[systemd-devel] PATCH: fix logind on xen

[Frederic Crozat]

Le vendredi 06 janvier 2012 à 18:16 +0100, Lennart Poettering a écrit :
> On Fri, 06.01.12 14:59, Frederic Crozat (fcrozat at suse.com) wrote:
> 
> > 
> > Le mardi 03 janvier 2012 à 21:52 +0100, Lennart Poettering a écrit :
> > > On Tue, 03.01.12 21:35, Lennart Poettering (lennart at poettering.net) wrote:
> > > 
> > > > > > currently, logind was enforcing the presence of /dev/tty0 to start
> > > > > > properly. This device is not present on Xen (when using xencons=tty) or
> > > > > > S/390.
> > > > > 
> > > > > Here is a regenerated version against master, since logind was moved to
> > > > > its own directory.
> > > > 
> > > > Thanks for rebasing the patch!
> > > > 
> > > > I fear the patch is not complete though. seat_read_active_vt() in
> > > > login/logind-seat.c needs to guard against the invalid fd in
> > > > s->manager->console_active_fd.
> > > > 
> > > > Also, I assume that if /dev/tty0 doesn't exist /dev/tty1, /dev/tty2 and
> > > > so on don't exist either. That means calls like seat_preallocate_vts()
> > > > need to be shortcut in this case, too.
> > > > 
> > > > logind currently implicitly and always creates a seat0, that exists
> > > > unconditionally, and that all hw that isn't assigned to anything else
> > > > belongs to. That notion is probably nothing we could or should get rid
> > > > off that easily, but that means that we need to make sure that a couple
> > > > of its operations become NOPs on the systems in question. Besides
> > > > seat_preallocate_vts() that's  seat_read_active_vt() and the whole logic
> > > > that watches VCSA devices (which should be shortcut, as if n_autovts was
> > > > 0, in manager_connect_udev()).
> > > 
> > > I have now commited a patch which reworks a lot of the logic there and
> > > tries to handle the no-VT case as gracefully as possible. We still
> > > implicitly create seat0, but we now stop advertising that it was
> > > multi-session capable. Hence we still end up with a seat, but only with
> > > the minimal properties that we need. This makes most of the other
> > > explicit checks unnecessary fortunately.
> > 
> > Hmm, I've tested this patch (I'm extracted the patche you did for it and
> > applied to our v37 package, thanks to git ;) and from what I see,
> > "console" login doesn't get any seat attached (but other login, like
> > over ssh are getting one), unlike my initial patch. So more work is
> > needed somehow.
> 
> "console" logins? What exactly is that? Logins on /dev/console? Where
> does /dev/console point to? i.e. what is the contents of
> /sys/class/tty/console/active if you do that?

Sorry, I was unclear. On Xen, you can login on a "text" terminal. When
doing that, /sys/class/tty/console/active outputs : tty-1 tty0

With your patch, no session is assigned to this particular login.

> Note that we consider serial logins much like ssh logins as "virtual",
> i.e. not physical, and hence with no seat assigned.

Wrong choice of word on my side, I was meaning : ssh login are seen as
"sessions without seats" in systemd-loginctl.

-- 
Frederic Crozat <fcrozat at suse.com>
SUSE