Skip to contentSkip to conversation

netfilter: ctnetlink: cleanup master conntrack assignation

PATCH Applied Pablo Neira Ayuso 2 participants 11 messages Mar 5, 2009 – Mar 18, 2009

Subsystems (2)
lore
Share & tools
lore

Pablo Neira Ayuso

author maintainer
#0 lore raw

PATCH 1/3 netfilter: ctnetlink: cleanup master conntrack assignation

This patch moves the assignation of the master conntrack to ctnetlink_create_conntrack(), which is where it really belongs. This patch is a cleanup.

---

 net/netfilter/nf_conntrack_netlink.c |   49 ++++++++++++++--------------------
 1 files changed, 20 insertions(+), 29 deletions(-)
net/netfilter/nf_conntrack_netlink.c +20 -29
index 41226f0..a240a66 100644
@@ -1234,9 +1234,9 @@ static int
 ctnetlink_create_conntrack(struct nlattr *cda[],
 			   struct nf_conntrack_tuple *otuple,
 			   struct nf_conntrack_tuple *rtuple,
-			   struct nf_conn *master_ct,
 			   u32 pid,
-			   int report)
+			   int report,
+			   u8 u3)
 {
 	struct nf_conn *ct;
 	int err = -EINVAL;
@@ -1347,7 +1347,22 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 #endif
 
 	/* setup master conntrack: this is a confirmed expectation */
-	if (master_ct) {
+	if (cda[CTA_TUPLE_MASTER]) {
+		struct nf_conntrack_tuple master;
+		struct nf_conntrack_tuple_hash *master_h;
+		struct nf_conn *master_ct;
+
+		err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
+		if (err < 0)
+			goto err;
+
+		master_h = __nf_conntrack_find(&init_net, &master);
+		if (master_h == NULL) {
+			err = -ENOENT;
+			goto err;
+		}
+		master_ct = nf_ct_tuplehash_to_ctrack(master_h);
+		nf_conntrack_get(&master_ct->ct_general);
 		__set_bit(IPS_EXPECTED_BIT, &ct->status);
 		ct->master = master_ct;
 	}
@@ -1395,39 +1410,15 @@ ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
 		h = __nf_conntrack_find(&init_net, &rtuple);
 
 	if (h == NULL) {
-		struct nf_conntrack_tuple master;
-		struct nf_conntrack_tuple_hash *master_h = NULL;
-		struct nf_conn *master_ct = NULL;
-
-		if (cda[CTA_TUPLE_MASTER]) {
-			err = ctnetlink_parse_tuple(cda,
-						    &master,
-						    CTA_TUPLE_MASTER,
-						    u3);
-			if (err < 0)
-				goto out_unlock;
-
-			master_h = __nf_conntrack_find(&init_net, &master);
-			if (master_h == NULL) {
-				err = -ENOENT;
-				goto out_unlock;
-			}
-			master_ct = nf_ct_tuplehash_to_ctrack(master_h);
-			nf_conntrack_get(&master_ct->ct_general);
-		}
-
 		err = -ENOENT;
 		if (nlh->nlmsg_flags & NLM_F_CREATE)
 			err = ctnetlink_create_conntrack(cda,
 							 &otuple,
 							 &rtuple,
-							 master_ct,
 							 NETLINK_CB(skb).pid,
-							 nlmsg_report(nlh));
+							 nlmsg_report(nlh),
+							 u3);
 		spin_unlock_bh(&nf_conntrack_lock);
-		if (err < 0 && master_ct)
-			nf_ct_put(master_ct);
-
 		return err;
 	}
 	/* implicit 'else' */
author maintainer
#1 lore raw

PATCH 2/3 netfilter: ctnetlink: cleanup conntrack update preliminary checkings

This patch moves the preliminary checkings that must be fulfilled to update a conntrack, which are the following:

* NAT manglings cannot be updated
* Changing the master conntrack is not allowed.

This patch is a cleanup.

