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/sched.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/proc_fs.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/stddef.h>
  22 #include <linux/slab.h>
  23 #include <linux/random.h>
  24 #include <linux/jhash.h>
  25 #include <linux/err.h>
  26 #include <linux/percpu.h>
  27 #include <linux/moduleparam.h>
  28 #include <linux/notifier.h>
  29 #include <linux/kernel.h>
  30 #include <linux/netdevice.h>
  31 #include <linux/socket.h>
  32 #include <linux/mm.h>
  33 #include <linux/rculist_nulls.h>
  34
  35 #include <net/netfilter/nf_conntrack.h>
  36 #include <net/netfilter/nf_conntrack_l3proto.h>
  37 #include <net/netfilter/nf_conntrack_l4proto.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_core.h>
  41 #include <net/netfilter/nf_conntrack_extend.h>
  42 #include <net/netfilter/nf_conntrack_acct.h>
  43 #include <net/netfilter/nf_conntrack_ecache.h>
  44 #include <net/netfilter/nf_nat.h>
  45 #include <net/netfilter/nf_nat_core.h>
  46
  47 #define NF_CONNTRACK_VERSION    "0.5.0"
  48
  49 int (*nfnetlink_parse_nat_setup_hook)(struct nf_conn *ct,
  50                                       enum nf_nat_manip_type manip,
  51                                       const struct nlattr *attr) __read_mostly;
  52 EXPORT_SYMBOL_GPL(nfnetlink_parse_nat_setup_hook);
  53
  54 DEFINE_SPINLOCK(nf_conntrack_lock);
  55 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
  56
  57 unsigned int nf_conntrack_htable_size __read_mostly;
  58 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
  59
  60 unsigned int nf_conntrack_max __read_mostly;
  61 EXPORT_SYMBOL_GPL(nf_conntrack_max);
  62
  63 struct nf_conn nf_conntrack_untracked __read_mostly;
  64 EXPORT_SYMBOL_GPL(nf_conntrack_untracked);
  65
  66 static struct kmem_cache *nf_conntrack_cachep __read_mostly;
  67
  68 static int nf_conntrack_hash_rnd_initted;
  69 static unsigned int nf_conntrack_hash_rnd;
  70
  71 static u_int32_t __hash_conntrack(const struct nf_conntrack_tuple *tuple,
  72                                   unsigned int size, unsigned int rnd)
  73 {
  74         unsigned int n;
  75         u_int32_t h;
  76
  77         /* The direction must be ignored, so we hash everything up to the
  78          * destination ports (which is a multiple of 4) and treat the last
  79          * three bytes manually.
  80          */
  81         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
  82         h = jhash2((u32 *)tuple, n,
  83                    rnd ^ (((__force __u16)tuple->dst.u.all << 16) |
  84                           tuple->dst.protonum));
  85
  86         return ((u64)h * size) >> 32;
  87 }
  88
  89 static inline u_int32_t hash_conntrack(const struct nf_conntrack_tuple *tuple)
  90 {
  91         return __hash_conntrack(tuple, nf_conntrack_htable_size,
  92                                 nf_conntrack_hash_rnd);
  93 }
  94
  95 bool
  96 nf_ct_get_tuple(const struct sk_buff *skb,
  97                 unsigned int nhoff,
  98                 unsigned int dataoff,
  99                 u_int16_t l3num,
 100                 u_int8_t protonum,
 101                 struct nf_conntrack_tuple *tuple,
 102                 const struct nf_conntrack_l3proto *l3proto,
 103                 const struct nf_conntrack_l4proto *l4proto)
 104 {
 105         memset(tuple, 0, sizeof(*tuple));
 106
 107         tuple->src.l3num = l3num;
 108         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 109                 return false;
 110
 111         tuple->dst.protonum = protonum;
 112         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 113
 114         return l4proto->pkt_to_tuple(skb, dataoff, tuple);
 115 }
 116 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 117
 118 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 119                        u_int16_t l3num, struct nf_conntrack_tuple *tuple)
 120 {
 121         struct nf_conntrack_l3proto *l3proto;
 122         struct nf_conntrack_l4proto *l4proto;
 123         unsigned int protoff;
 124         u_int8_t protonum;
 125         int ret;
 126
 127         rcu_read_lock();
 128
 129         l3proto = __nf_ct_l3proto_find(l3num);
 130         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
 131         if (ret != NF_ACCEPT) {
 132                 rcu_read_unlock();
 133                 return false;
 134         }
 135
 136         l4proto = __nf_ct_l4proto_find(l3num, protonum);
 137
 138         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, tuple,
 139                               l3proto, l4proto);
 140
 141         rcu_read_unlock();
 142         return ret;
 143 }
 144 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 145
 146 bool
 147 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 148                    const struct nf_conntrack_tuple *orig,
 149                    const struct nf_conntrack_l3proto *l3proto,
 150                    const struct nf_conntrack_l4proto *l4proto)
 151 {
 152         memset(inverse, 0, sizeof(*inverse));
 153
 154         inverse->src.l3num = orig->src.l3num;
 155         if (l3proto->invert_tuple(inverse, orig) == 0)
 156                 return false;
 157
 158         inverse->dst.dir = !orig->dst.dir;
 159
 160         inverse->dst.protonum = orig->dst.protonum;
 161         return l4proto->invert_tuple(inverse, orig);
 162 }
 163 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 164
 165 static void
 166 clean_from_lists(struct nf_conn *ct)
 167 {
 168         pr_debug("clean_from_lists(%p)\n", ct);
 169         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 170         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 171
 172         /* Destroy all pending expectations */
 173         nf_ct_remove_expectations(ct);
 174 }
 175
 176 static void
 177 destroy_conntrack(struct nf_conntrack *nfct)
 178 {
 179         struct nf_conn *ct = (struct nf_conn *)nfct;
 180         struct net *net = nf_ct_net(ct);
 181         struct nf_conntrack_l4proto *l4proto;
 182
 183         pr_debug("destroy_conntrack(%p)\n", ct);
 184         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 185         NF_CT_ASSERT(!timer_pending(&ct->timeout));
 186
 187         /* To make sure we don't get any weird locking issues here:
 188          * destroy_conntrack() MUST NOT be called with a write lock
 189          * to nf_conntrack_lock!!! -HW */
 190         rcu_read_lock();
 191         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 192         if (l4proto && l4proto->destroy)
 193                 l4proto->destroy(ct);
 194
 195         rcu_read_unlock();
 196
 197         spin_lock_bh(&nf_conntrack_lock);
 198         /* Expectations will have been removed in clean_from_lists,
 199          * except TFTP can create an expectation on the first packet,
 200          * before connection is in the list, so we need to clean here,
 201          * too. */
 202         nf_ct_remove_expectations(ct);
 203
 204         /* We overload first tuple to link into unconfirmed list. */
 205         if (!nf_ct_is_confirmed(ct)) {
 206                 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 207                 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 208         }
 209
 210         NF_CT_STAT_INC(net, delete);
 211         spin_unlock_bh(&nf_conntrack_lock);
 212
 213         if (ct->master)
 214                 nf_ct_put(ct->master);
 215
 216         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 217         nf_conntrack_free(ct);
 218 }
 219
 220 void nf_ct_delete_from_lists(struct nf_conn *ct)
 221 {
 222         struct net *net = nf_ct_net(ct);
 223
 224         nf_ct_helper_destroy(ct);
 225         spin_lock_bh(&nf_conntrack_lock);
 226         /* Inside lock so preempt is disabled on module removal path.
