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)
 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 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
 256                                             struct sk_buff *skb, int doff,
 257                                             __be32 saddr, __be16 sport,
 258                                             __be32 daddr, unsigned short hnum)
 259 {
 260         struct sock *reuse_sk = NULL;
 261         u32 phash;
 262
 263         if (sk->sk_reuseport) {
 264                 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 265                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 266         }
 267         return reuse_sk;
 268 }
 269
 270 /*
 271  * Here are some nice properties to exploit here. The BSD API
 272  * does not allow a listening sock to specify the remote port nor the
 273  * remote address for the connection. So always assume those are both
 274  * wildcarded during the search since they can never be otherwise.
 275  */
 276
 277 /* called with rcu_read_lock() : No refcount taken on the socket */
 278 static struct sock *inet_lhash2_lookup(struct net *net,
 279                                 struct inet_listen_hashbucket *ilb2,
 280                                 struct sk_buff *skb, int doff,
 281                                 const __be32 saddr, __be16 sport,
 282                                 const __be32 daddr, const unsigned short hnum,
 283                                 const int dif, const int sdif)
 284 {
 285         struct inet_connection_sock *icsk;
 286         struct sock *sk, *result = NULL;
 287         int score, hiscore = 0;
 288
 289         inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
 290                 sk = (struct sock *)icsk;
 291                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 292                 if (score > hiscore) {
 293                         result = lookup_reuseport(net, sk, skb, doff,
 294                                                   saddr, sport, daddr, hnum);
 295                         if (result)
 296                                 return result;
 297
 298                         result = sk;
 299                         hiscore = score;
 300                 }
 301         }
 302
 303         return result;
 304 }
 305
 306 static inline struct sock *inet_lookup_run_bpf(struct net *net,
 307                                                struct inet_hashinfo *hashinfo,
 308                                                struct sk_buff *skb, int doff,
 309                                                __be32 saddr, __be16 sport,
 310                                                __be32 daddr, u16 hnum)
 311 {
 312         struct sock *sk, *reuse_sk;
 313         bool no_reuseport;
 314
 315         if (hashinfo != &tcp_hashinfo)
 316                 return NULL; /* only TCP is supported */
 317
 318         no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP,
 319                                             saddr, sport, daddr, hnum, &sk);
 320         if (no_reuseport || IS_ERR_OR_NULL(sk))
 321                 return sk;
 322
 323         reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
 324         if (reuse_sk)
 325                 sk = reuse_sk;
 326         return sk;
 327 }
 328
 329 struct sock *__inet_lookup_listener(struct net *net,
 330                                     struct inet_hashinfo *hashinfo,
 331                                     struct sk_buff *skb, int doff,
 332                                     const __be32 saddr, __be16 sport,
 333                                     const __be32 daddr, const unsigned short hnum,
 334                                     const int dif, const int sdif)
 335 {
 336         struct inet_listen_hashbucket *ilb2;
 337         struct sock *result = NULL;
 338         unsigned int hash2;
 339
 340         /* Lookup redirect from BPF */
 341         if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
 342                 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
 343                                              saddr, sport, daddr, hnum);
 344                 if (result)
 345                         goto done;
 346         }
 347
 348         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 349         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 350
 351         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 352                                     saddr, sport, daddr, hnum,
 353                                     dif, sdif);
 354         if (result)
 355                 goto done;
 356
 357         /* Lookup lhash2 with INADDR_ANY */
 358         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 359         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 360
 361         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 362                                     saddr, sport, htonl(INADDR_ANY), hnum,
 363                                     dif, sdif);
 364 done:
 365         if (IS_ERR(result))
 366                 return NULL;
 367         return result;
 368 }
 369 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 370
 371 /* All sockets share common refcount, but have different destructors */
 372 void sock_gen_put(struct sock *sk)
 373 {
 374         if (!refcount_dec_and_test(&sk->sk_refcnt))
 375                 return;
 376
 377         if (sk->sk_state == TCP_TIME_WAIT)
 378                 inet_twsk_free(inet_twsk(sk));
 379         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 380                 reqsk_free(inet_reqsk(sk));
 381         else
 382                 sk_free(sk);
 383 }
 384 EXPORT_SYMBOL_GPL(sock_gen_put);
 385
 386 void sock_edemux(struct sk_buff *skb)
 387 {
 388         sock_gen_put(skb->sk);
 389 }
 390 EXPORT_SYMBOL(sock_edemux);
 391
 392 struct sock *__inet_lookup_established(struct net *net,
 393                                   struct inet_hashinfo *hashinfo,
 394                                   const __be32 saddr, const __be16 sport,
 395                                   const __be32 daddr, const u16 hnum,
 396                                   const int dif, const int sdif)
 397 {
 398         INET_ADDR_COOKIE(acookie, saddr, daddr);
 399         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 400         struct sock *sk;
 401         const struct hlist_nulls_node *node;
 402         /* Optimize here for direct hit, only listening connections can
 403          * have wildcards anyways.
