Skip to contentSkip to conversation

multiple neighbour cache tables for AF_INET

PATCH Wilfried Weissmann 3 participants 4 messages Jan 28, 2005 – Jan 29, 2005 Networking

lore
Share & tools
lore
#0 lore raw

Hi,

The kernels 2.4.28+ and 2.6.9+ with IPv4 and ATM-CLIP enabled have bugs in the neighbour cache code. neigh_delete() and neigh_add() only work properly if one cache table per address family exist. After ATM-CLIP installed a second cache table for AF_INET, neigh_delete() and neigh_add() only examine the first table (the ATM-CLIP table if IPv4 and ATM-CLIP are compiled into the kernel). neigh_dump_info() is also affected if the neigh_dump_table() call fails.

This bug causes "ip neigh flush dev <some ethernet device>" to run into an infinite loop when the arp cache is already populated (debian bug #282492). "ip neigh delete ..." commands for non ATM-CLIP entries fail with an EINVAL. This is because of the IPv4 neighbour table that is installed in arp.c is never examined when sending delete commands.

I have attached a minimally tested patch for 2.4.28 that fixes this problem.

Greetings,
Wilfried

Signature · 2 lines
10 GB Mailbox, 100 FreeSMS http://www.gmx.net/de/go/topmail
+++ GMX - die erste Adresse für Mail, Message, More +++
net/core/neighbour.c +46 -17
@@ -1351,7 +1351,6 @@ int neigh_delete(struct sk_buff *skb, st
 
 		if (tbl->family != ndm->ndm_family)
 			continue;
-		read_unlock(&neigh_tbl_lock);
 
 		err = -EINVAL;
 		if (nda[NDA_DST-1] == NULL ||
@@ -1360,18 +1359,28 @@ int neigh_delete(struct sk_buff *skb, st
 
 		if (ndm->ndm_flags&NTF_PROXY) {
 			err = pneigh_delete(tbl, RTA_DATA(nda[NDA_DST-1]), dev);
-			goto out;
+			if(err) {
+				continue;	/* maybe in another table */
+			}
+			else {
+				goto out;
+			}
 		}
 
-		if (dev == NULL)
-			return -EINVAL;
+		if (dev == NULL) {
+			goto out;
+		}
 
 		n = neigh_lookup(tbl, RTA_DATA(nda[NDA_DST-1]), dev);
 		if (n) {
 			err = neigh_update(n, NULL, NUD_FAILED, 1, 0);
 			neigh_release(n);
 		}
+		else {
+			continue;	/* maybe in another table */
+		}
 out:
+		read_unlock(&neigh_tbl_lock);
 		if (dev)
 			dev_put(dev);
 		return err;
@@ -1390,6 +1399,8 @@ int neigh_add(struct sk_buff *skb, struc
 	struct rtattr **nda = arg;
 	struct neigh_table *tbl;
 	struct net_device *dev = NULL;
+	int err = -EADDRNOTAVAIL;
+	int olderr;
 
 	if (ndm->ndm_ifindex) {
 		if ((dev = dev_get_by_index(ndm->ndm_ifindex)) == NULL)
@@ -1398,35 +1409,47 @@ int neigh_add(struct sk_buff *skb, struc
 
 	read_lock(&neigh_tbl_lock);
 	for (tbl=neigh_tables; tbl; tbl = tbl->next) {
-		int err = 0;
 		int override = 1;
 		struct neighbour *n;
 
 		if (tbl->family != ndm->ndm_family)
 			continue;
-		read_unlock(&neigh_tbl_lock);
 