---

 net/netfilter/nf_conntrack_netlink.c |   15 ++++-----------
 1 files changed, 4 insertions(+), 11 deletions(-)
net/netfilter/nf_conntrack_netlink.c +4 -11
index a240a66..d229370 100644
@@ -1168,6 +1168,10 @@ ctnetlink_change_conntrack(struct nf_conn *ct, struct nlattr *cda[])
 {
 	int err;
 
+	/* only allow NAT changes and master assignation for new conntracks */
+	if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
+		return -EOPNOTSUPP;
+
 	if (cda[CTA_HELP]) {
 		err = ctnetlink_change_helper(ct, cda);
 		if (err < 0)
@@ -1429,17 +1433,6 @@ ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
 	if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
 		struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 
-		/* we only allow nat config for new conntracks */
-		if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
-			err = -EOPNOTSUPP;
-			goto out_unlock;
-		}
-		/* can't link an existing conntrack to a master */
-		if (cda[CTA_TUPLE_MASTER]) {
-			err = -EOPNOTSUPP;
-			goto out_unlock;
-		}
-
 		err = ctnetlink_change_conntrack(ct, cda);
 		if (err == 0) {
 			nf_conntrack_get(&ct->ct_general);
#2 lore raw

Re: [PATCH 2/3] netfilter: ctnetlink: cleanup conntrack update preliminary checkings

Quoted 7 lines · Pablo Neira Ayuso

This patch moves the preliminary checkings that must be fulfilled to update a conntrack, which are the following:

* NAT manglings cannot be updated
* Changing the master conntrack is not allowed.

This patch is a cleanup.

Applied, thanks.

author maintainer
#3 lore raw

PATCH 3/3 netfilter: ctnetlink: move event reporting for new entries outside the lock

In reply to Pablo Neira Ayuso

This patch moves the event reporting outside the lock section. With this patch, the creation and update of entries is homogeneous from the event reporting perspective. Moreover, as the event reporting is done outside the lock section, the netlink broadcast delivery can benefit of the yield() call under congestion.

---

 net/netfilter/nf_conntrack_netlink.c |   41 +++++++++++++++++++---------------
 1 files changed, 23 insertions(+), 18 deletions(-)
net/netfilter/nf_conntrack_netlink.c +23 -18
index d229370..fa5b23f 100644
@@ -1234,12 +1234,10 @@ ctnetlink_event_report(struct nf_conn *ct, u32 pid, int report)
 				  report);
 }
 
-static int
+static struct nf_conn *
 ctnetlink_create_conntrack(struct nlattr *cda[],
 			   struct nf_conntrack_tuple *otuple,
 			   struct nf_conntrack_tuple *rtuple,
-			   u32 pid,
-			   int report,
 			   u8 u3)
 {
 	struct nf_conn *ct;
@@ -1248,7 +1246,7 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 
 	ct = nf_conntrack_alloc(&init_net, otuple, rtuple, GFP_ATOMIC);
 	if (IS_ERR(ct))
-		return -ENOMEM;
+		return ERR_PTR(-ENOMEM);
 
 	if (!cda[CTA_TIMEOUT])
 		goto err;
@@ -1371,18 +1369,14 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 		ct->master = master_ct;
 	}
 
-	nf_conntrack_get(&ct->ct_general);
 	add_timer(&ct->timeout);
 	nf_conntrack_hash_insert(ct);
 	rcu_read_unlock();
-	ctnetlink_event_report(ct, pid, report);
-	nf_ct_put(ct);
-
-	return 0;
 
+	return ct;
 err:
 	nf_conntrack_free(ct);
-	return err;
+	return ERR_PTR(err);
 }
 
 static int
@@ -1415,14 +1409,25 @@ ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
 
 	if (h == NULL) {
 		err = -ENOENT;
-		if (nlh->nlmsg_flags & NLM_F_CREATE)
-			err = ctnetlink_create_conntrack(cda,
-							 &otuple,
-							 &rtuple,
-							 NETLINK_CB(skb).pid,
-							 nlmsg_report(nlh),
-							 u3);
-		spin_unlock_bh(&nf_conntrack_lock);
+		if (nlh->nlmsg_flags & NLM_F_CREATE) {
+			struct nf_conn *ct;
+
+			ct = ctnetlink_create_conntrack(cda, &otuple,
+							&rtuple, u3);
+			if (IS_ERR(ct)) {
+				err = PTR_ERR(ct);
+				goto out_unlock;
+			}
+			err = 0;
+			nf_conntrack_get(&ct->ct_general);
+			spin_unlock_bh(&nf_conntrack_lock);
+			ctnetlink_event_report(ct,
+					       NETLINK_CB(skb).pid,
+					       nlmsg_report(nlh));
+			nf_ct_put(ct);
+		} else
+			spin_unlock_bh(&nf_conntrack_lock);
+
 		return err;
 	}
 	/* implicit 'else' */
#4 lore raw

Re: [PATCH 3/3] netfilter: ctnetlink: move event reporting for new entries outside the lock

Quoted 5 lines · Pablo Neira Ayuso

This patch moves the event reporting outside the lock section. With this patch, the creation and update of entries is homogeneous from the event reporting perspective. Moreover, as the event reporting is done outside the lock section, the netlink broadcast delivery can benefit of the yield() call under congestion.

