Message-ID: <CAL_JsqL5BwSCZDZPNW2-0kBOw0j9KZZ8WOdnqEmGKhSgquY6Pw@mail.gmail.com>
Date: Wed, 29 May 2024 11:46:55 -0500
From: Rob Herring <robh@...nel.org>
To: Anup Patel <apatel@...tanamicro.com>
Cc: Marc Zyngier <maz@...nel.org>, devicetree@...r.kernel.org, 
	linux-kernel@...r.kernel.org, linux-arm-kernel@...ts.infradead.org, 
	linux-riscv@...ts.infradead.org, Saravana Kannan <saravanak@...gle.com>
Subject: Re: [PATCH] of: property: Fix fw_devlink handling of interrupt-map

On Wed, May 29, 2024 at 8:51 AM Rob Herring <robh@...nel.org> wrote:
>
> On Wed, May 29, 2024 at 6:28 AM Anup Patel <apatel@...tanamicro.com> wrote:
> >
> > On Wed, May 29, 2024 at 4:15 PM Marc Zyngier <maz@...nel.org> wrote:
> > >
> > > On Wed, 29 May 2024 11:16:30 +0100,
> > > Anup Patel <apatel@...tanamicro.com> wrote:
> > > >
> > > > On Wed, May 29, 2024 at 12:03 PM Marc Zyngier <maz@...nel.org> wrote:
> > > > >
> > > > > On Wed, 29 May 2024 06:15:52 +0100,
> > > > > Anup Patel <apatel@...tanamicro.com> wrote:
> > > > > >
> > > > > > On Tue, May 28, 2024 at 10:11 PM Marc Zyngier <maz@...nel.org> wrote:
> > > > > > >
> > > > > > > Commit d976c6f4b32c ("of: property: Add fw_devlink support for
> > > > > > > interrupt-map property") tried to do what it says on the tin,
> > > > > > > but failed on a couple of points:
> > > > > > >
> > > > > > > - it confuses bytes and cells. Not a huge deal, except when it
> > > > > > >   comes to pointer arithmetic
> > > > > > >
> > > > > > > - it doesn't really handle anything but interrupt-maps that have
> > > > > > >   their parent #address-cells set to 0
> > > > > > >
> > > > > > > The combinations of the two leads to some serious fun on my M1
> > > > > > > box, with plenty of WARN-ON() firing all over the shop, and
> > > > > > > amusing values being generated for interrupt specifiers.
> > > > > > >
> > > > > > > Address both issues so that I can boot my machines again.
> > > > > > >
> > > > > > > Fixes: d976c6f4b32c ("of: property: Add fw_devlink support for interrupt-map property")
> > > > > > > Signed-off-by: Marc Zyngier <maz@...nel.org>
> > > > > > > Cc: Anup Patel <apatel@...tanamicro.com>
> > > > > > > Cc: Saravana Kannan <saravanak@...gle.com>
> > > > > > > Cc: Rob Herring (Arm) <robh@...nel.org>
> > > > > >
> > > > > > Thanks for the fix patch but unfortunately it breaks for RISC-V.
> > > > > >
> > > > > > > ---
> > > > > > >  drivers/of/property.c | 16 ++++++++++++++--
> > > > > > >  1 file changed, 14 insertions(+), 2 deletions(-)
> > > > > > >
> > > > > > > diff --git a/drivers/of/property.c b/drivers/of/property.c
> > > > > > > index 1c83e68f805b..9adebc63bea9 100644
> > > > > > > --- a/drivers/of/property.c
> > > > > > > +++ b/drivers/of/property.c
> > > > > > > @@ -1322,7 +1322,13 @@ static struct device_node *parse_interrupt_map(struct device_node *np,
> > > > > > >         addrcells = of_bus_n_addr_cells(np);
> > > > > > >
> > > > > > >         imap = of_get_property(np, "interrupt-map", &imaplen);
> > > > > > > -       if (!imap || imaplen <= (addrcells + intcells))
> > > > > > > +       imaplen /= sizeof(*imap);
> > > > > > > +
> > > > > > > +       /*
> > > > > > > +        * Check that we have enough runway for the child unit interrupt
> > > > > > > +        * specifier and a phandle. That's the bare minimum we can expect.
> > > > > > > +        */
> > > > > > > +       if (!imap || imaplen <= (addrcells + intcells + 1))
> > > > > > >                 return NULL;
> > > > > > >         imap_end = imap + imaplen;
> > > > > > >
> > > > > > > @@ -1346,8 +1352,14 @@ static struct device_node *parse_interrupt_map(struct device_node *np,
> > > > > > >                 if (!index)
> > > > > > >                         return sup_args.np;
> > > > > > >
> > > > > > > -               of_node_put(sup_args.np);
> > > > > > > +               /*
> > > > > > > +                * Account for the full parent unit interrupt specifier
> > > > > > > +                * (address cells, interrupt cells, and phandle).
> > > > > > > +                */
> > > > > > > +               imap += of_bus_n_addr_cells(sup_args.np);
> > > > > >
> > > > > > This breaks for RISC-V because we don't have "#address-cells"
> > > > > > property in interrupt controller DT node and of_bus_n_addr_cells()
> > > > > > retrieves "#address-cells" from the parent of interrupt controller.
> > > > >
> > > > > That's a feature, not a bug. #address-cells, AFAICT, applies to all
> > > > > child nodes until you set it otherwise.
> > > > >
> > > > > >
> > > > > > The of_irq_parse_raw() looks for "#address-cells" property
> > > > > > in the interrupt controller DT node only so we should do a
> > > > > > similar thing here as well.
> > > > >
> > > > > This looks more like a of_irq_parse_raw() bug than anything else.
> > > >
> > > > If we change of_irq_parse_raw() to use of_bus_n_addr_cells()
> > > > then it would still break for RISC-V.
> > >
> > > I'm not trying to "fix" riscv. I'm merely outlining that you are
> > > relying on both broken DTs and a buggy OS.
> > >
> > > >
> > > > Using of_bus_n_addr_cells() over here forces interrupt controller
> > > > DT nodes to have a "#address-cells" DT property. There are many
> > > > interrupt controller (e.g. RISC-V PLIC or RISC-V APLIC) where the
> > > > DT bindings don't require "#address-cells" DT property and existing
> > > > DTS files with such interrupt controllers don't have it either.
> > >
> > > It forces you to set #address-cells *if you rely on a different
> > > value in a child node*. It's not like the semantics are new.
> >
> > We don't have child nodes under the interrupt controller DT node
> > (for both RISC-V PLIC and APLIC) so we certainly don't require the
> > "#address-cells" property in the interrupt controller DT node.
>
> interrupt controller nodes which are referred to by interrupt-map
> require #address-cells. Period. Child nodes or not.

Now that I've re-read the code, the kernel will treat missing
#address-cells in the interrupt parent as 0. Looking in the parent
nodes for #address-cells only happens for the input address (i.e. the
address before the phandle). IOW, the first use of_bus_n_addr_cells()
was correct, but the 2nd use is not correct. That's not great, but
that's how this code passed down on stone tablets works...

As I commented on v1 of the original patch. I don't like parsing
interrupt-map in 2 places. It leads to problems exactly as this thread
has shown. The duplication was reduced some, but is still more than
I'd like. So I'm looking at how to refactor of_irq_parse_raw() to make
it work for what's needed here. Maybe I'll have a fix quickly. If not,
I'm going to revert the original commit.

Rob

Powered by blists - more mailing lists