *q)
> > +static void submit_exec_queue(struct xe_exec_queue *q, struct xe_sched_job *job)
> >   {
> >   	struct xe_guc *guc = exec_queue_to_guc(q);
> >   	struct xe_lrc *lrc = q->lrc[0];
> > @@ -1092,10 +1102,13 @@ static void submit_exec_queue(struct xe_exec_queue *q)
> >   	xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q));
> > -	if (xe_exec_queue_is_parallel(q))
> > -		wq_item_append(q);
> > -	else
> > -		xe_lrc_set_ring_tail(lrc, lrc->ring.tail);
> > +	if (!job->skip_emit || job->last_replay) {
> > +		if (xe_exec_queue_is_parallel(q))
> > +			wq_item_append(q);
> > +		else
> > +			xe_lrc_set_ring_tail(lrc, lrc->ring.tail);
> > +		job->last_replay = false;
> > +	}
> >   	if (exec_queue_suspended(q) && !xe_exec_queue_is_parallel(q))
> >   		return;
> > @@ -1148,8 +1161,10 @@ guc_exec_queue_run_job(struct drm_sched_job *drm_job)
> >   	if (!killed_or_banned_or_wedged && !xe_sched_job_is_error(job)) {
> >   		if (!exec_queue_registered(q))
> >   			register_exec_queue(q, GUC_CONTEXT_NORMAL);
> > -		q->ring_ops->emit_job(job);
> > -		submit_exec_queue(q);
> > +		if (!job->skip_emit)
> > +			q->ring_ops->emit_job(job);
> > +		submit_exec_queue(q, job);
> > +		job->skip_emit = false;
> >   	}
> >   	/*
> > @@ -1860,6 +1875,7 @@ static void __guc_exec_queue_process_msg_resume(struct xe_sched_msg *msg)
> >   #define RESUME		4
> >   #define OPCODE_MASK	0xf
> >   #define MSG_LOCKED	BIT(8)
> > +#define MSG_HEAD	BIT(9)
> >   static void guc_exec_queue_process_msg(struct xe_sched_msg *msg)
> >   {
> > @@ -1984,12 +2000,24 @@ static void guc_exec_queue_add_msg(struct xe_exec_queue *q, struct xe_sched_msg
> >   	msg->private_data = q;
> >   	trace_xe_sched_msg_add(msg);
> > -	if (opcode & MSG_LOCKED)
> > +	if (opcode & MSG_HEAD)
> > +		xe_sched_add_msg_head(&q->guc->sched, msg);
> > +	else if (opcode & MSG_LOCKED)
> >   		xe_sched_add_msg_locked(&q->guc->sched, msg);
> >   	else
> >   		xe_sched_add_msg(&q->guc->sched, msg);
> >   }
> > +static void guc_exec_queue_try_add_msg_head(struct xe_exec_queue *q,
> > +					    struct xe_sched_msg *msg,
> > +					    u32 opcode)
> > +{
> > +	if (!list_empty(&msg->link))
> > +		return;
> > +
> > +	guc_exec_queue_add_msg(q, msg, opcode | MSG_LOCKED | MSG_HEAD);
> > +}
> > +
> >   static bool guc_exec_queue_try_add_msg(struct xe_exec_queue *q,
> >   				       struct xe_sched_msg *msg,
> >   				       u32 opcode)
> > @@ -2264,6 +2292,93 @@ void xe_guc_submit_stop(struct xe_guc *guc)
> >   }
> > +/*
> > + * This function is quite complex but only real way to ensure no state is lost
> > + * during VF resume flows. The function scans the queue state, make adjustments
> > + * as needed, and queues jobs / messages which replayed upon unpause.
> > + */
> > +static void guc_exec_queue_pause(struct xe_guc *guc, struct xe_exec_queue *q)
> > +{
> > +	struct xe_gpu_scheduler *sched = &q->guc->sched;
> > +	struct xe_sched_job *job;
> > +	bool pending_enable, pending_disable, pending_resume;
> > +	int i;
> > +
> > +	lockdep_assert_held(&guc->submission_state.lock);
> > +
> > +	/* Stop scheduling + flush any DRM scheduler operations */
> > +	xe_sched_submission_stop(sched);
> > +	if (xe_exec_queue_is_lr(q))
> > +		cancel_work_sync(&q->guc->lr_tdr);
> > +	else
> > +		cancel_delayed_work_sync(&sched->base.work_tdr);
> 
> We're doing the same cancelling in `__guc_exec_queue_destroy_async()`, maybe
> close it into a function?
> 

I could but in follow up I'm going to drop &q->guc->lr_tdr and just use
DRM scheduler TDR so it will just be cancel_delayed_work_sync. I have
the code for this lying around and there a couple of upcoming features
this will get simplified if &q->guc->lr_tdr is droppped. 

