Re: [PATCH] TLS support for s390*.

[Jakub Jelinek <jakub at redhat dot com>]

On Thu, Jan 30, 2003 at 05:29:52PM +0100, Martin Schwidefsky wrote:
> 
> > BTW: It seems to work just fine if stack_addr passed to clone
> > is at least 8*16 bytes below end of particular mmaped area
> > (this is on s390x). My wild guess is that the register save
> > area is needed, which means clone.S should subtract 8*16
> > resp. 8*8 (on s390-32) bytes from stack_pointer before doing
> > the syscall.
> 
> Grr, it's the new clone implementation that stores %r6 on the
> stack. It reloads it after the svc because the father has to
> return with an unmodified %r6. But the restore of %r6 may not
> be done in the child! Four letter words again. Try to move
> the "lg %r6,48(%r15)" in clone.S after the two branches just
> before the "br %r14". That should do the trick.

Two more questions:

        lg      %r5,168(%r15)           /* load child_tid from stack */
        l       %r6,160(%r15)           /* load tls from stack */
        svc     SYS_ify(clone)
        lg      %r6,48(%r15)            /* restore %r6 */
        ltgr    %r2,%r2                 /* check return code */
        jgm     SYSCALL_ERROR_LABEL
        jz      thread_start
        br      %r14

1) Doesn't l (instead of lg) load just 32 bits from the stack
   (tls is void *)?
2) If the syscall fails, it should restore %r6 too, shouldn't it?
   Then it needs to be:
        svc     SYS_ify(clone)
        ltgr    %r2,%r2                 /* check return code */
        jz      thread_start
        lg      %r6,48(%r15)            /* restore %r6 */
        jgm     SYSCALL_ERROR_LABEL
        br      %r14

	Jakub