drm/i915: Only report a wakeup if the waiter was truly asleep
authorChris Wilson <chris@chris-wilson.co.uk>
Sat, 9 Dec 2017 12:47:10 +0000 (12:47 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Mon, 11 Dec 2017 17:23:02 +0000 (17:23 +0000)
If we attempt to wake up a waiter, who is currently checking the seqno
it will be in the TASK_INTERRUPTIBLE state and ttwu will report success.
However, it is actually awake and functioning -- so delay reporting the
actual wake up until it sleeps. This fixes some spurious claims of
missed_breadcrumbs when running under heavy load; i.e. sufficient load to
preempt away the newly woken waiter before they complete their checks.
However, it does so at the cost of a rare false negative; where the
waiter changes between the check and ttwu -- the only way to fix that
would be to extend the reporting from ttwu where the check could be done
atomically.

v2: Defend against !CONFIG_SMP
v3: Don't filter out calls to wake_up_process
v4: Drop risky microoptimisation to skip wakeups

Testcase: igt/drv_missed_irq # sanity check we do detect missed_breadcrumb()
Testcase: igt/gem_concurrent_blit # for generating false positives
References: https://bugs.freedesktop.org/show_bug.cgi?id=100007
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171209124710.1606-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
drivers/gpu/drm/i915/intel_breadcrumbs.c

index a7740696114dfbb225ff15f8ff6ee2a72e70b250..58c624f982d91c1383b8d51adbf7efea76eeef54 100644 (file)
 
 #include "i915_drv.h"
 
+#ifdef CONFIG_SMP
+#define task_asleep(tsk) ((tsk)->state & TASK_NORMAL && !(tsk)->on_cpu)
+#else
+#define task_asleep(tsk) ((tsk)->state & TASK_NORMAL)
+#endif
+
 static unsigned int __intel_breadcrumbs_wakeup(struct intel_breadcrumbs *b)
 {
        struct intel_wait *wait;
@@ -36,8 +42,20 @@ static unsigned int __intel_breadcrumbs_wakeup(struct intel_breadcrumbs *b)
 
        wait = b->irq_wait;
        if (wait) {
+               /*
+                * N.B. Since task_asleep() and ttwu are not atomic, the
+                * waiter may actually go to sleep after the check, causing
+                * us to suppress a valid wakeup. We prefer to reduce the
+                * number of false positive missed_breadcrumb() warnings
+                * at the expense of a few false negatives, as it it easy
+                * to trigger a false positive under heavy load. Enough
+                * signal should remain from genuine missed_breadcrumb()
+                * for us to detect in CI.
+                */
+               bool was_asleep = task_asleep(wait->tsk);
+
                result = ENGINE_WAKEUP_WAITER;
-               if (wake_up_process(wait->tsk))
+               if (wake_up_process(wait->tsk) && was_asleep)
                        result |= ENGINE_WAKEUP_ASLEEP;
        }
 
@@ -77,8 +95,8 @@ static noinline void missed_breadcrumb(struct intel_engine_cs *engine)
 
 static void intel_breadcrumbs_hangcheck(struct timer_list *t)
 {
-       struct intel_engine_cs *engine = from_timer(engine, t,
-                                                   breadcrumbs.hangcheck);
+       struct intel_engine_cs *engine =
+               from_timer(engine, t, breadcrumbs.hangcheck);
        struct intel_breadcrumbs *b = &engine->breadcrumbs;
 
        if (!b->irq_armed)
@@ -104,7 +122,7 @@ static void intel_breadcrumbs_hangcheck(struct timer_list *t)
         */
        if (intel_engine_wakeup(engine) & ENGINE_WAKEUP_ASLEEP) {
                missed_breadcrumb(engine);
-               mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
+               mod_timer(&b->fake_irq, jiffies + 1);
        } else {
                mod_timer(&b->hangcheck, wait_timeout());
        }