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 <linux/slab.h>
  14 #include <linux/module.h>
  15 #include <net/inet_hashtables.h>
  16 #include <net/inet_timewait_sock.h>
  17 #include <net/ip.h>
  18
  19
  20 /**
  21  *      inet_twsk_unhash - unhash a timewait socket from established hash
  22  *      @tw: timewait socket
  23  *
  24  *      unhash a timewait socket from established hash, if hashed.
  25  *      ehash lock must be held by caller.
  26  *      Returns 1 if caller should call inet_twsk_put() after lock release.
  27  */
  28 int inet_twsk_unhash(struct inet_timewait_sock *tw)
  29 {
  30         if (hlist_nulls_unhashed(&tw->tw_node))
  31                 return 0;
  32
  33         hlist_nulls_del_rcu(&tw->tw_node);
  34         sk_nulls_node_init(&tw->tw_node);
  35         /*
  36          * We cannot call inet_twsk_put() ourself under lock,
  37          * caller must call it for us.
  38          */
  39         return 1;
  40 }
  41
  42 /**
  43  *      inet_twsk_bind_unhash - unhash a timewait socket from bind hash
  44  *      @tw: timewait socket
  45  *      @hashinfo: hashinfo pointer
  46  *
  47  *      unhash a timewait socket from bind hash, if hashed.
  48  *      bind hash lock must be held by caller.
  49  *      Returns 1 if caller should call inet_twsk_put() after lock release.
  50  */
  51 int inet_twsk_bind_unhash(struct inet_timewait_sock *tw,
  52                           struct inet_hashinfo *hashinfo)
  53 {
  54         struct inet_bind_bucket *tb = tw->tw_tb;
  55
  56         if (!tb)
  57                 return 0;
  58
  59         __hlist_del(&tw->tw_bind_node);
  60         tw->tw_tb = NULL;
  61         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  62         /*
  63          * We cannot call inet_twsk_put() ourself under lock,
  64          * caller must call it for us.
  65          */
  66         return 1;
  67 }
  68
  69 /* Must be called with locally disabled BHs. */
  70 static void __inet_twsk_kill(struct inet_timewait_sock *tw,
  71                              struct inet_hashinfo *hashinfo)
  72 {
  73         struct inet_bind_hashbucket *bhead;
  74         int refcnt;
  75         /* Unlink from established hashes. */
  76         spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
  77
  78         spin_lock(lock);
  79         refcnt = inet_twsk_unhash(tw);
  80         spin_unlock(lock);
  81
  82         /* Disassociate with bind bucket. */
  83         bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), tw->tw_num,
  84                         hashinfo->bhash_size)];
  85
  86         spin_lock(&bhead->lock);
  87         refcnt += inet_twsk_bind_unhash(tw, hashinfo);
  88         spin_unlock(&bhead->lock);
  89
  90 #ifdef SOCK_REFCNT_DEBUG
  91         if (atomic_read(&tw->tw_refcnt) != 1) {
  92                 pr_debug("%s timewait_sock %p refcnt=%d\n",
  93                          tw->tw_prot->name, tw, atomic_read(&tw->tw_refcnt));
  94         }
  95 #endif
  96         while (refcnt) {
  97                 inet_twsk_put(tw);
  98                 refcnt--;
  99         }
 100 }
 101
 102 static noinline void inet_twsk_free(struct inet_timewait_sock *tw)
 103 {
 104         struct module *owner = tw->tw_prot->owner;
 105         twsk_destructor((struct sock *)tw);
 106 #ifdef SOCK_REFCNT_DEBUG
 107         pr_debug("%s timewait_sock %p released\n", tw->tw_prot->name, tw);
 108 #endif
 109         release_net(twsk_net(tw));
 110         kmem_cache_free(tw->tw_prot->twsk_prot->twsk_slab, tw);
 111         module_put(owner);
 112 }
 113
 114 void inet_twsk_put(struct inet_timewait_sock *tw)
 115 {
 116         if (atomic_dec_and_test(&tw->tw_refcnt))
 117                 inet_twsk_free(tw);
 118 }
 119 EXPORT_SYMBOL_GPL(inet_twsk_put);
 120
 121 /*
 122  * Enter the time wait state. This is called with locally disabled BH.
 123  * Essentially we whip up a timewait bucket, copy the relevant info into it
 124  * from the SK, and mess with hash chains and list linkage.
 125  */
 126 void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
 127                            struct inet_hashinfo *hashinfo)
 128 {
 129         const struct inet_sock *inet = inet_sk(sk);
 130         const struct inet_connection_sock *icsk = inet_csk(sk);
 131         struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
 132         spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 133         struct inet_bind_hashbucket *bhead;
 134         /* Step 1: Put TW into bind hash. Original socket stays there too.
 135            Note, that any socket with inet->num != 0 MUST be bound in
 136            binding cache, even if it is closed.
 137          */
 138         bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), inet->inet_num,
 139                         hashinfo->bhash_size)];
 140         spin_lock(&bhead->lock);
 141         tw->tw_tb = icsk->icsk_bind_hash;
 142         WARN_ON(!icsk->icsk_bind_hash);
 143         inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
 144         spin_unlock(&bhead->lock);
 145
 146         spin_lock(lock);
 147
 148         /*
 149          * Step 2: Hash TW into TIMEWAIT chain.
 150          * Should be done before removing sk from established chain
 151          * because readers are lockless and search established first.
 152          */
 153         inet_twsk_add_node_rcu(tw, &ehead->twchain);
 154
 155         /* Step 3: Remove SK from established hash. */
 156         if (__sk_nulls_del_node_init_rcu(sk))
 157                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 158
 159         /*
 160          * Notes :
 161          * - We initially set tw_refcnt to 0 in inet_twsk_alloc()
 162          * - We add one reference for the bhash link
 163          * - We add one reference for the ehash link
 164          * - We want this refcnt update done before allowing other
 165          *   threads to find this tw in ehash chain.
 166          */
 167         atomic_add(1 + 1 + 1, &tw->tw_refcnt);
 168
 169         spin_unlock(lock);
 170 }
 171 EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
 172
 173 struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
 174 {
 175         struct inet_timewait_sock *tw =
 176                 kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
 177                                  GFP_ATOMIC);
 178         if (tw != NULL) {
 179                 const struct inet_sock *inet = inet_sk(sk);
 180
 181                 kmemcheck_annotate_bitfield(tw, flags);
 182
 183                 /* Give us an identity. */
 184                 tw->tw_daddr        = inet->inet_daddr;
 185                 tw->tw_rcv_saddr    = inet->inet_rcv_saddr;
 186                 tw->tw_bound_dev_if = sk->sk_bound_dev_if;
 187                 tw->tw_tos          = inet->tos;
 188                 tw->tw_num          = inet->inet_num;
 189                 tw->tw_state        = TCP_TIME_WAIT;
 190                 tw->tw_substate     = state;
 191                 tw->tw_sport        = inet->inet_sport;
 192                 tw->tw_dport        = inet->inet_dport;
 193                 tw->tw_family       = sk->sk_family;
 194                 tw->tw_reuse        = sk->sk_reuse;
 195                 tw->tw_hash         = sk->sk_hash;
 196                 tw->tw_ipv6only     = 0;
 197                 tw->tw_transparent  = inet->transparent;
 198                 tw->tw_prot         = sk->sk_prot_creator;
 199                 twsk_net_set(tw, hold_net(sock_net(sk)));
 200                 /*
 201                  * Because we use RCU lookups, we should not set tw_refcnt
 202                  * to a non null value before everything is setup for this
 203                  * timewait socket.
 204                  */
 205                 atomic_set(&tw->tw_refcnt, 0);
 206                 inet_twsk_dead_node_init(tw);
 207                 __module_get(tw->tw_prot->owner);
 208         }
 209
 210         return tw;
 211 }
 212 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
 213
 214 /* Returns non-zero if quota exceeded.  */
 215 static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
 216                                     const int slot)
 217 {
 218         struct inet_timewait_sock *tw;
 219         unsigned int killed;
 220         int ret;
 221
 222         /* NOTE: compare this to previous version where lock
 223          * was released after detaching chain. It was racy,
 224          * because tw buckets are scheduled in not serialized context
 225          * in 2.3 (with netfilter), and with softnet it is common, because
 226          * soft irqs are not sequenced.
 227          */
 228         killed = 0;
 229         ret = 0;
 230 rescan:
 231         inet_twsk_for_each_inmate(tw, &twdr->cells[slot]) {
 232                 __inet_twsk_del_dead_node(tw);
 233                 spin_unlock(&twdr->death_lock);
 234                 __inet_twsk_kill(tw, twdr->hashinfo);
 235 #ifdef CONFIG_NET_NS
 236                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
 237 #endif
 238                 inet_twsk_put(tw);
 239                 killed++;
 240                 spin_lock(&twdr->death_lock);
 241                 if (killed > INET_TWDR_TWKILL_QUOTA) {
 242                         ret = 1;
 243                         break;
 244                 }
 245
 246                 /* While we dropped twdr->death_lock, another cpu may have
 247                  * killed off the next TW bucket in the list, therefore
 248                  * do a fresh re-read of the hlist head node with the
 249                  * lock reacquired.  We still use the hlist traversal
 250                  * macro in order to get the prefetches.
 251                  */
 252                 goto rescan;
 253         }
 254
 255         twdr->tw_count -= killed;
 256 #ifndef CONFIG_NET_NS
 257         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITED, killed);
 258 #endif
 259         return ret;
 260 }
 261
 262 void inet_twdr_hangman(unsigned long data)
 263 {
 264         struct inet_timewait_death_row *twdr;
 265         unsigned int need_timer;
 266
 267         twdr = (struct inet_timewait_death_row *)data;
 268         spin_lock(&twdr->death_lock);
 269
 270         if (twdr->tw_count == 0)
 271                 goto out;
 272
 273         need_timer = 0;
 274         if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
 275                 twdr->thread_slots |= (1 << twdr->slot);
 276                 schedule_work(&twdr->twkill_work);
 277                 need_timer = 1;
 278         } else {
 279                 /* We purged the entire slot, anything left?  */
 280                 if (twdr->tw_count)
 281                         need_timer = 1;
 282                 twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
 283         }
 284         if (need_timer)
 285                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 286 out:
 287         spin_unlock(&twdr->death_lock);
 288 }
 289 EXPORT_SYMBOL_GPL(inet_twdr_hangman);
 290
 291 void inet_twdr_twkill_work(struct work_struct *work)
 292 {
 293         struct inet_timewait_death_row *twdr =
 294                 container_of(work, struct inet_timewait_death_row, twkill_work);
 295         int i;
 296
 297         BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
 298                         (sizeof(twdr->thread_slots) * 8));
 299
 300         while (twdr->thread_slots) {
 301                 spin_lock_bh(&twdr->death_lock);
 302                 for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
 303                         if (!(twdr->thread_slots & (1 << i)))
 304                                 continue;
 305
 306                         while (inet_twdr_do_twkill_work(twdr, i) != 0) {
 307                                 if (need_resched()) {
 308                                         spin_unlock_bh(&twdr->death_lock);
 309                                         schedule();
 310                                         spin_lock_bh(&twdr->death_lock);
 311                                 }
 312                         }
 313
 314                         twdr->thread_slots &= ~(1 << i);
 315                 }
 316                 spin_unlock_bh(&twdr->death_lock);
 317         }
 318 }
 319 EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
 320
 321 /* These are always called from BH context.  See callers in
 322  * tcp_input.c to verify this.
