net/ipv6/inet6_hashtables.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Generic INET6 transport hashtables
   7  *
   8  * Authors:     Lotsa people, from code originally in tcp, generalised here
   9  *              by Arnaldo Carvalho de Melo <acme@mandriva.com>
  10  *
  11  *      This program is free software; you can redistribute it and/or
  12  *      modify it under the terms of the GNU General Public License
  13  *      as published by the Free Software Foundation; either version
  14  *      2 of the License, or (at your option) any later version.
  15  */
  16
  17 #include <linux/module.h>
  18 #include <linux/random.h>
  19
  20 #include <net/inet_connection_sock.h>
  21 #include <net/inet_hashtables.h>
  22 #include <net/inet6_hashtables.h>
  23 #include <net/ip.h>
  24
  25 void __inet6_hash(struct sock *sk)
  26 {
  27         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  28
  29         WARN_ON(!sk_unhashed(sk));
  30
  31         if (sk->sk_state == TCP_LISTEN) {
  32                 struct inet_listen_hashbucket *ilb;
  33
  34                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
  35                 spin_lock(&ilb->lock);
  36                 __sk_add_node(sk, &ilb->head);
  37                 spin_unlock(&ilb->lock);
  38         } else {
  39                 unsigned int hash;
  40                 struct hlist_nulls_head *list;
  41                 spinlock_t *lock;
  42
  43                 sk->sk_hash = hash = inet6_sk_ehashfn(sk);
  44                 list = &inet_ehash_bucket(hashinfo, hash)->chain;
  45                 lock = inet_ehash_lockp(hashinfo, hash);
  46                 spin_lock(lock);
  47                 __sk_nulls_add_node_rcu(sk, list);
  48                 spin_unlock(lock);
  49         }
  50
  51         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  52 }
  53 EXPORT_SYMBOL(__inet6_hash);
  54
  55 /*
  56  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so
  57  * we need not check it for TCP lookups anymore, thanks Alexey. -DaveM
  58  *
  59  * The sockhash lock must be held as a reader here.
  60  */
  61 struct sock *__inet6_lookup_established(struct net *net,
  62                                         struct inet_hashinfo *hashinfo,
  63                                            const struct in6_addr *saddr,
  64                                            const __be16 sport,
  65                                            const struct in6_addr *daddr,
  66                                            const u16 hnum,
  67                                            const int dif)
  68 {
  69         struct sock *sk;
  70         const struct hlist_nulls_node *node;
  71         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
  72         /* Optimize here for direct hit, only listening connections can
  73          * have wildcards anyways.
  74          */
  75         unsigned int hash = inet6_ehashfn(net, daddr, hnum, saddr, sport);
  76         unsigned int slot = hash & (hashinfo->ehash_size - 1);
  77         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
  78
  79
  80         rcu_read_lock();
  81 begin:
  82         sk_nulls_for_each_rcu(sk, node, &head->chain) {
  83                 /* For IPV6 do the cheaper port and family tests first. */
  84                 if (INET6_MATCH(sk, net, hash, saddr, daddr, ports, dif)) {
  85                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
  86                                 goto begintw;
  87                         if (!INET6_MATCH(sk, net, hash, saddr, daddr, ports, dif)) {
  88                                 sock_put(sk);
  89                                 goto begin;
  90                         }
  91                 goto out;
  92                 }
  93         }
  94         if (get_nulls_value(node) != slot)
  95                 goto begin;
  96
  97 begintw:
  98         /* Must check for a TIME_WAIT'er before going to listener hash. */
  99         sk_nulls_for_each_rcu(sk, node, &head->twchain) {
 100                 if (INET6_TW_MATCH(sk, net, hash, saddr, daddr, ports, dif)) {
 101                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) {
 102                                 sk = NULL;
 103                                 goto out;
 104                         }
 105                         if (!INET6_TW_MATCH(sk, net, hash, saddr, daddr, ports, dif)) {
 106                                 sock_put(sk);
 107                                 goto begintw;
 108                         }
 109                         goto out;
 110                 }
 111         }
 112         if (get_nulls_value(node) != slot)
 113                 goto begintw;
 114         sk = NULL;
 115 out:
 116         rcu_read_unlock();
 117         return sk;
 118 }
 119 EXPORT_SYMBOL(__inet6_lookup_established);
 120
 121 struct sock *inet6_lookup_listener(struct net *net,
 122                 struct inet_hashinfo *hashinfo, const struct in6_addr *daddr,
 123                 const unsigned short hnum, const int dif)
 124 {
 125         struct sock *sk;
 126         const struct hlist_node *node;
 127         struct sock *result = NULL;
 128         int score, hiscore = 0;
 129         struct inet_listen_hashbucket *ilb;
 130
 131         ilb = &hashinfo->listening_hash[inet_lhashfn(net, hnum)];
 132         spin_lock(&ilb->lock);
 133         sk_for_each(sk, node, &ilb->head) {
 134                 if (net_eq(sock_net(sk), net) && inet_sk(sk)->num == hnum &&
