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/bootmem.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 {
  70         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  71
  72         if (tb) {
  73                 write_pnet(&tb->ib_net, net);
  74                 tb->port      = snum;
  75                 tb->fastreuse = 0;
  76                 tb->fastreuseport = 0;
  77                 INIT_HLIST_HEAD(&tb->owners);
  78                 hlist_add_head(&tb->node, &head->chain);
  79         }
  80         return tb;
  81 }
  82
  83 /*
  84  * Caller must hold hashbucket lock for this tb with local BH disabled
  85  */
  86 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  87 {
  88         if (hlist_empty(&tb->owners)) {
  89                 __hlist_del(&tb->node);
  90                 kmem_cache_free(cachep, tb);
  91         }
  92 }
  93
  94 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  95                     const unsigned short snum)
  96 {
  97         inet_sk(sk)->inet_num = snum;
  98         sk_add_bind_node(sk, &tb->owners);
  99         inet_csk(sk)->icsk_bind_hash = tb;
 100 }
 101
 102 /*
 103  * Get rid of any references to a local port held by the given sock.
 104  */
 105 static void __inet_put_port(struct sock *sk)
 106 {
 107         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 108         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 109                         hashinfo->bhash_size);
 110         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 111         struct inet_bind_bucket *tb;
 112
 113         spin_lock(&head->lock);
 114         tb = inet_csk(sk)->icsk_bind_hash;
 115         __sk_del_bind_node(sk);
 116         inet_csk(sk)->icsk_bind_hash = NULL;
 117         inet_sk(sk)->inet_num = 0;
 118         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 119         spin_unlock(&head->lock);
 120 }
 121
 122 void inet_put_port(struct sock *sk)
 123 {
 124         local_bh_disable();
 125         __inet_put_port(sk);
 126         local_bh_enable();
 127 }
 128 EXPORT_SYMBOL(inet_put_port);
 129
 130 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 131 {
 132         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 133         unsigned short port = inet_sk(child)->inet_num;
 134         const int bhash = inet_bhashfn(sock_net(sk), port,
 135                         table->bhash_size);
 136         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 137         struct inet_bind_bucket *tb;
 138
 139         spin_lock(&head->lock);
 140         tb = inet_csk(sk)->icsk_bind_hash;
 141         if (unlikely(!tb)) {
 142                 spin_unlock(&head->lock);
 143                 return -ENOENT;
 144         }
 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                 inet_csk_update_fastreuse(tb, child);
 165         }
 166         inet_bind_hash(child, tb, port);
 167         spin_unlock(&head->lock);
 168
 169         return 0;
 170 }
 171 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 172
 173 static struct inet_listen_hashbucket *
 174 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 175 {
 176         u32 hash;
 177
 178 #if IS_ENABLED(CONFIG_IPV6)
 179         if (sk->sk_family == AF_INET6)
 180                 hash = ipv6_portaddr_hash(sock_net(sk),
 181                                           &sk->sk_v6_rcv_saddr,
 182                                           inet_sk(sk)->inet_num);
 183         else
 184 #endif
 185                 hash = ipv4_portaddr_hash(sock_net(sk),
 186                                           inet_sk(sk)->inet_rcv_saddr,
 187                                           inet_sk(sk)->inet_num);
 188         return inet_lhash2_bucket(h, hash);
 189 }
 190
 191 static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
 192 {
 193         struct inet_listen_hashbucket *ilb2;
 194
 195         if (!h->lhash2)
 196                 return;
 197
 198         ilb2 = inet_lhash2_bucket_sk(h, sk);
 199
 200         spin_lock(&ilb2->lock);
 201         if (sk->sk_reuseport && sk->sk_family == AF_INET6)
 202                 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 203                                    &ilb2->head);
 204         else
 205                 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 206                                    &ilb2->head);
 207         ilb2->count++;
 208         spin_unlock(&ilb2->lock);
 209 }
 210
 211 static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
 212 {
 213         struct inet_listen_hashbucket *ilb2;
 214
 215         if (!h->lhash2 ||
 216             WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
 217                 return;
 218
