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 net *net,
  32                                                  struct inet_bind_hashbucket *head,
  33                                                  const unsigned short snum)
  34 {
  35         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  36
  37         if (tb != NULL) {
  38                 tb->ib_net       = net;
  39                 tb->port      = snum;
  40                 tb->fastreuse = 0;
  41                 INIT_HLIST_HEAD(&tb->owners);
  42                 hlist_add_head(&tb->node, &head->chain);
  43         }
  44         return tb;
  45 }
  46
  47 /*
  48  * Caller must hold hashbucket lock for this tb with local BH disabled
  49  */
  50 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  51 {
  52         if (hlist_empty(&tb->owners)) {
  53                 __hlist_del(&tb->node);
  54                 kmem_cache_free(cachep, tb);
  55         }
  56 }
  57
  58 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  59                     const unsigned short snum)
  60 {
  61         inet_sk(sk)->num = snum;
  62         sk_add_bind_node(sk, &tb->owners);
  63         inet_csk(sk)->icsk_bind_hash = tb;
  64 }
  65
  66 /*
  67  * Get rid of any references to a local port held by the given sock.
  68  */
  69 static void __inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
  70 {
  71         const int bhash = inet_bhashfn(inet_sk(sk)->num, hashinfo->bhash_size);
  72         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
  73         struct inet_bind_bucket *tb;
  74
  75         spin_lock(&head->lock);
  76         tb = inet_csk(sk)->icsk_bind_hash;
  77         __sk_del_bind_node(sk);
  78         inet_csk(sk)->icsk_bind_hash = NULL;
  79         inet_sk(sk)->num = 0;
  80         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  81         spin_unlock(&head->lock);
  82 }
  83
  84 void inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
  85 {
  86         local_bh_disable();
  87         __inet_put_port(hashinfo, sk);
  88         local_bh_enable();
  89 }
  90
  91 EXPORT_SYMBOL(inet_put_port);
  92
  93 /*
  94  * This lock without WQ_FLAG_EXCLUSIVE is good on UP and it can be very bad on SMP.
  95  * Look, when several writers sleep and reader wakes them up, all but one
  96  * immediately hit write lock and grab all the cpus. Exclusive sleep solves
  97  * this, _but_ remember, it adds useless work on UP machines (wake up each
  98  * exclusive lock release). It should be ifdefed really.
  99  */
 100 void inet_listen_wlock(struct inet_hashinfo *hashinfo)
 101         __acquires(hashinfo->lhash_lock)
 102 {
 103         write_lock(&hashinfo->lhash_lock);
 104
 105         if (atomic_read(&hashinfo->lhash_users)) {
 106                 DEFINE_WAIT(wait);
 107
 108                 for (;;) {
 109                         prepare_to_wait_exclusive(&hashinfo->lhash_wait,
 110                                                   &wait, TASK_UNINTERRUPTIBLE);
 111                         if (!atomic_read(&hashinfo->lhash_users))
 112                                 break;
 113                         write_unlock_bh(&hashinfo->lhash_lock);
 114                         schedule();
 115                         write_lock_bh(&hashinfo->lhash_lock);
 116                 }
 117
 118                 finish_wait(&hashinfo->lhash_wait, &wait);
 119         }
 120 }
 121
 122 EXPORT_SYMBOL(inet_listen_wlock);
 123
 124 /*
 125  * Don't inline this cruft. Here are some nice properties to exploit here. The
 126  * BSD API does not allow a listening sock to specify the remote port nor the
 127  * remote address for the connection. So always assume those are both
 128  * wildcarded during the search since they can never be otherwise.
