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/inet_connection_sock.h>
  24 #include <net/inet_hashtables.h>
  25 #include <net/secure_seq.h>
  26 #include <net/ip.h>
  27
  28 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
  29                         const __u16 lport, const __be32 faddr,
  30                         const __be16 fport)
  31 {
  32         static u32 inet_ehash_secret __read_mostly;
  33
  34         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  35
  36         return __inet_ehashfn(laddr, lport, faddr, fport,
  37                               inet_ehash_secret + net_hash_mix(net));
  38 }
  39
  40 /* This function handles inet_sock, but also timewait and request sockets
  41  * for IPv4/IPv6.
  42  */
  43 u32 sk_ehashfn(const struct sock *sk)
  44 {
  45 #if IS_ENABLED(CONFIG_IPV6)
  46         if (sk->sk_family == AF_INET6 &&
  47             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  48                 return inet6_ehashfn(sock_net(sk),
  49                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  50                                      &sk->sk_v6_daddr, sk->sk_dport);
  51 #endif
  52         return inet_ehashfn(sock_net(sk),
  53                             sk->sk_rcv_saddr, sk->sk_num,
  54                             sk->sk_daddr, sk->sk_dport);
  55 }
  56
  57 /*
  58  * Allocate and initialize a new local port bind bucket.
  59  * The bindhash mutex for snum's hash chain must be held here.
  60  */
  61 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  62                                                  struct net *net,
  63                                                  struct inet_bind_hashbucket *head,
  64                                                  const unsigned short snum)
  65 {
  66         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  67
  68         if (tb) {
  69                 write_pnet(&tb->ib_net, net);
  70                 tb->port      = snum;
  71                 tb->fastreuse = 0;
  72                 tb->fastreuseport = 0;
  73                 tb->num_owners = 0;
  74                 INIT_HLIST_HEAD(&tb->owners);
  75                 hlist_add_head(&tb->node, &head->chain);
  76         }
  77         return tb;
  78 }
  79
  80 /*
  81  * Caller must hold hashbucket lock for this tb with local BH disabled
  82  */
  83 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  84 {
  85         if (hlist_empty(&tb->owners)) {
  86                 __hlist_del(&tb->node);
  87                 kmem_cache_free(cachep, tb);
  88         }
  89 }
  90
  91 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  92                     const unsigned short snum)
  93 {
  94         inet_sk(sk)->inet_num = snum;
  95         sk_add_bind_node(sk, &tb->owners);
  96         tb->num_owners++;
  97         inet_csk(sk)->icsk_bind_hash = tb;
  98 }
  99
 100 /*
 101  * Get rid of any references to a local port held by the given sock.
 102  */
 103 static void __inet_put_port(struct sock *sk)
 104 {
 105         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 106         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 107                         hashinfo->bhash_size);
 108         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 109         struct inet_bind_bucket *tb;
 110
 111         spin_lock(&head->lock);
 112         tb = inet_csk(sk)->icsk_bind_hash;
 113         __sk_del_bind_node(sk);
 114         tb->num_owners--;
 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                 inet_csk_update_fastreuse(tb, child);
 164         }
 165         inet_bind_hash(child, tb, port);
 166         spin_unlock(&head->lock);
 167
 168         return 0;
 169 }
 170 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 171
 172 static inline int compute_score(struct sock *sk, struct net *net,
 173                                 const unsigned short hnum, const __be32 daddr,
 174                                 const int dif)
 175 {
 176         int score = -1;
 177         struct inet_sock *inet = inet_sk(sk);
 178
 179         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 180                         !ipv6_only_sock(sk)) {
 181                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 182                 score = sk->sk_family == PF_INET ? 2 : 1;
 183                 if (rcv_saddr) {
 184                         if (rcv_saddr != daddr)
 185                                 return -1;
 186                         score += 4;
 187                 }
 188                 if (sk->sk_bound_dev_if) {
 189                         if (sk->sk_bound_dev_if != dif)
 190                                 return -1;
 191                         score += 4;
 192                 }
 193                 if (sk->sk_incoming_cpu == raw_smp_processor_id())
 194                         score++;
 195         }
 196         return score;
 197 }
 198
 199 /*
 200  * Don't inline this cruft. Here are some nice properties to exploit here. The
 201  * BSD API does not allow a listening sock to specify the remote port nor the
 202  * remote address for the connection. So always assume those are both
 203  * wildcarded during the search since they can never be otherwise.
