[PATCH v2] arm64/cpufeature: annotate arm64_use_ng_mappings with ro_after_init to prevent wrong idmap generation

[Ard Biesheuvel]

On Tue, 6 May 2025 at 10:16, Ryan Roberts <ryan.roberts at arm.com> wrote:
>
> On 06/05/2025 09:09, Yeoreum Yun wrote:
> > Hi Catalin,
> >
> >> On Sat, May 03, 2025 at 09:23:27PM +0100, Yeoreum Yun wrote:
> >>> Hi Catalin,
> >>>
> >>>> On Sat, May 03, 2025 at 11:16:12AM +0100, Catalin Marinas wrote:
> >>>>> On Fri, 02 May 2025 19:04:12 +0100, Yeoreum Yun wrote:
> >>>>>> create_init_idmap() could be called before .bss section initialization
> >>>>>> which is done in early_map_kernel().
> >>>>>> Therefore, data/test_prot could be set incorrectly by PTE_MAYBE_NG macro.
> >>>>>>
> >>>>>> PTE_MAYBE_NG macro set NG bit according to value of "arm64_use_ng_mappings".
> >>>>>> and this variable places in .bss section.
> >>>>>>
> >>>>>> [...]
> >>>>>
> >>>>> Applied to arm64 (for-next/fixes), with some slight tweaking of the
> >>>>> comment, thanks!
> >>>>>
> >>>>> [1/1] arm64/cpufeature: annotate arm64_use_ng_mappings with ro_after_init to prevent wrong idmap generation
> >>>>>       https://git.kernel.org/arm64/c/12657bcd1835
> >>>>
> >>>> I'm going to drop this for now. The kernel compiled with a clang 19.1.5
> >>>> version I have around (Debian sid) fails to boot, gets stuck early on:
> >>>>
> >>>> $ clang --version
> >>>> Debian clang version 19.1.5 (1)
> >>>> Target: aarch64-unknown-linux-gnu
> >>>> Thread model: posix
> >>>> InstalledDir: /usr/lib/llvm-19/bin
> >>>>
> >>>> I didn't have time to investigate, disassemble etc. I'll have a look
> >>>> next week.
> >>>
> >>> Just for your information.
> >>> When I see the debian package, clang 19.1.5-1 doesn't supply anymore:
> >>>  - https://ftp.debian.org/debian/pool/main/l/llvm-toolchain-19/
> >>>
> >>> and the default version for sid is below:
> >>>
> >>> $ clang-19 --version
> >>> Debian clang version 19.1.7 (3)
> >>> Target: aarch64-unknown-linux-gnu
> >>> Thread model: posix
> >>> InstalledDir: /usr/lib/llvm-19/bin
> >>>
> >>> When I tested with above version with arm64-linux's for-next/fixes
> >>> including this patch. it works well.
> >>
> >> It doesn't seem to be toolchain related. It fails with gcc as well from
> >> Debian stable but you'd need some older CPU (even if emulated, e.g.
> >> qemu). It fails with Cortex-A72 (guest on Raspberry Pi 4) but not
> >> Neoverse-N2. Also changing the annotation from __ro_after_init to
> >> __read_mostly also works.
>
> I think this is likely because __ro_after_init is also "ro before init" - i.e.
> if you try to write to it in the PI code an exception is generated due to it
> being mapped RO. Looks like early_map_kernel() is writiing to it.
>

Indeed.

> I've noticed a similar problem in the past and it would be nice to fix it so
> that PI code maps __ro_after_init RW.
>

The issue is that the store occurs via the ID map, which only consists
of one R-X and one RW- section. I'm not convinced that it's worth the
hassle to relax this.

If moving the variable to .data works, then let's just do that.