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/nsproxy.h>
  34 #include <linux/rculist_nulls.h>
  35
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_l3proto.h>
  38 #include <net/netfilter/nf_conntrack_l4proto.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_core.h>
  42 #include <net/netfilter/nf_conntrack_extend.h>
  43 #include <net/netfilter/nf_conntrack_acct.h>
  44 #include <net/netfilter/nf_conntrack_ecache.h>
  45 #include <net/netfilter/nf_nat.h>
  46 #include <net/netfilter/nf_nat_core.h>
  47
  48 #define NF_CONNTRACK_VERSION    "0.5.0"
  49
  50 int (*nfnetlink_parse_nat_setup_hook)(struct nf_conn *ct,
  51                                       enum nf_nat_manip_type manip,
  52                                       const struct nlattr *attr) __read_mostly;
  53 EXPORT_SYMBOL_GPL(nfnetlink_parse_nat_setup_hook);
  54
  55 DEFINE_SPINLOCK(nf_conntrack_lock);
  56 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
  57
  58 unsigned int nf_conntrack_htable_size __read_mostly;
  59 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
  60
  61 unsigned int nf_conntrack_max __read_mostly;
  62 EXPORT_SYMBOL_GPL(nf_conntrack_max);
  63
  64 struct nf_conn nf_conntrack_untracked __read_mostly;
  65 EXPORT_SYMBOL_GPL(nf_conntrack_untracked);
  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 net *net,
  89                                        const struct nf_conntrack_tuple *tuple)
  90 {
  91         return __hash_conntrack(tuple, net->ct.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(net, 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         struct net *net = nf_ct_net(ct);
 370         unsigned int hash, repl_hash;
 371
 372         hash = hash_conntrack(net, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 373         repl_hash = hash_conntrack(net, &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 374
 375         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 376 }
 377 EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
 378
 379 /* Confirm a connection given skb; places it in hash table */
 380 int
 381 __nf_conntrack_confirm(struct sk_buff *skb)
 382 {
 383         unsigned int hash, repl_hash;
 384         struct nf_conntrack_tuple_hash *h;
 385         struct nf_conn *ct;
 386         struct nf_conn_help *help;
 387         struct hlist_nulls_node *n;
 388         enum ip_conntrack_info ctinfo;
 389         struct net *net;
 390
 391         ct = nf_ct_get(skb, &ctinfo);
 392         net = nf_ct_net(ct);
 393
 394         /* ipt_REJECT uses nf_conntrack_attach to attach related
 395            ICMP/TCP RST packets in other direction.  Actual packet
 396            which created connection will be IP_CT_NEW or for an
 397            expected connection, IP_CT_RELATED. */
 398         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 399                 return NF_ACCEPT;
 400
 401         hash = hash_conntrack(net, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 402         repl_hash = hash_conntrack(net, &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 403
 404         /* We're not in hash table, and we refuse to set up related
 405            connections for unconfirmed conns.  But packet copies and
 406            REJECT will give spurious warnings here. */
 407         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 408
 409         /* No external references means noone else could have
 410            confirmed us. */
 411         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 412         pr_debug("Confirming conntrack %p\n", ct);
 413
 414         spin_lock_bh(&nf_conntrack_lock);
 415
 416         /* See if there's one in the list already, including reverse:
 417            NAT could have grabbed it without realizing, since we're
 418            not in the hash.  If there is, we lost race. */
 419         hlist_nulls_for_each_entry(h, n, &net->ct.hash[hash], hnnode)
 420                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 421                                       &h->tuple))
 422                         goto out;
 423         hlist_nulls_for_each_entry(h, n, &net->ct.hash[repl_hash], hnnode)
 424                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 425                                       &h->tuple))
 426                         goto out;
 427
 428         /* Remove from unconfirmed list */
 429         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 430
 431         /* Timer relative to confirmation time, not original
 432            setting time, otherwise we'd get timer wrap in
 433            weird delay cases. */
 434         ct->timeout.expires += jiffies;
 435         add_timer(&ct->timeout);
 436         atomic_inc(&ct->ct_general.use);
 437         set_bit(IPS_CONFIRMED_BIT, &ct->status);
 438
 439         /* Since the lookup is lockless, hash insertion must be done after
 440          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 441          * guarantee that no other CPU can find the conntrack before the above
 442          * stores are visible.
 443          */
 444         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 445         NF_CT_STAT_INC(net, insert);
 446         spin_unlock_bh(&nf_conntrack_lock);
 447
 448         help = nfct_help(ct);
 449         if (help && help->helper)
 450                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 451
 452         nf_conntrack_event_cache(master_ct(ct) ?
