On Sun, Apr 28, 2019 at 06:24:06PM +0200, Oleg Nesterov wrote:
> Thanks for cc'ing me...
>
> On 04/26, Christian Brauner wrote:
> >
> > On Thu, Apr 25, 2019 at 03:00:09PM -0400, Joel Fernandes (Google) wrote:
> > > +static unsigned int pidfd_poll(struct file *file, struct
> > > poll_table_struct *pts)
> > > +{
> > > + struct task_struct *task;
> > > + struct pid *pid;
> > > + int poll_flags = 0;
> > > +
> > > + /*
> > > + * tasklist_lock must be held because to avoid racing with
> > > + * changes in exit_state and wake up. Basically to avoid:
> > > + *
> > > + * P0: read exit_state = 0
> > > + * P1: write exit_state = EXIT_DEAD
> > > + * P1: Do a wake up - wq is empty, so do nothing
> > > + * P0: Queue for polling - wait forever.
> > > + */
> > > + read_lock(&tasklist_lock);
> > > + pid = file->private_data;
> > > + task = pid_task(pid, PIDTYPE_PID);
> > > + WARN_ON_ONCE(task && !thread_group_leader(task));
> > > +
> > > + if (!task || (task->exit_state && thread_group_empty(task)))
> > > + poll_flags = POLLIN | POLLRDNORM;
>
> Joel, I still can't understand why do we need tasklist... and I don't really
> understand the comment. The code looks as if you are trying to avoid
> poll_wait(),
> but this would be strange.
>
> OK, why can't pidfd_poll() do
>
> poll_wait(file, &pid->wait_pidfd, pts);
>
> rcu_read_lock();
> task = pid_task(pid, PIDTYPE_PID);
> if (!task || task->exit_state && thread_group_empty(task))
> poll_flags = POLLIN | ...;
> rcu_read_unlock();
>
> return poll_flags;
>
> ?
Oh that's much better Oleg, and would avoid the race I had in mind: Basically
I was acquiring the tasklist_lock to avoid a case where a polling task is not
woken up because it was added to the waitqueue too late. The reading of the
exit_state and the conditional adding to the wait queue, needed to be atomic.
Otherwise something like the following may be possible:
Task A (poller) Task B (exiting task being polled)
------------ ----------------
poll() called
exit_state is set to non-zero
read exit_state
wake_up_all()
add_wait_queue()
----------------------------------------------
However, in your code above, it is avoided because we get:
Task A (poller) Task B (exiting task being polled)
------------ ----------------
poll() called
add_wait_queue()
exit_state is set to non-zero
read exit_state
remove_wait_queue()
wake_up_all()
I don't see any other issues with your code above so I can try it out and
update the patches. Thanks.
> > > +static void do_notify_pidfd(struct task_struct *task)
> >
> > Maybe a short command that this helper can only be called when we know
> > that task is a thread-group leader wouldn't hurt so there's no confusion
> > later.
>
> Not really. If the task is traced, do_notify_parent() (and thus
> do_notify_pidfd())
> can be called to notify the debugger even if the task is not a leader and/or
> if
> it is not the last thread. The latter means a spurious wakeup for
> pidfd_poll().
Seems like you are replying to Christian's point. I agree with you.
> > > +{
> > > + struct pid *pid;
> > > +
> > > + lockdep_assert_held(&tasklist_lock);
> > > +
> > > + pid = get_task_pid(task, PIDTYPE_PID);
> > > + wake_up_all(&pid->wait_pidfd);
> > > + put_pid(pid);
>
> Why get/put?
Yes, pid_task() should do it. Will update it. Thanks!
- Joel
