kernel/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         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  94
  95         atomic_inc(&hashinfo->bsockets);
  96
  97         inet_sk(sk)->inet_num = snum;
  98         sk_add_bind_node(sk, &tb->owners);
  99         tb->num_owners++;
 100         inet_csk(sk)->icsk_bind_hash = tb;
 101 }
 102
 103 /*
 104  * Get rid of any references to a local port held by the given sock.
 105  */
 106 static void __inet_put_port(struct sock *sk)
 107 {
 108         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 109         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 110                         hashinfo->bhash_size);
 111         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 112         struct inet_bind_bucket *tb;
 113
 114         atomic_dec(&hashinfo->bsockets);
 115
 116         spin_lock(&head->lock);
 117         tb = inet_csk(sk)->icsk_bind_hash;
 118         __sk_del_bind_node(sk);
 119         tb->num_owners--;
 120         inet_csk(sk)->icsk_bind_hash = NULL;
 121         inet_sk(sk)->inet_num = 0;
 122         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 123         spin_unlock(&head->lock);
 124 }
 125
 126 void inet_put_port(struct sock *sk)
 127 {
 128         local_bh_disable();
 129         __inet_put_port(sk);
 130         local_bh_enable();
 131 }
 132 EXPORT_SYMBOL(inet_put_port);
 133
 134 int __inet_inherit_port(struct sock *sk, struct sock *child)
 135 {
 136         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 137         unsigned short port = inet_sk(child)->inet_num;
 138         const int bhash = inet_bhashfn(sock_net(sk), port,
 139                         table->bhash_size);
 140         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 141         struct inet_bind_bucket *tb;
 142
 143         spin_lock(&head->lock);
 144         tb = inet_csk(sk)->icsk_bind_hash;
 145         if (tb->port != port) {
 146                 /* NOTE: using tproxy and redirecting skbs to a proxy
 147                  * on a different listener port breaks the assumption
 148                  * that the listener socket's icsk_bind_hash is the same
 149                  * as that of the child socket. We have to look up or
 150                  * create a new bind bucket for the child here. */
 151                 inet_bind_bucket_for_each(tb, &head->chain) {
 152                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 153                             tb->port == port)
 154                                 break;
 155                 }
 156                 if (!tb) {
 157                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 158                                                      sock_net(sk), head, port);
 159                         if (!tb) {
 160                                 spin_unlock(&head->lock);
 161                                 return -ENOMEM;
 162                         }
 163                 }
 164         }
 165         inet_bind_hash(child, tb, port);
 166         spin_unlock(&head->lock);
 167
 168         return 0;
 169 }
 170 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 171
 172 static inline int compute_score(struct sock *sk, struct net *net,
 173                                 const unsigned short hnum, const __be32 daddr,
 174                                 const int dif)
 175 {
 176         int score = -1;
 177         struct inet_sock *inet = inet_sk(sk);
 178
 179         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 180                         !ipv6_only_sock(sk)) {
 181                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 182                 score = sk->sk_family == PF_INET ? 2 : 1;
 183                 if (rcv_saddr) {
 184                         if (rcv_saddr != daddr)
 185                                 return -1;
 186                         score += 4;
 187                 }
 188                 if (sk->sk_bound_dev_if) {
 189                         if (sk->sk_bound_dev_if != dif)
 190                                 return -1;
 191                         score += 4;
 192                 }
 193         }
 194         return score;
 195 }
 196
 197 /*
 198  * Don't inline this cruft. Here are some nice properties to exploit here. The
 199  * BSD API does not allow a listening sock to specify the remote port nor the
 200  * remote address for the connection. So always assume those are both
 201  * wildcarded during the search since they can never be otherwise.
