On Wed, 2015-08-19 at 17:16 -0500, Bjorn Helgaas wrote:
> On Tue, Aug 18, 2015 at 12:10:04PM -0700, David Miller wrote:
> > From: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> > Date: Tue, 18 Aug 2015 13:49:43 -0500
> >
> > > [+cc Dave, Eric, Ben, sparclinux]
> > >
> > > On Mon, Aug 17, 2015 at 06:47:58PM +0800, Yijing Wang wrote:
> > > > Commit d0751b98dfa3 ("PCI: Add dev->has_secondary_link to
> > > > track downstream PCIe links")assumed root port is always
> > > > the top device in pcie tree. But on sparc64 V245 and T2000tthe
> > > > pcie tree has no root port device in top level, the
> > > > upstream port device is connected directly to root bus.
> > > > So we may get NULL parent for this upstream port device.
Picking up that thread half way through...
> > > > Upstream port ------ Downstream port ------PCIe-PCI bridge
> > >
> > > This is an unusual, possibly even illegal, PCIe topology because
> > > it lacks a Root Port at the top of the hierarchy. From lspci
> > > output
> > > [1]
> > > collected by Meelis, the top-level devices are:
Nobody cares what is "illegal", what matters is what HW actually does :
-) Specs only matter as far as they get followed. In any case, it also
happens on powerpc, when running under a hypervisor such as IBM
PowerVM, PCIe devices appear directly under a virtual host bridge which
is not exposed as a PCIe root complex.
> > The root port is in the PCI controller on sparc64, and PCI
> > controllers
> > don't have actual real PCI configuration space available.
> >
> > I used to put dummy PCI bus nodes into the tree and do provide
> > dummy
> > software PCI config space methods, but that caused more harm than
> > good.
>
> I guess you're talking about what you removed with c26d3c013897
> ("sparc64: Stop creating dummy root PCI host controller devices."),
> which talks about duplicate sysfs and procfs nodes. Removing the
> dummy devices avoids the duplicate node problem, but it doesn't fix
> the root cause, so it's probably not the only possible resolution.
>
> We'll probably have to do something ugly like Yijing's patch just
> because v4.2 is imminent and I don't want it to be broken.
>
> I don't like it because we have these PCIe links where we only know
> about devices on one end of the link, and then we can't manage
> ASPM/MPS/etc. But we've had this situation for several years, so
> I guess nobody cares too much about those things on the affected
> mahines.
Correct, we have this situation on more than just sparc64 and we
shouldn't crash. Just don't touch ASPM/MPS/etc... in those cases.
Cheers,
Ben.
--
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
