Re: [PATCH -mm 3/3] proc: make task_sig() lockless

[Oleg Nesterov <oleg@redhat.com>]

On 04/09, Roland McGrath wrote:
>
> > Yes. From the changelog:
> >
> > 	Of course, this means we read pending/blocked/etc nonatomically,
> > 	but I hope this is OK for fs/proc.
> >
> > But I don't think the returned data could be "really" inconsistent
> > from the /bin/ps pov. Yes, it is possible that, say, some signal is
> > seen as both pending and ignored without ->siglock. Or we can report
> > user->sigpending != 0 while pending/shpending are empty.
> >
> > But this looks harmless to me. We never guaranteed /proc/pid/status
> > can't report the "intermediate" state, and I don't think we can
> > confuse the user-space.
> >
> > Do you agree? Or do you think this can make problems ?
>
> I'm not so sure.  Operations like sigprocmask and sigaction really have
> always been entirely atomic from the userland perspective before.  Now it
> becomes possible to read from /proc e.g. a blocked set that never existed
> as such (one word updated by sigprocmask but not yet the next word).

Yes, /proc/pid/status can report the intermediate state, I even sent
the updated changelog to document this.

But if you are not sure this is OK, I am worried. Do you think we should
drop this patch? If yes, I won't argue.

Oleg.