net/ipv4/inet_hashtables.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Generic INET transport hashtables
   8  *
   9  * Authors:     Lotsa people, from code originally in tcp
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/random.h>
  14 #include <linux/sched.h>
  15 #include <linux/slab.h>
  16 #include <linux/wait.h>
  17 #include <linux/vmalloc.h>
  18 #include <linux/memblock.h>
  19
  20 #include <net/addrconf.h>
  21 #include <net/inet_connection_sock.h>
  22 #include <net/inet_hashtables.h>
  23 #if IS_ENABLED(CONFIG_IPV6)
  24 #include <net/inet6_hashtables.h>
  25 #endif
  26 #include <net/secure_seq.h>
  27 #include <net/ip.h>
  28 #include <net/tcp.h>
  29 #include <net/sock_reuseport.h>
  30
  31 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 EXPORT_SYMBOL_GPL(inet_ehashfn);
  43
  44 /* This function handles inet_sock, but also timewait and request sockets
  45  * for IPv4/IPv6.
  46  */
  47 static u32 sk_ehashfn(const struct sock *sk)
  48 {
  49 #if IS_ENABLED(CONFIG_IPV6)
  50         if (sk->sk_family == AF_INET6 &&
  51             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  52                 return inet6_ehashfn(sock_net(sk),
  53                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  54                                      &sk->sk_v6_daddr, sk->sk_dport);
  55 #endif
  56         return inet_ehashfn(sock_net(sk),
  57                             sk->sk_rcv_saddr, sk->sk_num,
  58                             sk->sk_daddr, sk->sk_dport);
  59 }
  60
  61 /*
  62  * Allocate and initialize a new local port bind bucket.
  63  * The bindhash mutex for snum's hash chain must be held here.
  64  */
  65 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  66                                                  struct net *net,
  67                                                  struct inet_bind_hashbucket *head,
  68                                                  const unsigned short snum,
  69                                                  int l3mdev)
  70 {
  71         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  72
  73         if (tb) {
  74                 write_pnet(&tb->ib_net, net);
  75                 tb->l3mdev    = l3mdev;
  76                 tb->port      = snum;
  77                 tb->fastreuse = 0;
  78                 tb->fastreuseport = 0;
  79                 INIT_HLIST_HEAD(&tb->owners);
  80                 hlist_add_head(&tb->node, &head->chain);
  81         }
  82         return tb;
  83 }
  84
  85 /*
  86  * Caller must hold hashbucket lock for this tb with local BH disabled
  87  */
  88 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  89 {
  90         if (hlist_empty(&tb->owners)) {
  91                 __hlist_del(&tb->node);
  92                 kmem_cache_free(cachep, tb);
  93         }
  94 }
  95
  96 bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
  97                             unsigned short port, int l3mdev)
  98 {
  99         return net_eq(ib_net(tb), net) && tb->port == port &&
 100                 tb->l3mdev == l3mdev;
 101 }
 102
 103 static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb,
 104                                    struct net *net,
 105                                    struct inet_bind_hashbucket *head,
 106                                    unsigned short port, int l3mdev,
 107                                    const struct sock *sk)
 108 {
 109         write_pnet(&tb->ib_net, net);
 110         tb->l3mdev    = l3mdev;
 111         tb->port      = port;
 112 #if IS_ENABLED(CONFIG_IPV6)
 113         tb->family    = sk->sk_family;
 114         if (sk->sk_family == AF_INET6)
 115                 tb->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 116         else
 117 #endif
 118                 tb->rcv_saddr = sk->sk_rcv_saddr;
 119         INIT_HLIST_HEAD(&tb->owners);
 120         INIT_HLIST_HEAD(&tb->deathrow);
 121         hlist_add_head(&tb->node, &head->chain);
 122 }
 123
 124 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
 125                                                    struct net *net,
 126                                                    struct inet_bind_hashbucket *head,
 127                                                    unsigned short port,
 128                                                    int l3mdev,
 129                                                    const struct sock *sk)
 130 {
 131         struct inet_bind2_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
 132
 133         if (tb)
 134                 inet_bind2_bucket_init(tb, net, head, port, l3mdev, sk);
 135
 136         return tb;
 137 }
 138
 139 /* Caller must hold hashbucket lock for this tb with local BH disabled */
 140 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
 141 {
 142         if (hlist_empty(&tb->owners) && hlist_empty(&tb->deathrow)) {
 143                 __hlist_del(&tb->node);
 144                 kmem_cache_free(cachep, tb);
 145         }
 146 }
 147
 148 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
 149                                          const struct sock *sk)
 150 {
 151 #if IS_ENABLED(CONFIG_IPV6)
 152         if (sk->sk_family != tb2->family)
 153                 return false;
 154
 155         if (sk->sk_family == AF_INET6)
 156                 return ipv6_addr_equal(&tb2->v6_rcv_saddr,
 157                                        &sk->sk_v6_rcv_saddr);
 158 #endif
 159         return tb2->rcv_saddr == sk->sk_rcv_saddr;
 160 }
 161
 162 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 163                     struct inet_bind2_bucket *tb2, unsigned short port)
 164 {
 165         inet_sk(sk)->inet_num = port;
 166         sk_add_bind_node(sk, &tb->owners);
 167         inet_csk(sk)->icsk_bind_hash = tb;
 168         sk_add_bind2_node(sk, &tb2->owners);
 169         inet_csk(sk)->icsk_bind2_hash = tb2;
 170 }
 171
 172 /*
 173  * Get rid of any references to a local port held by the given sock.
