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 u32 inet_ehashfn(const 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 /* This function handles inet_sock, but also timewait and request sockets
  40  * for IPv4/IPv6.
  41  */
  42 u32 sk_ehashfn(const struct sock *sk)
  43 {
  44 #if IS_ENABLED(CONFIG_IPV6)
  45         if (sk->sk_family == AF_INET6 &&
  46             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  47                 return inet6_ehashfn(sock_net(sk),
  48                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  49                                      &sk->sk_v6_daddr, sk->sk_dport);
  50 #endif
  51         return inet_ehashfn(sock_net(sk),
  52                             sk->sk_rcv_saddr, sk->sk_num,
  53                             sk->sk_daddr, sk->sk_dport);
  54 }
  55
  56 /*
  57  * Allocate and initialize a new local port bind bucket.
  58  * The bindhash mutex for snum's hash chain must be held here.
  59  */
  60 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  61                                                  struct net *net,
  62                                                  struct inet_bind_hashbucket *head,
  63                                                  const unsigned short snum)
  64 {
  65         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  66
  67         if (tb) {
  68                 write_pnet(&tb->ib_net, net);
  69                 tb->port      = snum;
  70                 tb->fastreuse = 0;
  71                 tb->fastreuseport = 0;
  72                 tb->num_owners = 0;
  73                 INIT_HLIST_HEAD(&tb->owners);
  74                 hlist_add_head(&tb->node, &head->chain);
  75         }
  76         return tb;
  77 }
  78
  79 /*
  80  * Caller must hold hashbucket lock for this tb with local BH disabled
  81  */
  82 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  83 {
  84         if (hlist_empty(&tb->owners)) {
  85                 __hlist_del(&tb->node);
  86                 kmem_cache_free(cachep, tb);
  87         }
  88 }
  89
  90 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  91                     const unsigned short snum)
  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         spin_lock(&head->lock);
 111         tb = inet_csk(sk)->icsk_bind_hash;
 112         __sk_del_bind_node(sk);
 113         tb->num_owners--;
 114         inet_csk(sk)->icsk_bind_hash = NULL;
 115         inet_sk(sk)->inet_num = 0;
 116         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 117         spin_unlock(&head->lock);
 118 }
 119
 120 void inet_put_port(struct sock *sk)
 121 {
 122         local_bh_disable();
 123         __inet_put_port(sk);
 124         local_bh_enable();
 125 }
 126 EXPORT_SYMBOL(inet_put_port);
 127
 128 int __inet_inherit_port(struct sock *sk, struct sock *child)
 129 {
 130         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 131         unsigned short port = inet_sk(child)->inet_num;
 132         const int bhash = inet_bhashfn(sock_net(sk), port,
 133                         table->bhash_size);
 134         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 135         struct inet_bind_bucket *tb;
 136
 137         spin_lock(&head->lock);
 138         tb = inet_csk(sk)->icsk_bind_hash;
 139         if (tb->port != port) {
 140                 /* NOTE: using tproxy and redirecting skbs to a proxy
 141                  * on a different listener port breaks the assumption
 142                  * that the listener socket's icsk_bind_hash is the same
 143                  * as that of the child socket. We have to look up or
 144                  * create a new bind bucket for the child here. */
 145                 inet_bind_bucket_for_each(tb, &head->chain) {
 146                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 147                             tb->port == port)
 148                                 break;
 149                 }
 150                 if (!tb) {
 151                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 152                                                      sock_net(sk), head, port);
 153                         if (!tb) {
 154                                 spin_unlock(&head->lock);
 155                                 return -ENOMEM;
 156                         }
 157                 }
 158         }
 159         inet_bind_hash(child, tb, port);
 160         spin_unlock(&head->lock);
 161
 162         return 0;
 163 }
 164 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 165
 166 static inline int compute_score(struct sock *sk, struct net *net,
 167                                 const unsigned short hnum, const __be32 daddr,
 168                                 const int dif)
 169 {
 170         int score = -1;
 171         struct inet_sock *inet = inet_sk(sk);
 172
 173         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 174                         !ipv6_only_sock(sk)) {
 175                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 176                 score = sk->sk_family == PF_INET ? 2 : 1;
 177                 if (rcv_saddr) {
 178                         if (rcv_saddr != daddr)
 179                                 return -1;
 180                         score += 4;
 181                 }
 182                 if (sk->sk_bound_dev_if) {
 183                         if (sk->sk_bound_dev_if != dif)
 184                                 return -1;
 185                         score += 4;
 186                 }
 187         }
 188         return score;
 189 }
 190
 191 /*
 192  * Don't inline this cruft. Here are some nice properties to exploit here. The
 193  * BSD API does not allow a listening sock to specify the remote port nor the
 194  * remote address for the connection. So always assume those are both
 195  * wildcarded during the search since they can never be otherwise.
