drm/i915/gt: Make timeslice duration configurable
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 29 Oct 2019 09:16:32 +0000 (09:16 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Tue, 29 Oct 2019 16:23:55 +0000 (16:23 +0000)
Execlists uses a scheduling quantum (a timeslice) to alternate execution
between ready-to-run contexts of equal priority. This ensures that all
users (though only if they of equal importance) have the opportunity to
run and prevents livelocks where contexts may have implicit ordering due
to userspace semaphores. However, not all workloads necessarily benefit
from timeslicing and in the extreme some sysadmin may want to disable or
reduce the timeslicing granularity.

The timeslicing mechanism can be compiled out^W^W disabled (but should
DCE!) with

./scripts/config --set-val DRM_I915_TIMESLICE_DURATION 0

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191029091632.26281-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/Kconfig.profile
drivers/gpu/drm/i915/gt/intel_engine.h
drivers/gpu/drm/i915/gt/intel_engine_cs.c
drivers/gpu/drm/i915/gt/intel_engine_types.h
drivers/gpu/drm/i915/gt/intel_lrc.c
drivers/gpu/drm/i915/gt/selftest_lrc.c

index 8ab7af5eb31194f2c44a6e94349dc30a35e24102..1799537a32288aa30d582958dd18c6918bc51fd8 100644 (file)
@@ -59,3 +59,18 @@ config DRM_I915_STOP_TIMEOUT
          damage as the system is reset in order to recover. The corollary is
          that the reset itself may take longer and so be more disruptive to
          interactive or low latency workloads.
+
+config DRM_I915_TIMESLICE_DURATION
+       int "Scheduling quantum for userspace batches (ms, jiffy granularity)"
+       default 1 # milliseconds
+       help
+         When two user batches of equal priority are executing, we will
+         alternate execution of each batch to ensure forward progress of
+         all users. This is necessary in some cases where there may be
+         an implicit dependency between those batches that requires
+         concurrent execution in order for them to proceed, e.g. they
+         interact with each other via userspace semaphores. Each context
+         is scheduled for execution for the timeslice duration, before
+         switching to the next context.
+
+         May be 0 to disable timeslicing.
index c6895938b6269a70ea5bfb95e834dd54eb63b59c..bc3b72bfa9e3be8b39d4b1d9111bc219f67710e3 100644 (file)
@@ -329,10 +329,19 @@ void intel_engine_init_active(struct intel_engine_cs *engine,
 static inline bool
 intel_engine_has_preempt_reset(const struct intel_engine_cs *engine)
 {
-       if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
-               return 0;
+       if (!IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT))
+               return false;
 
        return intel_engine_has_preemption(engine);
 }
 
+static inline bool
+intel_engine_has_timeslices(const struct intel_engine_cs *engine)
+{
+       if (!IS_ACTIVE(CONFIG_DRM_I915_TIMESLICE_DURATION))
+               return false;
+
+       return intel_engine_has_semaphores(engine);
+}
+
 #endif /* _INTEL_RINGBUFFER_H_ */
index 355523114c712d2f40e6df36f16d3a42577c4fa2..f8113bc756c66304972ab4f5b96a8aa99742915f 100644 (file)
@@ -315,6 +315,8 @@ static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id)
                CONFIG_DRM_I915_PREEMPT_TIMEOUT;
        engine->props.stop_timeout_ms =
                CONFIG_DRM_I915_STOP_TIMEOUT;
+       engine->props.timeslice_duration_ms =
+               CONFIG_DRM_I915_TIMESLICE_DURATION;
 
        /*
         * To be overridden by the backend on setup. However to facilitate
index e8ea12b96755cbf34bb639db02f0896988b8765e..c5d1047a4bc5f482356ba5f87bc4ebf33d0c0d87 100644 (file)
@@ -523,6 +523,7 @@ struct intel_engine_cs {
                unsigned long heartbeat_interval_ms;
                unsigned long preempt_timeout_ms;
                unsigned long stop_timeout_ms;
+               unsigned long timeslice_duration_ms;
        } props;
 };
 
index 2f474c1f5c54a7bb99868d62a803a7d91ca05bdc..6fb3def5ba169fab0ba08e64574b105cd42bd376 100644 (file)
@@ -1467,7 +1467,7 @@ need_timeslice(struct intel_engine_cs *engine, const struct i915_request *rq)
 {
        int hint;
 
