[BUG] Caused by: sched/deadline: Move CPU frequency selection triggering points

From: Steven Rostedt
Date: Tue Jul 03 2018 - 10:54:56 EST


When looking to test SCHED_DEADLINE, I triggered a lockup. The lockup
appears to be caused by WARN_ON() done inside the scheduling path, and
I'm guessing it tried to grab the rq lock and caused a deadlock (all I
would get would be the "--- cut here ---" line, and then nothing after
that. But a bit of playing with the printks() I figured out it was
happening at:

assert_clock_updated() {
SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
}

I bisected it down to commit e0367b126 ("sched/deadline: Move CPU
frequency selection triggering points"). Reverting it indeed makes the
deadlock go away.

I commented out the WARN_ON that was being triggered, and that let the
system continue to run. I did the following change:

--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -914,7 +914,12 @@ static inline void assert_clock_updated(struct rq *rq)
* The only reason for not seeing a clock update since the
* last rq_pin_lock() is if we're currently skipping updates.
*/
- SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
+ if (rq->clock_update_flags < RQCF_ACT_SKIP)
+ trace_printk("WARN_ON: [%d] rq->clock_update_flags (%d) < %d)\n",
+ rq->cpu, rq->clock_update_flags, RQCF_ACT_SKIP);
+ else
+ trace_printk("GOOD: [%d] rq->clock_update_flags (%d) >= %d)\n",
+ rq->cpu, rq->clock_update_flags, RQCF_ACT_SKIP);
}

To see what was happening. I also, added trace_printk()s to all the
updates to clock_update_flags, and ran my test again.

Here's what I got:

deadline_test-1393 [002] 162.127132: bprint: push_dl_task.part.40: WARN_ON: [1] rq->clock_update_flags (0) < 2)
deadline_test-1393 [002] 162.127133: bprint: update_rq_clock: [1] clock was 0 now 4
deadline_test-1393 [002] 162.127134: bprint: rq_clock: GOOD: [1] rq->clock_update_flags (4) >= 2)

It appears that we hit this in this path:

push_dl_task {
add_running_bw() {
__add_running_bw() {
cpufreq_update_util() {
data->func(data, rq_clock(rq), flags);
rq_clock() {
assert_clock_updated()

And here the clock isn't updated and we get the splat.

Reverting the stated patch works because it added the call to
cpufreq_update_util() that does the rq_clock() at an inappropriate
time, which causes the splat.

I'm not sure what the right answer to this is. Reverting obviously
works, but I'm also guessing a proper placement of update_rq_clock()
may also work. I just don't know where that placement is, as I don't
understand the rq_clock() updates enough.

Help?

-- Steve