Re: [PATCH 4/5] watchdog: control hard lockup detection default

[Ulrich Obergfell]

>----- Original Message -----
>From: "Ingo Molnar" <mingo@xxxxxxxxxx>
>To: "Don Zickus" <dzickus@xxxxxxxxxx>
>Cc: akpm@xxxxxxxxxxxxxxxxxxxx, kvm@xxxxxxxxxxxxxxx, pbonzini@xxxxxxxxxx, mingo@xxxxxxxxxx, "LKML" <linux-kernel@xxxxxxxxxxxxxxx>, "Ulrich >Obergfell" <uobergfe@xxxxxxxxxx>, "Andrew Jones" <drjones@xxxxxxxxxx>
>Sent: Monday, August 18, 2014 11:16:44 AM
>Subject: Re: [PATCH 4/5] watchdog: control hard lockup detection default
>
>
> * Don Zickus <dzickus@xxxxxxxxxx> wrote:
>
>> The running kernel still has the ability to enable/disable at any
>> time with /proc/sys/kernel/nmi_watchdog us usual. However even
>> when the default has been overridden /proc/sys/kernel/nmi_watchdog
>> will initially show '1'. To truly turn it on one must disable/enable
>> it, i.e.
>>   echo 0 > /proc/sys/kernel/nmi_watchdog
>>   echo 1 > /proc/sys/kernel/nmi_watchdog
>
> This looks like a bug, why is this so?
>
> Thanks,
>
>	Ingo


This is because the hard lockup detector and the soft lockup detector are
enabled and disabled at the same time - there isn't a separate 'knob' for
each of them. Both are controlled via the 'watchdog_user_enabled' variable
which is 1 by default.

  lockup_detector_init
    if (watchdog_user_enabled)
        watchdog_enable_all_cpus
          smpboot_register_percpu_thread(&watchdog_threads)

At boot time, the above code path lauches a 'watchdog/N' thread for each
online CPU. The watchdog_enable() function is executed in the context of
these threads, and this attempts to enable the hard lockup detector and
the soft lockup detector. [Note: Soft lockup detection is implemented in
watchdog_timer_fn().]

  watchdog_enable
    hrtimer_init(hrtimer, ...)
    hrtimer->function = watchdog_timer_fn

    watchdog_nmi_enable
      perf_event_create_kernel_counter(..., watchdog_overflow_callback)

    hrtimer_start(hrtimer, ...)

On bare metal systems or in virtual environments where the hypervisor
does not emulate a PMU, watchdog_nmi_enable() can fail to allocate and
enable a PMU counter. This is reported by a console message:

  NMI watchdog: disabled (cpu0): hardware events not enabled

Hence, we can end up with a situation where the soft lockup detector is
enabled and the hard lockup detector is not enabled. However, the output
of 'cat /proc/sys/kernel/nmi_watchdog' is 1 because it merely shows the
state of the 'watchdog_user_enabled' variable.

The above is the behaviour even without the proposed patch. The patch
merely adds the following hunk in watchdog_nmi_enable() to 'fake' a
-ENOENT error return from perf_event_create_kernel_counter().

  +        if (!watchdog_hardlockup_detector_is_enabled()) {
  +                event = ERR_PTR(-ENOENT);
  +                goto handle_err;
  +        }

The patch does not break the output of 'cat /proc/sys/kernel/nmi_watchdog'
since the discrepancy between the output and the actual state of the hard
lockup detector is nothing new.


Regards,

Uli
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/