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