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  * (C) 2005-2012 Patrick McHardy <kaber@trash.net>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License version 2 as
  12  * published by the Free Software Foundation.
  13  */
  14
  15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  16
  17 #include <linux/types.h>
  18 #include <linux/netfilter.h>
  19 #include <linux/module.h>
  20 #include <linux/sched.h>
  21 #include <linux/skbuff.h>
  22 #include <linux/proc_fs.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/stddef.h>
  25 #include <linux/slab.h>
  26 #include <linux/random.h>
  27 #include <linux/jhash.h>
  28 #include <linux/err.h>
  29 #include <linux/percpu.h>
  30 #include <linux/moduleparam.h>
  31 #include <linux/notifier.h>
  32 #include <linux/kernel.h>
  33 #include <linux/netdevice.h>
  34 #include <linux/socket.h>
  35 #include <linux/mm.h>
  36 #include <linux/nsproxy.h>
  37 #include <linux/rculist_nulls.h>
  38
  39 #include <net/netfilter/nf_conntrack.h>
  40 #include <net/netfilter/nf_conntrack_l3proto.h>
  41 #include <net/netfilter/nf_conntrack_l4proto.h>
  42 #include <net/netfilter/nf_conntrack_expect.h>
  43 #include <net/netfilter/nf_conntrack_helper.h>
  44 #include <net/netfilter/nf_conntrack_seqadj.h>
  45 #include <net/netfilter/nf_conntrack_core.h>
  46 #include <net/netfilter/nf_conntrack_extend.h>
  47 #include <net/netfilter/nf_conntrack_acct.h>
  48 #include <net/netfilter/nf_conntrack_ecache.h>
  49 #include <net/netfilter/nf_conntrack_zones.h>
  50 #include <net/netfilter/nf_conntrack_timestamp.h>
  51 #include <net/netfilter/nf_conntrack_timeout.h>
  52 #include <net/netfilter/nf_conntrack_labels.h>
  53 #include <net/netfilter/nf_conntrack_synproxy.h>
  54 #include <net/netfilter/nf_nat.h>
  55 #include <net/netfilter/nf_nat_core.h>
  56 #include <net/netfilter/nf_nat_helper.h>
  57 #include <net/netns/hash.h>
  58
  59 #define NF_CONNTRACK_VERSION    "0.5.0"
  60
  61 int (*nfnetlink_parse_nat_setup_hook)(struct nf_conn *ct,
  62                                       enum nf_nat_manip_type manip,
  63                                       const struct nlattr *attr) __read_mostly;
  64 EXPORT_SYMBOL_GPL(nfnetlink_parse_nat_setup_hook);
  65
  66 __cacheline_aligned_in_smp spinlock_t nf_conntrack_locks[CONNTRACK_LOCKS];
  67 EXPORT_SYMBOL_GPL(nf_conntrack_locks);
  68
  69 __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
  70 EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
  71
  72 struct hlist_nulls_head *nf_conntrack_hash __read_mostly;
  73 EXPORT_SYMBOL_GPL(nf_conntrack_hash);
  74
  75 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
  76 static __read_mostly spinlock_t nf_conntrack_locks_all_lock;
  77 static __read_mostly seqcount_t nf_conntrack_generation;
  78 static __read_mostly DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
  79 static __read_mostly bool nf_conntrack_locks_all;
  80
  81 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
  82 {
  83         spin_lock(lock);
  84         while (unlikely(nf_conntrack_locks_all)) {
  85                 spin_unlock(lock);
  86
  87                 /*
  88                  * Order the 'nf_conntrack_locks_all' load vs. the
  89                  * spin_unlock_wait() loads below, to ensure
  90                  * that 'nf_conntrack_locks_all_lock' is indeed held:
  91                  */
  92                 smp_rmb(); /* spin_lock(&nf_conntrack_locks_all_lock) */
  93                 spin_unlock_wait(&nf_conntrack_locks_all_lock);
  94                 spin_lock(lock);
  95         }
  96 }
  97 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
  98
  99 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 100 {
 101         h1 %= CONNTRACK_LOCKS;
 102         h2 %= CONNTRACK_LOCKS;
 103         spin_unlock(&nf_conntrack_locks[h1]);
 104         if (h1 != h2)
 105                 spin_unlock(&nf_conntrack_locks[h2]);
 106 }
 107
 108 /* return true if we need to recompute hashes (in case hash table was resized) */
 109 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
 110                                      unsigned int h2, unsigned int sequence)
 111 {
 112         h1 %= CONNTRACK_LOCKS;
 113         h2 %= CONNTRACK_LOCKS;
 114         if (h1 <= h2) {
 115                 nf_conntrack_lock(&nf_conntrack_locks[h1]);
 116                 if (h1 != h2)
 117                         spin_lock_nested(&nf_conntrack_locks[h2],
 118                                          SINGLE_DEPTH_NESTING);
 119         } else {
 120                 nf_conntrack_lock(&nf_conntrack_locks[h2]);
 121                 spin_lock_nested(&nf_conntrack_locks[h1],
 122                                  SINGLE_DEPTH_NESTING);
 123         }
 124         if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
 125                 nf_conntrack_double_unlock(h1, h2);
 126                 return true;
 127         }
 128         return false;
 129 }
 130
 131 static void nf_conntrack_all_lock(void)
 132 {
 133         int i;
 134
 135         spin_lock(&nf_conntrack_locks_all_lock);
 136         nf_conntrack_locks_all = true;
 137
 138         /*
 139          * Order the above store of 'nf_conntrack_locks_all' against
 140          * the spin_unlock_wait() loads below, such that if
 141          * nf_conntrack_lock() observes 'nf_conntrack_locks_all'
 142          * we must observe nf_conntrack_locks[] held:
 143          */
 144         smp_mb(); /* spin_lock(&nf_conntrack_locks_all_lock) */
 145
 146         for (i = 0; i < CONNTRACK_LOCKS; i++) {
 147                 spin_unlock_wait(&nf_conntrack_locks[i]);
 148         }
 149 }
 150
 151 static void nf_conntrack_all_unlock(void)
 152 {
 153         /*
 154          * All prior stores must be complete before we clear
 155          * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
 156          * might observe the false value but not the entire
 157          * critical section:
 158          */
 159         smp_store_release(&nf_conntrack_locks_all, false);
 160         spin_unlock(&nf_conntrack_locks_all_lock);
 161 }
 162
 163 unsigned int nf_conntrack_htable_size __read_mostly;
 164 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
 165
 166 unsigned int nf_conntrack_max __read_mostly;
 167 EXPORT_SYMBOL_GPL(nf_conntrack_max);
 168
 169 DEFINE_PER_CPU(struct nf_conn, nf_conntrack_untracked);
 170 EXPORT_PER_CPU_SYMBOL(nf_conntrack_untracked);
 171
 172 static unsigned int nf_conntrack_hash_rnd __read_mostly;
 173
 174 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
 175                               const struct net *net)
 176 {
 177         unsigned int n;
 178         u32 seed;
 179
 180         get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
 181
 182         /* The direction must be ignored, so we hash everything up to the
 183          * destination ports (which is a multiple of 4) and treat the last
 184          * three bytes manually.
