After commit ac7c3e4ff401 ("compiler: enable CONFIG_OPTIMIZE_INLINING
forcibly"), a compiler may choose not to inline a function marked with
just 'inline'. If check_bugs() is not inlined into start_kernel(), which
occurs when building with clang after commit 9ea7e6b62c2b ("init: Mark
[arch_call_]rest_init() __noreturn"), modpost complains with:
WARNING: modpost: vmlinux.o: section mismatch in reference: check_bugs (section: .text) -> check_bugs32 (section: .init.text)
check_bugs() is only called from start_kernel(), which itself is marked
__init, so mark check_bugs() as __init as well to clear up the warning
and make everything work properly.
While there is currently no warning about check_bugs_early(), it could
have the same problem, as it is called from arch_setup() and calls
check_bugs64_early(), both marked __init. Mark it as __init for the same
reason as above.
Signed-off-by: Nathan Chancellor <[email protected]>
---
NOTE: This is based on v6.3-rc7, as the issue shows up due to a patch in
the tip tree, but this appears to be an ancient issue that could have
showed up at any point (hence why no explicit Fixes tag), so I chose a
base that should allow either the MIPS or tip folks to apply this patch.
Additionally, I was tempted to remove the explicit 'inline' since the
compiler is free to do whatever it wants anyways but this is a static
function in a header so we would need to add '__maybe_unused', which is
already added with 'inline' in a normal build so I just left it alone.
---
arch/mips/include/asm/bugs.h | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/arch/mips/include/asm/bugs.h b/arch/mips/include/asm/bugs.h
index d72dc6e1cf3c..9b9bf9bc7d24 100644
--- a/arch/mips/include/asm/bugs.h
+++ b/arch/mips/include/asm/bugs.h
@@ -24,13 +24,13 @@ extern void check_bugs64_early(void);
extern void check_bugs32(void);
extern void check_bugs64(void);
-static inline void check_bugs_early(void)
+static inline void __init check_bugs_early(void)
{
if (IS_ENABLED(CONFIG_CPU_R4X00_BUGS64))
check_bugs64_early();
}
-static inline void check_bugs(void)
+static inline void __init check_bugs(void)
{
unsigned int cpu = smp_processor_id();
---
base-commit: 6a8f57ae2eb07ab39a6f0ccad60c760743051026
change-id: 20230419-mips-check_bugs-init-attribute-026103bdb255
Best regards,
--
Nathan Chancellor <[email protected]>
On Wed, Apr 19, 2023 at 11:43 AM Nathan Chancellor <[email protected]> wrote:
>
> After commit ac7c3e4ff401 ("compiler: enable CONFIG_OPTIMIZE_INLINING
> forcibly"), a compiler may choose not to inline a function marked with
> just 'inline'. If check_bugs() is not inlined into start_kernel(), which
> occurs when building with clang after commit 9ea7e6b62c2b ("init: Mark
> [arch_call_]rest_init() __noreturn"), modpost complains with:
>
> WARNING: modpost: vmlinux.o: section mismatch in reference: check_bugs (section: .text) -> check_bugs32 (section: .init.text)
>
> check_bugs() is only called from start_kernel(), which itself is marked
> __init, so mark check_bugs() as __init as well to clear up the warning
> and make everything work properly.
>
> While there is currently no warning about check_bugs_early(), it could
> have the same problem, as it is called from arch_setup() and calls
> check_bugs64_early(), both marked __init. Mark it as __init for the same
> reason as above.
>
> Signed-off-by: Nathan Chancellor <[email protected]>
> ---
> NOTE: This is based on v6.3-rc7, as the issue shows up due to a patch in
> the tip tree, but this appears to be an ancient issue that could have
> showed up at any point (hence why no explicit Fixes tag), so I chose a
> base that should allow either the MIPS or tip folks to apply this patch.
Probably for the MIPS tree.
>
> Additionally, I was tempted to remove the explicit 'inline' since the
> compiler is free to do whatever it wants anyways but this is a static
> function in a header so we would need to add '__maybe_unused', which is
> already added with 'inline' in a normal build so I just left it alone.
> ---
> arch/mips/include/asm/bugs.h | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/mips/include/asm/bugs.h b/arch/mips/include/asm/bugs.h
> index d72dc6e1cf3c..9b9bf9bc7d24 100644
> --- a/arch/mips/include/asm/bugs.h
> +++ b/arch/mips/include/asm/bugs.h
> @@ -24,13 +24,13 @@ extern void check_bugs64_early(void);
> extern void check_bugs32(void);
> extern void check_bugs64(void);
>
> -static inline void check_bugs_early(void)
> +static inline void __init check_bugs_early(void)
> {
> if (IS_ENABLED(CONFIG_CPU_R4X00_BUGS64))
> check_bugs64_early();
> }
If the only call site is in arch/mips/kernel/setup.c, then perhaps we
can move the definition of check_bugs_early there and mark it static
__init and drop inline?
Unless we foresee other callers potentially in the future? Patch
otherwise LGTM. Thanks for the patch!
