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 #include <linux/vmalloc.h>
  22 #include <linux/memblock.h>
  23
  24 #include <net/addrconf.h>
  25 #include <net/inet_connection_sock.h>
  26 #include <net/inet_hashtables.h>
  27 #include <net/secure_seq.h>
  28 #include <net/ip.h>
  29 #include <net/tcp.h>
  30 #include <net/sock_reuseport.h>
  31
  32 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
  33                         const __u16 lport, const __be32 faddr,
  34                         const __be16 fport)
  35 {
  36         static u32 inet_ehash_secret __read_mostly;
  37
  38         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  39
  40         return __inet_ehashfn(laddr, lport, faddr, fport,
  41                               inet_ehash_secret + net_hash_mix(net));
  42 }
  43
  44 /* This function handles inet_sock, but also timewait and request sockets
  45  * for IPv4/IPv6.
  46  */
  47 static u32 sk_ehashfn(const struct sock *sk)
  48 {
  49 #if IS_ENABLED(CONFIG_IPV6)
  50         if (sk->sk_family == AF_INET6 &&
  51             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  52                 return inet6_ehashfn(sock_net(sk),
  53                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  54                                      &sk->sk_v6_daddr, sk->sk_dport);
  55 #endif
  56         return inet_ehashfn(sock_net(sk),
  57                             sk->sk_rcv_saddr, sk->sk_num,
  58                             sk->sk_daddr, sk->sk_dport);
  59 }
  60
  61 /*
  62  * Allocate and initialize a new local port bind bucket.
  63  * The bindhash mutex for snum's hash chain must be held here.
  64  */
  65 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  66                                                  struct net *net,
  67                                                  struct inet_bind_hashbucket *head,
  68                                                  const unsigned short snum,
  69                                                  int l3mdev)
  70 {
  71         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  72
  73         if (tb) {
  74                 write_pnet(&tb->ib_net, net);
  75                 tb->l3mdev    = l3mdev;
  76                 tb->port      = snum;
  77                 tb->fastreuse = 0;
  78                 tb->fastreuseport = 0;
  79                 INIT_HLIST_HEAD(&tb->owners);
  80                 hlist_add_head(&tb->node, &head->chain);
  81         }
  82         return tb;
  83 }
  84
  85 /*
  86  * Caller must hold hashbucket lock for this tb with local BH disabled
  87  */
  88 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  89 {
  90         if (hlist_empty(&tb->owners)) {
  91                 __hlist_del(&tb->node);
  92                 kmem_cache_free(cachep, tb);
  93         }
  94 }
  95
  96 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  97                     const unsigned short snum)
  98 {
  99         inet_sk(sk)->inet_num = snum;
 100         sk_add_bind_node(sk, &tb->owners);
 101         inet_csk(sk)->icsk_bind_hash = tb;
 102 }
 103
 104 /*
 105  * Get rid of any references to a local port held by the given sock.
 106  */
 107 static void __inet_put_port(struct sock *sk)
 108 {
 109         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 110         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 111                         hashinfo->bhash_size);
 112         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 113         struct inet_bind_bucket *tb;
 114
 115         spin_lock(&head->lock);
 116         tb = inet_csk(sk)->icsk_bind_hash;
 117         __sk_del_bind_node(sk);
 118         inet_csk(sk)->icsk_bind_hash = NULL;
 119         inet_sk(sk)->inet_num = 0;
 120         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 121         spin_unlock(&head->lock);
 122 }
 123
 124 void inet_put_port(struct sock *sk)
 125 {
 126         local_bh_disable();
 127         __inet_put_port(sk);
 128         local_bh_enable();
 129 }
 130 EXPORT_SYMBOL(inet_put_port);
 131
 132 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 133 {
 134         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 135         unsigned short port = inet_sk(child)->inet_num;
 136         const int bhash = inet_bhashfn(sock_net(sk), port,
 137                         table->bhash_size);
 138         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 139         struct inet_bind_bucket *tb;
 140         int l3mdev;
 141
 142         spin_lock(&head->lock);
 143         tb = inet_csk(sk)->icsk_bind_hash;
