Re: + cgroups-add-functionality-to-read-write-lock-clone_thread-forking-pe r-threadgroup.patch added to -mm tree

[Oleg Nesterov <oleg@redhat.com>]

In case I wasn't clear.

Let's suppose we have subthreads T1 and T2, and we have a reference to T1.
T1->thread_group->next == T2.

T1 dies, T1->thread_group->next is still T2.

T2 dies, rcu passed, its memory is freed and and re-used.
But T1->thread_group->next is still T2.

Now, we call threadgroup_fork_lock(T1), it sees T1->sighand == NULL and does

	rcu_read_lock();
	list_for_each_entry_rcu(T1->thread_group);

T1->thread_group->next points to nowhere.


Once again, I didn't actually read these patches, perhaps I missed something.

Oleg.

On 08/21, Oleg Nesterov wrote:
>
> On 08/20, Andrew Morton wrote:
> >
> > Subject: cgroups: add functionality to read/write lock CLONE_THREAD fork()ing per-threadgroup
> > From: Ben Blum <bblum@google.com>
> >
> > Add an rwsem that lives in a threadgroup's sighand_struct (next to the
> > sighand's atomic count, to piggyback on its cacheline), and two functions
> > in kernel/cgroup.c (for now) for easily+safely obtaining and releasing it.
> 
> Sorry. Currently I have no time to read these patched. Absolutely :/
> 
> But the very first change I noticed outside of cgroups.[ch] looks very wrong,
> 
> > +struct sighand_struct *threadgroup_fork_lock(struct task_struct *tsk)
> > +{
> > +	struct sighand_struct *sighand;
> > +	struct task_struct *p;
> > +
> > +	/* tasklist lock protects sighand_struct's disappearance in exit(). */
> > +	read_lock(&tasklist_lock);
> > +	if (likely(tsk->sighand)) {
> > +		/* simple case - check the thread we were given first */
> > +		sighand = tsk->sighand;
> > +	} else {
> > +		sighand = NULL;
> > +		/*
> > +		 * tsk is exiting; try to find another thread in the group
> > +		 * whose sighand pointer is still alive.
> > +		 */
> > +		rcu_read_lock();
> > +		list_for_each_entry_rcu(p, &tsk->thread_group, thread_group) {
> 
> If ->sighand == NULL we can't use list_for_each_entry_rcu(->thread_group),
> and rcu_read_lock() can't help.
> 
> The task was removed from ->thread_group, its ->next points to nowhere.
> 
> list_for_rcu(head) can _only_ work if we can trust head->next: it should
> point either to "head" (list_empty), or to the valid entry.
> 
> Please correct me if I missed something.
> 
> Otherwise, please send the changes which touch the process-management
> code separately. And please do not forget to CC people who work with
> this code ;)
> 
> Oleg.