 204  */
 205
 206
 207 struct sock *__inet_lookup_listener(struct net *net,
 208                                     struct inet_hashinfo *hashinfo,
 209                                     const __be32 saddr, __be16 sport,
 210                                     const __be32 daddr, const unsigned short hnum,
 211                                     const int dif)
 212 {
 213         struct sock *sk, *result;
 214         struct hlist_nulls_node *node;
 215         unsigned int hash = inet_lhashfn(net, hnum);
 216         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 217         int score, hiscore, matches = 0, reuseport = 0;
 218         u32 phash = 0;
 219
 220         rcu_read_lock();
 221 begin:
 222         result = NULL;
 223         hiscore = 0;
 224         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 225                 score = compute_score(sk, net, hnum, daddr, dif);
 226                 if (score > hiscore) {
 227                         result = sk;
 228                         hiscore = score;
 229                         reuseport = sk->sk_reuseport;
 230                         if (reuseport) {
 231                                 phash = inet_ehashfn(net, daddr, hnum,
 232                                                      saddr, sport);
 233                                 matches = 1;
 234                         }
 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         /*
 243          * if the nulls value we got at the end of this lookup is
 244          * not the expected one, we must restart lookup.
 245          * We probably met an item that was moved to another chain.
 246          */
 247         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 248                 goto begin;
 249         if (result) {
 250                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 251                         result = NULL;
 252                 else if (unlikely(compute_score(result, net, hnum, daddr,
 253                                   dif) < hiscore)) {
 254                         sock_put(result);
 255                         goto begin;
 256                 }
 257         }
 258         rcu_read_unlock();
 259         return result;
 260 }
 261 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 262
 263 /* All sockets share common refcount, but have different destructors */
 264 void sock_gen_put(struct sock *sk)
 265 {
 266         if (!atomic_dec_and_test(&sk->sk_refcnt))
 267                 return;
 268
 269         if (sk->sk_state == TCP_TIME_WAIT)
 270                 inet_twsk_free(inet_twsk(sk));
 271         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 272                 reqsk_free(inet_reqsk(sk));
 273         else
 274                 sk_free(sk);
 275 }
 276 EXPORT_SYMBOL_GPL(sock_gen_put);
 277
 278 void sock_edemux(struct sk_buff *skb)
 279 {
 280         sock_gen_put(skb->sk);
 281 }
 282 EXPORT_SYMBOL(sock_edemux);
 283
 284 struct sock *__inet_lookup_established(struct net *net,
 285                                   struct inet_hashinfo *hashinfo,
 286                                   const __be32 saddr, const __be16 sport,
 287                                   const __be32 daddr, const u16 hnum,
 288                                   const int dif)
 289 {
 290         INET_ADDR_COOKIE(acookie, saddr, daddr);
 291         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 292         struct sock *sk;
 293         const struct hlist_nulls_node *node;
 294         /* Optimize here for direct hit, only listening connections can
 295          * have wildcards anyways.
 296          */
 297         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 298         unsigned int slot = hash & hashinfo->ehash_mask;
 299         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 300
 301         rcu_read_lock();
 302 begin:
 303         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 304                 if (sk->sk_hash != hash)
 305                         continue;
 306                 if (likely(INET_MATCH(sk, net, acookie,
 307                                       saddr, daddr, ports, dif))) {
 308                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 309                                 goto out;
 310                         if (unlikely(!INET_MATCH(sk, net, acookie,
 311                                                  saddr, daddr, ports, dif))) {
 312                                 sock_gen_put(sk);
 313                                 goto begin;
 314                         }
 315                         goto found;
 316                 }
 317         }
 318         /*
 319          * if the nulls value we got at the end of this lookup is
 320          * not the expected one, we must restart lookup.