This one doesn't apply cleanly, probably because of the missing netlink_set_err patch. I've pushed out my current nf-next tree, please rediff against that tree.

#5 lore raw

Re: [PATCH 3/3] netfilter: ctnetlink: move event reporting for new entries outside the lock

Quoted 10 lines · Patrick McHardy
Quoted 5 lines · Pablo Neira Ayuso

This patch moves the event reporting outside the lock section. With this patch, the creation and update of entries is homogeneous from the event reporting perspective. Moreover, as the event reporting is done outside the lock section, the netlink broadcast delivery can benefit of the yield() call under congestion.

This one doesn't apply cleanly, probably because of the missing netlink_set_err patch. I've pushed out my current nf-next tree, please rediff against that tree.

It might also have been ordering, let me try again :)

#6 lore raw

Re: [PATCH 3/3] netfilter: ctnetlink: move event reporting for new entries outside the lock

In reply to Pablo Neira Ayuso
Quoted 5 lines · Pablo Neira Ayuso

This patch moves the event reporting outside the lock section. With this patch, the creation and update of entries is homogeneous from the event reporting perspective. Moreover, as the event reporting is done outside the lock section, the netlink broadcast delivery can benefit of the yield() call under congestion.

Applied, thanks.

#7 lore raw
In reply to Pablo Neira Ayuso
Quoted 3 lines · Pablo Neira Ayuso

This patch moves the assignation of the master conntrack to ctnetlink_create_conntrack(), which is where it really belongs. This patch is a cleanup.

Applied, thanks.

#8 lore raw
Quoted 6 lines · Patrick McHardy
Quoted 3 lines · Pablo Neira Ayuso

This patch moves the assignation of the master conntrack to ctnetlink_create_conntrack(), which is where it really belongs. This patch is a cleanup.

Applied, thanks.

I've added this patch on top to fix a RCU context imbalance.

This seems like a good opportunity to say this again: please (everyone) compile your code using sparse. It catches this type and more bugs.

There is a second bug introduced by this patch:

	add_timer(&ct->timeout);
	nf_conntrack_hash_insert(ct);
	rcu_read_unlock();

	return ct;

The conntrack lock is not held, it might crash or create double entries.

commit 0f5b3e85a3716efebb0150ebb7c6d022e2bf17d7

Author: Patrick McHardy <[email protected]>
Date:   Wed Mar 18 17:36:40 2009 +0100

netfilter: ctnetlink: fix rcu context imbalance

Introduced by 7ec47496 (netfilter: ctnetlink: cleanup master conntrack assignation):

net/netfilter/nf_conntrack_netlink.c:1275:2: warning: context imbalance in 'ctnetlink_create_conntrack' - different lock contexts for basic block

net/netfilter/nf_conntrack_netlink.c +23 -34
index 735ea9c..d1fe9d1 100644
@@ -1146,7 +1146,7 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 		return ERR_PTR(-ENOMEM);
 
 	if (!cda[CTA_TIMEOUT])
-		goto err;
+		goto err1;
 	ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
 
 	ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
@@ -1157,10 +1157,8 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
  		char *helpname;
  
  		err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
- 		if (err < 0) {
-			rcu_read_unlock();
-			goto err;
-		}
+ 		if (err < 0)
+			goto err2;
 
 		helper = __nf_conntrack_helper_find_byname(helpname);
 		if (helper == NULL) {
@@ -1168,28 +1166,26 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 #ifdef CONFIG_MODULES
 			if (request_module("nfct-helper-%s", helpname) < 0) {
 				err = -EOPNOTSUPP;
-				goto err;
+				goto err1;
 			}
 
 			rcu_read_lock();
 			helper = __nf_conntrack_helper_find_byname(helpname);
 			if (helper) {
-				rcu_read_unlock();
 				err = -EAGAIN;
-				goto err;
+				goto err2;
 			}
 			rcu_read_unlock();
 #endif
 			err = -EOPNOTSUPP;
-			goto err;
+			goto err1;
 		} else {
 			struct nf_conn_help *help;
 
 			help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
 			if (help == NULL) {
-				rcu_read_unlock();
 				err = -ENOMEM;
-				goto err;
+				goto err2;
 			}
 
