Re: [PATCH 3/3] pids: refactor vnr/nr_ns helpers to make them safe

[Louis Rilling <Louis.Rilling@kerlabs.com>]

On Fri, Jan 16, 2009 at 09:45:40PM +0100, Oleg Nesterov wrote:
> Hi Louis,
> 
> On 01/16, Louis Rilling wrote:
> >
> > On 16/01/09  6:55 +0100, Oleg Nesterov wrote:
> > > +			struct pid_namespace *ns)
> > >  {
> > > -	return pid_nr_ns(task_pid(tsk), ns);
> > > +	pid_t nr = 0;
> > > +
> > > +	rcu_read_lock();
> > > +	if (!ns)
> > > +		ns = current->nsproxy->pid_ns;
> > > +	if (likely(pid_alive(task))) {
> >
> > I don't see what this pid_alive() check buys you. Since tasklist_lock is not
> > enforced, nothing prevents another CPU from detaching the pid right after the
> > check.
> 
> pid_alive() should be renamed. We use it to make sure the task didn't pass
> __unhash_process().
> 
> Yes, you are right, nothing prevents another CPU from detaching the pid right
> after the  check. But this is fine: we read ->pids[].pid under rcu_read_lock(),
> and if it is NULL pid_nr_ns() returns. So, we don't need pid_alive() check at
> all.
> 
> However, we can not use task->group_leader unless we verify the task is still
> alive. That is why we need this check. We do not clear ->group_leader when
> the task exits, so we can't do
> 
> 		rcu_read_lock();
> 		if (task->group_leader)
> 			do_something(task->group_leader);
> 		rcu_unread_lock();
> 
> Instead we use pid_alive() before using ->group_leader.

Ok I see now. Since RCU is locked and pid_alive(task) has been true at
some point, task->group_leader cannot be freed because release_task()
does not release task->group_leader before releasing task. IOW
release_task() can't have started releasing task->group_leader before
rcu_read_lock().

Thank you for your explanation!

> 
> > I'm also a bit puzzled by your description with using tasklist_lock when task !=
> > current, and not seeing tasklist_lock anywhere in the patch. Does this mean that
> > "safe" is for "no access to freed memory is done, but caller has to take
> > tasklist_lock or may get 0 as return value"?
> 
> I am not sure I understand the question...
> 
> This patch doesn't use tasklist, it relies on rcu. With this patch the caller
> doesn't need tasklist/rcu to call these helpers (but of course, the caller
> must ensure that task_struct is stable).
> 
> But, whatever the caller does, it can get 0 as return value anyway if the
> task exists, this is correct. Or I misunderstood you?

My question was probably badly phrased. When reading the patch
description I thought that the point was to fix the helpers so that
tasklist_lock would be taken whenever task != current. Of course your
patch does not do this, and I'm perfectly fine with this.

Thanks,

Louis

-- 
Dr Louis Rilling			Kerlabs - IRISA
Skype: louis.rilling			Campus Universitaire de Beaulieu
Phone: (+33|0) 2 99 84 71 52		Avenue du General Leclerc
Fax: (+33|0) 2 99 84 71 71		35042 Rennes CEDEX - France
http://www.kerlabs.com/