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 struct conntrack_gc_work {
  76         struct delayed_work     dwork;
  77         u32                     last_bucket;
  78         bool                    exiting;
  79         long                    next_gc_run;
  80 };
  81
  82 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
  83 static __read_mostly spinlock_t nf_conntrack_locks_all_lock;
  84 static __read_mostly DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
  85 static __read_mostly bool nf_conntrack_locks_all;
  86
  87 /* every gc cycle scans at most 1/GC_MAX_BUCKETS_DIV part of table */
  88 #define GC_MAX_BUCKETS_DIV      128u
  89 /* upper bound of full table scan */
  90 #define GC_MAX_SCAN_JIFFIES     (16u * HZ)
  91 /* desired ratio of entries found to be expired */
  92 #define GC_EVICT_RATIO  50u
  93
  94 static struct conntrack_gc_work conntrack_gc_work;
  95
  96 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
  97 {
  98         spin_lock(lock);
  99         while (unlikely(nf_conntrack_locks_all)) {
 100                 spin_unlock(lock);
 101
 102                 /*
 103                  * Order the 'nf_conntrack_locks_all' load vs. the
 104                  * spin_unlock_wait() loads below, to ensure
 105                  * that 'nf_conntrack_locks_all_lock' is indeed held:
 106                  */
 107                 smp_rmb(); /* spin_lock(&nf_conntrack_locks_all_lock) */
 108                 spin_unlock_wait(&nf_conntrack_locks_all_lock);
 109                 spin_lock(lock);
 110         }
 111 }
 112 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
 113
 114 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 115 {
 116         h1 %= CONNTRACK_LOCKS;
 117         h2 %= CONNTRACK_LOCKS;
 118         spin_unlock(&nf_conntrack_locks[h1]);
 119         if (h1 != h2)
 120                 spin_unlock(&nf_conntrack_locks[h2]);
 121 }
 122
 123 /* return true if we need to recompute hashes (in case hash table was resized) */
 124 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
 125                                      unsigned int h2, unsigned int sequence)
 126 {
 127         h1 %= CONNTRACK_LOCKS;
 128         h2 %= CONNTRACK_LOCKS;
 129         if (h1 <= h2) {
 130                 nf_conntrack_lock(&nf_conntrack_locks[h1]);
 131                 if (h1 != h2)
 132                         spin_lock_nested(&nf_conntrack_locks[h2],
 133                                          SINGLE_DEPTH_NESTING);
 134         } else {
 135                 nf_conntrack_lock(&nf_conntrack_locks[h2]);
 136                 spin_lock_nested(&nf_conntrack_locks[h1],
 137                                  SINGLE_DEPTH_NESTING);
 138         }
 139         if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
 140                 nf_conntrack_double_unlock(h1, h2);
 141                 return true;
 142         }
 143         return false;
 144 }
 145
 146 static void nf_conntrack_all_lock(void)
 147 {
 148         int i;
 149
 150         spin_lock(&nf_conntrack_locks_all_lock);
 151         nf_conntrack_locks_all = true;
 152
 153         /*
 154          * Order the above store of 'nf_conntrack_locks_all' against
 155          * the spin_unlock_wait() loads below, such that if
 156          * nf_conntrack_lock() observes 'nf_conntrack_locks_all'
 157          * we must observe nf_conntrack_locks[] held:
 158          */
 159         smp_mb(); /* spin_lock(&nf_conntrack_locks_all_lock) */
 160
 161         for (i = 0; i < CONNTRACK_LOCKS; i++) {
 162                 spin_unlock_wait(&nf_conntrack_locks[i]);
 163         }
 164 }
 165
 166 static void nf_conntrack_all_unlock(void)
 167 {
 168         /*
 169          * All prior stores must be complete before we clear
 170          * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
 171          * might observe the false value but not the entire
 172          * critical section:
 173          */
 174         smp_store_release(&nf_conntrack_locks_all, false);
 175         spin_unlock(&nf_conntrack_locks_all_lock);
 176 }
 177
 178 unsigned int nf_conntrack_htable_size __read_mostly;
 179 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
 180
 181 unsigned int nf_conntrack_max __read_mostly;
 182 seqcount_t nf_conntrack_generation __read_mostly;
 183
 184 DEFINE_PER_CPU(struct nf_conn, nf_conntrack_untracked);
 185 EXPORT_PER_CPU_SYMBOL(nf_conntrack_untracked);
 186
 187 static unsigned int nf_conntrack_hash_rnd __read_mostly;
 188
 189 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
 190                               const struct net *net)
 191 {
 192         unsigned int n;
 193         u32 seed;
 194
 195         get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
 196
 197         /* The direction must be ignored, so we hash everything up to the
 198          * destination ports (which is a multiple of 4) and treat the last
 199          * three bytes manually.
