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

   1 /* Connection state tracking for netfilter.  This is separated from,
   2    but required by, the NAT layer; it can also be used by an iptables
   3    extension. */
   4
   5 /* (C) 1999-2001 Paul `Rusty' Russell
   6  * (C) 2002-2006 Netfilter Core Team <coreteam@netfilter.org>
   7  * (C) 2003,2004 USAGI/WIDE Project <http://www.linux-ipv6.org>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * 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         a = jhash((void *)tuple->src.u3.all, sizeof(tuple->src.u3.all),
 132                   ((tuple->src.l3num) << 16) | tuple->dst.protonum);
 133         b = jhash((void *)tuple->dst.u3.all, sizeof(tuple->dst.u3.all),
 134                         (tuple->src.u.all << 16) | tuple->dst.u.all);
 135
 136         return jhash_2words(a, b, rnd) % size;
 137 }
 138
 139 static inline u_int32_t hash_conntrack(const struct nf_conntrack_tuple *tuple)
 140 {
 141         return __hash_conntrack(tuple, nf_conntrack_htable_size,
 142                                 nf_conntrack_hash_rnd);
 143 }
 144
 145 int nf_conntrack_register_cache(u_int32_t features, const char *name,
 146                                 size_t size)
 147 {
 148         int ret = 0;
 149         char *cache_name;
 150         struct kmem_cache *cachep;
 151
 152         DEBUGP("nf_conntrack_register_cache: features=0x%x, name=%s, size=%d\n",
 153                features, name, size);
 154
 155         if (features < NF_CT_F_BASIC || features >= NF_CT_F_NUM) {
 156                 DEBUGP("nf_conntrack_register_cache: invalid features.: 0x%x\n",
 157                         features);
 158                 return -EINVAL;
 159         }
 160
 161         mutex_lock(&nf_ct_cache_mutex);
 162
 163         write_lock_bh(&nf_ct_cache_lock);
 164         /* e.g: multiple helpers are loaded */
 165         if (nf_ct_cache[features].use > 0) {
 166                 DEBUGP("nf_conntrack_register_cache: already resisterd.\n");
 167                 if ((!strncmp(nf_ct_cache[features].name, name,
 168                               NF_CT_FEATURES_NAMELEN))
 169                     && nf_ct_cache[features].size == size) {
 170                         DEBUGP("nf_conntrack_register_cache: reusing.\n");
 171                         nf_ct_cache[features].use++;
 172                         ret = 0;
 173                 } else
 174                         ret = -EBUSY;
 175
 176                 write_unlock_bh(&nf_ct_cache_lock);
 177                 mutex_unlock(&nf_ct_cache_mutex);
 178                 return ret;
 179         }
 180         write_unlock_bh(&nf_ct_cache_lock);
 181
 182         /*
 183          * The memory space for name of slab cache must be alive until
 184          * cache is destroyed.
 185          */
 186         cache_name = kmalloc(sizeof(char)*NF_CT_FEATURES_NAMELEN, GFP_ATOMIC);
 187         if (cache_name == NULL) {
 188                 DEBUGP("nf_conntrack_register_cache: can't alloc cache_name\n");
 189                 ret = -ENOMEM;
 190                 goto out_up_mutex;
 191         }
 192
 193         if (strlcpy(cache_name, name, NF_CT_FEATURES_NAMELEN)
 194                                                 >= NF_CT_FEATURES_NAMELEN) {
 195                 printk("nf_conntrack_register_cache: name too long\n");
 196                 ret = -EINVAL;
 197                 goto out_free_name;
 198         }
 199
 200         cachep = kmem_cache_create(cache_name, size, 0, 0,
 201                                    NULL, NULL);
 202         if (!cachep) {
 203                 printk("nf_conntrack_register_cache: Can't create slab cache "
 204                        "for the features = 0x%x\n", features);
 205                 ret = -ENOMEM;
 206                 goto out_free_name;
 207         }
 208
 209         write_lock_bh(&nf_ct_cache_lock);
 210         nf_ct_cache[features].use = 1;
 211         nf_ct_cache[features].size = size;
 212         nf_ct_cache[features].cachep = cachep;
 213         nf_ct_cache[features].name = cache_name;
 214         write_unlock_bh(&nf_ct_cache_lock);
 215
 216         goto out_up_mutex;
 217
 218 out_free_name:
 219         kfree(cache_name);
 220 out_up_mutex:
 221         mutex_unlock(&nf_ct_cache_mutex);
 222         return ret;
 223 }
 224 EXPORT_SYMBOL_GPL(nf_conntrack_register_cache);
 225
 226 /* FIXME: In the current, only nf_conntrack_cleanup() can call this function. */
 227 void nf_conntrack_unregister_cache(u_int32_t features)
 228 {
 229         struct kmem_cache *cachep;
 230         char *name;
 231
 232         /*
 233          * This assures that kmem_cache_create() isn't called before destroying
 234          * slab cache.
 235          */
 236         DEBUGP("nf_conntrack_unregister_cache: 0x%04x\n", features);
 237         mutex_lock(&nf_ct_cache_mutex);
 238
 239         write_lock_bh(&nf_ct_cache_lock);
 240         if (--nf_ct_cache[features].use > 0) {
 241                 write_unlock_bh(&nf_ct_cache_lock);
 242                 mutex_unlock(&nf_ct_cache_mutex);
 243                 return;
 244         }
 245         cachep = nf_ct_cache[features].cachep;
 246         name = nf_ct_cache[features].name;
 247         nf_ct_cache[features].cachep = NULL;
 248         nf_ct_cache[features].name = NULL;
 249         nf_ct_cache[features].size = 0;
 250         write_unlock_bh(&nf_ct_cache_lock);
 251
 252         synchronize_net();
 253
 254         kmem_cache_destroy(cachep);
 255         kfree(name);
 256
 257         mutex_unlock(&nf_ct_cache_mutex);
 258 }
 259 EXPORT_SYMBOL_GPL(nf_conntrack_unregister_cache);
 260
 261 int
 262 nf_ct_get_tuple(const struct sk_buff *skb,
 263                 unsigned int nhoff,
 264                 unsigned int dataoff,
 265                 u_int16_t l3num,
 266                 u_int8_t protonum,
 267                 struct nf_conntrack_tuple *tuple,
 268                 const struct nf_conntrack_l3proto *l3proto,
 269                 const struct nf_conntrack_l4proto *l4proto)
 270 {
 271         NF_CT_TUPLE_U_BLANK(tuple);
 272
 273         tuple->src.l3num = l3num;
 274         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 275                 return 0;
 276
 277         tuple->dst.protonum = protonum;
 278         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 279
 280         return l4proto->pkt_to_tuple(skb, dataoff, tuple);
 281 }
 282 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 283
 284 int
 285 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 286                    const struct nf_conntrack_tuple *orig,
 287                    const struct nf_conntrack_l3proto *l3proto,
 288                    const struct nf_conntrack_l4proto *l4proto)
 289 {
 290         NF_CT_TUPLE_U_BLANK(inverse);
 291
 292         inverse->src.l3num = orig->src.l3num;
 293         if (l3proto->invert_tuple(inverse, orig) == 0)
 294                 return 0;
 295
 296         inverse->dst.dir = !orig->dst.dir;
 297
 298         inverse->dst.protonum = orig->dst.protonum;
 299         return l4proto->invert_tuple(inverse, orig);
 300 }
 301 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 302
 303 static void
 304 clean_from_lists(struct nf_conn *ct)
 305 {
 306         DEBUGP("clean_from_lists(%p)\n", ct);
 307         list_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list);
 308         list_del(&ct->tuplehash[IP_CT_DIR_REPLY].list);
 309
 310         /* Destroy all pending expectations */
 311         nf_ct_remove_expectations(ct);
 312 }
 313
 314 static void
 315 destroy_conntrack(struct nf_conntrack *nfct)
 316 {
 317         struct nf_conn *ct = (struct nf_conn *)nfct;
 318         struct nf_conn_help *help = nfct_help(ct);
 319         struct nf_conntrack_l3proto *l3proto;
 320         struct nf_conntrack_l4proto *l4proto;
 321         typeof(nf_conntrack_destroyed) destroyed;
 322
 323         DEBUGP("destroy_conntrack(%p)\n", ct);
 324         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 325         NF_CT_ASSERT(!timer_pending(&ct->timeout));
 326
 327         nf_conntrack_event(IPCT_DESTROY, ct);
 328         set_bit(IPS_DYING_BIT, &ct->status);
 329
 330         if (help && help->helper && help->helper->destroy)
 331                 help->helper->destroy(ct);
 332
 333         /* To make sure we don't get any weird locking issues here:
 334          * destroy_conntrack() MUST NOT be called with a write lock
 335          * to nf_conntrack_lock!!! -HW */
 336         rcu_read_lock();
 337         l3proto = __nf_ct_l3proto_find(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.l3num);
 338         if (l3proto && l3proto->destroy)
 339                 l3proto->destroy(ct);
 340
 341         l4proto = __nf_ct_l4proto_find(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.l3num,
 342                                        ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.protonum);
 343         if (l4proto && l4proto->destroy)
 344                 l4proto->destroy(ct);
 345
 346         destroyed = rcu_dereference(nf_conntrack_destroyed);
 347         if (destroyed)
 348                 destroyed(ct);
 349
 350         rcu_read_unlock();
 351
 352         write_lock_bh(&nf_conntrack_lock);
 353         /* Expectations will have been removed in clean_from_lists,
 354          * except TFTP can create an expectation on the first packet,
 355          * before connection is in the list, so we need to clean here,
 356          * too. */
 357         nf_ct_remove_expectations(ct);
 358
 359         /* We overload first tuple to link into unconfirmed list. */
 360         if (!nf_ct_is_confirmed(ct)) {
 361                 BUG_ON(list_empty(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list));
 362                 list_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list);
 363         }
 364
 365         NF_CT_STAT_INC(delete);
 366         write_unlock_bh(&nf_conntrack_lock);
 367
 368         if (ct->master)
 369                 nf_ct_put(ct->master);
 370
 371         DEBUGP("destroy_conntrack: returning ct=%p to slab\n", ct);
 372         nf_conntrack_free(ct);
 373 }
 374
 375 static void death_by_timeout(unsigned long ul_conntrack)
 376 {
 377         struct nf_conn *ct = (void *)ul_conntrack;
 378
 379         write_lock_bh(&nf_conntrack_lock);
 380         /* Inside lock so preempt is disabled on module removal path.
