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 (tb->port != port) {
 141                 /* NOTE: using tproxy and redirecting skbs to a proxy
 142                  * on a different listener port breaks the assumption
 143                  * that the listener socket's icsk_bind_hash is the same
 144                  * as that of the child socket. We have to look up or
 145                  * create a new bind bucket for the child here. */
 146                 inet_bind_bucket_for_each(tb, &head->chain) {
 147                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 148                             tb->port == port)
 149                                 break;
 150                 }
 151                 if (!tb) {
 152                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 153                                                      sock_net(sk), head, port);
 154                         if (!tb) {
 155                                 spin_unlock(&head->lock);
 156                                 return -ENOMEM;
 157                         }
 158                 }
 159         }
 160         inet_bind_hash(child, tb, port);
 161         spin_unlock(&head->lock);
 162
 163         return 0;
 164 }
 165 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 166
 167 static inline int compute_score(struct sock *sk, struct net *net,
 168                                 const unsigned short hnum, const __be32 daddr,
 169                                 const int dif)
 170 {
 171         int score = -1;
 172         struct inet_sock *inet = inet_sk(sk);
 173
 174         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
 175                         !ipv6_only_sock(sk)) {
 176                 __be32 rcv_saddr = inet->inet_rcv_saddr;
 177                 score = sk->sk_family == PF_INET ? 2 : 1;
 178                 if (rcv_saddr) {
 179                         if (rcv_saddr != daddr)
 180                                 return -1;
 181                         score += 4;
 182                 }
 183                 if (sk->sk_bound_dev_if) {
 184                         if (sk->sk_bound_dev_if != dif)
 185                                 return -1;
 186                         score += 4;
 187                 }
 188                 if (sk->sk_incoming_cpu == raw_smp_processor_id())
 189                         score++;
 190         }
 191         return score;
 192 }
 193
 194 /*
 195  * Don't inline this cruft. Here are some nice properties to exploit here. The
 196  * BSD API does not allow a listening sock to specify the remote port nor the
 197  * remote address for the connection. So always assume those are both
 198  * wildcarded during the search since they can never be otherwise.
 199  */
 200
 201
 202 struct sock *__inet_lookup_listener(struct net *net,
 203                                     struct inet_hashinfo *hashinfo,
 204                                     const __be32 saddr, __be16 sport,
 205                                     const __be32 daddr, const unsigned short hnum,
 206                                     const int dif)
 207 {
 208         struct sock *sk, *result;
 209         struct hlist_nulls_node *node;
 210         unsigned int hash = inet_lhashfn(net, hnum);
 211         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
 212         int score, hiscore, matches = 0, reuseport = 0;
 213         u32 phash = 0;
 214
 215         rcu_read_lock();
 216 begin:
 217         result = NULL;
 218         hiscore = 0;
 219         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
 220                 score = compute_score(sk, net, hnum, daddr, dif);
 221                 if (score > hiscore) {
 222                         result = sk;
 223                         hiscore = score;
 224                         reuseport = sk->sk_reuseport;
 225                         if (reuseport) {
 226                                 phash = inet_ehashfn(net, daddr, hnum,
 227                                                      saddr, sport);
 228                                 matches = 1;
 229                         }
 230                 } else if (score == hiscore && reuseport) {
 231                         matches++;
 232                         if (reciprocal_scale(phash, matches) == 0)
 233                                 result = sk;
 234                         phash = next_pseudo_random32(phash);
 235                 }
 236         }
 237         /*
 238          * if the nulls value we got at the end of this lookup is
 239          * not the expected one, we must restart lookup.
 240          * We probably met an item that was moved to another chain.
