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
  23 #include <net/addrconf.h>
  24 #include <net/inet_connection_sock.h>
  25 #include <net/inet_hashtables.h>
  26 #include <net/secure_seq.h>
  27 #include <net/ip.h>
  28 #include <net/sock_reuseport.h>
  29
  30 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
  31                         const __u16 lport, const __be32 faddr,
  32                         const __be16 fport)
  33 {
  34         static u32 inet_ehash_secret __read_mostly;
  35
  36         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  37
  38         return __inet_ehashfn(laddr, lport, faddr, fport,
  39                               inet_ehash_secret + net_hash_mix(net));
  40 }
  41
  42 /* This function handles inet_sock, but also timewait and request sockets
  43  * for IPv4/IPv6.
  44  */
  45 u32 sk_ehashfn(const struct sock *sk)
  46 {
  47 #if IS_ENABLED(CONFIG_IPV6)
  48         if (sk->sk_family == AF_INET6 &&
  49             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  50                 return inet6_ehashfn(sock_net(sk),
  51                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  52                                      &sk->sk_v6_daddr, sk->sk_dport);
  53 #endif
  54         return inet_ehashfn(sock_net(sk),
  55                             sk->sk_rcv_saddr, sk->sk_num,
  56                             sk->sk_daddr, sk->sk_dport);
  57 }
  58
  59 /*
  60  * Allocate and initialize a new local port bind bucket.
  61  * The bindhash mutex for snum's hash chain must be held here.
  62  */
  63 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  64                                                  struct net *net,
  65                                                  struct inet_bind_hashbucket *head,
  66                                                  const unsigned short snum)
  67 {
  68         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  69
  70         if (tb) {
  71                 write_pnet(&tb->ib_net, net);
  72                 tb->port      = snum;
  73                 tb->fastreuse = 0;
  74                 tb->fastreuseport = 0;
  75                 tb->num_owners = 0;
  76                 INIT_HLIST_HEAD(&tb->owners);
  77                 hlist_add_head(&tb->node, &head->chain);
  78         }
  79         return tb;
  80 }
  81
  82 /*
  83  * Caller must hold hashbucket lock for this tb with local BH disabled
  84  */
  85 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  86 {
  87         if (hlist_empty(&tb->owners)) {
  88                 __hlist_del(&tb->node);
  89                 kmem_cache_free(cachep, tb);
  90         }
  91 }
  92
  93 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  94                     const unsigned short snum)
  95 {
  96         inet_sk(sk)->inet_num = snum;
  97         sk_add_bind_node(sk, &tb->owners);
  98         tb->num_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         tb->num_owners--;
 117         inet_csk(sk)->icsk_bind_hash = NULL;
 118         inet_sk(sk)->inet_num = 0;
 119         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 120         spin_unlock(&head->lock);
 121 }
 122
 123 void inet_put_port(struct sock *sk)
 124 {
 125         local_bh_disable();
 126         __inet_put_port(sk);
 127         local_bh_enable();
 128 }
 129 EXPORT_SYMBOL(inet_put_port);
 130
 131 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 132 {
 133         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 134         unsigned short port = inet_sk(child)->inet_num;
 135         const int bhash = inet_bhashfn(sock_net(sk), port,
 136                         table->bhash_size);
 137         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 138         struct inet_bind_bucket *tb;
 139
 140         spin_lock(&head->lock);
 141         tb = inet_csk(sk)->icsk_bind_hash;
 142         if (unlikely(!tb)) {
 143                 spin_unlock(&head->lock);
 144                 return -ENOENT;
 145         }
 146         if (tb->port != port) {
 147                 /* NOTE: using tproxy and redirecting skbs to a proxy
 148                  * on a different listener port breaks the assumption
 149                  * that the listener socket's icsk_bind_hash is the same
 150                  * as that of the child socket. We have to look up or
 151                  * create a new bind bucket for the child here. */
 152                 inet_bind_bucket_for_each(tb, &head->chain) {
 153                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 154                             tb->port == port)
 155                                 break;
 156                 }
 157                 if (!tb) {
 158                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 159                                                      sock_net(sk), head, port);
 160                         if (!tb) {
 161                                 spin_unlock(&head->lock);
 162                                 return -ENOMEM;
 163                         }
 164                 }
 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 inline int compute_score(struct sock *sk, struct net *net,
 174                                 const unsigned short hnum, const __be32 daddr,
 175                                 const int dif)
 176 {
 177         int score = -1;
 178         struct inet_sock *inet = inet_sk(sk);
 179
 180         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 181                         !ipv6_only_sock(sk)) {
 182                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 183                 score = sk->sk_family == PF_INET ? 2 : 1;
 184                 if (rcv_saddr) {
 185                         if (rcv_saddr != daddr)
 186                                 return -1;
 187                         score += 4;
 188                 }
 189                 if (sk->sk_bound_dev_if) {
 190                         if (sk->sk_bound_dev_if != dif)
 191                                 return -1;
 192                         score += 4;
 193                 }
 194                 if (sk->sk_incoming_cpu == raw_smp_processor_id())
 195                         score++;
 196         }
 197         return score;
 198 }
 199
 200 /*
 201  * Here are some nice properties to exploit here. The BSD API
 202  * does not allow a listening sock to specify the remote port nor the
 203  * remote address for the connection. So always assume those are both
 204  * wildcarded during the search since they can never be otherwise.
