Re: Q: cgroup: Questions about possible issues in cgroup locking

From: Mandeep Singh Baines
Date: Thu Jan 19 2012 - 13:18:24 EST


Oleg Nesterov (oleg@xxxxxxxxxx) wrote:
> On 01/18, Mandeep Singh Baines wrote:
> >
> > Oleg Nesterov (oleg@xxxxxxxxxx) wrote:
> > > On 01/13, Mandeep Singh Baines wrote:
> > > >
> > > > Case 3: g is some other thread
> > > >
> > > > In this case, g MUST be current
> > >
> > > Why? This is not true.
> >
> > Here is my thinking:
> >
> > The terminating condition, t != g, assumes that you can get back to
> > g. If g is unhashed, there is no guarantee you'll ever get back to it.
> > Holding a reference does not prevent unhashing.
> >
> > for_each_process avoids unhashing by starting and ending at init_task
> > (which can never be unhashed).
>
> Yes,
>
> > As you pointed out a while back, this doesn't work for:
> >
> > do_each_thread(g, t){
> > do_something(t);
> > } while_each_thread(g, t)
> >
> > because g can be unhashed.
> >
> > However, you should be able to use while_each_thread if you are current.
> > Being current would prevent 'g' from being unhashed.
>
> Ah, I misunderstood you. Yes, sure.
>
> > other than, do_each_thread/while_each_thread, all other callers
> > of while_each_thread() are starting at current. Otherwise, how do
> > you guarantee that it terminates.
>
> Hm, still can't understand...
>

On second thought. I think I've made some incorrect assumptions.

> > I see at least one example, coredump_wait() that uses while_each_thread
> > starting at current. I didn't find any cases where while_each_thread
> > starts anywhere other than current or group_leader.
>
> Probably you meant zap_threads/zap_process, not coredump_wait?
>
> zap_process() is fine, we hold ->siglock. But zap_threads does _not_

Ah. Missed the siglock.

> start at current, may be you misread the g == tsk->group_leader check
> in the main for_each_process() loop ? But most probably we simply
> misunderstand each other a bit, see below.
>
> However it starts at ->group_leader, so it won't suffer if we restrict
> the lockless while_each_thread_rcu().
>
> > > But I guess this doesn't matter, I think I am
> > > starting to understand why our discussion was a bit confusing.
> > >
> > > The problem is _not_ exec/de_thread by itself. The problem is that
> > > while_each_thread(g, t) can race with removing 'g' from the list.
> > > IOW, list_for_each_rcu-like things assume that the head of the list
> > > is "stable" but this is not true.
> > >
> > > And note that de_thread() does not matter in this sense, only
> > > __unhash_process()->list_del_rcu(&p->thread_group) does matter.
> > >
> > > Now. If 'g' is the group leader, we should only worry about exec,
> > > otherwise it can't disappear before other threads.
> > >
> > > But if it is not the group leader, it can simply exit and
> > > while_each_thread(g, t) can never stop by the same reason.
> > >
> >
> > I think we are on the same page. Your explanation is consistent with
> > my understanding.
> >
> > Some other thoughts:
> >
> > I suspect that other than do_each_thread()/while_each_thread() or
> > for_each_thread()/while_each_thread() where 'g' is the group_leader,
> > 'g' MUST be current. So the only cases to consider are:
>
> I didn't try to grep, but I do not know any example of the lockless
> while_each_thread() which starts at current.
>

Did a quick scan of all the while_each_thread()s under rcu_read_lock
and couldn't find one that starts at current.

> I guess this is the source of confusion.
>
> > > I'll try to recheck my thinking once again, what do you think? Anything
> > > else we could miss?
> >
> > Yeah, the ->group_leader solution seems the most promising. It seems
> > correct (ignoring barriers) and should work for all supported cases:
> >
> > 1) when g is group_leader
> > 2) when g is current
>
> OK, thanks.
>
> I'll try to investigate if we can remove
>
> leader->group_leader = tsk;
>
> from de_thread(). In fact I already thought about this change a long
> ago without any connection to while_each_thread(). This assignment
> looks "assymetrical" compared to other threads we kill. But we did
> have a reason (reasons?). Hopefully, the only really important reason
> was already removed by 087806b1.
>

Ah. So the leader->group_leader may have been necessary earlier in order
to prevent two tasks, old leader and new leader from both returning true
for thread_group_leader(tsk).

Regards,
Mandeep

> Oleg.
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/