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                 struct hlist_node *node;
 124                 inet_bind_bucket_for_each(tb, node, &head->chain) {
 125                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 126                             tb->port == port)
 127                                 break;
 128                 }
 129                 if (!node) {
 130                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 131                                                      sock_net(sk), head, port);
 132                         if (!tb) {
 133                                 spin_unlock(&head->lock);
 134                                 return -ENOMEM;
 135                         }
 136                 }
 137         }
 138         inet_bind_hash(child, tb, port);
 139         spin_unlock(&head->lock);
 140
 141         return 0;
 142 }
 143 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 144
 145 static inline int compute_score(struct sock *sk, struct net *net,
 146                                 const unsigned short hnum, const __be32 daddr,
 147                                 const int dif)
 148 {
 149         int score = -1;
 150         struct inet_sock *inet = inet_sk(sk);
 151
 152         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 153                         !ipv6_only_sock(sk)) {
 154                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 155                 score = sk->sk_family == PF_INET ? 2 : 1;
 156                 if (rcv_saddr) {
 157                         if (rcv_saddr != daddr)
 158                                 return -1;
 159                         score += 4;
 160                 }
 161                 if (sk->sk_bound_dev_if) {
 162                         if (sk->sk_bound_dev_if != dif)
 163                                 return -1;
 164                         score += 4;
 165                 }
 166         }
 167         return score;
 168 }
 169
 170 /*
 171  * Don't inline this cruft. Here are some nice properties to exploit here. The
 172  * BSD API does not allow a listening sock to specify the remote port nor the
 173  * remote address for the connection. So always assume those are both
 174  * wildcarded during the search since they can never be otherwise.
 175  */
 176
 177
 178 struct sock *__inet_lookup_listener(struct net *net,
 179                                     struct inet_hashinfo *hashinfo,
 180                                     const __be32 saddr, __be16 sport,
 181                                     const __be32 daddr, const unsigned short hnum,
 182                                     const int dif)
 183 {
 184         struct sock *sk, *result;
 185         struct hlist_nulls_node *node;
 186         unsigned int hash = inet_lhashfn(net, hnum);
 187         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 188         int score, hiscore, matches = 0, reuseport = 0;
 189         u32 phash = 0;
 190
 191         rcu_read_lock();
 192 begin:
 193         result = NULL;
 194         hiscore = 0;
 195         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 196                 score = compute_score(sk, net, hnum, daddr, dif);
 197                 if (score > hiscore) {
 198                         result = sk;
 199                         hiscore = score;
 200                         reuseport = sk->sk_reuseport;
 201                         if (reuseport) {
 202                                 phash = inet_ehashfn(net, daddr, hnum,
 203                                                      saddr, sport);
 204                                 matches = 1;
 205                         }
 206                 } else if (score == hiscore && reuseport) {
 207                         matches++;
 208                         if (((u64)phash * matches) >> 32 == 0)
 209                                 result = sk;
 210                         phash = next_pseudo_random32(phash);
 211                 }
 212         }
 213         /*
 214          * if the nulls value we got at the end of this lookup is
 215          * not the expected one, we must restart lookup.
 216          * We probably met an item that was moved to another chain.
 217          */
 218         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 219                 goto begin;
 220         if (result) {
 221                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 222                         result = NULL;
 223                 else if (unlikely(compute_score(result, net, hnum, daddr,
 224                                   dif) < hiscore)) {
 225                         sock_put(result);
 226                         goto begin;
 227                 }
 228         }
 229         rcu_read_unlock();
 230         return result;
 231 }
 232 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 233
 234 struct sock *__inet_lookup_established(struct net *net,
 235                                   struct inet_hashinfo *hashinfo,
 236                                   const __be32 saddr, const __be16 sport,
 237                                   const __be32 daddr, const u16 hnum,
 238                                   const int dif)
 239 {
 240         INET_ADDR_COOKIE(acookie, saddr, daddr)
 241         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 242         struct sock *sk;
 243         const struct hlist_nulls_node *node;
 244         /* Optimize here for direct hit, only listening connections can
 245          * have wildcards anyways.
