SAFE public projects git trees. - safe/jmp/linux-2.6/blob - net/netfilter/nf_conntrack_core.c

   1 /* Connection state tracking for netfilter.  This is separated from,
   2    but required by, the NAT layer; it can also be used by an iptables
   3    extension. */
   4
   5 /* (C) 1999-2001 Paul `Rusty' Russell
   6  * (C) 2002-2006 Netfilter Core Team <coreteam@netfilter.org>
   7  * (C) 2003,2004 USAGI/WIDE Project <http://www.linux-ipv6.org>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  */
  13
  14 #include <linux/types.h>
  15 #include <linux/netfilter.h>
  16 #include <linux/module.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/proc_fs.h>
  19 #include <linux/vmalloc.h>
  20 #include <linux/stddef.h>
  21 #include <linux/slab.h>
  22 #include <linux/random.h>
  23 #include <linux/jhash.h>
  24 #include <linux/err.h>
  25 #include <linux/percpu.h>
  26 #include <linux/moduleparam.h>
  27 #include <linux/notifier.h>
  28 #include <linux/kernel.h>
  29 #include <linux/netdevice.h>
  30 #include <linux/socket.h>
  31 #include <linux/mm.h>
  32 #include <linux/rculist_nulls.h>
  33
  34 #include <net/netfilter/nf_conntrack.h>
  35 #include <net/netfilter/nf_conntrack_l3proto.h>
  36 #include <net/netfilter/nf_conntrack_l4proto.h>
  37 #include <net/netfilter/nf_conntrack_expect.h>
  38 #include <net/netfilter/nf_conntrack_helper.h>
  39 #include <net/netfilter/nf_conntrack_core.h>
  40 #include <net/netfilter/nf_conntrack_extend.h>
  41 #include <net/netfilter/nf_conntrack_acct.h>
  42 #include <net/netfilter/nf_conntrack_ecache.h>
  43 #include <net/netfilter/nf_nat.h>
  44 #include <net/netfilter/nf_nat_core.h>
  45
  46 #define NF_CONNTRACK_VERSION    "0.5.0"
  47
  48 int (*nfnetlink_parse_nat_setup_hook)(struct nf_conn *ct,
  49                                       enum nf_nat_manip_type manip,
  50                                       const struct nlattr *attr) __read_mostly;
  51 EXPORT_SYMBOL_GPL(nfnetlink_parse_nat_setup_hook);
  52
  53 DEFINE_SPINLOCK(nf_conntrack_lock);
  54 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
  55
  56 unsigned int nf_conntrack_htable_size __read_mostly;
  57 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
  58
  59 unsigned int nf_conntrack_max __read_mostly;
  60 EXPORT_SYMBOL_GPL(nf_conntrack_max);
  61
  62 struct nf_conn nf_conntrack_untracked __read_mostly;
  63 EXPORT_SYMBOL_GPL(nf_conntrack_untracked);
  64
  65 static struct kmem_cache *nf_conntrack_cachep __read_mostly;
  66
  67 static int nf_conntrack_hash_rnd_initted;
  68 static unsigned int nf_conntrack_hash_rnd;
  69
  70 static u_int32_t __hash_conntrack(const struct nf_conntrack_tuple *tuple,
  71                                   unsigned int size, unsigned int rnd)
  72 {
  73         unsigned int n;
  74         u_int32_t h;
  75
  76         /* The direction must be ignored, so we hash everything up to the
  77          * destination ports (which is a multiple of 4) and treat the last
  78          * three bytes manually.
  79          */
  80         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
  81         h = jhash2((u32 *)tuple, n,
  82                    rnd ^ (((__force __u16)tuple->dst.u.all << 16) |
  83                           tuple->dst.protonum));
  84
  85         return ((u64)h * size) >> 32;
  86 }
  87
  88 static inline u_int32_t hash_conntrack(const struct nf_conntrack_tuple *tuple)
  89 {
  90         return __hash_conntrack(tuple, nf_conntrack_htable_size,
  91                                 nf_conntrack_hash_rnd);
  92 }
  93
  94 bool
  95 nf_ct_get_tuple(const struct sk_buff *skb,
  96                 unsigned int nhoff,
  97                 unsigned int dataoff,
  98                 u_int16_t l3num,
  99                 u_int8_t protonum,
 100                 struct nf_conntrack_tuple *tuple,
 101                 const struct nf_conntrack_l3proto *l3proto,
 102                 const struct nf_conntrack_l4proto *l4proto)
 103 {
 104         memset(tuple, 0, sizeof(*tuple));
 105
 106         tuple->src.l3num = l3num;
 107         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 108                 return false;
 109
 110         tuple->dst.protonum = protonum;
 111         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 112
 113         return l4proto->pkt_to_tuple(skb, dataoff, tuple);
 114 }
 115 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 116
 117 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 118                        u_int16_t l3num, struct nf_conntrack_tuple *tuple)
 119 {
 120         struct nf_conntrack_l3proto *l3proto;
 121         struct nf_conntrack_l4proto *l4proto;
 122         unsigned int protoff;
 123         u_int8_t protonum;
 124         int ret;
 125
 126         rcu_read_lock();
 127
 128         l3proto = __nf_ct_l3proto_find(l3num);
 129         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
 130         if (ret != NF_ACCEPT) {
 131                 rcu_read_unlock();
 132                 return false;
 133         }
 134
 135         l4proto = __nf_ct_l4proto_find(l3num, protonum);
 136
 137         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, tuple,
 138                               l3proto, l4proto);
 139
 140         rcu_read_unlock();
 141         return ret;
 142 }
 143 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 144
 145 bool
 146 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 147                    const struct nf_conntrack_tuple *orig,
 148                    const struct nf_conntrack_l3proto *l3proto,
 149                    const struct nf_conntrack_l4proto *l4proto)
 150 {
 151         memset(inverse, 0, sizeof(*inverse));
 152
 153         inverse->src.l3num = orig->src.l3num;
 154         if (l3proto->invert_tuple(inverse, orig) == 0)
 155                 return false;
 156
 157         inverse->dst.dir = !orig->dst.dir;
 158
 159         inverse->dst.protonum = orig->dst.protonum;
 160         return l4proto->invert_tuple(inverse, orig);
 161 }
 162 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 163
