net/ipv4/inet_hashtables.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Generic INET transport hashtables
   8  *
   9  * Authors:     Lotsa people, from code originally in tcp
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/random.h>
  14 #include <linux/sched.h>
  15 #include <linux/slab.h>
  16 #include <linux/wait.h>
  17 #include <linux/vmalloc.h>
  18 #include <linux/memblock.h>
  19
  20 #include <net/addrconf.h>
  21 #include <net/inet_connection_sock.h>
  22 #include <net/inet_hashtables.h>
  23 #if IS_ENABLED(CONFIG_IPV6)
  24 #include <net/inet6_hashtables.h>
  25 #endif
  26 #include <net/secure_seq.h>
  27 #include <net/ip.h>
  28 #include <net/tcp.h>
  29 #include <net/sock_reuseport.h>
  30
  31 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
  32                         const __u16 lport, const __be32 faddr,
  33                         const __be16 fport)
  34 {
  35         static u32 inet_ehash_secret __read_mostly;
  36
  37         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  38
  39         return __inet_ehashfn(laddr, lport, faddr, fport,
  40                               inet_ehash_secret + net_hash_mix(net));
  41 }
  42
  43 /* This function handles inet_sock, but also timewait and request sockets
  44  * for IPv4/IPv6.
  45  */
  46 static u32 sk_ehashfn(const struct sock *sk)
  47 {
  48 #if IS_ENABLED(CONFIG_IPV6)
  49         if (sk->sk_family == AF_INET6 &&
  50             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  51                 return inet6_ehashfn(sock_net(sk),
  52                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  53                                      &sk->sk_v6_daddr, sk->sk_dport);
  54 #endif
  55         return inet_ehashfn(sock_net(sk),
  56                             sk->sk_rcv_saddr, sk->sk_num,
  57                             sk->sk_daddr, sk->sk_dport);
  58 }
  59
  60 /*
  61  * Allocate and initialize a new local port bind bucket.
  62  * The bindhash mutex for snum's hash chain must be held here.
  63  */
  64 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  65                                                  struct net *net,
  66                                                  struct inet_bind_hashbucket *head,
  67                                                  const unsigned short snum,
  68                                                  int l3mdev)
  69 {
  70         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  71
  72         if (tb) {
  73                 write_pnet(&tb->ib_net, net);
  74                 tb->l3mdev    = l3mdev;
  75                 tb->port      = snum;
  76                 tb->fastreuse = 0;
  77                 tb->fastreuseport = 0;
  78                 INIT_HLIST_HEAD(&tb->owners);
  79                 hlist_add_head(&tb->node, &head->chain);
  80         }
  81         return tb;
  82 }
  83
  84 /*
  85  * Caller must hold hashbucket lock for this tb with local BH disabled
  86  */
  87 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  88 {
  89         if (hlist_empty(&tb->owners)) {
  90                 __hlist_del(&tb->node);
  91                 kmem_cache_free(cachep, tb);
  92         }
  93 }
  94
  95 bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
  96                             unsigned short port, int l3mdev)
  97 {
  98         return net_eq(ib_net(tb), net) && tb->port == port &&
  99                 tb->l3mdev == l3mdev;
 100 }
 101
 102 static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb,
 103                                    struct net *net,
 104                                    struct inet_bind_hashbucket *head,
 105                                    unsigned short port, int l3mdev,
 106                                    const struct sock *sk)
 107 {
 108         write_pnet(&tb->ib_net, net);
 109         tb->l3mdev    = l3mdev;
 110         tb->port      = port;
 111 #if IS_ENABLED(CONFIG_IPV6)
 112         tb->family    = sk->sk_family;
 113         if (sk->sk_family == AF_INET6)
 114                 tb->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 115         else
 116 #endif
 117                 tb->rcv_saddr = sk->sk_rcv_saddr;
 118         INIT_HLIST_HEAD(&tb->owners);
 119         INIT_HLIST_HEAD(&tb->deathrow);
 120         hlist_add_head(&tb->node, &head->chain);
 121 }
 122
 123 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
 124                                                    struct net *net,
 125                                                    struct inet_bind_hashbucket *head,
 126                                                    unsigned short port,
 127                                                    int l3mdev,
 128                                                    const struct sock *sk)
 129 {
 130         struct inet_bind2_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
 131
 132         if (tb)
 133                 inet_bind2_bucket_init(tb, net, head, port, l3mdev, sk);
 134
 135         return tb;
 136 }
 137
 138 /* Caller must hold hashbucket lock for this tb with local BH disabled */
 139 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
 140 {
 141         if (hlist_empty(&tb->owners) && hlist_empty(&tb->deathrow)) {
 142                 __hlist_del(&tb->node);
 143                 kmem_cache_free(cachep, tb);
 144         }
 145 }
 146
 147 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
 148                                          const struct sock *sk)
 149 {
 150 #if IS_ENABLED(CONFIG_IPV6)
 151         if (sk->sk_family != tb2->family)
 152                 return false;
 153
 154         if (sk->sk_family == AF_INET6)
 155                 return ipv6_addr_equal(&tb2->v6_rcv_saddr,
 156                                        &sk->sk_v6_rcv_saddr);
 157 #endif
 158         return tb2->rcv_saddr == sk->sk_rcv_saddr;
 159 }
 160
 161 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 162                     struct inet_bind2_bucket *tb2, unsigned short port)
 163 {
 164         inet_sk(sk)->inet_num = port;
 165         sk_add_bind_node(sk, &tb->owners);
 166         inet_csk(sk)->icsk_bind_hash = tb;
 167         sk_add_bind2_node(sk, &tb2->owners);
 168         inet_csk(sk)->icsk_bind2_hash = tb2;
 169 }
 170
 171 /*
 172  * Get rid of any references to a local port held by the given sock.