-       if (!intel_engine_has_semaphores(engine))
+       if (!intel_engine_has_timeslices(engine))
                return false;
 
        if (list_is_last(&rq->sched.link, &engine->active.requests))
@@ -1488,15 +1488,32 @@ switch_prio(struct intel_engine_cs *engine, const struct i915_request *rq)
        return rq_prio(list_next_entry(rq, sched.link));
 }
 
-static bool
-enable_timeslice(const struct intel_engine_execlists *execlists)
+static inline unsigned long
+timeslice(const struct intel_engine_cs *engine)
+{
+       return READ_ONCE(engine->props.timeslice_duration_ms);
+}
+
+static unsigned long
+active_timeslice(const struct intel_engine_cs *engine)
 {
-       const struct i915_request *rq = *execlists->active;
+       const struct i915_request *rq = *engine->execlists.active;
 
        if (i915_request_completed(rq))
-               return false;
+               return 0;
+
+       if (engine->execlists.switch_priority_hint < effective_prio(rq))
+               return 0;
+
+       return timeslice(engine);
+}
+
+static void set_timeslice(struct intel_engine_cs *engine)
+{
+       if (!intel_engine_has_timeslices(engine))
+               return;
 
-       return execlists->switch_priority_hint >= effective_prio(rq);
+       set_timer_ms(&engine->execlists.timer, active_timeslice(engine));
 }
 
 static void record_preemption(struct intel_engine_execlists *execlists)
@@ -1667,8 +1684,9 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                                 */
                                if (!execlists->timer.expires &&
                                    need_timeslice(engine, last))
-                                       mod_timer(&execlists->timer,
-                                                 jiffies + 1);
+                                       set_timer_ms(&execlists->timer,
+                                                    timeslice(engine));
+
                                return;
                        }
 
@@ -2092,10 +2110,7 @@ static void process_csb(struct intel_engine_cs *engine)
                                       execlists_num_ports(execlists) *
                                       sizeof(*execlists->pending));
 
-                       if (enable_timeslice(execlists))
-                               mod_timer(&execlists->timer, jiffies + 1);
-                       else
-                               cancel_timer(&execlists->timer);
+                       set_timeslice(engine);
 
                        WRITE_ONCE(execlists->pending[0], NULL);
                } else {
index 9507d750ae14df14eab291a10ca046095256b081..5e8c365548f0be9a35d6c73467ba6eb5d1dcc041 100644 (file)
@@ -440,6 +440,8 @@ static int live_timeslice_preempt(void *arg)
         * need to preempt the current task and replace it with another
         * ready task.
         */
+       if (!IS_ACTIVE(CONFIG_DRM_I915_TIMESLICE_DURATION))
+               return 0;
 
        obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
        if (IS_ERR(obj))
@@ -514,6 +516,11 @@ static void wait_for_submit(struct intel_engine_cs *engine,
        } while (!i915_request_is_active(rq));
 }
 
+static long timeslice_threshold(const struct intel_engine_cs *engine)
+{
+       return 2 * msecs_to_jiffies_timeout(timeslice(engine)) + 1;
+}
+
 static int live_timeslice_queue(void *arg)
 {
        struct intel_gt *gt = arg;
@@ -531,6 +538,8 @@ static int live_timeslice_queue(void *arg)
         * ELSP[1] is already occupied, so must rely on timeslicing to
         * eject ELSP[0] in favour of the queue.)
         */
+       if (!IS_ACTIVE(CONFIG_DRM_I915_TIMESLICE_DURATION))
+               return 0;
 
        obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
        if (IS_ERR(obj))
@@ -608,8 +617,8 @@ static int live_timeslice_queue(void *arg)
                        err = -EINVAL;
                }
 
-               /* Timeslice every jiffie, so within 2 we should signal */
-               if (i915_request_wait(rq, 0, 3) < 0) {
+               /* Timeslice every jiffy, so within 2 we should signal */
+               if (i915_request_wait(rq, 0, timeslice_threshold(engine)) < 0) {
                        struct drm_printer p =
                                drm_info_printer(gt->i915->drm.dev);
 
@@ -1383,7 +1392,7 @@ static int __cancel_hostile(struct live_preempt_cancel *arg)
        int err;
 
        /* Preempt cancel non-preemptible spinner in ELSP0 */
-       if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
+       if (!IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT))
                return 0;
 
        GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
@@ -2030,7 +2039,7 @@ static int live_preempt_timeout(void *arg)
         * Check that we force preemption to occur by cancelling the previous
         * context if it refuses to yield the GPU.
         */
-       if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
+       if (!IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT))
                return 0;
 
        if (!HAS_LOGICAL_RING_PREEMPTION(gt->i915))