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->num_owners = 0;
  43                 INIT_HLIST_HEAD(&tb->owners);
  44                 hlist_add_head(&tb->node, &head->chain);
  45         }
  46         return tb;
  47 }
  48
  49 /*
  50  * Caller must hold hashbucket lock for this tb with local BH disabled
  51  */
  52 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  53 {
  54         if (hlist_empty(&tb->owners)) {
  55                 __hlist_del(&tb->node);
  56                 release_net(ib_net(tb));
  57                 kmem_cache_free(cachep, tb);
  58         }
  59 }
  60
  61 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  62                     const unsigned short snum)
  63 {
  64         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  65
  66         atomic_inc(&hashinfo->bsockets);
  67
  68         inet_sk(sk)->inet_num = snum;
  69         sk_add_bind_node(sk, &tb->owners);
  70         tb->num_owners++;
  71         inet_csk(sk)->icsk_bind_hash = tb;
  72 }
  73
  74 /*
  75  * Get rid of any references to a local port held by the given sock.
  76  */
  77 static void __inet_put_port(struct sock *sk)
  78 {
  79         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  80         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
  81                         hashinfo->bhash_size);
  82         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
  83         struct inet_bind_bucket *tb;
  84
  85         atomic_dec(&hashinfo->bsockets);
  86
  87         spin_lock(&head->lock);
  88         tb = inet_csk(sk)->icsk_bind_hash;
  89         __sk_del_bind_node(sk);
  90         tb->num_owners--;
  91         inet_csk(sk)->icsk_bind_hash = NULL;
  92         inet_sk(sk)->inet_num = 0;
  93         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  94         spin_unlock(&head->lock);
  95 }
  96
  97 void inet_put_port(struct sock *sk)
  98 {
  99         local_bh_disable();
 100         __inet_put_port(sk);
 101         local_bh_enable();
 102 }
 103 EXPORT_SYMBOL(inet_put_port);
 104
 105 int __inet_inherit_port(struct sock *sk, struct sock *child)
 106 {
 107         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 108         unsigned short port = inet_sk(child)->inet_num;
 109         const int bhash = inet_bhashfn(sock_net(sk), port,
 110                         table->bhash_size);
 111         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 112         struct inet_bind_bucket *tb;
 113
 114         spin_lock(&head->lock);
 115         tb = inet_csk(sk)->icsk_bind_hash;
 116         if (tb->port != port) {
 117                 /* NOTE: using tproxy and redirecting skbs to a proxy
 118                  * on a different listener port breaks the assumption
 119                  * that the listener socket's icsk_bind_hash is the same
 120                  * as that of the child socket. We have to look up or
 121                  * create a new bind bucket for the child here. */
 122                 struct hlist_node *node;
 123                 inet_bind_bucket_for_each(tb, node, &head->chain) {
 124                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 125                             tb->port == port)
 126                                 break;
 127                 }
 128                 if (!node) {
 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 ? 1 : 0;
 155                 if (rcv_saddr) {
 156                         if (rcv_saddr != daddr)
 157                                 return -1;
 158                         score += 2;
 159                 }
 160                 if (sk->sk_bound_dev_if) {
 161                         if (sk->sk_bound_dev_if != dif)
 162                                 return -1;
 163                         score += 2;
 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 daddr, const unsigned short hnum,
 180                                     const int dif)
 181 {
 182         struct sock *sk, *result;
 183         struct hlist_nulls_node *node;
 184         unsigned int hash = inet_lhashfn(net, hnum);
 185         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 186         int score, hiscore;
 187
 188         rcu_read_lock();
 189 begin:
 190         result = NULL;
 191         hiscore = -1;
 192         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 193                 score = compute_score(sk, net, hnum, daddr, dif);
 194                 if (score > hiscore) {
 195                         result = sk;
 196                         hiscore = score;
 197                 }
 198         }
 199         /*
 200          * if the nulls value we got at the end of this lookup is
 201          * not the expected one, we must restart lookup.
 202          * We probably met an item that was moved to another chain.
