panic: APIC: Previous IPI is stuck

[Adrian Wontroba]

On Mon, Nov 15, 2004 at 04:49:56PM +1100, Andy Farkas wrote:
> [freebsd.org is rejecting my email (cant find hostname)
> so please feel free to copy this to the list]

So quoted in full.

> On Mon, 15 Nov 2004, Adrian Wontroba wrote:
> ...
> > The practice is that it it has now crashed three times in a couple of
> > days with "panic: APIC: Previous IPI is stuck", the most recent one
> > dragging me out from home early in a Monday morning.
> 
> /me raises hand
> 
> I still get panics too (5.3-STABLE cvsup'd last thursday).
> At one stage I thought it was fixed, but I was wrong.
> My box does not reboot itself either.
> 
> > Over in current there are a couple of threads starting in late September
> > where a few people are suffering this problem.  Like them, I'm using an
> > old (1997) Pentium Pro multiprocessor, in my case a 4 way Fujitsu M700.
> >
> > The machine is running with the SMP kernel (ie GENERIC + SMP), 4BSD
> > scheduler, without preemption.
> 
> Robert Watson has said it happens on his 4-way xeon box,
> so its not the "old hardware" thats to blame. (My box is
> an old Dell quad-ppro too). Something changed in the code
> around the end of August this year.
> 
> > I've set kern.sched.ipiwakeup.enabled=0 and crossed my fingers.
> 
> Doesn't help. I already tried. Panic will still happen.

Ah.  Will it last the day I wonder?

> > I'm a SMP novice.  Would the machine become stable if I switched to a
> > non-SMP kernel?  Reliability is more important than speed in this case,
> > and the opportunity for experimentation close to zero.  Creditability
> > has already been damaged by the gvinum RAID5 experience (8-(
> 
> A UP kernel will probably run forever. The IPI panic can
> only happen on SMP kernels.

Thanks.  I'll switch back to GENERIC.

> > I'm not knocking 5.3 - in all other respects it seems wonderful.
> 
> I'm not knocking 5.3 either, but it seems to its not quite
> stable. Its more of ".0" release, where things are still
> getting ironed out (like gvinum, which I also have problems
> with).

"RELENG_4: Time to die" - for all kinds of good reasons.  It was time
for 5-STABLE. The future release plan looks promising, but there
is still the age old problem - how do you get the more of the user
population to try out and find the problems in new versions before they
acquire -RELEASE status?

"Mea culpa" - I no longer have a "crash box".  Time to get my mail
off my own PPro (uniprocessor) box to free it up as such.  If I had
done this, I would have run into the vinum / gvinum issues in a less
embarrassing fashion.

> Stephan, you mentioned that the IPI code needs rewriting in order to
> fix this problem... how's it going?
>
> - andyf

-- 
Adrian Wontroba