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 static unsigned int inet_ehashfn(struct net *net, const __be32 laddr,
  28                                  const __u16 lport, const __be32 faddr,
  29                                  const __be16 fport)
  30 {
  31         static u32 inet_ehash_secret __read_mostly;
  32
  33         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  34
  35         return __inet_ehashfn(laddr, lport, faddr, fport,
  36                               inet_ehash_secret + net_hash_mix(net));
  37 }
  38
  39
  40 static unsigned int inet_sk_ehashfn(const struct sock *sk)
  41 {
  42         const struct inet_sock *inet = inet_sk(sk);
  43         const __be32 laddr = inet->inet_rcv_saddr;
  44         const __u16 lport = inet->inet_num;
  45         const __be32 faddr = inet->inet_daddr;
  46         const __be16 fport = inet->inet_dport;
  47         struct net *net = sock_net(sk);
  48
  49         return inet_ehashfn(net, laddr, lport, faddr, fport);
  50 }
  51
  52 /*
  53  * Allocate and initialize a new local port bind bucket.
  54  * The bindhash mutex for snum's hash chain must be held here.
  55  */
  56 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  57                                                  struct net *net,
  58                                                  struct inet_bind_hashbucket *head,
  59                                                  const unsigned short snum)
  60 {
  61         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  62
  63         if (tb != NULL) {
  64                 write_pnet(&tb->ib_net, net);
  65                 tb->port      = snum;
  66                 tb->fastreuse = 0;
  67                 tb->fastreuseport = 0;
  68                 tb->num_owners = 0;
  69                 INIT_HLIST_HEAD(&tb->owners);
  70                 hlist_add_head(&tb->node, &head->chain);
  71         }
  72         return tb;
  73 }
  74
  75 /*
  76  * Caller must hold hashbucket lock for this tb with local BH disabled
  77  */
  78 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  79 {
  80         if (hlist_empty(&tb->owners)) {
  81                 __hlist_del(&tb->node);
  82                 kmem_cache_free(cachep, tb);
  83         }
  84 }
  85
  86 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  87                     const unsigned short snum)
  88 {
  89         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  90
  91         atomic_inc(&hashinfo->bsockets);
  92
  93         inet_sk(sk)->inet_num = snum;
  94         sk_add_bind_node(sk, &tb->owners);
  95         tb->num_owners++;
  96         inet_csk(sk)->icsk_bind_hash = tb;
  97 }
  98
  99 /*
 100  * Get rid of any references to a local port held by the given sock.
 101  */
 102 static void __inet_put_port(struct sock *sk)
 103 {
 104         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 105         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 106                         hashinfo->bhash_size);
 107         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 108         struct inet_bind_bucket *tb;
 109
 110         atomic_dec(&hashinfo->bsockets);
 111
 112         spin_lock(&head->lock);
 113         tb = inet_csk(sk)->icsk_bind_hash;
 114         __sk_del_bind_node(sk);
 115         tb->num_owners--;
 116         inet_csk(sk)->icsk_bind_hash = NULL;
 117         inet_sk(sk)->inet_num = 0;
 118         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 119         spin_unlock(&head->lock);
 120 }
 121
 122 void inet_put_port(struct sock *sk)
 123 {
 124         local_bh_disable();
 125         __inet_put_port(sk);
 126         local_bh_enable();
 127 }
 128 EXPORT_SYMBOL(inet_put_port);
 129
 130 int __inet_inherit_port(struct sock *sk, struct sock *child)
 131 {
 132         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 133         unsigned short port = inet_sk(child)->inet_num;
 134         const int bhash = inet_bhashfn(sock_net(sk), port,
 135                         table->bhash_size);
 136         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 137         struct inet_bind_bucket *tb;
 138
 139         spin_lock(&head->lock);
 140         tb = inet_csk(sk)->icsk_bind_hash;
 141         if (tb->port != port) {
 142                 /* NOTE: using tproxy and redirecting skbs to a proxy
 143                  * on a different listener port breaks the assumption
 144                  * that the listener socket's icsk_bind_hash is the same
 145                  * as that of the child socket. We have to look up or
 146                  * create a new bind bucket for the child here. */
 147                 inet_bind_bucket_for_each(tb, &head->chain) {
 148                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 149                             tb->port == port)
 150                                 break;
 151                 }
 152                 if (!tb) {
 153                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 154                                                      sock_net(sk), head, port);
 155                         if (!tb) {
 156                                 spin_unlock(&head->lock);
 157                                 return -ENOMEM;
 158                         }
 159                 }
 160         }
 161         inet_bind_hash(child, tb, port);
 162         spin_unlock(&head->lock);
 163
 164         return 0;
 165 }
 166 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 167
 168 static inline int compute_score(struct sock *sk, struct net *net,
 169                                 const unsigned short hnum, const __be32 daddr,
 170                                 const int dif)
 171 {
 172         int score = -1;
 173         struct inet_sock *inet = inet_sk(sk);
 174
 175         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 176                         !ipv6_only_sock(sk)) {
 177                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 178                 score = sk->sk_family == PF_INET ? 2 : 1;
 179                 if (rcv_saddr) {
 180                         if (rcv_saddr != daddr)
 181                                 return -1;
 182                         score += 4;
 183                 }
 184                 if (sk->sk_bound_dev_if) {
 185                         if (sk->sk_bound_dev_if != dif)
 186                                 return -1;
 187                         score += 4;
 188                 }
 189         }
 190         return score;
 191 }
 192
 193 /*
 194  * Don't inline this cruft. Here are some nice properties to exploit here. The
 195  * BSD API does not allow a listening sock to specify the remote port nor the
 196  * remote address for the connection. So always assume those are both
 197  * wildcarded during the search since they can never be otherwise.