 185          */
 186         seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
 187         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
 188         return jhash2((u32 *)tuple, n, seed ^
 189                       (((__force __u16)tuple->dst.u.all << 16) |
 190                       tuple->dst.protonum));
 191 }
 192
 193 static u32 scale_hash(u32 hash)
 194 {
 195         return reciprocal_scale(hash, nf_conntrack_htable_size);
 196 }
 197
 198 static u32 __hash_conntrack(const struct net *net,
 199                             const struct nf_conntrack_tuple *tuple,
 200                             unsigned int size)
 201 {
 202         return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
 203 }
 204
 205 static u32 hash_conntrack(const struct net *net,
 206                           const struct nf_conntrack_tuple *tuple)
 207 {
 208         return scale_hash(hash_conntrack_raw(tuple, net));
 209 }
 210
 211 bool
 212 nf_ct_get_tuple(const struct sk_buff *skb,
 213                 unsigned int nhoff,
 214                 unsigned int dataoff,
 215                 u_int16_t l3num,
 216                 u_int8_t protonum,
 217                 struct net *net,
 218                 struct nf_conntrack_tuple *tuple,
 219                 const struct nf_conntrack_l3proto *l3proto,
 220                 const struct nf_conntrack_l4proto *l4proto)
 221 {
 222         memset(tuple, 0, sizeof(*tuple));
 223
 224         tuple->src.l3num = l3num;
 225         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 226                 return false;
 227
 228         tuple->dst.protonum = protonum;
 229         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 230
 231         return l4proto->pkt_to_tuple(skb, dataoff, net, tuple);
 232 }
 233 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 234
 235 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 236                        u_int16_t l3num,
 237                        struct net *net, struct nf_conntrack_tuple *tuple)
 238 {
 239         struct nf_conntrack_l3proto *l3proto;
 240         struct nf_conntrack_l4proto *l4proto;
 241         unsigned int protoff;
 242         u_int8_t protonum;
 243         int ret;
 244
 245         rcu_read_lock();
 246
 247         l3proto = __nf_ct_l3proto_find(l3num);
 248         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
 249         if (ret != NF_ACCEPT) {
 250                 rcu_read_unlock();
 251                 return false;
 252         }
 253
 254         l4proto = __nf_ct_l4proto_find(l3num, protonum);
 255
 256         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple,
 257                               l3proto, l4proto);
 258
 259         rcu_read_unlock();
 260         return ret;
 261 }
 262 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 263
 264 bool
 265 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 266                    const struct nf_conntrack_tuple *orig,
 267                    const struct nf_conntrack_l3proto *l3proto,
 268                    const struct nf_conntrack_l4proto *l4proto)
 269 {
 270         memset(inverse, 0, sizeof(*inverse));
 271
 272         inverse->src.l3num = orig->src.l3num;
 273         if (l3proto->invert_tuple(inverse, orig) == 0)
 274                 return false;
 275
 276         inverse->dst.dir = !orig->dst.dir;
 277
 278         inverse->dst.protonum = orig->dst.protonum;
 279         return l4proto->invert_tuple(inverse, orig);
 280 }
 281 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 282
 283 static void
 284 clean_from_lists(struct nf_conn *ct)
 285 {
 286         pr_debug("clean_from_lists(%p)\n", ct);
 287         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 288         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 289
 290         /* Destroy all pending expectations */
 291         nf_ct_remove_expectations(ct);
 292 }
 293
 294 /* must be called with local_bh_disable */
 295 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
 296 {
 297         struct ct_pcpu *pcpu;
 298
 299         /* add this conntrack to the (per cpu) dying list */
 300         ct->cpu = smp_processor_id();
 301         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 302
 303         spin_lock(&pcpu->lock);
 304         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 305                              &pcpu->dying);
 306         spin_unlock(&pcpu->lock);
 307 }
 308
 309 /* must be called with local_bh_disable */
 310 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
 311 {
 312         struct ct_pcpu *pcpu;
 313
 314         /* add this conntrack to the (per cpu) unconfirmed list */
 315         ct->cpu = smp_processor_id();
 316         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 317
 318         spin_lock(&pcpu->lock);
 319         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 320                              &pcpu->unconfirmed);
 321         spin_unlock(&pcpu->lock);
 322 }
 323
 324 /* must be called with local_bh_disable */
 325 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
 326 {
 327         struct ct_pcpu *pcpu;
 328
 329         /* We overload first tuple to link into unconfirmed or dying list.*/
 330         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 331
 332         spin_lock(&pcpu->lock);
 333         BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 334         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 335         spin_unlock(&pcpu->lock);
 336 }
 337
 338 /* Released via destroy_conntrack() */
 339 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
 340                                  const struct nf_conntrack_zone *zone,
 341                                  gfp_t flags)
 342 {
 343         struct nf_conn *tmpl;
 344
 345         tmpl = kzalloc(sizeof(*tmpl), flags);
 346         if (tmpl == NULL)
 347                 return NULL;
 348
 349         tmpl->status = IPS_TEMPLATE;
 350         write_pnet(&tmpl->ct_net, net);
 351         nf_ct_zone_add(tmpl, zone);
 352         atomic_set(&tmpl->ct_general.use, 0);
 353
 354         return tmpl;
 355 }
 356 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
 357
 358 void nf_ct_tmpl_free(struct nf_conn *tmpl)
 359 {
 360         nf_ct_ext_destroy(tmpl);
 361         nf_ct_ext_free(tmpl);
 362         kfree(tmpl);
 363 }
 364 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
 365
 366 static void
 367 destroy_conntrack(struct nf_conntrack *nfct)
 368 {
 369         struct nf_conn *ct = (struct nf_conn *)nfct;
 370         struct net *net = nf_ct_net(ct);
 371         struct nf_conntrack_l4proto *l4proto;
 372
 373         pr_debug("destroy_conntrack(%p)\n", ct);
 374         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 375         NF_CT_ASSERT(!timer_pending(&ct->timeout));
 376
 377         if (unlikely(nf_ct_is_template(ct))) {
 378                 nf_ct_tmpl_free(ct);
 379                 return;
 380         }
 381         rcu_read_lock();
 382         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 383         if (l4proto->destroy)
 384                 l4proto->destroy(ct);
 385
 386         rcu_read_unlock();
 387
 388         local_bh_disable();
 389         /* Expectations will have been removed in clean_from_lists,
 390          * except TFTP can create an expectation on the first packet,
 391          * before connection is in the list, so we need to clean here,
 392          * too.
 393          */
 394         nf_ct_remove_expectations(ct);
 395
 396         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 397
 398         NF_CT_STAT_INC(net, delete);
 399         local_bh_enable();
 400
 401         if (ct->master)
 402                 nf_ct_put(ct->master);
 403
 404         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 405         nf_conntrack_free(ct);
 406 }
 407
 408 static void nf_ct_delete_from_lists(struct nf_conn *ct)
 409 {
 410         struct net *net = nf_ct_net(ct);
 411         unsigned int hash, reply_hash;
 412         unsigned int sequence;
 413
 414         nf_ct_helper_destroy(ct);
 415
 416         local_bh_disable();
 417         do {
 418                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 419                 hash = hash_conntrack(net,
 420                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 421                 reply_hash = hash_conntrack(net,
 422                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 423         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 424
 425         clean_from_lists(ct);
 426         nf_conntrack_double_unlock(hash, reply_hash);
 427
 428         nf_ct_add_to_dying_list(ct);
 429
 430         NF_CT_STAT_INC(net, delete_list);
 431         local_bh_enable();
 432 }
 433
 434 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
 435 {
 436         struct nf_conn_tstamp *tstamp;
 437
 438         tstamp = nf_conn_tstamp_find(ct);
 439         if (tstamp && tstamp->stop == 0)
 440                 tstamp->stop = ktime_get_real_ns();
 441
 442         if (nf_ct_is_dying(ct))
 443                 goto delete;
 444
 445         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 446                                     portid, report) < 0) {
 447                 /* destroy event was not delivered */
 448                 nf_ct_delete_from_lists(ct);
 449                 nf_conntrack_ecache_delayed_work(nf_ct_net(ct));
 450                 return false;
 451         }
 452
 453         nf_conntrack_ecache_work(nf_ct_net(ct));
 454         set_bit(IPS_DYING_BIT, &ct->status);
 455  delete:
 456         nf_ct_delete_from_lists(ct);
 457         nf_ct_put(ct);
 458         return true;
 459 }
 460 EXPORT_SYMBOL_GPL(nf_ct_delete);
 461
 462 static void death_by_timeout(unsigned long ul_conntrack)
 463 {
 464         nf_ct_delete((struct nf_conn *)ul_conntrack, 0, 0);
 465 }
 466
 467 static inline bool
 468 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
 469                 const struct nf_conntrack_tuple *tuple,
 470                 const struct nf_conntrack_zone *zone,
 471                 const struct net *net)
 472 {
 473         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 474
 475         /* A conntrack can be recreated with the equal tuple,
 476          * so we need to check that the conntrack is confirmed
 477          */
 478         return nf_ct_tuple_equal(tuple, &h->tuple) &&
 479                nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
 480                nf_ct_is_confirmed(ct) &&
 481                net_eq(net, nf_ct_net(ct));
 482 }
 483
 484 /* must be called with rcu read lock held */
 485 void nf_conntrack_get_ht(struct hlist_nulls_head **hash, unsigned int *hsize)
 486 {
 487         struct hlist_nulls_head *hptr;
 488         unsigned int sequence, hsz;
 489
 490         do {
 491                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 492                 hsz = nf_conntrack_htable_size;
 493                 hptr = nf_conntrack_hash;
 494         } while (read_seqcount_retry(&nf_conntrack_generation, sequence));
 495
 496         *hash = hptr;
 497         *hsize = hsz;
 498 }
 499 EXPORT_SYMBOL_GPL(nf_conntrack_get_ht);
 500
 501 /*
 502  * Warning :
 503  * - Caller must take a reference on returned object
 504  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 505  */
 506 static struct nf_conntrack_tuple_hash *
 507 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
 508                       const struct nf_conntrack_tuple *tuple, u32 hash)
 509 {
 510         struct nf_conntrack_tuple_hash *h;
 511         struct hlist_nulls_head *ct_hash;
 512         struct hlist_nulls_node *n;
 513         unsigned int bucket, sequence;
 514
 515 begin:
 516         do {
 517                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 518                 bucket = scale_hash(hash);
 519                 ct_hash = nf_conntrack_hash;
 520         } while (read_seqcount_retry(&nf_conntrack_generation, sequence));
 521
 522         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
 523                 if (nf_ct_key_equal(h, tuple, zone, net)) {
 524                         NF_CT_STAT_INC_ATOMIC(net, found);
 525                         return h;
 526                 }
 527                 NF_CT_STAT_INC_ATOMIC(net, searched);
 528         }
 529         /*
 530          * if the nulls value we got at the end of this lookup is
 531          * not the expected one, we must restart lookup.
