Re: [PATCHv4 05/17] pps: access pps device by direct pointer

[Rodolfo Giometti <giometti@enneenne.com>]

On Sun, Nov 21, 2010 at 01:33:27AM +0300, Alexander Gordeev wrote:
> > > -	if (ret < 0) {
> > > +	if (pps == NULL) {
> > >  		pr_err("cannot register PPS source \"%s\"\n", info.path);
> > > -		return ret;
> > > +		return -ENOMEM;
> > >  	}
> > > -	tty->disc_data = (void *)(long)ret;
> > > +
> > > +	spin_lock_irq(&pps_ldisc_lock);
> > > +	tty->disc_data = pps;
> > > +	spin_unlock_irq(&pps_ldisc_lock);
> > 
> > Maybe this lock is useless... however, are we sure that before setting
> > tty->disc_data to pps its value is null? Otherwise the dcd_change may
> > be called with an oops! We cannot control serial port IRQ
> > generation... :-/
> 
> No, locking here is necessary.
> There is only one problem this spinlock protects us from: current tty
> code neither disables interrupts nor doesn't ensure there are no
> references to PPS ldisc from uart_handle_dcd_change() before closing it
> (and removing PPS source). It relies on flushing workqueue and disabling
> input. It worked good this way until dcd_change() was added which
> doesn't use workqueues and is called in atomic context so can't lock
> on mutex.
> 
> Imagine that (on SMP system) uart_handle_dcd_change() could obtain a
> reference to ldisc and call dcd_change() until actually calling
> pps_event(); then on another processor all the path from
> tty_ldisc_halt() until tty_ldisc_stop() is executed. And then
> pps_event() is called with illegal pps pointer.
> 
> I just thought you are right that disc_data can be set not NULL by
> another ldisc and it's a problem. But actually I just realised how to
> fix it completely. :)
> 
> I just have to add a spinlock to tty_struct, lock all the
> uart_handle_dcd_change() with it and add a "barrier" between
> tty_ldisc_halt() and tty_ldisc_close() i.e. just that:
> 
> ...
> spin_lock_irq();
> spin_unlock_irq();
> ...
> 
> This "barrier" will ensure that there is no references to ldisc from
> uart_handle_dcd_change(). It won't be able to obtain a new reference
> after tty_ldisc_halt() so will become completely sane. Not disabling
> interrupts won't be a problem because it won't be able to obtain an
> ldisc reference until tty_ldisc_enable() which is called only after the
> new ldisc is fully functional. If it's our ldisc than it will have both
> dcd_change defined and a valid pps pointer. If it's not our ldisc it
> won't have both so uart_handle_dcd_change() won't call dcd_change() at
> all.
> 
> I think I'll do that as a separate patch.

Excuse me but IMHO you should solve all your problems if you do the lock into
pps_tty_init/cleanup instead of into pps_tty_open/close.

   spin_lock_irq();
           err = tty_register_ldisc(N_PPS, &pps_ldisc_ops);
           if (err)
                   pr_err("can't register PPS line discipline\n");
           else
                   pr_info("PPS line discipline registered\n");
   spin_unlock_irq();

And the same into cleanup.

Ciao,

Rodolfo

-- 

GNU/Linux Solutions                  e-mail: giometti@enneenne.com
Linux Device Driver                          giometti@linux.it
Embedded Systems                     phone:  +39 349 2432127
UNIX programming                     skype:  rodolfo.giometti
Freelance ICT Italia - Consulente ICT Italia - www.consulenti-ict.it