Re: [PATCH 4/4] exit: Lockless iteration over task list in mm_update_next_owner()

[Kirill Tkhai]

On 26.04.2018 18:20, Peter Zijlstra wrote:
> On Thu, Apr 26, 2018 at 04:52:39PM +0300, Kirill Tkhai wrote:
>>>>
>>>> 1)for_each_process(g)            copy_process()
>>>>                                    p->mm = mm
>>>>     smp_rmb();                     smp_wmb() implied by alloc_pid()
>>>>     if (g->flags & PF_KTHREAD)     list_add_tail_rcu(&p->tasks, &init_task.tasks)
>>>>
>>>> 2)for_each_thread(g, c)          copy_process()
>>>>                                    p->mm = mm
>>>>     smp_rmb();                     smp_wmb() implied by alloc_pid()
>>>>     tmp = READ_ONCE(c->mm)         list_add_tail_rcu(&p->thread_node, ...)
> 
> For these two; what's the purpose of the smp_rmb()? which loads are
> ordered?

(1) and (2) make visible g->flags and c->mm in case we see process is linked
to the task list.

It seems in mm_update_next_owner() we may see result of list_add_tail_rcu(&p->thread_node, ...),
but p->mm = mm will not be visible.

Hm, also I've dived into list_add_tail_rcu() and it contains smp_wmb(). So, it seems
we do not need the barrier implied by alloc_pid().

Kirill