 241          */
 242         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
 243                 goto begin;
 244         if (result) {
 245                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
 246                         result = NULL;
 247                 else if (unlikely(compute_score(result, net, hnum, daddr,
 248                                   dif) < hiscore)) {
 249                         sock_put(result);
 250                         goto begin;
 251                 }
 252         }
 253         rcu_read_unlock();
 254         return result;
 255 }
 256 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 257
 258 /* All sockets share common refcount, but have different destructors */
 259 void sock_gen_put(struct sock *sk)
 260 {
 261         if (!atomic_dec_and_test(&sk->sk_refcnt))
 262                 return;
 263
 264         if (sk->sk_state == TCP_TIME_WAIT)
 265                 inet_twsk_free(inet_twsk(sk));
 266         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 267                 reqsk_free(inet_reqsk(sk));
 268         else
 269                 sk_free(sk);
 270 }
 271 EXPORT_SYMBOL_GPL(sock_gen_put);
 272
 273 void sock_edemux(struct sk_buff *skb)
 274 {
 275         sock_gen_put(skb->sk);
 276 }
 277 EXPORT_SYMBOL(sock_edemux);
 278
 279 struct sock *__inet_lookup_established(struct net *net,
 280                                   struct inet_hashinfo *hashinfo,
 281                                   const __be32 saddr, const __be16 sport,
 282                                   const __be32 daddr, const u16 hnum,
 283                                   const int dif)
 284 {
 285         INET_ADDR_COOKIE(acookie, saddr, daddr);
 286         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 287         struct sock *sk;
 288         const struct hlist_nulls_node *node;
 289         /* Optimize here for direct hit, only listening connections can
 290          * have wildcards anyways.
 291          */
 292         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 293         unsigned int slot = hash & hashinfo->ehash_mask;
 294         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 295
 296         rcu_read_lock();
 297 begin:
 298         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 299                 if (sk->sk_hash != hash)
 300                         continue;
 301                 if (likely(INET_MATCH(sk, net, acookie,
 302                                       saddr, daddr, ports, dif))) {
 303                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
 304                                 goto out;
 305                         if (unlikely(!INET_MATCH(sk, net, acookie,
 306                                                  saddr, daddr, ports, dif))) {
 307                                 sock_gen_put(sk);
 308                                 goto begin;
 309                         }
 310                         goto found;
 311                 }
 312         }
 313         /*
 314          * if the nulls value we got at the end of this lookup is
 315          * not the expected one, we must restart lookup.
 316          * We probably met an item that was moved to another chain.
 317          */
 318         if (get_nulls_value(node) != slot)
 319                 goto begin;
 320 out:
 321         sk = NULL;
 322 found:
 323         rcu_read_unlock();
 324         return sk;
 325 }
 326 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 327
 328 /* called with local bh disabled */
 329 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 330                                     struct sock *sk, __u16 lport,
 331                                     struct inet_timewait_sock **twp)
 332 {
 333         struct inet_hashinfo *hinfo = death_row->hashinfo;
 334         struct inet_sock *inet = inet_sk(sk);
 335         __be32 daddr = inet->inet_rcv_saddr;
 336         __be32 saddr = inet->inet_daddr;
 337         int dif = sk->sk_bound_dev_if;
 338         INET_ADDR_COOKIE(acookie, saddr, daddr);
 339         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 340         struct net *net = sock_net(sk);
 341         unsigned int hash = inet_ehashfn(net, daddr, lport,
 342                                          saddr, inet->inet_dport);
 343         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 344         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 345         struct sock *sk2;
 346         const struct hlist_nulls_node *node;
 347         struct inet_timewait_sock *tw = NULL;
 348
 349         spin_lock(lock);
 350
 351         sk_nulls_for_each(sk2, node, &head->chain) {
 352                 if (sk2->sk_hash != hash)
 353                         continue;
 354
 355                 if (likely(INET_MATCH(sk2, net, acookie,
 356                                          saddr, daddr, ports, dif))) {
 357                         if (sk2->sk_state == TCP_TIME_WAIT) {
 358                                 tw = inet_twsk(sk2);
 359                                 if (twsk_unique(sk, sk2, twp))
 360                                         break;
 361                         }
 362                         goto not_unique;
 363                 }
 364         }
 365
 366         /* Must record num and sport now. Otherwise we will see
 367          * in hash table socket with a funny identity.
