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, hold_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                 release_net(ib_net(tb));
  83                 kmem_cache_free(cachep, tb);
  84         }
  85 }
  86
  87 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  88                     const unsigned short snum)
  89 {
  90         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  91
  92         atomic_inc(&hashinfo->bsockets);
  93
  94         inet_sk(sk)->inet_num = snum;
  95         sk_add_bind_node(sk, &tb->owners);
  96         tb->num_owners++;
  97         inet_csk(sk)->icsk_bind_hash = tb;
  98 }
  99
 100 /*
 101  * Get rid of any references to a local port held by the given sock.
 102  */
 103 static void __inet_put_port(struct sock *sk)
 104 {
 105         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 106         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 107                         hashinfo->bhash_size);
 108         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 109         struct inet_bind_bucket *tb;
 110
 111         atomic_dec(&hashinfo->bsockets);
 112
 113         spin_lock(&head->lock);
 114         tb = inet_csk(sk)->icsk_bind_hash;
 115         __sk_del_bind_node(sk);
 116         tb->num_owners--;
 117         inet_csk(sk)->icsk_bind_hash = NULL;
 118         inet_sk(sk)->inet_num = 0;
 119         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 120         spin_unlock(&head->lock);
 121 }
 122
 123 void inet_put_port(struct sock *sk)
 124 {
 125         local_bh_disable();
 126         __inet_put_port(sk);
 127         local_bh_enable();
 128 }
 129 EXPORT_SYMBOL(inet_put_port);
 130
 131 int __inet_inherit_port(struct sock *sk, struct sock *child)
 132 {
 133         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 134         unsigned short port = inet_sk(child)->inet_num;
 135         const int bhash = inet_bhashfn(sock_net(sk), port,
 136                         table->bhash_size);
 137         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 138         struct inet_bind_bucket *tb;
 139
 140         spin_lock(&head->lock);
 141         tb = inet_csk(sk)->icsk_bind_hash;
 142         if (tb->port != port) {
 143                 /* NOTE: using tproxy and redirecting skbs to a proxy
 144                  * on a different listener port breaks the assumption
 145                  * that the listener socket's icsk_bind_hash is the same
 146                  * as that of the child socket. We have to look up or
 147                  * create a new bind bucket for the child here. */
 148                 inet_bind_bucket_for_each(tb, &head->chain) {
 149                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 150                             tb->port == port)
 151                                 break;
 152                 }
 153                 if (!tb) {
 154                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 155                                                      sock_net(sk), head, port);
 156                         if (!tb) {
 157                                 spin_unlock(&head->lock);
 158                                 return -ENOMEM;
 159                         }
 160                 }
 161         }
 162         inet_bind_hash(child, tb, port);
 163         spin_unlock(&head->lock);
 164
 165         return 0;
 166 }
 167 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 168
 169 static inline int compute_score(struct sock *sk, struct net *net,
 170                                 const unsigned short hnum, const __be32 daddr,
 171                                 const int dif)
 172 {
 173         int score = -1;
 174         struct inet_sock *inet = inet_sk(sk);
 175
 176         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 177                         !ipv6_only_sock(sk)) {
 178                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 179                 score = sk->sk_family == PF_INET ? 2 : 1;
 180                 if (rcv_saddr) {
 181                         if (rcv_saddr != daddr)
 182                                 return -1;
 183                         score += 4;
 184                 }
 185                 if (sk->sk_bound_dev_if) {
 186                         if (sk->sk_bound_dev_if != dif)
 187                                 return -1;
 188                         score += 4;
 189                 }
 190         }
 191         return score;
 192 }
 193
 194 /*
 195  * Don't inline this cruft. Here are some nice properties to exploit here. The
 196  * BSD API does not allow a listening sock to specify the remote port nor the
 197  * remote address for the connection. So always assume those are both
 198  * wildcarded during the search since they can never be otherwise.