 532          * We probably met an item that was moved to another chain.
 533          */
 534         if (get_nulls_value(n) != bucket) {
 535                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 536                 goto begin;
 537         }
 538
 539         return NULL;
 540 }
 541
 542 /* Find a connection corresponding to a tuple. */
 543 static struct nf_conntrack_tuple_hash *
 544 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 545                         const struct nf_conntrack_tuple *tuple, u32 hash)
 546 {
 547         struct nf_conntrack_tuple_hash *h;
 548         struct nf_conn *ct;
 549
 550         rcu_read_lock();
 551 begin:
 552         h = ____nf_conntrack_find(net, zone, tuple, hash);
 553         if (h) {
 554                 ct = nf_ct_tuplehash_to_ctrack(h);
 555                 if (unlikely(nf_ct_is_dying(ct) ||
 556                              !atomic_inc_not_zero(&ct->ct_general.use)))
 557                         h = NULL;
 558                 else {
 559                         if (unlikely(!nf_ct_key_equal(h, tuple, zone, net))) {
 560                                 nf_ct_put(ct);
 561                                 goto begin;
 562                         }
 563                 }
 564         }
 565         rcu_read_unlock();
 566
 567         return h;
 568 }
 569
 570 struct nf_conntrack_tuple_hash *
 571 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 572                       const struct nf_conntrack_tuple *tuple)
 573 {
 574         return __nf_conntrack_find_get(net, zone, tuple,
 575                                        hash_conntrack_raw(tuple, net));
 576 }
 577 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 578
 579 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 580                                        unsigned int hash,
 581                                        unsigned int reply_hash)
 582 {
 583         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 584                            &nf_conntrack_hash[hash]);
 585         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 586                            &nf_conntrack_hash[reply_hash]);
 587 }
 588
 589 int
 590 nf_conntrack_hash_check_insert(struct nf_conn *ct)
 591 {
 592         const struct nf_conntrack_zone *zone;
 593         struct net *net = nf_ct_net(ct);
 594         unsigned int hash, reply_hash;
 595         struct nf_conntrack_tuple_hash *h;
 596         struct hlist_nulls_node *n;
 597         unsigned int sequence;
 598
 599         zone = nf_ct_zone(ct);
 600
 601         local_bh_disable();
 602         do {
 603                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 604                 hash = hash_conntrack(net,
 605                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 606                 reply_hash = hash_conntrack(net,
 607                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 608         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 609
 610         /* See if there's one in the list already, including reverse */
 611         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 612                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 613                                     zone, net))
 614                         goto out;
 615
 616         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 617                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 618                                     zone, net))
 619                         goto out;
 620
 621         add_timer(&ct->timeout);
 622         smp_wmb();
 623         /* The caller holds a reference to this object */
 624         atomic_set(&ct->ct_general.use, 2);
 625         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 626         nf_conntrack_double_unlock(hash, reply_hash);
 627         NF_CT_STAT_INC(net, insert);
 628         local_bh_enable();
 629         return 0;
 630
 631 out:
 632         nf_conntrack_double_unlock(hash, reply_hash);
 633         NF_CT_STAT_INC(net, insert_failed);
 634         local_bh_enable();
 635         return -EEXIST;
 636 }
 637 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 638
 639 static inline void nf_ct_acct_update(struct nf_conn *ct,
 640                                      enum ip_conntrack_info ctinfo,
 641                                      unsigned int len)
 642 {
 643         struct nf_conn_acct *acct;
 644
 645         acct = nf_conn_acct_find(ct);
 646         if (acct) {
 647                 struct nf_conn_counter *counter = acct->counter;
 648
 649                 atomic64_inc(&counter[CTINFO2DIR(ctinfo)].packets);
 650                 atomic64_add(len, &counter[CTINFO2DIR(ctinfo)].bytes);
 651         }
 652 }
 653
 654 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
 655                              const struct nf_conn *loser_ct)
 656 {
 657         struct nf_conn_acct *acct;
 658
 659         acct = nf_conn_acct_find(loser_ct);
 660         if (acct) {
 661                 struct nf_conn_counter *counter = acct->counter;
 662                 unsigned int bytes;
 663
 664                 /* u32 should be fine since we must have seen one packet. */
 665                 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
 666                 nf_ct_acct_update(ct, ctinfo, bytes);
 667         }
 668 }
 669
 670 /* Resolve race on insertion if this protocol allows this. */
 671 static int nf_ct_resolve_clash(struct net *net, struct sk_buff *skb,
 672                                enum ip_conntrack_info ctinfo,
 673                                struct nf_conntrack_tuple_hash *h)
 674 {
 675         /* This is the conntrack entry already in hashes that won race. */
 676         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 677         struct nf_conntrack_l4proto *l4proto;
 678
 679         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 680         if (l4proto->allow_clash &&
 681             !nfct_nat(ct) &&
 682             !nf_ct_is_dying(ct) &&
 683             atomic_inc_not_zero(&ct->ct_general.use)) {
 684                 nf_ct_acct_merge(ct, ctinfo, (struct nf_conn *)skb->nfct);
 685                 nf_conntrack_put(skb->nfct);
 686                 /* Assign conntrack already in hashes to this skbuff. Don't
 687                  * modify skb->nfctinfo to ensure consistent stateful filtering.
 688                  */
 689                 skb->nfct = &ct->ct_general;
 690                 return NF_ACCEPT;
 691         }
 692         NF_CT_STAT_INC(net, drop);
 693         return NF_DROP;
 694 }
 695
 696 /* Confirm a connection given skb; places it in hash table */
 697 int
 698 __nf_conntrack_confirm(struct sk_buff *skb)
 699 {
 700         const struct nf_conntrack_zone *zone;
 701         unsigned int hash, reply_hash;
 702         struct nf_conntrack_tuple_hash *h;
 703         struct nf_conn *ct;
 704         struct nf_conn_help *help;
 705         struct nf_conn_tstamp *tstamp;
 706         struct hlist_nulls_node *n;
 707         enum ip_conntrack_info ctinfo;
 708         struct net *net;
 709         unsigned int sequence;
 710         int ret = NF_DROP;
 711
 712         ct = nf_ct_get(skb, &ctinfo);
 713         net = nf_ct_net(ct);
 714
 715         /* ipt_REJECT uses nf_conntrack_attach to attach related
 716            ICMP/TCP RST packets in other direction.  Actual packet
 717            which created connection will be IP_CT_NEW or for an
 718            expected connection, IP_CT_RELATED. */
 719         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 720                 return NF_ACCEPT;
 721
 722         zone = nf_ct_zone(ct);
 723         local_bh_disable();
 724
 725         do {
 726                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 727                 /* reuse the hash saved before */
 728                 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
 729                 hash = scale_hash(hash);
 730                 reply_hash = hash_conntrack(net,
 731                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 732
 733         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 734
 735         /* We're not in hash table, and we refuse to set up related
 736          * connections for unconfirmed conns.  But packet copies and
 737          * REJECT will give spurious warnings here.
