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 inline int compute_score(struct sock *sk, struct net *net,
 197                                 const unsigned short hnum, const __be32 daddr,
 198                                 const int dif, const int sdif)
 199 {
 200         int score = -1;
 201
 202         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 203                         !ipv6_only_sock(sk)) {
 204                 if (sk->sk_rcv_saddr != daddr)
 205                         return -1;
 206
 207                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 208                         return -1;
 209                 score =  sk->sk_bound_dev_if ? 2 : 1;
 210
 211                 if (sk->sk_family == PF_INET)
 212                         score++;
 213                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 214                         score++;
 215         }
 216         return score;
 217 }
 218
 219 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
 220                                             struct sk_buff *skb, int doff,
 221                                             __be32 saddr, __be16 sport,
 222                                             __be32 daddr, unsigned short hnum)
 223 {
 224         struct sock *reuse_sk = NULL;
 225         u32 phash;
 226
 227         if (sk->sk_reuseport) {
 228                 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 229                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 230         }
 231         return reuse_sk;
 232 }
 233
 234 /*
 235  * Here are some nice properties to exploit here. The BSD API
 236  * does not allow a listening sock to specify the remote port nor the
 237  * remote address for the connection. So always assume those are both
 238  * wildcarded during the search since they can never be otherwise.
 239  */
 240
 241 /* called with rcu_read_lock() : No refcount taken on the socket */
 242 static struct sock *inet_lhash2_lookup(struct net *net,
 243                                 struct inet_listen_hashbucket *ilb2,
 244                                 struct sk_buff *skb, int doff,
 245                                 const __be32 saddr, __be16 sport,
 246                                 const __be32 daddr, const unsigned short hnum,
 247                                 const int dif, const int sdif)
 248 {
 249         struct sock *sk, *result = NULL;
 250         struct hlist_nulls_node *node;
 251         int score, hiscore = 0;
 252
 253         sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
 254                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 255                 if (score > hiscore) {
 256                         result = lookup_reuseport(net, sk, skb, doff,
 257                                                   saddr, sport, daddr, hnum);
 258                         if (result)
 259                                 return result;
 260
 261                         result = sk;
 262                         hiscore = score;
 263                 }
 264         }
 265
 266         return result;
 267 }
 268
 269 static inline struct sock *inet_lookup_run_bpf(struct net *net,
 270                                                struct inet_hashinfo *hashinfo,
 271                                                struct sk_buff *skb, int doff,
 272                                                __be32 saddr, __be16 sport,
 273                                                __be32 daddr, u16 hnum, const int dif)
 274 {
 275         struct sock *sk, *reuse_sk;
 276         bool no_reuseport;
 277
 278         if (hashinfo != &tcp_hashinfo)
 279                 return NULL; /* only TCP is supported */
 280
 281         no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP, saddr, sport,
 282                                             daddr, hnum, dif, &sk);
 283         if (no_reuseport || IS_ERR_OR_NULL(sk))
 284                 return sk;
 285
 286         reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
 287         if (reuse_sk)
 288                 sk = reuse_sk;
 289         return sk;
 290 }
 291
 292 struct sock *__inet_lookup_listener(struct net *net,
 293                                     struct inet_hashinfo *hashinfo,
 294                                     struct sk_buff *skb, int doff,
 295                                     const __be32 saddr, __be16 sport,
 296                                     const __be32 daddr, const unsigned short hnum,
 297                                     const int dif, const int sdif)
 298 {
 299         struct inet_listen_hashbucket *ilb2;
 300         struct sock *result = NULL;
 301         unsigned int hash2;
 302
 303         /* Lookup redirect from BPF */
 304         if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
 305                 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
 306                                              saddr, sport, daddr, hnum, dif);
 307                 if (result)
 308                         goto done;
 309         }
 310
 311         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 312         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 313
 314         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 315                                     saddr, sport, daddr, hnum,
 316                                     dif, sdif);
 317         if (result)
 318                 goto done;
 319
 320         /* Lookup lhash2 with INADDR_ANY */
 321         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 322         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 323
