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                 tb->ib_net       = hold_net(net);
  39                 tb->port      = snum;
  40                 tb->fastreuse = 0;
  41                 INIT_HLIST_HEAD(&tb->owners);
  42                 hlist_add_head(&tb->node, &head->chain);
  43         }
  44         return tb;
  45 }
  46
  47 /*
  48  * Caller must hold hashbucket lock for this tb with local BH disabled
  49  */
  50 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  51 {
  52         if (hlist_empty(&tb->owners)) {
  53                 __hlist_del(&tb->node);
  54                 release_net(tb->ib_net);
  55                 kmem_cache_free(cachep, tb);
  56         }
  57 }
  58
  59 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  60                     const unsigned short snum)
  61 {
  62         inet_sk(sk)->num = snum;
  63         sk_add_bind_node(sk, &tb->owners);
  64         inet_csk(sk)->icsk_bind_hash = tb;
  65 }
  66
  67 /*
  68  * Get rid of any references to a local port held by the given sock.
  69  */
  70 static void __inet_put_port(struct sock *sk)
  71 {
  72         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  73         const int bhash = inet_bhashfn(inet_sk(sk)->num, hashinfo->bhash_size);
  74         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
  75         struct inet_bind_bucket *tb;
  76
  77         spin_lock(&head->lock);
  78         tb = inet_csk(sk)->icsk_bind_hash;
  79         __sk_del_bind_node(sk);
  80         inet_csk(sk)->icsk_bind_hash = NULL;
  81         inet_sk(sk)->num = 0;
  82         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  83         spin_unlock(&head->lock);
  84 }
  85
  86 void inet_put_port(struct sock *sk)
  87 {
  88         local_bh_disable();
  89         __inet_put_port(sk);
  90         local_bh_enable();
  91 }
  92
  93 EXPORT_SYMBOL(inet_put_port);
  94
  95 void __inet_inherit_port(struct sock *sk, struct sock *child)
  96 {
  97         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
  98         const int bhash = inet_bhashfn(inet_sk(child)->num, table->bhash_size);
  99         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 100         struct inet_bind_bucket *tb;
 101
 102         spin_lock(&head->lock);
 103         tb = inet_csk(sk)->icsk_bind_hash;
 104         sk_add_bind_node(child, &tb->owners);
 105         inet_csk(child)->icsk_bind_hash = tb;
 106         spin_unlock(&head->lock);
 107 }
 108
 109 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 110
 111 /*
 112  * This lock without WQ_FLAG_EXCLUSIVE is good on UP and it can be very bad on SMP.
 113  * Look, when several writers sleep and reader wakes them up, all but one
 114  * immediately hit write lock and grab all the cpus. Exclusive sleep solves
 115  * this, _but_ remember, it adds useless work on UP machines (wake up each
 116  * exclusive lock release). It should be ifdefed really.
 117  */
 118 void inet_listen_wlock(struct inet_hashinfo *hashinfo)
 119         __acquires(hashinfo->lhash_lock)
 120 {
 121         write_lock(&hashinfo->lhash_lock);
 122
 123         if (atomic_read(&hashinfo->lhash_users)) {
 124                 DEFINE_WAIT(wait);
 125
 126                 for (;;) {
 127                         prepare_to_wait_exclusive(&hashinfo->lhash_wait,
 128                                                   &wait, TASK_UNINTERRUPTIBLE);
 129                         if (!atomic_read(&hashinfo->lhash_users))
 130                                 break;
 131                         write_unlock_bh(&hashinfo->lhash_lock);
 132                         schedule();
 133                         write_lock_bh(&hashinfo->lhash_lock);
 134                 }
 135
 136                 finish_wait(&hashinfo->lhash_wait, &wait);
 137         }
 138 }
 139
 140 /*
 141  * Don't inline this cruft. Here are some nice properties to exploit here. The
 142  * BSD API does not allow a listening sock to specify the remote port nor the
 143  * remote address for the connection. So always assume those are both
 144  * wildcarded during the search since they can never be otherwise.