 205  */
 206
 207 /* called with rcu_read_lock() : No refcount taken on the socket */
 208 struct sock *__inet_lookup_listener(struct net *net,
 209                                     struct inet_hashinfo *hashinfo,
 210                                     struct sk_buff *skb, int doff,
 211                                     const __be32 saddr, __be16 sport,
 212                                     const __be32 daddr, const unsigned short hnum,
 213                                     const int dif)
 214 {
 215         unsigned int hash = inet_lhashfn(net, hnum);
 216         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 217         int score, hiscore = 0, matches = 0, reuseport = 0;
 218         struct sock *sk, *result = NULL;
 219         u32 phash = 0;
 220
 221         sk_for_each_rcu(sk, &ilb->head) {
 222                 score = compute_score(sk, net, hnum, daddr, dif);
 223                 if (score > hiscore) {
 224                         reuseport = sk->sk_reuseport;
 225                         if (reuseport) {
 226                                 phash = inet_ehashfn(net, daddr, hnum,
 227                                                      saddr, sport);
 228                                 result = reuseport_select_sock(sk, phash,
 229                                                                skb, doff);
 230                                 if (result)
 231                                         return result;
 232                                 matches = 1;
 233                         }
 234                         result = sk;
 235                         hiscore = score;
 236                 } else if (score == hiscore && reuseport) {
 237                         matches++;
 238                         if (reciprocal_scale(phash, matches) == 0)
 239                                 result = sk;
 240                         phash = next_pseudo_random32(phash);
 241                 }
 242         }
 243         return result;
 244 }
 245 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 246
 247 /* All sockets share common refcount, but have different destructors */
 248 void sock_gen_put(struct sock *sk)
 249 {
 250         if (!atomic_dec_and_test(&sk->sk_refcnt))
 251                 return;
 252
 253         if (sk->sk_state == TCP_TIME_WAIT)
 254                 inet_twsk_free(inet_twsk(sk));
 255         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 256                 reqsk_free(inet_reqsk(sk));
 257         else
 258                 sk_free(sk);
 259 }
 260 EXPORT_SYMBOL_GPL(sock_gen_put);
 261
 262 void sock_edemux(struct sk_buff *skb)
 263 {
 264         sock_gen_put(skb->sk);
 265 }
 266 EXPORT_SYMBOL(sock_edemux);
 267
 268 struct sock *__inet_lookup_established(struct net *net,
 269                                   struct inet_hashinfo *hashinfo,
 270                                   const __be32 saddr, const __be16 sport,
 271                                   const __be32 daddr, const u16 hnum,
 272                                   const int dif)
 273 {
 274         INET_ADDR_COOKIE(acookie, saddr, daddr);
 275         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 276         struct sock *sk;
 277         const struct hlist_nulls_node *node;
 278         /* Optimize here for direct hit, only listening connections can
 279          * have wildcards anyways.
 280          */
 281         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 282         unsigned int slot = hash & hashinfo->ehash_mask;
 283         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 284
 285 begin:
 286         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 287                 if (sk->sk_hash != hash)
 288                         continue;
 289                 if (likely(INET_MATCH(sk, net, acookie,
 290                                       saddr, daddr, ports, dif))) {
 291                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 292                                 goto out;
 293                         if (unlikely(!INET_MATCH(sk, net, acookie,
 294                                                  saddr, daddr, ports, dif))) {
 295                                 sock_gen_put(sk);
 296                                 goto begin;
 297                         }
 298                         goto found;
 299                 }
 300         }
 301         /*
 302          * if the nulls value we got at the end of this lookup is
 303          * not the expected one, we must restart lookup.
 304          * We probably met an item that was moved to another chain.
