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 #include <net/secure_seq.h>
  24 #include <net/ip.h>
  25 #include <net/tcp.h>
  26 #include <net/sock_reuseport.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 static 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                                                  int l3mdev)
  66 {
  67         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  68
  69         if (tb) {
  70                 write_pnet(&tb->ib_net, net);
  71                 tb->l3mdev    = l3mdev;
  72                 tb->port      = snum;
  73                 tb->fastreuse = 0;
  74                 tb->fastreuseport = 0;
  75                 INIT_HLIST_HEAD(&tb->owners);
  76                 hlist_add_head(&tb->node, &head->chain);
  77         }
  78         return tb;
  79 }
  80
  81 /*
  82  * Caller must hold hashbucket lock for this tb with local BH disabled
  83  */
  84 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  85 {
  86         if (hlist_empty(&tb->owners)) {
  87                 __hlist_del(&tb->node);
  88                 kmem_cache_free(cachep, tb);
  89         }
  90 }
  91
  92 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  93                     const unsigned short snum)
  94 {
  95         inet_sk(sk)->inet_num = snum;
  96         sk_add_bind_node(sk, &tb->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         inet_csk(sk)->icsk_bind_hash = NULL;
 115         inet_sk(sk)->inet_num = 0;
 116         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 117         spin_unlock(&head->lock);
 118 }
 119
 120 void inet_put_port(struct sock *sk)
 121 {
 122         local_bh_disable();
 123         __inet_put_port(sk);
 124         local_bh_enable();
 125 }
 126 EXPORT_SYMBOL(inet_put_port);
 127
 128 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 129 {
 130         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 131         unsigned short port = inet_sk(child)->inet_num;
 132         const int bhash = inet_bhashfn(sock_net(sk), port,
 133                         table->bhash_size);
 134         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 135         struct inet_bind_bucket *tb;
 136         int l3mdev;
 137
 138         spin_lock(&head->lock);
 139         tb = inet_csk(sk)->icsk_bind_hash;
 140         if (unlikely(!tb)) {
 141                 spin_unlock(&head->lock);
 142                 return -ENOENT;
 143         }
 144         if (tb->port != port) {
 145                 l3mdev = inet_sk_bound_l3mdev(sk);
 146
 147                 /* NOTE: using tproxy and redirecting skbs to a proxy
 148                  * on a different listener port breaks the assumption
 149                  * that the listener socket's icsk_bind_hash is the same
 150                  * as that of the child socket. We have to look up or
 151                  * create a new bind bucket for the child here. */
 152                 inet_bind_bucket_for_each(tb, &head->chain) {
 153                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 154                             tb->l3mdev == l3mdev && tb->port == port)
 155                                 break;
 156                 }
 157                 if (!tb) {
 158                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 159                                                      sock_net(sk), head, port,
 160                                                      l3mdev);
 161                         if (!tb) {
 162                                 spin_unlock(&head->lock);
 163                                 return -ENOMEM;
 164                         }
 165                 }
 166         }
 167         inet_bind_hash(child, tb, port);
 168         spin_unlock(&head->lock);
 169
 170         return 0;
 171 }
 172 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 173
 174 static struct inet_listen_hashbucket *
 175 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 176 {
 177         u32 hash;
 178
 179 #if IS_ENABLED(CONFIG_IPV6)
 180         if (sk->sk_family == AF_INET6)
 181                 hash = ipv6_portaddr_hash(sock_net(sk),
 182                                           &sk->sk_v6_rcv_saddr,
 183                                           inet_sk(sk)->inet_num);
 184         else
 185 #endif
 186                 hash = ipv4_portaddr_hash(sock_net(sk),
 187                                           inet_sk(sk)->inet_rcv_saddr,
 188                                           inet_sk(sk)->inet_num);
 189         return inet_lhash2_bucket(h, hash);
 190 }
 191
 192 static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
 193 {
 194         struct inet_listen_hashbucket *ilb2;
 195
 196         if (!h->lhash2)
 197                 return;
 198
 199         ilb2 = inet_lhash2_bucket_sk(h, sk);
 200
 201         spin_lock(&ilb2->lock);
 202         if (sk->sk_reuseport && sk->sk_family == AF_INET6)
 203                 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 204                                    &ilb2->head);
 205         else
 206                 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 207                                    &ilb2->head);
 208         ilb2->count++;
 209         spin_unlock(&ilb2->lock);
 210 }
 211
 212 static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
 213 {
 214         struct inet_listen_hashbucket *ilb2;
 215
 216         if (!h->lhash2 ||
 217             WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
 218                 return;
 219
