Re: [PATCH RFC v2 1/2] drivers: pci: fix pci_mmap_fits() implementation for procfs mmap

[Bjorn Helgaas]

On Tue, Nov 11, 2014 at 4:48 AM, Lorenzo Pieralisi
<lorenzo.pieralisi@xxxxxxx> wrote:
> On Mon, Nov 10, 2014 at 11:04:54PM +0000, Bjorn Helgaas wrote:
>> ...
>> Here's what I think I understand so far:
>>
>>   Applications can mmap PCI memory space via either sysfs or procfs (the
>>   procfs method is deprecated but still supported):
>>
>>     - In sysfs, there's a separate /sys/devices/pci*/.../resource* file
>>       for each device BAR, and the application opens the appropriate
>>       file and supplies the offset from the beginning of the BAR as the
>>       mmap(2) offset.
>>
>>     - In procfs, the application opens the single /proc/bus/pci/... file
>>       for the device.  On most platforms, it supplies the CPU physical
>>       address as the mmap(2) offset.  On a few platforms, such as SPARC,
>>       it supplies the bus address, i.e., a BAR value, instead.
>>
>> But I'm not sure I have this right.  If the procfs offset is either the
>> CPU physical address or the BAR value, then pci_resource_to_user()
>> should be (depending on the arch) either a no-op or use
>> pci_resource_to_bus().
>
> Exactly (pcibios_resource_to_bus() ?).
>
>> But that's not how it's implemented.  Maybe it *could* be?  If
>> pci_resource_to_user() gives you something that's not a CPU physical
>> address and not a bus address, what *does* it give you, and why would we
>> need this third kind of thing?
>
> Well, you need a per arch function implementation where to define if
> the conversion from CPU physical address to PCI bus should take place
> or not right ? As you mentioned above, if that should be a per-arch
> decision, there has to be a per-arch function to filter the resource
> in question, I guess that's my understanding behind pci_resource_to_user(),
> but I am not sure either, and understanding that was the primary reason
> for this patchset so comments are welcome.

I agree that we need pci_resource_to_user() because arches do
different things, so we can't just remove pci_resource_to_user() and
replace it with pci_resource_to_bus().  My point is that we have a
generic pci_resource_to_user() implementation that does nothing, and
if an arch *does* implement its own pci_resource_to_user(), it seems
like it should simply call pci_resource_to_user().

Bjorn
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/