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 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
  85                                                    struct net *net,
  86                                                    struct inet_bind2_hashbucket *head,
  87                                                    const unsigned short port,
  88                                                    int l3mdev,
  89                                                    const struct sock *sk)
  90 {
  91         struct inet_bind2_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  92
  93         if (tb) {
  94                 write_pnet(&tb->ib_net, net);
  95                 tb->l3mdev    = l3mdev;
  96                 tb->port      = port;
  97 #if IS_ENABLED(CONFIG_IPV6)
  98                 if (sk->sk_family == AF_INET6)
  99                         tb->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 100                 else
 101 #endif
 102                         tb->rcv_saddr = sk->sk_rcv_saddr;
 103                 INIT_HLIST_HEAD(&tb->owners);
 104                 hlist_add_head(&tb->node, &head->chain);
 105         }
 106         return tb;
 107 }
 108
 109 static bool bind2_bucket_addr_match(struct inet_bind2_bucket *tb2, struct sock *sk)
 110 {
 111 #if IS_ENABLED(CONFIG_IPV6)
 112         if (sk->sk_family == AF_INET6)
 113                 return ipv6_addr_equal(&tb2->v6_rcv_saddr,
 114                                        &sk->sk_v6_rcv_saddr);
 115 #endif
 116         return tb2->rcv_saddr == sk->sk_rcv_saddr;
 117 }
 118
 119 /*
 120  * Caller must hold hashbucket lock for this tb with local BH disabled
 121  */
 122 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
 123 {
 124         if (hlist_empty(&tb->owners)) {
 125                 __hlist_del(&tb->node);
 126                 kmem_cache_free(cachep, tb);
 127         }
 128 }
 129
 130 /* Caller must hold the lock for the corresponding hashbucket in the bhash table
 131  * with local BH disabled
 132  */
 133 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
 134 {
 135         if (hlist_empty(&tb->owners)) {
 136                 __hlist_del(&tb->node);
 137                 kmem_cache_free(cachep, tb);
 138         }
 139 }
 140
 141 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 142                     struct inet_bind2_bucket *tb2, const unsigned short snum)
 143 {
 144         inet_sk(sk)->inet_num = snum;
 145         sk_add_bind_node(sk, &tb->owners);
 146         inet_csk(sk)->icsk_bind_hash = tb;
 147         sk_add_bind2_node(sk, &tb2->owners);
 148         inet_csk(sk)->icsk_bind2_hash = tb2;
 149 }
 150
 151 /*
 152  * Get rid of any references to a local port held by the given sock.
 153  */
 154 static void __inet_put_port(struct sock *sk)
 155 {
 156         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 157         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
 158                         hashinfo->bhash_size);
 159         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 160         struct inet_bind2_bucket *tb2;
 161         struct inet_bind_bucket *tb;
 162
 163         spin_lock(&head->lock);
 164         tb = inet_csk(sk)->icsk_bind_hash;
 165         __sk_del_bind_node(sk);
 166         inet_csk(sk)->icsk_bind_hash = NULL;
 167         inet_sk(sk)->inet_num = 0;
 168         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 169
 170         if (inet_csk(sk)->icsk_bind2_hash) {
 171                 tb2 = inet_csk(sk)->icsk_bind2_hash;
 172                 __sk_del_bind2_node(sk);
 173                 inet_csk(sk)->icsk_bind2_hash = NULL;
 174                 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
 175         }
 176         spin_unlock(&head->lock);
 177 }
 178
 179 void inet_put_port(struct sock *sk)
 180 {
 181         local_bh_disable();
 182         __inet_put_port(sk);
 183         local_bh_enable();
 184 }
 185 EXPORT_SYMBOL(inet_put_port);
 186
 187 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 188 {
 189         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
 190         unsigned short port = inet_sk(child)->inet_num;
 191         const int bhash = inet_bhashfn(sock_net(sk), port,
 192                                        table->bhash_size);
 193         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 194         struct inet_bind2_hashbucket *head_bhash2;
 195         bool created_inet_bind_bucket = false;
 196         struct net *net = sock_net(sk);
 197         struct inet_bind2_bucket *tb2;
 198         struct inet_bind_bucket *tb;
 199         int l3mdev;
 200
 201         spin_lock(&head->lock);
 202         tb = inet_csk(sk)->icsk_bind_hash;
 203         tb2 = inet_csk(sk)->icsk_bind2_hash;
 204         if (unlikely(!tb || !tb2)) {
 205                 spin_unlock(&head->lock);
 206                 return -ENOENT;
