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