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/config.h>
  17 #include <linux/module.h>
  18 #include <linux/random.h>
  19 #include <linux/sched.h>
  20 #include <linux/slab.h>
  21 #include <linux/wait.h>
  22
  23 #include <net/inet_connection_sock.h>
  24 #include <net/inet_hashtables.h>
  25 #include <net/ip.h>
  26
  27 /*
  28  * Allocate and initialize a new local port bind bucket.
  29  * The bindhash mutex for snum's hash chain must be held here.
  30  */
  31 struct inet_bind_bucket *inet_bind_bucket_create(kmem_cache_t *cachep,
  32                                                  struct inet_bind_hashbucket *head,
  33                                                  const unsigned short snum)
  34 {
  35         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, SLAB_ATOMIC);
  36
  37         if (tb != NULL) {
  38                 tb->port      = snum;
  39                 tb->fastreuse = 0;
  40                 INIT_HLIST_HEAD(&tb->owners);
  41                 hlist_add_head(&tb->node, &head->chain);
  42         }
  43         return tb;
  44 }
  45
  46 EXPORT_SYMBOL(inet_bind_bucket_create);
  47
  48 /*
  49  * Caller must hold hashbucket lock for this tb with local BH disabled
  50  */
  51 void inet_bind_bucket_destroy(kmem_cache_t *cachep, struct inet_bind_bucket *tb)
  52 {
  53         if (hlist_empty(&tb->owners)) {
  54                 __hlist_del(&tb->node);
  55                 kmem_cache_free(cachep, tb);
  56         }
  57 }
  58
  59 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
  60                     const unsigned short snum)
  61 {
  62         inet_sk(sk)->num = snum;
  63         sk_add_bind_node(sk, &tb->owners);
  64         inet_csk(sk)->icsk_bind_hash = tb;
  65 }
  66
  67 EXPORT_SYMBOL(inet_bind_hash);
  68
  69 /*
  70  * Get rid of any references to a local port held by the given sock.
  71  */
  72 static void __inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
  73 {
  74         const int bhash = inet_bhashfn(inet_sk(sk)->num, hashinfo->bhash_size);
  75         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
  76         struct inet_bind_bucket *tb;
  77
  78         spin_lock(&head->lock);
  79         tb = inet_csk(sk)->icsk_bind_hash;
  80         __sk_del_bind_node(sk);
  81         inet_csk(sk)->icsk_bind_hash = NULL;
  82         inet_sk(sk)->num = 0;
  83         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  84         spin_unlock(&head->lock);
  85 }
  86
  87 void inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
  88 {
  89         local_bh_disable();
  90         __inet_put_port(hashinfo, sk);
  91         local_bh_enable();
  92 }
  93
  94 EXPORT_SYMBOL(inet_put_port);
  95
  96 /*
  97  * This lock without WQ_FLAG_EXCLUSIVE is good on UP and it can be very bad on SMP.
  98  * Look, when several writers sleep and reader wakes them up, all but one
  99  * immediately hit write lock and grab all the cpus. Exclusive sleep solves
 100  * this, _but_ remember, it adds useless work on UP machines (wake up each
 101  * exclusive lock release). It should be ifdefed really.
 102  */
 103 void inet_listen_wlock(struct inet_hashinfo *hashinfo)
 104 {
 105         write_lock(&hashinfo->lhash_lock);
 106
 107         if (atomic_read(&hashinfo->lhash_users)) {
 108                 DEFINE_WAIT(wait);
 109
 110                 for (;;) {
 111                         prepare_to_wait_exclusive(&hashinfo->lhash_wait,
 112                                                   &wait, TASK_UNINTERRUPTIBLE);
 113                         if (!atomic_read(&hashinfo->lhash_users))
 114                                 break;
 115                         write_unlock_bh(&hashinfo->lhash_lock);
 116                         schedule();
 117                         write_lock_bh(&hashinfo->lhash_lock);
 118                 }
 119
 120                 finish_wait(&hashinfo->lhash_wait, &wait);
 121         }
 122 }
 123
 124 EXPORT_SYMBOL(inet_listen_wlock);
 125
 126 /*
 127  * Don't inline this cruft. Here are some nice properties to exploit here. The
 128  * BSD API does not allow a listening sock to specify the remote port nor the
 129  * remote address for the connection. So always assume those are both
 130  * wildcarded during the search since they can never be otherwise.
