[Adeos-main] Re: SMP-problem with cvs-vulcano/3.4

[Philippe Gerum <rpm@xenomai.org>]

On Tue, 2006-10-10 at 11:29 +0200, Jan Kiszka wrote:
> Anders Zachrison wrote:
> > Reviving an old thread of mine, in the rtai-mail-list. As suggested by
> > Jan, I will also try to send it to the adeos list (if this doesn't
> > succeed, could you forward it, Jan?) (And last, for possible readers on
> > the adeos-list, pleas CC-me, as I'm not following that list).
> > 
> > * Jan Kiszka <jan.kiszka@domain.hid> [060922 12:37]:
> >> Jan Kiszka wrote:
> >>> Anders Zachrison wrote:
> >>>> ...  As a short summary, as soon as I patch the 2.6.17-kernels with
> >>>> the rtai-patch from vulcano-cvs (and apparently magma-cvs) the
> >>>> kernel stops booting. Quite often the call-trace shows calls to
> >>>> various ipipe-functions mainly setup and initialization functions.
> >>>> Other times it is around a call to smp_call_function.
> > 
> > A short summary for possible readers; as stated above, the computer
> > stops booting as soon as the 2.6.17.x kernel is patched with either an
> > adeos-ipipe patch or with the patch supplied by RTAI. It turned out the
> > the problem was MSI related (see the end of the mail).
> > 
> > It is an old SMP-box with double PII-400MHz processors. 
> > 
> > The kernel config is available at
> > http://hydra.ikp.liu.se/~andza/adeos_oops/config and a further
> > specification of the computer (from `lspci -vv`) is found at
> > http://hydra.ikp.liu.se/~andza/adeos_oops/lspci-vv.txt . The computer is
> > running Debian stable.
> > 
> >>> Could you do the Adeos project a favour and test the vanilla Ipipe
> >>> patch as well (http://download.gna.org/adeos/patches/v2.6/i386)?
> >>> Please report breakages with precise version number, system
> >>> description, and .config
> > 
> > I've used a 2.6.17.11-kernel (same problems occurs with a 2.6.17 one)
> > and I've patched it with the adeos-ipipe-2.6.17-i386-1.4-00.patch from
> > (http://download.gna.org/adeos/patches/v2.6/i386).
> > 
> >> ... and the oops if you can catch it (syslog, serial console, etc.).
> > 
> > Three oopses are available;
> > http://hydra.ikp.liu.se/~andza/adeos_oops/oops1.txt
> > http://hydra.ikp.liu.se/~andza/adeos_oops/oops2.txt
> > http://hydra.ikp.liu.se/~andza/adeos_oops/oops3.txt .
> > (Well, they should be all or most of the boot process, as received by
> > the serial console. However, at least the latter two lacks some parts of
> > the beginning, although the oopses at the end should be complete.)
> 
> Thanks for the oopses. They seem to point in the same direction Philippe
> already guessed: IRQ migration (the Linux code screams for some spinlock
> being acquired, which is probably not the case when the pipeline is active).
> 
> > 
> >>> to adeos-main@gna.org or to me (will forward then).
> >>>
> >>> Many thanks in advance!
> >>> Jan
> >>>
> > 
> >> BTW, there used to be problems around MSI with I-pipe patches, though
> >> my current information is that they are fixed. If you have MSI on your
> >> box, it might be worth trying to switch related kernel support off
> >> (CONFIG_PCI_MSI).
> > 
> > As stated in a previous reply, this was the case. Disableing the MSI in
> > the kernel config solved the problem, while disableing it in the kernel
> > boot parameter did not.
> > 
> > Anders
> > 
> 
> There were some commits to Adeos CVS recently regarding
> kernel/irq/migration.c [1], but the result looks unfinished to me (no
> more differences to vanilla?). Philippe, is this supposed to solve the
> issue?

No, this was a left over from a previous attempt to investigate the
issue. The issue is still pending.

> 
> Jan
> 
> [1]http://cvs.gna.org/cvsweb/ipipe/v2.6/2.6.17/kernel/irq/migration.c?cvsroot=adeos
> 
-- 
Philippe.