err.no Git - linux-2.6/blob - net/netfilter/nf_conntrack_core.c

   1 /* Connection state tracking for netfilter.  This is separated from,
   2    but required by, the NAT layer; it can also be used by an iptables
   3    extension. */
   4
   5 /* (C) 1999-2001 Paul `Rusty' Russell
   6  * (C) 2002-2006 Netfilter Core Team <coreteam@netfilter.org>
   7  * (C) 2003,2004 USAGI/WIDE Project <http://www.linux-ipv6.org>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * 23 Apr 2001: Harald Welte <laforge@gnumonks.org>
  14  *      - new API and handling of conntrack/nat helpers
  15  *      - now capable of multiple expectations for one master
  16  * 16 Jul 2002: Harald Welte <laforge@gnumonks.org>
  17  *      - add usage/reference counts to ip_conntrack_expect
  18  *      - export ip_conntrack[_expect]_{find_get,put} functions
  19  * 16 Dec 2003: Yasuyuki Kozakai @USAGI <yasuyuki.kozakai@toshiba.co.jp>
  20  *      - generalize L3 protocol denendent part.
  21  * 23 Mar 2004: Yasuyuki Kozakai @USAGI <yasuyuki.kozakai@toshiba.co.jp>
  22  *      - add support various size of conntrack structures.
  23  * 26 Jan 2006: Harald Welte <laforge@netfilter.org>
  24  *      - restructure nf_conn (introduce nf_conn_help)
  25  *      - redesign 'features' how they were originally intended
  26  * 26 Feb 2006: Pablo Neira Ayuso <pablo@eurodev.net>
  27  *      - add support for L3 protocol module load on demand.
  28  *
  29  * Derived from net/ipv4/netfilter/ip_conntrack_core.c
  30  */
  31
  32 #include <linux/types.h>
  33 #include <linux/netfilter.h>
  34 #include <linux/module.h>
  35 #include <linux/skbuff.h>
  36 #include <linux/proc_fs.h>
  37 #include <linux/vmalloc.h>
  38 #include <linux/stddef.h>
  39 #include <linux/slab.h>
  40 #include <linux/random.h>
  41 #include <linux/jhash.h>
  42 #include <linux/err.h>
  43 #include <linux/percpu.h>
  44 #include <linux/moduleparam.h>
  45 #include <linux/notifier.h>
  46 #include <linux/kernel.h>
  47 #include <linux/netdevice.h>
  48 #include <linux/socket.h>
  49 #include <linux/mm.h>
  50
  51 #include <net/netfilter/nf_conntrack.h>
  52 #include <net/netfilter/nf_conntrack_l3proto.h>
  53 #include <net/netfilter/nf_conntrack_l4proto.h>
  54 #include <net/netfilter/nf_conntrack_expect.h>
  55 #include <net/netfilter/nf_conntrack_helper.h>
  56 #include <net/netfilter/nf_conntrack_core.h>
  57
  58 #define NF_CONNTRACK_VERSION    "0.5.0"
  59
  60 #if 0
  61 #define DEBUGP printk
  62 #else
  63 #define DEBUGP(format, args...)
  64 #endif
  65
  66 DEFINE_RWLOCK(nf_conntrack_lock);
  67 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
  68
  69 /* nf_conntrack_standalone needs this */
  70 atomic_t nf_conntrack_count = ATOMIC_INIT(0);
  71 EXPORT_SYMBOL_GPL(nf_conntrack_count);
  72
  73 void (*nf_conntrack_destroyed)(struct nf_conn *conntrack);
  74 EXPORT_SYMBOL_GPL(nf_conntrack_destroyed);
  75
  76 unsigned int nf_conntrack_htable_size __read_mostly;
  77 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
  78
  79 int nf_conntrack_max __read_mostly;
  80 EXPORT_SYMBOL_GPL(nf_conntrack_max);
  81
  82 struct list_head *nf_conntrack_hash __read_mostly;
  83 EXPORT_SYMBOL_GPL(nf_conntrack_hash);
  84
  85 struct nf_conn nf_conntrack_untracked __read_mostly;
  86 EXPORT_SYMBOL_GPL(nf_conntrack_untracked);
  87
  88 unsigned int nf_ct_log_invalid __read_mostly;
  89 LIST_HEAD(unconfirmed);
  90 static int nf_conntrack_vmalloc __read_mostly;
  91
  92 static unsigned int nf_conntrack_next_id;
  93
  94 DEFINE_PER_CPU(struct ip_conntrack_stat, nf_conntrack_stat);
  95 EXPORT_PER_CPU_SYMBOL(nf_conntrack_stat);
  96
  97 /*
  98  * This scheme offers various size of "struct nf_conn" dependent on
  99  * features(helper, nat, ...)
 100  */
 101
 102 #define NF_CT_FEATURES_NAMELEN  256
 103 static struct {
 104         /* name of slab cache. printed in /proc/slabinfo */
 105         char *name;
 106
 107         /* size of slab cache */
 108         size_t size;
 109
 110         /* slab cache pointer */
 111         struct kmem_cache *cachep;
 112
 113         /* allocated slab cache + modules which uses this slab cache */
 114         int use;
 115
 116 } nf_ct_cache[NF_CT_F_NUM];
 117
 118 /* protect members of nf_ct_cache except of "use" */
 119 DEFINE_RWLOCK(nf_ct_cache_lock);
 120
 121 /* This avoids calling kmem_cache_create() with same name simultaneously */
 122 static DEFINE_MUTEX(nf_ct_cache_mutex);
 123
 124 static int nf_conntrack_hash_rnd_initted;
 125 static unsigned int nf_conntrack_hash_rnd;
 126
 127 static u_int32_t __hash_conntrack(const struct nf_conntrack_tuple *tuple,
 128                                   unsigned int size, unsigned int rnd)
 129 {
 130         unsigned int a, b;
 131
 132         a = jhash2(tuple->src.u3.all, ARRAY_SIZE(tuple->src.u3.all),
 133                    (tuple->src.l3num << 16) | tuple->dst.protonum);
 134         b = jhash2(tuple->dst.u3.all, ARRAY_SIZE(tuple->dst.u3.all),
 135                    (tuple->src.u.all << 16) | tuple->dst.u.all);
 136
 137         return jhash_2words(a, b, rnd) % size;
 138 }
 139
 140 static inline u_int32_t hash_conntrack(const struct nf_conntrack_tuple *tuple)
 141 {
 142         return __hash_conntrack(tuple, nf_conntrack_htable_size,
 143                                 nf_conntrack_hash_rnd);
 144 }
 145
 146 int nf_conntrack_register_cache(u_int32_t features, const char *name,
 147                                 size_t size)
 148 {
 149         int ret = 0;
 150         char *cache_name;
 151         struct kmem_cache *cachep;
 152
 153         DEBUGP("nf_conntrack_register_cache: features=0x%x, name=%s, size=%d\n",
 154                features, name, size);
 155
 156         if (features < NF_CT_F_BASIC || features >= NF_CT_F_NUM) {
 157                 DEBUGP("nf_conntrack_register_cache: invalid features.: 0x%x\n",
 158                         features);
 159                 return -EINVAL;
 160         }
 161
 162         mutex_lock(&nf_ct_cache_mutex);
 163
 164         write_lock_bh(&nf_ct_cache_lock);
 165         /* e.g: multiple helpers are loaded */
 166         if (nf_ct_cache[features].use > 0) {
 167                 DEBUGP("nf_conntrack_register_cache: already resisterd.\n");
 168                 if ((!strncmp(nf_ct_cache[features].name, name,
 169                               NF_CT_FEATURES_NAMELEN))
 170                     && nf_ct_cache[features].size == size) {
 171                         DEBUGP("nf_conntrack_register_cache: reusing.\n");
 172                         nf_ct_cache[features].use++;
 173                         ret = 0;
 174                 } else
 175                         ret = -EBUSY;
 176
 177                 write_unlock_bh(&nf_ct_cache_lock);
 178                 mutex_unlock(&nf_ct_cache_mutex);
 179                 return ret;
 180         }
 181         write_unlock_bh(&nf_ct_cache_lock);
 182
 183         /*
 184          * The memory space for name of slab cache must be alive until
 185          * cache is destroyed.
