Re: [PATCH] ftrace: do_each_pid_task() needs rcu lock

[Oleg Nesterov <oleg@redhat.com>]

On 02/03, Eric W. Biederman wrote:
>
> Oleg Nesterov <oleg@redhat.com> writes:
>
> > On 02/03, Oleg Nesterov wrote:
> >>
> >> "ftrace: use struct pid" commit 978f3a45d9499c7a447ca7615455cefb63d44165
> >> converted ftrace_pid_trace to "struct pid*". But we can't use
> >> do_each_pid_task() without rcu_read_lock() even if we know the pid
> >> itself can't go away (it was pinned in ftrace_pid_write). The exiting
> >> task can detach itself from this pid at any moment.
> >
> > Q: why do we use do_each_pid_task(PIDTYPE_PID) ? We can never have more
> > than 1 task in the loop. Perhaps,
>
> That is a bug in do_each_pid_task(PIDTYPE_PID).
> For ftrace we really want to grab all tasks with a given pid even
> in the crazy exec case.

Yes, I thought about de_thread() too. But we can't "fix" do_each_pid_task()
to avoid the race?

IOW. If we want to continue to trace the task with the same pid after
exec reliably, then we should do something like

	void ftrace_transfer_trace(struct task_struct *leader)
	{
		mutex_lock(&ftrace_start_lock);
		if (test_tsk_trace_trace(leader))
			set_tsk_trace_trace(current);
		mutex_unlock(&ftrace_start_lock);
	}

and, in de_thread,

		write_unlock_irq(&tasklist_lock);
	+
	+	ftrace_transfer_trace(leader);
	+
		release_task(leader);

No?

(the above is not right of course, we can race with clear_ftrace_pid(),
 just for illustration)

Oleg.