 381          * Otherwise we can get spurious warnings. */
 382         NF_CT_STAT_INC(delete_list);
 383         clean_from_lists(ct);
 384         write_unlock_bh(&nf_conntrack_lock);
 385         nf_ct_put(ct);
 386 }
 387
 388 struct nf_conntrack_tuple_hash *
 389 __nf_conntrack_find(const struct nf_conntrack_tuple *tuple,
 390                     const struct nf_conn *ignored_conntrack)
 391 {
 392         struct nf_conntrack_tuple_hash *h;
 393         unsigned int hash = hash_conntrack(tuple);
 394
 395         list_for_each_entry(h, &nf_conntrack_hash[hash], list) {
 396                 if (nf_ct_tuplehash_to_ctrack(h) != ignored_conntrack &&
 397                     nf_ct_tuple_equal(tuple, &h->tuple)) {
 398                         NF_CT_STAT_INC(found);
 399                         return h;
 400                 }
 401                 NF_CT_STAT_INC(searched);
 402         }
 403
 404         return NULL;
 405 }
 406 EXPORT_SYMBOL_GPL(__nf_conntrack_find);
 407
 408 /* Find a connection corresponding to a tuple. */
 409 struct nf_conntrack_tuple_hash *
 410 nf_conntrack_find_get(const struct nf_conntrack_tuple *tuple,
 411                       const struct nf_conn *ignored_conntrack)
 412 {
 413         struct nf_conntrack_tuple_hash *h;
 414
 415         read_lock_bh(&nf_conntrack_lock);
 416         h = __nf_conntrack_find(tuple, ignored_conntrack);
 417         if (h)
 418                 atomic_inc(&nf_ct_tuplehash_to_ctrack(h)->ct_general.use);
 419         read_unlock_bh(&nf_conntrack_lock);
 420
 421         return h;
 422 }
 423 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 424
 425 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 426                                        unsigned int hash,
 427                                        unsigned int repl_hash)
 428 {
 429         ct->id = ++nf_conntrack_next_id;
 430         list_add(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list,
 431                  &nf_conntrack_hash[hash]);
 432         list_add(&ct->tuplehash[IP_CT_DIR_REPLY].list,
 433                  &nf_conntrack_hash[repl_hash]);
 434 }
 435
 436 void nf_conntrack_hash_insert(struct nf_conn *ct)
 437 {
 438         unsigned int hash, repl_hash;
 439
 440         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 441         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 442
 443         write_lock_bh(&nf_conntrack_lock);
 444         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 445         write_unlock_bh(&nf_conntrack_lock);
 446 }
 447 EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
 448
 449 /* Confirm a connection given skb; places it in hash table */
 450 int
 451 __nf_conntrack_confirm(struct sk_buff **pskb)
 452 {
 453         unsigned int hash, repl_hash;
 454         struct nf_conntrack_tuple_hash *h;
 455         struct nf_conn *ct;
 456         struct nf_conn_help *help;
 457         enum ip_conntrack_info ctinfo;
 458
 459         ct = nf_ct_get(*pskb, &ctinfo);
 460
 461         /* ipt_REJECT uses nf_conntrack_attach to attach related
 462            ICMP/TCP RST packets in other direction.  Actual packet
 463            which created connection will be IP_CT_NEW or for an
 464            expected connection, IP_CT_RELATED. */
 465         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 466                 return NF_ACCEPT;
 467
 468         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 469         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 470
 471         /* We're not in hash table, and we refuse to set up related
 472            connections for unconfirmed conns.  But packet copies and
 473            REJECT will give spurious warnings here. */
 474         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 475
 476         /* No external references means noone else could have
 477            confirmed us. */
 478         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 479         DEBUGP("Confirming conntrack %p\n", ct);
 480
 481         write_lock_bh(&nf_conntrack_lock);
 482
 483         /* See if there's one in the list already, including reverse:
 484            NAT could have grabbed it without realizing, since we're
 485            not in the hash.  If there is, we lost race. */
 486         list_for_each_entry(h, &nf_conntrack_hash[hash], list)
 487                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 488                                       &h->tuple))
 489                         goto out;
 490         list_for_each_entry(h, &nf_conntrack_hash[repl_hash], list)
 491                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 492                                       &h->tuple))
 493                         goto out;
 494
 495         /* Remove from unconfirmed list */
 496         list_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].list);
 497
 498         __nf_conntrack_hash_insert(ct, hash, repl_hash);
 499         /* Timer relative to confirmation time, not original
 500            setting time, otherwise we'd get timer wrap in
 501            weird delay cases. */
 502         ct->timeout.expires += jiffies;
 503         add_timer(&ct->timeout);
 504         atomic_inc(&ct->ct_general.use);
 505         set_bit(IPS_CONFIRMED_BIT, &ct->status);
 506         NF_CT_STAT_INC(insert);
 507         write_unlock_bh(&nf_conntrack_lock);
 508         help = nfct_help(ct);
 509         if (help && help->helper)
 510                 nf_conntrack_event_cache(IPCT_HELPER, *pskb);
 511 #ifdef CONFIG_NF_NAT_NEEDED
 512         if (test_bit(IPS_SRC_NAT_DONE_BIT, &ct->status) ||
 513             test_bit(IPS_DST_NAT_DONE_BIT, &ct->status))
 514                 nf_conntrack_event_cache(IPCT_NATINFO, *pskb);
 515 #endif
 516         nf_conntrack_event_cache(master_ct(ct) ?