 202  */
 203
 204
 205 struct sock *__inet_lookup_listener(struct net *net,
 206                                     struct inet_hashinfo *hashinfo,
 207                                     const __be32 saddr, __be16 sport,
 208                                     const __be32 daddr, const unsigned short hnum,
 209                                     const int dif)
 210 {
 211         struct sock *sk, *result;
 212         struct hlist_nulls_node *node;
 213         unsigned int hash = inet_lhashfn(net, hnum);
 214         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 215         int score, hiscore, matches = 0, reuseport = 0;
 216         u32 phash = 0;
 217
 218         rcu_read_lock();
 219 begin:
 220         result = NULL;
 221         hiscore = 0;
 222         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 223                 score = compute_score(sk, net, hnum, daddr, dif);
 224                 if (score > hiscore) {
 225                         result = sk;
 226                         hiscore = score;
 227                         reuseport = sk->sk_reuseport;
 228                         if (reuseport) {
 229                                 phash = inet_ehashfn(net, daddr, hnum,
 230                                                      saddr, sport);
 231                                 matches = 1;
 232                         }
 233                 } else if (score == hiscore && reuseport) {
 234                         matches++;
 235                         if (reciprocal_scale(phash, matches) == 0)
 236                                 result = sk;
 237                         phash = next_pseudo_random32(phash);
 238                 }
 239         }
 240         /*
 241          * if the nulls value we got at the end of this lookup is
 242          * not the expected one, we must restart lookup.
 243          * We probably met an item that was moved to another chain.
 244          */
 245         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 246                 goto begin;
 247         if (result) {
 248                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 249                         result = NULL;
 250                 else if (unlikely(compute_score(result, net, hnum, daddr,
 251                                   dif) < hiscore)) {
 252                         sock_put(result);
 253                         goto begin;
 254                 }
 255         }
 256         rcu_read_unlock();
 257         return result;
 258 }
 259 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 260
 261 /* All sockets share common refcount, but have different destructors */
 262 void sock_gen_put(struct sock *sk)
 263 {
 264         if (!atomic_dec_and_test(&sk->sk_refcnt))
 265                 return;
 266
 267         if (sk->sk_state == TCP_TIME_WAIT)
 268                 inet_twsk_free(inet_twsk(sk));
 269         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 270                 reqsk_free(inet_reqsk(sk));
 271         else
 272                 sk_free(sk);
 273 }
 274 EXPORT_SYMBOL_GPL(sock_gen_put);
 275
 276 void sock_edemux(struct sk_buff *skb)
 277 {
 278         sock_gen_put(skb->sk);
 279 }
 280 EXPORT_SYMBOL(sock_edemux);
 281
 282 struct sock *__inet_lookup_established(struct net *net,
 283                                   struct inet_hashinfo *hashinfo,
 284                                   const __be32 saddr, const __be16 sport,
 285                                   const __be32 daddr, const u16 hnum,
 286                                   const int dif)
 287 {
 288         INET_ADDR_COOKIE(acookie, saddr, daddr);
 289         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 290         struct sock *sk;
 291         const struct hlist_nulls_node *node;
 292         /* Optimize here for direct hit, only listening connections can
 293          * have wildcards anyways.
 294          */
 295         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 296         unsigned int slot = hash & hashinfo->ehash_mask;
 297         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 298
 299         rcu_read_lock();
 300 begin:
 301         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 302                 if (sk->sk_hash != hash)
 303                         continue;
 304                 if (likely(INET_MATCH(sk, net, acookie,
 305                                       saddr, daddr, ports, dif))) {
 306                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 307                                 goto out;
 308                         if (unlikely(!INET_MATCH(sk, net, acookie,
 309                                                  saddr, daddr, ports, dif))) {
 310                                 sock_gen_put(sk);
 311                                 goto begin;
 312                         }
 313                         goto found;
 314                 }
 315         }
 316         /*
 317          * if the nulls value we got at the end of this lookup is
 318          * not the expected one, we must restart lookup.
 319          * We probably met an item that was moved to another chain.
 320          */
 321         if (get_nulls_value(node) != slot)
 322                 goto begin;
 323 out:
 324         sk = NULL;
 325 found:
 326         rcu_read_unlock();
 327         return sk;
 328 }
 329 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 330
 331 /* called with local bh disabled */
 332 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 333                                     struct sock *sk, __u16 lport,
 334                                     struct inet_timewait_sock **twp)
 335 {
 336         struct inet_hashinfo *hinfo = death_row->hashinfo;
 337         struct inet_sock *inet = inet_sk(sk);
 338         __be32 daddr = inet->inet_rcv_saddr;
 339         __be32 saddr = inet->inet_daddr;
 340         int dif = sk->sk_bound_dev_if;
 341         INET_ADDR_COOKIE(acookie, saddr, daddr);
 342         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 343         struct net *net = sock_net(sk);
 344         unsigned int hash = inet_ehashfn(net, daddr, lport,
 345                                          saddr, inet->inet_dport);
 346         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 347         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 348         struct sock *sk2;
 349         const struct hlist_nulls_node *node;
 350         struct inet_timewait_sock *tw = NULL;
 351         int twrefcnt = 0;
 352
 353         spin_lock(lock);
 354
 355         sk_nulls_for_each(sk2, node, &head->chain) {
 356                 if (sk2->sk_hash != hash)
 357                         continue;
 358
 359                 if (likely(INET_MATCH(sk2, net, acookie,
 360                                          saddr, daddr, ports, dif))) {
 361                         if (sk2->sk_state == TCP_TIME_WAIT) {
 362                                 tw = inet_twsk(sk2);
 363                                 if (twsk_unique(sk, sk2, twp))
 364                                         break;
 365                         }
 366                         goto not_unique;
 367                 }
 368         }
 369
 370         /* Must record num and sport now. Otherwise we will see
 371          * in hash table socket with a funny identity.