 174  */
 175 static void __inet_put_port(struct sock *sk)
 176 {
 177         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 178         struct inet_bind_hashbucket *head, *head2;
 179         struct net *net = sock_net(sk);
 180         struct inet_bind_bucket *tb;
 181         int bhash;
 182
 183         bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
 184         head = &hashinfo->bhash[bhash];
 185         head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
 186
 187         spin_lock(&head->lock);
 188         tb = inet_csk(sk)->icsk_bind_hash;
 189         __sk_del_bind_node(sk);
 190         inet_csk(sk)->icsk_bind_hash = NULL;
 191         inet_sk(sk)->inet_num = 0;
 192         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 193
 194         spin_lock(&head2->lock);
 195         if (inet_csk(sk)->icsk_bind2_hash) {
 196                 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
 197
 198                 __sk_del_bind2_node(sk);
 199                 inet_csk(sk)->icsk_bind2_hash = NULL;
 200                 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
 201         }
 202         spin_unlock(&head2->lock);
 203
 204         spin_unlock(&head->lock);
 205 }
 206
 207 void inet_put_port(struct sock *sk)
 208 {
 209         local_bh_disable();
 210         __inet_put_port(sk);
 211         local_bh_enable();
 212 }
 213 EXPORT_SYMBOL(inet_put_port);
 214
 215 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 216 {
 217         struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
 218         unsigned short port = inet_sk(child)->inet_num;
 219         struct inet_bind_hashbucket *head, *head2;
 220         bool created_inet_bind_bucket = false;
 221         struct net *net = sock_net(sk);
 222         bool update_fastreuse = false;
 223         struct inet_bind2_bucket *tb2;
 224         struct inet_bind_bucket *tb;
 225         int bhash, l3mdev;
 226
 227         bhash = inet_bhashfn(net, port, table->bhash_size);
 228         head = &table->bhash[bhash];
 229         head2 = inet_bhashfn_portaddr(table, child, net, port);
 230
 231         spin_lock(&head->lock);
 232         spin_lock(&head2->lock);
 233         tb = inet_csk(sk)->icsk_bind_hash;
 234         tb2 = inet_csk(sk)->icsk_bind2_hash;
 235         if (unlikely(!tb || !tb2)) {
 236                 spin_unlock(&head2->lock);
 237                 spin_unlock(&head->lock);
 238                 return -ENOENT;
 239         }
 240         if (tb->port != port) {
 241                 l3mdev = inet_sk_bound_l3mdev(sk);
 242
 243                 /* NOTE: using tproxy and redirecting skbs to a proxy
 244                  * on a different listener port breaks the assumption
 245                  * that the listener socket's icsk_bind_hash is the same
 246                  * as that of the child socket. We have to look up or
 247                  * create a new bind bucket for the child here. */
 248                 inet_bind_bucket_for_each(tb, &head->chain) {
 249                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 250                                 break;
 251                 }
 252                 if (!tb) {
 253                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 254                                                      net, head, port, l3mdev);
 255                         if (!tb) {
 256                                 spin_unlock(&head2->lock);
 257                                 spin_unlock(&head->lock);
 258                                 return -ENOMEM;
 259                         }
 260                         created_inet_bind_bucket = true;
 261                 }
 262                 update_fastreuse = true;
 263
 264                 goto bhash2_find;
 265         } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
 266                 l3mdev = inet_sk_bound_l3mdev(sk);
 267
 268 bhash2_find:
 269                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
 270                 if (!tb2) {
 271                         tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
 272                                                        net, head2, port,
 273                                                        l3mdev, child);
 274                         if (!tb2)
 275                                 goto error;
 276                 }
 277         }
 278         if (update_fastreuse)
 279                 inet_csk_update_fastreuse(tb, child);
 280         inet_bind_hash(child, tb, tb2, port);
 281         spin_unlock(&head2->lock);
 282         spin_unlock(&head->lock);
 283
 284         return 0;
 285
 286 error:
 287         if (created_inet_bind_bucket)
 288                 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
 289         spin_unlock(&head2->lock);
 290         spin_unlock(&head->lock);
 291         return -ENOMEM;
 292 }
 293 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 294
 295 static struct inet_listen_hashbucket *
 296 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 297 {
 298         u32 hash;
 299
 300 #if IS_ENABLED(CONFIG_IPV6)
 301         if (sk->sk_family == AF_INET6)
 302                 hash = ipv6_portaddr_hash(sock_net(sk),
 303                                           &sk->sk_v6_rcv_saddr,
 304                                           inet_sk(sk)->inet_num);
 305         else
 306 #endif
 307                 hash = ipv4_portaddr_hash(sock_net(sk),
 308                                           inet_sk(sk)->inet_rcv_saddr,
 309                                           inet_sk(sk)->inet_num);
 310         return inet_lhash2_bucket(h, hash);
 311 }
 312
 313 static inline int compute_score(struct sock *sk, struct net *net,
 314                                 const unsigned short hnum, const __be32 daddr,
 315                                 const int dif, const int sdif)
 316 {
 317         int score = -1;
 318
 319         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 320                         !ipv6_only_sock(sk)) {
 321                 if (sk->sk_rcv_saddr != daddr)
 322                         return -1;
 323
 324                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 325                         return -1;
 326                 score =  sk->sk_bound_dev_if ? 2 : 1;
 327
 328                 if (sk->sk_family == PF_INET)
 329                         score++;
 330                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 331                         score++;
 332         }
 333         return score;
 334 }
 335
 336 INDIRECT_CALLABLE_DECLARE(inet_ehashfn_t udp_ehashfn);
 337
 338 /**
 339  * inet_lookup_reuseport() - execute reuseport logic on AF_INET socket if necessary.