 144         if (unlikely(!tb)) {
 145                 spin_unlock(&head->lock);
 146                 return -ENOENT;
 147         }
 148         if (tb->port != port) {
 149                 l3mdev = inet_sk_bound_l3mdev(sk);
 150
 151                 /* NOTE: using tproxy and redirecting skbs to a proxy
 152                  * on a different listener port breaks the assumption
 153                  * that the listener socket's icsk_bind_hash is the same
 154                  * as that of the child socket. We have to look up or
 155                  * create a new bind bucket for the child here. */
 156                 inet_bind_bucket_for_each(tb, &head->chain) {
 157                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 158                             tb->l3mdev == l3mdev && tb->port == port)
 159                                 break;
 160                 }
 161                 if (!tb) {
 162                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 163                                                      sock_net(sk), head, port,
 164                                                      l3mdev);
 165                         if (!tb) {
 166                                 spin_unlock(&head->lock);
 167                                 return -ENOMEM;
 168                         }
 169                 }
 170         }
 171         inet_bind_hash(child, tb, port);
 172         spin_unlock(&head->lock);
 173
 174         return 0;
 175 }
 176 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 177
 178 static struct inet_listen_hashbucket *
 179 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 180 {
 181         u32 hash;
 182
 183 #if IS_ENABLED(CONFIG_IPV6)
 184         if (sk->sk_family == AF_INET6)
 185                 hash = ipv6_portaddr_hash(sock_net(sk),
 186                                           &sk->sk_v6_rcv_saddr,
 187                                           inet_sk(sk)->inet_num);
 188         else
 189 #endif
 190                 hash = ipv4_portaddr_hash(sock_net(sk),
 191                                           inet_sk(sk)->inet_rcv_saddr,
 192                                           inet_sk(sk)->inet_num);
 193         return inet_lhash2_bucket(h, hash);
 194 }
 195
 196 static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
 197 {
 198         struct inet_listen_hashbucket *ilb2;
 199
 200         if (!h->lhash2)
 201                 return;
 202
 203         ilb2 = inet_lhash2_bucket_sk(h, sk);
 204
 205         spin_lock(&ilb2->lock);
 206         if (sk->sk_reuseport && sk->sk_family == AF_INET6)
 207                 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 208                                    &ilb2->head);
 209         else
 210                 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 211                                    &ilb2->head);
 212         ilb2->count++;
 213         spin_unlock(&ilb2->lock);
 214 }
 215
 216 static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
 217 {
 218         struct inet_listen_hashbucket *ilb2;
 219
 220         if (!h->lhash2 ||
 221             WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
 222                 return;
 223
 224         ilb2 = inet_lhash2_bucket_sk(h, sk);
 225
 226         spin_lock(&ilb2->lock);
 227         hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
 228         ilb2->count--;
 229         spin_unlock(&ilb2->lock);
 230 }
 231
 232 static inline int compute_score(struct sock *sk, struct net *net,
 233                                 const unsigned short hnum, const __be32 daddr,
 234                                 const int dif, const int sdif, bool exact_dif)
 235 {
 236         int score = -1;
 237
 238         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 239                         !ipv6_only_sock(sk)) {
 240                 if (sk->sk_rcv_saddr != daddr)
 241                         return -1;
 242
 243                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 244                         return -1;
 245
 246                 score = sk->sk_family == PF_INET ? 2 : 1;
 247                 if (sk->sk_incoming_cpu == raw_smp_processor_id())
 248                         score++;
 249         }
 250         return score;
 251 }
 252
 253 /*
 254  * Here are some nice properties to exploit here. The BSD API
 255  * does not allow a listening sock to specify the remote port nor the
 256  * remote address for the connection. So always assume those are both
 257  * wildcarded during the search since they can never be otherwise.