 321          * We probably met an item that was moved to another chain.
 322          */
 323         if (get_nulls_value(node) != slot)
 324                 goto begin;
 325 out:
 326         sk = NULL;
 327 found:
 328         rcu_read_unlock();
 329         return sk;
 330 }
 331 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 332
 333 /* called with local bh disabled */
 334 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 335                                     struct sock *sk, __u16 lport,
 336                                     struct inet_timewait_sock **twp)
 337 {
 338         struct inet_hashinfo *hinfo = death_row->hashinfo;
 339         struct inet_sock *inet = inet_sk(sk);
 340         __be32 daddr = inet->inet_rcv_saddr;
 341         __be32 saddr = inet->inet_daddr;
 342         int dif = sk->sk_bound_dev_if;
 343         INET_ADDR_COOKIE(acookie, saddr, daddr);
 344         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 345         struct net *net = sock_net(sk);
 346         unsigned int hash = inet_ehashfn(net, daddr, lport,
 347                                          saddr, inet->inet_dport);
 348         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 349         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 350         struct sock *sk2;
 351         const struct hlist_nulls_node *node;
 352         struct inet_timewait_sock *tw = NULL;
 353
 354         spin_lock(lock);
 355
 356         sk_nulls_for_each(sk2, node, &head->chain) {
 357                 if (sk2->sk_hash != hash)
 358                         continue;
 359
 360                 if (likely(INET_MATCH(sk2, net, acookie,
 361                                          saddr, daddr, ports, dif))) {
 362                         if (sk2->sk_state == TCP_TIME_WAIT) {
 363                                 tw = inet_twsk(sk2);
 364                                 if (twsk_unique(sk, sk2, twp))
 365                                         break;
 366                         }
 367                         goto not_unique;
 368                 }
 369         }
 370
 371         /* Must record num and sport now. Otherwise we will see
 372          * in hash table socket with a funny identity.
 373          */
 374         inet->inet_num = lport;
 375         inet->inet_sport = htons(lport);
 376         sk->sk_hash = hash;
 377         WARN_ON(!sk_unhashed(sk));
 378         __sk_nulls_add_node_rcu(sk, &head->chain);
 379         if (tw) {
 380                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 381                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 382         }
 383         spin_unlock(lock);
 384         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 385
 386         if (twp) {
 387                 *twp = tw;
 388         } else if (tw) {
 389                 /* Silly. Should hash-dance instead... */
 390                 inet_twsk_deschedule_put(tw);
 391         }
 392         return 0;
 393
 394 not_unique:
 395         spin_unlock(lock);
 396         return -EADDRNOTAVAIL;
 397 }
 398
 399 static u32 inet_sk_port_offset(const struct sock *sk)
 400 {
 401         const struct inet_sock *inet = inet_sk(sk);
 402
 403         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 404                                           inet->inet_daddr,
 405                                           inet->inet_dport);
 406 }
 407
 408 /* insert a socket into ehash, and eventually remove another one
 409  * (The another one can be a SYN_RECV or TIMEWAIT
 410  */
 411 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
 412 {
 413         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 414         struct hlist_nulls_head *list;
 415         struct inet_ehash_bucket *head;
 416         spinlock_t *lock;
 417         bool ret = true;
 418
 419         WARN_ON_ONCE(!sk_unhashed(sk));
 420
 421         sk->sk_hash = sk_ehashfn(sk);
 422         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 423         list = &head->chain;
 424         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 425
 426         spin_lock(lock);
 427         if (osk) {
 428                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 429                 ret = sk_nulls_del_node_init_rcu(osk);