 145  */
 146 static struct sock *inet_lookup_listener_slow(struct net *net,
 147                                               const struct hlist_head *head,
 148                                               const __be32 daddr,
 149                                               const unsigned short hnum,
 150                                               const int dif)
 151 {
 152         struct sock *result = NULL, *sk;
 153         const struct hlist_node *node;
 154         int hiscore = -1;
 155
 156         sk_for_each(sk, node, head) {
 157                 const struct inet_sock *inet = inet_sk(sk);
 158
 159                 if (net_eq(sock_net(sk), net) && inet->num == hnum &&
 160                                 !ipv6_only_sock(sk)) {
 161                         const __be32 rcv_saddr = inet->rcv_saddr;
 162                         int score = sk->sk_family == PF_INET ? 1 : 0;
 163
 164                         if (rcv_saddr) {
 165                                 if (rcv_saddr != daddr)
 166                                         continue;
 167                                 score += 2;
 168                         }
 169                         if (sk->sk_bound_dev_if) {
 170                                 if (sk->sk_bound_dev_if != dif)
 171                                         continue;
 172                                 score += 2;
 173                         }
 174                         if (score == 5)
 175                                 return sk;
 176                         if (score > hiscore) {
 177                                 hiscore = score;
 178                                 result  = sk;
 179                         }
 180                 }
 181         }
 182         return result;
 183 }
 184
 185 /* Optimize the common listener case. */
 186 struct sock *__inet_lookup_listener(struct net *net,
 187                                     struct inet_hashinfo *hashinfo,
 188                                     const __be32 daddr, const unsigned short hnum,
 189                                     const int dif)
 190 {
 191         struct sock *sk = NULL;
 192         const struct hlist_head *head;
 193
 194         read_lock(&hashinfo->lhash_lock);
 195         head = &hashinfo->listening_hash[inet_lhashfn(hnum)];
 196         if (!hlist_empty(head)) {
 197                 const struct inet_sock *inet = inet_sk((sk = __sk_head(head)));
 198
 199                 if (inet->num == hnum && !sk->sk_node.next &&
 200                     (!inet->rcv_saddr || inet->rcv_saddr == daddr) &&
 201                     (sk->sk_family == PF_INET || !ipv6_only_sock(sk)) &&
 202                     !sk->sk_bound_dev_if && net_eq(sock_net(sk), net))
 203                         goto sherry_cache;
 204                 sk = inet_lookup_listener_slow(net, head, daddr, hnum, dif);
 205         }
 206         if (sk) {
 207 sherry_cache:
 208                 sock_hold(sk);
 209         }
 210         read_unlock(&hashinfo->lhash_lock);
 211         return sk;
 212 }
 213 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 214
 215 struct sock * __inet_lookup_established(struct net *net,
 216                                   struct inet_hashinfo *hashinfo,
 217                                   const __be32 saddr, const __be16 sport,
 218                                   const __be32 daddr, const u16 hnum,
 219                                   const int dif)
 220 {
 221         INET_ADDR_COOKIE(acookie, saddr, daddr)
 222         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 223         struct sock *sk;
 224         const struct hlist_node *node;
 225         /* Optimize here for direct hit, only listening connections can
 226          * have wildcards anyways.