 196  */
 197
 198
 199 struct sock *__inet_lookup_listener(struct net *net,
 200                                     struct inet_hashinfo *hashinfo,
 201                                     const __be32 saddr, __be16 sport,
 202                                     const __be32 daddr, const unsigned short hnum,
 203                                     const int dif)
 204 {
 205         struct sock *sk, *result;
 206         struct hlist_nulls_node *node;
 207         unsigned int hash = inet_lhashfn(net, hnum);
 208         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 209         int score, hiscore, matches = 0, reuseport = 0;
 210         u32 phash = 0;
 211
 212         rcu_read_lock();
 213 begin:
 214         result = NULL;
 215         hiscore = 0;
 216         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 217                 score = compute_score(sk, net, hnum, daddr, dif);
 218                 if (score > hiscore) {
 219                         result = sk;
 220                         hiscore = score;
 221                         reuseport = sk->sk_reuseport;
 222                         if (reuseport) {
 223                                 phash = inet_ehashfn(net, daddr, hnum,
 224                                                      saddr, sport);
 225                                 matches = 1;
 226                         }
 227                 } else if (score == hiscore && reuseport) {
 228                         matches++;
 229                         if (reciprocal_scale(phash, matches) == 0)
 230                                 result = sk;
 231                         phash = next_pseudo_random32(phash);
 232                 }
 233         }
 234         /*
 235          * if the nulls value we got at the end of this lookup is
 236          * not the expected one, we must restart lookup.
 237          * We probably met an item that was moved to another chain.
 238          */
 239         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 240                 goto begin;
 241         if (result) {
 242                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 243                         result = NULL;
 244                 else if (unlikely(compute_score(result, net, hnum, daddr,
 245                                   dif) < hiscore)) {
 246                         sock_put(result);
 247                         goto begin;
 248                 }
 249         }
 250         rcu_read_unlock();
 251         return result;
 252 }
 253 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 254
 255 /* All sockets share common refcount, but have different destructors */
 256 void sock_gen_put(struct sock *sk)
 257 {
 258         if (!atomic_dec_and_test(&sk->sk_refcnt))
 259                 return;
 260
 261         if (sk->sk_state == TCP_TIME_WAIT)
 262                 inet_twsk_free(inet_twsk(sk));
 263         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 264                 reqsk_free(inet_reqsk(sk));
 265         else
 266                 sk_free(sk);
 267 }
 268 EXPORT_SYMBOL_GPL(sock_gen_put);
 269
 270 void sock_edemux(struct sk_buff *skb)
 271 {
 272         sock_gen_put(skb->sk);
 273 }
 274 EXPORT_SYMBOL(sock_edemux);
 275
 276 struct sock *__inet_lookup_established(struct net *net,
 277                                   struct inet_hashinfo *hashinfo,
 278                                   const __be32 saddr, const __be16 sport,
 279                                   const __be32 daddr, const u16 hnum,
 280                                   const int dif)
 281 {
 282         INET_ADDR_COOKIE(acookie, saddr, daddr);
 283         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 284         struct sock *sk;
 285         const struct hlist_nulls_node *node;
 286         /* Optimize here for direct hit, only listening connections can
 287          * have wildcards anyways.
 288          */
 289         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 290         unsigned int slot = hash & hashinfo->ehash_mask;
 291         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 292
 293         rcu_read_lock();
 294 begin:
 295         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 296                 if (sk->sk_hash != hash)
 297                         continue;
 298                 if (likely(INET_MATCH(sk, net, acookie,
 299                                       saddr, daddr, ports, dif))) {
 300                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 301                                 goto out;
 302                         if (unlikely(!INET_MATCH(sk, net, acookie,
 303                                                  saddr, daddr, ports, dif))) {
 304                                 sock_gen_put(sk);
 305                                 goto begin;
 306                         }
 307                         goto found;
 308                 }
 309         }
 310         /*
 311          * if the nulls value we got at the end of this lookup is
 312          * not the expected one, we must restart lookup.
 313          * We probably met an item that was moved to another chain.