 131  */
 132 struct sock *__inet_lookup_listener(const struct hlist_head *head, const u32 daddr,
 133                                     const unsigned short hnum, 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 __u32 rcv_saddr = inet->rcv_saddr;
 144                         int score = sk->sk_family == PF_INET ? 1 : 0;
 145
 146                         if (inet_addr_match(sk->sk_nx_info, daddr, rcv_saddr))
 147                                 score += 2;
 148                         else
 149                                 continue;
 150                         if (sk->sk_bound_dev_if) {
 151                                 if (sk->sk_bound_dev_if != dif)
 152                                         continue;
 153                                 score += 2;
 154                         }
 155                         if (score == 5)
 156                                 return sk;
 157                         if (score > hiscore) {
 158                                 hiscore = score;
 159                                 result  = sk;
 160                         }
 161                 }
 162         }
 163         return result;
 164 }
 165
 166 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 167
 168 /* called with local bh disabled */
 169 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 170                                     struct sock *sk, __u16 lport,
 171                                     struct inet_timewait_sock **twp)
 172 {
 173         struct inet_hashinfo *hinfo = death_row->hashinfo;
 174         struct inet_sock *inet = inet_sk(sk);
 175         u32 daddr = inet->rcv_saddr;
 176         u32 saddr = inet->daddr;
 177         int dif = sk->sk_bound_dev_if;
 178         INET_ADDR_COOKIE(acookie, saddr, daddr)
 179         const __u32 ports = INET_COMBINED_PORTS(inet->dport, lport);
 180         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
 181         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 182         struct sock *sk2;
 183         const struct hlist_node *node;
 184         struct inet_timewait_sock *tw;
 185
 186         prefetch(head->chain.first);
 187         write_lock(&head->lock);
 188
 189         /* Check TIME-WAIT sockets first. */
 190         sk_for_each(sk2, node, &(head + hinfo->ehash_size)->chain) {
 191                 tw = inet_twsk(sk2);
 192
 193                 if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
 194                         if (twsk_unique(sk, sk2, twp))
 195                                 goto unique;
 196                         else
 197                                 goto not_unique;
 198                 }
 199         }
 200         tw = NULL;
 201
 202         /* And established part... */
 203         sk_for_each(sk2, node, &head->chain) {
 204                 if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
 205                         goto not_unique;
 206         }
 207
 208 unique:
 209         /* Must record num and sport now. Otherwise we will see
 210          * in hash table socket with a funny identity. */
 211         inet->num = lport;
 212         inet->sport = htons(lport);
 213         sk->sk_hash = hash;
 214         BUG_TRAP(sk_unhashed(sk));
 215         __sk_add_node(sk, &head->chain);
 216         sock_prot_inc_use(sk->sk_prot);
 217         write_unlock(&head->lock);
 218
 219         if (twp) {
 220                 *twp = tw;
 221                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 222         } else if (tw) {
 223                 /* Silly. Should hash-dance instead... */
 224                 inet_twsk_deschedule(tw, death_row);
 225                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 226
 227                 inet_twsk_put(tw);
 228         }
 229
 230         return 0;
 231
 232 not_unique:
 233         write_unlock(&head->lock);
 234         return -EADDRNOTAVAIL;
 235 }
 236
 237 static inline u32 inet_sk_port_offset(const struct sock *sk)
 238 {
 239         const struct inet_sock *inet = inet_sk(sk);
 240         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
 241                                           inet->dport);
 242 }
 243
 244 /*
 245  * Bind a port for a connect operation and hash it.
 246  */
 247 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 248                       struct sock *sk)
 249 {
 250         struct inet_hashinfo *hinfo = death_row->hashinfo;
 251         const unsigned short snum = inet_sk(sk)->num;
 252         struct inet_bind_hashbucket *head;
 253         struct inet_bind_bucket *tb;
 254         int ret;
 255
 256         if (!snum) {
 257                 int low = sysctl_local_port_range[0];
 258                 int high = sysctl_local_port_range[1];
 259                 int range = high - low;
 260                 int i;
 261                 int port;
 262                 static u32 hint;
 263                 u32 offset = hint + inet_sk_port_offset(sk);
 264                 struct hlist_node *node;
 265                 struct inet_timewait_sock *tw = NULL;
 266
 267                 local_bh_disable();
 268                 for (i = 1; i <= range; i++) {
 269                         port = low + (i + offset) % range;
 270                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
 271                         spin_lock(&head->lock);
 272
 273                         /* Does not bother with rcv_saddr checks,
 274                          * because the established check is already
 275                          * unique enough.
 276                          */
 277                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 278                                 if (tb->port == port) {
 279                                         BUG_TRAP(!hlist_empty(&tb->owners));
 280                                         if (tb->fastreuse >= 0)
 281                                                 goto next_port;
 282                                         if (!__inet_check_established(death_row,
 283                                                                       sk, port,
 284                                                                       &tw))
 285                                                 goto ok;
 286                                         goto next_port;
 287                                 }
 288                         }
 289
 290                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, head, port);
 291                         if (!tb) {
 292                                 spin_unlock(&head->lock);
 293                                 break;
 294                         }
 295                         tb->fastreuse = -1;
 296                         goto ok;
 297
 298                 next_port:
 299                         spin_unlock(&head->lock);
 300                 }
 301                 local_bh_enable();
 302
 303                 return -EADDRNOTAVAIL;
 304
 305 ok:
 306                 hint += i;
 307
 308                 /* Head lock still held and bh's disabled */
 309                 inet_bind_hash(sk, tb, port);
 310                 if (sk_unhashed(sk)) {
 311                         inet_sk(sk)->sport = htons(port);
 312                         __inet_hash(hinfo, sk, 0);
 313                 }
 314                 spin_unlock(&head->lock);
 315
 316                 if (tw) {
 317                         inet_twsk_deschedule(tw, death_row);;
 318                         inet_twsk_put(tw);
 319                 }
 320
 321                 ret = 0;
 322                 goto out;
 323         }
 324
 325         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
 326         tb  = inet_csk(sk)->icsk_bind_hash;
 327         spin_lock_bh(&head->lock);
 328         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 329                 __inet_hash(hinfo, sk, 0);
 330                 spin_unlock_bh(&head->lock);
 331                 return 0;
 332         } else {
 333                 spin_unlock(&head->lock);
 334                 /* No definite answer... Walk to established hash table */
 335                 ret = __inet_check_established(death_row, sk, snum, NULL);
 336 out:
 337                 local_bh_enable();
 338                 return ret;
 339         }
 340 }
 341
 342 EXPORT_SYMBOL_GPL(inet_hash_connect);