include/net/inet_hashtables.h

   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  * Authors:     Lotsa people, from code originally in tcp
   7  *
   8  *      This program is free software; you can redistribute it and/or
   9  *      modify it under the terms of the GNU General Public License
  10  *      as published by the Free Software Foundation; either version
  11  *      2 of the License, or (at your option) any later version.
  12  */
  13
  14 #ifndef _INET_HASHTABLES_H
  15 #define _INET_HASHTABLES_H
  16
  17 #include <linux/config.h>
  18
  19 #include <linux/interrupt.h>
  20 #include <linux/ipv6.h>
  21 #include <linux/list.h>
  22 #include <linux/slab.h>
  23 #include <linux/socket.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/types.h>
  26 #include <linux/wait.h>
  27
  28 #include <net/inet_connection_sock.h>
  29 #include <net/sock.h>
  30 #include <net/tcp_states.h>
  31
  32 #include <asm/atomic.h>
  33 #include <asm/byteorder.h>
  34
  35 /* This is for all connections with a full identity, no wildcards.
  36  * New scheme, half the table is for TIME_WAIT, the other half is
  37  * for the rest.  I'll experiment with dynamic table growth later.
  38  */
  39 struct inet_ehash_bucket {
  40         rwlock_t          lock;
  41         struct hlist_head chain;
  42 } __attribute__((__aligned__(8)));
  43
  44 /* There are a few simple rules, which allow for local port reuse by
  45  * an application.  In essence:
  46  *
  47  *      1) Sockets bound to different interfaces may share a local port.
  48  *         Failing that, goto test 2.
  49  *      2) If all sockets have sk->sk_reuse set, and none of them are in
  50  *         TCP_LISTEN state, the port may be shared.
  51  *         Failing that, goto test 3.
  52  *      3) If all sockets are bound to a specific inet_sk(sk)->rcv_saddr local
  53  *         address, and none of them are the same, the port may be
  54  *         shared.
  55  *         Failing this, the port cannot be shared.
  56  *
  57  * The interesting point, is test #2.  This is what an FTP server does
  58  * all day.  To optimize this case we use a specific flag bit defined
  59  * below.  As we add sockets to a bind bucket list, we perform a
  60  * check of: (newsk->sk_reuse && (newsk->sk_state != TCP_LISTEN))
  61  * As long as all sockets added to a bind bucket pass this test,
  62  * the flag bit will be set.
  63  * The resulting situation is that tcp_v[46]_verify_bind() can just check
  64  * for this flag bit, if it is set and the socket trying to bind has
  65  * sk->sk_reuse set, we don't even have to walk the owners list at all,
  66  * we return that it is ok to bind this socket to the requested local port.
  67  *
  68  * Sounds like a lot of work, but it is worth it.  In a more naive
  69  * implementation (ie. current FreeBSD etc.) the entire list of ports
  70  * must be walked for each data port opened by an ftp server.  Needless
  71  * to say, this does not scale at all.  With a couple thousand FTP
  72  * users logged onto your box, isn't it nice to know that new data
  73  * ports are created in O(1) time?  I thought so. ;-)   -DaveM
  74  */
  75 struct inet_bind_bucket {
  76         unsigned short          port;
  77         signed short            fastreuse;
  78         struct hlist_node       node;
  79         struct hlist_head       owners;
  80 };
  81
  82 #define inet_bind_bucket_for_each(tb, node, head) \
  83         hlist_for_each_entry(tb, node, head, node)
  84
  85 struct inet_bind_hashbucket {
  86         spinlock_t              lock;
  87         struct hlist_head       chain;
  88 };
  89
  90 /* This is for listening sockets, thus all sockets which possess wildcards. */
  91 #define INET_LHTABLE_SIZE       32      /* Yes, really, this is all you need. */
  92
  93 struct inet_hashinfo {
  94         /* This is for sockets with full identity only.  Sockets here will
  95          * always be without wildcards and will have the following invariant:
  96          *
  97          *          TCP_ESTABLISHED <= sk->sk_state < TCP_CLOSE
  98          *
  99          * First half of the table is for sockets not in TIME_WAIT, second half
 100          * is for TIME_WAIT sockets only.
 101          */
 102         struct inet_ehash_bucket        *ehash;
 103
 104         /* Ok, let's try this, I give up, we do need a local binding
 105          * TCP hash as well as the others for fast bind/connect.