 368          */
 369         inet->inet_num = lport;
 370         inet->inet_sport = htons(lport);
 371         sk->sk_hash = hash;
 372         WARN_ON(!sk_unhashed(sk));
 373         __sk_nulls_add_node_rcu(sk, &head->chain);
 374         if (tw) {
 375                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 376                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
 377         }
 378         spin_unlock(lock);
 379         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 380
 381         if (twp) {
 382                 *twp = tw;
 383         } else if (tw) {
 384                 /* Silly. Should hash-dance instead... */
 385                 inet_twsk_deschedule_put(tw);
 386         }
 387         return 0;
 388
 389 not_unique:
 390         spin_unlock(lock);
 391         return -EADDRNOTAVAIL;
 392 }
 393
 394 static u32 inet_sk_port_offset(const struct sock *sk)
 395 {
 396         const struct inet_sock *inet = inet_sk(sk);
 397
 398         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 399                                           inet->inet_daddr,
 400                                           inet->inet_dport);
 401 }
 402
 403 /* insert a socket into ehash, and eventually remove another one
 404  * (The another one can be a SYN_RECV or TIMEWAIT
 405  */
 406 int inet_ehash_insert(struct sock *sk, struct sock *osk)
 407 {
 408         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 409         struct hlist_nulls_head *list;
 410         struct inet_ehash_bucket *head;
 411         spinlock_t *lock;
 412         int ret = 0;
 413
 414         WARN_ON_ONCE(!sk_unhashed(sk));
 415
 416         sk->sk_hash = sk_ehashfn(sk);
 417         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 418         list = &head->chain;
 419         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 420
 421         spin_lock(lock);
 422         __sk_nulls_add_node_rcu(sk, list);
 423         if (osk) {
 424                 WARN_ON(sk->sk_hash != osk->sk_hash);
 425                 sk_nulls_del_node_init_rcu(osk);
 426         }
 427         spin_unlock(lock);
 428         return ret;
 429 }
 430
 431 void __inet_hash_nolisten(struct sock *sk, struct sock *osk)
 432 {
 433         inet_ehash_insert(sk, osk);
 434         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 435 }
 436 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 437
 438 void __inet_hash(struct sock *sk, struct sock *osk)
 439 {
 440         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 441         struct inet_listen_hashbucket *ilb;
 442
 443         if (sk->sk_state != TCP_LISTEN)
 444                 return __inet_hash_nolisten(sk, osk);
 445
 446         WARN_ON(!sk_unhashed(sk));
 447         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 448
 449         spin_lock(&ilb->lock);
 450         __sk_nulls_add_node_rcu(sk, &ilb->head);
 451         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 452         spin_unlock(&ilb->lock);
 453 }
 454 EXPORT_SYMBOL(__inet_hash);
 455
 456 void inet_hash(struct sock *sk)
 457 {
 458         if (sk->sk_state != TCP_CLOSE) {
 459                 local_bh_disable();
 460                 __inet_hash(sk, NULL);
 461                 local_bh_enable();
 462         }
 463 }
 464 EXPORT_SYMBOL_GPL(inet_hash);
 465
 466 void inet_unhash(struct sock *sk)
 467 {
 468         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 469         spinlock_t *lock;
 470         int done;
 471
 472         if (sk_unhashed(sk))
 473                 return;
 474
 475         if (sk->sk_state == TCP_LISTEN)
 476                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
 477         else
 478                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 479
 480         spin_lock_bh(lock);
 481         done = __sk_nulls_del_node_init_rcu(sk);
 482         if (done)
 483                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 484         spin_unlock_bh(lock);
 485 }
 486 EXPORT_SYMBOL_GPL(inet_unhash);
 487
 488 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 489                 struct sock *sk, u32 port_offset,
 490                 int (*check_established)(struct inet_timewait_death_row *,
 491                         struct sock *, __u16, struct inet_timewait_sock **))
 492 {
 493         struct inet_hashinfo *hinfo = death_row->hashinfo;
 494         const unsigned short snum = inet_sk(sk)->inet_num;
 495         struct inet_bind_hashbucket *head;
 496         struct inet_bind_bucket *tb;
 497         int ret;
 498         struct net *net = sock_net(sk);
 499
 500         if (!snum) {
 501                 int i, remaining, low, high, port;
 502                 static u32 hint;
 503                 u32 offset = hint + port_offset;
 504                 struct inet_timewait_sock *tw = NULL;
 505
 506                 inet_get_local_port_range(net, &low, &high);
 507                 remaining = (high - low) + 1;
 508
 509                 /* By starting with offset being an even number,
 510                  * we tend to leave about 50% of ports for other uses,
 511                  * like bind(0).
 512                  */
 513                 offset &= ~1;
 514
 515                 local_bh_disable();
 516                 for (i = 0; i < remaining; i++) {
 517                         port = low + (i + offset) % remaining;
 518                         if (inet_is_local_reserved_port(net, port))
 519                                 continue;
 520                         head = &hinfo->bhash[inet_bhashfn(net, port,
 521                                         hinfo->bhash_size)];
 522                         spin_lock(&head->lock);
 523
 524                         /* Does not bother with rcv_saddr checks,
 525                          * because the established check is already
 526                          * unique enough.