 340  * @net: network namespace.
 341  * @sk: AF_INET socket, must be in TCP_LISTEN state for TCP or TCP_CLOSE for UDP.
 342  * @skb: context for a potential SK_REUSEPORT program.
 343  * @doff: header offset.
 344  * @saddr: source address.
 345  * @sport: source port.
 346  * @daddr: destination address.
 347  * @hnum: destination port in host byte order.
 348  * @ehashfn: hash function used to generate the fallback hash.
 349  *
 350  * Return: NULL if sk doesn't have SO_REUSEPORT set, otherwise a pointer to
 351  *         the selected sock or an error.
 352  */
 353 struct sock *inet_lookup_reuseport(struct net *net, struct sock *sk,
 354                                    struct sk_buff *skb, int doff,
 355                                    __be32 saddr, __be16 sport,
 356                                    __be32 daddr, unsigned short hnum,
 357                                    inet_ehashfn_t *ehashfn)
 358 {
 359         struct sock *reuse_sk = NULL;
 360         u32 phash;
 361
 362         if (sk->sk_reuseport) {
 363                 phash = INDIRECT_CALL_2(ehashfn, udp_ehashfn, inet_ehashfn,
 364                                         net, daddr, hnum, saddr, sport);
 365                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 366         }
 367         return reuse_sk;
 368 }
 369 EXPORT_SYMBOL_GPL(inet_lookup_reuseport);
 370
 371 /*
 372  * Here are some nice properties to exploit here. The BSD API
 373  * does not allow a listening sock to specify the remote port nor the
 374  * remote address for the connection. So always assume those are both
 375  * wildcarded during the search since they can never be otherwise.
 376  */
 377
 378 /* called with rcu_read_lock() : No refcount taken on the socket */
 379 static struct sock *inet_lhash2_lookup(struct net *net,
 380                                 struct inet_listen_hashbucket *ilb2,
 381                                 struct sk_buff *skb, int doff,
 382                                 const __be32 saddr, __be16 sport,
 383                                 const __be32 daddr, const unsigned short hnum,
 384                                 const int dif, const int sdif)
 385 {
 386         struct sock *sk, *result = NULL;
 387         struct hlist_nulls_node *node;
 388         int score, hiscore = 0;
 389
 390         sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
 391                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 392                 if (score > hiscore) {
 393                         result = inet_lookup_reuseport(net, sk, skb, doff,
 394                                                        saddr, sport, daddr, hnum, inet_ehashfn);
 395                         if (result)
 396                                 return result;
 397
 398                         result = sk;
 399                         hiscore = score;
 400                 }
 401         }
 402
 403         return result;
 404 }
 405
 406 struct sock *inet_lookup_run_sk_lookup(struct net *net,
 407                                        int protocol,
 408                                        struct sk_buff *skb, int doff,
 409                                        __be32 saddr, __be16 sport,
 410                                        __be32 daddr, u16 hnum, const int dif,
 411                                        inet_ehashfn_t *ehashfn)
 412 {
 413         struct sock *sk, *reuse_sk;
 414         bool no_reuseport;
 415
 416         no_reuseport = bpf_sk_lookup_run_v4(net, protocol, saddr, sport,
 417                                             daddr, hnum, dif, &sk);
 418         if (no_reuseport || IS_ERR_OR_NULL(sk))
 419                 return sk;
 420
 421         reuse_sk = inet_lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum,
 422                                          ehashfn);
 423         if (reuse_sk)
 424                 sk = reuse_sk;
 425         return sk;
 426 }
 427
 428 struct sock *__inet_lookup_listener(struct net *net,
 429                                     struct inet_hashinfo *hashinfo,
 430                                     struct sk_buff *skb, int doff,
 431                                     const __be32 saddr, __be16 sport,
 432                                     const __be32 daddr, const unsigned short hnum,
 433                                     const int dif, const int sdif)
 434 {
 435         struct inet_listen_hashbucket *ilb2;
 436         struct sock *result = NULL;
 437         unsigned int hash2;
 438
 439         /* Lookup redirect from BPF */
 440         if (static_branch_unlikely(&bpf_sk_lookup_enabled) &&
 441             hashinfo == net->ipv4.tcp_death_row.hashinfo) {
 442                 result = inet_lookup_run_sk_lookup(net, IPPROTO_TCP, skb, doff,
 443                                                    saddr, sport, daddr, hnum, dif,
 444                                                    inet_ehashfn);
 445                 if (result)
 446                         goto done;
 447         }
 448
 449         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 450         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 451
 452         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 453                                     saddr, sport, daddr, hnum,
 454                                     dif, sdif);
 455         if (result)
 456                 goto done;
 457
 458         /* Lookup lhash2 with INADDR_ANY */
 459         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 460         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 461
 462         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 463                                     saddr, sport, htonl(INADDR_ANY), hnum,
 464                                     dif, sdif);
 465 done:
 466         if (IS_ERR(result))
 467                 return NULL;
 468         return result;
 469 }
 470 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 471