 164 static void
 165 clean_from_lists(struct nf_conn *ct)
 166 {
 167         pr_debug("clean_from_lists(%p)\n", ct);
 168         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 169         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 170
 171         /* Destroy all pending expectations */
 172         nf_ct_remove_expectations(ct);
 173 }
 174
 175 static void
 176 destroy_conntrack(struct nf_conntrack *nfct)
 177 {
 178         struct nf_conn *ct = (struct nf_conn *)nfct;
 179         struct net *net = nf_ct_net(ct);
 180         struct nf_conntrack_l4proto *l4proto;
 181
 182         pr_debug("destroy_conntrack(%p)\n", ct);
 183         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 184         NF_CT_ASSERT(!timer_pending(&ct->timeout));
 185
 186         /* To make sure we don't get any weird locking issues here:
 187          * destroy_conntrack() MUST NOT be called with a write lock
 188          * to nf_conntrack_lock!!! -HW */
 189         rcu_read_lock();
 190         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 191         if (l4proto && l4proto->destroy)
 192                 l4proto->destroy(ct);
 193
 194         rcu_read_unlock();
 195
 196         spin_lock_bh(&nf_conntrack_lock);
 197         /* Expectations will have been removed in clean_from_lists,
 198          * except TFTP can create an expectation on the first packet,
 199          * before connection is in the list, so we need to clean here,
 200          * too. */
 201         nf_ct_remove_expectations(ct);
 202
 203         /* We overload first tuple to link into unconfirmed list. */
 204         if (!nf_ct_is_confirmed(ct)) {
 205                 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 206                 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 207         }
 208
 209         NF_CT_STAT_INC(net, delete);
 210         spin_unlock_bh(&nf_conntrack_lock);
 211
 212         if (ct->master)
 213                 nf_ct_put(ct->master);
 214
 215         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 216         nf_conntrack_free(ct);
 217 }
 218
 219 void nf_ct_delete_from_lists(struct nf_conn *ct)
 220 {
 221         struct net *net = nf_ct_net(ct);
 222
 223         nf_ct_helper_destroy(ct);
 224         spin_lock_bh(&nf_conntrack_lock);
 225         /* Inside lock so preempt is disabled on module removal path.
 226          * Otherwise we can get spurious warnings. */
 227         NF_CT_STAT_INC(net, delete_list);
 228         clean_from_lists(ct);
 229         spin_unlock_bh(&nf_conntrack_lock);
 230 }
 231 EXPORT_SYMBOL_GPL(nf_ct_delete_from_lists);
 232
 233 static void death_by_event(unsigned long ul_conntrack)
 234 {
 235         struct nf_conn *ct = (void *)ul_conntrack;
 236         struct net *net = nf_ct_net(ct);
 237
 238         if (nf_conntrack_event(IPCT_DESTROY, ct) < 0) {
 239                 /* bad luck, let's retry again */
 240                 ct->timeout.expires = jiffies +
 241                         (random32() % net->ct.sysctl_events_retry_timeout);
 242                 add_timer(&ct->timeout);
 243                 return;
 244         }
 245         /* we've got the event delivered, now it's dying */
 246         set_bit(IPS_DYING_BIT, &ct->status);
 247         spin_lock(&nf_conntrack_lock);
 248         hlist_nulls_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 249         spin_unlock(&nf_conntrack_lock);
 250         nf_ct_put(ct);
 251 }
 252
 253 void nf_ct_insert_dying_list(struct nf_conn *ct)
 254 {
 255         struct net *net = nf_ct_net(ct);
 256
 257         /* add this conntrack to the dying list */
 258         spin_lock_bh(&nf_conntrack_lock);
 259         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 260                              &net->ct.dying);
 261         spin_unlock_bh(&nf_conntrack_lock);
 262         /* set a new timer to retry event delivery */
 263         setup_timer(&ct->timeout, death_by_event, (unsigned long)ct);
 264         ct->timeout.expires = jiffies +
 265                 (random32() % net->ct.sysctl_events_retry_timeout);
 266         add_timer(&ct->timeout);
 267 }
 268 EXPORT_SYMBOL_GPL(nf_ct_insert_dying_list);
 269
 270 static void death_by_timeout(unsigned long ul_conntrack)
 271 {
 272         struct nf_conn *ct = (void *)ul_conntrack;
 273
 274         if (!test_bit(IPS_DYING_BIT, &ct->status) &&
 275             unlikely(nf_conntrack_event(IPCT_DESTROY, ct) < 0)) {
 276                 /* destroy event was not delivered */
 277                 nf_ct_delete_from_lists(ct);
 278                 nf_ct_insert_dying_list(ct);
 279                 return;
 280         }
 281         set_bit(IPS_DYING_BIT, &ct->status);
 282         nf_ct_delete_from_lists(ct);
 283         nf_ct_put(ct);
 284 }
 285
 286 /*
 287  * Warning :
 288  * - Caller must take a reference on returned object
 289  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 290  * OR
 291  * - Caller must lock nf_conntrack_lock before calling this function
 292  */
 293 struct nf_conntrack_tuple_hash *
 294 __nf_conntrack_find(struct net *net, const struct nf_conntrack_tuple *tuple)
 295 {
 296         struct nf_conntrack_tuple_hash *h;
 297         struct hlist_nulls_node *n;
 298         unsigned int hash = hash_conntrack(tuple);
 299
 300         /* Disable BHs the entire time since we normally need to disable them
 301          * at least once for the stats anyway.
 302          */
 303         local_bh_disable();
 304 begin:
 305         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
 306                 if (nf_ct_tuple_equal(tuple, &h->tuple)) {
 307                         NF_CT_STAT_INC(net, found);
 308                         local_bh_enable();
 309                         return h;
 310                 }
 311                 NF_CT_STAT_INC(net, searched);
 312         }
 313         /*
 314          * if the nulls value we got at the end of this lookup is
 315          * not the expected one, we must restart lookup.
 316          * We probably met an item that was moved to another chain.
