Re: [PATCH 0/4] PCI legacy I/O port free driver (take4)

[Kenji Kaneshige <kaneshige.kenji@jp.fujitsu.com>]

Adam Belay wrote:
> On Thu, Mar 02, 2006 at 07:34:41PM +0000, Russell King wrote:
> 
>>On Thu, Mar 02, 2006 at 10:24:36AM -0700, Grant Grundler wrote:
>>
>>>On Thu, Mar 02, 2006 at 03:50:57PM +0000, Russell King wrote:
>>>
>>>>I've been wondering whether this "no_ioport" flag is the correct approach,
>>>>or whether it's adding to complexity when it isn't really required.
>>>
>>>I think it's the simplest solution to allowing a driver
>>>to indicate which resources it wants to use. It solves
>>>the problem of I/O Port resource allocation sufficiently
>>>well.
>>
>>I have another question (brought up by someone working on a series of
>>ARM machines which make heavy use of MMIO.)
>>
>>Why isn't pci_enable_device_bars() sufficient - why do we have to
>>have another interface to say "we don't want BARs XXX" ?
>>
>>Let's say that we have a device driver which does this sequence (with,
>>of course, error checking):
>>
>>	pci_enable_device_bars(dev, 1<<1);
>>	pci_request_regions(dev);
>>
>>(a) should PCI remember that only BAR 1 has been requested to be enabled,
>>    and as such shouldn't pci_request_regions() ignore BAR 0?
>>
>>(b) should the PCI driver pass into pci_request_regions() (or even
>>    pci_request_regions_bars()) a bitmask of the BARs it wants to have
>>    requested, and similarly for pci_release_regions().
>>
>>Basically, if BAR0 hasn't been enabled, has pci_request_regions() got
>>any business requesting it from the resource tree?
> 
> 
> I understand the point you're making, but I think this misrepresents what
> is actually happening.  From my understanding of the spec, it's not possible
> to disable individual bars (with the exception of the expansion ROM).  Rather
> there is one bit for IO enable and one bit for IOMMU enable.  Therefore, we
> can enable or disable all I/O ports, but there's really no in between.  If
> the device uses even one I/O port, it's still a huge loss because of the
> potential bridge window dependency.  Also, if a device has several I/O ports
> but the driver only wants to use one, all of the others must still be
> assigned.
> 

I see. I think you are right.

In addition to the fact that there is one bit for IO enable and one
bit for MMIO enable, I think we should not enable I/O port (or MMIO)
of the device if not all the I/O port (or MMIO) regions are assigned
to the device because we must build a consistent address mapping
before enabling it.

It seems that using pci_enable_device_bars() is not a good idea.
If there is no objection, I'll design and implement take6 again.

Any comments?

Thanks,
Kenji Kaneshige