ipv4: RCU conversion of ip_route_input_slow/ip_route_input_mc
authorEric Dumazet <eric.dumazet@gmail.com>
Wed, 2 Jun 2010 19:21:31 +0000 (19:21 +0000)
committerDavid S. Miller <davem@davemloft.net>
Thu, 3 Jun 2010 10:18:19 +0000 (03:18 -0700)
Avoid two atomic ops on struct in_device refcount per incoming packet,
if slow path taken, (or route cache disabled)

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/route.c

index d377b45..1cfe0d1 100644 (file)
@@ -1843,13 +1843,14 @@ static void rt_set_nexthop(struct rtable *rt, struct fib_result *res, u32 itag)
        rt->rt_type = res->type;
 }
 
+/* called in rcu_read_lock() section */
 static int ip_route_input_mc(struct sk_buff *skb, __be32 daddr, __be32 saddr,
                                u8 tos, struct net_device *dev, int our)
 {
-       unsigned hash;
+       unsigned int hash;
        struct rtable *rth;
        __be32 spec_dst;
-       struct in_device *in_dev = in_dev_get(dev);
+       struct in_device *in_dev = __in_dev_get_rcu(dev);
        u32 itag = 0;
        int err;
 
@@ -1914,18 +1915,14 @@ static int ip_route_input_mc(struct sk_buff *skb, __be32 daddr, __be32 saddr,
 #endif
        RT_CACHE_STAT_INC(in_slow_mc);
 
-       in_dev_put(in_dev);
        hash = rt_hash(daddr, saddr, dev->ifindex, rt_genid(dev_net(dev)));
        return rt_intern_hash(hash, rth, NULL, skb, dev->ifindex);
 
 e_nobufs:
-       in_dev_put(in_dev);
        return -ENOBUFS;
-
 e_inval:
-       err = -EINVAL;
+       return -EINVAL;
 e_err:
-       in_dev_put(in_dev);
        return err;
 }
 
@@ -2101,7 +2098,7 @@ static int ip_route_input_slow(struct sk_buff *skb, __be32 daddr, __be32 saddr,
                               u8 tos, struct net_device *dev)
 {
        struct fib_result res;
-       struct in_device *in_dev = in_dev_get(dev);
+       struct in_device *in_dev = __in_dev_get_rcu(dev);
        struct flowi fl = { .nl_u = { .ip4_u =
                                      { .daddr = daddr,
                                        .saddr = saddr,
@@ -2179,7 +2176,6 @@ static int ip_route_input_slow(struct sk_buff *skb, __be32 daddr, __be32 saddr,
 
        err = ip_mkroute_input(skb, &res, &fl, in_dev, daddr, saddr, tos);
 done:
-       in_dev_put(in_dev);
        if (free_res)
                fib_res_put(&res);
 out:   return err;
@@ -2288,16 +2284,18 @@ int ip_route_input_common(struct sk_buff *skb, __be32 daddr, __be32 saddr,
        unsigned        hash;
        int iif = dev->ifindex;
        struct net *net;
+       int res;
 
        net = dev_net(dev);
 
+       rcu_read_lock();
+
        if (!rt_caching(net))
                goto skip_cache;
 
        tos &= IPTOS_RT_MASK;
        hash = rt_hash(daddr, saddr, iif, rt_genid(net));
 
-       rcu_read_lock();
        for (rth = rcu_dereference(rt_hash_table[hash].chain); rth;
             rth = rcu_dereference(rth->u.dst.rt_next)) {
                if ((((__force u32)rth->fl.fl4_dst ^ (__force u32)daddr) |
@@ -2321,7 +2319,6 @@ int ip_route_input_common(struct sk_buff *skb, __be32 daddr, __be32 saddr,
                }
                RT_CACHE_STAT_INC(in_hlist_search);
        }
-       rcu_read_unlock();
 
 skip_cache:
        /* Multicast recognition logic is moved from route cache to here.
@@ -2336,12 +2333,11 @@ skip_cache:
           route cache entry is created eventually.
         */
        if (ipv4_is_multicast(daddr)) {
-               struct in_device *in_dev;
+               struct in_device *in_dev = __in_dev_get_rcu(dev);
 
-               rcu_read_lock();
-               if ((in_dev = __in_dev_get_rcu(dev)) != NULL) {
+               if (in_dev) {
                        int our = ip_check_mc(in_dev, daddr, saddr,
-                               ip_hdr(skb)->protocol);
+                                             ip_hdr(skb)->protocol);
                        if (our
 #ifdef CONFIG_IP_MROUTE
                                ||
@@ -2349,15 +2345,18 @@ skip_cache:
                             IN_DEV_MFORWARD(in_dev))
 #endif
                           ) {
+                               int res = ip_route_input_mc(skb, daddr, saddr,
+                                                           tos, dev, our);
                                rcu_read_unlock();
-                               return ip_route_input_mc(skb, daddr, saddr,
-                                                        tos, dev, our);
+                               return res;
                        }
                }
                rcu_read_unlock();
                return -EINVAL;
        }
-       return ip_route_input_slow(skb, daddr, saddr, tos, dev);
+       res = ip_route_input_slow(skb, daddr, saddr, tos, dev);
+       rcu_read_unlock();
+       return res;
 }
 EXPORT_SYMBOL(ip_route_input_common);