 517                                  IPCT_RELATED : IPCT_NEW, *pskb);
 518         return NF_ACCEPT;
 519
 520 out:
 521         NF_CT_STAT_INC(insert_failed);
 522         write_unlock_bh(&nf_conntrack_lock);
 523         return NF_DROP;
 524 }
 525 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 526
 527 /* Returns true if a connection correspondings to the tuple (required
 528    for NAT). */
 529 int
 530 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 531                          const struct nf_conn *ignored_conntrack)
 532 {
 533         struct nf_conntrack_tuple_hash *h;
 534
 535         read_lock_bh(&nf_conntrack_lock);
 536         h = __nf_conntrack_find(tuple, ignored_conntrack);
 537         read_unlock_bh(&nf_conntrack_lock);
 538
 539         return h != NULL;
 540 }
 541 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 542
 543 /* There's a small race here where we may free a just-assured
 544    connection.  Too bad: we're in trouble anyway. */
 545 static int early_drop(struct list_head *chain)
 546 {
 547         /* Traverse backwards: gives us oldest, which is roughly LRU */
 548         struct nf_conntrack_tuple_hash *h;
 549         struct nf_conn *ct = NULL, *tmp;
 550         int dropped = 0;
 551
 552         read_lock_bh(&nf_conntrack_lock);
 553         list_for_each_entry_reverse(h, chain, list) {
 554                 tmp = nf_ct_tuplehash_to_ctrack(h);
 555                 if (!test_bit(IPS_ASSURED_BIT, &tmp->status)) {
 556                         ct = tmp;
 557                         atomic_inc(&ct->ct_general.use);
 558                         break;
 559                 }
 560         }
 561         read_unlock_bh(&nf_conntrack_lock);
 562
 563         if (!ct)
 564                 return dropped;
 565
 566         if (del_timer(&ct->timeout)) {
 567                 death_by_timeout((unsigned long)ct);
 568                 dropped = 1;
 569                 NF_CT_STAT_INC_ATOMIC(early_drop);
 570         }
 571         nf_ct_put(ct);
 572         return dropped;
 573 }
 574
 575 static struct nf_conn *
 576 __nf_conntrack_alloc(const struct nf_conntrack_tuple *orig,
 577                      const struct nf_conntrack_tuple *repl,
 578                      const struct nf_conntrack_l3proto *l3proto,
 579                      u_int32_t features)
 580 {
 581         struct nf_conn *conntrack = NULL;
 582         struct nf_conntrack_helper *helper;
 583
 584         if (unlikely(!nf_conntrack_hash_rnd_initted)) {
 585                 get_random_bytes(&nf_conntrack_hash_rnd, 4);
 586                 nf_conntrack_hash_rnd_initted = 1;
 587         }
 588
 589         /* We don't want any race condition at early drop stage */
 590         atomic_inc(&nf_conntrack_count);
 591
 592         if (nf_conntrack_max
 593             && atomic_read(&nf_conntrack_count) > nf_conntrack_max) {
 594                 unsigned int hash = hash_conntrack(orig);
 595                 /* Try dropping from this hash chain. */
 596                 if (!early_drop(&nf_conntrack_hash[hash])) {
 597                         atomic_dec(&nf_conntrack_count);
 598                         if (net_ratelimit())
 599                                 printk(KERN_WARNING
 600                                        "nf_conntrack: table full, dropping"
 601                                        " packet.\n");
 602                         return ERR_PTR(-ENOMEM);
 603                 }
 604         }
 605
 606         /*  find features needed by this conntrack. */
 607         features |= l3proto->get_features(orig);
 608
 609         /* FIXME: protect helper list per RCU */
 610         read_lock_bh(&nf_conntrack_lock);
 611         helper = __nf_ct_helper_find(repl);
 612         /* NAT might want to assign a helper later */
 613         if (helper || features & NF_CT_F_NAT)
 614                 features |= NF_CT_F_HELP;
 615         read_unlock_bh(&nf_conntrack_lock);
 616
 617         DEBUGP("nf_conntrack_alloc: features=0x%x\n", features);
 618
 619         read_lock_bh(&nf_ct_cache_lock);
 620
 621         if (unlikely(!nf_ct_cache[features].use)) {
 622                 DEBUGP("nf_conntrack_alloc: not supported features = 0x%x\n",
 623                         features);
 624                 goto out;
 625         }
 626
 627         conntrack = kmem_cache_alloc(nf_ct_cache[features].cachep, GFP_ATOMIC);
 628         if (conntrack == NULL) {
 629                 DEBUGP("nf_conntrack_alloc: Can't alloc conntrack from cache\n");
 630                 goto out;
 631         }
 632
 633         memset(conntrack, 0, nf_ct_cache[features].size);
 634         conntrack->features = features;
 635         atomic_set(&conntrack->ct_general.use, 1);
 636         conntrack->ct_general.destroy = destroy_conntrack;
 637         conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
 638         conntrack->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
 639         /* Don't set timer yet: wait for confirmation */
 640         init_timer(&conntrack->timeout);
 641         conntrack->timeout.data = (unsigned long)conntrack;
 642         conntrack->timeout.function = death_by_timeout;
 643         read_unlock_bh(&nf_ct_cache_lock);
 644
 645         return conntrack;
 646 out:
 647         read_unlock_bh(&nf_ct_cache_lock);
 648         atomic_dec(&nf_conntrack_count);
 649         return conntrack;
 650 }
 651
 652 struct nf_conn *nf_conntrack_alloc(const struct nf_conntrack_tuple *orig,
 653                                    const struct nf_conntrack_tuple *repl)
 654 {
 655         struct nf_conntrack_l3proto *l3proto;
 656         struct nf_conn *ct;
 657
 658         rcu_read_lock();
 659         l3proto = __nf_ct_l3proto_find(orig->src.l3num);
 660         ct = __nf_conntrack_alloc(orig, repl, l3proto, 0);
 661         rcu_read_unlock();
 662
 663         return ct;
 664 }
 665 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
 666
 667 void nf_conntrack_free(struct nf_conn *conntrack)
 668 {
 669         u_int32_t features = conntrack->features;
 670         NF_CT_ASSERT(features >= NF_CT_F_BASIC && features < NF_CT_F_NUM);
 671         DEBUGP("nf_conntrack_free: features = 0x%x, conntrack=%p\n", features,
 672                conntrack);
 673         kmem_cache_free(nf_ct_cache[features].cachep, conntrack);
 674         atomic_dec(&nf_conntrack_count);
 675 }
 676 EXPORT_SYMBOL_GPL(nf_conntrack_free);
 677
 678 /* Allocate a new conntrack: we return -ENOMEM if classification
 679    failed due to stress.  Otherwise it really is unclassifiable. */
 680 static struct nf_conntrack_tuple_hash *
 681 init_conntrack(const struct nf_conntrack_tuple *tuple,
 682                struct nf_conntrack_l3proto *l3proto,
 683                struct nf_conntrack_l4proto *l4proto,
 684                struct sk_buff *skb,
 685                unsigned int dataoff)
 686 {
 687         struct nf_conn *conntrack;
 688         struct nf_conntrack_tuple repl_tuple;
 689         struct nf_conntrack_expect *exp;
 690         u_int32_t features = 0;
 691
 692         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
 693                 DEBUGP("Can't invert tuple.\n");
 694                 return NULL;
 695         }
 696
 697         read_lock_bh(&nf_conntrack_lock);
 698         exp = __nf_conntrack_expect_find(tuple);
 699         if (exp && exp->helper)
 700                 features = NF_CT_F_HELP;
 701         read_unlock_bh(&nf_conntrack_lock);
 702
 703         conntrack = __nf_conntrack_alloc(tuple, &repl_tuple, l3proto, features);
 704         if (conntrack == NULL || IS_ERR(conntrack)) {
 705                 DEBUGP("Can't allocate conntrack.\n");
 706                 return (struct nf_conntrack_tuple_hash *)conntrack;
 707         }
 708
 709         if (!l4proto->new(conntrack, skb, dataoff)) {
 710                 nf_conntrack_free(conntrack);
 711                 DEBUGP("init conntrack: can't track with proto module\n");
 712                 return NULL;
 713         }
 714
 715         write_lock_bh(&nf_conntrack_lock);
 716         exp = find_expectation(tuple);
 717
 718         if (exp) {
 719                 DEBUGP("conntrack: expectation arrives ct=%p exp=%p\n",
 720                         conntrack, exp);
 721                 /* Welcome, Mr. Bond.  We've been expecting you... */
 722                 __set_bit(IPS_EXPECTED_BIT, &conntrack->status);
 723                 conntrack->master = exp->master;
 724                 if (exp->helper)
 725                         nfct_help(conntrack)->helper = exp->helper;
 726 #ifdef CONFIG_NF_CONNTRACK_MARK
 727                 conntrack->mark = exp->master->mark;
 728 #endif
 729 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 730                 conntrack->secmark = exp->master->secmark;
 731 #endif
 732                 nf_conntrack_get(&conntrack->master->ct_general);
 733                 NF_CT_STAT_INC(expect_new);
 734         } else {
 735                 struct nf_conn_help *help = nfct_help(conntrack);
 736
 737                 if (help)
 738                         help->helper = __nf_ct_helper_find(&repl_tuple);
 739                 NF_CT_STAT_INC(new);
 740         }
 741
 742         /* Overload tuple linked list to put us in unconfirmed list. */
 743         list_add(&conntrack->tuplehash[IP_CT_DIR_ORIGINAL].list, &unconfirmed);
 744
 745         write_unlock_bh(&nf_conntrack_lock);
 746
 747         if (exp) {
 748                 if (exp->expectfn)
 749                         exp->expectfn(conntrack, exp);
 750                 nf_conntrack_expect_put(exp);
 751         }
 752
 753         return &conntrack->tuplehash[IP_CT_DIR_ORIGINAL];
 754 }
 755
 756 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
 757 static inline struct nf_conn *
 758 resolve_normal_ct(struct sk_buff *skb,
 759                   unsigned int dataoff,
 760                   u_int16_t l3num,
 761                   u_int8_t protonum,
 762                   struct nf_conntrack_l3proto *l3proto,
 763                   struct nf_conntrack_l4proto *l4proto,
 764                   int *set_reply,
 765                   enum ip_conntrack_info *ctinfo)
 766 {
 767         struct nf_conntrack_tuple tuple;
 768         struct nf_conntrack_tuple_hash *h;