 305          */
 306         if (get_nulls_value(node) != slot)
 307                 goto begin;
 308 out:
 309         sk = NULL;
 310 found:
 311         return sk;
 312 }
 313 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 314
 315 /* called with local bh disabled */
 316 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 317                                     struct sock *sk, __u16 lport,
 318                                     struct inet_timewait_sock **twp)
 319 {
 320         struct inet_hashinfo *hinfo = death_row->hashinfo;
 321         struct inet_sock *inet = inet_sk(sk);
 322         __be32 daddr = inet->inet_rcv_saddr;
 323         __be32 saddr = inet->inet_daddr;
 324         int dif = sk->sk_bound_dev_if;
 325         INET_ADDR_COOKIE(acookie, saddr, daddr);
 326         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 327         struct net *net = sock_net(sk);
 328         unsigned int hash = inet_ehashfn(net, daddr, lport,
 329                                          saddr, inet->inet_dport);
 330         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 331         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 332         struct sock *sk2;
 333         const struct hlist_nulls_node *node;
 334         struct inet_timewait_sock *tw = NULL;
 335
 336         spin_lock(lock);
 337
 338         sk_nulls_for_each(sk2, node, &head->chain) {
 339                 if (sk2->sk_hash != hash)
 340                         continue;
 341
 342                 if (likely(INET_MATCH(sk2, net, acookie,
 343                                          saddr, daddr, ports, dif))) {
 344                         if (sk2->sk_state == TCP_TIME_WAIT) {
 345                                 tw = inet_twsk(sk2);
 346                                 if (twsk_unique(sk, sk2, twp))
 347                                         break;
 348                         }
 349                         goto not_unique;
 350                 }
 351         }
 352
 353         /* Must record num and sport now. Otherwise we will see
 354          * in hash table socket with a funny identity.
 355          */
 356         inet->inet_num = lport;
 357         inet->inet_sport = htons(lport);
 358         sk->sk_hash = hash;
 359         WARN_ON(!sk_unhashed(sk));
 360         __sk_nulls_add_node_rcu(sk, &head->chain);
 361         if (tw) {
 362                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 363                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 364         }
 365         spin_unlock(lock);
 366         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 367
 368         if (twp) {
 369                 *twp = tw;
 370         } else if (tw) {
 371                 /* Silly. Should hash-dance instead... */
 372                 inet_twsk_deschedule_put(tw);
 373         }
 374         return 0;
 375
 376 not_unique:
 377         spin_unlock(lock);
 378         return -EADDRNOTAVAIL;
 379 }
 380
 381 static u32 inet_sk_port_offset(const struct sock *sk)
 382 {
 383         const struct inet_sock *inet = inet_sk(sk);
 384
 385         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 386                                           inet->inet_daddr,
 387                                           inet->inet_dport);
 388 }
 389
 390 /* insert a socket into ehash, and eventually remove another one
 391  * (The another one can be a SYN_RECV or TIMEWAIT
 392  */
 393 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
 394 {
 395         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 396         struct hlist_nulls_head *list;
 397         struct inet_ehash_bucket *head;
 398         spinlock_t *lock;
 399         bool ret = true;
 400
 401         WARN_ON_ONCE(!sk_unhashed(sk));
 402
 403         sk->sk_hash = sk_ehashfn(sk);
 404         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 405         list = &head->chain;
 406         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 407
 408         spin_lock(lock);
 409         if (osk) {
 410                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 411                 ret = sk_nulls_del_node_init_rcu(osk);
 412         }
 413         if (ret)
 414                 __sk_nulls_add_node_rcu(sk, list);
 415         spin_unlock(lock);
 416         return ret;
 417 }
 418
 419 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
 420 {
 421         bool ok = inet_ehash_insert(sk, osk);
 422
 423         if (ok) {
 424                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 425         } else {
 426                 percpu_counter_inc(sk->sk_prot->orphan_count);
 427                 sk->sk_state = TCP_CLOSE;
 428                 sock_set_flag(sk, SOCK_DEAD);
 429                 inet_csk_destroy_sock(sk);
 430         }
 431         return ok;
 432 }
 433 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 434
 435 static int inet_reuseport_add_sock(struct sock *sk,
 436                                    struct inet_listen_hashbucket *ilb,
 437                                    int (*saddr_same)(const struct sock *sk1,
 438                                                      const struct sock *sk2,
 439                                                      bool match_wildcard))
 440 {
 441         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 442         struct sock *sk2;
 443         kuid_t uid = sock_i_uid(sk);
