Re: interrupt locality for NUMA

[Jesse Barnes <jbarnes@engr.sgi.com>]

On Friday, August 13, 2004 8:41 am, Alex Williamson wrote:
> On Fri, 2004-08-13 at 08:32 -0700, Jesse Barnes wrote:
> > On Thursday, August 12, 2004 7:58 pm, Alex Williamson wrote:
> > >    This probably isn't ready for inclusion yet, but I wanted to see if
> > > anybody else could make use of it.  This works on HP sx1000 boxes setup
> > > for NUMA and I think it's ACPI namespace does the right thing.  All
> > > this does is walk through namespace looking for devices with an _MAT
> > > method that returns an IOSAPIC and also has a _PXM method to tell us
> > > the proximity domain where it lives.  The node data gets stored in the
> > > iosapic data structure because doing this lookup is pretty slow.  Does
> > > this jive with what other ACPI NUMA boxes are exporting in namespace?
> > > I'm hoping everyone will put the _PXM on the same device as the _MAT,
> > > but I'm wondering if I need to add support for looking on parent
> > > objects.  Thoughts?  Thanks,
> >
> > Matt Dobson is working on a pci_to_nodemask, might that be used instead? 
> > If we did it that way, we could put it in generic code, dependent on
> > CONFIG_NUMA or something, rather than keeping it ACPI specific.
>
>    We certainly need a pci to node mapping, but I'm not sure we want to
> use it for interrupt routing.  For instance, how would a non-pci serial
> port get assigned to the right node?  Not that this example is terribly
> important, but not everything is pci.

Yeah, that's true, maybe there's a convenient place to put the info in the 
device tree?  If not, that's fine, the ACPI approach is ok, we just won't be 
able to use it (we've already got our own code for this anyway).

Jesse