 220         ilb2 = inet_lhash2_bucket_sk(h, sk);
 221
 222         spin_lock(&ilb2->lock);
 223         hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
 224         ilb2->count--;
 225         spin_unlock(&ilb2->lock);
 226 }
 227
 228 static inline int compute_score(struct sock *sk, struct net *net,
 229                                 const unsigned short hnum, const __be32 daddr,
 230                                 const int dif, const int sdif, bool exact_dif)
 231 {
 232         int score = -1;
 233
 234         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 235                         !ipv6_only_sock(sk)) {
 236                 if (sk->sk_rcv_saddr != daddr)
 237                         return -1;
 238
 239                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 240                         return -1;
 241
 242                 score = sk->sk_family == PF_INET ? 2 : 1;
 243                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 244                         score++;
 245         }
 246         return score;
 247 }
 248
 249 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
 250                                             struct sk_buff *skb, int doff,
 251                                             __be32 saddr, __be16 sport,
 252                                             __be32 daddr, unsigned short hnum)
 253 {
 254         struct sock *reuse_sk = NULL;
 255         u32 phash;
 256
 257         if (sk->sk_reuseport) {
 258                 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 259                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 260         }
 261         return reuse_sk;
 262 }
 263
 264 /*
 265  * Here are some nice properties to exploit here. The BSD API
 266  * does not allow a listening sock to specify the remote port nor the
 267  * remote address for the connection. So always assume those are both
 268  * wildcarded during the search since they can never be otherwise.
 269  */
 270
 271 /* called with rcu_read_lock() : No refcount taken on the socket */
 272 static struct sock *inet_lhash2_lookup(struct net *net,
 273                                 struct inet_listen_hashbucket *ilb2,
 274                                 struct sk_buff *skb, int doff,
 275                                 const __be32 saddr, __be16 sport,
 276                                 const __be32 daddr, const unsigned short hnum,
 277                                 const int dif, const int sdif)
 278 {
 279         bool exact_dif = inet_exact_dif_match(net, skb);
 280         struct inet_connection_sock *icsk;
 281         struct sock *sk, *result = NULL;
 282         int score, hiscore = 0;
 283
 284         inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
 285                 sk = (struct sock *)icsk;
 286                 score = compute_score(sk, net, hnum, daddr,
 287                                       dif, sdif, exact_dif);
 288                 if (score > hiscore) {
 289                         result = lookup_reuseport(net, sk, skb, doff,
 290                                                   saddr, sport, daddr, hnum);
 291                         if (result)
 292                                 return result;
 293
 294                         result = sk;
 295                         hiscore = score;
 296                 }
 297         }
 298
 299         return result;
 300 }
 301
 302 static inline struct sock *inet_lookup_run_bpf(struct net *net,
 303                                                struct inet_hashinfo *hashinfo,
 304                                                struct sk_buff *skb, int doff,
 305                                                __be32 saddr, __be16 sport,
 306                                                __be32 daddr, u16 hnum)
 307 {
 308         struct sock *sk, *reuse_sk;
 309         bool no_reuseport;
 310
 311         if (hashinfo != &tcp_hashinfo)
 312                 return NULL; /* only TCP is supported */
 313
 314         no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP,
 315                                             saddr, sport, daddr, hnum, &sk);
 316         if (no_reuseport || IS_ERR_OR_NULL(sk))
 317                 return sk;
 318
 319         reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
 320         if (reuse_sk)
 321                 sk = reuse_sk;
 322         return sk;
 323 }
 324
 325 struct sock *__inet_lookup_listener(struct net *net,
 326                                     struct inet_hashinfo *hashinfo,
 327                                     struct sk_buff *skb, int doff,
 328                                     const __be32 saddr, __be16 sport,
 329                                     const __be32 daddr, const unsigned short hnum,
 330                                     const int dif, const int sdif)
 331 {
 332         struct inet_listen_hashbucket *ilb2;
 333         struct sock *result = NULL;
 334         unsigned int hash2;
 335
 336         /* Lookup redirect from BPF */
 337         if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
 338                 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
 339                                              saddr, sport, daddr, hnum);
 340                 if (result)
 341                         goto done;
 342         }
 343
 344         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 345         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 346
 347         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 348                                     saddr, sport, daddr, hnum,
