Pileus Git - ~andy/linux/blob - 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/secure_seq.h>
  24 #include <net/ip.h>
  25
  26 static unsigned int inet6_ehashfn(struct net *net,
  27                                   const struct in6_addr *laddr,
  28                                   const u16 lport,
  29                                   const struct in6_addr *faddr,
  30                                   const __be16 fport)
  31 {
  32         const u32 lhash = (__force u32)laddr->s6_addr32[3];
  33         const u32 fhash = __ipv6_addr_jhash(faddr, ipv6_hash_secret);
  34         return __inet6_ehashfn(lhash, lport, fhash, fport,
  35                                inet_ehash_secret + net_hash_mix(net));
  36 }
  37
  38 static int inet6_sk_ehashfn(const struct sock *sk)
  39 {
  40         const struct inet_sock *inet = inet_sk(sk);
  41         const struct in6_addr *laddr = &sk->sk_v6_rcv_saddr;
  42         const struct in6_addr *faddr = &sk->sk_v6_daddr;
  43         const __u16 lport = inet->inet_num;
  44         const __be16 fport = inet->inet_dport;
  45         struct net *net = sock_net(sk);
  46
  47         return inet6_ehashfn(net, laddr, lport, faddr, fport);
  48 }
  49
  50 int __inet6_hash(struct sock *sk, struct inet_timewait_sock *tw)
  51 {
  52         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  53         int twrefcnt = 0;
  54
  55         WARN_ON(!sk_unhashed(sk));
  56
  57         if (sk->sk_state == TCP_LISTEN) {
  58                 struct inet_listen_hashbucket *ilb;
  59
  60                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
  61                 spin_lock(&ilb->lock);
  62                 __sk_nulls_add_node_rcu(sk, &ilb->head);
  63                 spin_unlock(&ilb->lock);
  64         } else {
  65                 unsigned int hash;
  66                 struct hlist_nulls_head *list;
  67                 spinlock_t *lock;
  68
  69                 sk->sk_hash = hash = inet6_sk_ehashfn(sk);
  70                 list = &inet_ehash_bucket(hashinfo, hash)->chain;
  71                 lock = inet_ehash_lockp(hashinfo, hash);
  72                 spin_lock(lock);
  73                 __sk_nulls_add_node_rcu(sk, list);
  74                 if (tw) {
  75                         WARN_ON(sk->sk_hash != tw->tw_hash);
  76                         twrefcnt = inet_twsk_unhash(tw);
  77                 }
  78                 spin_unlock(lock);
  79         }
  80
  81         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  82         return twrefcnt;
  83 }
  84 EXPORT_SYMBOL(__inet6_hash);
  85
  86 /*
  87  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so
  88  * we need not check it for TCP lookups anymore, thanks Alexey. -DaveM
  89  *
  90  * The sockhash lock must be held as a reader here.
  91  */
  92 struct sock *__inet6_lookup_established(struct net *net,
  93                                         struct inet_hashinfo *hashinfo,
  94                                            const struct in6_addr *saddr,
  95                                            const __be16 sport,
  96                                            const struct in6_addr *daddr,
  97                                            const u16 hnum,
  98                                            const int dif)
  99 {
 100         struct sock *sk;
 101         const struct hlist_nulls_node *node;
 102         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 103         /* Optimize here for direct hit, only listening connections can
 104          * have wildcards anyways.
 105          */
 106         unsigned int hash = inet6_ehashfn(net, daddr, hnum, saddr, sport);
 107         unsigned int slot = hash & hashinfo->ehash_mask;
 108         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 109
 110
 111         rcu_read_lock();
 112 begin:
 113         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 114                 if (sk->sk_hash != hash)
 115                         continue;
 116                 if (!INET6_MATCH(sk, net, saddr, daddr, ports, dif))
 117                         continue;
 118                 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 119                         goto out;
 120
 121                 if (unlikely(!INET6_MATCH(sk, net, saddr, daddr, ports, dif))) {
 122                         sock_gen_put(sk);
 123                         goto begin;
 124                 }
 125                 goto found;
 126         }
 127         if (get_nulls_value(node) != slot)
 128                 goto begin;
 129 out:
 130         sk = NULL;
 131 found:
 132         rcu_read_unlock();
 133         return sk;
 134 }
 135 EXPORT_SYMBOL(__inet6_lookup_established);
 136
 137 static inline int compute_score(struct sock *sk, struct net *net,
 138                                 const unsigned short hnum,
 139                                 const struct in6_addr *daddr,
 140                                 const int dif)
 141 {
 142         int score = -1;
 143
 144         if (net_eq(sock_net(sk), net) && inet_sk(sk)->inet_num == hnum &&
 145             sk->sk_family == PF_INET6) {
 146
 147                 score = 1;
 148                 if (!ipv6_addr_any(&sk->sk_v6_rcv_saddr)) {
 149                         if (!ipv6_addr_equal(&sk->sk_v6_rcv_saddr, daddr))
 150                                 return -1;
 151                         score++;
 152                 }
 153                 if (sk->sk_bound_dev_if) {
 154                         if (sk->sk_bound_dev_if != dif)
 155                                 return -1;
 156                         score++;
 157                 }
 158         }
 159         return score;
 160 }
 161
 162 struct sock *inet6_lookup_listener(struct net *net,
 163                 struct inet_hashinfo *hashinfo, const struct in6_addr *saddr,
 164                 const __be16 sport, const struct in6_addr *daddr,
 165                 const unsigned short hnum, const int dif)
 166 {
 167         struct sock *sk;
 168         const struct hlist_nulls_node *node;
 169         struct sock *result;
 170         int score, hiscore, matches = 0, reuseport = 0;
 171         u32 phash = 0;
 172         unsigned int hash = inet_lhashfn(net, hnum);
 173         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 174
 175         rcu_read_lock();
 176 begin:
 177         result = NULL;
 178         hiscore = 0;
 179         sk_nulls_for_each(sk, node, &ilb->head) {
 180                 score = compute_score(sk, net, hnum, daddr, dif);
 181                 if (score > hiscore) {
 182                         hiscore = score;
 183                         result = sk;
 184                         reuseport = sk->sk_reuseport;
 185                         if (reuseport) {
 186                                 phash = inet6_ehashfn(net, daddr, hnum,
 187                                                       saddr, sport);
 188                                 matches = 1;
 189                         }
 190                 } else if (score == hiscore && reuseport) {
 191                         matches++;
 192                         if (((u64)phash * matches) >> 32 == 0)
 193                                 result = sk;
 194                         phash = next_pseudo_random32(phash);
 195                 }
 196         }
 197         /*
 198          * if the nulls value we got at the end of this lookup is
 199          * not the expected one, we must restart lookup.
