Re: device not available because of BAR 0 collisions

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Thu, 2011-04-28 at 13:29 -0400, Steven A. Falco wrote:
> On 04/27/2011 03:51 PM, Steven A. Falco wrote:
> > On 04/26/2011 07:39 PM, Benjamin Herrenschmidt wrote:
> >> On Tue, 2011-04-26 at 09:38 -0400, Steven A. Falco wrote:
> >>> On 04/25/2011 08:01 PM, Benjamin Herrenschmidt wrote:
> >>>> On Mon, 2011-04-25 at 16:10 -0400, Steven A. Falco wrote:
> >>>>> I'm getting an error message when trying to talk to some custom
> >>>>> hardware:
> >>>>>
> >>>>> dx83xx 0001:43:00.0: device not available because of BAR 0
> >>>>> [0xa1000000-0xa1ffffff] collisions
> 
> I believe I've gotten to the root cause of this issue.  Turns out the
> ASIC is reporting a class code of 0000, in violation of the PCI spec.
> 
> As a result, Linux is refusing to set up the ASIC, and so it remains
> with the settings made in U-Boot, which are in conflict with the
> allocation Linux makes for the bridge the ASIC is connected to.
> 
> So Ben, you are absolutely right that there was some left-over
> configuration.

Hrm. that's odd still, do you know where in linux is the code that tests
for the class code and decides to not setup the device resources ? I
can't see that...

> My choices appear to be:
> 
> 1) Fix the ASIC (yeah, right)
> 
> 2) Force Linux to use the U-Boot settings
> 
> 3) Hack Linux to set up a device with a bogus class.
> 
> I'm not sure why this hardware works in x86 - I guess it is less
> fussy.

x86 probably just re-uses whatever setting the BIOS does, but I'm still
a bit surprised by your class code story.

I supose you can do a PCI header quirk that overrides the class code in
struct pci_dev. Something like:

static void __devinit quirk_your_asic_class(struct pci_dev *dev)
{
	dev->class = foobar;
}
DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_xxx, PCI_DEVICE_ID_yyy, quirk_your_asic_class);

But I'd like to figure out where that is tested bcs I haven't found so
far...

> 	Steve
> 
> >>>>>
> >>>>> I see in setup-res.c that this message comes out when there is no
> >>>>> parent for
> >>>>> a device resource.
> >>>>
> >>>>  .../...
> >>>>
> >>>> It mostly happens in arch/powerpc/kernel/pci-common.c and the generic
> >>>> setup-res.c
> >>>>
> >>>> Try #define DEBUG at the top (before the #includes) of pci-common.c and
> >>>> pci_32.c (remove the exiting #undef in the last one) and send us the
> >>>> full dmesg log, along with the output of cat /proc/iomem
> >>
> >> Have you set any specific flags ? IE. Modified the value of
> >> ppc_pci_flags from what the 4xx code sets originally ?
> > 
> > For fun, I just tried changing:
> > 
> > ppc_pci_set_flags(PPC_PCI_REASSIGN_ALL_RSRC);
> > 
> > to:
> > 
> > ppc_pci_set_flags(PPC_PCI_PROBE_ONLY);
> > 
> > I realize that is the exact opposite of what you were suggesting, but
> > please bear with me for a bit.
> > 
> > I also changed the PCIE 0 ranges from:
> > 
> > ranges = <0x02000000 0x00000000 0x80000000 0x90000000 0x00000000 0x10000000
> >           0x01000000 0x00000000 0x00000000 0xe8010000 0x00000000 0x00010000>;
> > 
> > ranges = <0x02000000 0x00000000 0x90000000 0x90000000 0x00000000 0x10000000
> >           0x01000000 0x00000000 0x00000000 0xe8010000 0x00000000 0x00010000>;
> > 
> > I changed the ranges not because I wanted a 1:1 map, but because 90000000 is
> > what U-Boot chooses when it scans PCIe 1.
> > 
> > At this point, everything is working.  Here is /proc/iomap:
> > 
> > 90000000-9fffffff : /plb/pciex@0c0000000
> >   90000000-94ffffff : PCI Bus 0001:41
> >     90000000-9001ffff : 0001:41:00.0
> >     90100000-94ffffff : PCI Bus 0001:42
> >       90100000-92ffffff : PCI Bus 0001:43
> >         91000000-91ffffff : 0001:43:00.0  //<--- was missing before
> >         92000000-92ffffff : 0001:43:00.0  //<--- was missing before
> >       93000000-94ffffff : PCI Bus 0001:44
> >         93000000-93ffffff : 0001:44:00.0  //<--- was missing before
> >         94000000-94ffffff : 0001:44:00.0  //<--- was missing before
> > e0000000-e7ffffff : /plb/pciex@0a0000000
> >   e0000000-e7ffffff : PCI Bus 0000:01
> >     e0000000-e00fffff : 0000:01:00.0
> >     e0100000-e01fffff : 0000:01:00.0
> >     e4000000-e7ffffff : 0000:01:00.0
> > ef600200-ef600207 : serial
> > ef600300-ef600307 : serial
> > ef600600-ef600606 : spi_ppc4xx_of
> > ef6c0000-ef6cffff : dwc_otg.0
> >   ef6c0000-ef6cffff : dwc_otg
> > fc000000-ffffffff : fc000000.nor_flash
> > 
> > Now I see the bars for the ASICs (flagged above).  I could stop here,
> > and declare success, but I don't really like this solution, because it
> > requires me to be sure the dts has the same bus addresses that U-Boot
> > will choose.  Seems risky.
> > 
> > Tentative conclusion:  Either I still have something set wrong in my dts
> > or there is a bug in the Linux PCI bus mapping code.
> > 
> > 	Steve
> > 
> >>
> >> It does look to me like some of your device BARs have been setup already
> >> by the firmware in a way that conflict with the way you configure your
> >> ranges, and the kernel doesn't appear to detect nor try to remap that
> >> which would happen if you have the "probe only" flag set.
> >>
> >> IE. On your c0000000 bus, you have memory at 90000000 CPU space mapped
> >> to 80000000 PCI space. However, when probing, the kernel finds:
> >>
> >> pci 0001:41:00.0: reg 10 32bit mmio: [0x90000000-0x9001ffff]
> >>
> >> IE. A BAR was already set with a value of 90000000 PCI-side which is out
> >> of the bounds you have for your bus.
> >>
> >> Maybe you really want to configure that second bus to have CPU 90000000
> >> mapped to 90000000 PCI-side ? (IE. a 1:1 mapping). That would be
> >> something to fix in your "ranges" property.
> >>
> >> Cheers,
> >> Ben.
> > 
> > 
> 
>