ipv6: Fix sporadic sendmsg -EINVAL when sending to multicast groups.
Thanks to excellent diagnosis by Eduard Guzovsky.
The core problem is that on a network with lots of active
multicast traffic, the neighbour cache can fill up. If
we try to allocate a new route and thus neighbour cache
entry, the bog-standard GC attempt the neighbour layer does
in ineffective because route entries hold a reference
to the existing neighbour entries and GC can only liberate
entries with no references.
IPV4 already has a way to handle this, by doing a route cache
GC in such situations (when neigh attach returns -ENOBUFS).
So simply mimick this on the ipv6 side.
Tested-by: Eduard Guzovsky <eguzovsky@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/include/net/ndisc.h b/include/net/ndisc.h
index ce532f2..1459ed3 100644
--- a/include/net/ndisc.h
+++ b/include/net/ndisc.h
@@ -155,9 +155,9 @@
{
if (dev)
- return __neigh_lookup(&nd_tbl, addr, dev, 1);
+ return __neigh_lookup_errno(&nd_tbl, addr, dev);
- return NULL;
+ return ERR_PTR(-ENODEV);
}
diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index 18c486c..76f06b9 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -627,6 +627,9 @@
rt = ip6_rt_copy(ort);
if (rt) {
+ struct neighbour *neigh;
+ int attempts = !in_softirq();
+
if (!(rt->rt6i_flags&RTF_GATEWAY)) {
if (rt->rt6i_dst.plen != 128 &&
ipv6_addr_equal(&rt->rt6i_dst.addr, daddr))
@@ -646,7 +649,35 @@
}
#endif
- rt->rt6i_nexthop = ndisc_get_neigh(rt->rt6i_dev, &rt->rt6i_gateway);
+ retry:
+ neigh = ndisc_get_neigh(rt->rt6i_dev, &rt->rt6i_gateway);
+ if (IS_ERR(neigh)) {
+ struct net *net = dev_net(rt->rt6i_dev);
+ int saved_rt_min_interval =
+ net->ipv6.sysctl.ip6_rt_gc_min_interval;
+ int saved_rt_elasticity =
+ net->ipv6.sysctl.ip6_rt_gc_elasticity;
+
+ if (attempts-- > 0) {
+ net->ipv6.sysctl.ip6_rt_gc_elasticity = 1;
+ net->ipv6.sysctl.ip6_rt_gc_min_interval = 0;
+
+ ip6_dst_gc(net->ipv6.ip6_dst_ops);
+
+ net->ipv6.sysctl.ip6_rt_gc_elasticity =
+ saved_rt_elasticity;
+ net->ipv6.sysctl.ip6_rt_gc_min_interval =
+ saved_rt_min_interval;
+ goto retry;
+ }
+
+ if (net_ratelimit())
+ printk(KERN_WARNING
+ "Neighbour table overflow.\n");
+ dst_free(&rt->u.dst);
+ return NULL;
+ }
+ rt->rt6i_nexthop = neigh;
}
@@ -945,8 +976,11 @@
dev_hold(dev);
if (neigh)
neigh_hold(neigh);
- else
+ else {
neigh = ndisc_get_neigh(dev, addr);
+ if (IS_ERR(neigh))
+ neigh = NULL;
+ }
rt->rt6i_dev = dev;
rt->rt6i_idev = idev;
@@ -1887,6 +1921,7 @@
{
struct net *net = dev_net(idev->dev);
struct rt6_info *rt = ip6_dst_alloc(net->ipv6.ip6_dst_ops);
+ struct neighbour *neigh;
if (rt == NULL)
return ERR_PTR(-ENOMEM);
@@ -1909,11 +1944,18 @@
rt->rt6i_flags |= RTF_ANYCAST;
else
rt->rt6i_flags |= RTF_LOCAL;
- rt->rt6i_nexthop = ndisc_get_neigh(rt->rt6i_dev, &rt->rt6i_gateway);
- if (rt->rt6i_nexthop == NULL) {
+ neigh = ndisc_get_neigh(rt->rt6i_dev, &rt->rt6i_gateway);
+ if (IS_ERR(neigh)) {
dst_free(&rt->u.dst);
- return ERR_PTR(-ENOMEM);
+
+ /* We are casting this because that is the return
+ * value type. But an errno encoded pointer is the
+ * same regardless of the underlying pointer type,
+ * and that's what we are returning. So this is OK.
+ */
+ return (struct rt6_info *) neigh;
}
+ rt->rt6i_nexthop = neigh;
ipv6_addr_copy(&rt->rt6i_dst.addr, addr);
rt->rt6i_dst.plen = 128;