 769         struct nf_conn *ct;
 770
 771         if (!nf_ct_get_tuple(skb, (unsigned int)(skb->nh.raw - skb->data),
 772                              dataoff, l3num, protonum, &tuple, l3proto,
 773                              l4proto)) {
 774                 DEBUGP("resolve_normal_ct: Can't get tuple\n");
 775                 return NULL;
 776         }
 777
 778         /* look for tuple match */
 779         h = nf_conntrack_find_get(&tuple, NULL);
 780         if (!h) {
 781                 h = init_conntrack(&tuple, l3proto, l4proto, skb, dataoff);
 782                 if (!h)
 783                         return NULL;
 784                 if (IS_ERR(h))
 785                         return (void *)h;
 786         }
 787         ct = nf_ct_tuplehash_to_ctrack(h);
 788
 789         /* It exists; we have (non-exclusive) reference. */
 790         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
 791                 *ctinfo = IP_CT_ESTABLISHED + IP_CT_IS_REPLY;
 792                 /* Please set reply bit if this packet OK */
 793                 *set_reply = 1;
 794         } else {
 795                 /* Once we've had two way comms, always ESTABLISHED. */
 796                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
 797                         DEBUGP("nf_conntrack_in: normal packet for %p\n", ct);
 798                         *ctinfo = IP_CT_ESTABLISHED;
 799                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
 800                         DEBUGP("nf_conntrack_in: related packet for %p\n", ct);
 801                         *ctinfo = IP_CT_RELATED;
 802                 } else {
 803                         DEBUGP("nf_conntrack_in: new packet for %p\n", ct);
 804                         *ctinfo = IP_CT_NEW;
 805                 }
 806                 *set_reply = 0;
 807         }
 808         skb->nfct = &ct->ct_general;
 809         skb->nfctinfo = *ctinfo;
 810         return ct;
 811 }
 812
 813 unsigned int
 814 nf_conntrack_in(int pf, unsigned int hooknum, struct sk_buff **pskb)
 815 {
 816         struct nf_conn *ct;
 817         enum ip_conntrack_info ctinfo;
 818         struct nf_conntrack_l3proto *l3proto;
 819         struct nf_conntrack_l4proto *l4proto;
 820         unsigned int dataoff;
 821         u_int8_t protonum;
 822         int set_reply = 0;
 823         int ret;
 824
 825         /* Previously seen (loopback or untracked)?  Ignore. */
 826         if ((*pskb)->nfct) {
 827                 NF_CT_STAT_INC_ATOMIC(ignore);
 828                 return NF_ACCEPT;
 829         }
 830
 831         /* rcu_read_lock()ed by nf_hook_slow */
 832         l3proto = __nf_ct_l3proto_find((u_int16_t)pf);
 833
 834         if ((ret = l3proto->prepare(pskb, hooknum, &dataoff, &protonum)) <= 0) {
 835                 DEBUGP("not prepared to track yet or error occured\n");
 836                 return -ret;
 837         }
 838
 839         l4proto = __nf_ct_l4proto_find((u_int16_t)pf, protonum);
 840
 841         /* It may be an special packet, error, unclean...
 842          * inverse of the return code tells to the netfilter
 843          * core what to do with the packet. */
 844         if (l4proto->error != NULL &&
 845             (ret = l4proto->error(*pskb, dataoff, &ctinfo, pf, hooknum)) <= 0) {
 846                 NF_CT_STAT_INC_ATOMIC(error);
 847                 NF_CT_STAT_INC_ATOMIC(invalid);
 848                 return -ret;
 849         }
 850
 851         ct = resolve_normal_ct(*pskb, dataoff, pf, protonum, l3proto, l4proto,
 852                                &set_reply, &ctinfo);
 853         if (!ct) {
 854                 /* Not valid part of a connection */
 855                 NF_CT_STAT_INC_ATOMIC(invalid);
 856                 return NF_ACCEPT;
 857         }
 858
 859         if (IS_ERR(ct)) {
 860                 /* Too stressed to deal. */
 861                 NF_CT_STAT_INC_ATOMIC(drop);
 862                 return NF_DROP;
 863         }
 864
 865         NF_CT_ASSERT((*pskb)->nfct);
 866
 867         ret = l4proto->packet(ct, *pskb, dataoff, ctinfo, pf, hooknum);
 868         if (ret < 0) {
 869                 /* Invalid: inverse of the return code tells
 870                  * the netfilter core what to do */
 871                 DEBUGP("nf_conntrack_in: Can't track with proto module\n");
 872                 nf_conntrack_put((*pskb)->nfct);
 873                 (*pskb)->nfct = NULL;
 874                 NF_CT_STAT_INC_ATOMIC(invalid);
 875                 return -ret;
 876         }
 877
 878         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
 879                 nf_conntrack_event_cache(IPCT_STATUS, *pskb);
 880
 881         return ret;
 882 }
 883 EXPORT_SYMBOL_GPL(nf_conntrack_in);
 884
 885 int nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
 886                          const struct nf_conntrack_tuple *orig)
 887 {
 888         int ret;
 889
 890         rcu_read_lock();
 891         ret = nf_ct_invert_tuple(inverse, orig,
 892                                  __nf_ct_l3proto_find(orig->src.l3num),
 893                                  __nf_ct_l4proto_find(orig->src.l3num,