 258  */
 259
 260 /* called with rcu_read_lock() : No refcount taken on the socket */
 261 static struct sock *inet_lhash2_lookup(struct net *net,
 262                                 struct inet_listen_hashbucket *ilb2,
 263                                 struct sk_buff *skb, int doff,
 264                                 const __be32 saddr, __be16 sport,
 265                                 const __be32 daddr, const unsigned short hnum,
 266                                 const int dif, const int sdif)
 267 {
 268         bool exact_dif = inet_exact_dif_match(net, skb);
 269         struct inet_connection_sock *icsk;
 270         struct sock *sk, *result = NULL;
 271         int score, hiscore = 0;
 272         u32 phash = 0;
 273
 274         inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
 275                 sk = (struct sock *)icsk;
 276                 score = compute_score(sk, net, hnum, daddr,
 277                                       dif, sdif, exact_dif);
 278                 if (score > hiscore) {
 279                         if (sk->sk_reuseport) {
 280                                 phash = inet_ehashfn(net, daddr, hnum,
 281                                                      saddr, sport);
 282                                 result = reuseport_select_sock(sk, phash,
 283                                                                skb, doff);
 284                                 if (result)
 285                                         return result;
 286                         }
 287                         result = sk;
 288                         hiscore = score;
 289                 }
 290         }
 291
 292         return result;
 293 }
 294
 295 struct sock *__inet_lookup_listener(struct net *net,
 296                                     struct inet_hashinfo *hashinfo,
 297                                     struct sk_buff *skb, int doff,
 298                                     const __be32 saddr, __be16 sport,
 299                                     const __be32 daddr, const unsigned short hnum,
 300                                     const int dif, const int sdif)
 301 {
 302         struct inet_listen_hashbucket *ilb2;
 303         struct sock *result = NULL;
 304         unsigned int hash2;
 305
 306         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 307         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 308
 309         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 310                                     saddr, sport, daddr, hnum,
 311                                     dif, sdif);
 312         if (result)
 313                 goto done;
 314
 315         /* Lookup lhash2 with INADDR_ANY */
 316         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 317         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 318
 319         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 320                                     saddr, sport, htonl(INADDR_ANY), hnum,
 321                                     dif, sdif);
 322 done:
 323         if (unlikely(IS_ERR(result)))
 324                 return NULL;
 325         return result;
 326 }
 327 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 328
 329 /* All sockets share common refcount, but have different destructors */
 330 void sock_gen_put(struct sock *sk)
 331 {
 332         if (!refcount_dec_and_test(&sk->sk_refcnt))
 333                 return;
 334
 335         if (sk->sk_state == TCP_TIME_WAIT)
 336                 inet_twsk_free(inet_twsk(sk));
 337         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 338                 reqsk_free(inet_reqsk(sk));
 339         else
 340                 sk_free(sk);
 341 }
 342 EXPORT_SYMBOL_GPL(sock_gen_put);
 343
 344 void sock_edemux(struct sk_buff *skb)
 345 {
 346         sock_gen_put(skb->sk);
 347 }
 348 EXPORT_SYMBOL(sock_edemux);
 349
 350 struct sock *__inet_lookup_established(struct net *net,
 351                                   struct inet_hashinfo *hashinfo,
 352                                   const __be32 saddr, const __be16 sport,
 353                                   const __be32 daddr, const u16 hnum,
 354                                   const int dif, const int sdif)
 355 {
 356         INET_ADDR_COOKIE(acookie, saddr, daddr);
 357         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 358         struct sock *sk;
 359         const struct hlist_nulls_node *node;
 360         /* Optimize here for direct hit, only listening connections can
 361          * have wildcards anyways.
 362          */
 363         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 364         unsigned int slot = hash & hashinfo->ehash_mask;
 365         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 366
 367 begin:
 368         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 369                 if (sk->sk_hash != hash)
 370                         continue;
 371                 if (likely(INET_MATCH(sk, net, acookie,
 372                                       saddr, daddr, ports, dif, sdif))) {
 373                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 374                                 goto out;
 375                         if (unlikely(!INET_MATCH(sk, net, acookie,
 376                                                  saddr, daddr, ports,
 377                                                  dif, sdif))) {
 378                                 sock_gen_put(sk);
 379                                 goto begin;
 380                         }
 381                         goto found;
 382                 }
 383         }
 384         /*
 385          * if the nulls value we got at the end of this lookup is
 386          * not the expected one, we must restart lookup.
 387          * We probably met an item that was moved to another chain.
 388          */
 389         if (get_nulls_value(node) != slot)
 390                 goto begin;
 391 out:
 392         sk = NULL;
 393 found:
 394         return sk;
 395 }
 396 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 397
 398 /* called with local bh disabled */
 399 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 400                                     struct sock *sk, __u16 lport,
 401                                     struct inet_timewait_sock **twp)
 402 {
 403         struct inet_hashinfo *hinfo = death_row->hashinfo;
 404         struct inet_sock *inet = inet_sk(sk);
 405         __be32 daddr = inet->inet_rcv_saddr;
 406         __be32 saddr = inet->inet_daddr;
 407         int dif = sk->sk_bound_dev_if;
 408         struct net *net = sock_net(sk);
 409         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 410         INET_ADDR_COOKIE(acookie, saddr, daddr);
 411         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 412         unsigned int hash = inet_ehashfn(net, daddr, lport,
 413                                          saddr, inet->inet_dport);
 414         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 415         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 416         struct sock *sk2;
 417         const struct hlist_nulls_node *node;
 418         struct inet_timewait_sock *tw = NULL;
 419
 420         spin_lock(lock);
 421
 422         sk_nulls_for_each(sk2, node, &head->chain) {
 423                 if (sk2->sk_hash != hash)
 424                         continue;
 425
 426                 if (likely(INET_MATCH(sk2, net, acookie,
 427                                          saddr, daddr, ports, dif, sdif))) {
 428                         if (sk2->sk_state == TCP_TIME_WAIT) {
 429                                 tw = inet_twsk(sk2);
 430                                 if (twsk_unique(sk, sk2, twp))
 431                                         break;
 432                         }
 433                         goto not_unique;
 434                 }
 435         }
 436
 437         /* Must record num and sport now. Otherwise we will see
 438          * in hash table socket with a funny identity.