 199  */
 200
 201
 202 struct sock *__inet_lookup_listener(struct net *net,
 203                                     struct inet_hashinfo *hashinfo,
 204                                     const __be32 saddr, __be16 sport,
 205                                     const __be32 daddr, const unsigned short hnum,
 206                                     const int dif)
 207 {
 208         struct sock *sk, *result;
 209         struct hlist_nulls_node *node;
 210         unsigned int hash = inet_lhashfn(net, hnum);
 211         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 212         int score, hiscore, matches = 0, reuseport = 0;
 213         u32 phash = 0;
 214
 215         rcu_read_lock();
 216 begin:
 217         result = NULL;
 218         hiscore = 0;
 219         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 220                 score = compute_score(sk, net, hnum, daddr, dif);
 221                 if (score > hiscore) {
 222                         result = sk;
 223                         hiscore = score;
 224                         reuseport = sk->sk_reuseport;
 225                         if (reuseport) {
 226                                 phash = inet_ehashfn(net, daddr, hnum,
 227                                                      saddr, sport);
 228                                 matches = 1;
 229                         }
 230                 } else if (score == hiscore && reuseport) {
 231                         matches++;
 232                         if (((u64)phash * matches) >> 32 == 0)
 233                                 result = sk;
 234                         phash = next_pseudo_random32(phash);
 235                 }
 236         }
 237         /*
 238          * if the nulls value we got at the end of this lookup is
 239          * not the expected one, we must restart lookup.
 240          * We probably met an item that was moved to another chain.
 241          */
 242         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 243                 goto begin;
 244         if (result) {
 245                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 246                         result = NULL;
 247                 else if (unlikely(compute_score(result, net, hnum, daddr,
 248                                   dif) < hiscore)) {
 249                         sock_put(result);
 250                         goto begin;
 251                 }
 252         }
 253         rcu_read_unlock();
 254         return result;
 255 }
 256 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 257
 258 /* All sockets share common refcount, but have different destructors */
 259 void sock_gen_put(struct sock *sk)
 260 {
 261         if (!atomic_dec_and_test(&sk->sk_refcnt))
 262                 return;
 263
 264         if (sk->sk_state == TCP_TIME_WAIT)
 265                 inet_twsk_free(inet_twsk(sk));
 266         else
 267                 sk_free(sk);
 268 }
 269 EXPORT_SYMBOL_GPL(sock_gen_put);
 270
 271 struct sock *__inet_lookup_established(struct net *net,
 272                                   struct inet_hashinfo *hashinfo,
 273                                   const __be32 saddr, const __be16 sport,
 274                                   const __be32 daddr, const u16 hnum,
 275                                   const int dif)
 276 {
 277         INET_ADDR_COOKIE(acookie, saddr, daddr);
 278         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 279         struct sock *sk;
 280         const struct hlist_nulls_node *node;
 281         /* Optimize here for direct hit, only listening connections can
 282          * have wildcards anyways.
 283          */
 284         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 285         unsigned int slot = hash & hashinfo->ehash_mask;
 286         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 287
 288         rcu_read_lock();
 289 begin:
 290         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 291                 if (sk->sk_hash != hash)
 292                         continue;
 293                 if (likely(INET_MATCH(sk, net, acookie,
 294                                       saddr, daddr, ports, dif))) {
 295                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 296                                 goto out;
 297                         if (unlikely(!INET_MATCH(sk, net, acookie,
 298                                                  saddr, daddr, ports, dif))) {
 299                                 sock_gen_put(sk);
 300                                 goto begin;
 301                         }
 302                         goto found;
 303                 }
 304         }
 305         /*
 306          * if the nulls value we got at the end of this lookup is
 307          * not the expected one, we must restart lookup.
 308          * We probably met an item that was moved to another chain.
 309          */
 310         if (get_nulls_value(node) != slot)
 311                 goto begin;
 312 out:
 313         sk = NULL;
 314 found:
 315         rcu_read_unlock();
 316         return sk;
 317 }
 318 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 319
 320 /* called with local bh disabled */
 321 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 322                                     struct sock *sk, __u16 lport,
 323                                     struct inet_timewait_sock **twp)
 324 {
 325         struct inet_hashinfo *hinfo = death_row->hashinfo;
 326         struct inet_sock *inet = inet_sk(sk);
 327         __be32 daddr = inet->inet_rcv_saddr;
 328         __be32 saddr = inet->inet_daddr;
 329         int dif = sk->sk_bound_dev_if;
 330         INET_ADDR_COOKIE(acookie, saddr, daddr);
 331         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 332         struct net *net = sock_net(sk);
 333         unsigned int hash = inet_ehashfn(net, daddr, lport,
 334                                          saddr, inet->inet_dport);
 335         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 336         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 337         struct sock *sk2;
 338         const struct hlist_nulls_node *node;
 339         struct inet_timewait_sock *tw = NULL;
 340         int twrefcnt = 0;
 341
 342         spin_lock(lock);
 343
 344         sk_nulls_for_each(sk2, node, &head->chain) {
 345                 if (sk2->sk_hash != hash)
 346                         continue;
 347
 348                 if (likely(INET_MATCH(sk2, net, acookie,
 349                                          saddr, daddr, ports, dif))) {
 350                         if (sk2->sk_state == TCP_TIME_WAIT) {
 351                                 tw = inet_twsk(sk2);
 352                                 if (twsk_unique(sk, sk2, twp))
 353                                         break;
 354                         }
 355                         goto not_unique;
 356                 }
 357         }
 358
 359         /* Must record num and sport now. Otherwise we will see
 360          * in hash table socket with a funny identity.
