Re: [Bugme-new] [Bug 7952] New: slattach only works every other time

[Stephen Hemminger <shemminger@linux-foundation.org>]

On Mon, 12 Feb 2007 09:36:09 +0100
Jarek Poplawski <jarkao2@o2.pl> wrote:

> On 06-02-2007 22:57, Andrew Morton wrote:
> ...
> > First time slattach is run to set up a SLIP line all is ok.
> > If slattach process is then killed and restarted it fails with message:
> > SLIP_set_disc(1): File exists
> > Problem still occurs in 2.6.20rc6 kernel
> > 
> > dmesg shows:
> > object_add failed for sl0 with -EEXIST, don't try to register things 
> > with the same name in the same directory.
> >  [<c01b7b54>] kobject_add+0x147/0x16d
> >  [<c0211209>] class_device_add+0x9d/0x3b3
> >  [<c022829d>] register_netdevice+0x21a/0x2d0
> >  [<c8903213>] slip_open+0x3a1/0x4e2 [slip]
> >  [<c01fc709>] tty_ioctl+0x922/0xbac
> ... 
> > Steps to reproduce:
> > (requires a serial port but nothing needs to be attached to it):
> > # slattach -L -vd -p slip -s 115200 /dev/ttyS0
> ...
> > slip started on /dev/ttyS0 interface sl0
> > 
> > Above is OK, now kill process with CTRL-C
> > 
> > slattach: tty_set_speed: 0
> > # slattach -L -vd -p slip -s 115200 /dev/ttyS0
> ...
> > SLIP_set_disc(1): File exists
> ...
> > I believe this is called by this changeset :
> > http://www2.kernel.org/git/?p=linux/kernel/git/stable/linux-2.6.17.y.git;a=commit;h=b17a7c179dd3ce7d04373fddf660eda21efc9db9
> 
> I think Martin is probably right here.
> 
> It would be useful to check if time has anything to do
> with this and wait longer (e.g. >= 1 min.) before the
> second slattach.
> 
> Anyway, even if there is some other reason, the above
> trace shows (IMHO) some inconsistency in register/
> unregister_netdevice: if class_device_add is reached
> it means the name is valid (so was unregistered) and
> EEXIST from netdev_register_sysfs is wrong about the
> state of this device. So maybe there should be some
> warning plus some delayed action instead of register
> cancelled?
> 
> Regards,
> Jarek P. 

The problem is that the code in sl_alloc() tries to clear out
an net device by calling unregister_netdevice(), the device won't
actually disappear until after rtnl_unlock. This whole idea of
searching for unused devices is racy crap and needs to go.