 227          * Otherwise we can get spurious warnings. */
 228         NF_CT_STAT_INC(net, delete_list);
 229         clean_from_lists(ct);
 230         spin_unlock_bh(&nf_conntrack_lock);
 231 }
 232 EXPORT_SYMBOL_GPL(nf_ct_delete_from_lists);
 233
 234 static void death_by_event(unsigned long ul_conntrack)
 235 {
 236         struct nf_conn *ct = (void *)ul_conntrack;
 237         struct net *net = nf_ct_net(ct);
 238
 239         if (nf_conntrack_event(IPCT_DESTROY, ct) < 0) {
 240                 /* bad luck, let's retry again */
 241                 ct->timeout.expires = jiffies +
 242                         (random32() % net->ct.sysctl_events_retry_timeout);
 243                 add_timer(&ct->timeout);
 244                 return;
 245         }
 246         /* we've got the event delivered, now it's dying */
 247         set_bit(IPS_DYING_BIT, &ct->status);
 248         spin_lock(&nf_conntrack_lock);
 249         hlist_nulls_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 250         spin_unlock(&nf_conntrack_lock);
 251         nf_ct_put(ct);
 252 }
 253
 254 void nf_ct_insert_dying_list(struct nf_conn *ct)
 255 {
 256         struct net *net = nf_ct_net(ct);
 257
 258         /* add this conntrack to the dying list */
 259         spin_lock_bh(&nf_conntrack_lock);
 260         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 261                              &net->ct.dying);
 262         spin_unlock_bh(&nf_conntrack_lock);
 263         /* set a new timer to retry event delivery */
 264         setup_timer(&ct->timeout, death_by_event, (unsigned long)ct);
 265         ct->timeout.expires = jiffies +
 266                 (random32() % net->ct.sysctl_events_retry_timeout);
 267         add_timer(&ct->timeout);
 268 }
 269 EXPORT_SYMBOL_GPL(nf_ct_insert_dying_list);
 270
 271 static void death_by_timeout(unsigned long ul_conntrack)
 272 {
 273         struct nf_conn *ct = (void *)ul_conntrack;
 274
 275         if (!test_bit(IPS_DYING_BIT, &ct->status) &&
 276             unlikely(nf_conntrack_event(IPCT_DESTROY, ct) < 0)) {
 277                 /* destroy event was not delivered */
 278                 nf_ct_delete_from_lists(ct);
 279                 nf_ct_insert_dying_list(ct);
 280                 return;
 281         }
 282         set_bit(IPS_DYING_BIT, &ct->status);
 283         nf_ct_delete_from_lists(ct);
 284         nf_ct_put(ct);
 285 }
 286
 287 /*
 288  * Warning :
 289  * - Caller must take a reference on returned object
 290  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 291  * OR
 292  * - Caller must lock nf_conntrack_lock before calling this function
 293  */
 294 struct nf_conntrack_tuple_hash *
 295 __nf_conntrack_find(struct net *net, const struct nf_conntrack_tuple *tuple)
 296 {
 297         struct nf_conntrack_tuple_hash *h;
 298         struct hlist_nulls_node *n;
 299         unsigned int hash = hash_conntrack(tuple);
 300
 301         /* Disable BHs the entire time since we normally need to disable them
 302          * at least once for the stats anyway.
 303          */
 304         local_bh_disable();
 305 begin:
 306         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
 307                 if (nf_ct_tuple_equal(tuple, &h->tuple)) {
 308                         NF_CT_STAT_INC(net, found);
 309                         local_bh_enable();
 310                         return h;
 311                 }
 312                 NF_CT_STAT_INC(net, searched);
 313         }
 314         /*
 315          * if the nulls value we got at the end of this lookup is
 316          * not the expected one, we must restart lookup.
 317          * We probably met an item that was moved to another chain.
