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/hotdata.h>
  27 #include <net/ip.h>
  28 #include <net/rps.h>
  29 #include <net/secure_seq.h>
  30 #include <net/sock_reuseport.h>
  31 #include <net/tcp.h>
  32
  33 u32 inet_ehashfn(const struct net *net, const __be32 laddr,
  34                  const __u16 lport, const __be32 faddr,
  35                  const __be16 fport)
  36 {
  37         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  38
  39         return lport + __inet_ehashfn(laddr, 0, faddr, fport,
  40                                       inet_ehash_secret + net_hash_mix(net));
  41 }
  42 EXPORT_SYMBOL_GPL(inet_ehashfn);
  43
  44 /* This function handles inet_sock, but also timewait and request sockets
  45  * for IPv4/IPv6.
  46  */
  47 static u32 sk_ehashfn(const struct sock *sk)
  48 {
  49 #if IS_ENABLED(CONFIG_IPV6)
  50         if (sk->sk_family == AF_INET6 &&
  51             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  52                 return inet6_ehashfn(sock_net(sk),
  53                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  54                                      &sk->sk_v6_daddr, sk->sk_dport);
  55 #endif
  56         return inet_ehashfn(sock_net(sk),
  57                             sk->sk_rcv_saddr, sk->sk_num,
  58                             sk->sk_daddr, sk->sk_dport);
  59 }
  60
  61 /*
  62  * Allocate and initialize a new local port bind bucket.
  63  * The bindhash mutex for snum's hash chain must be held here.
  64  */
  65 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  66                                                  struct net *net,
  67                                                  struct inet_bind_hashbucket *head,
  68                                                  const unsigned short snum,
  69                                                  int l3mdev)
  70 {
  71         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  72
  73         if (tb) {
  74                 write_pnet(&tb->ib_net, net);
  75                 tb->l3mdev    = l3mdev;
  76                 tb->port      = snum;
  77                 tb->fastreuse = 0;
  78                 tb->fastreuseport = 0;
  79                 INIT_HLIST_HEAD(&tb->bhash2);
  80                 hlist_add_head_rcu(&tb->node, &head->chain);
  81         }
  82         return tb;
  83 }
  84
  85 /*
  86  * Caller must hold hashbucket lock for this tb with local BH disabled
  87  */
  88 void inet_bind_bucket_destroy(struct inet_bind_bucket *tb)
  89 {
  90         if (hlist_empty(&tb->bhash2)) {
  91                 hlist_del_rcu(&tb->node);
  92                 kfree_rcu(tb, rcu);
  93         }
  94 }
  95
  96 bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
  97                             unsigned short port, int l3mdev)
  98 {
  99         return net_eq(ib_net(tb), net) && tb->port == port &&
 100                 tb->l3mdev == l3mdev;
 101 }
 102
 103 static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb2,
 104                                    struct net *net,
 105                                    struct inet_bind_hashbucket *head,
 106                                    struct inet_bind_bucket *tb,
 107                                    const struct sock *sk)
 108 {
 109         write_pnet(&tb2->ib_net, net);
 110         tb2->l3mdev = tb->l3mdev;
 111         tb2->port = tb->port;
 112 #if IS_ENABLED(CONFIG_IPV6)
 113         BUILD_BUG_ON(USHRT_MAX < (IPV6_ADDR_ANY | IPV6_ADDR_MAPPED));
 114         if (sk->sk_family == AF_INET6) {
 115                 tb2->addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
 116                 tb2->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 117         } else {
 118                 tb2->addr_type = IPV6_ADDR_MAPPED;
 119                 ipv6_addr_set_v4mapped(sk->sk_rcv_saddr, &tb2->v6_rcv_saddr);
 120         }
 121 #else
 122         tb2->rcv_saddr = sk->sk_rcv_saddr;
 123 #endif
 124         INIT_HLIST_HEAD(&tb2->owners);
 125         hlist_add_head(&tb2->node, &head->chain);
 126         hlist_add_head(&tb2->bhash_node, &tb->bhash2);
 127 }
 128
 129 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
 130                                                    struct net *net,
 131                                                    struct inet_bind_hashbucket *head,
 132                                                    struct inet_bind_bucket *tb,
 133                                                    const struct sock *sk)
 134 {
 135         struct inet_bind2_bucket *tb2 = kmem_cache_alloc(cachep, GFP_ATOMIC);
 136
 137         if (tb2)
 138                 inet_bind2_bucket_init(tb2, net, head, tb, sk);
 139
 140         return tb2;
 141 }
 142
 143 /* Caller must hold hashbucket lock for this tb with local BH disabled */
 144 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
 145 {
 146         if (hlist_empty(&tb->owners)) {
 147                 __hlist_del(&tb->node);
 148                 __hlist_del(&tb->bhash_node);
 149                 kmem_cache_free(cachep, tb);
 150         }
 151 }
 152
 153 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
 154                                          const struct sock *sk)
 155 {
 156 #if IS_ENABLED(CONFIG_IPV6)
 157         if (sk->sk_family == AF_INET6)
 158                 return ipv6_addr_equal(&tb2->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
 159
 160         if (tb2->addr_type != IPV6_ADDR_MAPPED)
 161                 return false;
 162 #endif
 163         return tb2->rcv_saddr == sk->sk_rcv_saddr;
 164 }
 165
 166 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 167                     struct inet_bind2_bucket *tb2, unsigned short port)
 168 {
 169         inet_sk(sk)->inet_num = port;
 170         inet_csk(sk)->icsk_bind_hash = tb;
 171         inet_csk(sk)->icsk_bind2_hash = tb2;
 172         sk_add_bind_node(sk, &tb2->owners);
 173 }
 174
 175 /*
 176  * Get rid of any references to a local port held by the given sock.
 177  */
 178 static void __inet_put_port(struct sock *sk)
 179 {
 180         struct inet_hashinfo *hashinfo = tcp_get_hashinfo(sk);
 181         struct inet_bind_hashbucket *head, *head2;
 182         struct net *net = sock_net(sk);
 183         struct inet_bind_bucket *tb;
 184         int bhash;
 185
 186         bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
 187         head = &hashinfo->bhash[bhash];
 188         head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
 189
 190         spin_lock(&head->lock);
 191         tb = inet_csk(sk)->icsk_bind_hash;
 192         inet_csk(sk)->icsk_bind_hash = NULL;
 193         inet_sk(sk)->inet_num = 0;
 194
 195         spin_lock(&head2->lock);
 196         if (inet_csk(sk)->icsk_bind2_hash) {
 197                 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
 198
 199                 __sk_del_bind_node(sk);
 200                 inet_csk(sk)->icsk_bind2_hash = NULL;
 201                 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
 202         }
 203         spin_unlock(&head2->lock);
 204
 205         inet_bind_bucket_destroy(tb);
 206         spin_unlock(&head->lock);
 207 }
 208
 209 void inet_put_port(struct sock *sk)
 210 {
 211         local_bh_disable();
 212         __inet_put_port(sk);
 213         local_bh_enable();
 214 }
 215 EXPORT_SYMBOL(inet_put_port);
 216
 217 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 218 {
 219         struct inet_hashinfo *table = tcp_get_hashinfo(sk);
 220         unsigned short port = inet_sk(child)->inet_num;
 221         struct inet_bind_hashbucket *head, *head2;
 222         bool created_inet_bind_bucket = false;
 223         struct net *net = sock_net(sk);
 224         bool update_fastreuse = false;
 225         struct inet_bind2_bucket *tb2;
 226         struct inet_bind_bucket *tb;
 227         int bhash, l3mdev;
 228
 229         bhash = inet_bhashfn(net, port, table->bhash_size);
 230         head = &table->bhash[bhash];
 231         head2 = inet_bhashfn_portaddr(table, child, net, port);
 232
 233         spin_lock(&head->lock);
 234         spin_lock(&head2->lock);
 235         tb = inet_csk(sk)->icsk_bind_hash;
 236         tb2 = inet_csk(sk)->icsk_bind2_hash;
 237         if (unlikely(!tb || !tb2)) {
 238                 spin_unlock(&head2->lock);
 239                 spin_unlock(&head->lock);
 240                 return -ENOENT;
 241         }
 242         if (tb->port != port) {
 243                 l3mdev = inet_sk_bound_l3mdev(sk);
 244
 245                 /* NOTE: using tproxy and redirecting skbs to a proxy
 246                  * on a different listener port breaks the assumption
 247                  * that the listener socket's icsk_bind_hash is the same
 248                  * as that of the child socket. We have to look up or
 249                  * create a new bind bucket for the child here. */
 250                 inet_bind_bucket_for_each(tb, &head->chain) {
 251                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 252                                 break;
 253                 }
 254                 if (!tb) {
 255                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 256                                                      net, head, port, l3mdev);
 257                         if (!tb) {
 258                                 spin_unlock(&head2->lock);
 259                                 spin_unlock(&head->lock);
 260                                 return -ENOMEM;
 261                         }
 262                         created_inet_bind_bucket = true;
 263                 }
 264                 update_fastreuse = true;
 265
 266                 goto bhash2_find;
 267         } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
 268                 l3mdev = inet_sk_bound_l3mdev(sk);
 269
 270 bhash2_find:
 271                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
 272                 if (!tb2) {
 273                         tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
 274                                                        net, head2, tb, child);
 275                         if (!tb2)
 276                                 goto error;
 277                 }
 278         }
 279         if (update_fastreuse)
 280                 inet_csk_update_fastreuse(tb, child);
 281         inet_bind_hash(child, tb, tb2, port);
 282         spin_unlock(&head2->lock);
 283         spin_unlock(&head->lock);
 284
 285         return 0;
 286
 287 error:
 288         if (created_inet_bind_bucket)
 289                 inet_bind_bucket_destroy(tb);
 290         spin_unlock(&head2->lock);
 291         spin_unlock(&head->lock);
 292         return -ENOMEM;
 293 }
 294 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 295
 296 static struct inet_listen_hashbucket *
 297 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 298 {
 299         u32 hash;
 300
 301 #if IS_ENABLED(CONFIG_IPV6)
 302         if (sk->sk_family == AF_INET6)
 303                 hash = ipv6_portaddr_hash(sock_net(sk),
 304                                           &sk->sk_v6_rcv_saddr,
 305                                           inet_sk(sk)->inet_num);
 306         else
 307 #endif
 308                 hash = ipv4_portaddr_hash(sock_net(sk),
 309                                           inet_sk(sk)->inet_rcv_saddr,
 310                                           inet_sk(sk)->inet_num);
 311         return inet_lhash2_bucket(h, hash);
 312 }
 313
 314 static inline int compute_score(struct sock *sk, const struct net *net,
 315                                 const unsigned short hnum, const __be32 daddr,
 316                                 const int dif, const int sdif)
 317 {
 318         int score = -1;
 319
 320         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 321                         !ipv6_only_sock(sk)) {
 322                 if (sk->sk_rcv_saddr != daddr)
 323                         return -1;
 324
 325                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 326                         return -1;
 327                 score =  sk->sk_bound_dev_if ? 2 : 1;
 328
 329                 if (sk->sk_family == PF_INET)
 330                         score++;
 331                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 332                         score++;
 333         }
 334         return score;
 335 }
 336
 337 /**
 338  * inet_lookup_reuseport() - execute reuseport logic on AF_INET socket if necessary.
 339  * @net: network namespace.
 340  * @sk: AF_INET socket, must be in TCP_LISTEN state for TCP or TCP_CLOSE for UDP.
 341  * @skb: context for a potential SK_REUSEPORT program.
 342  * @doff: header offset.
 343  * @saddr: source address.
 344  * @sport: source port.
 345  * @daddr: destination address.
 346  * @hnum: destination port in host byte order.
 347  * @ehashfn: hash function used to generate the fallback hash.
 348  *
 349  * Return: NULL if sk doesn't have SO_REUSEPORT set, otherwise a pointer to
 350  *         the selected sock or an error.
 351  */
 352 struct sock *inet_lookup_reuseport(const struct net *net, struct sock *sk,
 353                                    struct sk_buff *skb, int doff,
 354                                    __be32 saddr, __be16 sport,
 355                                    __be32 daddr, unsigned short hnum,
 356                                    inet_ehashfn_t *ehashfn)
 357 {
 358         struct sock *reuse_sk = NULL;
 359         u32 phash;
 360
 361         if (sk->sk_reuseport) {
 362                 phash = INDIRECT_CALL_2(ehashfn, udp_ehashfn, inet_ehashfn,
 363                                         net, daddr, hnum, saddr, sport);
 364                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 365         }
 366         return reuse_sk;
 367 }
 368 EXPORT_SYMBOL_GPL(inet_lookup_reuseport);
 369
 370 /*
 371  * Here are some nice properties to exploit here. The BSD API
 372  * does not allow a listening sock to specify the remote port nor the
 373  * remote address for the connection. So always assume those are both
 374  * wildcarded during the search since they can never be otherwise.
 375  */
 376
 377 /* called with rcu_read_lock() : No refcount taken on the socket */
 378 static struct sock *inet_lhash2_lookup(const struct net *net,
 379                                 struct inet_listen_hashbucket *ilb2,
 380                                 struct sk_buff *skb, int doff,
 381                                 const __be32 saddr, __be16 sport,
 382                                 const __be32 daddr, const unsigned short hnum,
 383                                 const int dif, const int sdif)
 384 {
 385         struct sock *sk, *result = NULL;
 386         struct hlist_nulls_node *node;
 387         int score, hiscore = 0;
 388
 389         sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
 390                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 391                 if (score > hiscore) {
 392                         result = inet_lookup_reuseport(net, sk, skb, doff,
 393                                                        saddr, sport, daddr, hnum, inet_ehashfn);
 394                         if (result)
 395                                 return result;
 396
 397                         result = sk;
 398                         hiscore = score;
 399                 }
 400         }
 401
 402         return result;
 403 }
 404
 405 struct sock *inet_lookup_run_sk_lookup(const struct net *net,
 406                                        int protocol,
 407                                        struct sk_buff *skb, int doff,
 408                                        __be32 saddr, __be16 sport,
 409                                        __be32 daddr, u16 hnum, const int dif,
 410                                        inet_ehashfn_t *ehashfn)
 411 {
 412         struct sock *sk, *reuse_sk;
 413         bool no_reuseport;
 414
 415         no_reuseport = bpf_sk_lookup_run_v4(net, protocol, saddr, sport,
 416                                             daddr, hnum, dif, &sk);
 417         if (no_reuseport || IS_ERR_OR_NULL(sk))
 418                 return sk;
 419
 420         reuse_sk = inet_lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum,
 421                                          ehashfn);
 422         if (reuse_sk)
 423                 sk = reuse_sk;
 424         return sk;
 425 }
 426
 427 struct sock *__inet_lookup_listener(const struct net *net,
 428                                     struct inet_hashinfo *hashinfo,
 429                                     struct sk_buff *skb, int doff,
 430                                     const __be32 saddr, __be16 sport,
 431                                     const __be32 daddr, const unsigned short hnum,
 432                                     const int dif, const int sdif)
 433 {
 434         struct inet_listen_hashbucket *ilb2;
 435         struct sock *result = NULL;
 436         unsigned int hash2;
 437
 438         /* Lookup redirect from BPF */
 439         if (static_branch_unlikely(&bpf_sk_lookup_enabled) &&
 440             hashinfo == net->ipv4.tcp_death_row.hashinfo) {
 441                 result = inet_lookup_run_sk_lookup(net, IPPROTO_TCP, skb, doff,
 442                                                    saddr, sport, daddr, hnum, dif,
 443                                                    inet_ehashfn);
 444                 if (result)
 445                         goto done;
 446         }
 447
 448         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 449         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 450
 451         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 452                                     saddr, sport, daddr, hnum,
 453                                     dif, sdif);
 454         if (result)
 455                 goto done;
 456
 457         /* Lookup lhash2 with INADDR_ANY */
 458         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 459         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 460
 461         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 462                                     saddr, sport, htonl(INADDR_ANY), hnum,
 463                                     dif, sdif);
 464 done:
 465         if (IS_ERR(result))
 466                 return NULL;
 467         return result;
 468 }
 469 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 470
 471 /* All sockets share common refcount, but have different destructors */
 472 void sock_gen_put(struct sock *sk)
 473 {
 474         if (!refcount_dec_and_test(&sk->sk_refcnt))
 475                 return;
 476
 477         if (sk->sk_state == TCP_TIME_WAIT)
 478                 inet_twsk_free(inet_twsk(sk));
 479         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 480                 reqsk_free(inet_reqsk(sk));
 481         else
 482                 sk_free(sk);
 483 }
 484 EXPORT_SYMBOL_GPL(sock_gen_put);
 485
 486 void sock_edemux(struct sk_buff *skb)
 487 {
 488         sock_gen_put(skb->sk);
 489 }
 490 EXPORT_SYMBOL(sock_edemux);
 491
 492 struct sock *__inet_lookup_established(const struct net *net,
 493                                   struct inet_hashinfo *hashinfo,
 494                                   const __be32 saddr, const __be16 sport,
 495                                   const __be32 daddr, const u16 hnum,
 496                                   const int dif, const int sdif)
 497 {
 498         INET_ADDR_COOKIE(acookie, saddr, daddr);
 499         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 500         struct sock *sk;
 501         const struct hlist_nulls_node *node;
 502         /* Optimize here for direct hit, only listening connections can
 503          * have wildcards anyways.