 186          */
 187         cache_name = kmalloc(sizeof(char)*NF_CT_FEATURES_NAMELEN, GFP_ATOMIC);
 188         if (cache_name == NULL) {
 189                 DEBUGP("nf_conntrack_register_cache: can't alloc cache_name\n");
 190                 ret = -ENOMEM;
 191                 goto out_up_mutex;
 192         }
 193
 194         if (strlcpy(cache_name, name, NF_CT_FEATURES_NAMELEN)
 195                                                 >= NF_CT_FEATURES_NAMELEN) {
 196                 printk("nf_conntrack_register_cache: name too long\n");
 197                 ret = -EINVAL;
 198                 goto out_free_name;
 199         }
 200
 201         cachep = kmem_cache_create(cache_name, size, 0, 0,
 202                                    NULL, NULL);
 203         if (!cachep) {
 204                 printk("nf_conntrack_register_cache: Can't create slab cache "
 205                        "for the features = 0x%x\n", features);
 206                 ret = -ENOMEM;
 207                 goto out_free_name;
 208         }
 209
 210         write_lock_bh(&nf_ct_cache_lock);
 211         nf_ct_cache[features].use = 1;
 212         nf_ct_cache[features].size = size;
 213         nf_ct_cache[features].cachep = cachep;
 214         nf_ct_cache[features].name = cache_name;
 215         write_unlock_bh(&nf_ct_cache_lock);
 216
 217         goto out_up_mutex;
 218
 219 out_free_name:
 220         kfree(cache_name);
 221 out_up_mutex:
 222         mutex_unlock(&nf_ct_cache_mutex);
 223         return ret;
 224 }
 225 EXPORT_SYMBOL_GPL(nf_conntrack_register_cache);
 226
 227 /* FIXME: In the current, only nf_conntrack_cleanup() can call this function. */
 228 void nf_conntrack_unregister_cache(u_int32_t features)
 229 {
 230         struct kmem_cache *cachep;
 231         char *name;
 232
 233         /*
 234          * This assures that kmem_cache_create() isn't called before destroying
 235          * slab cache.
 236          */
 237         DEBUGP("nf_conntrack_unregister_cache: 0x%04x\n", features);
 238         mutex_lock(&nf_ct_cache_mutex);
 239
 240         write_lock_bh(&nf_ct_cache_lock);
 241         if (--nf_ct_cache[features].use > 0) {
 242                 write_unlock_bh(&nf_ct_cache_lock);
 243                 mutex_unlock(&nf_ct_cache_mutex);
 244                 return;
 245         }
 246         cachep = nf_ct_cache[features].cachep;
 247         name = nf_ct_cache[features].name;
 248         nf_ct_cache[features].cachep = NULL;
 249         nf_ct_cache[features].name = NULL;
 250         nf_ct_cache[features].size = 0;
 251         write_unlock_bh(&nf_ct_cache_lock);
 252
 253         synchronize_net();
 254
 255         kmem_cache_destroy(cachep);
 256         kfree(name);
 257
 258         mutex_unlock(&nf_ct_cache_mutex);
 259 }
 260 EXPORT_SYMBOL_GPL(nf_conntrack_unregister_cache);
 261
 262 int
 263 nf_ct_get_tuple(const struct sk_buff *skb,
 264                 unsigned int nhoff,
 265                 unsigned int dataoff,
 266                 u_int16_t l3num,
 267                 u_int8_t protonum,
 268                 struct nf_conntrack_tuple *tuple,
 269                 const struct nf_conntrack_l3proto *l3proto,
 270                 const struct nf_conntrack_l4proto *l4proto)
 271 {
 272         NF_CT_TUPLE_U_BLANK(tuple);
 273
 274         tuple->src.l3num = l3num;
 275         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 276                 return 0;
 277
 278         tuple->dst.protonum = protonum;
 279         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 280
 281         return l4proto->pkt_to_tuple(skb, dataoff, tuple);
 282 }
 283 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 284
 285 int
 286 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 287                    const struct nf_conntrack_tuple *orig,
 288                    const struct nf_conntrack_l3proto *l3proto,
 289                    const struct nf_conntrack_l4proto *l4proto)
 290 {
 291         NF_CT_TUPLE_U_BLANK(inverse);
 292
 293         inverse->src.l3num = orig->src.l3num;
 294         if (l3proto->invert_tuple(inverse, orig) == 0)
 295                 return 0;
 296
 297         inverse->dst.dir = !orig->dst.dir;
 298
 299         inverse->dst.protonum = orig->dst.protonum;
 300         return l4proto->invert_tuple(inverse, orig);
 301 }
 302 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 303
 304 static void
 305 clean_from_lists(struct nf_conn *ct)
 306 {
 307         DEBUGP("clean_from_lists(%p)\n", ct);
 308         list_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list);
 309         list_del(&ct->tuplehash[IP_CT_DIR_REPLY].list);
 310
 311         /* Destroy all pending expectations */
 312         nf_ct_remove_expectations(ct);
 313 }
 314
 315 static void
 316 destroy_conntrack(struct nf_conntrack *nfct)
 317 {
 318         struct nf_conn *ct = (struct nf_conn *)nfct;
 319         struct nf_conn_help *help = nfct_help(ct);
 320         struct nf_conntrack_l3proto *l3proto;
 321         struct nf_conntrack_l4proto *l4proto;
 322         typeof(nf_conntrack_destroyed) destroyed;
 323
 324         DEBUGP("destroy_conntrack(%p)\n", ct);
 325         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 326         NF_CT_ASSERT(!timer_pending(&ct->timeout));
 327
 328         nf_conntrack_event(IPCT_DESTROY, ct);
 329         set_bit(IPS_DYING_BIT, &ct->status);
 330
 331         if (help && help->helper && help->helper->destroy)
 332                 help->helper->destroy(ct);
 333
 334         /* To make sure we don't get any weird locking issues here:
 335          * destroy_conntrack() MUST NOT be called with a write lock
 336          * to nf_conntrack_lock!!! -HW */
 337         rcu_read_lock();
 338         l3proto = __nf_ct_l3proto_find(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.l3num);
 339         if (l3proto && l3proto->destroy)
 340                 l3proto->destroy(ct);
 341
 342         l4proto = __nf_ct_l4proto_find(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.l3num,
 343                                        ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.protonum);
 344         if (l4proto && l4proto->destroy)
 345                 l4proto->destroy(ct);
 346
 347         destroyed = rcu_dereference(nf_conntrack_destroyed);
 348         if (destroyed)
 349                 destroyed(ct);
 350
 351         rcu_read_unlock();
 352
 353         write_lock_bh(&nf_conntrack_lock);
 354         /* Expectations will have been removed in clean_from_lists,
 355          * except TFTP can create an expectation on the first packet,
 356          * before connection is in the list, so we need to clean here,
 357          * too. */
