HEAD panic with ofw_pcibus.c 1.21 on Blade 100

[Marius Strobl]

On Mon, Sep 01, 2008 at 05:20:44PM +0100, Gavin Atkinson wrote:
> On Mon, 2008-09-01 at 15:20 +0100, Gavin Atkinson wrote:
> > Hi all,
> > 
> > My Blade 100 now panics on boot with HEAD, and I've tracked it down to
> > sys/sparc64/pci/ofw_pcibus.c 1.21 (SVN r182108) by marius at .
> > Specifically, this version now configures bridges differently, and not
> > setting "Master Abort Mode" prevents the panic:
> > 
> > Index: src/sys/sparc64/pci/ofw_pcibus.c
> > ===================================================================
> > RCS file: /home/ncvs/src/sys/sparc64/pci/ofw_pcibus.c,v
> > retrieving revision 1.21
> > diff -u -r1.21 ofw_pcibus.c
> > --- src/sys/sparc64/pci/ofw_pcibus.c    24 Aug 2008 15:05:46 -0000      1.21
> > +++ src/sys/sparc64/pci/ofw_pcibus.c    1 Sep 2008 14:09:27 -0000
> > @@ -140,7 +140,7 @@
> >             PCIM_HDRTYPE) == PCIM_HDRTYPE_BRIDGE) {
> >                 reg = PCIB_READ_CONFIG(bridge, busno, slot, func,
> >                     PCIR_BRIDGECTL_1, 1);
> > -               reg |= PCIB_BCR_MASTER_ABORT_MODE | PCIB_BCR_SERR_ENABLE |
> > +               reg |= /* PCIB_BCR_MASTER_ABORT_MODE | */ PCIB_BCR_SERR_ENABLE |
> >                     PCIB_BCR_PERR_ENABLE;
> >  #ifdef OFW_PCI_DEBUG
> >                 device_printf(bridge,
> 
> [snip]
> 
> > Any suggestions?  Are we missing some code necessary to support master
> > mode aborts?
> 
> After further research (mainly involving eyeballing
> pci_pbm_err_handler() in OpenSolaris), it looks like we are indeed
> missing code to handle them.  Therefore, until this code is written, I
> suspect the patch above is actually correct.
> 

While not setting master abort mode on PCI-PCI-bridges might
hide your problem, the right place for ignoring master and
(in this case) target aborts, both of which are fatal in
general though, would be the host-PCI-bridge. Similarly,
support for peeking and poking of I/O and memory space like
OpenSolaris apparently has (the associated recovery handlers
probably are the code you're refering to) should be implemented
PCI-bus wide and not just grounded at PCI-PCI-bridges. I don't
think there's a real need to go through the hoops to support
these in FreeBSD though. The blind bus access ahc(4) ISA
front-end does is also what hangs B100 during boot even with
master abort mode in the PCI-PCI-bridges I think.
We're looking at several problems here though and IMO the
first one is that ahc(4) shouldn't try to identify cards
on LPC(-like) busses and the respective code also only should
be compiled in on architectures where machine actually can
have ISA slots (which currently is only i386 AFAICT).

Marius