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