Re: [PATCH RFC] clocksource: Detect a watchdog overflow

[John Stultz]

On Thu, Apr 7, 2016 at 1:14 AM, Gratian Crisan <gratian.crisan@xxxxxx> wrote:
> John Stultz writes:
>> So I'm sympathetic to this issue, because I remember seeing similar
>> problems w/ runaway SCHED_FIFO tasks w/ PREEMPT_RT.
>
> Yeah, a runaway rt thread can easily do it. That's just bad design. In
> our case it was a bit more subtle bc. it was a combination of high
> priority interrupts and rt threads that would occasionally stack up to
> delay the timer softirq long enough to cause the watchdog wrap.

So in the last discussion, I believe Thomas and others were skeptical
because we really shouldn't be blocking tasks from running for such a
long time. Instead of trying to turn off the watchdog, instead they
were suggesting we ensure we don't get into such a state where things
are delayed so unexpectedly long.


>> However, its really difficult to create a solution without opening new
>> cases where bad clocksources will be mis-identified as good (which
>> your solution seems to suffer as well, measuring the time past with a
>> known bad clocksource can easily result in large deltas, which will be
>> ignored if the watchdog has a short interval).
>
> Fair point. Ultimately you have to trust one of the clocksources. I
> guess I was naive in thinking that the main clocksource can't drift more
> than what the watchdog clocksource can measure within the
> WATCHDOG_INTERVAL. I'm glad I don't have to deal with hardware that
> lobotomized.

Another thought might be to try to add a third longer-running clock
into the mix. Possibly a very rough fallback check against something
like the RTC to see if the interval was really long enough to have the
watchdog wrap.

> Would a simple solution that exposes the config option for the
> clocksource wathchdog[1] (and defaults it to on) be an acceptable
> alternative? It will work for us because we test the stability of the
> main clocksource - part of the hardware bring-up.

So there is already the tsc=reliable boot option, which I believe
disables the watchdog. So I'm not sure the build time option makes the
most sense.

>> A previous effort on this was made here, and there's a resulting
>> thread that didn't come to resolution:
>>     https://lkml.org/lkml/2015/8/17/542
>
> Sorry I've missed it.
>
>> Way back I had tried to come up with an approach where if the time
>> delta was large, it was divided by the watchdog interval, and then we
>> just compared the remainder with the current watchdog delta to see if
>> they matched (closely enough). Unfortunately this didn't work out for
>> me then, but perhaps it deserves a second try?
>
> I've entertained that idea too but I think I was trying to optimize
> things too early and do everything with the mult/shift math. That first
> attempt failed but I do need to try harder because it would be a better
> general solution.

Yea. I'd much prefer a general solution.

thanks
-john