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 static 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
  43 /* This function handles inet_sock, but also timewait and request sockets
  44  * for IPv4/IPv6.
  45  */
  46 static u32 sk_ehashfn(const struct sock *sk)
  47 {
  48 #if IS_ENABLED(CONFIG_IPV6)
  49         if (sk->sk_family == AF_INET6 &&
  50             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  51                 return inet6_ehashfn(sock_net(sk),
  52                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  53                                      &sk->sk_v6_daddr, sk->sk_dport);
  54 #endif
  55         return inet_ehashfn(sock_net(sk),
  56                             sk->sk_rcv_saddr, sk->sk_num,
  57                             sk->sk_daddr, sk->sk_dport);
  58 }
  59
  60 /*
  61  * Allocate and initialize a new local port bind bucket.
  62  * The bindhash mutex for snum's hash chain must be held here.
  63  */
  64 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  65                                                  struct net *net,
  66                                                  struct inet_bind_hashbucket *head,
  67                                                  const unsigned short snum,
  68                                                  int l3mdev)
  69 {
  70         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  71
  72         if (tb) {
  73                 write_pnet(&tb->ib_net, net);
  74                 tb->l3mdev    = l3mdev;
  75                 tb->port      = snum;
  76                 tb->fastreuse = 0;
  77                 tb->fastreuseport = 0;
  78                 INIT_HLIST_HEAD(&tb->owners);
  79                 hlist_add_head(&tb->node, &head->chain);
  80         }
  81         return tb;
  82 }
  83
  84 /*
  85  * Caller must hold hashbucket lock for this tb with local BH disabled
  86  */
  87 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  88 {
  89         if (hlist_empty(&tb->owners)) {
  90                 __hlist_del(&tb->node);
  91                 kmem_cache_free(cachep, tb);
  92         }
  93 }
  94
  95 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  96                     const unsigned short snum)
  97 {
  98         inet_sk(sk)->inet_num = snum;
  99         sk_add_bind_node(sk, &tb->owners);
 100         inet_csk(sk)->icsk_bind_hash = tb;
 101 }
 102
 103 /*
 104  * Get rid of any references to a local port held by the given sock.
 105  */
 106 static void __inet_put_port(struct sock *sk)
 107 {
 108         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 109         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 110                         hashinfo->bhash_size);
 111         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 112         struct inet_bind_bucket *tb;
 113
 114         spin_lock(&head->lock);
 115         tb = inet_csk(sk)->icsk_bind_hash;
 116         __sk_del_bind_node(sk);
 117         inet_csk(sk)->icsk_bind_hash = NULL;
 118         inet_sk(sk)->inet_num = 0;
 119         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 120         spin_unlock(&head->lock);
 121 }
 122
 123 void inet_put_port(struct sock *sk)
 124 {
 125         local_bh_disable();
 126         __inet_put_port(sk);
 127         local_bh_enable();
 128 }
 129 EXPORT_SYMBOL(inet_put_port);
 130
 131 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 132 {
 133         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 134         unsigned short port = inet_sk(child)->inet_num;
 135         const int bhash = inet_bhashfn(sock_net(sk), port,
 136                         table->bhash_size);
 137         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 138         struct inet_bind_bucket *tb;
 139         int l3mdev;
 140
 141         spin_lock(&head->lock);
 142         tb = inet_csk(sk)->icsk_bind_hash;
 143         if (unlikely(!tb)) {
 144                 spin_unlock(&head->lock);
 145                 return -ENOENT;
 146         }
 147         if (tb->port != port) {
 148                 l3mdev = inet_sk_bound_l3mdev(sk);
 149
 150                 /* NOTE: using tproxy and redirecting skbs to a proxy
 151                  * on a different listener port breaks the assumption