 404          */
 405         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 406         unsigned int slot = hash & hashinfo->ehash_mask;
 407         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 408
 409 begin:
 410         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 411                 if (sk->sk_hash != hash)
 412                         continue;
 413                 if (likely(INET_MATCH(net, sk, acookie, ports, dif, sdif))) {
 414                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 415                                 goto out;
 416                         if (unlikely(!INET_MATCH(net, sk, acookie,
 417                                                  ports, dif, sdif))) {
 418                                 sock_gen_put(sk);
 419                                 goto begin;
 420                         }
 421                         goto found;
 422                 }
 423         }
 424         /*
 425          * if the nulls value we got at the end of this lookup is
 426          * not the expected one, we must restart lookup.
 427          * We probably met an item that was moved to another chain.
 428          */
 429         if (get_nulls_value(node) != slot)
 430                 goto begin;
 431 out:
 432         sk = NULL;
 433 found:
 434         return sk;
 435 }
 436 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 437
 438 /* called with local bh disabled */
 439 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 440                                     struct sock *sk, __u16 lport,
 441                                     struct inet_timewait_sock **twp)
 442 {
 443         struct inet_hashinfo *hinfo = death_row->hashinfo;
 444         struct inet_sock *inet = inet_sk(sk);
 445         __be32 daddr = inet->inet_rcv_saddr;
 446         __be32 saddr = inet->inet_daddr;
 447         int dif = sk->sk_bound_dev_if;
 448         struct net *net = sock_net(sk);
 449         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 450         INET_ADDR_COOKIE(acookie, saddr, daddr);
 451         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 452         unsigned int hash = inet_ehashfn(net, daddr, lport,
 453                                          saddr, inet->inet_dport);
 454         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 455         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 456         struct sock *sk2;
 457         const struct hlist_nulls_node *node;
 458         struct inet_timewait_sock *tw = NULL;
 459
 460         spin_lock(lock);
 461
 462         sk_nulls_for_each(sk2, node, &head->chain) {
 463                 if (sk2->sk_hash != hash)
 464                         continue;
 465
 466                 if (likely(INET_MATCH(net, sk2, acookie, ports, dif, sdif))) {
 467                         if (sk2->sk_state == TCP_TIME_WAIT) {
 468                                 tw = inet_twsk(sk2);
 469                                 if (twsk_unique(sk, sk2, twp))
 470                                         break;
 471                         }
 472                         goto not_unique;
 473                 }
 474         }
 475
 476         /* Must record num and sport now. Otherwise we will see
 477          * in hash table socket with a funny identity.
