drm/i915/gt: Incorporate the virtual engine into timeslicing
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 19 May 2020 13:20:46 +0000 (14:20 +0100)
committerJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Mon, 15 Jun 2020 09:35:53 +0000 (12:35 +0300)
It was quite the oversight to only factor in the normal queue to decide
the timeslicing switch priority. By leaving out the next virtual request
from the priority decision, we would not timeslice the current engine if
there was an available virtual request.

Testcase: igt/gem_exec_balancer/sliced
Fixes: 3df2deed411e ("drm/i915/execlists: Enable timeslice on partial virtual engine dequeue")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200519132046.22443-3-chris@chris-wilson.co.uk
(cherry picked from commit 6ad249ba59badc7ff157d4db1f835748f0e2c9b6)
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
drivers/gpu/drm/i915/gt/intel_lrc.c

index 87e6c5bdd2dcef7e94a7e4095ad7e5b8438c5158..0b58e6c1e235247c178131d756badfc7b7c62962 100644 (file)
@@ -1895,7 +1895,8 @@ static void defer_active(struct intel_engine_cs *engine)
 
 static bool
 need_timeslice(const struct intel_engine_cs *engine,
-              const struct i915_request *rq)
+              const struct i915_request *rq,
+              const struct rb_node *rb)
 {
        int hint;
 
@@ -1903,6 +1904,24 @@ need_timeslice(const struct intel_engine_cs *engine,
                return false;
 
        hint = engine->execlists.queue_priority_hint;
+
+       if (rb) {
+               const struct virtual_engine *ve =
+                       rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
+               const struct intel_engine_cs *inflight =
+                       intel_context_inflight(&ve->context);
+
+               if (!inflight || inflight == engine) {
+                       struct i915_request *next;
+
+                       rcu_read_lock();
+                       next = READ_ONCE(ve->request);
+                       if (next)
+                               hint = max(hint, rq_prio(next));
+                       rcu_read_unlock();
+               }
+       }
+
        if (!list_is_last(&rq->sched.link, &engine->active.requests))
                hint = max(hint, rq_prio(list_next_entry(rq, sched.link)));
 
@@ -1977,10 +1996,9 @@ static void set_timeslice(struct intel_engine_cs *engine)
        set_timer_ms(&engine->execlists.timer, duration);
 }
 
-static void start_timeslice(struct intel_engine_cs *engine)
+static void start_timeslice(struct intel_engine_cs *engine, int prio)
 {
        struct intel_engine_execlists *execlists = &engine->execlists;
-       const int prio = queue_prio(execlists);
        unsigned long duration;
 
        if (!intel_engine_has_timeslices(engine))
@@ -2140,7 +2158,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                        __unwind_incomplete_requests(engine);
 
                        last = NULL;
-               } else if (need_timeslice(engine, last) &&
+               } else if (need_timeslice(engine, last, rb) &&
                           timeslice_expired(execlists, last)) {
                        if (i915_request_completed(last)) {
                                tasklet_hi_schedule(&execlists->tasklet);
@@ -2188,7 +2206,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                                 * Even if ELSP[1] is occupied and not worthy
                                 * of timeslices, our queue might be.
                                 */
-                               start_timeslice(engine);
+                               start_timeslice(engine, queue_prio(execlists));
                                return;
                        }
                }
@@ -2223,7 +2241,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 
                        if (last && !can_merge_rq(last, rq)) {
                                spin_unlock(&ve->base.active.lock);
-                               start_timeslice(engine);
+                               start_timeslice(engine, rq_prio(rq));
                                return; /* leave this for another sibling */
                        }