 358         nf_ct_remove_expectations(ct);
 359
 360         /* We overload first tuple to link into unconfirmed list. */
 361         if (!nf_ct_is_confirmed(ct)) {
 362                 BUG_ON(list_empty(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list));
 363                 list_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list);
 364         }
 365
 366         NF_CT_STAT_INC(delete);
 367         write_unlock_bh(&nf_conntrack_lock);
 368
 369         if (ct->master)
 370                 nf_ct_put(ct->master);
 371
 372         DEBUGP("destroy_conntrack: returning ct=%p to slab\n", ct);
 373         nf_conntrack_free(ct);
 374 }
 375
 376 static void death_by_timeout(unsigned long ul_conntrack)
 377 {
 378         struct nf_conn *ct = (void *)ul_conntrack;
 379
 380         write_lock_bh(&nf_conntrack_lock);
 381         /* Inside lock so preempt is disabled on module removal path.
 382          * Otherwise we can get spurious warnings. */
 383         NF_CT_STAT_INC(delete_list);
 384         clean_from_lists(ct);
 385         write_unlock_bh(&nf_conntrack_lock);
 386         nf_ct_put(ct);
 387 }
 388
 389 struct nf_conntrack_tuple_hash *
 390 __nf_conntrack_find(const struct nf_conntrack_tuple *tuple,
 391                     const struct nf_conn *ignored_conntrack)
 392 {
 393         struct nf_conntrack_tuple_hash *h;
 394         unsigned int hash = hash_conntrack(tuple);
 395
 396         list_for_each_entry(h, &nf_conntrack_hash[hash], list) {
 397                 if (nf_ct_tuplehash_to_ctrack(h) != ignored_conntrack &&
 398                     nf_ct_tuple_equal(tuple, &h->tuple)) {
 399                         NF_CT_STAT_INC(found);
 400                         return h;
 401                 }
 402                 NF_CT_STAT_INC(searched);
 403         }
 404
 405         return NULL;
 406 }
 407 EXPORT_SYMBOL_GPL(__nf_conntrack_find);
 408
 409 /* Find a connection corresponding to a tuple. */
 410 struct nf_conntrack_tuple_hash *
 411 nf_conntrack_find_get(const struct nf_conntrack_tuple *tuple,
 412                       const struct nf_conn *ignored_conntrack)
 413 {
 414         struct nf_conntrack_tuple_hash *h;
 415
 416         read_lock_bh(&nf_conntrack_lock);
 417         h = __nf_conntrack_find(tuple, ignored_conntrack);
 418         if (h)
 419                 atomic_inc(&nf_ct_tuplehash_to_ctrack(h)->ct_general.use);
 420         read_unlock_bh(&nf_conntrack_lock);
 421
 422         return h;
 423 }
 424 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 425
 426 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 427                                        unsigned int hash,
 428                                        unsigned int repl_hash)
 429 {
 430         ct->id = ++nf_conntrack_next_id;
 431         list_add(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list,
 432                  &nf_conntrack_hash[hash]);
 433         list_add(&ct->tuplehash[IP_CT_DIR_REPLY].list,
 434                  &nf_conntrack_hash[repl_hash]);
 435 }
 436
 437 void nf_conntrack_hash_insert(struct nf_conn *ct)
 438 {
 439         unsigned int hash, repl_hash;
 440
 441         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 442         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 443
 444         write_lock_bh(&nf_conntrack_lock);
 445         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 446         write_unlock_bh(&nf_conntrack_lock);
 447 }
 448 EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
 449
 450 /* Confirm a connection given skb; places it in hash table */
 451 int
 452 __nf_conntrack_confirm(struct sk_buff **pskb)
 453 {
 454         unsigned int hash, repl_hash;
 455         struct nf_conntrack_tuple_hash *h;
 456         struct nf_conn *ct;
 457         struct nf_conn_help *help;
 458         enum ip_conntrack_info ctinfo;
 459
 460         ct = nf_ct_get(*pskb, &ctinfo);
 461
 462         /* ipt_REJECT uses nf_conntrack_attach to attach related
 463            ICMP/TCP RST packets in other direction.  Actual packet
 464            which created connection will be IP_CT_NEW or for an
 465            expected connection, IP_CT_RELATED. */
 466         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 467                 return NF_ACCEPT;
 468
 469         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 470         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 471
 472         /* We're not in hash table, and we refuse to set up related
 473            connections for unconfirmed conns.  But packet copies and
 474            REJECT will give spurious warnings here. */
 475         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 476
 477         /* No external references means noone else could have
 478            confirmed us. */
 479         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 480         DEBUGP("Confirming conntrack %p\n", ct);
 481
 482         write_lock_bh(&nf_conntrack_lock);
 483
 484         /* See if there's one in the list already, including reverse:
 485            NAT could have grabbed it without realizing, since we're
 486            not in the hash.  If there is, we lost race. */
 487         list_for_each_entry(h, &nf_conntrack_hash[hash], list)
 488                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 489                                       &h->tuple))
 490                         goto out;
 491         list_for_each_entry(h, &nf_conntrack_hash[repl_hash], list)
 492                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 493                                       &h->tuple))
 494                         goto out;
 495
 496         /* Remove from unconfirmed list */
 497         list_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list);
 498
 499         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 500         /* Timer relative to confirmation time, not original
 501            setting time, otherwise we'd get timer wrap in
 502            weird delay cases. */
 503         ct->timeout.expires += jiffies;
 504         add_timer(&ct->timeout);
 505         atomic_inc(&ct->ct_general.use);
 506         set_bit(IPS_CONFIRMED_BIT, &ct->status);
 507         NF_CT_STAT_INC(insert);
 508         write_unlock_bh(&nf_conntrack_lock);
 509         help = nfct_help(ct);
 510         if (help && help->helper)
 511                 nf_conntrack_event_cache(IPCT_HELPER, *pskb);
 512 #ifdef CONFIG_NF_NAT_NEEDED
 513         if (test_bit(IPS_SRC_NAT_DONE_BIT, &ct->status) ||
 514             test_bit(IPS_DST_NAT_DONE_BIT, &ct->status))
 515                 nf_conntrack_event_cache(IPCT_NATINFO, *pskb);
 516 #endif
 517         nf_conntrack_event_cache(master_ct(ct) ?