 894                                                       orig->dst.protonum));
 895         rcu_read_unlock();
 896         return ret;
 897 }
 898 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
 899
 900 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
 901    implicitly racy: see __nf_conntrack_confirm */
 902 void nf_conntrack_alter_reply(struct nf_conn *ct,
 903                               const struct nf_conntrack_tuple *newreply)
 904 {
 905         struct nf_conn_help *help = nfct_help(ct);
 906
 907         write_lock_bh(&nf_conntrack_lock);
 908         /* Should be unconfirmed, so not in hash table yet */
 909         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 910
 911         DEBUGP("Altering reply tuple of %p to ", ct);
 912         NF_CT_DUMP_TUPLE(newreply);
 913
 914         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
 915         if (!ct->master && help && help->expecting == 0)
 916                 help->helper = __nf_ct_helper_find(newreply);
 917         write_unlock_bh(&nf_conntrack_lock);
 918 }
 919 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
 920
 921 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
 922 void __nf_ct_refresh_acct(struct nf_conn *ct,
 923                           enum ip_conntrack_info ctinfo,
 924                           const struct sk_buff *skb,
 925                           unsigned long extra_jiffies,
 926                           int do_acct)
 927 {
 928         int event = 0;
 929
 930         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
 931         NF_CT_ASSERT(skb);
 932
 933         write_lock_bh(&nf_conntrack_lock);
 934
 935         /* Only update if this is not a fixed timeout */
 936         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status)) {
 937                 write_unlock_bh(&nf_conntrack_lock);
 938                 return;
 939         }
 940
 941         /* If not in hash table, timer will not be active yet */
 942         if (!nf_ct_is_confirmed(ct)) {
 943                 ct->timeout.expires = extra_jiffies;
 944                 event = IPCT_REFRESH;
 945         } else {
 946                 unsigned long newtime = jiffies + extra_jiffies;
 947
 948                 /* Only update the timeout if the new timeout is at least
 949                    HZ jiffies from the old timeout. Need del_timer for race
 950                    avoidance (may already be dying). */
 951                 if (newtime - ct->timeout.expires >= HZ
 952                     && del_timer(&ct->timeout)) {
 953                         ct->timeout.expires = newtime;
 954                         add_timer(&ct->timeout);
 955                         event = IPCT_REFRESH;
 956                 }
 957         }
 958
 959 #ifdef CONFIG_NF_CT_ACCT
 960         if (do_acct) {
 961                 ct->counters[CTINFO2DIR(ctinfo)].packets++;
 962                 ct->counters[CTINFO2DIR(ctinfo)].bytes +=
 963                         skb->len - (unsigned int)(skb->nh.raw - skb->data);
 964
 965                 if ((ct->counters[CTINFO2DIR(ctinfo)].packets & 0x80000000)
 966                     || (ct->counters[CTINFO2DIR(ctinfo)].bytes & 0x80000000))
 967                         event |= IPCT_COUNTER_FILLING;
 968         }
 969 #endif
 970
 971         write_unlock_bh(&nf_conntrack_lock);
 972
 973         /* must be unlocked when calling event cache */
 974         if (event)
 975                 nf_conntrack_event_cache(event, skb);
 976 }
 977 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
 978
 979 #if defined(CONFIG_NF_CT_NETLINK) || \
 980     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                         goto found;
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_l4proto_unregister(&nf_conntrack_l4proto_generic);
1157
1158         /* free l3proto protocol tables */
1159         for (i = 0; i < PF_MAX; i++)
1160                 if (nf_ct_protos[i]) {
1161                         kfree(nf_ct_protos[i]);
1162                         nf_ct_protos[i] = NULL;
1163                 }
1164 }
1165
1166 static struct list_head *alloc_hashtable(int size, int *vmalloced)
1167 {
1168         struct list_head *hash;
1169         unsigned int i;
1170
1171         *vmalloced = 0;
1172         hash = (void*)__get_free_pages(GFP_KERNEL,
1173                                        get_order(sizeof(struct list_head)
1174                                                  * size));
1175         if (!hash) {
1176                 *vmalloced = 1;
1177                 printk(KERN_WARNING "nf_conntrack: falling back to vmalloc.\n");
1178                 hash = vmalloc(sizeof(struct list_head) * size);
1179         }
1180
1181         if (hash)
1182                 for (i = 0; i < size; i++)
1183                         INIT_LIST_HEAD(&hash[i]);
1184
1185         return hash;
1186 }
1187
1188 int set_hashsize(const char *val, struct kernel_param *kp)
1189 {
1190         int i, bucket, hashsize, vmalloced;
1191         int old_vmalloced, old_size;
1192         int rnd;
1193         struct list_head *hash, *old_hash;