 106          */
 107         struct inet_bind_hashbucket     *bhash;
 108
 109         int                             bhash_size;
 110         int                             ehash_size;
 111
 112         /* All sockets in TCP_LISTEN state will be in here.  This is the only
 113          * table where wildcard'd TCP sockets can exist.  Hash function here
 114          * is just local port number.
 115          */
 116         struct hlist_head               listening_hash[INET_LHTABLE_SIZE];
 117
 118         /* All the above members are written once at bootup and
 119          * never written again _or_ are predominantly read-access.
 120          *
 121          * Now align to a new cache line as all the following members
 122          * are often dirty.
 123          */
 124         rwlock_t                        lhash_lock ____cacheline_aligned;
 125         atomic_t                        lhash_users;
 126         wait_queue_head_t               lhash_wait;
 127         spinlock_t                      portalloc_lock;
 128         kmem_cache_t                    *bind_bucket_cachep;
 129         int                             port_rover;
 130 };
 131
 132 static inline int inet_ehashfn(const __u32 laddr, const __u16 lport,
 133                                const __u32 faddr, const __u16 fport,
 134                                const int ehash_size)
 135 {
 136         int h = (laddr ^ lport) ^ (faddr ^ fport);
 137         h ^= h >> 16;
 138         h ^= h >> 8;
 139         return h & (ehash_size - 1);
 140 }
 141
 142 static inline int inet_sk_ehashfn(const struct sock *sk, const int ehash_size)
 143 {
 144         const struct inet_sock *inet = inet_sk(sk);
 145         const __u32 laddr = inet->rcv_saddr;
 146         const __u16 lport = inet->num;
 147         const __u32 faddr = inet->daddr;
 148         const __u16 fport = inet->dport;
 149
 150         return inet_ehashfn(laddr, lport, faddr, fport, ehash_size);
 151 }
 152
 153 extern struct inet_bind_bucket *
 154                     inet_bind_bucket_create(kmem_cache_t *cachep,
 155                                             struct inet_bind_hashbucket *head,
 156                                             const unsigned short snum);
 157 extern void inet_bind_bucket_destroy(kmem_cache_t *cachep,
 158                                      struct inet_bind_bucket *tb);
 159
 160 static inline int inet_bhashfn(const __u16 lport, const int bhash_size)
 161 {
 162         return lport & (bhash_size - 1);
 163 }
 164
 165 extern void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 166                            const unsigned short snum);
 167
 168 /* These can have wildcards, don't try too hard. */
 169 static inline int inet_lhashfn(const unsigned short num)
 170 {
 171         return num & (INET_LHTABLE_SIZE - 1);
 172 }
 173
 174 static inline int inet_sk_listen_hashfn(const struct sock *sk)
 175 {
 176         return inet_lhashfn(inet_sk(sk)->num);
 177 }
 178
 179 /* Caller must disable local BH processing. */
 180 static inline void __inet_inherit_port(struct inet_hashinfo *table,
 181                                        struct sock *sk, struct sock *child)
 182 {
 183         const int bhash = inet_bhashfn(inet_sk(child)->num, table->bhash_size);
 184         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 185         struct inet_bind_bucket *tb;
 186
 187         spin_lock(&head->lock);
 188         tb = inet_csk(sk)->icsk_bind_hash;
 189         sk_add_bind_node(child, &tb->owners);
 190         inet_csk(child)->icsk_bind_hash = tb;
 191         spin_unlock(&head->lock);
 192 }
 193
 194 static inline void inet_inherit_port(struct inet_hashinfo *table,
 195                                      struct sock *sk, struct sock *child)
 196 {
 197         local_bh_disable();
 198         __inet_inherit_port(table, sk, child);
 199         local_bh_enable();
 200 }
 201
 202 extern void inet_put_port(struct inet_hashinfo *table, struct sock *sk);
 203
 204 extern void inet_listen_wlock(struct inet_hashinfo *hashinfo);
 205
 206 /*
 207  * - We may sleep inside this lock.
 208  * - If sleeping is not required (or called from BH),
 209  *   use plain read_(un)lock(&inet_hashinfo.lhash_lock).
 210  */
 211 static inline void inet_listen_lock(struct inet_hashinfo *hashinfo)
 212 {
 213         /* read_lock synchronizes to candidates to writers */
 214         read_lock(&hashinfo->lhash_lock);
 215         atomic_inc(&hashinfo->lhash_users);
 216         read_unlock(&hashinfo->lhash_lock);
 217 }
 218
 219 static inline void inet_listen_unlock(struct inet_hashinfo *hashinfo)
 220 {
 221         if (atomic_dec_and_test(&hashinfo->lhash_users))
 222                 wake_up(&hashinfo->lhash_wait);
 223 }
 224
 225 static inline void __inet_hash(struct inet_hashinfo *hashinfo,
 226                                struct sock *sk, const int listen_possible)
 227 {
 228         struct hlist_head *list;
 229         rwlock_t *lock;
 230
 231         BUG_TRAP(sk_unhashed(sk));
 232         if (listen_possible && sk->sk_state == TCP_LISTEN) {
 233                 list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 234                 lock = &hashinfo->lhash_lock;
 235                 inet_listen_wlock(hashinfo);
 236         } else {
 237                 sk->sk_hashent = inet_sk_ehashfn(sk, hashinfo->ehash_size);
 238                 list = &hashinfo->ehash[sk->sk_hashent].chain;
 239                 lock = &hashinfo->ehash[sk->sk_hashent].lock;
 240                 write_lock(lock);
 241         }
 242         __sk_add_node(sk, list);
 243         sock_prot_inc_use(sk->sk_prot);
 244         write_unlock(lock);
 245         if (listen_possible && sk->sk_state == TCP_LISTEN)
 246                 wake_up(&hashinfo->lhash_wait);
 247 }
 248
 249 static inline void inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk)
 250 {
 251         if (sk->sk_state != TCP_CLOSE) {
 252                 local_bh_disable();
 253                 __inet_hash(hashinfo, sk, 1);
 254                 local_bh_enable();
 255         }
 256 }
 257
 258 static inline void inet_unhash(struct inet_hashinfo *hashinfo, struct sock *sk)
 259 {
 260         rwlock_t *lock;
 261
 262         if (sk_unhashed(sk))
 263                 goto out;
 264
 265         if (sk->sk_state == TCP_LISTEN) {
 266                 local_bh_disable();
 267                 inet_listen_wlock(hashinfo);
 268                 lock = &hashinfo->lhash_lock;
 269         } else {
 270                 struct inet_ehash_bucket *head = &hashinfo->ehash[sk->sk_hashent];
 271                 lock = &head->lock;
 272                 write_lock_bh(&head->lock);
 273         }
 274
 275         if (__sk_del_node_init(sk))
 276                 sock_prot_dec_use(sk->sk_prot);
 277         write_unlock_bh(lock);
 278 out:
 279         if (sk->sk_state == TCP_LISTEN)
 280                 wake_up(&hashinfo->lhash_wait);
 281 }
 282
 283 extern struct sock *__inet_lookup_listener(const struct hlist_head *head,
 284                                            const u32 daddr,
 285                                            const unsigned short hnum,
 286                                            const int dif);
 287
 288 /* Optimize the common listener case. */
 289 static inline struct sock *
 290                 inet_lookup_listener(struct inet_hashinfo *hashinfo,
 291                                      const u32 daddr,
 292                                      const unsigned short hnum, const int dif)
 293 {
 294         struct sock *sk = NULL;
 295         const struct hlist_head *head;
 296
 297         read_lock(&hashinfo->lhash_lock);
 298         head = &hashinfo->listening_hash[inet_lhashfn(hnum)];
 299         if (!hlist_empty(head)) {
 300                 const struct inet_sock *inet = inet_sk((sk = __sk_head(head)));
 301
 302                 if (inet->num == hnum && !sk->sk_node.next &&
 303                     (!inet->rcv_saddr || inet->rcv_saddr == daddr) &&
 304                     (sk->sk_family == PF_INET || !ipv6_only_sock(sk)) &&
 305                     !sk->sk_bound_dev_if)
 306                         goto sherry_cache;
 307                 sk = __inet_lookup_listener(head, daddr, hnum, dif);
 308         }
 309         if (sk) {
 310 sherry_cache:
 311                 sock_hold(sk);
 312         }
 313         read_unlock(&hashinfo->lhash_lock);
 314         return sk;
 315 }
 316
 317 /* Socket demux engine toys. */
 318 #ifdef __BIG_ENDIAN
 319 #define INET_COMBINED_PORTS(__sport, __dport) \
 320         (((__u32)(__sport) << 16) | (__u32)(__dport))
 321 #else /* __LITTLE_ENDIAN */
 322 #define INET_COMBINED_PORTS(__sport, __dport) \
 323         (((__u32)(__dport) << 16) | (__u32)(__sport))
 324 #endif
 325
 326 #if (BITS_PER_LONG == 64)
 327 #ifdef __BIG_ENDIAN
 328 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
