net/ipv4/inet_hashtables.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Generic INET transport hashtables
   7  *
   8  * Authors:     Lotsa people, from code originally in tcp
   9  *
  10  *      This program is free software; you can redistribute it and/or
  11  *      modify it under the terms of the GNU General Public License
  12  *      as published by the Free Software Foundation; either version
  13  *      2 of the License, or (at your option) any later version.
  14  */
  15
  16 #include <linux/module.h>
  17 #include <linux/random.h>
  18 #include <linux/sched.h>
  19 #include <linux/slab.h>
  20 #include <linux/wait.h>
  21
  22 #include <net/inet_connection_sock.h>
  23 #include <net/inet_hashtables.h>
  24 #include <net/ip.h>
  25
  26 /*
  27  * Allocate and initialize a new local port bind bucket.
  28  * The bindhash mutex for snum's hash chain must be held here.
  29  */
  30 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  31                                                  struct inet_bind_hashbucket *head,
  32                                                  const unsigned short snum)
  33 {
  34         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  35
  36         if (tb != NULL) {
  37                 tb->port      = snum;
  38                 tb->fastreuse = 0;
  39                 INIT_HLIST_HEAD(&tb->owners);
  40                 hlist_add_head(&tb->node, &head->chain);
  41         }
  42         return tb;
  43 }
  44
  45 /*
  46  * Caller must hold hashbucket lock for this tb with local BH disabled
  47  */
  48 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  49 {
  50         if (hlist_empty(&tb->owners)) {
  51                 __hlist_del(&tb->node);
  52                 kmem_cache_free(cachep, tb);
  53         }
  54 }
  55
  56 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  57                     const unsigned short snum)
  58 {
  59         inet_sk(sk)->num = snum;
  60         sk_add_bind_node(sk, &tb->owners);
  61         inet_csk(sk)->icsk_bind_hash = tb;
  62 }
  63
  64 /*
  65  * Get rid of any references to a local port held by the given sock.
  66  */
  67 static void __inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
  68 {
  69         const int bhash = inet_bhashfn(inet_sk(sk)->num, hashinfo->bhash_size);
  70         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
  71         struct inet_bind_bucket *tb;
  72
  73         spin_lock(&head->lock);
  74         tb = inet_csk(sk)->icsk_bind_hash;
  75         __sk_del_bind_node(sk);
  76         inet_csk(sk)->icsk_bind_hash = NULL;
  77         inet_sk(sk)->num = 0;
  78         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  79         spin_unlock(&head->lock);
  80 }
  81
  82 void inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
  83 {
  84         local_bh_disable();
  85         __inet_put_port(hashinfo, sk);
  86         local_bh_enable();
  87 }
  88
  89 EXPORT_SYMBOL(inet_put_port);
  90
  91 /*
  92  * This lock without WQ_FLAG_EXCLUSIVE is good on UP and it can be very bad on SMP.
  93  * Look, when several writers sleep and reader wakes them up, all but one
  94  * immediately hit write lock and grab all the cpus. Exclusive sleep solves
  95  * this, _but_ remember, it adds useless work on UP machines (wake up each
  96  * exclusive lock release). It should be ifdefed really.
  97  */
  98 void inet_listen_wlock(struct inet_hashinfo *hashinfo)
  99         __acquires(hashinfo->lhash_lock)
 100 {
 101         write_lock(&hashinfo->lhash_lock);
 102
 103         if (atomic_read(&hashinfo->lhash_users)) {
 104                 DEFINE_WAIT(wait);
 105
 106                 for (;;) {
 107                         prepare_to_wait_exclusive(&hashinfo->lhash_wait,
 108                                                   &wait, TASK_UNINTERRUPTIBLE);
 109                         if (!atomic_read(&hashinfo->lhash_users))
 110                                 break;
 111                         write_unlock_bh(&hashinfo->lhash_lock);
 112                         schedule();
 113                         write_lock_bh(&hashinfo->lhash_lock);
 114                 }
 115
 116                 finish_wait(&hashinfo->lhash_wait, &wait);
 117         }
 118 }
 119
 120 EXPORT_SYMBOL(inet_listen_wlock);
 121
 122 /*
 123  * Don't inline this cruft. Here are some nice properties to exploit here. The
 124  * BSD API does not allow a listening sock to specify the remote port nor the
 125  * remote address for the connection. So always assume those are both
 126  * wildcarded during the search since they can never be otherwise.
 127  */
 128 static struct sock *inet_lookup_listener_slow(const struct hlist_head *head,
 129                                               const __be32 daddr,
 130                                               const unsigned short hnum,
 131                                               const int dif)
 132 {
 133         struct sock *result = NULL, *sk;
 134         const struct hlist_node *node;
 135         int hiscore = -1;
 136
 137         sk_for_each(sk, node, head) {
 138                 const struct inet_sock *inet = inet_sk(sk);
 139
 140                 if (inet->num == hnum && !ipv6_only_sock(sk)) {
 141                         const __be32 rcv_saddr = inet->rcv_saddr;
 142                         int score = sk->sk_family == PF_INET ? 1 : 0;
 143
 144                         if (rcv_saddr) {
 145                                 if (rcv_saddr != daddr)
 146                                         continue;
 147                                 score += 2;
 148                         }
 149                         if (sk->sk_bound_dev_if) {
 150                                 if (sk->sk_bound_dev_if != dif)
 151                                         continue;
 152                                 score += 2;
 153                         }
 154                         if (score == 5)
 155                                 return sk;
 156                         if (score > hiscore) {
 157                                 hiscore = score;
 158                                 result  = sk;
 159                         }
 160                 }
 161         }
 162         return result;
 163 }
 164
 165 /* Optimize the common listener case. */
 166 struct sock *__inet_lookup_listener(struct inet_hashinfo *hashinfo,
 167                                     const __be32 daddr, const unsigned short hnum,
 168                                     const int dif)
 169 {
 170         struct sock *sk = NULL;
 171         const struct hlist_head *head;
 172
 173         read_lock(&hashinfo->lhash_lock);
 174         head = &hashinfo->listening_hash[inet_lhashfn(hnum)];
 175         if (!hlist_empty(head)) {
 176                 const struct inet_sock *inet = inet_sk((sk = __sk_head(head)));
 177
 178                 if (inet->num == hnum && !sk->sk_node.next &&
 179                     (!inet->rcv_saddr || inet->rcv_saddr == daddr) &&
 180                     (sk->sk_family == PF_INET || !ipv6_only_sock(sk)) &&
 181                     !sk->sk_bound_dev_if)
 182                         goto sherry_cache;
 183                 sk = inet_lookup_listener_slow(head, daddr, hnum, dif);
 184         }
 185         if (sk) {
 186 sherry_cache:
 187                 sock_hold(sk);
 188         }
 189         read_unlock(&hashinfo->lhash_lock);
 190         return sk;
 191 }
 192 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 193
 194 struct sock * __inet_lookup_established(struct inet_hashinfo *hashinfo,
 195                                   const __be32 saddr, const __be16 sport,
 196                                   const __be32 daddr, const u16 hnum,
 197                                   const int dif)
 198 {
 199         INET_ADDR_COOKIE(acookie, saddr, daddr)
 200         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 201         struct sock *sk;
 202         const struct hlist_node *node;
 203         /* Optimize here for direct hit, only listening connections can
 204          * have wildcards anyways.
