net/ipv4/inet_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 INET transport hashtables
   7  *
   8  * Authors:     Lotsa people, from code originally in tcp
   9  *
  10  *      This program is free software; you can redistribute it and/or
  11  *      modify it under the terms of the GNU General Public License
  12  *      as published by the Free Software Foundation; either version
  13  *      2 of the License, or (at your option) any later version.
  14  */
  15
  16 #include <linux/module.h>
  17 #include <linux/random.h>
  18 #include <linux/sched.h>
  19 #include <linux/slab.h>
  20 #include <linux/wait.h>
  21
  22 #include <net/inet_connection_sock.h>
  23 #include <net/inet_hashtables.h>
  24 #include <net/secure_seq.h>
  25 #include <net/ip.h>
  26
  27 /*
  28  * Allocate and initialize a new local port bind bucket.
  29  * The bindhash mutex for snum's hash chain must be held here.
  30  */
  31 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  32                                                  struct net *net,
  33                                                  struct inet_bind_hashbucket *head,
  34                                                  const unsigned short snum)
  35 {
  36         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  37
  38         if (tb != NULL) {
  39                 write_pnet(&tb->ib_net, hold_net(net));
  40                 tb->port      = snum;
  41                 tb->fastreuse = 0;
  42                 tb->fastreuseport = 0;
  43                 tb->num_owners = 0;
  44                 INIT_HLIST_HEAD(&tb->owners);
  45                 hlist_add_head(&tb->node, &head->chain);
  46         }
  47         return tb;
  48 }
  49
  50 /*
  51  * Caller must hold hashbucket lock for this tb with local BH disabled
  52  */
  53 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  54 {
  55         if (hlist_empty(&tb->owners)) {
  56                 __hlist_del(&tb->node);
  57                 release_net(ib_net(tb));
  58                 kmem_cache_free(cachep, tb);
  59         }
  60 }
  61
  62 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  63                     const unsigned short snum)
  64 {
  65         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  66
  67         atomic_inc(&hashinfo->bsockets);
  68
  69         inet_sk(sk)->inet_num = snum;
  70         sk_add_bind_node(sk, &tb->owners);
  71         tb->num_owners++;
  72         inet_csk(sk)->icsk_bind_hash = tb;
  73 }
  74
  75 /*
  76  * Get rid of any references to a local port held by the given sock.
  77  */
  78 static void __inet_put_port(struct sock *sk)
  79 {
  80         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  81         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
  82                         hashinfo->bhash_size);
  83         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
  84         struct inet_bind_bucket *tb;
  85
  86         atomic_dec(&hashinfo->bsockets);
  87
  88         spin_lock(&head->lock);
  89         tb = inet_csk(sk)->icsk_bind_hash;
  90         __sk_del_bind_node(sk);
  91         tb->num_owners--;
  92         inet_csk(sk)->icsk_bind_hash = NULL;
  93         inet_sk(sk)->inet_num = 0;
  94         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  95         spin_unlock(&head->lock);
  96 }
  97
  98 void inet_put_port(struct sock *sk)
  99 {
 100         local_bh_disable();
 101         __inet_put_port(sk);
 102         local_bh_enable();
 103 }
 104 EXPORT_SYMBOL(inet_put_port);
 105
 106 int __inet_inherit_port(struct sock *sk, struct sock *child)
 107 {
 108         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 109         unsigned short port = inet_sk(child)->inet_num;
 110         const int bhash = inet_bhashfn(sock_net(sk), port,
 111                         table->bhash_size);
 112         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 113         struct inet_bind_bucket *tb;
 114
 115         spin_lock(&head->lock);
 116         tb = inet_csk(sk)->icsk_bind_hash;
 117         if (tb->port != port) {
 118                 /* NOTE: using tproxy and redirecting skbs to a proxy
 119                  * on a different listener port breaks the assumption
 120                  * that the listener socket's icsk_bind_hash is the same
 121                  * as that of the child socket. We have to look up or
 122                  * create a new bind bucket for the child here. */
 123                 inet_bind_bucket_for_each(tb, &head->chain) {
 124                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 125                             tb->port == port)
 126                                 break;
 127                 }
 128                 if (!tb) {
 129                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 130                                                      sock_net(sk), head, port);
 131                         if (!tb) {
 132                                 spin_unlock(&head->lock);
 133                                 return -ENOMEM;
 134                         }
 135                 }
 136         }
 137         inet_bind_hash(child, tb, port);
 138         spin_unlock(&head->lock);
 139
 140         return 0;
 141 }
 142 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 143
 144 static inline int compute_score(struct sock *sk, struct net *net,
 145                                 const unsigned short hnum, const __be32 daddr,
 146                                 const int dif)
 147 {
 148         int score = -1;
 149         struct inet_sock *inet = inet_sk(sk);
 150
 151         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 152                         !ipv6_only_sock(sk)) {
 153                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 154                 score = sk->sk_family == PF_INET ? 2 : 1;
 155                 if (rcv_saddr) {
 156                         if (rcv_saddr != daddr)
 157                                 return -1;
 158                         score += 4;
 159                 }
 160                 if (sk->sk_bound_dev_if) {
 161                         if (sk->sk_bound_dev_if != dif)
 162                                 return -1;
 163                         score += 4;
 164                 }
 165         }
 166         return score;
 167 }
 168
 169 /*
 170  * Don't inline this cruft. Here are some nice properties to exploit here. The
 171  * BSD API does not allow a listening sock to specify the remote port nor the
 172  * remote address for the connection. So always assume those are both
 173  * wildcarded during the search since they can never be otherwise.
