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
  18 #include <linux/interrupt.h>
  19 #include <linux/ipv6.h>
  20 #include <linux/list.h>
  21 #include <linux/slab.h>
  22 #include <linux/socket.h>
  23 #include <linux/spinlock.h>
  24 #include <linux/types.h>
  25 #include <linux/wait.h>
  26
  27 #include <net/inet_connection_sock.h>
  28 #include <net/inet_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  * One chain is dedicated to TIME_WAIT sockets.
  37  * I'll experiment with dynamic table growth later.
  38  */
  39 struct inet_ehash_bucket {
  40         struct hlist_head chain;
  41         struct hlist_head twchain;
  42 };
  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          * TIME_WAIT sockets use a separate chain (twchain).
 100          */
 101         struct inet_ehash_bucket        *ehash;
 102         rwlock_t                        *ehash_locks;
 103         unsigned int                    ehash_size;
 104         unsigned int                    ehash_locks_mask;
 105
 106         /* Ok, let's try this, I give up, we do need a local binding
 107          * TCP hash as well as the others for fast bind/connect.
 108          */
 109         struct inet_bind_hashbucket     *bhash;
 110
 111         unsigned int                    bhash_size;
 112         /* Note : 4 bytes padding on 64 bit arches */
 113
 114         /* All sockets in TCP_LISTEN state will be in here.  This is the only
 115          * table where wildcard'd TCP sockets can exist.  Hash function here
 116          * is just local port number.
 117          */
 118         struct hlist_head               listening_hash[INET_LHTABLE_SIZE];
 119
 120         /* All the above members are written once at bootup and
 121          * never written again _or_ are predominantly read-access.
 122          *
 123          * Now align to a new cache line as all the following members
 124          * are often dirty.
 125          */
 126         rwlock_t                        lhash_lock ____cacheline_aligned;
 127         atomic_t                        lhash_users;
 128         wait_queue_head_t               lhash_wait;
 129         struct kmem_cache                       *bind_bucket_cachep;
 130 };
 131
 132 static inline struct inet_ehash_bucket *inet_ehash_bucket(
 133         struct inet_hashinfo *hashinfo,
 134         unsigned int hash)
 135 {
 136         return &hashinfo->ehash[hash & (hashinfo->ehash_size - 1)];
 137 }
 138
 139 static inline rwlock_t *inet_ehash_lockp(
 140         struct inet_hashinfo *hashinfo,
 141         unsigned int hash)
 142 {
 143         return &hashinfo->ehash_locks[hash & hashinfo->ehash_locks_mask];
 144 }
 145
 146 static inline int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 147 {
 148         unsigned int i, size = 256;
 149 #if defined(CONFIG_PROVE_LOCKING)
 150         unsigned int nr_pcpus = 2;
 151 #else
 152         unsigned int nr_pcpus = num_possible_cpus();
 153 #endif
 154         if (nr_pcpus >= 4)
 155                 size = 512;
 156         if (nr_pcpus >= 8)
 157                 size = 1024;
 158         if (nr_pcpus >= 16)
 159                 size = 2048;
 160         if (nr_pcpus >= 32)
 161                 size = 4096;
 162         if (sizeof(rwlock_t) != 0) {
 163 #ifdef CONFIG_NUMA
 164                 if (size * sizeof(rwlock_t) > PAGE_SIZE)
 165                         hashinfo->ehash_locks = vmalloc(size * sizeof(rwlock_t));
 166                 else
 167 #endif
 168                 hashinfo->ehash_locks = kmalloc(size * sizeof(rwlock_t),
 169                                                 GFP_KERNEL);
 170                 if (!hashinfo->ehash_locks)
 171                         return ENOMEM;
 172                 for (i = 0; i < size; i++)
 173                         rwlock_init(&hashinfo->ehash_locks[i]);
 174         }
 175         hashinfo->ehash_locks_mask = size - 1;
 176         return 0;
 177 }
 178
 179 static inline void inet_ehash_locks_free(struct inet_hashinfo *hashinfo)
 180 {
 181         if (hashinfo->ehash_locks) {
 182 #ifdef CONFIG_NUMA
 183                 unsigned int size = (hashinfo->ehash_locks_mask + 1) *
 184                                                         sizeof(rwlock_t);
 185                 if (size > PAGE_SIZE)
 186                         vfree(hashinfo->ehash_locks);
 187                 else
 188 #else
 189                 kfree(hashinfo->ehash_locks);
 190 #endif
 191                 hashinfo->ehash_locks = NULL;
 192         }
 193 }
 194
 195 extern struct inet_bind_bucket *
 196                     inet_bind_bucket_create(struct kmem_cache *cachep,
 197                                             struct inet_bind_hashbucket *head,
 198                                             const unsigned short snum);
 199 extern void inet_bind_bucket_destroy(struct kmem_cache *cachep,
 200                                      struct inet_bind_bucket *tb);
 201
 202 static inline int inet_bhashfn(const __u16 lport, const int bhash_size)
 203 {
 204         return lport & (bhash_size - 1);
 205 }
 206
 207 extern void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 208                            const unsigned short snum);
 209
 210 /* These can have wildcards, don't try too hard. */
 211 static inline int inet_lhashfn(const unsigned short num)
 212 {
 213         return num & (INET_LHTABLE_SIZE - 1);
 214 }
 215
 216 static inline int inet_sk_listen_hashfn(const struct sock *sk)
 217 {
 218         return inet_lhashfn(inet_sk(sk)->num);
 219 }
 220
 221 /* Caller must disable local BH processing. */
 222 static inline void __inet_inherit_port(struct inet_hashinfo *table,
 223                                        struct sock *sk, struct sock *child)
 224 {
 225         const int bhash = inet_bhashfn(inet_sk(child)->num, table->bhash_size);
 226         struct inet_bind_hashbucket *head = &table->bhash[bhash];
 227         struct inet_bind_bucket *tb;
 228
 229         spin_lock(&head->lock);
 230         tb = inet_csk(sk)->icsk_bind_hash;
 231         sk_add_bind_node(child, &tb->owners);
 232         inet_csk(child)->icsk_bind_hash = tb;
 233         spin_unlock(&head->lock);
 234 }
 235
 236 static inline void inet_inherit_port(struct inet_hashinfo *table,
 237                                      struct sock *sk, struct sock *child)
 238 {
 239         local_bh_disable();
 240         __inet_inherit_port(table, sk, child);
 241         local_bh_enable();
 242 }
 243
 244 extern void inet_put_port(struct inet_hashinfo *table, struct sock *sk);
 245
 246 extern void inet_listen_wlock(struct inet_hashinfo *hashinfo);
 247
 248 /*
 249  * - We may sleep inside this lock.
 250  * - If sleeping is not required (or called from BH),
 251  *   use plain read_(un)lock(&inet_hashinfo.lhash_lock).
 252  */
 253 static inline void inet_listen_lock(struct inet_hashinfo *hashinfo)
 254 {
 255         /* read_lock synchronizes to candidates to writers */
 256         read_lock(&hashinfo->lhash_lock);
 257         atomic_inc(&hashinfo->lhash_users);
 258         read_unlock(&hashinfo->lhash_lock);
 259 }
 260
 261 static inline void inet_listen_unlock(struct inet_hashinfo *hashinfo)
 262 {
 263         if (atomic_dec_and_test(&hashinfo->lhash_users))
 264                 wake_up(&hashinfo->lhash_wait);
 265 }
 266
 267 static inline void __inet_hash(struct inet_hashinfo *hashinfo,
 268                                struct sock *sk, const int listen_possible)
 269 {
 270         struct hlist_head *list;
 271         rwlock_t *lock;
 272
 273         BUG_TRAP(sk_unhashed(sk));
 274         if (listen_possible && sk->sk_state == TCP_LISTEN) {
 275                 list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
 276                 lock = &hashinfo->lhash_lock;
 277                 inet_listen_wlock(hashinfo);
 278         } else {
 279                 struct inet_ehash_bucket *head;
 280                 sk->sk_hash = inet_sk_ehashfn(sk);
 281                 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 282                 list = &head->chain;
 283                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 284                 write_lock(lock);
 285         }
 286         __sk_add_node(sk, list);
 287         sock_prot_inc_use(sk->sk_prot);
 288         write_unlock(lock);
 289         if (listen_possible && sk->sk_state == TCP_LISTEN)
 290                 wake_up(&hashinfo->lhash_wait);
 291 }
 292
 293 static inline void inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk)
 294 {
 295         if (sk->sk_state != TCP_CLOSE) {
 296                 local_bh_disable();
 297                 __inet_hash(hashinfo, sk, 1);
 298                 local_bh_enable();
 299         }
 300 }
 301
 302 static inline void inet_unhash(struct inet_hashinfo *hashinfo, struct sock *sk)
 303 {
 304         rwlock_t *lock;
 305
 306         if (sk_unhashed(sk))
 307                 goto out;
 308
 309         if (sk->sk_state == TCP_LISTEN) {
 310                 local_bh_disable();
 311                 inet_listen_wlock(hashinfo);
 312                 lock = &hashinfo->lhash_lock;
 313         } else {
 314                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 315                 write_lock_bh(lock);
 316         }
 317
 318         if (__sk_del_node_init(sk))
 319                 sock_prot_dec_use(sk->sk_prot);
 320         write_unlock_bh(lock);
 321 out:
 322         if (sk->sk_state == TCP_LISTEN)
 323                 wake_up(&hashinfo->lhash_wait);
 324 }
 325
 326 extern struct sock *__inet_lookup_listener(struct inet_hashinfo *hashinfo,
 327                                            const __be32 daddr,
 328                                            const unsigned short hnum,
 329                                            const int dif);
 330
 331 static inline struct sock *inet_lookup_listener(struct inet_hashinfo *hashinfo,
 332                                                 __be32 daddr, __be16 dport, int dif)
 333 {
 334         return __inet_lookup_listener(hashinfo, daddr, ntohs(dport), dif);
 335 }
 336
 337 /* Socket demux engine toys. */
 338 /* What happens here is ugly; there's a pair of adjacent fields in
 339    struct inet_sock; __be16 dport followed by __u16 num.  We want to
 340    search by pair, so we combine the keys into a single 32bit value
 341    and compare with 32bit value read from &...->dport.  Let's at least
 342    make sure that it's not mixed with anything else...
 343    On 64bit targets we combine comparisons with pair of adjacent __be32
 344    fields in the same way.
 345 */
 346 typedef __u32 __bitwise __portpair;
 347 #ifdef __BIG_ENDIAN
 348 #define INET_COMBINED_PORTS(__sport, __dport) \
 349         ((__force __portpair)(((__force __u32)(__be16)(__sport) << 16) | (__u32)(__dport)))
 350 #else /* __LITTLE_ENDIAN */
 351 #define INET_COMBINED_PORTS(__sport, __dport) \
 352         ((__force __portpair)(((__u32)(__dport) << 16) | (__force __u32)(__be16)(__sport)))
 353 #endif
 354
 355 #if (BITS_PER_LONG == 64)
 356 typedef __u64 __bitwise __addrpair;
 357 #ifdef __BIG_ENDIAN
 358 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
 359         const __addrpair __name = (__force __addrpair) ( \
 360                                    (((__force __u64)(__be32)(__saddr)) << 32) | \
 361                                    ((__force __u64)(__be32)(__daddr)));
 362 #else /* __LITTLE_ENDIAN */
 363 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
 364         const __addrpair __name = (__force __addrpair) ( \
 365                                    (((__force __u64)(__be32)(__daddr)) << 32) | \
 366                                    ((__force __u64)(__be32)(__saddr)));
 367 #endif /* __BIG_ENDIAN */
 368 #define INET_MATCH(__sk, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
 369         (((__sk)->sk_hash == (__hash))                          &&      \
 370          ((*((__addrpair *)&(inet_sk(__sk)->daddr))) == (__cookie))     &&      \
 371          ((*((__portpair *)&(inet_sk(__sk)->dport))) == (__ports))      &&      \
 372          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 373 #define INET_TW_MATCH(__sk, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
 374         (((__sk)->sk_hash == (__hash))                          &&      \
 375          ((*((__addrpair *)&(inet_twsk(__sk)->tw_daddr))) == (__cookie)) &&     \
 376          ((*((__portpair *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&      \
 377          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 378 #else /* 32-bit arch */
 379 #define INET_ADDR_COOKIE(__name, __saddr, __daddr)
 380 #define INET_MATCH(__sk, __hash, __cookie, __saddr, __daddr, __ports, __dif)    \
 381         (((__sk)->sk_hash == (__hash))                          &&      \
 382          (inet_sk(__sk)->daddr          == (__saddr))           &&      \
 383          (inet_sk(__sk)->rcv_saddr      == (__daddr))           &&      \
 384          ((*((__portpair *)&(inet_sk(__sk)->dport))) == (__ports))      &&      \
 385          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 386 #define INET_TW_MATCH(__sk, __hash,__cookie, __saddr, __daddr, __ports, __dif)  \
 387         (((__sk)->sk_hash == (__hash))                          &&      \
 388          (inet_twsk(__sk)->tw_daddr     == (__saddr))           &&      \
 389          (inet_twsk(__sk)->tw_rcv_saddr == (__daddr))           &&      \
 390          ((*((__portpair *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&      \
 391          (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
 392 #endif /* 64-bit arch */
 393
 394 /*
 395  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so we need
 396  * not check it for lookups anymore, thanks Alexey. -DaveM
 397  *
 398  * Local BH must be disabled here.
