Re: while_each_thread() under rcu_read_lock() is broken?

[Oleg Nesterov <oleg@redhat.com>]

On 06/18, Roland McGrath wrote:
>
> I think you're right.  I can't see what would prevent that race.

How sad.

> So for_each_process

for_each_process() looks fine. It uses init_task as the anchor,
it can't go away, it is swapper.

> and do_each_thread are safe only under
> read_lock(&tasklist_lock) and while_each_thread is only safe under
> either that or siglock.

Yes,

(Also a few places using next_thread in
> similar loops outside those macros.)

I hope that most (all?) of next_thread() users can be converted to
use while_each_thread().

> Perhaps we could move those del's from __unhash_process to
> __put_task_struct (or just delayed_put_task_struct?)

This needs write_lock_irq(tasklist), we can't take it in atomic
context. And I bet this change (at least right now) has other
implications.

> I think de_thread() in exec-by-nonleader is the only case where this
> can happen, right?  So then perhaps we could make it call release_task
> only via call_rcu?

Hmm, perhaps...  I am already sleeping, will try to check this idea
tomorrow. At first glance, it looks promising to me. And I see the
email from Paul which is too late to read for me today ;)

In any case, I _think_ we can fix while_each_thread(), say XXX(t)
from the previous email. But then we should audit the users like
zap_threads() which assume we should not miss any "interesting" task.
Probably zap_threads() is fine because of mmap_sem, but I can't
think properly now.

Oleg.