PCIBIOS_MIN_IO/MEM problem

PCIBIOS_MIN_IO/MEM problem

[Benjamin Herrenschmidt]

Hi !

I have a problem on platforms with multiple PCI domains and allocating
resources.

The mechanism of having the arch set PCIBIOS_MIN_IO and PCIBIOS_MIN_MEM
to avoid allocating things over the legacy areas among other things
doesn't work well for us.

The reason is that on platforms where resource addresses are remapped
because the PCI bus isn't mapped 1:1, the values of the ranges we try
to avoid (typically 0...4K for IO space and legacy VGA memory) is
different from one domain to the next.

Right now, I'm doing a workaround for .28 that will reserve those
regions on some platforms before the allocation occurs, but it's a bit
fishy, especially it gives strange results when using things like legacy
IDE PCI cards that really want to allocate in those areas.

Ideally, we would need to pass the pci_bus pointer to those macros.

Thus, is there any objection to me replacing those accross the board
with something like that ?

	resource_size_t pcibios_min_assign_io(struct pci_bus *bus)
	resource_size_t pcibios_min_assign_mem(struct pci_bus *bus)

Please holler if you have objections before I start fixing everybody :-)

Cheers,
Ben.

Re: PCIBIOS_MIN_IO/MEM problem

[Jesse Barnes]

On Monday, November 10, 2008 8:56 pm Benjamin Herrenschmidt wrote:
> Hi !
>
> I have a problem on platforms with multiple PCI domains and allocating
> resources.
>
> The mechanism of having the arch set PCIBIOS_MIN_IO and PCIBIOS_MIN_MEM
> to avoid allocating things over the legacy areas among other things
> doesn't work well for us.
>
> The reason is that on platforms where resource addresses are remapped
> because the PCI bus isn't mapped 1:1, the values of the ranges we try
> to avoid (typically 0...4K for IO space and legacy VGA memory) is
> different from one domain to the next.
>
> Right now, I'm doing a workaround for .28 that will reserve those
> regions on some platforms before the allocation occurs, but it's a bit
> fishy, especially it gives strange results when using things like legacy
> IDE PCI cards that really want to allocate in those areas.
>
> Ideally, we would need to pass the pci_bus pointer to those macros.
>
> Thus, is there any objection to me replacing those accross the board
> with something like that ?
>
> 	resource_size_t pcibios_min_assign_io(struct pci_bus *bus)
> 	resource_size_t pcibios_min_assign_mem(struct pci_bus *bus)
>
> Please holler if you have objections before I start fixing everybody :-)

Ah yeah, having per-bus ranges makes sense.  No objections here.

Thanks,
Jesse

Re: PCIBIOS_MIN_IO/MEM problem

[Jesse Barnes]

On Monday, November 10, 2008 8:56 pm Benjamin Herrenschmidt wrote:
> Hi !
>
> I have a problem on platforms with multiple PCI domains and allocating
> resources.
>
> The mechanism of having the arch set PCIBIOS_MIN_IO and PCIBIOS_MIN_MEM
> to avoid allocating things over the legacy areas among other things
> doesn't work well for us.
>
> The reason is that on platforms where resource addresses are remapped
> because the PCI bus isn't mapped 1:1, the values of the ranges we try
> to avoid (typically 0...4K for IO space and legacy VGA memory) is
> different from one domain to the next.
>
> Right now, I'm doing a workaround for .28 that will reserve those
> regions on some platforms before the allocation occurs, but it's a bit
> fishy, especially it gives strange results when using things like legacy
> IDE PCI cards that really want to allocate in those areas.
>
> Ideally, we would need to pass the pci_bus pointer to those macros.
>
> Thus, is there any objection to me replacing those accross the board
> with something like that ?
>
> 	resource_size_t pcibios_min_assign_io(struct pci_bus *bus)
> 	resource_size_t pcibios_min_assign_mem(struct pci_bus *bus)
>
> Please holler if you have objections before I start fixing everybody :-)

Ah yeah, having per-bus ranges makes sense.  No objections here.

Thanks,
Jesse