Re: PROBLEM: with daemon.c after y2038 on 32-bits Kernel

[Arnd Bergmann <arnd@kernel.org>]

On 17/05/2022 09:51, Arnaud Panaïotis wrote:
> Hello,
>
> I'm working for a client to generate embedded 32-bits Linux Kernel working after y2038 issue.
>
> I generated a 5.15 Kernel thought Buildroot with Coreutils 9.0, GCC 11.2.0, Binutils 2.37, Glibc 2.34-9 and CFLAGS  -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64  -D_TIME_BITS=64.
>
> I encounter an issue while working with OpenSSH (I initially contacted them before).

To clarify: did you build just openssh with  -D_TIME_BITS=64, or did
you build the user space
this way?

> After 2038, /usr/sbin/sshd does not create an error but it child does generate this one:
> daemon() failed: Value too large for defined data type
>
> This happend here in sshd.c:
>
> 2019         /*
> 2020          * If not in debugging mode, not started from inetd and not already
> 2021          * daemonized (eg re-exec via SIGHUP), disconnect from the controlling
> 2022          * terminal, and fork.  The original process exits.
> 2023          */
> 2024         already_daemon = daemonized();
> 2025         if (!(debug_flag || inetd_flag || no_daemon_flag || already_daemon)) {
> 2026
> 2027                 if (daemon(0, 0) == -1)
> 2028                         fatal("daemon() failed: %.200s", strerror(errno));

My guess is that there are parts of glibc that are not fully
y2038-safe at the moment, but
merely provide the interfaces for time64 applications.


In the glibc code, I see

int
daemon (int nochdir, int noclose)
{
...
                if ((fd = __open_nocancel(_PATH_DEVNULL, O_RDWR, 0)) != -1
                    && (__builtin_expect (__fstat64 (fd, &st), 0)
                        == 0)) {
...
                             } else {
                        __close_nocancel_nostatus (fd);
                        return -1;
                }
     return (0);
}

__fstatat64 (int fd, const char *file, struct stat64 *buf, int flags)
{
  struct __stat64_t64 st_t64;
  return __fstatat64_time64 (fd, file, &st_t64, flags)
         ?: __cp_stat64_t64_stat64 (&st_t64, buf);
}

If I'm reading this correctly, daemon() internally uses the time32
version of 'stat', which
fails for files with out-of-range timestamps. Are you able to rebuild
the ssh binary
(or your entire distro, if that's easier) against musl-1.2.x instead
of glibc to see if the
same thing happens there?

       Arnd

> To reproduce:
>
> # date -s "2040-05-12"
> # hwclock --systohc
> # reboot
> # /usr/sbin/sshd
>
> Note this error occurs only after the reboot, and setting a date before 2038 also require a reboot to remove the error.
>
> strace and gdb trace linked.
>
> Let me know if you need additional information.