 444
 445         sk_for_each_rcu(sk2, &ilb->head) {
 446                 if (sk2 != sk &&
 447                     sk2->sk_family == sk->sk_family &&
 448                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 449                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 450                     inet_csk(sk2)->icsk_bind_hash == tb &&
 451                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 452                     saddr_same(sk, sk2, false))
 453                         return reuseport_add_sock(sk, sk2);
 454         }
 455
 456         /* Initial allocation may have already happened via setsockopt */
 457         if (!rcu_access_pointer(sk->sk_reuseport_cb))
 458                 return reuseport_alloc(sk);
 459         return 0;
 460 }
 461
 462 int __inet_hash(struct sock *sk, struct sock *osk,
 463                  int (*saddr_same)(const struct sock *sk1,
 464                                    const struct sock *sk2,
 465                                    bool match_wildcard))
 466 {
 467         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 468         struct inet_listen_hashbucket *ilb;
 469         int err = 0;
 470
 471         if (sk->sk_state != TCP_LISTEN) {
 472                 inet_ehash_nolisten(sk, osk);
 473                 return 0;
 474         }
 475         WARN_ON(!sk_unhashed(sk));
 476         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 477
 478         spin_lock(&ilb->lock);
 479         if (sk->sk_reuseport) {
 480                 err = inet_reuseport_add_sock(sk, ilb, saddr_same);
 481                 if (err)
 482                         goto unlock;
 483         }
 484         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 485                 sk->sk_family == AF_INET6)
 486                 hlist_add_tail_rcu(&sk->sk_node, &ilb->head);
 487         else
 488                 hlist_add_head_rcu(&sk->sk_node, &ilb->head);
 489         sock_set_flag(sk, SOCK_RCU_FREE);
 490         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 491 unlock:
 492         spin_unlock(&ilb->lock);
 493
 494         return err;
 495 }
 496 EXPORT_SYMBOL(__inet_hash);
 497
 498 int inet_hash(struct sock *sk)
 499 {
 500         int err = 0;
 501
 502         if (sk->sk_state != TCP_CLOSE) {
 503                 local_bh_disable();
 504                 err = __inet_hash(sk, NULL, ipv4_rcv_saddr_equal);
 505                 local_bh_enable();
 506         }
 507
 508         return err;
 509 }
 510 EXPORT_SYMBOL_GPL(inet_hash);
 511
 512 void inet_unhash(struct sock *sk)
 513 {
 514         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 515         spinlock_t *lock;
 516         bool listener = false;
 517         int done;
 518
 519         if (sk_unhashed(sk))
 520                 return;
 521
 522         if (sk->sk_state == TCP_LISTEN) {
 523                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 524                 listener = true;
 525         } else {
 526                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 527         }
 528         spin_lock_bh(lock);
 529         if (rcu_access_pointer(sk->sk_reuseport_cb))
 530                 reuseport_detach_sock(sk);
 531         if (listener)
 532                 done = __sk_del_node_init(sk);
 533         else
 534                 done = __sk_nulls_del_node_init_rcu(sk);
 535         if (done)
 536                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 537         spin_unlock_bh(lock);
 538 }
 539 EXPORT_SYMBOL_GPL(inet_unhash);
 540
 541 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 542                 struct sock *sk, u32 port_offset,
 543                 int (*check_established)(struct inet_timewait_death_row *,
 544                         struct sock *, __u16, struct inet_timewait_sock **))
 545 {
 546         struct inet_hashinfo *hinfo = death_row->hashinfo;
 547         struct inet_timewait_sock *tw = NULL;
 548         struct inet_bind_hashbucket *head;
 549         int port = inet_sk(sk)->inet_num;
 550         struct net *net = sock_net(sk);
 551         struct inet_bind_bucket *tb;
 552         u32 remaining, offset;
 553         int ret, i, low, high;
 554         static u32 hint;
 555
 556         if (port) {
 557                 head = &hinfo->bhash[inet_bhashfn(net, port,
 558                                                   hinfo->bhash_size)];
 559                 tb = inet_csk(sk)->icsk_bind_hash;
 560                 spin_lock_bh(&head->lock);
 561                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 562                         inet_ehash_nolisten(sk, NULL);
 563                         spin_unlock_bh(&head->lock);
 564                         return 0;
 565                 }
 566                 spin_unlock(&head->lock);
 567                 /* No definite answer... Walk to established hash table */
 568                 ret = check_established(death_row, sk, port, NULL);
 569                 local_bh_enable();
 570                 return ret;
 571         }
 572
 573         inet_get_local_port_range(net, &low, &high);
 574         high++; /* [32768, 60999] -> [32768, 61000[ */
 575         remaining = high - low;
 576         if (likely(remaining > 1))
 577                 remaining &= ~1U;
 578
 579         offset = (hint + port_offset) % remaining;
 580         /* In first pass we try ports of @low parity.