 			/* not in hash table yet so not strictly necessary */
@@ -1198,44 +1194,34 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 	} else {
 		/* try an implicit helper assignation */
 		err = __nf_ct_try_assign_helper(ct, GFP_ATOMIC);
-		if (err < 0) {
-			rcu_read_unlock();
-			goto err;
-		}
+		if (err < 0)
+			goto err2;
 	}
 
 	if (cda[CTA_STATUS]) {
 		err = ctnetlink_change_status(ct, cda);
-		if (err < 0) {
-			rcu_read_unlock();
-			goto err;
-		}
+		if (err < 0)
+			goto err2;
 	}
 
 	if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
 		err = ctnetlink_change_nat(ct, cda);
-		if (err < 0) {
-			rcu_read_unlock();
-			goto err;
-		}
+		if (err < 0)
+			goto err2;
 	}
 
 #ifdef CONFIG_NF_NAT_NEEDED
 	if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
 		err = ctnetlink_change_nat_seq_adj(ct, cda);
-		if (err < 0) {
-			rcu_read_unlock();
-			goto err;
-		}
+		if (err < 0)
+			goto err2;
 	}
 #endif
 
 	if (cda[CTA_PROTOINFO]) {
 		err = ctnetlink_change_protoinfo(ct, cda);
-		if (err < 0) {
-			rcu_read_unlock();
-			goto err;
-		}
+		if (err < 0)
+			goto err2;
 	}
 
 	nf_ct_acct_ext_add(ct, GFP_ATOMIC);
@@ -1253,12 +1239,12 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 
 		err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
 		if (err < 0)
-			goto err;
+			goto err2;
 
 		master_h = __nf_conntrack_find(&init_net, &master);
 		if (master_h == NULL) {
 			err = -ENOENT;
-			goto err;
+			goto err2;
 		}
 		master_ct = nf_ct_tuplehash_to_ctrack(master_h);
 		nf_conntrack_get(&master_ct->ct_general);
@@ -1271,7 +1257,10 @@ ctnetlink_create_conntrack(struct nlattr *cda[],
 	rcu_read_unlock();
 
 	return ct;
-err:
+
+err2:
+	rcu_read_unlock();
+err1:
 	nf_conntrack_free(ct);
 	return ERR_PTR(err);
 }
author maintainer
#9 lore raw
Quoted 9 lines · Patrick McHardy
Quoted 6 lines · Patrick McHardy
Quoted 3 lines · Pablo Neira Ayuso

This patch moves the assignation of the master conntrack to ctnetlink_create_conntrack(), which is where it really belongs. This patch is a cleanup.

Applied, thanks.

I've added this patch on top to fix a RCU context imbalance.

Oh, stupid mistake, sorry for that.

Quoted 2 lines

This seems like a good opportunity to say this again: please (everyone) compile your code using sparse. It catches this type and more bugs.

I'll do.

Quoted 9 lines

There is a second bug introduced by this patch:

    add_timer(&ct->timeout);
    nf_conntrack_hash_insert(ct);
    rcu_read_unlock();

    return ct;

The conntrack lock is not held, it might crash or create double entries.

Hm, but that code is inside ctnetlink_create_conntrack() which is called with the conntrack lock held.

if (nlh->nlmsg_flags & NLM_F_CREATE)
err = ctnetlink_create_conntrack(cda,
&otuple, &rtuple,
- master_ct,
NETLINK_CB(skb).pid,

-                                                 nlmsg_report(nlh));
+                                                 nlmsg_report(nlh),
+                                                 u3);
                spin_unlock_bh(&nf_conntrack_lock);
Signature · 1 line
"Los honestos son inadaptados sociales" -- Les Luthiers
#10 lore raw
Quoted 13 lines · Pablo Neira Ayuso
Quoted 2 lines · Patrick McHardy

The conntrack lock is not held, it might crash or create double entries.

Hm, but that code is inside ctnetlink_create_conntrack() which is called with the conntrack lock held.

if (nlh->nlmsg_flags & NLM_F_CREATE)
err = ctnetlink_create_conntrack(cda,
&otuple, &rtuple,

Review

Quoted 6 lines
(quoted diff) +2 -2
-                                                 master_ct,
                                                  NETLINK_CB(skb).pid,
-                                                 nlmsg_report(nlh));
+                                                 nlmsg_report(nlh),
+                                                 u3);
                spin_unlock_bh(&nf_conntrack_lock);

Right, I missed that, sorry.