Re: [PATCH v4 1/2] pid: Replace pid bitmap implementation with IDR API

[Oleg Nesterov]

On 10/10, Gargi Sharma wrote:
>
> On Tue, Oct 10, 2017 at 12:50 PM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> > On 10/09, Andrew Morton wrote:
> >>
> >> Especially here.  I don't think pidmap_lock is held.  Is that IDR
> >> iteration safe?
> >
> > Yes, this doesn't look right, we need rcu_read_lock() or pidmap_lock.
> >
> > And, we also need rcu_read_lock() for another reason, to protect "struct pid".
>
> Ah, I missed this. From what I understood idr_for_each_entry_continue
> should be safe

Without rcu? Why?

> because calls idr_get_next which in turn calls
> radix_tree_iter_find to find the next populated entry in the idr.

and then it does rcu_dereference_raw(*slot). Without rcu or pidmap this
slot can got away if we race with free_pid().

> If
> the pid that you are looking up the task for is deleted, task will get
> a NULL from pid_task and no signal to kill will be sent.

pid->tasks is protected by tasklist_lock, but the pid itself can go away
without rcu lock.

So I think you need to take rcu_read_lock() right after tasklist_lock.

> > Gargi, I suggested to use idr_for_each_entry_continue(), but now I am wondering
> > if we should use idr_for_each() instead. IIUC this would be a bit faster? Not
> > that I think this is really important...
>
> I can run benchmarks with idr_for_each to see how much speed up is
> achieved and then we can go with whatever we think is better. How does
> that sounds?

Up to you ;) I am fine either way.

Oleg.