 207         }
 208         if (tb->port != port) {
 209                 l3mdev = inet_sk_bound_l3mdev(sk);
 210
 211                 /* NOTE: using tproxy and redirecting skbs to a proxy
 212                  * on a different listener port breaks the assumption
 213                  * that the listener socket's icsk_bind_hash is the same
 214                  * as that of the child socket. We have to look up or
 215                  * create a new bind bucket for the child here. */
 216                 inet_bind_bucket_for_each(tb, &head->chain) {
 217                         if (check_bind_bucket_match(tb, net, port, l3mdev))
 218                                 break;
 219                 }
 220                 if (!tb) {
 221                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 222                                                      net, head, port, l3mdev);
 223                         if (!tb) {
 224                                 spin_unlock(&head->lock);
 225                                 return -ENOMEM;
 226                         }
 227                         created_inet_bind_bucket = true;
 228                 }
 229                 inet_csk_update_fastreuse(tb, child);
 230
 231                 goto bhash2_find;
 232         } else if (!bind2_bucket_addr_match(tb2, child)) {
 233                 l3mdev = inet_sk_bound_l3mdev(sk);
 234
 235 bhash2_find:
 236                 tb2 = inet_bind2_bucket_find(table, net, port, l3mdev, child,
 237                                              &head_bhash2);
 238                 if (!tb2) {
 239                         tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
 240                                                        net, head_bhash2, port,
 241                                                        l3mdev, child);
 242                         if (!tb2)
 243                                 goto error;
 244                 }
 245         }
 246         inet_bind_hash(child, tb, tb2, port);
 247         spin_unlock(&head->lock);
 248
 249         return 0;
 250
 251 error:
 252         if (created_inet_bind_bucket)
 253                 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
 254         spin_unlock(&head->lock);
 255         return -ENOMEM;
 256 }
 257 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 258
 259 static struct inet_listen_hashbucket *
 260 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 261 {
 262         u32 hash;
 263
 264 #if IS_ENABLED(CONFIG_IPV6)
 265         if (sk->sk_family == AF_INET6)
 266                 hash = ipv6_portaddr_hash(sock_net(sk),
 267                                           &sk->sk_v6_rcv_saddr,
 268                                           inet_sk(sk)->inet_num);
 269         else
 270 #endif
 271                 hash = ipv4_portaddr_hash(sock_net(sk),
 272                                           inet_sk(sk)->inet_rcv_saddr,
 273                                           inet_sk(sk)->inet_num);
 274         return inet_lhash2_bucket(h, hash);
 275 }
 276
 277 static inline int compute_score(struct sock *sk, struct net *net,
 278                                 const unsigned short hnum, const __be32 daddr,
 279                                 const int dif, const int sdif)
 280 {
 281         int score = -1;
 282
 283         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 284                         !ipv6_only_sock(sk)) {
 285                 if (sk->sk_rcv_saddr != daddr)
 286                         return -1;
 287
 288                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 289                         return -1;
 290                 score =  sk->sk_bound_dev_if ? 2 : 1;
 291
 292                 if (sk->sk_family == PF_INET)
 293                         score++;
 294                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 295                         score++;
 296         }
 297         return score;
 298 }
 299
 300 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
 301                                             struct sk_buff *skb, int doff,
 302                                             __be32 saddr, __be16 sport,
 303                                             __be32 daddr, unsigned short hnum)
 304 {
 305         struct sock *reuse_sk = NULL;
 306         u32 phash;
 307
 308         if (sk->sk_reuseport) {
 309                 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 310                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 311         }
 312         return reuse_sk;
 313 }
 314
 315 /*
 316  * Here are some nice properties to exploit here. The BSD API
 317  * does not allow a listening sock to specify the remote port nor the
 318  * remote address for the connection. So always assume those are both
 319  * wildcarded during the search since they can never be otherwise.