 317          */
 318         if (get_nulls_value(n) != hash)
 319                 goto begin;
 320         local_bh_enable();
 321
 322         return NULL;
 323 }
 324 EXPORT_SYMBOL_GPL(__nf_conntrack_find);
 325
 326 /* Find a connection corresponding to a tuple. */
 327 struct nf_conntrack_tuple_hash *
 328 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_tuple *tuple)
 329 {
 330         struct nf_conntrack_tuple_hash *h;
 331         struct nf_conn *ct;
 332
 333         rcu_read_lock();
 334 begin:
 335         h = __nf_conntrack_find(net, tuple);
 336         if (h) {
 337                 ct = nf_ct_tuplehash_to_ctrack(h);
 338                 if (unlikely(nf_ct_is_dying(ct) ||
 339                              !atomic_inc_not_zero(&ct->ct_general.use)))
 340                         h = NULL;
 341                 else {
 342                         if (unlikely(!nf_ct_tuple_equal(tuple, &h->tuple))) {
 343                                 nf_ct_put(ct);
 344                                 goto begin;
 345                         }
 346                 }
 347         }
 348         rcu_read_unlock();
 349
 350         return h;
 351 }
 352 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 353
 354 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 355                                        unsigned int hash,
 356                                        unsigned int repl_hash)
 357 {
 358         struct net *net = nf_ct_net(ct);
 359
 360         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 361                            &net->ct.hash[hash]);
 362         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 363                            &net->ct.hash[repl_hash]);
 364 }
 365
 366 void nf_conntrack_hash_insert(struct nf_conn *ct)
 367 {
 368         unsigned int hash, repl_hash;
 369
 370         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 371         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 372
 373         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 374 }
 375 EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
 376
 377 /* Confirm a connection given skb; places it in hash table */
 378 int
 379 __nf_conntrack_confirm(struct sk_buff *skb)
 380 {
 381         unsigned int hash, repl_hash;
 382         struct nf_conntrack_tuple_hash *h;
 383         struct nf_conn *ct;
 384         struct nf_conn_help *help;
 385         struct hlist_nulls_node *n;
 386         enum ip_conntrack_info ctinfo;
 387         struct net *net;
 388
 389         ct = nf_ct_get(skb, &ctinfo);
 390         net = nf_ct_net(ct);
 391
 392         /* ipt_REJECT uses nf_conntrack_attach to attach related
 393            ICMP/TCP RST packets in other direction.  Actual packet
 394            which created connection will be IP_CT_NEW or for an
 395            expected connection, IP_CT_RELATED. */
 396         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 397                 return NF_ACCEPT;
 398
 399         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 400         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 401
 402         /* We're not in hash table, and we refuse to set up related
 403            connections for unconfirmed conns.  But packet copies and
 404            REJECT will give spurious warnings here. */
 405         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 406
 407         /* No external references means noone else could have
 408            confirmed us. */
 409         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 410         pr_debug("Confirming conntrack %p\n", ct);
 411
 412         spin_lock_bh(&nf_conntrack_lock);
 413
 414         /* See if there's one in the list already, including reverse:
 415            NAT could have grabbed it without realizing, since we're
 416            not in the hash.  If there is, we lost race. */
 417         hlist_nulls_for_each_entry(h, n, &net->ct.hash[hash], hnnode)
 418                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 419                                       &h->tuple))
 420                         goto out;
 421         hlist_nulls_for_each_entry(h, n, &net->ct.hash[repl_hash], hnnode)
 422                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 423                                       &h->tuple))
 424                         goto out;
 425
 426         /* Remove from unconfirmed list */
 427         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 428
 429         /* Timer relative to confirmation time, not original
 430            setting time, otherwise we'd get timer wrap in
 431            weird delay cases. */
 432         ct->timeout.expires += jiffies;
 433         add_timer(&ct->timeout);
 434         atomic_inc(&ct->ct_general.use);
 435         set_bit(IPS_CONFIRMED_BIT, &ct->status);
 436
 437         /* Since the lookup is lockless, hash insertion must be done after
 438          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 439          * guarantee that no other CPU can find the conntrack before the above
 440          * stores are visible.
 441          */
 442         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 443         NF_CT_STAT_INC(net, insert);
 444         spin_unlock_bh(&nf_conntrack_lock);
 445
 446         help = nfct_help(ct);
 447         if (help && help->helper)
 448                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 449
 450         nf_conntrack_event_cache(master_ct(ct) ?
 451                                  IPCT_RELATED : IPCT_NEW, ct);
 452         return NF_ACCEPT;
 453
 454 out:
 455         NF_CT_STAT_INC(net, insert_failed);
 456         spin_unlock_bh(&nf_conntrack_lock);
 457         return NF_DROP;
 458 }
 459 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 460
 461 /* Returns true if a connection correspondings to the tuple (required
 462    for NAT). */
 463 int
 464 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 465                          const struct nf_conn *ignored_conntrack)
 466 {
 467         struct net *net = nf_ct_net(ignored_conntrack);
 468         struct nf_conntrack_tuple_hash *h;
 469         struct hlist_nulls_node *n;
 470         unsigned int hash = hash_conntrack(tuple);
 471
 472         /* Disable BHs the entire time since we need to disable them at
 473          * least once for the stats anyway.
