Re: [PATCH v3] PCI/ACPI: xgene: Add ECAM quirk for X-Gene PCIe controller

[Bjorn Helgaas]

On Fri, Dec 02, 2016 at 07:33:46PM -0500, Jon Masters wrote:
> On 12/02/2016 06:39 PM, Bjorn Helgaas wrote:
> > On Thu, Dec 01, 2016 at 11:08:23PM -0500, Jon Masters wrote:
> 
> >> Let's see if I summarized this correctly...
> >>
> >> 1. The MMIO registers for the host bridge itself need to be described
> >>    somewhere, especially if we need to find those in a quirk and poke
> >>    them. Since those registers are very much part of the bridge device,
> >>    it makes sense for them to be in the _CRS for PNP0A08/PNP0A03.
> >>
> >> 2. The address space covering these registers MUST be described as a
> >>    ResourceConsumer in order to avoid accidentally exposing them as
> >>    available for use by downstream devices on the PCI bus.
> >>
> >> 3. The ACPI specification allows for resources of the type "Memory32Fixed".
> >>    This is a macro that doesn't have the notion of a producer or consumer.
> >>    HOWEVER various interpretations seem to be that this could/should
> >>    default to being interpreted as a consumed region.
> > 
> > I agree; I think that per spec, Memory24, Memory32, Memory32Fixed, IO,
> > and FixedIO should all be for consumed resources, not for bridge
> > windows, since they don't have the notion of producer.
> 
> Ok. If we ultimately codify this somewhere as the general Linux kernel
> consensus (Rafael?) then we can also go and get the various ARM server
> specs updated to reflect this in (for e.g.) reference firmware builds.
> 
> > I'm pretty sure there's x86 firmware in the field that uses these for
> > windows, so I think we have to accept that usage, at least on x86.
> 
> Ok. I was pondering how to even go about finding that out, but even if
> I scheduled a job across RH's infra to look, that would be a drop in
> the bucket of possible machines that might be out there doing this.

Hmmm, when researching this, I thought I came across a change
specifically for a machine that used Memory32Fixed this way, but I
can't find it now.

The only thing I did find was some old experiments with Windows that
showed it interpreting a Memory32Fixed region as a window and putting
PCI devices in it: https://bugzilla.kernel.org/show_bug.cgi?id=15817
But that was a synthetic example with qemu, not a real machine in the
field.

> > Even without this patch, I don't think it's a show-stopper to have
> > Linux mistakenly thinking this region is routed to PCI, because the
> > driver does reserve it and the PCI core will never try to use it.
> 
> Ok. So are you happy with pulling in Duc's v4 patch and retaining
> status quo on the bridge resources for 4.10?

Yes, I think it looks good.  I'll finish packaging things up and
repost the current series.

Bjorn