On Mon, Apr 30, 2012 at 2:10 PM, Sameer Nanda <[email protected]> wrote:
> On Sun, Apr 29, 2012 at 11:12 PM, Srivatsa S. Bhat
> <[email protected]> wrote:
>> On 04/27/2012 11:40 PM, Sameer Nanda wrote:
>>
>>> On the suspend/resume path the boot CPU does not go though an
>>> offline->online transition. This breaks the NMI detector
>>> post-resume since it depends on PMU state that is lost when
>>> the system gets suspended.
>>>
>>> Fix this by forcing a CPU offline->online transition for the
>>> lockup detector on the boot CPU during resume.
>>>
>>> Signed-off-by: Sameer Nanda <[email protected]>
>>> ---
>>> To provide more context, we enable NMI watchdog on
>>> Chrome OS. We have seen several reports of systems freezing
>>> up completely which indicated that the NMI watchdog was not
>>> firing for some reason.
>>>
>>> Debugging further, we found a simple way of repro'ing system
>>> freezes -- issuing the command 'tasket 1 sh -c "echo nmilockup > /proc/breakme"'
>>> after the system has been suspended/resumed one or more times.
>>>
>>> With this patch in place, the system freeze result in panics,
>>> as expected. These panics provide a nice stack trace for us
>>> to debug the actual issue causing the freeze.
>>>
>>>
>>> include/linux/sched.h | 4 ++++
>>> kernel/power/suspend.c | 3 +++
>>> kernel/watchdog.c | 16 ++++++++++++++++
>>> 3 files changed, 23 insertions(+), 0 deletions(-)
>>>
>>> diff --git a/include/linux/sched.h b/include/linux/sched.h
>>> index 81a173c..118cc38 100644
>>> --- a/include/linux/sched.h
>>> +++ b/include/linux/sched.h
>>> @@ -317,6 +317,7 @@ extern int proc_dowatchdog_thresh(struct ctl_table *table, int write,
>>> size_t *lenp, loff_t *ppos);
>>> extern unsigned int softlockup_panic;
>>> void lockup_detector_init(void);
>>> +void lockup_detector_bootcpu_resume(void);
>>> #else
>>> static inline void touch_softlockup_watchdog(void)
>>> {
>>> @@ -330,6 +331,9 @@ static inline void touch_all_softlockup_watchdogs(void)
>>> static inline void lockup_detector_init(void)
>>> {
>>> }
>>> +static inline void lockup_detector_bootcpu_resume(void)
>>> +{
>>> +}
>>> #endif
>>>
>>> #ifdef CONFIG_DETECT_HUNG_TASK
>>> diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c
>>> index 396d262..0d262a8 100644
>>> --- a/kernel/power/suspend.c
>>> +++ b/kernel/power/suspend.c
>>> @@ -177,6 +177,9 @@ static int suspend_enter(suspend_state_t state, bool *wakeup)
>>> arch_suspend_enable_irqs();
>>> BUG_ON(irqs_disabled());
>>>
>>> + /* Kick the lockup detector */
>>> + lockup_detector_bootcpu_resume();
>>> +
>>> Enable_cpus:
>>> enable_nonboot_cpus();
>>>
>>> diff --git a/kernel/watchdog.c b/kernel/watchdog.c
>>> index df30ee0..dd2ac93 100644
>>> --- a/kernel/watchdog.c
>>> +++ b/kernel/watchdog.c
>>> @@ -585,6 +585,22 @@ static struct notifier_block __cpuinitdata cpu_nfb = {
>>> .notifier_call = cpu_callback
>>> };
>>>
>>> +void lockup_detector_bootcpu_resume(void)
>>> +{
>>> + void *cpu = (void *)(long)smp_processor_id();
>>> +
>>> + /*
>>> + * On the suspend/resume path the boot CPU does not go though the
>>> + * offline->online transition. This breaks the NMI detector post
>>> + * resume. Force an offline->online transition for the boot CPU on
>>> + * resume.
>>> + */
>>> + cpu_callback(&cpu_nfb, CPU_DEAD, cpu);
>>> + cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
>>> +
>>
>>
>> I have a couple of comments about this:
>>
>> 1. Strictly speaking, we should be using the _FROZEN variants here (since the
>> tasks are still frozen).
>>
>> Like, cpu_callback(&cpu_nfb, CPU_DEAD_FROZEN, cpu);
>> and cpu_callback(&cpu_nfb, CPU_ONLINE_FROZEN, cpu);
>>
>> Right now, since the same action is taken for either variant (ie., with or without
>> _FROZEN), it really doesn't matter. But still, good to be on the safer side no?
>
> Agreed that the _FROZEN counterparts are a better fit here since the
> tasks are still frozen. Let me make this change.
>
>>
>> 2. Why are we skipping the CPU_UP_PREPARE_FROZEN callback?
>
> Mainly because the hrtimer_init has already been done at kernel init
> time. But, this seems to be a good idea since the non-boot CPUs do
> transition through the CPU_UP_PREPARE_FROZEN phase on the way up
> during resume so it makes sense to keep the boot CPU path symmetrical.
>
> Let me make this change also.
Just sent the updated patch incorporating these two changes as well as
the earlier feedback from akpm.
>
>>
>> 3. How about hibernation? We don't hit this problem there?
>
> I am not too familiar with hibernation path and don't have a setup to
> test it either so can't really answer this one.
>
>>
>>> + return;
>>> +}
>>> +
>>> void __init lockup_detector_init(void)
>>> {
>>> void *cpu = (void *)(long)smp_processor_id();
>>
>>
>>
>> Regards,
>> Srivatsa S. Bhat
>>
>
>
>
> --
> Sameer
--
Sameer