Pileus Git - ~andy/linux/blob - net/ipv4/inet_timewait_sock.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 TIME_WAIT sockets functions
   7  *
   8  *              From code orinally in TCP
   9  */
  10
  11 #include <linux/kernel.h>
  12 #include <linux/kmemcheck.h>
  13 #include <net/inet_hashtables.h>
  14 #include <net/inet_timewait_sock.h>
  15 #include <net/ip.h>
  16
  17
  18 /*
  19  * unhash a timewait socket from established hash
  20  * lock must be hold by caller
  21  */
  22 int inet_twsk_unhash(struct inet_timewait_sock *tw)
  23 {
  24         if (hlist_nulls_unhashed(&tw->tw_node))
  25                 return 0;
  26
  27         hlist_nulls_del_rcu(&tw->tw_node);
  28         sk_nulls_node_init(&tw->tw_node);
  29         return 1;
  30 }
  31
  32 /* Must be called with locally disabled BHs. */
  33 static void __inet_twsk_kill(struct inet_timewait_sock *tw,
  34                              struct inet_hashinfo *hashinfo)
  35 {
  36         struct inet_bind_hashbucket *bhead;
  37         struct inet_bind_bucket *tb;
  38         int refcnt;
  39         /* Unlink from established hashes. */
  40         spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
  41
  42         spin_lock(lock);
  43         refcnt = inet_twsk_unhash(tw);
  44         spin_unlock(lock);
  45
  46         /* Disassociate with bind bucket. */
  47         bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), tw->tw_num,
  48                         hashinfo->bhash_size)];
  49         spin_lock(&bhead->lock);
  50         tb = tw->tw_tb;
  51         if (tb) {
  52                 __hlist_del(&tw->tw_bind_node);
  53                 tw->tw_tb = NULL;
  54                 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  55                 refcnt++;
  56         }
  57         spin_unlock(&bhead->lock);
  58 #ifdef SOCK_REFCNT_DEBUG
  59         if (atomic_read(&tw->tw_refcnt) != 1) {
  60                 printk(KERN_DEBUG "%s timewait_sock %p refcnt=%d\n",
  61                        tw->tw_prot->name, tw, atomic_read(&tw->tw_refcnt));
  62         }
  63 #endif
  64         while (refcnt) {
  65                 inet_twsk_put(tw);
  66                 refcnt--;
  67         }
  68 }
  69
  70 static noinline void inet_twsk_free(struct inet_timewait_sock *tw)
  71 {
  72         struct module *owner = tw->tw_prot->owner;
  73         twsk_destructor((struct sock *)tw);
  74 #ifdef SOCK_REFCNT_DEBUG
  75         pr_debug("%s timewait_sock %p released\n", tw->tw_prot->name, tw);
  76 #endif
  77         release_net(twsk_net(tw));
  78         kmem_cache_free(tw->tw_prot->twsk_prot->twsk_slab, tw);
  79         module_put(owner);
  80 }
  81
  82 void inet_twsk_put(struct inet_timewait_sock *tw)
  83 {
  84         if (atomic_dec_and_test(&tw->tw_refcnt))
  85                 inet_twsk_free(tw);
  86 }
  87 EXPORT_SYMBOL_GPL(inet_twsk_put);
  88
  89 /*
  90  * Enter the time wait state. This is called with locally disabled BH.
  91  * Essentially we whip up a timewait bucket, copy the relevant info into it
  92  * from the SK, and mess with hash chains and list linkage.
  93  */
  94 void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
  95                            struct inet_hashinfo *hashinfo)
  96 {
  97         const struct inet_sock *inet = inet_sk(sk);
  98         const struct inet_connection_sock *icsk = inet_csk(sk);
  99         struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
 100         spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 101         struct inet_bind_hashbucket *bhead;
 102         /* Step 1: Put TW into bind hash. Original socket stays there too.
 103            Note, that any socket with inet->num != 0 MUST be bound in
 104            binding cache, even if it is closed.
 105          */
 106         bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), inet->inet_num,
 107                         hashinfo->bhash_size)];
 108         spin_lock(&bhead->lock);
 109         tw->tw_tb = icsk->icsk_bind_hash;
 110         WARN_ON(!icsk->icsk_bind_hash);
 111         inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
 112         atomic_inc(&tw->tw_refcnt);
 113         spin_unlock(&bhead->lock);
 114
 115         spin_lock(lock);
 116
 117         /*
 118          * Step 2: Hash TW into TIMEWAIT chain.
 119          * Should be done before removing sk from established chain
 120          * because readers are lockless and search established first.
 121          */
 122         atomic_inc(&tw->tw_refcnt);
 123         inet_twsk_add_node_rcu(tw, &ehead->twchain);
 124
 125         /* Step 3: Remove SK from established hash. */
 126         if (__sk_nulls_del_node_init_rcu(sk))
 127                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 128
 129         spin_unlock(lock);
 130 }
 131
 132 EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
 133
 134 struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
 135 {
 136         struct inet_timewait_sock *tw =
 137                 kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
 138                                  GFP_ATOMIC);
 139         if (tw != NULL) {
 140                 const struct inet_sock *inet = inet_sk(sk);
 141
 142                 kmemcheck_annotate_bitfield(tw, flags);
 143
 144                 /* Give us an identity. */
 145                 tw->tw_daddr        = inet->inet_daddr;
 146                 tw->tw_rcv_saddr    = inet->inet_rcv_saddr;
 147                 tw->tw_bound_dev_if = sk->sk_bound_dev_if;
 148                 tw->tw_num          = inet->inet_num;
 149                 tw->tw_state        = TCP_TIME_WAIT;
 150                 tw->tw_substate     = state;
 151                 tw->tw_sport        = inet->inet_sport;
 152                 tw->tw_dport        = inet->inet_dport;
 153                 tw->tw_family       = sk->sk_family;
 154                 tw->tw_reuse        = sk->sk_reuse;
 155                 tw->tw_hash         = sk->sk_hash;
 156                 tw->tw_ipv6only     = 0;
 157                 tw->tw_transparent  = inet->transparent;
 158                 tw->tw_prot         = sk->sk_prot_creator;
 159                 twsk_net_set(tw, hold_net(sock_net(sk)));
 160                 atomic_set(&tw->tw_refcnt, 1);
 161                 inet_twsk_dead_node_init(tw);
 162                 __module_get(tw->tw_prot->owner);
 163         }
 164
 165         return tw;
 166 }
 167
 168 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
 169
 170 /* Returns non-zero if quota exceeded.  */
 171 static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
 172                                     const int slot)
 173 {
 174         struct inet_timewait_sock *tw;
 175         struct hlist_node *node;
 176         unsigned int killed;
 177         int ret;
 178
 179         /* NOTE: compare this to previous version where lock
 180          * was released after detaching chain. It was racy,
 181          * because tw buckets are scheduled in not serialized context
 182          * in 2.3 (with netfilter), and with softnet it is common, because
 183          * soft irqs are not sequenced.
 184          */
 185         killed = 0;
 186         ret = 0;
 187 rescan:
 188         inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
 189                 __inet_twsk_del_dead_node(tw);
 190                 spin_unlock(&twdr->death_lock);
 191                 __inet_twsk_kill(tw, twdr->hashinfo);
 192 #ifdef CONFIG_NET_NS
 193                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
 194 #endif
 195                 inet_twsk_put(tw);
 196                 killed++;
 197                 spin_lock(&twdr->death_lock);
 198                 if (killed > INET_TWDR_TWKILL_QUOTA) {
 199                         ret = 1;
 200                         break;
 201                 }
 202
 203                 /* While we dropped twdr->death_lock, another cpu may have
 204                  * killed off the next TW bucket in the list, therefore
 205                  * do a fresh re-read of the hlist head node with the
 206                  * lock reacquired.  We still use the hlist traversal
 207                  * macro in order to get the prefetches.
 208                  */
 209                 goto rescan;
 210         }
 211
 212         twdr->tw_count -= killed;
 213 #ifndef CONFIG_NET_NS
 214         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITED, killed);
 215 #endif
 216         return ret;
 217 }
 218
 219 void inet_twdr_hangman(unsigned long data)
 220 {
 221         struct inet_timewait_death_row *twdr;
 222         int unsigned need_timer;
 223
 224         twdr = (struct inet_timewait_death_row *)data;
 225         spin_lock(&twdr->death_lock);
 226
 227         if (twdr->tw_count == 0)
 228                 goto out;
 229
 230         need_timer = 0;
 231         if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
 232                 twdr->thread_slots |= (1 << twdr->slot);
 233                 schedule_work(&twdr->twkill_work);
 234                 need_timer = 1;
 235         } else {
 236                 /* We purged the entire slot, anything left?  */
 237                 if (twdr->tw_count)
 238                         need_timer = 1;
 239                 twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
 240         }
 241         if (need_timer)
 242                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 243 out:
 244         spin_unlock(&twdr->death_lock);
 245 }
 246
 247 EXPORT_SYMBOL_GPL(inet_twdr_hangman);
 248
 249 void inet_twdr_twkill_work(struct work_struct *work)
 250 {
 251         struct inet_timewait_death_row *twdr =
 252                 container_of(work, struct inet_timewait_death_row, twkill_work);
 253         int i;
 254
 255         BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
 256                         (sizeof(twdr->thread_slots) * 8));
 257
 258         while (twdr->thread_slots) {
 259                 spin_lock_bh(&twdr->death_lock);
 260                 for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
 261                         if (!(twdr->thread_slots & (1 << i)))
 262                                 continue;
 263
 264                         while (inet_twdr_do_twkill_work(twdr, i) != 0) {
 265                                 if (need_resched()) {
 266                                         spin_unlock_bh(&twdr->death_lock);
 267                                         schedule();
 268                                         spin_lock_bh(&twdr->death_lock);
 269                                 }
 270                         }
 271
 272                         twdr->thread_slots &= ~(1 << i);
 273                 }
 274                 spin_unlock_bh(&twdr->death_lock);
 275         }
 276 }
 277
 278 EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
 279
 280 /* These are always called from BH context.  See callers in
 281  * tcp_input.c to verify this.