 504          */
 505         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 506         unsigned int slot = hash & hashinfo->ehash_mask;
 507         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 508
 509 begin:
 510         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 511                 if (sk->sk_hash != hash)
 512                         continue;
 513                 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
 514                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 515                                 goto out;
 516                         if (unlikely(!inet_match(net, sk, acookie,
 517                                                  ports, dif, sdif))) {
 518                                 sock_gen_put(sk);
 519                                 goto begin;
 520                         }
 521                         goto found;
 522                 }
 523         }
 524         /*
 525          * if the nulls value we got at the end of this lookup is
 526          * not the expected one, we must restart lookup.
 527          * We probably met an item that was moved to another chain.
 528          */
 529         if (get_nulls_value(node) != slot)
 530                 goto begin;
 531 out:
 532         sk = NULL;
 533 found:
 534         return sk;
 535 }
 536 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 537
 538 /* called with local bh disabled */
 539 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 540                                     struct sock *sk, __u16 lport,
 541                                     struct inet_timewait_sock **twp,
 542                                     bool rcu_lookup,
 543                                     u32 hash)
 544 {
 545         struct inet_hashinfo *hinfo = death_row->hashinfo;
 546         struct inet_sock *inet = inet_sk(sk);
 547         __be32 daddr = inet->inet_rcv_saddr;
 548         __be32 saddr = inet->inet_daddr;
 549         int dif = sk->sk_bound_dev_if;
 550         struct net *net = sock_net(sk);
 551         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 552         INET_ADDR_COOKIE(acookie, saddr, daddr);
 553         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 554         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 555         struct inet_timewait_sock *tw = NULL;
 556         const struct hlist_nulls_node *node;
 557         struct sock *sk2;
 558         spinlock_t *lock;
 559
 560         if (rcu_lookup) {
 561                 sk_nulls_for_each(sk2, node, &head->chain) {
 562                         if (sk2->sk_hash != hash ||
 563                             !inet_match(net, sk2, acookie, ports, dif, sdif))
 564                                 continue;
 565                         if (sk2->sk_state == TCP_TIME_WAIT)
 566                                 break;
 567                         return -EADDRNOTAVAIL;
 568                 }
 569                 return 0;
 570         }
 571
 572         lock = inet_ehash_lockp(hinfo, hash);
 573         spin_lock(lock);
 574
 575         sk_nulls_for_each(sk2, node, &head->chain) {
 576                 if (sk2->sk_hash != hash)
 577                         continue;
 578
 579                 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
 580                         if (sk2->sk_state == TCP_TIME_WAIT) {
 581                                 tw = inet_twsk(sk2);
 582                                 if (sk->sk_protocol == IPPROTO_TCP &&
 583                                     tcp_twsk_unique(sk, sk2, twp))
 584                                         break;
 585                         }
 586                         goto not_unique;
 587                 }
 588         }
 589
 590         /* Must record num and sport now. Otherwise we will see
 591          * in hash table socket with a funny identity.
 592          */
 593         inet->inet_num = lport;
 594         inet->inet_sport = htons(lport);
 595         sk->sk_hash = hash;
 596         WARN_ON(!sk_unhashed(sk));
 597         __sk_nulls_add_node_rcu(sk, &head->chain);
 598         if (tw) {
 599                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 600                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 601         }
 602         spin_unlock(lock);
 603         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 604
 605         if (twp) {
 606                 *twp = tw;
 607         } else if (tw) {
 608                 /* Silly. Should hash-dance instead... */
 609                 inet_twsk_deschedule_put(tw);
 610         }
 611         return 0;
 612
 613 not_unique:
 614         spin_unlock(lock);
 615         return -EADDRNOTAVAIL;
 616 }
 617
 618 static u64 inet_sk_port_offset(const struct sock *sk)
 619 {
 620         const struct inet_sock *inet = inet_sk(sk);
 621
 622         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 623                                           inet->inet_daddr,
 624                                           inet->inet_dport);
 625 }
 626
 627 /* Searches for an exsiting socket in the ehash bucket list.
 628  * Returns true if found, false otherwise.
 629  */
 630 static bool inet_ehash_lookup_by_sk(struct sock *sk,
 631                                     struct hlist_nulls_head *list)
 632 {
 633         const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
 634         const int sdif = sk->sk_bound_dev_if;
 635         const int dif = sk->sk_bound_dev_if;
 636         const struct hlist_nulls_node *node;
 637         struct net *net = sock_net(sk);
 638         struct sock *esk;
 639
 640         INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
 641
 642         sk_nulls_for_each_rcu(esk, node, list) {
 643                 if (esk->sk_hash != sk->sk_hash)
 644                         continue;
 645                 if (sk->sk_family == AF_INET) {
 646                         if (unlikely(inet_match(net, esk, acookie,
 647                                                 ports, dif, sdif))) {
 648                                 return true;
 649                         }
 650                 }
 651 #if IS_ENABLED(CONFIG_IPV6)
 652                 else if (sk->sk_family == AF_INET6) {
 653                         if (unlikely(inet6_match(net, esk,
 654                                                  &sk->sk_v6_daddr,
 655                                                  &sk->sk_v6_rcv_saddr,
 656                                                  ports, dif, sdif))) {
 657                                 return true;
 658                         }
 659                 }
 660 #endif
 661         }
 662         return false;
 663 }
 664
 665 /* Insert a socket into ehash, and eventually remove another one
 666  * (The another one can be a SYN_RECV or TIMEWAIT)
 667  * If an existing socket already exists, socket sk is not inserted,
 668  * and sets found_dup_sk parameter to true.
