Re: [Qemu-devel] [Xen-devel] [XEN][RFC PATCH 03/15] hvm-pci: Handle PCI config space in Xen

[Jan Beulich]

>>> On 26.03.12 at 14:20, Julien Grall <address@hidden> wrote:
> On 03/23/2012 08:29 AM, Jan Beulich wrote:
>> Is there a reasonably low enforced boundary on the number
>> of devices? Otherwise, a linear lookup would seem overly
>> simple to me.
>>    
> The maximum of bdf is 2^16 => 65536.
> Which kind of structure do you advice ? Array ? Hash Table ?

Radix tree, especially if you fold in the segment number.

>> Further, with how PCI_CMP_BDF() is defined, you're doing the
>> wrong thing here anyway - bit 31 is required to be set for the
>> port CFC access to be a config space one. Plus there's an AMD
>> extension to this interface, so I think other than shifting out
>> the low 8 bits and checking that the high bit is set, you shouldn't
>> do any other masking here.
>>    
> Actually in config address register the 24-30 bits are reserved.
> So, do I need to mask it ?

Not necessarily - I'd suggest considering the part of the address
(which should generally result in a mismatch on any comparison. This
so that this ...

> Moreover what is the AMD extension ?

... can work without additional code. For an implementation, please
have a look at current Linux'es arch/x86/pci/direct.c - bits 24...27
are used for extended config space register accesses (which will be
needed for advanced PCIe or PCI-X functionality, and which may
be particularly important as long as we don't emulate MMCFG - at
least I don't think we do).

Jan