 152                  * that the listener socket's icsk_bind_hash is the same
 153                  * as that of the child socket. We have to look up or
 154                  * create a new bind bucket for the child here. */
 155                 inet_bind_bucket_for_each(tb, &head->chain) {
 156                         if (net_eq(ib_net(tb), sock_net(sk)) &&
 157                             tb->l3mdev == l3mdev && tb->port == port)
 158                                 break;
 159                 }
 160                 if (!tb) {
 161                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 162                                                      sock_net(sk), head, port,
 163                                                      l3mdev);
 164                         if (!tb) {
 165                                 spin_unlock(&head->lock);
 166                                 return -ENOMEM;
 167                         }
 168                 }
 169                 inet_csk_update_fastreuse(tb, child);
 170         }
 171         inet_bind_hash(child, tb, port);
 172         spin_unlock(&head->lock);
 173
 174         return 0;
 175 }
 176 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 177
 178 static struct inet_listen_hashbucket *
 179 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 180 {
 181         u32 hash;
 182
 183 #if IS_ENABLED(CONFIG_IPV6)
 184         if (sk->sk_family == AF_INET6)
 185                 hash = ipv6_portaddr_hash(sock_net(sk),
 186                                           &sk->sk_v6_rcv_saddr,
 187                                           inet_sk(sk)->inet_num);
 188         else
 189 #endif
 190                 hash = ipv4_portaddr_hash(sock_net(sk),
 191                                           inet_sk(sk)->inet_rcv_saddr,
 192                                           inet_sk(sk)->inet_num);
 193         return inet_lhash2_bucket(h, hash);
 194 }
 195
 196 static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
 197 {
 198         struct inet_listen_hashbucket *ilb2;
 199
 200         if (!h->lhash2)
 201                 return;
 202
 203         ilb2 = inet_lhash2_bucket_sk(h, sk);
 204
 205         spin_lock(&ilb2->lock);
 206         if (sk->sk_reuseport && sk->sk_family == AF_INET6)
 207                 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 208                                    &ilb2->head);
 209         else
 210                 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
 211                                    &ilb2->head);
 212         ilb2->count++;
 213         spin_unlock(&ilb2->lock);
 214 }
 215
 216 static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
 217 {
 218         struct inet_listen_hashbucket *ilb2;
 219
 220         if (!h->lhash2 ||
 221             WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
 222                 return;
 223
 224         ilb2 = inet_lhash2_bucket_sk(h, sk);
 225
 226         spin_lock(&ilb2->lock);
 227         hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
 228         ilb2->count--;
 229         spin_unlock(&ilb2->lock);
 230 }
 231
 232 static inline int compute_score(struct sock *sk, struct net *net,
 233                                 const unsigned short hnum, const __be32 daddr,
 234                                 const int dif, const int sdif, bool exact_dif)
 235 {
 236         int score = -1;
 237
 238         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 239                         !ipv6_only_sock(sk)) {
 240                 if (sk->sk_rcv_saddr != daddr)
 241                         return -1;
 242
 243                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 244                         return -1;
 245                 score =  sk->sk_bound_dev_if ? 2 : 1;
 246
 247                 if (sk->sk_family == PF_INET)
 248                         score++;
 249                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 250                         score++;
 251         }
 252         return score;
 253 }
 254
 255 /*
 256  * Here are some nice properties to exploit here. The BSD API
 257  * does not allow a listening sock to specify the remote port nor the
 258  * remote address for the connection. So always assume those are both
 259  * wildcarded during the search since they can never be otherwise.
 260  */
 261
 262 /* called with rcu_read_lock() : No refcount taken on the socket */
 263 static struct sock *inet_lhash2_lookup(struct net *net,
 264                                 struct inet_listen_hashbucket *ilb2,
 265                                 struct sk_buff *skb, int doff,
 266                                 const __be32 saddr, __be16 sport,
 267                                 const __be32 daddr, const unsigned short hnum,
 268                                 const int dif, const int sdif)
 269 {
 270         bool exact_dif = inet_exact_dif_match(net, skb);
 271         struct inet_connection_sock *icsk;
 272         struct sock *sk, *result = NULL;
 273         int score, hiscore = 0;
 274         u32 phash = 0;
 275
 276         inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
 277                 sk = (struct sock *)icsk;
 278                 score = compute_score(sk, net, hnum, daddr,
 279                                       dif, sdif, exact_dif);
 280                 if (score > hiscore) {
 281                         if (sk->sk_reuseport) {
 282                                 phash = inet_ehashfn(net, daddr, hnum,
 283                                                      saddr, sport);
 284                                 result = reuseport_select_sock(sk, phash,
 285                                                                skb, doff);
 286                                 if (result)
 287                                         return result;
 288                         }
 289                         result = sk;
 290                         hiscore = score;
 291                 }
 292         }
 293
 294         return result;
 295 }
 296
 297 struct sock *__inet_lookup_listener(struct net *net,
 298                                     struct inet_hashinfo *hashinfo,
 299                                     struct sk_buff *skb, int doff,
 300                                     const __be32 saddr, __be16 sport,
 301                                     const __be32 daddr, const unsigned short hnum,
 302                                     const int dif, const int sdif)
 303 {
 304         struct inet_listen_hashbucket *ilb2;
 305         struct sock *result = NULL;
 306         unsigned int hash2;
 307
 308         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 309         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 310
 311         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 312                                     saddr, sport, daddr, hnum,
 313                                     dif, sdif);
 314         if (result)
 315                 goto done;
 316
 317         /* Lookup lhash2 with INADDR_ANY */
 318         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 319         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 320
 321         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 322                                     saddr, sport, htonl(INADDR_ANY), hnum,
 323                                     dif, sdif);
 324 done:
 325         if (IS_ERR(result))
 326                 return NULL;
 327         return result;
 328 }
 329 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 330
 331 /* All sockets share common refcount, but have different destructors */
 332 void sock_gen_put(struct sock *sk)
 333 {
 334         if (!refcount_dec_and_test(&sk->sk_refcnt))
 335                 return;
 336
 337         if (sk->sk_state == TCP_TIME_WAIT)
 338                 inet_twsk_free(inet_twsk(sk));
 339         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 340                 reqsk_free(inet_reqsk(sk));
 341         else
 342                 sk_free(sk);
 343 }
 344 EXPORT_SYMBOL_GPL(sock_gen_put);
 345
 346 void sock_edemux(struct sk_buff *skb)
 347 {
 348         sock_gen_put(skb->sk);
 349 }
 350 EXPORT_SYMBOL(sock_edemux);
 351
 352 struct sock *__inet_lookup_established(struct net *net,
 353                                   struct inet_hashinfo *hashinfo,
 354                                   const __be32 saddr, const __be16 sport,
 355                                   const __be32 daddr, const u16 hnum,
 356                                   const int dif, const int sdif)
 357 {
 358         INET_ADDR_COOKIE(acookie, saddr, daddr);
 359         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 360         struct sock *sk;
 361         const struct hlist_nulls_node *node;
 362         /* Optimize here for direct hit, only listening connections can
 363          * have wildcards anyways.
 364          */
 365         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 366         unsigned int slot = hash & hashinfo->ehash_mask;
 367         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 368
 369 begin:
 370         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 371                 if (sk->sk_hash != hash)
 372                         continue;
 373                 if (likely(INET_MATCH(sk, net, acookie,
 374                                       saddr, daddr, ports, dif, sdif))) {
 375                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 376                                 goto out;
 377                         if (unlikely(!INET_MATCH(sk, net, acookie,
 378                                                  saddr, daddr, ports,
 379                                                  dif, sdif))) {
 380                                 sock_gen_put(sk);
 381                                 goto begin;
 382                         }
 383                         goto found;
 384                 }
 385         }
 386         /*
 387          * if the nulls value we got at the end of this lookup is
 388          * not the expected one, we must restart lookup.
 389          * We probably met an item that was moved to another chain.
 390          */
 391         if (get_nulls_value(node) != slot)
 392                 goto begin;
 393 out:
 394         sk = NULL;
 395 found:
 396         return sk;
 397 }
 398 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 399
 400 /* called with local bh disabled */
 401 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 402                                     struct sock *sk, __u16 lport,
 403                                     struct inet_timewait_sock **twp)
 404 {
 405         struct inet_hashinfo *hinfo = death_row->hashinfo;
 406         struct inet_sock *inet = inet_sk(sk);
 407         __be32 daddr = inet->inet_rcv_saddr;
 408         __be32 saddr = inet->inet_daddr;
 409         int dif = sk->sk_bound_dev_if;
 410         struct net *net = sock_net(sk);
 411         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 412         INET_ADDR_COOKIE(acookie, saddr, daddr);
 413         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 414         unsigned int hash = inet_ehashfn(net, daddr, lport,
 415                                          saddr, inet->inet_dport);
 416         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 417         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 418         struct sock *sk2;
 419         const struct hlist_nulls_node *node;
 420         struct inet_timewait_sock *tw = NULL;
 421
 422         spin_lock(lock);
 423
 424         sk_nulls_for_each(sk2, node, &head->chain) {
 425                 if (sk2->sk_hash != hash)
 426                         continue;
 427
 428                 if (likely(INET_MATCH(sk2, net, acookie,
 429                                          saddr, daddr, ports, dif, sdif))) {
 430                         if (sk2->sk_state == TCP_TIME_WAIT) {
 431                                 tw = inet_twsk(sk2);
 432                                 if (twsk_unique(sk, sk2, twp))
 433                                         break;
 434                         }
 435                         goto not_unique;
 436                 }
 437         }
 438
 439         /* Must record num and sport now. Otherwise we will see
 440          * in hash table socket with a funny identity.
 441          */
 442         inet->inet_num = lport;
 443         inet->inet_sport = htons(lport);
 444         sk->sk_hash = hash;
 445         WARN_ON(!sk_unhashed(sk));
 446         __sk_nulls_add_node_rcu(sk, &head->chain);
 447         if (tw) {
 448                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 449                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 450         }
 451         spin_unlock(lock);
 452         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 453
 454         if (twp) {
 455                 *twp = tw;
 456         } else if (tw) {
 457                 /* Silly. Should hash-dance instead... */
 458                 inet_twsk_deschedule_put(tw);
 459         }
 460         return 0;
 461
 462 not_unique:
 463         spin_unlock(lock);
 464         return -EADDRNOTAVAIL;
 465 }
 466
 467 static u64 inet_sk_port_offset(const struct sock *sk)
 468 {
 469         const struct inet_sock *inet = inet_sk(sk);
 470
 471         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 472                                           inet->inet_daddr,
 473                                           inet->inet_dport);
 474 }
 475
 476 /* Searches for an exsiting socket in the ehash bucket list.
 477  * Returns true if found, false otherwise.
 478  */
 479 static bool inet_ehash_lookup_by_sk(struct sock *sk,
 480                                     struct hlist_nulls_head *list)
 481 {
 482         const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
 483         const int sdif = sk->sk_bound_dev_if;
 484         const int dif = sk->sk_bound_dev_if;
 485         const struct hlist_nulls_node *node;
 486         struct net *net = sock_net(sk);
 487         struct sock *esk;
 488
 489         INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
 490
 491         sk_nulls_for_each_rcu(esk, node, list) {
 492                 if (esk->sk_hash != sk->sk_hash)
 493                         continue;
 494                 if (sk->sk_family == AF_INET) {
 495                         if (unlikely(INET_MATCH(esk, net, acookie,
 496                                                 sk->sk_daddr,
 497                                                 sk->sk_rcv_saddr,
 498                                                 ports, dif, sdif))) {
 499                                 return true;
 500                         }
 501                 }
 502 #if IS_ENABLED(CONFIG_IPV6)
 503                 else if (sk->sk_family == AF_INET6) {
 504                         if (unlikely(INET6_MATCH(esk, net,
 505                                                  &sk->sk_v6_daddr,
 506                                                  &sk->sk_v6_rcv_saddr,
 507                                                  ports, dif, sdif))) {
 508                                 return true;
 509                         }
 510                 }
 511 #endif
 512         }
 513         return false;
 514 }
 515
 516 /* Insert a socket into ehash, and eventually remove another one
 517  * (The another one can be a SYN_RECV or TIMEWAIT)
 518  * If an existing socket already exists, socket sk is not inserted,
 519  * and sets found_dup_sk parameter to true.
 520  */
 521 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 522 {
 523         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 524         struct hlist_nulls_head *list;
 525         struct inet_ehash_bucket *head;
 526         spinlock_t *lock;
 527         bool ret = true;
 528
 529         WARN_ON_ONCE(!sk_unhashed(sk));
 530
 531         sk->sk_hash = sk_ehashfn(sk);
 532         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 533         list = &head->chain;
 534         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 535
 536         spin_lock(lock);
 537         if (osk) {
 538                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 539                 ret = sk_nulls_del_node_init_rcu(osk);
 540         } else if (found_dup_sk) {
 541                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 542                 if (*found_dup_sk)
 543                         ret = false;
 544         }
 545
 546         if (ret)
 547                 __sk_nulls_add_node_rcu(sk, list);
 548
 549         spin_unlock(lock);
 550
 551         return ret;
 552 }
 553
 554 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 555 {
 556         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 557
 558         if (ok) {
 559                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 560         } else {
 561                 percpu_counter_inc(sk->sk_prot->orphan_count);
 562                 inet_sk_set_state(sk, TCP_CLOSE);
 563                 sock_set_flag(sk, SOCK_DEAD);
 564                 inet_csk_destroy_sock(sk);
 565         }
 566         return ok;
 567 }
 568 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 569
 570 static int inet_reuseport_add_sock(struct sock *sk,
 571                                    struct inet_listen_hashbucket *ilb)
 572 {
 573         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 574         const struct hlist_nulls_node *node;
 575         struct sock *sk2;
 576         kuid_t uid = sock_i_uid(sk);
 577
 578         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 579                 if (sk2 != sk &&
 580                     sk2->sk_family == sk->sk_family &&
 581                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 582                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 583                     inet_csk(sk2)->icsk_bind_hash == tb &&
 584                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 585                     inet_rcv_saddr_equal(sk, sk2, false))
 586                         return reuseport_add_sock(sk, sk2,
 587                                                   inet_rcv_saddr_any(sk));
 588         }
 589
 590         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 591 }
 592
 593 int __inet_hash(struct sock *sk, struct sock *osk)
 594 {
 595         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 596         struct inet_listen_hashbucket *ilb;
 597         int err = 0;
 598
 599         if (sk->sk_state != TCP_LISTEN) {
 600                 inet_ehash_nolisten(sk, osk, NULL);
 601                 return 0;
 602         }
 603         WARN_ON(!sk_unhashed(sk));
 604         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 605
 606         spin_lock(&ilb->lock);
 607         if (sk->sk_reuseport) {
 608                 err = inet_reuseport_add_sock(sk, ilb);
 609                 if (err)
 610                         goto unlock;
 611         }
 612         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 613                 sk->sk_family == AF_INET6)
 614                 __sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
 615         else
 616                 __sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
 617         inet_hash2(hashinfo, sk);
 618         ilb->count++;
 619         sock_set_flag(sk, SOCK_RCU_FREE);
 620         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 621 unlock:
 622         spin_unlock(&ilb->lock);
 623
 624         return err;
 625 }
 626 EXPORT_SYMBOL(__inet_hash);
 627
 628 int inet_hash(struct sock *sk)
 629 {
 630         int err = 0;
 631
 632         if (sk->sk_state != TCP_CLOSE) {
 633                 local_bh_disable();
 634                 err = __inet_hash(sk, NULL);
 635                 local_bh_enable();
 636         }
 637
 638         return err;
 639 }
 640 EXPORT_SYMBOL_GPL(inet_hash);
 641
 642 void inet_unhash(struct sock *sk)
 643 {
 644         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 645         struct inet_listen_hashbucket *ilb = NULL;
 646         spinlock_t *lock;
 647
 648         if (sk_unhashed(sk))
 649                 return;
 650
 651         if (sk->sk_state == TCP_LISTEN) {
 652                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 653                 lock = &ilb->lock;
 654         } else {
 655                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 656         }
 657         spin_lock_bh(lock);
 658         if (sk_unhashed(sk))
 659                 goto unlock;
 660
 661         if (rcu_access_pointer(sk->sk_reuseport_cb))
 662                 reuseport_detach_sock(sk);
 663         if (ilb) {
 664                 inet_unhash2(hashinfo, sk);
 665                 ilb->count--;
 666         }
 667         __sk_nulls_del_node_init_rcu(sk);
 668         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 669 unlock:
 670         spin_unlock_bh(lock);
 671 }
 672 EXPORT_SYMBOL_GPL(inet_unhash);
 673
 674 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 675  * Note that we use 32bit integers (vs RFC 'short integers')
 676  * because 2^16 is not a multiple of num_ephemeral and this
 677  * property might be used by clever attacker.
