[PATCH] arm64/cpufeature: annotate arm64_use_ng_mappings with ro_after_init to prevent wrong idmap generation

[Ard Biesheuvel]

On Sat, 3 May 2025 at 12:06, Catalin Marinas <catalin.marinas at arm.com> wrote:
>
> On Fri, May 02, 2025 at 07:14:12PM +0100, Yeoreum Yun wrote:
> > > On Fri, May 02, 2025 at 06:41:33PM +0200, Ard Biesheuvel wrote:
> > > > Making arm64_use_ng_mappings __ro_after_init seems like a useful
> > > > change by itself, so I am not objecting to that. But we don't solve it
> > > > more fundamentally, please at least add a big fat comment why it is
> > > > important that the variable remains there.
> > >
> > > Maybe something like the section reference checker we use for __init -
> > > verify that the early C code does not refer anything in the BSS section.
> >
> > Maybe but it would be better to be checked at compile time (I don't
> > know it's possible) otherwise, early C code writer should check
> > mandatroy by calling is_kernel_bss_data() (not exist) for data it refers.
>
> This would be compile time (or rather final link time). See
> scripts/mod/modpost.c (the sectioncheck[] array) on how we check if, for
> example, a .text section references a .init one. We could move the whole
> pi code to its own section (e.g. .init.nommu.*) and add modpost checks
> for references to the bss or other sections.
>

We can take care of this locally in image-vars.h, where each variable
used by the startup code is made available to it explicitly.

Sending out a separate series shortly.