 320  */
 321
 322 /* called with rcu_read_lock() : No refcount taken on the socket */
 323 static struct sock *inet_lhash2_lookup(struct net *net,
 324                                 struct inet_listen_hashbucket *ilb2,
 325                                 struct sk_buff *skb, int doff,
 326                                 const __be32 saddr, __be16 sport,
 327                                 const __be32 daddr, const unsigned short hnum,
 328                                 const int dif, const int sdif)
 329 {
 330         struct sock *sk, *result = NULL;
 331         struct hlist_nulls_node *node;
 332         int score, hiscore = 0;
 333
 334         sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
 335                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 336                 if (score > hiscore) {
 337                         result = lookup_reuseport(net, sk, skb, doff,
 338                                                   saddr, sport, daddr, hnum);
 339                         if (result)
 340                                 return result;
 341
 342                         result = sk;
 343                         hiscore = score;
 344                 }
 345         }
 346
 347         return result;
 348 }
 349
 350 static inline struct sock *inet_lookup_run_bpf(struct net *net,
 351                                                struct inet_hashinfo *hashinfo,
 352                                                struct sk_buff *skb, int doff,
 353                                                __be32 saddr, __be16 sport,
 354                                                __be32 daddr, u16 hnum, const int dif)
 355 {
 356         struct sock *sk, *reuse_sk;
 357         bool no_reuseport;
 358
 359         if (hashinfo != &tcp_hashinfo)
 360                 return NULL; /* only TCP is supported */
 361
 362         no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP, saddr, sport,
 363                                             daddr, hnum, dif, &sk);
 364         if (no_reuseport || IS_ERR_OR_NULL(sk))
 365                 return sk;
 366
 367         reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
 368         if (reuse_sk)
 369                 sk = reuse_sk;
 370         return sk;
 371 }
 372
 373 struct sock *__inet_lookup_listener(struct net *net,
 374                                     struct inet_hashinfo *hashinfo,
 375                                     struct sk_buff *skb, int doff,
 376                                     const __be32 saddr, __be16 sport,
 377                                     const __be32 daddr, const unsigned short hnum,
 378                                     const int dif, const int sdif)
 379 {
 380         struct inet_listen_hashbucket *ilb2;
 381         struct sock *result = NULL;
 382         unsigned int hash2;
 383
 384         /* Lookup redirect from BPF */
 385         if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
 386                 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
 387                                              saddr, sport, daddr, hnum, dif);
 388                 if (result)
 389                         goto done;
 390         }
 391
 392         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 393         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 394
 395         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 396                                     saddr, sport, daddr, hnum,
 397                                     dif, sdif);
 398         if (result)
 399                 goto done;
 400
 401         /* Lookup lhash2 with INADDR_ANY */
 402         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 403         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 404
 405         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 406                                     saddr, sport, htonl(INADDR_ANY), hnum,
 407                                     dif, sdif);
 408 done:
 409         if (IS_ERR(result))
 410                 return NULL;
 411         return result;
 412 }
 413 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 414
 415 /* All sockets share common refcount, but have different destructors */
 416 void sock_gen_put(struct sock *sk)
 417 {
 418         if (!refcount_dec_and_test(&sk->sk_refcnt))
 419                 return;
 420
 421         if (sk->sk_state == TCP_TIME_WAIT)
 422                 inet_twsk_free(inet_twsk(sk));
 423         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 424                 reqsk_free(inet_reqsk(sk));
 425         else
 426                 sk_free(sk);
 427 }
 428 EXPORT_SYMBOL_GPL(sock_gen_put);
 429
 430 void sock_edemux(struct sk_buff *skb)
 431 {
 432         sock_gen_put(skb->sk);
 433 }
 434 EXPORT_SYMBOL(sock_edemux);
 435
 436 struct sock *__inet_lookup_established(struct net *net,
 437                                   struct inet_hashinfo *hashinfo,
 438                                   const __be32 saddr, const __be16 sport,
 439                                   const __be32 daddr, const u16 hnum,
 440                                   const int dif, const int sdif)
 441 {
 442         INET_ADDR_COOKIE(acookie, saddr, daddr);
 443         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 444         struct sock *sk;
 445         const struct hlist_nulls_node *node;
 446         /* Optimize here for direct hit, only listening connections can
 447          * have wildcards anyways.