 200          * We probably met an item that was moved to another chain.
 201          */
 202         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 203                 goto begin;
 204         if (result) {
 205                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 206                         result = NULL;
 207                 else if (unlikely(compute_score(result, net, hnum, daddr,
 208                                   dif) < hiscore)) {
 209                         sock_put(result);
 210                         goto begin;
 211                 }
 212         }
 213         rcu_read_unlock();
 214         return result;
 215 }
 216
 217 EXPORT_SYMBOL_GPL(inet6_lookup_listener);
 218
 219 struct sock *inet6_lookup(struct net *net, struct inet_hashinfo *hashinfo,
 220                           const struct in6_addr *saddr, const __be16 sport,
 221                           const struct in6_addr *daddr, const __be16 dport,
 222                           const int dif)
 223 {
 224         struct sock *sk;
 225
 226         local_bh_disable();
 227         sk = __inet6_lookup(net, hashinfo, saddr, sport, daddr, ntohs(dport), dif);
 228         local_bh_enable();
 229
 230         return sk;
 231 }
 232
 233 EXPORT_SYMBOL_GPL(inet6_lookup);
 234
 235 static int __inet6_check_established(struct inet_timewait_death_row *death_row,
 236                                      struct sock *sk, const __u16 lport,
 237                                      struct inet_timewait_sock **twp)
 238 {
 239         struct inet_hashinfo *hinfo = death_row->hashinfo;
 240         struct inet_sock *inet = inet_sk(sk);
 241         const struct in6_addr *daddr = &sk->sk_v6_rcv_saddr;
 242         const struct in6_addr *saddr = &sk->sk_v6_daddr;
 243         const int dif = sk->sk_bound_dev_if;
 244         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 245         struct net *net = sock_net(sk);
 246         const unsigned int hash = inet6_ehashfn(net, daddr, lport, saddr,
 247                                                 inet->inet_dport);
 248         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 249         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 250         struct sock *sk2;
 251         const struct hlist_nulls_node *node;
 252         struct inet_timewait_sock *tw = NULL;
 253         int twrefcnt = 0;
 254
 255         spin_lock(lock);
 256
 257         sk_nulls_for_each(sk2, node, &head->chain) {
 258                 if (sk2->sk_hash != hash)
 259                         continue;
 260
 261                 if (likely(INET6_MATCH(sk2, net, saddr, daddr, ports, dif))) {
 262                         if (sk2->sk_state == TCP_TIME_WAIT) {
 263                                 tw = inet_twsk(sk2);
 264                                 if (twsk_unique(sk, sk2, twp))
 265                                         break;
 266                         }
 267                         goto not_unique;
 268                 }
 269         }
 270
 271         /* Must record num and sport now. Otherwise we will see
 272          * in hash table socket with a funny identity.
 273          */
 274         inet->inet_num = lport;
 275         inet->inet_sport = htons(lport);
 276         sk->sk_hash = hash;
 277         WARN_ON(!sk_unhashed(sk));
 278         __sk_nulls_add_node_rcu(sk, &head->chain);
 279         if (tw) {
 280                 twrefcnt = inet_twsk_unhash(tw);
 281                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 282         }
 283         spin_unlock(lock);
 284         if (twrefcnt)
 285                 inet_twsk_put(tw);
 286         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 287
 288         if (twp) {
 289                 *twp = tw;
 290         } else if (tw) {
 291                 /* Silly. Should hash-dance instead... */
 292                 inet_twsk_deschedule(tw, death_row);
 293
 294                 inet_twsk_put(tw);
 295         }
 296         return 0;
 297
 298 not_unique:
 299         spin_unlock(lock);
 300         return -EADDRNOTAVAIL;
 301 }
 302
 303 static inline u32 inet6_sk_port_offset(const struct sock *sk)
 304 {
 305         const struct inet_sock *inet = inet_sk(sk);
 306
 307         return secure_ipv6_port_ephemeral(sk->sk_v6_rcv_saddr.s6_addr32,
 308                                           sk->sk_v6_daddr.s6_addr32,
 309                                           inet->inet_dport);
 310 }
 311
 312 int inet6_hash_connect(struct inet_timewait_death_row *death_row,
 313                        struct sock *sk)
 314 {
 315         return __inet_hash_connect(death_row, sk, inet6_sk_port_offset(sk),
 316                         __inet6_check_established, __inet6_hash);
 317 }
 318
 319 EXPORT_SYMBOL_GPL(inet6_hash_connect);