 282  */
 283
 284 /* This is for handling early-kills of TIME_WAIT sockets. */
 285 void inet_twsk_deschedule(struct inet_timewait_sock *tw,
 286                           struct inet_timewait_death_row *twdr)
 287 {
 288         spin_lock(&twdr->death_lock);
 289         if (inet_twsk_del_dead_node(tw)) {
 290                 inet_twsk_put(tw);
 291                 if (--twdr->tw_count == 0)
 292                         del_timer(&twdr->tw_timer);
 293         }
 294         spin_unlock(&twdr->death_lock);
 295         __inet_twsk_kill(tw, twdr->hashinfo);
 296 }
 297
 298 EXPORT_SYMBOL(inet_twsk_deschedule);
 299
 300 void inet_twsk_schedule(struct inet_timewait_sock *tw,
 301                        struct inet_timewait_death_row *twdr,
 302                        const int timeo, const int timewait_len)
 303 {
 304         struct hlist_head *list;
 305         int slot;
 306
 307         /* timeout := RTO * 3.5
 308          *
 309          * 3.5 = 1+2+0.5 to wait for two retransmits.
 310          *
 311          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
 312          * our ACK acking that FIN can be lost. If N subsequent retransmitted
 313          * FINs (or previous seqments) are lost (probability of such event
 314          * is p^(N+1), where p is probability to lose single packet and
 315          * time to detect the loss is about RTO*(2^N - 1) with exponential
 316          * backoff). Normal timewait length is calculated so, that we
 317          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
 318          * [ BTW Linux. following BSD, violates this requirement waiting
 319          *   only for 60sec, we should wait at least for 240 secs.
 320          *   Well, 240 consumes too much of resources 8)
 321          * ]
 322          * This interval is not reduced to catch old duplicate and
 323          * responces to our wandering segments living for two MSLs.
 324          * However, if we use PAWS to detect
 325          * old duplicates, we can reduce the interval to bounds required
 326          * by RTO, rather than MSL. So, if peer understands PAWS, we
 327          * kill tw bucket after 3.5*RTO (it is important that this number
 328          * is greater than TS tick!) and detect old duplicates with help
 329          * of PAWS.