 174  */
 175
 176
 177 struct sock *__inet_lookup_listener(struct net *net,
 178                                     struct inet_hashinfo *hashinfo,
 179                                     const __be32 saddr, __be16 sport,
 180                                     const __be32 daddr, const unsigned short hnum,
 181                                     const int dif)
 182 {
 183         struct sock *sk, *result;
 184         struct hlist_nulls_node *node;
 185         unsigned int hash = inet_lhashfn(net, hnum);
 186         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 187         int score, hiscore, matches = 0, reuseport = 0;
 188         u32 phash = 0;
 189
 190         rcu_read_lock();
 191 begin:
 192         result = NULL;
 193         hiscore = 0;
 194         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 195                 score = compute_score(sk, net, hnum, daddr, dif);
 196                 if (score > hiscore) {
 197                         result = sk;
 198                         hiscore = score;
 199                         reuseport = sk->sk_reuseport;
 200                         if (reuseport) {
 201                                 phash = inet_ehashfn(net, daddr, hnum,
 202                                                      saddr, sport);
 203                                 matches = 1;
 204                         }
 205                 } else if (score == hiscore && reuseport) {
 206                         matches++;
 207                         if (((u64)phash * matches) >> 32 == 0)
 208                                 result = sk;
 209                         phash = next_pseudo_random32(phash);
 210                 }
 211         }
 212         /*
 213          * if the nulls value we got at the end of this lookup is
 214          * not the expected one, we must restart lookup.
 215          * We probably met an item that was moved to another chain.
 216          */
 217         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 218                 goto begin;
 219         if (result) {
 220                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 221                         result = NULL;
 222                 else if (unlikely(compute_score(result, net, hnum, daddr,
 223                                   dif) < hiscore)) {
 224                         sock_put(result);
 225                         goto begin;
 226                 }
 227         }
 228         rcu_read_unlock();
 229         return result;
 230 }
 231 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 232
 233 /* All sockets share common refcount, but have different destructors */
 234 void sock_gen_put(struct sock *sk)
 235 {
 236         if (!atomic_dec_and_test(&sk->sk_refcnt))
 237                 return;
 238
 239         if (sk->sk_state == TCP_TIME_WAIT)
 240                 inet_twsk_free(inet_twsk(sk));
 241         else
 242                 sk_free(sk);
 243 }
 244 EXPORT_SYMBOL_GPL(sock_gen_put);
 245
 246 struct sock *__inet_lookup_established(struct net *net,
 247                                   struct inet_hashinfo *hashinfo,
 248                                   const __be32 saddr, const __be16 sport,
 249                                   const __be32 daddr, const u16 hnum,
 250                                   const int dif)
 251 {
 252         INET_ADDR_COOKIE(acookie, saddr, daddr)
 253         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 254         struct sock *sk;
 255         const struct hlist_nulls_node *node;
 256         /* Optimize here for direct hit, only listening connections can
 257          * have wildcards anyways.
 258          */
 259         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 260         unsigned int slot = hash & hashinfo->ehash_mask;
 261         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 262
 263         rcu_read_lock();
 264 begin:
 265         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 266                 if (sk->sk_hash != hash)
 267                         continue;
 268                 if (likely(INET_MATCH(sk, net, acookie,
 269                                       saddr, daddr, ports, dif))) {
 270                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 271                                 goto out;
 272                         if (unlikely(!INET_MATCH(sk, net, acookie,
 273                                                  saddr, daddr, ports, dif))) {
 274                                 sock_gen_put(sk);
 275                                 goto begin;
 276                         }
 277                         goto found;
 278                 }
 279         }
 280         /*
 281          * if the nulls value we got at the end of this lookup is
 282          * not the expected one, we must restart lookup.
 283          * We probably met an item that was moved to another chain.
