sched/deadline: Always calculate end of period on sched_yield()
authorPeter Zijlstra <peterz@infradead.org>
Tue, 23 Feb 2016 12:28:22 +0000 (13:28 +0100)
committerIngo Molnar <mingo@kernel.org>
Mon, 29 Feb 2016 08:41:51 +0000 (09:41 +0100)
Steven noticed that occasionally a sched_yield() call would not result
in a wait for the next period edge as expected.

It turns out that when we call update_curr_dl() and end up with
delta_exec <= 0, we will bail early and fail to throttle.

Further inspection of the yield code revealed that yield_task_dl()
clearing dl.runtime is wrong too, it will not account the last bit of
runtime which could result in dl.runtime < 0, which in turn means that
replenish would gift us with too much runtime.

Fix both issues by not relying on the dl.runtime value for yield.

Reported-by: Steven Rostedt <rostedt@goodmis.org>
Tested-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Clark Williams <williams@redhat.com>
Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Juri Lelli <juri.lelli@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20160223122822.GP6357@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/deadline.c

index 57b939c81bce5d06fa587df8915f05affbe22b82..04a569cdd613df57a9bd84d86ce674f08068bde2 100644 (file)
@@ -399,6 +399,9 @@ static void replenish_dl_entity(struct sched_dl_entity *dl_se,
                dl_se->runtime = pi_se->dl_runtime;
        }
 
+       if (dl_se->dl_yielded && dl_se->runtime > 0)
+               dl_se->runtime = 0;
+
        /*
         * We keep moving the deadline away until we get some
         * available runtime for the entity. This ensures correct
@@ -735,8 +738,11 @@ static void update_curr_dl(struct rq *rq)
         * approach need further study.
         */
        delta_exec = rq_clock_task(rq) - curr->se.exec_start;
-       if (unlikely((s64)delta_exec <= 0))
+       if (unlikely((s64)delta_exec <= 0)) {
+               if (unlikely(dl_se->dl_yielded))
+                       goto throttle;
                return;
+       }
 
        schedstat_set(curr->se.statistics.exec_max,
                      max(curr->se.statistics.exec_max, delta_exec));
@@ -749,8 +755,10 @@ static void update_curr_dl(struct rq *rq)
 
        sched_rt_avg_update(rq, delta_exec);
 
-       dl_se->runtime -= dl_se->dl_yielded ? 0 : delta_exec;
-       if (dl_runtime_exceeded(dl_se)) {
+       dl_se->runtime -= delta_exec;
+
+throttle:
+       if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
                dl_se->dl_throttled = 1;
                __dequeue_task_dl(rq, curr, 0);
                if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
@@ -994,18 +1002,14 @@ static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
  */
 static void yield_task_dl(struct rq *rq)
 {
-       struct task_struct *p = rq->curr;
-
        /*
         * We make the task go to sleep until its current deadline by
         * forcing its runtime to zero. This way, update_curr_dl() stops
         * it and the bandwidth timer will wake it up and will give it
         * new scheduling parameters (thanks to dl_yielded=1).
         */
-       if (p->dl.runtime > 0) {
-               rq->curr->dl.dl_yielded = 1;
-               p->dl.runtime = 0;
-       }
+       rq->curr->dl.dl_yielded = 1;
+
        update_rq_clock(rq);
        update_curr_dl(rq);
        /*