1194         struct nf_conntrack_tuple_hash *h;
1195
1196         /* On boot, we can set this without any fancy locking. */
1197         if (!nf_conntrack_htable_size)
1198                 return param_set_uint(val, kp);
1199
1200         hashsize = simple_strtol(val, NULL, 0);
1201         if (!hashsize)
1202                 return -EINVAL;
1203
1204         hash = alloc_hashtable(hashsize, &vmalloced);
1205         if (!hash)
1206                 return -ENOMEM;
1207
1208         /* We have to rehahs for the new table anyway, so we also can
1209          * use a newrandom seed */
1210         get_random_bytes(&rnd, 4);
1211
1212         write_lock_bh(&nf_conntrack_lock);
1213         for (i = 0; i < nf_conntrack_htable_size; i++) {
1214                 while (!list_empty(&nf_conntrack_hash[i])) {
1215                         h = list_entry(nf_conntrack_hash[i].next,
1216                                        struct nf_conntrack_tuple_hash, list);
1217                         list_del(&h->list);
1218                         bucket = __hash_conntrack(&h->tuple, hashsize, rnd);
1219                         list_add_tail(&h->list, &hash[bucket]);
1220                 }
1221         }
1222         old_size = nf_conntrack_htable_size;
1223         old_vmalloced = nf_conntrack_vmalloc;
1224         old_hash = nf_conntrack_hash;
1225
1226         nf_conntrack_htable_size = hashsize;
1227         nf_conntrack_vmalloc = vmalloced;
1228         nf_conntrack_hash = hash;
1229         nf_conntrack_hash_rnd = rnd;
1230         write_unlock_bh(&nf_conntrack_lock);
1231
1232         free_conntrack_hash(old_hash, old_vmalloced, old_size);
1233         return 0;
1234 }
1235
1236 module_param_call(hashsize, set_hashsize, param_get_uint,
1237                   &nf_conntrack_htable_size, 0600);
1238
1239 int __init nf_conntrack_init(void)
1240 {
1241         unsigned int i;
1242         int ret;
1243
1244         /* Idea from tcp.c: use 1/16384 of memory.  On i386: 32MB
1245          * machine has 256 buckets.  >= 1GB machines have 8192 buckets. */
1246         if (!nf_conntrack_htable_size) {
1247                 nf_conntrack_htable_size
1248                         = (((num_physpages << PAGE_SHIFT) / 16384)
1249                            / sizeof(struct list_head));
1250                 if (num_physpages > (1024 * 1024 * 1024 / PAGE_SIZE))
1251                         nf_conntrack_htable_size = 8192;
1252                 if (nf_conntrack_htable_size < 16)
1253                         nf_conntrack_htable_size = 16;
1254         }
1255         nf_conntrack_max = 8 * nf_conntrack_htable_size;
1256
1257         printk("nf_conntrack version %s (%u buckets, %d max)\n",
1258                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1259                nf_conntrack_max);
1260
1261         nf_conntrack_hash = alloc_hashtable(nf_conntrack_htable_size,
1262                                             &nf_conntrack_vmalloc);
1263         if (!nf_conntrack_hash) {
1264                 printk(KERN_ERR "Unable to create nf_conntrack_hash\n");
1265                 goto err_out;
1266         }
1267
1268         ret = nf_conntrack_register_cache(NF_CT_F_BASIC, "nf_conntrack:basic",
1269                                           sizeof(struct nf_conn));
1270         if (ret < 0) {
1271                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
1272                 goto err_free_hash;
1273         }
1274
1275         nf_conntrack_expect_cachep = kmem_cache_create("nf_conntrack_expect",
1276                                         sizeof(struct nf_conntrack_expect),
1277                                         0, 0, NULL, NULL);
1278         if (!nf_conntrack_expect_cachep) {
1279                 printk(KERN_ERR "Unable to create nf_expect slab cache\n");
1280                 goto err_free_conntrack_slab;
1281         }
1282
1283         ret = nf_conntrack_l4proto_register(&nf_conntrack_l4proto_generic);
1284         if (ret < 0)
1285                 goto out_free_expect_slab;
1286
1287         /* Don't NEED lock here, but good form anyway. */
1288         write_lock_bh(&nf_conntrack_lock);
1289         for (i = 0; i < AF_MAX; i++)
1290                 nf_ct_l3protos[i] = &nf_conntrack_l3proto_generic;
1291         write_unlock_bh(&nf_conntrack_lock);
1292
1293         /* For use by REJECT target */
1294         rcu_assign_pointer(ip_ct_attach, __nf_conntrack_attach);
1295
1296         /* Set up fake conntrack:
1297             - to never be deleted, not in any hashes */
1298         atomic_set(&nf_conntrack_untracked.ct_general.use, 1);
1299         /*  - and look it like as a confirmed connection */
1300         set_bit(IPS_CONFIRMED_BIT, &nf_conntrack_untracked.status);
1301
1302         return ret;
1303
1304 out_free_expect_slab:
1305         kmem_cache_destroy(nf_conntrack_expect_cachep);
1306 err_free_conntrack_slab:
1307         nf_conntrack_unregister_cache(NF_CT_F_BASIC);
1308 err_free_hash:
1309         free_conntrack_hash(nf_conntrack_hash, nf_conntrack_vmalloc,
1310                             nf_conntrack_htable_size);
1311 err_out:
1312         return -ENOMEM;
1313 }