 738          */
 739         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 740
 741         /* No external references means no one else could have
 742          * confirmed us.
 743          */
 744         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 745         pr_debug("Confirming conntrack %p\n", ct);
 746         /* We have to check the DYING flag after unlink to prevent
 747          * a race against nf_ct_get_next_corpse() possibly called from
 748          * user context, else we insert an already 'dead' hash, blocking
 749          * further use of that particular connection -JM.
 750          */
 751         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 752
 753         if (unlikely(nf_ct_is_dying(ct))) {
 754                 nf_ct_add_to_dying_list(ct);
 755                 goto dying;
 756         }
 757
 758         /* See if there's one in the list already, including reverse:
 759            NAT could have grabbed it without realizing, since we're
 760            not in the hash.  If there is, we lost race. */
 761         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 762                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 763                                     zone, net))
 764                         goto out;
 765
 766         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 767                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 768                                     zone, net))
 769                         goto out;
 770
 771         /* Timer relative to confirmation time, not original
 772            setting time, otherwise we'd get timer wrap in
 773            weird delay cases. */
 774         ct->timeout.expires += jiffies;
 775         add_timer(&ct->timeout);
 776         atomic_inc(&ct->ct_general.use);
 777         ct->status |= IPS_CONFIRMED;
 778
 779         /* set conntrack timestamp, if enabled. */
 780         tstamp = nf_conn_tstamp_find(ct);
 781         if (tstamp) {
 782                 if (skb->tstamp.tv64 == 0)
 783                         __net_timestamp(skb);
 784
 785                 tstamp->start = ktime_to_ns(skb->tstamp);
 786         }
 787         /* Since the lookup is lockless, hash insertion must be done after
 788          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 789          * guarantee that no other CPU can find the conntrack before the above
 790          * stores are visible.
 791          */
 792         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 793         nf_conntrack_double_unlock(hash, reply_hash);
 794         NF_CT_STAT_INC(net, insert);
 795         local_bh_enable();
 796
 797         help = nfct_help(ct);
 798         if (help && help->helper)
 799                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 800
 801         nf_conntrack_event_cache(master_ct(ct) ?
 802                                  IPCT_RELATED : IPCT_NEW, ct);
 803         return NF_ACCEPT;
 804
 805 out:
 806         nf_ct_add_to_dying_list(ct);
 807         ret = nf_ct_resolve_clash(net, skb, ctinfo, h);
 808 dying:
 809         nf_conntrack_double_unlock(hash, reply_hash);
 810         NF_CT_STAT_INC(net, insert_failed);
 811         local_bh_enable();
 812         return ret;
 813 }
 814 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 815
 816 /* Returns true if a connection correspondings to the tuple (required
 817    for NAT). */
 818 int
 819 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 820                          const struct nf_conn *ignored_conntrack)
 821 {
 822         struct net *net = nf_ct_net(ignored_conntrack);
 823         const struct nf_conntrack_zone *zone;
 824         struct nf_conntrack_tuple_hash *h;
 825         struct hlist_nulls_head *ct_hash;
 826         unsigned int hash, sequence;
 827         struct hlist_nulls_node *n;
 828         struct nf_conn *ct;
 829
 830         zone = nf_ct_zone(ignored_conntrack);
 831
 832         rcu_read_lock();
 833         do {
 834                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 835                 hash = hash_conntrack(net, tuple);
 836                 ct_hash = nf_conntrack_hash;
 837         } while (read_seqcount_retry(&nf_conntrack_generation, sequence));
 838
 839         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
 840                 ct = nf_ct_tuplehash_to_ctrack(h);
 841                 if (ct != ignored_conntrack &&
 842                     nf_ct_key_equal(h, tuple, zone, net)) {
 843                         NF_CT_STAT_INC_ATOMIC(net, found);
 844                         rcu_read_unlock();
 845                         return 1;
 846                 }
 847                 NF_CT_STAT_INC_ATOMIC(net, searched);
 848         }
 849         rcu_read_unlock();
 850
 851         return 0;
 852 }
 853 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 854
 855 #define NF_CT_EVICTION_RANGE    8
 856
 857 /* There's a small race here where we may free a just-assured
 858    connection.  Too bad: we're in trouble anyway. */
 859 static unsigned int early_drop_list(struct net *net,
 860                                     struct hlist_nulls_head *head)
 861 {
 862         struct nf_conntrack_tuple_hash *h;
 863         struct hlist_nulls_node *n;
 864         unsigned int drops = 0;
 865         struct nf_conn *tmp;
 866
 867         hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
 868                 tmp = nf_ct_tuplehash_to_ctrack(h);
 869
 870                 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
 871                     !net_eq(nf_ct_net(tmp), net) ||
 872                     nf_ct_is_dying(tmp))
 873                         continue;
 874
 875                 if (!atomic_inc_not_zero(&tmp->ct_general.use))
 876                         continue;
 877
 878                 /* kill only if still in same netns -- might have moved due to
 879                  * SLAB_DESTROY_BY_RCU rules.
 880                  *
 881                  * We steal the timer reference.  If that fails timer has
 882                  * already fired or someone else deleted it. Just drop ref
 883                  * and move to next entry.
 884                  */
 885                 if (net_eq(nf_ct_net(tmp), net) &&
 886                     nf_ct_is_confirmed(tmp) &&
 887                     del_timer(&tmp->timeout) &&
 888                     nf_ct_delete(tmp, 0, 0))
 889                         drops++;
 890
 891                 nf_ct_put(tmp);
 892         }
 893
 894         return drops;
 895 }
 896
 897 static noinline int early_drop(struct net *net, unsigned int _hash)
 898 {
 899         unsigned int i;
 900
 901         for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
 902                 struct hlist_nulls_head *ct_hash;
 903                 unsigned hash, sequence, drops;
 904
 905                 rcu_read_lock();
 906                 do {
 907                         sequence = read_seqcount_begin(&nf_conntrack_generation);
 908                         hash = scale_hash(_hash++);
 909                         ct_hash = nf_conntrack_hash;
 910                 } while (read_seqcount_retry(&nf_conntrack_generation, sequence));
 911
 912                 drops = early_drop_list(net, &ct_hash[hash]);
 913                 rcu_read_unlock();
 914
 915                 if (drops) {
 916                         NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
 917                         return true;
 918                 }
 919         }
 920
 921         return false;
 922 }
 923
 924 static struct nf_conn *
 925 __nf_conntrack_alloc(struct net *net,
 926                      const struct nf_conntrack_zone *zone,
 927                      const struct nf_conntrack_tuple *orig,
 928                      const struct nf_conntrack_tuple *repl,
 929                      gfp_t gfp, u32 hash)
 930 {
 931         struct nf_conn *ct;
 932
 933         /* We don't want any race condition at early drop stage */
 934         atomic_inc(&net->ct.count);
 935
 936         if (nf_conntrack_max &&
 937             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
 938                 if (!early_drop(net, hash)) {
 939                         atomic_dec(&net->ct.count);
 940                         net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
 941                         return ERR_PTR(-ENOMEM);
 942                 }
 943         }
 944
 945         /*
 946          * Do not use kmem_cache_zalloc(), as this cache uses
 947          * SLAB_DESTROY_BY_RCU.