 448          */
 449         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 450         unsigned int slot = hash & hashinfo->ehash_mask;
 451         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 452
 453 begin:
 454         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 455                 if (sk->sk_hash != hash)
 456                         continue;
 457                 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
 458                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 459                                 goto out;
 460                         if (unlikely(!inet_match(net, sk, acookie,
 461                                                  ports, dif, sdif))) {
 462                                 sock_gen_put(sk);
 463                                 goto begin;
 464                         }
 465                         goto found;
 466                 }
 467         }
 468         /*
 469          * if the nulls value we got at the end of this lookup is
 470          * not the expected one, we must restart lookup.
 471          * We probably met an item that was moved to another chain.
 472          */
 473         if (get_nulls_value(node) != slot)
 474                 goto begin;
 475 out:
 476         sk = NULL;
 477 found:
 478         return sk;
 479 }
 480 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 481
 482 /* called with local bh disabled */
 483 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 484                                     struct sock *sk, __u16 lport,
 485                                     struct inet_timewait_sock **twp)
 486 {
 487         struct inet_hashinfo *hinfo = death_row->hashinfo;
 488         struct inet_sock *inet = inet_sk(sk);
 489         __be32 daddr = inet->inet_rcv_saddr;
 490         __be32 saddr = inet->inet_daddr;
 491         int dif = sk->sk_bound_dev_if;
 492         struct net *net = sock_net(sk);
 493         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 494         INET_ADDR_COOKIE(acookie, saddr, daddr);
 495         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 496         unsigned int hash = inet_ehashfn(net, daddr, lport,
 497                                          saddr, inet->inet_dport);
 498         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 499         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 500         struct sock *sk2;
 501         const struct hlist_nulls_node *node;
 502         struct inet_timewait_sock *tw = NULL;
 503
 504         spin_lock(lock);
 505
 506         sk_nulls_for_each(sk2, node, &head->chain) {
 507                 if (sk2->sk_hash != hash)
 508                         continue;
 509
 510                 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
 511                         if (sk2->sk_state == TCP_TIME_WAIT) {
 512                                 tw = inet_twsk(sk2);
 513                                 if (twsk_unique(sk, sk2, twp))
 514                                         break;
 515                         }
 516                         goto not_unique;
 517                 }
 518         }
 519
 520         /* Must record num and sport now. Otherwise we will see
 521          * in hash table socket with a funny identity.
 522          */
 523         inet->inet_num = lport;
 524         inet->inet_sport = htons(lport);
 525         sk->sk_hash = hash;
 526         WARN_ON(!sk_unhashed(sk));
 527         __sk_nulls_add_node_rcu(sk, &head->chain);
 528         if (tw) {
 529                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 530                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 531         }
 532         spin_unlock(lock);
 533         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 534
 535         if (twp) {
 536                 *twp = tw;
 537         } else if (tw) {
 538                 /* Silly. Should hash-dance instead... */
 539                 inet_twsk_deschedule_put(tw);
 540         }
 541         return 0;
 542
 543 not_unique:
 544         spin_unlock(lock);
 545         return -EADDRNOTAVAIL;
 546 }
 547
 548 static u64 inet_sk_port_offset(const struct sock *sk)
 549 {
 550         const struct inet_sock *inet = inet_sk(sk);
 551
 552         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 553                                           inet->inet_daddr,
 554                                           inet->inet_dport);
 555 }
 556
 557 /* Searches for an exsiting socket in the ehash bucket list.
 558  * Returns true if found, false otherwise.
 559  */
 560 static bool inet_ehash_lookup_by_sk(struct sock *sk,
 561                                     struct hlist_nulls_head *list)
 562 {
 563         const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
 564         const int sdif = sk->sk_bound_dev_if;
 565         const int dif = sk->sk_bound_dev_if;
 566         const struct hlist_nulls_node *node;
 567         struct net *net = sock_net(sk);
 568         struct sock *esk;
 569
 570         INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
 571
 572         sk_nulls_for_each_rcu(esk, node, list) {
 573                 if (esk->sk_hash != sk->sk_hash)
 574                         continue;
 575                 if (sk->sk_family == AF_INET) {
 576                         if (unlikely(inet_match(net, esk, acookie,
 577                                                 ports, dif, sdif))) {
 578                                 return true;
 579                         }
 580                 }
 581 #if IS_ENABLED(CONFIG_IPV6)
 582                 else if (sk->sk_family == AF_INET6) {
 583                         if (unlikely(inet6_match(net, esk,
 584                                                  &sk->sk_v6_daddr,
 585                                                  &sk->sk_v6_rcv_saddr,
 586                                                  ports, dif, sdif))) {
 587                                 return true;
 588                         }
 589                 }
 590 #endif
 591         }
 592         return false;
 593 }
 594
 595 /* Insert a socket into ehash, and eventually remove another one
 596  * (The another one can be a SYN_RECV or TIMEWAIT)
 597  * If an existing socket already exists, socket sk is not inserted,
 598  * and sets found_dup_sk parameter to true.