> > +
> > +	pending_enable = exec_queue_pending_enable(q);
> > +	pending_resume = exec_queue_pending_resume(q);
> > +
> > +	if (pending_enable && pending_resume)
> > +		q->guc->needs_resume = true;
> > +
> > +	if (pending_enable && !pending_resume &&
> > +	    !exec_queue_pending_tdr_exit(q)) {
> > +		clear_exec_queue_registered(q);
> > +		if (xe_exec_queue_is_lr(q))
> > +			xe_exec_queue_put(q);
> > +	}
> > +
> > +	if (pending_enable) {
> > +		clear_exec_queue_enabled(q);
> > +		clear_exec_queue_pending_resume(q);
> > +		clear_exec_queue_pending_tdr_exit(q);
> > +		clear_exec_queue_pending_enable(q);
> > +	}
> > +
> > +	if (exec_queue_destroyed(q) && exec_queue_registered(q)) {
> > +		clear_exec_queue_destroyed(q);
> > +		if (exec_queue_extra_ref(q))
> > +			xe_exec_queue_put(q);
> > +		else
> > +			q->guc->needs_cleanup = true;
> > +		clear_exec_queue_extra_ref(q);
> > +	}
> > +
> > +	pending_disable = exec_queue_pending_disable(q);
> > +
> > +	if (pending_disable && exec_queue_suspended(q)) {
> > +		clear_exec_queue_suspended(q);
> > +		q->guc->needs_suspend = true;
> > +	}
> > +
> > +	if (pending_disable) {
> > +		if (!pending_enable)
> > +			set_exec_queue_enabled(q);
> > +		clear_exec_queue_pending_disable(q);
> > +		clear_exec_queue_check_timeout(q);
> > +	}
> 
> maybe we can close the above into a separate function as well?
> 
> ie. guc_exec_queue_undo_unfinished_state_change()?
> 

Do you mean all the parsing of exec queue state into a single function?
That seems reasonable. I don't really want to abstract each individual
if statement here as that in IMO that is too much abstraction and hard
to figure out what is exactly going on.

> guc_exec_queue_revert_pending_state_change()?
> 
> That would make this function easier to read, but also describe what we're
> doing.
> 
> Then, a counterfunction could be ripped out of guc_exec_queue_unpause().
> 
> > +
> > +	q->guc->resume_time = 0;
> > +
> > +	if (xe_exec_queue_is_parallel(q)) {
> > +		struct xe_device *xe = guc_to_xe(guc);
> > +		struct iosys_map map = xe_lrc_parallel_map(q->lrc[0]);
> > +
> > +		for (i = 0; i < WQ_SIZE / sizeof(u32); ++i)
> > +			parallel_write(xe, map, wq[i],
> > +				       FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_NOOP) |
> > +				       FIELD_PREP(WQ_LEN_MASK, 0));
> 
> ok so for parallel wq we're NOP'ing everything and adding the items back at
> new positions? Maybe a comment here would help in understanding that.
> 

Yes, the GuC didn't like messing with the head / tail. I can add a
comment.

