Re: [Qemu-devel] [PATCH 4/4] pci: use pci_config_header in pci.c

[Anthony Liguori <anthony@codemonkey.ws>]

Gerd Hoffmann wrote:
> Anthony Liguori wrote:
>   
>> Because the config structure is loaded after pci_register device and
>> because the MMIO callbacks into the config space?  If you look at
>> pci_default_read_config(), it actually does conversion from le to host
>> CPU because that's what MMIO callbacks expect.
>>     
>
> Yep.
>
>   
>> If you simply removed
>> this, it would work.
>>     
>
> Ah, *now* I see what you mean.
>
> Well.  That assumes the guests always use the correct in{b,w,l}
> instruction to access the config space values, i.e. never ever try to
> use two inb reads for a 16bit value for example.  Or a inw for the lower
> 16 bits of a 32bit value because they know the high bits are zero anyway.
>
> I have my doubts that this is a sane expectation.
>   

The config space seems organized by data size so you could always just 
take care of this in the accessors.  That is, if you get a byte access 
to a 32-bit value, you would do something like:

(cpu_to_le32(d->config + (addr & ~3)) >> (8 * (addr % 4))) & 0xFF;

Not the prettiest thing in the world, but it would work.

>> Of course, you'll have to audit each caller of
>> pci_register_device() and make sure their config's get changed too.
>>     
>
> Sure, that would be a *big* task & patch.
>   

Actually, it's not.  I did a quick grep, noone seems to pass in a 
config_read accessor.  Only config_write is ever passed and it only 
seems to be to used to add callbacks when particular config values are 
written.  pci_default_config_write is still used to do the actual write.

Regards,

Anthony Liguori

> cheers,
>   Gerd
>
>
>