 205          */
 206         unsigned int hash = inet_ehashfn(daddr, hnum, saddr, sport);
 207         struct inet_ehash_bucket *head = inet_ehash_bucket(hashinfo, hash);
 208         rwlock_t *lock = inet_ehash_lockp(hashinfo, hash);
 209
 210         prefetch(head->chain.first);
 211         read_lock(lock);
 212         sk_for_each(sk, node, &head->chain) {
 213                 if (INET_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 214                         goto hit; /* You sunk my battleship! */
 215         }
 216
 217         /* Must check for a TIME_WAIT'er before going to listener hash. */
 218         sk_for_each(sk, node, &head->twchain) {
 219                 if (INET_TW_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 220                         goto hit;
 221         }
 222         sk = NULL;
 223 out:
 224         read_unlock(lock);
 225         return sk;
 226 hit:
 227         sock_hold(sk);
 228         goto out;
 229 }
 230 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 231
 232 /* called with local bh disabled */
 233 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 234                                     struct sock *sk, __u16 lport,
 235                                     struct inet_timewait_sock **twp)
 236 {
 237         struct inet_hashinfo *hinfo = death_row->hashinfo;
 238         struct inet_sock *inet = inet_sk(sk);
 239         __be32 daddr = inet->rcv_saddr;
 240         __be32 saddr = inet->daddr;
 241         int dif = sk->sk_bound_dev_if;
 242         INET_ADDR_COOKIE(acookie, saddr, daddr)
 243         const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
 244         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
 245         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 246         rwlock_t *lock = inet_ehash_lockp(hinfo, hash);
 247         struct sock *sk2;
 248         const struct hlist_node *node;
 249         struct inet_timewait_sock *tw;
 250
 251         prefetch(head->chain.first);
 252         write_lock(lock);
 253
 254         /* Check TIME-WAIT sockets first. */
 255         sk_for_each(sk2, node, &head->twchain) {
 256                 tw = inet_twsk(sk2);
 257
 258                 if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
 259                         if (twsk_unique(sk, sk2, twp))
 260                                 goto unique;
 261                         else
 262                                 goto not_unique;
 263                 }
 264         }
 265         tw = NULL;
 266
 267         /* And established part... */
 268         sk_for_each(sk2, node, &head->chain) {
 269                 if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
 270                         goto not_unique;
 271         }
 272
 273 unique:
 274         /* Must record num and sport now. Otherwise we will see
 275          * in hash table socket with a funny identity. */
 276         inet->num = lport;
 277         inet->sport = htons(lport);
 278         sk->sk_hash = hash;
 279         BUG_TRAP(sk_unhashed(sk));
 280         __sk_add_node(sk, &head->chain);
 281         sock_prot_inuse_add(sk->sk_prot, 1);
 282         write_unlock(lock);
 283
 284         if (twp) {
 285                 *twp = tw;
 286                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 287         } else if (tw) {
 288                 /* Silly. Should hash-dance instead... */
 289                 inet_twsk_deschedule(tw, death_row);
 290                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 291
 292                 inet_twsk_put(tw);
 293         }
 294
 295         return 0;
 296
 297 not_unique:
 298         write_unlock(lock);
 299         return -EADDRNOTAVAIL;
 300 }
 301
 302 static inline u32 inet_sk_port_offset(const struct sock *sk)
 303 {
 304         const struct inet_sock *inet = inet_sk(sk);
 305         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
 306                                           inet->dport);
 307 }
 308
 309 void __inet_hash_nolisten(struct inet_hashinfo *hashinfo, struct sock *sk)
 310 {
 311         struct hlist_head *list;
 312         rwlock_t *lock;
 313         struct inet_ehash_bucket *head;
 314
 315         BUG_TRAP(sk_unhashed(sk));
 316
 317         sk->sk_hash = inet_sk_ehashfn(sk);
 318         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 319         list = &head->chain;
 320         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 321
 322         write_lock(lock);
 323         __sk_add_node(sk, list);
 324         sock_prot_inuse_add(sk->sk_prot, 1);
 325         write_unlock(lock);
 326 }
 327 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 328
 329 void __inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk)
 330 {
 331         struct hlist_head *list;
 332         rwlock_t *lock;
 333
 334         if (sk->sk_state != TCP_LISTEN) {
 335                 __inet_hash_nolisten(hashinfo, sk);
 336                 return;
 337         }
 338
 339         BUG_TRAP(sk_unhashed(sk));
 340         list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 341         lock = &hashinfo->lhash_lock;
 342
 343         inet_listen_wlock(hashinfo);
 344         __sk_add_node(sk, list);
 345         sock_prot_inuse_add(sk->sk_prot, 1);
 346         write_unlock(lock);
 347         wake_up(&hashinfo->lhash_wait);
 348 }
 349 EXPORT_SYMBOL_GPL(__inet_hash);
 350
 351 /*
 352  * Bind a port for a connect operation and hash it.
 353  */
 354 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 355                       struct sock *sk)
 356 {
 357         struct inet_hashinfo *hinfo = death_row->hashinfo;
 358         const unsigned short snum = inet_sk(sk)->num;
 359         struct inet_bind_hashbucket *head;
 360         struct inet_bind_bucket *tb;
 361         int ret;
 362
 363         if (!snum) {
 364                 int i, remaining, low, high, port;
 365                 static u32 hint;
 366                 u32 offset = hint + inet_sk_port_offset(sk);
 367                 struct hlist_node *node;
 368                 struct inet_timewait_sock *tw = NULL;
 369
 370                 inet_get_local_port_range(&low, &high);
 371                 remaining = (high - low) + 1;
 372
 373                 local_bh_disable();
 374                 for (i = 1; i <= remaining; i++) {
 375                         port = low + (i + offset) % remaining;
 376                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
 377                         spin_lock(&head->lock);
 378
 379                         /* Does not bother with rcv_saddr checks,
 380                          * because the established check is already
 381                          * unique enough.
 382                          */
 383                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 384                                 if (tb->port == port) {
 385                                         BUG_TRAP(!hlist_empty(&tb->owners));
 386                                         if (tb->fastreuse >= 0)
 387                                                 goto next_port;
 388                                         if (!__inet_check_established(death_row,
 389                                                                       sk, port,
 390                                                                       &tw))
 391                                                 goto ok;
 392                                         goto next_port;
 393                                 }
 394                         }
 395
 396                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, head, port);
 397                         if (!tb) {
 398                                 spin_unlock(&head->lock);
 399                                 break;
 400                         }
 401                         tb->fastreuse = -1;
 402                         goto ok;
 403
 404                 next_port:
 405                         spin_unlock(&head->lock);
 406                 }
 407                 local_bh_enable();
 408
 409                 return -EADDRNOTAVAIL;
 410
 411 ok:
 412                 hint += i;
 413
 414                 /* Head lock still held and bh's disabled */
 415                 inet_bind_hash(sk, tb, port);
 416                 if (sk_unhashed(sk)) {
 417                         inet_sk(sk)->sport = htons(port);
 418                         __inet_hash_nolisten(hinfo, sk);
 419                 }
 420                 spin_unlock(&head->lock);
 421
 422                 if (tw) {
 423                         inet_twsk_deschedule(tw, death_row);
 424                         inet_twsk_put(tw);
 425                 }
 426
 427                 ret = 0;
 428                 goto out;
 429         }
 430
 431         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
 432         tb  = inet_csk(sk)->icsk_bind_hash;
 433         spin_lock_bh(&head->lock);
 434         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 435                 __inet_hash_nolisten(hinfo, sk);
 436                 spin_unlock_bh(&head->lock);
 437                 return 0;
 438         } else {
 439                 spin_unlock(&head->lock);
 440                 /* No definite answer... Walk to established hash table */
 441                 ret = __inet_check_established(death_row, sk, snum, NULL);
 442 out:
 443                 local_bh_enable();
 444                 return ret;
 445         }
 446 }
 447
 448 EXPORT_SYMBOL_GPL(inet_hash_connect);