 200          */
 201         seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
 202         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
 203         return jhash2((u32 *)tuple, n, seed ^
 204                       (((__force __u16)tuple->dst.u.all << 16) |
 205                       tuple->dst.protonum));
 206 }
 207
 208 static u32 scale_hash(u32 hash)
 209 {
 210         return reciprocal_scale(hash, nf_conntrack_htable_size);
 211 }
 212
 213 static u32 __hash_conntrack(const struct net *net,
 214                             const struct nf_conntrack_tuple *tuple,
 215                             unsigned int size)
 216 {
 217         return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
 218 }
 219
 220 static u32 hash_conntrack(const struct net *net,
 221                           const struct nf_conntrack_tuple *tuple)
 222 {
 223         return scale_hash(hash_conntrack_raw(tuple, net));
 224 }
 225
 226 bool
 227 nf_ct_get_tuple(const struct sk_buff *skb,
 228                 unsigned int nhoff,
 229                 unsigned int dataoff,
 230                 u_int16_t l3num,
 231                 u_int8_t protonum,
 232                 struct net *net,
 233                 struct nf_conntrack_tuple *tuple,
 234                 const struct nf_conntrack_l3proto *l3proto,
 235                 const struct nf_conntrack_l4proto *l4proto)
 236 {
 237         memset(tuple, 0, sizeof(*tuple));
 238
 239         tuple->src.l3num = l3num;
 240         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 241                 return false;
 242
 243         tuple->dst.protonum = protonum;
 244         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 245
 246         return l4proto->pkt_to_tuple(skb, dataoff, net, tuple);
 247 }
 248 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 249
 250 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 251                        u_int16_t l3num,
 252                        struct net *net, struct nf_conntrack_tuple *tuple)
 253 {
 254         struct nf_conntrack_l3proto *l3proto;
 255         struct nf_conntrack_l4proto *l4proto;
 256         unsigned int protoff;
 257         u_int8_t protonum;
 258         int ret;
 259
 260         rcu_read_lock();
 261
 262         l3proto = __nf_ct_l3proto_find(l3num);
 263         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
 264         if (ret != NF_ACCEPT) {
 265                 rcu_read_unlock();
 266                 return false;
 267         }
 268
 269         l4proto = __nf_ct_l4proto_find(l3num, protonum);
 270
 271         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple,
 272                               l3proto, l4proto);
 273
 274         rcu_read_unlock();
 275         return ret;
 276 }
 277 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 278
 279 bool
 280 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 281                    const struct nf_conntrack_tuple *orig,
 282                    const struct nf_conntrack_l3proto *l3proto,
 283                    const struct nf_conntrack_l4proto *l4proto)
 284 {
 285         memset(inverse, 0, sizeof(*inverse));
 286
 287         inverse->src.l3num = orig->src.l3num;
 288         if (l3proto->invert_tuple(inverse, orig) == 0)
 289                 return false;
 290
 291         inverse->dst.dir = !orig->dst.dir;
 292
 293         inverse->dst.protonum = orig->dst.protonum;
 294         return l4proto->invert_tuple(inverse, orig);
 295 }
 296 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 297
 298 static void
 299 clean_from_lists(struct nf_conn *ct)
 300 {
 301         pr_debug("clean_from_lists(%p)\n", ct);
 302         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 303         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 304
 305         /* Destroy all pending expectations */
 306         nf_ct_remove_expectations(ct);
 307 }
 308
 309 /* must be called with local_bh_disable */
 310 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
 311 {
 312         struct ct_pcpu *pcpu;
 313
 314         /* add this conntrack to the (per cpu) dying 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->dying);
 321         spin_unlock(&pcpu->lock);
 322 }
 323
 324 /* must be called with local_bh_disable */
 325 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
 326 {
 327         struct ct_pcpu *pcpu;
 328
 329         /* add this conntrack to the (per cpu) unconfirmed list */
 330         ct->cpu = smp_processor_id();
 331         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 332
 333         spin_lock(&pcpu->lock);
 334         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 335                              &pcpu->unconfirmed);
 336         spin_unlock(&pcpu->lock);
 337 }
 338
 339 /* must be called with local_bh_disable */
 340 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
 341 {
 342         struct ct_pcpu *pcpu;
 343
 344         /* We overload first tuple to link into unconfirmed or dying list.*/
 345         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 346
 347         spin_lock(&pcpu->lock);
 348         BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 349         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 350         spin_unlock(&pcpu->lock);
 351 }
 352
 353 /* Released via destroy_conntrack() */
 354 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
 355                                  const struct nf_conntrack_zone *zone,
 356                                  gfp_t flags)
 357 {
 358         struct nf_conn *tmpl;
 359
 360         tmpl = kzalloc(sizeof(*tmpl), flags);
 361         if (tmpl == NULL)
 362                 return NULL;
 363
 364         tmpl->status = IPS_TEMPLATE;
 365         write_pnet(&tmpl->ct_net, net);
 366         nf_ct_zone_add(tmpl, zone);
 367         atomic_set(&tmpl->ct_general.use, 0);
 368
 369         return tmpl;
 370 }
 371 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
 372
 373 void nf_ct_tmpl_free(struct nf_conn *tmpl)
 374 {
 375         nf_ct_ext_destroy(tmpl);
 376         nf_ct_ext_free(tmpl);
 377         kfree(tmpl);
 378 }
 379 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
 380
 381 static void
 382 destroy_conntrack(struct nf_conntrack *nfct)
 383 {
 384         struct nf_conn *ct = (struct nf_conn *)nfct;
 385         struct nf_conntrack_l4proto *l4proto;
 386
 387         pr_debug("destroy_conntrack(%p)\n", ct);
 388         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 389
 390         if (unlikely(nf_ct_is_template(ct))) {
 391                 nf_ct_tmpl_free(ct);
 392                 return;
 393         }
 394         rcu_read_lock();
 395         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 396         if (l4proto->destroy)
 397                 l4proto->destroy(ct);
 398
 399         rcu_read_unlock();
 400
 401         local_bh_disable();
 402         /* Expectations will have been removed in clean_from_lists,
 403          * except TFTP can create an expectation on the first packet,
 404          * before connection is in the list, so we need to clean here,
 405          * too.
 406          */
 407         nf_ct_remove_expectations(ct);
 408
 409         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 410
 411         local_bh_enable();
 412
 413         if (ct->master)
 414                 nf_ct_put(ct->master);
 415
 416         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 417         nf_conntrack_free(ct);
 418 }
 419
 420 static void nf_ct_delete_from_lists(struct nf_conn *ct)
 421 {
 422         struct net *net = nf_ct_net(ct);
 423         unsigned int hash, reply_hash;
 424         unsigned int sequence;
 425
 426         nf_ct_helper_destroy(ct);
 427
 428         local_bh_disable();
 429         do {
 430                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 431                 hash = hash_conntrack(net,
 432                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 433                 reply_hash = hash_conntrack(net,
 434                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 435         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 436
 437         clean_from_lists(ct);
 438         nf_conntrack_double_unlock(hash, reply_hash);
 439
 440         nf_ct_add_to_dying_list(ct);
 441
 442         local_bh_enable();
 443 }
 444
 445 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
 446 {
 447         struct nf_conn_tstamp *tstamp;
 448
 449         if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
 450                 return false;
 451
 452         tstamp = nf_conn_tstamp_find(ct);
 453         if (tstamp && tstamp->stop == 0)
 454                 tstamp->stop = ktime_get_real_ns();
 455
 456         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 457                                     portid, report) < 0) {
 458                 /* destroy event was not delivered. nf_ct_put will
 459                  * be done by event cache worker on redelivery.
