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