lkml.org 
[lkml]   [2010]   [Nov]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCHv4 05/17] pps: access pps device by direct pointer
    В Sun, 21 Nov 2010 09:26:58 +0100
    Rodolfo Giometti <giometti@enneenne.com> пишет:

    > 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.

    Sorry, I don't understand how this is supposed to help. Anyway, a new
    patch is underway.

    --
    Alexander
    [unhandled content-type:application/pgp-signature]
    \
     
     \ /
      Last update: 2010-11-22 16:03    [W:2.983 / U:0.156 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site