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/addrconf.h>
  21 #include <net/inet_connection_sock.h>
  22 #include <net/inet_hashtables.h>
  23 #include <net/inet6_hashtables.h>
  24 #include <net/secure_seq.h>
  25 #include <net/ip.h>
  26 #include <net/sock_reuseport.h>
  27
  28 u32 inet6_ehashfn(const struct net *net,
  29                   const struct in6_addr *laddr, const u16 lport,
  30                   const struct in6_addr *faddr, 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 /*
  48  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so
  49  * we need not check it for TCP lookups anymore, thanks Alexey. -DaveM
  50  *
  51  * The sockhash lock must be held as a reader here.
  52  */
  53 struct sock *__inet6_lookup_established(struct net *net,
  54                                         struct inet_hashinfo *hashinfo,
  55                                            const struct in6_addr *saddr,
  56                                            const __be16 sport,
  57                                            const struct in6_addr *daddr,
  58                                            const u16 hnum,
  59                                            const int dif, const int sdif)
  60 {
  61         struct sock *sk;
  62         const struct hlist_nulls_node *node;
  63         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
  64         /* Optimize here for direct hit, only listening connections can
  65          * have wildcards anyways.
  66          */
  67         unsigned int hash = inet6_ehashfn(net, daddr, hnum, saddr, sport);
  68         unsigned int slot = hash & hashinfo->ehash_mask;
  69         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
  70
  71
  72 begin:
  73         sk_nulls_for_each_rcu(sk, node, &head->chain) {
  74                 if (sk->sk_hash != hash)
  75                         continue;
  76                 if (!INET6_MATCH(sk, net, saddr, daddr, ports, dif, sdif))
  77                         continue;
  78                 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
  79                         goto out;
  80
  81                 if (unlikely(!INET6_MATCH(sk, net, saddr, daddr, ports, dif, sdif))) {
  82                         sock_gen_put(sk);
  83                         goto begin;
  84                 }
  85                 goto found;
  86         }
  87         if (get_nulls_value(node) != slot)
  88                 goto begin;
  89 out:
  90         sk = NULL;
  91 found:
  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, const int sdif, bool exact_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 || exact_dif) {
 113                         bool dev_match = (sk->sk_bound_dev_if == dif ||
 114                                           sk->sk_bound_dev_if == sdif);
 115
 116                         if (!dev_match)
 117                                 return -1;
 118                         if (sk->sk_bound_dev_if)
 119                                 score++;
 120                 }
 121                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 122                         score++;
 123         }
 124         return score;
 125 }
 126
 127 /* called with rcu_read_lock() */
 128 static struct sock *inet6_lhash2_lookup(struct net *net,
 129                 struct inet_listen_hashbucket *ilb2,
 130                 struct sk_buff *skb, int doff,
 131                 const struct in6_addr *saddr,
 132                 const __be16 sport, const struct in6_addr *daddr,
 133                 const unsigned short hnum, const int dif, const int sdif)
 134 {
 135         bool exact_dif = inet6_exact_dif_match(net, skb);
 136         struct inet_connection_sock *icsk;
 137         struct sock *sk, *result = NULL;
 138         int score, hiscore = 0;
 139         u32 phash = 0;
 140
 141         inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
 142                 sk = (struct sock *)icsk;
 143                 score = compute_score(sk, net, hnum, daddr, dif, sdif,
 144                                       exact_dif);
 145                 if (score > hiscore) {
 146                         if (sk->sk_reuseport) {
 147                                 phash = inet6_ehashfn(net, daddr, hnum,
 148                                                       saddr, sport);
 149                                 result = reuseport_select_sock(sk, phash,
 150                                                                skb, doff);
 151                                 if (result)
 152                                         return result;
 153                         }
 154                         result = sk;
 155                         hiscore = score;
 156                 }
 157         }
 158
 159         return result;
 160 }
 161
 162 struct sock *inet6_lookup_listener(struct net *net,
 163                 struct inet_hashinfo *hashinfo,
 164                 struct sk_buff *skb, int doff,
 165                 const struct in6_addr *saddr,
 166                 const __be16 sport, const struct in6_addr *daddr,
 167                 const unsigned short hnum, const int dif, const int sdif)
 168 {
 169         unsigned int hash = inet_lhashfn(net, hnum);
 170         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 171         bool exact_dif = inet6_exact_dif_match(net, skb);
 172         struct inet_listen_hashbucket *ilb2;
 173         struct sock *sk, *result = NULL;
 174         struct hlist_nulls_node *node;
 175         int score, hiscore = 0;
 176         unsigned int hash2;
 177         u32 phash = 0;
 178
 179         if (ilb->count <= 10 || !hashinfo->lhash2)
 180                 goto port_lookup;
 181
 182         /* Too many sk in the ilb bucket (which is hashed by port alone).
 183          * Try lhash2 (which is hashed by port and addr) instead.