 318          */
 319         if (get_nulls_value(n) != hash)
 320                 goto begin;
 321         local_bh_enable();
 322
 323         return NULL;
 324 }
 325 EXPORT_SYMBOL_GPL(__nf_conntrack_find);
 326
 327 /* Find a connection corresponding to a tuple. */
 328 struct nf_conntrack_tuple_hash *
 329 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_tuple *tuple)
 330 {
 331         struct nf_conntrack_tuple_hash *h;
 332         struct nf_conn *ct;
 333
 334         rcu_read_lock();
 335 begin:
 336         h = __nf_conntrack_find(net, tuple);
 337         if (h) {
 338                 ct = nf_ct_tuplehash_to_ctrack(h);
 339                 if (unlikely(nf_ct_is_dying(ct) ||
 340                              !atomic_inc_not_zero(&ct->ct_general.use)))
 341                         h = NULL;
 342                 else {
 343                         if (unlikely(!nf_ct_tuple_equal(tuple, &h->tuple))) {
 344                                 nf_ct_put(ct);
 345                                 goto begin;
 346                         }
 347                 }
 348         }
 349         rcu_read_unlock();
 350
 351         return h;
 352 }
 353 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 354
 355 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 356                                        unsigned int hash,
 357                                        unsigned int repl_hash)
 358 {
 359         struct net *net = nf_ct_net(ct);
 360
 361         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 362                            &net->ct.hash[hash]);
 363         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 364                            &net->ct.hash[repl_hash]);
 365 }
 366
 367 void nf_conntrack_hash_insert(struct nf_conn *ct)
 368 {
 369         unsigned int hash, repl_hash;
 370
 371         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 372         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 373
 374         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 375 }
 376 EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
 377
 378 /* Confirm a connection given skb; places it in hash table */
 379 int
 380 __nf_conntrack_confirm(struct sk_buff *skb)
 381 {
 382         unsigned int hash, repl_hash;
 383         struct nf_conntrack_tuple_hash *h;
 384         struct nf_conn *ct;
 385         struct nf_conn_help *help;
 386         struct hlist_nulls_node *n;
 387         enum ip_conntrack_info ctinfo;
 388         struct net *net;
 389
 390         ct = nf_ct_get(skb, &ctinfo);
 391         net = nf_ct_net(ct);
 392
 393         /* ipt_REJECT uses nf_conntrack_attach to attach related
 394            ICMP/TCP RST packets in other direction.  Actual packet
 395            which created connection will be IP_CT_NEW or for an
 396            expected connection, IP_CT_RELATED. */
 397         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 398                 return NF_ACCEPT;
 399
 400         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 401         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 402
 403         /* We're not in hash table, and we refuse to set up related
 404            connections for unconfirmed conns.  But packet copies and
 405            REJECT will give spurious warnings here. */
 406         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 407
 408         /* No external references means noone else could have
 409            confirmed us. */
 410         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 411         pr_debug("Confirming conntrack %p\n", ct);
 412
 413         spin_lock_bh(&nf_conntrack_lock);
 414
 415         /* See if there's one in the list already, including reverse:
 416            NAT could have grabbed it without realizing, since we're
 417            not in the hash.  If there is, we lost race. */
 418         hlist_nulls_for_each_entry(h, n, &net->ct.hash[hash], hnnode)
 419                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 420                                       &h->tuple))
 421                         goto out;
 422         hlist_nulls_for_each_entry(h, n, &net->ct.hash[repl_hash], hnnode)
 423                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 424                                       &h->tuple))
 425                         goto out;
 426
 427         /* Remove from unconfirmed list */
 428         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 429
 430         /* Timer relative to confirmation time, not original
 431            setting time, otherwise we'd get timer wrap in
 432            weird delay cases. */
 433         ct->timeout.expires += jiffies;
 434         add_timer(&ct->timeout);
 435         atomic_inc(&ct->ct_general.use);
 436         set_bit(IPS_CONFIRMED_BIT, &ct->status);
 437
 438         /* Since the lookup is lockless, hash insertion must be done after
 439          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 440          * guarantee that no other CPU can find the conntrack before the above
 441          * stores are visible.
 442          */
 443         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 444         NF_CT_STAT_INC(net, insert);
 445         spin_unlock_bh(&nf_conntrack_lock);
 446
 447         help = nfct_help(ct);
 448         if (help && help->helper)
 449                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 450
 451         nf_conntrack_event_cache(master_ct(ct) ?
 452                                  IPCT_RELATED : IPCT_NEW, ct);
 453         return NF_ACCEPT;
 454
 455 out:
 456         NF_CT_STAT_INC(net, insert_failed);
 457         spin_unlock_bh(&nf_conntrack_lock);
 458         return NF_DROP;
 459 }
 460 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 461
 462 /* Returns true if a connection correspondings to the tuple (required
 463    for NAT). */
 464 int
 465 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 466                          const struct nf_conn *ignored_conntrack)
 467 {
 468         struct net *net = nf_ct_net(ignored_conntrack);
 469         struct nf_conntrack_tuple_hash *h;
 470         struct hlist_nulls_node *n;
 471         unsigned int hash = hash_conntrack(tuple);
 472
 473         /* Disable BHs the entire time since we need to disable them at
 474          * least once for the stats anyway.
 475          */
 476         rcu_read_lock_bh();
 477         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
 478                 if (nf_ct_tuplehash_to_ctrack(h) != ignored_conntrack &&
 479                     nf_ct_tuple_equal(tuple, &h->tuple)) {
 480                         NF_CT_STAT_INC(net, found);
 481                         rcu_read_unlock_bh();
 482                         return 1;
 483                 }
 484                 NF_CT_STAT_INC(net, searched);
 485         }
 486         rcu_read_unlock_bh();
 487
 488         return 0;
 489 }
 490 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 491
 492 #define NF_CT_EVICTION_RANGE    8
 493
 494 /* There's a small race here where we may free a just-assured
 495    connection.  Too bad: we're in trouble anyway. */
 496 static noinline int early_drop(struct net *net, unsigned int hash)
 497 {
 498         /* Use oldest entry, which is roughly LRU */
 499         struct nf_conntrack_tuple_hash *h;
 500         struct nf_conn *ct = NULL, *tmp;
 501         struct hlist_nulls_node *n;
 502         unsigned int i, cnt = 0;
 503         int dropped = 0;
 504
 505         rcu_read_lock();
 506         for (i = 0; i < nf_conntrack_htable_size; i++) {
 507                 hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash],
 508                                          hnnode) {
 509                         tmp = nf_ct_tuplehash_to_ctrack(h);
 510                         if (!test_bit(IPS_ASSURED_BIT, &tmp->status))
 511                                 ct = tmp;
 512                         cnt++;
 513                 }
 514
 515                 if (ct != NULL) {
 516                         if (likely(!nf_ct_is_dying(ct) &&
 517                                    atomic_inc_not_zero(&ct->ct_general.use)))
 518                                 break;
 519                         else
 520                                 ct = NULL;
 521                 }
 522
 523                 if (cnt >= NF_CT_EVICTION_RANGE)
 524                         break;
 525
 526                 hash = (hash + 1) % nf_conntrack_htable_size;
 527         }
 528         rcu_read_unlock();
 529
 530         if (!ct)
 531                 return dropped;
 532
 533         if (del_timer(&ct->timeout)) {
 534                 death_by_timeout((unsigned long)ct);
 535                 dropped = 1;
 536                 NF_CT_STAT_INC_ATOMIC(net, early_drop);
 537         }
 538         nf_ct_put(ct);
 539         return dropped;
 540 }
 541
 542 struct nf_conn *nf_conntrack_alloc(struct net *net,
 543                                    const struct nf_conntrack_tuple *orig,
 544                                    const struct nf_conntrack_tuple *repl,
 545                                    gfp_t gfp)
 546 {
 547         struct nf_conn *ct;
 548
 549         if (unlikely(!nf_conntrack_hash_rnd_initted)) {
 550                 get_random_bytes(&nf_conntrack_hash_rnd,
 551                                 sizeof(nf_conntrack_hash_rnd));
 552                 nf_conntrack_hash_rnd_initted = 1;
 553         }
 554
 555         /* We don't want any race condition at early drop stage */
 556         atomic_inc(&net->ct.count);
 557
 558         if (nf_conntrack_max &&
 559             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
 560                 unsigned int hash = hash_conntrack(orig);
 561                 if (!early_drop(net, hash)) {
 562                         atomic_dec(&net->ct.count);
 563                         if (net_ratelimit())
 564                                 printk(KERN_WARNING
 565                                        "nf_conntrack: table full, dropping"
 566                                        " packet.\n");
 567                         return ERR_PTR(-ENOMEM);
 568                 }
 569         }
 570
 571         /*
 572          * Do not use kmem_cache_zalloc(), as this cache uses
 573          * SLAB_DESTROY_BY_RCU.
 574          */
 575         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
 576         if (ct == NULL) {
 577                 pr_debug("nf_conntrack_alloc: Can't alloc conntrack.\n");
 578                 atomic_dec(&net->ct.count);
 579                 return ERR_PTR(-ENOMEM);
 580         }
 581         /*
 582          * Let ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.next
 583          * and ct->tuplehash[IP_CT_DIR_REPLY].hnnode.next unchanged.
 584          */
 585         memset(&ct->tuplehash[IP_CT_DIR_MAX], 0,
 586                sizeof(*ct) - offsetof(struct nf_conn, tuplehash[IP_CT_DIR_MAX]));
 587         spin_lock_init(&ct->lock);
 588         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
 589         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
 590         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
 591         ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev = NULL;
 592         /* Don't set timer yet: wait for confirmation */
 593         setup_timer(&ct->timeout, death_by_timeout, (unsigned long)ct);
 594 #ifdef CONFIG_NET_NS
 595         ct->ct_net = net;
 596 #endif
 597
 598         /*
 599          * changes to lookup keys must be done before setting refcnt to 1
 600          */
 601         smp_wmb();
 602         atomic_set(&ct->ct_general.use, 1);
 603         return ct;
 604 }
 605 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
 606
 607 void nf_conntrack_free(struct nf_conn *ct)
 608 {
 609         struct net *net = nf_ct_net(ct);
 610
 611         nf_ct_ext_destroy(ct);
 612         atomic_dec(&net->ct.count);
 613         nf_ct_ext_free(ct);
 614         kmem_cache_free(nf_conntrack_cachep, ct);
 615 }
 616 EXPORT_SYMBOL_GPL(nf_conntrack_free);
 617
 618 /* Allocate a new conntrack: we return -ENOMEM if classification
 619    failed due to stress.  Otherwise it really is unclassifiable. */
 620 static struct nf_conntrack_tuple_hash *
 621 init_conntrack(struct net *net, struct nf_conn *tmpl,
 622                const struct nf_conntrack_tuple *tuple,
 623                struct nf_conntrack_l3proto *l3proto,
 624                struct nf_conntrack_l4proto *l4proto,
 625                struct sk_buff *skb,
 626                unsigned int dataoff)
 627 {
 628         struct nf_conn *ct;
 629         struct nf_conn_help *help;
 630         struct nf_conntrack_tuple repl_tuple;
 631         struct nf_conntrack_ecache *ecache;
 632         struct nf_conntrack_expect *exp;
 633
 634         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
 635                 pr_debug("Can't invert tuple.\n");
 636                 return NULL;
 637         }
 638
 639         ct = nf_conntrack_alloc(net, tuple, &repl_tuple, GFP_ATOMIC);
 640         if (IS_ERR(ct)) {
 641                 pr_debug("Can't allocate conntrack.\n");
 642                 return (struct nf_conntrack_tuple_hash *)ct;
 643         }
 644
 645         if (!l4proto->new(ct, skb, dataoff)) {
 646                 nf_conntrack_free(ct);
 647                 pr_debug("init conntrack: can't track with proto module\n");
 648                 return NULL;
 649         }
 650
 651         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
 652
 653         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
 654         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
 655                                  ecache ? ecache->expmask : 0,
 656                              GFP_ATOMIC);
 657
 658         spin_lock_bh(&nf_conntrack_lock);
 659         exp = nf_ct_find_expectation(net, tuple);
 660         if (exp) {
 661                 pr_debug("conntrack: expectation arrives ct=%p exp=%p\n",
 662                          ct, exp);
 663                 /* Welcome, Mr. Bond.  We've been expecting you... */
 664                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
 665                 ct->master = exp->master;
 666                 if (exp->helper) {
 667                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
 668                         if (help)
 669                                 rcu_assign_pointer(help->helper, exp->helper);
 670                 }
 671
 672 #ifdef CONFIG_NF_CONNTRACK_MARK
 673                 ct->mark = exp->master->mark;
 674 #endif
 675 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 676                 ct->secmark = exp->master->secmark;
 677 #endif
 678                 nf_conntrack_get(&ct->master->ct_general);
 679                 NF_CT_STAT_INC(net, expect_new);
 680         } else {
 681                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
 682                 NF_CT_STAT_INC(net, new);
 683         }
 684
 685         /* Overload tuple linked list to put us in unconfirmed list. */
 686         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 687                        &net->ct.unconfirmed);
 688
 689         spin_unlock_bh(&nf_conntrack_lock);
 690
 691         if (exp) {
 692                 if (exp->expectfn)
 693                         exp->expectfn(ct, exp);
 694                 nf_ct_expect_put(exp);
 695         }
 696
 697         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
 698 }
 699
 700 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
 701 static inline struct nf_conn *
 702 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
 703                   struct sk_buff *skb,
 704                   unsigned int dataoff,
 705                   u_int16_t l3num,
 706                   u_int8_t protonum,
 707                   struct nf_conntrack_l3proto *l3proto,
 708                   struct nf_conntrack_l4proto *l4proto,
 709                   int *set_reply,
 710                   enum ip_conntrack_info *ctinfo)
 711 {
 712         struct nf_conntrack_tuple tuple;
 713         struct nf_conntrack_tuple_hash *h;
 714         struct nf_conn *ct;
 715
 716         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
 717                              dataoff, l3num, protonum, &tuple, l3proto,
 718                              l4proto)) {
 719                 pr_debug("resolve_normal_ct: Can't get tuple\n");
 720                 return NULL;
 721         }
 722
 723         /* look for tuple match */
 724         h = nf_conntrack_find_get(net, &tuple);
 725         if (!h) {
 726                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
 727                                    skb, dataoff);
 728                 if (!h)
 729                         return NULL;
 730                 if (IS_ERR(h))
 731                         return (void *)h;
