net: fix neigh_delete to handle mult. tables
PATCH 1/1 net: fix neigh_delete to handle mult. tables
The neigh_delete assumes that there is only one table with the desired
protocol family for the address being deleted. This assumption is not
always valid. For example, in a kernel configured with both the
typical Ethernet configuration (CONFIG_NET_ETHERNET=y etc.) as well
as ATM (CONFIG_ATM=y and CONFIG_ATM_CLIP=y), there are two tables, namely arp_tbl and clip_tbl, both with family AF_INET.
The problem fixed by this patch is that the "for" loop in neigh_delete will not iterate further once it finds one table with a matching family. The code changes refactor neigh_delete such that it will continue
to iterate through the neighbor tables until it finds a match or finishes the tables.
To see the symptom of the problem, you must have a kernel with more than one AF_INET transport configured, then try to delete a neighbor table entry. The delete will fail with EINVAL. For example:
[root@bonefish ~]# ip neigh list dev eth4 192.168.4.24 lladdr 00:90:fb:a1:07:2f nud reachable [root@bonefish ~]# ip neigh delete 192.168.4.24 dev eth4 RTNETLINK answers: Invalid argument [root@bonefish ~]# ip neigh list dev eth4 192.168.4.24 lladdr 00:90:fb:a1:07:2f nud reachable
One side effect of this change is that neigh_lookup is now called while a read lock is held on neigh_tbl_lock. neigh_lookup does take read locks on the tbl->lock set. However, this does not introduce any new lock order dependencies, since we already have that precedent set in the neightbl_set function.
---
I have built this patch against the 2.6.17-rc1 kernel.
I tested this patch with the simple test case described above, plus let it run for a few days on the box to make sure it behaves OK. I also verified that I can properly delete IPV6 neighbor entries.
--- linux-2.6.17-rc1/net/core/neighbour.c.orig 2006-04-11 16:28:10.000000000 -0 400
+++ linux-2.6.17-rc1/net/core/neighbour.c 2006-04-13 10:01:19.000000000 -0
400
@@ -1430,48 +1430,54 @@ int neigh_delete(struct sk_buff *skb, st
struct rtattr **nda = arg;
struct neigh_table *tbl;
struct net_device *dev = NULL;- int err = -ENODEV;
+ int err;
if (ndm->ndm_ifindex &&
- (dev = dev_get_by_index(ndm->ndm_ifindex)) == NULL)
- goto out;
+ (dev = dev_get_by_index(ndm->ndm_ifindex)) == NULL)
+ return -ENODEV;
read_lock(&neigh_tbl_lock);
- for (tbl = neigh_tables; tbl; tbl = tbl->next) {
+ for (tbl = neigh_tables; ; tbl = tbl->next) {
struct rtattr *dst_attr = nda[NDA_DST - 1];
struct neighbour *n;+ if (!tbl) {
+ read_unlock(&neigh_tbl_lock);
+ err = -EADDRNOTAVAIL;
+ break;
+ }
if (tbl->family != ndm->ndm_family)
continue;- read_unlock(&neigh_tbl_lock);
err = -EINVAL;
- if (!dst_attr || RTA_PAYLOAD(dst_attr) < tbl->key_len)
- goto out_dev_put;
+ if (!dst_attr || RTA_PAYLOAD(dst_attr) < tbl->key_len) {
+ read_unlock(&neigh_tbl_lock);
+ break;
+ }
if (ndm->ndm_flags & NTF_PROXY) {
+ read_unlock(&neigh_tbl_lock);
err = pneigh_delete(tbl, RTA_DATA(dst_attr), dev);
- goto out_dev_put;
+ break;
}- if (!dev)
- goto out;
+ if (!dev) {
+ read_unlock(&neigh_tbl_lock);
+ break;
+ }
n = neigh_lookup(tbl, RTA_DATA(dst_attr), dev);
if (n) {
- err = neigh_update(n, NULL, NUD_FAILED, + read_unlock(&neigh_tbl_lock);
+ err = neigh_update(n, NULL, NUD_FAILED,
NEIGH_UPDATE_F_OVERRIDE|
NEIGH_UPDATE_F_ADMIN);
neigh_release(n);
+ break;
}
- goto out_dev_put;
}
- read_unlock(&neigh_tbl_lock);
- err = -EADDRNOTAVAIL;
-out_dev_put:
if (dev)
dev_put(dev);
-out:
return err;
}In article <[email protected]> (at Thu, 20 Apr 2006 23:31:24 -0400), Don Law <[email protected]> says:
Quoted 5 lines
One side effect of this change is that
neigh_lookupis now called while a read lock is held onneigh_tbl_lock.neigh_lookupdoes take read locks on thetbl->lockset. However, this does not introduce any new lock order dependencies, since we already have that precedent set in theneightbl_setfunction.