 284          */
 285         if (get_nulls_value(node) != slot)
 286                 goto begin;
 287 out:
 288         sk = NULL;
 289 found:
 290         rcu_read_unlock();
 291         return sk;
 292 }
 293 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 294
 295 /* called with local bh disabled */
 296 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 297                                     struct sock *sk, __u16 lport,
 298                                     struct inet_timewait_sock **twp)
 299 {
 300         struct inet_hashinfo *hinfo = death_row->hashinfo;
 301         struct inet_sock *inet = inet_sk(sk);
 302         __be32 daddr = inet->inet_rcv_saddr;
 303         __be32 saddr = inet->inet_daddr;
 304         int dif = sk->sk_bound_dev_if;
 305         INET_ADDR_COOKIE(acookie, saddr, daddr)
 306         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 307         struct net *net = sock_net(sk);
 308         unsigned int hash = inet_ehashfn(net, daddr, lport,
 309                                          saddr, inet->inet_dport);
 310         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 311         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 312         struct sock *sk2;
 313         const struct hlist_nulls_node *node;
 314         struct inet_timewait_sock *tw = NULL;
 315         int twrefcnt = 0;
 316
 317         spin_lock(lock);
 318
 319         sk_nulls_for_each(sk2, node, &head->chain) {
 320                 if (sk2->sk_hash != hash)
 321                         continue;
 322
 323                 if (likely(INET_MATCH(sk2, net, acookie,
 324                                          saddr, daddr, ports, dif))) {
 325                         if (sk2->sk_state == TCP_TIME_WAIT) {
 326                                 tw = inet_twsk(sk2);
 327                                 if (twsk_unique(sk, sk2, twp))
 328                                         break;
 329                         }
 330                         goto not_unique;
 331                 }
 332         }
 333
 334         /* Must record num and sport now. Otherwise we will see
 335          * in hash table socket with a funny identity.
 336          */
 337         inet->inet_num = lport;
 338         inet->inet_sport = htons(lport);
 339         sk->sk_hash = hash;
 340         WARN_ON(!sk_unhashed(sk));
 341         __sk_nulls_add_node_rcu(sk, &head->chain);
 342         if (tw) {
 343                 twrefcnt = inet_twsk_unhash(tw);
 344                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 345         }
 346         spin_unlock(lock);
 347         if (twrefcnt)
 348                 inet_twsk_put(tw);
 349         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 350
 351         if (twp) {
 352                 *twp = tw;
 353         } else if (tw) {
 354                 /* Silly. Should hash-dance instead... */
 355                 inet_twsk_deschedule(tw, death_row);
 356
 357                 inet_twsk_put(tw);
 358         }
 359         return 0;
 360
 361 not_unique:
 362         spin_unlock(lock);
 363         return -EADDRNOTAVAIL;
 364 }
 365
 366 static inline u32 inet_sk_port_offset(const struct sock *sk)
 367 {
 368         const struct inet_sock *inet = inet_sk(sk);
 369         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 370                                           inet->inet_daddr,
 371                                           inet->inet_dport);
 372 }
 373
 374 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
 375 {
 376         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 377         struct hlist_nulls_head *list;
 378         spinlock_t *lock;
 379         struct inet_ehash_bucket *head;
 380         int twrefcnt = 0;
 381
 382         WARN_ON(!sk_unhashed(sk));
 383
 384         sk->sk_hash = inet_sk_ehashfn(sk);
 385         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 386         list = &head->chain;
 387         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 388
 389         spin_lock(lock);
 390         __sk_nulls_add_node_rcu(sk, list);
 391         if (tw) {
 392                 WARN_ON(sk->sk_hash != tw->tw_hash);
 393                 twrefcnt = inet_twsk_unhash(tw);
 394         }
 395         spin_unlock(lock);
 396         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 397         return twrefcnt;
 398 }
 399 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 400
 401 static void __inet_hash(struct sock *sk)
 402 {
 403         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 404         struct inet_listen_hashbucket *ilb;
 405
 406         if (sk->sk_state != TCP_LISTEN) {
 407                 __inet_hash_nolisten(sk, NULL);
 408                 return;
 409         }
 410
 411         WARN_ON(!sk_unhashed(sk));
 412         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 413
 414         spin_lock(&ilb->lock);
 415         __sk_nulls_add_node_rcu(sk, &ilb->head);
 416         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 417         spin_unlock(&ilb->lock);
 418 }
 419
 420 void inet_hash(struct sock *sk)
 421 {
 422         if (sk->sk_state != TCP_CLOSE) {
 423                 local_bh_disable();
 424                 __inet_hash(sk);
 425                 local_bh_enable();
 426         }
 427 }
 428 EXPORT_SYMBOL_GPL(inet_hash);
 429
 430 void inet_unhash(struct sock *sk)
 431 {
 432         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 433         spinlock_t *lock;
 434         int done;
 435
 436         if (sk_unhashed(sk))
 437                 return;
 438
 439         if (sk->sk_state == TCP_LISTEN)
 440                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 441         else
 442                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 443
 444         spin_lock_bh(lock);
 445         done = __sk_nulls_del_node_init_rcu(sk);