 246          */
 247         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 248         unsigned int slot = hash & hashinfo->ehash_mask;
 249         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 250
 251         rcu_read_lock();
 252 begin:
 253         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 254                 if (sk->sk_hash != hash)
 255                         continue;
 256                 if (likely(INET_MATCH(sk, net, acookie,
 257                                       saddr, daddr, ports, dif))) {
 258                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 259                                 goto begintw;
 260                         if (unlikely(!INET_MATCH(sk, net, acookie,
 261                                                  saddr, daddr, ports, dif))) {
 262                                 sock_put(sk);
 263                                 goto begin;
 264                         }
 265                         goto out;
 266                 }
 267         }
 268         /*
 269          * if the nulls value we got at the end of this lookup is
 270          * not the expected one, we must restart lookup.
 271          * We probably met an item that was moved to another chain.
 272          */
 273         if (get_nulls_value(node) != slot)
 274                 goto begin;
 275
 276 begintw:
 277         /* Must check for a TIME_WAIT'er before going to listener hash. */
 278         sk_nulls_for_each_rcu(sk, node, &head->twchain) {
 279                 if (sk->sk_hash != hash)
 280                         continue;
 281                 if (likely(INET_TW_MATCH(sk, net, acookie,
 282                                          saddr, daddr, ports,
 283                                          dif))) {
 284                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) {
 285                                 sk = NULL;
 286                                 goto out;
 287                         }
 288                         if (unlikely(!INET_TW_MATCH(sk, net, acookie,
 289                                                     saddr, daddr, ports,
 290                                                     dif))) {
 291                                 sock_put(sk);
 292                                 goto begintw;
 293                         }
 294                         goto out;
 295                 }
 296         }
 297         /*
 298          * if the nulls value we got at the end of this lookup is
 299          * not the expected one, we must restart lookup.
 300          * We probably met an item that was moved to another chain.
 301          */
 302         if (get_nulls_value(node) != slot)
 303                 goto begintw;
 304         sk = NULL;
 305 out:
 306         rcu_read_unlock();
 307         return sk;
 308 }
 309 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 310
 311 /* called with local bh disabled */
 312 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 313                                     struct sock *sk, __u16 lport,
 314                                     struct inet_timewait_sock **twp)
 315 {
 316         struct inet_hashinfo *hinfo = death_row->hashinfo;
 317         struct inet_sock *inet = inet_sk(sk);
 318         __be32 daddr = inet->inet_rcv_saddr;
 319         __be32 saddr = inet->inet_daddr;
 320         int dif = sk->sk_bound_dev_if;
 321         INET_ADDR_COOKIE(acookie, saddr, daddr)
 322         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 323         struct net *net = sock_net(sk);
 324         unsigned int hash = inet_ehashfn(net, daddr, lport,
 325                                          saddr, inet->inet_dport);
 326         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 327         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 328         struct sock *sk2;
 329         const struct hlist_nulls_node *node;
 330         struct inet_timewait_sock *tw;
 331         int twrefcnt = 0;
 332
 333         spin_lock(lock);
 334
 335         /* Check TIME-WAIT sockets first. */
 336         sk_nulls_for_each(sk2, node, &head->twchain) {
 337                 if (sk2->sk_hash != hash)
 338                         continue;
 339
 340                 if (likely(INET_TW_MATCH(sk2, net, acookie,
 341                                          saddr, daddr, ports, dif))) {
 342                         tw = inet_twsk(sk2);
 343                         if (twsk_unique(sk, sk2, twp))
 344                                 goto unique;
 345                         else
 346                                 goto not_unique;
 347                 }
 348         }
 349         tw = NULL;
 350
 351         /* And established part... */
 352         sk_nulls_for_each(sk2, node, &head->chain) {
 353                 if (sk2->sk_hash != hash)
 354                         continue;
 355                 if (likely(INET_MATCH(sk2, net, acookie,
 356                                       saddr, daddr, ports, dif)))
 357                         goto not_unique;
 358         }
 359
 360 unique:
 361         /* Must record num and sport now. Otherwise we will see
 362          * in hash table socket with a funny identity. */
 363         inet->inet_num = lport;
 364         inet->inet_sport = htons(lport);
 365         sk->sk_hash = hash;
 366         WARN_ON(!sk_unhashed(sk));
 367         __sk_nulls_add_node_rcu(sk, &head->chain);