 453                                  IPCT_RELATED : IPCT_NEW, ct);
 454         return NF_ACCEPT;
 455
 456 out:
 457         NF_CT_STAT_INC(net, insert_failed);
 458         spin_unlock_bh(&nf_conntrack_lock);
 459         return NF_DROP;
 460 }
 461 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 462
 463 /* Returns true if a connection correspondings to the tuple (required
 464    for NAT). */
 465 int
 466 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 467                          const struct nf_conn *ignored_conntrack)
 468 {
 469         struct net *net = nf_ct_net(ignored_conntrack);
 470         struct nf_conntrack_tuple_hash *h;
 471         struct hlist_nulls_node *n;
 472         unsigned int hash = hash_conntrack(net, tuple);
 473
 474         /* Disable BHs the entire time since we need to disable them at
 475          * least once for the stats anyway.
 476          */
 477         rcu_read_lock_bh();
 478         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
 479                 if (nf_ct_tuplehash_to_ctrack(h) != ignored_conntrack &&
 480                     nf_ct_tuple_equal(tuple, &h->tuple)) {
 481                         NF_CT_STAT_INC(net, found);
 482                         rcu_read_unlock_bh();
 483                         return 1;
 484                 }
 485                 NF_CT_STAT_INC(net, searched);
 486         }
 487         rcu_read_unlock_bh();
 488
 489         return 0;
 490 }
 491 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 492
 493 #define NF_CT_EVICTION_RANGE    8
 494
 495 /* There's a small race here where we may free a just-assured
 496    connection.  Too bad: we're in trouble anyway. */
 497 static noinline int early_drop(struct net *net, unsigned int hash)
 498 {
 499         /* Use oldest entry, which is roughly LRU */
 500         struct nf_conntrack_tuple_hash *h;
 501         struct nf_conn *ct = NULL, *tmp;
 502         struct hlist_nulls_node *n;
 503         unsigned int i, cnt = 0;
 504         int dropped = 0;
 505
 506         rcu_read_lock();
 507         for (i = 0; i < net->ct.htable_size; i++) {
 508                 hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash],
 509                                          hnnode) {
 510                         tmp = nf_ct_tuplehash_to_ctrack(h);
 511                         if (!test_bit(IPS_ASSURED_BIT, &tmp->status))
 512                                 ct = tmp;
 513                         cnt++;
 514                 }
 515
 516                 if (ct != NULL) {
 517                         if (likely(!nf_ct_is_dying(ct) &&
 518                                    atomic_inc_not_zero(&ct->ct_general.use)))
 519                                 break;
 520                         else
 521                                 ct = NULL;
 522                 }
 523
 524                 if (cnt >= NF_CT_EVICTION_RANGE)
 525                         break;
 526
 527                 hash = (hash + 1) % net->ct.htable_size;
 528         }
 529         rcu_read_unlock();
 530
 531         if (!ct)
 532                 return dropped;
 533
 534         if (del_timer(&ct->timeout)) {
 535                 death_by_timeout((unsigned long)ct);
 536                 dropped = 1;
 537                 NF_CT_STAT_INC_ATOMIC(net, early_drop);
 538         }
 539         nf_ct_put(ct);
 540         return dropped;
 541 }
 542
 543 struct nf_conn *nf_conntrack_alloc(struct net *net,
 544                                    const struct nf_conntrack_tuple *orig,
 545                                    const struct nf_conntrack_tuple *repl,
 546                                    gfp_t gfp)
 547 {
 548         struct nf_conn *ct;
 549
 550         if (unlikely(!nf_conntrack_hash_rnd_initted)) {
 551                 get_random_bytes(&nf_conntrack_hash_rnd,
 552                                 sizeof(nf_conntrack_hash_rnd));
 553                 nf_conntrack_hash_rnd_initted = 1;
 554         }
 555
 556         /* We don't want any race condition at early drop stage */
 557         atomic_inc(&net->ct.count);
 558
 559         if (nf_conntrack_max &&
 560             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
 561                 unsigned int hash = hash_conntrack(net, orig);
 562                 if (!early_drop(net, hash)) {
 563                         atomic_dec(&net->ct.count);
 564                         if (net_ratelimit())
 565                                 printk(KERN_WARNING
 566                                        "nf_conntrack: table full, dropping"
 567                                        " packet.\n");
 568                         return ERR_PTR(-ENOMEM);
 569                 }
 570         }
 571
 572         /*
 573          * Do not use kmem_cache_zalloc(), as this cache uses
 574          * SLAB_DESTROY_BY_RCU.
 575          */
 576         ct = kmem_cache_alloc(net->ct.nf_conntrack_cachep, gfp);
 577         if (ct == NULL) {
 578                 pr_debug("nf_conntrack_alloc: Can't alloc conntrack.\n");
 579                 atomic_dec(&net->ct.count);
 580                 return ERR_PTR(-ENOMEM);
 581         }
 582         /*
 583          * Let ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.next
 584          * and ct->tuplehash[IP_CT_DIR_REPLY].hnnode.next unchanged.