 581          * inet_csk_get_port() does the opposite choice.
 582          */
 583         offset &= ~1U;
 584 other_parity_scan:
 585         port = low + offset;
 586         for (i = 0; i < remaining; i += 2, port += 2) {
 587                 if (unlikely(port >= high))
 588                         port -= remaining;
 589                 if (inet_is_local_reserved_port(net, port))
 590                         continue;
 591                 head = &hinfo->bhash[inet_bhashfn(net, port,
 592                                                   hinfo->bhash_size)];
 593                 spin_lock_bh(&head->lock);
 594
 595                 /* Does not bother with rcv_saddr checks, because
 596                  * the established check is already unique enough.
 597                  */
 598                 inet_bind_bucket_for_each(tb, &head->chain) {
 599                         if (net_eq(ib_net(tb), net) && tb->port == port) {
 600                                 if (tb->fastreuse >= 0 ||
 601                                     tb->fastreuseport >= 0)
 602                                         goto next_port;
 603                                 WARN_ON(hlist_empty(&tb->owners));
 604                                 if (!check_established(death_row, sk,
 605                                                        port, &tw))
 606                                         goto ok;
 607                                 goto next_port;
 608                         }
 609                 }
 610
 611                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 612                                              net, head, port);
 613                 if (!tb) {
 614                         spin_unlock_bh(&head->lock);
 615                         return -ENOMEM;
 616                 }
 617                 tb->fastreuse = -1;
 618                 tb->fastreuseport = -1;
 619                 goto ok;
 620 next_port:
 621                 spin_unlock_bh(&head->lock);
 622                 cond_resched();
 623         }
 624
 625         offset++;
 626         if ((offset & 1) && remaining > 1)
 627                 goto other_parity_scan;
 628
 629         return -EADDRNOTAVAIL;
 630
 631 ok:
 632         hint += i + 2;
 633
 634         /* Head lock still held and bh's disabled */
 635         inet_bind_hash(sk, tb, port);
 636         if (sk_unhashed(sk)) {
 637                 inet_sk(sk)->inet_sport = htons(port);
 638                 inet_ehash_nolisten(sk, (struct sock *)tw);
 639         }
 640         if (tw)
 641                 inet_twsk_bind_unhash(tw, hinfo);
 642         spin_unlock(&head->lock);
 643         if (tw)
 644                 inet_twsk_deschedule_put(tw);
 645         local_bh_enable();
 646         return 0;
 647 }
 648
 649 /*
 650  * Bind a port for a connect operation and hash it.
 651  */
 652 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 653                       struct sock *sk)
 654 {
 655         u32 port_offset = 0;
 656
 657         if (!inet_sk(sk)->inet_num)
 658                 port_offset = inet_sk_port_offset(sk);
 659         return __inet_hash_connect(death_row, sk, port_offset,
 660                                    __inet_check_established);
 661 }
 662 EXPORT_SYMBOL_GPL(inet_hash_connect);
 663
 664 void inet_hashinfo_init(struct inet_hashinfo *h)
 665 {
 666         int i;
 667
 668         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 669                 spin_lock_init(&h->listening_hash[i].lock);
 670                 INIT_HLIST_HEAD(&h->listening_hash[i].head);
 671         }
 672 }
 673 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 674
 675 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 676 {
 677         unsigned int locksz = sizeof(spinlock_t);
 678         unsigned int i, nblocks = 1;
 679
 680         if (locksz != 0) {
 681                 /* allocate 2 cache lines or at least one spinlock per cpu */
 682                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 683                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 684
 685                 /* no more locks than number of hash buckets */
 686                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 687
 688                 hashinfo->ehash_locks = kmalloc_array(nblocks, locksz,
 689                                                       GFP_KERNEL | __GFP_NOWARN);
 690                 if (!hashinfo->ehash_locks)
 691                         hashinfo->ehash_locks = vmalloc(nblocks * locksz);
 692
 693                 if (!hashinfo->ehash_locks)
 694                         return -ENOMEM;
 695
 696                 for (i = 0; i < nblocks; i++)
 697                         spin_lock_init(&hashinfo->ehash_locks[i]);
 698         }
 699         hashinfo->ehash_locks_mask = nblocks - 1;
 700         return 0;
 701 }
 702 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);