drm/i915/execlists: Direct submission of new requests (avoid tasklet/ksoftirqd)
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 28 Jun 2018 20:12:11 +0000 (21:12 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 28 Jun 2018 21:55:10 +0000 (22:55 +0100)
Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a
bottom half"), we came to the conclusion that running our CSB processing
and ELSP submission from inside the irq handler was a bad idea. A really
bad idea as we could impose nearly 1s latency on other users of the
system, on average! Deferring our work to a tasklet allowed us to do the
processing with irqs enabled, reducing the impact to an average of about
50us.

We have since eradicated the use of forcewaked mmio from inside the CSB
processing and ELSP submission, bringing the impact down to around 5us
(on Kabylake); an order of magnitude better than our measurements 2
years ago on Broadwell and only about 2x worse on average than the
gem_syslatency on an unladen system.

In this iteration of the tasklet-vs-direct submission debate, we seek a
compromise where by we submit new requests immediately to the HW but
defer processing the CS interrupt onto a tasklet. We gain the advantage
of low-latency and ksoftirqd avoidance when waking up the HW, while
avoiding the system-wide starvation of our CS irq-storms.

Comparing the impact on the maximum latency observed (that is the time
stolen from an RT process) over a 120s interval, repeated several times
(using gem_syslatency, similar to RT's cyclictest) while the system is
fully laden with i915 nops, we see that direct submission an actually
improve the worse case.

Maximum latency in microseconds of a third party RT thread
(gem_syslatency -t 120 -f 2)
  x Always using tasklets (a couple of >1000us outliers removed)
  + Only using tasklets from CS irq, direct submission of requests
+------------------------------------------------------------------------+
|          +                                                             |
|          +                                                             |
|          +                                                             |
|          +       +                                                     |
|          + +     +                                                     |
|       +  + +     +  x     x     x                                      |
|      +++ + +     +  x  x  x  x  x  x                                   |
|      +++ + ++  + +  *x x  x  x  x  x                                   |
|      +++ + ++  + *  *x x  *  x  x  x                                   |
|    + +++ + ++  * * +*xxx  *  x  x  xx                                  |
|    * +++ + ++++* *x+**xx+ *  x  x xxxx x                               |
|   **x++++*++**+*x*x****x+ * +x xx xxxx x          x                    |
|x* ******+***************++*+***xxxxxx* xx*x     xxx +                x+|
|             |__________MA___________|                                  |
|      |______M__A________|                                              |
+------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x 118            91           186           124     125.28814     16.279137
+ 120            92           187           109     112.00833     13.458617
Difference at 95.0% confidence
-13.2798 +/- 3.79219
-10.5994% +/- 3.02677%
(Student's t, pooled s = 14.9237)

However the mean latency is adversely affected:

Mean latency in microseconds of a third party RT thread
(gem_syslatency -t 120 -f 1)
  x Always using tasklets
  + Only using tasklets from CS irq, direct submission of requests
+------------------------------------------------------------------------+
|           xxxxxx                                        +   ++         |
|           xxxxxx                                        +   ++         |
|           xxxxxx                                      + +++ ++         |
|           xxxxxxx                                     +++++ ++         |
|           xxxxxxx                                     +++++ ++         |
|           xxxxxxx                                     +++++ +++        |
|           xxxxxxx                                   + ++++++++++       |
|           xxxxxxxx                                 ++ ++++++++++       |
|           xxxxxxxx                                 ++ ++++++++++       |
|          xxxxxxxxxx                                +++++++++++++++     |
|         xxxxxxxxxxx    x                           +++++++++++++++     |
|x       xxxxxxxxxxxxx   x           +            + ++++++++++++++++++  +|
|           |__A__|                                                      |
|                                                      |____A___|        |
+------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x 120         3.506         3.727         3.631     3.6321417    0.02773109
+ 120         3.834         4.149         4.039     4.0375167   0.041221676
Difference at 95.0% confidence
0.405375 +/- 0.00888913
11.1608% +/- 0.244735%
(Student's t, pooled s = 0.03513)

However, since the mean latency corresponds to the amount of irqsoff
processing we have to do for a CS interrupt, we only need to speed that
up to benefit not just system latency but our own throughput.

v2: Remember to defer submissions when under reset.
v4: Only use direct submission for new requests
v5: Be aware that with mixing direct tasklet evaluation and deferred
tasklets, we may end up idling before running the deferred tasklet.
v6: Remove the redudant likely() from tasklet_is_enabled(), restrict the
annotation to reset_in_progress().
v7: Take the full timeline.lock when enabling perf_pmu stats as the
tasklet is no longer a valid guard. A consequence is that the stats are
now only valid for engines also using the timeline.lock to process
state.