 323  */
 324
 325 /* This is for handling early-kills of TIME_WAIT sockets. */
 326 void inet_twsk_deschedule(struct inet_timewait_sock *tw,
 327                           struct inet_timewait_death_row *twdr)
 328 {
 329         spin_lock(&twdr->death_lock);
 330         if (inet_twsk_del_dead_node(tw)) {
 331                 inet_twsk_put(tw);
 332                 if (--twdr->tw_count == 0)
 333                         del_timer(&twdr->tw_timer);
 334         }
 335         spin_unlock(&twdr->death_lock);
 336         __inet_twsk_kill(tw, twdr->hashinfo);
 337 }
 338 EXPORT_SYMBOL(inet_twsk_deschedule);
 339
 340 void inet_twsk_schedule(struct inet_timewait_sock *tw,
 341                        struct inet_timewait_death_row *twdr,
 342                        const int timeo, const int timewait_len)
 343 {
 344         struct hlist_head *list;
 345         int slot;
 346
 347         /* timeout := RTO * 3.5
 348          *
 349          * 3.5 = 1+2+0.5 to wait for two retransmits.
 350          *
 351          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
 352          * our ACK acking that FIN can be lost. If N subsequent retransmitted
 353          * FINs (or previous seqments) are lost (probability of such event
 354          * is p^(N+1), where p is probability to lose single packet and
 355          * time to detect the loss is about RTO*(2^N - 1) with exponential
 356          * backoff). Normal timewait length is calculated so, that we
 357          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
 358          * [ BTW Linux. following BSD, violates this requirement waiting
 359          *   only for 60sec, we should wait at least for 240 secs.
 360          *   Well, 240 consumes too much of resources 8)
 361          * ]
 362          * This interval is not reduced to catch old duplicate and
 363          * responces to our wandering segments living for two MSLs.
 364          * However, if we use PAWS to detect
 365          * old duplicates, we can reduce the interval to bounds required
 366          * by RTO, rather than MSL. So, if peer understands PAWS, we
 367          * kill tw bucket after 3.5*RTO (it is important that this number
 368          * is greater than TS tick!) and detect old duplicates with help
 369          * of PAWS.
 370          */
 371         slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
 372
 373         spin_lock(&twdr->death_lock);
 374
 375         /* Unlink it, if it was scheduled */
 376         if (inet_twsk_del_dead_node(tw))
 377                 twdr->tw_count--;
 378         else
 379                 atomic_inc(&tw->tw_refcnt);
 380
 381         if (slot >= INET_TWDR_RECYCLE_SLOTS) {
 382                 /* Schedule to slow timer */
 383                 if (timeo >= timewait_len) {
 384                         slot = INET_TWDR_TWKILL_SLOTS - 1;
 385                 } else {
 386                         slot = DIV_ROUND_UP(timeo, twdr->period);
 387                         if (slot >= INET_TWDR_TWKILL_SLOTS)
 388                                 slot = INET_TWDR_TWKILL_SLOTS - 1;
 389                 }
 390                 tw->tw_ttd = jiffies + timeo;
 391                 slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
 392                 list = &twdr->cells[slot];
 393         } else {
 394                 tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
 395
 396                 if (twdr->twcal_hand < 0) {
 397                         twdr->twcal_hand = 0;
 398                         twdr->twcal_jiffie = jiffies;
 399                         twdr->twcal_timer.expires = twdr->twcal_jiffie +
 400                                               (slot << INET_TWDR_RECYCLE_TICK);
 401                         add_timer(&twdr->twcal_timer);
 402                 } else {
 403                         if (time_after(twdr->twcal_timer.expires,
 404                                        jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
 405                                 mod_timer(&twdr->twcal_timer,
 406                                           jiffies + (slot << INET_TWDR_RECYCLE_TICK));
 407                         slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
 408                 }
 409                 list = &twdr->twcal_row[slot];
 410         }
 411
 412         hlist_add_head(&tw->tw_death_node, list);
 413
 414         if (twdr->tw_count++ == 0)
 415                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 416         spin_unlock(&twdr->death_lock);
