Re: [PATCH 3/4] mm, oom: do not rely on TIF_MEMDIE for exit_oom_victim

From: Oleg Nesterov
Date: Tue Oct 04 2016 - 12:22:31 EST


On 10/04, Michal Hocko wrote:
>
> -void release_task(struct task_struct *p)
> +bool release_task(struct task_struct *p)
> {
> struct task_struct *leader;
> int zap_leader;
> + bool last = false;
> repeat:
> /* don't need to get the RCU readlock here - the process is dead and
> * can't be modifying its own credentials. But shut RCU-lockdep up */
> @@ -197,8 +198,10 @@ void release_task(struct task_struct *p)
> * then we are the one who should release the leader.
> */
> zap_leader = do_notify_parent(leader, leader->exit_signal);
> - if (zap_leader)
> + if (zap_leader) {
> leader->exit_state = EXIT_DEAD;
> + last = true;
> + }
> }

This looks strange... it won't return true if "p" is the group leader.

> @@ -584,12 +587,15 @@ static void forget_original_parent(struct task_struct *father,
> /*
> * Send signals to all our closest relatives so that they know
> * to properly mourn us..
> + *
> + * Returns true if this is the last thread from the thread group
> */
> -static void exit_notify(struct task_struct *tsk, int group_dead)
> +static bool exit_notify(struct task_struct *tsk, int group_dead)
> {
> bool autoreap;
> struct task_struct *p, *n;
> LIST_HEAD(dead);
> + bool last = false;
>
> write_lock_irq(&tasklist_lock);
> forget_original_parent(tsk, &dead);
> @@ -606,6 +612,7 @@ static void exit_notify(struct task_struct *tsk, int group_dead)
> } else if (thread_group_leader(tsk)) {
> autoreap = thread_group_empty(tsk) &&
> do_notify_parent(tsk, tsk->exit_signal);
> + last = thread_group_empty(tsk);

so this can't detect the multi-threaded group exit, and ...

> list_for_each_entry_safe(p, n, &dead, ptrace_entry) {
> list_del_init(&p->ptrace_entry);
> - release_task(p);
> + if (release_task(p) && p == tsk)
> + last = true;

this can only happen if this process auto-reaps itself. Not to mention
that exit_notify() will never return true if traced.

No, this doesn't look right.

Oleg.