PCI BAR initialisation..?

PCI BAR initialisation..?

[ashish anand]

I need a little explanation for following from pci experts.
As per the pci specs , to determine the resource requirement by  BAR in pci config space
we write 0xffffffff and then read back , and from the data read we decide the size and
alignment requirement for pci address in either pci i/o or memory spaces.

In case the firmware hasn't initialised the pci devices BAR in slot we have to handle
it  at our own as I have to do for sandpoint + mcp750 with DINK32 as firmware.

But by any chance if somebody has come across a random feature of pci BAR.
that is if you write the required pci addresses in BAR after writing 0xffffffff and
reading it back , and to confirm if the actual value has reached the BAR , we read it again
but then it returns funny values.

more surprisingly if i move the code in some other function then everything behaves correctly.
I am more then sure that nothing relevant is happening betwen two places.

my sequence is like this,.

1.pcibios_write_config_dword(bus,devfn, PCI_BASE_ADDRESS_1, 0xffffffff);
2.pcibios_read_config_dword( bus,devfn, PCI_BASE_ADDRESS_1, &add);

3.now I decode the value of add to determine the size and alignment requirement
  let us say decoded value is "ADD"

4.pcibios_write_config_dword(bus,devfn, PCI_BASE_ADDRESS_1,ADD);

5.now confirming the bar value by,
    pcibios_read_config_dword(..., ..., base address register , &add);

but value of add is not the actual one written.

sometimes this behaviour gets automatically corrected and then i can read the actual value
written.

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: PCI BAR initialisation..?

[ashish anand]

I have mpc107 pci-host bridge.
motherboard is sandpoint + mcp750 processor.
pci_config_access method is indirect and config_address is 0xfec00000
config_data is 0xfee00000 rather than 0xcfc and 0xcf8.

i wonder if it would have been a case of memory reordering in that case
In bios how can we always believe on resource assignment of devices.
does it just a matter of chance that when we assign through bios this
reordering issue doesn't affect us..



Paul Mackerras wrote:
>
> ashish anand writes:
>
> > But by any chance if somebody has come across a random feature of pci BAR.
> > that is if you write the required pci addresses in BAR after writing 0xffffffff and
> > reading it back , and to confirm if the actual value has reached the BAR , we read it again
> > but then it returns funny values.
> >
> > more surprisingly if i move the code in some other function then everything behaves correctly.
> > I am more then sure that nothing relevant is happening betwen two places.
>
> Hmmm, that sounds like maybe the config space reads and writes are
> getting reordered, which would be a Bad Thing.
>
> What sort of PCI host bridge does your system have?
>
> If you place an mb() call after each pci config read or write, does
> that fix the problem?
>
> Paul.

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: PCI BAR initialisation..?

[Paul Mackerras]

ashish anand writes:

> But by any chance if somebody has come across a random feature of pci BAR.
> that is if you write the required pci addresses in BAR after writing 0xffffffff and
> reading it back , and to confirm if the actual value has reached the BAR , we read it again
> but then it returns funny values.
>
> more surprisingly if i move the code in some other function then everything behaves correctly.
> I am more then sure that nothing relevant is happening betwen two places.

Hmmm, that sounds like maybe the config space reads and writes are
getting reordered, which would be a Bad Thing.

What sort of PCI host bridge does your system have?

If you place an mb() call after each pci config read or write, does
that fix the problem?

Paul.

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/