Re: [Bug #15124] PCI host bridge windows ignored (works with pci=use_crs)

From: Bjorn Helgaas
Date: Wed Jan 27 2010 - 15:46:12 EST


On Wednesday 27 January 2010 09:53:37 am Jesse Barnes wrote:
> On Wed, 27 Jan 2010 09:45:15 -0700
> Bjorn Helgaas <bjorn.helgaas@xxxxxx> wrote:
>
> > On Tuesday 26 January 2010 03:57:31 pm Yinghai Lu wrote:
> > > [PATCH] x86/pci: don't use ioh resource if only have one ioh
> > >
> > > some system could use reosurce out of IOH resources when only one ioh is there.
> > >
> > > could be BIOS have wrong IOH resources and not enable them.
> >
> > The subtractive decode theory makes sense and would explain what's
> > happening, but I don't like this patch.
> >
> > If we assume that this really is a subtractive decode issue, this
> > patch approaches it the wrong way. We need to know whether a
> > particular host bridge is configured for subtractive decode. This
> > patch tests whether we have more than one host bridge, which is quite
> > a different question.
> >
> > Imagine these system configurations:
> >
> > 1) a single host bridge with subtractive decode
> > 2) a single host bridge with only positive decode
> > 3) multiple host bridges with subtractive decode enabled on one
> > 4) multiple host bridges with only positive decode
> >
> > This patch will break if we encounter configs 2 or 3. In config 2,
> > this patch assumes the bridge performs subtractive decode, so we
> > think the bridge forwards more address space than it actually does.
> > If we try to use that address space, the device will never see the
> > accesses. In config 3, this patch assumes there's no subtractive
> > decode, so we would see Jeff's problem all over again.
>
> Right, but OTOH:
> - multiple IOH has already been tested with the intel_bus.c code
> - we want to move to using _CRS data in these cases instead

> So do you have any objection to applying this patch for 2.6.33 and then
> moving away from intel_bus.c in .34 (assuming we can get _CRS working
> well on the same machines where intel_bus.c was needed)?

Without intel_bus.c, we essentially assume config 1 all the time.
If we keep intel_bus.c and this patch for .33, things should work
for configs 1 and 4. Adding support for config 4 is good.

The bad part is that for config 4, intel_bus.c covers up any defects
in the _CRS or the Linux code that interprets it. The reason Yinghai
added intel_bus.c in the first place was to work around a defect in
this area[1]. Keeping it will make it harder to fix the underlying
issue that keeps us from turning on _CRS for that box.

Bjorn

[1] http://lkml.org/lkml/2009/10/6/371
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/