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