On Thu, Dec 8, 2022 at 4:17 AM Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
>
> When building arm64 allmodconfig + ThinLTO with clang and a proposed
> modpost update to account for -ffuncton-sections, the following warning
> appears:
>
> WARNING: modpost: vmlinux.o: section mismatch in reference: padata_work_init (section: .text.padata_work_init) -> padata_mt_helper (section: .init.text)
> WARNING: modpost: vmlinux.o: section mismatch in reference: padata_work_init (section: .text.padata_work_init) -> padata_mt_helper (section: .init.text)
>
> LLVM has optimized padata_work_init() to include the address of
> padata_mt_helper() directly, which causes modpost to complain since
> padata_work_init() is not __init, whereas padata_mt_helper() is. In
> reality, padata_work_init() is only called with padata_mt_helper() as
> the work_fn argument in code that is __init, so this warning will not
> result in any problems. Silence it with __ref, which makes it clear to
> modpost that padata_work_init() can only use padata_mt_helper() in
> __init code.
>
> Suggested-by: Daniel Jordan <daniel.m.jordan@xxxxxxxxxx>
> Signed-off-by: Nathan Chancellor <nathan@xxxxxxxxxx>
> ---
> Cc: Steffen Klassert <steffen.klassert@xxxxxxxxxxx>
> Cc: Daniel Jordan <daniel.m.jordan@xxxxxxxxxx>
> Cc: linux-crypto@xxxxxxxxxxxxxxx
> ---
> kernel/padata.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/padata.c b/kernel/padata.c
> index e5819bb8bd1d..4c3137fe8449 100644
> --- a/kernel/padata.c
> +++ b/kernel/padata.c
> @@ -83,8 +83,8 @@ static struct padata_work *padata_work_alloc(void)
> return pw;
> }
>
> -static void padata_work_init(struct padata_work *pw, work_func_t work_fn,
> - void *data, int flags)
> +static __ref void padata_work_init(struct padata_work *pw, work_func_t work_fn,
> + void *data, int flags)
> {
> if (flags & PADATA_WORK_ONSTACK)
> INIT_WORK_ONSTACK(&pw->pw_work, work_fn);
>
> base-commit: 76dcd734eca23168cb008912c0f69ff408905235
> --
> 2.38.1
>
It took me a while to understand why LTO can embed
padata_mt_helper's address into padata_work_init().
There are 3 call-sites to padata_work_init().
(1) __init padata_work_alloc_mt()
--> padata_work_init(..., padata_mt_helper, ...)
(2) padata_do_parallel()
--> padata_work_init(..., padata_parallel_worker, ...)
(3) __init padata_do_multithreaded()
--> padata_work_init(..., padata_mt_helper, ...)
The function call (2) is squashed away.
With only (1) and (3) remaining, the 2nd parameter to
padata_work_init() is always padata_mt_helper,
therefore LLVM embeds padata_mt_hlper's address
directly into padata_work_init().
I am not sure if the compiler should do this level of optimization
because kernel/padata.c does not seem to be a special case.
Perhaps, we might be hit with more cases that need __ref annotation,
which is only required by LTO.
One note is that, we could discard padata_work_init()
because (1) and (3) are both annotated as __init.
So, another way of fixing is
static __always_inline void padata_work_init(...)
because the compiler would determine padata_work_init()
would be small enough if the caller and callee belonged to
the same section.
I do not have a strong opinion.
Honestly, I do not know what the best approach would be to fix this.
If we go with the __ref annotation, I can pick this, but
at least can you add some comments?
include/linux/init.h says:
"optimally document why the __ref is needed and why it's OK"
I think this is the case that needs some comments
because LTO optimization looks too tricky to me.
--
Best Regards
Masahiro Yamada
