Re: 3.9+ tty lockdep trace.

[Peter Hurley <peter@hurleysoftware.com>]

On 05/07/2013 02:00 PM, Dave Jones wrote:
> On Tue, May 07, 2013 at 01:35:36PM -0400, Peter Hurley wrote:
>   > On 05/07/2013 11:04 AM, Dave Jones wrote:
>   > > On Tue, May 07, 2013 at 10:14:09AM -0400, Peter Hurley wrote:
>   > >
>   > >   > >> Just hit this again, by switching tty's during bootup.
>   > >   > >> The previous time was switching from X to console iirc, so that seems to be
>   > >   > >> the key to triggering this.
>   > >   > >
>   > >   > > Sorry for not getting back to this sooner, merge window fun...
>   > >   > >
>   > >   > > Anyway, Peter, could this be something that your ldisc patches are
>   > >   > > triggering?
>   > >   >
>   > >   > Yes, although early analysis also suggests the lockdep splat is possible
>   > >   > regardless, just much less likely.
>   > >   >
>   > >   > I looked at this over the weekend, but I wasn't able to determine (in
>   > >   > the limited time I had available) if the splat was valid; ie., if the
>   > >   > cpu states could actually be concurrent and result in deadlock.
>   > >   >
>   > >   > I need to study the VT_DISALLOCATE ioctl in more detail which I plan
>   > >   > to do this week.
>   > >
>   > > I don't know if it proves/disproves anything, but shortly before the dump
>   > > gets printed, everything is kind of locked up. I can switch tty's, but
>   > > the gettys on them are dead to the world until after the dump.
>   > >
>   > > ISTR a pause for a while when I hit this switching from X too.
>   >
>   > It would help to know the steps to reproduce. For example, you mention
>   > unresponsive gettys but the VT_DISALLOCATE ioctl doesn't work (and isn't
>   > designed to) on open ttys.
>
> Hmm, dunno. I booted up, and tried to log in on tty1 (no X on that box), and after
> I entered my password everything just hung for a while. I flipped to tty2
> to try and log in as root, but after changing tty, I just got a blinking cursor.
> 10 seconds or so later, lockdep spew.

Ok, Dave, thanks for the extra information; that makes sense now.

systemd optionally uses the VT_DISALLOCATE ioctl to free the VT. So when
getty on tty2 died because of some error, systemd exercises the VT_DISALLOCATE
ioctl which generates the lockdep splat.

I'm still trying to figure out if the states can actually overlap; they're
not supposed to, but the VT_DISALLOCATE ioctl is not particularly
defensive and I suspect that overlap is possible.

It's somewhat unfortunate that systemd chose to use a really antiquated
and fundamentally unsafe ioctl like VT_DISALLOCATE. Even when I fix this
lockdep splat, bad things may still be possible.

I'm certain that the lockdep warning has nothing to do with your
login delay; if you want to send me sysrq+t output when that happens again,
I'd be willing to look over it for the underlying problem.

Regards,
Peter Hurley