Testcase: igt/gem_exec_latency/*rthog*
References: 27af5eea54d1 ("drm/i915: Move execlists irq handler to a bottom half")
Suggested-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180628201211.13837-9-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem.h
drivers/gpu/drm/i915/intel_engine_cs.c
drivers/gpu/drm/i915/intel_lrc.c

index 261da577829a61b1c84ae8ab2d91bc6d307a6e0d..e465929568726c31a39dfb6037da594c1a93f3ac 100644 (file)
@@ -88,4 +88,9 @@ static inline void __tasklet_enable_sync_once(struct tasklet_struct *t)
                tasklet_kill(t);
 }
 
+static inline bool __tasklet_is_enabled(const struct tasklet_struct *t)
+{
+       return !atomic_read(&t->count);
+}
+
 #endif /* __I915_GEM_H__ */
index ace93958689ef78528e0df8de8d0da47c02fd096..01862884a4360f6bedbb211c448fa2cce454e943 100644 (file)
@@ -1619,8 +1619,8 @@ int intel_enable_engine_stats(struct intel_engine_cs *engine)
        if (!intel_engine_supports_stats(engine))
                return -ENODEV;
 
-       tasklet_disable(&execlists->tasklet);
-       write_seqlock_irqsave(&engine->stats.lock, flags);
+       spin_lock_irqsave(&engine->timeline.lock, flags);
+       write_seqlock(&engine->stats.lock);
 
        if (unlikely(engine->stats.enabled == ~0)) {
                err = -EBUSY;
@@ -1644,8 +1644,8 @@ int intel_enable_engine_stats(struct intel_engine_cs *engine)
        }
 
 unlock:
-       write_sequnlock_irqrestore(&engine->stats.lock, flags);
-       tasklet_enable(&execlists->tasklet);
+       write_sequnlock(&engine->stats.lock);
+       spin_unlock_irqrestore(&engine->timeline.lock, flags);
 
        return err;
 }
index d835da128a172e5f8fb80d6e85081b5555d6b638..6ab6ddb103d128fb2bed3fb2df76e6b249748af4 100644 (file)
@@ -563,12 +563,14 @@ static void complete_preempt_context(struct intel_engine_execlists *execlists)
        GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT));
 
        execlists_cancel_port_requests(execlists);
-       execlists_unwind_incomplete_requests(execlists);
+       __unwind_incomplete_requests(container_of(execlists,
+                                                 struct intel_engine_cs,
+                                                 execlists));
 
        execlists_clear_active(execlists, EXECLISTS_ACTIVE_PREEMPT);
 }
 
-static void __execlists_dequeue(struct intel_engine_cs *engine)
+static void execlists_dequeue(struct intel_engine_cs *engine)
 {
        struct intel_engine_execlists * const execlists = &engine->execlists;
        struct execlist_port *port = execlists->port;
@@ -578,9 +580,8 @@ static void __execlists_dequeue(struct intel_engine_cs *engine)
        struct rb_node *rb;
        bool submit = false;
 
-       lockdep_assert_held(&engine->timeline.lock);
-
-       /* Hardware submission is through 2 ports. Conceptually each port
+       /*
+        * Hardware submission is through 2 ports. Conceptually each port
         * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is
         * static for a context, and unique to each, so we only execute
         * requests belonging to a single context from each ring. RING_HEAD
@@ -770,15 +771,6 @@ done:
                   !port_isset(engine->execlists.port));
 }
 
-static void execlists_dequeue(struct intel_engine_cs *engine)
-{
-       unsigned long flags;
-
-       spin_lock_irqsave(&engine->timeline.lock, flags);
-       __execlists_dequeue(engine);
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
-}
-
 void
 execlists_cancel_port_requests(struct intel_engine_execlists * const execlists)
 {
@@ -958,6 +950,12 @@ static void execlists_cancel_requests(struct intel_engine_cs *engine)
        spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }
 
+static inline bool
+reset_in_progress(const struct intel_engine_execlists *execlists)
+{
+       return unlikely(!__tasklet_is_enabled(&execlists->tasklet));
+}
+
 static void process_csb(struct intel_engine_cs *engine)
 {
        struct intel_engine_execlists * const execlists = &engine->execlists;
@@ -1109,18 +1107,9 @@ static void process_csb(struct intel_engine_cs *engine)
        execlists->csb_head = head;
 }
 
-/*
- * Check the unread Context Status Buffers and manage the submission of new
- * contexts to the ELSP accordingly.
- */
-static void execlists_submission_tasklet(unsigned long data)
+static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
 {
-       struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
-
-       GEM_TRACE("%s awake?=%d, active=%x\n",
-                 engine->name,
-                 engine->i915->gt.awake,
-                 engine->execlists.active);
+       lockdep_assert_held(&engine->timeline.lock);
 
