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/tcp.h>
  29 #include <net/sock_reuseport.h>
  30
  31 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
  32                         const __u16 lport, const __be32 faddr,
  33                         const __be16 fport)
  34 {
  35         static u32 inet_ehash_secret __read_mostly;
  36
  37         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  38
  39         return __inet_ehashfn(laddr, lport, faddr, fport,
  40                               inet_ehash_secret + net_hash_mix(net));
  41 }
  42
  43 /* This function handles inet_sock, but also timewait and request sockets
  44  * for IPv4/IPv6.
  45  */
  46 u32 sk_ehashfn(const struct sock *sk)
  47 {
  48 #if IS_ENABLED(CONFIG_IPV6)
  49         if (sk->sk_family == AF_INET6 &&
  50             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  51                 return inet6_ehashfn(sock_net(sk),
  52                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  53                                      &sk->sk_v6_daddr, sk->sk_dport);
  54 #endif
  55         return inet_ehashfn(sock_net(sk),
  56                             sk->sk_rcv_saddr, sk->sk_num,
  57                             sk->sk_daddr, sk->sk_dport);
  58 }
  59
  60 /*
  61  * Allocate and initialize a new local port bind bucket.
  62  * The bindhash mutex for snum's hash chain must be held here.
  63  */
  64 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  65                                                  struct net *net,
  66                                                  struct inet_bind_hashbucket *head,
  67                                                  const unsigned short snum)
  68 {
  69         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  70
  71         if (tb) {
  72                 write_pnet(&tb->ib_net, net);
  73                 tb->port      = snum;
  74                 tb->fastreuse = 0;
  75                 tb->fastreuseport = 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         inet_csk(sk)->icsk_bind_hash = tb;
  99 }
 100
 101 /*
 102  * Get rid of any references to a local port held by the given sock.
 103  */
 104 static void __inet_put_port(struct sock *sk)
 105 {
 106         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 107         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 108                         hashinfo->bhash_size);
 109         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 110         struct inet_bind_bucket *tb;
 111
 112         spin_lock(&head->lock);
 113         tb = inet_csk(sk)->icsk_bind_hash;
 114         __sk_del_bind_node(sk);
 115         inet_csk(sk)->icsk_bind_hash = NULL;
 116         inet_sk(sk)->inet_num = 0;
 117         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 118         spin_unlock(&head->lock);
 119 }
 120
 121 void inet_put_port(struct sock *sk)
 122 {
 123         local_bh_disable();
 124         __inet_put_port(sk);
 125         local_bh_enable();
 126 }
 127 EXPORT_SYMBOL(inet_put_port);
 128
 129 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 130 {
 131         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 132         unsigned short port = inet_sk(child)->inet_num;
 133         const int bhash = inet_bhashfn(sock_net(sk), port,
 134                         table->bhash_size);
 135         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 136         struct inet_bind_bucket *tb;
 137
 138         spin_lock(&head->lock);
 139         tb = inet_csk(sk)->icsk_bind_hash;
 140         if (unlikely(!tb)) {
 141                 spin_unlock(&head->lock);
 142                 return -ENOENT;
 143         }
 144         if (tb->port != port) {
 145                 /* NOTE: using tproxy and redirecting skbs to a proxy
 146                  * on a different listener port breaks the assumption
 147                  * that the listener socket's icsk_bind_hash is the same
 148                  * as that of the child socket. We have to look up or
 149                  * create a new bind bucket for the child here. */
 150                 inet_bind_bucket_for_each(tb, &head->chain) {
 151                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 152                             tb->port == port)
 153                                 break;
 154                 }
 155                 if (!tb) {
 156                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 157                                                      sock_net(sk), head, port);
 158                         if (!tb) {
 159                                 spin_unlock(&head->lock);
 160                                 return -ENOMEM;
 161                         }
 162                 }
 163         }
 164         inet_bind_hash(child, tb, port);
 165         spin_unlock(&head->lock);
 166
 167         return 0;
 168 }
 169 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 170
 171 static inline int compute_score(struct sock *sk, struct net *net,
 172                                 const unsigned short hnum, const __be32 daddr,
 173                                 const int dif, bool exact_dif)
 174 {
 175         int score = -1;
 176         struct inet_sock *inet = inet_sk(sk);
 177
 178         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 179                         !ipv6_only_sock(sk)) {
 180                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 181                 score = sk->sk_family == PF_INET ? 2 : 1;
 182                 if (rcv_saddr) {
 183                         if (rcv_saddr != daddr)
 184                                 return -1;
 185                         score += 4;
 186                 }
 187                 if (sk->sk_bound_dev_if || exact_dif) {
 188                         if (sk->sk_bound_dev_if != dif)
 189                                 return -1;
 190                         score += 4;
 191                 }
 192                 if (sk->sk_incoming_cpu == raw_smp_processor_id())
 193                         score++;
 194         }
 195         return score;
 196 }
 197
 198 /*
 199  * Here are some nice properties to exploit here. The BSD API
 200  * does not allow a listening sock to specify the remote port nor the
 201  * remote address for the connection. So always assume those are both
 202  * wildcarded during the search since they can never be otherwise.