 349                                     dif, sdif);
 350         if (result)
 351                 goto done;
 352
 353         /* Lookup lhash2 with INADDR_ANY */
 354         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 355         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 356
 357         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 358                                     saddr, sport, htonl(INADDR_ANY), hnum,
 359                                     dif, sdif);
 360 done:
 361         if (IS_ERR(result))
 362                 return NULL;
 363         return result;
 364 }
 365 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 366
 367 /* All sockets share common refcount, but have different destructors */
 368 void sock_gen_put(struct sock *sk)
 369 {
 370         if (!refcount_dec_and_test(&sk->sk_refcnt))
 371                 return;
 372
 373         if (sk->sk_state == TCP_TIME_WAIT)
 374                 inet_twsk_free(inet_twsk(sk));
 375         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 376                 reqsk_free(inet_reqsk(sk));
 377         else
 378                 sk_free(sk);
 379 }
 380 EXPORT_SYMBOL_GPL(sock_gen_put);
 381
 382 void sock_edemux(struct sk_buff *skb)
 383 {
 384         sock_gen_put(skb->sk);
 385 }
 386 EXPORT_SYMBOL(sock_edemux);
 387
 388 struct sock *__inet_lookup_established(struct net *net,
 389                                   struct inet_hashinfo *hashinfo,
 390                                   const __be32 saddr, const __be16 sport,
 391                                   const __be32 daddr, const u16 hnum,
 392                                   const int dif, const int sdif)
 393 {
 394         INET_ADDR_COOKIE(acookie, saddr, daddr);
 395         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 396         struct sock *sk;
 397         const struct hlist_nulls_node *node;
 398         /* Optimize here for direct hit, only listening connections can
 399          * have wildcards anyways.
 400          */
 401         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 402         unsigned int slot = hash & hashinfo->ehash_mask;
 403         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 404
 405 begin:
 406         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 407                 if (sk->sk_hash != hash)
 408                         continue;
 409                 if (likely(INET_MATCH(sk, net, acookie,
 410                                       saddr, daddr, ports, dif, sdif))) {
 411                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 412                                 goto out;
 413                         if (unlikely(!INET_MATCH(sk, net, acookie,
 414                                                  saddr, daddr, ports,
 415                                                  dif, sdif))) {
 416                                 sock_gen_put(sk);
 417                                 goto begin;
 418                         }
 419                         goto found;
 420                 }
 421         }
 422         /*
 423          * if the nulls value we got at the end of this lookup is
 424          * not the expected one, we must restart lookup.
 425          * We probably met an item that was moved to another chain.
 426          */
 427         if (get_nulls_value(node) != slot)
 428                 goto begin;
 429 out:
 430         sk = NULL;
 431 found:
 432         return sk;
 433 }
 434 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 435
 436 /* called with local bh disabled */
 437 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 438                                     struct sock *sk, __u16 lport,
 439                                     struct inet_timewait_sock **twp)
 440 {
 441         struct inet_hashinfo *hinfo = death_row->hashinfo;
 442         struct inet_sock *inet = inet_sk(sk);
 443         __be32 daddr = inet->inet_rcv_saddr;
 444         __be32 saddr = inet->inet_daddr;
 445         int dif = sk->sk_bound_dev_if;
 446         struct net *net = sock_net(sk);
 447         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 448         INET_ADDR_COOKIE(acookie, saddr, daddr);
 449         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 450         unsigned int hash = inet_ehashfn(net, daddr, lport,
 451                                          saddr, inet->inet_dport);
 452         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 453         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 454         struct sock *sk2;
 455         const struct hlist_nulls_node *node;
 456         struct inet_timewait_sock *tw = NULL;
 457
 458         spin_lock(lock);
 459
 460         sk_nulls_for_each(sk2, node, &head->chain) {
 461                 if (sk2->sk_hash != hash)
 462                         continue;
 463
 464                 if (likely(INET_MATCH(sk2, net, acookie,
 465                                          saddr, daddr, ports, dif, sdif))) {
 466                         if (sk2->sk_state == TCP_TIME_WAIT) {
 467                                 tw = inet_twsk(sk2);
 468                                 if (twsk_unique(sk, sk2, twp))
 469                                         break;
 470                         }
 471                         goto not_unique;
 472                 }
 473         }
 474
 475         /* Must record num and sport now. Otherwise we will see
 476          * in hash table socket with a funny identity.
 477          */
 478         inet->inet_num = lport;
 479         inet->inet_sport = htons(lport);
 480         sk->sk_hash = hash;
 481         WARN_ON(!sk_unhashed(sk));
 482         __sk_nulls_add_node_rcu(sk, &head->chain);
 483         if (tw) {
 484                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 485                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 486         }
 487         spin_unlock(lock);
 488         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 489
 490         if (twp) {
 491                 *twp = tw;
 492         } else if (tw) {
 493                 /* Silly. Should hash-dance instead... */
 494                 inet_twsk_deschedule_put(tw);
 495         }
 496         return 0;
 497
 498 not_unique:
 499         spin_unlock(lock);
 500         return -EADDRNOTAVAIL;
 501 }
 502
 503 static u32 inet_sk_port_offset(const struct sock *sk)
 504 {
 505         const struct inet_sock *inet = inet_sk(sk);
 506
 507         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 508                                           inet->inet_daddr,
 509                                           inet->inet_dport);
 510 }
 511
 512 /* insert a socket into ehash, and eventually remove another one
 513  * (The another one can be a SYN_RECV or TIMEWAIT
 514  */
 515 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
 516 {
 517         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 518         struct hlist_nulls_head *list;
 519         struct inet_ehash_bucket *head;
 520         spinlock_t *lock;
 521         bool ret = true;
 522
 523         WARN_ON_ONCE(!sk_unhashed(sk));
 524
 525         sk->sk_hash = sk_ehashfn(sk);
 526         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 527         list = &head->chain;
 528         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 529
 530         spin_lock(lock);
 531         if (osk) {
 532                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 533                 ret = sk_nulls_del_node_init_rcu(osk);
 534         }
 535         if (ret)
 536                 __sk_nulls_add_node_rcu(sk, list);
 537         spin_unlock(lock);
 538         return ret;
 539 }
 540
 541 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
 542 {
 543         bool ok = inet_ehash_insert(sk, osk);
 544
 545         if (ok) {
 546                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 547         } else {
 548                 percpu_counter_inc(sk->sk_prot->orphan_count);
 549                 inet_sk_set_state(sk, TCP_CLOSE);
 550                 sock_set_flag(sk, SOCK_DEAD);
 551                 inet_csk_destroy_sock(sk);
 552         }
 553         return ok;
 554 }
 555 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 556
 557 static int inet_reuseport_add_sock(struct sock *sk,
 558                                    struct inet_listen_hashbucket *ilb)
 559 {
 560         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 561         const struct hlist_nulls_node *node;
 562         struct sock *sk2;
 563         kuid_t uid = sock_i_uid(sk);
 564
 565         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 566                 if (sk2 != sk &&
 567                     sk2->sk_family == sk->sk_family &&
 568                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 569                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 570                     inet_csk(sk2)->icsk_bind_hash == tb &&
 571                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 572                     inet_rcv_saddr_equal(sk, sk2, false))
 573                         return reuseport_add_sock(sk, sk2,
 574                                                   inet_rcv_saddr_any(sk));
 575         }
 576
 577         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 578 }
 579
 580 int __inet_hash(struct sock *sk, struct sock *osk)
 581 {
 582         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 583         struct inet_listen_hashbucket *ilb;
 584         int err = 0;
 585
 586         if (sk->sk_state != TCP_LISTEN) {
 587                 inet_ehash_nolisten(sk, osk);
 588                 return 0;
 589         }
 590         WARN_ON(!sk_unhashed(sk));
 591         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 592
 593         spin_lock(&ilb->lock);
 594         if (sk->sk_reuseport) {
 595                 err = inet_reuseport_add_sock(sk, ilb);
 596                 if (err)
 597                         goto unlock;
 598         }
 599         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 600                 sk->sk_family == AF_INET6)
