netfilter: nf_conntrack: fix RCU race in nf_conntrack_find_get
authorAndrey Vagin <avagin@openvz.org>
Wed, 29 Jan 2014 18:34:14 +0000 (19:34 +0100)
committerPablo Neira Ayuso <pablo@netfilter.org>
Wed, 5 Feb 2014 12:16:18 +0000 (13:16 +0100)
Lets look at destroy_conntrack:

hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
...
nf_conntrack_free(ct)
kmem_cache_free(net->ct.nf_conntrack_cachep, ct);

net->ct.nf_conntrack_cachep is created with SLAB_DESTROY_BY_RCU.

The hash is protected by rcu, so readers look up conntracks without
locks.
A conntrack is removed from the hash, but in this moment a few readers
still can use the conntrack. Then this conntrack is released and another
thread creates conntrack with the same address and the equal tuple.
After this a reader starts to validate the conntrack:
* It's not dying, because a new conntrack was created
* nf_ct_tuple_equal() returns true.

But this conntrack is not initialized yet, so it can not be used by two
threads concurrently. In this case BUG_ON may be triggered from
nf_nat_setup_info().

Florian Westphal suggested to check the confirm bit too. I think it's
right.

task 1 task 2 task 3
nf_conntrack_find_get
 ____nf_conntrack_find
destroy_conntrack
 hlist_nulls_del_rcu
 nf_conntrack_free
 kmem_cache_free
__nf_conntrack_alloc
 kmem_cache_alloc
 memset(&ct->tuplehash[IP_CT_DIR_MAX],
 if (nf_ct_is_dying(ct))
 if (!nf_ct_tuple_equal()

I'm not sure, that I have ever seen this race condition in a real life.
Currently we are investigating a bug, which is reproduced on a few nodes.
In our case one conntrack is initialized from a few tasks concurrently,
we don't have any other explanation for this.

<2>[46267.083061] kernel BUG at net/ipv4/netfilter/nf_nat_core.c:322!
...
<4>[46267.083951] RIP: 0010:[<ffffffffa01e00a4>]  [<ffffffffa01e00a4>] nf_nat_setup_info+0x564/0x590 [nf_nat]
...
<4>[46267.085549] Call Trace:
<4>[46267.085622]  [<ffffffffa023421b>] alloc_null_binding+0x5b/0xa0 [iptable_nat]
<4>[46267.085697]  [<ffffffffa02342bc>] nf_nat_rule_find+0x5c/0x80 [iptable_nat]
<4>[46267.085770]  [<ffffffffa0234521>] nf_nat_fn+0x111/0x260 [iptable_nat]
<4>[46267.085843]  [<ffffffffa0234798>] nf_nat_out+0x48/0xd0 [iptable_nat]
<4>[46267.085919]  [<ffffffff814841b9>] nf_iterate+0x69/0xb0
<4>[46267.085991]  [<ffffffff81494e70>] ? ip_finish_output+0x0/0x2f0
<4>[46267.086063]  [<ffffffff81484374>] nf_hook_slow+0x74/0x110
<4>[46267.086133]  [<ffffffff81494e70>] ? ip_finish_output+0x0/0x2f0
<4>[46267.086207]  [<ffffffff814b5890>] ? dst_output+0x0/0x20
<4>[46267.086277]  [<ffffffff81495204>] ip_output+0xa4/0xc0
<4>[46267.086346]  [<ffffffff814b65a4>] raw_sendmsg+0x8b4/0x910
<4>[46267.086419]  [<ffffffff814c10fa>] inet_sendmsg+0x4a/0xb0
<4>[46267.086491]  [<ffffffff814459aa>] ? sock_update_classid+0x3a/0x50
<4>[46267.086562]  [<ffffffff81444d67>] sock_sendmsg+0x117/0x140
<4>[46267.086638]  [<ffffffff8151997b>] ? _spin_unlock_bh+0x1b/0x20
<4>[46267.086712]  [<ffffffff8109d370>] ? autoremove_wake_function+0x0/0x40
<4>[46267.086785]  [<ffffffff81495e80>] ? do_ip_setsockopt+0x90/0xd80
<4>[46267.086858]  [<ffffffff8100be0e>] ? call_function_interrupt+0xe/0x20
<4>[46267.086936]  [<ffffffff8118cb10>] ? ub_slab_ptr+0x20/0x90
<4>[46267.087006]  [<ffffffff8118cb10>] ? ub_slab_ptr+0x20/0x90
<4>[46267.087081]  [<ffffffff8118f2e8>] ? kmem_cache_alloc+0xd8/0x1e0
<4>[46267.087151]  [<ffffffff81445599>] sys_sendto+0x139/0x190
<4>[46267.087229]  [<ffffffff81448c0d>] ? sock_setsockopt+0x16d/0x6f0
<4>[46267.087303]  [<ffffffff810efa47>] ? audit_syscall_entry+0x1d7/0x200
<4>[46267.087378]  [<ffffffff810ef795>] ? __audit_syscall_exit+0x265/0x290
<4>[46267.087454]  [<ffffffff81474885>] ? compat_sys_setsockopt+0x75/0x210
<4>[46267.087531]  [<ffffffff81474b5f>] compat_sys_socketcall+0x13f/0x210
<4>[46267.087607]  [<ffffffff8104dea3>] ia32_sysret+0x0/0x5
<4>[46267.087676] Code: 91 20 e2 01 75 29 48 89 de 4c 89 f7 e8 56 fa ff ff 85 c0 0f 84 68 fc ff ff 0f b6 4d c6 41 8b 45 00 e9 4d fb ff ff e8 7c 19 e9 e0 <0f> 0b eb fe f6 05 17 91 20 e2 80 74 ce 80 3d 5f 2e 00 00 00 74
<1>[46267.088023] RIP  [<ffffffffa01e00a4>] nf_nat_setup_info+0x564/0x590

Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Florian Westphal <fw@strlen.de>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Cc: Patrick McHardy <kaber@trash.net>
Cc: Jozsef Kadlecsik <kadlec@blackhole.kfki.hu>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: Andrey Vagin <avagin@openvz.org>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
net/netfilter/nf_conntrack_core.c

index 8824ed0ccc9cd544e4799484ea7158f1f8db9c67..4d1fb5d094c3ba98f40870a71ea44b0fe8f37f03 100644 (file)
@@ -312,6 +312,21 @@ static void death_by_timeout(unsigned long ul_conntrack)
        nf_ct_delete((struct nf_conn *)ul_conntrack, 0, 0);
 }
 
+static inline bool
+nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
+                       const struct nf_conntrack_tuple *tuple,
+                       u16 zone)
+{
+       struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
+
+       /* A conntrack can be recreated with the equal tuple,
+        * so we need to check that the conntrack is confirmed
+        */
+       return nf_ct_tuple_equal(tuple, &h->tuple) &&
+               nf_ct_zone(ct) == zone &&
+               nf_ct_is_confirmed(ct);
+}
+
 /*
  * Warning :
  * - Caller must take a reference on returned object
@@ -333,8 +348,7 @@ ____nf_conntrack_find(struct net *net, u16 zone,
        local_bh_disable();
 begin:
        hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[bucket], hnnode) {
-               if (nf_ct_tuple_equal(tuple, &h->tuple) &&
-                   nf_ct_zone(nf_ct_tuplehash_to_ctrack(h)) == zone) {
+               if (nf_ct_key_equal(h, tuple, zone)) {
                        NF_CT_STAT_INC(net, found);
                        local_bh_enable();
                        return h;
@@ -372,8 +386,7 @@ begin:
                             !atomic_inc_not_zero(&ct->ct_general.use)))
                        h = NULL;
                else {
-                       if (unlikely(!nf_ct_tuple_equal(tuple, &h->tuple) ||
-                                    nf_ct_zone(ct) != zone)) {
+                       if (unlikely(!nf_ct_key_equal(h, tuple, zone))) {
                                nf_ct_put(ct);
                                goto begin;
                        }