 430         }
 431         if (ret)
 432                 __sk_nulls_add_node_rcu(sk, list);
 433         spin_unlock(lock);
 434         return ret;
 435 }
 436
 437 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
 438 {
 439         bool ok = inet_ehash_insert(sk, osk);
 440
 441         if (ok) {
 442                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 443         } else {
 444                 percpu_counter_inc(sk->sk_prot->orphan_count);
 445                 sk->sk_state = TCP_CLOSE;
 446                 sock_set_flag(sk, SOCK_DEAD);
 447                 inet_csk_destroy_sock(sk);
 448         }
 449         return ok;
 450 }
 451 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 452
 453 void __inet_hash(struct sock *sk, struct sock *osk)
 454 {
 455         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 456         struct inet_listen_hashbucket *ilb;
 457
 458         if (sk->sk_state != TCP_LISTEN) {
 459                 inet_ehash_nolisten(sk, osk);
 460                 return;
 461         }
 462         WARN_ON(!sk_unhashed(sk));
 463         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 464
 465         spin_lock(&ilb->lock);
 466         __sk_nulls_add_node_rcu(sk, &ilb->head);
 467         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 468         spin_unlock(&ilb->lock);
 469 }
 470 EXPORT_SYMBOL(__inet_hash);
 471
 472 void inet_hash(struct sock *sk)
 473 {
 474         if (sk->sk_state != TCP_CLOSE) {
 475                 local_bh_disable();
 476                 __inet_hash(sk, NULL);
 477                 local_bh_enable();
 478         }
 479 }
 480 EXPORT_SYMBOL_GPL(inet_hash);
 481
 482 void inet_unhash(struct sock *sk)
 483 {
 484         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 485         spinlock_t *lock;
 486         int done;
 487
 488         if (sk_unhashed(sk))
 489                 return;
 490
 491         if (sk->sk_state == TCP_LISTEN)
 492                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 493         else
 494                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 495
 496         spin_lock_bh(lock);
 497         done = __sk_nulls_del_node_init_rcu(sk);
 498         if (done)
 499                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 500         spin_unlock_bh(lock);
 501 }
 502 EXPORT_SYMBOL_GPL(inet_unhash);
 503
 504 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 505                 struct sock *sk, u32 port_offset,
 506                 int (*check_established)(struct inet_timewait_death_row *,
 507                         struct sock *, __u16, struct inet_timewait_sock **))
 508 {
 509         struct inet_hashinfo *hinfo = death_row->hashinfo;
 510         const unsigned short snum = inet_sk(sk)->inet_num;
 511         struct inet_bind_hashbucket *head;
 512         struct inet_bind_bucket *tb;
 513         int ret;
 514         struct net *net = sock_net(sk);
 515
 516         if (!snum) {
 517                 int i, remaining, low, high, port;
 518                 static u32 hint;
 519                 u32 offset = hint + port_offset;
 520                 struct inet_timewait_sock *tw = NULL;
 521
 522                 inet_get_local_port_range(net, &low, &high);
 523                 remaining = (high - low) + 1;
 524
 525                 /* By starting with offset being an even number,
 526                  * we tend to leave about 50% of ports for other uses,
 527                  * like bind(0).
 528                  */
 529                 offset &= ~1;
 530
 531                 local_bh_disable();
 532                 for (i = 0; i < remaining; i++) {
 533                         port = low + (i + offset) % remaining;
 534                         if (inet_is_local_reserved_port(net, port))
 535                                 continue;
 536                         head = &hinfo->bhash[inet_bhashfn(net, port,
 537                                         hinfo->bhash_size)];
 538                         spin_lock(&head->lock);
 539
 540                         /* Does not bother with rcv_saddr checks,
 541                          * because the established check is already
 542                          * unique enough.