-		err = -EINVAL;
 		if (nda[NDA_DST-1] == NULL ||
-		    nda[NDA_DST-1]->rta_len != RTA_LENGTH(tbl->key_len))
+		    nda[NDA_DST-1]->rta_len != RTA_LENGTH(tbl->key_len)) {
+			err = -EINVAL;
 			goto out;
+		}
+
 		if (ndm->ndm_flags&NTF_PROXY) {
 			err = -ENOBUFS;
-			if (pneigh_lookup(tbl, RTA_DATA(nda[NDA_DST-1]), dev, 1))
+			if (pneigh_lookup(tbl, RTA_DATA(nda[NDA_DST-1]), dev, 1)) {
 				err = 0;
+				goto out;
+			}
+			else {
+				continue;	/* maybe in another table */
+			}
+		}
+		if (dev == NULL) {
+			err = -EINVAL;
 			goto out;
 		}
-		if (dev == NULL)
-			return -EINVAL;
-		err = -EINVAL;
+
 		if (nda[NDA_LLADDR-1] != NULL &&
-		    nda[NDA_LLADDR-1]->rta_len != RTA_LENGTH(dev->addr_len))
+		    nda[NDA_LLADDR-1]->rta_len != RTA_LENGTH(dev->addr_len)) {
+			err = -EINVAL;
 			goto out;
+		}
+
+		olderr=err;
 		err = 0;
 		n = neigh_lookup(tbl, RTA_DATA(nda[NDA_DST-1]), dev);
 		if (n) {
-			if (nlh->nlmsg_flags&NLM_F_EXCL)
+			if (nlh->nlmsg_flags&NLM_F_EXCL) {
 				err = -EEXIST;
+				goto outneigh;
+			}
 			override = nlh->nlmsg_flags&NLM_F_REPLACE;
 		} else if (!(nlh->nlmsg_flags&NLM_F_CREATE))
 			err = -ENOENT;
@@ -1442,9 +1465,16 @@ int neigh_add(struct sk_buff *skb, struc
 					   ndm->ndm_state,
 					   override, 0);
 		}
+		else {
+			err=olderr;
+			continue;	/* maybe in another table */
+		}
+
+outneigh:
 		if (n)
 			neigh_release(n);
 out:
+		read_unlock(&neigh_tbl_lock);
 		if (dev)
 			dev_put(dev);
 		return err;
@@ -1453,7 +1483,7 @@ out:
 
 	if (dev)
 		dev_put(dev);
-	return -EADDRNOTAVAIL;
+	return err;
 }
 
 
@@ -1547,8 +1577,7 @@ int neigh_dump_info(struct sk_buff *skb,
 			continue;
 		if (t > s_t)
 			memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
-		if (neigh_dump_table(tbl, skb, cb) < 0) 
-			break;
+		neigh_dump_table(tbl, skb, cb);
 	}
 	read_unlock(&neigh_tbl_lock);
 
#1 lore raw
Quoted 8 lines · Wilfried Weissmann

The kernels 2.4.28+ and 2.6.9+ with IPv4 and ATM-CLIP enabled have bugs in the neighbour cache code. neigh_delete() and neigh_add() only work properly if one cache table per address family exist. After ATM-CLIP installed a second cache table for AF_INET, neigh_delete() and neigh_add() only examine the first table (the ATM-CLIP table if IPv4 and ATM-CLIP are compiled into the kernel). neigh_dump_info() is also affected if the neigh_dump_table() call fails.

Indeed, this has been the case for a very long time.

IMHO you need to give the user a way to specify which table they want to operate on. If they don't specify one, then the current behaviour of choosing the first table found is reasonble.

Cheers,

Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
#2 lore raw

In article <[email protected]> (at Sat, 29 Jan 2005 09:19:49 +1100), Herbert Xu <[email protected]> says:

Quoted 3 lines

IMHO you need to give the user a way to specify which table they want to operate on. If they don't specify one, then the current behaviour of choosing the first table found is reasonble.

We have dev. Isn't is sufficient?

--yoshfuji

#3 lore raw
Quoted 7 lines · On Sat, Jan 29, 2005 at 07:46:05AM +0900, YOSHIFUJI Hideaki / ?$B5HF#1QL@ wrote:

In article <[email protected]> (at Sat, 29 Jan 2005 09:19:49 +1100), Herbert Xu <[email protected]> says:

Quoted 3 lines

IMHO you need to give the user a way to specify which table they want to operate on. If they don't specify one, then the current behaviour of choosing the first table found is reasonble.

We have dev. Isn't is sufficient?

It could be used for neigh_add/neigh_delete. We'll need to add a way to query whether a given table is the right one for a device.

For dump it isn't the same. However, perhaps it's not too important to query a specific table.

Cheers,

Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt