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(kmem_cache_t *cachep,
  31                                                  struct inet_bind_hashbucket *head,
  32                                                  const unsigned short snum)
  33 {
  34         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, SLAB_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(kmem_cache_t *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 struct sock *__inet_lookup_listener(const struct hlist_head *head, const u32 daddr,
 128                                     const unsigned short hnum, const int dif)
 129 {
 130         struct sock *result = NULL, *sk;
 131         const struct hlist_node *node;
 132         int hiscore = -1;
 133
 134         sk_for_each(sk, node, head) {
 135                 const struct inet_sock *inet = inet_sk(sk);
 136
 137                 if (inet->num == hnum && !ipv6_only_sock(sk)) {
 138                         const __u32 rcv_saddr = inet->rcv_saddr;
 139                         int score = sk->sk_family == PF_INET ? 1 : 0;
 140
 141                         if (inet_addr_match(sk->sk_nx_info, daddr, rcv_saddr))
 142                                 score += 2;
 143                         else
 144                                 continue;
 145                         if (sk->sk_bound_dev_if) {
 146                                 if (sk->sk_bound_dev_if != dif)
 147                                         continue;
 148                                 score += 2;
 149                         }
 150                         if (score == 5)
 151                                 return sk;
 152                         if (score > hiscore) {
 153                                 hiscore = score;
 154                                 result  = sk;
 155                         }
 156                 }
 157         }
 158         return result;
 159 }
 160
 161 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 162
 163 /* called with local bh disabled */
 164 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 165                                     struct sock *sk, __u16 lport,
 166                                     struct inet_timewait_sock **twp)
 167 {
 168         struct inet_hashinfo *hinfo = death_row->hashinfo;
 169         struct inet_sock *inet = inet_sk(sk);
 170         u32 daddr = inet->rcv_saddr;
 171         u32 saddr = inet->daddr;
 172         int dif = sk->sk_bound_dev_if;
 173         INET_ADDR_COOKIE(acookie, saddr, daddr)
 174         const __u32 ports = INET_COMBINED_PORTS(inet->dport, lport);
 175         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
 176         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 177         struct sock *sk2;
 178         const struct hlist_node *node;
 179         struct inet_timewait_sock *tw;
 180
 181         prefetch(head->chain.first);
 182         write_lock(&head->lock);
 183
 184         /* Check TIME-WAIT sockets first. */
 185         sk_for_each(sk2, node, &(head + hinfo->ehash_size)->chain) {
 186                 tw = inet_twsk(sk2);
 187
 188                 if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
 189                         if (twsk_unique(sk, sk2, twp))
 190                                 goto unique;
 191                         else
 192                                 goto not_unique;
 193                 }
 194         }
 195         tw = NULL;
 196
 197         /* And established part... */
 198         sk_for_each(sk2, node, &head->chain) {
 199                 if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
 200                         goto not_unique;
 201         }
 202
 203 unique:
 204         /* Must record num and sport now. Otherwise we will see
 205          * in hash table socket with a funny identity. */
 206         inet->num = lport;
 207         inet->sport = htons(lport);
 208         sk->sk_hash = hash;
 209         BUG_TRAP(sk_unhashed(sk));
 210         __sk_add_node(sk, &head->chain);
 211         sock_prot_inc_use(sk->sk_prot);
 212         write_unlock(&head->lock);
 213
 214         if (twp) {
 215                 *twp = tw;
 216                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 217         } else if (tw) {
 218                 /* Silly. Should hash-dance instead... */
 219                 inet_twsk_deschedule(tw, death_row);
 220                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 221
 222                 inet_twsk_put(tw);
 223         }
 224
 225         return 0;
 226
 227 not_unique:
 228         write_unlock(&head->lock);
 229         return -EADDRNOTAVAIL;
 230 }
 231
 232 static inline u32 inet_sk_port_offset(const struct sock *sk)
 233 {
 234         const struct inet_sock *inet = inet_sk(sk);
 235         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
 236                                           inet->dport);
 237 }
 238
 239 /*
 240  * Bind a port for a connect operation and hash it.
 241  */
 242 int inet_hash_connect(struct inet_timewait_death_row *death_row,
 243                       struct sock *sk)
 244 {
 245         struct inet_hashinfo *hinfo = death_row->hashinfo;
 246         const unsigned short snum = inet_sk(sk)->num;
 247         struct inet_bind_hashbucket *head;
 248         struct inet_bind_bucket *tb;
 249         int ret;
 250
 251         if (!snum) {
 252                 int low = sysctl_local_port_range[0];
 253                 int high = sysctl_local_port_range[1];
 254                 int range = high - low;
 255                 int i;
 256                 int port;
 257                 static u32 hint;
 258                 u32 offset = hint + inet_sk_port_offset(sk);
 259                 struct hlist_node *node;
 260                 struct inet_timewait_sock *tw = NULL;
 261
 262                 local_bh_disable();
 263                 for (i = 1; i <= range; i++) {
 264                         port = low + (i + offset) % range;
 265                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
 266                         spin_lock(&head->lock);
 267
 268                         /* Does not bother with rcv_saddr checks,
 269                          * because the established check is already
 270                          * unique enough.
 271                          */
 272                         inet_bind_bucket_for_each(tb, node, &head->chain) {
 273                                 if (tb->port == port) {
 274                                         BUG_TRAP(!hlist_empty(&tb->owners));
 275                                         if (tb->fastreuse >= 0)
 276                                                 goto next_port;
 277                                         if (!__inet_check_established(death_row,
 278                                                                       sk, port,
 279                                                                       &tw))
 280                                                 goto ok;
 281                                         goto next_port;
 282                                 }
 283                         }
 284
 285                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, head, port);
 286                         if (!tb) {
 287                                 spin_unlock(&head->lock);
 288                                 break;
 289                         }
 290                         tb->fastreuse = -1;
 291                         goto ok;
 292
 293                 next_port:
 294                         spin_unlock(&head->lock);
 295                 }
 296                 local_bh_enable();
 297
 298                 return -EADDRNOTAVAIL;
 299
 300 ok:
 301                 hint += i;
 302
 303                 /* Head lock still held and bh's disabled */
 304                 inet_bind_hash(sk, tb, port);
 305                 if (sk_unhashed(sk)) {
 306                         inet_sk(sk)->sport = htons(port);
 307                         __inet_hash(hinfo, sk, 0);
 308                 }
 309                 spin_unlock(&head->lock);
 310
 311                 if (tw) {
 312                         inet_twsk_deschedule(tw, death_row);
 313                         inet_twsk_put(tw);
 314                 }
 315
 316                 ret = 0;
 317                 goto out;
 318         }
 319
 320         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
 321         tb  = inet_csk(sk)->icsk_bind_hash;
 322         spin_lock_bh(&head->lock);
 323         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 324                 __inet_hash(hinfo, sk, 0);
 325                 spin_unlock_bh(&head->lock);
 326                 return 0;
 327         } else {
 328                 spin_unlock(&head->lock);
 329                 /* No definite answer... Walk to established hash table */
 330                 ret = __inet_check_established(death_row, sk, snum, NULL);
 331 out:
 332                 local_bh_enable();
 333                 return ret;
 334         }
 335 }
 336
 337 EXPORT_SYMBOL_GPL(inet_hash_connect);