drm/i915/execlists: Avoid kicking priority on the current context
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 25 Sep 2018 08:31:59 +0000 (09:31 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 27 Sep 2018 12:45:20 +0000 (13:45 +0100)
If the request is currently on the HW (in port 0), then we do not need
to kick the submission tasklet to evaluate whether we should be
preempting itself in order to execute it again.

In the case that was annoying me:

   execlists_schedule: rq(18:211173).prio=0 -> 2
   need_preempt: last(18:211174).prio=0, queue.prio=2

We are bumping the priority of the first of a pair of requests running
in the current context. Then when evaluating preempt, we would see that
that our priority request is higher than the last executing request in
ELSP0 and so trigger preemption, not realising that our intended request
was already executing.

v2: As we assume state of the execlists->port[] that is only valid while
we hold the timeline lock we have to repeat some earlier tests that on
the validity of the node.
v3: Wrap guc submission under the timeline.lock as is now the way of all
things.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180925083205.2229-2-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/intel_guc_submission.c
drivers/gpu/drm/i915/intel_lrc.c

index a81f04d46e87650b7185a6508afe9d4f72567ddd..4874a212754c142f9a4db64474fd6507f41a951e 100644 (file)
@@ -791,19 +791,8 @@ done:
 
 static void guc_dequeue(struct intel_engine_cs *engine)
 {
-       unsigned long flags;
-       bool submit;
-
-       local_irq_save(flags);
-
-       spin_lock(&engine->timeline.lock);
-       submit = __guc_dequeue(engine);
-       spin_unlock(&engine->timeline.lock);
-
-       if (submit)
+       if (__guc_dequeue(engine))
                guc_submit(engine);
-
-       local_irq_restore(flags);
 }
 
 static void guc_submission_tasklet(unsigned long data)
@@ -812,6 +801,9 @@ static void guc_submission_tasklet(unsigned long data)
        struct intel_engine_execlists * const execlists = &engine->execlists;
        struct execlist_port *port = execlists->port;
        struct i915_request *rq;
+       unsigned long flags;
+
+       spin_lock_irqsave(&engine->timeline.lock, flags);
 
        rq = port_request(port);
        while (rq && i915_request_completed(rq)) {
@@ -835,6 +827,8 @@ static void guc_submission_tasklet(unsigned long data)
 
        if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT))
                guc_dequeue(engine);
+
+       spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }
 
 static struct i915_request *
index 593999c1e2aad33b3f38ddddbc45020e2fab30ba..15345e74d8ce1224907ba91401569fe3b604e3ac 100644 (file)
@@ -356,13 +356,8 @@ execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists)
 {
        struct intel_engine_cs *engine =
                container_of(execlists, typeof(*engine), execlists);
-       unsigned long flags;
-
-       spin_lock_irqsave(&engine->timeline.lock, flags);
 
        __unwind_incomplete_requests(engine);
-
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }
 
 static inline void
@@ -1233,9 +1228,13 @@ static void execlists_schedule(struct i915_request *request,
 
                engine = sched_lock_engine(node, engine);
 
+               /* Recheck after acquiring the engine->timeline.lock */
                if (prio <= node->attr.priority)
                        continue;
 
+               if (i915_sched_node_signaled(node))
+                       continue;
+
                node->attr.priority = prio;
                if (!list_empty(&node->link)) {
                        if (last != engine) {
@@ -1244,14 +1243,34 @@ static void execlists_schedule(struct i915_request *request,
                        }
                        GEM_BUG_ON(pl->priority != prio);
                        list_move_tail(&node->link, &pl->requests);
+               } else {
+                       /*
+                        * If the request is not in the priolist queue because
+                        * it is not yet runnable, then it doesn't contribute
+                        * to our preemption decisions. On the other hand,
+                        * if the request is on the HW, it too is not in the
+                        * queue; but in that case we may still need to reorder
+                        * the inflight requests.
+                        */
+                       if (!i915_sw_fence_done(&sched_to_request(node)->submit))
+                               continue;
                }
 
-               if (prio > engine->execlists.queue_priority &&
-                   i915_sw_fence_done(&sched_to_request(node)->submit)) {
-                       /* defer submission until after all of our updates */
-                       __update_queue(engine, prio);
-                       tasklet_hi_schedule(&engine->execlists.tasklet);
-               }
+               if (prio <= engine->execlists.queue_priority)
+                       continue;
+
+               /*
+                * If we are already the currently executing context, don't
+                * bother evaluating if we should preempt ourselves.
+                */
+               if (sched_to_request(node)->global_seqno &&
+                   i915_seqno_passed(port_request(engine->execlists.port)->global_seqno,
+                                     sched_to_request(node)->global_seqno))
+                       continue;
+
+               /* Defer (tasklet) submission until after all of our updates. */
+               __update_queue(engine, prio);
+               tasklet_hi_schedule(&engine->execlists.tasklet);
        }
 
        spin_unlock_irq(&engine->timeline.lock);