 203  */
 204
 205 /* called with rcu_read_lock() : No refcount taken on the socket */
 206 struct sock *__inet_lookup_listener(struct net *net,
 207                                     struct inet_hashinfo *hashinfo,
 208                                     struct sk_buff *skb, int doff,
 209                                     const __be32 saddr, __be16 sport,
 210                                     const __be32 daddr, const unsigned short hnum,
 211                                     const int dif)
 212 {
 213         unsigned int hash = inet_lhashfn(net, hnum);
 214         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 215         int score, hiscore = 0, matches = 0, reuseport = 0;
 216         bool exact_dif = inet_exact_dif_match(net, skb);
 217         struct sock *sk, *result = NULL;
 218         u32 phash = 0;
 219
 220         sk_for_each_rcu(sk, &ilb->head) {
 221                 score = compute_score(sk, net, hnum, daddr, dif, exact_dif);
 222                 if (score > hiscore) {
 223                         reuseport = sk->sk_reuseport;
 224                         if (reuseport) {
 225                                 phash = inet_ehashfn(net, daddr, hnum,
 226                                                      saddr, sport);
 227                                 result = reuseport_select_sock(sk, phash,
 228                                                                skb, doff);
 229                                 if (result)
 230                                         return result;
 231                                 matches = 1;
 232                         }
 233                         result = sk;
 234                         hiscore = score;
 235                 } else if (score == hiscore && reuseport) {
 236                         matches++;
 237                         if (reciprocal_scale(phash, matches) == 0)
 238                                 result = sk;
 239                         phash = next_pseudo_random32(phash);
 240                 }
 241         }
 242         return result;
 243 }
 244 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 245
 246 /* All sockets share common refcount, but have different destructors */
 247 void sock_gen_put(struct sock *sk)
 248 {
 249         if (!atomic_dec_and_test(&sk->sk_refcnt))
 250                 return;
 251
 252         if (sk->sk_state == TCP_TIME_WAIT)
 253                 inet_twsk_free(inet_twsk(sk));
 254         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 255                 reqsk_free(inet_reqsk(sk));
 256         else
 257                 sk_free(sk);
 258 }
 259 EXPORT_SYMBOL_GPL(sock_gen_put);
 260
 261 void sock_edemux(struct sk_buff *skb)
 262 {
 263         sock_gen_put(skb->sk);
 264 }
 265 EXPORT_SYMBOL(sock_edemux);
 266
 267 struct sock *__inet_lookup_established(struct net *net,
 268                                   struct inet_hashinfo *hashinfo,
 269                                   const __be32 saddr, const __be16 sport,
 270                                   const __be32 daddr, const u16 hnum,
 271                                   const int dif)
 272 {
 273         INET_ADDR_COOKIE(acookie, saddr, daddr);
 274         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 275         struct sock *sk;
 276         const struct hlist_nulls_node *node;
 277         /* Optimize here for direct hit, only listening connections can
 278          * have wildcards anyways.
 279          */
 280         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 281         unsigned int slot = hash & hashinfo->ehash_mask;
 282         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 283
 284 begin:
 285         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 286                 if (sk->sk_hash != hash)
 287                         continue;
 288                 if (likely(INET_MATCH(sk, net, acookie,
 289                                       saddr, daddr, ports, dif))) {
 290                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 291                                 goto out;
 292                         if (unlikely(!INET_MATCH(sk, net, acookie,
 293                                                  saddr, daddr, ports, dif))) {
 294                                 sock_gen_put(sk);
 295                                 goto begin;
 296                         }
 297                         goto found;
 298                 }
 299         }
 300         /*
 301          * if the nulls value we got at the end of this lookup is
 302          * not the expected one, we must restart lookup.
 303          * We probably met an item that was moved to another chain.