 314          */
 315         if (get_nulls_value(node) != slot)
 316                 goto begin;
 317 out:
 318         sk = NULL;
 319 found:
 320         rcu_read_unlock();
 321         return sk;
 322 }
 323 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 324
 325 /* called with local bh disabled */
 326 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 327                                     struct sock *sk, __u16 lport,
 328                                     struct inet_timewait_sock **twp)
 329 {
 330         struct inet_hashinfo *hinfo = death_row->hashinfo;
 331         struct inet_sock *inet = inet_sk(sk);
 332         __be32 daddr = inet->inet_rcv_saddr;
 333         __be32 saddr = inet->inet_daddr;
 334         int dif = sk->sk_bound_dev_if;
 335         INET_ADDR_COOKIE(acookie, saddr, daddr);
 336         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 337         struct net *net = sock_net(sk);
 338         unsigned int hash = inet_ehashfn(net, daddr, lport,
 339                                          saddr, inet->inet_dport);
 340         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 341         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 342         struct sock *sk2;
 343         const struct hlist_nulls_node *node;
 344         struct inet_timewait_sock *tw = NULL;
 345         int twrefcnt = 0;
 346
 347         spin_lock(lock);
 348
 349         sk_nulls_for_each(sk2, node, &head->chain) {
 350                 if (sk2->sk_hash != hash)
 351                         continue;
 352
 353                 if (likely(INET_MATCH(sk2, net, acookie,
 354                                          saddr, daddr, ports, dif))) {
 355                         if (sk2->sk_state == TCP_TIME_WAIT) {
 356                                 tw = inet_twsk(sk2);
 357                                 if (twsk_unique(sk, sk2, twp))
 358                                         break;
 359                         }
 360                         goto not_unique;
 361                 }
 362         }
 363
 364         /* Must record num and sport now. Otherwise we will see
 365          * in hash table socket with a funny identity.
 366          */
 367         inet->inet_num = lport;
 368         inet->inet_sport = htons(lport);
 369         sk->sk_hash = hash;
 370         WARN_ON(!sk_unhashed(sk));
 371         __sk_nulls_add_node_rcu(sk, &head->chain);
 372         if (tw) {
 373                 twrefcnt = inet_twsk_unhash(tw);
 374                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 375         }
 376         spin_unlock(lock);
 377         if (twrefcnt)
 378                 inet_twsk_put(tw);
 379         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 380
 381         if (twp) {
 382                 *twp = tw;
 383         } else if (tw) {
 384                 /* Silly. Should hash-dance instead... */
 385                 inet_twsk_deschedule(tw);
 386
 387                 inet_twsk_put(tw);
 388         }
 389         return 0;
 390
 391 not_unique:
 392         spin_unlock(lock);
 393         return -EADDRNOTAVAIL;
 394 }
 395
 396 static inline u32 inet_sk_port_offset(const struct sock *sk)
 397 {
 398         const struct inet_sock *inet = inet_sk(sk);
 399         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 400                                           inet->inet_daddr,
 401                                           inet->inet_dport);
 402 }
 403
 404 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
 405 {
 406         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 407         struct hlist_nulls_head *list;
 408         struct inet_ehash_bucket *head;
 409         spinlock_t *lock;
 410         int twrefcnt = 0;
 411
 412         WARN_ON(!sk_unhashed(sk));
 413
 414         sk->sk_hash = sk_ehashfn(sk);
 415         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 416         list = &head->chain;
 417         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 418
 419         spin_lock(lock);
 420         __sk_nulls_add_node_rcu(sk, list);
 421         if (tw) {
 422                 WARN_ON(sk->sk_hash != tw->tw_hash);
 423                 twrefcnt = inet_twsk_unhash(tw);
 424         }
 425         spin_unlock(lock);
 426         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 427         return twrefcnt;
 428 }
 429 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 430
 431 int __inet_hash(struct sock *sk, struct inet_timewait_sock *tw)
 432 {
 433         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 434         struct inet_listen_hashbucket *ilb;
 435
 436         if (sk->sk_state != TCP_LISTEN)
 437                 return __inet_hash_nolisten(sk, tw);
 438
 439         WARN_ON(!sk_unhashed(sk));
 440         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 441
 442         spin_lock(&ilb->lock);
 443         __sk_nulls_add_node_rcu(sk, &ilb->head);
 444         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 445         spin_unlock(&ilb->lock);
 446         return 0;
 447 }
 448 EXPORT_SYMBOL(__inet_hash);
 449
 450 void inet_hash(struct sock *sk)
 451 {
 452         if (sk->sk_state != TCP_CLOSE) {
 453                 local_bh_disable();
 454                 __inet_hash(sk, NULL);
 455                 local_bh_enable();
 456         }
 457 }
 458 EXPORT_SYMBOL_GPL(inet_hash);
 459
 460 void inet_unhash(struct sock *sk)
 461 {
 462         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 463         spinlock_t *lock;
 464         int done;
 465
 466         if (sk_unhashed(sk))
 467                 return;
 468
 469         if (sk->sk_state == TCP_LISTEN)
 470                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 471         else
 472                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 473