 361          */
 362         inet->inet_num = lport;
 363         inet->inet_sport = htons(lport);
 364         sk->sk_hash = hash;
 365         WARN_ON(!sk_unhashed(sk));
 366         __sk_nulls_add_node_rcu(sk, &head->chain);
 367         if (tw) {
 368                 twrefcnt = inet_twsk_unhash(tw);
 369                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 370         }
 371         spin_unlock(lock);
 372         if (twrefcnt)
 373                 inet_twsk_put(tw);
 374         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 375
 376         if (twp) {
 377                 *twp = tw;
 378         } else if (tw) {
 379                 /* Silly. Should hash-dance instead... */
 380                 inet_twsk_deschedule(tw, death_row);
 381
 382                 inet_twsk_put(tw);
 383         }
 384         return 0;
 385
 386 not_unique:
 387         spin_unlock(lock);
 388         return -EADDRNOTAVAIL;
 389 }
 390
 391 static inline u32 inet_sk_port_offset(const struct sock *sk)
 392 {
 393         const struct inet_sock *inet = inet_sk(sk);
 394         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 395                                           inet->inet_daddr,
 396                                           inet->inet_dport);
 397 }
 398
 399 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
 400 {
 401         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 402         struct hlist_nulls_head *list;
 403         spinlock_t *lock;
 404         struct inet_ehash_bucket *head;
 405         int twrefcnt = 0;
 406
 407         WARN_ON(!sk_unhashed(sk));
 408
 409         sk->sk_hash = inet_sk_ehashfn(sk);
 410         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 411         list = &head->chain;
 412         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 413
 414         spin_lock(lock);
 415         __sk_nulls_add_node_rcu(sk, list);
 416         if (tw) {
 417                 WARN_ON(sk->sk_hash != tw->tw_hash);
 418                 twrefcnt = inet_twsk_unhash(tw);
 419         }
 420         spin_unlock(lock);
 421         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 422         return twrefcnt;
 423 }
 424 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 425
 426 static void __inet_hash(struct sock *sk)
 427 {
 428         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 429         struct inet_listen_hashbucket *ilb;
 430
 431         if (sk->sk_state != TCP_LISTEN) {
 432                 __inet_hash_nolisten(sk, NULL);
 433                 return;
 434         }
 435
 436         WARN_ON(!sk_unhashed(sk));
 437         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 438
 439         spin_lock(&ilb->lock);
 440         __sk_nulls_add_node_rcu(sk, &ilb->head);
 441         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 442         spin_unlock(&ilb->lock);
 443 }
 444
 445 void inet_hash(struct sock *sk)
 446 {
 447         if (sk->sk_state != TCP_CLOSE) {
 448                 local_bh_disable();
 449                 __inet_hash(sk);
 450                 local_bh_enable();
 451         }
 452 }
 453 EXPORT_SYMBOL_GPL(inet_hash);
 454
 455 void inet_unhash(struct sock *sk)
 456 {
 457         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 458         spinlock_t *lock;
 459         int done;
 460
 461         if (sk_unhashed(sk))
 462                 return;
 463
 464         if (sk->sk_state == TCP_LISTEN)
 465                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 466         else
 467                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 468
 469         spin_lock_bh(lock);
 470         done = __sk_nulls_del_node_init_rcu(sk);
 471         if (done)
 472                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 473         spin_unlock_bh(lock);
 474 }
 475 EXPORT_SYMBOL_GPL(inet_unhash);
 476
 477 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 478                 struct sock *sk, u32 port_offset,
 479                 int (*check_established)(struct inet_timewait_death_row *,
 480                         struct sock *, __u16, struct inet_timewait_sock **),
 481                 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
 482 {
 483         struct inet_hashinfo *hinfo = death_row->hashinfo;
 484         const unsigned short snum = inet_sk(sk)->inet_num;
 485         struct inet_bind_hashbucket *head;
 486         struct inet_bind_bucket *tb;
 487         int ret;
 488         struct net *net = sock_net(sk);
 489         int twrefcnt = 1;
 490
 491         if (!snum) {
 492                 int i, remaining, low, high, port;
 493                 static u32 hint;
 494                 u32 offset = hint + port_offset;
 495                 struct inet_timewait_sock *tw = NULL;
