Knowing the round trip time of an engine is useful for tracking the
health of the system as well as providing a metric for the baseline
responsiveness of the engine. We can use the latter metric for
automatically tuning our waits in selftests and when idling so we don't
confuse a slower system with a dead one.
Upon idling the engine, we send one last pulse to switch the context
away from precious user state to the volatile kernel context. We know
the engine is idle at this point, and the pulse is non-preemptible, so
this provides us with a good measurement of the round trip time. It also
provides us with faster engine parking for ringbuffer submission, which
is a welcome bonus (e.g. softer-rc6).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Stuart Summers <stuart.summers@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191219105043.4169050-1-chris@chris-wilson.co.uk
Link: https://patchwork.freedesktop.org/patch/msgid/20191219124353.8607-2-chris@chris-wilson.co.uk
/* Nothing to do here, execute in order of dependencies */
engine->schedule = NULL;
+ ewma__engine_latency_init(&engine->latency);
seqlock_init(&engine->stats.lock);
ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
drm_printf(m, "\tBarriers?: %s\n",
yesno(!llist_empty(&engine->barrier_tasks)));
+ drm_printf(m, "\tLatency: %luus\n",
+ ewma__engine_latency_read(&engine->latency));
rcu_read_lock();
rq = READ_ONCE(engine->heartbeat.systole);
#endif /* !IS_ENABLED(CONFIG_LOCKDEP) */
+static void duration(struct dma_fence *fence, struct dma_fence_cb *cb)
+{
+ struct i915_request *rq = to_request(fence);
+
+ ewma__engine_latency_add(&rq->engine->latency,
+ ktime_us_delta(rq->fence.timestamp,
+ rq->duration.emitted));
+}
+
static void
__queue_and_release_pm(struct i915_request *rq,
struct intel_timeline *tl,
/* Install ourselves as a preemption barrier */
rq->sched.attr.priority = I915_PRIORITY_BARRIER;
- __i915_request_commit(rq);
+ if (likely(!__i915_request_commit(rq))) { /* engine should be idle! */
+ /*
+ * Use an interrupt for precise measurement of duration,
+ * otherwise we rely on someone else retiring all the requests
+ * which may delay the signaling (i.e. we will likely wait
+ * until the background request retirement running every
+ * second or two).
+ */
+ BUILD_BUG_ON(sizeof(rq->duration) > sizeof(rq->submitq));
+ dma_fence_add_callback(&rq->fence, &rq->duration.cb, duration);
+ rq->duration.emitted = ktime_get();
+ }
/* Expose ourselves to the world */
__queue_and_release_pm(rq, ce->timeline, engine);
#ifndef __INTEL_ENGINE_TYPES__
#define __INTEL_ENGINE_TYPES__
+#include <linux/average.h>
#include <linux/hashtable.h>
#include <linux/irq_work.h>
#include <linux/kref.h>
#define INVALID_ENGINE ((enum intel_engine_id)-1)
};
+/* A simple estimator for the round-trip latency of an engine */
+DECLARE_EWMA(_engine_latency, 6, 4)
+
struct st_preempt_hang {
struct completion completion;
unsigned int count;
struct intel_timeline *timeline;
} legacy;
+ /*
+ * We track the average duration of the idle pulse on parking the
+ * engine to keep an estimate of the how the fast the engine is
+ * under ideal conditions.
+ */
+ struct ewma__engine_latency latency;
+
/* Rather than have every client wait upon all user interrupts,
* with the herd waking after every interrupt and each doing the
* heavyweight seqno dance, we delegate the task (of being the
union {
wait_queue_entry_t submitq;
struct i915_sw_dma_fence_cb dmaq;
+ struct i915_request_duration_cb {
+ struct dma_fence_cb cb;
+ ktime_t emitted;
+ } duration;
};
struct list_head execute_cb;
struct i915_sw_fence semaphore;