 129  */
 130 static struct sock *inet_lookup_listener_slow(const struct hlist_head *head,
 131                                               const __be32 daddr,
 132                                               const unsigned short hnum,
 133                                               const int dif)
 134 {
 135         struct sock *result = NULL, *sk;
 136         const struct hlist_node *node;
 137         int hiscore = -1;
 138
 139         sk_for_each(sk, node, head) {
 140                 const struct inet_sock *inet = inet_sk(sk);
 141
 142                 if (inet->num == hnum && !ipv6_only_sock(sk)) {
 143                         const __be32 rcv_saddr = inet->rcv_saddr;
 144                         int score = sk->sk_family == PF_INET ? 1 : 0;
 145
 146                         if (rcv_saddr) {
 147                                 if (rcv_saddr != daddr)
 148                                         continue;
 149                                 score += 2;
 150                         }
 151                         if (sk->sk_bound_dev_if) {
 152                                 if (sk->sk_bound_dev_if != dif)
 153                                         continue;
 154                                 score += 2;
 155                         }
 156                         if (score == 5)
 157                                 return sk;
 158                         if (score > hiscore) {
 159                                 hiscore = score;
 160                                 result  = sk;
 161                         }
 162                 }
 163         }
 164         return result;
 165 }
 166
 167 /* Optimize the common listener case. */
 168 struct sock *__inet_lookup_listener(struct inet_hashinfo *hashinfo,
 169                                     const __be32 daddr, const unsigned short hnum,
 170                                     const int dif)
 171 {
 172         struct sock *sk = NULL;
 173         const struct hlist_head *head;
 174
 175         read_lock(&hashinfo->lhash_lock);
 176         head = &hashinfo->listening_hash[inet_lhashfn(hnum)];
 177         if (!hlist_empty(head)) {
 178                 const struct inet_sock *inet = inet_sk((sk = __sk_head(head)));
 179
 180                 if (inet->num == hnum && !sk->sk_node.next &&
 181                     (!inet->rcv_saddr || inet->rcv_saddr == daddr) &&
 182                     (sk->sk_family == PF_INET || !ipv6_only_sock(sk)) &&
 183                     !sk->sk_bound_dev_if)
 184                         goto sherry_cache;
 185                 sk = inet_lookup_listener_slow(head, daddr, hnum, dif);
 186         }
 187         if (sk) {
 188 sherry_cache:
 189                 sock_hold(sk);
 190         }
 191         read_unlock(&hashinfo->lhash_lock);
 192         return sk;
 193 }
 194 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 195
 196 struct sock * __inet_lookup_established(struct inet_hashinfo *hashinfo,
 197                                   const __be32 saddr, const __be16 sport,
 198                                   const __be32 daddr, const u16 hnum,
 199                                   const int dif)
 200 {
 201         INET_ADDR_COOKIE(acookie, saddr, daddr)
 202         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 203         struct sock *sk;
 204         const struct hlist_node *node;
 205         /* Optimize here for direct hit, only listening connections can
 206          * have wildcards anyways.
 207          */
 208         unsigned int hash = inet_ehashfn(daddr, hnum, saddr, sport);
 209         struct inet_ehash_bucket *head = inet_ehash_bucket(hashinfo, hash);
 210         rwlock_t *lock = inet_ehash_lockp(hashinfo, hash);
 211
 212         prefetch(head->chain.first);
 213         read_lock(lock);
 214         sk_for_each(sk, node, &head->chain) {
 215                 if (INET_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 216                         goto hit; /* You sunk my battleship! */
 217         }
 218
 219         /* Must check for a TIME_WAIT'er before going to listener hash. */
 220         sk_for_each(sk, node, &head->twchain) {
 221                 if (INET_TW_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 222                         goto hit;
 223         }
 224         sk = NULL;
 225 out:
 226         read_unlock(lock);
 227         return sk;
 228 hit:
 229         sock_hold(sk);
 230         goto out;
 231 }
 232 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 233
 234 /* called with local bh disabled */
 235 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 236                                     struct sock *sk, __u16 lport,
 237                                     struct inet_timewait_sock **twp)
 238 {
 239         struct inet_hashinfo *hinfo = death_row->hashinfo;
 240         struct inet_sock *inet = inet_sk(sk);
 241         __be32 daddr = inet->rcv_saddr;
 242         __be32 saddr = inet->daddr;
 243         int dif = sk->sk_bound_dev_if;
 244         INET_ADDR_COOKIE(acookie, saddr, daddr)
 245         const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
 246         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
 247         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 248         rwlock_t *lock = inet_ehash_lockp(hinfo, hash);
 249         struct sock *sk2;
 250         const struct hlist_node *node;
 251         struct inet_timewait_sock *tw;
 252
 253         prefetch(head->chain.first);
 254         write_lock(lock);
 255
 256         /* Check TIME-WAIT sockets first. */
 257         sk_for_each(sk2, node, &head->twchain) {
 258                 tw = inet_twsk(sk2);
 259
 260                 if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
 261                         if (twsk_unique(sk, sk2, twp))
 262                                 goto unique;
 263                         else
 264                                 goto not_unique;
 265                 }
 266         }
 267         tw = NULL;
 268
 269         /* And established part... */
 270         sk_for_each(sk2, node, &head->chain) {
 271                 if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
 272                         goto not_unique;
 273         }
 274
 275 unique:
 276         /* Must record num and sport now. Otherwise we will see
 277          * in hash table socket with a funny identity. */
 278         inet->num = lport;
 279         inet->sport = htons(lport);
 280         sk->sk_hash = hash;
 281         BUG_TRAP(sk_unhashed(sk));
 282         __sk_add_node(sk, &head->chain);
 283         sock_prot_inuse_add(sk->sk_prot, 1);
 284         write_unlock(lock);
 285
 286         if (twp) {
 287                 *twp = tw;