 368         if (tw) {
 369                 twrefcnt = inet_twsk_unhash(tw);
 370                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 371         }
 372         spin_unlock(lock);
 373         if (twrefcnt)
 374                 inet_twsk_put(tw);
 375         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 376
 377         if (twp) {
 378                 *twp = tw;
 379         } else if (tw) {
 380                 /* Silly. Should hash-dance instead... */
 381                 inet_twsk_deschedule(tw, death_row);
 382
 383                 inet_twsk_put(tw);
 384         }
 385         return 0;
 386
 387 not_unique:
 388         spin_unlock(lock);
 389         return -EADDRNOTAVAIL;
 390 }
 391
 392 static inline u32 inet_sk_port_offset(const struct sock *sk)
 393 {
 394         const struct inet_sock *inet = inet_sk(sk);
 395         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 396                                           inet->inet_daddr,
 397                                           inet->inet_dport);
 398 }
 399
 400 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
 401 {
 402         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 403         struct hlist_nulls_head *list;
 404         spinlock_t *lock;
 405         struct inet_ehash_bucket *head;
 406         int twrefcnt = 0;
 407
 408         WARN_ON(!sk_unhashed(sk));
 409
 410         sk->sk_hash = inet_sk_ehashfn(sk);
 411         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 412         list = &head->chain;
 413         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 414
 415         spin_lock(lock);
 416         __sk_nulls_add_node_rcu(sk, list);
 417         if (tw) {
 418                 WARN_ON(sk->sk_hash != tw->tw_hash);
 419                 twrefcnt = inet_twsk_unhash(tw);
 420         }
 421         spin_unlock(lock);
 422         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 423         return twrefcnt;
 424 }
 425 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 426
 427 static void __inet_hash(struct sock *sk)
 428 {
 429         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 430         struct inet_listen_hashbucket *ilb;
 431
 432         if (sk->sk_state != TCP_LISTEN) {
 433                 __inet_hash_nolisten(sk, NULL);
 434                 return;
 435         }
 436
 437         WARN_ON(!sk_unhashed(sk));
 438         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 439
 440         spin_lock(&ilb->lock);
 441         __sk_nulls_add_node_rcu(sk, &ilb->head);
 442         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 443         spin_unlock(&ilb->lock);
 444 }
 445
 446 void inet_hash(struct sock *sk)
 447 {
 448         if (sk->sk_state != TCP_CLOSE) {
 449                 local_bh_disable();
 450                 __inet_hash(sk);
 451                 local_bh_enable();
 452         }
 453 }
 454 EXPORT_SYMBOL_GPL(inet_hash);
 455
 456 void inet_unhash(struct sock *sk)
 457 {
 458         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 459         spinlock_t *lock;
 460         int done;
 461
 462         if (sk_unhashed(sk))
 463                 return;
 464
 465         if (sk->sk_state == TCP_LISTEN)
 466                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 467         else
 468                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 469
 470         spin_lock_bh(lock);
 471         done =__sk_nulls_del_node_init_rcu(sk);
 472         if (done)
 473                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 474         spin_unlock_bh(lock);
 475 }
 476 EXPORT_SYMBOL_GPL(inet_unhash);
 477
 478 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 479                 struct sock *sk, u32 port_offset,
 480                 int (*check_established)(struct inet_timewait_death_row *,
 481                         struct sock *, __u16, struct inet_timewait_sock **),
 482                 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
 483 {
 484         struct inet_hashinfo *hinfo = death_row->hashinfo;
 485         const unsigned short snum = inet_sk(sk)->inet_num;
 486         struct inet_bind_hashbucket *head;
 487         struct inet_bind_bucket *tb;
 488         int ret;
 489         struct net *net = sock_net(sk);
 490         int twrefcnt = 1;
 491
 492         if (!snum) {
 493                 int i, remaining, low, high, port;
 494                 static u32 hint;
 495                 u32 offset = hint + port_offset;
 496                 struct hlist_node *node;
 497                 struct inet_timewait_sock *tw = NULL;
 498
 499                 inet_get_local_port_range(&low, &high);
 500                 remaining = (high - low) + 1;
 501
 502                 local_bh_disable();
 503                 for (i = 1; i <= remaining; i++) {
 504                         port = low + (i + offset) % remaining;
 505                         if (inet_is_reserved_local_port(port))
 506                                 continue;
 507                         head = &hinfo->bhash[inet_bhashfn(net, port,
 508                                         hinfo->bhash_size)];
 509                         spin_lock(&head->lock);
 510
 511                         /* Does not bother with rcv_saddr checks,
 512                          * because the established check is already
 513                          * unique enough.