 496
 497                 inet_get_local_port_range(net, &low, &high);
 498                 remaining = (high - low) + 1;
 499
 500                 local_bh_disable();
 501                 for (i = 1; i <= remaining; i++) {
 502                         port = low + (i + offset) % remaining;
 503                         if (inet_is_local_reserved_port(net, port))
 504                                 continue;
 505                         head = &hinfo->bhash[inet_bhashfn(net, port,
 506                                         hinfo->bhash_size)];
 507                         spin_lock(&head->lock);
 508
 509                         /* Does not bother with rcv_saddr checks,
 510                          * because the established check is already
 511                          * unique enough.
 512                          */
 513                         inet_bind_bucket_for_each(tb, &head->chain) {
 514                                 if (net_eq(ib_net(tb), net) &&
 515                                     tb->port == port) {
 516                                         if (tb->fastreuse >= 0 ||
 517                                             tb->fastreuseport >= 0)
 518                                                 goto next_port;
 519                                         WARN_ON(hlist_empty(&tb->owners));
 520                                         if (!check_established(death_row, sk,
 521                                                                 port, &tw))
 522                                                 goto ok;
 523                                         goto next_port;
 524                                 }
 525                         }
 526
 527                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 528                                         net, head, port);
 529                         if (!tb) {
 530                                 spin_unlock(&head->lock);
 531                                 break;
 532                         }
 533                         tb->fastreuse = -1;
 534                         tb->fastreuseport = -1;
 535                         goto ok;
 536
 537                 next_port:
 538                         spin_unlock(&head->lock);
 539                 }
 540                 local_bh_enable();
 541
 542                 return -EADDRNOTAVAIL;
 543
 544 ok:
 545                 hint += i;
 546
 547                 /* Head lock still held and bh's disabled */
 548                 inet_bind_hash(sk, tb, port);
 549                 if (sk_unhashed(sk)) {
 550                         inet_sk(sk)->inet_sport = htons(port);
 551                         twrefcnt += hash(sk, tw);
 552                 }
 553                 if (tw)
 554                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
 555                 spin_unlock(&head->lock);
 556
 557                 if (tw) {
 558                         inet_twsk_deschedule(tw, death_row);
 559                         while (twrefcnt) {
 560                                 twrefcnt--;
 561                                 inet_twsk_put(tw);
 562                         }
 563                 }
 564
 565                 ret = 0;
 566                 goto out;
 567         }
 568
 569         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 570         tb  = inet_csk(sk)->icsk_bind_hash;
 571         spin_lock_bh(&head->lock);
 572         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 573                 hash(sk, NULL);
 574                 spin_unlock_bh(&head->lock);
 575                 return 0;
 576         } else {
 577                 spin_unlock(&head->lock);
 578                 /* No definite answer... Walk to established hash table */
 579                 ret = check_established(death_row, sk, snum, NULL);
 580 out:
 581                 local_bh_enable();
 582                 return ret;
 583         }
 584 }
 585
 586 /*
 587  * Bind a port for a connect operation and hash it.
 588  */
 589 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 590                       struct sock *sk)
 591 {
 592         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 593                         __inet_check_established, __inet_hash_nolisten);
 594 }
 595 EXPORT_SYMBOL_GPL(inet_hash_connect);
 596
 597 void inet_hashinfo_init(struct inet_hashinfo *h)
 598 {
 599         int i;
 600
 601         atomic_set(&h->bsockets, 0);
 602         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 603                 spin_lock_init(&h->listening_hash[i].lock);
 604                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 605                                       i + LISTENING_NULLS_BASE);
 606                 }
 607 }
 608 EXPORT_SYMBOL_GPL(inet_hashinfo_init);