Re: [PATCH 17/47] drm/i915/guc: Extend deregistration fence to schedule disable

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

On Fri, Jul 09, 2021 at 03:59:11PM -0700, John Harrison wrote:
> On 6/24/2021 00:04, Matthew Brost wrote:
> > Extend the deregistration context fence to fence whne a GuC context has
> > scheduling disable pending.
> > 
> > Cc: John Harrison <john.c.harrison@intel.com>
> > Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> > ---
> >   .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 37 +++++++++++++++----
> >   1 file changed, 30 insertions(+), 7 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> > index 0386ccd5a481..0a6ccdf32316 100644
> > --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> > +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> > @@ -918,7 +918,19 @@ static void guc_context_sched_disable(struct intel_context *ce)
> >   		goto unpin;
> >   	spin_lock_irqsave(&ce->guc_state.lock, flags);
> > +
> > +	/*
> > +	 * We have to check if the context has been pinned again as another pin
> > +	 * operation is allowed to pass this function. Checking the pin count
> > +	 * here synchronizes this function with guc_request_alloc ensuring a
> > +	 * request doesn't slip through the 'context_pending_disable' fence.
> > +	 */
> The pin count is an atomic so doesn't need the spinlock. Also the above

How about?

/*
 * We have to check if the context has been pinned again as another pin
 * operation is allowed to pass this function. Checking the pin count,
 * within ce->guc_state.lock, synchronizes this function with
 * guc_request_alloc ensuring a request doesn't slip through the
 * 'context_pending_disable' fence. Checking within the spin lock (can't
 * sleep) ensures another process doesn't pin this context and generate
 * a request before we set the 'context_pending_disable' flag here.
 */

Matt

> comment 'checking the pin count here synchronizes ...' seems wrong. Isn't
> the point that acquiring the spinlock is what synchronises with
> guc_request_alloc? So the comment should be before the spinlock acquire and
> should mention using the spinlock for this purpose?
> 
> John.
> 
> 
> > +	if (unlikely(atomic_add_unless(&ce->pin_count, -2, 2))) {
> > +		spin_unlock_irqrestore(&ce->guc_state.lock, flags);
> > +		return;
> > +	}
> >   	guc_id = prep_context_pending_disable(ce);
> > +
> >   	spin_unlock_irqrestore(&ce->guc_state.lock, flags);
> >   	with_intel_runtime_pm(runtime_pm, wakeref)
> > @@ -1123,19 +1135,22 @@ static int guc_request_alloc(struct i915_request *rq)
> >   out:
> >   	/*
> >   	 * We block all requests on this context if a G2H is pending for a
> > -	 * context deregistration as the GuC will fail a context registration
> > -	 * while this G2H is pending. Once a G2H returns, the fence is released
> > -	 * that is blocking these requests (see guc_signal_context_fence).
> > +	 * schedule disable or context deregistration as the GuC will fail a
> > +	 * schedule enable or context registration if either G2H is pending
> > +	 * respectfully. Once a G2H returns, the fence is released that is
> > +	 * blocking these requests (see guc_signal_context_fence).
> >   	 *
> > -	 * We can safely check the below field outside of the lock as it isn't
> > -	 * possible for this field to transition from being clear to set but
> > +	 * We can safely check the below fields outside of the lock as it isn't
> > +	 * possible for these fields to transition from being clear to set but
> >   	 * converse is possible, hence the need for the check within the lock.
> >   	 */
> > -	if (likely(!context_wait_for_deregister_to_register(ce)))
> > +	if (likely(!context_wait_for_deregister_to_register(ce) &&
> > +		   !context_pending_disable(ce)))
> >   		return 0;
> >   	spin_lock_irqsave(&ce->guc_state.lock, flags);
> > -	if (context_wait_for_deregister_to_register(ce)) {
> > +	if (context_wait_for_deregister_to_register(ce) ||
> > +	    context_pending_disable(ce)) {
> >   		i915_sw_fence_await(&rq->submit);
> >   		list_add_tail(&rq->guc_fence_link, &ce->guc_state.fences);
> > @@ -1484,10 +1499,18 @@ int intel_guc_sched_done_process_msg(struct intel_guc *guc,
> >   	if (context_pending_enable(ce)) {
> >   		clr_context_pending_enable(ce);
> >   	} else if (context_pending_disable(ce)) {
> > +		/*
> > +		 * Unpin must be done before __guc_signal_context_fence,
> > +		 * otherwise a race exists between the requests getting
> > +		 * submitted + retired before this unpin completes resulting in
> > +		 * the pin_count going to zero and the context still being
> > +		 * enabled.
> > +		 */
> >   		intel_context_sched_disable_unpin(ce);
> >   		spin_lock_irqsave(&ce->guc_state.lock, flags);
> >   		clr_context_pending_disable(ce);
> > +		__guc_signal_context_fence(ce);
> >   		spin_unlock_irqrestore(&ce->guc_state.lock, flags);
> >   	}
>