release/src-rt-6.x/linux/linux-2.6/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 {
 100         write_lock(&hashinfo->lhash_lock);
 101
 102         if (atomic_read(&hashinfo->lhash_users)) {
 103                 DEFINE_WAIT(wait);
 104
 105                 for (;;) {
 106                         prepare_to_wait_exclusive(&hashinfo->lhash_wait,
 107                                                   &wait, TASK_UNINTERRUPTIBLE);
 108                         if (!atomic_read(&hashinfo->lhash_users))
 109                                 break;
 110                         write_unlock_bh(&hashinfo->lhash_lock);
 111                         schedule();
 112                         write_lock_bh(&hashinfo->lhash_lock);
 113                 }
 114
 115                 finish_wait(&hashinfo->lhash_wait, &wait);
 116         }
 117 }
 118
 119 EXPORT_SYMBOL(inet_listen_wlock);
 120
 121 /*
 122  * Don't inline this cruft. Here are some nice properties to exploit here. The
 123  * BSD API does not allow a listening sock to specify the remote port nor the
 124  * remote address for the connection. So always assume those are both
 125  * wildcarded during the search since they can never be otherwise.
 126  */
 127 static struct sock *inet_lookup_listener_slow(const struct hlist_head *head,
 128                                               const __be32 daddr,
 129                                               const unsigned short hnum,
 130                                               const int dif)
 131 {
 132         struct sock *result = NULL, *sk;
 133         const struct hlist_node *node;
 134         int hiscore = -1;
 135
 136         sk_for_each(sk, node, head) {
 137                 const struct inet_sock *inet = inet_sk(sk);
 138
 139                 if (inet->num == hnum && !ipv6_only_sock(sk)) {
 140                         const __be32 rcv_saddr = inet->rcv_saddr;
 141                         int score = sk->sk_family == PF_INET ? 1 : 0;
 142
 143                         if (rcv_saddr) {
 144                                 if (rcv_saddr != daddr)
 145                                         continue;
 146                                 score += 2;
 147                         }
 148                         if (sk->sk_bound_dev_if) {
 149                                 if (sk->sk_bound_dev_if != dif)
 150                                         continue;
 151                                 score += 2;
 152                         }
 153                         if (score == 5)
 154                                 return sk;
 155                         if (score > hiscore) {
 156                                 hiscore = score;
 157                                 result  = sk;
 158                         }
 159                 }
 160         }
 161         return result;
 162 }
 163
 164 /* Optimize the common listener case. */
 165 struct sock *__inet_lookup_listener(struct inet_hashinfo *hashinfo,
 166                                     const __be32 daddr, const unsigned short hnum,
 167                                     const int dif)
 168 {
 169         struct sock *sk = NULL;
 170         const struct hlist_head *head;
 171
 172         read_lock(&hashinfo->lhash_lock);
 173         head = &hashinfo->listening_hash[inet_lhashfn(hnum)];
 174         if (!hlist_empty(head)) {
 175                 const struct inet_sock *inet = inet_sk((sk = __sk_head(head)));
 176
 177                 if (inet->num == hnum && !sk->sk_node.next &&
 178                     (!inet->rcv_saddr || inet->rcv_saddr == daddr) &&
 179                     (sk->sk_family == PF_INET || !ipv6_only_sock(sk)) &&
 180                     !sk->sk_bound_dev_if)
 181                         goto sherry_cache;
 182                 sk = inet_lookup_listener_slow(head, daddr, hnum, dif);
 183         }
 184         if (sk) {
 185 sherry_cache:
 186                 sock_hold(sk);
 187         }
 188         read_unlock(&hashinfo->lhash_lock);
 189         return sk;
 190 }
 191 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 192
 193 struct sock * __inet_lookup_established(struct inet_hashinfo *hashinfo,
 194                                   const __be32 saddr, const __be16 sport,
 195                                   const __be32 daddr, const u16 hnum,
 196                                   const int dif)
 197 {
 198         INET_ADDR_COOKIE(acookie, saddr, daddr)
 199         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 200         struct sock *sk;
 201         const struct hlist_node *node;
 202         /* Optimize here for direct hit, only listening connections can
 203          * have wildcards anyways.
