Re: [PATCH v3] kallsyms: strip CLANG CFI postfix ".cfi_jt"

[Sami Tolvanen]

Hi,

On Thu, Jul 29, 2021 at 1:54 PM Padmanabha Srinivasaiah
<treasure4paddy@xxxxxxxxx> wrote:
>
> Clang CFI adds a postfix ".cfi_jt" to a symbols of extern functions.
> For example this breaks syscall tracer that doesn't expect such postfix,
> so strip out the postfix from the expanded symbol.
>
> Signed-off-by: Padmanabha Srinivasaiah <treasure4paddy@xxxxxxxxx>
> ---
>
> Change in v3:
>   - Modified commit message to indicate fix is for Clang CFI postfix
>   - Rebased on recent patch from ndesaulniers@xxxxxxxxxx.
>   - Fix is enabled even for CONFIG_LTO_CLANG
>
> Change in v2:
>   - Use existing routine in kallsyms to strip postfix ".cfi_jt" from
>     extern function name.
>   - Modified the commit message accordingly
>
>  kernel/kallsyms.c | 21 +++++++++++++++------
>  1 file changed, 15 insertions(+), 6 deletions(-)
>
> diff --git a/kernel/kallsyms.c b/kernel/kallsyms.c
> index 5cabe4dd3ff4..67d015854cbd 100644
> --- a/kernel/kallsyms.c
> +++ b/kernel/kallsyms.c
> @@ -174,7 +174,8 @@ static bool cleanup_symbol_name(char *s)
>          * foo.llvm.974640843467629774. This can break hooking of static
>          * functions with kprobes.
>          */
> -       if (!IS_ENABLED(CONFIG_LTO_CLANG_THIN))
> +       if (!(IS_ENABLED(CONFIG_LTO_CLANG) ||
> +             IS_ENABLED(CONFIG_LTO_CLANG_THIN)))

This is redundant. LTO_CLANG is selected for both LTO modes, so
there's no need to also check for LTO_CLANG_THIN here.

>                 return false;
>
>         res = strstr(s, ".llvm.");

However, we should probably check for ".llvm." only with LTO_CLANG_THIN.

> @@ -184,16 +185,24 @@ static bool cleanup_symbol_name(char *s)
>         }
>
>         /*
> -        * LLVM appends a hash to static function names when ThinLTO and CFI
> -        * are both enabled, i.e. foo() becomes
> -        * foo$707af9a22804d33c81801f27dcfe489b. This causes confusion and
> -        * potentially breaks user space tools, so we strip the suffix from
> -        * expanded symbol names.
> +        * LLVM appends a hash to static function names when both
> +        * (Thin/FULL) LTO and CFI are enabled, i.e. foo() becomes
> +        * foo$707af9a22804d33c81801f27dcfe489b.

That's not quite right, the hash is only appended with ThinLTO. I
would leave this comment untouched.

> +        *
> +        * In case of non static function symbol <funcsym>,
> +        * the local jump table will have entry as <funcsym>.cfi_jt.
> +        *
> +        * This causes confusion and potentially breaks
> +        * user space tools and some built-in components.
> +        * So we strip the suffix from expanded symbol names.
>          */
>         if (!IS_ENABLED(CONFIG_CFI_CLANG))
>                 return false;
>
>         res = strrchr(s, '$');
> +       if (!res)
> +               res = strstr(s, ".cfi_jt");

And add a comment about stripping .cfi_jt from jump table symbols
before this part.

> +
>         if (res) {
>                 *res = '\0';
>                 return true;
> --
> 2.17.1
>

Sami