[PATCH][BUGFIX] cgroups: more safe tasklist locking in cgroup_attach_proc

Ben Blum bblum at andrew.cmu.edu
Mon Aug 15 16:09:00 PDT 2011


On Mon, Aug 15, 2011 at 07:04:15PM -0400, Ben Blum wrote:
> On Tue, Aug 16, 2011 at 12:50:06AM +0200, Frederic Weisbecker wrote:
> > On Mon, Aug 15, 2011 at 08:49:57PM +0200, Oleg Nesterov wrote:
> > > On 07/29, Ben Blum wrote:
> > > >
> > > > According to this thread - https://lkml.org/lkml/2011/7/27/243 - RCU is
> > > > not sufficient to guarantee the tasklist is stable w.r.t. de_thread and
> > > > exit. Taking tasklist_lock for reading, instead of rcu_read_lock,
> > > > ensures proper exclusion.
> > > 
> > > Yes.
> > > 
> > > So far I still think we should fix while_each_thread() so that it works
> > > under rcu_read_lock() "as exepected", I'll try to think more.
> > > 
> > > But whatever we do with while_each_thread(), this can't help
> > > cgroup_attach_proc(), it needs the locking.
> > > 
> > > > -	rcu_read_lock();
> > > > +	read_lock(&tasklist_lock);
> > > >  	if (!thread_group_leader(leader)) {
> > > 
> > > Agreed, this should work.
> > > 
> > > But can't we avoid the global list? thread_group_leader() or not, we do
> > > not really care. We only need to ensure we can safely find all threads.
> > > 
> > > How about the patch below?
> > > 
> > > 
> > > With or without this/your patch this leader can die right after we
> > > drop the lock. ss->can_attach(leader) and ss->attach(leader) look
> > > suspicious. If a sub-thread execs, this task_struct has nothing to
> > > do with the threadgroup.
> > > 
> > > 
> > > 
> > > Also. This is off-topic, but... Why cgroup_attach_proc() and
> > > cgroup_attach_task() do ->attach_task() + cgroup_task_migrate()
> > > in the different order? cgroup_attach_proc() looks wrong even
> > > if currently doesn't matter.
> > 
> > Right. As we concluded in our off-list discussion, if there
> > is no strong reason for that, I'm going to fix that in my task
> > counter patchset because there it really matters. If we can't
> > migrate the thread because it has already exited, we really
> > don't want to call ->attach_task() but rather cancel_attach_task().
> > 
> > Thanks.
> > 
> 
> Yes. Um, this must have been a mistake on my part. The lines of code
> should be the other way around. It should be done in a separate bugfix
> patch, though, so it goes through faster...
> 
> -- Ben
> 

also, there is no cancel_attach_task, afaict. cancel_attach is only for
memcg, for the whole operation at once - unless you are changing this,
in which case feel free to modify on top of the patch I'm about to send
out.


More information about the Containers mailing list