 219         ilb2 = inet_lhash2_bucket_sk(h, sk);
 220
 221         spin_lock(&ilb2->lock);
 222         hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
 223         ilb2->count--;
 224         spin_unlock(&ilb2->lock);
 225 }
 226
 227 static inline int compute_score(struct sock *sk, struct net *net,
 228                                 const unsigned short hnum, const __be32 daddr,
 229                                 const int dif, const int sdif, bool exact_dif)
 230 {
 231         int score = -1;
 232         struct inet_sock *inet = inet_sk(sk);
 233
 234         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 235                         !ipv6_only_sock(sk)) {
 236                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 237                 score = sk->sk_family == PF_INET ? 2 : 1;
 238                 if (rcv_saddr) {
 239                         if (rcv_saddr != daddr)
 240                                 return -1;
 241                         score += 4;
 242                 }
 243                 if (sk->sk_bound_dev_if || exact_dif) {
 244                         bool dev_match = (sk->sk_bound_dev_if == dif ||
 245                                           sk->sk_bound_dev_if == sdif);
 246
 247                         if (!dev_match)
 248                                 return -1;
 249                         if (sk->sk_bound_dev_if)
 250                                 score += 4;
 251                 }
 252                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 253                         score++;
 254         }
 255         return score;
 256 }
 257
 258 /*
 259  * Here are some nice properties to exploit here. The BSD API
 260  * does not allow a listening sock to specify the remote port nor the
 261  * remote address for the connection. So always assume those are both
 262  * wildcarded during the search since they can never be otherwise.
 263  */
 264
 265 /* called with rcu_read_lock() : No refcount taken on the socket */
 266 static struct sock *inet_lhash2_lookup(struct net *net,
 267                                 struct inet_listen_hashbucket *ilb2,
 268                                 struct sk_buff *skb, int doff,
 269                                 const __be32 saddr, __be16 sport,
 270                                 const __be32 daddr, const unsigned short hnum,
 271                                 const int dif, const int sdif)
 272 {
 273         bool exact_dif = inet_exact_dif_match(net, skb);
 274         struct inet_connection_sock *icsk;
 275         struct sock *sk, *result = NULL;
 276         int score, hiscore = 0;
 277         u32 phash = 0;
 278
 279         inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
 280                 sk = (struct sock *)icsk;
 281                 score = compute_score(sk, net, hnum, daddr,
 282                                       dif, sdif, exact_dif);
 283                 if (score > hiscore) {
 284                         if (sk->sk_reuseport) {
 285                                 phash = inet_ehashfn(net, daddr, hnum,
 286                                                      saddr, sport);
 287                                 result = reuseport_select_sock(sk, phash,
 288                                                                skb, doff);
 289                                 if (result)
 290                                         return result;
 291                         }
 292                         result = sk;
 293                         hiscore = score;
 294                 }
 295         }
 296
 297         return result;
 298 }
 299
 300 struct sock *__inet_lookup_listener(struct net *net,
 301                                     struct inet_hashinfo *hashinfo,
 302                                     struct sk_buff *skb, int doff,
 303                                     const __be32 saddr, __be16 sport,
 304                                     const __be32 daddr, const unsigned short hnum,
 305                                     const int dif, const int sdif)
 306 {
 307         unsigned int hash = inet_lhashfn(net, hnum);
 308         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 309         bool exact_dif = inet_exact_dif_match(net, skb);
 310         struct inet_listen_hashbucket *ilb2;
 311         struct sock *sk, *result = NULL;
 312         struct hlist_nulls_node *node;
 313         int score, hiscore = 0;
 314         unsigned int hash2;
 315         u32 phash = 0;
 316
 317         if (ilb->count <= 10 || !hashinfo->lhash2)
 318                 goto port_lookup;
 319
 320         /* Too many sk in the ilb bucket (which is hashed by port alone).
 321          * Try lhash2 (which is hashed by port and addr) instead.
