Re: Regression: ONE CPU fails bootup at Re: [3.2.0-RC7] BUG: unable to handle kernel NULL pointer dereference at 0000000000000598 1.478005] IP: [<ffffffff8107a6c4>] queue_work_on+0x4/0x30

[Stefan Bader <stefan.bader@canonical.com>]

On 04.01.2012 13:25, Stefan Bader wrote:
> On 04.01.2012 09:17, Stefan Bader wrote:
>> On 04.01.2012 01:53, John Stultz wrote:
>>> On Wed, 2012-01-04 at 11:31 +1100, NeilBrown wrote:
>>>> On Tue, 03 Jan 2012 15:09:48 -0800 John Stultz <john.stultz@linaro.org> wrote:
>>>>> >From the stack trace, we've kicked off a rtc_timer_do_work, probably
>>>>> from the rtc_initialize_alarm() schedule_work call added in Neil's
>>>>> patch.  From there, we call __rtc_set_alarm -> cmos_set_alarm ->
>>>>> cmos_rq_disable -> cmos_checkintr -> rtc_update_irq -> schedule_work.
>>
>> Sorry, I was off for the evening a while after sending this out. And I just
>> started, so a few thing I will be doing later but have not yet had time.
>>
>> Over night I had still be thinking on this and maybe one important fact I had
>> been ignoring. This really has only been observed on paravirt guests on Xen as
>> far as I know. And one thing that I should have pointed out is that
>>
>> [    0.792634] rtc_cmos rtc_cmos: rtc core: registered rtc_cmos as rtc0
>> [    0.792725] rtc_cmos: probe of rtc_cmos failed with error -38
>>
>> So first the registration is done and the first line is the last thing printed
>> in the registration function. Then, and that line always comes after, the probe,
>> which looks like being done asynchronously, detects that the rtc is not
>> implemented. I would assume that this causes the rtc to be unregistered again
>> and that is probably the point where, under the right circumstances, the worker
>> triggered by the initialize alarm is trying to set another alarm. Probably while
>> some of the elements of the structure started to be torn down. I need to check
>> on that code path, yet. So right now its more a guess.
>>
>>>>>
>>>>> So, what it looks to me is that in cmos_checkintr, we grab the cmos->rtc
>>>>> and pass that along. Unfortunately, since the cmos->rtc value isn't set
>>>>> until after rtc_device_register() returns its null at that point. So
>>>>> your patch isn't really fixing the issue, but just reducing the race
>>>>> window for the second cpu to schedule the work.
>>>>>
>>>>> Sigh. I'd guess dropping the schedule_work call from
>>>>> rtc_initialize_alarm() is the right approach (see below). When reviewing
>>>>> Neil's patch it seemed like a good idea there, but it seems off to me
>>>>> now.
>>>>>
>>>>> Neil, any thoughts on the following? Can you expand on the condition you
>>>>> were worried about in around that call?
>>>>
>>>> If you set an alarm in the future, then shutdown and boot again after that
>>>> time, then you will end up with a timer_queue node which is in the past.
>>>
>>> Thanks for explaining this again.
>>>
>>> Hrm. It seems the easy answer is to simply not add alarms that are in
>>> the past.  Further, I'm a bit perplexed, as if they are in the past, the
>>> enabled flag shouldn't be set.   __rtc_read_alarm() does check the
>>> current time, so maybe we can make sure we don't return old values? I
>>> guess I assumed __rtc_read_alarm() avoided returning stale values, but
>>> apparently not.
>>>
>>>> When this happens the queue gets stuck.  That entry-in-the-past won't get
>>>> removed until and interrupt happens and an interrupt won't happen because the
>>>> RTC only triggers an interrupt when the alarm is "now".
>>>>
>>>> So you'll find that e.g. "hwclock" will always tell you that 'select' timed
>>>> out.
>>>>
>>>> So we force the interrupt work to happen at the start just in case.
>>>
>>> Unfortunately its too early. 
>>>
>>>> Did you see my proposed patch which converted those calls to do the work
>>>> in-process rather than passing it to a worker-thread?  I think that is a
>>>> clean fix.
>>>
>>> I don't think I saw it today. Was it from before the holidays?
>>>
>>
>> I fear I caused a bit of confusion there. Neil responded to my initial mail
>> which was done as a reply to the mail announcing this patch for stable (which
>> just was the first thread I could get hold of).
>> I will try Neil's patch as well. And in parallel try to see whether the theory I
>> had this night makes sense. If it does, then it is only indirectly that the work
>> is scheduled too early. In that case just the teardown needs to make sure that
>> no work is being run while removal. Well, maybe the question is whether there
>> should be a delay in running the irq work until the device really, really is
>> completely set up... But that sounds a bit more complicated.
> 
> By now I tried Neil's proposed patch and unfortunately that makes things rather
> worse. I also played around with the idea of the unregistration race. Maybe
> there also is one (that cancel_work_sync should be called before unregistering
> the device) but definitely it is not what happens at least in the one CPU case.
> I added some more printk's and the crash happens before even the rtc core class
> has been fully registered. And no unregister is call has been made either.
> 
> Which may point to execution of the irq worker (including a schedule_work)
> before the rtc-cmos parts are finished... Would explain why moving the
> initialize call further down does at least narrow the window for it to happen...
> The only thing I do not understand then is why that seems only to happen on Xen
> guests...
> 

Darn, guess I understand now... So cmos-rtc will call rtc_device_register from
within do_cmos_probe and set cmos_rtc.rtc with the pointer that
rtc_device_register returns.
But when having rtc_initialize_alarm being called earlier in
rtc_device_register, and that queues the work, there is a chance it will fire
even before that function returns.
And then cmos_checkintr will dereference the null pointer still stored in
cmos_rtc.rtc for calling rtc_update_irq...

-Stefan
> -Stefan
> 
>>> Even so, at this point, I don't know if we have enough time for testing,
>>> so I'm thinking we either just drop the problematic sched_work call or
>>> revert the whole thing and try again for 3.3
>>
>> That was the reason I was in a bit of hurry to get this back to you. Especially
>> since this patch had been marked as stable material and sooner or later will or
>> would be added to all the stable releases it applies to.
>>
>> Thanks,
>> Stefan
>>
>>> thanks
>>> -john
>>>
>>>
>>
>