 329         const __u64 __name = (((__u64)(__saddr)) << 32) | ((__u64)(__daddr));
 330 #else /* __LITTLE_ENDIAN */
 331 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
 332         const __u64 __name = (((__u64)(__daddr)) << 32) | ((__u64)(__saddr));
 333 #endif /* __BIG_ENDIAN */
 334 #define INET_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif)\
 335         (((*((__u64 *)&(inet_sk(__sk)->daddr))) == (__cookie))  &&      \
 336          ((*((__u32 *)&(inet_sk(__sk)->dport))) == (__ports))   &&      \
 337          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 338 #define INET_TW_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif)\
 339         (((*((__u64 *)&(inet_twsk(__sk)->tw_daddr))) == (__cookie)) &&  \
 340          ((*((__u32 *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&   \
 341          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 342 #else /* 32-bit arch */
 343 #define INET_ADDR_COOKIE(__name, __saddr, __daddr)
 344 #define INET_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif)    \
 345         ((inet_sk(__sk)->daddr          == (__saddr))           &&      \
 346          (inet_sk(__sk)->rcv_saddr      == (__daddr))           &&      \
 347          ((*((__u32 *)&(inet_sk(__sk)->dport))) == (__ports))   &&      \
 348          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 349 #define INET_TW_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif) \
 350         ((inet_twsk(__sk)->tw_daddr     == (__saddr))           &&      \
 351          (inet_twsk(__sk)->tw_rcv_saddr == (__daddr))           &&      \
 352          ((*((__u32 *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&   \
 353          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 354 #endif /* 64-bit arch */
 355
 356 /*
 357  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so we need
 358  * not check it for lookups anymore, thanks Alexey. -DaveM
 359  *
 360  * Local BH must be disabled here.
 361  */
 362 static inline struct sock *
 363         __inet_lookup_established(struct inet_hashinfo *hashinfo,
 364                                   const u32 saddr, const u16 sport,
 365                                   const u32 daddr, const u16 hnum,
 366                                   const int dif)
 367 {
 368         INET_ADDR_COOKIE(acookie, saddr, daddr)
 369         const __u32 ports = INET_COMBINED_PORTS(sport, hnum);
 370         struct sock *sk;
 371         const struct hlist_node *node;
 372         /* Optimize here for direct hit, only listening connections can
 373          * have wildcards anyways.
 374          */
 375         const int hash = inet_ehashfn(daddr, hnum, saddr, sport, hashinfo->ehash_size);
 376         struct inet_ehash_bucket *head = &hashinfo->ehash[hash];
 377
 378         read_lock(&head->lock);
 379         sk_for_each(sk, node, &head->chain) {
 380                 if (INET_MATCH(sk, acookie, saddr, daddr, ports, dif))
 381                         goto hit; /* You sunk my battleship! */
 382         }
 383
 384         /* Must check for a TIME_WAIT'er before going to listener hash. */
 385         sk_for_each(sk, node, &(head + hashinfo->ehash_size)->chain) {
 386                 if (INET_TW_MATCH(sk, acookie, saddr, daddr, ports, dif))
 387                         goto hit;
 388         }
 389         sk = NULL;
 390 out:
 391         read_unlock(&head->lock);
 392         return sk;
 393 hit:
 394         sock_hold(sk);
 395         goto out;
 396 }
 397
 398 static inline struct sock *__inet_lookup(struct inet_hashinfo *hashinfo,
 399                                          const u32 saddr, const u16 sport,
 400                                          const u32 daddr, const u16 hnum,
 401                                          const int dif)
 402 {
 403         struct sock *sk = __inet_lookup_established(hashinfo, saddr, sport, daddr,
 404                                                     hnum, dif);
 405         return sk ? : inet_lookup_listener(hashinfo, daddr, hnum, dif);
 406 }
 407
 408 static inline struct sock *inet_lookup(struct inet_hashinfo *hashinfo,
 409                                        const u32 saddr, const u16 sport,
 410                                        const u32 daddr, const u16 dport,
 411                                        const int dif)
 412 {
 413         struct sock *sk;
 414
 415         local_bh_disable();
 416         sk = __inet_lookup(hashinfo, saddr, sport, daddr, ntohs(dport), dif);
 417         local_bh_enable();
 418
 419         return sk;
 420 }
 421 #endif /* _INET_HASHTABLES_H */