netfilter: ctnetlink: fix soft lockup when netlink adds new entries (v2)
authorJozsef Kadlecsik <kadlec@blackhole.kfki.hu>
Fri, 24 Feb 2012 10:45:49 +0000 (11:45 +0100)
committerPablo Neira Ayuso <pablo@netfilter.org>
Fri, 24 Feb 2012 11:24:15 +0000 (12:24 +0100)
Marcell Zambo and Janos Farago noticed and reported that when
new conntrack entries are added via netlink and the conntrack table
gets full, soft lockup happens. This is because the nf_conntrack_lock
is held while nf_conntrack_alloc is called, which is in turn wants
to lock nf_conntrack_lock while evicting entries from the full table.

The patch fixes the soft lockup with limiting the holding of the
nf_conntrack_lock to the minimum, where it's absolutely required.
It required to extend (and thus change) nf_conntrack_hash_insert
so that it makes sure conntrack and ctnetlink do not add the same entry
twice to the conntrack table.

Signed-off-by: Jozsef Kadlecsik <kadlec@blackhole.kfki.hu>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
include/net/netfilter/nf_conntrack.h
net/netfilter/nf_conntrack_core.c
net/netfilter/nf_conntrack_netlink.c

index 8a2b0ae..ab86036 100644 (file)
@@ -209,7 +209,7 @@ extern struct nf_conntrack_tuple_hash *
 __nf_conntrack_find(struct net *net, u16 zone,
                    const struct nf_conntrack_tuple *tuple);
 
-extern void nf_conntrack_hash_insert(struct nf_conn *ct);
+extern int nf_conntrack_hash_check_insert(struct nf_conn *ct);
 extern void nf_ct_delete_from_lists(struct nf_conn *ct);
 extern void nf_ct_insert_dying_list(struct nf_conn *ct);
 
index 76613f5..ed86a3b 100644 (file)
@@ -404,19 +404,49 @@ static void __nf_conntrack_hash_insert(struct nf_conn *ct,
                           &net->ct.hash[repl_hash]);
 }
 
-void nf_conntrack_hash_insert(struct nf_conn *ct)
+int
+nf_conntrack_hash_check_insert(struct nf_conn *ct)
 {
        struct net *net = nf_ct_net(ct);
        unsigned int hash, repl_hash;
+       struct nf_conntrack_tuple_hash *h;
+       struct hlist_nulls_node *n;
        u16 zone;
 
        zone = nf_ct_zone(ct);
-       hash = hash_conntrack(net, zone, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
-       repl_hash = hash_conntrack(net, zone, &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
+       hash = hash_conntrack(net, zone,
+                             &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
+       repl_hash = hash_conntrack(net, zone,
+                                  &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
+
+       spin_lock_bh(&nf_conntrack_lock);
 
+       /* See if there's one in the list already, including reverse */
+       hlist_nulls_for_each_entry(h, n, &net->ct.hash[hash], hnnode)
+               if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
+                                     &h->tuple) &&
+                   zone == nf_ct_zone(nf_ct_tuplehash_to_ctrack(h)))
+                       goto out;
+       hlist_nulls_for_each_entry(h, n, &net->ct.hash[repl_hash], hnnode)
+               if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
+                                     &h->tuple) &&
+                   zone == nf_ct_zone(nf_ct_tuplehash_to_ctrack(h)))
+                       goto out;
+
+       add_timer(&ct->timeout);
+       nf_conntrack_get(&ct->ct_general);
        __nf_conntrack_hash_insert(ct, hash, repl_hash);
+       NF_CT_STAT_INC(net, insert);
+       spin_unlock_bh(&nf_conntrack_lock);
+
+       return 0;
+
+out:
+       NF_CT_STAT_INC(net, insert_failed);
+       spin_unlock_bh(&nf_conntrack_lock);
+       return -EEXIST;
 }
-EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
+EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 
 /* Confirm a connection given skb; places it in hash table */
 int
index 9307b03..30c9d4c 100644 (file)
@@ -1367,15 +1367,12 @@ ctnetlink_create_conntrack(struct net *net, u16 zone,
                                                    nf_ct_protonum(ct));
                if (helper == NULL) {
                        rcu_read_unlock();
-                       spin_unlock_bh(&nf_conntrack_lock);
 #ifdef CONFIG_MODULES
                        if (request_module("nfct-helper-%s", helpname) < 0) {
-                               spin_lock_bh(&nf_conntrack_lock);
                                err = -EOPNOTSUPP;
                                goto err1;
                        }
 