 322          */
 323
 324         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 325         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 326         if (ilb2->count > ilb->count)
 327                 goto port_lookup;
 328
 329         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 330                                     saddr, sport, daddr, hnum,
 331                                     dif, sdif);
 332         if (result)
 333                 goto done;
 334
 335         /* Lookup lhash2 with INADDR_ANY */
 336
 337         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 338         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 339         if (ilb2->count > ilb->count)
 340                 goto port_lookup;
 341
 342         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 343                                     saddr, sport, daddr, hnum,
 344                                     dif, sdif);
 345         goto done;
 346
 347 port_lookup:
 348         sk_nulls_for_each_rcu(sk, node, &ilb->nulls_head) {
 349                 score = compute_score(sk, net, hnum, daddr,
 350                                       dif, sdif, exact_dif);
 351                 if (score > hiscore) {
 352                         if (sk->sk_reuseport) {
 353                                 phash = inet_ehashfn(net, daddr, hnum,
 354                                                      saddr, sport);
 355                                 result = reuseport_select_sock(sk, phash,
 356                                                                skb, doff);
 357                                 if (result)
 358                                         goto done;
 359                         }
 360                         result = sk;
 361                         hiscore = score;
 362                 }
 363         }
 364 done:
 365         if (unlikely(IS_ERR(result)))
 366                 return NULL;
 367         return result;
 368 }
 369 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 370
 371 /* All sockets share common refcount, but have different destructors */
 372 void sock_gen_put(struct sock *sk)
 373 {
 374         if (!refcount_dec_and_test(&sk->sk_refcnt))
 375                 return;
 376
 377         if (sk->sk_state == TCP_TIME_WAIT)
 378                 inet_twsk_free(inet_twsk(sk));
 379         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 380                 reqsk_free(inet_reqsk(sk));
 381         else
 382                 sk_free(sk);
 383 }
 384 EXPORT_SYMBOL_GPL(sock_gen_put);
 385
 386 void sock_edemux(struct sk_buff *skb)
 387 {
 388         sock_gen_put(skb->sk);
 389 }
 390 EXPORT_SYMBOL(sock_edemux);
 391
 392 struct sock *__inet_lookup_established(struct net *net,
 393                                   struct inet_hashinfo *hashinfo,
 394                                   const __be32 saddr, const __be16 sport,
 395                                   const __be32 daddr, const u16 hnum,
 396                                   const int dif, const int sdif)
 397 {
 398         INET_ADDR_COOKIE(acookie, saddr, daddr);
 399         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 400         struct sock *sk;
 401         const struct hlist_nulls_node *node;
 402         /* Optimize here for direct hit, only listening connections can
 403          * have wildcards anyways.
 404          */
 405         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 406         unsigned int slot = hash & hashinfo->ehash_mask;
 407         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 408
 409 begin:
 410         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 411                 if (sk->sk_hash != hash)
 412                         continue;
 413                 if (likely(INET_MATCH(sk, net, acookie,
 414                                       saddr, daddr, ports, dif, sdif))) {
 415                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 416                                 goto out;
 417                         if (unlikely(!INET_MATCH(sk, net, acookie,
 418                                                  saddr, daddr, ports,
 419                                                  dif, sdif))) {
 420                                 sock_gen_put(sk);
 421                                 goto begin;
 422                         }
 423                         goto found;
 424                 }
 425         }
 426         /*
 427          * if the nulls value we got at the end of this lookup is
 428          * not the expected one, we must restart lookup.
 429          * We probably met an item that was moved to another chain.
 430          */
 431         if (get_nulls_value(node) != slot)
 432                 goto begin;
 433 out:
 434         sk = NULL;
 435 found:
 436         return sk;
 437 }
 438 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 439
 440 /* called with local bh disabled */
 441 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 442                                     struct sock *sk, __u16 lport,
 443                                     struct inet_timewait_sock **twp)
 444 {
 445         struct inet_hashinfo *hinfo = death_row->hashinfo;
 446         struct inet_sock *inet = inet_sk(sk);
 447         __be32 daddr = inet->inet_rcv_saddr;
 448         __be32 saddr = inet->inet_daddr;
 449         int dif = sk->sk_bound_dev_if;
 450         struct net *net = sock_net(sk);
 451         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 452         INET_ADDR_COOKIE(acookie, saddr, daddr);
 453         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 454         unsigned int hash = inet_ehashfn(net, daddr, lport,
 455                                          saddr, inet->inet_dport);
 456         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 457         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 458         struct sock *sk2;
 459         const struct hlist_nulls_node *node;
 460         struct inet_timewait_sock *tw = NULL;
 461
 462         spin_lock(lock);
 463
 464         sk_nulls_for_each(sk2, node, &head->chain) {
 465                 if (sk2->sk_hash != hash)
 466                         continue;
 467
 468                 if (likely(INET_MATCH(sk2, net, acookie,
 469                                          saddr, daddr, ports, dif, sdif))) {
 470                         if (sk2->sk_state == TCP_TIME_WAIT) {
 471                                 tw = inet_twsk(sk2);
 472                                 if (twsk_unique(sk, sk2, twp))
 473                                         break;
 474                         }
 475                         goto not_unique;
 476                 }
 477         }
 478
 479         /* Must record num and sport now. Otherwise we will see
 480          * in hash table socket with a funny identity.
