Re: [PATCH v3 2/2] Make hard lockup detection use timestamps

[Don Zickus <dzickus@redhat.com>]

On Fri, Jul 29, 2011 at 04:12:32PM -0700, ZAK Magnus wrote:
> Are you saying that any call to printk() will touch the watchdogs? I
> wasn't aware of that and it doesn't seem to comply with my
> observations too well, either. Then again, at the moment I don't
> understand some of the things I'm currently seeing so I could just be
> wrong.

I believe the serial console write is the source of all the
touch_nmi_watchdogs that result from a printk.

drivers/tty/serial/8250.c::serial8250_console_write()

Cheers,
Don

> 
> On Fri, Jul 29, 2011 at 1:55 PM, Don Zickus <dzickus@redhat.com> wrote:
> > On Thu, Jul 28, 2011 at 05:16:00PM -0700, ZAK Magnus wrote:
> >> No news?
> >>
> >> I've been testing and looking into issues and I realized dump_stack()
> >> calls touch_nmi_watchdog(). That wrecks what the patch is trying to do
> >> so I'm changing it to save the trace and print it later after the
> >> stall has completed. This would also resolve some other things you
> >> were saying weren't so good. Hopefully the logic is similar enough
> >> that some things you may have learned still apply.
> >
> > So yeah, the acting of printing was resesting the softlockup counter and
> > delaying it forever.  In parallel, rcu has its own stall detector that was
> > going off after a minute or two.
> >
> > Once I routed the printk to trace_printk and disabled dump_stack,
> > everything started working as expected.
> >
> > Now the question is how to avoid shooting ourselves in the foot by
> > printk'ing a message without resetting the hard/soft lock watchdogs.
> >
> > I'll have to think about how to do that.  If you can come up with any
> > ideas let me know.
> >
> > We almost need a quiet dump_stack that dumps to a buffer instead of the
> > console.  But I am not sure that is worth the effort.
> >
> > Hmm.
> >
> > Cheers,
> > Don
> >