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 <net/inet_hashtables.h>
  13 #include <net/inet_timewait_sock.h>
  14 #include <net/ip.h>
  15
  16 /* Must be called with locally disabled BHs. */
  17 static void __inet_twsk_kill(struct inet_timewait_sock *tw,
  18                              struct inet_hashinfo *hashinfo)
  19 {
  20         struct inet_bind_hashbucket *bhead;
  21         struct inet_bind_bucket *tb;
  22         /* Unlink from established hashes. */
  23         rwlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
  24
  25         write_lock(lock);
  26         if (hlist_unhashed(&tw->tw_node)) {
  27                 write_unlock(lock);
  28                 return;
  29         }
  30         __hlist_del(&tw->tw_node);
  31         sk_node_init(&tw->tw_node);
  32         write_unlock(lock);
  33
  34         /* Disassociate with bind bucket. */
  35         bhead = &hashinfo->bhash[inet_bhashfn(tw->tw_num, hashinfo->bhash_size)];
  36         spin_lock(&bhead->lock);
  37         tb = tw->tw_tb;
  38         __hlist_del(&tw->tw_bind_node);
  39         tw->tw_tb = NULL;
  40         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
  41         spin_unlock(&bhead->lock);
  42 #ifdef SOCK_REFCNT_DEBUG
  43         if (atomic_read(&tw->tw_refcnt) != 1) {
  44                 printk(KERN_DEBUG "%s timewait_sock %p refcnt=%d\n",
  45                        tw->tw_prot->name, tw, atomic_read(&tw->tw_refcnt));
  46         }
  47 #endif
  48         inet_twsk_put(tw);
  49 }
  50
  51 /*
  52  * Enter the time wait state. This is called with locally disabled BH.
  53  * Essentially we whip up a timewait bucket, copy the relevant info into it
  54  * from the SK, and mess with hash chains and list linkage.
  55  */
  56 void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
  57                            struct inet_hashinfo *hashinfo)
  58 {
  59         const struct inet_sock *inet = inet_sk(sk);
  60         const struct inet_connection_sock *icsk = inet_csk(sk);
  61         struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
  62         rwlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
  63         struct inet_bind_hashbucket *bhead;
  64         /* Step 1: Put TW into bind hash. Original socket stays there too.
  65            Note, that any socket with inet->num != 0 MUST be bound in
  66            binding cache, even if it is closed.
  67          */
  68         bhead = &hashinfo->bhash[inet_bhashfn(inet->num, hashinfo->bhash_size)];
  69         spin_lock(&bhead->lock);
  70         tw->tw_tb = icsk->icsk_bind_hash;
  71         BUG_TRAP(icsk->icsk_bind_hash);
  72         inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
  73         spin_unlock(&bhead->lock);
  74
  75         write_lock(lock);
  76
  77         /* Step 2: Remove SK from established hash. */
  78         if (__sk_del_node_init(sk))
  79                 sock_prot_dec_use(sk->sk_prot);
  80
  81         /* Step 3: Hash TW into TIMEWAIT chain. */
  82         inet_twsk_add_node(tw, &ehead->twchain);
  83         atomic_inc(&tw->tw_refcnt);
  84
  85         write_unlock(lock);
  86 }
  87
  88 EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
  89
  90 struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
  91 {
  92         struct inet_timewait_sock *tw =
  93                 kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
  94                                  GFP_ATOMIC);
  95         if (tw != NULL) {
  96                 const struct inet_sock *inet = inet_sk(sk);
  97
  98                 /* Give us an identity. */
  99                 tw->tw_daddr        = inet->daddr;
 100                 tw->tw_rcv_saddr    = inet->rcv_saddr;
 101                 tw->tw_bound_dev_if = sk->sk_bound_dev_if;
 102                 tw->tw_num          = inet->num;
 103                 tw->tw_state        = TCP_TIME_WAIT;
 104                 tw->tw_substate     = state;
 105                 tw->tw_sport        = inet->sport;
 106                 tw->tw_dport        = inet->dport;
 107                 tw->tw_family       = sk->sk_family;
 108                 tw->tw_reuse        = sk->sk_reuse;
 109                 tw->tw_hash         = sk->sk_hash;
 110                 tw->tw_ipv6only     = 0;
 111                 tw->tw_prot         = sk->sk_prot_creator;
 112                 atomic_set(&tw->tw_refcnt, 1);
 113                 inet_twsk_dead_node_init(tw);
 114                 __module_get(tw->tw_prot->owner);
 115         }
 116
 117         return tw;
 118 }
 119
 120 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
 121
 122 /* Returns non-zero if quota exceeded.  */
 123 static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
 124                                     const int slot)
 125 {
 126         struct inet_timewait_sock *tw;
 127         struct hlist_node *node;
 128         unsigned int killed;
 129         int ret;
 130
 131         /* NOTE: compare this to previous version where lock
 132          * was released after detaching chain. It was racy,
 133          * because tw buckets are scheduled in not serialized context
 134          * in 2.3 (with netfilter), and with softnet it is common, because
 135          * soft irqs are not sequenced.