 417 }
 418 EXPORT_SYMBOL_GPL(inet_twsk_schedule);
 419
 420 void inet_twdr_twcal_tick(unsigned long data)
 421 {
 422         struct inet_timewait_death_row *twdr;
 423         int n, slot;
 424         unsigned long j;
 425         unsigned long now = jiffies;
 426         int killed = 0;
 427         int adv = 0;
 428
 429         twdr = (struct inet_timewait_death_row *)data;
 430
 431         spin_lock(&twdr->death_lock);
 432         if (twdr->twcal_hand < 0)
 433                 goto out;
 434
 435         slot = twdr->twcal_hand;
 436         j = twdr->twcal_jiffie;
 437
 438         for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
 439                 if (time_before_eq(j, now)) {
 440                         struct hlist_node *safe;
 441                         struct inet_timewait_sock *tw;
 442
 443                         inet_twsk_for_each_inmate_safe(tw, safe,
 444                                                        &twdr->twcal_row[slot]) {
 445                                 __inet_twsk_del_dead_node(tw);
 446                                 __inet_twsk_kill(tw, twdr->hashinfo);
 447 #ifdef CONFIG_NET_NS
 448                                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
 449 #endif
 450                                 inet_twsk_put(tw);
 451                                 killed++;
 452                         }
 453                 } else {
 454                         if (!adv) {
 455                                 adv = 1;
 456                                 twdr->twcal_jiffie = j;
 457                                 twdr->twcal_hand = slot;
 458                         }
 459
 460                         if (!hlist_empty(&twdr->twcal_row[slot])) {
 461                                 mod_timer(&twdr->twcal_timer, j);
 462                                 goto out;
 463                         }
 464                 }
 465                 j += 1 << INET_TWDR_RECYCLE_TICK;
 466                 slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
 467         }
 468         twdr->twcal_hand = -1;
 469
 470 out:
 471         if ((twdr->tw_count -= killed) == 0)
 472                 del_timer(&twdr->tw_timer);
 473 #ifndef CONFIG_NET_NS
 474         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITKILLED, killed);
 475 #endif
 476         spin_unlock(&twdr->death_lock);
 477 }
 478 EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);
 479
 480 void inet_twsk_purge(struct inet_hashinfo *hashinfo,
 481                      struct inet_timewait_death_row *twdr, int family)
 482 {
 483         struct inet_timewait_sock *tw;
 484         struct sock *sk;
 485         struct hlist_nulls_node *node;
 486         unsigned int slot;
 487
 488         for (slot = 0; slot <= hashinfo->ehash_mask; slot++) {
 489                 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 490 restart_rcu:
 491                 rcu_read_lock();
 492 restart:
 493                 sk_nulls_for_each_rcu(sk, node, &head->twchain) {
 494                         tw = inet_twsk(sk);
 495                         if ((tw->tw_family != family) ||
 496                                 atomic_read(&twsk_net(tw)->count))
 497                                 continue;
 498
 499                         if (unlikely(!atomic_inc_not_zero(&tw->tw_refcnt)))
 500                                 continue;
 501
 502                         if (unlikely((tw->tw_family != family) ||
 503                                      atomic_read(&twsk_net(tw)->count))) {
 504                                 inet_twsk_put(tw);
 505                                 goto restart;
 506                         }
 507
 508                         rcu_read_unlock();
 509                         local_bh_disable();
 510                         inet_twsk_deschedule(tw, twdr);
 511                         local_bh_enable();
 512                         inet_twsk_put(tw);
 513                         goto restart_rcu;
 514                 }
 515                 /* If the nulls value we got at the end of this lookup is
 516                  * not the expected one, we must restart lookup.
 517                  * We probably met an item that was moved to another chain.
 518                  */
 519                 if (get_nulls_value(node) != slot)
 520                         goto restart;
 521                 rcu_read_unlock();
 522         }
 523 }
 524 EXPORT_SYMBOL_GPL(inet_twsk_purge);