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         static u32 inet6_ehash_secret __read_mostly;
  33         static u32 ipv6_hash_secret __read_mostly;
  34
  35         u32 lhash, fhash;
  36
  37         net_get_random_once(&inet6_ehash_secret, sizeof(inet6_ehash_secret));
  38         net_get_random_once(&ipv6_hash_secret, sizeof(ipv6_hash_secret));
  39
  40         lhash = (__force u32)laddr->s6_addr32[3];
  41         fhash = __ipv6_addr_jhash(faddr, ipv6_hash_secret);
  42
  43         return __inet6_ehashfn(lhash, lport, fhash, fport,
  44                                inet6_ehash_secret + net_hash_mix(net));
  45 }
  46
  47 static int inet6_sk_ehashfn(const struct sock *sk)
  48 {
  49         const struct inet_sock *inet = inet_sk(sk);
  50         const struct in6_addr *laddr = &sk->sk_v6_rcv_saddr;
  51         const struct in6_addr *faddr = &sk->sk_v6_daddr;
  52         const __u16 lport = inet->inet_num;
  53         const __be16 fport = inet->inet_dport;
  54         struct net *net = sock_net(sk);
  55
  56         return inet6_ehashfn(net, laddr, lport, faddr, fport);
  57 }
  58
  59 int __inet6_hash(struct sock *sk, struct inet_timewait_sock *tw)
  60 {
  61         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  62         int twrefcnt = 0;
  63
  64         WARN_ON(!sk_unhashed(sk));
  65
  66         if (sk->sk_state == TCP_LISTEN) {
  67                 struct inet_listen_hashbucket *ilb;
  68
  69                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
  70                 spin_lock(&ilb->lock);
  71                 __sk_nulls_add_node_rcu(sk, &ilb->head);
  72                 spin_unlock(&ilb->lock);
  73         } else {
  74                 unsigned int hash;
  75                 struct hlist_nulls_head *list;
  76                 spinlock_t *lock;
  77
  78                 sk->sk_hash = hash = inet6_sk_ehashfn(sk);
  79                 list = &inet_ehash_bucket(hashinfo, hash)->chain;
  80                 lock = inet_ehash_lockp(hashinfo, hash);
  81                 spin_lock(lock);
  82                 __sk_nulls_add_node_rcu(sk, list);
  83                 if (tw) {
  84                         WARN_ON(sk->sk_hash != tw->tw_hash);
  85                         twrefcnt = inet_twsk_unhash(tw);
  86                 }
  87                 spin_unlock(lock);
  88         }
  89
  90         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  91         return twrefcnt;
  92 }
  93 EXPORT_SYMBOL(__inet6_hash);
  94
  95 /*
  96  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so
  97  * we need not check it for TCP lookups anymore, thanks Alexey. -DaveM
  98  *
  99  * The sockhash lock must be held as a reader here.
 100  */
 101 struct sock *__inet6_lookup_established(struct net *net,
 102                                         struct inet_hashinfo *hashinfo,
 103                                            const struct in6_addr *saddr,
 104                                            const __be16 sport,
 105                                            const struct in6_addr *daddr,
 106                                            const u16 hnum,
 107                                            const int dif)
 108 {
 109         struct sock *sk;
 110         const struct hlist_nulls_node *node;
 111         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 112         /* Optimize here for direct hit, only listening connections can
 113          * have wildcards anyways.
 114          */
 115         unsigned int hash = inet6_ehashfn(net, daddr, hnum, saddr, sport);
 116         unsigned int slot = hash & hashinfo->ehash_mask;
 117         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 118
 119
 120         rcu_read_lock();
 121 begin:
 122         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 123                 if (sk->sk_hash != hash)
 124                         continue;
 125                 if (!INET6_MATCH(sk, net, saddr, daddr, ports, dif))
 126                         continue;
 127                 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 128                         goto out;
 129
 130                 if (unlikely(!INET6_MATCH(sk, net, saddr, daddr, ports, dif))) {
 131                         sock_gen_put(sk);
 132                         goto begin;
 133                 }
 134                 goto found;
 135         }
 136         if (get_nulls_value(node) != slot)
 137                 goto begin;
 138 out:
 139         sk = NULL;
 140 found:
 141         rcu_read_unlock();
 142         return sk;
 143 }
 144 EXPORT_SYMBOL(__inet6_lookup_established);
 145
 146 static inline int compute_score(struct sock *sk, struct net *net,
 147                                 const unsigned short hnum,
 148                                 const struct in6_addr *daddr,
 149                                 const int dif)
 150 {
 151         int score = -1;
 152
 153         if (net_eq(sock_net(sk), net) && inet_sk(sk)->inet_num == hnum &&
 154             sk->sk_family == PF_INET6) {
 155
 156                 score = 1;
 157                 if (!ipv6_addr_any(&sk->sk_v6_rcv_saddr)) {
 158                         if (!ipv6_addr_equal(&sk->sk_v6_rcv_saddr, daddr))
 159                                 return -1;
 160                         score++;
 161                 }
 162                 if (sk->sk_bound_dev_if) {
 163                         if (sk->sk_bound_dev_if != dif)
 164                                 return -1;
 165                         score++;
 166                 }
 167         }
 168         return score;
 169 }
 170
 171 struct sock *inet6_lookup_listener(struct net *net,
 172                 struct inet_hashinfo *hashinfo, const struct in6_addr *saddr,
 173                 const __be16 sport, const struct in6_addr *daddr,
 174                 const unsigned short hnum, const int dif)
 175 {
 176         struct sock *sk;
 177         const struct hlist_nulls_node *node;
 178         struct sock *result;
 179         int score, hiscore, matches = 0, reuseport = 0;
 180         u32 phash = 0;
 181         unsigned int hash = inet_lhashfn(net, hnum);
 182         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 183
 184         rcu_read_lock();
 185 begin:
 186         result = NULL;
 187         hiscore = 0;
 188         sk_nulls_for_each(sk, node, &ilb->head) {
 189                 score = compute_score(sk, net, hnum, daddr, dif);
 190                 if (score > hiscore) {
 191                         hiscore = score;
 192                         result = sk;
 193                         reuseport = sk->sk_reuseport;
 194                         if (reuseport) {
 195                                 phash = inet6_ehashfn(net, daddr, hnum,
 196                                                       saddr, sport);
 197                                 matches = 1;
 198                         }
 199                 } else if (score == hiscore && reuseport) {
 200                         matches++;
 201                         if (reciprocal_scale(phash, matches) == 0)
 202                                 result = sk;
 203                         phash = next_pseudo_random32(phash);
 204                 }
 205         }
 206         /*
 207          * if the nulls value we got at the end of this lookup is
 208          * not the expected one, we must restart lookup.