 198  */
 199
 200
 201 struct sock *__inet_lookup_listener(struct net *net,
 202                                     struct inet_hashinfo *hashinfo,
 203                                     const __be32 saddr, __be16 sport,
 204                                     const __be32 daddr, const unsigned short hnum,
 205                                     const int dif)
 206 {
 207         struct sock *sk, *result;
 208         struct hlist_nulls_node *node;
 209         unsigned int hash = inet_lhashfn(net, hnum);
 210         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 211         int score, hiscore, matches = 0, reuseport = 0;
 212         u32 phash = 0;
 213
 214         rcu_read_lock();
 215 begin:
 216         result = NULL;
 217         hiscore = 0;
 218         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 219                 score = compute_score(sk, net, hnum, daddr, dif);
 220                 if (score > hiscore) {
 221                         result = sk;
 222                         hiscore = score;
 223                         reuseport = sk->sk_reuseport;
 224                         if (reuseport) {
 225                                 phash = inet_ehashfn(net, daddr, hnum,
 226                                                      saddr, sport);
 227                                 matches = 1;
 228                         }
 229                 } else if (score == hiscore && reuseport) {
 230                         matches++;
 231                         if (reciprocal_scale(phash, matches) == 0)
 232                                 result = sk;
 233                         phash = next_pseudo_random32(phash);
 234                 }
 235         }
 236         /*
 237          * if the nulls value we got at the end of this lookup is
 238          * not the expected one, we must restart lookup.
 239          * We probably met an item that was moved to another chain.
 240          */
 241         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 242                 goto begin;
 243         if (result) {
 244                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 245                         result = NULL;
 246                 else if (unlikely(compute_score(result, net, hnum, daddr,
 247                                   dif) < hiscore)) {
 248                         sock_put(result);
 249                         goto begin;
 250                 }
 251         }
 252         rcu_read_unlock();
 253         return result;
 254 }
 255 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 256
 257 /* All sockets share common refcount, but have different destructors */
 258 void sock_gen_put(struct sock *sk)
 259 {
 260         if (!atomic_dec_and_test(&sk->sk_refcnt))
 261                 return;
 262
 263         if (sk->sk_state == TCP_TIME_WAIT)
 264                 inet_twsk_free(inet_twsk(sk));
 265         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 266                 reqsk_free(inet_reqsk(sk));
 267         else
 268                 sk_free(sk);
 269 }
 270 EXPORT_SYMBOL_GPL(sock_gen_put);
 271
 272 struct sock *__inet_lookup_established(struct net *net,
 273                                   struct inet_hashinfo *hashinfo,
 274                                   const __be32 saddr, const __be16 sport,
 275                                   const __be32 daddr, const u16 hnum,
 276                                   const int dif)
 277 {
 278         INET_ADDR_COOKIE(acookie, saddr, daddr);
 279         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 280         struct sock *sk;
 281         const struct hlist_nulls_node *node;
 282         /* Optimize here for direct hit, only listening connections can
 283          * have wildcards anyways.
 284          */
 285         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 286         unsigned int slot = hash & hashinfo->ehash_mask;
 287         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 288
 289         rcu_read_lock();
 290 begin:
 291         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 292                 if (sk->sk_hash != hash)
 293                         continue;
 294                 if (likely(INET_MATCH(sk, net, acookie,
 295                                       saddr, daddr, ports, dif))) {
 296                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 297                                 goto out;
 298                         if (unlikely(!INET_MATCH(sk, net, acookie,
 299                                                  saddr, daddr, ports, dif))) {
 300                                 sock_gen_put(sk);
 301                                 goto begin;
 302                         }
 303                         goto found;
 304                 }
 305         }
 306         /*
 307          * if the nulls value we got at the end of this lookup is
 308          * not the expected one, we must restart lookup.
 309          * We probably met an item that was moved to another chain.