 481          */
 482         inet->inet_num = lport;
 483         inet->inet_sport = htons(lport);
 484         sk->sk_hash = hash;
 485         WARN_ON(!sk_unhashed(sk));
 486         __sk_nulls_add_node_rcu(sk, &head->chain);
 487         if (tw) {
 488                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 489                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 490         }
 491         spin_unlock(lock);
 492         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 493
 494         if (twp) {
 495                 *twp = tw;
 496         } else if (tw) {
 497                 /* Silly. Should hash-dance instead... */
 498                 inet_twsk_deschedule_put(tw);
 499         }
 500         return 0;
 501
 502 not_unique:
 503         spin_unlock(lock);
 504         return -EADDRNOTAVAIL;
 505 }
 506
 507 static u32 inet_sk_port_offset(const struct sock *sk)
 508 {
 509         const struct inet_sock *inet = inet_sk(sk);
 510
 511         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 512                                           inet->inet_daddr,
 513                                           inet->inet_dport);
 514 }
 515
 516 /* insert a socket into ehash, and eventually remove another one
 517  * (The another one can be a SYN_RECV or TIMEWAIT
 518  */
 519 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
 520 {
 521         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 522         struct hlist_nulls_head *list;
 523         struct inet_ehash_bucket *head;
 524         spinlock_t *lock;
 525         bool ret = true;
 526
 527         WARN_ON_ONCE(!sk_unhashed(sk));
 528
 529         sk->sk_hash = sk_ehashfn(sk);
 530         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 531         list = &head->chain;
 532         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 533
 534         spin_lock(lock);
 535         if (osk) {
 536                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 537                 ret = sk_nulls_del_node_init_rcu(osk);
 538         }
 539         if (ret)
 540                 __sk_nulls_add_node_rcu(sk, list);
 541         spin_unlock(lock);
 542         return ret;
 543 }
 544
 545 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
 546 {
 547         bool ok = inet_ehash_insert(sk, osk);
 548
 549         if (ok) {
 550                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 551         } else {
 552                 percpu_counter_inc(sk->sk_prot->orphan_count);
 553                 inet_sk_set_state(sk, TCP_CLOSE);
 554                 sock_set_flag(sk, SOCK_DEAD);
 555                 inet_csk_destroy_sock(sk);
 556         }
 557         return ok;
 558 }
 559 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 560
 561 static int inet_reuseport_add_sock(struct sock *sk,
 562                                    struct inet_listen_hashbucket *ilb)
 563 {
 564         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 565         const struct hlist_nulls_node *node;
 566         struct sock *sk2;
 567         kuid_t uid = sock_i_uid(sk);
 568
 569         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 570                 if (sk2 != sk &&
 571                     sk2->sk_family == sk->sk_family &&
 572                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 573                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 574                     inet_csk(sk2)->icsk_bind_hash == tb &&
 575                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 576                     inet_rcv_saddr_equal(sk, sk2, false))
 577                         return reuseport_add_sock(sk, sk2,
 578                                                   inet_rcv_saddr_any(sk));
 579         }
 580
 581         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 582 }
 583
 584 int __inet_hash(struct sock *sk, struct sock *osk)
 585 {
 586         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 587         struct inet_listen_hashbucket *ilb;
 588         int err = 0;
 589
 590         if (sk->sk_state != TCP_LISTEN) {
 591                 inet_ehash_nolisten(sk, osk);
 592                 return 0;
 593         }
 594         WARN_ON(!sk_unhashed(sk));
 595         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 596
 597         spin_lock(&ilb->lock);
 598         if (sk->sk_reuseport) {
 599                 err = inet_reuseport_add_sock(sk, ilb);
 600                 if (err)
 601                         goto unlock;
 602         }