 669  */
 670 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 671 {
 672         struct inet_hashinfo *hashinfo = tcp_get_hashinfo(sk);
 673         struct inet_ehash_bucket *head;
 674         struct hlist_nulls_head *list;
 675         spinlock_t *lock;
 676         bool ret = true;
 677
 678         WARN_ON_ONCE(!sk_unhashed(sk));
 679
 680         sk->sk_hash = sk_ehashfn(sk);
 681         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 682         list = &head->chain;
 683         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 684
 685         spin_lock(lock);
 686         if (osk) {
 687                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 688                 ret = sk_nulls_del_node_init_rcu(osk);
 689         } else if (found_dup_sk) {
 690                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 691                 if (*found_dup_sk)
 692                         ret = false;
 693         }
 694
 695         if (ret)
 696                 __sk_nulls_add_node_rcu(sk, list);
 697
 698         spin_unlock(lock);
 699
 700         return ret;
 701 }
 702
 703 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 704 {
 705         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 706
 707         if (ok) {
 708                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 709         } else {
 710                 this_cpu_inc(*sk->sk_prot->orphan_count);
 711                 inet_sk_set_state(sk, TCP_CLOSE);
 712                 sock_set_flag(sk, SOCK_DEAD);
 713                 inet_csk_destroy_sock(sk);
 714         }
 715         return ok;
 716 }
 717 EXPORT_IPV6_MOD(inet_ehash_nolisten);
 718
 719 static int inet_reuseport_add_sock(struct sock *sk,
 720                                    struct inet_listen_hashbucket *ilb)
 721 {
 722         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 723         const struct hlist_nulls_node *node;
 724         struct sock *sk2;
 725         kuid_t uid = sock_i_uid(sk);
 726
 727         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 728                 if (sk2 != sk &&
 729                     sk2->sk_family == sk->sk_family &&
 730                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 731                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 732                     inet_csk(sk2)->icsk_bind_hash == tb &&
 733                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 734                     inet_rcv_saddr_equal(sk, sk2, false))
 735                         return reuseport_add_sock(sk, sk2,
 736                                                   inet_rcv_saddr_any(sk));
 737         }
 738
 739         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 740 }
 741
 742 int __inet_hash(struct sock *sk, struct sock *osk)
 743 {
 744         struct inet_hashinfo *hashinfo = tcp_get_hashinfo(sk);
 745         struct inet_listen_hashbucket *ilb2;
 746         int err = 0;
 747
 748         if (sk->sk_state != TCP_LISTEN) {
 749                 local_bh_disable();
 750                 inet_ehash_nolisten(sk, osk, NULL);
 751                 local_bh_enable();
 752                 return 0;
 753         }
 754         WARN_ON(!sk_unhashed(sk));
 755         ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 756
 757         spin_lock(&ilb2->lock);
 758         if (sk->sk_reuseport) {
 759                 err = inet_reuseport_add_sock(sk, ilb2);
 760                 if (err)
 761                         goto unlock;
 762         }
 763         sock_set_flag(sk, SOCK_RCU_FREE);
 764         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 765                 sk->sk_family == AF_INET6)
 766                 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
 767         else
 768                 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
 769         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 770 unlock:
 771         spin_unlock(&ilb2->lock);
 772
 773         return err;
 774 }
 775 EXPORT_IPV6_MOD(__inet_hash);
 776
 777 int inet_hash(struct sock *sk)
 778 {
 779         int err = 0;
 780
 781         if (sk->sk_state != TCP_CLOSE)
 782                 err = __inet_hash(sk, NULL);
 783
 784         return err;
 785 }
 786
 787 void inet_unhash(struct sock *sk)
 788 {
 789         struct inet_hashinfo *hashinfo = tcp_get_hashinfo(sk);
 790
 791         if (sk_unhashed(sk))
 792                 return;
 793
 794         sock_rps_delete_flow(sk);
 795         if (sk->sk_state == TCP_LISTEN) {
 796                 struct inet_listen_hashbucket *ilb2;
 797
 798                 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 799                 /* Don't disable bottom halves while acquiring the lock to
 800                  * avoid circular locking dependency on PREEMPT_RT.
 801                  */
 802                 spin_lock(&ilb2->lock);
 803                 if (sk_unhashed(sk)) {
 804                         spin_unlock(&ilb2->lock);
 805                         return;
 806                 }
 807
 808                 if (rcu_access_pointer(sk->sk_reuseport_cb))
 809                         reuseport_stop_listen_sock(sk);
 810
 811                 __sk_nulls_del_node_init_rcu(sk);
 812                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 813                 spin_unlock(&ilb2->lock);
 814         } else {
 815                 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 816
 817                 spin_lock_bh(lock);
 818                 if (sk_unhashed(sk)) {
 819                         spin_unlock_bh(lock);
 820                         return;
 821                 }
 822                 __sk_nulls_del_node_init_rcu(sk);
 823                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 824                 spin_unlock_bh(lock);
 825         }
 826 }
 827 EXPORT_IPV6_MOD(inet_unhash);
 828
 829 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
 830                                     const struct net *net, unsigned short port,
 831                                     int l3mdev, const struct sock *sk)
 832 {
 833         if (!net_eq(ib2_net(tb), net) || tb->port != port ||
 834             tb->l3mdev != l3mdev)
 835                 return false;
 836
 837         return inet_bind2_bucket_addr_match(tb, sk);
 838 }
 839
 840 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
 841                                       unsigned short port, int l3mdev, const struct sock *sk)
 842 {
 843         if (!net_eq(ib2_net(tb), net) || tb->port != port ||
 844             tb->l3mdev != l3mdev)
 845                 return false;
 846
 847 #if IS_ENABLED(CONFIG_IPV6)
 848         if (tb->addr_type == IPV6_ADDR_ANY)
 849                 return true;
 850
 851         if (tb->addr_type != IPV6_ADDR_MAPPED)
 852                 return false;
 853
 854         if (sk->sk_family == AF_INET6 &&
 855             !ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr))
 856                 return false;
 857 #endif
 858         return tb->rcv_saddr == 0;
 859 }
 860
 861 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
 862 struct inet_bind2_bucket *
 863 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
 864                        unsigned short port, int l3mdev, const struct sock *sk)
 865 {
 866         struct inet_bind2_bucket *bhash2 = NULL;
 867
 868         inet_bind_bucket_for_each(bhash2, &head->chain)
 869                 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
 870                         break;
 871
 872         return bhash2;
 873 }
 874
 875 struct inet_bind_hashbucket *
 876 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
 877 {
 878         struct inet_hashinfo *hinfo = tcp_get_hashinfo(sk);
 879         u32 hash;
 880
 881 #if IS_ENABLED(CONFIG_IPV6)
 882         if (sk->sk_family == AF_INET6)
 883                 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
 884         else
 885 #endif
 886                 hash = ipv4_portaddr_hash(net, 0, port);
 887
 888         return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
 889 }
 890
 891 static void inet_update_saddr(struct sock *sk, void *saddr, int family)
 892 {
 893         if (family == AF_INET) {
 894                 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
 895                 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
 896         }
 897 #if IS_ENABLED(CONFIG_IPV6)
 898         else {
 899                 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
 900         }
 901 #endif
 902 }
 903
 904 static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
 905 {
 906         struct inet_hashinfo *hinfo = tcp_get_hashinfo(sk);
 907         struct inet_bind_hashbucket *head, *head2;
 908         struct inet_bind2_bucket *tb2, *new_tb2;
 909         int l3mdev = inet_sk_bound_l3mdev(sk);
 910         int port = inet_sk(sk)->inet_num;
 911         struct net *net = sock_net(sk);
 912         int bhash;
 913
 914         if (!inet_csk(sk)->icsk_bind2_hash) {
 915                 /* Not bind()ed before. */
 916                 if (reset)
 917                         inet_reset_saddr(sk);
 918                 else
 919                         inet_update_saddr(sk, saddr, family);
 920
 921                 return 0;
 922         }
 923
 924         /* Allocate a bind2 bucket ahead of time to avoid permanently putting
 925          * the bhash2 table in an inconsistent state if a new tb2 bucket
 926          * allocation fails.
 927          */
 928         new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
 929         if (!new_tb2) {
 930                 if (reset) {
 931                         /* The (INADDR_ANY, port) bucket might have already
 932                          * been freed, then we cannot fixup icsk_bind2_hash,
 933                          * so we give up and unlink sk from bhash/bhash2 not
 934                          * to leave inconsistency in bhash2.
 935                          */
 936                         inet_put_port(sk);
 937                         inet_reset_saddr(sk);
 938                 }
 939
 940                 return -ENOMEM;
 941         }
 942
 943         bhash = inet_bhashfn(net, port, hinfo->bhash_size);
 944         head = &hinfo->bhash[bhash];
 945         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 946
 947         /* If we change saddr locklessly, another thread
 948          * iterating over bhash might see corrupted address.
 949          */
 950         spin_lock_bh(&head->lock);
 951
 952         spin_lock(&head2->lock);
 953         __sk_del_bind_node(sk);
 954         inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
 955         spin_unlock(&head2->lock);
 956
 957         if (reset)
 958                 inet_reset_saddr(sk);
 959         else
 960                 inet_update_saddr(sk, saddr, family);
 961
 962         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 963
 964         spin_lock(&head2->lock);
 965         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 966         if (!tb2) {
 967                 tb2 = new_tb2;
 968                 inet_bind2_bucket_init(tb2, net, head2, inet_csk(sk)->icsk_bind_hash, sk);
 969         }
 970         inet_csk(sk)->icsk_bind2_hash = tb2;
 971         sk_add_bind_node(sk, &tb2->owners);
 972         spin_unlock(&head2->lock);
 973
 974         spin_unlock_bh(&head->lock);
 975
 976         if (tb2 != new_tb2)
 977                 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
 978
 979         return 0;
 980 }
 981
 982 int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
 983 {
 984         return __inet_bhash2_update_saddr(sk, saddr, family, false);
 985 }
 986 EXPORT_IPV6_MOD(inet_bhash2_update_saddr);
 987
 988 void inet_bhash2_reset_saddr(struct sock *sk)
 989 {
 990         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
 991                 __inet_bhash2_update_saddr(sk, NULL, 0, true);
 992 }
 993 EXPORT_IPV6_MOD(inet_bhash2_reset_saddr);
 994
 995 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 996  * Note that we use 32bit integers (vs RFC 'short integers')
 997  * because 2^16 is not a multiple of num_ephemeral and this
 998  * property might be used by clever attacker.
 999  *
1000  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
1001  * attacks were since demonstrated, thus we use 65536 by default instead
1002  * to really give more isolation and privacy, at the expense of 256kB
1003  * of kernel memory.
1004  */
1005 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
1006 static u32 *table_perturb;
1007
1008 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
1009                 struct sock *sk, u64 port_offset,
1010                 u32 hash_port0,
1011                 int (*check_established)(struct inet_timewait_death_row *,
1012                         struct sock *, __u16, struct inet_timewait_sock **,
1013                         bool rcu_lookup, u32 hash))
1014 {
1015         struct inet_hashinfo *hinfo = death_row->hashinfo;
1016         struct inet_bind_hashbucket *head, *head2;
1017         struct inet_timewait_sock *tw = NULL;
1018         int port = inet_sk(sk)->inet_num;
1019         struct net *net = sock_net(sk);
1020         struct inet_bind2_bucket *tb2;
1021         struct inet_bind_bucket *tb;
1022         bool tb_created = false;
1023         u32 remaining, offset;
1024         int ret, i, low, high;
1025         bool local_ports;
1026         int step, l3mdev;
1027         u32 index;
1028
1029         if (port) {
1030                 local_bh_disable();
1031                 ret = check_established(death_row, sk, port, NULL, false,
1032                                         hash_port0 + port);
1033                 local_bh_enable();
1034                 return ret;
1035         }
1036
1037         l3mdev = inet_sk_bound_l3mdev(sk);
1038
1039         local_ports = inet_sk_get_local_port_range(sk, &low, &high);
1040         step = local_ports ? 1 : 2;
1041
1042         high++; /* [32768, 60999] -> [32768, 61000[ */
1043         remaining = high - low;
1044         if (!local_ports && remaining > 1)
1045                 remaining &= ~1U;
1046
1047         get_random_sleepable_once(table_perturb,
1048                                   INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1049         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1050
1051         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1052         offset %= remaining;
1053
1054         /* In first pass we try ports of @low parity.
1055          * inet_csk_get_port() does the opposite choice.
1056          */
1057         if (!local_ports)
1058                 offset &= ~1U;
1059 other_parity_scan:
1060         port = low + offset;
1061         for (i = 0; i < remaining; i += step, port += step) {
1062                 if (unlikely(port >= high))
1063                         port -= remaining;
1064                 if (inet_is_local_reserved_port(net, port))
1065                         continue;
1066                 head = &hinfo->bhash[inet_bhashfn(net, port,
1067                                                   hinfo->bhash_size)];
1068                 rcu_read_lock();
1069                 hlist_for_each_entry_rcu(tb, &head->chain, node) {
1070                         if (!inet_bind_bucket_match(tb, net, port, l3mdev))
1071                                 continue;
1072                         if (tb->fastreuse >= 0 || tb->fastreuseport >= 0) {
1073                                 rcu_read_unlock();
1074                                 goto next_port;
1075                         }
1076                         if (!check_established(death_row, sk, port, &tw, true,
1077                                                hash_port0 + port))
1078                                 break;
1079                         rcu_read_unlock();
1080                         goto next_port;
1081                 }
1082                 rcu_read_unlock();
1083
1084                 spin_lock_bh(&head->lock);
1085
1086                 /* Does not bother with rcv_saddr checks, because
1087                  * the established check is already unique enough.
1088                  */
1089                 inet_bind_bucket_for_each(tb, &head->chain) {
1090                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1091                                 if (tb->fastreuse >= 0 ||
1092                                     tb->fastreuseport >= 0)
1093                                         goto next_port_unlock;
1094                                 WARN_ON(hlist_empty(&tb->bhash2));
1095                                 if (!check_established(death_row, sk,
1096                                                        port, &tw, false,
1097                                                        hash_port0 + port))
1098                                         goto ok;
1099                                 goto next_port_unlock;
1100                         }
1101                 }
1102
1103                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1104                                              net, head, port, l3mdev);
1105                 if (!tb) {
1106                         spin_unlock_bh(&head->lock);
1107                         return -ENOMEM;
1108                 }
1109                 tb_created = true;
1110                 tb->fastreuse = -1;
1111                 tb->fastreuseport = -1;
1112                 goto ok;
1113 next_port_unlock:
1114                 spin_unlock_bh(&head->lock);
1115 next_port:
1116                 cond_resched();
1117         }
1118
1119         if (!local_ports) {
1120                 offset++;
1121                 if ((offset & 1) && remaining > 1)
1122                         goto other_parity_scan;
1123         }
1124         return -EADDRNOTAVAIL;
1125
1126 ok:
1127         /* Find the corresponding tb2 bucket since we need to
1128          * add the socket to the bhash2 table as well
1129          */
1130         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1131         spin_lock(&head2->lock);
1132
1133         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1134         if (!tb2) {
1135                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1136                                                head2, tb, sk);
1137                 if (!tb2)
1138                         goto error;
1139         }
1140
1141         /* Here we want to add a little bit of randomness to the next source
1142          * port that will be chosen. We use a max() with a random here so that
1143          * on low contention the randomness is maximal and on high contention
1144          * it may be inexistent.