 203          */
 204         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 205                 goto begin;
 206         if (result) {
 207                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 208                         result = NULL;
 209                 else if (unlikely(compute_score(result, net, hnum, daddr,
 210                                   dif) < hiscore)) {
 211                         sock_put(result);
 212                         goto begin;
 213                 }
 214         }
 215         rcu_read_unlock();
 216         return result;
 217 }
 218 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 219
 220 struct sock *__inet_lookup_established(struct net *net,
 221                                   struct inet_hashinfo *hashinfo,
 222                                   const __be32 saddr, const __be16 sport,
 223                                   const __be32 daddr, const u16 hnum,
 224                                   const int dif)
 225 {
 226         INET_ADDR_COOKIE(acookie, saddr, daddr)
 227         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 228         struct sock *sk;
 229         const struct hlist_nulls_node *node;
 230         /* Optimize here for direct hit, only listening connections can
 231          * have wildcards anyways.
 232          */
 233         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 234         unsigned int slot = hash & hashinfo->ehash_mask;
 235         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 236
 237         rcu_read_lock();
 238 begin:
 239         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 240                 if (sk->sk_hash != hash)
 241                         continue;
 242                 if (likely(INET_MATCH(sk, net, acookie,
 243                                       saddr, daddr, ports, dif))) {
 244                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 245                                 goto begintw;
 246                         if (unlikely(!INET_MATCH(sk, net, acookie,
 247                                                  saddr, daddr, ports, dif))) {
 248                                 sock_put(sk);
 249                                 goto begin;
 250                         }
 251                         goto out;
 252                 }
 253         }
 254         /*
 255          * if the nulls value we got at the end of this lookup is
 256          * not the expected one, we must restart lookup.
 257          * We probably met an item that was moved to another chain.
 258          */
 259         if (get_nulls_value(node) != slot)
 260                 goto begin;
 261
 262 begintw:
 263         /* Must check for a TIME_WAIT'er before going to listener hash. */
 264         sk_nulls_for_each_rcu(sk, node, &head->twchain) {
 265                 if (sk->sk_hash != hash)
 266                         continue;
 267                 if (likely(INET_TW_MATCH(sk, net, acookie,
 268                                          saddr, daddr, ports,
 269                                          dif))) {
 270                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) {
 271                                 sk = NULL;
 272                                 goto out;
 273                         }
 274                         if (unlikely(!INET_TW_MATCH(sk, net, acookie,
 275                                                     saddr, daddr, ports,
 276                                                     dif))) {
 277                                 sock_put(sk);
 278                                 goto begintw;
 279                         }
 280                         goto out;
 281                 }
 282         }
 283         /*
 284          * if the nulls value we got at the end of this lookup is
 285          * not the expected one, we must restart lookup.
 286          * We probably met an item that was moved to another chain.
 287          */
 288         if (get_nulls_value(node) != slot)
 289                 goto begintw;
 290         sk = NULL;
 291 out:
 292         rcu_read_unlock();
 293         return sk;
 294 }
 295 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 296
 297 /* called with local bh disabled */
 298 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 299                                     struct sock *sk, __u16 lport,
 300                                     struct inet_timewait_sock **twp)
 301 {
 302         struct inet_hashinfo *hinfo = death_row->hashinfo;
 303         struct inet_sock *inet = inet_sk(sk);
 304         __be32 daddr = inet->inet_rcv_saddr;
 305         __be32 saddr = inet->inet_daddr;
 306         int dif = sk->sk_bound_dev_if;
 307         INET_ADDR_COOKIE(acookie, saddr, daddr)
 308         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 309         struct net *net = sock_net(sk);
 310         unsigned int hash = inet_ehashfn(net, daddr, lport,
 311                                          saddr, inet->inet_dport);
 312         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 313         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 314         struct sock *sk2;
 315         const struct hlist_nulls_node *node;
 316         struct inet_timewait_sock *tw;
 317         int twrefcnt = 0;
 318
 319         spin_lock(lock);