>
> -static inline void check_bugs(void)
> +static inline void __init check_bugs(void)
> {
> unsigned int cpu = smp_processor_id();
>
>
> ---
> base-commit: 6a8f57ae2eb07ab39a6f0ccad60c760743051026
> change-id: 20230419-mips-check_bugs-init-attribute-026103bdb255
>
> Best regards,
> --
> Nathan Chancellor <[email protected]>
>
--
Thanks,
~Nick Desaulniers
On Wed, Apr 19, 2023 at 03:03:31PM -0700, Nick Desaulniers wrote:
> On Wed, Apr 19, 2023 at 11:43 AM Nathan Chancellor <[email protected]> wrote:
> >
> > After commit ac7c3e4ff401 ("compiler: enable CONFIG_OPTIMIZE_INLINING
> > forcibly"), a compiler may choose not to inline a function marked with
> > just 'inline'. If check_bugs() is not inlined into start_kernel(), which
> > occurs when building with clang after commit 9ea7e6b62c2b ("init: Mark
> > [arch_call_]rest_init() __noreturn"), modpost complains with:
> >
> > WARNING: modpost: vmlinux.o: section mismatch in reference: check_bugs (section: .text) -> check_bugs32 (section: .init.text)
> >
> > check_bugs() is only called from start_kernel(), which itself is marked
> > __init, so mark check_bugs() as __init as well to clear up the warning
> > and make everything work properly.
> >
> > While there is currently no warning about check_bugs_early(), it could
> > have the same problem, as it is called from arch_setup() and calls
> > check_bugs64_early(), both marked __init. Mark it as __init for the same
> > reason as above.
> >
> > Signed-off-by: Nathan Chancellor <[email protected]>
> > ---
> > NOTE: This is based on v6.3-rc7, as the issue shows up due to a patch in
> > the tip tree, but this appears to be an ancient issue that could have
> > showed up at any point (hence why no explicit Fixes tag), so I chose a
> > base that should allow either the MIPS or tip folks to apply this patch.
>
> Probably for the MIPS tree.
>
> >
> > Additionally, I was tempted to remove the explicit 'inline' since the
> > compiler is free to do whatever it wants anyways but this is a static
> > function in a header so we would need to add '__maybe_unused', which is
> > already added with 'inline' in a normal build so I just left it alone.
> > ---
> > arch/mips/include/asm/bugs.h | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/mips/include/asm/bugs.h b/arch/mips/include/asm/bugs.h
> > index d72dc6e1cf3c..9b9bf9bc7d24 100644
> > --- a/arch/mips/include/asm/bugs.h
> > +++ b/arch/mips/include/asm/bugs.h
> > @@ -24,13 +24,13 @@ extern void check_bugs64_early(void);
> > extern void check_bugs32(void);
> > extern void check_bugs64(void);
> >
> > -static inline void check_bugs_early(void)
> > +static inline void __init check_bugs_early(void)
> > {
> > if (IS_ENABLED(CONFIG_CPU_R4X00_BUGS64))
> > check_bugs64_early();
> > }
>
> If the only call site is in arch/mips/kernel/setup.c, then perhaps we
> can move the definition of check_bugs_early there and mark it static
> __init and drop inline?
Sure, we could even go a step further and just copy the body into the
one call site ourselves, I see little reason for this to be a dedicated
function. That is probably best done in a separate patch altogether in
lieu of just adding __init.
> Unless we foresee other callers potentially in the future? Patch
> otherwise LGTM. Thanks for the patch!
Seems unlikely, I did a super scientific survey of several different
revisions (:P) and check_bugs_early() is called once in each one:
v3.0:arch/mips/kernel/setup.c: check_bugs_early();
v3.5:arch/mips/kernel/setup.c: check_bugs_early();
v3.10:arch/mips/kernel/setup.c: check_bugs_early();
v4.0:arch/mips/kernel/setup.c: check_bugs_early();
v4.5:arch/mips/kernel/setup.c: check_bugs_early();
v5.0:arch/mips/kernel/setup.c: check_bugs_early();
v4.10:arch/mips/kernel/setup.c: check_bugs_early();
v5.5:arch/mips/kernel/setup.c: check_bugs_early();
v5.10:arch/mips/kernel/setup.c: check_bugs_early();
v6.0:arch/mips/kernel/setup.c: check_bugs_early();
I'll do whatever Thomas would prefer.
> >
> > -static inline void check_bugs(void)
> > +static inline void __init check_bugs(void)
> > {
> > unsigned int cpu = smp_processor_id();
> >
> >
> > ---
> > base-commit: 6a8f57ae2eb07ab39a6f0ccad60c760743051026
> > change-id: 20230419-mips-check_bugs-init-attribute-026103bdb255
> >
> > Best regards,
> > --
> > Nathan Chancellor <[email protected]>
> >
>
>
> --
> Thanks,
> ~Nick Desaulniers
On Wed, Apr 19, 2023 at 04:37:10PM -0700, Nathan Chancellor wrote:
> > > diff --git a/arch/mips/include/asm/bugs.h b/arch/mips/include/asm/bugs.h
> > > index d72dc6e1cf3c..9b9bf9bc7d24 100644
> > > --- a/arch/mips/include/asm/bugs.h
> > > +++ b/arch/mips/include/asm/bugs.h
> > > @@ -24,13 +24,13 @@ extern void check_bugs64_early(void);
> > > extern void check_bugs32(void);
> > > extern void check_bugs64(void);
> > >
> > > -static inline void check_bugs_early(void)
> > > +static inline void __init check_bugs_early(void)
> > > {
> > > if (IS_ENABLED(CONFIG_CPU_R4X00_BUGS64))
> > > check_bugs64_early();
> > > }
> >
> > If the only call site is in arch/mips/kernel/setup.c, then perhaps we
> > can move the definition of check_bugs_early there and mark it static
> > __init and drop inline?
>
> Sure, we could even go a step further and just copy the body into the
> one call site ourselves, I see little reason for this to be a dedicated
> function. That is probably best done in a separate patch altogether in
> lieu of just adding __init.
I don't expect any new checks added to check_bugs_early so let's
move the whole function into the call site.
Thomas.
--
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea. [ RFC1925, 2.3 ]