 514                          */
 515                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 516                                 if (net_eq(ib_net(tb), net) &&
 517                                     tb->port == port) {
 518                                         if (tb->fastreuse >= 0 ||
 519                                             tb->fastreuseport >= 0)
 520                                                 goto next_port;
 521                                         WARN_ON(hlist_empty(&tb->owners));
 522                                         if (!check_established(death_row, sk,
 523                                                                 port, &tw))
 524                                                 goto ok;
 525                                         goto next_port;
 526                                 }
 527                         }
 528
 529                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 530                                         net, head, port);
 531                         if (!tb) {
 532                                 spin_unlock(&head->lock);
 533                                 break;
 534                         }
 535                         tb->fastreuse = -1;
 536                         tb->fastreuseport = -1;
 537                         goto ok;
 538
 539                 next_port:
 540                         spin_unlock(&head->lock);
 541                 }
 542                 local_bh_enable();
 543
 544                 return -EADDRNOTAVAIL;
 545
 546 ok:
 547                 hint += i;
 548
 549                 /* Head lock still held and bh's disabled */
 550                 inet_bind_hash(sk, tb, port);
 551                 if (sk_unhashed(sk)) {
 552                         inet_sk(sk)->inet_sport = htons(port);
 553                         twrefcnt += hash(sk, tw);
 554                 }
 555                 if (tw)
 556                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
 557                 spin_unlock(&head->lock);
 558
 559                 if (tw) {
 560                         inet_twsk_deschedule(tw, death_row);
 561                         while (twrefcnt) {
 562                                 twrefcnt--;
 563                                 inet_twsk_put(tw);
 564                         }
 565                 }
 566
 567                 ret = 0;
 568                 goto out;
 569         }
 570
 571         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 572         tb  = inet_csk(sk)->icsk_bind_hash;
 573         spin_lock_bh(&head->lock);
 574         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 575                 hash(sk, NULL);
 576                 spin_unlock_bh(&head->lock);
 577                 return 0;
 578         } else {
 579                 spin_unlock(&head->lock);
 580                 /* No definite answer... Walk to established hash table */
 581                 ret = check_established(death_row, sk, snum, NULL);
 582 out:
 583                 local_bh_enable();
 584                 return ret;
 585         }
 586 }
 587
 588 /*
 589  * Bind a port for a connect operation and hash it.
 590  */
 591 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 592                       struct sock *sk)
 593 {
 594         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 595                         __inet_check_established, __inet_hash_nolisten);
 596 }
 597 EXPORT_SYMBOL_GPL(inet_hash_connect);
 598
 599 void inet_hashinfo_init(struct inet_hashinfo *h)
 600 {
 601         int i;
 602
 603         atomic_set(&h->bsockets, 0);
 604         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 605                 spin_lock_init(&h->listening_hash[i].lock);
 606                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 607                                       i + LISTENING_NULLS_BASE);
 608                 }
 609 }
 610 EXPORT_SYMBOL_GPL(inet_hashinfo_init);