Στις 2018-10-23 17:30, Rob Herring έγραψε:
On Mon, Oct 22, 2018 at 5:55 PM Palmer Dabbelt <palmer@xxxxxxxxxx>
wrote:
On Mon, 22 Oct 2018 15:42:13 PDT (-0700), robh@xxxxxxxxxx wrote:
> On Mon, Oct 15, 2018 at 05:20:10PM +0300, Nick Kossifidis wrote:
>> The /chosen node is optional so we should handle CMDLINE regardless
>> the presence of /chosen/bootargs. Move handling of CMDLINE in
>> early_init_dt_scan() instead.
>
> I looked at this a while back. I'm not sure this behavior can be changed
> without breaking some MIPS platforms that could be relying on the
> current behavior. But trying to make sense of the MIPS code is a
> challenge and they have some other issues in this area.
>
> Can't this be fixed by making /chosen manditory? I'd expect ultimately
> you are always going to need it.
>
> I'd rather not resort to making this per arch. There's also some effort
> to consolidate cmd line handling[1].
I'd rather make /chosen mandatory on RISC-V than to have per-arch
handling, as
like you've said there's already too much duplication. That said, it
does seem
like a bug to me because the behavior seems somewhat arbitrary -- an
empty
/chosen node causing the built-in command-line argument handling to go
off the
rails just smells so buggy.
Yes. Probably need to do some archaeology on this code to figure out
some of the expectations.
If that's the case, could we at least have something like
"CONFIG_OF_CHOSEN_IS_MANDATORY" that provides a warning when there is
no
/chosen node and is set on architecture where the spec mandates
/chosen?
I'd be okay to make it a warning unconditionally. At least then we can
find the cases that deviate and either fix them or understand their
expectations.
Rob
I don't think we can make /chosen node mandatory since it's not
specified as such
by the spec
(https://github.com/devicetree-org/devicetree-specification/releases/tag/v0.2).
No matter what we say from the kernel side, the device tree provider is
not expected
to always provide the /chosen mode. Also device tree is not the only
provider of
a command line, we might get a command line from different places
per-arch (e.g. atags
on arm) or through EFI/ACPI/kexec as well. That's why my initial
approach for RISC-V
was on the arch-specific code.
We shouldn't try to handle the built-in CMDLINE on each of the possible
command line
providers and if we do we should at least make sure we don't depend on
the presence
of a provided command line (which is the issue in this case).
I believe the best approach is to consolidate all the different CMDLINE
approaches
for the various archs / providers and clean up the mess instead of
re-implementing
the same thing again and again. I saw the patch you mentioned and it's a
start.
I'll look more into it and try to come up with a similar one.
Nick
