Re: 2.6.24-rc6: possible recursive locking detected

[Davide Libenzi <davidel@xmailserver.org>]

On Sat, 5 Jan 2008, Peter Zijlstra wrote:

> 
> On Sat, 2008-01-05 at 17:53 +0100, Peter Zijlstra wrote:
> > On Sat, 2008-01-05 at 18:12 +1100, Herbert Xu wrote:
> > > On Fri, Jan 04, 2008 at 09:30:49AM +0100, Ingo Molnar wrote:
> > > >
> > > > > > [ 1310.670986] =============================================
> > > > > > [ 1310.671690] [ INFO: possible recursive locking detected ]
> > > > > > [ 1310.672097] 2.6.24-rc6 #1
> > > > > > [ 1310.672421] ---------------------------------------------
> > > > > > [ 1310.672828] FahCore_a0.exe/3692 is trying to acquire lock:
> > > > > > [ 1310.673238]  (&q->lock){++..}, at: [<c011544b>] __wake_up+0x1b/0x50
> > > > > > [ 1310.673869]
> > > > > > [ 1310.673870] but task is already holding lock:
> > > > > > [ 1310.674567]  (&q->lock){++..}, at: [<c011544b>] __wake_up+0x1b/0x50
> > > > > > [ 1310.675267]
> > > > > > [ 1310.675268] other info that might help us debug this:
> > > > > > [ 1310.675952] 5 locks held by FahCore_a0.exe/3692:
> > > > > > [ 1310.676334]  #0:  (rcu_read_lock){..--}, at: [<c038b620>] net_rx_action+0x60/0x1b0
> > > > > > [ 1310.677251]  #1:  (rcu_read_lock){..--}, at: [<c0388d60>] netif_receive_skb+0x100/0x470
> > > > > > [ 1310.677924]  #2:  (rcu_read_lock){..--}, at: [<c03a7fb2>] ip_local_deliver_finish+0x32/0x210
> > > > > > [ 1310.678460]  #3:  (clock-AF_INET){-.-?}, at: [<c038164e>] sock_def_readable+0x1e/0x80
> > > > > > [ 1310.679250]  #4:  (&q->lock){++..}, at: [<c011544b>] __wake_up+0x1b/0x50
> > > 
> > > The net part might just be a red herring, since the problem is that
> > > __wake_up is somehow reentering itself.
> > 
> > /*
> >  * Perform a safe wake up of the poll wait list. The problem is that
> >  * with the new callback'd wake up system, it is possible that the
> >  * poll callback is reentered from inside the call to wake_up() done
> >  * on the poll wait queue head. The rule is that we cannot reenter the
> >  * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
> >  * and we cannot reenter the same wait queue head at all. This will
> >  * enable to have a hierarchy of epoll file descriptor of no more than
> >  * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
> >  * because this one gets called by the poll callback, that in turn is called
> >  * from inside a wake_up(), that might be called from irq context.
> >  */
> > 
> > Seems to suggest that the epoll code can indeed recurse into wakeup.
> > 
> > Davide, Johannes, any ideas?
> 
> Since EP_MAX_POLLWAKE_NESTS < MAX_LOCKDEP_SUBCLASSES we could perhaps do
> something like:
> 
>   wake_up_nested(..., wake_nests);
> 
> although I'm not quite sure that is correct, my understanding of this
> code is still fragile at best.

I remember I talked with Arjan about this time ago. Basically, since 1) 
you can drop an epoll fd inside another epoll fd 2) callback-based wakeups 
are used, you can see a wake_up() from inside another wake_up(), but they 
will never refer to the same lock instance.
Think about:

	dfd = socket(...);
	efd1 = epoll_create();
	efd2 = epoll_create();
	epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
	epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);

When a packet arrives to the device underneath "dfd", the net code will 
issue a wake_up() on its poll wake list. Epoll (efd1) has installed a 
callback wakeup entry on that queue, and the wake_up() performed by the 
"dfd" net code will end up in ep_poll_callback(). At this point epoll 
(efd1) notices that it may have some event ready, so it needs to wake up 
the waiters on its poll wait list (efd2). So it calls ep_poll_safewake() 
that ends up in another wake_up(), after having checked about the 
recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to 
avoid stack blasting. Never hit the same queue, to avoid loops like:

	epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
	epoll_ctl(efd3, EPOLL_CTL_ADD, efd2, ...);
	epoll_ctl(efd4, EPOLL_CTL_ADD, efd3, ...);
	epoll_ctl(efd1, EPOLL_CTL_ADD, efd4, ...);

The code "if (tncur->wq == wq || ..." prevents re-entering the same 
queue/lock.
I don't know how the lockdep code works, so I can't say about 
wake_up_nested(). Although I have a feeling is not enough in this case.
A solution may be to move the call to ep_poll_safewake() (that'd become a 
simple wake_up()) inside a tasklet or whatever is today trendy for delayed 
work. But his kinda scares me to be honest, since epoll has already a 
bunch of places where it could be asynchronously hit (plus performance 
regression will need to be verified).



- Davide