 474          */
 475         rcu_read_lock_bh();
 476         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
 477                 if (nf_ct_tuplehash_to_ctrack(h) != ignored_conntrack &&
 478                     nf_ct_tuple_equal(tuple, &h->tuple)) {
 479                         NF_CT_STAT_INC(net, found);
 480                         rcu_read_unlock_bh();
 481                         return 1;
 482                 }
 483                 NF_CT_STAT_INC(net, searched);
 484         }
 485         rcu_read_unlock_bh();
 486
 487         return 0;
 488 }
 489 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 490
 491 #define NF_CT_EVICTION_RANGE    8
 492
 493 /* There's a small race here where we may free a just-assured
 494    connection.  Too bad: we're in trouble anyway. */
 495 static noinline int early_drop(struct net *net, unsigned int hash)
 496 {
 497         /* Use oldest entry, which is roughly LRU */
 498         struct nf_conntrack_tuple_hash *h;
 499         struct nf_conn *ct = NULL, *tmp;
 500         struct hlist_nulls_node *n;
 501         unsigned int i, cnt = 0;
 502         int dropped = 0;
 503
 504         rcu_read_lock();
 505         for (i = 0; i < nf_conntrack_htable_size; i++) {
 506                 hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash],
 507                                          hnnode) {
 508                         tmp = nf_ct_tuplehash_to_ctrack(h);
 509                         if (!test_bit(IPS_ASSURED_BIT, &tmp->status))
 510                                 ct = tmp;
 511                         cnt++;
 512                 }
 513
 514                 if (ct && unlikely(nf_ct_is_dying(ct) ||
 515                                    !atomic_inc_not_zero(&ct->ct_general.use)))
 516                         ct = NULL;
 517                 if (ct || cnt >= NF_CT_EVICTION_RANGE)
 518                         break;
 519                 hash = (hash + 1) % nf_conntrack_htable_size;
 520         }
 521         rcu_read_unlock();
 522
 523         if (!ct)
 524                 return dropped;
 525
 526         if (del_timer(&ct->timeout)) {
 527                 death_by_timeout((unsigned long)ct);
 528                 dropped = 1;
 529                 NF_CT_STAT_INC_ATOMIC(net, early_drop);
 530         }
 531         nf_ct_put(ct);
 532         return dropped;
 533 }
 534
 535 struct nf_conn *nf_conntrack_alloc(struct net *net,
 536                                    const struct nf_conntrack_tuple *orig,
 537                                    const struct nf_conntrack_tuple *repl,
 538                                    gfp_t gfp)
 539 {
 540         struct nf_conn *ct;
 541
 542         if (unlikely(!nf_conntrack_hash_rnd_initted)) {
 543                 get_random_bytes(&nf_conntrack_hash_rnd,
 544                                 sizeof(nf_conntrack_hash_rnd));
 545                 nf_conntrack_hash_rnd_initted = 1;
 546         }
 547
 548         /* We don't want any race condition at early drop stage */
 549         atomic_inc(&net->ct.count);
 550
 551         if (nf_conntrack_max &&
 552             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
 553                 unsigned int hash = hash_conntrack(orig);
 554                 if (!early_drop(net, hash)) {
 555                         atomic_dec(&net->ct.count);
 556                         if (net_ratelimit())
 557                                 printk(KERN_WARNING
 558                                        "nf_conntrack: table full, dropping"
 559                                        " packet.\n");
 560                         return ERR_PTR(-ENOMEM);
 561                 }
 562         }
 563
 564         /*
 565          * Do not use kmem_cache_zalloc(), as this cache uses
 566          * SLAB_DESTROY_BY_RCU.
 567          */
 568         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
 569         if (ct == NULL) {
 570                 pr_debug("nf_conntrack_alloc: Can't alloc conntrack.\n");
 571                 atomic_dec(&net->ct.count);
 572                 return ERR_PTR(-ENOMEM);
 573         }
 574         /*
 575          * Let ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.next
 576          * and ct->tuplehash[IP_CT_DIR_REPLY].hnnode.next unchanged.
 577          */
 578         memset(&ct->tuplehash[IP_CT_DIR_MAX], 0,
 579                sizeof(*ct) - offsetof(struct nf_conn, tuplehash[IP_CT_DIR_MAX]));
 580         spin_lock_init(&ct->lock);
 581         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
 582         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
 583         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
 584         ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev = NULL;
 585         /* Don't set timer yet: wait for confirmation */
 586         setup_timer(&ct->timeout, death_by_timeout, (unsigned long)ct);
 587 #ifdef CONFIG_NET_NS
 588         ct->ct_net = net;
 589 #endif
 590
 591         /*
 592          * changes to lookup keys must be done before setting refcnt to 1
 593          */
 594         smp_wmb();
 595         atomic_set(&ct->ct_general.use, 1);
 596         return ct;
 597 }
 598 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
 599
 600 void nf_conntrack_free(struct nf_conn *ct)
 601 {
 602         struct net *net = nf_ct_net(ct);
 603
 604         nf_ct_ext_destroy(ct);
 605         atomic_dec(&net->ct.count);
 606         nf_ct_ext_free(ct);
 607         kmem_cache_free(nf_conntrack_cachep, ct);
 608 }
 609 EXPORT_SYMBOL_GPL(nf_conntrack_free);
 610
 611 /* Allocate a new conntrack: we return -ENOMEM if classification
 612    failed due to stress.  Otherwise it really is unclassifiable. */
 613 static struct nf_conntrack_tuple_hash *
 614 init_conntrack(struct net *net,
 615                const struct nf_conntrack_tuple *tuple,
 616                struct nf_conntrack_l3proto *l3proto,
 617                struct nf_conntrack_l4proto *l4proto,
 618                struct sk_buff *skb,
 619                unsigned int dataoff)
 620 {
 621         struct nf_conn *ct;
 622         struct nf_conn_help *help;
 623         struct nf_conntrack_tuple repl_tuple;
 624         struct nf_conntrack_expect *exp;
 625
 626         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
 627                 pr_debug("Can't invert tuple.\n");
 628                 return NULL;
 629         }
 630
 631         ct = nf_conntrack_alloc(net, tuple, &repl_tuple, GFP_ATOMIC);
 632         if (IS_ERR(ct)) {
 633                 pr_debug("Can't allocate conntrack.\n");
 634                 return (struct nf_conntrack_tuple_hash *)ct;
 635         }
 636
 637         if (!l4proto->new(ct, skb, dataoff)) {
 638                 nf_conntrack_free(ct);
 639                 pr_debug("init conntrack: can't track with proto module\n");
 640                 return NULL;
 641         }
 642
 643         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
 644         nf_ct_ecache_ext_add(ct, GFP_ATOMIC);
 645
 646         spin_lock_bh(&nf_conntrack_lock);
 647         exp = nf_ct_find_expectation(net, tuple);
 648         if (exp) {
 649                 pr_debug("conntrack: expectation arrives ct=%p exp=%p\n",
 650                          ct, exp);
 651                 /* Welcome, Mr. Bond.  We've been expecting you... */
 652                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
 653                 ct->master = exp->master;
 654                 if (exp->helper) {
 655                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
 656                         if (help)
 657                                 rcu_assign_pointer(help->helper, exp->helper);
 658                 }
 659
 660 #ifdef CONFIG_NF_CONNTRACK_MARK
 661                 ct->mark = exp->master->mark;
 662 #endif
 663 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 664                 ct->secmark = exp->master->secmark;
 665 #endif
 666                 nf_conntrack_get(&ct->master->ct_general);
 667                 NF_CT_STAT_INC(net, expect_new);
 668         } else {
 669                 __nf_ct_try_assign_helper(ct, GFP_ATOMIC);
 670                 NF_CT_STAT_INC(net, new);
 671         }
 672
 673         /* Overload tuple linked list to put us in unconfirmed list. */
 674         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 675                        &net->ct.unconfirmed);
 676
 677         spin_unlock_bh(&nf_conntrack_lock);
 678
 679         if (exp) {
 680                 if (exp->expectfn)
 681                         exp->expectfn(ct, exp);
 682                 nf_ct_expect_put(exp);
 683         }
 684
 685         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
 686 }
 687
 688 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
 689 static inline struct nf_conn *
 690 resolve_normal_ct(struct net *net,
 691                   struct sk_buff *skb,
 692                   unsigned int dataoff,
 693                   u_int16_t l3num,
 694                   u_int8_t protonum,
 695                   struct nf_conntrack_l3proto *l3proto,
 696                   struct nf_conntrack_l4proto *l4proto,
 697                   int *set_reply,
 698                   enum ip_conntrack_info *ctinfo)
 699 {
 700         struct nf_conntrack_tuple tuple;
 701         struct nf_conntrack_tuple_hash *h;
 702         struct nf_conn *ct;
 703
 704         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
 705                              dataoff, l3num, protonum, &tuple, l3proto,
 706                              l4proto)) {
 707                 pr_debug("resolve_normal_ct: Can't get tuple\n");
 708                 return NULL;
 709         }
 710
 711         /* look for tuple match */
 712         h = nf_conntrack_find_get(net, &tuple);
 713         if (!h) {
 714                 h = init_conntrack(net, &tuple, l3proto, l4proto, skb, dataoff);
 715                 if (!h)
 716                         return NULL;
 717                 if (IS_ERR(h))
 718                         return (void *)h;
 719         }
 720         ct = nf_ct_tuplehash_to_ctrack(h);
 721
 722         /* It exists; we have (non-exclusive) reference. */
 723         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
 724                 *ctinfo = IP_CT_ESTABLISHED + IP_CT_IS_REPLY;
 725                 /* Please set reply bit if this packet OK */
 726                 *set_reply = 1;
 727         } else {
 728                 /* Once we've had two way comms, always ESTABLISHED. */
 729                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
 730                         pr_debug("nf_conntrack_in: normal packet for %p\n", ct);
 731                         *ctinfo = IP_CT_ESTABLISHED;
 732                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
 733                         pr_debug("nf_conntrack_in: related packet for %p\n",
 734                                  ct);
 735                         *ctinfo = IP_CT_RELATED;
 736                 } else {
 737                         pr_debug("nf_conntrack_in: new packet for %p\n", ct);
 738                         *ctinfo = IP_CT_NEW;
 739                 }
 740                 *set_reply = 0;
 741         }
 742         skb->nfct = &ct->ct_general;
 743         skb->nfctinfo = *ctinfo;
 744         return ct;
 745 }
 746
 747 unsigned int
 748 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
 749                 struct sk_buff *skb)
 750 {
 751         struct nf_conn *ct;
 752         enum ip_conntrack_info ctinfo;
 753         struct nf_conntrack_l3proto *l3proto;
 754         struct nf_conntrack_l4proto *l4proto;
 755         unsigned int dataoff;
 756         u_int8_t protonum;
 757         int set_reply = 0;
 758         int ret;
 759
 760         /* Previously seen (loopback or untracked)?  Ignore. */
 761         if (skb->nfct) {
 762                 NF_CT_STAT_INC_ATOMIC(net, ignore);
 763                 return NF_ACCEPT;
 764         }
 765
 766         /* rcu_read_lock()ed by nf_hook_slow */
 767         l3proto = __nf_ct_l3proto_find(pf);
 768         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
 769                                    &dataoff, &protonum);
 770         if (ret <= 0) {
 771                 pr_debug("not prepared to track yet or error occured\n");
 772                 NF_CT_STAT_INC_ATOMIC(net, error);
 773                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 774                 return -ret;
 775         }
 776
 777         l4proto = __nf_ct_l4proto_find(pf, protonum);
 778
 779         /* It may be an special packet, error, unclean...
 780          * inverse of the return code tells to the netfilter
 781          * core what to do with the packet. */
 782         if (l4proto->error != NULL) {
 783                 ret = l4proto->error(net, skb, dataoff, &ctinfo, pf, hooknum);
 784                 if (ret <= 0) {
 785                         NF_CT_STAT_INC_ATOMIC(net, error);
 786                         NF_CT_STAT_INC_ATOMIC(net, invalid);
 787                         return -ret;
 788                 }
 789         }
 790
 791         ct = resolve_normal_ct(net, skb, dataoff, pf, protonum,
 792                                l3proto, l4proto, &set_reply, &ctinfo);
 793         if (!ct) {
 794                 /* Not valid part of a connection */
 795                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 796                 return NF_ACCEPT;
 797         }
 798
 799         if (IS_ERR(ct)) {
 800                 /* Too stressed to deal. */
 801                 NF_CT_STAT_INC_ATOMIC(net, drop);
 802                 return NF_DROP;
 803         }
 804
 805         NF_CT_ASSERT(skb->nfct);
 806
 807         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum);
 808         if (ret <= 0) {
 809                 /* Invalid: inverse of the return code tells
 810                  * the netfilter core what to do */
 811                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
 812                 nf_conntrack_put(skb->nfct);
 813                 skb->nfct = NULL;
 814                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 815                 if (ret == -NF_DROP)
 816                         NF_CT_STAT_INC_ATOMIC(net, drop);
 817                 return -ret;
 818         }
 819
 820         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
 821                 nf_conntrack_event_cache(IPCT_STATUS, ct);
 822
 823         return ret;
 824 }
 825 EXPORT_SYMBOL_GPL(nf_conntrack_in);
 826
 827 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
 828                           const struct nf_conntrack_tuple *orig)