 543                          */
 544                         inet_bind_bucket_for_each(tb, &head->chain) {
 545                                 if (net_eq(ib_net(tb), net) &&
 546                                     tb->port == port) {
 547                                         if (tb->fastreuse >= 0 ||
 548                                             tb->fastreuseport >= 0)
 549                                                 goto next_port;
 550                                         WARN_ON(hlist_empty(&tb->owners));
 551                                         if (!check_established(death_row, sk,
 552                                                                 port, &tw))
 553                                                 goto ok;
 554                                         goto next_port;
 555                                 }
 556                         }
 557
 558                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 559                                         net, head, port);
 560                         if (!tb) {
 561                                 spin_unlock(&head->lock);
 562                                 break;
 563                         }
 564                         tb->fastreuse = -1;
 565                         tb->fastreuseport = -1;
 566                         goto ok;
 567
 568                 next_port:
 569                         spin_unlock(&head->lock);
 570                 }
 571                 local_bh_enable();
 572
 573                 return -EADDRNOTAVAIL;
 574
 575 ok:
 576                 hint += (i + 2) & ~1;
 577
 578                 /* Head lock still held and bh's disabled */
 579                 inet_bind_hash(sk, tb, port);
 580                 if (sk_unhashed(sk)) {
 581                         inet_sk(sk)->inet_sport = htons(port);
 582                         inet_ehash_nolisten(sk, (struct sock *)tw);
 583                 }
 584                 if (tw)
 585                         inet_twsk_bind_unhash(tw, hinfo);
 586                 spin_unlock(&head->lock);
 587
 588                 if (tw)
 589                         inet_twsk_deschedule_put(tw);
 590
 591                 ret = 0;
 592                 goto out;
 593         }
 594
 595         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 596         tb  = inet_csk(sk)->icsk_bind_hash;
 597         spin_lock_bh(&head->lock);
 598         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 599                 inet_ehash_nolisten(sk, NULL);
 600                 spin_unlock_bh(&head->lock);
 601                 return 0;
 602         } else {
 603                 spin_unlock(&head->lock);
 604                 /* No definite answer... Walk to established hash table */
 605                 ret = check_established(death_row, sk, snum, NULL);
 606 out:
 607                 local_bh_enable();
 608                 return ret;
 609         }
 610 }
 611
 612 /*
 613  * Bind a port for a connect operation and hash it.
 614  */
 615 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 616                       struct sock *sk)
 617 {
 618         u32 port_offset = 0;
 619
 620         if (!inet_sk(sk)->inet_num)
 621                 port_offset = inet_sk_port_offset(sk);
 622         return __inet_hash_connect(death_row, sk, port_offset,
 623                                    __inet_check_established);
 624 }
 625 EXPORT_SYMBOL_GPL(inet_hash_connect);
 626
 627 void inet_hashinfo_init(struct inet_hashinfo *h)
 628 {
 629         int i;
 630
 631         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 632                 spin_lock_init(&h->listening_hash[i].lock);
 633                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 634                                       i + LISTENING_NULLS_BASE);
 635                 }
 636 }
 637 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 638
 639 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 640 {
 641         unsigned int locksz = sizeof(spinlock_t);
 642         unsigned int i, nblocks = 1;
 643
 644         if (locksz != 0) {
 645                 /* allocate 2 cache lines or at least one spinlock per cpu */
 646                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 647                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 648
 649                 /* no more locks than number of hash buckets */
 650                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 651
 652                 hashinfo->ehash_locks = kmalloc_array(nblocks, locksz,
 653                                                       GFP_KERNEL | __GFP_NOWARN);
 654                 if (!hashinfo->ehash_locks)
 655                         hashinfo->ehash_locks = vmalloc(nblocks * locksz);
 656
 657                 if (!hashinfo->ehash_locks)
 658                         return -ENOMEM;
 659
 660                 for (i = 0; i < nblocks; i++)
 661                         spin_lock_init(&hashinfo->ehash_locks[i]);
 662         }
 663         hashinfo->ehash_locks_mask = nblocks - 1;
 664         return 0;
 665 }
 666 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);