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 u32 inet6_ehashfn(const struct net *net,
  27                   const struct in6_addr *laddr, const u16 lport,
  28                   const struct in6_addr *faddr, const __be16 fport)
  29 {
  30         static u32 inet6_ehash_secret __read_mostly;
  31         static u32 ipv6_hash_secret __read_mostly;
  32
  33         u32 lhash, fhash;
  34
  35         net_get_random_once(&inet6_ehash_secret, sizeof(inet6_ehash_secret));
  36         net_get_random_once(&ipv6_hash_secret, sizeof(ipv6_hash_secret));
  37
  38         lhash = (__force u32)laddr->s6_addr32[3];
  39         fhash = __ipv6_addr_jhash(faddr, ipv6_hash_secret);
  40
  41         return __inet6_ehashfn(lhash, lport, fhash, fport,
  42                                inet6_ehash_secret + net_hash_mix(net));
  43 }
  44
  45 /*
  46  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so
  47  * we need not check it for TCP lookups anymore, thanks Alexey. -DaveM
  48  *
  49  * The sockhash lock must be held as a reader here.
  50  */
  51 struct sock *__inet6_lookup_established(struct net *net,
  52                                         struct inet_hashinfo *hashinfo,
  53                                            const struct in6_addr *saddr,
  54                                            const __be16 sport,
  55                                            const struct in6_addr *daddr,
  56                                            const u16 hnum,
  57                                            const int dif)
  58 {
  59         struct sock *sk;
  60         const struct hlist_nulls_node *node;
  61         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
  62         /* Optimize here for direct hit, only listening connections can
  63          * have wildcards anyways.
  64          */
  65         unsigned int hash = inet6_ehashfn(net, daddr, hnum, saddr, sport);
  66         unsigned int slot = hash & hashinfo->ehash_mask;
  67         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
  68
  69
  70         rcu_read_lock();
  71 begin:
  72         sk_nulls_for_each_rcu(sk, node, &head->chain) {
  73                 if (sk->sk_hash != hash)
  74                         continue;
  75                 if (!INET6_MATCH(sk, net, saddr, daddr, ports, dif))
  76                         continue;
  77                 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
  78                         goto out;
  79
  80                 if (unlikely(!INET6_MATCH(sk, net, saddr, daddr, ports, dif))) {
  81                         sock_gen_put(sk);
  82                         goto begin;
  83                 }
  84                 goto found;
  85         }
  86         if (get_nulls_value(node) != slot)
  87                 goto begin;
  88 out:
  89         sk = NULL;
  90 found:
  91         rcu_read_unlock();
  92         return sk;
  93 }
  94 EXPORT_SYMBOL(__inet6_lookup_established);
  95
  96 static inline int compute_score(struct sock *sk, struct net *net,
  97                                 const unsigned short hnum,
  98                                 const struct in6_addr *daddr,
  99                                 const int dif)
 100 {
 101         int score = -1;
 102
 103         if (net_eq(sock_net(sk), net) && inet_sk(sk)->inet_num == hnum &&
 104             sk->sk_family == PF_INET6) {
 105
 106                 score = 1;
 107                 if (!ipv6_addr_any(&sk->sk_v6_rcv_saddr)) {
 108                         if (!ipv6_addr_equal(&sk->sk_v6_rcv_saddr, daddr))
 109                                 return -1;
 110                         score++;
 111                 }
 112                 if (sk->sk_bound_dev_if) {
 113                         if (sk->sk_bound_dev_if != dif)
 114                                 return -1;
 115                         score++;
 116                 }
 117                 if (sk->sk_incoming_cpu == raw_smp_processor_id())
 118                         score++;
 119         }
 120         return score;
 121 }
 122
 123 struct sock *inet6_lookup_listener(struct net *net,
 124                 struct inet_hashinfo *hashinfo, const struct in6_addr *saddr,
 125                 const __be16 sport, const struct in6_addr *daddr,
 126                 const unsigned short hnum, const int dif)
 127 {
 128         struct sock *sk;
 129         const struct hlist_nulls_node *node;
 130         struct sock *result;