 173  */
 174 static void __inet_put_port(struct sock *sk)
 175 {
 176         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 177         struct inet_bind_hashbucket *head, *head2;
 178         struct net *net = sock_net(sk);
 179         struct inet_bind_bucket *tb;
 180         int bhash;
 181
 182         bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
 183         head = &hashinfo->bhash[bhash];
 184         head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
 185
 186         spin_lock(&head->lock);
 187         tb = inet_csk(sk)->icsk_bind_hash;
 188         __sk_del_bind_node(sk);
 189         inet_csk(sk)->icsk_bind_hash = NULL;
 190         inet_sk(sk)->inet_num = 0;
 191         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 192
 193         spin_lock(&head2->lock);
 194         if (inet_csk(sk)->icsk_bind2_hash) {
 195                 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
 196
 197                 __sk_del_bind2_node(sk);
 198                 inet_csk(sk)->icsk_bind2_hash = NULL;
 199                 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
 200         }
 201         spin_unlock(&head2->lock);
 202
 203         spin_unlock(&head->lock);
 204 }
 205
 206 void inet_put_port(struct sock *sk)
 207 {
 208         local_bh_disable();
 209         __inet_put_port(sk);
 210         local_bh_enable();
 211 }
 212 EXPORT_SYMBOL(inet_put_port);
 213
 214 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 215 {
 216         struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
 217         unsigned short port = inet_sk(child)->inet_num;
 218         struct inet_bind_hashbucket *head, *head2;
 219         bool created_inet_bind_bucket = false;
 220         struct net *net = sock_net(sk);
 221         bool update_fastreuse = false;
 222         struct inet_bind2_bucket *tb2;
 223         struct inet_bind_bucket *tb;
 224         int bhash, l3mdev;
 225
 226         bhash = inet_bhashfn(net, port, table->bhash_size);
 227         head = &table->bhash[bhash];
 228         head2 = inet_bhashfn_portaddr(table, child, net, port);
 229
 230         spin_lock(&head->lock);
 231         spin_lock(&head2->lock);
 232         tb = inet_csk(sk)->icsk_bind_hash;
 233         tb2 = inet_csk(sk)->icsk_bind2_hash;
 234         if (unlikely(!tb || !tb2)) {
 235                 spin_unlock(&head2->lock);
 236                 spin_unlock(&head->lock);
 237                 return -ENOENT;
 238         }
 239         if (tb->port != port) {
 240                 l3mdev = inet_sk_bound_l3mdev(sk);
 241
 242                 /* NOTE: using tproxy and redirecting skbs to a proxy
 243                  * on a different listener port breaks the assumption
 244                  * that the listener socket's icsk_bind_hash is the same
 245                  * as that of the child socket. We have to look up or
 246                  * create a new bind bucket for the child here. */
 247                 inet_bind_bucket_for_each(tb, &head->chain) {
 248                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 249                                 break;
 250                 }
 251                 if (!tb) {
 252                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 253                                                      net, head, port, l3mdev);
 254                         if (!tb) {
 255                                 spin_unlock(&head2->lock);
 256                                 spin_unlock(&head->lock);
 257                                 return -ENOMEM;
 258                         }
 259                         created_inet_bind_bucket = true;
 260                 }
 261                 update_fastreuse = true;
 262
 263                 goto bhash2_find;
 264         } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
 265                 l3mdev = inet_sk_bound_l3mdev(sk);
 266
 267 bhash2_find:
 268                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
 269                 if (!tb2) {
 270                         tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
 271                                                        net, head2, port,
 272                                                        l3mdev, child);
 273                         if (!tb2)
 274                                 goto error;
 275                 }
 276         }
 277         if (update_fastreuse)
 278                 inet_csk_update_fastreuse(tb, child);
 279         inet_bind_hash(child, tb, tb2, port);
 280         spin_unlock(&head2->lock);
 281         spin_unlock(&head->lock);
 282
 283         return 0;
 284
 285 error:
 286         if (created_inet_bind_bucket)
 287                 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
 288         spin_unlock(&head2->lock);
 289         spin_unlock(&head->lock);
 290         return -ENOMEM;
 291 }
 292 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 293
 294 static struct inet_listen_hashbucket *
 295 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 296 {
 297         u32 hash;
 298
 299 #if IS_ENABLED(CONFIG_IPV6)
 300         if (sk->sk_family == AF_INET6)
 301                 hash = ipv6_portaddr_hash(sock_net(sk),
 302                                           &sk->sk_v6_rcv_saddr,
 303                                           inet_sk(sk)->inet_num);
 304         else
 305 #endif
 306                 hash = ipv4_portaddr_hash(sock_net(sk),
 307                                           inet_sk(sk)->inet_rcv_saddr,
 308                                           inet_sk(sk)->inet_num);
 309         return inet_lhash2_bucket(h, hash);
 310 }
 311
 312 static inline int compute_score(struct sock *sk, struct net *net,
 313                                 const unsigned short hnum, const __be32 daddr,
 314                                 const int dif, const int sdif)
 315 {
 316         int score = -1;
 317
 318         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 319                         !ipv6_only_sock(sk)) {
 320                 if (sk->sk_rcv_saddr != daddr)
 321                         return -1;
 322
 323                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 324                         return -1;
 325                 score =  sk->sk_bound_dev_if ? 2 : 1;
 326
 327                 if (sk->sk_family == PF_INET)
 328                         score++;
 329                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 330                         score++;
 331         }
 332         return score;
 333 }
 334
 335 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
 336                                             struct sk_buff *skb, int doff,
 337                                             __be32 saddr, __be16 sport,
 338                                             __be32 daddr, unsigned short hnum)
 339 {
 340         struct sock *reuse_sk = NULL;
 341         u32 phash;
 342
 343         if (sk->sk_reuseport) {
 344                 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 345                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 346         }
 347         return reuse_sk;
 348 }
 349
 350 /*
 351  * Here are some nice properties to exploit here. The BSD API
 352  * does not allow a listening sock to specify the remote port nor the
 353  * remote address for the connection. So always assume those are both
 354  * wildcarded during the search since they can never be otherwise.
 355  */
 356
 357 /* called with rcu_read_lock() : No refcount taken on the socket */
 358 static struct sock *inet_lhash2_lookup(struct net *net,
 359                                 struct inet_listen_hashbucket *ilb2,
 360                                 struct sk_buff *skb, int doff,
 361                                 const __be32 saddr, __be16 sport,
 362                                 const __be32 daddr, const unsigned short hnum,
 363                                 const int dif, const int sdif)
 364 {
 365         struct sock *sk, *result = NULL;
 366         struct hlist_nulls_node *node;
 367         int score, hiscore = 0;
 368
 369         sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
 370                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 371                 if (score > hiscore) {
 372                         result = lookup_reuseport(net, sk, skb, doff,
 373                                                   saddr, sport, daddr, hnum);
 374                         if (result)
 375                                 return result;
 376
 377                         result = sk;
 378                         hiscore = score;
 379                 }
 380         }
 381
 382         return result;
 383 }
 384
 385 static inline struct sock *inet_lookup_run_bpf(struct net *net,
 386                                                struct inet_hashinfo *hashinfo,
 387                                                struct sk_buff *skb, int doff,
 388                                                __be32 saddr, __be16 sport,
 389                                                __be32 daddr, u16 hnum, const int dif)
 390 {
 391         struct sock *sk, *reuse_sk;
 392         bool no_reuseport;
 393
 394         if (hashinfo != net->ipv4.tcp_death_row.hashinfo)
 395                 return NULL; /* only TCP is supported */
 396
 397         no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP, saddr, sport,
 398                                             daddr, hnum, dif, &sk);
 399         if (no_reuseport || IS_ERR_OR_NULL(sk))
 400                 return sk;
 401
 402         reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
 403         if (reuse_sk)
 404                 sk = reuse_sk;
 405         return sk;
 406 }
 407
 408 struct sock *__inet_lookup_listener(struct net *net,
 409                                     struct inet_hashinfo *hashinfo,
 410                                     struct sk_buff *skb, int doff,
 411                                     const __be32 saddr, __be16 sport,
 412                                     const __be32 daddr, const unsigned short hnum,
 413                                     const int dif, const int sdif)
 414 {
 415         struct inet_listen_hashbucket *ilb2;
 416         struct sock *result = NULL;
 417         unsigned int hash2;
 418
 419         /* Lookup redirect from BPF */
 420         if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
 421                 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
 422                                              saddr, sport, daddr, hnum, dif);
 423                 if (result)
 424                         goto done;
 425         }
 426
 427         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 428         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 429
 430         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 431                                     saddr, sport, daddr, hnum,
 432                                     dif, sdif);
 433         if (result)
 434                 goto done;
 435
 436         /* Lookup lhash2 with INADDR_ANY */
 437         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 438         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 439
 440         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 441                                     saddr, sport, htonl(INADDR_ANY), hnum,
 442                                     dif, sdif);
 443 done:
 444         if (IS_ERR(result))
 445                 return NULL;
 446         return result;
 447 }
 448 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 449
 450 /* All sockets share common refcount, but have different destructors */
 451 void sock_gen_put(struct sock *sk)
 452 {
 453         if (!refcount_dec_and_test(&sk->sk_refcnt))
 454                 return;
 455
 456         if (sk->sk_state == TCP_TIME_WAIT)
 457                 inet_twsk_free(inet_twsk(sk));
 458         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 459                 reqsk_free(inet_reqsk(sk));
 460         else
 461                 sk_free(sk);
 462 }
 463 EXPORT_SYMBOL_GPL(sock_gen_put);
 464
 465 void sock_edemux(struct sk_buff *skb)
 466 {
 467         sock_gen_put(skb->sk);
 468 }
 469 EXPORT_SYMBOL(sock_edemux);
 470
 471 struct sock *__inet_lookup_established(struct net *net,
 472                                   struct inet_hashinfo *hashinfo,
 473                                   const __be32 saddr, const __be16 sport,
 474                                   const __be32 daddr, const u16 hnum,
 475                                   const int dif, const int sdif)
 476 {
 477         INET_ADDR_COOKIE(acookie, saddr, daddr);
 478         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 479         struct sock *sk;
 480         const struct hlist_nulls_node *node;
 481         /* Optimize here for direct hit, only listening connections can
 482          * have wildcards anyways.
 483          */
 484         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 485         unsigned int slot = hash & hashinfo->ehash_mask;
 486         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 487
 488 begin:
 489         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 490                 if (sk->sk_hash != hash)
 491                         continue;
 492                 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
 493                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 494                                 goto out;
 495                         if (unlikely(!inet_match(net, sk, acookie,
 496                                                  ports, dif, sdif))) {
 497                                 sock_gen_put(sk);
 498                                 goto begin;
 499                         }
 500                         goto found;
 501                 }
 502         }
 503         /*
 504          * if the nulls value we got at the end of this lookup is
 505          * not the expected one, we must restart lookup.
 506          * We probably met an item that was moved to another chain.
 507          */
 508         if (get_nulls_value(node) != slot)
 509                 goto begin;
 510 out:
 511         sk = NULL;
 512 found:
 513         return sk;
 514 }
 515 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 516
 517 /* called with local bh disabled */
 518 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 519                                     struct sock *sk, __u16 lport,
 520                                     struct inet_timewait_sock **twp)
 521 {
 522         struct inet_hashinfo *hinfo = death_row->hashinfo;
 523         struct inet_sock *inet = inet_sk(sk);
 524         __be32 daddr = inet->inet_rcv_saddr;
 525         __be32 saddr = inet->inet_daddr;
 526         int dif = sk->sk_bound_dev_if;
 527         struct net *net = sock_net(sk);
 528         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 529         INET_ADDR_COOKIE(acookie, saddr, daddr);
 530         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 531         unsigned int hash = inet_ehashfn(net, daddr, lport,
 532                                          saddr, inet->inet_dport);
 533         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 534         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 535         struct sock *sk2;
 536         const struct hlist_nulls_node *node;
 537         struct inet_timewait_sock *tw = NULL;
 538
 539         spin_lock(lock);
 540
 541         sk_nulls_for_each(sk2, node, &head->chain) {
 542                 if (sk2->sk_hash != hash)
 543                         continue;
 544
 545                 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
 546                         if (sk2->sk_state == TCP_TIME_WAIT) {
 547                                 tw = inet_twsk(sk2);
 548                                 if (twsk_unique(sk, sk2, twp))
 549                                         break;
 550                         }
 551                         goto not_unique;
 552                 }
 553         }
 554
 555         /* Must record num and sport now. Otherwise we will see
 556          * in hash table socket with a funny identity.
 557          */
 558         inet->inet_num = lport;
 559         inet->inet_sport = htons(lport);
 560         sk->sk_hash = hash;
 561         WARN_ON(!sk_unhashed(sk));
 562         __sk_nulls_add_node_rcu(sk, &head->chain);
 563         if (tw) {
 564                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 565                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 566         }
 567         spin_unlock(lock);
 568         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 569
 570         if (twp) {
 571                 *twp = tw;
 572         } else if (tw) {
 573                 /* Silly. Should hash-dance instead... */
 574                 inet_twsk_deschedule_put(tw);
 575         }
 576         return 0;
 577
 578 not_unique:
 579         spin_unlock(lock);
 580         return -EADDRNOTAVAIL;
 581 }
 582
 583 static u64 inet_sk_port_offset(const struct sock *sk)
 584 {
 585         const struct inet_sock *inet = inet_sk(sk);
 586
 587         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 588                                           inet->inet_daddr,
 589                                           inet->inet_dport);
 590 }
 591
 592 /* Searches for an exsiting socket in the ehash bucket list.
 593  * Returns true if found, false otherwise.
 594  */
 595 static bool inet_ehash_lookup_by_sk(struct sock *sk,
 596                                     struct hlist_nulls_head *list)
 597 {
 598         const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
 599         const int sdif = sk->sk_bound_dev_if;
 600         const int dif = sk->sk_bound_dev_if;
 601         const struct hlist_nulls_node *node;
 602         struct net *net = sock_net(sk);
 603         struct sock *esk;
 604
 605         INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
 606
 607         sk_nulls_for_each_rcu(esk, node, list) {
 608                 if (esk->sk_hash != sk->sk_hash)
 609                         continue;
 610                 if (sk->sk_family == AF_INET) {
 611                         if (unlikely(inet_match(net, esk, acookie,
 612                                                 ports, dif, sdif))) {
 613                                 return true;
 614                         }
 615                 }
 616 #if IS_ENABLED(CONFIG_IPV6)
 617                 else if (sk->sk_family == AF_INET6) {
 618                         if (unlikely(inet6_match(net, esk,
 619                                                  &sk->sk_v6_daddr,
 620                                                  &sk->sk_v6_rcv_saddr,
 621                                                  ports, dif, sdif))) {
 622                                 return true;
 623                         }
 624                 }
 625 #endif
 626         }
 627         return false;
 628 }
 629
 630 /* Insert a socket into ehash, and eventually remove another one
 631  * (The another one can be a SYN_RECV or TIMEWAIT)
 632  * If an existing socket already exists, socket sk is not inserted,
 633  * and sets found_dup_sk parameter to true.
 634  */
 635 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 636 {
 637         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 638         struct inet_ehash_bucket *head;
 639         struct hlist_nulls_head *list;
 640         spinlock_t *lock;
 641         bool ret = true;
 642
 643         WARN_ON_ONCE(!sk_unhashed(sk));
 644
 645         sk->sk_hash = sk_ehashfn(sk);
 646         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 647         list = &head->chain;
 648         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 649
 650         spin_lock(lock);
 651         if (osk) {
 652                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 653                 ret = sk_nulls_del_node_init_rcu(osk);
 654         } else if (found_dup_sk) {
 655                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 656                 if (*found_dup_sk)
 657                         ret = false;
 658         }
 659
 660         if (ret)
 661                 __sk_nulls_add_node_rcu(sk, list);
 662
 663         spin_unlock(lock);
 664
 665         return ret;
 666 }
 667
 668 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 669 {
 670         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 671
 672         if (ok) {
 673                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 674         } else {
 675                 this_cpu_inc(*sk->sk_prot->orphan_count);
 676                 inet_sk_set_state(sk, TCP_CLOSE);
 677                 sock_set_flag(sk, SOCK_DEAD);
 678                 inet_csk_destroy_sock(sk);
 679         }
 680         return ok;
 681 }
 682 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 683
 684 static int inet_reuseport_add_sock(struct sock *sk,
 685                                    struct inet_listen_hashbucket *ilb)
 686 {
 687         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 688         const struct hlist_nulls_node *node;
 689         struct sock *sk2;
 690         kuid_t uid = sock_i_uid(sk);
 691
 692         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 693                 if (sk2 != sk &&
 694                     sk2->sk_family == sk->sk_family &&
 695                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 696                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 697                     inet_csk(sk2)->icsk_bind_hash == tb &&
 698                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 699                     inet_rcv_saddr_equal(sk, sk2, false))
 700                         return reuseport_add_sock(sk, sk2,
 701                                                   inet_rcv_saddr_any(sk));
 702         }
 703
 704         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 705 }
 706
 707 int __inet_hash(struct sock *sk, struct sock *osk)
 708 {
 709         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 710         struct inet_listen_hashbucket *ilb2;
 711         int err = 0;
 712
 713         if (sk->sk_state != TCP_LISTEN) {
 714                 local_bh_disable();
 715                 inet_ehash_nolisten(sk, osk, NULL);
 716                 local_bh_enable();
 717                 return 0;
 718         }
 719         WARN_ON(!sk_unhashed(sk));
 720         ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 721
 722         spin_lock(&ilb2->lock);
 723         if (sk->sk_reuseport) {
 724                 err = inet_reuseport_add_sock(sk, ilb2);
 725                 if (err)
 726                         goto unlock;
 727         }
 728         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 729                 sk->sk_family == AF_INET6)
 730                 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
 731         else
 732                 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
 733         sock_set_flag(sk, SOCK_RCU_FREE);
 734         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 735 unlock:
 736         spin_unlock(&ilb2->lock);
 737
 738         return err;
 739 }
 740 EXPORT_SYMBOL(__inet_hash);
 741
 742 int inet_hash(struct sock *sk)
 743 {
 744         int err = 0;
 745
 746         if (sk->sk_state != TCP_CLOSE)
 747                 err = __inet_hash(sk, NULL);
 748
 749         return err;
 750 }
 751 EXPORT_SYMBOL_GPL(inet_hash);
 752
 753 void inet_unhash(struct sock *sk)
 754 {
 755         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 756
 757         if (sk_unhashed(sk))
 758                 return;
 759
 760         if (sk->sk_state == TCP_LISTEN) {
 761                 struct inet_listen_hashbucket *ilb2;
 762
 763                 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 764                 /* Don't disable bottom halves while acquiring the lock to
 765                  * avoid circular locking dependency on PREEMPT_RT.
 766                  */
 767                 spin_lock(&ilb2->lock);
 768                 if (sk_unhashed(sk)) {
 769                         spin_unlock(&ilb2->lock);
 770                         return;
 771                 }
 772
 773                 if (rcu_access_pointer(sk->sk_reuseport_cb))
 774                         reuseport_stop_listen_sock(sk);
 775
 776                 __sk_nulls_del_node_init_rcu(sk);
 777                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 778                 spin_unlock(&ilb2->lock);
 779         } else {
 780                 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 781
 782                 spin_lock_bh(lock);
 783                 if (sk_unhashed(sk)) {
 784                         spin_unlock_bh(lock);
 785                         return;
 786                 }
 787                 __sk_nulls_del_node_init_rcu(sk);
 788                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 789                 spin_unlock_bh(lock);
 790         }
 791 }
 792 EXPORT_SYMBOL_GPL(inet_unhash);
 793
 794 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
 795                                     const struct net *net, unsigned short port,
 796                                     int l3mdev, const struct sock *sk)
 797 {
 798 #if IS_ENABLED(CONFIG_IPV6)
 799         if (sk->sk_family != tb->family)
 800                 return false;
 801
 802         if (sk->sk_family == AF_INET6)
 803                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 804                         tb->l3mdev == l3mdev &&
 805                         ipv6_addr_equal(&tb->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
 806         else
 807 #endif
 808                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 809                         tb->l3mdev == l3mdev && tb->rcv_saddr == sk->sk_rcv_saddr;
 810 }
 811
 812 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
 813                                       unsigned short port, int l3mdev, const struct sock *sk)
 814 {
 815 #if IS_ENABLED(CONFIG_IPV6)
 816         if (sk->sk_family != tb->family) {
 817                 if (sk->sk_family == AF_INET)
 818                         return net_eq(ib2_net(tb), net) && tb->port == port &&
 819                                 tb->l3mdev == l3mdev &&
 820                                 ipv6_addr_any(&tb->v6_rcv_saddr);
 821
 822                 return false;
 823         }
 824
 825         if (sk->sk_family == AF_INET6)
 826                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 827                         tb->l3mdev == l3mdev &&
 828                         ipv6_addr_any(&tb->v6_rcv_saddr);
 829         else
 830 #endif
 831                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 832                         tb->l3mdev == l3mdev && tb->rcv_saddr == 0;
 833 }
 834
 835 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
 836 struct inet_bind2_bucket *
 837 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
 838                        unsigned short port, int l3mdev, const struct sock *sk)
 839 {
 840         struct inet_bind2_bucket *bhash2 = NULL;
 841
 842         inet_bind_bucket_for_each(bhash2, &head->chain)
 843                 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
 844                         break;
 845
 846         return bhash2;
 847 }
 848
 849 struct inet_bind_hashbucket *
 850 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
 851 {
 852         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 853         u32 hash;
 854
 855 #if IS_ENABLED(CONFIG_IPV6)
 856         if (sk->sk_family == AF_INET6)
 857                 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
 858         else
 859 #endif
 860                 hash = ipv4_portaddr_hash(net, 0, port);
 861
 862         return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
 863 }
 864
 865 static void inet_update_saddr(struct sock *sk, void *saddr, int family)
 866 {
 867         if (family == AF_INET) {
 868                 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
 869                 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
 870         }
 871 #if IS_ENABLED(CONFIG_IPV6)
 872         else {
 873                 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
 874         }
 875 #endif
 876 }
 877
 878 static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
 879 {
 880         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 881         struct inet_bind_hashbucket *head, *head2;
 882         struct inet_bind2_bucket *tb2, *new_tb2;
 883         int l3mdev = inet_sk_bound_l3mdev(sk);
 884         int port = inet_sk(sk)->inet_num;
 885         struct net *net = sock_net(sk);
 886         int bhash;
 887
 888         if (!inet_csk(sk)->icsk_bind2_hash) {
 889                 /* Not bind()ed before. */
 890                 if (reset)
 891                         inet_reset_saddr(sk);
 892                 else
 893                         inet_update_saddr(sk, saddr, family);
 894
 895                 return 0;
 896         }
 897
 898         /* Allocate a bind2 bucket ahead of time to avoid permanently putting
 899          * the bhash2 table in an inconsistent state if a new tb2 bucket
 900          * allocation fails.
 901          */
 902         new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
 903         if (!new_tb2) {
 904                 if (reset) {
 905                         /* The (INADDR_ANY, port) bucket might have already
 906                          * been freed, then we cannot fixup icsk_bind2_hash,
 907                          * so we give up and unlink sk from bhash/bhash2 not
 908                          * to leave inconsistency in bhash2.
 909                          */
 910                         inet_put_port(sk);
 911                         inet_reset_saddr(sk);
 912                 }
 913
 914                 return -ENOMEM;
 915         }
 916
 917         bhash = inet_bhashfn(net, port, hinfo->bhash_size);
 918         head = &hinfo->bhash[bhash];
 919         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 920
 921         /* If we change saddr locklessly, another thread
 922          * iterating over bhash might see corrupted address.
