ipv6: Avoid double dst_free

It is a prep work to get dst freeing from fib tree undergo
a rcu grace period.

The following is a common paradigm:
if (ip6_del_rt(rt))
	dst_free(rt)

which means, if rt cannot be deleted from the fib tree, dst_free(rt) now.
1. We don't know the ip6_del_rt(rt) failure is because it
   was not managed by fib tree (e.g. DST_NOCACHE) or it had already been
   removed from the fib tree.
2. If rt had been managed by the fib tree, ip6_del_rt(rt) failure means
   dst_free(rt) has been called already.  A second
   dst_free(rt) is not always obviously safe.  The rt may have
   been destroyed already.
3. If rt is a DST_NOCACHE, dst_free(rt) should not be called.
4. It is a stopper to make dst freeing from fib tree undergo a
   rcu grace period.

This patch is to use a DST_NOCACHE flag to indicate a rt is
not managed by the fib tree.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/ipv6/addrconf.c

@@ -5127,13 +5127,12 @@ static void __ipv6_ifa_notify(int event, struct inet6_ifaddr *ifp)
 
 			rt = addrconf_get_prefix_route(&ifp->peer_addr, 128,
 						       ifp->idev->dev, 0, 0);
-			if (rt && ip6_del_rt(rt))
-				dst_free(&rt->dst);
+			if (rt)
+				ip6_del_rt(rt);
 		}
 		dst_hold(&ifp->rt->dst);
 
-		if (ip6_del_rt(ifp->rt))
-			dst_free(&ifp->rt->dst);
+		ip6_del_rt(ifp->rt);
 
 		rt_genid_bump_ipv6(net);
 		break;

net/ipv6/ip6_fib.c

@@ -933,6 +933,10 @@ int fib6_add(struct fib6_node *root, struct rt6_info *rt,
 	int replace_required = 0;
 	int sernum = fib6_new_sernum(info->nl_net);
 
+	if (WARN_ON_ONCE((rt->dst.flags & DST_NOCACHE) &&
+			 !atomic_read(&rt->dst.__refcnt)))
+		return -EINVAL;
+
 	if (info->nlh) {
 		if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
 			allow_create = 0;
@@ -1025,6 +1029,7 @@ int fib6_add(struct fib6_node *root, struct rt6_info *rt,
 		fib6_start_gc(info->nl_net, rt);
 		if (!(rt->rt6i_flags & RTF_CACHE))
 			fib6_prune_clones(info->nl_net, pn);
+		rt->dst.flags &= ~DST_NOCACHE;
 	}
 
 out:
@@ -1049,7 +1054,8 @@ int fib6_add(struct fib6_node *root, struct rt6_info *rt,
 			atomic_inc(&pn->leaf->rt6i_ref);
 		}
 #endif
-		dst_free(&rt->dst);
+		if (!(rt->dst.flags & DST_NOCACHE))
+			dst_free(&rt->dst);
 	}
 	return err;
 
@@ -1060,7 +1066,8 @@ int fib6_add(struct fib6_node *root, struct rt6_info *rt,
 st_failure:
 	if (fn && !(fn->fn_flags & (RTN_RTINFO|RTN_ROOT)))
 		fib6_repair_tree(info->nl_net, fn);
-	dst_free(&rt->dst);
+	if (!(rt->dst.flags & DST_NOCACHE))
+		dst_free(&rt->dst);
 	return err;
 #endif
 }

net/ipv6/route.c

@@ -1322,8 +1322,7 @@ static void ip6_link_failure(struct sk_buff *skb)
 	if (rt) {
 		if (rt->rt6i_flags & RTF_CACHE) {
 			dst_hold(&rt->dst);
-			if (ip6_del_rt(rt))
-				dst_free(&rt->dst);
+			ip6_del_rt(rt);
 		} else if (rt->rt6i_node && (rt->rt6i_flags & RTF_DEFAULT)) {
 			rt->rt6i_node->fn_sernum = -1;
 		}
@@ -2028,7 +2027,8 @@ static int __ip6_del_rt(struct rt6_info *rt, struct nl_info *info)
 	struct fib6_table *table;
 	struct net *net = dev_net(rt->dst.dev);
 
-	if (rt == net->ipv6.ip6_null_entry) {
+	if (rt == net->ipv6.ip6_null_entry ||
+	    rt->dst.flags & DST_NOCACHE) {
 		err = -ENOENT;
 		goto out;
 	}
@@ -2515,6 +2515,7 @@ struct rt6_info *addrconf_dst_alloc(struct inet6_dev *idev,
 	rt->rt6i_dst.addr = *addr;
 	rt->rt6i_dst.plen = 128;
 	rt->rt6i_table = fib6_get_table(net, RT6_TABLE_LOCAL);
+	rt->dst.flags |= DST_NOCACHE;
 
 	atomic_set(&rt->dst.__refcnt, 1);