Hi, Ammar
> The ABI mandates that the %esp register must be a multiple of 16 when
> executing a 'call' instruction.
>
> Commit 2ab446336b17 ("tools/nolibc: i386: shrink _start with _start_c")
> simplified the _start function, but it didn't take care of the %esp
> alignment, causing SIGSEGV on SSE and AVX programs that use aligned move
> instruction (e.g., movdqa, movaps, and vmovdqa).
>
As Thomas suggested and you replied, since there is no public report
link, what about uses the link which has test code at the end of above
paragraph.
...
instruction (e.g., movdqa, movaps, and vmovdqa) [1].
> The 'and $-16, %esp' aligns the %esp at a multiple of 16. Then 'push
> %eax' will subtract the %esp by 4; thus, it breaks the 16-byte
> alignment. Make sure the %esp is correctly aligned after the push by
> subtracting 12 before the push.
>
And at the end of commit message:
[1]: https://lore.kernel.org/lkml/[email protected]/
This test code may be important for future change verification.
Thanks,
Zhangjin
> Cc: Zhangjin Wu <[email protected]>
> Fixes: 2ab446336b17aad362c6decee29b4efd83a01979 ("tools/nolibc: i386: shrink _start with _start_c")
> Reported-by: Nicholas Rosenberg <[email protected]>
> Acked-by: Thomas Weißschuh <[email protected]>
> Signed-off-by: Ammar Faizi <[email protected]>
> ---
> tools/include/nolibc/arch-i386.h | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/tools/include/nolibc/arch-i386.h b/tools/include/nolibc/arch-i386.h
> index 64415b9fac77f996..8079974424fa18b0 100644
> --- a/tools/include/nolibc/arch-i386.h
> +++ b/tools/include/nolibc/arch-i386.h
> @@ -167,7 +167,8 @@ void __attribute__((weak, noreturn, optimize("Os", "omit-frame-pointer"))) __no_
> __asm__ volatile (
> "xor %ebp, %ebp\n" /* zero the stack frame */
> "mov %esp, %eax\n" /* save stack pointer to %eax, as arg1 of _start_c */
> - "and $-16, %esp\n" /* last pushed argument must be 16-byte aligned */
> + "and $-16, %esp\n" /* align stack to 16 bytes */
> + "sub $(16 - 4), %esp\n" /* push %eax breaks 16-byte alignment, so sub 12 bytes */
> "push %eax\n" /* push arg1 on stack to support plain stack modes too */
> "call _start_c\n" /* transfer to c runtime */
> "hlt\n" /* ensure it does not return */
> --
> Ammar Faizi
On Tue, Aug 29, 2023 at 07:17:10PM +0700, Ammar Faizi wrote:
> On Tue, Aug 29, 2023 at 08:21:47AM +0200, Willy Tarreau wrote:
> > Indeed, good catch! However if we want to do it cleany (i.e not punch a
> > 16 to 28 byes hole in the stack), we should do this instead:
> >
> > add $12, %esp /* the stack must be aligned to 16 ... */
> > and $-16, %esp /* ... bytes after eax is pushed and ... */
> > sub $12, %esp /* ... before the call. */
> >
> > This will only add 0 to 12 bytes depending on the existing alignment.
> >
> > What do you think ?
>
> Good point. I'll send a v3 revision tomorrow.
OK thanks!
> I just saw that Linus has pulled the PR from Shuah that contains this
> bug. IOW, I missed this fix for the 6.6 merge window. Let's see if it
> can go to 6.6-rc2. Or maybe sooner than that.
No worries, we all know that -rc1 gets more exposure than individual
branches and raises bugs like this one.
Cheers,
Willy