 204          */
 205         unsigned int hash = inet_ehashfn(daddr, hnum, saddr, sport);
 206         struct inet_ehash_bucket *head = inet_ehash_bucket(hashinfo, hash);
 207
 208         prefetch(head->chain.first);
 209         read_lock(&head->lock);
 210         sk_for_each(sk, node, &head->chain) {
 211                 if (INET_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 212                         goto hit; /* You sunk my battleship! */
 213         }
 214
 215         /* Must check for a TIME_WAIT'er before going to listener hash. */
 216         sk_for_each(sk, node, &head->twchain) {
 217                 if (INET_TW_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 218                         goto hit;
 219         }
 220         sk = NULL;
 221 out:
 222         read_unlock(&head->lock);
 223         return sk;
 224 hit:
 225         sock_hold(sk);
 226         goto out;
 227 }
 228 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 229
 230 /* called with local bh disabled */
 231 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 232                                     struct sock *sk, __u16 lport,
 233                                     struct inet_timewait_sock **twp)
 234 {
 235         struct inet_hashinfo *hinfo = death_row->hashinfo;
 236         struct inet_sock *inet = inet_sk(sk);
 237         __be32 daddr = inet->rcv_saddr;
 238         __be32 saddr = inet->daddr;
 239         int dif = sk->sk_bound_dev_if;
 240         INET_ADDR_COOKIE(acookie, saddr, daddr)
 241         const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
 242         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
 243         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 244         struct sock *sk2;
 245         const struct hlist_node *node;
 246         struct inet_timewait_sock *tw;
 247
 248         prefetch(head->chain.first);
 249         write_lock(&head->lock);
 250
 251         /* Check TIME-WAIT sockets first. */
 252         sk_for_each(sk2, node, &head->twchain) {
 253                 tw = inet_twsk(sk2);
 254
 255                 if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
 256                         if (twsk_unique(sk, sk2, twp))
 257                                 goto unique;
 258                         else
 259                                 goto not_unique;
 260                 }
 261         }
 262         tw = NULL;
 263
 264         /* And established part... */
 265         sk_for_each(sk2, node, &head->chain) {
 266                 if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
 267                         goto not_unique;
 268         }
 269
 270 unique:
 271         /* Must record num and sport now. Otherwise we will see
 272          * in hash table socket with a funny identity. */
 273         inet->num = lport;
 274         inet->sport = htons(lport);
 275         sk->sk_hash = hash;
 276         BUG_TRAP(sk_unhashed(sk));
 277         __sk_add_node(sk, &head->chain);
 278         sock_prot_inc_use(sk->sk_prot);
 279         write_unlock(&head->lock);
 280
 281         if (twp) {
 282                 *twp = tw;
 283                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 284         } else if (tw) {
 285                 /* Silly. Should hash-dance instead... */
 286                 inet_twsk_deschedule(tw, death_row);
 287                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 288
 289                 inet_twsk_put(tw);
 290         }
 291
 292         return 0;
 293
 294 not_unique:
 295         write_unlock(&head->lock);
 296         return -EADDRNOTAVAIL;
 297 }
 298
 299 static inline u32 inet_sk_port_offset(const struct sock *sk)
 300 {
 301         const struct inet_sock *inet = inet_sk(sk);
 302         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
 303                                           inet->dport);
 304 }
 305
 306 void __inet_hash_nolisten(struct inet_hashinfo *hashinfo, struct sock *sk)
 307 {
 308         struct hlist_head *list;
 309         rwlock_t *lock;
 310         struct inet_ehash_bucket *head;
 311
 312         BUG_TRAP(sk_unhashed(sk));
 313
 314         sk->sk_hash = inet_sk_ehashfn(sk);
 315         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 316         list = &head->chain;
 317         lock = &head->lock;
 318
 319         write_lock(lock);
 320         __sk_add_node(sk, list);
 321         sock_prot_inc_use(sk->sk_prot);
 322         write_unlock(lock);
 323 }
 324 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
 325
 326 void __inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk)
 327 {
 328         struct hlist_head *list;
 329         rwlock_t *lock;
 330
 331         if (sk->sk_state != TCP_LISTEN) {
 332                 __inet_hash_nolisten(hashinfo, sk);
 333                 return;
 334         }
 335
 336         BUG_TRAP(sk_unhashed(sk));
