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