Re: [PATCH] modpost: fix section mismatch message for R_ARM_ABS32

[Masahiro Yamada]

On Thu, May 18, 2023 at 6:41 AM Nick Desaulniers
<ndesaulniers@xxxxxxxxxx> wrote:
>
> On Sun, May 14, 2023 at 5:54 PM Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> >
> > The section mismatch check does not show proper warning messages for ARM.
> >
> > Here, very simple test code.
> >
> >     #include <linux/init.h>
> >
> >     static int __initdata foo;
> >
> >     void set_foo(int x)
> >     {
> >             foo = x;
> >     }
> >
> >     int get_foo(int x)
> >     {
> >             return foo;
> >     }
> >
> > If I compile it for ARM, modpost does not show the symbol name.
> >
> >   WARNING: modpost: vmlinux.o: section mismatch in reference: set_foo (section: .text) -> (unknown) (section: .init.data)
> >   WARNING: modpost: vmlinux.o: section mismatch in reference: get_foo (section: .text) -> (unknown) (section: .init.data)
> >
> > If I compile it for other architectures, modpost shows the correct symbol name.
> >
> >   WARNING: modpost: vmlinux.o: section mismatch in reference: set_foo (section: .text) -> foo (section: .init.data)
> >   WARNING: modpost: vmlinux.o: section mismatch in reference: get_foo (section: .text) -> foo (section: .init.data)
> >
> > For R_ARM_ABS32, addend_arm_rel() sets r->r_addend to a wrong value.
> >
> > arch/arm/kernel/module.c handles R_ARM_ABS32 as follows:
> >
> >         case R_ARM_ABS32:
> >         case R_ARM_TARGET1:
> >                 *(u32 *)loc += sym->st_value;
> >
> > I just mimicked it in modpost.
> >
> > Fixes: 56a974fa2d59 ("kbuild: make better section mismatch reports on arm")
> > Signed-off-by: Masahiro Yamada <masahiroy@xxxxxxxxxx>
> > ---
> >
> >  scripts/mod/modpost.c | 7 ++++---
> >  1 file changed, 4 insertions(+), 3 deletions(-)
> >
> > diff --git a/scripts/mod/modpost.c b/scripts/mod/modpost.c
> > index d4531d09984d..c93780d93caf 100644
> > --- a/scripts/mod/modpost.c
> > +++ b/scripts/mod/modpost.c
> > @@ -1460,12 +1460,13 @@ static int addend_386_rel(struct elf_info *elf, Elf_Shdr *sechdr, Elf_Rela *r)
> >  static int addend_arm_rel(struct elf_info *elf, Elf_Shdr *sechdr, Elf_Rela *r)
> >  {
> >         unsigned int r_typ = ELF_R_TYPE(r->r_info);
> > +       unsigned int *location = reloc_location(elf, sechdr, r);
>
> If `location` is only used in one case of the switch, consider
> computing `location` only in that case.


I really suspect the other case labels are also wrong.

For example, see R_ARM_PC24 in arch/arm/kernel/module.c

The offset is encoded in the instruction.
If you can compute the addend without reading the instruction,
I do not know how.

Anyway, I will fix another breakage.
It will need 'location' as well.









>
> > +       Elf_Sym *sym;
> >
> >         switch (r_typ) {
> >         case R_ARM_ABS32:
> > -               /* From ARM ABI: (S + A) | T */
> > -               r->r_addend = (int)(long)
> > -                             (elf->symtab_start + ELF_R_SYM(r->r_info));
> > +               sym = elf->symtab_start + ELF_R_SYM(r->r_info);
> > +               r->r_addend = TO_NATIVE(*location) + sym->st_value;
> >                 break;
> >         case R_ARM_PC24:
> >         case R_ARM_CALL:
> > --
> > 2.39.2
> >
>
>
> --
> Thanks,
> ~Nick Desaulniers



-- 
Best Regards
Masahiro Yamada