Matt

> -Tomasz
> 
> > +	}
> > +
> > +	job = xe_sched_first_pending_job(sched);
> > +	if (job) {
> > +		/*
> > +		 * Adjust software tail so jobs submitted overwrite previous
> > +		 * position in ring buffer with new GGTT addresses.
> > +		 */
> > +		for (i = 0; i < q->width; ++i)
> > +			q->lrc[i]->ring.tail = job->ptrs[i].head;
> > +	}
> > +}
> > +
> >   /**
> >    * xe_guc_submit_pause - Stop further runs of submission tasks on given GuC.
> >    * @guc: the &xe_guc struct instance whose scheduler is to be disabled
> > @@ -2273,8 +2388,12 @@ 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)
> > -		xe_sched_submission_stop_async(&q->guc->sched);
> > +		guc_exec_queue_pause(guc, q);
> > +	mutex_unlock(&guc->submission_state.lock);
> >   }
> >   static void guc_exec_queue_start(struct xe_exec_queue *q)
> > @@ -2323,11 +2442,87 @@ int xe_guc_submit_start(struct xe_guc *guc)
> >   	return 0;
> >   }
> > -static void guc_exec_queue_unpause(struct xe_exec_queue *q)
> > +static void guc_exec_queue_unpause_prepare(struct xe_guc *guc,
> > +					   struct xe_exec_queue *q)
> >   {
> >   	struct xe_gpu_scheduler *sched = &q->guc->sched;
> > +	struct drm_sched_job *s_job;
> > +	struct xe_sched_job *job = NULL;
> > +
> > +	list_for_each_entry(s_job, &sched->base.pending_list, list) {
> > +		job = to_xe_sched_job(s_job);
> > +
> > +		q->ring_ops->emit_job(job);
> > +		job->skip_emit = true;
> > +	}
> > +
> > +	if (job)
> > +		job->last_replay = true;
> > +}
> > +
> > +/**
> > + * xe_guc_submit_unpause_prepare - Prepare unpause submission tasks on given GuC.
> > + * @guc: the &xe_guc struct instance whose scheduler is to be prepared for unpause
> > + */
> > +void xe_guc_submit_unpause_prepare(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)
> > +		guc_exec_queue_unpause_prepare(guc, q);
> > +	mutex_unlock(&guc->submission_state.lock);
> > +}
> > +
> > +static void guc_exec_queue_unpause(struct xe_guc *guc, struct xe_exec_queue *q)
> > +{
> > +	struct xe_gpu_scheduler *sched = &q->guc->sched;
> > +	struct xe_sched_msg *msg;
> > +	bool needs_tdr = exec_queue_killed_or_banned_or_wedged(q);
> > +
> > +	lockdep_assert_held(&guc->submission_state.lock);
> > +
> > +	xe_sched_resubmit_jobs(sched);
> > +
> > +	if (q->guc->needs_cleanup) {
> > +		msg = q->guc->static_msgs + STATIC_MSG_CLEANUP;
> > +
> > +		guc_exec_queue_add_msg(q, msg, CLEANUP);
> > +		q->guc->needs_cleanup = false;
> > +	}
> > +
> > +	if (q->guc->needs_suspend) {
> > +		msg = q->guc->static_msgs + STATIC_MSG_SUSPEND;
> > +
> > +		xe_sched_msg_lock(sched);
> > +		guc_exec_queue_try_add_msg_head(q, msg, SUSPEND);
> > +		xe_sched_msg_unlock(sched);
> > +
> > +		q->guc->needs_suspend = false;
> > +	}
> > +
> > +	/*
> > +	 * The resume must be in the message queue before the suspend as it is
> > +	 * not possible for a resume to be issued if a suspend pending is, but
> > +	 * the inverse is possible.
> > +	 */
> > +	if (q->guc->needs_resume) {
> > +		msg = q->guc->static_msgs + STATIC_MSG_RESUME;
> > +
> > +		xe_sched_msg_lock(sched);
> > +		guc_exec_queue_try_add_msg_head(q, msg, RESUME);
> > +		xe_sched_msg_unlock(sched);
> > +
> > +		q->guc->needs_resume = false;
> > +	}
> >   	xe_sched_submission_start(sched);
> > +	if (needs_tdr)
> > +		xe_guc_exec_queue_trigger_cleanup(q);
> > +	xe_sched_submission_resume_tdr(sched);
> >   }
> >   /**
> > @@ -2339,10 +2534,10 @@ void xe_guc_submit_unpause(struct xe_guc *guc)
> >   	struct xe_exec_queue *q;
> >   	unsigned long index;
> > +	mutex_lock(&guc->submission_state.lock);
> >   	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q)
> > -		guc_exec_queue_unpause(q);
> > -
> > -	wake_up_all(&guc->ct.wq);
> > +		guc_exec_queue_unpause(guc, q);
> > +	mutex_unlock(&guc->submission_state.lock);
> >   }
> >   /**
> > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.h b/drivers/gpu/drm/xe/xe_guc_submit.h
> > index fe82c317048e..b49a2748ec46 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_submit.h
> > +++ b/drivers/gpu/drm/xe/xe_guc_submit.h
> > @@ -22,6 +22,7 @@ void xe_guc_submit_stop(struct xe_guc *guc);
> >   int xe_guc_submit_start(struct xe_guc *guc);
> >   void xe_guc_submit_pause(struct xe_guc *guc);
> >   void xe_guc_submit_unpause(struct xe_guc *guc);
> > +void xe_guc_submit_unpause_prepare(struct xe_guc *guc);
> >   void xe_guc_submit_pause_abort(struct xe_guc *guc);
> >   void xe_guc_submit_wedge(struct xe_guc *guc);
> > diff --git a/drivers/gpu/drm/xe/xe_sched_job_types.h b/drivers/gpu/drm/xe/xe_sched_job_types.h
> > index 7ce58765a34a..13e7a12b03ad 100644
> > --- a/drivers/gpu/drm/xe/xe_sched_job_types.h
> > +++ b/drivers/gpu/drm/xe/xe_sched_job_types.h
> > @@ -63,6 +63,10 @@ struct xe_sched_job {
> >   	bool ring_ops_flush_tlb;
> >   	/** @ggtt: mapped in ggtt. */
> >   	bool ggtt;
> > +	/** @skip_emit: skip emitting the job */
> > +	bool skip_emit;
> > +	/** @last_replay: last job being replayed */
> > +	bool last_replay;
> >   	/** @ptrs: per instance pointers. */
> >   	struct xe_job_ptrs ptrs[];
> >   };