 732         }
 733         ct = nf_ct_tuplehash_to_ctrack(h);
 734
 735         /* It exists; we have (non-exclusive) reference. */
 736         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
 737                 *ctinfo = IP_CT_ESTABLISHED + IP_CT_IS_REPLY;
 738                 /* Please set reply bit if this packet OK */
 739                 *set_reply = 1;
 740         } else {
 741                 /* Once we've had two way comms, always ESTABLISHED. */
 742                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
 743                         pr_debug("nf_conntrack_in: normal packet for %p\n", ct);
 744                         *ctinfo = IP_CT_ESTABLISHED;
 745                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
 746                         pr_debug("nf_conntrack_in: related packet for %p\n",
 747                                  ct);
 748                         *ctinfo = IP_CT_RELATED;
 749                 } else {
 750                         pr_debug("nf_conntrack_in: new packet for %p\n", ct);
 751                         *ctinfo = IP_CT_NEW;
 752                 }
 753                 *set_reply = 0;
 754         }
 755         skb->nfct = &ct->ct_general;
 756         skb->nfctinfo = *ctinfo;
 757         return ct;
 758 }
 759
 760 unsigned int
 761 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
 762                 struct sk_buff *skb)
 763 {
 764         struct nf_conn *ct, *tmpl = NULL;
 765         enum ip_conntrack_info ctinfo;
 766         struct nf_conntrack_l3proto *l3proto;
 767         struct nf_conntrack_l4proto *l4proto;
 768         unsigned int dataoff;
 769         u_int8_t protonum;
 770         int set_reply = 0;
 771         int ret;
 772
 773         if (skb->nfct) {
 774                 /* Previously seen (loopback or untracked)?  Ignore. */
 775                 tmpl = (struct nf_conn *)skb->nfct;
 776                 if (!nf_ct_is_template(tmpl)) {
 777                         NF_CT_STAT_INC_ATOMIC(net, ignore);
 778                         return NF_ACCEPT;
 779                 }
 780                 skb->nfct = NULL;
 781         }
 782
 783         /* rcu_read_lock()ed by nf_hook_slow */
 784         l3proto = __nf_ct_l3proto_find(pf);
 785         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
 786                                    &dataoff, &protonum);
 787         if (ret <= 0) {
 788                 pr_debug("not prepared to track yet or error occured\n");
 789                 NF_CT_STAT_INC_ATOMIC(net, error);
 790                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 791                 ret = -ret;
 792                 goto out;
 793         }
 794
 795         l4proto = __nf_ct_l4proto_find(pf, protonum);
 796
 797         /* It may be an special packet, error, unclean...
 798          * inverse of the return code tells to the netfilter
 799          * core what to do with the packet. */
 800         if (l4proto->error != NULL) {
 801                 ret = l4proto->error(net, skb, dataoff, &ctinfo, pf, hooknum);
 802                 if (ret <= 0) {
 803                         NF_CT_STAT_INC_ATOMIC(net, error);
 804                         NF_CT_STAT_INC_ATOMIC(net, invalid);
 805                         ret = -ret;
 806                         goto out;
 807                 }
 808         }
 809
 810         ct = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
 811                                l3proto, l4proto, &set_reply, &ctinfo);
 812         if (!ct) {
 813                 /* Not valid part of a connection */
 814                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 815                 ret = NF_ACCEPT;
 816                 goto out;
 817         }
 818
 819         if (IS_ERR(ct)) {
 820                 /* Too stressed to deal. */
 821                 NF_CT_STAT_INC_ATOMIC(net, drop);
 822                 ret = NF_DROP;
 823                 goto out;
 824         }
 825
 826         NF_CT_ASSERT(skb->nfct);
 827
 828         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum);
 829         if (ret <= 0) {
 830                 /* Invalid: inverse of the return code tells
 831                  * the netfilter core what to do */
 832                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
 833                 nf_conntrack_put(skb->nfct);
 834                 skb->nfct = NULL;
 835                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 836                 if (ret == -NF_DROP)
 837                         NF_CT_STAT_INC_ATOMIC(net, drop);
 838                 ret = -ret;
 839                 goto out;
 840         }
 841
 842         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
 843                 nf_conntrack_event_cache(IPCT_REPLY, ct);
 844 out:
 845         if (tmpl)
 846                 nf_ct_put(tmpl);
 847
 848         return ret;
 849 }
 850 EXPORT_SYMBOL_GPL(nf_conntrack_in);
 851
 852 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
 853                           const struct nf_conntrack_tuple *orig)
 854 {
 855         bool ret;
 856
 857         rcu_read_lock();
 858         ret = nf_ct_invert_tuple(inverse, orig,
 859                                  __nf_ct_l3proto_find(orig->src.l3num),
 860                                  __nf_ct_l4proto_find(orig->src.l3num,
 861                                                       orig->dst.protonum));
 862         rcu_read_unlock();
 863         return ret;
 864 }
 865 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
 866
 867 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
 868    implicitly racy: see __nf_conntrack_confirm */
 869 void nf_conntrack_alter_reply(struct nf_conn *ct,
 870                               const struct nf_conntrack_tuple *newreply)
 871 {
 872         struct nf_conn_help *help = nfct_help(ct);
 873
 874         /* Should be unconfirmed, so not in hash table yet */
 875         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 876
 877         pr_debug("Altering reply tuple of %p to ", ct);
 878         nf_ct_dump_tuple(newreply);
 879
 880         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
 881         if (ct->master || (help && !hlist_empty(&help->expectations)))
 882                 return;
 883
 884         rcu_read_lock();
 885         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
 886         rcu_read_unlock();
 887 }
 888 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
 889
 890 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
 891 void __nf_ct_refresh_acct(struct nf_conn *ct,
 892                           enum ip_conntrack_info ctinfo,
 893                           const struct sk_buff *skb,
 894                           unsigned long extra_jiffies,
 895                           int do_acct)
 896 {
 897         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
 898         NF_CT_ASSERT(skb);
 899
 900         /* Only update if this is not a fixed timeout */
 901         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
 902                 goto acct;
 903
 904         /* If not in hash table, timer will not be active yet */
 905         if (!nf_ct_is_confirmed(ct)) {
 906                 ct->timeout.expires = extra_jiffies;
 907         } else {
 908                 unsigned long newtime = jiffies + extra_jiffies;
 909
 910                 /* Only update the timeout if the new timeout is at least
 911                    HZ jiffies from the old timeout. Need del_timer for race
 912                    avoidance (may already be dying). */
 913                 if (newtime - ct->timeout.expires >= HZ)
 914                         mod_timer_pending(&ct->timeout, newtime);
 915         }
 916
 917 acct:
 918         if (do_acct) {
 919                 struct nf_conn_counter *acct;
 920
 921                 acct = nf_conn_acct_find(ct);
 922                 if (acct) {
 923                         spin_lock_bh(&ct->lock);
 924                         acct[CTINFO2DIR(ctinfo)].packets++;
 925                         acct[CTINFO2DIR(ctinfo)].bytes +=
 926                                 skb->len - skb_network_offset(skb);
 927                         spin_unlock_bh(&ct->lock);
 928                 }
 929         }
 930 }
 931 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
 932
 933 bool __nf_ct_kill_acct(struct nf_conn *ct,
 934                        enum ip_conntrack_info ctinfo,
 935                        const struct sk_buff *skb,
 936                        int do_acct)
 937 {
 938         if (do_acct) {
 939                 struct nf_conn_counter *acct;
 940
 941                 acct = nf_conn_acct_find(ct);
 942                 if (acct) {
 943                         spin_lock_bh(&ct->lock);
 944                         acct[CTINFO2DIR(ctinfo)].packets++;
 945                         acct[CTINFO2DIR(ctinfo)].bytes +=
 946                                 skb->len - skb_network_offset(skb);
 947                         spin_unlock_bh(&ct->lock);
 948                 }
 949         }
 950
 951         if (del_timer(&ct->timeout)) {
 952                 ct->timeout.function((unsigned long)ct);
 953                 return true;
 954         }
 955         return false;
 956 }
 957 EXPORT_SYMBOL_GPL(__nf_ct_kill_acct);
 958
 959 #if defined(CONFIG_NF_CT_NETLINK) || defined(CONFIG_NF_CT_NETLINK_MODULE)
 960
 961 #include <linux/netfilter/nfnetlink.h>
 962 #include <linux/netfilter/nfnetlink_conntrack.h>
 963 #include <linux/mutex.h>
 964
 965 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
 966  * in ip_conntrack_core, since we don't want the protocols to autoload
 967  * or depend on ctnetlink */
 968 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
 969                                const struct nf_conntrack_tuple *tuple)
 970 {
 971         NLA_PUT_BE16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port);
 972         NLA_PUT_BE16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port);
 973         return 0;
 974
 975 nla_put_failure:
 976         return -1;
 977 }
 978 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
 979
 980 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
 981         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
 982         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
 983 };
 984 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
 985
 986 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
 987                                struct nf_conntrack_tuple *t)
 988 {
 989         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
 990                 return -EINVAL;
 991
 992         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
 993         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
 994
 995         return 0;
 996 }
 997 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
 998
 999 int nf_ct_port_nlattr_tuple_size(void)
1000 {
1001         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1002 }
1003 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1004 #endif
1005
1006 /* Used by ipt_REJECT and ip6t_REJECT. */
1007 static void nf_conntrack_attach(struct sk_buff *nskb, struct sk_buff *skb)
1008 {
1009         struct nf_conn *ct;
1010         enum ip_conntrack_info ctinfo;
1011
1012         /* This ICMP is in reverse direction to the packet which caused it */
1013         ct = nf_ct_get(skb, &ctinfo);
1014         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1015                 ctinfo = IP_CT_RELATED + IP_CT_IS_REPLY;
1016         else
1017                 ctinfo = IP_CT_RELATED;
1018
1019         /* Attach to new skbuff, and increment count */
1020         nskb->nfct = &ct->ct_general;
1021         nskb->nfctinfo = ctinfo;
1022         nf_conntrack_get(nskb->nfct);
1023 }
1024
1025 /* Bring out ya dead! */
1026 static struct nf_conn *
1027 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1028                 void *data, unsigned int *bucket)
1029 {
1030         struct nf_conntrack_tuple_hash *h;
1031         struct nf_conn *ct;
1032         struct hlist_nulls_node *n;
1033
1034         spin_lock_bh(&nf_conntrack_lock);
1035         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1036                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[*bucket], hnnode) {
1037                         ct = nf_ct_tuplehash_to_ctrack(h);
1038                         if (iter(ct, data))
1039                                 goto found;
1040                 }
1041         }
1042         hlist_nulls_for_each_entry(h, n, &net->ct.unconfirmed, hnnode) {
1043                 ct = nf_ct_tuplehash_to_ctrack(h);
1044                 if (iter(ct, data))
1045                         set_bit(IPS_DYING_BIT, &ct->status);
1046         }
1047         spin_unlock_bh(&nf_conntrack_lock);
1048         return NULL;
1049 found:
1050         atomic_inc(&ct->ct_general.use);
1051         spin_unlock_bh(&nf_conntrack_lock);
1052         return ct;
1053 }
1054
1055 void nf_ct_iterate_cleanup(struct net *net,
1056                            int (*iter)(struct nf_conn *i, void *data),
1057                            void *data)
1058 {
1059         struct nf_conn *ct;
1060         unsigned int bucket = 0;
1061
1062         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1063                 /* Time to push up daises... */
1064                 if (del_timer(&ct->timeout))
1065                         death_by_timeout((unsigned long)ct);
1066                 /* ... else the timer will get him soon. */
1067
1068                 nf_ct_put(ct);
1069         }
1070 }
1071 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1072
1073 struct __nf_ct_flush_report {
1074         u32 pid;
1075         int report;
1076 };
1077
1078 static int kill_report(struct nf_conn *i, void *data)
1079 {
1080         struct __nf_ct_flush_report *fr = (struct __nf_ct_flush_report *)data;
1081
1082         /* If we fail to deliver the event, death_by_timeout() will retry */
1083         if (nf_conntrack_event_report(IPCT_DESTROY, i,
1084                                       fr->pid, fr->report) < 0)
1085                 return 1;
1086
1087         /* Avoid the delivery of the destroy event in death_by_timeout(). */
1088         set_bit(IPS_DYING_BIT, &i->status);
1089         return 1;
1090 }
1091
1092 static int kill_all(struct nf_conn *i, void *data)
1093 {
1094         return 1;
1095 }
1096
1097 void nf_ct_free_hashtable(void *hash, int vmalloced, unsigned int size)
1098 {
1099         if (vmalloced)
1100                 vfree(hash);
1101         else
1102                 free_pages((unsigned long)hash,
1103                            get_order(sizeof(struct hlist_head) * size));
1104 }
1105 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1106
1107 void nf_conntrack_flush_report(struct net *net, u32 pid, int report)
1108 {
1109         struct __nf_ct_flush_report fr = {
1110                 .pid    = pid,
1111                 .report = report,
1112         };
1113         nf_ct_iterate_cleanup(net, kill_report, &fr);
1114 }
1115 EXPORT_SYMBOL_GPL(nf_conntrack_flush_report);
1116
1117 static void nf_ct_release_dying_list(struct net *net)
1118 {
1119         struct nf_conntrack_tuple_hash *h;
1120         struct nf_conn *ct;
1121         struct hlist_nulls_node *n;
1122
1123         spin_lock_bh(&nf_conntrack_lock);
1124         hlist_nulls_for_each_entry(h, n, &net->ct.dying, hnnode) {
1125                 ct = nf_ct_tuplehash_to_ctrack(h);
1126                 /* never fails to remove them, no listeners at this point */
1127                 nf_ct_kill(ct);
1128         }
1129         spin_unlock_bh(&nf_conntrack_lock);
1130 }
1131
1132 static void nf_conntrack_cleanup_init_net(void)
1133 {
1134         nf_conntrack_helper_fini();
1135         nf_conntrack_proto_fini();
1136         kmem_cache_destroy(nf_conntrack_cachep);
1137 }
1138
1139 static void nf_conntrack_cleanup_net(struct net *net)
1140 {
1141  i_see_dead_people:
1142         nf_ct_iterate_cleanup(net, kill_all, NULL);
1143         nf_ct_release_dying_list(net);
1144         if (atomic_read(&net->ct.count) != 0) {
1145                 schedule();
1146                 goto i_see_dead_people;
1147         }
1148         /* wait until all references to nf_conntrack_untracked are dropped */
1149         while (atomic_read(&nf_conntrack_untracked.ct_general.use) > 1)
1150                 schedule();
1151
1152         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1153                              nf_conntrack_htable_size);
1154         nf_conntrack_ecache_fini(net);
1155         nf_conntrack_acct_fini(net);
1156         nf_conntrack_expect_fini(net);
1157         free_percpu(net->ct.stat);
1158 }
1159
1160 /* Mishearing the voices in his head, our hero wonders how he's
1161    supposed to kill the mall. */
1162 void nf_conntrack_cleanup(struct net *net)
1163 {
1164         if (net_eq(net, &init_net))
1165                 rcu_assign_pointer(ip_ct_attach, NULL);
1166
1167         /* This makes sure all current packets have passed through
1168            netfilter framework.  Roll on, two-stage module
1169            delete... */
1170         synchronize_net();
1171
1172         nf_conntrack_cleanup_net(net);
1173
1174         if (net_eq(net, &init_net)) {
1175                 rcu_assign_pointer(nf_ct_destroy, NULL);
1176                 nf_conntrack_cleanup_init_net();
1177         }
1178 }
1179
1180 void *nf_ct_alloc_hashtable(unsigned int *sizep, int *vmalloced, int nulls)
1181 {
1182         struct hlist_nulls_head *hash;
1183         unsigned int nr_slots, i;
1184         size_t sz;
1185
1186         *vmalloced = 0;
1187
1188         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1189         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1190         sz = nr_slots * sizeof(struct hlist_nulls_head);
1191         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1192                                         get_order(sz));
1193         if (!hash) {
1194                 *vmalloced = 1;
1195                 printk(KERN_WARNING "nf_conntrack: falling back to vmalloc.\n");
1196                 hash = __vmalloc(sz, GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL);
1197         }
1198
1199         if (hash && nulls)
1200                 for (i = 0; i < nr_slots; i++)
1201                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1202
1203         return hash;
1204 }
1205 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1206
1207 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1208 {
1209         int i, bucket, vmalloced, old_vmalloced;
1210         unsigned int hashsize, old_size;
1211         int rnd;
1212         struct hlist_nulls_head *hash, *old_hash;
1213         struct nf_conntrack_tuple_hash *h;
1214
1215         /* On boot, we can set this without any fancy locking. */
1216         if (!nf_conntrack_htable_size)
1217                 return param_set_uint(val, kp);
1218
1219         hashsize = simple_strtoul(val, NULL, 0);
1220         if (!hashsize)
1221                 return -EINVAL;
1222
1223         hash = nf_ct_alloc_hashtable(&hashsize, &vmalloced, 1);
1224         if (!hash)
1225                 return -ENOMEM;
1226
1227         /* We have to rehahs for the new table anyway, so we also can
1228          * use a newrandom seed */
1229         get_random_bytes(&rnd, sizeof(rnd));
1230
1231         /* Lookups in the old hash might happen in parallel, which means we
1232          * might get false negatives during connection lookup. New connections
1233          * created because of a false negative won't make it into the hash
1234          * though since that required taking the lock.
