@@ -59,3 +59,21 @@ 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.
+
+ This is adjustable via
+ /sys/class/drm/card?/engine/*/timeslice_duration_ms
+
+ May be 0 to disable timeslicing.
@@ -390,4 +390,13 @@ intel_engine_has_preempt_reset(const struct intel_engine_cs *engine)
return intel_engine_has_preemption(engine);
}
+static inline bool
+intel_engine_has_timeslices(const struct intel_engine_cs *engine)
+{
+ if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
+ return 0;
+
+ return intel_engine_has_semaphores(engine);
+}
+
#endif /* _INTEL_RINGBUFFER_H_ */
@@ -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
@@ -142,6 +142,48 @@ all_caps_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
static struct kobj_attribute all_caps_attr =
__ATTR(known_capabilities, 0444, all_caps_show, NULL);
+static ssize_t
+timeslice_store(struct kobject *kobj, struct kobj_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct intel_engine_cs *engine = kobj_to_engine(kobj);
+ unsigned long long duration;
+ int err;
+
+ /*
+ * 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.
+ */
+
+ err = kstrtoull(buf, 0, &duration);
+ if (err)
+ return err;
+
+ if (duration > jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT))
+ return -EINVAL;
+
+ WRITE_ONCE(engine->props.timeslice_duration_ms, duration);
+
+ if (execlists_active(&engine->execlists))
+ set_timer_ms(&engine->execlists.timer, duration);
+
+ return count;
+}
+
+static ssize_t
+timeslice_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
+{
+ struct intel_engine_cs *engine = kobj_to_engine(kobj);
+
+ return sprintf(buf, "%lu\n", engine->props.timeslice_duration_ms);
+}
+
+static struct kobj_attribute timeslice_duration_attr =
+__ATTR(timeslice_duration_ms, 0644, timeslice_show, timeslice_store);
+
static void kobj_engine_release(struct kobject *kobj)
{
kfree(kobj);
@@ -206,6 +248,10 @@ void intel_engines_add_sysfs(struct drm_i915_private *i915)
if (sysfs_create_files(kobj, files))
goto err_object;
+ if (intel_engine_has_timeslices(engine) &&
+ sysfs_create_file(kobj, ×lice_duration_attr.attr))
+ goto err_engine;
+
if (0) {
err_object:
kobject_put(kobj);
@@ -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;
};
@@ -1419,7 +1419,7 @@ 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))
@@ -1440,15 +1440,32 @@ 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)
@@ -1619,8 +1636,9 @@ 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;
}
@@ -2044,10 +2062,7 @@ 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 {
@@ -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 (!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 (!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);
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. The timeslicing mechanism can be compiled out with ./scripts/config --set-val DRM_I915_TIMESLICE_DURATION 0 The timeslice duration can be adjusted per-engine using, /sys/class/drm/card?/engine/*/timeslice_duration_ms Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> --- drivers/gpu/drm/i915/Kconfig.profile | 18 ++++++++ drivers/gpu/drm/i915/gt/intel_engine.h | 9 ++++ drivers/gpu/drm/i915/gt/intel_engine_cs.c | 2 + drivers/gpu/drm/i915/gt/intel_engine_sysfs.c | 46 ++++++++++++++++++++ drivers/gpu/drm/i915/gt/intel_engine_types.h | 1 + drivers/gpu/drm/i915/gt/intel_lrc.c | 39 ++++++++++++----- drivers/gpu/drm/i915/gt/selftest_lrc.c | 13 +++++- 7 files changed, 114 insertions(+), 14 deletions(-)