 439          */
 440         inet->inet_num = lport;
 441         inet->inet_sport = htons(lport);
 442         sk->sk_hash = hash;
 443         WARN_ON(!sk_unhashed(sk));
 444         __sk_nulls_add_node_rcu(sk, &head->chain);
 445         if (tw) {
 446                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 447                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 448         }
 449         spin_unlock(lock);
 450         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 451
 452         if (twp) {
 453                 *twp = tw;
 454         } else if (tw) {
 455                 /* Silly. Should hash-dance instead... */
 456                 inet_twsk_deschedule_put(tw);
 457         }
 458         return 0;
 459
 460 not_unique:
 461         spin_unlock(lock);
 462         return -EADDRNOTAVAIL;
 463 }
 464
 465 static u32 inet_sk_port_offset(const struct sock *sk)
 466 {
 467         const struct inet_sock *inet = inet_sk(sk);
 468
 469         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 470                                           inet->inet_daddr,
 471                                           inet->inet_dport);
 472 }
 473
 474 /* insert a socket into ehash, and eventually remove another one
 475  * (The another one can be a SYN_RECV or TIMEWAIT
 476  */
 477 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
 478 {
 479         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 480         struct hlist_nulls_head *list;
 481         struct inet_ehash_bucket *head;
 482         spinlock_t *lock;
 483         bool ret = true;
 484
 485         WARN_ON_ONCE(!sk_unhashed(sk));
 486
 487         sk->sk_hash = sk_ehashfn(sk);
 488         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 489         list = &head->chain;
 490         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 491
 492         spin_lock(lock);
 493         if (osk) {
 494                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 495                 ret = sk_nulls_del_node_init_rcu(osk);
 496         }
 497         if (ret)
 498                 __sk_nulls_add_node_rcu(sk, list);
 499         spin_unlock(lock);
 500         return ret;
 501 }
 502
 503 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
 504 {
 505         bool ok = inet_ehash_insert(sk, osk);
 506
 507         if (ok) {
 508                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 509         } else {
 510                 percpu_counter_inc(sk->sk_prot->orphan_count);
 511                 inet_sk_set_state(sk, TCP_CLOSE);
 512                 sock_set_flag(sk, SOCK_DEAD);
 513                 inet_csk_destroy_sock(sk);
 514         }
 515         return ok;
 516 }
 517 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 518
 519 static int inet_reuseport_add_sock(struct sock *sk,
 520                                    struct inet_listen_hashbucket *ilb)
 521 {
 522         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 523         struct sock *sk2;
 524         kuid_t uid = sock_i_uid(sk);
 525
 526         sk_for_each_rcu(sk2, &ilb->head) {
 527                 if (sk2 != sk &&
 528                     sk2->sk_family == sk->sk_family &&
 529                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 530                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 531                     inet_csk(sk2)->icsk_bind_hash == tb &&
 532                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 533                     inet_rcv_saddr_equal(sk, sk2, false))
 534                         return reuseport_add_sock(sk, sk2,
 535                                                   inet_rcv_saddr_any(sk));
 536         }
 537
 538         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 539 }
 540
 541 int __inet_hash(struct sock *sk, struct sock *osk)
 542 {
 543         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 544         struct inet_listen_hashbucket *ilb;
 545         int err = 0;
 546
 547         if (sk->sk_state != TCP_LISTEN) {
 548                 inet_ehash_nolisten(sk, osk);
 549                 return 0;
 550         }
 551         WARN_ON(!sk_unhashed(sk));
 552         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 553
 554         spin_lock(&ilb->lock);
 555         if (sk->sk_reuseport) {
 556                 err = inet_reuseport_add_sock(sk, ilb);
 557                 if (err)
 558                         goto unlock;
 559         }
 560         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 561                 sk->sk_family == AF_INET6)