 923          */
 924         spin_lock_bh(&head->lock);
 925
 926         spin_lock(&head2->lock);
 927         __sk_del_bind2_node(sk);
 928         inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
 929         spin_unlock(&head2->lock);
 930
 931         if (reset)
 932                 inet_reset_saddr(sk);
 933         else
 934                 inet_update_saddr(sk, saddr, family);
 935
 936         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 937
 938         spin_lock(&head2->lock);
 939         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 940         if (!tb2) {
 941                 tb2 = new_tb2;
 942                 inet_bind2_bucket_init(tb2, net, head2, port, l3mdev, sk);
 943         }
 944         sk_add_bind2_node(sk, &tb2->owners);
 945         inet_csk(sk)->icsk_bind2_hash = tb2;
 946         spin_unlock(&head2->lock);
 947
 948         spin_unlock_bh(&head->lock);
 949
 950         if (tb2 != new_tb2)
 951                 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
 952
 953         return 0;
 954 }
 955
 956 int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
 957 {
 958         return __inet_bhash2_update_saddr(sk, saddr, family, false);
 959 }
 960 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
 961
 962 void inet_bhash2_reset_saddr(struct sock *sk)
 963 {
 964         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
 965                 __inet_bhash2_update_saddr(sk, NULL, 0, true);
 966 }
 967 EXPORT_SYMBOL_GPL(inet_bhash2_reset_saddr);
 968
 969 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 970  * Note that we use 32bit integers (vs RFC 'short integers')
 971  * because 2^16 is not a multiple of num_ephemeral and this
 972  * property might be used by clever attacker.
 973  *
 974  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
 975  * attacks were since demonstrated, thus we use 65536 by default instead
 976  * to really give more isolation and privacy, at the expense of 256kB
 977  * of kernel memory.
 978  */
 979 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
 980 static u32 *table_perturb;
 981
 982 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 983                 struct sock *sk, u64 port_offset,
 984                 int (*check_established)(struct inet_timewait_death_row *,
 985                         struct sock *, __u16, struct inet_timewait_sock **))
 986 {
 987         struct inet_hashinfo *hinfo = death_row->hashinfo;
 988         struct inet_bind_hashbucket *head, *head2;
 989         struct inet_timewait_sock *tw = NULL;
 990         int port = inet_sk(sk)->inet_num;
 991         struct net *net = sock_net(sk);
 992         struct inet_bind2_bucket *tb2;
 993         struct inet_bind_bucket *tb;
 994         bool tb_created = false;
 995         u32 remaining, offset;
 996         int ret, i, low, high;
 997         int l3mdev;
 998         u32 index;
 999
1000         if (port) {
1001                 local_bh_disable();
1002                 ret = check_established(death_row, sk, port, NULL);
1003                 local_bh_enable();
1004                 return ret;
1005         }
1006
1007         l3mdev = inet_sk_bound_l3mdev(sk);
1008
1009         inet_sk_get_local_port_range(sk, &low, &high);
1010         high++; /* [32768, 60999] -> [32768, 61000[ */
1011         remaining = high - low;
1012         if (likely(remaining > 1))
1013                 remaining &= ~1U;
1014
1015         get_random_sleepable_once(table_perturb,
1016                                   INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1017         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1018
1019         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1020         offset %= remaining;
1021
1022         /* In first pass we try ports of @low parity.
1023          * inet_csk_get_port() does the opposite choice.
1024          */
1025         offset &= ~1U;
1026 other_parity_scan:
1027         port = low + offset;
1028         for (i = 0; i < remaining; i += 2, port += 2) {
1029                 if (unlikely(port >= high))
1030                         port -= remaining;
1031                 if (inet_is_local_reserved_port(net, port))
1032                         continue;
1033                 head = &hinfo->bhash[inet_bhashfn(net, port,
1034                                                   hinfo->bhash_size)];
1035                 spin_lock_bh(&head->lock);
1036
1037                 /* Does not bother with rcv_saddr checks, because
1038                  * the established check is already unique enough.
1039                  */
1040                 inet_bind_bucket_for_each(tb, &head->chain) {
1041                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1042                                 if (tb->fastreuse >= 0 ||
1043                                     tb->fastreuseport >= 0)
1044                                         goto next_port;
1045                                 WARN_ON(hlist_empty(&tb->owners));
1046                                 if (!check_established(death_row, sk,
1047                                                        port, &tw))
1048                                         goto ok;
1049                                 goto next_port;
1050                         }
1051                 }
1052
1053                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1054                                              net, head, port, l3mdev);
1055                 if (!tb) {
1056                         spin_unlock_bh(&head->lock);
1057                         return -ENOMEM;
1058                 }
1059                 tb_created = true;
1060                 tb->fastreuse = -1;
1061                 tb->fastreuseport = -1;
1062                 goto ok;
1063 next_port:
1064                 spin_unlock_bh(&head->lock);
1065                 cond_resched();
1066         }
1067
1068         offset++;
1069         if ((offset & 1) && remaining > 1)
1070                 goto other_parity_scan;
1071
1072         return -EADDRNOTAVAIL;
1073
1074 ok:
1075         /* Find the corresponding tb2 bucket since we need to
1076          * add the socket to the bhash2 table as well
1077          */
1078         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1079         spin_lock(&head2->lock);
1080
1081         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1082         if (!tb2) {
1083                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1084                                                head2, port, l3mdev, sk);
1085                 if (!tb2)
1086                         goto error;
1087         }
1088
1089         /* Here we want to add a little bit of randomness to the next source
1090          * port that will be chosen. We use a max() with a random here so that
1091          * on low contention the randomness is maximal and on high contention
1092          * it may be inexistent.