 131         int score, hiscore, matches = 0, reuseport = 0;
 132         u32 phash = 0;
 133         unsigned int hash = inet_lhashfn(net, hnum);
 134         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 135
 136         rcu_read_lock();
 137 begin:
 138         result = NULL;
 139         hiscore = 0;
 140         sk_nulls_for_each(sk, node, &ilb->head) {
 141                 score = compute_score(sk, net, hnum, daddr, dif);
 142                 if (score > hiscore) {
 143                         hiscore = score;
 144                         result = sk;
 145                         reuseport = sk->sk_reuseport;
 146                         if (reuseport) {
 147                                 phash = inet6_ehashfn(net, daddr, hnum,
 148                                                       saddr, sport);
 149                                 matches = 1;
 150                         }
 151                 } else if (score == hiscore && reuseport) {
 152                         matches++;
 153                         if (reciprocal_scale(phash, matches) == 0)
 154                                 result = sk;
 155                         phash = next_pseudo_random32(phash);
 156                 }
 157         }
 158         /*
 159          * if the nulls value we got at the end of this lookup is
 160          * not the expected one, we must restart lookup.
 161          * We probably met an item that was moved to another chain.
 162          */
 163         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 164                 goto begin;
 165         if (result) {
 166                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 167                         result = NULL;
 168                 else if (unlikely(compute_score(result, net, hnum, daddr,
 169                                   dif) < hiscore)) {
 170                         sock_put(result);
 171                         goto begin;
 172                 }
 173         }
 174         rcu_read_unlock();
 175         return result;
 176 }
 177 EXPORT_SYMBOL_GPL(inet6_lookup_listener);
 178
 179 struct sock *inet6_lookup(struct net *net, struct inet_hashinfo *hashinfo,
 180                           const struct in6_addr *saddr, const __be16 sport,
 181                           const struct in6_addr *daddr, const __be16 dport,
 182                           const int dif)
 183 {
 184         struct sock *sk;
 185
 186         local_bh_disable();
 187         sk = __inet6_lookup(net, hashinfo, saddr, sport, daddr, ntohs(dport), dif);
 188         local_bh_enable();
 189
 190         return sk;
 191 }
 192 EXPORT_SYMBOL_GPL(inet6_lookup);
 193
 194 static int __inet6_check_established(struct inet_timewait_death_row *death_row,
 195                                      struct sock *sk, const __u16 lport,
 196                                      struct inet_timewait_sock **twp)
 197 {
 198         struct inet_hashinfo *hinfo = death_row->hashinfo;
 199         struct inet_sock *inet = inet_sk(sk);
 200         const struct in6_addr *daddr = &sk->sk_v6_rcv_saddr;
 201         const struct in6_addr *saddr = &sk->sk_v6_daddr;
 202         const int dif = sk->sk_bound_dev_if;
 203         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 204         struct net *net = sock_net(sk);
 205         const unsigned int hash = inet6_ehashfn(net, daddr, lport, saddr,
 206                                                 inet->inet_dport);
 207         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 208         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 209         struct sock *sk2;
 210         const struct hlist_nulls_node *node;
 211         struct inet_timewait_sock *tw = NULL;
 212
 213         spin_lock(lock);
 214
 215         sk_nulls_for_each(sk2, node, &head->chain) {
 216                 if (sk2->sk_hash != hash)
 217                         continue;
 218
 219                 if (likely(INET6_MATCH(sk2, net, saddr, daddr, ports, dif))) {
 220                         if (sk2->sk_state == TCP_TIME_WAIT) {
 221                                 tw = inet_twsk(sk2);
 222                                 if (twsk_unique(sk, sk2, twp))
 223                                         break;
 224                         }
 225                         goto not_unique;
 226                 }
 227         }
 228
 229         /* Must record num and sport now. Otherwise we will see
 230          * in hash table socket with a funny identity.