 585          */
 586         memset(&ct->tuplehash[IP_CT_DIR_MAX], 0,
 587                sizeof(*ct) - offsetof(struct nf_conn, tuplehash[IP_CT_DIR_MAX]));
 588         spin_lock_init(&ct->lock);
 589         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
 590         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
 591         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
 592         ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev = NULL;
 593         /* Don't set timer yet: wait for confirmation */
 594         setup_timer(&ct->timeout, death_by_timeout, (unsigned long)ct);
 595 #ifdef CONFIG_NET_NS
 596         ct->ct_net = net;
 597 #endif
 598
 599         /*
 600          * changes to lookup keys must be done before setting refcnt to 1
 601          */
 602         smp_wmb();
 603         atomic_set(&ct->ct_general.use, 1);
 604         return ct;
 605 }
 606 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
 607
 608 void nf_conntrack_free(struct nf_conn *ct)
 609 {
 610         struct net *net = nf_ct_net(ct);
 611
 612         nf_ct_ext_destroy(ct);
 613         atomic_dec(&net->ct.count);
 614         nf_ct_ext_free(ct);
 615         kmem_cache_free(net->ct.nf_conntrack_cachep, ct);
 616 }
 617 EXPORT_SYMBOL_GPL(nf_conntrack_free);
 618
 619 /* Allocate a new conntrack: we return -ENOMEM if classification
 620    failed due to stress.  Otherwise it really is unclassifiable. */
 621 static struct nf_conntrack_tuple_hash *
 622 init_conntrack(struct net *net, struct nf_conn *tmpl,
 623                const struct nf_conntrack_tuple *tuple,
 624                struct nf_conntrack_l3proto *l3proto,
 625                struct nf_conntrack_l4proto *l4proto,
 626                struct sk_buff *skb,
 627                unsigned int dataoff)
 628 {
 629         struct nf_conn *ct;
 630         struct nf_conn_help *help;
 631         struct nf_conntrack_tuple repl_tuple;
 632         struct nf_conntrack_ecache *ecache;
 633         struct nf_conntrack_expect *exp;
 634
 635         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
 636                 pr_debug("Can't invert tuple.\n");
 637                 return NULL;
 638         }
 639
 640         ct = nf_conntrack_alloc(net, tuple, &repl_tuple, GFP_ATOMIC);
 641         if (IS_ERR(ct)) {
 642                 pr_debug("Can't allocate conntrack.\n");
 643                 return (struct nf_conntrack_tuple_hash *)ct;
 644         }
 645
 646         if (!l4proto->new(ct, skb, dataoff)) {
 647                 nf_conntrack_free(ct);
 648                 pr_debug("init conntrack: can't track with proto module\n");
 649                 return NULL;
 650         }
 651
 652         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
 653
 654         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
 655         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
 656                                  ecache ? ecache->expmask : 0,
 657                              GFP_ATOMIC);
 658
 659         spin_lock_bh(&nf_conntrack_lock);
 660         exp = nf_ct_find_expectation(net, tuple);
 661         if (exp) {
 662                 pr_debug("conntrack: expectation arrives ct=%p exp=%p\n",
 663                          ct, exp);
 664                 /* Welcome, Mr. Bond.  We've been expecting you... */
 665                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
 666                 ct->master = exp->master;
 667                 if (exp->helper) {
 668                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
 669                         if (help)
 670                                 rcu_assign_pointer(help->helper, exp->helper);
 671                 }
 672
 673 #ifdef CONFIG_NF_CONNTRACK_MARK
 674                 ct->mark = exp->master->mark;
 675 #endif
 676 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 677                 ct->secmark = exp->master->secmark;
 678 #endif
 679                 nf_conntrack_get(&ct->master->ct_general);
 680                 NF_CT_STAT_INC(net, expect_new);
 681         } else {
 682                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
 683                 NF_CT_STAT_INC(net, new);
 684         }
 685
 686         /* Overload tuple linked list to put us in unconfirmed list. */
 687         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 688                        &net->ct.unconfirmed);
 689
 690         spin_unlock_bh(&nf_conntrack_lock);
 691
 692         if (exp) {
 693                 if (exp->expectfn)
 694                         exp->expectfn(ct, exp);
 695                 nf_ct_expect_put(exp);
 696         }
 697
 698         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
 699 }
 700
 701 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
 702 static inline struct nf_conn *
 703 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
 704                   struct sk_buff *skb,
 705                   unsigned int dataoff,
 706                   u_int16_t l3num,
 707                   u_int8_t protonum,
 708                   struct nf_conntrack_l3proto *l3proto,
 709                   struct nf_conntrack_l4proto *l4proto,
 710                   int *set_reply,
 711                   enum ip_conntrack_info *ctinfo)
 712 {
 713         struct nf_conntrack_tuple tuple;
 714         struct nf_conntrack_tuple_hash *h;
 715         struct nf_conn *ct;
 716
 717         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
 718                              dataoff, l3num, protonum, &tuple, l3proto,
 719                              l4proto)) {
 720                 pr_debug("resolve_normal_ct: Can't get tuple\n");
 721                 return NULL;
 722         }
 723
 724         /* look for tuple match */
 725         h = nf_conntrack_find_get(net, &tuple);
 726         if (!h) {
 727                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
 728                                    skb, dataoff);
 729                 if (!h)
 730                         return NULL;
 731                 if (IS_ERR(h))
 732                         return (void *)h;
 733         }
 734         ct = nf_ct_tuplehash_to_ctrack(h);
 735
 736         /* It exists; we have (non-exclusive) reference. */
 737         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
 738                 *ctinfo = IP_CT_ESTABLISHED + IP_CT_IS_REPLY;
 739                 /* Please set reply bit if this packet OK */
 740                 *set_reply = 1;
 741         } else {
 742                 /* Once we've had two way comms, always ESTABLISHED. */
 743                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
 744                         pr_debug("nf_conntrack_in: normal packet for %p\n", ct);
 745                         *ctinfo = IP_CT_ESTABLISHED;
 746                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
 747                         pr_debug("nf_conntrack_in: related packet for %p\n",
 748                                  ct);
 749                         *ctinfo = IP_CT_RELATED;
 750                 } else {
 751                         pr_debug("nf_conntrack_in: new packet for %p\n", ct);
 752                         *ctinfo = IP_CT_NEW;
 753                 }
 754                 *set_reply = 0;
 755         }
 756         skb->nfct = &ct->ct_general;
 757         skb->nfctinfo = *ctinfo;
 758         return ct;
 759 }
 760
 761 unsigned int
 762 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
 763                 struct sk_buff *skb)
 764 {
 765         struct nf_conn *ct, *tmpl = NULL;
 766         enum ip_conntrack_info ctinfo;
 767         struct nf_conntrack_l3proto *l3proto;
 768         struct nf_conntrack_l4proto *l4proto;
 769         unsigned int dataoff;
 770         u_int8_t protonum;
 771         int set_reply = 0;
 772         int ret;
 773
 774         if (skb->nfct) {
 775                 /* Previously seen (loopback or untracked)?  Ignore. */
 776                 tmpl = (struct nf_conn *)skb->nfct;
 777                 if (!nf_ct_is_template(tmpl)) {
 778                         NF_CT_STAT_INC_ATOMIC(net, ignore);
 779                         return NF_ACCEPT;
 780                 }
 781                 skb->nfct = NULL;
 782         }
 783
 784         /* rcu_read_lock()ed by nf_hook_slow */
 785         l3proto = __nf_ct_l3proto_find(pf);
 786         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
 787                                    &dataoff, &protonum);
 788         if (ret <= 0) {
 789                 pr_debug("not prepared to track yet or error occured\n");
 790                 NF_CT_STAT_INC_ATOMIC(net, error);
 791                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 792                 ret = -ret;
 793                 goto out;
 794         }
 795
 796         l4proto = __nf_ct_l4proto_find(pf, protonum);
 797
 798         /* It may be an special packet, error, unclean...
 799          * inverse of the return code tells to the netfilter
 800          * core what to do with the packet. */
 801         if (l4proto->error != NULL) {
 802                 ret = l4proto->error(net, tmpl, skb, dataoff, &ctinfo,
 803                                      pf, hooknum);
 804                 if (ret <= 0) {
 805                         NF_CT_STAT_INC_ATOMIC(net, error);
 806                         NF_CT_STAT_INC_ATOMIC(net, invalid);
 807                         ret = -ret;
 808                         goto out;
 809                 }
 810         }
 811
 812         ct = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
 813                                l3proto, l4proto, &set_reply, &ctinfo);
 814         if (!ct) {
 815                 /* Not valid part of a connection */
 816                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 817                 ret = NF_ACCEPT;
 818                 goto out;
 819         }
 820
 821         if (IS_ERR(ct)) {
 822                 /* Too stressed to deal. */
 823                 NF_CT_STAT_INC_ATOMIC(net, drop);
 824                 ret = NF_DROP;
 825                 goto out;
 826         }
 827
 828         NF_CT_ASSERT(skb->nfct);
 829
 830         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum);
 831         if (ret <= 0) {
 832                 /* Invalid: inverse of the return code tells
 833                  * the netfilter core what to do */
 834                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
 835                 nf_conntrack_put(skb->nfct);
 836                 skb->nfct = NULL;
 837                 NF_CT_STAT_INC_ATOMIC(net, invalid);
 838                 if (ret == -NF_DROP)
 839                         NF_CT_STAT_INC_ATOMIC(net, drop);
 840                 ret = -ret;
 841                 goto out;
 842         }
 843
 844         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
 845                 nf_conntrack_event_cache(IPCT_REPLY, ct);
 846 out:
 847         if (tmpl)
 848                 nf_ct_put(tmpl);
 849
 850         return ret;
 851 }
 852 EXPORT_SYMBOL_GPL(nf_conntrack_in);
 853
 854 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
 855                           const struct nf_conntrack_tuple *orig)
 856 {
 857         bool ret;
 858
 859         rcu_read_lock();
 860         ret = nf_ct_invert_tuple(inverse, orig,
 861                                  __nf_ct_l3proto_find(orig->src.l3num),
 862                                  __nf_ct_l4proto_find(orig->src.l3num,
 863                                                       orig->dst.protonum));
 864         rcu_read_unlock();
 865         return ret;
 866 }
 867 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
 868
 869 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
 870    implicitly racy: see __nf_conntrack_confirm */
 871 void nf_conntrack_alter_reply(struct nf_conn *ct,
 872                               const struct nf_conntrack_tuple *newreply)
 873 {
 874         struct nf_conn_help *help = nfct_help(ct);
 875
 876         /* Should be unconfirmed, so not in hash table yet */
 877         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 878
 879         pr_debug("Altering reply tuple of %p to ", ct);
 880         nf_ct_dump_tuple(newreply);
 881
 882         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
 883         if (ct->master || (help && !hlist_empty(&help->expectations)))
 884                 return;
 885
 886         rcu_read_lock();
 887         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
 888         rcu_read_unlock();
 889 }
 890 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
 891
 892 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
 893 void __nf_ct_refresh_acct(struct nf_conn *ct,
 894                           enum ip_conntrack_info ctinfo,
 895                           const struct sk_buff *skb,
 896                           unsigned long extra_jiffies,
 897                           int do_acct)
 898 {
 899         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
 900         NF_CT_ASSERT(skb);
 901
 902         /* Only update if this is not a fixed timeout */
 903         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
 904                 goto acct;
 905
 906         /* If not in hash table, timer will not be active yet */
 907         if (!nf_ct_is_confirmed(ct)) {
 908                 ct->timeout.expires = extra_jiffies;
 909         } else {
 910                 unsigned long newtime = jiffies + extra_jiffies;
 911
 912                 /* Only update the timeout if the new timeout is at least
 913                    HZ jiffies from the old timeout. Need del_timer for race
 914                    avoidance (may already be dying). */
 915                 if (newtime - ct->timeout.expires >= HZ)
 916                         mod_timer_pending(&ct->timeout, newtime);
 917         }
 918
 919 acct:
 920         if (do_acct) {
 921                 struct nf_conn_counter *acct;
 922
 923                 acct = nf_conn_acct_find(ct);
 924                 if (acct) {
 925                         spin_lock_bh(&ct->lock);
 926                         acct[CTINFO2DIR(ctinfo)].packets++;
 927                         acct[CTINFO2DIR(ctinfo)].bytes +=
 928                                 skb->len - skb_network_offset(skb);
 929                         spin_unlock_bh(&ct->lock);
 930                 }
 931         }
 932 }
 933 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
 934
 935 bool __nf_ct_kill_acct(struct nf_conn *ct,
 936                        enum ip_conntrack_info ctinfo,
 937                        const struct sk_buff *skb,
 938                        int do_acct)
 939 {
 940         if (do_acct) {
 941                 struct nf_conn_counter *acct;
 942
 943                 acct = nf_conn_acct_find(ct);
 944                 if (acct) {
 945                         spin_lock_bh(&ct->lock);
 946                         acct[CTINFO2DIR(ctinfo)].packets++;
 947                         acct[CTINFO2DIR(ctinfo)].bytes +=
 948                                 skb->len - skb_network_offset(skb);
 949                         spin_unlock_bh(&ct->lock);
 950                 }
 951         }
 952
 953         if (del_timer(&ct->timeout)) {
 954                 ct->timeout.function((unsigned long)ct);
 955                 return true;
 956         }
 957         return false;
 958 }
 959 EXPORT_SYMBOL_GPL(__nf_ct_kill_acct);
 960
 961 #if defined(CONFIG_NF_CT_NETLINK) || defined(CONFIG_NF_CT_NETLINK_MODULE)
 962
 963 #include <linux/netfilter/nfnetlink.h>
 964 #include <linux/netfilter/nfnetlink_conntrack.h>
 965 #include <linux/mutex.h>
 966
 967 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
 968  * in ip_conntrack_core, since we don't want the protocols to autoload
 969  * or depend on ctnetlink */
 970 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
 971                                const struct nf_conntrack_tuple *tuple)
 972 {
 973         NLA_PUT_BE16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port);
 974         NLA_PUT_BE16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port);
 975         return 0;
 976
 977 nla_put_failure:
 978         return -1;
 979 }
 980 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
 981
 982 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
 983         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
 984         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
 985 };
 986 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
 987
 988 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
 989                                struct nf_conntrack_tuple *t)
 990 {
 991         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
 992                 return -EINVAL;
 993
 994         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
