sched/deadline: Fix stale yield state
authorPeter Zijlstra <peterz@infradead.org>
Wed, 26 Nov 2014 00:44:03 +0000 (08:44 +0800)
committerIngo Molnar <mingo@kernel.org>
Wed, 4 Feb 2015 06:52:26 +0000 (07:52 +0100)
When we fail to start the deadline timer in update_curr_dl(), we
forget to clear ->dl_yielded, resulting in wrecked time keeping.

Since the natural place to clear both ->dl_yielded and ->dl_throttled
is in replenish_dl_entity(); both are after all waiting for that event;
make it so.

Luckily since 67dfa1b756f2 ("sched/deadline: Implement
cancel_dl_timer() to use in switched_from_dl()") the
task_on_rq_queued() condition in dl_task_timer() must be true, and can
therefore call enqueue_task_dl() unconditionally.

Reported-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Kirill Tkhai <ktkhai@parallels.com>
Cc: Juri Lelli <juri.lelli@arm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1416962647-76792-4-git-send-email-wanpeng.li@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/deadline.c

index 7b684f9..a027799 100644 (file)
@@ -350,6 +350,11 @@ static void replenish_dl_entity(struct sched_dl_entity *dl_se,
                dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
                dl_se->runtime = pi_se->dl_runtime;
        }
+
+       if (dl_se->dl_yielded)
+               dl_se->dl_yielded = 0;
+       if (dl_se->dl_throttled)
+               dl_se->dl_throttled = 0;
 }
 
 /*
@@ -536,23 +541,19 @@ again:
 
        sched_clock_tick();
        update_rq_clock(rq);
-       dl_se->dl_throttled = 0;
-       dl_se->dl_yielded = 0;
-       if (task_on_rq_queued(p)) {
-               enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
-               if (dl_task(rq->curr))
-                       check_preempt_curr_dl(rq, p, 0);
-               else
-                       resched_curr(rq);
+       enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
+       if (dl_task(rq->curr))
+               check_preempt_curr_dl(rq, p, 0);
+       else
+               resched_curr(rq);
 #ifdef CONFIG_SMP
-               /*
-                * Queueing this task back might have overloaded rq,
-                * check if we need to kick someone away.
-                */
-               if (has_pushable_dl_tasks(rq))
-                       push_dl_task(rq);
+       /*
+        * Queueing this task back might have overloaded rq,
+        * check if we need to kick someone away.
+        */
+       if (has_pushable_dl_tasks(rq))
+               push_dl_task(rq);
 #endif
-       }
 unlock:
        raw_spin_unlock(&rq->lock);
 
@@ -613,10 +614,9 @@ static void update_curr_dl(struct rq *rq)
 
        dl_se->runtime -= dl_se->dl_yielded ? 0 : delta_exec;
        if (dl_runtime_exceeded(rq, dl_se)) {
+               dl_se->dl_throttled = 1;
                __dequeue_task_dl(rq, curr, 0);
-               if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted)))
-                       dl_se->dl_throttled = 1;
-               else
+               if (unlikely(!start_dl_timer(dl_se, curr->dl.dl_boosted)))
                        enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
 
                if (!is_leftmost(curr, &rq->dl))
@@ -853,7 +853,7 @@ static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
         * its rq, the bandwidth timer callback (which clearly has not
         * run yet) will take care of this.
         */
-       if (p->dl.dl_throttled)
+       if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH))
                return;
 
        enqueue_dl_entity(&p->dl, pi_se, flags);