Re: net: possible deadlock in skb_queue_tail
From: Florian Westphal
Date: Thu Feb 23 2017 - 21:58:41 EST
Andrey Konovalov <andreyknvl@xxxxxxxxxx> wrote:
[ CC Paolo ]
> I've got the following error report while fuzzing the kernel with syzkaller.
>
> On commit c470abd4fde40ea6a0846a2beab642a578c0b8cd (4.10).
>
> Unfortunately I can't reproduce it.
This needs NETLINK_BROADCAST_ERROR enabled on a netlink socket
that then subscribes to netfilter conntrack (ctnetlink) events.
probably syzkaller did this by accident -- impressive.
(one task is the ctnetlink event redelivery worker
which won't be scheduled otherwise).
> ======================================================
> [ INFO: possible circular locking dependency detected ]
> 4.10.0-rc8+ #201 Not tainted
> -------------------------------------------------------
> kworker/0:2/1404 is trying to acquire lock:
> (&(&list->lock)->rlock#3){+.-...}, at: [<ffffffff8335b23f>]
> skb_queue_tail+0xcf/0x2f0 net/core/skbuff.c:2478
>
> but task is already holding lock:
> (&(&pcpu->lock)->rlock){+.-...}, at: [<ffffffff8366b55f>] spin_lock
> include/linux/spinlock.h:302 [inline]
> (&(&pcpu->lock)->rlock){+.-...}, at: [<ffffffff8366b55f>]
> ecache_work_evict_list+0xaf/0x590
> net/netfilter/nf_conntrack_ecache.c:48
>
> which lock already depends on the new lock.
Cong is correct, this is a false positive.
However we should fix this splat.
Paolo, this happens since 7c13f97ffde63cc792c49ec1513f3974f2f05229
('udp: do fwd memory scheduling on dequeue'), before this
commit kfree_skb() was invoked outside of the locked section in
first_packet_length().
cpu 0 call chain:
- first_packet_length (hold udp sk_receive_queue lock)
- kfree_skb
- nf_conntrack_destroy
- spin_lock(net->ct.pcpu->lock)
cpu 1 call chain:
- ecache_work_evict_list
- spin_lock( net->ct.pcpu->lock)
- nf_conntrack_event
- aquire netlink socket sk_receive_queue
So this could only ever deadlock if a netlink socket
calls kfree_skb while holding its sk_receive_queue lock, but afaics
this is never the case.
There are two ways to avoid this splat (other than lockdep annotation):
1. re-add the list to first_packet_length() and free the
skbs outside of locked section.
2. change ecache_work_evict_list to not call nf_conntrack_event()
while holding the pcpu lock.
doing #2 might be a good idea anyway to avoid potential deadlock
when kfree_skb gets invoked while other cpu holds its sk_receive_queue
lock, I'll have a look if this is feasible.