 227          */
 228         unsigned int hash = inet_ehashfn(daddr, hnum, saddr, sport);
 229         struct inet_ehash_bucket *head = inet_ehash_bucket(hashinfo, hash);
 230         rwlock_t *lock = inet_ehash_lockp(hashinfo, hash);
 231
 232         prefetch(head->chain.first);
 233         read_lock(lock);
 234         sk_for_each(sk, node, &head->chain) {
 235                 if (INET_MATCH(sk, net, hash, acookie,
 236                                         saddr, daddr, ports, dif))
 237                         goto hit; /* You sunk my battleship! */
 238         }
 239
 240         /* Must check for a TIME_WAIT'er before going to listener hash. */
 241         sk_for_each(sk, node, &head->twchain) {
 242                 if (INET_TW_MATCH(sk, net, hash, acookie,
 243                                         saddr, daddr, ports, dif))
 244                         goto hit;
 245         }
 246         sk = NULL;
 247 out:
 248         read_unlock(lock);
 249         return sk;
 250 hit:
 251         sock_hold(sk);
 252         goto out;
 253 }
 254 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 255
 256 /* called with local bh disabled */
 257 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 258                                     struct sock *sk, __u16 lport,
 259                                     struct inet_timewait_sock **twp)
 260 {
 261         struct inet_hashinfo *hinfo = death_row->hashinfo;
 262         struct inet_sock *inet = inet_sk(sk);
 263         __be32 daddr = inet->rcv_saddr;
 264         __be32 saddr = inet->daddr;
 265         int dif = sk->sk_bound_dev_if;
 266         INET_ADDR_COOKIE(acookie, saddr, daddr)
 267         const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
 268         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
 269         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 270         rwlock_t *lock = inet_ehash_lockp(hinfo, hash);
 271         struct sock *sk2;
 272         const struct hlist_node *node;
 273         struct inet_timewait_sock *tw;
 274         struct net *net = sock_net(sk);
 275
 276         prefetch(head->chain.first);
 277         write_lock(lock);
 278
 279         /* Check TIME-WAIT sockets first. */
 280         sk_for_each(sk2, node, &head->twchain) {
 281                 tw = inet_twsk(sk2);
 282
 283                 if (INET_TW_MATCH(sk2, net, hash, acookie,
 284                                         saddr, daddr, ports, dif)) {
 285                         if (twsk_unique(sk, sk2, twp))
 286                                 goto unique;
 287                         else
 288                                 goto not_unique;
 289                 }
 290         }
 291         tw = NULL;
 292
 293         /* And established part... */
 294         sk_for_each(sk2, node, &head->chain) {
 295                 if (INET_MATCH(sk2, net, hash, acookie,
 296                                         saddr, daddr, ports, dif))
 297                         goto not_unique;
 298         }
 299
 300 unique:
 301         /* Must record num and sport now. Otherwise we will see
 302          * in hash table socket with a funny identity. */
 303         inet->num = lport;
 304         inet->sport = htons(lport);
 305         sk->sk_hash = hash;
 306         BUG_TRAP(sk_unhashed(sk));
 307         __sk_add_node(sk, &head->chain);
 308         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 309         write_unlock(lock);
 310
 311         if (twp) {
 312                 *twp = tw;
 313                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 314         } else if (tw) {
 315                 /* Silly. Should hash-dance instead... */
 316                 inet_twsk_deschedule(tw, death_row);
 317                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 318
 319                 inet_twsk_put(tw);
 320         }
 321
 322         return 0;
 323
 324 not_unique:
 325         write_unlock(lock);
 326         return -EADDRNOTAVAIL;
 327 }
 328
 329 static inline u32 inet_sk_port_offset(const struct sock *sk)
 330 {
 331         const struct inet_sock *inet = inet_sk(sk);
 332         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
 333                                           inet->dport);
 334 }
 335
 336 void __inet_hash_nolisten(struct sock *sk)
 337 {
 338         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 339         struct hlist_head *list;
 340         rwlock_t *lock;
 341         struct inet_ehash_bucket *head;
 342
 343         BUG_TRAP(sk_unhashed(sk));
 344
 345         sk->sk_hash = inet_sk_ehashfn(sk);
 346         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 347         list = &head->chain;
 348         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 349
 350         write_lock(lock);
 351         __sk_add_node(sk, list);
 352         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 353         write_unlock(lock);
 354 }
 355 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 356
 357 static void __inet_hash(struct sock *sk)
 358 {
 359         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 360         struct hlist_head *list;
 361         rwlock_t *lock;
 362
 363         if (sk->sk_state != TCP_LISTEN) {
 364                 __inet_hash_nolisten(sk);
 365                 return;
 366         }
 367
 368         BUG_TRAP(sk_unhashed(sk));
 369         list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 370         lock = &hashinfo->lhash_lock;
 371
 372         inet_listen_wlock(hashinfo);
 373         __sk_add_node(sk, list);
 374         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 375         write_unlock(lock);
 376         wake_up(&hashinfo->lhash_wait);