 518                                  IPCT_RELATED : IPCT_NEW, *pskb);
 519         return NF_ACCEPT;
 520
 521 out:
 522         NF_CT_STAT_INC(insert_failed);
 523         write_unlock_bh(&nf_conntrack_lock);
 524         return NF_DROP;
 525 }
 526 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 527
 528 /* Returns true if a connection correspondings to the tuple (required
 529    for NAT). */
 530 int
 531 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 532                          const struct nf_conn *ignored_conntrack)
 533 {
 534         struct nf_conntrack_tuple_hash *h;
 535
 536         read_lock_bh(&nf_conntrack_lock);
 537         h = __nf_conntrack_find(tuple, ignored_conntrack);
 538         read_unlock_bh(&nf_conntrack_lock);
 539
 540         return h != NULL;
 541 }
 542 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 543
 544 /* There's a small race here where we may free a just-assured
 545    connection.  Too bad: we're in trouble anyway. */
 546 static int early_drop(struct list_head *chain)
 547 {
 548         /* Traverse backwards: gives us oldest, which is roughly LRU */
 549         struct nf_conntrack_tuple_hash *h;
 550         struct nf_conn *ct = NULL, *tmp;
 551         int dropped = 0;
 552
 553         read_lock_bh(&nf_conntrack_lock);
 554         list_for_each_entry_reverse(h, chain, list) {
 555                 tmp = nf_ct_tuplehash_to_ctrack(h);
 556                 if (!test_bit(IPS_ASSURED_BIT, &tmp->status)) {
 557                         ct = tmp;
 558                         atomic_inc(&ct->ct_general.use);
 559                         break;
 560                 }
 561         }
 562         read_unlock_bh(&nf_conntrack_lock);
 563
 564         if (!ct)
 565                 return dropped;
 566
 567         if (del_timer(&ct->timeout)) {
 568                 death_by_timeout((unsigned long)ct);
 569                 dropped = 1;
 570                 NF_CT_STAT_INC_ATOMIC(early_drop);
 571         }
 572         nf_ct_put(ct);
 573         return dropped;
 574 }
 575
 576 static struct nf_conn *
 577 __nf_conntrack_alloc(const struct nf_conntrack_tuple *orig,
 578                      const struct nf_conntrack_tuple *repl,
 579                      const struct nf_conntrack_l3proto *l3proto,
 580                      u_int32_t features)
 581 {
 582         struct nf_conn *conntrack = NULL;
 583         struct nf_conntrack_helper *helper;
 584
 585         if (unlikely(!nf_conntrack_hash_rnd_initted)) {
 586                 get_random_bytes(&nf_conntrack_hash_rnd, 4);
 587                 nf_conntrack_hash_rnd_initted = 1;
 588         }
 589
 590         /* We don't want any race condition at early drop stage */
 591         atomic_inc(&nf_conntrack_count);
 592
 593         if (nf_conntrack_max
 594             && atomic_read(&nf_conntrack_count) > nf_conntrack_max) {
 595                 unsigned int hash = hash_conntrack(orig);
 596                 /* Try dropping from this hash chain. */
 597                 if (!early_drop(&nf_conntrack_hash[hash])) {
 598                         atomic_dec(&nf_conntrack_count);
 599                         if (net_ratelimit())
 600                                 printk(KERN_WARNING
 601                                        "nf_conntrack: table full, dropping"
 602                                        " packet.\n");
 603                         return ERR_PTR(-ENOMEM);
 604                 }
 605         }
 606
 607         /*  find features needed by this conntrack. */
 608         features |= l3proto->get_features(orig);
 609
 610         /* FIXME: protect helper list per RCU */
 611         read_lock_bh(&nf_conntrack_lock);
 612         helper = __nf_ct_helper_find(repl);
 613         /* NAT might want to assign a helper later */
 614         if (helper || features & NF_CT_F_NAT)
 615                 features |= NF_CT_F_HELP;
 616         read_unlock_bh(&nf_conntrack_lock);
 617
 618         DEBUGP("nf_conntrack_alloc: features=0x%x\n", features);
 619
 620         read_lock_bh(&nf_ct_cache_lock);
 621
 622         if (unlikely(!nf_ct_cache[features].use)) {
 623                 DEBUGP("nf_conntrack_alloc: not supported features = 0x%x\n",
 624                         features);
 625                 goto out;
 626         }
 627
 628         conntrack = kmem_cache_alloc(nf_ct_cache[features].cachep, GFP_ATOMIC);
 629         if (conntrack == NULL) {
 630                 DEBUGP("nf_conntrack_alloc: Can't alloc conntrack from cache\n");
 631                 goto out;
 632         }
 633
 634         memset(conntrack, 0, nf_ct_cache[features].size);
 635         conntrack->features = features;
 636         atomic_set(&conntrack->ct_general.use, 1);
 637         conntrack->ct_general.destroy = destroy_conntrack;
 638         conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
 639         conntrack->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
 640         /* Don't set timer yet: wait for confirmation */
 641         init_timer(&conntrack->timeout);
 642         conntrack->timeout.data = (unsigned long)conntrack;
 643         conntrack->timeout.function = death_by_timeout;
 644         read_unlock_bh(&nf_ct_cache_lock);
 645
 646         return conntrack;
 647 out:
 648         read_unlock_bh(&nf_ct_cache_lock);
 649         atomic_dec(&nf_conntrack_count);
 650         return conntrack;
 651 }
 652
 653 struct nf_conn *nf_conntrack_alloc(const struct nf_conntrack_tuple *orig,
 654                                    const struct nf_conntrack_tuple *repl)
 655 {
 656         struct nf_conntrack_l3proto *l3proto;
 657         struct nf_conn *ct;
 658
 659         rcu_read_lock();
 660         l3proto = __nf_ct_l3proto_find(orig->src.l3num);
 661         ct = __nf_conntrack_alloc(orig, repl, l3proto, 0);
 662         rcu_read_unlock();
 663
 664         return ct;
 665 }
 666 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
 667
 668 void nf_conntrack_free(struct nf_conn *conntrack)
 669 {
 670         u_int32_t features = conntrack->features;
 671         NF_CT_ASSERT(features >= NF_CT_F_BASIC && features < NF_CT_F_NUM);
 672         DEBUGP("nf_conntrack_free: features = 0x%x, conntrack=%p\n", features,
 673                conntrack);
 674         kmem_cache_free(nf_ct_cache[features].cachep, conntrack);
 675         atomic_dec(&nf_conntrack_count);
 676 }
 677 EXPORT_SYMBOL_GPL(nf_conntrack_free);
 678
 679 /* Allocate a new conntrack: we return -ENOMEM if classification
 680    failed due to stress.  Otherwise it really is unclassifiable. */
 681 static struct nf_conntrack_tuple_hash *
 682 init_conntrack(const struct nf_conntrack_tuple *tuple,
