Re: [PATCH] ARM: dts: zynq: Add address-cells property to interrupt controllers

[Rob Herring]

On Wed, Feb 3, 2021 at 12:03 PM Rob Herring <robh+dt@xxxxxxxxxx> wrote:
>
> 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).

Actually, Sparc doesn't even use this code. Turns out PowerPC is a bit
more complicated.

I traced where the '2' in this code came from. PowerPC had a mixture
of the default being 1 or 2. For the interrupt parsing code, it was 1
(from prom_n_addr_cells()) before commit 0ebfff1491ef and 2
(hardcoded) after it. That's not the only place that a default was
set. The early_init_dt_scan_root() function at that time defaulted to
2. Now it's 1 as we added per arch default defines which used the '1'
from prom_n_addr_cells() (now of_n_addr_cells()). So in conclusion,
PowerPC has had a mixture of defaults and no one cared since 2006 when
it changed. I'm inclined to rip out these defaults and just fail.

Rob