 310          */
 311         if (get_nulls_value(node) != slot)
 312                 goto begin;
 313 out:
 314         sk = NULL;
 315 found:
 316         rcu_read_unlock();
 317         return sk;
 318 }
 319 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 320
 321 /* called with local bh disabled */
 322 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 323                                     struct sock *sk, __u16 lport,
 324                                     struct inet_timewait_sock **twp)
 325 {
 326         struct inet_hashinfo *hinfo = death_row->hashinfo;
 327         struct inet_sock *inet = inet_sk(sk);
 328         __be32 daddr = inet->inet_rcv_saddr;
 329         __be32 saddr = inet->inet_daddr;
 330         int dif = sk->sk_bound_dev_if;
 331         INET_ADDR_COOKIE(acookie, saddr, daddr);
 332         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 333         struct net *net = sock_net(sk);
 334         unsigned int hash = inet_ehashfn(net, daddr, lport,
 335                                          saddr, inet->inet_dport);
 336         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 337         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 338         struct sock *sk2;
 339         const struct hlist_nulls_node *node;
 340         struct inet_timewait_sock *tw = NULL;
 341         int twrefcnt = 0;
 342
 343         spin_lock(lock);
 344
 345         sk_nulls_for_each(sk2, node, &head->chain) {
 346                 if (sk2->sk_hash != hash)
 347                         continue;
 348
 349                 if (likely(INET_MATCH(sk2, net, acookie,
 350                                          saddr, daddr, ports, dif))) {
 351                         if (sk2->sk_state == TCP_TIME_WAIT) {
 352                                 tw = inet_twsk(sk2);
 353                                 if (twsk_unique(sk, sk2, twp))
 354                                         break;
 355                         }
 356                         goto not_unique;
 357                 }
 358         }
 359
 360         /* Must record num and sport now. Otherwise we will see
 361          * in hash table socket with a funny identity.
 362          */
 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 inet_timewait_sock *tw = NULL;
 497
 498                 inet_get_local_port_range(net, &low, &high);
 499                 remaining = (high - low) + 1;
 500
 501                 local_bh_disable();
 502                 for (i = 1; i <= remaining; i++) {
 503                         port = low + (i + offset) % remaining;
 504                         if (inet_is_local_reserved_port(net, port))
 505                                 continue;
 506                         head = &hinfo->bhash[inet_bhashfn(net, port,
 507                                         hinfo->bhash_size)];
 508                         spin_lock(&head->lock);
 509
 510                         /* Does not bother with rcv_saddr checks,
 511                          * because the established check is already
 512                          * unique enough.
 513                          */
 514                         inet_bind_bucket_for_each(tb, &head->chain) {
 515                                 if (net_eq(ib_net(tb), net) &&
 516                                     tb->port == port) {
 517                                         if (tb->fastreuse >= 0 ||
 518                                             tb->fastreuseport >= 0)
 519                                                 goto next_port;
 520                                         WARN_ON(hlist_empty(&tb->owners));
 521                                         if (!check_established(death_row, sk,
 522                                                                 port, &tw))
 523                                                 goto ok;
 524                                         goto next_port;
 525                                 }
 526                         }
 527
 528                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 529                                         net, head, port);
 530                         if (!tb) {
 531                                 spin_unlock(&head->lock);
 532                                 break;
 533                         }
 534                         tb->fastreuse = -1;
 535                         tb->fastreuseport = -1;
 536                         goto ok;
 537
 538                 next_port:
 539                         spin_unlock(&head->lock);
 540                 }
 541                 local_bh_enable();
 542
 543                 return -EADDRNOTAVAIL;
 544
 545 ok:
 546                 hint += i;
 547
 548                 /* Head lock still held and bh's disabled */
 549                 inet_bind_hash(sk, tb, port);
 550                 if (sk_unhashed(sk)) {
 551                         inet_sk(sk)->inet_sport = htons(port);
 552                         twrefcnt += hash(sk, tw);
 553                 }
 554                 if (tw)
 555                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
 556                 spin_unlock(&head->lock);
 557
 558                 if (tw) {
 559                         inet_twsk_deschedule(tw, death_row);
 560                         while (twrefcnt) {
 561                                 twrefcnt--;
 562                                 inet_twsk_put(tw);
 563                         }
 564                 }
 565
 566                 ret = 0;
 567                 goto out;
 568         }
 569
 570         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 571         tb  = inet_csk(sk)->icsk_bind_hash;
 572         spin_lock_bh(&head->lock);
 573         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 574                 hash(sk, NULL);
 575                 spin_unlock_bh(&head->lock);
 576                 return 0;
 577         } else {
 578                 spin_unlock(&head->lock);
 579                 /* No definite answer... Walk to established hash table */
 580                 ret = check_established(death_row, sk, snum, NULL);
 581 out:
 582                 local_bh_enable();
 583                 return ret;
 584         }
 585 }
 586
 587 /*
 588  * Bind a port for a connect operation and hash it.
 589  */
 590 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 591                       struct sock *sk)
 592 {
 593         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 594                         __inet_check_established, __inet_hash_nolisten);
 595 }
 596 EXPORT_SYMBOL_GPL(inet_hash_connect);
 597
 598 void inet_hashinfo_init(struct inet_hashinfo *h)
 599 {
 600         int i;
 601
 602         atomic_set(&h->bsockets, 0);
 603         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 604                 spin_lock_init(&h->listening_hash[i].lock);
 605                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 606                                       i + LISTENING_NULLS_BASE);
 607                 }
 608 }
 609 EXPORT_SYMBOL_GPL(inet_hashinfo_init);