Re: [PATCH 1/1] freezer: fix wait_event_freezable/__thaw_task races

[Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

Hi,

On 09/08, Tejun Heo wrote:
>
> Hello,
>
> On Wed, Sep 07, 2011 at 08:22:17PM +0200, Oleg Nesterov wrote:
> > wait_event_freezable() and wait_event_freezable_timeout() stop
> > the waiting if try_to_freeze() fails. This is not right, we can
> > race with __thaw_task() and in this case
> >
> > 	- wait_event_freezable() returns the wrong ERESTARTSYS
> >
> > 	- wait_event_freezable_timeout() can return the positive
> > 	  value while condition == F
>
> Indeed, nice catch.  This one actually is pretty dangerous.  We
> probably want to make a separate fix for this and backport it to
> -stable?

And I forgot to mention that wait_event_freezable_timeout() doesn't
handle -ERESTARTSYS at all.

But I don't think -stable needs this fix. According to grep, nobody
check the returned value, and none of the callers plays with signals.

> > Change the code to always check __retval/condition before return.
> >
> > Note: with or without this patch the timeout logic looks strange,
> > probably we should recalc timeout if try_to_freeze() returns T.
> >
> > Signed-off-by: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
>
> Yeap, with freezable_with_signal gone, this looks correct & simpler to
> me

I don't really understand this... set_freezable_with_signal() has a
lot of problems, yes... But even if it wasn't removed this fix makes
sense anyway, afaics.

If freezable_with_signal caller does wait_event_freezable_timeout(),
__retval becomes -ERESTARTSYS after freeze_task(). The next iteration
will return 0 with the KERN_ERR message from schedule_timeout().

> but it would be nice if you can sprinkle some documentation while
> at it. :)

But they already have the comment ;) What can I add?

Oleg.