 472 /* All sockets share common refcount, but have different destructors */
 473 void sock_gen_put(struct sock *sk)
 474 {
 475         if (!refcount_dec_and_test(&sk->sk_refcnt))
 476                 return;
 477
 478         if (sk->sk_state == TCP_TIME_WAIT)
 479                 inet_twsk_free(inet_twsk(sk));
 480         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 481                 reqsk_free(inet_reqsk(sk));
 482         else
 483                 sk_free(sk);
 484 }
 485 EXPORT_SYMBOL_GPL(sock_gen_put);
 486
 487 void sock_edemux(struct sk_buff *skb)
 488 {
 489         sock_gen_put(skb->sk);
 490 }
 491 EXPORT_SYMBOL(sock_edemux);
 492
 493 struct sock *__inet_lookup_established(struct net *net,
 494                                   struct inet_hashinfo *hashinfo,
 495                                   const __be32 saddr, const __be16 sport,
 496                                   const __be32 daddr, const u16 hnum,
 497                                   const int dif, const int sdif)
 498 {
 499         INET_ADDR_COOKIE(acookie, saddr, daddr);
 500         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 501         struct sock *sk;
 502         const struct hlist_nulls_node *node;
 503         /* Optimize here for direct hit, only listening connections can
 504          * have wildcards anyways.
 505          */
 506         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 507         unsigned int slot = hash & hashinfo->ehash_mask;
 508         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 509
 510 begin:
 511         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 512                 if (sk->sk_hash != hash)
 513                         continue;
 514                 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
 515                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 516                                 goto out;
 517                         if (unlikely(!inet_match(net, sk, acookie,
 518                                                  ports, dif, sdif))) {
 519                                 sock_gen_put(sk);
 520                                 goto begin;
 521                         }
 522                         goto found;
 523                 }
 524         }
 525         /*
 526          * if the nulls value we got at the end of this lookup is
 527          * not the expected one, we must restart lookup.
 528          * We probably met an item that was moved to another chain.
 529          */
 530         if (get_nulls_value(node) != slot)
 531                 goto begin;
 532 out:
 533         sk = NULL;
 534 found:
 535         return sk;
 536 }
 537 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 538
 539 /* called with local bh disabled */
 540 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 541                                     struct sock *sk, __u16 lport,
 542                                     struct inet_timewait_sock **twp)
 543 {
 544         struct inet_hashinfo *hinfo = death_row->hashinfo;
 545         struct inet_sock *inet = inet_sk(sk);
 546         __be32 daddr = inet->inet_rcv_saddr;
 547         __be32 saddr = inet->inet_daddr;
 548         int dif = sk->sk_bound_dev_if;
 549         struct net *net = sock_net(sk);
 550         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 551         INET_ADDR_COOKIE(acookie, saddr, daddr);
 552         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 553         unsigned int hash = inet_ehashfn(net, daddr, lport,
 554                                          saddr, inet->inet_dport);
 555         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 556         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 557         struct sock *sk2;
 558         const struct hlist_nulls_node *node;
 559         struct inet_timewait_sock *tw = NULL;
 560
 561         spin_lock(lock);
 562
 563         sk_nulls_for_each(sk2, node, &head->chain) {
 564                 if (sk2->sk_hash != hash)
 565                         continue;
 566
 567                 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
 568                         if (sk2->sk_state == TCP_TIME_WAIT) {
 569                                 tw = inet_twsk(sk2);
 570                                 if (twsk_unique(sk, sk2, twp))
 571                                         break;
 572                         }
 573                         goto not_unique;
 574                 }
 575         }
 576
 577         /* Must record num and sport now. Otherwise we will see
 578          * in hash table socket with a funny identity.
 579          */
 580         inet->inet_num = lport;
 581         inet->inet_sport = htons(lport);
 582         sk->sk_hash = hash;
 583         WARN_ON(!sk_unhashed(sk));
 584         __sk_nulls_add_node_rcu(sk, &head->chain);
 585         if (tw) {
 586                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 587                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 588         }
 589         spin_unlock(lock);
 590         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 591
 592         if (twp) {
 593                 *twp = tw;
 594         } else if (tw) {
 595                 /* Silly. Should hash-dance instead... */
 596                 inet_twsk_deschedule_put(tw);
 597         }
 598         return 0;
 599
 600 not_unique:
 601         spin_unlock(lock);
 602         return -EADDRNOTAVAIL;
 603 }
 604
 605 static u64 inet_sk_port_offset(const struct sock *sk)
 606 {
 607         const struct inet_sock *inet = inet_sk(sk);
 608
 609         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 610                                           inet->inet_daddr,
 611                                           inet->inet_dport);
 612 }
 613
 614 /* Searches for an exsiting socket in the ehash bucket list.
 615  * Returns true if found, false otherwise.