 320
 321         /* Check TIME-WAIT sockets first. */
 322         sk_nulls_for_each(sk2, node, &head->twchain) {
 323                 if (sk2->sk_hash != hash)
 324                         continue;
 325
 326                 if (likely(INET_TW_MATCH(sk2, net, acookie,
 327                                          saddr, daddr, ports, dif))) {
 328                         tw = inet_twsk(sk2);
 329                         if (twsk_unique(sk, sk2, twp))
 330                                 goto unique;
 331                         else
 332                                 goto not_unique;
 333                 }
 334         }
 335         tw = NULL;
 336
 337         /* And established part... */
 338         sk_nulls_for_each(sk2, node, &head->chain) {
 339                 if (sk2->sk_hash != hash)
 340                         continue;
 341                 if (likely(INET_MATCH(sk2, net, acookie,
 342                                       saddr, daddr, ports, dif)))
 343                         goto not_unique;
 344         }
 345
 346 unique:
 347         /* Must record num and sport now. Otherwise we will see
 348          * in hash table socket with a funny identity. */
 349         inet->inet_num = lport;
 350         inet->inet_sport = htons(lport);
 351         sk->sk_hash = hash;
 352         WARN_ON(!sk_unhashed(sk));
 353         __sk_nulls_add_node_rcu(sk, &head->chain);
 354         if (tw) {
 355                 twrefcnt = inet_twsk_unhash(tw);
 356                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 357         }
 358         spin_unlock(lock);
 359         if (twrefcnt)
 360                 inet_twsk_put(tw);
 361         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 362
 363         if (twp) {
 364                 *twp = tw;
 365         } else if (tw) {
 366                 /* Silly. Should hash-dance instead... */
 367                 inet_twsk_deschedule(tw, death_row);
 368
 369                 inet_twsk_put(tw);
 370         }
 371         return 0;
 372
 373 not_unique:
 374         spin_unlock(lock);
 375         return -EADDRNOTAVAIL;
 376 }
 377
 378 static inline u32 inet_sk_port_offset(const struct sock *sk)
 379 {
 380         const struct inet_sock *inet = inet_sk(sk);
 381         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 382                                           inet->inet_daddr,
 383                                           inet->inet_dport);
 384 }
 385
 386 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
 387 {
 388         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 389         struct hlist_nulls_head *list;
 390         spinlock_t *lock;
 391         struct inet_ehash_bucket *head;
 392         int twrefcnt = 0;
 393
 394         WARN_ON(!sk_unhashed(sk));
 395
 396         sk->sk_hash = inet_sk_ehashfn(sk);
 397         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 398         list = &head->chain;
 399         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 400
 401         spin_lock(lock);
 402         __sk_nulls_add_node_rcu(sk, list);
 403         if (tw) {
 404                 WARN_ON(sk->sk_hash != tw->tw_hash);
 405                 twrefcnt = inet_twsk_unhash(tw);
 406         }
 407         spin_unlock(lock);
 408         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 409         return twrefcnt;
 410 }
 411 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 412
 413 static void __inet_hash(struct sock *sk)
 414 {
 415         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 416         struct inet_listen_hashbucket *ilb;
 417
 418         if (sk->sk_state != TCP_LISTEN) {
 419                 __inet_hash_nolisten(sk, NULL);
 420                 return;
 421         }
 422
 423         WARN_ON(!sk_unhashed(sk));
 424         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 425
 426         spin_lock(&ilb->lock);
 427         __sk_nulls_add_node_rcu(sk, &ilb->head);
 428         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 429         spin_unlock(&ilb->lock);
 430 }
 431
 432 void inet_hash(struct sock *sk)
 433 {
 434         if (sk->sk_state != TCP_CLOSE) {
 435                 local_bh_disable();
 436                 __inet_hash(sk);
 437                 local_bh_enable();
 438         }
 439 }
 440 EXPORT_SYMBOL_GPL(inet_hash);
 441
 442 void inet_unhash(struct sock *sk)
 443 {
 444         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 445         spinlock_t *lock;
 446         int done;
 447
 448         if (sk_unhashed(sk))
 449                 return;
 450
 451         if (sk->sk_state == TCP_LISTEN)
 452                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 453         else
 454                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 455
 456         spin_lock_bh(lock);
 457         done =__sk_nulls_del_node_init_rcu(sk);
 458         if (done)
 459                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 460         spin_unlock_bh(lock);
 461 }
 462 EXPORT_SYMBOL_GPL(inet_unhash);
 463
 464 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 465                 struct sock *sk, u32 port_offset,
 466                 int (*check_established)(struct inet_timewait_death_row *,
 467                         struct sock *, __u16, struct inet_timewait_sock **),