 603         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 604                 sk->sk_family == AF_INET6)
 605                 __sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
 606         else
 607                 __sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
 608         inet_hash2(hashinfo, sk);
 609         ilb->count++;
 610         sock_set_flag(sk, SOCK_RCU_FREE);
 611         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 612 unlock:
 613         spin_unlock(&ilb->lock);
 614
 615         return err;
 616 }
 617 EXPORT_SYMBOL(__inet_hash);
 618
 619 int inet_hash(struct sock *sk)
 620 {
 621         int err = 0;
 622
 623         if (sk->sk_state != TCP_CLOSE) {
 624                 local_bh_disable();
 625                 err = __inet_hash(sk, NULL);
 626                 local_bh_enable();
 627         }
 628
 629         return err;
 630 }
 631 EXPORT_SYMBOL_GPL(inet_hash);
 632
 633 void inet_unhash(struct sock *sk)
 634 {
 635         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 636         struct inet_listen_hashbucket *ilb = NULL;
 637         spinlock_t *lock;
 638
 639         if (sk_unhashed(sk))
 640                 return;
 641
 642         if (sk->sk_state == TCP_LISTEN) {
 643                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 644                 lock = &ilb->lock;
 645         } else {
 646                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 647         }
 648         spin_lock_bh(lock);
 649         if (sk_unhashed(sk))
 650                 goto unlock;
 651
 652         if (rcu_access_pointer(sk->sk_reuseport_cb))
 653                 reuseport_detach_sock(sk);
 654         if (ilb) {
 655                 inet_unhash2(hashinfo, sk);
 656                 ilb->count--;
 657         }
 658         __sk_nulls_del_node_init_rcu(sk);
 659         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 660 unlock:
 661         spin_unlock_bh(lock);
 662 }
 663 EXPORT_SYMBOL_GPL(inet_unhash);
 664
 665 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 666                 struct sock *sk, u32 port_offset,
 667                 int (*check_established)(struct inet_timewait_death_row *,
 668                         struct sock *, __u16, struct inet_timewait_sock **))
 669 {
 670         struct inet_hashinfo *hinfo = death_row->hashinfo;
 671         struct inet_timewait_sock *tw = NULL;
 672         struct inet_bind_hashbucket *head;
 673         int port = inet_sk(sk)->inet_num;
 674         struct net *net = sock_net(sk);
 675         struct inet_bind_bucket *tb;
 676         u32 remaining, offset;
 677         int ret, i, low, high;
 678         static u32 hint;
 679
 680         if (port) {
 681                 head = &hinfo->bhash[inet_bhashfn(net, port,
 682                                                   hinfo->bhash_size)];
 683                 tb = inet_csk(sk)->icsk_bind_hash;
 684                 spin_lock_bh(&head->lock);
 685                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 686                         inet_ehash_nolisten(sk, NULL);
 687                         spin_unlock_bh(&head->lock);
 688                         return 0;
 689                 }
 690                 spin_unlock(&head->lock);
 691                 /* No definite answer... Walk to established hash table */
 692                 ret = check_established(death_row, sk, port, NULL);
 693                 local_bh_enable();
 694                 return ret;
 695         }
 696
 697         inet_get_local_port_range(net, &low, &high);
 698         high++; /* [32768, 60999] -> [32768, 61000[ */
 699         remaining = high - low;
 700         if (likely(remaining > 1))
 701                 remaining &= ~1U;
 702
 703         offset = (hint + port_offset) % remaining;
 704         /* In first pass we try ports of @low parity.
 705          * inet_csk_get_port() does the opposite choice.
 706          */
 707         offset &= ~1U;
 708 other_parity_scan:
 709         port = low + offset;
 710         for (i = 0; i < remaining; i += 2, port += 2) {
 711                 if (unlikely(port >= high))
 712                         port -= remaining;
 713                 if (inet_is_local_reserved_port(net, port))
 714                         continue;
 715                 head = &hinfo->bhash[inet_bhashfn(net, port,
 716                                                   hinfo->bhash_size)];
 717                 spin_lock_bh(&head->lock);
 718
 719                 /* Does not bother with rcv_saddr checks, because
 720                  * the established check is already unique enough.
