Re: [PATCH 1/5] sched: Convert thread_group_cputime() to use for_each_thread()

[Oleg Nesterov <oleg@redhat.com>]

On 04/09, Peter Zijlstra wrote:
>
> On Wed, Apr 09, 2014 at 06:11:18PM +0200, Frederic Weisbecker wrote:
> > do_each_thread/while_each_thread iterators are deprecated by
> > for_each_thread/for_each_process_thread() APIs.
> >
> > Lets convert the callers in the cputime code accounting. The ultimate
> > goal is to remove the struct task_struct::thread_group field and
> > the corresponding do_each_thread/while_each_thread iterators that are
> > RCU unsafe.
> >
> > It also makes thread_group_cputime() eventually RCU-safe.
>
> this fails to explain how the current code is broken.

while_each_thread(g, t) will loop forever if g exits and removes itself
from ->thread_group. This can happen even if it is the group leader,
de_thread() can do this.

Another problem is that it is used wrongly very often, even if
while_each_thread() was fine people forget to check pid_alive() to ensure
it didn't exit even before we take rcu_read_lock().

for_each_thread(p, t) is always safe. Unless p's task_struct can't go away,
of course.

But there is a difference. Ignoring the bug above

	p = g;
	do {
		printk("pid=%d\n", p->pid);
	} while_each_thread(g, p);

always prints at least one pid.

	for_each_thread(g, p)
		printk("pid=%d\n", p->pid);

can print nothing if g has already exited.

Oleg.