Re: [PATCHv2] backing-dev: fix wakeup timer races with bdi_unregister()

[Rabin Vincent <rabin@rab.in>]

On Fri, Jan 20, 2012 at 03:15:32PM +0900, Namjae Jeon wrote:
> 2012/1/20 Rabin Vincent <rabin@rab.in>:
> > On Fri, Jan 20, 2012 at 05:16, Namjae Jeon <linkinjeon@gmail.com> wrote:
> >>>                bdi_debug_unregister(bdi);
> >>> -               device_unregister(bdi->dev);
> >>> +
> >>> +               spin_lock_bh(&bdi->wb_lock);
> >>>                bdi->dev = NULL;
> >>> +               spin_unlock_bh(&bdi->wb_lock);
> >> Hi.
> >> Would you explain me why you add spinlock in here ?
> >
> > wakeup_timer_fn() does the following, where the
> > trace_writeback_wake_forker_thread() also accesses bdi->dev.
> > It does this under the wb_lock:
> >
> >        } else if (bdi->dev) {
> >                /*
> >                 * When bdi tasks are inactive for long time, they are killed.
> >                 * In this case we have to wake-up the forker thread which
> >                 * should create and run the bdi thread.
> >                 */
> >                trace_writeback_wake_forker_thread(bdi);
> >
> > If we don't have the lock above, the bdi->dev could potentially be
> > cleared after the check but before the tracepoint is hit, leading to a
> > NULL pointer dereference.
> Is there no possibility trace_writeback_wake_forker_thread is called
> after spin_unlock of bdi->de= null ?

wakeup_timer_fn() holds the wb_lock across the check for bdi->dev !=
NULL and the call to trace_writeback_wake_forker_thread(), so no.