Re: [PATCH 5/6] PCI: acpiphp: look _RMV method a bit deeper in the hierarhcy

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Tuesday, July 02, 2013 11:09:19 AM Bjorn Helgaas wrote:
> On Tue, Jul 2, 2013 at 4:44 AM, Kirill A. Shutemov
> <kirill.shutemov@linux.intel.com> wrote:
> > Mika Westerberg wrote:
> >> The acpiphp driver finds out whether the device is hotpluggable by checking
> >> whether it has _RMV method behind it (and if it returns 1). However, at
> >> least Acer Aspire S5 with Thunderbolt host router has this method placed
> >> behind device called EPUP (endpoint upstream port?) and not directly behind
> >> the root port as can be seen from the ASL code below:
> >>
> >> Device (RP05)
> >> {
> >>       ...
> >>       Device (HRUP)
> >>       {
> >>               Name (_ADR, Zero)
> >>               Name (_PRW, Package (0x02)
> >>               {
> >>                       0x09,
> >>                       0x04
> >>               })
> >>               Device (HRDN)
> >>               {
> >>                       Name (_ADR, 0x00040000)
> >>                       Name (_PRW, Package (0x02)
> >>                       {
> >>                               0x09,
> >>                               0x04
> >>                       })
> >>                       Device (EPUP)
> >>                       {
> >>                               Name (_ADR, Zero)
> >>                               Method (_RMV, 0, NotSerialized)
> >>                               {
> >>                                       Return (One)
> >>                               }
> >>                       }
> >>               }
> >>       }
> >>
> >> If we want to support such machines we must look for the _RMV method a bit
> >> deeper in the hierarchy. Fix this by changing pcihp_is_ejectable() to check
> >> few more devices down from the root port.
> >
> > We found that this approach is broken. We've got false positive: host
> > bridge itself was detected as hotplugable slot %) I think it's not
> > acceptable.
> >
> > Mika has tried few more approaches, but we haven't found anything better
> > then hardcoded path like in original workaround patch[1]. It's not generic
> > at all, but safe from false positives.
> >
> > Any thoughts?
> >
> > [1] http://article.gmane.org/gmane.linux.kernel.pci/19102
> 
> The changelog of this patch ([1]) says "Correct ACPI PCI hotplug
> implementation should have _RMV method in a PCI slot (device under pci
> bridge). In Acer Aspire S5 case we have it deeper in hierarchy ..."
> 
> This is exactly what I mean about acpiphp being brittle because of the
> assumptions it makes about the ACPI namespace.  Is there actually
> something in the spec that requires the _RMV method to be where
> pcihp_is_ejectable() expects it?
> 
> I'm not 100% dead-set against merging the workaround with hard-coded
> path, but I still don't think it's a good idea.  It "fixes" it for one
> particular machine, but there will likely be other machines that
> require similar fixes in the future.  It makes it harder for somebody
> to clean up the design later, because that person won't have an Aspire
> S5 to test.  It makes it less likely that somebody *will* clean it up
> later, because "everything is already working."
> 
> That's why my preference (given infinite resources) would be to rework
> acpiphp now, while people are interested and can test it.

Well, I have a plan to do just that, but quite frankly I'd prefer to make
things physically work first and *then* to make them nicer.

> I'm sure this would be a major redesign of acpiphp and its interaction with
> pciehp, but I think it's something we're going to have to do at some
> point.

Yes, we need to do that pretty much as soon as reasonably possible, but let's
not hold back changes allowing people to use their hardware with Linux because
of that.  We very well may need to spend a couple of releases on it.

Thanks,
Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.