 446         if (done)
 447                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 448         spin_unlock_bh(lock);
 449 }
 450 EXPORT_SYMBOL_GPL(inet_unhash);
 451
 452 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 453                 struct sock *sk, u32 port_offset,
 454                 int (*check_established)(struct inet_timewait_death_row *,
 455                         struct sock *, __u16, struct inet_timewait_sock **),
 456                 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
 457 {
 458         struct inet_hashinfo *hinfo = death_row->hashinfo;
 459         const unsigned short snum = inet_sk(sk)->inet_num;
 460         struct inet_bind_hashbucket *head;
 461         struct inet_bind_bucket *tb;
 462         int ret;
 463         struct net *net = sock_net(sk);
 464         int twrefcnt = 1;
 465
 466         if (!snum) {
 467                 int i, remaining, low, high, port;
 468                 static u32 hint;
 469                 u32 offset = hint + port_offset;
 470                 struct inet_timewait_sock *tw = NULL;
 471
 472                 inet_get_local_port_range(net, &low, &high);
 473                 remaining = (high - low) + 1;
 474
 475                 local_bh_disable();
 476                 for (i = 1; i <= remaining; i++) {
 477                         port = low + (i + offset) % remaining;
 478                         if (inet_is_reserved_local_port(port))
 479                                 continue;
 480                         head = &hinfo->bhash[inet_bhashfn(net, port,
 481                                         hinfo->bhash_size)];
 482                         spin_lock(&head->lock);
 483
 484                         /* Does not bother with rcv_saddr checks,
 485                          * because the established check is already
 486                          * unique enough.
 487                          */
 488                         inet_bind_bucket_for_each(tb, &head->chain) {
 489                                 if (net_eq(ib_net(tb), net) &&
 490                                     tb->port == port) {
 491                                         if (tb->fastreuse >= 0 ||
 492                                             tb->fastreuseport >= 0)
 493                                                 goto next_port;
 494                                         WARN_ON(hlist_empty(&tb->owners));
 495                                         if (!check_established(death_row, sk,
 496                                                                 port, &tw))
 497                                                 goto ok;
 498                                         goto next_port;
 499                                 }
 500                         }
 501
 502                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 503                                         net, head, port);
 504                         if (!tb) {
 505                                 spin_unlock(&head->lock);
 506                                 break;
 507                         }
 508                         tb->fastreuse = -1;
 509                         tb->fastreuseport = -1;
 510                         goto ok;
 511
 512                 next_port:
 513                         spin_unlock(&head->lock);
 514                 }
 515                 local_bh_enable();
 516
 517                 return -EADDRNOTAVAIL;
 518
 519 ok:
 520                 hint += i;
 521
 522                 /* Head lock still held and bh's disabled */
 523                 inet_bind_hash(sk, tb, port);
 524                 if (sk_unhashed(sk)) {
 525                         inet_sk(sk)->inet_sport = htons(port);
 526                         twrefcnt += hash(sk, tw);
 527                 }
 528                 if (tw)
 529                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
 530                 spin_unlock(&head->lock);
 531
 532                 if (tw) {
 533                         inet_twsk_deschedule(tw, death_row);
 534                         while (twrefcnt) {
 535                                 twrefcnt--;
 536                                 inet_twsk_put(tw);
 537                         }
 538                 }
 539
 540                 ret = 0;
 541                 goto out;
 542         }
 543
 544         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 545         tb  = inet_csk(sk)->icsk_bind_hash;
 546         spin_lock_bh(&head->lock);
 547         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 548                 hash(sk, NULL);
 549                 spin_unlock_bh(&head->lock);
 550                 return 0;
 551         } else {
 552                 spin_unlock(&head->lock);
 553                 /* No definite answer... Walk to established hash table */
 554                 ret = check_established(death_row, sk, snum, NULL);
 555 out:
 556                 local_bh_enable();
 557                 return ret;
 558         }
 559 }
 560
 561 /*
 562  * Bind a port for a connect operation and hash it.
 563  */
 564 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 565                       struct sock *sk)
 566 {
 567         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 568                         __inet_check_established, __inet_hash_nolisten);
 569 }
 570 EXPORT_SYMBOL_GPL(inet_hash_connect);
 571
 572 void inet_hashinfo_init(struct inet_hashinfo *h)
 573 {
 574         int i;
 575
 576         atomic_set(&h->bsockets, 0);
 577         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 578                 spin_lock_init(&h->listening_hash[i].lock);
 579                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 580                                       i + LISTENING_NULLS_BASE);
 581                 }
 582 }
 583 EXPORT_SYMBOL_GPL(inet_hashinfo_init);