Re: [patch v2 04/10] kdump: Trigger kdump via panic notifier chainon s390

[Vivek Goyal]

On Mon, Aug 08, 2011 at 07:47:39PM +0200, Michael Holzheu wrote:
> Hello Vivek,
> 
> On Thu, 2011-08-04 at 17:14 -0400, Vivek Goyal wrote:
> > On Wed, Aug 03, 2011 at 11:50:39AM +0200, Michael Holzheu wrote:
> 
> [snip]
>  
> > > Only if the user has specified panic action stand-alone dump, we do the
> > > detour via the stand-alone dump tools.
> > 
> > If a user decides to load kdump kernel to capture dump, then why does it
> > still make sense to set panic action as "stand-alone dump tools". One
> > could argue that user loaded kdump kernel but not necessarily wants
> > that mechanism to use, in that case dump-tools does not have to jump
> > to kdump kernel at all. 
> 
> The user on s390 can currently freely configure the panic action. If we
> would always do kdump on panic, when kdump is active, we would have to
> ensure in sysfs that the user can't change the setting any more.
> Technically we could do that of course.
> 
> One use case we had in mind was that the s390 Linux administrator does
> not have to learn new things when using kdump. kdump can be used as
> extension of the already existing mechanism. Today users configure
> stand-alone dump via sysfs. With kdump + trigger via dump tools they
> could do it still the same way. And also for manual dump they just IPL
> the dump tool as they are used to do it and if kdump is fine, kdump will
> be triggered. Nothing new has to be learned.
> 
> If users only want to use kdump without stand-alone dump tools as on
> other architectures also this is possible. Then the panic action will be
> just kdump.

IMHO, it makes sense to introduce a new trigger action "kdump" and let
user configure that instead of masking everything behind dump tools.

In fact, we don't have to even configure it. Like other architectures, 
we can always call crash_kexec() upon panic() if there is a crash
kernel loaded. That way, panic() path remains simple and there are
no arch specific #ifdefs.

> 
> > > 
> > > > like other architectures and jump to stand
> > > > alone kernel only if some piece of code is corrupted and that action
> > > > failed.
> > > > 
> > > > What's the point of jumping to stand alone kenrel in case of panic()
> > > > and then re-enter it back to original kernel using crash_kexec(). Sound
> > > > like a very odd design choice to me.
> > > > 
> > > > I am now I am repeating this question umpteen time simply because
> > > > I never got a good answer except "we have to do it this way".
> > > 
> > > Sometimes communication is really hard and frustrating.
> > > ... but at least we are still communicating.
> > > 
> > > Ok very last try:
> > > 
> > > * We can use the same mechanism for manual dump and automatic dump on
> > > panic: IPL the stand-alone dump tools.
> > 
> > So manual dump/intervention is only required if automatic dump failed?
> 
> Manual intervention is required only if panic code does not make it to
> the "IPL stand-alone dump tools" code.
> 
> > 
> > > kdump check and backup
> > > stand-alone dump is implemented only in the stand-alone dump code.
> > 
> > My argument is that why stand alone dump is trying to trigger kdump
> > at all? Shouldn't it all be part of loading kdump kernel and user
> > setting panic() action to kdump?
> 
> To summarize: Our approach was to do it in the stand-alone dump tools
> code for both the manual and the automatic on panic case:
> 
> panic ------+                                 +- valid -> kdump
>             +-> IPL dump tools -> try kdump --+
> hard hang --+                                 +- invalid -> stand-alone dump
> 
> Your suggestion looks like the following:
> 
> panic --> try kdump +-- valid ---> kdump
>                     |
>                     +-- invalid -> IPL dump tools --> stand-alone dump
> 
>                                             +- valid -> kdump
> hard hang --> IPL dump tools -> try kdump --+ 
>                                             +- invalid -> stand-alone dump

I think thew whole notion of jumping from dump tools to kdump is not
a very good design as it enforces us to pass additional state to dump
tools and also the awkward notion of re-enter the crashed kernel
(crash_kexec() being called from dump tools).

Is there a notion of NMI on s390. What's the x86 NMI equivalent on s390.

> Is that what you are suggesting? We can do it that way, too. Then
> we would not need the #ifdef CONFIG_S390 in panic().

How about keeping it simple in first round.

- Always use kdump for capturing dump is kdump kernel is loaded. If user
  does not want to use kdump, don't load kdump kenrel.

- Use NMI equivalent to handle the hard hang case. In case there is no
  equivalent, how about keeping it simple in first go and always use
  dump-tools to capture full dump and improve upon it in a later patchset.

  (This is no worse then today as today you don't have any filtering
   mechanism).

Thanks
Vivek
--
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/