 135                                 sk->sk_family == PF_INET6) {
 136                         const struct ipv6_pinfo *np = inet6_sk(sk);
 137
 138                         score = 1;
 139                         if (!ipv6_addr_any(&np->rcv_saddr)) {
 140                                 if (!ipv6_addr_equal(&np->rcv_saddr, daddr))
 141                                         continue;
 142                                 score++;
 143                         }
 144                         if (sk->sk_bound_dev_if) {
 145                                 if (sk->sk_bound_dev_if != dif)
 146                                         continue;
 147                                 score++;
 148                         }
 149                         if (score == 3) {
 150                                 result = sk;
 151                                 break;
 152                         }
 153                         if (score > hiscore) {
 154                                 hiscore = score;
 155                                 result = sk;
 156                         }
 157                 }
 158         }
 159         if (result)
 160                 sock_hold(result);
 161         spin_unlock(&ilb->lock);
 162         return result;
 163 }
 164
 165 EXPORT_SYMBOL_GPL(inet6_lookup_listener);
 166
 167 struct sock *inet6_lookup(struct net *net, struct inet_hashinfo *hashinfo,
 168                           const struct in6_addr *saddr, const __be16 sport,
 169                           const struct in6_addr *daddr, const __be16 dport,
 170                           const int dif)
 171 {
 172         struct sock *sk;
 173
 174         local_bh_disable();
 175         sk = __inet6_lookup(net, hashinfo, saddr, sport, daddr, ntohs(dport), dif);
 176         local_bh_enable();
 177
 178         return sk;
 179 }
 180
 181 EXPORT_SYMBOL_GPL(inet6_lookup);
 182
 183 static int __inet6_check_established(struct inet_timewait_death_row *death_row,
 184                                      struct sock *sk, const __u16 lport,
 185                                      struct inet_timewait_sock **twp)
 186 {
 187         struct inet_hashinfo *hinfo = death_row->hashinfo;
 188         struct inet_sock *inet = inet_sk(sk);
 189         const struct ipv6_pinfo *np = inet6_sk(sk);
 190         const struct in6_addr *daddr = &np->rcv_saddr;
 191         const struct in6_addr *saddr = &np->daddr;
 192         const int dif = sk->sk_bound_dev_if;
 193         const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
 194         struct net *net = sock_net(sk);
 195         const unsigned int hash = inet6_ehashfn(net, daddr, lport, saddr,
 196                                                 inet->dport);
 197         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 198         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 199         struct sock *sk2;
 200         const struct hlist_nulls_node *node;
 201         struct inet_timewait_sock *tw;
 202
 203         spin_lock(lock);
 204
 205         /* Check TIME-WAIT sockets first. */
 206         sk_nulls_for_each(sk2, node, &head->twchain) {
 207                 tw = inet_twsk(sk2);
 208
 209                 if (INET6_TW_MATCH(sk2, net, hash, saddr, daddr, ports, dif)) {
 210                         if (twsk_unique(sk, sk2, twp))
 211                                 goto unique;
 212                         else
 213                                 goto not_unique;
 214                 }
 215         }
 216         tw = NULL;
 217
 218         /* And established part... */
 219         sk_nulls_for_each(sk2, node, &head->chain) {
 220                 if (INET6_MATCH(sk2, net, hash, saddr, daddr, ports, dif))
 221                         goto not_unique;
 222         }
 223
 224 unique:
 225         /* Must record num and sport now. Otherwise we will see
 226          * in hash table socket with a funny identity. */
 227         inet->num = lport;
 228         inet->sport = htons(lport);
 229         WARN_ON(!sk_unhashed(sk));
 230         __sk_nulls_add_node_rcu(sk, &head->chain);
 231         sk->sk_hash = hash;
 232         spin_unlock(lock);
 233         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 234
 235         if (twp != NULL) {
 236                 *twp = tw;
 237                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITRECYCLED);
 238         } else if (tw != NULL) {
 239                 /* Silly. Should hash-dance instead... */
 240                 inet_twsk_deschedule(tw, death_row);
 241                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITRECYCLED);
 242
 243                 inet_twsk_put(tw);
 244         }
 245         return 0;
 246
 247 not_unique:
 248         spin_unlock(lock);
 249         return -EADDRNOTAVAIL;
 250 }
 251
 252 static inline u32 inet6_sk_port_offset(const struct sock *sk)
 253 {
 254         const struct inet_sock *inet = inet_sk(sk);
 255         const struct ipv6_pinfo *np = inet6_sk(sk);
 256         return secure_ipv6_port_ephemeral(np->rcv_saddr.s6_addr32,
 257                                           np->daddr.s6_addr32,
 258                                           inet->dport);
 259 }
 260
 261 int inet6_hash_connect(struct inet_timewait_death_row *death_row,
 262                        struct sock *sk)
 263 {
 264         return __inet_hash_connect(death_row, sk, inet6_sk_port_offset(sk),
 265                         __inet6_check_established, __inet6_hash);
 266 }
 267
 268 EXPORT_SYMBOL_GPL(inet6_hash_connect);