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

["H.J. Lu" <hjl.tools@gmail.com>]

On Sat, Mar 16, 2024 at 3:05 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Sat, Mar 16, 2024 at 10:51 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > On Sat, Mar 16, 2024 at 10:42 AM Florian Weimer <fweimer@redhat.com> wrote:
> > >
> > > * H. J. Lu:
> > >
> > > > Please verify if this is the right testcase.
> > >
> > > Test case works (fails without my fix, succeeds with my fix).  Some
> > > comments below.
> > >
> > > > diff --git a/sysdeps/x86_64/tst-gnu2-tls2-x86-64-mod0.S b/sysdeps/x86_64/tst-gnu2-tls2-x86-64-mod0.S
> > > > new file mode 100644
> > > > index 0000000000..8129b28061
> > > > --- /dev/null
> > > > +++ b/sysdeps/x86_64/tst-gnu2-tls2-x86-64-mod0.S
> > > > @@ -0,0 +1,57 @@
> > >
> > > > +     .text
> > > > +     .p2align 4
> > > > +     .globl  apply_tls
> > > > +     .type   apply_tls, @function
> > > > +apply_tls:
> > > > +     .cfi_startproc
> > >
> > > Missing CET marker.
> > >
> > > > +     subq    $24, %rsp
> > > > +     .cfi_def_cfa_offset 32
> > > > +     movdqu  (%rdi), %xmm0
> > > > +     movq    %fs:40, %rax
> > > > +     movq    %rax, 8(%rsp)
> > > > +     xorl    %eax, %eax
> > > > +     leaq    tls_var0@TLSDESC(%rip), %rax
> > > > +     call    *tls_var0@TLSCALL(%rax)
> > > > +     addq    %fs:0, %rax
> > > > +     movups  %xmm0, (%rax)
> > > > +     movdqu  16(%rdi), %xmm1
> > > > +     movups  %xmm1, 16(%rax)
> > > > +     movq    8(%rsp), %rdx
> > > > +     subq    %fs:40, %rdx
> > > > +     jne     .L5
> > > > +     addq    $24, %rsp
> > > > +     .cfi_remember_state
> > > > +     .cfi_def_cfa_offset 8
> > > > +     ret
> > > > +.L5:
> > > > +     .cfi_restore_state
> > > > +     call    __stack_chk_fail@PLT
> > >
> > > Not sure if we need this?
> > >
> > > Maybe add some comment what exactly this subtest is exercising?
> > >
> > > These are present in the other TLS modules as well.
> > >
> > > > diff --git a/sysdeps/x86_64/tst-gnu2-tls2-x86-64-mod1.S b/sysdeps/x86_64/tst-gnu2-tls2-x86-64-mod1.S
> > > > new file mode 100644
> > > > index 0000000000..af4b7ca761
> > > > --- /dev/null
> > > > +++ b/sysdeps/x86_64/tst-gnu2-tls2-x86-64-mod1.S
> > >
> > > > +/* Select an offset which will cause _dl_tlsdesc_dynamic_xsavec to
> > > > +   clobber %rbx.  */
> > > > +#define OFFSET (56 + 16 + 16 + 16)
> > > > +
> > > > +     .text
> > > > +     .p2align 4
> > > > +     .globl  apply_tls
> > > > +     .type   apply_tls, @function
> > > > +apply_tls:
> > > > +     .cfi_startproc
> > > > +     pushq   %rbp
> > > > +     .cfi_def_cfa_offset 16
> > > > +     .cfi_offset 6, -16
> > > > +     movq    %rsp, %rbp
> > > > +     .cfi_def_cfa_register 6
> > > > +     /* Align stack to 64 bytes.  */
> > > > +     andq    $-64, %rsp
> > > > +     pushq   %rbx
> > > > +     subq    $OFFSET, %rsp
> > >
> > > The offset could be loaded from a global variable or something like
> > > that.  We should exercise a wide range of stack alignments—the
> > > individual tests are cheap.  And maybe check extra registers.
> >
> > I will clean it up with a different fix.
> >
>
> I submitted a patch with a testase:
>
> https://patchwork.sourceware.org/project/glibc/list/?series=31963
>
> My patch allocates 64 more bytes to avoid clobbering saved RDI,
> RSI and RBX values on stack by xsave.  It avoids 2 stack
> adjustments.  Either my fix or Florian's fix should fix the issue.
> I don't have a strong preference as long as my testcase is
> included.
>
>
I think my testcase may fail on AMD AVX CPUs without the
fix.  On Intel AVX CPUs, the state size is 960 bytes.  But the
last 128 bytes may be unused.

-- 
H.J.