 304          */
 305         if (get_nulls_value(node) != slot)
 306                 goto begin;
 307 out:
 308         sk = NULL;
 309 found:
 310         return sk;
 311 }
 312 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 313
 314 /* called with local bh disabled */
 315 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 316                                     struct sock *sk, __u16 lport,
 317                                     struct inet_timewait_sock **twp)
 318 {
 319         struct inet_hashinfo *hinfo = death_row->hashinfo;
 320         struct inet_sock *inet = inet_sk(sk);
 321         __be32 daddr = inet->inet_rcv_saddr;
 322         __be32 saddr = inet->inet_daddr;
 323         int dif = sk->sk_bound_dev_if;
 324         INET_ADDR_COOKIE(acookie, saddr, daddr);
 325         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 326         struct net *net = sock_net(sk);
 327         unsigned int hash = inet_ehashfn(net, daddr, lport,
 328                                          saddr, inet->inet_dport);
 329         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 330         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 331         struct sock *sk2;
 332         const struct hlist_nulls_node *node;
 333         struct inet_timewait_sock *tw = NULL;
 334
 335         spin_lock(lock);
 336
 337         sk_nulls_for_each(sk2, node, &head->chain) {
 338                 if (sk2->sk_hash != hash)
 339                         continue;
 340
 341                 if (likely(INET_MATCH(sk2, net, acookie,
 342                                          saddr, daddr, ports, dif))) {
 343                         if (sk2->sk_state == TCP_TIME_WAIT) {
 344                                 tw = inet_twsk(sk2);
 345                                 if (twsk_unique(sk, sk2, twp))
 346                                         break;
 347                         }
 348                         goto not_unique;
 349                 }
 350         }
 351
 352         /* Must record num and sport now. Otherwise we will see
 353          * in hash table socket with a funny identity.
 354          */
 355         inet->inet_num = lport;
 356         inet->inet_sport = htons(lport);
 357         sk->sk_hash = hash;
 358         WARN_ON(!sk_unhashed(sk));
 359         __sk_nulls_add_node_rcu(sk, &head->chain);
 360         if (tw) {
 361                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 362                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 363         }
 364         spin_unlock(lock);
 365         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 366
 367         if (twp) {
 368                 *twp = tw;
 369         } else if (tw) {
 370                 /* Silly. Should hash-dance instead... */
 371                 inet_twsk_deschedule_put(tw);
 372         }
 373         return 0;
 374
 375 not_unique:
 376         spin_unlock(lock);
 377         return -EADDRNOTAVAIL;
 378 }
 379
 380 static u32 inet_sk_port_offset(const struct sock *sk)
 381 {
 382         const struct inet_sock *inet = inet_sk(sk);
 383
 384         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 385                                           inet->inet_daddr,
 386                                           inet->inet_dport);
 387 }
 388
 389 /* insert a socket into ehash, and eventually remove another one
 390  * (The another one can be a SYN_RECV or TIMEWAIT
 391  */
 392 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
 393 {
 394         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 395         struct hlist_nulls_head *list;
 396         struct inet_ehash_bucket *head;
 397         spinlock_t *lock;
 398         bool ret = true;
 399
 400         WARN_ON_ONCE(!sk_unhashed(sk));
 401
 402         sk->sk_hash = sk_ehashfn(sk);
 403         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 404         list = &head->chain;
 405         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 406
 407         spin_lock(lock);
 408         if (osk) {
 409                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 410                 ret = sk_nulls_del_node_init_rcu(osk);
 411         }
 412         if (ret)
 413                 __sk_nulls_add_node_rcu(sk, list);
 414         spin_unlock(lock);
 415         return ret;
 416 }
 417
 418 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
 419 {
 420         bool ok = inet_ehash_insert(sk, osk);
 421
 422         if (ok) {
 423                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 424         } else {
 425                 percpu_counter_inc(sk->sk_prot->orphan_count);
 426                 sk->sk_state = TCP_CLOSE;
 427                 sock_set_flag(sk, SOCK_DEAD);
 428                 inet_csk_destroy_sock(sk);
 429         }
 430         return ok;
 431 }
 432 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 433
 434 static int inet_reuseport_add_sock(struct sock *sk,
 435                                    struct inet_listen_hashbucket *ilb)
 436 {
 437         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 438         struct sock *sk2;
 439         kuid_t uid = sock_i_uid(sk);
 440
