Hi!
> > If frozen is atomic_t, do we need memory barrier?
>
> I think so. For example on x86-64 atomic_read() is just a read.
I'm not sure, but for x86-64 barriers are nops, anyway, IIRC.
> > > @@ -128,6 +135,21 @@ static unsigned int try_to_freeze_tasks(
> > > } while_each_thread(g, p);
> > > read_unlock(&tasklist_lock);
> > > yield(); /* Yield is okay here */
> > > + if (!todo) {
> > > + /* Make sure that none of the stopped processes has
> > > + * received the continuation signal after we checked
> > > + * last time.
> > > + */
> >
> > I do not like the counting idea; it should be simpler to just check if
> > all the processes are still stopped.
>
> I thought about that but didn't invent anything reasonable enough.
>
> > But I'm not sure if this is enough. What if signal is being delivered
> > on another CPU while freezing, still being delivered while this second
> > check runs, and then SIGCONT is delivered?
>
> Hm, is this possible in practice? I mean, if todo is 0 and nr_stopped doesn't
> change, then there are no processes that can send the SIGCONT (unless someone
> creates a kernel thread with PF_NOFREEZE that will do just that).
No, SIGCONT was sent before freezing even started, and for some reason
takes long time on other CPU.
[Of course it is going to be quite hard to hit that race _in practice_
and mdelay(1000) before check would solve it for practical
purposes...?]
> Anyway, for now I've no idea how to fix this properly. Will think about it
> tomorrow.
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