 683                struct nf_conntrack_l3proto *l3proto,
 684                struct nf_conntrack_l4proto *l4proto,
 685                struct sk_buff *skb,
 686                unsigned int dataoff)
 687 {
 688         struct nf_conn *conntrack;
 689         struct nf_conntrack_tuple repl_tuple;
 690         struct nf_conntrack_expect *exp;
 691         u_int32_t features = 0;
 692
 693         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
 694                 DEBUGP("Can't invert tuple.\n");
 695                 return NULL;
 696         }
 697
 698         read_lock_bh(&nf_conntrack_lock);
 699         exp = __nf_conntrack_expect_find(tuple);
 700         if (exp && exp->helper)
 701                 features = NF_CT_F_HELP;
 702         read_unlock_bh(&nf_conntrack_lock);
 703
 704         conntrack = __nf_conntrack_alloc(tuple, &repl_tuple, l3proto, features);
 705         if (conntrack == NULL || IS_ERR(conntrack)) {
 706                 DEBUGP("Can't allocate conntrack.\n");
 707                 return (struct nf_conntrack_tuple_hash *)conntrack;
 708         }
 709
 710         if (!l4proto->new(conntrack, skb, dataoff)) {
 711                 nf_conntrack_free(conntrack);
 712                 DEBUGP("init conntrack: can't track with proto module\n");
 713                 return NULL;
 714         }
 715
 716         write_lock_bh(&nf_conntrack_lock);
 717         exp = find_expectation(tuple);
 718
 719         if (exp) {
 720                 DEBUGP("conntrack: expectation arrives ct=%p exp=%p\n",
 721                         conntrack, exp);
 722                 /* Welcome, Mr. Bond.  We've been expecting you... */
 723                 __set_bit(IPS_EXPECTED_BIT, &conntrack->status);
 724                 conntrack->master = exp->master;
 725                 if (exp->helper)
 726                         nfct_help(conntrack)->helper = exp->helper;
 727 #ifdef CONFIG_NF_CONNTRACK_MARK
 728                 conntrack->mark = exp->master->mark;
 729 #endif
 730 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 731                 conntrack->secmark = exp->master->secmark;
 732 #endif
 733                 nf_conntrack_get(&conntrack->master->ct_general);
 734                 NF_CT_STAT_INC(expect_new);
 735         } else {
 736                 struct nf_conn_help *help = nfct_help(conntrack);
 737
 738                 if (help)
 739                         help->helper = __nf_ct_helper_find(&repl_tuple);
 740                 NF_CT_STAT_INC(new);
 741         }
 742
 743         /* Overload tuple linked list to put us in unconfirmed list. */
 744         list_add(&conntrack->tuplehash[IP_CT_DIR_ORIGINAL].list, &unconfirmed);
 745
 746         write_unlock_bh(&nf_conntrack_lock);
 747
 748         if (exp) {
 749                 if (exp->expectfn)
 750                         exp->expectfn(conntrack, exp);
 751                 nf_conntrack_expect_put(exp);
 752         }
 753
 754         return &conntrack->tuplehash[IP_CT_DIR_ORIGINAL];
 755 }
 756
 757 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
 758 static inline struct nf_conn *
 759 resolve_normal_ct(struct sk_buff *skb,
 760                   unsigned int dataoff,
 761                   u_int16_t l3num,
 762                   u_int8_t protonum,
 763                   struct nf_conntrack_l3proto *l3proto,
 764                   struct nf_conntrack_l4proto *l4proto,
 765                   int *set_reply,
 766                   enum ip_conntrack_info *ctinfo)
 767 {
 768         struct nf_conntrack_tuple tuple;
 769         struct nf_conntrack_tuple_hash *h;
 770         struct nf_conn *ct;
 771
 772         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
 773                              dataoff, l3num, protonum, &tuple, l3proto,
 774                              l4proto)) {
 775                 DEBUGP("resolve_normal_ct: Can't get tuple\n");
 776                 return NULL;
 777         }
 778
 779         /* look for tuple match */
 780         h = nf_conntrack_find_get(&tuple, NULL);
 781         if (!h) {
 782                 h = init_conntrack(&tuple, l3proto, l4proto, skb, dataoff);
 783                 if (!h)
 784                         return NULL;
 785                 if (IS_ERR(h))
 786                         return (void *)h;
 787         }
 788         ct = nf_ct_tuplehash_to_ctrack(h);
 789
 790         /* It exists; we have (non-exclusive) reference. */
 791         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
 792                 *ctinfo = IP_CT_ESTABLISHED + IP_CT_IS_REPLY;
 793                 /* Please set reply bit if this packet OK */
 794                 *set_reply = 1;
 795         } else {
 796                 /* Once we've had two way comms, always ESTABLISHED. */
 797                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
 798                         DEBUGP("nf_conntrack_in: normal packet for %p\n", ct);
 799                         *ctinfo = IP_CT_ESTABLISHED;
 800                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
 801                         DEBUGP("nf_conntrack_in: related packet for %p\n", ct);
 802                         *ctinfo = IP_CT_RELATED;
 803                 } else {
 804                         DEBUGP("nf_conntrack_in: new packet for %p\n", ct);
 805                         *ctinfo = IP_CT_NEW;
 806                 }
 807                 *set_reply = 0;
 808         }
 809         skb->nfct = &ct->ct_general;
 810         skb->nfctinfo = *ctinfo;
 811         return ct;
 812 }
 813
 814 unsigned int
 815 nf_conntrack_in(int pf, unsigned int hooknum, struct sk_buff **pskb)
 816 {
 817         struct nf_conn *ct;
 818         enum ip_conntrack_info ctinfo;
 819         struct nf_conntrack_l3proto *l3proto;
 820         struct nf_conntrack_l4proto *l4proto;
 821         unsigned int dataoff;
 822         u_int8_t protonum;
 823         int set_reply = 0;
 824         int ret;
 825
 826         /* Previously seen (loopback or untracked)?  Ignore. */
 827         if ((*pskb)->nfct) {
 828                 NF_CT_STAT_INC_ATOMIC(ignore);
 829                 return NF_ACCEPT;
 830         }
 831
 832         /* rcu_read_lock()ed by nf_hook_slow */
 833         l3proto = __nf_ct_l3proto_find((u_int16_t)pf);
 834
 835         if ((ret = l3proto->prepare(pskb, hooknum, &dataoff, &protonum)) <= 0) {
 836                 DEBUGP("not prepared to track yet or error occured\n");
 837                 return -ret;
 838         }
 839
 840         l4proto = __nf_ct_l4proto_find((u_int16_t)pf, protonum);
 841
 842         /* It may be an special packet, error, unclean...