 601                 __sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
 602         else
 603                 __sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
 604         inet_hash2(hashinfo, sk);
 605         ilb->count++;
 606         sock_set_flag(sk, SOCK_RCU_FREE);
 607         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 608 unlock:
 609         spin_unlock(&ilb->lock);
 610
 611         return err;
 612 }
 613 EXPORT_SYMBOL(__inet_hash);
 614
 615 int inet_hash(struct sock *sk)
 616 {
 617         int err = 0;
 618
 619         if (sk->sk_state != TCP_CLOSE) {
 620                 local_bh_disable();
 621                 err = __inet_hash(sk, NULL);
 622                 local_bh_enable();
 623         }
 624
 625         return err;
 626 }
 627 EXPORT_SYMBOL_GPL(inet_hash);
 628
 629 void inet_unhash(struct sock *sk)
 630 {
 631         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 632         struct inet_listen_hashbucket *ilb = NULL;
 633         spinlock_t *lock;
 634
 635         if (sk_unhashed(sk))
 636                 return;
 637
 638         if (sk->sk_state == TCP_LISTEN) {
 639                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 640                 lock = &ilb->lock;
 641         } else {
 642                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 643         }
 644         spin_lock_bh(lock);
 645         if (sk_unhashed(sk))
 646                 goto unlock;
 647
 648         if (rcu_access_pointer(sk->sk_reuseport_cb))
 649                 reuseport_detach_sock(sk);
 650         if (ilb) {
 651                 inet_unhash2(hashinfo, sk);
 652                 ilb->count--;
 653         }
 654         __sk_nulls_del_node_init_rcu(sk);
 655         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 656 unlock:
 657         spin_unlock_bh(lock);
 658 }
 659 EXPORT_SYMBOL_GPL(inet_unhash);
 660
 661 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 662                 struct sock *sk, u32 port_offset,
 663                 int (*check_established)(struct inet_timewait_death_row *,
 664                         struct sock *, __u16, struct inet_timewait_sock **))
 665 {
 666         struct inet_hashinfo *hinfo = death_row->hashinfo;
 667         struct inet_timewait_sock *tw = NULL;
 668         struct inet_bind_hashbucket *head;
 669         int port = inet_sk(sk)->inet_num;
 670         struct net *net = sock_net(sk);
 671         struct inet_bind_bucket *tb;
 672         u32 remaining, offset;
 673         int ret, i, low, high;
 674         static u32 hint;
 675         int l3mdev;
 676
 677         if (port) {
 678                 head = &hinfo->bhash[inet_bhashfn(net, port,
 679                                                   hinfo->bhash_size)];
 680                 tb = inet_csk(sk)->icsk_bind_hash;
 681                 spin_lock_bh(&head->lock);
 682                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 683                         inet_ehash_nolisten(sk, NULL);
 684                         spin_unlock_bh(&head->lock);
 685                         return 0;
 686                 }
 687                 spin_unlock(&head->lock);
 688                 /* No definite answer... Walk to established hash table */
 689                 ret = check_established(death_row, sk, port, NULL);
 690                 local_bh_enable();
 691                 return ret;
 692         }
 693
 694         l3mdev = inet_sk_bound_l3mdev(sk);
 695
 696         inet_get_local_port_range(net, &low, &high);
 697         high++; /* [32768, 60999] -> [32768, 61000[ */
 698         remaining = high - low;
 699         if (likely(remaining > 1))
 700                 remaining &= ~1U;
 701
 702         offset = (hint + port_offset) % remaining;
 703         /* In first pass we try ports of @low parity.
 704          * inet_csk_get_port() does the opposite choice.
 705          */
 706         offset &= ~1U;
 707 other_parity_scan:
 708         port = low + offset;
 709         for (i = 0; i < remaining; i += 2, port += 2) {
 710                 if (unlikely(port >= high))
 711                         port -= remaining;
 712                 if (inet_is_local_reserved_port(net, port))
 713                         continue;
 714                 head = &hinfo->bhash[inet_bhashfn(net, port,
 715                                                   hinfo->bhash_size)];
 716                 spin_lock_bh(&head->lock);
 717
 718                 /* Does not bother with rcv_saddr checks, because
 719                  * the established check is already unique enough.