 599  */
 600 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 601 {
 602         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 603         struct hlist_nulls_head *list;
 604         struct inet_ehash_bucket *head;
 605         spinlock_t *lock;
 606         bool ret = true;
 607
 608         WARN_ON_ONCE(!sk_unhashed(sk));
 609
 610         sk->sk_hash = sk_ehashfn(sk);
 611         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 612         list = &head->chain;
 613         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 614
 615         spin_lock(lock);
 616         if (osk) {
 617                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 618                 ret = sk_nulls_del_node_init_rcu(osk);
 619         } else if (found_dup_sk) {
 620                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 621                 if (*found_dup_sk)
 622                         ret = false;
 623         }
 624
 625         if (ret)
 626                 __sk_nulls_add_node_rcu(sk, list);
 627
 628         spin_unlock(lock);
 629
 630         return ret;
 631 }
 632
 633 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 634 {
 635         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 636
 637         if (ok) {
 638                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 639         } else {
 640                 this_cpu_inc(*sk->sk_prot->orphan_count);
 641                 inet_sk_set_state(sk, TCP_CLOSE);
 642                 sock_set_flag(sk, SOCK_DEAD);
 643                 inet_csk_destroy_sock(sk);
 644         }
 645         return ok;
 646 }
 647 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 648
 649 static int inet_reuseport_add_sock(struct sock *sk,
 650                                    struct inet_listen_hashbucket *ilb)
 651 {
 652         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 653         const struct hlist_nulls_node *node;
 654         struct sock *sk2;
 655         kuid_t uid = sock_i_uid(sk);
 656
 657         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 658                 if (sk2 != sk &&
 659                     sk2->sk_family == sk->sk_family &&
 660                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 661                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 662                     inet_csk(sk2)->icsk_bind_hash == tb &&
 663                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 664                     inet_rcv_saddr_equal(sk, sk2, false))
 665                         return reuseport_add_sock(sk, sk2,
 666                                                   inet_rcv_saddr_any(sk));
 667         }
 668
 669         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 670 }
 671
 672 int __inet_hash(struct sock *sk, struct sock *osk)
 673 {
 674         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 675         struct inet_listen_hashbucket *ilb2;
 676         int err = 0;
 677
 678         if (sk->sk_state != TCP_LISTEN) {
 679                 local_bh_disable();
 680                 inet_ehash_nolisten(sk, osk, NULL);
 681                 local_bh_enable();
 682                 return 0;
 683         }
 684         WARN_ON(!sk_unhashed(sk));
 685         ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 686
 687         spin_lock(&ilb2->lock);
 688         if (sk->sk_reuseport) {
 689                 err = inet_reuseport_add_sock(sk, ilb2);
 690                 if (err)
 691                         goto unlock;
 692         }
 693         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 694                 sk->sk_family == AF_INET6)
 695                 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
 696         else
 697                 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
 698         sock_set_flag(sk, SOCK_RCU_FREE);
 699         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 700 unlock:
 701         spin_unlock(&ilb2->lock);
 702
 703         return err;
 704 }
 705 EXPORT_SYMBOL(__inet_hash);
 706
 707 int inet_hash(struct sock *sk)
 708 {
 709         int err = 0;
 710
 711         if (sk->sk_state != TCP_CLOSE)
 712                 err = __inet_hash(sk, NULL);
 713
 714         return err;
 715 }
 716 EXPORT_SYMBOL_GPL(inet_hash);
 717
 718 void inet_unhash(struct sock *sk)
 719 {
 720         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 721
 722         if (sk_unhashed(sk))
 723                 return;
 724
 725         if (sk->sk_state == TCP_LISTEN) {
 726                 struct inet_listen_hashbucket *ilb2;
 727
 728                 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 729                 /* Don't disable bottom halves while acquiring the lock to
 730                  * avoid circular locking dependency on PREEMPT_RT.
