Re: [Intel-gfx] [PATCH 7/8] drm/i915/gt: Decouple inflight virtual engines

[Chris Wilson <chris@chris-wilson.co.uk>]

Quoting Tvrtko Ursulin (2020-05-18 15:55:46)
> 
> On 18/05/2020 14:00, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2020-05-18 13:53:29)
> >>
> >> On 18/05/2020 09:14, Chris Wilson wrote:
> >>> Once a virtual engine has been bound to a sibling, it will remain bound
> >>> until we finally schedule out the last active request. We can not rebind
> >>> the context to a new sibling while it is inflight as the context save
> >>> will conflict, hence we wait. As we cannot then use any other sibliing
> >>> while the context is inflight, only kick the bound sibling while it
> >>> inflight and upon scheduling out the kick the rest (so that we can swap
> >>> engines on timeslicing if the previously bound engine becomes
> >>> oversubscribed).
> >>>
> >>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> >>> ---
> >>>    drivers/gpu/drm/i915/gt/intel_lrc.c | 30 +++++++++++++----------------
> >>>    1 file changed, 13 insertions(+), 17 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
> >>> index 7a5ac3375225..fe8f3518d6b8 100644
> >>> --- a/drivers/gpu/drm/i915/gt/intel_lrc.c
> >>> +++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
> >>> @@ -1398,9 +1398,8 @@ execlists_schedule_in(struct i915_request *rq, int idx)
> >>>    static void kick_siblings(struct i915_request *rq, struct intel_context *ce)
> >>>    {
> >>>        struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
> >>> -     struct i915_request *next = READ_ONCE(ve->request);
> >>>    
> >>> -     if (next == rq || (next && next->execution_mask & ~rq->execution_mask))
> >>> +     if (READ_ONCE(ve->request))
> >>>                tasklet_hi_schedule(&ve->base.execlists.tasklet);
> >>>    }
> >>>    
> >>> @@ -1821,18 +1820,14 @@ first_virtual_engine(struct intel_engine_cs *engine)
> >>>                        rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
> >>>                struct i915_request *rq = READ_ONCE(ve->request);
> >>>    
> >>> -             if (!rq) { /* lazily cleanup after another engine handled rq */
> >>> +             /* lazily cleanup after another engine handled rq */
> >>> +             if (!rq || !virtual_matches(ve, rq, engine)) {
> >>>                        rb_erase_cached(rb, &el->virtual);
> >>>                        RB_CLEAR_NODE(rb);
> >>>                        rb = rb_first_cached(&el->virtual);
> >>>                        continue;
> >>>                }
> >>>    
> >>> -             if (!virtual_matches(ve, rq, engine)) {
> >>> -                     rb = rb_next(rb);
> >>> -                     continue;
> >>> -             }
> >>> -
> >>>                return ve;
> >>>        }
> >>>    
> >>> @@ -5478,7 +5473,6 @@ static void virtual_submission_tasklet(unsigned long data)
> >>>        if (unlikely(!mask))
> >>>                return;
> >>>    
> >>> -     local_irq_disable();
> >>>        for (n = 0; n < ve->num_siblings; n++) {
> >>>                struct intel_engine_cs *sibling = READ_ONCE(ve->siblings[n]);
> >>>                struct ve_node * const node = &ve->nodes[sibling->id];
> >>> @@ -5488,20 +5482,19 @@ static void virtual_submission_tasklet(unsigned long data)
> >>>                if (!READ_ONCE(ve->request))
> >>>                        break; /* already handled by a sibling's tasklet */
> >>>    
> >>> +             spin_lock_irq(&sibling->active.lock);
> >>> +
> >>>                if (unlikely(!(mask & sibling->mask))) {
> >>>                        if (!RB_EMPTY_NODE(&node->rb)) {
> >>> -                             spin_lock(&sibling->active.lock);
> >>>                                rb_erase_cached(&node->rb,
> >>>                                                &sibling->execlists.virtual);
> >>>                                RB_CLEAR_NODE(&node->rb);
> >>> -                             spin_unlock(&sibling->active.lock);
> >>>                        }
> >>> -                     continue;
> >>> -             }
> >>>    
> >>> -             spin_lock(&sibling->active.lock);
> >>> +                     goto unlock_engine;
> >>> +             }
> >>>    
> >>> -             if (!RB_EMPTY_NODE(&node->rb)) {
> >>> +             if (unlikely(!RB_EMPTY_NODE(&node->rb))) {
> >>>                        /*
> >>>                         * Cheat and avoid rebalancing the tree if we can
> >>>                         * reuse this node in situ.
> >>> @@ -5541,9 +5534,12 @@ static void virtual_submission_tasklet(unsigned long data)
> >>>                if (first && prio >= sibling->execlists.queue_priority_hint)
> >>>                        tasklet_hi_schedule(&sibling->execlists.tasklet);
> >>>    
> >>> -             spin_unlock(&sibling->active.lock);
> >>> +unlock_engine:
> >>> +             spin_unlock_irq(&sibling->active.lock);
> >>> +
> >>> +             if (intel_context_inflight(&ve->context))
> >>> +                     break;
> >>
> >> So virtual request may not be added to all siblings any longer. Will it
> >> still be able to schedule it on any if time slicing kicks in under these
> >> conditions?
> > 
> > Yes.
> >   
> >> This is equivalent to the hunk in first_virtual_engine which also
> >> removes it from all other siblings.
> >>
> >> I guess it's inline with what the commit messages says - that new
> >> sibling will be picked upon time slicing. I just don't quite see the
> >> path which would do it. Only path which shuffles the siblings array
> >> around is in dequeue, and dequeue on other that the engine which first
> >> picked it will not happen any more. I must be missing something..
> > 
> > It's all on the execlists_schedule_out. During timeslicing we call
> > unwind_incomplete_requests which moves the requests back to the priotree
> > (and in this patch back to the virtual engine).
> > 
> > But... We cannot use the virtual request on any other engine until it has
> > been scheduled out. That only happens after we complete execlists_dequeue()
> > and submit a new ELSP[]. Once the HW acks the change, we call
> > execlists_schedule_out on the virtual_request.
> > 
> > Now we known that intel_context_inflight() will return false so any
> > engine can pick up the request, and so it's time to kick the virtual
> > tasklet and in turn kick all the siblings.
> > 
> > So timeslicing works by not submitting the virtual request again and
> > when it expires on this sibling[0], we wake up all the other siblings
> > and the first that is idle wins the race.
> 
> If a virtual request is on hw and timeslice expires:
> 
> 1. Unwinds the request.
>        -> kicks the virtual tasklet
> 2. Virtual tasklet runs and puts the request back on siblings.
>        -> kicks the physical tasklets
> 3. Siblings tasklet runs and submits the request.
> 
> So two tasklets latency even if no other runnable requests?

Yes. It's worse than that when you look at it. Since the other
execlists_submission_tasklet will run *before* we schedule out, they do
nothing as they cannot use the virtual request. So they will only
function when we kick them again after the schedule-out.

This was "bug" number 3 I mentioned you found with your tiny snippet.
In the next patch we optimise away the ineffective timeslice preemptions
of the virtual request. The inter-engine inefficiencies are inherent to
the system though -- we can only schedule the other engines once we have
scheduled out after the context save of the virtual engine.
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx