linux/net/ipv4/inet_timewait_sock.c
<<
>>
Prefs
   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 */
  28int 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 */
  51int 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. */
  70static 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
 102static 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
 114void inet_twsk_put(struct inet_timewait_sock *tw)
 115{
 116        if (atomic_dec_and_test(&tw->tw_refcnt))
 117                inet_twsk_free(tw);
 118}
 119EXPORT_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 */
 126void __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}
 171EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
 172
 173struct 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}
 212EXPORT_SYMBOL_GPL(inet_twsk_alloc);
 213
 214/* Returns non-zero if quota exceeded.  */
 215static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
 216                                    const int slot)
 217{
 218        struct inet_timewait_sock *tw;
 219        struct hlist_node *node;
 220        unsigned int killed;
 221        int ret;
 222
 223        /* NOTE: compare this to previous version where lock
 224         * was released after detaching chain. It was racy,
 225         * because tw buckets are scheduled in not serialized context
 226         * in 2.3 (with netfilter), and with softnet it is common, because
 227         * soft irqs are not sequenced.
 228         */
 229        killed = 0;
 230        ret = 0;
 231rescan:
 232        inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
 233                __inet_twsk_del_dead_node(tw);
 234                spin_unlock(&twdr->death_lock);
 235                __inet_twsk_kill(tw, twdr->hashinfo);
 236#ifdef CONFIG_NET_NS
 237                NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
 238#endif
 239                inet_twsk_put(tw);
 240                killed++;
 241                spin_lock(&twdr->death_lock);
 242                if (killed > INET_TWDR_TWKILL_QUOTA) {
 243                        ret = 1;
 244                        break;
 245                }
 246
 247                /* While we dropped twdr->death_lock, another cpu may have
 248                 * killed off the next TW bucket in the list, therefore
 249                 * do a fresh re-read of the hlist head node with the
 250                 * lock reacquired.  We still use the hlist traversal
 251                 * macro in order to get the prefetches.
 252                 */
 253                goto rescan;
 254        }
 255
 256        twdr->tw_count -= killed;
 257#ifndef CONFIG_NET_NS
 258        NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITED, killed);
 259#endif
 260        return ret;
 261}
 262
 263void inet_twdr_hangman(unsigned long data)
 264{
 265        struct inet_timewait_death_row *twdr;
 266        unsigned int need_timer;
 267
 268        twdr = (struct inet_timewait_death_row *)data;
 269        spin_lock(&twdr->death_lock);
 270
 271        if (twdr->tw_count == 0)
 272                goto out;
 273
 274        need_timer = 0;
 275        if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
 276                twdr->thread_slots |= (1 << twdr->slot);
 277                schedule_work(&twdr->twkill_work);
 278                need_timer = 1;
 279        } else {
 280                /* We purged the entire slot, anything left?  */
 281                if (twdr->tw_count)
 282                        need_timer = 1;
 283                twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
 284        }
 285        if (need_timer)
 286                mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 287out:
 288        spin_unlock(&twdr->death_lock);
 289}
 290EXPORT_SYMBOL_GPL(inet_twdr_hangman);
 291
 292void inet_twdr_twkill_work(struct work_struct *work)
 293{
 294        struct inet_timewait_death_row *twdr =
 295                container_of(work, struct inet_timewait_death_row, twkill_work);
 296        int i;
 297
 298        BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
 299                        (sizeof(twdr->thread_slots) * 8));
 300
 301        while (twdr->thread_slots) {
 302                spin_lock_bh(&twdr->death_lock);
 303                for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
 304                        if (!(twdr->thread_slots & (1 << i)))
 305                                continue;
 306
 307                        while (inet_twdr_do_twkill_work(twdr, i) != 0) {
 308                                if (need_resched()) {
 309                                        spin_unlock_bh(&twdr->death_lock);
 310                                        schedule();
 311                                        spin_lock_bh(&twdr->death_lock);
 312                                }
 313                        }
 314
 315                        twdr->thread_slots &= ~(1 << i);
 316                }
 317                spin_unlock_bh(&twdr->death_lock);
 318        }
 319}
 320EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
 321
 322/* These are always called from BH context.  See callers in
 323 * tcp_input.c to verify this.
 324 */
 325
 326/* This is for handling early-kills of TIME_WAIT sockets. */
 327void inet_twsk_deschedule(struct inet_timewait_sock *tw,
 328                          struct inet_timewait_death_row *twdr)
 329{
 330        spin_lock(&twdr->death_lock);
 331        if (inet_twsk_del_dead_node(tw)) {
 332                inet_twsk_put(tw);
 333                if (--twdr->tw_count == 0)
 334                        del_timer(&twdr->tw_timer);
 335        }
 336        spin_unlock(&twdr->death_lock);
 337        __inet_twsk_kill(tw, twdr->hashinfo);
 338}
 339EXPORT_SYMBOL(inet_twsk_deschedule);
 340
 341void inet_twsk_schedule(struct inet_timewait_sock *tw,
 342                       struct inet_timewait_death_row *twdr,
 343                       const int timeo, const int timewait_len)
 344{
 345        struct hlist_head *list;
 346        int slot;
 347
 348        /* timeout := RTO * 3.5
 349         *
 350         * 3.5 = 1+2+0.5 to wait for two retransmits.
 351         *
 352         * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
 353         * our ACK acking that FIN can be lost. If N subsequent retransmitted
 354         * FINs (or previous seqments) are lost (probability of such event
 355         * is p^(N+1), where p is probability to lose single packet and
 356         * time to detect the loss is about RTO*(2^N - 1) with exponential
 357         * backoff). Normal timewait length is calculated so, that we
 358         * waited at least for one retransmitted FIN (maximal RTO is 120sec).
 359         * [ BTW Linux. following BSD, violates this requirement waiting
 360         *   only for 60sec, we should wait at least for 240 secs.
 361         *   Well, 240 consumes too much of resources 8)
 362         * ]
 363         * This interval is not reduced to catch old duplicate and
 364         * responces to our wandering segments living for two MSLs.
 365         * However, if we use PAWS to detect
 366         * old duplicates, we can reduce the interval to bounds required
 367         * by RTO, rather than MSL. So, if peer understands PAWS, we
 368         * kill tw bucket after 3.5*RTO (it is important that this number
 369         * is greater than TS tick!) and detect old duplicates with help
 370         * of PAWS.
 371         */
 372        slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
 373
 374        spin_lock(&twdr->death_lock);
 375
 376        /* Unlink it, if it was scheduled */
 377        if (inet_twsk_del_dead_node(tw))
 378                twdr->tw_count--;
 379        else
 380                atomic_inc(&tw->tw_refcnt);
 381
 382        if (slot >= INET_TWDR_RECYCLE_SLOTS) {
 383                /* Schedule to slow timer */
 384                if (timeo >= timewait_len) {
 385                        slot = INET_TWDR_TWKILL_SLOTS - 1;
 386                } else {
 387                        slot = DIV_ROUND_UP(timeo, twdr->period);
 388                        if (slot >= INET_TWDR_TWKILL_SLOTS)
 389                                slot = INET_TWDR_TWKILL_SLOTS - 1;
 390                }
 391                tw->tw_ttd = jiffies + timeo;
 392                slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
 393                list = &twdr->cells[slot];
 394        } else {
 395                tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
 396
 397                if (twdr->twcal_hand < 0) {
 398                        twdr->twcal_hand = 0;
 399                        twdr->twcal_jiffie = jiffies;
 400                        twdr->twcal_timer.expires = twdr->twcal_jiffie +
 401                                              (slot << INET_TWDR_RECYCLE_TICK);
 402                        add_timer(&twdr->twcal_timer);
 403                } else {
 404                        if (time_after(twdr->twcal_timer.expires,
 405                                       jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
 406                                mod_timer(&twdr->twcal_timer,
 407                                          jiffies + (slot << INET_TWDR_RECYCLE_TICK));
 408                        slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
 409                }
 410                list = &twdr->twcal_row[slot];
 411        }
 412
 413        hlist_add_head(&tw->tw_death_node, list);
 414
 415        if (twdr->tw_count++ == 0)
 416                mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 417        spin_unlock(&twdr->death_lock);
 418}
 419EXPORT_SYMBOL_GPL(inet_twsk_schedule);
 420
 421void inet_twdr_twcal_tick(unsigned long data)
 422{
 423        struct inet_timewait_death_row *twdr;
 424        int n, slot;
 425        unsigned long j;
 426        unsigned long now = jiffies;
 427        int killed = 0;
 428        int adv = 0;
 429
 430        twdr = (struct inet_timewait_death_row *)data;
 431
 432        spin_lock(&twdr->death_lock);
 433        if (twdr->twcal_hand < 0)
 434                goto out;
 435
 436        slot = twdr->twcal_hand;
 437        j = twdr->twcal_jiffie;
 438
 439        for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
 440                if (time_before_eq(j, now)) {
 441                        struct hlist_node *node, *safe;
 442                        struct inet_timewait_sock *tw;
 443
 444                        inet_twsk_for_each_inmate_safe(tw, node, safe,
 445                                                       &twdr->twcal_row[slot]) {
 446                                __inet_twsk_del_dead_node(tw);
 447                                __inet_twsk_kill(tw, twdr->hashinfo);
 448#ifdef CONFIG_NET_NS
 449                                NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
 450#endif
 451                                inet_twsk_put(tw);
 452                                killed++;
 453                        }
 454                } else {
 455                        if (!adv) {
 456                                adv = 1;
 457                                twdr->twcal_jiffie = j;
 458                                twdr->twcal_hand = slot;
 459                        }
 460
 461                        if (!hlist_empty(&twdr->twcal_row[slot])) {
 462                                mod_timer(&twdr->twcal_timer, j);
 463                                goto out;
 464                        }
 465                }
 466                j += 1 << INET_TWDR_RECYCLE_TICK;
 467                slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
 468        }
 469        twdr->twcal_hand = -1;
 470
 471out:
 472        if ((twdr->tw_count -= killed) == 0)
 473                del_timer(&twdr->tw_timer);
 474#ifndef CONFIG_NET_NS
 475        NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITKILLED, killed);
 476#endif
 477        spin_unlock(&twdr->death_lock);
 478}
 479EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);
 480
 481void inet_twsk_purge(struct inet_hashinfo *hashinfo,
 482                     struct inet_timewait_death_row *twdr, int family)
 483{
 484        struct inet_timewait_sock *tw;
 485        struct sock *sk;
 486        struct hlist_nulls_node *node;
 487        unsigned int slot;
 488
 489        for (slot = 0; slot <= hashinfo->ehash_mask; slot++) {
 490                struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 491restart_rcu:
 492                rcu_read_lock();
 493restart:
 494                sk_nulls_for_each_rcu(sk, node, &head->twchain) {
 495                        tw = inet_twsk(sk);
 496                        if ((tw->tw_family != family) ||
 497                                atomic_read(&twsk_net(tw)->count))
 498                                continue;
 499
 500                        if (unlikely(!atomic_inc_not_zero(&tw->tw_refcnt)))
 501                                continue;
 502
 503                        if (unlikely((tw->tw_family != family) ||
 504                                     atomic_read(&twsk_net(tw)->count))) {
 505                                inet_twsk_put(tw);
 506                                goto restart;
 507                        }
 508
 509                        rcu_read_unlock();
 510                        local_bh_disable();
 511                        inet_twsk_deschedule(tw, twdr);
 512                        local_bh_enable();
 513                        inet_twsk_put(tw);
 514                        goto restart_rcu;
 515                }
 516                /* If the nulls value we got at the end of this lookup is
 517                 * not the expected one, we must restart lookup.
 518                 * We probably met an item that was moved to another chain.
 519                 */
 520                if (get_nulls_value(node) != slot)
 521                        goto restart;
 522                rcu_read_unlock();
 523        }
 524}
 525EXPORT_SYMBOL_GPL(inet_twsk_purge);
 526
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.