 324         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 325                                     saddr, sport, htonl(INADDR_ANY), hnum,
 326                                     dif, sdif);
 327 done:
 328         if (IS_ERR(result))
 329                 return NULL;
 330         return result;
 331 }
 332 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 333
 334 /* All sockets share common refcount, but have different destructors */
 335 void sock_gen_put(struct sock *sk)
 336 {
 337         if (!refcount_dec_and_test(&sk->sk_refcnt))
 338                 return;
 339
 340         if (sk->sk_state == TCP_TIME_WAIT)
 341                 inet_twsk_free(inet_twsk(sk));
 342         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 343                 reqsk_free(inet_reqsk(sk));
 344         else
 345                 sk_free(sk);
 346 }
 347 EXPORT_SYMBOL_GPL(sock_gen_put);
 348
 349 void sock_edemux(struct sk_buff *skb)
 350 {
 351         sock_gen_put(skb->sk);
 352 }
 353 EXPORT_SYMBOL(sock_edemux);
 354
 355 struct sock *__inet_lookup_established(struct net *net,
 356                                   struct inet_hashinfo *hashinfo,
 357                                   const __be32 saddr, const __be16 sport,
 358                                   const __be32 daddr, const u16 hnum,
 359                                   const int dif, const int sdif)
 360 {
 361         INET_ADDR_COOKIE(acookie, saddr, daddr);
 362         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 363         struct sock *sk;
 364         const struct hlist_nulls_node *node;
 365         /* Optimize here for direct hit, only listening connections can
 366          * have wildcards anyways.
 367          */
 368         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 369         unsigned int slot = hash & hashinfo->ehash_mask;
 370         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 371
 372 begin:
 373         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 374                 if (sk->sk_hash != hash)
 375                         continue;
 376                 if (likely(INET_MATCH(net, sk, acookie, ports, dif, sdif))) {
 377                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 378                                 goto out;
 379                         if (unlikely(!INET_MATCH(net, sk, acookie,
 380                                                  ports, dif, sdif))) {
 381                                 sock_gen_put(sk);
 382                                 goto begin;
 383                         }
 384                         goto found;
 385                 }
 386         }
 387         /*
 388          * if the nulls value we got at the end of this lookup is
 389          * not the expected one, we must restart lookup.
 390          * We probably met an item that was moved to another chain.
 391          */
 392         if (get_nulls_value(node) != slot)
 393                 goto begin;
 394 out:
 395         sk = NULL;
 396 found:
 397         return sk;
 398 }
 399 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 400
 401 /* called with local bh disabled */
 402 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 403                                     struct sock *sk, __u16 lport,
 404                                     struct inet_timewait_sock **twp)
 405 {
 406         struct inet_hashinfo *hinfo = death_row->hashinfo;
 407         struct inet_sock *inet = inet_sk(sk);
 408         __be32 daddr = inet->inet_rcv_saddr;
 409         __be32 saddr = inet->inet_daddr;
 410         int dif = sk->sk_bound_dev_if;
 411         struct net *net = sock_net(sk);
 412         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 413         INET_ADDR_COOKIE(acookie, saddr, daddr);
 414         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 415         unsigned int hash = inet_ehashfn(net, daddr, lport,
 416                                          saddr, inet->inet_dport);
 417         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 418         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 419         struct sock *sk2;
 420         const struct hlist_nulls_node *node;
 421         struct inet_timewait_sock *tw = NULL;
 422
 423         spin_lock(lock);
 424
 425         sk_nulls_for_each(sk2, node, &head->chain) {
 426                 if (sk2->sk_hash != hash)
 427                         continue;
 428
 429                 if (likely(INET_MATCH(net, sk2, acookie, 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(net, esk, acookie,
 496                                                 ports, dif, sdif))) {
 497                                 return true;
 498                         }
 499                 }
 500 #if IS_ENABLED(CONFIG_IPV6)
 501                 else if (sk->sk_family == AF_INET6) {
 502                         if (unlikely(inet6_match(net, esk,
 503                                                  &sk->sk_v6_daddr,
 504                                                  &sk->sk_v6_rcv_saddr,
 505                                                  ports, dif, sdif))) {
 506                                 return true;
 507                         }
 508                 }
 509 #endif
 510         }
 511         return false;
 512 }
 513
 514 /* Insert a socket into ehash, and eventually remove another one
 515  * (The another one can be a SYN_RECV or TIMEWAIT)
 516  * If an existing socket already exists, socket sk is not inserted,
 517  * and sets found_dup_sk parameter to true.
 518  */
 519 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 520 {
 521         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 522         struct hlist_nulls_head *list;
 523         struct inet_ehash_bucket *head;
 524         spinlock_t *lock;
 525         bool ret = true;
 526
 527         WARN_ON_ONCE(!sk_unhashed(sk));
 528
 529         sk->sk_hash = sk_ehashfn(sk);
 530         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 531         list = &head->chain;
 532         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 533
 534         spin_lock(lock);
 535         if (osk) {
 536                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 537                 ret = sk_nulls_del_node_init_rcu(osk);
 538         } else if (found_dup_sk) {
 539                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 540                 if (*found_dup_sk)
 541                         ret = false;
 542         }
 543
 544         if (ret)
 545                 __sk_nulls_add_node_rcu(sk, list);
 546
 547         spin_unlock(lock);
 548
 549         return ret;
 550 }
 551
 552 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 553 {
 554         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 555
 556         if (ok) {
 557                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 558         } else {
 559                 this_cpu_inc(*sk->sk_prot->orphan_count);
 560                 inet_sk_set_state(sk, TCP_CLOSE);
 561                 sock_set_flag(sk, SOCK_DEAD);
 562                 inet_csk_destroy_sock(sk);
 563         }
 564         return ok;
 565 }
 566 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 567
 568 static int inet_reuseport_add_sock(struct sock *sk,
 569                                    struct inet_listen_hashbucket *ilb)
 570 {
 571         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 572         const struct hlist_nulls_node *node;
 573         struct sock *sk2;
 574         kuid_t uid = sock_i_uid(sk);
 575
 576         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 577                 if (sk2 != sk &&
 578                     sk2->sk_family == sk->sk_family &&
 579                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 580                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 581                     inet_csk(sk2)->icsk_bind_hash == tb &&
 582                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 583                     inet_rcv_saddr_equal(sk, sk2, false))
 584                         return reuseport_add_sock(sk, sk2,
 585                                                   inet_rcv_saddr_any(sk));
 586         }
 587
 588         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 589 }
 590
 591 int __inet_hash(struct sock *sk, struct sock *osk)
 592 {
 593         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 594         struct inet_listen_hashbucket *ilb2;
 595         int err = 0;
 596
 597         if (sk->sk_state != TCP_LISTEN) {
 598                 local_bh_disable();
 599                 inet_ehash_nolisten(sk, osk, NULL);
 600                 local_bh_enable();
 601                 return 0;
 602         }
 603         WARN_ON(!sk_unhashed(sk));
 604         ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 605
 606         spin_lock(&ilb2->lock);
 607         if (sk->sk_reuseport) {
 608                 err = inet_reuseport_add_sock(sk, ilb2);
 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, &ilb2->nulls_head);
 615         else
 616                 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
 617         sock_set_flag(sk, SOCK_RCU_FREE);
 618         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 619 unlock:
 620         spin_unlock(&ilb2->lock);
 621
 622         return err;
 623 }
 624 EXPORT_SYMBOL(__inet_hash);
 625
 626 int inet_hash(struct sock *sk)
 627 {
 628         int err = 0;
 629
 630         if (sk->sk_state != TCP_CLOSE)
 631                 err = __inet_hash(sk, NULL);
 632
 633         return err;
 634 }
 635 EXPORT_SYMBOL_GPL(inet_hash);
 636
 637 void inet_unhash(struct sock *sk)
 638 {
 639         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 640
 641         if (sk_unhashed(sk))
 642                 return;
 643
 644         if (sk->sk_state == TCP_LISTEN) {
 645                 struct inet_listen_hashbucket *ilb2;
 646
 647                 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 648                 /* Don't disable bottom halves while acquiring the lock to
 649                  * avoid circular locking dependency on PREEMPT_RT.