 720                  */
 721                 inet_bind_bucket_for_each(tb, &head->chain) {
 722                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 723                             tb->port == port) {
 724                                 if (tb->fastreuse >= 0 ||
 725                                     tb->fastreuseport >= 0)
 726                                         goto next_port;
 727                                 WARN_ON(hlist_empty(&tb->owners));
 728                                 if (!check_established(death_row, sk,
 729                                                        port, &tw))
 730                                         goto ok;
 731                                 goto next_port;
 732                         }
 733                 }
 734
 735                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 736                                              net, head, port, l3mdev);
 737                 if (!tb) {
 738                         spin_unlock_bh(&head->lock);
 739                         return -ENOMEM;
 740                 }
 741                 tb->fastreuse = -1;
 742                 tb->fastreuseport = -1;
 743                 goto ok;
 744 next_port:
 745                 spin_unlock_bh(&head->lock);
 746                 cond_resched();
 747         }
 748
 749         offset++;
 750         if ((offset & 1) && remaining > 1)
 751                 goto other_parity_scan;
 752
 753         return -EADDRNOTAVAIL;
 754
 755 ok:
 756         hint += i + 2;
 757
 758         /* Head lock still held and bh's disabled */
 759         inet_bind_hash(sk, tb, port);
 760         if (sk_unhashed(sk)) {
 761                 inet_sk(sk)->inet_sport = htons(port);
 762                 inet_ehash_nolisten(sk, (struct sock *)tw);
 763         }
 764         if (tw)
 765                 inet_twsk_bind_unhash(tw, hinfo);
 766         spin_unlock(&head->lock);
 767         if (tw)
 768                 inet_twsk_deschedule_put(tw);
 769         local_bh_enable();
 770         return 0;
 771 }
 772
 773 /*
 774  * Bind a port for a connect operation and hash it.
 775  */
 776 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 777                       struct sock *sk)
 778 {
 779         u32 port_offset = 0;
 780
 781         if (!inet_sk(sk)->inet_num)
 782                 port_offset = inet_sk_port_offset(sk);
 783         return __inet_hash_connect(death_row, sk, port_offset,
 784                                    __inet_check_established);
 785 }
 786 EXPORT_SYMBOL_GPL(inet_hash_connect);
 787
 788 void inet_hashinfo_init(struct inet_hashinfo *h)
 789 {
 790         int i;
 791
 792         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 793                 spin_lock_init(&h->listening_hash[i].lock);
 794                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
 795                                       i + LISTENING_NULLS_BASE);
 796                 h->listening_hash[i].count = 0;
 797         }
 798
 799         h->lhash2 = NULL;
 800 }
 801 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 802
 803 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
 804 {
 805         int i;
 806
 807         for (i = 0; i <= h->lhash2_mask; i++) {
 808                 spin_lock_init(&h->lhash2[i].lock);
 809                 INIT_HLIST_HEAD(&h->lhash2[i].head);
 810                 h->lhash2[i].count = 0;
 811         }
 812 }
 813
 814 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
 815                                 unsigned long numentries, int scale,
 816                                 unsigned long low_limit,
 817                                 unsigned long high_limit)
 818 {
 819         h->lhash2 = alloc_large_system_hash(name,
 820                                             sizeof(*h->lhash2),
 821                                             numentries,
 822                                             scale,
 823                                             0,
 824                                             NULL,
 825                                             &h->lhash2_mask,
 826                                             low_limit,
 827                                             high_limit);
 828         init_hashinfo_lhash2(h);
 829 }
 830
 831 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
 832 {
 833         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
 834         if (!h->lhash2)
 835                 return -ENOMEM;
 836
 837         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
 838         /* INET_LHTABLE_SIZE must be a power of 2 */
 839         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
 840
 841         init_hashinfo_lhash2(h);
 842         return 0;
 843 }
 844 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
 845
 846 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 847 {
 848         unsigned int locksz = sizeof(spinlock_t);
 849         unsigned int i, nblocks = 1;
 850
 851         if (locksz != 0) {
 852                 /* allocate 2 cache lines or at least one spinlock per cpu */
 853                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 854                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 855
 856                 /* no more locks than number of hash buckets */
 857                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 858
 859                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
 860                 if (!hashinfo->ehash_locks)
 861                         return -ENOMEM;
 862
 863                 for (i = 0; i < nblocks; i++)
 864                         spin_lock_init(&hashinfo->ehash_locks[i]);
 865         }
 866         hashinfo->ehash_locks_mask = nblocks - 1;
 867         return 0;
 868 }
 869 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);