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 static 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, const int sdif, 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                         bool dev_match = (sk->sk_bound_dev_if == dif ||
 189                                           sk->sk_bound_dev_if == sdif);
 190
 191                         if (exact_dif && !dev_match)
 192                                 return -1;
 193                         if (sk->sk_bound_dev_if && dev_match)
 194                                 score += 4;
 195                 }
 196                 if (sk->sk_incoming_cpu == raw_smp_processor_id())
 197                         score++;
 198         }
 199         return score;
 200 }
 201
 202 /*
 203  * Here are some nice properties to exploit here. The BSD API
 204  * does not allow a listening sock to specify the remote port nor the
 205  * remote address for the connection. So always assume those are both
 206  * wildcarded during the search since they can never be otherwise.
 207  */
 208
 209 /* called with rcu_read_lock() : No refcount taken on the socket */
 210 struct sock *__inet_lookup_listener(struct net *net,
 211                                     struct inet_hashinfo *hashinfo,
 212                                     struct sk_buff *skb, int doff,
 213                                     const __be32 saddr, __be16 sport,
 214                                     const __be32 daddr, const unsigned short hnum,
 215                                     const int dif, const int sdif)
 216 {
 217         unsigned int hash = inet_lhashfn(net, hnum);
 218         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 219         int score, hiscore = 0, matches = 0, reuseport = 0;
 220         bool exact_dif = inet_exact_dif_match(net, skb);
 221         struct sock *sk, *result = NULL;
 222         u32 phash = 0;
 223
 224         sk_for_each_rcu(sk, &ilb->head) {
 225                 score = compute_score(sk, net, hnum, daddr,
 226                                       dif, sdif, exact_dif);
 227                 if (score > hiscore) {
 228                         reuseport = sk->sk_reuseport;
 229                         if (reuseport) {
 230                                 phash = inet_ehashfn(net, daddr, hnum,
 231                                                      saddr, sport);
 232                                 result = reuseport_select_sock(sk, phash,
 233                                                                skb, doff);
 234                                 if (result)
 235                                         return result;
 236                                 matches = 1;
 237                         }
 238                         result = sk;
 239                         hiscore = score;
 240                 } else if (score == hiscore && reuseport) {
 241                         matches++;
 242                         if (reciprocal_scale(phash, matches) == 0)
 243                                 result = sk;
 244                         phash = next_pseudo_random32(phash);
 245                 }
 246         }
 247         return result;
 248 }
 249 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 250
 251 /* All sockets share common refcount, but have different destructors */
 252 void sock_gen_put(struct sock *sk)
 253 {
 254         if (!refcount_dec_and_test(&sk->sk_refcnt))
 255                 return;
 256
 257         if (sk->sk_state == TCP_TIME_WAIT)
 258                 inet_twsk_free(inet_twsk(sk));
 259         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 260                 reqsk_free(inet_reqsk(sk));
 261         else
 262                 sk_free(sk);
 263 }
 264 EXPORT_SYMBOL_GPL(sock_gen_put);
 265
 266 void sock_edemux(struct sk_buff *skb)
 267 {
 268         sock_gen_put(skb->sk);
 269 }
 270 EXPORT_SYMBOL(sock_edemux);
 271
 272 struct sock *__inet_lookup_established(struct net *net,
 273                                   struct inet_hashinfo *hashinfo,
 274                                   const __be32 saddr, const __be16 sport,
 275                                   const __be32 daddr, const u16 hnum,
 276                                   const int dif, const int sdif)
 277 {
 278         INET_ADDR_COOKIE(acookie, saddr, daddr);
 279         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 280         struct sock *sk;
 281         const struct hlist_nulls_node *node;
 282         /* Optimize here for direct hit, only listening connections can
 283          * have wildcards anyways.
 284          */
 285         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 286         unsigned int slot = hash & hashinfo->ehash_mask;
 287         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 288
 289 begin:
 290         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 291                 if (sk->sk_hash != hash)
 292                         continue;
 293                 if (likely(INET_MATCH(sk, net, acookie,
 294                                       saddr, daddr, ports, dif, sdif))) {
 295                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 296                                 goto out;
 297                         if (unlikely(!INET_MATCH(sk, net, acookie,
 298                                                  saddr, daddr, ports,
 299                                                  dif, sdif))) {
 300                                 sock_gen_put(sk);
 301                                 goto begin;
 302                         }
 303                         goto found;
 304                 }
 305         }
 306         /*
 307          * if the nulls value we got at the end of this lookup is
 308          * not the expected one, we must restart lookup.
 309          * We probably met an item that was moved to another chain.