 460                  */
 461                 nf_ct_delete_from_lists(ct);
 462                 nf_conntrack_ecache_delayed_work(nf_ct_net(ct));
 463                 return false;
 464         }
 465
 466         nf_conntrack_ecache_work(nf_ct_net(ct));
 467         nf_ct_delete_from_lists(ct);
 468         nf_ct_put(ct);
 469         return true;
 470 }
 471 EXPORT_SYMBOL_GPL(nf_ct_delete);
 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 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
 491 static void nf_ct_gc_expired(struct nf_conn *ct)
 492 {
 493         if (!atomic_inc_not_zero(&ct->ct_general.use))
 494                 return;
 495
 496         if (nf_ct_should_gc(ct))
 497                 nf_ct_kill(ct);
 498
 499         nf_ct_put(ct);
 500 }
 501
 502 /*
 503  * Warning :
 504  * - Caller must take a reference on returned object
 505  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 506  */
 507 static struct nf_conntrack_tuple_hash *
 508 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
 509                       const struct nf_conntrack_tuple *tuple, u32 hash)
 510 {
 511         struct nf_conntrack_tuple_hash *h;
 512         struct hlist_nulls_head *ct_hash;
 513         struct hlist_nulls_node *n;
 514         unsigned int bucket, hsize;
 515
 516 begin:
 517         nf_conntrack_get_ht(&ct_hash, &hsize);
 518         bucket = reciprocal_scale(hash, hsize);
 519
 520         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
 521                 struct nf_conn *ct;
 522
 523                 ct = nf_ct_tuplehash_to_ctrack(h);
 524                 if (nf_ct_is_expired(ct)) {
 525                         nf_ct_gc_expired(ct);
 526                         continue;
 527                 }
 528
 529                 if (nf_ct_is_dying(ct))
 530                         continue;
 531
 532                 if (nf_ct_key_equal(h, tuple, zone, net))
 533                         return h;
 534         }
 535         /*
 536          * if the nulls value we got at the end of this lookup is
 537          * not the expected one, we must restart lookup.
 538          * We probably met an item that was moved to another chain.
 539          */
 540         if (get_nulls_value(n) != bucket) {
 541                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 542                 goto begin;
 543         }
 544
 545         return NULL;
 546 }
 547
 548 /* Find a connection corresponding to a tuple. */
 549 static struct nf_conntrack_tuple_hash *
 550 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 551                         const struct nf_conntrack_tuple *tuple, u32 hash)
 552 {
 553         struct nf_conntrack_tuple_hash *h;
 554         struct nf_conn *ct;
 555
 556         rcu_read_lock();
 557 begin:
 558         h = ____nf_conntrack_find(net, zone, tuple, hash);
 559         if (h) {
 560                 ct = nf_ct_tuplehash_to_ctrack(h);
 561                 if (unlikely(nf_ct_is_dying(ct) ||
 562                              !atomic_inc_not_zero(&ct->ct_general.use)))
 563                         h = NULL;
 564                 else {
 565                         if (unlikely(!nf_ct_key_equal(h, tuple, zone, net))) {
 566                                 nf_ct_put(ct);
 567                                 goto begin;
 568                         }
 569                 }
 570         }
 571         rcu_read_unlock();
 572
 573         return h;
 574 }
 575
 576 struct nf_conntrack_tuple_hash *
 577 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 578                       const struct nf_conntrack_tuple *tuple)
 579 {
 580         return __nf_conntrack_find_get(net, zone, tuple,
 581                                        hash_conntrack_raw(tuple, net));
 582 }
 583 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 584
 585 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 586                                        unsigned int hash,
 587                                        unsigned int reply_hash)
 588 {
 589         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 590                            &nf_conntrack_hash[hash]);
 591         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 592                            &nf_conntrack_hash[reply_hash]);
 593 }
 594
 595 int
 596 nf_conntrack_hash_check_insert(struct nf_conn *ct)
 597 {
 598         const struct nf_conntrack_zone *zone;
 599         struct net *net = nf_ct_net(ct);
 600         unsigned int hash, reply_hash;
 601         struct nf_conntrack_tuple_hash *h;
 602         struct hlist_nulls_node *n;
 603         unsigned int sequence;
 604
 605         zone = nf_ct_zone(ct);
 606
 607         local_bh_disable();
 608         do {
 609                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 610                 hash = hash_conntrack(net,
 611                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 612                 reply_hash = hash_conntrack(net,
 613                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 614         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 615
 616         /* See if there's one in the list already, including reverse */
 617         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 618                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 619                                     zone, net))
 620                         goto out;
 621
 622         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 623                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 624                                     zone, net))
 625                         goto out;
 626
 627         smp_wmb();
 628         /* The caller holds a reference to this object */
 629         atomic_set(&ct->ct_general.use, 2);
 630         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 631         nf_conntrack_double_unlock(hash, reply_hash);
 632         NF_CT_STAT_INC(net, insert);
 633         local_bh_enable();
 634         return 0;
 635
 636 out:
 637         nf_conntrack_double_unlock(hash, reply_hash);
 638         NF_CT_STAT_INC(net, insert_failed);
 639         local_bh_enable();
 640         return -EEXIST;
 641 }
 642 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 643
 644 static inline void nf_ct_acct_update(struct nf_conn *ct,
 645                                      enum ip_conntrack_info ctinfo,
 646                                      unsigned int len)
 647 {
 648         struct nf_conn_acct *acct;
 649
 650         acct = nf_conn_acct_find(ct);
 651         if (acct) {
 652                 struct nf_conn_counter *counter = acct->counter;
 653
 654                 atomic64_inc(&counter[CTINFO2DIR(ctinfo)].packets);
 655                 atomic64_add(len, &counter[CTINFO2DIR(ctinfo)].bytes);
 656         }
 657 }
 658
 659 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
 660                              const struct nf_conn *loser_ct)
 661 {
 662         struct nf_conn_acct *acct;
 663
 664         acct = nf_conn_acct_find(loser_ct);
 665         if (acct) {
 666                 struct nf_conn_counter *counter = acct->counter;
 667                 unsigned int bytes;
 668
 669                 /* u32 should be fine since we must have seen one packet. */
 670                 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
 671                 nf_ct_acct_update(ct, ctinfo, bytes);
 672         }
 673 }
 674
 675 /* Resolve race on insertion if this protocol allows this. */
 676 static int nf_ct_resolve_clash(struct net *net, struct sk_buff *skb,
 677                                enum ip_conntrack_info ctinfo,
 678                                struct nf_conntrack_tuple_hash *h)
 679 {
 680         /* This is the conntrack entry already in hashes that won race. */
 681         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 682         struct nf_conntrack_l4proto *l4proto;
 683
 684         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 685         if (l4proto->allow_clash &&
 686             !nfct_nat(ct) &&
 687             !nf_ct_is_dying(ct) &&
 688             atomic_inc_not_zero(&ct->ct_general.use)) {
 689                 nf_ct_acct_merge(ct, ctinfo, (struct nf_conn *)skb->nfct);
 690                 nf_conntrack_put(skb->nfct);
 691                 /* Assign conntrack already in hashes to this skbuff. Don't
 692                  * modify skb->nfctinfo to ensure consistent stateful filtering.