 372          */
 373         inet->inet_num = lport;
 374         inet->inet_sport = htons(lport);
 375         sk->sk_hash = hash;
 376         WARN_ON(!sk_unhashed(sk));
 377         __sk_nulls_add_node_rcu(sk, &head->chain);
 378         if (tw) {
 379                 twrefcnt = inet_twsk_unhash(tw);
 380                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 381         }
 382         spin_unlock(lock);
 383         if (twrefcnt)
 384                 inet_twsk_put(tw);
 385         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 386
 387         if (twp) {
 388                 *twp = tw;
 389         } else if (tw) {
 390                 /* Silly. Should hash-dance instead... */
 391                 inet_twsk_deschedule(tw);
 392
 393                 inet_twsk_put(tw);
 394         }
 395         return 0;
 396
 397 not_unique:
 398         spin_unlock(lock);
 399         return -EADDRNOTAVAIL;
 400 }
 401
 402 static inline u32 inet_sk_port_offset(const struct sock *sk)
 403 {
 404         const struct inet_sock *inet = inet_sk(sk);
 405         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 406                                           inet->inet_daddr,
 407                                           inet->inet_dport);
 408 }
 409
 410 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
 411 {
 412         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 413         struct hlist_nulls_head *list;
 414         struct inet_ehash_bucket *head;
 415         spinlock_t *lock;
 416         int twrefcnt = 0;
 417
 418         WARN_ON(!sk_unhashed(sk));
 419
 420         sk->sk_hash = sk_ehashfn(sk);
 421         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 422         list = &head->chain;
 423         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 424
 425         spin_lock(lock);
 426         __sk_nulls_add_node_rcu(sk, list);
 427         if (tw) {
 428                 WARN_ON(sk->sk_hash != tw->tw_hash);
 429                 twrefcnt = inet_twsk_unhash(tw);
 430         }
 431         spin_unlock(lock);
 432         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 433         return twrefcnt;
 434 }
 435 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 436
 437 int __inet_hash(struct sock *sk, struct inet_timewait_sock *tw)
 438 {
 439         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 440         struct inet_listen_hashbucket *ilb;
 441
 442         if (sk->sk_state != TCP_LISTEN)
 443                 return __inet_hash_nolisten(sk, tw);
 444
 445         WARN_ON(!sk_unhashed(sk));
 446         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 447
 448         spin_lock(&ilb->lock);
 449         __sk_nulls_add_node_rcu(sk, &ilb->head);
 450         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 451         spin_unlock(&ilb->lock);
 452         return 0;
 453 }
 454 EXPORT_SYMBOL(__inet_hash);
 455
 456 void inet_hash(struct sock *sk)
 457 {
 458         if (sk->sk_state != TCP_CLOSE) {
 459                 local_bh_disable();
 460                 __inet_hash(sk, NULL);
 461                 local_bh_enable();
 462         }
 463 }
 464 EXPORT_SYMBOL_GPL(inet_hash);
 465
 466 void inet_unhash(struct sock *sk)
 467 {
 468         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 469         spinlock_t *lock;
 470         int done;
 471
 472         if (sk_unhashed(sk))
 473                 return;
 474
 475         if (sk->sk_state == TCP_LISTEN)
 476                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 477         else
 478                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 479
 480         spin_lock_bh(lock);
 481         done = __sk_nulls_del_node_init_rcu(sk);
 482         if (done)
 483                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 484         spin_unlock_bh(lock);
 485 }
 486 EXPORT_SYMBOL_GPL(inet_unhash);
 487
 488 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 489                 struct sock *sk, u32 port_offset,
 490                 int (*check_established)(struct inet_timewait_death_row *,
 491                         struct sock *, __u16, struct inet_timewait_sock **))
 492 {
 493         struct inet_hashinfo *hinfo = death_row->hashinfo;
 494         const unsigned short snum = inet_sk(sk)->inet_num;
 495         struct inet_bind_hashbucket *head;
 496         struct inet_bind_bucket *tb;
 497         int ret;
 498         struct net *net = sock_net(sk);
 499         int twrefcnt = 1;
 500
 501         if (!snum) {
 502                 int i, remaining, low, high, port;
 503                 static u32 hint;
 504                 u32 offset = hint + port_offset;
 505                 struct inet_timewait_sock *tw = NULL;
 506
