Bug 16713 - [s390x] Member "sa_flags" does not have the correct type

[Stefan Liebler]

Hi,

yes the kernel and glibc struct differs on s390. This is intended in 
order to get POSIX conform. I got kernel agreement before i changed the 
struct. The kernel has no problem with the __glibc_reserved0 variable 
and its random content.

Bye
Stefan

On 09/10/2014 01:31 AM, Mark Wielaard wrote:
> Hi,
>
> [Sorry for the duplicate Stefan, I got the list name wrong on the CC.]
>
> To fix the following bug:
> https://sourceware.org/bugzilla/show_bug.cgi?id=16713
> There was this simple change:
>
> --- a/sysdeps/unix/sysv/linux/s390/bits/sigaction.h
> +++ b/sysdeps/unix/sysv/linux/s390/bits/sigaction.h
> @@ -43,7 +43,8 @@ struct sigaction
>   #endif
>
>       /* Special flags.  */
> -    unsigned long int sa_flags;
> +    int __glibc_reserved0;
> +    int sa_flags;
>
>       /* Restore handler.  */
>       void (*sa_restorer) (void);
>
> But means glibc and the kernel don't agree anymore on the structure
> layout of struct sigaction. The kernel is still using unsigned long
> sa_flags on s390. I guess this still works out because all current
> SA_... flag values fit in an int. But it does mean the kernel cannot
> expand the flag field values anymore in the future.
>
> Because valgrind follows the kernel definition the above change causes
> valgrind memcheck to complain now whenever only the sa_flags value (as
> seen by user space) is assigned. For example in nptl-init.c:
>
>    struct sigaction sa;
>    sa.sa_sigaction = sigcancel_handler;
>    sa.sa_flags = SA_SIGINFO;
>    __sigemptyset (&sa.sa_mask);
>
>    (void) __libc_sigaction (SIGCANCEL, &sa, NULL);
>
> valgrind will now warn:
>
>   Syscall param rt_sigaction(act->sa_flags) points to uninitialised byte(s)
>      at 0x42EECC8: __libc_sigaction (sigaction.c:42)
>      by 0x42E2253: __pthread_initialize_minimal (nptl-init.c:381)
>      by 0x42E15EF: ??? (in /usr/lib64/libpthread-2.19.90.so)
>
> Because the '__glibc_reserved0' part of sa_flags as the kernel will see
> them is indeed undefined.
>
> We can easily change valgrind to follow the layout as now used by
> glibc/user space on s390 instead of the kernel layout and just ignore
> the __glibc_reserved0 field. But I wanted to first make sure this
> change/difference between glibc/kernel was intended.
>
> Maybe glibc wants to pass the struct as intended by the kernel and needs
> to create a "bridge" struct that properly initializes the kernel's view
> of sa_flags?
>
> Thanks,
>
> Mark
>
>