Re: AX25 patches and how it affects the end user

[Ralf Baechle DL5RB <ralf@linux-mips.org>]

On Wed, Jan 18, 2006 at 07:26:45AM -0500, Chuck Hast wrote:

> > > After applying mkiss patch to kernel 2.6.15.1 I compiled it for a 3 GHz
> > > Xeon P4 configuring it for SMP and multithread plus lock options.
> > >
> > > The SMP kernel seems very sensitive to AX25 configuration errors and it
> > > locks up quite soon in that case when loading applications.
> > >
> > > However when ax25 is carefully initialized, mkiss, kissattach, ax25ipd
> > > and ROSE/FPAC switch software suite) are running without problem.
> > >
> > > But there is still a spinlock lockup when shutting down the system.
> > >
> > > Here is a copy of the sequence I made by hand (subject to errors) :
> >
> > Thanks.  Btw, in some cases digital cameras have served well to catch
> > messages from screens without types - just make sure the images aren't
> > larger than necessary to be readable.
> >
> > > Spinlock lockup on CPU#0, kissattach / 5048, f8d68714
> > > EIP <c01e3ba5>
> > > rose_remove_neigh + 0x30/0xb0 [rose]
> > > rose_rt_device_down + 0xeb / 0x120 [rose]
> > > rose_device_event + 0x42/0x50 [rose]
> > > notifier_call_chain + 0x2/0x50 [rose]
> > > dev_close + 0x7b/0xb0
> > > unregister_netdevice + 0x19e/0x250
> > > unregister_netdev+0x16/0x1d
> > > mkiss_close + 0x4a /0xa0 [mkiss]
> > > release_dev....
> > > tty_release ...
> > >
> > > Hope this can help.
> > > Please suggest any more test to be done.
> >
> > I think the ROSE routing code is beyond recovery.  The current data
> > structures requires extensive locking code that is easily prone to
> > deadlocks like this.  It also is slow - fortunately nobody is using ROSE
> > on highspeed links ...
> >
> > Anyway, the problem is pretty obvious in your traceback and I'll cook a
> > patch for you to test.
> >
> 
> What do you consider high speed links? We will be doing so in Florida
> as we move the switches around the state to Linux. Some of these devices
> will be linked over 802.11 type links.

That certainly would be a fast link.  There are not only the limits of
the code itself but also the design limits of the connected mode protocols
themselves which become a theoretical limit for what is possible with
AX.25.

  Ralf