 441         sk_for_each_rcu(sk2, &ilb->head) {
 442                 if (sk2 != sk &&
 443                     sk2->sk_family == sk->sk_family &&
 444                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 445                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 446                     inet_csk(sk2)->icsk_bind_hash == tb &&
 447                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 448                     inet_rcv_saddr_equal(sk, sk2, false))
 449                         return reuseport_add_sock(sk, sk2);
 450         }
 451
 452         /* Initial allocation may have already happened via setsockopt */
 453         if (!rcu_access_pointer(sk->sk_reuseport_cb))
 454                 return reuseport_alloc(sk);
 455         return 0;
 456 }
 457
 458 int __inet_hash(struct sock *sk, struct sock *osk)
 459 {
 460         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 461         struct inet_listen_hashbucket *ilb;
 462         int err = 0;
 463
 464         if (sk->sk_state != TCP_LISTEN) {
 465                 inet_ehash_nolisten(sk, osk);
 466                 return 0;
 467         }
 468         WARN_ON(!sk_unhashed(sk));
 469         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 470
 471         spin_lock(&ilb->lock);
 472         if (sk->sk_reuseport) {
 473                 err = inet_reuseport_add_sock(sk, ilb);
 474                 if (err)
 475                         goto unlock;
 476         }
 477         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 478                 sk->sk_family == AF_INET6)
 479                 hlist_add_tail_rcu(&sk->sk_node, &ilb->head);
 480         else
 481                 hlist_add_head_rcu(&sk->sk_node, &ilb->head);
 482         sock_set_flag(sk, SOCK_RCU_FREE);
 483         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 484 unlock:
 485         spin_unlock(&ilb->lock);
 486
 487         return err;
 488 }
 489 EXPORT_SYMBOL(__inet_hash);
 490
 491 int inet_hash(struct sock *sk)
 492 {
 493         int err = 0;
 494
 495         if (sk->sk_state != TCP_CLOSE) {
 496                 local_bh_disable();
 497                 err = __inet_hash(sk, NULL);
 498                 local_bh_enable();
 499         }
 500
 501         return err;
 502 }
 503 EXPORT_SYMBOL_GPL(inet_hash);
 504
 505 void inet_unhash(struct sock *sk)
 506 {
 507         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 508         spinlock_t *lock;
 509         bool listener = false;
 510         int done;
 511
 512         if (sk_unhashed(sk))
 513                 return;
 514
 515         if (sk->sk_state == TCP_LISTEN) {
 516                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 517                 listener = true;
 518         } else {
 519                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 520         }
 521         spin_lock_bh(lock);
 522         if (rcu_access_pointer(sk->sk_reuseport_cb))
 523                 reuseport_detach_sock(sk);
 524         if (listener)
 525                 done = __sk_del_node_init(sk);
 526         else
 527                 done = __sk_nulls_del_node_init_rcu(sk);
 528         if (done)
 529                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 530         spin_unlock_bh(lock);
 531 }
 532 EXPORT_SYMBOL_GPL(inet_unhash);
 533
 534 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 535                 struct sock *sk, u32 port_offset,
 536                 int (*check_established)(struct inet_timewait_death_row *,
 537                         struct sock *, __u16, struct inet_timewait_sock **))
 538 {
 539         struct inet_hashinfo *hinfo = death_row->hashinfo;
 540         struct inet_timewait_sock *tw = NULL;
 541         struct inet_bind_hashbucket *head;
 542         int port = inet_sk(sk)->inet_num;
 543         struct net *net = sock_net(sk);
 544         struct inet_bind_bucket *tb;
 545         u32 remaining, offset;
 546         int ret, i, low, high;
 547         static u32 hint;
 548
 549         if (port) {
 550                 head = &hinfo->bhash[inet_bhashfn(net, port,
 551                                                   hinfo->bhash_size)];
 552                 tb = inet_csk(sk)->icsk_bind_hash;
 553                 spin_lock_bh(&head->lock);
 554                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 555                         inet_ehash_nolisten(sk, NULL);
 556                         spin_unlock_bh(&head->lock);
 557                         return 0;
 558                 }
 559                 spin_unlock(&head->lock);
 560                 /* No definite answer... Walk to established hash table */
 561                 ret = check_established(death_row, sk, port, NULL);
 562                 local_bh_enable();
 563                 return ret;
 564         }
 565
 566         inet_get_local_port_range(net, &low, &high);
 567         high++; /* [32768, 60999] -> [32768, 61000[ */
 568         remaining = high - low;
 569         if (likely(remaining > 1))
 570                 remaining &= ~1U;
 571
 572         offset = (hint + port_offset) % remaining;
 573         /* In first pass we try ports of @low parity.