-                       spin_lock_bh(&nf_conntrack_lock);
                        rcu_read_lock();
                        helper = __nf_conntrack_helper_find(helpname,
                                                            nf_ct_l3num(ct),
@@ -1468,8 +1465,10 @@ ctnetlink_create_conntrack(struct net *net, u16 zone,
        if (tstamp)
                tstamp->start = ktime_to_ns(ktime_get_real());
 
-       add_timer(&ct->timeout);
-       nf_conntrack_hash_insert(ct);
+       err = nf_conntrack_hash_check_insert(ct);
+       if (err < 0)
+               goto err2;
+
        rcu_read_unlock();
 
        return ct;
@@ -1490,6 +1489,7 @@ ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
        struct nf_conntrack_tuple otuple, rtuple;
        struct nf_conntrack_tuple_hash *h = NULL;
        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
+       struct nf_conn *ct;
        u_int8_t u3 = nfmsg->nfgen_family;
        u16 zone;
        int err;
@@ -1510,27 +1510,22 @@ ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
                        return err;
        }
 
-       spin_lock_bh(&nf_conntrack_lock);
        if (cda[CTA_TUPLE_ORIG])
-               h = __nf_conntrack_find(net, zone, &otuple);
+               h = nf_conntrack_find_get(net, zone, &otuple);
        else if (cda[CTA_TUPLE_REPLY])
-               h = __nf_conntrack_find(net, zone, &rtuple);
+               h = nf_conntrack_find_get(net, zone, &rtuple);
 
        if (h == NULL) {
                err = -ENOENT;
                if (nlh->nlmsg_flags & NLM_F_CREATE) {
-                       struct nf_conn *ct;
                        enum ip_conntrack_events events;
 
                        ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
                                                        &rtuple, u3);
-                       if (IS_ERR(ct)) {
-                               err = PTR_ERR(ct);
-                               goto out_unlock;
-                       }
+                       if (IS_ERR(ct))
+                               return PTR_ERR(ct);
+
                        err = 0;
-                       nf_conntrack_get(&ct->ct_general);
-                       spin_unlock_bh(&nf_conntrack_lock);
                        if (test_bit(IPS_EXPECTED_BIT, &ct->status))
                                events = IPCT_RELATED;
                        else
@@ -1545,23 +1540,19 @@ ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
                                                      ct, NETLINK_CB(skb).pid,
                                                      nlmsg_report(nlh));
                        nf_ct_put(ct);
-               } else
-                       spin_unlock_bh(&nf_conntrack_lock);
+               }
 
                return err;
        }
        /* implicit 'else' */
 
-       /* We manipulate the conntrack inside the global conntrack table lock,
-        * so there's no need to increase the refcount */
        err = -EEXIST;
+       ct = nf_ct_tuplehash_to_ctrack(h);
        if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
-               struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
-
+               spin_lock_bh(&nf_conntrack_lock);
                err = ctnetlink_change_conntrack(ct, cda);
+               spin_unlock_bh(&nf_conntrack_lock);
                if (err == 0) {
-                       nf_conntrack_get(&ct->ct_general);
-                       spin_unlock_bh(&nf_conntrack_lock);
                        nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
                                                      (1 << IPCT_ASSURED) |
                                                      (1 << IPCT_HELPER) |
@@ -1570,15 +1561,10 @@ ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
                                                      (1 << IPCT_MARK),
                                                      ct, NETLINK_CB(skb).pid,
                                                      nlmsg_report(nlh));
-                       nf_ct_put(ct);
-               } else
-                       spin_unlock_bh(&nf_conntrack_lock);
-
-               return err;
+               }
        }
 
-out_unlock:
-       spin_unlock_bh(&nf_conntrack_lock);
+       nf_ct_put(ct);
        return err;
 }