 468                 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
 469 {
 470         struct inet_hashinfo *hinfo = death_row->hashinfo;
 471         const unsigned short snum = inet_sk(sk)->inet_num;
 472         struct inet_bind_hashbucket *head;
 473         struct inet_bind_bucket *tb;
 474         int ret;
 475         struct net *net = sock_net(sk);
 476         int twrefcnt = 1;
 477
 478         if (!snum) {
 479                 int i, remaining, low, high, port;
 480                 static u32 hint;
 481                 u32 offset = hint + port_offset;
 482                 struct hlist_node *node;
 483                 struct inet_timewait_sock *tw = NULL;
 484
 485                 inet_get_local_port_range(&low, &high);
 486                 remaining = (high - low) + 1;
 487
 488                 local_bh_disable();
 489                 for (i = 1; i <= remaining; i++) {
 490                         port = low + (i + offset) % remaining;
 491                         if (inet_is_reserved_local_port(port))
 492                                 continue;
 493                         head = &hinfo->bhash[inet_bhashfn(net, port,
 494                                         hinfo->bhash_size)];
 495                         spin_lock(&head->lock);
 496
 497                         /* Does not bother with rcv_saddr checks,
 498                          * because the established check is already
 499                          * unique enough.
 500                          */
 501                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 502                                 if (net_eq(ib_net(tb), net) &&
 503                                     tb->port == port) {
 504                                         if (tb->fastreuse >= 0)
 505                                                 goto next_port;
 506                                         WARN_ON(hlist_empty(&tb->owners));
 507                                         if (!check_established(death_row, sk,
 508                                                                 port, &tw))
 509                                                 goto ok;
 510                                         goto next_port;
 511                                 }
 512                         }
 513
 514                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 515                                         net, head, port);
 516                         if (!tb) {
 517                                 spin_unlock(&head->lock);
 518                                 break;
 519                         }
 520                         tb->fastreuse = -1;
 521                         goto ok;
 522
 523                 next_port:
 524                         spin_unlock(&head->lock);
 525                 }
 526                 local_bh_enable();
 527
 528                 return -EADDRNOTAVAIL;
 529
 530 ok:
 531                 hint += i;
 532
 533                 /* Head lock still held and bh's disabled */
 534                 inet_bind_hash(sk, tb, port);
 535                 if (sk_unhashed(sk)) {
 536                         inet_sk(sk)->inet_sport = htons(port);
 537                         twrefcnt += hash(sk, tw);
 538                 }
 539                 if (tw)
 540                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
 541                 spin_unlock(&head->lock);
 542
 543                 if (tw) {
 544                         inet_twsk_deschedule(tw, death_row);
 545                         while (twrefcnt) {
 546                                 twrefcnt--;
 547                                 inet_twsk_put(tw);
 548                         }
 549                 }
 550
 551                 ret = 0;
 552                 goto out;
 553         }
 554
 555         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 556         tb  = inet_csk(sk)->icsk_bind_hash;
 557         spin_lock_bh(&head->lock);
 558         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 559                 hash(sk, NULL);
 560                 spin_unlock_bh(&head->lock);
 561                 return 0;
 562         } else {
 563                 spin_unlock(&head->lock);
 564                 /* No definite answer... Walk to established hash table */
 565                 ret = check_established(death_row, sk, snum, NULL);
 566 out:
 567                 local_bh_enable();
 568                 return ret;
 569         }
 570 }
 571
 572 /*
 573  * Bind a port for a connect operation and hash it.
 574  */
 575 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 576                       struct sock *sk)
 577 {
 578         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 579                         __inet_check_established, __inet_hash_nolisten);
 580 }
 581 EXPORT_SYMBOL_GPL(inet_hash_connect);
 582
 583 void inet_hashinfo_init(struct inet_hashinfo *h)
 584 {
 585         int i;
 586
 587         atomic_set(&h->bsockets, 0);
 588         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 589                 spin_lock_init(&h->listening_hash[i].lock);
 590                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 591                                       i + LISTENING_NULLS_BASE);
 592                 }
 593 }
 594 EXPORT_SYMBOL_GPL(inet_hashinfo_init);