 829 {
 830         bool ret;
 831
 832         rcu_read_lock();
 833         ret = nf_ct_invert_tuple(inverse, orig,
 834                                  __nf_ct_l3proto_find(orig->src.l3num),
 835                                  __nf_ct_l4proto_find(orig->src.l3num,
 836                                                       orig->dst.protonum));
 837         rcu_read_unlock();
 838         return ret;
 839 }
 840 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
 841
 842 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
 843    implicitly racy: see __nf_conntrack_confirm */
 844 void nf_conntrack_alter_reply(struct nf_conn *ct,
 845                               const struct nf_conntrack_tuple *newreply)
 846 {
 847         struct nf_conn_help *help = nfct_help(ct);
 848
 849         /* Should be unconfirmed, so not in hash table yet */
 850         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 851
 852         pr_debug("Altering reply tuple of %p to ", ct);
 853         nf_ct_dump_tuple(newreply);
 854
 855         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
 856         if (ct->master || (help && !hlist_empty(&help->expectations)))
 857                 return;
 858
 859         rcu_read_lock();
 860         __nf_ct_try_assign_helper(ct, GFP_ATOMIC);
 861         rcu_read_unlock();
 862 }
 863 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
 864
 865 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
 866 void __nf_ct_refresh_acct(struct nf_conn *ct,
 867                           enum ip_conntrack_info ctinfo,
 868                           const struct sk_buff *skb,
 869                           unsigned long extra_jiffies,
 870                           int do_acct)
 871 {
 872         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
 873         NF_CT_ASSERT(skb);
 874
 875         /* Only update if this is not a fixed timeout */
 876         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
 877                 goto acct;
 878
 879         /* If not in hash table, timer will not be active yet */
 880         if (!nf_ct_is_confirmed(ct)) {
 881                 ct->timeout.expires = extra_jiffies;
 882         } else {
 883                 unsigned long newtime = jiffies + extra_jiffies;
 884
 885                 /* Only update the timeout if the new timeout is at least
 886                    HZ jiffies from the old timeout. Need del_timer for race
 887                    avoidance (may already be dying). */
 888                 if (newtime - ct->timeout.expires >= HZ)
 889                         mod_timer_pending(&ct->timeout, newtime);
 890         }
 891
 892 acct:
 893         if (do_acct) {
 894                 struct nf_conn_counter *acct;
 895
 896                 acct = nf_conn_acct_find(ct);
 897                 if (acct) {
 898                         spin_lock_bh(&ct->lock);
 899                         acct[CTINFO2DIR(ctinfo)].packets++;
 900                         acct[CTINFO2DIR(ctinfo)].bytes +=
 901                                 skb->len - skb_network_offset(skb);
 902                         spin_unlock_bh(&ct->lock);
 903                 }
 904         }
 905 }
 906 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
 907
 908 bool __nf_ct_kill_acct(struct nf_conn *ct,
 909                        enum ip_conntrack_info ctinfo,
 910                        const struct sk_buff *skb,
 911                        int do_acct)
 912 {
 913         if (do_acct) {
 914                 struct nf_conn_counter *acct;
 915
 916                 acct = nf_conn_acct_find(ct);
 917                 if (acct) {
 918                         spin_lock_bh(&ct->lock);
 919                         acct[CTINFO2DIR(ctinfo)].packets++;
 920                         acct[CTINFO2DIR(ctinfo)].bytes +=
 921                                 skb->len - skb_network_offset(skb);
 922                         spin_unlock_bh(&ct->lock);
 923                 }
 924         }
 925
 926         if (del_timer(&ct->timeout)) {
 927                 ct->timeout.function((unsigned long)ct);
 928                 return true;
 929         }
 930         return false;
 931 }
 932 EXPORT_SYMBOL_GPL(__nf_ct_kill_acct);
 933
 934 #if defined(CONFIG_NF_CT_NETLINK) || defined(CONFIG_NF_CT_NETLINK_MODULE)
 935
 936 #include <linux/netfilter/nfnetlink.h>
 937 #include <linux/netfilter/nfnetlink_conntrack.h>
 938 #include <linux/mutex.h>
 939
 940 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
 941  * in ip_conntrack_core, since we don't want the protocols to autoload
 942  * or depend on ctnetlink */
 943 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
 944                                const struct nf_conntrack_tuple *tuple)
 945 {
 946         NLA_PUT_BE16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port);
 947         NLA_PUT_BE16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port);
 948         return 0;
 949
 950 nla_put_failure:
 951         return -1;
 952 }
 953 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
 954
 955 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
 956         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
 957         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
 958 };
 959 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
 960
 961 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
 962                                struct nf_conntrack_tuple *t)
 963 {
 964         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
 965                 return -EINVAL;
 966
 967         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
 968         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
 969
 970         return 0;
 971 }
 972 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
 973
 974 int nf_ct_port_nlattr_tuple_size(void)
 975 {
 976         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
 977 }
 978 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
 979 #endif
 980
 981 /* Used by ipt_REJECT and ip6t_REJECT. */
 982 static void nf_conntrack_attach(struct sk_buff *nskb, struct sk_buff *skb)
 983 {
 984         struct nf_conn *ct;
 985         enum ip_conntrack_info ctinfo;
 986
 987         /* This ICMP is in reverse direction to the packet which caused it */
 988         ct = nf_ct_get(skb, &ctinfo);
 989         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
 990                 ctinfo = IP_CT_RELATED + IP_CT_IS_REPLY;
 991         else
 992                 ctinfo = IP_CT_RELATED;
 993
 994         /* Attach to new skbuff, and increment count */
 995         nskb->nfct = &ct->ct_general;
 996         nskb->nfctinfo = ctinfo;
 997         nf_conntrack_get(nskb->nfct);
 998 }
 999
1000 /* Bring out ya dead! */
1001 static struct nf_conn *
1002 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1003                 void *data, unsigned int *bucket)
1004 {
1005         struct nf_conntrack_tuple_hash *h;
1006         struct nf_conn *ct;
1007         struct hlist_nulls_node *n;
1008
1009         spin_lock_bh(&nf_conntrack_lock);
1010         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1011                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[*bucket], hnnode) {
1012                         ct = nf_ct_tuplehash_to_ctrack(h);