 995         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
 996
 997         return 0;
 998 }
 999 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1000
1001 int nf_ct_port_nlattr_tuple_size(void)
1002 {
1003         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1004 }
1005 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1006 #endif
1007
1008 /* Used by ipt_REJECT and ip6t_REJECT. */
1009 static void nf_conntrack_attach(struct sk_buff *nskb, struct sk_buff *skb)
1010 {
1011         struct nf_conn *ct;
1012         enum ip_conntrack_info ctinfo;
1013
1014         /* This ICMP is in reverse direction to the packet which caused it */
1015         ct = nf_ct_get(skb, &ctinfo);
1016         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1017                 ctinfo = IP_CT_RELATED + IP_CT_IS_REPLY;
1018         else
1019                 ctinfo = IP_CT_RELATED;
1020
1021         /* Attach to new skbuff, and increment count */
1022         nskb->nfct = &ct->ct_general;
1023         nskb->nfctinfo = ctinfo;
1024         nf_conntrack_get(nskb->nfct);
1025 }
1026
1027 /* Bring out ya dead! */
1028 static struct nf_conn *
1029 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1030                 void *data, unsigned int *bucket)
1031 {
1032         struct nf_conntrack_tuple_hash *h;
1033         struct nf_conn *ct;
1034         struct hlist_nulls_node *n;
1035
1036         spin_lock_bh(&nf_conntrack_lock);
1037         for (; *bucket < net->ct.htable_size; (*bucket)++) {
1038                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[*bucket], hnnode) {
1039                         ct = nf_ct_tuplehash_to_ctrack(h);
1040                         if (iter(ct, data))
1041                                 goto found;
1042                 }
1043         }
1044         hlist_nulls_for_each_entry(h, n, &net->ct.unconfirmed, hnnode) {
1045                 ct = nf_ct_tuplehash_to_ctrack(h);
1046                 if (iter(ct, data))
1047                         set_bit(IPS_DYING_BIT, &ct->status);
1048         }
1049         spin_unlock_bh(&nf_conntrack_lock);
1050         return NULL;
1051 found:
1052         atomic_inc(&ct->ct_general.use);
1053         spin_unlock_bh(&nf_conntrack_lock);
1054         return ct;
1055 }
1056
1057 void nf_ct_iterate_cleanup(struct net *net,
1058                            int (*iter)(struct nf_conn *i, void *data),
1059                            void *data)
1060 {
1061         struct nf_conn *ct;
1062         unsigned int bucket = 0;
1063
1064         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1065                 /* Time to push up daises... */
1066                 if (del_timer(&ct->timeout))
1067                         death_by_timeout((unsigned long)ct);
1068                 /* ... else the timer will get him soon. */
1069
1070                 nf_ct_put(ct);
1071         }
1072 }
1073 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1074
1075 struct __nf_ct_flush_report {
1076         u32 pid;
1077         int report;
1078 };
1079
1080 static int kill_report(struct nf_conn *i, void *data)
1081 {
1082         struct __nf_ct_flush_report *fr = (struct __nf_ct_flush_report *)data;
1083
1084         /* If we fail to deliver the event, death_by_timeout() will retry */
1085         if (nf_conntrack_event_report(IPCT_DESTROY, i,
1086                                       fr->pid, fr->report) < 0)
1087                 return 1;
1088
1089         /* Avoid the delivery of the destroy event in death_by_timeout(). */
1090         set_bit(IPS_DYING_BIT, &i->status);
1091         return 1;
1092 }
1093
1094 static int kill_all(struct nf_conn *i, void *data)
1095 {
1096         return 1;
1097 }
1098
1099 void nf_ct_free_hashtable(void *hash, int vmalloced, unsigned int size)
1100 {
1101         if (vmalloced)
1102                 vfree(hash);
1103         else
1104                 free_pages((unsigned long)hash,
1105                            get_order(sizeof(struct hlist_head) * size));
1106 }
1107 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1108
1109 void nf_conntrack_flush_report(struct net *net, u32 pid, int report)
1110 {
1111         struct __nf_ct_flush_report fr = {
1112                 .pid    = pid,
1113                 .report = report,
1114         };
1115         nf_ct_iterate_cleanup(net, kill_report, &fr);
1116 }
1117 EXPORT_SYMBOL_GPL(nf_conntrack_flush_report);
1118
1119 static void nf_ct_release_dying_list(struct net *net)
1120 {
1121         struct nf_conntrack_tuple_hash *h;
1122         struct nf_conn *ct;
1123         struct hlist_nulls_node *n;
1124
1125         spin_lock_bh(&nf_conntrack_lock);
1126         hlist_nulls_for_each_entry(h, n, &net->ct.dying, hnnode) {
1127                 ct = nf_ct_tuplehash_to_ctrack(h);
1128                 /* never fails to remove them, no listeners at this point */
1129                 nf_ct_kill(ct);
1130         }
1131         spin_unlock_bh(&nf_conntrack_lock);
1132 }
1133
1134 static void nf_conntrack_cleanup_init_net(void)
1135 {
1136         /* wait until all references to nf_conntrack_untracked are dropped */
1137         while (atomic_read(&nf_conntrack_untracked.ct_general.use) > 1)
1138                 schedule();
1139
1140         nf_conntrack_helper_fini();
1141         nf_conntrack_proto_fini();
1142 }
1143
1144 static void nf_conntrack_cleanup_net(struct net *net)
1145 {
1146  i_see_dead_people:
1147         nf_ct_iterate_cleanup(net, kill_all, NULL);
1148         nf_ct_release_dying_list(net);
1149         if (atomic_read(&net->ct.count) != 0) {
1150                 schedule();
1151                 goto i_see_dead_people;
1152         }