 678  *
 679  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
 680  * attacks were since demonstrated, thus we use 65536 by default instead
 681  * to really give more isolation and privacy, at the expense of 256kB
 682  * of kernel memory.
 683  */
 684 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
 685 static u32 *table_perturb;
 686
 687 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 688                 struct sock *sk, u64 port_offset,
 689                 int (*check_established)(struct inet_timewait_death_row *,
 690                         struct sock *, __u16, struct inet_timewait_sock **))
 691 {
 692         struct inet_hashinfo *hinfo = death_row->hashinfo;
 693         struct inet_timewait_sock *tw = NULL;
 694         struct inet_bind_hashbucket *head;
 695         int port = inet_sk(sk)->inet_num;
 696         struct net *net = sock_net(sk);
 697         struct inet_bind_bucket *tb;
 698         u32 remaining, offset;
 699         int ret, i, low, high;
 700         int l3mdev;
 701         u32 index;
 702
 703         if (port) {
 704                 local_bh_disable();
 705                 ret = check_established(death_row, sk, port, NULL);
 706                 local_bh_enable();
 707                 return ret;
 708         }
 709
 710         l3mdev = inet_sk_bound_l3mdev(sk);
 711
 712         inet_get_local_port_range(net, &low, &high);
 713         high++; /* [32768, 60999] -> [32768, 61000[ */
 714         remaining = high - low;
 715         if (likely(remaining > 1))
 716                 remaining &= ~1U;
 717
 718         get_random_slow_once(table_perturb,
 719                              INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
 720         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
 721
 722         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
 723         offset %= remaining;
 724
 725         /* In first pass we try ports of @low parity.
 726          * inet_csk_get_port() does the opposite choice.
 727          */
 728         offset &= ~1U;
 729 other_parity_scan:
 730         port = low + offset;
 731         for (i = 0; i < remaining; i += 2, port += 2) {
 732                 if (unlikely(port >= high))
 733                         port -= remaining;
 734                 if (inet_is_local_reserved_port(net, port))
 735                         continue;
 736                 head = &hinfo->bhash[inet_bhashfn(net, port,
 737                                                   hinfo->bhash_size)];
 738                 spin_lock_bh(&head->lock);
 739
 740                 /* Does not bother with rcv_saddr checks, because
 741                  * the established check is already unique enough.
 742                  */
 743                 inet_bind_bucket_for_each(tb, &head->chain) {
 744                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 745                             tb->port == port) {
 746                                 if (tb->fastreuse >= 0 ||
 747                                     tb->fastreuseport >= 0)
 748                                         goto next_port;
 749                                 WARN_ON(hlist_empty(&tb->owners));
 750                                 if (!check_established(death_row, sk,
 751                                                        port, &tw))
 752                                         goto ok;
 753                                 goto next_port;
 754                         }
 755                 }
 756
 757                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 758                                              net, head, port, l3mdev);
 759                 if (!tb) {
 760                         spin_unlock_bh(&head->lock);
 761                         return -ENOMEM;
 762                 }
 763                 tb->fastreuse = -1;
 764                 tb->fastreuseport = -1;
 765                 goto ok;
 766 next_port:
 767                 spin_unlock_bh(&head->lock);
 768                 cond_resched();
 769         }
 770
 771         offset++;
 772         if ((offset & 1) && remaining > 1)
 773                 goto other_parity_scan;
 774
 775         return -EADDRNOTAVAIL;
 776
 777 ok:
 778         /* Here we want to add a little bit of randomness to the next source
 779          * port that will be chosen. We use a max() with a random here so that
 780          * on low contention the randomness is maximal and on high contention
 781          * it may be inexistent.