 693                  */
 694                 skb->nfct = &ct->ct_general;
 695                 return NF_ACCEPT;
 696         }
 697         NF_CT_STAT_INC(net, drop);
 698         return NF_DROP;
 699 }
 700
 701 /* Confirm a connection given skb; places it in hash table */
 702 int
 703 __nf_conntrack_confirm(struct sk_buff *skb)
 704 {
 705         const struct nf_conntrack_zone *zone;
 706         unsigned int hash, reply_hash;
 707         struct nf_conntrack_tuple_hash *h;
 708         struct nf_conn *ct;
 709         struct nf_conn_help *help;
 710         struct nf_conn_tstamp *tstamp;
 711         struct hlist_nulls_node *n;
 712         enum ip_conntrack_info ctinfo;
 713         struct net *net;
 714         unsigned int sequence;
 715         int ret = NF_DROP;
 716
 717         ct = nf_ct_get(skb, &ctinfo);
 718         net = nf_ct_net(ct);
 719
 720         /* ipt_REJECT uses nf_conntrack_attach to attach related
 721            ICMP/TCP RST packets in other direction.  Actual packet
 722            which created connection will be IP_CT_NEW or for an
 723            expected connection, IP_CT_RELATED. */
 724         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 725                 return NF_ACCEPT;
 726
 727         zone = nf_ct_zone(ct);
 728         local_bh_disable();
 729
 730         do {
 731                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 732                 /* reuse the hash saved before */
 733                 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
 734                 hash = scale_hash(hash);
 735                 reply_hash = hash_conntrack(net,
 736                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 737
 738         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 739
 740         /* We're not in hash table, and we refuse to set up related
 741          * connections for unconfirmed conns.  But packet copies and
 742          * REJECT will give spurious warnings here.
 743          */
 744         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 745
 746         /* No external references means no one else could have
 747          * confirmed us.
 748          */
 749         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 750         pr_debug("Confirming conntrack %p\n", ct);
 751         /* We have to check the DYING flag after unlink to prevent
 752          * a race against nf_ct_get_next_corpse() possibly called from
 753          * user context, else we insert an already 'dead' hash, blocking
 754          * further use of that particular connection -JM.
 755          */
 756         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 757
 758         if (unlikely(nf_ct_is_dying(ct))) {
 759                 nf_ct_add_to_dying_list(ct);
 760                 goto dying;
 761         }
 762
 763         /* See if there's one in the list already, including reverse:
 764            NAT could have grabbed it without realizing, since we're
 765            not in the hash.  If there is, we lost race. */
 766         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 767                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 768                                     zone, net))
 769                         goto out;
 770
 771         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 772                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 773                                     zone, net))
 774                         goto out;
 775
 776         /* Timer relative to confirmation time, not original
 777            setting time, otherwise we'd get timer wrap in
 778            weird delay cases. */
 779         ct->timeout += nfct_time_stamp;
 780         atomic_inc(&ct->ct_general.use);
 781         ct->status |= IPS_CONFIRMED;
 782
 783         /* set conntrack timestamp, if enabled. */
 784         tstamp = nf_conn_tstamp_find(ct);
 785         if (tstamp) {
 786                 if (skb->tstamp == 0)
 787                         __net_timestamp(skb);
 788
 789                 tstamp->start = ktime_to_ns(skb->tstamp);
 790         }
 791         /* Since the lookup is lockless, hash insertion must be done after
 792          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 793          * guarantee that no other CPU can find the conntrack before the above
 794          * stores are visible.
 795          */
 796         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 797         nf_conntrack_double_unlock(hash, reply_hash);
 798         local_bh_enable();
 799
 800         help = nfct_help(ct);
 801         if (help && help->helper)
 802                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 803
 804         nf_conntrack_event_cache(master_ct(ct) ?