 616  */
 617 static bool inet_ehash_lookup_by_sk(struct sock *sk,
 618                                     struct hlist_nulls_head *list)
 619 {
 620         const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
 621         const int sdif = sk->sk_bound_dev_if;
 622         const int dif = sk->sk_bound_dev_if;
 623         const struct hlist_nulls_node *node;
 624         struct net *net = sock_net(sk);
 625         struct sock *esk;
 626
 627         INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
 628
 629         sk_nulls_for_each_rcu(esk, node, list) {
 630                 if (esk->sk_hash != sk->sk_hash)
 631                         continue;
 632                 if (sk->sk_family == AF_INET) {
 633                         if (unlikely(inet_match(net, esk, acookie,
 634                                                 ports, dif, sdif))) {
 635                                 return true;
 636                         }
 637                 }
 638 #if IS_ENABLED(CONFIG_IPV6)
 639                 else if (sk->sk_family == AF_INET6) {
 640                         if (unlikely(inet6_match(net, esk,
 641                                                  &sk->sk_v6_daddr,
 642                                                  &sk->sk_v6_rcv_saddr,
 643                                                  ports, dif, sdif))) {
 644                                 return true;
 645                         }
 646                 }
 647 #endif
 648         }
 649         return false;
 650 }
 651
 652 /* Insert a socket into ehash, and eventually remove another one
 653  * (The another one can be a SYN_RECV or TIMEWAIT)
 654  * If an existing socket already exists, socket sk is not inserted,
 655  * and sets found_dup_sk parameter to true.
 656  */
 657 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 658 {
 659         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 660         struct inet_ehash_bucket *head;
 661         struct hlist_nulls_head *list;
 662         spinlock_t *lock;
 663         bool ret = true;
 664
 665         WARN_ON_ONCE(!sk_unhashed(sk));
 666
 667         sk->sk_hash = sk_ehashfn(sk);
 668         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 669         list = &head->chain;
 670         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 671
 672         spin_lock(lock);
 673         if (osk) {
 674                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 675                 ret = sk_nulls_del_node_init_rcu(osk);
 676         } else if (found_dup_sk) {
 677                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 678                 if (*found_dup_sk)
 679                         ret = false;
 680         }
 681
 682         if (ret)
 683                 __sk_nulls_add_node_rcu(sk, list);
 684
 685         spin_unlock(lock);
 686
 687         return ret;
 688 }
 689
 690 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 691 {
 692         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 693
 694         if (ok) {
 695                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 696         } else {
 697                 this_cpu_inc(*sk->sk_prot->orphan_count);
 698                 inet_sk_set_state(sk, TCP_CLOSE);
 699                 sock_set_flag(sk, SOCK_DEAD);
 700                 inet_csk_destroy_sock(sk);
 701         }
 702         return ok;
 703 }
 704 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 705
 706 static int inet_reuseport_add_sock(struct sock *sk,
 707                                    struct inet_listen_hashbucket *ilb)
 708 {
 709         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 710         const struct hlist_nulls_node *node;
 711         struct sock *sk2;
 712         kuid_t uid = sock_i_uid(sk);
 713
 714         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 715                 if (sk2 != sk &&
 716                     sk2->sk_family == sk->sk_family &&
 717                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 718                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 719                     inet_csk(sk2)->icsk_bind_hash == tb &&
 720                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 721                     inet_rcv_saddr_equal(sk, sk2, false))
 722                         return reuseport_add_sock(sk, sk2,
 723                                                   inet_rcv_saddr_any(sk));
 724         }
 725
 726         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 727 }
 728
 729 int __inet_hash(struct sock *sk, struct sock *osk)
 730 {
 731         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 732         struct inet_listen_hashbucket *ilb2;
 733         int err = 0;
 734
 735         if (sk->sk_state != TCP_LISTEN) {
 736                 local_bh_disable();
 737                 inet_ehash_nolisten(sk, osk, NULL);
 738                 local_bh_enable();
 739                 return 0;
 740         }
 741         WARN_ON(!sk_unhashed(sk));
 742         ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 743
 744         spin_lock(&ilb2->lock);
 745         if (sk->sk_reuseport) {
 746                 err = inet_reuseport_add_sock(sk, ilb2);
 747                 if (err)
 748                         goto unlock;
 749         }
 750         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 751                 sk->sk_family == AF_INET6)
 752                 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
 753         else
 754                 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
 755         sock_set_flag(sk, SOCK_RCU_FREE);
 756         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 757 unlock:
 758         spin_unlock(&ilb2->lock);
 759
 760         return err;
 761 }
 762 EXPORT_SYMBOL(__inet_hash);
 763
 764 int inet_hash(struct sock *sk)
 765 {
 766         int err = 0;
 767
 768         if (sk->sk_state != TCP_CLOSE)
 769                 err = __inet_hash(sk, NULL);
 770
 771         return err;
 772 }
 773 EXPORT_SYMBOL_GPL(inet_hash);
 774
 775 void inet_unhash(struct sock *sk)
 776 {
 777         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 778
 779         if (sk_unhashed(sk))
 780                 return;
 781
 782         if (sk->sk_state == TCP_LISTEN) {
 783                 struct inet_listen_hashbucket *ilb2;
 784
 785                 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 786                 /* Don't disable bottom halves while acquiring the lock to
 787                  * avoid circular locking dependency on PREEMPT_RT.