1145          */
1146         i = max_t(int, i, get_random_u32_below(8) * step);
1147         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + step);
1148
1149         /* Head lock still held and bh's disabled */
1150         inet_bind_hash(sk, tb, tb2, port);
1151
1152         if (sk_unhashed(sk)) {
1153                 inet_sk(sk)->inet_sport = htons(port);
1154                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1155         }
1156         if (tw)
1157                 inet_twsk_bind_unhash(tw, hinfo);
1158
1159         spin_unlock(&head2->lock);
1160         spin_unlock(&head->lock);
1161
1162         if (tw)
1163                 inet_twsk_deschedule_put(tw);
1164         local_bh_enable();
1165         return 0;
1166
1167 error:
1168         if (sk_hashed(sk)) {
1169                 spinlock_t *lock = inet_ehash_lockp(hinfo, sk->sk_hash);
1170
1171                 sock_prot_inuse_add(net, sk->sk_prot, -1);
1172
1173                 spin_lock(lock);
1174                 __sk_nulls_del_node_init_rcu(sk);
1175                 spin_unlock(lock);
1176
1177                 sk->sk_hash = 0;
1178                 inet_sk(sk)->inet_sport = 0;
1179                 inet_sk(sk)->inet_num = 0;
1180
1181                 if (tw)
1182                         inet_twsk_bind_unhash(tw, hinfo);
1183         }
1184
1185         spin_unlock(&head2->lock);
1186         if (tb_created)
1187                 inet_bind_bucket_destroy(tb);
1188         spin_unlock(&head->lock);
1189
1190         if (tw)
1191                 inet_twsk_deschedule_put(tw);
1192
1193         local_bh_enable();
1194
1195         return -ENOMEM;
1196 }
1197
1198 /*
1199  * Bind a port for a connect operation and hash it.
1200  */
1201 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1202                       struct sock *sk)
1203 {
1204         const struct inet_sock *inet = inet_sk(sk);
1205         const struct net *net = sock_net(sk);
1206         u64 port_offset = 0;
1207         u32 hash_port0;
1208
1209         if (!inet_sk(sk)->inet_num)
1210                 port_offset = inet_sk_port_offset(sk);
1211
1212         hash_port0 = inet_ehashfn(net, inet->inet_rcv_saddr, 0,
1213                                   inet->inet_daddr, inet->inet_dport);
1214
1215         return __inet_hash_connect(death_row, sk, port_offset, hash_port0,
1216                                    __inet_check_established);
1217 }
1218
1219 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1220 {
1221         int i;
1222
1223         for (i = 0; i <= h->lhash2_mask; i++) {
1224                 spin_lock_init(&h->lhash2[i].lock);
1225                 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1226                                       i + LISTENING_NULLS_BASE);
1227         }
1228 }
1229
1230 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1231                                 unsigned long numentries, int scale,
1232                                 unsigned long low_limit,
1233                                 unsigned long high_limit)
1234 {
1235         h->lhash2 = alloc_large_system_hash(name,
1236                                             sizeof(*h->lhash2),
1237                                             numentries,
1238                                             scale,
1239                                             0,
1240                                             NULL,
1241                                             &h->lhash2_mask,
1242                                             low_limit,
1243                                             high_limit);
1244         init_hashinfo_lhash2(h);
1245
1246         /* this one is used for source ports of outgoing connections */
1247         table_perturb = alloc_large_system_hash("Table-perturb",
1248                                                 sizeof(*table_perturb),
1249                                                 INET_TABLE_PERTURB_SIZE,
1250                                                 0, 0, NULL, NULL,
1251                                                 INET_TABLE_PERTURB_SIZE,
1252                                                 INET_TABLE_PERTURB_SIZE);
1253 }
1254
1255 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1256 {
1257         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1258         if (!h->lhash2)
1259                 return -ENOMEM;
1260
1261         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1262         /* INET_LHTABLE_SIZE must be a power of 2 */
1263         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1264
1265         init_hashinfo_lhash2(h);
1266         return 0;
1267 }
1268
1269 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1270 {
1271         unsigned int locksz = sizeof(spinlock_t);
1272         unsigned int i, nblocks = 1;
1273         spinlock_t *ptr = NULL;
1274
1275         if (locksz == 0)
1276                 goto set_mask;
1277
1278         /* Allocate 2 cache lines or at least one spinlock per cpu. */
1279         nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U) * num_possible_cpus();
1280
1281         /* At least one page per NUMA node. */
1282         nblocks = max(nblocks, num_online_nodes() * PAGE_SIZE / locksz);
1283
1284         nblocks = roundup_pow_of_two(nblocks);
1285
1286         /* No more locks than number of hash buckets. */
1287         nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1288
1289         if (num_online_nodes() > 1) {
1290                 /* Use vmalloc() to allow NUMA policy to spread pages
1291                  * on all available nodes if desired.
1292                  */
1293                 ptr = vmalloc_array(nblocks, locksz);
1294         }
1295         if (!ptr) {
1296                 ptr = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1297                 if (!ptr)
1298                         return -ENOMEM;
1299         }
1300         for (i = 0; i < nblocks; i++)
1301                 spin_lock_init(&ptr[i]);
1302         hashinfo->ehash_locks = ptr;
1303 set_mask:
1304         hashinfo->ehash_locks_mask = nblocks - 1;
1305         return 0;
1306 }
1307
1308 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1309                                                  unsigned int ehash_entries)
1310 {
1311         struct inet_hashinfo *new_hashinfo;
1312         int i;
1313
1314         new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1315         if (!new_hashinfo)
1316                 goto err;
1317
1318         new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1319                                            GFP_KERNEL_ACCOUNT);
1320         if (!new_hashinfo->ehash)
1321                 goto free_hashinfo;
1322
1323         new_hashinfo->ehash_mask = ehash_entries - 1;
1324
1325         if (inet_ehash_locks_alloc(new_hashinfo))
1326                 goto free_ehash;
1327
1328         for (i = 0; i < ehash_entries; i++)
1329                 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1330
1331         new_hashinfo->pernet = true;
1332
1333         return new_hashinfo;
1334
1335 free_ehash:
1336         vfree(new_hashinfo->ehash);
1337 free_hashinfo:
1338         kfree(new_hashinfo);
1339 err:
1340         return NULL;
1341 }
1342
1343 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1344 {
1345         if (!hashinfo->pernet)
1346                 return;
1347
1348         inet_ehash_locks_free(hashinfo);
1349         vfree(hashinfo->ehash);
1350         kfree(hashinfo);
1351 }