1013                         if (iter(ct, data))
1014                                 goto found;
1015                 }
1016         }
1017         hlist_nulls_for_each_entry(h, n, &net->ct.unconfirmed, hnnode) {
1018                 ct = nf_ct_tuplehash_to_ctrack(h);
1019                 if (iter(ct, data))
1020                         set_bit(IPS_DYING_BIT, &ct->status);
1021         }
1022         spin_unlock_bh(&nf_conntrack_lock);
1023         return NULL;
1024 found:
1025         atomic_inc(&ct->ct_general.use);
1026         spin_unlock_bh(&nf_conntrack_lock);
1027         return ct;
1028 }
1029
1030 void nf_ct_iterate_cleanup(struct net *net,
1031                            int (*iter)(struct nf_conn *i, void *data),
1032                            void *data)
1033 {
1034         struct nf_conn *ct;
1035         unsigned int bucket = 0;
1036
1037         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1038                 /* Time to push up daises... */
1039                 if (del_timer(&ct->timeout))
1040                         death_by_timeout((unsigned long)ct);
1041                 /* ... else the timer will get him soon. */
1042
1043                 nf_ct_put(ct);
1044         }
1045 }
1046 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1047
1048 struct __nf_ct_flush_report {
1049         u32 pid;
1050         int report;
1051 };
1052
1053 static int kill_report(struct nf_conn *i, void *data)
1054 {
1055         struct __nf_ct_flush_report *fr = (struct __nf_ct_flush_report *)data;
1056
1057         /* If we fail to deliver the event, death_by_timeout() will retry */
1058         if (nf_conntrack_event_report(IPCT_DESTROY, i,
1059                                       fr->pid, fr->report) < 0)
1060                 return 1;
1061
1062         /* Avoid the delivery of the destroy event in death_by_timeout(). */
1063         set_bit(IPS_DYING_BIT, &i->status);
1064         return 1;
1065 }
1066
1067 static int kill_all(struct nf_conn *i, void *data)
1068 {
1069         return 1;
1070 }
1071
1072 void nf_ct_free_hashtable(void *hash, int vmalloced, unsigned int size)
1073 {
1074         if (vmalloced)
1075                 vfree(hash);
1076         else
1077                 free_pages((unsigned long)hash,
1078                            get_order(sizeof(struct hlist_head) * size));
1079 }
1080 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1081
1082 void nf_conntrack_flush_report(struct net *net, u32 pid, int report)
1083 {
1084         struct __nf_ct_flush_report fr = {
1085                 .pid    = pid,
1086                 .report = report,
1087         };
1088         nf_ct_iterate_cleanup(net, kill_report, &fr);
1089 }
1090 EXPORT_SYMBOL_GPL(nf_conntrack_flush_report);
1091
1092 static void nf_ct_release_dying_list(struct net *net)
1093 {
1094         struct nf_conntrack_tuple_hash *h;
1095         struct nf_conn *ct;
1096         struct hlist_nulls_node *n;
1097
1098         spin_lock_bh(&nf_conntrack_lock);
1099         hlist_nulls_for_each_entry(h, n, &net->ct.dying, hnnode) {
1100                 ct = nf_ct_tuplehash_to_ctrack(h);
1101                 /* never fails to remove them, no listeners at this point */
1102                 nf_ct_kill(ct);
1103         }
1104         spin_unlock_bh(&nf_conntrack_lock);
1105 }
1106
1107 static void nf_conntrack_cleanup_init_net(void)
1108 {
1109         nf_conntrack_helper_fini();
1110         nf_conntrack_proto_fini();
1111         kmem_cache_destroy(nf_conntrack_cachep);
1112 }
1113
1114 static void nf_conntrack_cleanup_net(struct net *net)
1115 {
1116  i_see_dead_people:
1117         nf_ct_iterate_cleanup(net, kill_all, NULL);
1118         nf_ct_release_dying_list(net);
1119         if (atomic_read(&net->ct.count) != 0) {
1120                 schedule();
1121                 goto i_see_dead_people;
1122         }
1123         /* wait until all references to nf_conntrack_untracked are dropped */
1124         while (atomic_read(&nf_conntrack_untracked.ct_general.use) > 1)
1125                 schedule();
1126
1127         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1128                              nf_conntrack_htable_size);
1129         nf_conntrack_ecache_fini(net);
1130         nf_conntrack_acct_fini(net);
1131         nf_conntrack_expect_fini(net);
1132         free_percpu(net->ct.stat);
1133 }
1134
1135 /* Mishearing the voices in his head, our hero wonders how he's
1136    supposed to kill the mall. */
1137 void nf_conntrack_cleanup(struct net *net)
1138 {
1139         if (net_eq(net, &init_net))
1140                 rcu_assign_pointer(ip_ct_attach, NULL);
1141
1142         /* This makes sure all current packets have passed through
1143            netfilter framework.  Roll on, two-stage module
1144            delete... */
1145         synchronize_net();
1146
1147         nf_conntrack_cleanup_net(net);
1148
1149         if (net_eq(net, &init_net)) {
1150                 rcu_assign_pointer(nf_ct_destroy, NULL);
1151                 nf_conntrack_cleanup_init_net();
1152         }
1153 }
1154
1155 void *nf_ct_alloc_hashtable(unsigned int *sizep, int *vmalloced, int nulls)
1156 {
1157         struct hlist_nulls_head *hash;
1158         unsigned int nr_slots, i;
1159         size_t sz;
1160
1161         *vmalloced = 0;
1162
1163         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1164         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1165         sz = nr_slots * sizeof(struct hlist_nulls_head);
1166         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1167                                         get_order(sz));
1168         if (!hash) {
1169                 *vmalloced = 1;
1170                 printk(KERN_WARNING "nf_conntrack: falling back to vmalloc.\n");
1171                 hash = __vmalloc(sz, GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL);
1172         }
1173
1174         if (hash && nulls)
1175                 for (i = 0; i < nr_slots; i++)
1176                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1177
1178         return hash;
1179 }
1180 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1181
1182 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1183 {
1184         int i, bucket, vmalloced, old_vmalloced;
1185         unsigned int hashsize, old_size;
1186         int rnd;
1187         struct hlist_nulls_head *hash, *old_hash;
1188         struct nf_conntrack_tuple_hash *h;
1189
1190         /* On boot, we can set this without any fancy locking. */
1191         if (!nf_conntrack_htable_size)
1192                 return param_set_uint(val, kp);
1193
1194         hashsize = simple_strtoul(val, NULL, 0);
1195         if (!hashsize)
1196                 return -EINVAL;
1197
1198         hash = nf_ct_alloc_hashtable(&hashsize, &vmalloced, 1);
1199         if (!hash)
1200                 return -ENOMEM;
1201
1202         /* We have to rehahs for the new table anyway, so we also can
1203          * use a newrandom seed */
1204         get_random_bytes(&rnd, sizeof(rnd));
1205
1206         /* Lookups in the old hash might happen in parallel, which means we
1207          * might get false negatives during connection lookup. New connections
1208          * created because of a false negative won't make it into the hash
1209          * though since that required taking the lock.
