Re: [RFC] cell: saving spus information for kexec crash

Re: [RFC] cell: saving spus information for kexec crash

[Anton Blanchard]

 
Hi,

> This patch adds support for investigating spus information after a
> kernel crash event, through kdump vmcore file.
> Implementation is based on xmon code, but the new functionality was
> kept independent.

This brings up a point: it looks like kdump doesnt capture the FP and
altivec regs. Could be important for some bugs.

This brings up yet another point, it looks like our userspace core dumps
dont capture altivec or SPU state.

Anton

Re: [RFC] cell: saving spus information for kexec crash

[Arnd Bergmann]

On Friday 11 May 2007, Anton Blanchard wrote:

> > This patch adds support for investigating spus information after a
> > kernel crash event, through kdump vmcore file.
> > Implementation is based on xmon code, but the new functionality was
> > kept independent.
> 
> This brings up a point: it looks like kdump doesnt capture the FP and
> altivec regs. Could be important for some bugs.

Yes, and there are more things that might be worth adding:

- SLB contents (as noticed by Luke Browning)
- some SPRs
- state of some device drivers

We have briefly discussed the idea of having generic notifiers into
which arbitrary subsystems can register, but unless we have more than
one or two users, it's probably not worth it.

> This brings up yet another point, it looks like our userspace core dumps
> dont capture altivec or SPU state.

Actually, the userspace core dumps do capture the SPU state as of 2.6.21,
thanks to the work of Dwayne McConnell :-)

	Arnd <><

Re: [Cbe-oss-dev] [RFC] cell: saving spus information for kexec crash

[Benjamin Herrenschmidt]

On Fri, 2007-05-11 at 12:22 -0500, Anton Blanchard wrote:
> Hi,
> 
> > This patch adds support for investigating spus information after a
> > kernel crash event, through kdump vmcore file.
> > Implementation is based on xmon code, but the new functionality was
> > kept independent.
> 
> This brings up a point: it looks like kdump doesnt capture the FP and
> altivec regs. Could be important for some bugs.
> 
> This brings up yet another point, it looks like our userspace core dumps
> dont capture altivec or SPU state.

That's interesting: We do have a dump_task_altivec() but it looks like
its never called ... ouch. Maybe we can use the XFPREGS stuff or should
we add our own hook in binfmt_elf.c for that ?

As for SPU state, there's code to dump the SPEs in a separate section,
though I don't know if it's been merged yet (no up to date git tree
at hand right now). It uses different macros off binfmt_elf.

Ben.