Re: [RFC PATCH 2/5] can: slcan: remove legacy infrastructure

[Max Staudt <max@enpas.org>]

On Mon, 25 Jul 2022 08:40:24 +0200
Dario Binacchi <dario.binacchi@amarulasolutions.com> wrote:

> > > @@ -883,72 +786,50 @@ static int slcan_open(struct tty_struct *tty)
> > >       if (!tty->ops->write)
> > >               return -EOPNOTSUPP;
> > >
> > > -     /* RTnetlink lock is misused here to serialize concurrent
> > > -      * opens of slcan channels. There are better ways, but it is
> > > -      * the simplest one.
> > > -      */
> > > -     rtnl_lock();
> > > +     dev = alloc_candev(sizeof(*sl), 1);
> > > +     if (!dev)
> > > +             return -ENFILE;
> > >
> > > -     /* Collect hanged up channels. */
> > > -     slc_sync();
> > > +     sl = netdev_priv(dev);
> > >
> > > -     sl = tty->disc_data;
> > > +     /* Configure TTY interface */
> > > +     tty->receive_room = 65536; /* We don't flow control */
> > > +     sl->rcount   = 0;
> > > +     sl->xleft    = 0;  
> >
> > I suggest moving the zeroing to slc_open() - i.e. to the netdev open
> > function. As a bonus, you can then remove the same two assignments from
> > slc_close() (see above). They are only used when netif_running(), with
> > appropiate guards already in place as far as I can see.  
> 
> I think it is better to keep the code as it is, since at the entry of
> the netdev
> open function, netif_running already returns true (it is set to true by the
> calling function) and therefore it would be less safe to reset the
> rcount and xleft
> fields.

Wow, great catch!

I wonder why __LINK_STATE_START is set before ->ndo_open() is called...?


Since the drivers are similar, I've checked can327. It is unaffected,
because the counters are additionally guarded by a spinlock. Same in
slcan, where netdev_close() takes the spinlock to reset the counters.

So you *could* move them to netdev_open() *if* they are always guarded
by the slcan lock.

Or, leave it as it is, as it seems to be correct. Your choice :)


Thank you!

Max