 231          */
 232         inet->inet_num = lport;
 233         inet->inet_sport = htons(lport);
 234         sk->sk_hash = hash;
 235         WARN_ON(!sk_unhashed(sk));
 236         __sk_nulls_add_node_rcu(sk, &head->chain);
 237         if (tw) {
 238                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 239                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 240         }
 241         spin_unlock(lock);
 242         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 243
 244         if (twp) {
 245                 *twp = tw;
 246         } else if (tw) {
 247                 /* Silly. Should hash-dance instead... */
 248                 inet_twsk_deschedule_put(tw);
 249         }
 250         return 0;
 251
 252 not_unique:
 253         spin_unlock(lock);
 254         return -EADDRNOTAVAIL;
 255 }
 256
 257 static u32 inet6_sk_port_offset(const struct sock *sk)
 258 {
 259         const struct inet_sock *inet = inet_sk(sk);
 260
 261         return secure_ipv6_port_ephemeral(sk->sk_v6_rcv_saddr.s6_addr32,
 262                                           sk->sk_v6_daddr.s6_addr32,
 263                                           inet->inet_dport);
 264 }
 265
 266 int inet6_hash_connect(struct inet_timewait_death_row *death_row,
 267                        struct sock *sk)
 268 {
 269         u32 port_offset = 0;
 270
 271         if (!inet_sk(sk)->inet_num)
 272                 port_offset = inet6_sk_port_offset(sk);
 273         return __inet_hash_connect(death_row, sk, port_offset,
 274                                    __inet6_check_established);
 275 }
 276 EXPORT_SYMBOL_GPL(inet6_hash_connect);
 277
 278 int inet6_hash(struct sock *sk)
 279 {
 280         if (sk->sk_state != TCP_CLOSE) {
 281                 local_bh_disable();
 282                 __inet_hash(sk, NULL);
 283                 local_bh_enable();
 284         }
 285
 286         return 0;
 287 }
 288 EXPORT_SYMBOL_GPL(inet6_hash);
 289
 290 /* match_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses if IPv6
 291  *                          only, and any IPv4 addresses if not IPv6 only
 292  * match_wildcard == false: addresses must be exactly the same, i.e.
 293  *                          IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
 294  *                          and 0.0.0.0 equals to 0.0.0.0 only
 295  */
 296 int ipv6_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
 297                          bool match_wildcard)
 298 {
 299         const struct in6_addr *sk2_rcv_saddr6 = inet6_rcv_saddr(sk2);
 300         int sk2_ipv6only = inet_v6_ipv6only(sk2);
 301         int addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
 302         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
 303
 304         /* if both are mapped, treat as IPv4 */
 305         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
 306                 if (!sk2_ipv6only) {
 307                         if (sk->sk_rcv_saddr == sk2->sk_rcv_saddr)
 308                                 return 1;
 309                         if (!sk->sk_rcv_saddr || !sk2->sk_rcv_saddr)
 310                                 return match_wildcard;
 311                 }
 312                 return 0;
 313         }
 314
 315         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
 316                 return 1;
 317
 318         if (addr_type2 == IPV6_ADDR_ANY && match_wildcard &&
 319             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
 320                 return 1;
 321
 322         if (addr_type == IPV6_ADDR_ANY && match_wildcard &&
 323             !(ipv6_only_sock(sk) && addr_type2 == IPV6_ADDR_MAPPED))
 324                 return 1;
 325
 326         if (sk2_rcv_saddr6 &&
 327             ipv6_addr_equal(&sk->sk_v6_rcv_saddr, sk2_rcv_saddr6))
 328                 return 1;
 329
 330         return 0;
 331 }
 332 EXPORT_SYMBOL_GPL(ipv6_rcv_saddr_equal);