        /*
         * We can skip acquiring intel_runtime_pm_get() here as it was taken
@@ -1137,6 +1126,28 @@ static void execlists_submission_tasklet(unsigned long data)
                execlists_dequeue(engine);
 }
 
+/*
+ * Check the unread Context Status Buffers and manage the submission of new
+ * contexts to the ELSP accordingly.
+ */
+static void execlists_submission_tasklet(unsigned long data)
+{
+       struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
+       unsigned long flags;
+
+       GEM_TRACE("%s awake?=%d, active=%x\n",
+                 engine->name,
+                 engine->i915->gt.awake,
+                 engine->execlists.active);
+
+       spin_lock_irqsave(&engine->timeline.lock, flags);
+
+       if (engine->i915->gt.awake) /* we may be delayed until after we idle! */
+               __execlists_submission_tasklet(engine);
+
+       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+}
+
 static void queue_request(struct intel_engine_cs *engine,
                          struct i915_sched_node *node,
                          int prio)
@@ -1145,16 +1156,30 @@ static void queue_request(struct intel_engine_cs *engine,
                      &lookup_priolist(engine, prio)->requests);
 }
 
-static void __submit_queue(struct intel_engine_cs *engine, int prio)
+static void __update_queue(struct intel_engine_cs *engine, int prio)
 {
        engine->execlists.queue_priority = prio;
-       tasklet_hi_schedule(&engine->execlists.tasklet);
+}
+
+static void __submit_queue_imm(struct intel_engine_cs *engine)
+{
+       struct intel_engine_execlists * const execlists = &engine->execlists;
+
+       if (reset_in_progress(execlists))
+               return; /* defer until we restart the engine following reset */
+
+       if (execlists->tasklet.func == execlists_submission_tasklet)
+               __execlists_submission_tasklet(engine);
+       else
+               tasklet_hi_schedule(&execlists->tasklet);
 }
 
 static void submit_queue(struct intel_engine_cs *engine, int prio)
 {
-       if (prio > engine->execlists.queue_priority)
-               __submit_queue(engine, prio);
+       if (prio > engine->execlists.queue_priority) {
+               __update_queue(engine, prio);
+               __submit_queue_imm(engine);
+       }
 }
 
 static void execlists_submit_request(struct i915_request *request)
@@ -1166,11 +1191,12 @@ static void execlists_submit_request(struct i915_request *request)
        spin_lock_irqsave(&engine->timeline.lock, flags);
 
        queue_request(engine, &request->sched, rq_prio(request));
-       submit_queue(engine, rq_prio(request));
 
        GEM_BUG_ON(!engine->execlists.first);
        GEM_BUG_ON(list_empty(&request->sched.link));
 
+       submit_queue(engine, rq_prio(request));
+
        spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }
 
@@ -1297,8 +1323,11 @@ static void execlists_schedule(struct i915_request *request,
                }
 
                if (prio > engine->execlists.queue_priority &&
-                   i915_sw_fence_done(&sched_to_request(node)->submit))
-                       __submit_queue(engine, prio);
+                   i915_sw_fence_done(&sched_to_request(node)->submit)) {
+                       /* defer submission until after all of our updates */
+                       __update_queue(engine, prio);
+                       tasklet_hi_schedule(&engine->execlists.tasklet);
+               }
        }
 
        spin_unlock_irq(&engine->timeline.lock);
@@ -1879,6 +1908,7 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
 {
        struct intel_engine_execlists * const execlists = &engine->execlists;
        struct i915_request *request, *active;
+       unsigned long flags;
 
        GEM_TRACE("%s\n", engine->name);
 
@@ -1893,6 +1923,8 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
         */
        __tasklet_disable_sync_once(&execlists->tasklet);
 
+       spin_lock_irqsave(&engine->timeline.lock, flags);
+
        /*
         * We want to flush the pending context switches, having disabled
         * the tasklet above, we can assume exclusive access to the execlists.
@@ -1910,15 +1942,12 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
        active = NULL;
        request = port_request(execlists->port);
        if (request) {
-               unsigned long flags;
-
                /*
                 * Prevent the breadcrumb from advancing before we decide
                 * which request is currently active.
                 */
                intel_engine_stop_cs(engine);
 
-               spin_lock_irqsave(&engine->timeline.lock, flags);
                list_for_each_entry_from_reverse(request,
                                                 &engine->timeline.requests,
                                                 link) {
@@ -1928,9 +1957,10 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
 
                        active = request;
                }
-               spin_unlock_irqrestore(&engine->timeline.lock, flags);
        }
 
+       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+
        return active;
 }