 377 }
 378
 379 void inet_hash(struct sock *sk)
 380 {
 381         if (sk->sk_state != TCP_CLOSE) {
 382                 local_bh_disable();
 383                 __inet_hash(sk);
 384                 local_bh_enable();
 385         }
 386 }
 387 EXPORT_SYMBOL_GPL(inet_hash);
 388
 389 void inet_unhash(struct sock *sk)
 390 {
 391         rwlock_t *lock;
 392         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 393
 394         if (sk_unhashed(sk))
 395                 goto out;
 396
 397         if (sk->sk_state == TCP_LISTEN) {
 398                 local_bh_disable();
 399                 inet_listen_wlock(hashinfo);
 400                 lock = &hashinfo->lhash_lock;
 401         } else {
 402                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 403                 write_lock_bh(lock);
 404         }
 405
 406         if (__sk_del_node_init(sk))
 407                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 408         write_unlock_bh(lock);
 409 out:
 410         if (sk->sk_state == TCP_LISTEN)
 411                 wake_up(&hashinfo->lhash_wait);
 412 }
 413 EXPORT_SYMBOL_GPL(inet_unhash);
 414
 415 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 416                 struct sock *sk, u32 port_offset,
 417                 int (*check_established)(struct inet_timewait_death_row *,
 418                         struct sock *, __u16, struct inet_timewait_sock **),
 419                 void (*hash)(struct sock *sk))
 420 {
 421         struct inet_hashinfo *hinfo = death_row->hashinfo;
 422         const unsigned short snum = inet_sk(sk)->num;
 423         struct inet_bind_hashbucket *head;
 424         struct inet_bind_bucket *tb;
 425         int ret;
 426         struct net *net = sock_net(sk);
 427
 428         if (!snum) {
 429                 int i, remaining, low, high, port;
 430                 static u32 hint;
 431                 u32 offset = hint + port_offset;
 432                 struct hlist_node *node;
 433                 struct inet_timewait_sock *tw = NULL;
 434
 435                 inet_get_local_port_range(&low, &high);
 436                 remaining = (high - low) + 1;
 437
 438                 local_bh_disable();
 439                 for (i = 1; i <= remaining; i++) {
 440                         port = low + (i + offset) % remaining;
 441                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
 442                         spin_lock(&head->lock);
 443
 444                         /* Does not bother with rcv_saddr checks,
 445                          * because the established check is already
 446                          * unique enough.
 447                          */
 448                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 449                                 if (tb->ib_net == net && tb->port == port) {
 450                                         BUG_TRAP(!hlist_empty(&tb->owners));
 451                                         if (tb->fastreuse >= 0)
 452                                                 goto next_port;
 453                                         if (!check_established(death_row, sk,
 454                                                                 port, &tw))
 455                                                 goto ok;
 456                                         goto next_port;
 457                                 }
 458                         }
 459
 460                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 461                                         net, head, port);
 462                         if (!tb) {
 463                                 spin_unlock(&head->lock);
 464                                 break;
 465                         }
 466                         tb->fastreuse = -1;
 467                         goto ok;
 468
 469                 next_port:
 470                         spin_unlock(&head->lock);
 471                 }
 472                 local_bh_enable();
 473
 474                 return -EADDRNOTAVAIL;
 475
 476 ok:
 477                 hint += i;
 478
 479                 /* Head lock still held and bh's disabled */
 480                 inet_bind_hash(sk, tb, port);
 481                 if (sk_unhashed(sk)) {
 482                         inet_sk(sk)->sport = htons(port);
 483                         hash(sk);
 484                 }
 485                 spin_unlock(&head->lock);
 486
 487                 if (tw) {
 488                         inet_twsk_deschedule(tw, death_row);
 489                         inet_twsk_put(tw);
 490                 }
 491
 492                 ret = 0;
 493                 goto out;
 494         }
 495
 496         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
 497         tb  = inet_csk(sk)->icsk_bind_hash;
 498         spin_lock_bh(&head->lock);
 499         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 500                 hash(sk);
 501                 spin_unlock_bh(&head->lock);
 502                 return 0;
 503         } else {
 504                 spin_unlock(&head->lock);
 505                 /* No definite answer... Walk to established hash table */
 506                 ret = check_established(death_row, sk, snum, NULL);
 507 out:
 508                 local_bh_enable();
 509                 return ret;
 510         }
 511 }
 512
 513 /*
 514  * Bind a port for a connect operation and hash it.
 515  */
 516 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 517                       struct sock *sk)
 518 {
 519         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 520                         __inet_check_established, __inet_hash_nolisten);
 521 }
 522
 523 EXPORT_SYMBOL_GPL(inet_hash_connect);