On Mon, 30 Mar 2026 08:04:23 -0700
Breno Leitao <[email protected]> wrote:

> On Mon, Mar 30, 2026 at 06:15:17AM -0700, Breno Leitao wrote:
> > On Fri, Mar 27, 2026 at 10:37:44PM +0900, Masami Hiramatsu wrote:
> > > On Fri, 27 Mar 2026 03:06:41 -0700
> > > Breno Leitao <[email protected]> wrote:
> >
> > > > > To fix this, we need to change setup_arch() for each architecture so
> > > > > that it calls this bootconfig_apply_early_params().
> > > >
> > > > Could we instead integrate this into parse_early_param() itself? That
> > > > approach would avoid the need to modify each architecture individually.
> > >
> > > Ah, indeed.
> >
> > I investigated integrating bootconfig into parse_early_param() and hit a
> > blocker: xbc_init() and xbc_make_cmdline() depend on memblock_alloc(), but 
> > on
> > most architectures (x86, arm64, arm, s390, riscv) parse_early_param() is 
> > called
> > from setup_arch() _before_ memblock is initialized.
> 
> That said, I'd like to propose a simpler approach as a first step:
> 
> 1) Keep calling bootconfig_apply_early_params() from setup_boot_config().
>    This is the least intrusive approach and expands bootconfig support to
>    additional early boot parameters.
> 
> 2) Document that architecture-specific early parameters might be ignored.
>    If a parameter is consumed early enough (during setup_arch()), it will
>    not see the bootconfig value.

We have to carefully do this because what parameter is arch specific or not
depends on architecture and undocumented.

How about introducing a new Kconfig, which supports early params by
embedded bootconfig?

> 
> 3) Ensure that early bootconfig parameters don't overwrite the boot command
>    line. For example, if the boot command line has foo=bar and bootconfig
>    later has foo=baz, the command line value should take precedence.
>    This prevents early boot code (in setup_arch()) from seeing a parameter
>    value that will be changed later.

OK, this also needs to be considered. Currently we just pass the bootconfig
parameters right before bootloader given parameters as "extra_command_line"
if "bootconfig" in cmdline or CONFIG_BOOT_CONFIG_FORCE=y.

[boot_config(.kernel)]<command_line>[ -- 
[boot_config(.init)][init_command_line]]

This is currently expected behavior. The bootconfig parameters are
expected to be overridden by command_line or command_line are appended.

If we change this for early params, we also should change the expected
output of /proc/cmdline too. I think we have 2 options;

 - As before, we expect the parameters provided by the boot configuration
   to be processed first and then overridden later by the command line.

Or,

 - ignore all parameters which is given from the command line, this also
   updates existing setup_boot_config() (means xbc_snprint_cmdline() ).

Anyway, this behavior change will also be a bit critical... We have
to announce it.

> 
> 
> If that is OK, that is what I have right now:

Let me comment on it.

> diff --git a/Documentation/admin-guide/bootconfig.rst 
> b/Documentation/admin-guide/bootconfig.rst
> index f712758472d5c..6ed852a0c66d8 100644
> --- a/Documentation/admin-guide/bootconfig.rst
> +++ b/Documentation/admin-guide/bootconfig.rst
> @@ -169,6 +169,15 @@ Boot Kernel With a Boot Config
>  There are two options to boot the kernel with bootconfig: attaching the
>  bootconfig to the initrd image or embedding it in the kernel itself.
>  
> +Early options (those registered with ``early_param()``) may only be
> +specified in the embedded bootconfig, because the initrd is not yet
> +available when early parameters are processed.
> +
> +Note that embedded bootconfig is parsed after ``setup_arch()``, so
> +early options that are consumed during architecture initialization
> +(e.g., ``mem=``, ``memmap=``, ``earlycon``, ``noapic``, ``nolapic``,
> +``acpi=``, ``numa=``, ``iommu=``) may not take effect from bootconfig.
> +

This is easy to explain, but it's quite troublesome for users to
determine which parameters are unavailable. Currently we can identify
it by `git grep early_param -- arch/${ARCH}`. But it is setup in
setup_arch() we need to track the source code. (Or ask AI :))


>  Attaching a Boot Config to Initrd
>  ---------------------------------
>  
> diff --git a/init/Kconfig b/init/Kconfig
> index 7484cd703bc1a..34adcc1feb9b6 100644
> --- a/init/Kconfig
> +++ b/init/Kconfig
> @@ -1525,6 +1525,16 @@ config BOOT_CONFIG_EMBED
>         image. But if the system doesn't support initrd, this option will
>         help you by embedding a bootconfig file while building the kernel.
>  
> +       Unlike bootconfig attached to initrd, the embedded bootconfig also
> +       supports early options (those registered with early_param()). Any
> +       kernel.* key in the embedded bootconfig is applied before
> +       parse_early_param() runs.  Early options in initrd bootconfig will
> +       not be applied.  Early options consumed during setup_arch() (e.g.
> +       mem=, memmap=, earlycon, noapic, acpi=, numa=, iommu=) may not
> +       take effect.  If the same early option
> +       appears in both bootconfig and the kernel command line, the
> +       command line value takes precedence.
> +

As a compromise, how about making this a separate Kconfig?

config BOOT_CONFIG_EMBED_EARLY_PARAM
        bool "Support early_params in embedded bootconfig"

and afterwards, we can introduce 

        depends on ARCH_BOOT_CONFIG_EMBED_EARLY_PARAM


Thank you,

-- 
Masami Hiramatsu (Google) <[email protected]>

Reply via email to