 562                 hlist_add_tail_rcu(&sk->sk_node, &ilb->head);
 563         else
 564                 hlist_add_head_rcu(&sk->sk_node, &ilb->head);
 565         inet_hash2(hashinfo, sk);
 566         ilb->count++;
 567         sock_set_flag(sk, SOCK_RCU_FREE);
 568         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 569 unlock:
 570         spin_unlock(&ilb->lock);
 571
 572         return err;
 573 }
 574 EXPORT_SYMBOL(__inet_hash);
 575
 576 int inet_hash(struct sock *sk)
 577 {
 578         int err = 0;
 579
 580         if (sk->sk_state != TCP_CLOSE) {
 581                 local_bh_disable();
 582                 err = __inet_hash(sk, NULL);
 583                 local_bh_enable();
 584         }
 585
 586         return err;
 587 }
 588 EXPORT_SYMBOL_GPL(inet_hash);
 589
 590 void inet_unhash(struct sock *sk)
 591 {
 592         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 593         struct inet_listen_hashbucket *ilb = NULL;
 594         spinlock_t *lock;
 595
 596         if (sk_unhashed(sk))
 597                 return;
 598
 599         if (sk->sk_state == TCP_LISTEN) {
 600                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 601                 lock = &ilb->lock;
 602         } else {
 603                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 604         }
 605         spin_lock_bh(lock);
 606         if (sk_unhashed(sk))
 607                 goto unlock;
 608
 609         if (rcu_access_pointer(sk->sk_reuseport_cb))
 610                 reuseport_detach_sock(sk);
 611         if (ilb) {
 612                 inet_unhash2(hashinfo, sk);
 613                  __sk_del_node_init(sk);
 614                  ilb->count--;
 615         } else {
 616                 __sk_nulls_del_node_init_rcu(sk);
 617         }
 618         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 619 unlock:
 620         spin_unlock_bh(lock);
 621 }
 622 EXPORT_SYMBOL_GPL(inet_unhash);
 623
 624 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 625                 struct sock *sk, u32 port_offset,
 626                 int (*check_established)(struct inet_timewait_death_row *,
 627                         struct sock *, __u16, struct inet_timewait_sock **))
 628 {
 629         struct inet_hashinfo *hinfo = death_row->hashinfo;
 630         struct inet_timewait_sock *tw = NULL;
 631         struct inet_bind_hashbucket *head;
 632         int port = inet_sk(sk)->inet_num;
 633         struct net *net = sock_net(sk);
 634         struct inet_bind_bucket *tb;
 635         u32 remaining, offset;
 636         int ret, i, low, high;
 637         static u32 hint;
 638         int l3mdev;
 639
 640         if (port) {
 641                 head = &hinfo->bhash[inet_bhashfn(net, port,
 642                                                   hinfo->bhash_size)];
 643                 tb = inet_csk(sk)->icsk_bind_hash;
 644                 spin_lock_bh(&head->lock);
 645                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 646                         inet_ehash_nolisten(sk, NULL);
 647                         spin_unlock_bh(&head->lock);
 648                         return 0;
 649                 }
 650                 spin_unlock(&head->lock);
 651                 /* No definite answer... Walk to established hash table */
 652                 ret = check_established(death_row, sk, port, NULL);
 653                 local_bh_enable();
 654                 return ret;
 655         }
 656
 657         l3mdev = inet_sk_bound_l3mdev(sk);
 658
 659         inet_get_local_port_range(net, &low, &high);
 660         high++; /* [32768, 60999] -> [32768, 61000[ */
 661         remaining = high - low;
 662         if (likely(remaining > 1))
 663                 remaining &= ~1U;
 664
 665         offset = (hint + port_offset) % remaining;
 666         /* In first pass we try ports of @low parity.
 667          * inet_csk_get_port() does the opposite choice.
 668          */
 669         offset &= ~1U;
 670 other_parity_scan:
 671         port = low + offset;
 672         for (i = 0; i < remaining; i += 2, port += 2) {
 673                 if (unlikely(port >= high))
 674                         port -= remaining;
 675                 if (inet_is_local_reserved_port(net, port))
 676                         continue;
 677                 head = &hinfo->bhash[inet_bhashfn(net, port,
 678                                                   hinfo->bhash_size)];
 679                 spin_lock_bh(&head->lock);
 680
 681                 /* Does not bother with rcv_saddr checks, because
 682                  * the established check is already unique enough.