1153
1154         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1155                              net->ct.htable_size);
1156         nf_conntrack_ecache_fini(net);
1157         nf_conntrack_acct_fini(net);
1158         nf_conntrack_expect_fini(net);
1159         kmem_cache_destroy(net->ct.nf_conntrack_cachep);
1160         kfree(net->ct.slabname);
1161         free_percpu(net->ct.stat);
1162 }
1163
1164 /* Mishearing the voices in his head, our hero wonders how he's
1165    supposed to kill the mall. */
1166 void nf_conntrack_cleanup(struct net *net)
1167 {
1168         if (net_eq(net, &init_net))
1169                 rcu_assign_pointer(ip_ct_attach, NULL);
1170
1171         /* This makes sure all current packets have passed through
1172            netfilter framework.  Roll on, two-stage module
1173            delete... */
1174         synchronize_net();
1175
1176         nf_conntrack_cleanup_net(net);
1177
1178         if (net_eq(net, &init_net)) {
1179                 rcu_assign_pointer(nf_ct_destroy, NULL);
1180                 nf_conntrack_cleanup_init_net();
1181         }
1182 }
1183
1184 void *nf_ct_alloc_hashtable(unsigned int *sizep, int *vmalloced, int nulls)
1185 {
1186         struct hlist_nulls_head *hash;
1187         unsigned int nr_slots, i;
1188         size_t sz;
1189
1190         *vmalloced = 0;
1191
1192         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1193         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1194         sz = nr_slots * sizeof(struct hlist_nulls_head);
1195         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1196                                         get_order(sz));
1197         if (!hash) {
1198                 *vmalloced = 1;
1199                 printk(KERN_WARNING "nf_conntrack: falling back to vmalloc.\n");
1200                 hash = __vmalloc(sz, GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL);
1201         }
1202
1203         if (hash && nulls)
1204                 for (i = 0; i < nr_slots; i++)
1205                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1206
1207         return hash;
1208 }
1209 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1210
1211 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1212 {
1213         int i, bucket, vmalloced, old_vmalloced;
1214         unsigned int hashsize, old_size;
1215         struct hlist_nulls_head *hash, *old_hash;
1216         struct nf_conntrack_tuple_hash *h;
1217
1218         if (current->nsproxy->net_ns != &init_net)
1219                 return -EOPNOTSUPP;
1220
1221         /* On boot, we can set this without any fancy locking. */
1222         if (!nf_conntrack_htable_size)
1223                 return param_set_uint(val, kp);
1224
1225         hashsize = simple_strtoul(val, NULL, 0);
1226         if (!hashsize)
1227                 return -EINVAL;
1228
1229         hash = nf_ct_alloc_hashtable(&hashsize, &vmalloced, 1);
1230         if (!hash)
1231                 return -ENOMEM;
1232
1233         /* Lookups in the old hash might happen in parallel, which means we
1234          * might get false negatives during connection lookup. New connections
1235          * created because of a false negative won't make it into the hash
1236          * though since that required taking the lock.
1237          */
1238         spin_lock_bh(&nf_conntrack_lock);
1239         for (i = 0; i < init_net.ct.htable_size; i++) {
1240                 while (!hlist_nulls_empty(&init_net.ct.hash[i])) {
1241                         h = hlist_nulls_entry(init_net.ct.hash[i].first,
1242                                         struct nf_conntrack_tuple_hash, hnnode);
1243                         hlist_nulls_del_rcu(&h->hnnode);
1244                         bucket = __hash_conntrack(&h->tuple, hashsize,
1245                                                   nf_conntrack_hash_rnd);
1246                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1247                 }
1248         }
1249         old_size = init_net.ct.htable_size;
1250         old_vmalloced = init_net.ct.hash_vmalloc;
1251         old_hash = init_net.ct.hash;
1252
1253         init_net.ct.htable_size = nf_conntrack_htable_size = hashsize;
1254         init_net.ct.hash_vmalloc = vmalloced;
1255         init_net.ct.hash = hash;
1256         spin_unlock_bh(&nf_conntrack_lock);
1257
1258         nf_ct_free_hashtable(old_hash, old_vmalloced, old_size);
1259         return 0;
1260 }
1261 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1262
1263 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1264                   &nf_conntrack_htable_size, 0600);
1265
1266 static int nf_conntrack_init_init_net(void)
1267 {
1268         int max_factor = 8;
1269         int ret;
1270
1271         /* Idea from tcp.c: use 1/16384 of memory.  On i386: 32MB
1272          * machine has 512 buckets. >= 1GB machines have 16384 buckets. */
1273         if (!nf_conntrack_htable_size) {
1274                 nf_conntrack_htable_size
1275                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1276                            / sizeof(struct hlist_head));
1277                 if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1278                         nf_conntrack_htable_size = 16384;
1279                 if (nf_conntrack_htable_size < 32)
1280                         nf_conntrack_htable_size = 32;
1281
1282                 /* Use a max. factor of four by default to get the same max as
1283                  * with the old struct list_heads. When a table size is given
1284                  * we use the old value of 8 to avoid reducing the max.