I don't think this is good change.
What you need to fix your problem is to remove the last "goto out_dev_put;" isn't it?
--yoshfuji
In article <[email protected]> (at Fri, 21 Apr 2006 12:55:43 +0900 (JST)), YOSHIFUJI Hideaki <[email protected]> says:
Quoted 8 lines
In article <[email protected]> (at Thu, 20 Apr 2006 23:31:24 -0400), Don Law <[email protected]> says:
Quoted 5 lines
One side effect of this change is that
neigh_lookupis now called while a read lock is held onneigh_tbl_lock.neigh_lookupdoes take read locks on thetbl->lockset. However, this does not introduce any new lock order dependencies, since we already have that precedent set in theneightbl_setfunction.
Quoted 4 lines
I don't think this is good change.
What you need to fix your problem is to remove the last "goto
out_dev_put;" isn't it?
Hmm, okay, I notice we need to hold lock for pneigh_delete(). How about this? Also handles proxy neighbors.
net/core/neighbour.c +13 -10
@@ -1430,6 +1430,8 @@ int neigh_delete(struct sk_buff *skb, st (dev = dev_get_by_index(ndm->ndm_ifindex)) == NULL) goto out; + err = -EADDRNOTAVAIL ; +read_lock (&neigh_tbl_lock);for (tbl = neigh_tables; tbl; tbl = tbl->next) {struct rtattr *dst_attr = nda[NDA_DST -1 ];
@@ -1437,32 +1439,33 @@ int neigh_delete(struct sk_buff *skb, stif (tbl->family != ndm->ndm_family)continue ; -read_unlock (&neigh_tbl_lock); - err = -EINVAL ; -if (!dst_attr ||RTA_PAYLOAD (dst_attr) < tbl->key_len) - goto out_dev_put; +if (!dst_attr ||RTA_PAYLOAD (dst_attr) < tbl->key_len) { + err = -EINVAL ; +break ; + }if (ndm->ndm_flags &NTF_PROXY ) { - err =pneigh_delete (tbl,RTA_DATA (dst_attr), dev); - goto out_dev_put; +if (!pneigh_delete (tbl,RTA_DATA (dst_attr), dev)) { +read_unlock (&neigh_tbl_lock); +break ; + } }if (!dev) - goto out; +continue ; n =neigh_lookup (tbl,RTA_DATA (dst_attr), dev);if (n) { +read_unlock (&neigh_tbl_lock); err =neigh_update (n,NULL ,NUD_FAILED ,NEIGH_UPDATE_F_OVERRIDE |NEIGH_UPDATE_F_ADMIN );neigh_release (n); +break ; } - goto out_dev_put; }read_unlock (&neigh_tbl_lock); - err = -EADDRNOTAVAIL ; -out_dev_put:if (dev)dev_put (dev); out:
Signature · 2 lines
YOSHIFUJI Hideaki @ USAGI Project <[email protected]> GPG-FP : 9022 65EB 1ECF 3AD1 0BDF 80D8 4807 F894 E062 0EEA