 948          */
 949         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
 950         if (ct == NULL)
 951                 goto out;
 952
 953         spin_lock_init(&ct->lock);
 954         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
 955         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
 956         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
 957         /* save hash for reusing when confirming */
 958         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
 959         ct->status = 0;
 960         /* Don't set timer yet: wait for confirmation */
 961         setup_timer(&ct->timeout, death_by_timeout, (unsigned long)ct);
 962         write_pnet(&ct->ct_net, net);
 963         memset(&ct->__nfct_init_offset[0], 0,
 964                offsetof(struct nf_conn, proto) -
 965                offsetof(struct nf_conn, __nfct_init_offset[0]));
 966
 967         nf_ct_zone_add(ct, zone);
 968
 969         /* Because we use RCU lookups, we set ct_general.use to zero before
 970          * this is inserted in any list.
 971          */
 972         atomic_set(&ct->ct_general.use, 0);
 973         return ct;
 974 out:
 975         atomic_dec(&net->ct.count);
 976         return ERR_PTR(-ENOMEM);
 977 }
 978
 979 struct nf_conn *nf_conntrack_alloc(struct net *net,
 980                                    const struct nf_conntrack_zone *zone,
 981                                    const struct nf_conntrack_tuple *orig,
 982                                    const struct nf_conntrack_tuple *repl,
 983                                    gfp_t gfp)
 984 {
 985         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
 986 }
 987 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
 988
 989 void nf_conntrack_free(struct nf_conn *ct)
 990 {
 991         struct net *net = nf_ct_net(ct);
 992
 993         /* A freed object has refcnt == 0, that's
 994          * the golden rule for SLAB_DESTROY_BY_RCU
 995          */
 996         NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 0);
 997
 998         nf_ct_ext_destroy(ct);
 999         nf_ct_ext_free(ct);
1000         kmem_cache_free(nf_conntrack_cachep, ct);
1001         smp_mb__before_atomic();
1002         atomic_dec(&net->ct.count);
1003 }
1004 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1005
1006
1007 /* Allocate a new conntrack: we return -ENOMEM if classification
1008    failed due to stress.  Otherwise it really is unclassifiable. */
1009 static struct nf_conntrack_tuple_hash *
1010 init_conntrack(struct net *net, struct nf_conn *tmpl,
1011                const struct nf_conntrack_tuple *tuple,
1012                struct nf_conntrack_l3proto *l3proto,
1013                struct nf_conntrack_l4proto *l4proto,
1014                struct sk_buff *skb,
1015                unsigned int dataoff, u32 hash)
1016 {
1017         struct nf_conn *ct;
1018         struct nf_conn_help *help;
1019         struct nf_conntrack_tuple repl_tuple;
1020         struct nf_conntrack_ecache *ecache;
1021         struct nf_conntrack_expect *exp = NULL;
1022         const struct nf_conntrack_zone *zone;
1023         struct nf_conn_timeout *timeout_ext;
1024         struct nf_conntrack_zone tmp;
1025         unsigned int *timeouts;
1026
1027         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
1028                 pr_debug("Can't invert tuple.\n");
1029                 return NULL;
1030         }
1031
1032         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1033         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1034                                   hash);
1035         if (IS_ERR(ct))
1036                 return (struct nf_conntrack_tuple_hash *)ct;
1037
1038         if (tmpl && nfct_synproxy(tmpl)) {
1039                 nfct_seqadj_ext_add(ct);
1040                 nfct_synproxy_ext_add(ct);
1041         }
1042
1043         timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1044         if (timeout_ext) {
1045                 timeouts = nf_ct_timeout_data(timeout_ext);
1046                 if (unlikely(!timeouts))
1047                         timeouts = l4proto->get_timeouts(net);
1048         } else {
1049                 timeouts = l4proto->get_timeouts(net);
1050         }
1051
1052         if (!l4proto->new(ct, skb, dataoff, timeouts)) {
1053                 nf_conntrack_free(ct);
1054                 pr_debug("can't track with proto module\n");
1055                 return NULL;
1056         }
1057
1058         if (timeout_ext)
1059                 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1060                                       GFP_ATOMIC);
1061
1062         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1063         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1064         nf_ct_labels_ext_add(ct);
1065
1066         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1067         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1068                                  ecache ? ecache->expmask : 0,
1069                              GFP_ATOMIC);
1070
1071         local_bh_disable();
1072         if (net->ct.expect_count) {
1073                 spin_lock(&nf_conntrack_expect_lock);
1074                 exp = nf_ct_find_expectation(net, zone, tuple);
1075                 if (exp) {
1076                         pr_debug("expectation arrives ct=%p exp=%p\n",
1077                                  ct, exp);
1078                         /* Welcome, Mr. Bond.  We've been expecting you... */
1079                         __set_bit(IPS_EXPECTED_BIT, &ct->status);
1080                         /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1081                         ct->master = exp->master;
1082                         if (exp->helper) {
1083                                 help = nf_ct_helper_ext_add(ct, exp->helper,
1084                                                             GFP_ATOMIC);
1085                                 if (help)
1086                                         rcu_assign_pointer(help->helper, exp->helper);
1087                         }
1088
1089 #ifdef CONFIG_NF_CONNTRACK_MARK
1090                         ct->mark = exp->master->mark;
1091 #endif
1092 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1093                         ct->secmark = exp->master->secmark;
1094 #endif
1095                         NF_CT_STAT_INC(net, expect_new);
1096                 }
1097                 spin_unlock(&nf_conntrack_expect_lock);
1098         }
1099         if (!exp) {
1100                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1101                 NF_CT_STAT_INC(net, new);
1102         }
1103
1104         /* Now it is inserted into the unconfirmed list, bump refcount */
1105         nf_conntrack_get(&ct->ct_general);
1106         nf_ct_add_to_unconfirmed_list(ct);
1107
1108         local_bh_enable();
1109
1110         if (exp) {
1111                 if (exp->expectfn)
1112                         exp->expectfn(ct, exp);
1113                 nf_ct_expect_put(exp);
1114         }
1115
1116         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1117 }
1118
1119 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
1120 static inline struct nf_conn *
1121 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
1122                   struct sk_buff *skb,
1123                   unsigned int dataoff,
1124                   u_int16_t l3num,
1125                   u_int8_t protonum,
1126                   struct nf_conntrack_l3proto *l3proto,
1127                   struct nf_conntrack_l4proto *l4proto,
1128                   int *set_reply,
1129                   enum ip_conntrack_info *ctinfo)
1130 {
1131         const struct nf_conntrack_zone *zone;
1132         struct nf_conntrack_tuple tuple;
1133         struct nf_conntrack_tuple_hash *h;
1134         struct nf_conntrack_zone tmp;
1135         struct nf_conn *ct;
1136         u32 hash;
1137
1138         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1139                              dataoff, l3num, protonum, net, &tuple, l3proto,
1140                              l4proto)) {
1141                 pr_debug("Can't get tuple\n");
1142                 return NULL;
1143         }
1144
1145         /* look for tuple match */
1146         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1147         hash = hash_conntrack_raw(&tuple, net);
1148         h = __nf_conntrack_find_get(net, zone, &tuple, hash);
1149         if (!h) {
1150                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
1151                                    skb, dataoff, hash);
1152                 if (!h)
1153                         return NULL;
1154                 if (IS_ERR(h))
1155                         return (void *)h;
1156         }
1157         ct = nf_ct_tuplehash_to_ctrack(h);
1158
1159         /* It exists; we have (non-exclusive) reference. */
1160         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1161                 *ctinfo = IP_CT_ESTABLISHED_REPLY;
1162                 /* Please set reply bit if this packet OK */
1163                 *set_reply = 1;
1164         } else {
1165                 /* Once we've had two way comms, always ESTABLISHED. */
1166                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1167                         pr_debug("normal packet for %p\n", ct);
1168                         *ctinfo = IP_CT_ESTABLISHED;
1169                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1170                         pr_debug("related packet for %p\n", ct);
1171                         *ctinfo = IP_CT_RELATED;
1172                 } else {
1173                         pr_debug("new packet for %p\n", ct);
1174                         *ctinfo = IP_CT_NEW;
1175                 }
1176                 *set_reply = 0;
1177         }
1178         skb->nfct = &ct->ct_general;
1179         skb->nfctinfo = *ctinfo;
1180         return ct;
1181 }
1182
1183 unsigned int
1184 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
1185                 struct sk_buff *skb)
1186 {
1187         struct nf_conn *ct, *tmpl = NULL;
1188         enum ip_conntrack_info ctinfo;
1189         struct nf_conntrack_l3proto *l3proto;
1190         struct nf_conntrack_l4proto *l4proto;
1191         unsigned int *timeouts;
1192         unsigned int dataoff;
1193         u_int8_t protonum;
1194         int set_reply = 0;
1195         int ret;
1196
1197         if (skb->nfct) {
1198                 /* Previously seen (loopback or untracked)?  Ignore. */
1199                 tmpl = (struct nf_conn *)skb->nfct;
1200                 if (!nf_ct_is_template(tmpl)) {
1201                         NF_CT_STAT_INC_ATOMIC(net, ignore);
1202                         return NF_ACCEPT;
1203                 }
1204                 skb->nfct = NULL;
1205         }
1206
1207         /* rcu_read_lock()ed by nf_hook_slow */
1208         l3proto = __nf_ct_l3proto_find(pf);
1209         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
1210                                    &dataoff, &protonum);
1211         if (ret <= 0) {
1212                 pr_debug("not prepared to track yet or error occurred\n");
1213                 NF_CT_STAT_INC_ATOMIC(net, error);
1214                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1215                 ret = -ret;
1216                 goto out;
1217         }
1218
1219         l4proto = __nf_ct_l4proto_find(pf, protonum);
1220
1221         /* It may be an special packet, error, unclean...