 788                  */
 789                 spin_lock(&ilb2->lock);
 790                 if (sk_unhashed(sk)) {
 791                         spin_unlock(&ilb2->lock);
 792                         return;
 793                 }
 794
 795                 if (rcu_access_pointer(sk->sk_reuseport_cb))
 796                         reuseport_stop_listen_sock(sk);
 797
 798                 __sk_nulls_del_node_init_rcu(sk);
 799                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 800                 spin_unlock(&ilb2->lock);
 801         } else {
 802                 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 803
 804                 spin_lock_bh(lock);
 805                 if (sk_unhashed(sk)) {
 806                         spin_unlock_bh(lock);
 807                         return;
 808                 }
 809                 __sk_nulls_del_node_init_rcu(sk);
 810                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 811                 spin_unlock_bh(lock);
 812         }
 813 }
 814 EXPORT_SYMBOL_GPL(inet_unhash);
 815
 816 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
 817                                     const struct net *net, unsigned short port,
 818                                     int l3mdev, const struct sock *sk)
 819 {
 820 #if IS_ENABLED(CONFIG_IPV6)
 821         if (sk->sk_family != tb->family)
 822                 return false;
 823
 824         if (sk->sk_family == AF_INET6)
 825                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 826                         tb->l3mdev == l3mdev &&
 827                         ipv6_addr_equal(&tb->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
 828         else
 829 #endif
 830                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 831                         tb->l3mdev == l3mdev && tb->rcv_saddr == sk->sk_rcv_saddr;
 832 }
 833
 834 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
 835                                       unsigned short port, int l3mdev, const struct sock *sk)
 836 {
 837 #if IS_ENABLED(CONFIG_IPV6)
 838         if (sk->sk_family != tb->family) {
 839                 if (sk->sk_family == AF_INET)
 840                         return net_eq(ib2_net(tb), net) && tb->port == port &&
 841                                 tb->l3mdev == l3mdev &&
 842                                 ipv6_addr_any(&tb->v6_rcv_saddr);
 843
 844                 return false;
 845         }
 846
 847         if (sk->sk_family == AF_INET6)
 848                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 849                         tb->l3mdev == l3mdev &&
 850                         ipv6_addr_any(&tb->v6_rcv_saddr);
 851         else
 852 #endif
 853                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 854                         tb->l3mdev == l3mdev && tb->rcv_saddr == 0;
 855 }
 856
 857 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
 858 struct inet_bind2_bucket *
 859 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
 860                        unsigned short port, int l3mdev, const struct sock *sk)
 861 {
 862         struct inet_bind2_bucket *bhash2 = NULL;
 863
 864         inet_bind_bucket_for_each(bhash2, &head->chain)
 865                 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
 866                         break;
 867
 868         return bhash2;
 869 }
 870
 871 struct inet_bind_hashbucket *
 872 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
 873 {
 874         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 875         u32 hash;
 876
 877 #if IS_ENABLED(CONFIG_IPV6)
 878         if (sk->sk_family == AF_INET6)
 879                 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
 880         else
 881 #endif
 882                 hash = ipv4_portaddr_hash(net, 0, port);
 883
 884         return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
 885 }
 886
 887 static void inet_update_saddr(struct sock *sk, void *saddr, int family)
 888 {
 889         if (family == AF_INET) {
 890                 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
 891                 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
 892         }
 893 #if IS_ENABLED(CONFIG_IPV6)
 894         else {
 895                 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
 896         }
 897 #endif
 898 }
 899
 900 static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
 901 {
 902         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 903         struct inet_bind_hashbucket *head, *head2;
 904         struct inet_bind2_bucket *tb2, *new_tb2;
 905         int l3mdev = inet_sk_bound_l3mdev(sk);
 906         int port = inet_sk(sk)->inet_num;
 907         struct net *net = sock_net(sk);
 908         int bhash;
 909
 910         if (!inet_csk(sk)->icsk_bind2_hash) {
 911                 /* Not bind()ed before. */
 912                 if (reset)
 913                         inet_reset_saddr(sk);
 914                 else
 915                         inet_update_saddr(sk, saddr, family);
 916
 917                 return 0;
 918         }
 919
 920         /* Allocate a bind2 bucket ahead of time to avoid permanently putting
 921          * the bhash2 table in an inconsistent state if a new tb2 bucket
 922          * allocation fails.
 923          */
 924         new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
 925         if (!new_tb2) {
 926                 if (reset) {
 927                         /* The (INADDR_ANY, port) bucket might have already
 928                          * been freed, then we cannot fixup icsk_bind2_hash,
 929                          * so we give up and unlink sk from bhash/bhash2 not
 930                          * to leave inconsistency in bhash2.
 931                          */
 932                         inet_put_port(sk);
 933                         inet_reset_saddr(sk);
 934                 }
 935
 936                 return -ENOMEM;
 937         }
 938
 939         bhash = inet_bhashfn(net, port, hinfo->bhash_size);
 940         head = &hinfo->bhash[bhash];
 941         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 942
 943         /* If we change saddr locklessly, another thread
 944          * iterating over bhash might see corrupted address.