 683                  */
 684                 inet_bind_bucket_for_each(tb, &head->chain) {
 685                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 686                             tb->port == port) {
 687                                 if (tb->fastreuse >= 0 ||
 688                                     tb->fastreuseport >= 0)
 689                                         goto next_port;
 690                                 WARN_ON(hlist_empty(&tb->owners));
 691                                 if (!check_established(death_row, sk,
 692                                                        port, &tw))
 693                                         goto ok;
 694                                 goto next_port;
 695                         }
 696                 }
 697
 698                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 699                                              net, head, port, l3mdev);
 700                 if (!tb) {
 701                         spin_unlock_bh(&head->lock);
 702                         return -ENOMEM;
 703                 }
 704                 tb->fastreuse = -1;
 705                 tb->fastreuseport = -1;
 706                 goto ok;
 707 next_port:
 708                 spin_unlock_bh(&head->lock);
 709                 cond_resched();
 710         }
 711
 712         offset++;
 713         if ((offset & 1) && remaining > 1)
 714                 goto other_parity_scan;
 715
 716         return -EADDRNOTAVAIL;
 717
 718 ok:
 719         hint += i + 2;
 720
 721         /* Head lock still held and bh's disabled */
 722         inet_bind_hash(sk, tb, port);
 723         if (sk_unhashed(sk)) {
 724                 inet_sk(sk)->inet_sport = htons(port);
 725                 inet_ehash_nolisten(sk, (struct sock *)tw);
 726         }
 727         if (tw)
 728                 inet_twsk_bind_unhash(tw, hinfo);
 729         spin_unlock(&head->lock);
 730         if (tw)
 731                 inet_twsk_deschedule_put(tw);
 732         local_bh_enable();
 733         return 0;
 734 }
 735
 736 /*
 737  * Bind a port for a connect operation and hash it.
 738  */
 739 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 740                       struct sock *sk)
 741 {
 742         u32 port_offset = 0;
 743
 744         if (!inet_sk(sk)->inet_num)
 745                 port_offset = inet_sk_port_offset(sk);
 746         return __inet_hash_connect(death_row, sk, port_offset,
 747                                    __inet_check_established);
 748 }
 749 EXPORT_SYMBOL_GPL(inet_hash_connect);
 750
 751 void inet_hashinfo_init(struct inet_hashinfo *h)
 752 {
 753         int i;
 754
 755         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 756                 spin_lock_init(&h->listening_hash[i].lock);
 757                 INIT_HLIST_HEAD(&h->listening_hash[i].head);
 758                 h->listening_hash[i].count = 0;
 759         }
 760
 761         h->lhash2 = NULL;
 762 }
 763 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 764
 765 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
 766                                 unsigned long numentries, int scale,
 767                                 unsigned long low_limit,
 768                                 unsigned long high_limit)
 769 {
 770         unsigned int i;
 771
 772         h->lhash2 = alloc_large_system_hash(name,
 773                                             sizeof(*h->lhash2),
 774                                             numentries,
 775                                             scale,
 776                                             0,
 777                                             NULL,
 778                                             &h->lhash2_mask,
 779                                             low_limit,
 780                                             high_limit);
 781
 782         for (i = 0; i <= h->lhash2_mask; i++) {
 783                 spin_lock_init(&h->lhash2[i].lock);
 784                 INIT_HLIST_HEAD(&h->lhash2[i].head);
 785                 h->lhash2[i].count = 0;
 786         }
 787 }
 788 EXPORT_SYMBOL_GPL(inet_hashinfo2_init);
 789
 790 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 791 {
 792         unsigned int locksz = sizeof(spinlock_t);
 793         unsigned int i, nblocks = 1;
 794
 795         if (locksz != 0) {
 796                 /* allocate 2 cache lines or at least one spinlock per cpu */
 797                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 798                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 799
 800                 /* no more locks than number of hash buckets */
 801                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 802
 803                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
 804                 if (!hashinfo->ehash_locks)
 805                         return -ENOMEM;
 806
 807                 for (i = 0; i < nblocks; i++)
 808                         spin_lock_init(&hashinfo->ehash_locks[i]);
 809         }
 810         hashinfo->ehash_locks_mask = nblocks - 1;
 811         return 0;
 812 }
 813 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);