Re: Panic on boot. How do I get a kernel dump.

Peter Jeremy wrote:
> On 2007-Apr-27 14:00:21 -0400, Kris Kennaway <kris_at_obsecurity.org> wrote:
>> On Fri, Apr 27, 2007 at 08:31:16AM -0700, Tom Cumming wrote:
>>> One possibility is to hard code dumpdev in the kernel, then boot that
>>> kernel.
>> I don't know many different ways I can say "there is no way to do it".
> 
> I think Tom is saying that he needs to do something that _used_ to be
> possible.  Normally the Project is careful to avoid regressions so
> this was a surprise to me as well.  (It looks like it's demise wasn't
> clearly spelt out at the time).
> 
> Since dumpdev is now intertwined with geom and the geom tasting is
> quite late in the boot process, I agree that the current crashdump
> code does not seem amenable to use early in the boot process.
> 
> Having a kernel crash before it reaches userland is not unheard of.
> Whilst it may be possible to debug this using DDB or remote GDB in
> some cases, I can think of two cases where this is not practical:
> 1) It is a production server that can't be left down for extended periods.
> 2) It is a remote system without remote console access.
> 
> Lets put the original question slightly differently:  How can the
> kernel state be saved if the kernel crashes before it's possible to
> invoke dumpon(8)?  IMHO, "there is no way to do it" is not a
> satisfactory answer for the reasons above.
> 

Implement network crashdumps.  This involves writing a new, separate,
stripped down network stack, dealing with network configuration
headaches, etc.  It it will address the problem, though, albeit with a
lot of development effort and pain.

Scott