 945          */
 946         spin_lock_bh(&head->lock);
 947
 948         spin_lock(&head2->lock);
 949         __sk_del_bind2_node(sk);
 950         inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
 951         spin_unlock(&head2->lock);
 952
 953         if (reset)
 954                 inet_reset_saddr(sk);
 955         else
 956                 inet_update_saddr(sk, saddr, family);
 957
 958         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 959
 960         spin_lock(&head2->lock);
 961         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 962         if (!tb2) {
 963                 tb2 = new_tb2;
 964                 inet_bind2_bucket_init(tb2, net, head2, port, l3mdev, sk);
 965         }
 966         sk_add_bind2_node(sk, &tb2->owners);
 967         inet_csk(sk)->icsk_bind2_hash = tb2;
 968         spin_unlock(&head2->lock);
 969
 970         spin_unlock_bh(&head->lock);
 971
 972         if (tb2 != new_tb2)
 973                 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
 974
 975         return 0;
 976 }
 977
 978 int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
 979 {
 980         return __inet_bhash2_update_saddr(sk, saddr, family, false);
 981 }
 982 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
 983
 984 void inet_bhash2_reset_saddr(struct sock *sk)
 985 {
 986         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
 987                 __inet_bhash2_update_saddr(sk, NULL, 0, true);
 988 }
 989 EXPORT_SYMBOL_GPL(inet_bhash2_reset_saddr);
 990
 991 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 992  * Note that we use 32bit integers (vs RFC 'short integers')
 993  * because 2^16 is not a multiple of num_ephemeral and this
 994  * property might be used by clever attacker.
 995  *
 996  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
 997  * attacks were since demonstrated, thus we use 65536 by default instead
 998  * to really give more isolation and privacy, at the expense of 256kB
 999  * of kernel memory.
1000  */
1001 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
1002 static u32 *table_perturb;
1003
1004 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
1005                 struct sock *sk, u64 port_offset,
1006                 int (*check_established)(struct inet_timewait_death_row *,
1007                         struct sock *, __u16, struct inet_timewait_sock **))
1008 {
1009         struct inet_hashinfo *hinfo = death_row->hashinfo;
1010         struct inet_bind_hashbucket *head, *head2;
1011         struct inet_timewait_sock *tw = NULL;
1012         int port = inet_sk(sk)->inet_num;
1013         struct net *net = sock_net(sk);
1014         struct inet_bind2_bucket *tb2;
1015         struct inet_bind_bucket *tb;
1016         bool tb_created = false;
1017         u32 remaining, offset;
1018         int ret, i, low, high;
1019         int l3mdev;
1020         u32 index;
1021
1022         if (port) {
1023                 local_bh_disable();
1024                 ret = check_established(death_row, sk, port, NULL);
1025                 local_bh_enable();
1026                 return ret;
1027         }
1028
1029         l3mdev = inet_sk_bound_l3mdev(sk);
1030
1031         inet_sk_get_local_port_range(sk, &low, &high);
1032         high++; /* [32768, 60999] -> [32768, 61000[ */
1033         remaining = high - low;
1034         if (likely(remaining > 1))
1035                 remaining &= ~1U;
1036
1037         get_random_sleepable_once(table_perturb,
1038                                   INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1039         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1040
1041         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1042         offset %= remaining;
1043
1044         /* In first pass we try ports of @low parity.
1045          * inet_csk_get_port() does the opposite choice.
1046          */
1047         offset &= ~1U;
1048 other_parity_scan:
1049         port = low + offset;
1050         for (i = 0; i < remaining; i += 2, port += 2) {
1051                 if (unlikely(port >= high))
1052                         port -= remaining;
1053                 if (inet_is_local_reserved_port(net, port))
1054                         continue;
1055                 head = &hinfo->bhash[inet_bhashfn(net, port,
1056                                                   hinfo->bhash_size)];
1057                 spin_lock_bh(&head->lock);
1058
1059                 /* Does not bother with rcv_saddr checks, because
1060                  * the established check is already unique enough.
1061                  */
1062                 inet_bind_bucket_for_each(tb, &head->chain) {
1063                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1064                                 if (tb->fastreuse >= 0 ||
1065                                     tb->fastreuseport >= 0)
1066                                         goto next_port;
1067                                 WARN_ON(hlist_empty(&tb->owners));
1068                                 if (!check_established(death_row, sk,
1069                                                        port, &tw))
1070                                         goto ok;
1071                                 goto next_port;
1072                         }
1073                 }
1074
1075                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1076                                              net, head, port, l3mdev);
1077                 if (!tb) {
1078                         spin_unlock_bh(&head->lock);
1079                         return -ENOMEM;
1080                 }
1081                 tb_created = true;
1082                 tb->fastreuse = -1;
1083                 tb->fastreuseport = -1;
1084                 goto ok;
1085 next_port:
1086                 spin_unlock_bh(&head->lock);
1087                 cond_resched();
1088         }
1089
1090         offset++;
1091         if ((offset & 1) && remaining > 1)
1092                 goto other_parity_scan;
1093
1094         return -EADDRNOTAVAIL;
1095
1096 ok:
1097         /* Find the corresponding tb2 bucket since we need to
1098          * add the socket to the bhash2 table as well
1099          */
1100         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1101         spin_lock(&head2->lock);
1102
1103         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1104         if (!tb2) {
1105                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1106                                                head2, port, l3mdev, sk);
1107                 if (!tb2)
1108                         goto error;
1109         }
1110
1111         /* Here we want to add a little bit of randomness to the next source
1112          * port that will be chosen. We use a max() with a random here so that
1113          * on low contention the randomness is maximal and on high contention
1114          * it may be inexistent.
