Re: [PATCH 4/6] of: Create of_root if no dtb provided by firmware

[Rob Herring]

On Thu, Jan 18, 2024 at 2:46 AM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> Hi Rob,
>
> On Wed, Jan 17, 2024 at 6:41 PM Rob Herring <robh@xxxxxxxxxx> wrote:
> > On Tue, Jan 16, 2024 at 05:18:15PM -0800, Stephen Boyd wrote:
> > > Quoting Rob Herring (2024-01-15 12:32:30)
> > > > On Fri, Jan 12, 2024 at 12:07:47PM -0800, Stephen Boyd wrote:
> > > > > diff --git a/drivers/of/Kconfig b/drivers/of/Kconfig
> > > > > index da9826accb1b..9628e48baa15 100644
> > > > > --- a/drivers/of/Kconfig
> > > > > +++ b/drivers/of/Kconfig
> > > > > @@ -54,9 +54,14 @@ config OF_FLATTREE
> > > > >       select CRC32
> > > > >
> > > > >  config OF_EARLY_FLATTREE
> > > > > -     bool
> > > > > +     bool "Functions for accessing Flat Devicetree (FDT) early in boot"
> > > >
> > > > I think we could instead just get rid of this kconfig option. Or
> > > > always enable with CONFIG_OF (except on Sparc). The only cost of
> > > > enabling it is init section functions which get freed anyways.
> > >
> > > Getting rid of it is a more massive change. It can be the default and
> > > kept hidden instead? If it can't be selected on Sparc then it should be
> > > hidden there anyway.
> >
> > The easier option is certainly fine for this series. I just don't want
> > it visible.
> >
> > > > >       select DMA_DECLARE_COHERENT if HAS_DMA && HAS_IOMEM
> > > > >       select OF_FLATTREE
> > > > > +     help
> > > > > +       Normally selected by platforms that process an FDT that has been
> > > > > +       passed to the kernel by the bootloader.  If the bootloader does not
> > > > > +       pass an FDT to the kernel and you need an empty devicetree that
> > > > > +       contains only a root node to exist, then say Y here.
> > > > >
> > > > >  config OF_PROMTREE
> > > > >       bool
> > > [...]
> > > > > @@ -195,6 +191,17 @@ static inline int of_node_check_flag(const struct device_node *n, unsigned long
> > > > >       return test_bit(flag, &n->_flags);
> > > > >  }
> > > > >
> > > > > +/**
> > > > > + * of_have_populated_dt() - Has DT been populated by bootloader
> > > > > + *
> > > > > + * Return: True if a DTB has been populated by the bootloader and it isn't the
> > > > > + * empty builtin one. False otherwise.
> > > > > + */
> > > > > +static inline bool of_have_populated_dt(void)
> > > > > +{
> > > > > +     return of_root != NULL && !of_node_check_flag(of_root, OF_EMPTY_ROOT);
> > > >
> > > > Just a side comment, but I think many/all callers of this function could
> > > > just be removed.
> > > >
> > > > I don't love new flags. Another possible way to handle this would be
> > > > checking for "compatible" being present in the root node. I guess this
> > > > is fine as-is for now at least.
> > >
> > > Ok. I can add a check for a compatible property. That's probably better
> > > anyway. Should there be a compatible property there to signal that this
> > > DT isn't compatible with anything? I worry about DT overlays injecting a
> > > compatible string into the root node, but maybe that is already
> > > prevented.
> >
> > I worry about DT overlays injecting anything...
> >
> > I don't think it is explicitly forbidden, but I have asked that any
> > general purpose interface to apply overlays be restricted to nodes
> > explicitly allowed (e.g. downstream of a connector node). For now, the
> > places (i.e. drivers) overlays are applied are limited.
> >
> > We could probably restrict the root node to new nodes only and no new
> > or changed properties.
>
> Changing (<wild dream>or appending to</wild dream>) the root
> "compatible" and/or "model" properties is useful in case of large
> extension boards, though.  This is also the case for DTBs created from
> a base DTB and one or more overlays using fdtoverlay.

I think appending by adding another compatible value could be okay.
Removing or appending to an existing entry is not. We don't want the
following sequence to be possible:

of_machine_is_compatible("foo") --> true
apply overlay
of_machine_is_compatible("foo") --> false

For Stephen's case, it's going from no root compatible at all to
something. I don't think your case would apply here. To put it another
way, if we've booted with ACPI, compatible in the root node is not
valid.


> For the latter, see also the following threads, where you weren't
> (but probably should have been) CCed:
>
> [1] "[PATCH v9 2/2] arm64: boot: Support Flat Image Tree"
>      https://lore.kernel.org/all/20231202035511.487946-3-sjg@xxxxxxxxxxxx
> [2] "Proposal: FIT support for extension boards / overlays"
>     https://lore.kernel.org/all/CAPnjgZ06s64C2ux1rABNAnMv3q4W++sjhNGCO_uPMH_9sTF7Mw@xxxxxxxxxxxxxx

That all seems pretty orthogonal to the issues here.

Rob