 805                                  IPCT_RELATED : IPCT_NEW, ct);
 806         return NF_ACCEPT;
 807
 808 out:
 809         nf_ct_add_to_dying_list(ct);
 810         ret = nf_ct_resolve_clash(net, skb, ctinfo, h);
 811 dying:
 812         nf_conntrack_double_unlock(hash, reply_hash);
 813         NF_CT_STAT_INC(net, insert_failed);
 814         local_bh_enable();
 815         return ret;
 816 }
 817 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 818
 819 /* Returns true if a connection correspondings to the tuple (required
 820    for NAT). */
 821 int
 822 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 823                          const struct nf_conn *ignored_conntrack)
 824 {
 825         struct net *net = nf_ct_net(ignored_conntrack);
 826         const struct nf_conntrack_zone *zone;
 827         struct nf_conntrack_tuple_hash *h;
 828         struct hlist_nulls_head *ct_hash;
 829         unsigned int hash, hsize;
 830         struct hlist_nulls_node *n;
 831         struct nf_conn *ct;
 832
 833         zone = nf_ct_zone(ignored_conntrack);
 834
 835         rcu_read_lock();
 836  begin:
 837         nf_conntrack_get_ht(&ct_hash, &hsize);
 838         hash = __hash_conntrack(net, tuple, hsize);
 839
 840         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
 841                 ct = nf_ct_tuplehash_to_ctrack(h);
 842
 843                 if (ct == ignored_conntrack)
 844                         continue;
 845
 846                 if (nf_ct_is_expired(ct)) {
 847                         nf_ct_gc_expired(ct);
 848                         continue;
 849                 }
 850
 851                 if (nf_ct_key_equal(h, tuple, zone, net)) {
 852                         NF_CT_STAT_INC_ATOMIC(net, found);
 853                         rcu_read_unlock();
 854                         return 1;
 855                 }
 856         }
 857
 858         if (get_nulls_value(n) != hash) {
 859                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 860                 goto begin;
 861         }
 862
 863         rcu_read_unlock();
 864
 865         return 0;
 866 }
 867 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 868
 869 #define NF_CT_EVICTION_RANGE    8
 870
 871 /* There's a small race here where we may free a just-assured
 872    connection.  Too bad: we're in trouble anyway. */
 873 static unsigned int early_drop_list(struct net *net,
 874                                     struct hlist_nulls_head *head)
 875 {
 876         struct nf_conntrack_tuple_hash *h;
 877         struct hlist_nulls_node *n;
 878         unsigned int drops = 0;
 879         struct nf_conn *tmp;
 880
 881         hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
 882                 tmp = nf_ct_tuplehash_to_ctrack(h);
 883
 884                 if (nf_ct_is_expired(tmp)) {
 885                         nf_ct_gc_expired(tmp);
 886                         continue;
 887                 }
 888
 889                 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
 890                     !net_eq(nf_ct_net(tmp), net) ||
 891                     nf_ct_is_dying(tmp))
 892                         continue;
 893
 894                 if (!atomic_inc_not_zero(&tmp->ct_general.use))
 895                         continue;
 896
 897                 /* kill only if still in same netns -- might have moved due to
 898                  * SLAB_DESTROY_BY_RCU rules.
 899                  *
 900                  * We steal the timer reference.  If that fails timer has
 901                  * already fired or someone else deleted it. Just drop ref
 902                  * and move to next entry.
 903                  */
 904                 if (net_eq(nf_ct_net(tmp), net) &&
 905                     nf_ct_is_confirmed(tmp) &&
 906                     nf_ct_delete(tmp, 0, 0))
 907                         drops++;
 908
 909                 nf_ct_put(tmp);
 910         }
 911
 912         return drops;
 913 }
 914
 915 static noinline int early_drop(struct net *net, unsigned int _hash)
 916 {
 917         unsigned int i;
 918
 919         for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
 920                 struct hlist_nulls_head *ct_hash;
 921                 unsigned int hash, hsize, drops;
 922
 923                 rcu_read_lock();
 924                 nf_conntrack_get_ht(&ct_hash, &hsize);
 925                 hash = reciprocal_scale(_hash++, hsize);
 926
 927                 drops = early_drop_list(net, &ct_hash[hash]);
 928                 rcu_read_unlock();
 929
 930                 if (drops) {
 931                         NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
 932                         return true;
 933                 }
 934         }
 935
 936         return false;
 937 }
 938
 939 static void gc_worker(struct work_struct *work)
 940 {
 941         unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
 942         unsigned int i, goal, buckets = 0, expired_count = 0;
 943         struct conntrack_gc_work *gc_work;
 944         unsigned int ratio, scanned = 0;
 945         unsigned long next_run;
 946
 947         gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
 948
 949         goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
 950         i = gc_work->last_bucket;
 951
 952         do {
 953                 struct nf_conntrack_tuple_hash *h;
 954                 struct hlist_nulls_head *ct_hash;
 955                 struct hlist_nulls_node *n;
 956                 unsigned int hashsz;
 957                 struct nf_conn *tmp;
 958
 959                 i++;
 960                 rcu_read_lock();
 961
 962                 nf_conntrack_get_ht(&ct_hash, &hashsz);
 963                 if (i >= hashsz)
 964                         i = 0;
 965
 966                 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
 967                         tmp = nf_ct_tuplehash_to_ctrack(h);
 968
 969                         scanned++;
 970                         if (nf_ct_is_expired(tmp)) {
 971                                 nf_ct_gc_expired(tmp);
 972                                 expired_count++;
 973                                 continue;
 974                         }
 975                 }
 976
 977                 /* could check get_nulls_value() here and restart if ct
 978                  * was moved to another chain.  But given gc is best-effort
 979                  * we will just continue with next hash slot.
 980                  */
 981                 rcu_read_unlock();
 982                 cond_resched_rcu_qs();
 983         } while (++buckets < goal);
 984
 985         if (gc_work->exiting)
 986                 return;
 987
 988         /*
 989          * Eviction will normally happen from the packet path, and not
 990          * from this gc worker.
 991          *
 992          * This worker is only here to reap expired entries when system went
 993          * idle after a busy period.
 994          *
 995          * The heuristics below are supposed to balance conflicting goals:
 996          *
 997          * 1. Minimize time until we notice a stale entry
 998          * 2. Maximize scan intervals to not waste cycles
 999          *
1000          * Normally, expire ratio will be close to 0.
1001          *
1002          * As soon as a sizeable fraction of the entries have expired
1003          * increase scan frequency.
1004          */
1005         ratio = scanned ? expired_count * 100 / scanned : 0;
1006         if (ratio > GC_EVICT_RATIO) {
1007                 gc_work->next_gc_run = min_interval;
1008         } else {
1009                 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1010
1011                 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1012
1013                 gc_work->next_gc_run += min_interval;
1014                 if (gc_work->next_gc_run > max)
1015                         gc_work->next_gc_run = max;
1016         }
1017
1018         next_run = gc_work->next_gc_run;
1019         gc_work->last_bucket = i;
1020         queue_delayed_work(system_long_wq, &gc_work->dwork, next_run);
1021 }
1022
1023 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1024 {
1025         INIT_DELAYED_WORK(&gc_work->dwork, gc_worker);
1026         gc_work->next_gc_run = HZ;
1027         gc_work->exiting = false;
1028 }
1029
1030 static struct nf_conn *
1031 __nf_conntrack_alloc(struct net *net,
1032                      const struct nf_conntrack_zone *zone,
1033                      const struct nf_conntrack_tuple *orig,
1034                      const struct nf_conntrack_tuple *repl,
1035                      gfp_t gfp, u32 hash)
1036 {
1037         struct nf_conn *ct;
1038
1039         /* We don't want any race condition at early drop stage */
1040         atomic_inc(&net->ct.count);
1041
1042         if (nf_conntrack_max &&
1043             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
1044                 if (!early_drop(net, hash)) {
1045                         atomic_dec(&net->ct.count);
1046                         net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1047                         return ERR_PTR(-ENOMEM);
1048                 }
1049         }
1050
1051         /*
1052          * Do not use kmem_cache_zalloc(), as this cache uses
1053          * SLAB_DESTROY_BY_RCU.
1054          */
1055         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1056         if (ct == NULL)
1057                 goto out;
1058
1059         spin_lock_init(&ct->lock);
1060         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1061         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1062         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1063         /* save hash for reusing when confirming */
1064         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1065         ct->status = 0;
1066         write_pnet(&ct->ct_net, net);
1067         memset(&ct->__nfct_init_offset[0], 0,
1068                offsetof(struct nf_conn, proto) -
1069                offsetof(struct nf_conn, __nfct_init_offset[0]));
1070
1071         nf_ct_zone_add(ct, zone);
1072
1073         /* Because we use RCU lookups, we set ct_general.use to zero before
1074          * this is inserted in any list.
1075          */
1076         atomic_set(&ct->ct_general.use, 0);
1077         return ct;
1078 out:
1079         atomic_dec(&net->ct.count);
1080         return ERR_PTR(-ENOMEM);
1081 }
1082
1083 struct nf_conn *nf_conntrack_alloc(struct net *net,
1084                                    const struct nf_conntrack_zone *zone,
1085                                    const struct nf_conntrack_tuple *orig,
1086                                    const struct nf_conntrack_tuple *repl,
1087                                    gfp_t gfp)
1088 {
1089         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1090 }
1091 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1092
1093 void nf_conntrack_free(struct nf_conn *ct)
1094 {
1095         struct net *net = nf_ct_net(ct);
1096
1097         /* A freed object has refcnt == 0, that's
1098          * the golden rule for SLAB_DESTROY_BY_RCU
1099          */
1100         NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 0);
1101
1102         nf_ct_ext_destroy(ct);
1103         nf_ct_ext_free(ct);
1104         kmem_cache_free(nf_conntrack_cachep, ct);
1105         smp_mb__before_atomic();
1106         atomic_dec(&net->ct.count);
1107 }
1108 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1109
1110
1111 /* Allocate a new conntrack: we return -ENOMEM if classification
1112    failed due to stress.  Otherwise it really is unclassifiable. */
1113 static struct nf_conntrack_tuple_hash *
1114 init_conntrack(struct net *net, struct nf_conn *tmpl,
1115                const struct nf_conntrack_tuple *tuple,
1116                struct nf_conntrack_l3proto *l3proto,
1117                struct nf_conntrack_l4proto *l4proto,
1118                struct sk_buff *skb,
1119                unsigned int dataoff, u32 hash)
1120 {
1121         struct nf_conn *ct;
1122         struct nf_conn_help *help;
1123         struct nf_conntrack_tuple repl_tuple;
1124         struct nf_conntrack_ecache *ecache;
1125         struct nf_conntrack_expect *exp = NULL;
1126         const struct nf_conntrack_zone *zone;
1127         struct nf_conn_timeout *timeout_ext;
1128         struct nf_conntrack_zone tmp;
1129         unsigned int *timeouts;
1130
1131         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
1132                 pr_debug("Can't invert tuple.\n");
1133                 return NULL;
1134         }
1135
1136         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1137         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1138                                   hash);
1139         if (IS_ERR(ct))
1140                 return (struct nf_conntrack_tuple_hash *)ct;
1141
1142         if (!nf_ct_add_synproxy(ct, tmpl)) {
1143                 nf_conntrack_free(ct);
1144                 return ERR_PTR(-ENOMEM);
1145         }
1146
1147         timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1148         if (timeout_ext) {
1149                 timeouts = nf_ct_timeout_data(timeout_ext);
1150                 if (unlikely(!timeouts))
1151                         timeouts = l4proto->get_timeouts(net);
1152         } else {
1153                 timeouts = l4proto->get_timeouts(net);
1154         }
1155
1156         if (!l4proto->new(ct, skb, dataoff, timeouts)) {
1157                 nf_conntrack_free(ct);
1158                 pr_debug("can't track with proto module\n");
1159                 return NULL;
1160         }
1161
1162         if (timeout_ext)
1163                 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1164                                       GFP_ATOMIC);
1165
1166         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1167         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1168         nf_ct_labels_ext_add(ct);
1169
