Re: [PATCH] sysrq, kdump: fix regression, revert "simplify sysrq-c handler"

[Vivek Goyal <vgoyal@redhat.com>]

On Wed, Jul 22, 2009 at 07:10:49AM -0400, Neil Horman wrote:
> On Wed, Jul 22, 2009 at 11:01:29AM +0900, Hidetoshi Seto wrote:
> > Neil Horman wrote:
> > > None of this answers Erics question, what is it that you could do before, that
> > > you couldn't do now?
> > 
> > One is, as Ohmichi-san pointed, triggering kdump via echo c > /proc/sysrq-trigger.
> > 
> > In contrast to oops via SysRq-c from keyboard interrupt which results in
> > panic due to in_interrupt(), oops via echo-c will not become panic unless
> > panic_on_oops.
> > 
> > So in other words, we could expect same effect in both of echo-c and SysRq-c
> > before, but now we cannot because it depends on the panic_on_oops.
> > Isn't it a regression?
> > 
> Only if you blindly consider a change in behavior to be a regression.  consider
> that previously executing a sysrq-c did the same thing if you did a echo c >
> /proc/sysrq-trigger on a keyboard sysrq-c, but did different things based on
> weather or not your had a kexec kernel loaded.
> 
> > Whether kdump should be executed on oops (which is not panic) or not is a
> > separate thing.
> > 
> > > There are reasons to want to have a convenient way to
> > > crash the kernel, other than to test kdump (several distributions have augmented
> > > sysrq-c to do this for some time to test other previous dump mechanisms and
> > > features), so while its not been upstream, saying that its well known to test
> > > kdump without causing an oops is a bit of a misleading statement.
> > 
> > Let make me sure the difference between 'crash', 'oops', and 'panic'.
> > At least 'oops' is not panic, as is obvious from the name of panic_on_oops.
> > And it seems you are using 'crash' and 'oops' in mixture.
> > 
> I'm perfectly well aware of the difference, I just assert theres value to having
> sysrq-c be able to test both paths, especially given that we already have the
> sysrq-c sysctl available to toggle behavior for just this case.
> 
> > If you mean 'crash' as 'panic', my complaint is echo-c does not panic while
> > SysRq-c does panic.  So if possible I'd like to suggest a change like:
> >
> See above, I think theres value to having sysrq-c be able to do both, although I
> agree the method by which it triggers both is a bit muddled.
>  
> >  static void sysrq_handle_crash(int key, struct tty_struct *tty)
> >  {
> > -	char *killer = NULL;
> > -	*killer = 1;
> > +	panic("SysRq-triggered panic!\n");
> >  }
> > 

> Well, this removes the ability from sysrq-c to test the oops handling path, but
> I suppose it does buy us consistent behavior between the keyboard and proc
> interfaces, which is likely more important.  I can agree to that.  Perhaps we
> can create another sysctl to test the oops path later.
> 

Can't we just set panic_on_oops = 1 in sysrq_handle_crash()? This will
make sure that we test oops path as well as have consistent behavior
between two methods of sysrc-c invocation.

Thanks
Vivek

> > I agree that causing a real crash(panic) is better way to test crashdump than
> > calling the entry function of the crashdump directly, and also that opening
> > the path for other dump mechanisms is welcomed.
> >  
> Ok, so we're in line there :)
> 
> > > It seems to
> > > me that right now your major complaint is that the documentation is out of date,
> > > and you're having to do things slightly differently to get the same behavioral
> > > results.  Would it solve your issue, if we simply updated the documentation to
> > > illustrate how it works now?
> > 
> > Of course the documentation should be updated asap.
> > But I think the major complaint is about a difference in the behaviors of SysRq-c
> > and "echo c > /proc/sysrq-trigger".
> > 
> Ok, I can agree with that.  I'd support a change like what you have above to
> bring the keyboard and proc interface behavior in line.
> 
> Regards
> Neil
> 
> 
> > 
> > Thanks,
> > H.Seto
> > 
> >