net/netfilter/nf_conntrack_core.c

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