 288                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 289         } else if (tw) {
 290                 /* Silly. Should hash-dance instead... */
 291                 inet_twsk_deschedule(tw, death_row);
 292                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 293
 294                 inet_twsk_put(tw);
 295         }
 296
 297         return 0;
 298
 299 not_unique:
 300         write_unlock(lock);
 301         return -EADDRNOTAVAIL;
 302 }
 303
 304 static inline u32 inet_sk_port_offset(const struct sock *sk)
 305 {
 306         const struct inet_sock *inet = inet_sk(sk);
 307         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
 308                                           inet->dport);
 309 }
 310
 311 void __inet_hash_nolisten(struct inet_hashinfo *hashinfo, struct sock *sk)
 312 {
 313         struct hlist_head *list;
 314         rwlock_t *lock;
 315         struct inet_ehash_bucket *head;
 316
 317         BUG_TRAP(sk_unhashed(sk));
 318
 319         sk->sk_hash = inet_sk_ehashfn(sk);
 320         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 321         list = &head->chain;
 322         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 323
 324         write_lock(lock);
 325         __sk_add_node(sk, list);
 326         sock_prot_inuse_add(sk->sk_prot, 1);
 327         write_unlock(lock);
 328 }
 329 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 330
 331 void __inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk)
 332 {
 333         struct hlist_head *list;
 334         rwlock_t *lock;
 335
 336         if (sk->sk_state != TCP_LISTEN) {
 337                 __inet_hash_nolisten(hashinfo, sk);
 338                 return;
 339         }
 340
 341         BUG_TRAP(sk_unhashed(sk));
 342         list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 343         lock = &hashinfo->lhash_lock;
 344
 345         inet_listen_wlock(hashinfo);
 346         __sk_add_node(sk, list);
 347         sock_prot_inuse_add(sk->sk_prot, 1);
 348         write_unlock(lock);
 349         wake_up(&hashinfo->lhash_wait);
 350 }
 351 EXPORT_SYMBOL_GPL(__inet_hash);
 352
 353 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 354                 struct sock *sk,
 355                 int (*check_established)(struct inet_timewait_death_row *,
 356                         struct sock *, __u16, struct inet_timewait_sock **),
 357                 void (*hash)(struct inet_hashinfo *, struct sock *))
 358 {
 359         struct inet_hashinfo *hinfo = death_row->hashinfo;
 360         const unsigned short snum = inet_sk(sk)->num;
 361         struct inet_bind_hashbucket *head;
 362         struct inet_bind_bucket *tb;
 363         int ret;
 364         struct net *net = sk->sk_net;
 365
 366         if (!snum) {
 367                 int i, remaining, low, high, port;
 368                 static u32 hint;
 369                 u32 offset = hint + inet_sk_port_offset(sk);
 370                 struct hlist_node *node;
 371                 struct inet_timewait_sock *tw = NULL;
 372
 373                 inet_get_local_port_range(&low, &high);
 374                 remaining = (high - low) + 1;
 375
 376                 local_bh_disable();
 377                 for (i = 1; i <= remaining; i++) {
 378                         port = low + (i + offset) % remaining;
 379                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
 380                         spin_lock(&head->lock);
 381
 382                         /* Does not bother with rcv_saddr checks,
 383                          * because the established check is already
 384                          * unique enough.
 385                          */
 386                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 387                                 if (tb->ib_net == net && tb->port == port) {
 388                                         BUG_TRAP(!hlist_empty(&tb->owners));
 389                                         if (tb->fastreuse >= 0)
 390                                                 goto next_port;
 391                                         if (!check_established(death_row, sk,
 392                                                                 port, &tw))
 393                                                 goto ok;
 394                                         goto next_port;
 395                                 }
 396                         }
 397
 398                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 399                                         net, head, port);
 400                         if (!tb) {
 401                                 spin_unlock(&head->lock);
 402                                 break;
 403                         }
 404                         tb->fastreuse = -1;
 405                         goto ok;
 406
 407                 next_port:
 408                         spin_unlock(&head->lock);
 409                 }
 410                 local_bh_enable();
 411
 412                 return -EADDRNOTAVAIL;
 413
 414 ok:
 415                 hint += i;
 416
 417                 /* Head lock still held and bh's disabled */
 418                 inet_bind_hash(sk, tb, port);
 419                 if (sk_unhashed(sk)) {
 420                         inet_sk(sk)->sport = htons(port);
 421                         hash(hinfo, sk);
 422                 }
 423                 spin_unlock(&head->lock);
 424
 425                 if (tw) {
 426                         inet_twsk_deschedule(tw, death_row);
 427                         inet_twsk_put(tw);
 428                 }
 429
 430                 ret = 0;
 431                 goto out;
 432         }
 433
 434         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
 435         tb  = inet_csk(sk)->icsk_bind_hash;
 436         spin_lock_bh(&head->lock);
 437         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 438                 hash(hinfo, sk);
 439                 spin_unlock_bh(&head->lock);
 440                 return 0;
 441         } else {
 442                 spin_unlock(&head->lock);
 443                 /* No definite answer... Walk to established hash table */
 444                 ret = check_established(death_row, sk, snum, NULL);
 445 out:
 446                 local_bh_enable();
 447                 return ret;
 448         }
 449 }
 450 EXPORT_SYMBOL_GPL(__inet_hash_connect);
 451
 452 /*
 453  * Bind a port for a connect operation and hash it.
 454  */
 455 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 456                       struct sock *sk)
 457 {
 458         return __inet_hash_connect(death_row, sk,
 459                         __inet_check_established, __inet_hash_nolisten);
 460 }
 461
 462 EXPORT_SYMBOL_GPL(inet_hash_connect);