1115          */
1116         i = max_t(int, i, get_random_u32_below(8) * 2);
1117         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
1118
1119         /* Head lock still held and bh's disabled */
1120         inet_bind_hash(sk, tb, tb2, port);
1121
1122         if (sk_unhashed(sk)) {
1123                 inet_sk(sk)->inet_sport = htons(port);
1124                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1125         }
1126         if (tw)
1127                 inet_twsk_bind_unhash(tw, hinfo);
1128
1129         spin_unlock(&head2->lock);
1130         spin_unlock(&head->lock);
1131
1132         if (tw)
1133                 inet_twsk_deschedule_put(tw);
1134         local_bh_enable();
1135         return 0;
1136
1137 error:
1138         spin_unlock(&head2->lock);
1139         if (tb_created)
1140                 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1141         spin_unlock_bh(&head->lock);
1142         return -ENOMEM;
1143 }
1144
1145 /*
1146  * Bind a port for a connect operation and hash it.
1147  */
1148 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1149                       struct sock *sk)
1150 {
1151         u64 port_offset = 0;
1152
1153         if (!inet_sk(sk)->inet_num)
1154                 port_offset = inet_sk_port_offset(sk);
1155         return __inet_hash_connect(death_row, sk, port_offset,
1156                                    __inet_check_established);
1157 }
1158 EXPORT_SYMBOL_GPL(inet_hash_connect);
1159
1160 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1161 {
1162         int i;
1163
1164         for (i = 0; i <= h->lhash2_mask; i++) {
1165                 spin_lock_init(&h->lhash2[i].lock);
1166                 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1167                                       i + LISTENING_NULLS_BASE);
1168         }
1169 }
1170
1171 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1172                                 unsigned long numentries, int scale,
1173                                 unsigned long low_limit,
1174                                 unsigned long high_limit)
1175 {
1176         h->lhash2 = alloc_large_system_hash(name,
1177                                             sizeof(*h->lhash2),
1178                                             numentries,
1179                                             scale,
1180                                             0,
1181                                             NULL,
1182                                             &h->lhash2_mask,
1183                                             low_limit,
1184                                             high_limit);
1185         init_hashinfo_lhash2(h);
1186
1187         /* this one is used for source ports of outgoing connections */
1188         table_perturb = alloc_large_system_hash("Table-perturb",
1189                                                 sizeof(*table_perturb),
1190                                                 INET_TABLE_PERTURB_SIZE,
1191                                                 0, 0, NULL, NULL,
1192                                                 INET_TABLE_PERTURB_SIZE,
1193                                                 INET_TABLE_PERTURB_SIZE);
1194 }
1195
1196 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1197 {
1198         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1199         if (!h->lhash2)
1200                 return -ENOMEM;
1201
1202         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1203         /* INET_LHTABLE_SIZE must be a power of 2 */
1204         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1205
1206         init_hashinfo_lhash2(h);
1207         return 0;
1208 }
1209 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1210
1211 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1212 {
1213         unsigned int locksz = sizeof(spinlock_t);
1214         unsigned int i, nblocks = 1;
1215
1216         if (locksz != 0) {
1217                 /* allocate 2 cache lines or at least one spinlock per cpu */
1218                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1219                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1220
1221                 /* no more locks than number of hash buckets */
1222                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1223
1224                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1225                 if (!hashinfo->ehash_locks)
1226                         return -ENOMEM;
1227
1228                 for (i = 0; i < nblocks; i++)
1229                         spin_lock_init(&hashinfo->ehash_locks[i]);
1230         }
1231         hashinfo->ehash_locks_mask = nblocks - 1;
1232         return 0;
1233 }
1234 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1235
1236 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1237                                                  unsigned int ehash_entries)
1238 {
1239         struct inet_hashinfo *new_hashinfo;
1240         int i;
1241
1242         new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1243         if (!new_hashinfo)
1244                 goto err;
1245
1246         new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1247                                            GFP_KERNEL_ACCOUNT);
1248         if (!new_hashinfo->ehash)
1249                 goto free_hashinfo;
1250
1251         new_hashinfo->ehash_mask = ehash_entries - 1;
1252
1253         if (inet_ehash_locks_alloc(new_hashinfo))
1254                 goto free_ehash;
1255
1256         for (i = 0; i < ehash_entries; i++)
1257                 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1258
1259         new_hashinfo->pernet = true;
1260
1261         return new_hashinfo;
1262
1263 free_ehash:
1264         vfree(new_hashinfo->ehash);
1265 free_hashinfo:
1266         kfree(new_hashinfo);
1267 err:
1268         return NULL;
1269 }
1270 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1271
1272 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1273 {
1274         if (!hashinfo->pernet)
1275                 return;
1276
1277         inet_ehash_locks_free(hashinfo);
1278         vfree(hashinfo->ehash);
1279         kfree(hashinfo);
1280 }
1281 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);