Re: Subject: [PATCH 1/1] PCI: of: avoid warning for 4 GiB non-prefetchable

[Bjorn Helgaas <helgaas@kernel.org>]

On Fri, Feb 28, 2025 at 05:01:51PM -0600, Rob Herring wrote:
> On Fri, Feb 28, 2025 at 12:27 PM Bjorn Helgaas <helgaas@kernel.org> wrote:
> > On Thu, Nov 14, 2024 at 02:05:08PM +0000, Wannes Bouwen (Nokia) wrote:
> > > Subject: [PATCH 1/1] PCI: of: avoid warning for 4 GiB non-prefetchable
> > > windows.
> > >
> > > According to the PCIe spec, non-prefetchable memory supports only 32-bit
> > > BAR registers and are hence limited to 4 GiB. In the kernel there is a
> > > check that prints a warning if a non-prefetchable resource exceeds the
> > > 32-bit limit.
> > >
> > > This check however prints a warning when a 4 GiB window on the host
> > > bridge is used. This is perfectly possible according to the PCIe spec,
> > > so in my opinion the warning is a bit too strict. This changeset
> > > subtracts 1 from the resource_size to avoid printing a warning in the
> > > case of a 4 GiB non-prefetchable window.
> > >
> > > Signed-off-by: Wannes Bouwen <wannes.bouwen@nokia.com>
> > > ---
> > >  drivers/pci/of.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/pci/of.c b/drivers/pci/of.c
> > > index dacea3fc5128..ccbb1f1c2212 100644
> > > --- a/drivers/pci/of.c
> > > +++ b/drivers/pci/of.c
> > > @@ -622,7 +622,7 @@ static int pci_parse_request_of_pci_ranges(struct device *dev,
> > >             res_valid |= !(res->flags & IORESOURCE_PREFETCH);
> > >
> > >             if (!(res->flags & IORESOURCE_PREFETCH))
> > > -               if (upper_32_bits(resource_size(res)))
> > > +               if (upper_32_bits(resource_size(res) - 1))
> > >                     dev_warn(dev, "Memory resource size exceeds max for 32 bits\n");
> >
> > I guess this relies on the fact that BARs must be a power of two in
> > size, right?  So anything where the upper 32 bits of the size are
> > non-zero is either 0x1_0000_0000 (4GiB window that we shouldn't warn
> > about), or 0x2_0000_0000 or bigger (where we *do* want to warn about
> > it).
> >
> > But it looks like this is used for host bridge resources, which are
> > windows, not BARs, so they don't have to be a power of two size.  A
> > window of size 0x1_8000_0000 is perfectly legal and would fit the
> > criteria for this warning, but this patch would turn off the warning.
> 
> 0x1_8000_0000 - 1 = 0x1_7fff_ffff
> 
> So that would still work. Maybe you read it as the subtract being
> after upper_32_bits()?

Right, sorry.  I guess a better example would be something like this:

  [mem 0x2000_0000-0x21ff_ffff] -> [pci 0x0_ff00_0000-0x1_00ff_ffff]

where the size is only 0x0200_0000, so we wouldn't warn about it, but
half of the window is above 4G on PCI.

> > I don't really understand this warning in the first place, though.  It
> > was added by fede8526cc48 ("PCI: of: Warn if non-prefetchable memory
> > aperture size is > 32-bit").  But I think the real issue would be
> > related to the highest address, not the size.  For example, an
> > aperture of 0x0_c000_0000 - 0x1_4000_0000 is only 0x8000_0000 in size,
> > but the upper half of it it would be invalid for non-prefetchable
> > 32-bit BARs.
> 
> Are we talking CPU addresses or PCI addresses? For CPU addresses, it
> would be perfectly fine to be above 4G as long as PCI addresses are
> below 4G, right?

Yes, CPU addresses can be above 4G; all that matters for this is the
PCI address.

I think what's important is the largest PCI address in the window, not
the size.