 310          */
 311         if (get_nulls_value(node) != slot)
 312                 goto begin;
 313 out:
 314         sk = NULL;
 315 found:
 316         return sk;
 317 }
 318 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 319
 320 /* called with local bh disabled */
 321 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 322                                     struct sock *sk, __u16 lport,
 323                                     struct inet_timewait_sock **twp)
 324 {
 325         struct inet_hashinfo *hinfo = death_row->hashinfo;
 326         struct inet_sock *inet = inet_sk(sk);
 327         __be32 daddr = inet->inet_rcv_saddr;
 328         __be32 saddr = inet->inet_daddr;
 329         int dif = sk->sk_bound_dev_if;
 330         struct net *net = sock_net(sk);
 331         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 332         INET_ADDR_COOKIE(acookie, saddr, daddr);
 333         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 334         unsigned int hash = inet_ehashfn(net, daddr, lport,
 335                                          saddr, inet->inet_dport);
 336         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 337         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 338         struct sock *sk2;
 339         const struct hlist_nulls_node *node;
 340         struct inet_timewait_sock *tw = NULL;
 341
 342         spin_lock(lock);
 343
 344         sk_nulls_for_each(sk2, node, &head->chain) {
 345                 if (sk2->sk_hash != hash)
 346                         continue;
 347
 348                 if (likely(INET_MATCH(sk2, net, acookie,
 349                                          saddr, daddr, ports, dif, sdif))) {
 350                         if (sk2->sk_state == TCP_TIME_WAIT) {
 351                                 tw = inet_twsk(sk2);
 352                                 if (twsk_unique(sk, sk2, twp))
 353                                         break;
 354                         }
 355                         goto not_unique;
 356                 }
 357         }
 358
 359         /* Must record num and sport now. Otherwise we will see
 360          * in hash table socket with a funny identity.
 361          */
 362         inet->inet_num = lport;
 363         inet->inet_sport = htons(lport);
 364         sk->sk_hash = hash;
 365         WARN_ON(!sk_unhashed(sk));
 366         __sk_nulls_add_node_rcu(sk, &head->chain);
 367         if (tw) {
 368                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 369                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 370         }
 371         spin_unlock(lock);
 372         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 373
 374         if (twp) {
 375                 *twp = tw;
 376         } else if (tw) {
 377                 /* Silly. Should hash-dance instead... */
 378                 inet_twsk_deschedule_put(tw);
 379         }
 380         return 0;
 381
 382 not_unique:
 383         spin_unlock(lock);
 384         return -EADDRNOTAVAIL;
 385 }
 386
 387 static u32 inet_sk_port_offset(const struct sock *sk)
 388 {
 389         const struct inet_sock *inet = inet_sk(sk);
 390
 391         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 392                                           inet->inet_daddr,
 393                                           inet->inet_dport);
 394 }
 395
 396 /* insert a socket into ehash, and eventually remove another one
 397  * (The another one can be a SYN_RECV or TIMEWAIT
 398  */
 399 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
 400 {
 401         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 402         struct hlist_nulls_head *list;
 403         struct inet_ehash_bucket *head;
 404         spinlock_t *lock;
 405         bool ret = true;
 406
 407         WARN_ON_ONCE(!sk_unhashed(sk));
 408
 409         sk->sk_hash = sk_ehashfn(sk);
 410         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 411         list = &head->chain;
 412         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 413
 414         spin_lock(lock);
 415         if (osk) {
 416                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 417                 ret = sk_nulls_del_node_init_rcu(osk);
 418         }
 419         if (ret)
 420                 __sk_nulls_add_node_rcu(sk, list);
 421         spin_unlock(lock);
 422         return ret;
 423 }
 424
 425 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
 426 {
 427         bool ok = inet_ehash_insert(sk, osk);
 428
 429         if (ok) {
 430                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 431         } else {
 432                 percpu_counter_inc(sk->sk_prot->orphan_count);
 433                 sk->sk_state = TCP_CLOSE;
 434                 sock_set_flag(sk, SOCK_DEAD);
 435                 inet_csk_destroy_sock(sk);
 436         }
 437         return ok;
 438 }
 439 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 440
 441 static int inet_reuseport_add_sock(struct sock *sk,
 442                                    struct inet_listen_hashbucket *ilb)
 443 {
 444         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 445         struct sock *sk2;
 446         kuid_t uid = sock_i_uid(sk);
 447
 448         sk_for_each_rcu(sk2, &ilb->head) {
 449                 if (sk2 != sk &&
 450                     sk2->sk_family == sk->sk_family &&
 451                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 452                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 453                     inet_csk(sk2)->icsk_bind_hash == tb &&
 454                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 455                     inet_rcv_saddr_equal(sk, sk2, false))
 456                         return reuseport_add_sock(sk, sk2);
 457         }
 458
 459         /* Initial allocation may have already happened via setsockopt */
 460         if (!rcu_access_pointer(sk->sk_reuseport_cb))
 461                 return reuseport_alloc(sk);
 462         return 0;
 463 }
 464
 465 int __inet_hash(struct sock *sk, struct sock *osk)
 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);
 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);
 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 = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
 689                 if (!hashinfo->ehash_locks)
 690                         return -ENOMEM;
 691
 692                 for (i = 0; i < nblocks; i++)
 693                         spin_lock_init(&hashinfo->ehash_locks[i]);
 694         }
 695         hashinfo->ehash_locks_mask = nblocks - 1;
 696         return 0;
 697 }
 698 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);