 399  */
 400 static inline struct sock *
 401         __inet_lookup_established(struct inet_hashinfo *hashinfo,
 402                                   const __be32 saddr, const __be16 sport,
 403                                   const __be32 daddr, const u16 hnum,
 404                                   const int dif)
 405 {
 406         INET_ADDR_COOKIE(acookie, saddr, daddr)
 407         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 408         struct sock *sk;
 409         const struct hlist_node *node;
 410         /* Optimize here for direct hit, only listening connections can
 411          * have wildcards anyways.
 412          */
 413         unsigned int hash = inet_ehashfn(daddr, hnum, saddr, sport);
 414         struct inet_ehash_bucket *head = inet_ehash_bucket(hashinfo, hash);
 415         rwlock_t *lock = inet_ehash_lockp(hashinfo, hash);
 416
 417         prefetch(head->chain.first);
 418         read_lock(lock);
 419         sk_for_each(sk, node, &head->chain) {
 420                 if (INET_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 421                         goto hit; /* You sunk my battleship! */
 422         }
 423
 424         /* Must check for a TIME_WAIT'er before going to listener hash. */
 425         sk_for_each(sk, node, &head->twchain) {
 426                 if (INET_TW_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
 427                         goto hit;
 428         }
 429         sk = NULL;
 430 out:
 431         read_unlock(lock);
 432         return sk;
 433 hit:
 434         sock_hold(sk);
 435         goto out;
 436 }
 437
 438 static inline struct sock *
 439         inet_lookup_established(struct inet_hashinfo *hashinfo,
 440                                 const __be32 saddr, const __be16 sport,
 441                                 const __be32 daddr, const __be16 dport,
 442                                 const int dif)
 443 {
 444         return __inet_lookup_established(hashinfo, saddr, sport, daddr,
 445                                          ntohs(dport), dif);
 446 }
 447
 448 static inline struct sock *__inet_lookup(struct inet_hashinfo *hashinfo,
 449                                          const __be32 saddr, const __be16 sport,
 450                                          const __be32 daddr, const __be16 dport,
 451                                          const int dif)
 452 {
 453         u16 hnum = ntohs(dport);
 454         struct sock *sk = __inet_lookup_established(hashinfo, saddr, sport, daddr,
 455                                                     hnum, dif);
 456         return sk ? : __inet_lookup_listener(hashinfo, daddr, hnum, dif);
 457 }
 458
 459 static inline struct sock *inet_lookup(struct inet_hashinfo *hashinfo,
 460                                        const __be32 saddr, const __be16 sport,
 461                                        const __be32 daddr, const __be16 dport,
 462                                        const int dif)
 463 {
 464         struct sock *sk;
 465
 466         local_bh_disable();
 467         sk = __inet_lookup(hashinfo, saddr, sport, daddr, dport, dif);
 468         local_bh_enable();
 469
 470         return sk;
 471 }
 472
 473 extern int inet_hash_connect(struct inet_timewait_death_row *death_row,
 474                              struct sock *sk);
 475 #endif /* _INET_HASHTABLES_H */