 527                          */
 528                         inet_bind_bucket_for_each(tb, &head->chain) {
 529                                 if (net_eq(ib_net(tb), net) &&
 530                                     tb->port == port) {
 531                                         if (tb->fastreuse >= 0 ||
 532                                             tb->fastreuseport >= 0)
 533                                                 goto next_port;
 534                                         WARN_ON(hlist_empty(&tb->owners));
 535                                         if (!check_established(death_row, sk,
 536                                                                 port, &tw))
 537                                                 goto ok;
 538                                         goto next_port;
 539                                 }
 540                         }
 541
 542                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 543                                         net, head, port);
 544                         if (!tb) {
 545                                 spin_unlock(&head->lock);
 546                                 break;
 547                         }
 548                         tb->fastreuse = -1;
 549                         tb->fastreuseport = -1;
 550                         goto ok;
 551
 552                 next_port:
 553                         spin_unlock(&head->lock);
 554                 }
 555                 local_bh_enable();
 556
 557                 return -EADDRNOTAVAIL;
 558
 559 ok:
 560                 hint += (i + 2) & ~1;
 561
 562                 /* Head lock still held and bh's disabled */
 563                 inet_bind_hash(sk, tb, port);
 564                 if (sk_unhashed(sk)) {
 565                         inet_sk(sk)->inet_sport = htons(port);
 566                         __inet_hash_nolisten(sk, (struct sock *)tw);
 567                 }
 568                 if (tw)
 569                         inet_twsk_bind_unhash(tw, hinfo);
 570                 spin_unlock(&head->lock);
 571
 572                 if (tw)
 573                         inet_twsk_deschedule_put(tw);
 574
 575                 ret = 0;
 576                 goto out;
 577         }
 578
 579         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
 580         tb  = inet_csk(sk)->icsk_bind_hash;
 581         spin_lock_bh(&head->lock);
 582         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 583                 __inet_hash_nolisten(sk, NULL);
 584                 spin_unlock_bh(&head->lock);
 585                 return 0;
 586         } else {
 587                 spin_unlock(&head->lock);
 588                 /* No definite answer... Walk to established hash table */
 589                 ret = check_established(death_row, sk, snum, NULL);
 590 out:
 591                 local_bh_enable();
 592                 return ret;
 593         }
 594 }
 595
 596 /*
 597  * Bind a port for a connect operation and hash it.
 598  */
 599 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 600                       struct sock *sk)
 601 {
 602         u32 port_offset = 0;
 603
 604         if (!inet_sk(sk)->inet_num)
 605                 port_offset = inet_sk_port_offset(sk);
 606         return __inet_hash_connect(death_row, sk, port_offset,
 607                                    __inet_check_established);
 608 }
 609 EXPORT_SYMBOL_GPL(inet_hash_connect);
 610
 611 void inet_hashinfo_init(struct inet_hashinfo *h)
 612 {
 613         int i;
 614
 615         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 616                 spin_lock_init(&h->listening_hash[i].lock);
 617                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
 618                                       i + LISTENING_NULLS_BASE);
 619                 }
 620 }
 621 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 622
 623 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 624 {
 625         unsigned int locksz = sizeof(spinlock_t);
 626         unsigned int i, nblocks = 1;
 627
 628         if (locksz != 0) {
 629                 /* allocate 2 cache lines or at least one spinlock per cpu */
 630                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 631                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 632
 633                 /* no more locks than number of hash buckets */
 634                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 635
 636                 hashinfo->ehash_locks = kmalloc_array(nblocks, locksz,
 637                                                       GFP_KERNEL | __GFP_NOWARN);
 638                 if (!hashinfo->ehash_locks)
 639                         hashinfo->ehash_locks = vmalloc(nblocks * locksz);
 640
 641                 if (!hashinfo->ehash_locks)
 642                         return -ENOMEM;
 643
 644                 for (i = 0; i < nblocks; i++)
 645                         spin_lock_init(&hashinfo->ehash_locks[i]);
 646         }
 647         hashinfo->ehash_locks_mask = nblocks - 1;
 648         return 0;
 649 }
 650 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);