 650                  */
 651                 spin_lock(&ilb2->lock);
 652                 if (sk_unhashed(sk)) {
 653                         spin_unlock(&ilb2->lock);
 654                         return;
 655                 }
 656
 657                 if (rcu_access_pointer(sk->sk_reuseport_cb))
 658                         reuseport_stop_listen_sock(sk);
 659
 660                 __sk_nulls_del_node_init_rcu(sk);
 661                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 662                 spin_unlock(&ilb2->lock);
 663         } else {
 664                 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 665
 666                 spin_lock_bh(lock);
 667                 if (sk_unhashed(sk)) {
 668                         spin_unlock_bh(lock);
 669                         return;
 670                 }
 671                 __sk_nulls_del_node_init_rcu(sk);
 672                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 673                 spin_unlock_bh(lock);
 674         }
 675 }
 676 EXPORT_SYMBOL_GPL(inet_unhash);
 677
 678 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 679  * Note that we use 32bit integers (vs RFC 'short integers')
 680  * because 2^16 is not a multiple of num_ephemeral and this
 681  * property might be used by clever attacker.
 682  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
 683  * attacks were since demonstrated, thus we use 65536 instead to really
 684  * give more isolation and privacy, at the expense of 256kB of kernel
 685  * memory.
 686  */
 687 #define INET_TABLE_PERTURB_SHIFT 16
 688 #define INET_TABLE_PERTURB_SIZE (1 << INET_TABLE_PERTURB_SHIFT)
 689 static u32 *table_perturb;
 690
 691 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 692                 struct sock *sk, u64 port_offset,
 693                 int (*check_established)(struct inet_timewait_death_row *,
 694                         struct sock *, __u16, struct inet_timewait_sock **))
 695 {
 696         struct inet_hashinfo *hinfo = death_row->hashinfo;
 697         struct inet_timewait_sock *tw = NULL;
 698         struct inet_bind_hashbucket *head;
 699         int port = inet_sk(sk)->inet_num;
 700         struct net *net = sock_net(sk);
 701         struct inet_bind_bucket *tb;
 702         u32 remaining, offset;
 703         int ret, i, low, high;
 704         int l3mdev;
 705         u32 index;
 706
 707         if (port) {
 708                 head = &hinfo->bhash[inet_bhashfn(net, port,
 709                                                   hinfo->bhash_size)];
 710                 tb = inet_csk(sk)->icsk_bind_hash;
 711                 spin_lock_bh(&head->lock);
 712                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 713                         inet_ehash_nolisten(sk, NULL, NULL);
 714                         spin_unlock_bh(&head->lock);
 715                         return 0;
 716                 }
 717                 spin_unlock(&head->lock);
 718                 /* No definite answer... Walk to established hash table */
 719                 ret = check_established(death_row, sk, port, NULL);
 720                 local_bh_enable();
 721                 return ret;
 722         }
 723
 724         l3mdev = inet_sk_bound_l3mdev(sk);
 725
 726         inet_get_local_port_range(net, &low, &high);
 727         high++; /* [32768, 60999] -> [32768, 61000[ */
 728         remaining = high - low;
 729         if (likely(remaining > 1))
 730                 remaining &= ~1U;
 731
 732         net_get_random_once(table_perturb,
 733                             INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
 734         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
 735
 736         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
 737         offset %= remaining;
 738
 739         /* In first pass we try ports of @low parity.
 740          * inet_csk_get_port() does the opposite choice.
 741          */
 742         offset &= ~1U;
 743 other_parity_scan:
 744         port = low + offset;
 745         for (i = 0; i < remaining; i += 2, port += 2) {
 746                 if (unlikely(port >= high))
 747                         port -= remaining;
 748                 if (inet_is_local_reserved_port(net, port))
 749                         continue;
 750                 head = &hinfo->bhash[inet_bhashfn(net, port,
 751                                                   hinfo->bhash_size)];
 752                 spin_lock_bh(&head->lock);
 753
 754                 /* Does not bother with rcv_saddr checks, because
 755                  * the established check is already unique enough.