 843          * inverse of the return code tells to the netfilter
 844          * core what to do with the packet. */
 845         if (l4proto->error != NULL &&
 846             (ret = l4proto->error(*pskb, dataoff, &ctinfo, pf, hooknum)) <= 0) {
 847                 NF_CT_STAT_INC_ATOMIC(error);
 848                 NF_CT_STAT_INC_ATOMIC(invalid);
 849                 return -ret;
 850         }
 851
 852         ct = resolve_normal_ct(*pskb, dataoff, pf, protonum, l3proto, l4proto,
 853                                &set_reply, &ctinfo);
 854         if (!ct) {
 855                 /* Not valid part of a connection */
 856                 NF_CT_STAT_INC_ATOMIC(invalid);
 857                 return NF_ACCEPT;
 858         }
 859
 860         if (IS_ERR(ct)) {
 861                 /* Too stressed to deal. */
 862                 NF_CT_STAT_INC_ATOMIC(drop);
 863                 return NF_DROP;
 864         }
 865
 866         NF_CT_ASSERT((*pskb)->nfct);
 867
 868         ret = l4proto->packet(ct, *pskb, dataoff, ctinfo, pf, hooknum);
 869         if (ret < 0) {
 870                 /* Invalid: inverse of the return code tells
 871                  * the netfilter core what to do */
 872                 DEBUGP("nf_conntrack_in: Can't track with proto module\n");
 873                 nf_conntrack_put((*pskb)->nfct);
 874                 (*pskb)->nfct = NULL;
 875                 NF_CT_STAT_INC_ATOMIC(invalid);
 876                 return -ret;
 877         }
 878
 879         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
 880                 nf_conntrack_event_cache(IPCT_STATUS, *pskb);
 881
 882         return ret;
 883 }
 884 EXPORT_SYMBOL_GPL(nf_conntrack_in);
 885
 886 int nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
 887                          const struct nf_conntrack_tuple *orig)
 888 {
 889         int ret;
 890
 891         rcu_read_lock();
 892         ret = nf_ct_invert_tuple(inverse, orig,
 893                                  __nf_ct_l3proto_find(orig->src.l3num),
 894                                  __nf_ct_l4proto_find(orig->src.l3num,
 895                                                       orig->dst.protonum));
 896         rcu_read_unlock();
 897         return ret;
 898 }
 899 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
 900
 901 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
 902    implicitly racy: see __nf_conntrack_confirm */
 903 void nf_conntrack_alter_reply(struct nf_conn *ct,
 904                               const struct nf_conntrack_tuple *newreply)
 905 {
 906         struct nf_conn_help *help = nfct_help(ct);
 907
 908         write_lock_bh(&nf_conntrack_lock);
 909         /* Should be unconfirmed, so not in hash table yet */
 910         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 911
 912         DEBUGP("Altering reply tuple of %p to ", ct);
 913         NF_CT_DUMP_TUPLE(newreply);
 914
 915         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
 916         if (!ct->master && help && help->expecting == 0)
 917                 help->helper = __nf_ct_helper_find(newreply);
 918         write_unlock_bh(&nf_conntrack_lock);
 919 }
 920 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
 921
 922 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
 923 void __nf_ct_refresh_acct(struct nf_conn *ct,
 924                           enum ip_conntrack_info ctinfo,
 925                           const struct sk_buff *skb,
 926                           unsigned long extra_jiffies,
 927                           int do_acct)
 928 {
 929         int event = 0;
 930
 931         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
 932         NF_CT_ASSERT(skb);
 933
 934         write_lock_bh(&nf_conntrack_lock);
 935
 936         /* Only update if this is not a fixed timeout */
 937         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status)) {
 938                 write_unlock_bh(&nf_conntrack_lock);
 939                 return;
 940         }
 941
 942         /* If not in hash table, timer will not be active yet */
 943         if (!nf_ct_is_confirmed(ct)) {
 944                 ct->timeout.expires = extra_jiffies;
 945                 event = IPCT_REFRESH;
 946         } else {
 947                 unsigned long newtime = jiffies + extra_jiffies;
 948
 949                 /* Only update the timeout if the new timeout is at least
 950                    HZ jiffies from the old timeout. Need del_timer for race
 951                    avoidance (may already be dying). */
 952                 if (newtime - ct->timeout.expires >= HZ
 953                     && del_timer(&ct->timeout)) {
 954                         ct->timeout.expires = newtime;
 955                         add_timer(&ct->timeout);
 956                         event = IPCT_REFRESH;
 957                 }
 958         }
 959
 960 #ifdef CONFIG_NF_CT_ACCT
 961         if (do_acct) {
 962                 ct->counters[CTINFO2DIR(ctinfo)].packets++;
 963                 ct->counters[CTINFO2DIR(ctinfo)].bytes +=
 964                         skb->len - skb_network_offset(skb);
 965
 966                 if ((ct->counters[CTINFO2DIR(ctinfo)].packets & 0x80000000)
 967                     || (ct->counters[CTINFO2DIR(ctinfo)].bytes & 0x80000000))
 968                         event |= IPCT_COUNTER_FILLING;
 969         }
 970 #endif
 971
 972         write_unlock_bh(&nf_conntrack_lock);
 973
 974         /* must be unlocked when calling event cache */
 975         if (event)
 976                 nf_conntrack_event_cache(event, skb);
 977 }
 978 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
 979
 980 #if defined(CONFIG_NF_CT_NETLINK) || defined(CONFIG_NF_CT_NETLINK_MODULE)
 981
 982 #include <linux/netfilter/nfnetlink.h>
 983 #include <linux/netfilter/nfnetlink_conntrack.h>
 984 #include <linux/mutex.h>
 985
 986
 987 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
 988  * in ip_conntrack_core, since we don't want the protocols to autoload