 574          * inet_csk_get_port() does the opposite choice.
 575          */
 576         offset &= ~1U;
 577 other_parity_scan:
 578         port = low + offset;
 579         for (i = 0; i < remaining; i += 2, port += 2) {
 580                 if (unlikely(port >= high))
 581                         port -= remaining;
 582                 if (inet_is_local_reserved_port(net, port))
 583                         continue;
 584                 head = &hinfo->bhash[inet_bhashfn(net, port,
 585                                                   hinfo->bhash_size)];
 586                 spin_lock_bh(&head->lock);
 587
 588                 /* Does not bother with rcv_saddr checks, because
 589                  * the established check is already unique enough.
 590                  */
 591                 inet_bind_bucket_for_each(tb, &head->chain) {
 592                         if (net_eq(ib_net(tb), net) && tb->port == port) {
 593                                 if (tb->fastreuse >= 0 ||
 594                                     tb->fastreuseport >= 0)
 595                                         goto next_port;
 596                                 WARN_ON(hlist_empty(&tb->owners));
 597                                 if (!check_established(death_row, sk,
 598                                                        port, &tw))
 599                                         goto ok;
 600                                 goto next_port;
 601                         }
 602                 }
 603
 604                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 605                                              net, head, port);
 606                 if (!tb) {
 607                         spin_unlock_bh(&head->lock);
 608                         return -ENOMEM;
 609                 }
 610                 tb->fastreuse = -1;
 611                 tb->fastreuseport = -1;
 612                 goto ok;
 613 next_port:
 614                 spin_unlock_bh(&head->lock);
 615                 cond_resched();
 616         }
 617
 618         offset++;
 619         if ((offset & 1) && remaining > 1)
 620                 goto other_parity_scan;
 621
 622         return -EADDRNOTAVAIL;
 623
 624 ok:
 625         hint += i + 2;
 626
 627         /* Head lock still held and bh's disabled */
 628         inet_bind_hash(sk, tb, port);
 629         if (sk_unhashed(sk)) {
 630                 inet_sk(sk)->inet_sport = htons(port);
 631                 inet_ehash_nolisten(sk, (struct sock *)tw);
 632         }
 633         if (tw)
 634                 inet_twsk_bind_unhash(tw, hinfo);
 635         spin_unlock(&head->lock);
 636         if (tw)
 637                 inet_twsk_deschedule_put(tw);
 638         local_bh_enable();
 639         return 0;
 640 }
 641
 642 /*
 643  * Bind a port for a connect operation and hash it.
 644  */
 645 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 646                       struct sock *sk)
 647 {
 648         u32 port_offset = 0;
 649
 650         if (!inet_sk(sk)->inet_num)
 651                 port_offset = inet_sk_port_offset(sk);
 652         return __inet_hash_connect(death_row, sk, port_offset,
 653                                    __inet_check_established);
 654 }
 655 EXPORT_SYMBOL_GPL(inet_hash_connect);
 656
 657 void inet_hashinfo_init(struct inet_hashinfo *h)
 658 {
 659         int i;
 660
 661         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 662                 spin_lock_init(&h->listening_hash[i].lock);
 663                 INIT_HLIST_HEAD(&h->listening_hash[i].head);
 664         }
 665 }
 666 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 667
 668 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 669 {
 670         unsigned int locksz = sizeof(spinlock_t);
 671         unsigned int i, nblocks = 1;
 672
 673         if (locksz != 0) {
 674                 /* allocate 2 cache lines or at least one spinlock per cpu */
 675                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 676                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 677
 678                 /* no more locks than number of hash buckets */
 679                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 680
 681                 hashinfo->ehash_locks = kmalloc_array(nblocks, locksz,
 682                                                       GFP_KERNEL | __GFP_NOWARN);
 683                 if (!hashinfo->ehash_locks)
 684                         hashinfo->ehash_locks = vmalloc(nblocks * locksz);
 685
 686                 if (!hashinfo->ehash_locks)
 687                         return -ENOMEM;
 688
 689                 for (i = 0; i < nblocks; i++)
 690                         spin_lock_init(&hashinfo->ehash_locks[i]);
 691         }
 692         hashinfo->ehash_locks_mask = nblocks - 1;
 693         return 0;
 694 }
 695 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);