1222          * inverse of the return code tells to the netfilter
1223          * core what to do with the packet. */
1224         if (l4proto->error != NULL) {
1225                 ret = l4proto->error(net, tmpl, skb, dataoff, &ctinfo,
1226                                      pf, hooknum);
1227                 if (ret <= 0) {
1228                         NF_CT_STAT_INC_ATOMIC(net, error);
1229                         NF_CT_STAT_INC_ATOMIC(net, invalid);
1230                         ret = -ret;
1231                         goto out;
1232                 }
1233                 /* ICMP[v6] protocol trackers may assign one conntrack. */
1234                 if (skb->nfct)
1235                         goto out;
1236         }
1237
1238         ct = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
1239                                l3proto, l4proto, &set_reply, &ctinfo);
1240         if (!ct) {
1241                 /* Not valid part of a connection */
1242                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1243                 ret = NF_ACCEPT;
1244                 goto out;
1245         }
1246
1247         if (IS_ERR(ct)) {
1248                 /* Too stressed to deal. */
1249                 NF_CT_STAT_INC_ATOMIC(net, drop);
1250                 ret = NF_DROP;
1251                 goto out;
1252         }
1253
1254         NF_CT_ASSERT(skb->nfct);
1255
1256         /* Decide what timeout policy we want to apply to this flow. */
1257         timeouts = nf_ct_timeout_lookup(net, ct, l4proto);
1258
1259         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum, timeouts);
1260         if (ret <= 0) {
1261                 /* Invalid: inverse of the return code tells
1262                  * the netfilter core what to do */
1263                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1264                 nf_conntrack_put(skb->nfct);
1265                 skb->nfct = NULL;
1266                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1267                 if (ret == -NF_DROP)
1268                         NF_CT_STAT_INC_ATOMIC(net, drop);
1269                 ret = -ret;
1270                 goto out;
1271         }
1272
1273         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1274                 nf_conntrack_event_cache(IPCT_REPLY, ct);
1275 out:
1276         if (tmpl) {
1277                 /* Special case: we have to repeat this hook, assign the
1278                  * template again to this packet. We assume that this packet
1279                  * has no conntrack assigned. This is used by nf_ct_tcp. */
1280                 if (ret == NF_REPEAT)
1281                         skb->nfct = (struct nf_conntrack *)tmpl;
1282                 else
1283                         nf_ct_put(tmpl);
1284         }
1285
1286         return ret;
1287 }
1288 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1289
1290 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
1291                           const struct nf_conntrack_tuple *orig)
1292 {
1293         bool ret;
1294
1295         rcu_read_lock();
1296         ret = nf_ct_invert_tuple(inverse, orig,
1297                                  __nf_ct_l3proto_find(orig->src.l3num),
1298                                  __nf_ct_l4proto_find(orig->src.l3num,
1299                                                       orig->dst.protonum));
1300         rcu_read_unlock();
1301         return ret;
1302 }
1303 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
1304
1305 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
1306    implicitly racy: see __nf_conntrack_confirm */
1307 void nf_conntrack_alter_reply(struct nf_conn *ct,
1308                               const struct nf_conntrack_tuple *newreply)
1309 {
1310         struct nf_conn_help *help = nfct_help(ct);
1311
1312         /* Should be unconfirmed, so not in hash table yet */
1313         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
1314
1315         pr_debug("Altering reply tuple of %p to ", ct);
1316         nf_ct_dump_tuple(newreply);
1317
1318         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1319         if (ct->master || (help && !hlist_empty(&help->expectations)))
1320                 return;
1321
1322         rcu_read_lock();
1323         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1324         rcu_read_unlock();
1325 }
1326 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1327
1328 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1329 void __nf_ct_refresh_acct(struct nf_conn *ct,
1330                           enum ip_conntrack_info ctinfo,
1331                           const struct sk_buff *skb,
1332                           unsigned long extra_jiffies,
1333                           int do_acct)
1334 {
1335         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
1336         NF_CT_ASSERT(skb);
1337
1338         /* Only update if this is not a fixed timeout */
1339         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1340                 goto acct;
1341
1342         /* If not in hash table, timer will not be active yet */
1343         if (!nf_ct_is_confirmed(ct)) {
1344                 ct->timeout.expires = extra_jiffies;
1345         } else {
1346                 unsigned long newtime = jiffies + extra_jiffies;
1347
1348                 /* Only update the timeout if the new timeout is at least
1349                    HZ jiffies from the old timeout. Need del_timer for race
1350                    avoidance (may already be dying). */
1351                 if (newtime - ct->timeout.expires >= HZ)
1352                         mod_timer_pending(&ct->timeout, newtime);
1353         }
1354
1355 acct:
1356         if (do_acct)
1357                 nf_ct_acct_update(ct, ctinfo, skb->len);
1358 }
1359 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1360
1361 bool __nf_ct_kill_acct(struct nf_conn *ct,
1362                        enum ip_conntrack_info ctinfo,
1363                        const struct sk_buff *skb,
1364                        int do_acct)
1365 {
1366         if (do_acct)
1367                 nf_ct_acct_update(ct, ctinfo, skb->len);
1368
1369         if (del_timer(&ct->timeout)) {
1370                 ct->timeout.function((unsigned long)ct);
1371                 return true;
1372         }
1373         return false;
1374 }
1375 EXPORT_SYMBOL_GPL(__nf_ct_kill_acct);
1376
1377 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1378
1379 #include <linux/netfilter/nfnetlink.h>
1380 #include <linux/netfilter/nfnetlink_conntrack.h>
1381 #include <linux/mutex.h>
1382
1383 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
1384  * in ip_conntrack_core, since we don't want the protocols to autoload
1385  * or depend on ctnetlink */
1386 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1387                                const struct nf_conntrack_tuple *tuple)
1388 {
1389         if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1390             nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1391                 goto nla_put_failure;
1392         return 0;
1393
1394 nla_put_failure:
1395         return -1;
1396 }
1397 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1398
1399 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1400         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
1401         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
1402 };
1403 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1404
1405 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1406                                struct nf_conntrack_tuple *t)
1407 {
1408         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
1409                 return -EINVAL;
1410
1411         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1412         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1413
1414         return 0;
1415 }
1416 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1417
1418 int nf_ct_port_nlattr_tuple_size(void)
1419 {
1420         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1421 }
1422 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1423 #endif
1424
1425 /* Used by ipt_REJECT and ip6t_REJECT. */
1426 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
1427 {
1428         struct nf_conn *ct;
1429         enum ip_conntrack_info ctinfo;
1430
1431         /* This ICMP is in reverse direction to the packet which caused it */
1432         ct = nf_ct_get(skb, &ctinfo);
1433         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1434                 ctinfo = IP_CT_RELATED_REPLY;
