On Wed, Feb 3, 2021 at 10:49 AM Rob Herring <robh+dt@xxxxxxxxxx> wrote:
>
> On Wed, Feb 3, 2021 at 8:44 AM Andre Przywara <andre.przywara@xxxxxxx> wrote:
> >
> > On Wed, 3 Feb 2021 15:15:19 +0100
> > Michal Simek <michal.simek@xxxxxxxxxx> wrote:
> >
> > > On 2/3/21 3:12 PM, Rob Herring wrote:
> > > > On Wed, Feb 3, 2021 at 1:01 AM Michal Simek <michal.simek@xxxxxxxxxx> wrote:
> > > >>
> > > >>
> > > >>
> > > >> On 2/1/21 6:41 PM, Rob Herring wrote:
> > > >>> On Mon, Feb 1, 2021 at 8:27 AM Michal Simek <michal.simek@xxxxxxxxxx> wrote:
> > > >>>>
> > > >>>> The commit 3eb619b2f7d8 ("scripts/dtc: Update to upstream version
> > > >>>> v1.6.0-11-g9d7888cbf19c") updated dtc version which also contained DTC
> > > >>>> commit
> > > >>>> "81e0919a3e21 checks: Add interrupt provider test"
> > > >>>> where reasons for this checking are mentioned as
> > > >>>> "A missing #address-cells property is less critical, but creates
> > > >>>> ambiguities when used in interrupt-map properties, so warn about this as
> > > >>>> well now."
> > > >>>>
> > > >>>> Add address-cells property to gic and gpio nodes to get rid of this warning.
> > > >>>> The similar change has been done for ZynqMP too.
> > > >>>
> > > >>> FYI, we're going to make this check dependent on having an
> > > >>> interrupt-map property. So adding these isn't necessary.
> > > >>
> > > >> Good to know. Is there going to be report if interrupt-map doesn't
> > > >> exist? Which can end up with reverting these changes?
> > > >
> > > > You mean a warning if '#address-cells' is present and interrupt-map is
> > > > not? No, that would cause lots of warnings.
> > >
> > > yep.
> >
> > Why would we do that? That sounds dangerous and would be broken if the
> > IRQ controller is in a generic .dtsi (as it usually is), but the
> > interrupt map is only in *some* of the board .dts files.
> >
> > What is the problem of just putting #address-cells = <0>; in the
> > IRQ controller node, after checking that there currently no interrupt
> > maps in use and no IRQ children? And be safe for good? That's 16 bytes
> > in the DTB, IIUC.
>
> Because I don't think we need a bunch of warning fix patches to add
> these everywhere. Also, the need for #address-cells pretty much makes
> no sense on any modern system. It is a relic from days when the bus
> (address) topology and interrupt topology were related.
>
> > Because otherwise we have that lovely ambiguity between the
> > implicit default #address-cells = 2; and the assumed default of 0.
> >
> > And that's why I think we also cannot *automatically* add an #ac = <0>;
> > property, because that would change behaviour.
>
> I'd rather try to limit where we assume the default of 2. My guess is
> that's only some combination of old PowerPC and/or Sparc and no FDT
> based DT.
Actually, after reviewing of_irq_parse_raw() again, I think you're
mixing the 2 different #address-cells involved. Let's review which
#*-cells applies to parts of interrupt-map:
interrupt-map = <[ac current node or parent] [ic current node] [parent
intc phandle] [ac parent intc] [ic parent intc]>;
For [ac current node or parent], we start in the 'interrupt-map' node
(because it's the interrupt parent). From there, we walk up the tree
to find #address-cells. Worst case is we find none and take the
default of 2. First, dtc has pretty much always made no root
#address-cells a warning. Second, Linux has notion of a default and
that varies by arch and isn't used here. Only Sparc defaults to 2 (see
of_private.h) which means we should never hit the default on PowerPC
or Arm (or anything else).
The #address-cells the fix here addresses is the [parent intc
phandle]'s for [ac parent intc] cells. This default is 0 (see
newaddrsize in of_irq_parse_raw()). So really, we only need to be
checking for #address-cells in nodes with interrupt-map.
Rob