1170         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1171         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1172                                  ecache ? ecache->expmask : 0,
1173                              GFP_ATOMIC);
1174
1175         local_bh_disable();
1176         if (net->ct.expect_count) {
1177                 spin_lock(&nf_conntrack_expect_lock);
1178                 exp = nf_ct_find_expectation(net, zone, tuple);
1179                 if (exp) {
1180                         pr_debug("expectation arrives ct=%p exp=%p\n",
1181                                  ct, exp);
1182                         /* Welcome, Mr. Bond.  We've been expecting you... */
1183                         __set_bit(IPS_EXPECTED_BIT, &ct->status);
1184                         /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1185                         ct->master = exp->master;
1186                         if (exp->helper) {
1187                                 help = nf_ct_helper_ext_add(ct, exp->helper,
1188                                                             GFP_ATOMIC);
1189                                 if (help)
1190                                         rcu_assign_pointer(help->helper, exp->helper);
1191                         }
1192
1193 #ifdef CONFIG_NF_CONNTRACK_MARK
1194                         ct->mark = exp->master->mark;
1195 #endif
1196 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1197                         ct->secmark = exp->master->secmark;
1198 #endif
1199                         NF_CT_STAT_INC(net, expect_new);
1200                 }
1201                 spin_unlock(&nf_conntrack_expect_lock);
1202         }
1203         if (!exp)
1204                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1205
1206         /* Now it is inserted into the unconfirmed list, bump refcount */
1207         nf_conntrack_get(&ct->ct_general);
1208         nf_ct_add_to_unconfirmed_list(ct);
1209
1210         local_bh_enable();
1211
1212         if (exp) {
1213                 if (exp->expectfn)
1214                         exp->expectfn(ct, exp);
1215                 nf_ct_expect_put(exp);
1216         }
1217
1218         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1219 }
1220
1221 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
1222 static inline struct nf_conn *
1223 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
1224                   struct sk_buff *skb,
1225                   unsigned int dataoff,
1226                   u_int16_t l3num,
1227                   u_int8_t protonum,
1228                   struct nf_conntrack_l3proto *l3proto,
1229                   struct nf_conntrack_l4proto *l4proto,
1230                   int *set_reply,
1231                   enum ip_conntrack_info *ctinfo)
1232 {
1233         const struct nf_conntrack_zone *zone;
1234         struct nf_conntrack_tuple tuple;
1235         struct nf_conntrack_tuple_hash *h;
1236         struct nf_conntrack_zone tmp;
1237         struct nf_conn *ct;
1238         u32 hash;
1239
1240         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1241                              dataoff, l3num, protonum, net, &tuple, l3proto,
1242                              l4proto)) {
1243                 pr_debug("Can't get tuple\n");
1244                 return NULL;
1245         }
1246
1247         /* look for tuple match */
1248         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1249         hash = hash_conntrack_raw(&tuple, net);
1250         h = __nf_conntrack_find_get(net, zone, &tuple, hash);
1251         if (!h) {
1252                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
1253                                    skb, dataoff, hash);
1254                 if (!h)
1255                         return NULL;
1256                 if (IS_ERR(h))
1257                         return (void *)h;
1258         }
1259         ct = nf_ct_tuplehash_to_ctrack(h);
1260
1261         /* It exists; we have (non-exclusive) reference. */
1262         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1263                 *ctinfo = IP_CT_ESTABLISHED_REPLY;
1264                 /* Please set reply bit if this packet OK */
1265                 *set_reply = 1;
1266         } else {
1267                 /* Once we've had two way comms, always ESTABLISHED. */
1268                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1269                         pr_debug("normal packet for %p\n", ct);
1270                         *ctinfo = IP_CT_ESTABLISHED;
1271                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1272                         pr_debug("related packet for %p\n", ct);
1273                         *ctinfo = IP_CT_RELATED;
1274                 } else {
1275                         pr_debug("new packet for %p\n", ct);
1276                         *ctinfo = IP_CT_NEW;
1277                 }
1278                 *set_reply = 0;
1279         }
1280         skb->nfct = &ct->ct_general;
1281         skb->nfctinfo = *ctinfo;
1282         return ct;
1283 }
1284
1285 unsigned int
1286 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
1287                 struct sk_buff *skb)
1288 {
1289         struct nf_conn *ct, *tmpl = NULL;
1290         enum ip_conntrack_info ctinfo;
1291         struct nf_conntrack_l3proto *l3proto;
1292         struct nf_conntrack_l4proto *l4proto;
1293         unsigned int *timeouts;
1294         unsigned int dataoff;
1295         u_int8_t protonum;
1296         int set_reply = 0;
1297         int ret;
1298
1299         if (skb->nfct) {
1300                 /* Previously seen (loopback or untracked)?  Ignore. */
1301                 tmpl = (struct nf_conn *)skb->nfct;
1302                 if (!nf_ct_is_template(tmpl)) {
1303                         NF_CT_STAT_INC_ATOMIC(net, ignore);
1304                         return NF_ACCEPT;
1305                 }
1306                 skb->nfct = NULL;
1307         }
1308
1309         /* rcu_read_lock()ed by nf_hook_thresh */
1310         l3proto = __nf_ct_l3proto_find(pf);
1311         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
1312                                    &dataoff, &protonum);
1313         if (ret <= 0) {
1314                 pr_debug("not prepared to track yet or error occurred\n");
1315                 NF_CT_STAT_INC_ATOMIC(net, error);
1316                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1317                 ret = -ret;
1318                 goto out;
1319         }
1320
1321         l4proto = __nf_ct_l4proto_find(pf, protonum);
1322
1323         /* It may be an special packet, error, unclean...