 184          */
 185
 186         hash2 = ipv6_portaddr_hash(net, daddr, hnum);
 187         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 188         if (ilb2->count > ilb->count)
 189                 goto port_lookup;
 190
 191         result = inet6_lhash2_lookup(net, ilb2, skb, doff,
 192                                      saddr, sport, daddr, hnum,
 193                                      dif, sdif);
 194         if (result)
 195                 goto done;
 196
 197         /* Lookup lhash2 with in6addr_any */
 198
 199         hash2 = ipv6_portaddr_hash(net, &in6addr_any, hnum);
 200         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 201         if (ilb2->count > ilb->count)
 202                 goto port_lookup;
 203
 204         result = inet6_lhash2_lookup(net, ilb2, skb, doff,
 205                                      saddr, sport, daddr, hnum,
 206                                      dif, sdif);
 207         goto done;
 208
 209 port_lookup:
 210         sk_nulls_for_each(sk, node, &ilb->nulls_head) {
 211                 score = compute_score(sk, net, hnum, daddr, dif, sdif, exact_dif);
 212                 if (score > hiscore) {
 213                         if (sk->sk_reuseport) {
 214                                 phash = inet6_ehashfn(net, daddr, hnum,
 215                                                       saddr, sport);
 216                                 result = reuseport_select_sock(sk, phash,
 217                                                                skb, doff);
 218                                 if (result)
 219                                         goto done;
 220                         }
 221                         result = sk;
 222                         hiscore = score;
 223                 }
 224         }
 225 done:
 226         if (unlikely(IS_ERR(result)))
 227                 return NULL;
 228         return result;
 229 }
 230 EXPORT_SYMBOL_GPL(inet6_lookup_listener);
 231
 232 struct sock *inet6_lookup(struct net *net, struct inet_hashinfo *hashinfo,
 233                           struct sk_buff *skb, int doff,
 234                           const struct in6_addr *saddr, const __be16 sport,
 235                           const struct in6_addr *daddr, const __be16 dport,
 236                           const int dif)
 237 {
 238         struct sock *sk;
 239         bool refcounted;
 240
 241         sk = __inet6_lookup(net, hashinfo, skb, doff, saddr, sport, daddr,
 242                             ntohs(dport), dif, 0, &refcounted);
 243         if (sk && !refcounted && !refcount_inc_not_zero(&sk->sk_refcnt))
 244                 sk = NULL;
 245         return sk;
 246 }
 247 EXPORT_SYMBOL_GPL(inet6_lookup);
 248
 249 static int __inet6_check_established(struct inet_timewait_death_row *death_row,
 250                                      struct sock *sk, const __u16 lport,
 251                                      struct inet_timewait_sock **twp)
 252 {
 253         struct inet_hashinfo *hinfo = death_row->hashinfo;
 254         struct inet_sock *inet = inet_sk(sk);
 255         const struct in6_addr *daddr = &sk->sk_v6_rcv_saddr;
 256         const struct in6_addr *saddr = &sk->sk_v6_daddr;
 257         const int dif = sk->sk_bound_dev_if;
 258         struct net *net = sock_net(sk);
 259         const int sdif = l3mdev_master_ifindex_by_index(net, dif);
 260         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 261         const unsigned int hash = inet6_ehashfn(net, daddr, lport, saddr,
 262                                                 inet->inet_dport);
 263         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 264         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 265         struct sock *sk2;
 266         const struct hlist_nulls_node *node;
 267         struct inet_timewait_sock *tw = NULL;
 268
 269         spin_lock(lock);
 270
 271         sk_nulls_for_each(sk2, node, &head->chain) {
 272                 if (sk2->sk_hash != hash)
 273                         continue;
 274
 275                 if (likely(INET6_MATCH(sk2, net, saddr, daddr, ports,
 276                                        dif, sdif))) {
 277                         if (sk2->sk_state == TCP_TIME_WAIT) {
 278                                 tw = inet_twsk(sk2);
 279                                 if (twsk_unique(sk, sk2, twp))
 280                                         break;
 281                         }
 282                         goto not_unique;
 283                 }
 284         }
 285
 286         /* Must record num and sport now. Otherwise we will see
 287          * in hash table socket with a funny identity.
 288          */
 289         inet->inet_num = lport;
 290         inet->inet_sport = htons(lport);
 291         sk->sk_hash = hash;
 292         WARN_ON(!sk_unhashed(sk));
 293         __sk_nulls_add_node_rcu(sk, &head->chain);
 294         if (tw) {
 295                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 296                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 297         }
 298         spin_unlock(lock);
 299         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 300
 301         if (twp) {
 302                 *twp = tw;
 303         } else if (tw) {
 304                 /* Silly. Should hash-dance instead... */
 305                 inet_twsk_deschedule_put(tw);
 306         }
 307         return 0;
 308
 309 not_unique:
 310         spin_unlock(lock);
 311         return -EADDRNOTAVAIL;
 312 }
 313
 314 static u64 inet6_sk_port_offset(const struct sock *sk)
 315 {
 316         const struct inet_sock *inet = inet_sk(sk);
 317
 318         return secure_ipv6_port_ephemeral(sk->sk_v6_rcv_saddr.s6_addr32,
 319                                           sk->sk_v6_daddr.s6_addr32,
 320                                           inet->inet_dport);
 321 }
 322
 323 int inet6_hash_connect(struct inet_timewait_death_row *death_row,
 324                        struct sock *sk)
 325 {
 326         u64 port_offset = 0;
 327
 328         if (!inet_sk(sk)->inet_num)
 329                 port_offset = inet6_sk_port_offset(sk);
 330         return __inet_hash_connect(death_row, sk, port_offset,
 331                                    __inet6_check_established);
 332 }
 333 EXPORT_SYMBOL_GPL(inet6_hash_connect);
 334
 335 int inet6_hash(struct sock *sk)
 336 {
 337         int err = 0;
 338
 339         if (sk->sk_state != TCP_CLOSE) {
 340                 local_bh_disable();
 341                 err = __inet_hash(sk, NULL);
 342                 local_bh_enable();
 343         }
 344
 345         return err;
 346 }
 347 EXPORT_SYMBOL_GPL(inet6_hash);