1093          */
1094         i = max_t(int, i, get_random_u32_below(8) * 2);
1095         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
1096
1097         /* Head lock still held and bh's disabled */
1098         inet_bind_hash(sk, tb, tb2, port);
1099
1100         if (sk_unhashed(sk)) {
1101                 inet_sk(sk)->inet_sport = htons(port);
1102                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1103         }
1104         if (tw)
1105                 inet_twsk_bind_unhash(tw, hinfo);
1106
1107         spin_unlock(&head2->lock);
1108         spin_unlock(&head->lock);
1109
1110         if (tw)
1111                 inet_twsk_deschedule_put(tw);
1112         local_bh_enable();
1113         return 0;
1114
1115 error:
1116         spin_unlock(&head2->lock);
1117         if (tb_created)
1118                 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1119         spin_unlock_bh(&head->lock);
1120         return -ENOMEM;
1121 }
1122
1123 /*
1124  * Bind a port for a connect operation and hash it.
1125  */
1126 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1127                       struct sock *sk)
1128 {
1129         u64 port_offset = 0;
1130
1131         if (!inet_sk(sk)->inet_num)
1132                 port_offset = inet_sk_port_offset(sk);
1133         return __inet_hash_connect(death_row, sk, port_offset,
1134                                    __inet_check_established);
1135 }
1136 EXPORT_SYMBOL_GPL(inet_hash_connect);
1137
1138 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1139 {
1140         int i;
1141
1142         for (i = 0; i <= h->lhash2_mask; i++) {
1143                 spin_lock_init(&h->lhash2[i].lock);
1144                 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1145                                       i + LISTENING_NULLS_BASE);
1146         }
1147 }
1148
1149 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1150                                 unsigned long numentries, int scale,
1151                                 unsigned long low_limit,
1152                                 unsigned long high_limit)
1153 {
1154         h->lhash2 = alloc_large_system_hash(name,
1155                                             sizeof(*h->lhash2),
1156                                             numentries,
1157                                             scale,
1158                                             0,
1159                                             NULL,
1160                                             &h->lhash2_mask,
1161                                             low_limit,
1162                                             high_limit);
1163         init_hashinfo_lhash2(h);
1164
1165         /* this one is used for source ports of outgoing connections */
1166         table_perturb = alloc_large_system_hash("Table-perturb",
1167                                                 sizeof(*table_perturb),
1168                                                 INET_TABLE_PERTURB_SIZE,
1169                                                 0, 0, NULL, NULL,
1170                                                 INET_TABLE_PERTURB_SIZE,
1171                                                 INET_TABLE_PERTURB_SIZE);
1172 }
1173
1174 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1175 {
1176         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1177         if (!h->lhash2)
1178                 return -ENOMEM;
1179
1180         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1181         /* INET_LHTABLE_SIZE must be a power of 2 */
1182         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1183
1184         init_hashinfo_lhash2(h);
1185         return 0;
1186 }
1187 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1188
1189 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1190 {
1191         unsigned int locksz = sizeof(spinlock_t);
1192         unsigned int i, nblocks = 1;
1193
1194         if (locksz != 0) {
1195                 /* allocate 2 cache lines or at least one spinlock per cpu */
1196                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1197                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1198
1199                 /* no more locks than number of hash buckets */
1200                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1201
1202                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1203                 if (!hashinfo->ehash_locks)
1204                         return -ENOMEM;
1205
1206                 for (i = 0; i < nblocks; i++)
1207                         spin_lock_init(&hashinfo->ehash_locks[i]);
1208         }
1209         hashinfo->ehash_locks_mask = nblocks - 1;
1210         return 0;
1211 }
1212 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1213
1214 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1215                                                  unsigned int ehash_entries)
1216 {
1217         struct inet_hashinfo *new_hashinfo;
1218         int i;
1219
1220         new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1221         if (!new_hashinfo)
1222                 goto err;
1223
1224         new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1225                                            GFP_KERNEL_ACCOUNT);
1226         if (!new_hashinfo->ehash)
1227                 goto free_hashinfo;
1228
1229         new_hashinfo->ehash_mask = ehash_entries - 1;
1230
1231         if (inet_ehash_locks_alloc(new_hashinfo))
1232                 goto free_ehash;
1233
1234         for (i = 0; i < ehash_entries; i++)
1235                 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1236
1237         new_hashinfo->pernet = true;
1238
1239         return new_hashinfo;
1240
1241 free_ehash:
1242         vfree(new_hashinfo->ehash);
1243 free_hashinfo:
1244         kfree(new_hashinfo);
1245 err:
1246         return NULL;
1247 }
1248 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1249
1250 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1251 {
1252         if (!hashinfo->pernet)
1253                 return;
1254
1255         inet_ehash_locks_free(hashinfo);
1256         vfree(hashinfo->ehash);
1257         kfree(hashinfo);
1258 }
1259 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);