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 (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 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 168
 169 /* called with local bh disabled */
 170 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 171                                     struct sock *sk, __u16 lport,
 172                                     struct inet_timewait_sock **twp)
 173 {
 174         struct inet_hashinfo *hinfo = death_row->hashinfo;
 175         struct inet_sock *inet = inet_sk(sk);
 176         u32 daddr = inet->rcv_saddr;
 177         u32 saddr = inet->daddr;
 178         int dif = sk->sk_bound_dev_if;
 179         INET_ADDR_COOKIE(acookie, saddr, daddr)
 180         const __u32 ports = INET_COMBINED_PORTS(inet->dport, lport);
 181         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
 182         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 183         struct sock *sk2;
 184         const struct hlist_node *node;
 185         struct inet_timewait_sock *tw;
 186
 187         prefetch(head->chain.first);
 188         write_lock(&head->lock);
 189
 190         /* Check TIME-WAIT sockets first. */
 191         sk_for_each(sk2, node, &(head + hinfo->ehash_size)->chain) {
 192                 tw = inet_twsk(sk2);
 193
 194                 if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
 195                         if (twsk_unique(sk, sk2, twp))
 196                                 goto unique;
 197                         else
 198                                 goto not_unique;
 199                 }
 200         }
 201         tw = NULL;
 202
 203         /* And established part... */
 204         sk_for_each(sk2, node, &head->chain) {
 205                 if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
 206                         goto not_unique;
 207         }
 208
 209 unique:
 210         /* Must record num and sport now. Otherwise we will see
 211          * in hash table socket with a funny identity. */
 212         inet->num = lport;
 213         inet->sport = htons(lport);
 214         sk->sk_hash = hash;
 215         BUG_TRAP(sk_unhashed(sk));
 216         __sk_add_node(sk, &head->chain);
 217         sock_prot_inc_use(sk->sk_prot);
 218         write_unlock(&head->lock);
 219
 220         if (twp) {
 221                 *twp = tw;
 222                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 223         } else if (tw) {
 224                 /* Silly. Should hash-dance instead... */
 225                 inet_twsk_deschedule(tw, death_row);
 226                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 227
 228                 inet_twsk_put(tw);
 229         }
 230
 231         return 0;
 232
 233 not_unique:
 234         write_unlock(&head->lock);
 235         return -EADDRNOTAVAIL;
 236 }
 237
 238 static inline u32 inet_sk_port_offset(const struct sock *sk)
 239 {
 240         const struct inet_sock *inet = inet_sk(sk);
 241         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
 242                                           inet->dport);
 243 }
 244
 245 /*
 246  * Bind a port for a connect operation and hash it.
 247  */
 248 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 249                       struct sock *sk)
 250 {
 251         struct inet_hashinfo *hinfo = death_row->hashinfo;
 252         const unsigned short snum = inet_sk(sk)->num;
 253         struct inet_bind_hashbucket *head;
 254         struct inet_bind_bucket *tb;
 255         int ret;
 256
 257         if (!snum) {
 258                 int low = sysctl_local_port_range[0];
 259                 int high = sysctl_local_port_range[1];
 260                 int range = high - low;
 261                 int i;
 262                 int port;
 263                 static u32 hint;
 264                 u32 offset = hint + inet_sk_port_offset(sk);
 265                 struct hlist_node *node;
 266                 struct inet_timewait_sock *tw = NULL;
 267
 268                 local_bh_disable();
 269                 for (i = 1; i <= range; i++) {
 270                         port = low + (i + offset) % range;
 271                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
 272                         spin_lock(&head->lock);
 273
 274                         /* Does not bother with rcv_saddr checks,
 275                          * because the established check is already
 276                          * unique enough.
 277                          */
 278                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 279                                 if (tb->port == port) {
 280                                         BUG_TRAP(!hlist_empty(&tb->owners));
 281                                         if (tb->fastreuse >= 0)
 282                                                 goto next_port;
 283                                         if (!__inet_check_established(death_row,
 284                                                                       sk, port,
 285                                                                       &tw))
 286                                                 goto ok;
 287                                         goto next_port;
 288                                 }
 289                         }
 290
 291                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, head, port);
 292                         if (!tb) {
 293                                 spin_unlock(&head->lock);
 294                                 break;
 295                         }
 296                         tb->fastreuse = -1;
 297                         goto ok;
 298
 299                 next_port:
 300                         spin_unlock(&head->lock);
 301                 }
 302                 local_bh_enable();
 303
 304                 return -EADDRNOTAVAIL;
 305
 306 ok:
 307                 hint += i;
 308
 309                 /* Head lock still held and bh's disabled */
 310                 inet_bind_hash(sk, tb, port);
 311                 if (sk_unhashed(sk)) {
 312                         inet_sk(sk)->sport = htons(port);
 313                         __inet_hash(hinfo, sk, 0);
 314                 }
 315                 spin_unlock(&head->lock);
 316
 317                 if (tw) {
 318                         inet_twsk_deschedule(tw, death_row);
 319                         inet_twsk_put(tw);
 320                 }
 321
 322                 ret = 0;
 323                 goto out;
 324         }
 325
 326         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
 327         tb  = inet_csk(sk)->icsk_bind_hash;
 328         spin_lock_bh(&head->lock);
 329         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 330                 __inet_hash(hinfo, sk, 0);
 331                 spin_unlock_bh(&head->lock);
 332                 return 0;
 333         } else {
 334                 spin_unlock(&head->lock);
 335                 /* No definite answer... Walk to established hash table */
 336                 ret = __inet_check_established(death_row, sk, snum, NULL);
 337 out:
 338                 local_bh_enable();
 339                 return ret;
 340         }
 341 }
 342
 343 EXPORT_SYMBOL_GPL(inet_hash_connect);