Re: epoll oops.

[Oleg Nesterov <oleg@redhat.com>]

On 10/16, Eric Wong wrote:
>
> Oleg Nesterov <oleg@redhat.com> wrote:
> > Yes. Before that 971316f0503a hack epoll can't even know if the task
> > which did signalfd_poll() exits and frees the active signalfd_wqh.
> > If for example that task forked a child before exit.
> >
> > And the whole RCU logic is only needed if exit/ep_remove_wait_queue
> > actually race with each other.
>
> Is there any chance this oops is caused by (or at least more easily
> exposed by) commit 91cf5ab60ff82ecf4550a596867787c1e360dd3f ?
> (epoll: add a reschedule point in ep_free())
>
> I thought 91cf5ab would be benign, except...
>
> > Yes, ugly, agreed. d80e731ecab4 even tries to docunent that this all
> > is the hack.
>
> .. the following sentence from d80e731ecab4 caught my eye:
>
>     It also assumes that nobody can take tasklist_lock under epoll
>     locks, this seems to be true.

This just reminds that with this patch __wake_up/ep_poll_callback can
be called under write_lock(tasklist).

> I haven't been able to trace if cond_resched() can take tasklist_lock.

No, it can't hold the non-sleepable rwlock_t. And the sentence above
doesn't mean the locks like epmutex, it is mostlt about ep->lock.

Oleg.