 731                  */
 732                 spin_lock(&ilb2->lock);
 733                 if (sk_unhashed(sk)) {
 734                         spin_unlock(&ilb2->lock);
 735                         return;
 736                 }
 737
 738                 if (rcu_access_pointer(sk->sk_reuseport_cb))
 739                         reuseport_stop_listen_sock(sk);
 740
 741                 __sk_nulls_del_node_init_rcu(sk);
 742                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 743                 spin_unlock(&ilb2->lock);
 744         } else {
 745                 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 746
 747                 spin_lock_bh(lock);
 748                 if (sk_unhashed(sk)) {
 749                         spin_unlock_bh(lock);
 750                         return;
 751                 }
 752                 __sk_nulls_del_node_init_rcu(sk);
 753                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 754                 spin_unlock_bh(lock);
 755         }
 756 }
 757 EXPORT_SYMBOL_GPL(inet_unhash);
 758
 759 static bool check_bind2_bucket_match(struct inet_bind2_bucket *tb,
 760                                      struct net *net, unsigned short port,
 761                                      int l3mdev, struct sock *sk)
 762 {
 763 #if IS_ENABLED(CONFIG_IPV6)
 764         if (sk->sk_family == AF_INET6)
 765                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 766                         tb->l3mdev == l3mdev &&
 767                         ipv6_addr_equal(&tb->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
 768         else
 769 #endif
 770                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 771                         tb->l3mdev == l3mdev && tb->rcv_saddr == sk->sk_rcv_saddr;
 772 }
 773
 774 bool check_bind2_bucket_match_nulladdr(struct inet_bind2_bucket *tb,
 775                                        struct net *net, const unsigned short port,
 776                                        int l3mdev, const struct sock *sk)
 777 {
 778 #if IS_ENABLED(CONFIG_IPV6)
 779         struct in6_addr nulladdr = {};
 780
 781         if (sk->sk_family == AF_INET6)
 782                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 783                         tb->l3mdev == l3mdev &&
 784                         ipv6_addr_equal(&tb->v6_rcv_saddr, &nulladdr);
 785         else
 786 #endif
 787                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 788                         tb->l3mdev == l3mdev && tb->rcv_saddr == 0;
 789 }
 790
 791 static struct inet_bind2_hashbucket *
 792 inet_bhashfn_portaddr(struct inet_hashinfo *hinfo, const struct sock *sk,
 793                       const struct net *net, unsigned short port)
 794 {
 795         u32 hash;
 796
 797 #if IS_ENABLED(CONFIG_IPV6)
 798         if (sk->sk_family == AF_INET6)
 799                 hash = ipv6_portaddr_hash(net, &sk->sk_v6_rcv_saddr, port);
 800         else
 801 #endif
 802                 hash = ipv4_portaddr_hash(net, sk->sk_rcv_saddr, port);
 803         return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
 804 }
 805
 806 /* This should only be called when the spinlock for the socket's corresponding
 807  * bind_hashbucket is held
 808  */
 809 struct inet_bind2_bucket *
 810 inet_bind2_bucket_find(struct inet_hashinfo *hinfo, struct net *net,
 811                        const unsigned short port, int l3mdev, struct sock *sk,
 812                        struct inet_bind2_hashbucket **head)
 813 {
 814         struct inet_bind2_bucket *bhash2 = NULL;
 815         struct inet_bind2_hashbucket *h;
 816
 817         h = inet_bhashfn_portaddr(hinfo, sk, net, port);
 818         inet_bind_bucket_for_each(bhash2, &h->chain) {
 819                 if (check_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
 820                         break;
 821         }
 822
 823         if (head)
 824                 *head = h;
 825
 826         return bhash2;
 827 }
 828
 829 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 830  * Note that we use 32bit integers (vs RFC 'short integers')
 831  * because 2^16 is not a multiple of num_ephemeral and this
 832  * property might be used by clever attacker.
 833  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
 834  * attacks were since demonstrated, thus we use 65536 instead to really
 835  * give more isolation and privacy, at the expense of 256kB of kernel
 836  * memory.
 837  */
 838 #define INET_TABLE_PERTURB_SHIFT 16
 839 #define INET_TABLE_PERTURB_SIZE (1 << INET_TABLE_PERTURB_SHIFT)
 840 static u32 *table_perturb;
 841
 842 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 843                 struct sock *sk, u64 port_offset,
 844                 int (*check_established)(struct inet_timewait_death_row *,
 845                         struct sock *, __u16, struct inet_timewait_sock **))
 846 {
 847         struct inet_hashinfo *hinfo = death_row->hashinfo;
 848         struct inet_timewait_sock *tw = NULL;
 849         struct inet_bind2_hashbucket *head2;
 850         struct inet_bind_hashbucket *head;
 851         int port = inet_sk(sk)->inet_num;
 852         struct net *net = sock_net(sk);
 853         struct inet_bind2_bucket *tb2;
 854         struct inet_bind_bucket *tb;
 855         bool tb_created = false;
 856         u32 remaining, offset;
 857         int ret, i, low, high;
 858         int l3mdev;
 859         u32 index;
 860
 861         if (port) {
 862                 head = &hinfo->bhash[inet_bhashfn(net, port,
 863                                                   hinfo->bhash_size)];
 864                 tb = inet_csk(sk)->icsk_bind_hash;
 865                 spin_lock_bh(&head->lock);
 866                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 867                         inet_ehash_nolisten(sk, NULL, NULL);
 868                         spin_unlock_bh(&head->lock);
 869                         return 0;
 870                 }
 871                 spin_unlock(&head->lock);
 872                 /* No definite answer... Walk to established hash table */
 873                 ret = check_established(death_row, sk, port, NULL);
 874                 local_bh_enable();
 875                 return ret;
 876         }
 877
 878         l3mdev = inet_sk_bound_l3mdev(sk);
 879
 880         inet_get_local_port_range(net, &low, &high);
 881         high++; /* [32768, 60999] -> [32768, 61000[ */
 882         remaining = high - low;
 883         if (likely(remaining > 1))
 884                 remaining &= ~1U;
 885
 886         net_get_random_once(table_perturb,
 887                             INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
 888         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
 889
 890         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
 891         offset %= remaining;
 892
 893         /* In first pass we try ports of @low parity.
 894          * inet_csk_get_port() does the opposite choice.
 895          */
 896         offset &= ~1U;
 897 other_parity_scan:
 898         port = low + offset;
 899         for (i = 0; i < remaining; i += 2, port += 2) {
 900                 if (unlikely(port >= high))
 901                         port -= remaining;
 902                 if (inet_is_local_reserved_port(net, port))
 903                         continue;
 904                 head = &hinfo->bhash[inet_bhashfn(net, port,
 905                                                   hinfo->bhash_size)];
 906                 spin_lock_bh(&head->lock);
 907
 908                 /* Does not bother with rcv_saddr checks, because
 909                  * the established check is already unique enough.
 910                  */
 911                 inet_bind_bucket_for_each(tb, &head->chain) {
 912                         if (check_bind_bucket_match(tb, net, port, l3mdev)) {
 913                                 if (tb->fastreuse >= 0 ||
 914                                     tb->fastreuseport >= 0)
 915                                         goto next_port;
 916                                 WARN_ON(hlist_empty(&tb->owners));
 917                                 if (!check_established(death_row, sk,
 918                                                        port, &tw))
 919                                         goto ok;
 920                                 goto next_port;
 921                         }
 922                 }
 923
 924                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 925                                              net, head, port, l3mdev);
 926                 if (!tb) {
 927                         spin_unlock_bh(&head->lock);
 928                         return -ENOMEM;
 929                 }
 930                 tb_created = true;
 931                 tb->fastreuse = -1;
 932                 tb->fastreuseport = -1;
 933                 goto ok;
 934 next_port:
 935                 spin_unlock_bh(&head->lock);
 936                 cond_resched();
 937         }
 938
 939         offset++;
 940         if ((offset & 1) && remaining > 1)
 941                 goto other_parity_scan;
 942
 943         return -EADDRNOTAVAIL;
 944
 945 ok:
 946         /* Find the corresponding tb2 bucket since we need to
 947          * add the socket to the bhash2 table as well
 948          */
 949         tb2 = inet_bind2_bucket_find(hinfo, net, port, l3mdev, sk, &head2);
 950         if (!tb2) {
 951                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
 952                                                head2, port, l3mdev, sk);
 953                 if (!tb2)
 954                         goto error;
 955         }
 956
 957         /* Here we want to add a little bit of randomness to the next source
 958          * port that will be chosen. We use a max() with a random here so that
 959          * on low contention the randomness is maximal and on high contention
 960          * it may be inexistent.