 507                 inet_get_local_port_range(net, &low, &high);
 508                 remaining = (high - low) + 1;
 509
 510                 local_bh_disable();
 511                 for (i = 1; i <= remaining; i++) {
 512                         port = low + (i + offset) % remaining;
 513                         if (inet_is_local_reserved_port(net, port))
 514                                 continue;
 515                         head = &hinfo->bhash[inet_bhashfn(net, port,
 516                                         hinfo->bhash_size)];
 517                         spin_lock(&head->lock);
 518
 519                         /* Does not bother with rcv_saddr checks,
 520                          * because the established check is already
 521                          * unique enough.
 522                          */
 523                         inet_bind_bucket_for_each(tb, &head->chain) {
 524                                 if (net_eq(ib_net(tb), net) &&
 525                                     tb->port == port) {
 526                                         if (tb->fastreuse >= 0 ||
 527                                             tb->fastreuseport >= 0)
 528                                                 goto next_port;
 529                                         WARN_ON(hlist_empty(&tb->owners));
 530                                         if (!check_established(death_row, sk,
 531                                                                 port, &tw))
 532                                                 goto ok;
 533                                         goto next_port;
 534                                 }
 535                         }
 536
 537                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 538                                         net, head, port);
 539                         if (!tb) {
 540                                 spin_unlock(&head->lock);
 541                                 break;
 542                         }
 543                         tb->fastreuse = -1;
 544                         tb->fastreuseport = -1;
 545                         goto ok;
 546
 547                 next_port:
 548                         spin_unlock(&head->lock);
 549                 }
 550                 local_bh_enable();
 551
 552                 return -EADDRNOTAVAIL;
 553
 554 ok:
 555                 hint += i;
 556
 557                 /* Head lock still held and bh's disabled */
 558                 inet_bind_hash(sk, tb, port);
 559                 if (sk_unhashed(sk)) {
 560                         inet_sk(sk)->inet_sport = htons(port);
 561                         twrefcnt += __inet_hash_nolisten(sk, tw);
 562                 }
 563                 if (tw)
 564                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
 565                 spin_unlock(&head->lock);
 566
 567                 if (tw) {
 568                         inet_twsk_deschedule(tw);
 569                         while (twrefcnt) {
 570                                 twrefcnt--;
 571                                 inet_twsk_put(tw);
 572                         }
 573                 }
 574
 575                 ret = 0;
 576                 goto out;
 577         }
 578
 579         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 580         tb  = inet_csk(sk)->icsk_bind_hash;
 581         spin_lock_bh(&head->lock);
 582         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 583                 __inet_hash_nolisten(sk, NULL);
 584                 spin_unlock_bh(&head->lock);
 585                 return 0;
 586         } else {
 587                 spin_unlock(&head->lock);
 588                 /* No definite answer... Walk to established hash table */
 589                 ret = check_established(death_row, sk, snum, NULL);
 590 out:
 591                 local_bh_enable();
 592                 return ret;
 593         }
 594 }
 595
 596 /*
 597  * Bind a port for a connect operation and hash it.
 598  */
 599 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 600                       struct sock *sk)
 601 {
 602         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 603                                    __inet_check_established);
 604 }
 605 EXPORT_SYMBOL_GPL(inet_hash_connect);
 606
 607 void inet_hashinfo_init(struct inet_hashinfo *h)
 608 {
 609         int i;
 610
 611         atomic_set(&h->bsockets, 0);
 612         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 613                 spin_lock_init(&h->listening_hash[i].lock);
 614                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 615                                       i + LISTENING_NULLS_BASE);
 616                 }
 617 }
 618 EXPORT_SYMBOL_GPL(inet_hashinfo_init);