Re: [PATCH 01/11] objtool: Generic annotation infrastructure

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Dec 04, 2023 at 10:37:03AM +0100, Peter Zijlstra wrote:
> Avoid endless .discard.foo sections for each annotation, create a
> single .discard.annotate section that takes an annotation type along
> with the instruction.
> 
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> ---
> --- a/include/linux/objtool.h
> +++ b/include/linux/objtool.h
> @@ -57,6 +57,13 @@
>  	".long 998b\n\t"						\
>  	".popsection\n\t"
>  
> +#define ASM_ANNOTATE(x)						\
> +	"911:\n\t"						\
> +	".pushsection .discard.annotate,\"M\",@progbits,8\n\t"	\
> +	".long 911b - .\n\t"					\
> +	".long " __stringify(x) "\n\t"				\
> +	".popsection\n\t"
> +
>  #else /* __ASSEMBLY__ */
>  
>  /*
> @@ -146,6 +153,14 @@
>  	.popsection
>  .endm
>  
> +.macro ANNOTATE type:req
> +.Lhere_\@:
> +	.pushsection .discard.annotate,"M",@progbits,8
> +	.long	.Lhere_\@ - .
> +	.long	\type
> +	.popsection
> +.endm
> +
>  #endif /* __ASSEMBLY__ */
>  
>  #else /* !CONFIG_OBJTOOL */
> @@ -167,6 +182,8 @@
>  .endm
>  .macro REACHABLE
>  .endm
> +.macro ANNOTATE
> +.endm
>  #endif
>  
>  #endif /* CONFIG_OBJTOOL */
> --- a/tools/objtool/check.c
> +++ b/tools/objtool/check.c
> @@ -2308,6 +2308,41 @@ static int read_unwind_hints(struct objt
>  	return 0;
>  }
>  
> +static int read_annotate(struct objtool_file *file, void (*func)(int type, struct instruction *insn))
> +{
> +	struct section *rsec, *sec;
> +	struct instruction *insn;
> +	struct reloc *reloc;
> +	int type;
> +
> +	rsec = find_section_by_name(file->elf, ".rela.discard.annotate");
> +	if (!rsec)
> +		return 0;
> +
> +	sec = find_section_by_name(file->elf, ".discard.annotate");
> +	if (!sec)
> +		return 0;
> +
> +	for_each_reloc(rsec, reloc) {
> +		insn = find_insn(file, reloc->sym->sec,
> +				 reloc->sym->offset + reloc_addend(reloc));
> +		if (!insn) {
> +			WARN("bad .discard.annotate entry: %d", reloc_idx(reloc));
> +			return -1;
> +		}
> +
> +		type = *(u32 *)(sec->data->d_buf + (reloc_idx(reloc) * sec->sh.sh_entsize) + 4);
> +
> +		func(type, insn);
> +	}
> +
> +	return 0;
> +}

So... ld.lld hates this :-(

From an LLVM=-19 build we can see that:

$ readelf -WS tmp-build/arch/x86/kvm/vmx/vmenter.o | grep annotate
  [13] .discard.annotate PROGBITS        0000000000000000 00028c 000018 08   M  0   0  1

$ readelf -WS tmp-build/arch/x86/kvm/kvm-intel.o | grep annotate
  [ 3] .discard.annotate PROGBITS        0000000000000000 069fe0 0089d0 00   M  0   0  1

Which tells us that the translation unit itself has a sh_entsize of 8,
while the linked object has sh_entsize of 0.

This then completely messes up the indexing objtool does, which relies
on it being a sane number.

GCC/binutils very much does not do this, it retains the 8.

Dear clang folks, help?