 961          */
 962         i = max_t(int, i, (prandom_u32() & 7) * 2);
 963         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
 964
 965         /* Head lock still held and bh's disabled */
 966         inet_bind_hash(sk, tb, tb2, port);
 967         if (sk_unhashed(sk)) {
 968                 inet_sk(sk)->inet_sport = htons(port);
 969                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
 970         }
 971         if (tw)
 972                 inet_twsk_bind_unhash(tw, hinfo);
 973         spin_unlock(&head->lock);
 974         if (tw)
 975                 inet_twsk_deschedule_put(tw);
 976         local_bh_enable();
 977         return 0;
 978
 979 error:
 980         if (tb_created)
 981                 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
 982         spin_unlock_bh(&head->lock);
 983         return -ENOMEM;
 984 }
 985
 986 /*
 987  * Bind a port for a connect operation and hash it.
 988  */
 989 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 990                       struct sock *sk)
 991 {
 992         u64 port_offset = 0;
 993
 994         if (!inet_sk(sk)->inet_num)
 995                 port_offset = inet_sk_port_offset(sk);
 996         return __inet_hash_connect(death_row, sk, port_offset,
 997                                    __inet_check_established);
 998 }
 999 EXPORT_SYMBOL_GPL(inet_hash_connect);
1000
1001 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1002 {
1003         int i;
1004
1005         for (i = 0; i <= h->lhash2_mask; i++) {
1006                 spin_lock_init(&h->lhash2[i].lock);
1007                 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1008                                       i + LISTENING_NULLS_BASE);
1009         }
1010 }
1011
1012 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1013                                 unsigned long numentries, int scale,
1014                                 unsigned long low_limit,
1015                                 unsigned long high_limit)
1016 {
1017         h->lhash2 = alloc_large_system_hash(name,
1018                                             sizeof(*h->lhash2),
1019                                             numentries,
1020                                             scale,
1021                                             0,
1022                                             NULL,
1023                                             &h->lhash2_mask,
1024                                             low_limit,
1025                                             high_limit);
1026         init_hashinfo_lhash2(h);
1027
1028         /* this one is used for source ports of outgoing connections */
1029         table_perturb = alloc_large_system_hash("Table-perturb",
1030                                                 sizeof(*table_perturb),
1031                                                 INET_TABLE_PERTURB_SIZE,
1032                                                 0, 0, NULL, NULL,
1033                                                 INET_TABLE_PERTURB_SIZE,
1034                                                 INET_TABLE_PERTURB_SIZE);
1035 }
1036
1037 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1038 {
1039         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1040         if (!h->lhash2)
1041                 return -ENOMEM;
1042
1043         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1044         /* INET_LHTABLE_SIZE must be a power of 2 */
1045         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1046
1047         init_hashinfo_lhash2(h);
1048         return 0;
1049 }
1050 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1051
1052 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1053 {
1054         unsigned int locksz = sizeof(spinlock_t);
1055         unsigned int i, nblocks = 1;
1056
1057         if (locksz != 0) {
1058                 /* allocate 2 cache lines or at least one spinlock per cpu */
1059                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1060                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1061
1062                 /* no more locks than number of hash buckets */
1063                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1064
1065                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1066                 if (!hashinfo->ehash_locks)
1067                         return -ENOMEM;
1068
1069                 for (i = 0; i < nblocks; i++)
1070                         spin_lock_init(&hashinfo->ehash_locks[i]);
1071         }
1072         hashinfo->ehash_locks_mask = nblocks - 1;
1073         return 0;
1074 }
1075 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);