 337         list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 338         lock = &hashinfo->lhash_lock;
 339
 340         inet_listen_wlock(hashinfo);
 341         __sk_add_node(sk, list);
 342         sock_prot_inc_use(sk->sk_prot);
 343         write_unlock(lock);
 344         wake_up(&hashinfo->lhash_wait);
 345 }
 346 EXPORT_SYMBOL_GPL(__inet_hash);
 347
 348 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 349                 struct sock *sk, u32 port_offset,
 350                 int (*check_established)(struct inet_timewait_death_row *,
 351                         struct sock *, __u16, struct inet_timewait_sock **),
 352                 void (*hash)(struct inet_hashinfo *, struct sock *))
 353 {
 354         struct inet_hashinfo *hinfo = death_row->hashinfo;
 355         const unsigned short snum = inet_sk(sk)->num;
 356         struct inet_bind_hashbucket *head;
 357         struct inet_bind_bucket *tb;
 358         int ret;
 359
 360         if (!snum) {
 361                 int low = sysctl_local_port_range[0];
 362                 int high = sysctl_local_port_range[1];
 363                 int range = high - low;
 364                 int i;
 365                 int port;
 366                 static u32 hint;
 367                 u32 offset = hint + port_offset;
 368                 struct hlist_node *node;
 369                 struct inet_timewait_sock *tw = NULL;
 370
 371                 local_bh_disable();
 372                 for (i = 1; i <= range; i++) {
 373                         port = low + (i + offset) % range;
 374                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
 375                         spin_lock(&head->lock);
 376
 377                         /* Does not bother with rcv_saddr checks,
 378                          * because the established check is already
 379                          * unique enough.
 380                          */
 381                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 382                                 if (tb->port == port) {
 383                                         BUG_TRAP(!hlist_empty(&tb->owners));
 384                                         if (tb->fastreuse >= 0)
 385                                                 goto next_port;
 386                                         if (!check_established(death_row, sk,
 387                                                                 port, &tw))
 388                                                 goto ok;
 389                                         goto next_port;
 390                                 }
 391                         }
 392
 393                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, head, port);
 394                         if (!tb) {
 395                                 spin_unlock(&head->lock);
 396                                 break;
 397                         }
 398                         tb->fastreuse = -1;
 399                         goto ok;
 400
 401                 next_port:
 402                         spin_unlock(&head->lock);
 403                 }
 404                 local_bh_enable();
 405
 406                 return -EADDRNOTAVAIL;
 407
 408 ok:
 409                 hint += i;
 410
 411                 /* Head lock still held and bh's disabled */
 412                 inet_bind_hash(sk, tb, port);
 413                 if (sk_unhashed(sk)) {
 414                         inet_sk(sk)->sport = htons(port);
 415                         hash(hinfo, sk);
 416                 }
 417                 spin_unlock(&head->lock);
 418
 419                 if (tw) {
 420                         inet_twsk_deschedule(tw, death_row);
 421                         inet_twsk_put(tw);
 422                 }
 423
 424                 ret = 0;
 425                 goto out;
 426         }
 427
 428         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
 429         tb  = inet_csk(sk)->icsk_bind_hash;
 430         spin_lock_bh(&head->lock);
 431         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 432                 hash(hinfo, sk);
 433                 spin_unlock_bh(&head->lock);
 434                 return 0;
 435         } else {
 436                 spin_unlock(&head->lock);
 437                 /* No definite answer... Walk to established hash table */
 438                 ret = check_established(death_row, sk, snum, NULL);
 439 out:
 440                 local_bh_enable();
 441                 return ret;
 442         }
 443 }
 444 EXPORT_SYMBOL_GPL(__inet_hash_connect);
 445
 446 /*
 447  * Bind a port for a connect operation and hash it.
 448  */
 449 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 450                       struct sock *sk)
 451 {
 452         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
 453                         __inet_check_established, __inet_hash_nolisten);
 454 }
 455
 456 EXPORT_SYMBOL_GPL(inet_hash_connect);