1285                  * entries. */
1286                 max_factor = 4;
1287         }
1288         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1289
1290         printk("nf_conntrack version %s (%u buckets, %d max)\n",
1291                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1292                nf_conntrack_max);
1293
1294         ret = nf_conntrack_proto_init();
1295         if (ret < 0)
1296                 goto err_proto;
1297
1298         ret = nf_conntrack_helper_init();
1299         if (ret < 0)
1300                 goto err_helper;
1301
1302         /* Set up fake conntrack: to never be deleted, not in any hashes */
1303 #ifdef CONFIG_NET_NS
1304         nf_conntrack_untracked.ct_net = &init_net;
1305 #endif
1306         atomic_set(&nf_conntrack_untracked.ct_general.use, 1);
1307         /*  - and look it like as a confirmed connection */
1308         set_bit(IPS_CONFIRMED_BIT, &nf_conntrack_untracked.status);
1309
1310         return 0;
1311
1312 err_helper:
1313         nf_conntrack_proto_fini();
1314 err_proto:
1315         return ret;
1316 }
1317
1318 /*
1319  * We need to use special "null" values, not used in hash table
1320  */
1321 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1322 #define DYING_NULLS_VAL         ((1<<30)+1)
1323
1324 static int nf_conntrack_init_net(struct net *net)
1325 {
1326         int ret;
1327
1328         atomic_set(&net->ct.count, 0);
1329         INIT_HLIST_NULLS_HEAD(&net->ct.unconfirmed, UNCONFIRMED_NULLS_VAL);
1330         INIT_HLIST_NULLS_HEAD(&net->ct.dying, DYING_NULLS_VAL);
1331         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
1332         if (!net->ct.stat) {
1333                 ret = -ENOMEM;
1334                 goto err_stat;
1335         }
1336
1337         net->ct.slabname = kasprintf(GFP_KERNEL, "nf_conntrack_%p", net);
1338         if (!net->ct.slabname) {
1339                 ret = -ENOMEM;
1340                 goto err_slabname;
1341         }
1342
1343         net->ct.nf_conntrack_cachep = kmem_cache_create(net->ct.slabname,
1344                                                         sizeof(struct nf_conn), 0,
1345                                                         SLAB_DESTROY_BY_RCU, NULL);
1346         if (!net->ct.nf_conntrack_cachep) {
1347                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
1348                 ret = -ENOMEM;
1349                 goto err_cache;
1350         }
1351
1352         net->ct.htable_size = nf_conntrack_htable_size;
1353         net->ct.hash = nf_ct_alloc_hashtable(&net->ct.htable_size,
1354                                              &net->ct.hash_vmalloc, 1);
1355         if (!net->ct.hash) {
1356                 ret = -ENOMEM;
1357                 printk(KERN_ERR "Unable to create nf_conntrack_hash\n");
1358                 goto err_hash;
1359         }
1360         ret = nf_conntrack_expect_init(net);
1361         if (ret < 0)
1362                 goto err_expect;
1363         ret = nf_conntrack_acct_init(net);
1364         if (ret < 0)
1365                 goto err_acct;
1366         ret = nf_conntrack_ecache_init(net);
1367         if (ret < 0)
1368                 goto err_ecache;
1369
1370         return 0;
1371
1372 err_ecache:
1373         nf_conntrack_acct_fini(net);
1374 err_acct:
1375         nf_conntrack_expect_fini(net);
1376 err_expect:
1377         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1378                              net->ct.htable_size);
1379 err_hash:
1380         kmem_cache_destroy(net->ct.nf_conntrack_cachep);
1381 err_cache:
1382         kfree(net->ct.slabname);
1383 err_slabname:
1384         free_percpu(net->ct.stat);
1385 err_stat:
1386         return ret;
1387 }
1388
1389 s16 (*nf_ct_nat_offset)(const struct nf_conn *ct,
1390                         enum ip_conntrack_dir dir,
1391                         u32 seq);
1392 EXPORT_SYMBOL_GPL(nf_ct_nat_offset);
1393
1394 int nf_conntrack_init(struct net *net)
1395 {
1396         int ret;
1397
1398         if (net_eq(net, &init_net)) {
1399                 ret = nf_conntrack_init_init_net();
1400                 if (ret < 0)
1401                         goto out_init_net;
1402         }
1403         ret = nf_conntrack_init_net(net);
1404         if (ret < 0)
1405                 goto out_net;
1406
1407         if (net_eq(net, &init_net)) {
1408                 /* For use by REJECT target */
1409                 rcu_assign_pointer(ip_ct_attach, nf_conntrack_attach);
1410                 rcu_assign_pointer(nf_ct_destroy, destroy_conntrack);
1411
1412                 /* Howto get NAT offsets */
1413                 rcu_assign_pointer(nf_ct_nat_offset, NULL);
1414         }
1415         return 0;
1416
1417 out_net:
1418         if (net_eq(net, &init_net))
1419                 nf_conntrack_cleanup_init_net();
1420 out_init_net:
1421         return ret;
1422 }