 136          */
 137         killed = 0;
 138         ret = 0;
 139 rescan:
 140         inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
 141                 __inet_twsk_del_dead_node(tw);
 142                 spin_unlock(&twdr->death_lock);
 143                 __inet_twsk_kill(tw, twdr->hashinfo);
 144                 inet_twsk_put(tw);
 145                 killed++;
 146                 spin_lock(&twdr->death_lock);
 147                 if (killed > INET_TWDR_TWKILL_QUOTA) {
 148                         ret = 1;
 149                         break;
 150                 }
 151
 152                 /* While we dropped twdr->death_lock, another cpu may have
 153                  * killed off the next TW bucket in the list, therefore
 154                  * do a fresh re-read of the hlist head node with the
 155                  * lock reacquired.  We still use the hlist traversal
 156                  * macro in order to get the prefetches.
 157                  */
 158                 goto rescan;
 159         }
 160
 161         twdr->tw_count -= killed;
 162         NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITED, killed);
 163
 164         return ret;
 165 }
 166
 167 void inet_twdr_hangman(unsigned long data)
 168 {
 169         struct inet_timewait_death_row *twdr;
 170         int unsigned need_timer;
 171
 172         twdr = (struct inet_timewait_death_row *)data;
 173         spin_lock(&twdr->death_lock);
 174
 175         if (twdr->tw_count == 0)
 176                 goto out;
 177
 178         need_timer = 0;
 179         if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
 180                 twdr->thread_slots |= (1 << twdr->slot);
 181                 schedule_work(&twdr->twkill_work);
 182                 need_timer = 1;
 183         } else {
 184                 /* We purged the entire slot, anything left?  */
 185                 if (twdr->tw_count)
 186                         need_timer = 1;
 187         }
 188         twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
 189         if (need_timer)
 190                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 191 out:
 192         spin_unlock(&twdr->death_lock);
 193 }
 194
 195 EXPORT_SYMBOL_GPL(inet_twdr_hangman);
 196
 197 void inet_twdr_twkill_work(struct work_struct *work)
 198 {
 199         struct inet_timewait_death_row *twdr =
 200                 container_of(work, struct inet_timewait_death_row, twkill_work);
 201         int i;
 202
 203         BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
 204                         (sizeof(twdr->thread_slots) * 8));
 205
 206         while (twdr->thread_slots) {
 207                 spin_lock_bh(&twdr->death_lock);
 208                 for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
 209                         if (!(twdr->thread_slots & (1 << i)))
 210                                 continue;
 211
 212                         while (inet_twdr_do_twkill_work(twdr, i) != 0) {
 213                                 if (need_resched()) {
 214                                         spin_unlock_bh(&twdr->death_lock);
 215                                         schedule();
 216                                         spin_lock_bh(&twdr->death_lock);
 217                                 }
 218                         }
 219
 220                         twdr->thread_slots &= ~(1 << i);
 221                 }
 222                 spin_unlock_bh(&twdr->death_lock);
 223         }
 224 }
 225
 226 EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
 227
 228 /* These are always called from BH context.  See callers in
 229  * tcp_input.c to verify this.
 230  */
 231
 232 /* This is for handling early-kills of TIME_WAIT sockets. */
 233 void inet_twsk_deschedule(struct inet_timewait_sock *tw,
 234                           struct inet_timewait_death_row *twdr)
 235 {
 236         spin_lock(&twdr->death_lock);
 237         if (inet_twsk_del_dead_node(tw)) {
 238                 inet_twsk_put(tw);
 239                 if (--twdr->tw_count == 0)
 240                         del_timer(&twdr->tw_timer);
 241         }
 242         spin_unlock(&twdr->death_lock);
 243         __inet_twsk_kill(tw, twdr->hashinfo);
 244 }
 245
 246 EXPORT_SYMBOL(inet_twsk_deschedule);
 247
 248 void inet_twsk_schedule(struct inet_timewait_sock *tw,
 249                        struct inet_timewait_death_row *twdr,
 250                        const int timeo, const int timewait_len)
 251 {
 252         struct hlist_head *list;
 253         int slot;
 254
 255         /* timeout := RTO * 3.5
 256          *
 257          * 3.5 = 1+2+0.5 to wait for two retransmits.
 258          *
 259          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
 260          * our ACK acking that FIN can be lost. If N subsequent retransmitted
 261          * FINs (or previous seqments) are lost (probability of such event
 262          * is p^(N+1), where p is probability to lose single packet and
 263          * time to detect the loss is about RTO*(2^N - 1) with exponential
 264          * backoff). Normal timewait length is calculated so, that we
 265          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
 266          * [ BTW Linux. following BSD, violates this requirement waiting
 267          *   only for 60sec, we should wait at least for 240 secs.
 268          *   Well, 240 consumes too much of resources 8)
 269          * ]
 270          * This interval is not reduced to catch old duplicate and
 271          * responces to our wandering segments living for two MSLs.
 272          * However, if we use PAWS to detect
 273          * old duplicates, we can reduce the interval to bounds required
 274          * by RTO, rather than MSL. So, if peer understands PAWS, we
 275          * kill tw bucket after 3.5*RTO (it is important that this number
 276          * is greater than TS tick!) and detect old duplicates with help
 277          * of PAWS.
 278          */
 279         slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
 280
 281         spin_lock(&twdr->death_lock);
 282
 283         /* Unlink it, if it was scheduled */
 284         if (inet_twsk_del_dead_node(tw))
 285                 twdr->tw_count--;
 286         else
 287                 atomic_inc(&tw->tw_refcnt);
 288
 289         if (slot >= INET_TWDR_RECYCLE_SLOTS) {
 290                 /* Schedule to slow timer */
 291                 if (timeo >= timewait_len) {
 292                         slot = INET_TWDR_TWKILL_SLOTS - 1;
 293                 } else {
 294                         slot = DIV_ROUND_UP(timeo, twdr->period);
 295                         if (slot >= INET_TWDR_TWKILL_SLOTS)
 296                                 slot = INET_TWDR_TWKILL_SLOTS - 1;
 297                 }
 298                 tw->tw_ttd = jiffies + timeo;
 299                 slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
 300                 list = &twdr->cells[slot];
 301         } else {
 302                 tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
 303
 304                 if (twdr->twcal_hand < 0) {
 305                         twdr->twcal_hand = 0;
 306                         twdr->twcal_jiffie = jiffies;
 307                         twdr->twcal_timer.expires = twdr->twcal_jiffie +
 308                                               (slot << INET_TWDR_RECYCLE_TICK);
 309                         add_timer(&twdr->twcal_timer);
 310                 } else {
 311                         if (time_after(twdr->twcal_timer.expires,
 312                                        jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
 313                                 mod_timer(&twdr->twcal_timer,
 314                                           jiffies + (slot << INET_TWDR_RECYCLE_TICK));
 315                         slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
 316                 }
 317                 list = &twdr->twcal_row[slot];
 318         }
 319
 320         hlist_add_head(&tw->tw_death_node, list);
 321
 322         if (twdr->tw_count++ == 0)
 323                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
 324         spin_unlock(&twdr->death_lock);
 325 }
 326
 327 EXPORT_SYMBOL_GPL(inet_twsk_schedule);
 328
 329 void inet_twdr_twcal_tick(unsigned long data)
 330 {
 331         struct inet_timewait_death_row *twdr;
 332         int n, slot;
 333         unsigned long j;
 334         unsigned long now = jiffies;
 335         int killed = 0;
 336         int adv = 0;
 337
 338         twdr = (struct inet_timewait_death_row *)data;
 339
 340         spin_lock(&twdr->death_lock);
 341         if (twdr->twcal_hand < 0)
 342                 goto out;
 343
 344         slot = twdr->twcal_hand;
 345         j = twdr->twcal_jiffie;
 346
 347         for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
 348                 if (time_before_eq(j, now)) {
 349                         struct hlist_node *node, *safe;
 350                         struct inet_timewait_sock *tw;
 351
 352                         inet_twsk_for_each_inmate_safe(tw, node, safe,
 353                                                        &twdr->twcal_row[slot]) {
 354                                 __inet_twsk_del_dead_node(tw);
 355                                 __inet_twsk_kill(tw, twdr->hashinfo);
 356                                 inet_twsk_put(tw);
 357                                 killed++;
 358                         }
 359                 } else {
 360                         if (!adv) {
 361                                 adv = 1;
 362                                 twdr->twcal_jiffie = j;
 363                                 twdr->twcal_hand = slot;
 364                         }
 365
 366                         if (!hlist_empty(&twdr->twcal_row[slot])) {
 367                                 mod_timer(&twdr->twcal_timer, j);
 368                                 goto out;
 369                         }
 370                 }
 371                 j += 1 << INET_TWDR_RECYCLE_TICK;
 372                 slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
 373         }
 374         twdr->twcal_hand = -1;
 375
 376 out:
 377         if ((twdr->tw_count -= killed) == 0)
 378                 del_timer(&twdr->tw_timer);
 379         NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITKILLED, killed);
 380         spin_unlock(&twdr->death_lock);
 381 }
 382
 383 EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);