 330          */
 331         slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
 332
 333         spin_lock(&twdr->death_lock);
 334
 335         /* Unlink it, if it was scheduled */
 336         if (inet_twsk_del_dead_node(tw))
 337                 twdr->tw_count--;
 338         else
 339                 atomic_inc(&tw->tw_refcnt);
 340
 341         if (slot >= INET_TWDR_RECYCLE_SLOTS) {
 342                 /* Schedule to slow timer */
 343                 if (timeo >= timewait_len) {
 344                         slot = INET_TWDR_TWKILL_SLOTS - 1;
 345                 } else {
 346                         slot = DIV_ROUND_UP(timeo, twdr->period);
 347                         if (slot >= INET_TWDR_TWKILL_SLOTS)
 348                                 slot = INET_TWDR_TWKILL_SLOTS - 1;
 349                 }
 350                 tw->tw_ttd = jiffies + timeo;
 351                 slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
 352                 list = &twdr->cells[slot];
 353         } else {
 354                 tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
 355
 356                 if (twdr->twcal_hand < 0) {
 357                         twdr->twcal_hand = 0;
 358                         twdr->twcal_jiffie = jiffies;
 359                         twdr->twcal_timer.expires = twdr->twcal_jiffie +
 360                                               (slot << INET_TWDR_RECYCLE_TICK);
 361                         add_timer(&twdr->twcal_timer);
 362                 } else {
 363                         if (time_after(twdr->twcal_timer.expires,
 364                                        jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
 365                                 mod_timer(&twdr->twcal_timer,
 366                                           jiffies + (slot << INET_TWDR_RECYCLE_TICK));
 367                         slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
 368                 }
 369                 list = &twdr->twcal_row[slot];
 370         }
 371
 372         hlist_add_head(&tw->tw_death_node, list);
 373
 374         if (twdr->tw_count++ == 0)
 375                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 376         spin_unlock(&twdr->death_lock);
 377 }
 378
 379 EXPORT_SYMBOL_GPL(inet_twsk_schedule);
 380
 381 void inet_twdr_twcal_tick(unsigned long data)
 382 {
 383         struct inet_timewait_death_row *twdr;
 384         int n, slot;
 385         unsigned long j;
 386         unsigned long now = jiffies;
 387         int killed = 0;
 388         int adv = 0;
 389
 390         twdr = (struct inet_timewait_death_row *)data;
 391
 392         spin_lock(&twdr->death_lock);
 393         if (twdr->twcal_hand < 0)
 394                 goto out;
 395
 396         slot = twdr->twcal_hand;
 397         j = twdr->twcal_jiffie;
 398
 399         for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
 400                 if (time_before_eq(j, now)) {
 401                         struct hlist_node *node, *safe;
 402                         struct inet_timewait_sock *tw;
 403
 404                         inet_twsk_for_each_inmate_safe(tw, node, safe,
 405                                                        &twdr->twcal_row[slot]) {
 406                                 __inet_twsk_del_dead_node(tw);
 407                                 __inet_twsk_kill(tw, twdr->hashinfo);
 408 #ifdef CONFIG_NET_NS
 409                                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
 410 #endif
 411                                 inet_twsk_put(tw);
 412                                 killed++;
 413                         }
 414                 } else {
 415                         if (!adv) {
 416                                 adv = 1;
 417                                 twdr->twcal_jiffie = j;
 418                                 twdr->twcal_hand = slot;
 419                         }
 420
 421                         if (!hlist_empty(&twdr->twcal_row[slot])) {
 422                                 mod_timer(&twdr->twcal_timer, j);
 423                                 goto out;
 424                         }
 425                 }
 426                 j += 1 << INET_TWDR_RECYCLE_TICK;
 427                 slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
 428         }
 429         twdr->twcal_hand = -1;
 430
 431 out:
 432         if ((twdr->tw_count -= killed) == 0)
 433                 del_timer(&twdr->tw_timer);
 434 #ifndef CONFIG_NET_NS
 435         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITKILLED, killed);
 436 #endif
 437         spin_unlock(&twdr->death_lock);
 438 }
 439
 440 EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);
 441
 442 void inet_twsk_purge(struct inet_hashinfo *hashinfo,
 443                      struct inet_timewait_death_row *twdr, int family)
 444 {
 445         struct inet_timewait_sock *tw;
 446         struct sock *sk;
 447         struct hlist_nulls_node *node;
 448         unsigned int slot;
 449
 450         for (slot = 0; slot <= hashinfo->ehash_mask; slot++) {
 451                 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 452 restart_rcu:
 453                 rcu_read_lock();
 454 restart:
 455                 sk_nulls_for_each_rcu(sk, node, &head->twchain) {
 456                         tw = inet_twsk(sk);
 457                         if ((tw->tw_family != family) ||
 458                                 atomic_read(&twsk_net(tw)->count))
 459                                 continue;
 460
 461                         if (unlikely(!atomic_inc_not_zero(&tw->tw_refcnt)))
 462                                 continue;
 463
 464                         if (unlikely((tw->tw_family != family) ||
 465                                      atomic_read(&twsk_net(tw)->count))) {
 466                                 inet_twsk_put(tw);
 467                                 goto restart;
 468                         }
 469
 470                         rcu_read_unlock();
 471                         inet_twsk_deschedule(tw, twdr);
 472                         inet_twsk_put(tw);
 473                         goto restart_rcu;
 474                 }
 475                 /* If the nulls value we got at the end of this lookup is
 476                  * not the expected one, we must restart lookup.
 477                  * We probably met an item that was moved to another chain.
 478                  */
 479                 if (get_nulls_value(node) != slot)
 480                         goto restart;
 481                 rcu_read_unlock();
 482         }
 483 }
 484 EXPORT_SYMBOL_GPL(inet_twsk_purge);