On Fri, Nov 07, 2014 at 02:00:56PM +0000, Rob Herring wrote:
> On Fri, Nov 7, 2014 at 4:17 AM, Liviu Dudau <Liviu.Dudau@xxxxxxx> wrote:
> > On Thu, Nov 06, 2014 at 07:46:33PM +0000, Rob Herring wrote:
> >> On Thu, Nov 6, 2014 at 9:30 AM, Liviu Dudau <Liviu.Dudau@xxxxxxx> wrote:
> >> > On Thu, Nov 06, 2014 at 02:57:35PM +0000, Rob Herring wrote:
> >> >> On Thu, Nov 6, 2014 at 4:05 AM, Liviu Dudau <Liviu.Dudau@xxxxxxx> wrote:
> >> >> > On Wed, Nov 05, 2014 at 11:17:43PM +0000, Bjorn Helgaas wrote:
> >> >> >> On Tue, Nov 04, 2014 at 12:47:40PM +0100, Lucas Stach wrote:
> >> >> >> > This property was added by 41e5c0f81d3e
> >> >> >> > (of/pci: Add pci_get_new_domain_nr() and of_get_pci_domain_nr())
> >> >> >> > without the required binding documentation. As this property
> >> >> >> > will be supported by a number of host bridge drivers going forward,
> >> >> >> > add it to the common PCI binding doc.
> >> >> >> >
> >> >> >> > Signed-off-by: Lucas Stach <l.stach@xxxxxxxxxxxxxx>
> >> >> >>
> >> >> >> I merged 41e5c0f81d3e through my tree, and I could merge something like
> >> >> >> this if a consensus develops with some acks. But I'll just let you guys
> >> >> >> handle it unless you poke me again.
> >> >> >
> >> >> > While I think the "linux,pci-domain" property *must* be documented, I
> >> >> > would like to get a consensus first on the usage. If we agree that
> >> >> > the property is mandatory to all host bridge drivers that use OF then
> >> >> > we need to patch existing drivers (partially done through Lorenzo's
> >> >> > patches, but other arches are ignoring it). If we say all *new* drivers
> >> >> > need to use it then we also need to come up with a strategy on how to
> >> >> > deal with old vs new school drivers.
> >> >> >
> >> >> > My preferred approach is the 3rd way: "linux,pci-domain" becomes part of
> >> >> > the core PCI infrastructure (and we find the common ground with ACPI).
> >> >> > That way the host bridge drivers don't have to do anything, but the DT
> >> >> > creators have to specify a value.
> >> >>
> >> >> I'm okay with it being in the core. It was the mixture of using the
> >> >> property and automatic numbering that I had issues with. Any mixture
> >> >> whether in DT or in drivers should be an error. Also, I think having a
> >> >> mixture of root bus host drivers would be rare, so I'm not too
> >> >> concerned about some drivers supporting the property and others not.
> >> >> In any case, these issues are all with the kernel and not really the
> >> >> concern for the binding. For the binding, simply all hosts set the
> >> >> domain or none of them do.
> >> >
> >> > Repeating what you've said to verify my understanding: you are OK with
> >> > the "linux,pci-domain" being handled in the PCI framework and mandatory
> >> > to all OF-based host bridges and architectures. Failure to include
> >> > the property should be an error and no host bridge driver should default
> >> > to the auto-generation of domain numbers.
> >> >
> >> > Is that correct?
> >>
> >> Not exactly. It is only mandatory when you have multiple root buses.
> >> But we can't say an existing dtb is in error, so it has to remain
> >> optional for compatibility. Also, given it is a Linux property, you
> >> can't really say it is mandatory for all PCI bindings from a DT
> >> perspective.
> >>
> >> While we could have issues in theory if this is handled in the
> >> drivers, I don't think we will in practice as having root buses with
> >> different drivers is unlikely.
> >
> > That's correct. However, I would like to bring to you attention that as
> > long as the property is treated as optional in certain cases we will need,
> > in the framework code, to mix the assignment of a domain number coming from
> > parsing of the property with the auto-numbering scheme in order to support
> > old DT files. And that was one of your major concerns when reviewing the
> > series. Any suggestions or clarifications?
>
> We have to support both allocation schemes, but not on the same
> system. A mixture on a given system is an error. So the cases to
> handle are like this:
>
> 1 root bus w/ domain -> use domain prop
> 1 root bus w/o domain -> auto numbering
> N root buses w/ domains -> use domain prop
> N root buses w/o domains -> auto numbering
> N root buses w/ and w/o domains -> error
Agree.
>
> I also had issues just around the implementation details, but we can
> discuss those when there is another version.
Well, there is a version already in arch/arm64/kernel/pci.c:
pci_bus_assign_domain_nr() and I believe it works on all the cases
listed above. Question is if you agree on making this the default
implementation for OF-based architectures. Lorenzo has already
copied it into arch/arm.
Best regards,
Liviu
>
> Rob
>
--
====================
| I would like to |
| fix the world, |
| but they're not |
| giving me the |
\ source code! /
---------------
¯\_(ツ)_/¯
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