1235          */
1236         spin_lock_bh(&nf_conntrack_lock);
1237         for (i = 0; i < nf_conntrack_htable_size; i++) {
1238                 while (!hlist_nulls_empty(&init_net.ct.hash[i])) {
1239                         h = hlist_nulls_entry(init_net.ct.hash[i].first,
1240                                         struct nf_conntrack_tuple_hash, hnnode);
1241                         hlist_nulls_del_rcu(&h->hnnode);
1242                         bucket = __hash_conntrack(&h->tuple, hashsize, rnd);
1243                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1244                 }
1245         }
1246         old_size = nf_conntrack_htable_size;
1247         old_vmalloced = init_net.ct.hash_vmalloc;
1248         old_hash = init_net.ct.hash;
1249
1250         nf_conntrack_htable_size = hashsize;
1251         init_net.ct.hash_vmalloc = vmalloced;
1252         init_net.ct.hash = hash;
1253         nf_conntrack_hash_rnd = rnd;
1254         spin_unlock_bh(&nf_conntrack_lock);
1255
1256         nf_ct_free_hashtable(old_hash, old_vmalloced, old_size);
1257         return 0;
1258 }
1259 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1260
1261 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1262                   &nf_conntrack_htable_size, 0600);
1263
1264 static int nf_conntrack_init_init_net(void)
1265 {
1266         int max_factor = 8;
1267         int ret;
1268
1269         /* Idea from tcp.c: use 1/16384 of memory.  On i386: 32MB
1270          * machine has 512 buckets. >= 1GB machines have 16384 buckets. */
1271         if (!nf_conntrack_htable_size) {
1272                 nf_conntrack_htable_size
1273                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1274                            / sizeof(struct hlist_head));
1275                 if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1276                         nf_conntrack_htable_size = 16384;
1277                 if (nf_conntrack_htable_size < 32)
1278                         nf_conntrack_htable_size = 32;
1279
1280                 /* Use a max. factor of four by default to get the same max as
1281                  * with the old struct list_heads. When a table size is given
1282                  * we use the old value of 8 to avoid reducing the max.
1283                  * entries. */
1284                 max_factor = 4;
1285         }
1286         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1287
1288         printk("nf_conntrack version %s (%u buckets, %d max)\n",
1289                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1290                nf_conntrack_max);
1291
1292         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
1293                                                 sizeof(struct nf_conn),
1294                                                 0, SLAB_DESTROY_BY_RCU, NULL);
1295         if (!nf_conntrack_cachep) {
1296                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
1297                 ret = -ENOMEM;
1298                 goto err_cache;
1299         }
1300
1301         ret = nf_conntrack_proto_init();
1302         if (ret < 0)
1303                 goto err_proto;
1304
1305         ret = nf_conntrack_helper_init();
1306         if (ret < 0)
1307                 goto err_helper;
1308
1309         return 0;
1310
1311 err_helper:
1312         nf_conntrack_proto_fini();
1313 err_proto:
1314         kmem_cache_destroy(nf_conntrack_cachep);
1315 err_cache:
1316         return ret;
1317 }
1318
1319 /*
1320  * We need to use special "null" values, not used in hash table
1321  */
1322 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1323 #define DYING_NULLS_VAL         ((1<<30)+1)
1324
1325 static int nf_conntrack_init_net(struct net *net)
1326 {
1327         int ret;
1328
1329         atomic_set(&net->ct.count, 0);
1330         INIT_HLIST_NULLS_HEAD(&net->ct.unconfirmed, UNCONFIRMED_NULLS_VAL);
1331         INIT_HLIST_NULLS_HEAD(&net->ct.dying, DYING_NULLS_VAL);
1332         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
1333         if (!net->ct.stat) {
1334                 ret = -ENOMEM;
1335                 goto err_stat;
1336         }
1337         net->ct.hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size,
1338                                              &net->ct.hash_vmalloc, 1);
1339         if (!net->ct.hash) {
1340                 ret = -ENOMEM;
1341                 printk(KERN_ERR "Unable to create nf_conntrack_hash\n");
1342                 goto err_hash;
1343         }
1344         ret = nf_conntrack_expect_init(net);
1345         if (ret < 0)
1346                 goto err_expect;
1347         ret = nf_conntrack_acct_init(net);
1348         if (ret < 0)
1349                 goto err_acct;
1350         ret = nf_conntrack_ecache_init(net);
1351         if (ret < 0)
1352                 goto err_ecache;
1353
1354         /* Set up fake conntrack:
1355             - to never be deleted, not in any hashes */
1356 #ifdef CONFIG_NET_NS
1357         nf_conntrack_untracked.ct_net = &init_net;
1358 #endif
1359         atomic_set(&nf_conntrack_untracked.ct_general.use, 1);
1360         /*  - and look it like as a confirmed connection */
1361         set_bit(IPS_CONFIRMED_BIT, &nf_conntrack_untracked.status);
1362
1363         return 0;
1364
1365 err_ecache:
1366         nf_conntrack_acct_fini(net);
1367 err_acct:
1368         nf_conntrack_expect_fini(net);
1369 err_expect:
1370         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1371                              nf_conntrack_htable_size);
1372 err_hash:
1373         free_percpu(net->ct.stat);
1374 err_stat:
1375         return ret;
1376 }
1377
1378 s16 (*nf_ct_nat_offset)(const struct nf_conn *ct,
1379                         enum ip_conntrack_dir dir,
1380                         u32 seq);
1381 EXPORT_SYMBOL_GPL(nf_ct_nat_offset);
1382
1383 int nf_conntrack_init(struct net *net)
1384 {
1385         int ret;
1386
1387         if (net_eq(net, &init_net)) {
1388                 ret = nf_conntrack_init_init_net();
1389                 if (ret < 0)
1390                         goto out_init_net;
1391         }
1392         ret = nf_conntrack_init_net(net);
1393         if (ret < 0)
1394                 goto out_net;
1395
1396         if (net_eq(net, &init_net)) {
1397                 /* For use by REJECT target */
1398                 rcu_assign_pointer(ip_ct_attach, nf_conntrack_attach);
1399                 rcu_assign_pointer(nf_ct_destroy, destroy_conntrack);
1400
1401                 /* Howto get NAT offsets */
1402                 rcu_assign_pointer(nf_ct_nat_offset, NULL);
1403         }
1404         return 0;
1405
1406 out_net:
1407         if (net_eq(net, &init_net))
1408                 nf_conntrack_cleanup_init_net();
1409 out_init_net:
1410         return ret;
1411 }