[RFC]Kexec based crash dumping

[RFC]Kexec based crash dumping

[Hariprasad Nellitheertha]

Hi,

The patches that follow contain the initial implementation for kexec based
crash dumping that we are working on. I had sent this to the fastboot mailing 
list a couple of weeks ago and this set of patches includes the changes made as
per feedback from Andrew, Eric and others.

Main Idea

- Whenever a panic occurs, reboot to a new kernel using kexec using a small
  amount of memory (16MB). The rest of the memory is preserved across the
  reboot.
- In the second kernel, the memory contents from the failed kernel is 
  available as an ELF format file for write-out.

Details on the design and implementation and on how to setup this facility
are available in the first of the patches that follow. The patches have
been made for the 2.6.8.1 kernel.

Kindly review these patches and provide feedback.

Thanks to Martin Bligh and Suparna for the design ideas and to Adam Litke
who hacked up most of the memory preserving reboot code and the dump
device abstraction code.

Regards, Hari
-- 
Hariprasad Nellitheertha
Linux Technology Center
India Software Labs
IBM India, Bangalore

Re: [RFC]Kexec based crash dumping

[Andrew Morton]

Hariprasad Nellitheertha <hari@in.ibm.com> wrote:
>
> The patches that follow contain the initial implementation for kexec based
> crash dumping that we are working on.

It seems to be coming together nicely.

Where do we stand with support for other architectures?  Do you expect that
each architecture will involve a lot of work?

And how much of the i386 implementation do you expect x86_64 can
reuse?

Re: [RFC]Kexec based crash dumping

[Hariprasad Nellitheertha]

Hi Andrew,

On Tue, Aug 17, 2004 at 03:44:36PM -0700, Andrew Morton wrote:
> Hariprasad Nellitheertha <hari@in.ibm.com> wrote:
> >
> > The patches that follow contain the initial implementation for kexec based
> > crash dumping that we are working on.
> 
> It seems to be coming together nicely.
> 
> Where do we stand with support for other architectures?  Do you expect that
> each architecture will involve a lot of work?

I don't think so. The main architecture dependent components are register
snapshotting and silencing of other cpus. These are not new problems to
solve as other projects such as LKCD, KDB have already done this for
most archs.

> 
> And how much of the i386 implementation do you expect x86_64 can
> reuse?

We should be able to re-use most of the code. The x86_64 port should be 
ready pretty quickly once kexec itself is available.

Regards, Hari
-- 
Hariprasad Nellitheertha
Linux Technology Center
India Software Labs
IBM India, Bangalore

Re: [RFC]Kexec based crash dumping

[Itsuro Oda]

Hi.

It is interesting.

We(*) also make crash dump acording to similer concept.
    (*) NTT Data Corporation & VA Linux Systems Japan K.K.

Our main idea is the following.

- prepare a kernel which does only dump real memory to block
  device. ("dump mini kernel")
- pre-allocate the memory (4MB is enough) used by the dump mini
  kernel and pre-load the dump mini kernel.
- when crash occur exec the dump mini kernel.
- the dump mini kernel stands in and only uses pre-allocated
  area.

We made mkexec(mini kernel exec) based on kexec code but 
simplify a lot. Mkexec can co-exist kexec and can be kernel
module. We made also "dump mini kernel". 
(now i386 only)

We will open our code shortly. (maybe I think)
I think there are many items to be able to cooperate.

Thank you.
-- 
Itsuro ODA <oda@valinux.co.jp>

Re: [RFC]Kexec based crash dumping

[Tonnerre]

[-- Attachment #1: Type: text/plain, Size: 613 bytes --]

Salut,

On Thu, Aug 19, 2004 at 08:45:00AM +0900, Itsuro Oda wrote:
> - prepare a kernel which does only dump real memory to block
>   device. ("dump mini kernel")
> - pre-allocate the memory (4MB is enough) used by the dump mini
>   kernel and pre-load the dump mini kernel.
> - when crash occur exec the dump mini kernel.
> - the dump mini kernel stands in and only uses pre-allocated
>   area.

One question, what  happens in your concept when  some stubborn zombie
kernel driver  (say Nvidia) comes  along and overwrites  random memory
areas? Will the booting of the dump kernel simply fail?

			    Tonnerre

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: [RFC]Kexec based crash dumping

[Itsuro Oda]

Hi,

> areas? Will the booting of the dump kernel simply fail?
yes.

The pre-allocation area used by the dump kernel is set 
write protection after loading the dump kernel. It protects
from writing through virtual address space, but does not
protect through DMA writing.

Thank you

On Fri, 20 Aug 2004 16:28:40 +0200
Tonnerre <tonnerre@thundrix.ch> wrote:

> Salut,
> 
> On Thu, Aug 19, 2004 at 08:45:00AM +0900, Itsuro Oda wrote:
> > - prepare a kernel which does only dump real memory to block
> >   device. ("dump mini kernel")
> > - pre-allocate the memory (4MB is enough) used by the dump mini
> >   kernel and pre-load the dump mini kernel.
> > - when crash occur exec the dump mini kernel.
> > - the dump mini kernel stands in and only uses pre-allocated
> >   area.
> 
> One question, what  happens in your concept when  some stubborn zombie
> kernel driver  (say Nvidia) comes  along and overwrites  random memory
> areas? Will the booting of the dump kernel simply fail?
> 
> 			    Tonnerre

-- 
Itsuro ODA <oda@valinux.co.jp>