[PATCH v2] kernel/panic/kexec: fix "crash_kexec_post_notifiers" option issue in oops path

[Hidehiro Kawai]

Hello all,

(2015/03/24 23:32), Vivek Goyal wrote:
> On Tue, Mar 24, 2015 at 05:27:10AM -0500, Eric W. Biederman wrote:
>> Ingo Molnar <mingo at kernel.org> writes:
>>
>>> * Masami Hiramatsu <masami.hiramatsu.pt at hitachi.com> wrote:
>>>
>>>>>
>>>>>   f06e5153f4ae ("kernel/panic.c: add "crash_kexec_post_notifiers" option for kdump after panic_notifers")
>>>>>
>>>>> Was that crash_kexec() was called unconditionally after notifiers were 
>>>>> called, which should be fixed via the simple patch below (untested). 
>>>>> Looks much simpler than your fix.
>>>>
>>>> No, Daisuke's patch is not for that case. [...]
>>>
>>> Yet the actual bug is in that commit, 'crash_kexec_post_notifiers' was 
>>> clearly not a no-op in the default case, against expectations.
>>>
>>> So the first step should be to restore the original behavior (my 
>>> patch), then should any new tweaks be added.
>>
>> Honestly I think the proper fix is to simply revert f06e5153f4ae.
>>
>> It was clearly not properly tested by the people who wanted it because
>> they came back quite a while later with additional bleh.
>>
>> I think this pretty much counts as hitting the code doesn't work let's
>> remove it threshold.
> 
> IMHO, we should give users flexibility of running panic notifiers before
> crash_kexec(). Different people have been asking for it since last 7-8
> years and it is a pretty small code in kernel so no major maintenance
> headache. 
> 
> Agreed that this might be very unreliable, but if users want to shoot
> themseleves in the foot, it is their choice. This will not be upstream
> default and I am hoping that distributions don't make it their default
> either.

We are going to use panic notifier to write SEL record, and actually
it seems to be unreliable.  At least I found two problems in IPMI driver
code while testing Hatayama-san's patch, and they will cause an infinite
loop.  I think users wouldn't notice this bug because most of users use
kdump and there is no difference on display between the infinite loop
case and successful case.

Anyway, we need to harden panic notifier callee.  I will post bug fix
patches for IPMI driver ASAP.

Best regards,

Hidehiro Kawai
Hitachi, Yokohama Research Laboratory