Re: [PATCH v11 4/5] core: Add kernel message dumper to call on oopses and panics

["Shargorodsky Atal (EXT-Teleca/Helsinki)" <ext-atal.shargorodsky@nokia.com>]

On Mon, 2009-10-26 at 08:41 +0100, ext Simon Kagstrom wrote:
> On Fri, 23 Oct 2009 18:53:22 +0300
> "Shargorodsky Atal (EXT-Teleca/Helsinki)" <ext-atal.shargorodsky@nokia.com> wrote:
> 
> > 1. If somebody writes a module that uses dumper for uploading the
> > oopses/panics logs via some pay-per-byte medium,  since he has no way
> > to know in a module if the panic_on_oops flag is set, he'll have
> > to upload both oops and the following panic, because he does not
> > know for sure that the panic was caused by the oops. Hence he
> > pays twice for the same information, right?
> > 
> > I can think of a couple of way to figure it out in the module
> > itself, but I could not think of any clean way to do it.
> 
> This is correct, and the mtdoops driver has some provisions to handle
> this. First, there is a parameter to the module to specify whether
> oopses should be dumped at all - I added this for the particular case
> that someone has panic_on_oops set.
> 

It takes care of most of the situations, but panic_on_oops
can be changed any time, even after the module is loaded.

While I think that exporting oops_on_panic is a wrong thing to do,
I believe that dumpers differ a bit from the rest of the modules in
that aspect and should be at least hinted about this flag setting.
Does it not make sense?

> Second, it does not dump oopses directly anyway, but puts it in a work
> queue. That way, if panic_on_oops is set, it will store the panic but
> the oops (called from the workqueue) will not get written anyway.
> 

AFAIK, mtdoops does not put oopses in a work queue. And if by any chance
it does, then I think it's wrong and might lead to missed oopses, as
the oops might be because of the work queues themselves, or it might
look to the kernel like some non-fatal fault, but actually it's a
sign of a much more catastrophic failure - IOMMU device garbaging
memory, for instance.

But anyway, I was not talking about mtdoops. In fact, I was not
talking about any particular module, I just described some situation
which looks a bit problematic to me.

> > 2. We tried to use panic notifiers mechanism to print additional
> > information that we want to see in mtdoops logs and it worked well,
> > but having the kmsg_dump(KMSG_DUMP_PANIC) before the
> > atomic_notifier_call_chain() breaks this functionality.
> > Can we the call kmsg_dump() after the notifiers had been invoked?
> 
> Well, it depends I think. The code currently looks like this:
> 
> 	kmsg_dump(KMSG_DUMP_PANIC);
> 	/*
> 	 * If we have crashed and we have a crash kernel loaded let it handle
> 	 * everything else.
> 	 * Do we want to call this before we try to display a message?
> 	 */
> 	crash_kexec(NULL);
> 	[... Comments removed]
> 	atomic_notifier_call_chain(&panic_notifier_list, 0, buf);
> 
> And moving kdump_msg() after crash_kexec() will make us miss the
> message if we have a kexec crash kernel as well. I realise that these
> two approaches might be complementary and are not likely to be used at
> the same time, but it's still something to think about.
> 
> Then again, maybe it's possible to move the panic notifiers above
> crash_kexec() as well, which would solve things nicely.
> 

Which leaves me no choice but just ask the question, as it bothering me
for some time: does anybody know why we try to crash_kexec() at so early
stage?


> // Simon