1435         else
1436                 ctinfo = IP_CT_RELATED;
1437
1438         /* Attach to new skbuff, and increment count */
1439         nskb->nfct = &ct->ct_general;
1440         nskb->nfctinfo = ctinfo;
1441         nf_conntrack_get(nskb->nfct);
1442 }
1443
1444 /* Bring out ya dead! */
1445 static struct nf_conn *
1446 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1447                 void *data, unsigned int *bucket)
1448 {
1449         struct nf_conntrack_tuple_hash *h;
1450         struct nf_conn *ct;
1451         struct hlist_nulls_node *n;
1452         int cpu;
1453         spinlock_t *lockp;
1454
1455         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1456                 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
1457                 local_bh_disable();
1458                 nf_conntrack_lock(lockp);
1459                 if (*bucket < nf_conntrack_htable_size) {
1460                         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
1461                                 if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1462                                         continue;
1463                                 ct = nf_ct_tuplehash_to_ctrack(h);
1464                                 if (net_eq(nf_ct_net(ct), net) &&
1465                                     iter(ct, data))
1466                                         goto found;
1467                         }
1468                 }
1469                 spin_unlock(lockp);
1470                 local_bh_enable();
1471                 cond_resched();
1472         }
1473
1474         for_each_possible_cpu(cpu) {
1475                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1476
1477                 spin_lock_bh(&pcpu->lock);
1478                 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
1479                         ct = nf_ct_tuplehash_to_ctrack(h);
1480                         if (iter(ct, data))
1481                                 set_bit(IPS_DYING_BIT, &ct->status);
1482                 }
1483                 spin_unlock_bh(&pcpu->lock);
1484                 cond_resched();
1485         }
1486         return NULL;
1487 found:
1488         atomic_inc(&ct->ct_general.use);
1489         spin_unlock(lockp);
1490         local_bh_enable();
1491         return ct;
1492 }
1493
1494 void nf_ct_iterate_cleanup(struct net *net,
1495                            int (*iter)(struct nf_conn *i, void *data),
1496                            void *data, u32 portid, int report)
1497 {
1498         struct nf_conn *ct;
1499         unsigned int bucket = 0;
1500
1501         might_sleep();
1502
1503         if (atomic_read(&net->ct.count) == 0)
1504                 return;
1505
1506         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1507                 /* Time to push up daises... */
1508                 if (del_timer(&ct->timeout))
1509                         nf_ct_delete(ct, portid, report);
1510
1511                 /* ... else the timer will get him soon. */
1512
1513                 nf_ct_put(ct);
1514                 cond_resched();
1515         }
1516 }
1517 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1518
1519 static int kill_all(struct nf_conn *i, void *data)
1520 {
1521         return 1;
1522 }
1523
1524 void nf_ct_free_hashtable(void *hash, unsigned int size)
1525 {
1526         if (is_vmalloc_addr(hash))
1527                 vfree(hash);
1528         else
1529                 free_pages((unsigned long)hash,
1530                            get_order(sizeof(struct hlist_head) * size));
1531 }
1532 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1533
1534 static int untrack_refs(void)
1535 {
1536         int cnt = 0, cpu;
1537
1538         for_each_possible_cpu(cpu) {
1539                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1540
1541                 cnt += atomic_read(&ct->ct_general.use) - 1;
1542         }
1543         return cnt;
1544 }
1545
1546 void nf_conntrack_cleanup_start(void)
1547 {
1548         RCU_INIT_POINTER(ip_ct_attach, NULL);
1549 }
1550
1551 void nf_conntrack_cleanup_end(void)
1552 {
1553         RCU_INIT_POINTER(nf_ct_destroy, NULL);
1554         while (untrack_refs() > 0)
1555                 schedule();
1556
1557         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1558
1559         nf_conntrack_proto_fini();
1560         nf_conntrack_seqadj_fini();
1561         nf_conntrack_labels_fini();
1562         nf_conntrack_helper_fini();
1563         nf_conntrack_timeout_fini();
1564         nf_conntrack_ecache_fini();
1565         nf_conntrack_tstamp_fini();
1566         nf_conntrack_acct_fini();
1567         nf_conntrack_expect_fini();
1568
1569         kmem_cache_destroy(nf_conntrack_cachep);
1570 }
1571
1572 /*
1573  * Mishearing the voices in his head, our hero wonders how he's
1574  * supposed to kill the mall.
1575  */
1576 void nf_conntrack_cleanup_net(struct net *net)
1577 {
1578         LIST_HEAD(single);
1579
1580         list_add(&net->exit_list, &single);
1581         nf_conntrack_cleanup_net_list(&single);
1582 }
1583
1584 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
1585 {
1586         int busy;
1587         struct net *net;
1588
1589         /*
1590          * This makes sure all current packets have passed through
1591          *  netfilter framework.  Roll on, two-stage module
1592          *  delete...
1593          */
1594         synchronize_net();
1595 i_see_dead_people:
1596         busy = 0;
1597         list_for_each_entry(net, net_exit_list, exit_list) {
1598                 nf_ct_iterate_cleanup(net, kill_all, NULL, 0, 0);
1599                 if (atomic_read(&net->ct.count) != 0)
1600                         busy = 1;
1601         }
1602         if (busy) {
1603                 schedule();
1604                 goto i_see_dead_people;
1605         }
1606
1607         list_for_each_entry(net, net_exit_list, exit_list) {
1608                 nf_conntrack_proto_pernet_fini(net);
1609                 nf_conntrack_helper_pernet_fini(net);
1610                 nf_conntrack_ecache_pernet_fini(net);
1611                 nf_conntrack_tstamp_pernet_fini(net);
1612                 nf_conntrack_acct_pernet_fini(net);
1613                 nf_conntrack_expect_pernet_fini(net);
1614                 free_percpu(net->ct.stat);
1615                 free_percpu(net->ct.pcpu_lists);
1616         }
1617 }
1618
1619 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
1620 {
1621         struct hlist_nulls_head *hash;
1622         unsigned int nr_slots, i;
1623         size_t sz;
1624
1625         if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1626                 return NULL;
1627
1628         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1629         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1630
1631         if (nr_slots > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1632                 return NULL;
1633
1634         sz = nr_slots * sizeof(struct hlist_nulls_head);
1635         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1636                                         get_order(sz));
1637         if (!hash)
1638                 hash = vzalloc(sz);
1639
1640         if (hash && nulls)
1641                 for (i = 0; i < nr_slots; i++)
1642                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1643
1644         return hash;
1645 }
1646 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1647
1648 int nf_conntrack_hash_resize(unsigned int hashsize)
1649 {
1650         int i, bucket;
1651         unsigned int old_size;
1652         struct hlist_nulls_head *hash, *old_hash;
1653         struct nf_conntrack_tuple_hash *h;
1654         struct nf_conn *ct;
1655
1656         if (!hashsize)
1657                 return -EINVAL;
1658
1659         hash = nf_ct_alloc_hashtable(&hashsize, 1);
1660         if (!hash)
1661                 return -ENOMEM;
1662
1663         old_size = nf_conntrack_htable_size;
1664         if (old_size == hashsize) {
1665                 nf_ct_free_hashtable(hash, hashsize);
1666                 return 0;
1667         }
1668
1669         local_bh_disable();
1670         nf_conntrack_all_lock();
1671         write_seqcount_begin(&nf_conntrack_generation);
1672
1673         /* Lookups in the old hash might happen in parallel, which means we
1674          * might get false negatives during connection lookup. New connections
1675          * created because of a false negative won't make it into the hash
1676          * though since that required taking the locks.