 989  * or depend on ctnetlink */
 990 int nf_ct_port_tuple_to_nfattr(struct sk_buff *skb,
 991                                const struct nf_conntrack_tuple *tuple)
 992 {
 993         NFA_PUT(skb, CTA_PROTO_SRC_PORT, sizeof(u_int16_t),
 994                 &tuple->src.u.tcp.port);
 995         NFA_PUT(skb, CTA_PROTO_DST_PORT, sizeof(u_int16_t),
 996                 &tuple->dst.u.tcp.port);
 997         return 0;
 998
 999 nfattr_failure:
1000         return -1;
1001 }
1002 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nfattr);
1003
1004 static const size_t cta_min_proto[CTA_PROTO_MAX] = {
1005         [CTA_PROTO_SRC_PORT-1]  = sizeof(u_int16_t),
1006         [CTA_PROTO_DST_PORT-1]  = sizeof(u_int16_t)
1007 };
1008
1009 int nf_ct_port_nfattr_to_tuple(struct nfattr *tb[],
1010                                struct nf_conntrack_tuple *t)
1011 {
1012         if (!tb[CTA_PROTO_SRC_PORT-1] || !tb[CTA_PROTO_DST_PORT-1])
1013                 return -EINVAL;
1014
1015         if (nfattr_bad_size(tb, CTA_PROTO_MAX, cta_min_proto))
1016                 return -EINVAL;
1017
1018         t->src.u.tcp.port = *(__be16 *)NFA_DATA(tb[CTA_PROTO_SRC_PORT-1]);
1019         t->dst.u.tcp.port = *(__be16 *)NFA_DATA(tb[CTA_PROTO_DST_PORT-1]);
1020
1021         return 0;
1022 }
1023 EXPORT_SYMBOL_GPL(nf_ct_port_nfattr_to_tuple);
1024 #endif
1025
1026 /* Used by ipt_REJECT and ip6t_REJECT. */
1027 void __nf_conntrack_attach(struct sk_buff *nskb, struct sk_buff *skb)
1028 {
1029         struct nf_conn *ct;
1030         enum ip_conntrack_info ctinfo;
1031
1032         /* This ICMP is in reverse direction to the packet which caused it */
1033         ct = nf_ct_get(skb, &ctinfo);
1034         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1035                 ctinfo = IP_CT_RELATED + IP_CT_IS_REPLY;
1036         else
1037                 ctinfo = IP_CT_RELATED;
1038
1039         /* Attach to new skbuff, and increment count */
1040         nskb->nfct = &ct->ct_general;
1041         nskb->nfctinfo = ctinfo;
1042         nf_conntrack_get(nskb->nfct);
1043 }
1044 EXPORT_SYMBOL_GPL(__nf_conntrack_attach);
1045
1046 static inline int
1047 do_iter(const struct nf_conntrack_tuple_hash *i,
1048         int (*iter)(struct nf_conn *i, void *data),
1049         void *data)
1050 {
1051         return iter(nf_ct_tuplehash_to_ctrack(i), data);
1052 }
1053
1054 /* Bring out ya dead! */
1055 static struct nf_conn *
1056 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
1057                 void *data, unsigned int *bucket)
1058 {
1059         struct nf_conntrack_tuple_hash *h;
1060         struct nf_conn *ct;
1061
1062         write_lock_bh(&nf_conntrack_lock);
1063         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1064                 list_for_each_entry(h, &nf_conntrack_hash[*bucket], list) {
1065                         ct = nf_ct_tuplehash_to_ctrack(h);
1066                         if (iter(ct, data))
1067                                 goto found;
1068                 }
1069         }
1070         list_for_each_entry(h, &unconfirmed, list) {
1071                 ct = nf_ct_tuplehash_to_ctrack(h);
1072                 if (iter(ct, data))
1073                         set_bit(IPS_DYING_BIT, &ct->status);
1074         }
1075         write_unlock_bh(&nf_conntrack_lock);
1076         return NULL;
1077 found:
1078         atomic_inc(&ct->ct_general.use);
1079         write_unlock_bh(&nf_conntrack_lock);
1080         return ct;
1081 }
1082
1083 void
1084 nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data), void *data)
1085 {
1086         struct nf_conn *ct;
1087         unsigned int bucket = 0;
1088
1089         while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
1090                 /* Time to push up daises... */
1091                 if (del_timer(&ct->timeout))
1092                         death_by_timeout((unsigned long)ct);
1093                 /* ... else the timer will get him soon. */
1094
1095                 nf_ct_put(ct);
1096         }
1097 }
1098 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1099
1100 static int kill_all(struct nf_conn *i, void *data)
1101 {
1102         return 1;
1103 }
1104
1105 static void free_conntrack_hash(struct list_head *hash, int vmalloced, int size)
1106 {
1107         if (vmalloced)
1108                 vfree(hash);
1109         else
1110                 free_pages((unsigned long)hash,
1111                            get_order(sizeof(struct list_head) * size));
1112 }
1113
1114 void nf_conntrack_flush(void)
1115 {
1116         nf_ct_iterate_cleanup(kill_all, NULL);
1117 }
1118 EXPORT_SYMBOL_GPL(nf_conntrack_flush);
1119
1120 /* Mishearing the voices in his head, our hero wonders how he's
1121    supposed to kill the mall. */
1122 void nf_conntrack_cleanup(void)
1123 {
1124         int i;
1125
1126         rcu_assign_pointer(ip_ct_attach, NULL);
1127
1128         /* This makes sure all current packets have passed through
1129            netfilter framework.  Roll on, two-stage module
1130            delete... */
1131         synchronize_net();
1132
1133         nf_ct_event_cache_flush();
1134  i_see_dead_people:
1135         nf_conntrack_flush();
1136         if (atomic_read(&nf_conntrack_count) != 0) {
1137                 schedule();
1138                 goto i_see_dead_people;
1139         }
1140         /* wait until all references to nf_conntrack_untracked are dropped */
1141         while (atomic_read(&nf_conntrack_untracked.ct_general.use) > 1)
1142                 schedule();
1143
1144         for (i = 0; i < NF_CT_F_NUM; i++) {
1145                 if (nf_ct_cache[i].use == 0)
1146                         continue;
1147
1148                 NF_CT_ASSERT(nf_ct_cache[i].use == 1);
1149                 nf_ct_cache[i].use = 1;
1150                 nf_conntrack_unregister_cache(i);
1151         }
1152         kmem_cache_destroy(nf_conntrack_expect_cachep);
1153         free_conntrack_hash(nf_conntrack_hash, nf_conntrack_vmalloc,