 782          */
 783         i = max_t(int, i, (prandom_u32() & 7) * 2);
 784         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
 785
 786         /* Head lock still held and bh's disabled */
 787         inet_bind_hash(sk, tb, port);
 788         if (sk_unhashed(sk)) {
 789                 inet_sk(sk)->inet_sport = htons(port);
 790                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
 791         }
 792         if (tw)
 793                 inet_twsk_bind_unhash(tw, hinfo);
 794         spin_unlock(&head->lock);
 795         if (tw)
 796                 inet_twsk_deschedule_put(tw);
 797         local_bh_enable();
 798         return 0;
 799 }
 800
 801 /*
 802  * Bind a port for a connect operation and hash it.
 803  */
 804 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 805                       struct sock *sk)
 806 {
 807         u64 port_offset = 0;
 808
 809         if (!inet_sk(sk)->inet_num)
 810                 port_offset = inet_sk_port_offset(sk);
 811         return __inet_hash_connect(death_row, sk, port_offset,
 812                                    __inet_check_established);
 813 }
 814 EXPORT_SYMBOL_GPL(inet_hash_connect);
 815
 816 void inet_hashinfo_init(struct inet_hashinfo *h)
 817 {
 818         int i;
 819
 820         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 821                 spin_lock_init(&h->listening_hash[i].lock);
 822                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
 823                                       i + LISTENING_NULLS_BASE);
 824                 h->listening_hash[i].count = 0;
 825         }
 826
 827         h->lhash2 = NULL;
 828 }
 829 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 830
 831 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
 832 {
 833         int i;
 834
 835         for (i = 0; i <= h->lhash2_mask; i++) {
 836                 spin_lock_init(&h->lhash2[i].lock);
 837                 INIT_HLIST_HEAD(&h->lhash2[i].head);
 838                 h->lhash2[i].count = 0;
 839         }
 840 }
 841
 842 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
 843                                 unsigned long numentries, int scale,
 844                                 unsigned long low_limit,
 845                                 unsigned long high_limit)
 846 {
 847         h->lhash2 = alloc_large_system_hash(name,
 848                                             sizeof(*h->lhash2),
 849                                             numentries,
 850                                             scale,
 851                                             0,
 852                                             NULL,
 853                                             &h->lhash2_mask,
 854                                             low_limit,
 855                                             high_limit);
 856         init_hashinfo_lhash2(h);
 857
 858         /* this one is used for source ports of outgoing connections */
 859         table_perturb = kmalloc_array(INET_TABLE_PERTURB_SIZE,
 860                                       sizeof(*table_perturb), GFP_KERNEL);
 861         if (!table_perturb)
 862                 panic("TCP: failed to alloc table_perturb");
 863 }
 864
 865 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
 866 {
 867         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
 868         if (!h->lhash2)
 869                 return -ENOMEM;
 870
 871         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
 872         /* INET_LHTABLE_SIZE must be a power of 2 */
 873         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
 874
 875         init_hashinfo_lhash2(h);
 876         return 0;
 877 }
 878 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
 879
 880 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 881 {
 882         unsigned int locksz = sizeof(spinlock_t);
 883         unsigned int i, nblocks = 1;
 884
 885         if (locksz != 0) {
 886                 /* allocate 2 cache lines or at least one spinlock per cpu */
 887                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 888                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 889
 890                 /* no more locks than number of hash buckets */
 891                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 892
 893                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
 894                 if (!hashinfo->ehash_locks)
 895                         return -ENOMEM;
 896
 897                 for (i = 0; i < nblocks; i++)
 898                         spin_lock_init(&hashinfo->ehash_locks[i]);
 899         }
 900         hashinfo->ehash_locks_mask = nblocks - 1;
 901         return 0;
 902 }
 903 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);