Re: [PATCH v2] Track hard and soft "short lockups" or "stalls."

[Don Zickus <dzickus@redhat.com>]

On Mon, Jul 18, 2011 at 02:45:55PM -0700, ZAK Magnus wrote:
> Okay, great. I'm eager to hear anything you may discover, good or bad. By
> the way, would you mind sharing a bit about how you do your testing for
> this?

Sorry for getting back to you late, busy week.

Most of the testing I do is from the lkdtm module

modprobe lkdtm
mount -t debugfs none /sys/kernel/debug
cd /sys/kernel/debug/provoke-crashing/
service cpuspeed stop
echo HARDLOCKUP > DIRECT #or SOFTLOCKUP or HUNG_TASK

I then count to 10 seconds to make sure the timer is within reason.

So I did the above test and noticed the panic looked funny because it spit
out the 

new worst hard stall seen on CPU#0: 3 interrupts missed

and then 

new worst hard stall seen on CPU#0: 4 interrupts missed

and then finally the HARDLOCKUP message

I am not sure that is what we want as it confuses people as to where the
panic really is.

What if you moved the 'update_hardstall()' to just underneath the zero'ing
out of the hrtimer_interrupts_missed?  This only then prints out the
interrupts missed line when you know the end point.  And avoids printing
it all together in the case of a true HARDLOCKUP.  Like the patch below

diff --git a/kernel/watchdog.c b/kernel/watchdog.c
index 7d37cc2..ba41a74 100644
--- a/kernel/watchdog.c
+++ b/kernel/watchdog.c
@@ -238,13 +238,14 @@ static int is_hardlockup(int this_cpu)
 
 	if (hrint_saved == hrint)
 		ints_missed = per_cpu(hrtimer_interrupts_missed, this_cpu)++;
-	else
+	else {
 		__this_cpu_write(hrtimer_interrupts_missed, 0);
+		update_hardstall(ints_missed, this_cpu);
+	}
 
 	if (ints_missed >= hardlockup_thresh)
 		return 1;
 
-	update_hardstall(ints_missed, this_cpu);
 	return 0;
 }
 #endif

The softlockup case probably needs the same.

Thoughts?

Cheers,
Don