 478          */
 479         inet->inet_num = lport;
 480         inet->inet_sport = htons(lport);
 481         sk->sk_hash = hash;
 482         WARN_ON(!sk_unhashed(sk));
 483         __sk_nulls_add_node_rcu(sk, &head->chain);
 484         if (tw) {
 485                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 486                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 487         }
 488         spin_unlock(lock);
 489         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 490
 491         if (twp) {
 492                 *twp = tw;
 493         } else if (tw) {
 494                 /* Silly. Should hash-dance instead... */
 495                 inet_twsk_deschedule_put(tw);
 496         }
 497         return 0;
 498
 499 not_unique:
 500         spin_unlock(lock);
 501         return -EADDRNOTAVAIL;
 502 }
 503
 504 static u64 inet_sk_port_offset(const struct sock *sk)
 505 {
 506         const struct inet_sock *inet = inet_sk(sk);
 507
 508         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 509                                           inet->inet_daddr,
 510                                           inet->inet_dport);
 511 }
 512
 513 /* Searches for an exsiting socket in the ehash bucket list.
 514  * Returns true if found, false otherwise.
 515  */
 516 static bool inet_ehash_lookup_by_sk(struct sock *sk,
 517                                     struct hlist_nulls_head *list)
 518 {
 519         const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
 520         const int sdif = sk->sk_bound_dev_if;
 521         const int dif = sk->sk_bound_dev_if;
 522         const struct hlist_nulls_node *node;
 523         struct net *net = sock_net(sk);
 524         struct sock *esk;
 525
 526         INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
 527
 528         sk_nulls_for_each_rcu(esk, node, list) {
 529                 if (esk->sk_hash != sk->sk_hash)
 530                         continue;
 531                 if (sk->sk_family == AF_INET) {
 532                         if (unlikely(INET_MATCH(net, esk, acookie,
 533                                                 ports, dif, sdif))) {
 534                                 return true;
 535                         }
 536                 }
 537 #if IS_ENABLED(CONFIG_IPV6)
 538                 else if (sk->sk_family == AF_INET6) {
 539                         if (unlikely(inet6_match(net, esk,
 540                                                  &sk->sk_v6_daddr,
 541                                                  &sk->sk_v6_rcv_saddr,
 542                                                  ports, dif, sdif))) {
 543                                 return true;
 544                         }
 545                 }
 546 #endif
 547         }
 548         return false;
 549 }
 550
 551 /* Insert a socket into ehash, and eventually remove another one
 552  * (The another one can be a SYN_RECV or TIMEWAIT)
 553  * If an existing socket already exists, socket sk is not inserted,
 554  * and sets found_dup_sk parameter to true.
 555  */
 556 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 557 {
 558         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 559         struct hlist_nulls_head *list;
 560         struct inet_ehash_bucket *head;
 561         spinlock_t *lock;
 562         bool ret = true;
 563
 564         WARN_ON_ONCE(!sk_unhashed(sk));
 565
 566         sk->sk_hash = sk_ehashfn(sk);
 567         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 568         list = &head->chain;
 569         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 570
 571         spin_lock(lock);
 572         if (osk) {
 573                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 574                 ret = sk_nulls_del_node_init_rcu(osk);
 575         } else if (found_dup_sk) {
 576                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 577                 if (*found_dup_sk)
 578                         ret = false;
 579         }
 580
 581         if (ret)
 582                 __sk_nulls_add_node_rcu(sk, list);
 583
 584         spin_unlock(lock);
 585
 586         return ret;
 587 }
 588
 589 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 590 {
 591         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 592
 593         if (ok) {
 594                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 595         } else {
 596                 this_cpu_inc(*sk->sk_prot->orphan_count);
 597                 inet_sk_set_state(sk, TCP_CLOSE);
 598                 sock_set_flag(sk, SOCK_DEAD);
 599                 inet_csk_destroy_sock(sk);
 600         }
 601         return ok;
 602 }
 603 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 604
 605 static int inet_reuseport_add_sock(struct sock *sk,
 606                                    struct inet_listen_hashbucket *ilb)
 607 {
 608         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 609         const struct hlist_nulls_node *node;
 610         struct sock *sk2;
 611         kuid_t uid = sock_i_uid(sk);
 612
 613         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 614                 if (sk2 != sk &&
 615                     sk2->sk_family == sk->sk_family &&
 616                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 617                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 618                     inet_csk(sk2)->icsk_bind_hash == tb &&
 619                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 620                     inet_rcv_saddr_equal(sk, sk2, false))
 621                         return reuseport_add_sock(sk, sk2,
 622                                                   inet_rcv_saddr_any(sk));
 623         }
 624
 625         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 626 }
 627
 628 int __inet_hash(struct sock *sk, struct sock *osk)
 629 {
 630         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 631         struct inet_listen_hashbucket *ilb;
 632         int err = 0;
 633
 634         if (sk->sk_state != TCP_LISTEN) {
 635                 local_bh_disable();
 636                 inet_ehash_nolisten(sk, osk, NULL);
 637                 local_bh_enable();
 638                 return 0;
 639         }
 640         WARN_ON(!sk_unhashed(sk));
 641         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 642
 643         spin_lock(&ilb->lock);
 644         if (sk->sk_reuseport) {
 645                 err = inet_reuseport_add_sock(sk, ilb);
 646                 if (err)
 647                         goto unlock;
 648         }
 649         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 650                 sk->sk_family == AF_INET6)
 651                 __sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
 652         else
 653                 __sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
 654         inet_hash2(hashinfo, sk);
 655         ilb->count++;
 656         sock_set_flag(sk, SOCK_RCU_FREE);
 657         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 658 unlock:
 659         spin_unlock(&ilb->lock);
 660
 661         return err;
 662 }
 663 EXPORT_SYMBOL(__inet_hash);
 664
 665 int inet_hash(struct sock *sk)
 666 {
 667         int err = 0;
 668
 669         if (sk->sk_state != TCP_CLOSE)
 670                 err = __inet_hash(sk, NULL);
 671
 672         return err;
 673 }
 674 EXPORT_SYMBOL_GPL(inet_hash);
 675
 676 static void __inet_unhash(struct sock *sk, struct inet_listen_hashbucket *ilb)
 677 {
 678         if (sk_unhashed(sk))
 679                 return;
 680
 681         if (rcu_access_pointer(sk->sk_reuseport_cb))
 682                 reuseport_detach_sock(sk);
 683         if (ilb) {
 684                 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 685
 686                 inet_unhash2(hashinfo, sk);
 687                 ilb->count--;
 688         }
 689         __sk_nulls_del_node_init_rcu(sk);
 690         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 691 }
 692
 693 void inet_unhash(struct sock *sk)
 694 {
 695         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 696
 697         if (sk_unhashed(sk))
 698                 return;
 699
 700         if (sk->sk_state == TCP_LISTEN) {
 701                 struct inet_listen_hashbucket *ilb;
 702
 703                 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 704                 /* Don't disable bottom halves while acquiring the lock to
 705                  * avoid circular locking dependency on PREEMPT_RT.
 706                  */
 707                 spin_lock(&ilb->lock);
 708                 __inet_unhash(sk, ilb);
 709                 spin_unlock(&ilb->lock);
 710         } else {
 711                 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 712
 713                 spin_lock_bh(lock);
 714                 __inet_unhash(sk, NULL);
 715                 spin_unlock_bh(lock);
 716         }
 717 }
 718 EXPORT_SYMBOL_GPL(inet_unhash);
 719
 720 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 721  * Note that we use 32bit integers (vs RFC 'short integers')
 722  * because 2^16 is not a multiple of num_ephemeral and this
 723  * property might be used by clever attacker.
 724  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
 725  * attacks were since demonstrated, thus we use 65536 instead to really
 726  * give more isolation and privacy, at the expense of 256kB of kernel
 727  * memory.
