On Thu, 4 Apr 2019 08:00:24 +0100
Okash Khawaja <okash.khawaja@xxxxxxxxx> wrote:
> On Thu, 4 Apr 2019 00:21:34 +0200
> Samuel Thibault <samuel.thibault@xxxxxxxxxxxx> wrote:
>
> > But can't just just get the kref in set_selection before using
> > cmpxchg? and if that fails, put it back. If it succeeded, the work
> > schedule will happen, and either it will manage to take the tty and
> > put its kref, or be canceled, and the kref will be put as well.
>
> Thinking about it, we can also make use of return value of
> cancel_work_sync(). It returns true if work was pending. So, if no
> work was pending then we don't need to do anything as the kref would
> be balanced and the tty would be set to NULL by the time
> cancel_work_sync() returns. This would make speakup_cancel_selection()
> simpler:
>
> void speakup_cancel_selection(void)
> {
> struct tty_struct *tty;
> bool was_pending;
>
> was_pending = cancel_work_sync(&speakup_sel_work.work);
> if (!was_pending)
> return;
>
> tty_kref_put(speakup_sel_work.tty);
> speakup_sel_work.tty = NULL;
> }
>
> We will still need to get kref before cmpxchg() as you suggested
> above.
I think using xchg() in speakup_cancel_work() is better because it
covers a wider region (from cmpxchg() in speakup_set_selection()
onwards) as compared to checking return value of cancel_work_sync()
(which covers the region from schedule_work_on() in
speakup_set_selection()).
I'll send v3 in a bit.
Thanks,
Okash
_______________________________________________
Speakup mailing list
Speakup@xxxxxxxxxxxxxxxxx
http://linux-speakup.org/cgi-bin/mailman/listinfo/speakup