 474         spin_lock_bh(lock);
 475         done = __sk_nulls_del_node_init_rcu(sk);
 476         if (done)
 477                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 478         spin_unlock_bh(lock);
 479 }
 480 EXPORT_SYMBOL_GPL(inet_unhash);
 481
 482 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 483                 struct sock *sk, u32 port_offset,
 484                 int (*check_established)(struct inet_timewait_death_row *,
 485                         struct sock *, __u16, struct inet_timewait_sock **))
 486 {
 487         struct inet_hashinfo *hinfo = death_row->hashinfo;
 488         const unsigned short snum = inet_sk(sk)->inet_num;
 489         struct inet_bind_hashbucket *head;
 490         struct inet_bind_bucket *tb;
 491         int ret;
 492         struct net *net = sock_net(sk);
 493         int twrefcnt = 1;
 494
 495         if (!snum) {
 496                 int i, remaining, low, high, port;
 497                 static u32 hint;
 498                 u32 offset = hint + port_offset;
 499                 struct inet_timewait_sock *tw = NULL;
 500
 501                 inet_get_local_port_range(net, &low, &high);
 502                 remaining = (high - low) + 1;
 503
 504                 local_bh_disable();
 505                 for (i = 1; i <= remaining; i++) {
 506                         port = low + (i + offset) % remaining;
 507                         if (inet_is_local_reserved_port(net, port))
 508                                 continue;
 509                         head = &hinfo->bhash[inet_bhashfn(net, port,
 510                                         hinfo->bhash_size)];
 511                         spin_lock(&head->lock);
 512
 513                         /* Does not bother with rcv_saddr checks,
 514                          * because the established check is already
 515                          * unique enough.
 516                          */
 517                         inet_bind_bucket_for_each(tb, &head->chain) {
 518                                 if (net_eq(ib_net(tb), net) &&
 519                                     tb->port == port) {
 520                                         if (tb->fastreuse >= 0 ||
 521                                             tb->fastreuseport >= 0)
 522                                                 goto next_port;
 523                                         WARN_ON(hlist_empty(&tb->owners));
 524                                         if (!check_established(death_row, sk,
 525                                                                 port, &tw))
 526                                                 goto ok;
 527                                         goto next_port;
 528                                 }
 529                         }
 530
 531                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 532                                         net, head, port);
 533                         if (!tb) {
 534                                 spin_unlock(&head->lock);
 535                                 break;
 536                         }
 537                         tb->fastreuse = -1;
 538                         tb->fastreuseport = -1;
 539                         goto ok;
 540
 541                 next_port:
 542                         spin_unlock(&head->lock);
 543                 }
 544                 local_bh_enable();
 545
 546                 return -EADDRNOTAVAIL;
 547
 548 ok:
 549                 hint += i;
 550
 551                 /* Head lock still held and bh's disabled */
 552                 inet_bind_hash(sk, tb, port);
 553                 if (sk_unhashed(sk)) {
 554                         inet_sk(sk)->inet_sport = htons(port);
 555                         twrefcnt += __inet_hash_nolisten(sk, tw);
 556                 }
 557                 if (tw)
 558                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
 559                 spin_unlock(&head->lock);
 560
 561                 if (tw) {
 562                         inet_twsk_deschedule(tw);
 563                         while (twrefcnt) {
 564                                 twrefcnt--;
 565                                 inet_twsk_put(tw);
 566                         }
 567                 }
 568
 569                 ret = 0;
 570                 goto out;
 571         }
 572
 573         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 574         tb  = inet_csk(sk)->icsk_bind_hash;
 575         spin_lock_bh(&head->lock);
 576         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 577                 __inet_hash_nolisten(sk, NULL);
 578                 spin_unlock_bh(&head->lock);
 579                 return 0;
 580         } else {
 581                 spin_unlock(&head->lock);
 582                 /* No definite answer... Walk to established hash table */
 583                 ret = check_established(death_row, sk, snum, NULL);
 584 out:
 585                 local_bh_enable();
 586                 return ret;
 587         }
 588 }
 589
 590 /*
 591  * Bind a port for a connect operation and hash it.
 592  */
 593 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 594                       struct sock *sk)
 595 {
 596         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 597                                    __inet_check_established);
 598 }
 599 EXPORT_SYMBOL_GPL(inet_hash_connect);
 600
 601 void inet_hashinfo_init(struct inet_hashinfo *h)
 602 {
 603         int i;
 604
 605         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 606                 spin_lock_init(&h->listening_hash[i].lock);
 607                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 608                                       i + LISTENING_NULLS_BASE);
 609                 }
 610 }
 611 EXPORT_SYMBOL_GPL(inet_hashinfo_init);