[PATCH] of: treat PCI config space as IORESOURCE_MEM type

[Liviu Dudau]

On Fri, May 30, 2014 at 05:16:52PM -0600, Bjorn Helgaas wrote:
> On Fri, May 30, 2014 at 5:11 PM, Liviu Dudau <liviu at dudau.co.uk> wrote:
> > On Fri, May 30, 2014 at 03:45:05PM -0500, Kumar Gala wrote:
> >>
> >> On May 29, 2014, at 8:41 PM, Liviu Dudau <liviu at dudau.co.uk> wrote:
> >>
> >> > On Thu, May 29, 2014 at 07:29:31PM -0600, Bjorn Helgaas wrote:
> >> >> On Thu, May 29, 2014 at 6:56 PM, Liviu Dudau <liviu at dudau.co.uk> wrote:
> >> >>> On Thu, May 29, 2014 at 03:51:28PM -0500, Kumar Gala wrote:
> >> >>>>
> >> >>>> On May 29, 2014, at 3:44 PM, Rob Herring <robherring2 at gmail.com> wrote:
> >> >>>>
> >> >>>>> On Thu, May 29, 2014 at 11:03 AM, Kumar Gala <galak at codeaurora.org> wrote:
> >> >>>>>> If we have a PCI config space specified in something like a ranges
> >> >>>>>> property we should treat it as memory type resource.
> >> >>>>>
> >> >>>>> Config space should not be in ranges[1]. We have some cases that are,
> >> >>>>> but we don't want new ones.
> >> >>>>
> >> >>>> For the cases we have I agree, however an ECAM based cfg seems completely legit.
> >> >>>>
> >> >>>>>> Signed-off-by: Kumar Gala <galak at codeaurora.org>
> >> >>>>>> ---
> >> >>>>>> drivers/of/address.c | 3 +++
> >> >>>>>> 1 file changed, 3 insertions(+)
> >> >>>>>>
> >> >>>>>> diff --git a/drivers/of/address.c b/drivers/of/address.c
> >> >>>>>> index cb4242a..4e7ee59 100644
> >> >>>>>> --- a/drivers/of/address.c
> >> >>>>>> +++ b/drivers/of/address.c
> >> >>>>>> @@ -122,6 +122,9 @@ static unsigned int of_bus_pci_get_flags(const __be32 *addr)
> >> >>>>>>       u32 w = be32_to_cpup(addr);
> >> >>>>>>
> >> >>>>>>       switch((w >> 24) & 0x03) {
> >> >>>>>> +       case 0x00: /* cfg space */
> >> >>>>>> +               flags |= IORESOURCE_MEM;
> >> >>>>>> +               break;
> >> >>>>>
> >> >>>>> How would you then distinguish actual memory ranges?
> >> >>>>
> >> >>>> One assumes you are still looking at pci_space as part of of_pci_range
> >> >>>
> >> >>> That doesn't happen when you start scanning the bus. The existing code will
> >> >>> use the IORESOURCE_MEM for allocating memory space for devices, which is
> >> >>> not what you want. Did you test your patch on any PCI system? I'm pretty
> >> >>> sure that with my patch series that tries to make a generic framework for
> >> >>> host controllers this will fail.
> >> >>>
> >> >>> We really need a IORESOURCE_CFG flag for this space.
> >> >>
> >> >> Maybe, but I'm not convinced yet.  The existing IORESOURCE_TYPE_BITS
> >> >> types are for things that are mutually exclusive address spaces.  I
> >> >> think this discussion is about ECAM, where the CPU side is definitely
> >> >> in the same address space (IORESOURCE_MEM) as RAM, APICs, host bridge
> >> >> apertures, device MMIO, etc.  The ECAM area must appear in the
> >> >> iomem_resource tree so we avoid it when allocating other areas.
> >> >
> >> > Agree, I'm only concerned that if this ECAM config space gets added to
> >> > the list of pci_host_bridge windows it will be indistinguishable from
> >> > IORESOURCE_MEM resources and pci_create_root_bus() will add it to the
> >> > bus and allow devices present on that bus to be assigned addresses from
> >> > that range. Which might not be what one wants for certain BARs.
> >> >
> >> > I've had an aborted attempt to parse ECAM ranges in one version of my
> >> > series (granted, I was trying to hack the IORESOURCE_TYPE_BITS as well)
> >> > and things got horribly wrong quickly. I could give this patch a go with
> >> > my series tomorrow when I'm in the office and report back.
> >>
> >> We need to fix the parsing code to be smarter about this case.
> >
> > Wow, what a sweeping statement! Did you not understand that the issue is not
> > the parsing code but the way the rest of the core code uses an IORESOURCE_MEM
> > once you have parsed it into a resource structure and added it to the list
> > of pci_host_bridge_windows?
> 
> Why do you want to add the ECAM area to the list of host bridge
> windows?  My intent was that the windows tell the core what resources
> are available for devices behind the bridge.

I don't *want*, it is just that with my series that enhances Andrew's
parses of the ranges they all come together as host bridge windows. And it is
quite natural to put them all together when creating the root bus as the
space needs to be added to the iomem_resource tree anyway and that will happen
without special casing.

But maybe I'm wrong with that idea. What I know for sure is that wherever you
are going to pass the ECAM range converted to an IORESOURCE_MEM, the existing
code will not be able to distinguish it from a normal IORESOURCE_MEM and it
will treat it as a non-prefetcheable memory area. Is that how we want to treat
the ECFG space or do we want to special case it?

Best regards,
Liviu

> 
> Bjorn
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
> the body of a message to majordomo at vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 

-- 
-------------------
   .oooO
   (   )
    \ (  Oooo.
     \_) (   )
          ) /
         (_/

 One small step
   for me ...