64-bit time_t and __WORDSIZE_TIME64_COMPAT32

[Florian Weimer]

* Thorsten Kukuk via Libc-alpha:

> I think we don't lose anything if btmp and lastlog don't exist anymore.
> Querying utmp for the runlevel in systemd times is also obsolete.
> Statistical data how many users are logged in if you login or if you
> call uptime is also only nice to have, nothing more.

This is something that systemd could do, in a far more reliable fashion
(or the kernel audit subsystem, I assume).

> So the main consumer of utmp is glibc itself: getlogin()/getlogin_r()
> Is there any idea how to replace that, if utmp/wtmp gets deprecated?

We use /proc/self/loginuid already, and try to resolve that UID using
NSS.  We still have fallback to __getutline_r.  I think it would be
really helpful if you could discover why that fallback is used.  I would
have thought it's dead code.

Is it because /proc/self/loginuid is missing, or because NSS fails for
some reason?  If it's about /proc/self/loginuid, maybe we can look at
the owner of /dev/tty instead.

> That's the only use case for which I don't see an alternate currently.
> The musl libc way to use an environment variable is really not safe...

The manual page has strong guidance that getlogin is not to be trusted.
/proc/self/loginuid is probably quite hard to fool, though, but the
fallback code is another matter.

> Another real usecase is to find the TTYs for wall and wall alike
> messages by tools like shutdown, systemd and similar.

Again, this seems to be more in the realm of the init service.

> Functions like login()/logout() are only used by openssh. So should be
> solveable, too.

Also FTP servers.  But we have seen quite a few problems due to account
database updates because the data structure isn't great, and isn't
really suitable to handling PTYs.

Thanks,
Florian