 728  */
 729 #define INET_TABLE_PERTURB_SHIFT 16
 730 #define INET_TABLE_PERTURB_SIZE (1 << INET_TABLE_PERTURB_SHIFT)
 731 static u32 *table_perturb;
 732
 733 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 734                 struct sock *sk, u64 port_offset,
 735                 int (*check_established)(struct inet_timewait_death_row *,
 736                         struct sock *, __u16, struct inet_timewait_sock **))
 737 {
 738         struct inet_hashinfo *hinfo = death_row->hashinfo;
 739         struct inet_timewait_sock *tw = NULL;
 740         struct inet_bind_hashbucket *head;
 741         int port = inet_sk(sk)->inet_num;
 742         struct net *net = sock_net(sk);
 743         struct inet_bind_bucket *tb;
 744         u32 remaining, offset;
 745         int ret, i, low, high;
 746         int l3mdev;
 747         u32 index;
 748
 749         if (port) {
 750                 head = &hinfo->bhash[inet_bhashfn(net, port,
 751                                                   hinfo->bhash_size)];
 752                 tb = inet_csk(sk)->icsk_bind_hash;
 753                 spin_lock_bh(&head->lock);
 754                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 755                         inet_ehash_nolisten(sk, NULL, NULL);
 756                         spin_unlock_bh(&head->lock);
 757                         return 0;
 758                 }
 759                 spin_unlock(&head->lock);
 760                 /* No definite answer... Walk to established hash table */
 761                 ret = check_established(death_row, sk, port, NULL);
 762                 local_bh_enable();
 763                 return ret;
 764         }
 765
 766         l3mdev = inet_sk_bound_l3mdev(sk);
 767
 768         inet_get_local_port_range(net, &low, &high);
 769         high++; /* [32768, 60999] -> [32768, 61000[ */
 770         remaining = high - low;
 771         if (likely(remaining > 1))
 772                 remaining &= ~1U;
 773
 774         get_random_slow_once(table_perturb,
 775                              INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
 776         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
 777
 778         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
 779         offset %= remaining;
 780
 781         /* In first pass we try ports of @low parity.
 782          * inet_csk_get_port() does the opposite choice.
 783          */
 784         offset &= ~1U;
 785 other_parity_scan:
 786         port = low + offset;
 787         for (i = 0; i < remaining; i += 2, port += 2) {
 788                 if (unlikely(port >= high))
 789                         port -= remaining;
 790                 if (inet_is_local_reserved_port(net, port))
 791                         continue;
 792                 head = &hinfo->bhash[inet_bhashfn(net, port,
 793                                                   hinfo->bhash_size)];
 794                 spin_lock_bh(&head->lock);
 795
 796                 /* Does not bother with rcv_saddr checks, because
 797                  * the established check is already unique enough.
 798                  */
 799                 inet_bind_bucket_for_each(tb, &head->chain) {
 800                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 801                             tb->port == port) {
 802                                 if (tb->fastreuse >= 0 ||
 803                                     tb->fastreuseport >= 0)
 804                                         goto next_port;
 805                                 WARN_ON(hlist_empty(&tb->owners));
 806                                 if (!check_established(death_row, sk,
 807                                                        port, &tw))
 808                                         goto ok;
 809                                 goto next_port;
 810                         }
 811                 }
 812
 813                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 814                                              net, head, port, l3mdev);
 815                 if (!tb) {
 816                         spin_unlock_bh(&head->lock);
 817                         return -ENOMEM;
 818                 }
 819                 tb->fastreuse = -1;
 820                 tb->fastreuseport = -1;
 821                 goto ok;
 822 next_port:
 823                 spin_unlock_bh(&head->lock);
 824                 cond_resched();
 825         }
 826
 827         offset++;
 828         if ((offset & 1) && remaining > 1)
 829                 goto other_parity_scan;
 830
 831         return -EADDRNOTAVAIL;
 832
 833 ok:
 834         /* Here we want to add a little bit of randomness to the next source
 835          * port that will be chosen. We use a max() with a random here so that
 836          * on low contention the randomness is maximal and on high contention
 837          * it may be inexistent.
