[PATCH] x86-64: Stack alignment in _dl_tlsdesc_dynamic and red zone usage (bug 31501)

[H.J. Lu]

On Sat, Mar 16, 2024 at 7:33 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> In sysdeps/x86_64/dl-tlsdesc-dynamic.h, the stack pointer is
> realigned for some variants (notably _dl_tlsdesc_dynamic_xsavec).
> This realignment does not take into account that the function has
> already used part of the red zone at this point, thus clobbering
> the initally saved register values located there if the stack
> alignment inherited from the caller is unfortunate.
>
> (Note: I do not know to write a good test case for this in the existing
> framework.  We saw this as a random LTO plugin crash when building GCC
> with -mtls-dialect=gnu2.  The existing tests on pass on x86_64-linux-gnu
> with this change here.)

Will a different STATE_SAVE_OFFSET for TLS descriptor work?

> ---
>  sysdeps/x86_64/dl-tlsdesc-dynamic.h | 4 ++++
>  1 file changed, 4 insertions(+)
>
> diff --git a/sysdeps/x86_64/dl-tlsdesc-dynamic.h b/sysdeps/x86_64/dl-tlsdesc-dynamic.h
> index 9f02cfc3eb..8e49e7eece 100644
> --- a/sysdeps/x86_64/dl-tlsdesc-dynamic.h
> +++ b/sysdeps/x86_64/dl-tlsdesc-dynamic.h
> @@ -83,6 +83,8 @@ _dl_tlsdesc_dynamic:
>  2:
>  #if DL_RUNTIME_RESOLVE_REALIGN_STACK
>         movq    %rbx, -24(%rsp)
> +       subq    $24, %rsp
> +       cfi_adjust_cfa_offset(24)
>         mov     %RSP_LP, %RBX_LP
>         cfi_def_cfa_register(%rbx)
>         and     $-STATE_SAVE_ALIGNMENT, %RSP_LP
> @@ -153,6 +155,8 @@ _dl_tlsdesc_dynamic:
>  #if DL_RUNTIME_RESOLVE_REALIGN_STACK
>         mov     %RBX_LP, %RSP_LP
>         cfi_def_cfa_register(%rsp)
> +       addq    $24, %rsp
> +       cfi_adjust_cfa_offset(-24)
>         movq    -24(%rsp), %rbx
>         cfi_restore(%rbx)
>  #else
>
> base-commit: 5ebc24f785dc0dff494a93ca82a369497c3cdc68
>


-- 
H.J.