Re: The trouble with __weak and objtool got worse

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Apr 15, 2022 at 11:21:30AM -0700, Josh Poimboeuf wrote:

> If LLVM assembler doesn't support this option then we may have to go
> with something like this?  I can't seem to recreate so I'm not able to
> test.
> 
> We'd also need some objtool checks to make sure the non-section-symbol
> fallback isn't being done for a weak symbol.  Or just remove the
> fallback altogether and force section symbols whereever they're needed,
> similar to the below (untested).
> 
> 
> diff --git a/include/linux/static_call_types.h b/include/linux/static_call_types.h
> index 5a00b8b2cf9f..77040ce575fa 100644
> --- a/include/linux/static_call_types.h
> +++ b/include/linux/static_call_types.h
> @@ -52,9 +52,14 @@ struct static_call_site {
>  #define __STATIC_CALL_ADDRESSABLE(name) \
>  	__ADDRESSABLE(STATIC_CALL_KEY(name))
>  
> +extern unsigned long __addressable_ip;
> +
> +#define __STATIC_CALL_SITE_ADDRESSABLE() __addressable_ip = _THIS_IP_;
> +
>  #define __static_call(name)						\
>  ({									\
>  	__STATIC_CALL_ADDRESSABLE(name);				\
> +	__STATIC_CALL_SITE_ADDRESSABLE();				\
>  	__raw_static_call(name);					\
>  })
>  
> diff --git a/kernel/static_call_inline.c b/kernel/static_call_inline.c
> index dc5665b62814..f6e3e0463efb 100644
> --- a/kernel/static_call_inline.c
> +++ b/kernel/static_call_inline.c
> @@ -17,6 +17,8 @@ extern struct static_call_tramp_key __start_static_call_tramp_key[],
>  
>  static bool static_call_initialized;
>  
> +unsigned long __section(".discard.addressable") __addressable_ip;
> +
>  /* mutex to protect key modules/sites */
>  static DEFINE_MUTEX(static_call_mutex);

This *might* work for static_call. But what about things like
.retpoline_sites?

If we get an indirect inside the weak function and at
a different place inside the non-weak function, then we end up with two
patch sites in the non-weak function.

And since all of that is compiler generated, we don't have anything to
stick anything in to alleviate trouble.

(in fact, I didn't observe this with static_call, I just picked it for
the example because it was somewhat easier)