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