 721                  */
 722                 inet_bind_bucket_for_each(tb, &head->chain) {
 723                         if (net_eq(ib_net(tb), net) && tb->port == port) {
 724                                 if (tb->fastreuse >= 0 ||
 725                                     tb->fastreuseport >= 0)
 726                                         goto next_port;
 727                                 WARN_ON(hlist_empty(&tb->owners));
 728                                 if (!check_established(death_row, sk,
 729                                                        port, &tw))
 730                                         goto ok;
 731                                 goto next_port;
 732                         }
 733                 }
 734
 735                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 736                                              net, head, port);
 737                 if (!tb) {
 738                         spin_unlock_bh(&head->lock);
 739                         return -ENOMEM;
 740                 }
 741                 tb->fastreuse = -1;
 742                 tb->fastreuseport = -1;
 743                 goto ok;
 744 next_port:
 745                 spin_unlock_bh(&head->lock);
 746                 cond_resched();
 747         }
 748
 749         offset++;
 750         if ((offset & 1) && remaining > 1)
 751                 goto other_parity_scan;
 752
 753         return -EADDRNOTAVAIL;
 754
 755 ok:
 756         hint += i + 2;
 757
 758         /* Head lock still held and bh's disabled */
 759         inet_bind_hash(sk, tb, port);
 760         if (sk_unhashed(sk)) {
 761                 inet_sk(sk)->inet_sport = htons(port);
 762                 inet_ehash_nolisten(sk, (struct sock *)tw);
 763         }
 764         if (tw)
 765                 inet_twsk_bind_unhash(tw, hinfo);
 766         spin_unlock(&head->lock);
 767         if (tw)
 768                 inet_twsk_deschedule_put(tw);
 769         local_bh_enable();
 770         return 0;
 771 }
 772
 773 /*
 774  * Bind a port for a connect operation and hash it.
 775  */
 776 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 777                       struct sock *sk)
 778 {
 779         u32 port_offset = 0;
 780
 781         if (!inet_sk(sk)->inet_num)
 782                 port_offset = inet_sk_port_offset(sk);
 783         return __inet_hash_connect(death_row, sk, port_offset,
 784                                    __inet_check_established);
 785 }
 786 EXPORT_SYMBOL_GPL(inet_hash_connect);
 787
 788 void inet_hashinfo_init(struct inet_hashinfo *h)
 789 {
 790         int i;
 791
 792         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 793                 spin_lock_init(&h->listening_hash[i].lock);
 794                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
 795                                       i + LISTENING_NULLS_BASE);
 796                 h->listening_hash[i].count = 0;
 797         }
 798
 799         h->lhash2 = NULL;
 800 }
 801 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 802
 803 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
 804                                 unsigned long numentries, int scale,
 805                                 unsigned long low_limit,
 806                                 unsigned long high_limit)
 807 {
 808         unsigned int i;
 809
 810         h->lhash2 = alloc_large_system_hash(name,
 811                                             sizeof(*h->lhash2),
 812                                             numentries,
 813                                             scale,
 814                                             0,
 815                                             NULL,
 816                                             &h->lhash2_mask,
 817                                             low_limit,
 818                                             high_limit);
 819
 820         for (i = 0; i <= h->lhash2_mask; i++) {
 821                 spin_lock_init(&h->lhash2[i].lock);
 822                 INIT_HLIST_HEAD(&h->lhash2[i].head);
 823                 h->lhash2[i].count = 0;
 824         }
 825 }
 826
 827 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 828 {
 829         unsigned int locksz = sizeof(spinlock_t);
 830         unsigned int i, nblocks = 1;
 831
 832         if (locksz != 0) {
 833                 /* allocate 2 cache lines or at least one spinlock per cpu */
 834                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 835                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 836
 837                 /* no more locks than number of hash buckets */
 838                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 839
 840                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
 841                 if (!hashinfo->ehash_locks)
 842                         return -ENOMEM;
 843
 844                 for (i = 0; i < nblocks; i++)
 845                         spin_lock_init(&hashinfo->ehash_locks[i]);
 846         }
 847         hashinfo->ehash_locks_mask = nblocks - 1;
 848         return 0;
 849 }
 850 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);