Re: [PATCH v5 3/3] pci, pci-thunder-ecam: Add driver for ThunderX-pass1 on-chip devices

From: Bjorn Helgaas
Date: Tue Feb 09 2016 - 11:26:40 EST


On Tue, Feb 09, 2016 at 10:25:33AM +0100, Arnd Bergmann wrote:
> On Monday 08 February 2016 17:24:30 Bjorn Helgaas wrote:
> > > >
> > > >I assume your system conforms to expectations like these; I'm just
> > > >pointing them out because you mentioned buses with multiple devices on
> > > >them, which is definitely something one doesn't expect in PCIe.
> > >
> > > The topology we have is currently working with the kernel's core PCI
> > > code. I don't really want to get into discussing what the
> > > definition of PCIe is. We have multiple devices (more than 32) on a
> > > single bus, and they have PCI Express and ARI Capabilities. Is that
> > > PCIe? I don't know.
> >
> > I don't need to know the details of your topology. As long as it
> > conforms to the PCIe spec, it should be fine. If it *doesn't* conform
> > to the spec, but things currently seem to work, that's less fine,
> > because a future Linux change is liable to break something for you.
> >
> > I was a little concerned about your statement that "there are multiple
> > devices residing on each bus, so from that point of view it cannot be
> > PCIe." That made it sound like you're doing something outside the
> > spec. If you're just using regular multi-function devices or ARI,
> > then I don't see any issue (or any reason to say it can't be PCIe).
>
> It doesn't conform to the PCIe port spec, because there are no external
> ports but just integrated devices in the host bridge.

Is there a spec section you have in mind? Based on sec 1.3.1, I don't
think there's a requirement to have PCI Express Ports (is that what
you mean by "external ports"?)

Root Complex Integrated Endpoints (sec 1.3.2.3) are clearly supported
and they would not be behind a Root Port. If you're using those, I
hope they're correctly identified via the PCIe capability Device/Port
Type (sec 7.8.2) because we rely on that type to figure out whether
the link-related registers are implemented.

The spec does include rules related to peer-to-peer transactions, MPS,
ASPM, error reporting, etc., and Linux relies on those, so I think it
would be important to get those right.

> For this special
> case, I don't think it matters at all from the point of view of the DT
> binding whether we call the node name "pci" or "pcie".

And the PCI core doesn't even know the node name, it doesn't matter
there either.

Bjorn