1677          */
1678
1679         for (i = 0; i < nf_conntrack_htable_size; i++) {
1680                 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
1681                         h = hlist_nulls_entry(nf_conntrack_hash[i].first,
1682                                               struct nf_conntrack_tuple_hash, hnnode);
1683                         ct = nf_ct_tuplehash_to_ctrack(h);
1684                         hlist_nulls_del_rcu(&h->hnnode);
1685                         bucket = __hash_conntrack(nf_ct_net(ct),
1686                                                   &h->tuple, hashsize);
1687                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1688                 }
1689         }
1690         old_size = nf_conntrack_htable_size;
1691         old_hash = nf_conntrack_hash;
1692
1693         nf_conntrack_hash = hash;
1694         nf_conntrack_htable_size = hashsize;
1695
1696         write_seqcount_end(&nf_conntrack_generation);
1697         nf_conntrack_all_unlock();
1698         local_bh_enable();
1699
1700         synchronize_net();
1701         nf_ct_free_hashtable(old_hash, old_size);
1702         return 0;
1703 }
1704
1705 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1706 {
1707         unsigned int hashsize;
1708         int rc;
1709
1710         if (current->nsproxy->net_ns != &init_net)
1711                 return -EOPNOTSUPP;
1712
1713         /* On boot, we can set this without any fancy locking. */
1714         if (!nf_conntrack_htable_size)
1715                 return param_set_uint(val, kp);
1716
1717         rc = kstrtouint(val, 0, &hashsize);
1718         if (rc)
1719                 return rc;
1720
1721         return nf_conntrack_hash_resize(hashsize);
1722 }
1723 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1724
1725 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1726                   &nf_conntrack_htable_size, 0600);
1727
1728 void nf_ct_untracked_status_or(unsigned long bits)
1729 {
1730         int cpu;
1731
1732         for_each_possible_cpu(cpu)
1733                 per_cpu(nf_conntrack_untracked, cpu).status |= bits;
1734 }
1735 EXPORT_SYMBOL_GPL(nf_ct_untracked_status_or);
1736
1737 int nf_conntrack_init_start(void)
1738 {
1739         int max_factor = 8;
1740         int ret = -ENOMEM;
1741         int i, cpu;
1742
1743         seqcount_init(&nf_conntrack_generation);
1744
1745         for (i = 0; i < CONNTRACK_LOCKS; i++)
1746                 spin_lock_init(&nf_conntrack_locks[i]);
1747
1748         if (!nf_conntrack_htable_size) {
1749                 /* Idea from tcp.c: use 1/16384 of memory.
1750                  * On i386: 32MB machine has 512 buckets.
1751                  * >= 1GB machines have 16384 buckets.
1752                  * >= 4GB machines have 65536 buckets.
1753                  */
1754                 nf_conntrack_htable_size
1755                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1756                            / sizeof(struct hlist_head));
1757                 if (totalram_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
1758                         nf_conntrack_htable_size = 65536;
1759                 else if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1760                         nf_conntrack_htable_size = 16384;
1761                 if (nf_conntrack_htable_size < 32)
1762                         nf_conntrack_htable_size = 32;
1763
1764                 /* Use a max. factor of four by default to get the same max as
1765                  * with the old struct list_heads. When a table size is given
1766                  * we use the old value of 8 to avoid reducing the max.
1767                  * entries. */
1768                 max_factor = 4;
1769         }
1770
1771         nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
1772         if (!nf_conntrack_hash)
1773                 return -ENOMEM;
1774
1775         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1776
1777         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
1778                                                 sizeof(struct nf_conn), 0,
1779                                                 SLAB_DESTROY_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
1780         if (!nf_conntrack_cachep)
1781                 goto err_cachep;
1782
1783         printk(KERN_INFO "nf_conntrack version %s (%u buckets, %d max)\n",
1784                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1785                nf_conntrack_max);
1786
1787         ret = nf_conntrack_expect_init();
1788         if (ret < 0)
1789                 goto err_expect;
1790
1791         ret = nf_conntrack_acct_init();
1792         if (ret < 0)
1793                 goto err_acct;
1794
1795         ret = nf_conntrack_tstamp_init();
1796         if (ret < 0)
1797                 goto err_tstamp;
1798
1799         ret = nf_conntrack_ecache_init();
1800         if (ret < 0)
1801                 goto err_ecache;
1802
1803         ret = nf_conntrack_timeout_init();
1804         if (ret < 0)
1805                 goto err_timeout;
1806
1807         ret = nf_conntrack_helper_init();
1808         if (ret < 0)
1809                 goto err_helper;
1810
1811         ret = nf_conntrack_labels_init();
1812         if (ret < 0)
1813                 goto err_labels;
1814
1815         ret = nf_conntrack_seqadj_init();
1816         if (ret < 0)
1817                 goto err_seqadj;
1818
1819         ret = nf_conntrack_proto_init();
1820         if (ret < 0)
1821                 goto err_proto;
1822
1823         /* Set up fake conntrack: to never be deleted, not in any hashes */
1824         for_each_possible_cpu(cpu) {
1825                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1826                 write_pnet(&ct->ct_net, &init_net);
1827                 atomic_set(&ct->ct_general.use, 1);
1828         }
1829         /*  - and look it like as a confirmed connection */
1830         nf_ct_untracked_status_or(IPS_CONFIRMED | IPS_UNTRACKED);
1831         return 0;
1832
1833 err_proto:
1834         nf_conntrack_seqadj_fini();
1835 err_seqadj:
1836         nf_conntrack_labels_fini();
1837 err_labels:
1838         nf_conntrack_helper_fini();
1839 err_helper:
1840         nf_conntrack_timeout_fini();
1841 err_timeout:
1842         nf_conntrack_ecache_fini();
1843 err_ecache:
1844         nf_conntrack_tstamp_fini();
1845 err_tstamp:
1846         nf_conntrack_acct_fini();
1847 err_acct:
1848         nf_conntrack_expect_fini();
1849 err_expect:
1850         kmem_cache_destroy(nf_conntrack_cachep);
1851 err_cachep:
1852         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1853         return ret;
1854 }
1855
1856 void nf_conntrack_init_end(void)
1857 {
1858         /* For use by REJECT target */
1859         RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
1860         RCU_INIT_POINTER(nf_ct_destroy, destroy_conntrack);
1861 }
1862
1863 /*
1864  * We need to use special "null" values, not used in hash table
1865  */
1866 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1867 #define DYING_NULLS_VAL         ((1<<30)+1)
1868 #define TEMPLATE_NULLS_VAL      ((1<<30)+2)
1869
1870 int nf_conntrack_init_net(struct net *net)
1871 {
1872         int ret = -ENOMEM;
1873         int cpu;
1874
1875         atomic_set(&net->ct.count, 0);
1876
1877         net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
1878         if (!net->ct.pcpu_lists)
1879                 goto err_stat;
1880
1881         for_each_possible_cpu(cpu) {
1882                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1883
1884                 spin_lock_init(&pcpu->lock);
1885                 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
1886                 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
1887         }
1888
1889         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
1890         if (!net->ct.stat)
1891                 goto err_pcpu_lists;
1892
1893         ret = nf_conntrack_expect_pernet_init(net);
1894         if (ret < 0)
1895                 goto err_expect;
1896         ret = nf_conntrack_acct_pernet_init(net);
1897         if (ret < 0)
1898                 goto err_acct;
1899         ret = nf_conntrack_tstamp_pernet_init(net);
1900         if (ret < 0)
1901                 goto err_tstamp;
1902         ret = nf_conntrack_ecache_pernet_init(net);
1903         if (ret < 0)
1904                 goto err_ecache;
1905         ret = nf_conntrack_helper_pernet_init(net);
1906         if (ret < 0)
1907                 goto err_helper;
1908         ret = nf_conntrack_proto_pernet_init(net);
1909         if (ret < 0)
1910                 goto err_proto;
1911         return 0;
1912
1913 err_proto:
1914         nf_conntrack_helper_pernet_fini(net);
1915 err_helper:
1916         nf_conntrack_ecache_pernet_fini(net);
1917 err_ecache:
1918         nf_conntrack_tstamp_pernet_fini(net);
1919 err_tstamp:
1920         nf_conntrack_acct_pernet_fini(net);
1921 err_acct:
1922         nf_conntrack_expect_pernet_fini(net);
1923 err_expect:
1924         free_percpu(net->ct.stat);
1925 err_pcpu_lists:
1926         free_percpu(net->ct.pcpu_lists);
1927 err_stat:
1928         return ret;
1929 }