Re: [PATCH] panic.c: export panic_on_oops

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 12 Oct 2009 20:23:46 +0200 Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > On Mon, 12 Oct 2009 08:36:38 -0700 (PDT) Linus Torvalds <torvalds@linux-foundation.org> wrote:
> > 
> > > I agree with the "save kernel buffer on panic" thing, but I disagree with 
> > > making it anything new, and hooking into "printk()" or the console 
> > > subsystem AT ALL. That's just bogus, stupid, and WRONG.
> > > 
> > > What you can do is to just flush the 'log_buf' buffer (or as much of it as 
> > > you want - the buffer may be a megabyte in size, and maybe you only want 
> > > to flush the last 8kB or something like that) on oops. And _not_ mix this 
> > > up with anything else.
> > 
> > What he said.  I did it that way in the Digeo kernel back in 2002. 
> > Worked good.
> > 
> > Doing it via a console is rather weird.  It will need core kernel 
> > changes to do it properly.
> > 
> > Perhaps oops_enter() is a good place to mark the start of the log, and 
> > flush it within oops_exit().
> 
> Simplest would be to do the last 2K in oops_exit()? That gives the oops, 
> and the history leading up to it. Since the blocking is 2K, the extra 
> log output is for free.
> 
> (unless the oops is larger than 2K - but that is rather rare.)
> 

Some oops traces can be pretty large.  Perhaps "oops_enter minus 1k up
to oops_exit".

The digeo kernel later got changed to package the oops into a binary
record format for logging to nvram.  iirc that was for space reasons,
and for ease of downstream analysis.

That kernel used to be downloadable but I can't immediately find it. 
Probably not very interesting anyway.