 838          */
 839         i = max_t(int, i, (prandom_u32() & 7) * 2);
 840         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
 841
 842         /* Head lock still held and bh's disabled */
 843         inet_bind_hash(sk, tb, port);
 844         if (sk_unhashed(sk)) {
 845                 inet_sk(sk)->inet_sport = htons(port);
 846                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
 847         }
 848         if (tw)
 849                 inet_twsk_bind_unhash(tw, hinfo);
 850         spin_unlock(&head->lock);
 851         if (tw)
 852                 inet_twsk_deschedule_put(tw);
 853         local_bh_enable();
 854         return 0;
 855 }
 856
 857 /*
 858  * Bind a port for a connect operation and hash it.
 859  */
 860 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 861                       struct sock *sk)
 862 {
 863         u64 port_offset = 0;
 864
 865         if (!inet_sk(sk)->inet_num)
 866                 port_offset = inet_sk_port_offset(sk);
 867         return __inet_hash_connect(death_row, sk, port_offset,
 868                                    __inet_check_established);
 869 }
 870 EXPORT_SYMBOL_GPL(inet_hash_connect);
 871
 872 void inet_hashinfo_init(struct inet_hashinfo *h)
 873 {
 874         int i;
 875
 876         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
 877                 spin_lock_init(&h->listening_hash[i].lock);
 878                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
 879                                       i + LISTENING_NULLS_BASE);
 880                 h->listening_hash[i].count = 0;
 881         }
 882
 883         h->lhash2 = NULL;
 884 }
 885 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
 886
 887 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
 888 {
 889         int i;
 890
 891         for (i = 0; i <= h->lhash2_mask; i++) {
 892                 spin_lock_init(&h->lhash2[i].lock);
 893                 INIT_HLIST_HEAD(&h->lhash2[i].head);
 894                 h->lhash2[i].count = 0;
 895         }
 896 }
 897
 898 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
 899                                 unsigned long numentries, int scale,
 900                                 unsigned long low_limit,
 901                                 unsigned long high_limit)
 902 {
 903         h->lhash2 = alloc_large_system_hash(name,
 904                                             sizeof(*h->lhash2),
 905                                             numentries,
 906                                             scale,
 907                                             0,
 908                                             NULL,
 909                                             &h->lhash2_mask,
 910                                             low_limit,
 911                                             high_limit);
 912         init_hashinfo_lhash2(h);
 913
 914         /* this one is used for source ports of outgoing connections */
 915         table_perturb = kmalloc_array(INET_TABLE_PERTURB_SIZE,
 916                                       sizeof(*table_perturb), GFP_KERNEL);
 917         if (!table_perturb)
 918                 panic("TCP: failed to alloc table_perturb");
 919 }
 920
 921 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
 922 {
 923         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
 924         if (!h->lhash2)
 925                 return -ENOMEM;
 926
 927         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
 928         /* INET_LHTABLE_SIZE must be a power of 2 */
 929         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
 930
 931         init_hashinfo_lhash2(h);
 932         return 0;
 933 }
 934 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
 935
 936 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 937 {
 938         unsigned int locksz = sizeof(spinlock_t);
 939         unsigned int i, nblocks = 1;
 940
 941         if (locksz != 0) {
 942                 /* allocate 2 cache lines or at least one spinlock per cpu */
 943                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
 944                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
 945
 946                 /* no more locks than number of hash buckets */
 947                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 948
 949                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
 950                 if (!hashinfo->ehash_locks)
 951                         return -ENOMEM;
 952
 953                 for (i = 0; i < nblocks; i++)
 954                         spin_lock_init(&hashinfo->ehash_locks[i]);
 955         }
 956         hashinfo->ehash_locks_mask = nblocks - 1;
 957         return 0;
 958 }
 959 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);