 756                  */
 757                 inet_bind_bucket_for_each(tb, &head->chain) {
 758                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 759                             tb->port == port) {
 760                                 if (tb->fastreuse >= 0 ||
 761                                     tb->fastreuseport >= 0)
 762                                         goto next_port;
 763                                 WARN_ON(hlist_empty(&tb->owners));
 764                                 if (!check_established(death_row, sk,
 765                                                        port, &tw))
 766                                         goto ok;
 767                                 goto next_port;
 768                         }
 769                 }
 770
 771                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 772                                              net, head, port, l3mdev);
 773                 if (!tb) {
 774                         spin_unlock_bh(&head->lock);
 775                         return -ENOMEM;
 776                 }
 777                 tb->fastreuse = -1;
 778                 tb->fastreuseport = -1;
 779                 goto ok;
 780 next_port:
 781                 spin_unlock_bh(&head->lock);
 782                 cond_resched();
 783         }
 784
 785         offset++;
 786         if ((offset & 1) && remaining > 1)
 787                 goto other_parity_scan;
 788
 789         return -EADDRNOTAVAIL;
 790
 791 ok:
 792         /* Here we want to add a little bit of randomness to the next source
 793          * port that will be chosen. We use a max() with a random here so that
 794          * on low contention the randomness is maximal and on high contention
 795          * it may be inexistent.
 796          */
 797         i = max_t(int, i, (prandom_u32() & 7) * 2);
 798         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
 799
 800         /* Head lock still held and bh's disabled */
 801         inet_bind_hash(sk, tb, port);
 802         if (sk_unhashed(sk)) {
 803                 inet_sk(sk)->inet_sport = htons(port);
 804                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
 805         }
 806         if (tw)
 807                 inet_twsk_bind_unhash(tw, hinfo);
 808         spin_unlock(&head->lock);
 809         if (tw)
 810                 inet_twsk_deschedule_put(tw);
 811         local_bh_enable();
 812         return 0;
 813 }
 814
 815 /*
 816  * Bind a port for a connect operation and hash it.
 817  */
 818 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 819                       struct sock *sk)
 820 {
 821         u64 port_offset = 0;
 822
 823         if (!inet_sk(sk)->inet_num)
 824                 port_offset = inet_sk_port_offset(sk);
 825         return __inet_hash_connect(death_row, sk, port_offset,
 826                                    __inet_check_established);
 827 }
 828 EXPORT_SYMBOL_GPL(inet_hash_connect);
 829
 830 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
 831 {
 832         int i;
 833
 834         for (i = 0; i <= h->lhash2_mask; i++) {
 835                 spin_lock_init(&h->lhash2[i].lock);
 836                 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
 837                                       i + LISTENING_NULLS_BASE);
 838         }
 839 }
 840
 841 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
 842                                 unsigned long numentries, int scale,
 843                                 unsigned long low_limit,
 844                                 unsigned long high_limit)
 845 {
 846         h->lhash2 = alloc_large_system_hash(name,
 847                                             sizeof(*h->lhash2),
 848                                             numentries,
 849                                             scale,
 850                                             0,
 851                                             NULL,
 852                                             &h->lhash2_mask,
 853                                             low_limit,
 854                                             high_limit);
 855         init_hashinfo_lhash2(h);
 856
 857         /* this one is used for source ports of outgoing connections */
 858         table_perturb = kmalloc_array(INET_TABLE_PERTURB_SIZE,
 859                                       sizeof(*table_perturb), GFP_KERNEL);
 860         if (!table_perturb)
 861                 panic("TCP: failed to alloc table_perturb");
 862 }
 863
 864 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
 865 {
 866         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
 867         if (!h->lhash2)
 868                 return -ENOMEM;
 869
 870         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
 871         /* INET_LHTABLE_SIZE must be a power of 2 */
 872         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
 873
 874         init_hashinfo_lhash2(h);
 875         return 0;
 876 }
 877 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
 878
 879 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 880 {
 881         unsigned int locksz = sizeof(spinlock_t);
 882         unsigned int i, nblocks = 1;
 883
 884         if (locksz != 0) {
 885                 /* allocate 2 cache lines or at least one spinlock per cpu */
 886                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 887                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 888
 889                 /* no more locks than number of hash buckets */
 890                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 891
 892                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
 893                 if (!hashinfo->ehash_locks)
 894                         return -ENOMEM;
 895
 896                 for (i = 0; i < nblocks; i++)
 897                         spin_lock_init(&hashinfo->ehash_locks[i]);
 898         }
 899         hashinfo->ehash_locks_mask = nblocks - 1;
 900         return 0;
 901 }
 902 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);