1154                             nf_conntrack_htable_size);
1155
1156         nf_conntrack_proto_fini();
1157 }
1158
1159 static struct list_head *alloc_hashtable(int size, int *vmalloced)
1160 {
1161         struct list_head *hash;
1162         unsigned int i;
1163
1164         *vmalloced = 0;
1165         hash = (void*)__get_free_pages(GFP_KERNEL,
1166                                        get_order(sizeof(struct list_head)
1167                                                  * size));
1168         if (!hash) {
1169                 *vmalloced = 1;
1170                 printk(KERN_WARNING "nf_conntrack: falling back to vmalloc.\n");
1171                 hash = vmalloc(sizeof(struct list_head) * size);
1172         }
1173
1174         if (hash)
1175                 for (i = 0; i < size; i++)
1176                         INIT_LIST_HEAD(&hash[i]);
1177
1178         return hash;
1179 }
1180
1181 int set_hashsize(const char *val, struct kernel_param *kp)
1182 {
1183         int i, bucket, hashsize, vmalloced;
1184         int old_vmalloced, old_size;
1185         int rnd;
1186         struct list_head *hash, *old_hash;
1187         struct nf_conntrack_tuple_hash *h;
1188
1189         /* On boot, we can set this without any fancy locking. */
1190         if (!nf_conntrack_htable_size)
1191                 return param_set_uint(val, kp);
1192
1193         hashsize = simple_strtol(val, NULL, 0);
1194         if (!hashsize)
1195                 return -EINVAL;
1196
1197         hash = alloc_hashtable(hashsize, &vmalloced);
1198         if (!hash)
1199                 return -ENOMEM;
1200
1201         /* We have to rehahs for the new table anyway, so we also can
1202          * use a newrandom seed */
1203         get_random_bytes(&rnd, 4);
1204
1205         write_lock_bh(&nf_conntrack_lock);
1206         for (i = 0; i < nf_conntrack_htable_size; i++) {
1207                 while (!list_empty(&nf_conntrack_hash[i])) {
1208                         h = list_entry(nf_conntrack_hash[i].next,
1209                                        struct nf_conntrack_tuple_hash, list);
1210                         list_del(&h->list);
1211                         bucket = __hash_conntrack(&h->tuple, hashsize, rnd);
1212                         list_add_tail(&h->list, &hash[bucket]);
1213                 }
1214         }
1215         old_size = nf_conntrack_htable_size;
1216         old_vmalloced = nf_conntrack_vmalloc;
1217         old_hash = nf_conntrack_hash;
1218
1219         nf_conntrack_htable_size = hashsize;
1220         nf_conntrack_vmalloc = vmalloced;
1221         nf_conntrack_hash = hash;
1222         nf_conntrack_hash_rnd = rnd;
1223         write_unlock_bh(&nf_conntrack_lock);
1224
1225         free_conntrack_hash(old_hash, old_vmalloced, old_size);
1226         return 0;
1227 }
1228
1229 module_param_call(hashsize, set_hashsize, param_get_uint,
1230                   &nf_conntrack_htable_size, 0600);
1231
1232 int __init nf_conntrack_init(void)
1233 {
1234         int ret;
1235
1236         /* Idea from tcp.c: use 1/16384 of memory.  On i386: 32MB
1237          * machine has 256 buckets.  >= 1GB machines have 8192 buckets. */
1238         if (!nf_conntrack_htable_size) {
1239                 nf_conntrack_htable_size
1240                         = (((num_physpages << PAGE_SHIFT) / 16384)
1241                            / sizeof(struct list_head));
1242                 if (num_physpages > (1024 * 1024 * 1024 / PAGE_SIZE))
1243                         nf_conntrack_htable_size = 8192;
1244                 if (nf_conntrack_htable_size < 16)
1245                         nf_conntrack_htable_size = 16;
1246         }
1247         nf_conntrack_max = 8 * nf_conntrack_htable_size;
1248
1249         printk("nf_conntrack version %s (%u buckets, %d max)\n",
1250                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1251                nf_conntrack_max);
1252
1253         nf_conntrack_hash = alloc_hashtable(nf_conntrack_htable_size,
1254                                             &nf_conntrack_vmalloc);
1255         if (!nf_conntrack_hash) {
1256                 printk(KERN_ERR "Unable to create nf_conntrack_hash\n");
1257                 goto err_out;
1258         }
1259
1260         ret = nf_conntrack_register_cache(NF_CT_F_BASIC, "nf_conntrack:basic",
1261                                           sizeof(struct nf_conn));
1262         if (ret < 0) {
1263                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
1264                 goto err_free_hash;
1265         }
1266
1267         nf_conntrack_expect_cachep = kmem_cache_create("nf_conntrack_expect",
1268                                         sizeof(struct nf_conntrack_expect),
1269                                         0, 0, NULL, NULL);
1270         if (!nf_conntrack_expect_cachep) {
1271                 printk(KERN_ERR "Unable to create nf_expect slab cache\n");
1272                 goto err_free_conntrack_slab;
1273         }
1274
1275         ret = nf_conntrack_proto_init();
1276         if (ret < 0)
1277                 goto out_free_expect_slab;
1278
1279         /* For use by REJECT target */
1280         rcu_assign_pointer(ip_ct_attach, __nf_conntrack_attach);
1281
1282         /* Set up fake conntrack:
1283             - to never be deleted, not in any hashes */
1284         atomic_set(&nf_conntrack_untracked.ct_general.use, 1);
1285         /*  - and look it like as a confirmed connection */
1286         set_bit(IPS_CONFIRMED_BIT, &nf_conntrack_untracked.status);
1287
1288         return ret;
1289
1290 out_free_expect_slab:
1291         kmem_cache_destroy(nf_conntrack_expect_cachep);
1292 err_free_conntrack_slab:
1293         nf_conntrack_unregister_cache(NF_CT_F_BASIC);
1294 err_free_hash:
1295         free_conntrack_hash(nf_conntrack_hash, nf_conntrack_vmalloc,
1296                             nf_conntrack_htable_size);
1297 err_out:
1298         return -ENOMEM;
1299 }