1210          */
1211         spin_lock_bh(&nf_conntrack_lock);
1212         for (i = 0; i < nf_conntrack_htable_size; i++) {
1213                 while (!hlist_nulls_empty(&init_net.ct.hash[i])) {
1214                         h = hlist_nulls_entry(init_net.ct.hash[i].first,
1215                                         struct nf_conntrack_tuple_hash, hnnode);
1216                         hlist_nulls_del_rcu(&h->hnnode);
1217                         bucket = __hash_conntrack(&h->tuple, hashsize, rnd);
1218                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1219                 }
1220         }
1221         old_size = nf_conntrack_htable_size;
1222         old_vmalloced = init_net.ct.hash_vmalloc;
1223         old_hash = init_net.ct.hash;
1224
1225         nf_conntrack_htable_size = hashsize;
1226         init_net.ct.hash_vmalloc = vmalloced;
1227         init_net.ct.hash = hash;
1228         nf_conntrack_hash_rnd = rnd;
1229         spin_unlock_bh(&nf_conntrack_lock);
1230
1231         nf_ct_free_hashtable(old_hash, old_vmalloced, old_size);
1232         return 0;
1233 }
1234 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1235
1236 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1237                   &nf_conntrack_htable_size, 0600);
1238
1239 static int nf_conntrack_init_init_net(void)
1240 {
1241         int max_factor = 8;
1242         int ret;
1243
1244         /* Idea from tcp.c: use 1/16384 of memory.  On i386: 32MB
1245          * machine has 512 buckets. >= 1GB machines have 16384 buckets. */
1246         if (!nf_conntrack_htable_size) {
1247                 nf_conntrack_htable_size
1248                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1249                            / sizeof(struct hlist_head));
1250                 if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1251                         nf_conntrack_htable_size = 16384;
1252                 if (nf_conntrack_htable_size < 32)
1253                         nf_conntrack_htable_size = 32;
1254
1255                 /* Use a max. factor of four by default to get the same max as
1256                  * with the old struct list_heads. When a table size is given
1257                  * we use the old value of 8 to avoid reducing the max.
1258                  * entries. */
1259                 max_factor = 4;
1260         }
1261         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1262
1263         printk("nf_conntrack version %s (%u buckets, %d max)\n",
1264                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1265                nf_conntrack_max);
1266
1267         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
1268                                                 sizeof(struct nf_conn),
1269                                                 0, SLAB_DESTROY_BY_RCU, NULL);
1270         if (!nf_conntrack_cachep) {
1271                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
1272                 ret = -ENOMEM;
1273                 goto err_cache;
1274         }
1275
1276         ret = nf_conntrack_proto_init();
1277         if (ret < 0)
1278                 goto err_proto;
1279
1280         ret = nf_conntrack_helper_init();
1281         if (ret < 0)
1282                 goto err_helper;
1283
1284         return 0;
1285
1286 err_helper:
1287         nf_conntrack_proto_fini();
1288 err_proto:
1289         kmem_cache_destroy(nf_conntrack_cachep);
1290 err_cache:
1291         return ret;
1292 }
1293
1294 /*
1295  * We need to use special "null" values, not used in hash table
1296  */
1297 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1298 #define DYING_NULLS_VAL         ((1<<30)+1)
1299
1300 static int nf_conntrack_init_net(struct net *net)
1301 {
1302         int ret;
1303
1304         atomic_set(&net->ct.count, 0);
1305         INIT_HLIST_NULLS_HEAD(&net->ct.unconfirmed, UNCONFIRMED_NULLS_VAL);
1306         INIT_HLIST_NULLS_HEAD(&net->ct.dying, DYING_NULLS_VAL);
1307         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
1308         if (!net->ct.stat) {
1309                 ret = -ENOMEM;
1310                 goto err_stat;
1311         }
1312         net->ct.hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size,
1313                                              &net->ct.hash_vmalloc, 1);
1314         if (!net->ct.hash) {
1315                 ret = -ENOMEM;
1316                 printk(KERN_ERR "Unable to create nf_conntrack_hash\n");
1317                 goto err_hash;
1318         }
1319         ret = nf_conntrack_expect_init(net);
1320         if (ret < 0)
1321                 goto err_expect;
1322         ret = nf_conntrack_acct_init(net);
1323         if (ret < 0)
1324                 goto err_acct;
1325         ret = nf_conntrack_ecache_init(net);
1326         if (ret < 0)
1327                 goto err_ecache;
1328
1329         /* Set up fake conntrack:
1330             - to never be deleted, not in any hashes */
1331 #ifdef CONFIG_NET_NS
1332         nf_conntrack_untracked.ct_net = &init_net;
1333 #endif
1334         atomic_set(&nf_conntrack_untracked.ct_general.use, 1);
1335         /*  - and look it like as a confirmed connection */
1336         set_bit(IPS_CONFIRMED_BIT, &nf_conntrack_untracked.status);
1337
1338         return 0;
1339
1340 err_ecache:
1341         nf_conntrack_acct_fini(net);
1342 err_acct:
1343         nf_conntrack_expect_fini(net);
1344 err_expect:
1345         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1346                              nf_conntrack_htable_size);
1347 err_hash:
1348         free_percpu(net->ct.stat);
1349 err_stat:
1350         return ret;
1351 }
1352
1353 int nf_conntrack_init(struct net *net)
1354 {
1355         int ret;
1356
1357         if (net_eq(net, &init_net)) {
1358                 ret = nf_conntrack_init_init_net();
1359                 if (ret < 0)
1360                         goto out_init_net;
1361         }
1362         ret = nf_conntrack_init_net(net);
1363         if (ret < 0)
1364                 goto out_net;
1365
1366         if (net_eq(net, &init_net)) {
1367                 /* For use by REJECT target */
1368                 rcu_assign_pointer(ip_ct_attach, nf_conntrack_attach);
1369                 rcu_assign_pointer(nf_ct_destroy, destroy_conntrack);
1370         }
1371         return 0;
1372
1373 out_net:
1374         if (net_eq(net, &init_net))
1375                 nf_conntrack_cleanup_init_net();
1376 out_init_net:
1377         return ret;
1378 }