[Devel] Re: [PATCH 14/15] Destroy pid namespace on init's death

[Oleg Nesterov]

On 07/26, Pavel Emelyanov wrote:
>
> @@ -895,6 +915,7 @@ fastcall NORET_TYPE void do_exit(long co
> {
> 	struct task_struct *tsk = current;
> 	int group_dead;
> +	struct pid_namespace *pid_ns = tsk->nsproxy->pid_ns;
> 
> 	profile_task_exit(tsk);
> 
> @@ -905,9 +926,10 @@ fastcall NORET_TYPE void do_exit(long co
> 	if (unlikely(!tsk->pid))
> 		panic("Attempted to kill the idle task!");
> 	if (unlikely(tsk == task_child_reaper(tsk))) {
> -		if (task_active_pid_ns(tsk) != &init_pid_ns)
> -			task_active_pid_ns(tsk)->child_reaper =
> -					init_pid_ns.child_reaper;
> +		if (pid_ns != &init_pid_ns) {
> +			zap_pid_ns_processes(pid_ns);
> +			pid_ns->child_reaper = init_pid_ns.child_reaper;
> +		}
> 		else
> 			panic("Attempted to kill init!");

No, no, this is wrong. Yes, the current code is buggy too, I'll send
the fix.

I think this code should be moved below under the "if (group_dead)",
and we should use tsk->group_leader.

> +void zap_pid_ns_processes(struct pid_namespace *pid_ns)
> +{
> +	int i;
> +	int nr;
> +	int nfree;
> +	int options = WNOHANG|WEXITED|__WALL;
> +
> +repeat:
> +	/*
> +	 * We know pid == 1 is terminating. Find remaining pid_ts
> +	 * in the namespace, signal them and then wait for them
> +	 * exit.
> +	 */
> +	nr = next_pidmap(pid_ns, 1);
> +	while (nr > 0) {
> +		kill_proc_info(SIGKILL, SEND_SIG_PRIV, nr);
> +		nr = next_pidmap(pid_ns, nr);
> +	}
> +
> +	nr = next_pidmap(pid_ns, 1);
> +	while (nr > 0) {
> +		do_wait(nr, options, NULL, NULL, NULL);

When the first child of init exits, it sends SIGCHLD. After that,
do_wait() will never sleep, so we are doing a busy-wait loop.
Not good, especially when we have a niced child, can livelock.

> +		nr = next_pidmap(pid_ns, nr);
> +	}
> +
> +	nfree = 0;
> +	for (i = 0; i < PIDMAP_ENTRIES; i++)
> +		nfree += atomic_read(&pid_ns->pidmap[i].nr_free);
> +
> +	/*
> +	 * If pidmap has entries for processes other than 0 and 1, retry.
> +	 */
> +	if (nfree < (BITS_PER_PAGE * PIDMAP_ENTRIES - 2))
> +		goto repeat;

This doesn't look right.

Suppose that some "struct pid" was pinned from the parent namespace.
In that case zap_pid_ns_processes() will burn CPU until put_pid(), bad.

I think we can rely on forget_original_child() and do something like
this:

	zap_active_ns_processes(void)
	{
		// kill all tasks in our ns and below
		kill(-1, SIGKILL);

		do {
			clear_thread_flag(TIF_SIGPENDING);
		} while (wait(NULL) != -ECHLD);
	}

Oleg.