 209          * We probably met an item that was moved to another chain.
 210          */
 211         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 212                 goto begin;
 213         if (result) {
 214                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 215                         result = NULL;
 216                 else if (unlikely(compute_score(result, net, hnum, daddr,
 217                                   dif) < hiscore)) {
 218                         sock_put(result);
 219                         goto begin;
 220                 }
 221         }
 222         rcu_read_unlock();
 223         return result;
 224 }
 225 EXPORT_SYMBOL_GPL(inet6_lookup_listener);
 226
 227 struct sock *inet6_lookup(struct net *net, struct inet_hashinfo *hashinfo,
 228                           const struct in6_addr *saddr, const __be16 sport,
 229                           const struct in6_addr *daddr, const __be16 dport,
 230                           const int dif)
 231 {
 232         struct sock *sk;
 233
 234         local_bh_disable();
 235         sk = __inet6_lookup(net, hashinfo, saddr, sport, daddr, ntohs(dport), dif);
 236         local_bh_enable();
 237
 238         return sk;
 239 }
 240 EXPORT_SYMBOL_GPL(inet6_lookup);
 241
 242 static int __inet6_check_established(struct inet_timewait_death_row *death_row,
 243                                      struct sock *sk, const __u16 lport,
 244                                      struct inet_timewait_sock **twp)
 245 {
 246         struct inet_hashinfo *hinfo = death_row->hashinfo;
 247         struct inet_sock *inet = inet_sk(sk);
 248         const struct in6_addr *daddr = &sk->sk_v6_rcv_saddr;
 249         const struct in6_addr *saddr = &sk->sk_v6_daddr;
 250         const int dif = sk->sk_bound_dev_if;
 251         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 252         struct net *net = sock_net(sk);
 253         const unsigned int hash = inet6_ehashfn(net, daddr, lport, saddr,
 254                                                 inet->inet_dport);
 255         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 256         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 257         struct sock *sk2;
 258         const struct hlist_nulls_node *node;
 259         struct inet_timewait_sock *tw = NULL;
 260         int twrefcnt = 0;
 261
 262         spin_lock(lock);
 263
 264         sk_nulls_for_each(sk2, node, &head->chain) {
 265                 if (sk2->sk_hash != hash)
 266                         continue;
 267
 268                 if (likely(INET6_MATCH(sk2, net, saddr, daddr, ports, dif))) {
 269                         if (sk2->sk_state == TCP_TIME_WAIT) {
 270                                 tw = inet_twsk(sk2);
 271                                 if (twsk_unique(sk, sk2, twp))
 272                                         break;
 273                         }
 274                         goto not_unique;
 275                 }
 276         }
 277
 278         /* Must record num and sport now. Otherwise we will see
 279          * in hash table socket with a funny identity.
 280          */
 281         inet->inet_num = lport;
 282         inet->inet_sport = htons(lport);
 283         sk->sk_hash = hash;
 284         WARN_ON(!sk_unhashed(sk));
 285         __sk_nulls_add_node_rcu(sk, &head->chain);
 286         if (tw) {
 287                 twrefcnt = inet_twsk_unhash(tw);
 288                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 289         }
 290         spin_unlock(lock);
 291         if (twrefcnt)
 292                 inet_twsk_put(tw);
 293         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 294
 295         if (twp) {
 296                 *twp = tw;
 297         } else if (tw) {
 298                 /* Silly. Should hash-dance instead... */
 299                 inet_twsk_deschedule(tw, death_row);
 300
 301                 inet_twsk_put(tw);
 302         }
 303         return 0;
 304
 305 not_unique:
 306         spin_unlock(lock);
 307         return -EADDRNOTAVAIL;
 308 }
 309
 310 static inline u32 inet6_sk_port_offset(const struct sock *sk)
 311 {
 312         const struct inet_sock *inet = inet_sk(sk);
 313
 314         return secure_ipv6_port_ephemeral(sk->sk_v6_rcv_saddr.s6_addr32,
 315                                           sk->sk_v6_daddr.s6_addr32,
 316                                           inet->inet_dport);
 317 }
 318
 319 int inet6_hash_connect(struct inet_timewait_death_row *death_row,
 320                        struct sock *sk)
 321 {
 322         return __inet_hash_connect(death_row, sk, inet6_sk_port_offset(sk),
 323                         __inet6_check_established, __inet6_hash);
 324 }
 325 EXPORT_SYMBOL_GPL(inet6_hash_connect);