drm/i915: Add kick_backend function to i915_sched_engine
authorMatthew Brost <matthew.brost@intel.com>
Fri, 18 Jun 2021 01:06:36 +0000 (18:06 -0700)
committerMatt Roper <matthew.d.roper@intel.com>
Fri, 18 Jun 2021 22:14:09 +0000 (15:14 -0700)
Not all back-ends require a kick after a  scheduling update, so make the
kick a call-back function that the  back-end can opt-in to. Also move
the current kick function from the scheduler to the execlists file as it
is specific to that back-end.

Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210618010638.98941-7-matthew.brost@intel.com
drivers/gpu/drm/i915/gt/intel_execlists_submission.c
drivers/gpu/drm/i915/i915_scheduler.c
drivers/gpu/drm/i915/i915_scheduler_types.h

index 8a3d401..9487d9e 100644 (file)
@@ -3116,10 +3116,61 @@ static bool can_preempt(struct intel_engine_cs *engine)
        return engine->class != RENDER_CLASS;
 }
 
+static void kick_execlists(const struct i915_request *rq, int prio)
+{
+       struct intel_engine_cs *engine = rq->engine;
+       struct i915_sched_engine *sched_engine = engine->sched_engine;
+       const struct i915_request *inflight;
+
+       /*
+        * We only need to kick the tasklet once for the high priority
+        * new context we add into the queue.
+        */
+       if (prio <= sched_engine->queue_priority_hint)
+               return;
+
+       rcu_read_lock();
+
+       /* Nothing currently active? We're overdue for a submission! */
+       inflight = execlists_active(&engine->execlists);
+       if (!inflight)
+               goto unlock;
+
+       /*
+        * If we are already the currently executing context, don't
+        * bother evaluating if we should preempt ourselves.
+        */
+       if (inflight->context == rq->context)
+               goto unlock;
+
+       ENGINE_TRACE(engine,
+                    "bumping queue-priority-hint:%d for rq:%llx:%lld, inflight:%llx:%lld prio %d\n",
+                    prio,
+                    rq->fence.context, rq->fence.seqno,
+                    inflight->fence.context, inflight->fence.seqno,
+                    inflight->sched.attr.priority);
+
+       sched_engine->queue_priority_hint = prio;
+
+       /*
+        * Allow preemption of low -> normal -> high, but we do
+        * not allow low priority tasks to preempt other low priority
+        * tasks under the impression that latency for low priority
+        * tasks does not matter (as much as background throughput),
+        * so kiss.
+        */
+       if (prio >= max(I915_PRIORITY_NORMAL, rq_prio(inflight)))
+               tasklet_hi_schedule(&engine->execlists.tasklet);
+
+unlock:
+       rcu_read_unlock();
+}
+
 static void execlists_set_default_submission(struct intel_engine_cs *engine)
 {
        engine->submit_request = execlists_submit_request;
        engine->sched_engine->schedule = i915_schedule;
+       engine->sched_engine->kick_backend = kick_execlists;
        engine->execlists.tasklet.callback = execlists_submission_tasklet;
 }
 
@@ -3702,6 +3753,7 @@ intel_execlists_create_virtual(struct intel_engine_cs **siblings,
        ve->base.request_alloc = execlists_request_alloc;
 
        ve->base.sched_engine->schedule = i915_schedule;
+       ve->base.sched_engine->kick_backend = kick_execlists;
        ve->base.submit_request = virtual_submit_request;
        ve->base.bond_execute = virtual_bond_execute;
 
index 4bc6969..035b88f 100644 (file)
@@ -157,65 +157,6 @@ sched_lock_engine(const struct i915_sched_node *node,
        return locked;
 }
 
-static inline int rq_prio(const struct i915_request *rq)
-{
-       return rq->sched.attr.priority;
-}
-
-static inline bool need_preempt(int prio, int active)
-{
-       /*
-        * Allow preemption of low -> normal -> high, but we do
-        * not allow low priority tasks to preempt other low priority
-        * tasks under the impression that latency for low priority
-        * tasks does not matter (as much as background throughput),
-        * so kiss.
-        */
-       return prio >= max(I915_PRIORITY_NORMAL, active);
-}
-
-static void kick_submission(struct intel_engine_cs *engine,
-                           const struct i915_request *rq,
-                           int prio)
-{
-       const struct i915_request *inflight;
-
-       /*
-        * We only need to kick the tasklet once for the high priority
-        * new context we add into the queue.
-        */
-       if (prio <= engine->sched_engine->queue_priority_hint)
-               return;
-
-       rcu_read_lock();
-
-       /* Nothing currently active? We're overdue for a submission! */
-       inflight = execlists_active(&engine->execlists);
-       if (!inflight)
-               goto unlock;
-
-       /*
-        * If we are already the currently executing context, don't
-        * bother evaluating if we should preempt ourselves.
-        */
-       if (inflight->context == rq->context)
-               goto unlock;
-
-       ENGINE_TRACE(engine,
-                    "bumping queue-priority-hint:%d for rq:%llx:%lld, inflight:%llx:%lld prio %d\n",
-                    prio,
-                    rq->fence.context, rq->fence.seqno,
-                    inflight->fence.context, inflight->fence.seqno,
-                    inflight->sched.attr.priority);
-
-       engine->sched_engine->queue_priority_hint = prio;
-       if (need_preempt(prio, rq_prio(inflight)))
-               tasklet_hi_schedule(&engine->execlists.tasklet);
-
-unlock:
-       rcu_read_unlock();
-}
-
 static void __i915_schedule(struct i915_sched_node *node,
                            const struct i915_sched_attr *attr)
 {
@@ -335,7 +276,8 @@ static void __i915_schedule(struct i915_sched_node *node,
                }
 
                /* Defer (tasklet) submission until after all of our updates. */
-               kick_submission(engine, node_to_request(node), prio);
+               if (engine->sched_engine->kick_backend)
+                       engine->sched_engine->kick_backend(node_to_request(node), prio);
        }
 
        spin_unlock(&engine->sched_engine->lock);
index 0c1e417..8bd07d0 100644 (file)
@@ -153,6 +153,12 @@ struct i915_sched_engine {
         */
        bool no_priolist;
 
+       /**
+        * @kick_backend: kick backend after a request's priority has changed
+        */
+       void    (*kick_backend)(const struct i915_request *rq,
+                               int prio);
+
        /**
         * @schedule: adjust priority of request
         *