1324          * inverse of the return code tells to the netfilter
1325          * core what to do with the packet. */
1326         if (l4proto->error != NULL) {
1327                 ret = l4proto->error(net, tmpl, skb, dataoff, &ctinfo,
1328                                      pf, hooknum);
1329                 if (ret <= 0) {
1330                         NF_CT_STAT_INC_ATOMIC(net, error);
1331                         NF_CT_STAT_INC_ATOMIC(net, invalid);
1332                         ret = -ret;
1333                         goto out;
1334                 }
1335                 /* ICMP[v6] protocol trackers may assign one conntrack. */
1336                 if (skb->nfct)
1337                         goto out;
1338         }
1339 repeat:
1340         ct = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
1341                                l3proto, l4proto, &set_reply, &ctinfo);
1342         if (!ct) {
1343                 /* Not valid part of a connection */
1344                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1345                 ret = NF_ACCEPT;
1346                 goto out;
1347         }
1348
1349         if (IS_ERR(ct)) {
1350                 /* Too stressed to deal. */
1351                 NF_CT_STAT_INC_ATOMIC(net, drop);
1352                 ret = NF_DROP;
1353                 goto out;
1354         }
1355
1356         NF_CT_ASSERT(skb->nfct);
1357
1358         /* Decide what timeout policy we want to apply to this flow. */
1359         timeouts = nf_ct_timeout_lookup(net, ct, l4proto);
1360
1361         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum, timeouts);
1362         if (ret <= 0) {
1363                 /* Invalid: inverse of the return code tells
1364                  * the netfilter core what to do */
1365                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1366                 nf_conntrack_put(skb->nfct);
1367                 skb->nfct = NULL;
1368                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1369                 if (ret == -NF_DROP)
1370                         NF_CT_STAT_INC_ATOMIC(net, drop);
1371                 /* Special case: TCP tracker reports an attempt to reopen a
1372                  * closed/aborted connection. We have to go back and create a
1373                  * fresh conntrack.
1374                  */
1375                 if (ret == -NF_REPEAT)
1376                         goto repeat;
1377                 ret = -ret;
1378                 goto out;
1379         }
1380
1381         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1382                 nf_conntrack_event_cache(IPCT_REPLY, ct);
1383 out:
1384         if (tmpl)
1385                 nf_ct_put(tmpl);
1386
1387         return ret;
1388 }
1389 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1390
1391 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
1392                           const struct nf_conntrack_tuple *orig)
1393 {
1394         bool ret;
1395
1396         rcu_read_lock();
1397         ret = nf_ct_invert_tuple(inverse, orig,
1398                                  __nf_ct_l3proto_find(orig->src.l3num),
1399                                  __nf_ct_l4proto_find(orig->src.l3num,
1400                                                       orig->dst.protonum));
1401         rcu_read_unlock();
1402         return ret;
1403 }
1404 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
1405
1406 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
1407    implicitly racy: see __nf_conntrack_confirm */
1408 void nf_conntrack_alter_reply(struct nf_conn *ct,
1409                               const struct nf_conntrack_tuple *newreply)
1410 {
1411         struct nf_conn_help *help = nfct_help(ct);
1412
1413         /* Should be unconfirmed, so not in hash table yet */
1414         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
1415
1416         pr_debug("Altering reply tuple of %p to ", ct);
1417         nf_ct_dump_tuple(newreply);
1418
1419         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1420         if (ct->master || (help && !hlist_empty(&help->expectations)))
1421                 return;
1422
1423         rcu_read_lock();
1424         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1425         rcu_read_unlock();
1426 }
1427 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1428
1429 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1430 void __nf_ct_refresh_acct(struct nf_conn *ct,
1431                           enum ip_conntrack_info ctinfo,
1432                           const struct sk_buff *skb,
1433                           unsigned long extra_jiffies,
1434                           int do_acct)
1435 {
1436         NF_CT_ASSERT(skb);
1437
1438         /* Only update if this is not a fixed timeout */
1439         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1440                 goto acct;
1441
1442         /* If not in hash table, timer will not be active yet */
1443         if (nf_ct_is_confirmed(ct))
1444                 extra_jiffies += nfct_time_stamp;
1445
1446         ct->timeout = extra_jiffies;
1447 acct:
1448         if (do_acct)
1449                 nf_ct_acct_update(ct, ctinfo, skb->len);
1450 }
1451 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1452
1453 bool nf_ct_kill_acct(struct nf_conn *ct,
1454                      enum ip_conntrack_info ctinfo,
1455                      const struct sk_buff *skb)
1456 {
1457         nf_ct_acct_update(ct, ctinfo, skb->len);
1458
1459         return nf_ct_delete(ct, 0, 0);
1460 }
1461 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1462
1463 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1464
1465 #include <linux/netfilter/nfnetlink.h>
1466 #include <linux/netfilter/nfnetlink_conntrack.h>
1467 #include <linux/mutex.h>
1468
1469 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
1470  * in ip_conntrack_core, since we don't want the protocols to autoload
1471  * or depend on ctnetlink */
1472 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1473                                const struct nf_conntrack_tuple *tuple)
1474 {
1475         if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1476             nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1477                 goto nla_put_failure;
1478         return 0;
1479
1480 nla_put_failure:
1481         return -1;
1482 }
1483 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1484
1485 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1486         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
1487         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
1488 };
1489 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1490
1491 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1492                                struct nf_conntrack_tuple *t)
1493 {
1494         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
1495                 return -EINVAL;
1496
1497         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1498         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1499
1500         return 0;
1501 }
1502 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1503
1504 int nf_ct_port_nlattr_tuple_size(void)
1505 {
1506         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1507 }
1508 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1509 #endif
1510
1511 /* Used by ipt_REJECT and ip6t_REJECT. */
1512 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
1513 {
1514         struct nf_conn *ct;
1515         enum ip_conntrack_info ctinfo;
1516
1517         /* This ICMP is in reverse direction to the packet which caused it */
1518         ct = nf_ct_get(skb, &ctinfo);
1519         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1520                 ctinfo = IP_CT_RELATED_REPLY;
1521         else
1522                 ctinfo = IP_CT_RELATED;
1523
1524         /* Attach to new skbuff, and increment count */
1525         nskb->nfct = &ct->ct_general;
1526         nskb->nfctinfo = ctinfo;
1527         nf_conntrack_get(nskb->nfct);
1528 }
1529
1530 /* Bring out ya dead! */
1531 static struct nf_conn *
1532 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1533                 void *data, unsigned int *bucket)
1534 {
1535         struct nf_conntrack_tuple_hash *h;
1536         struct nf_conn *ct;
1537         struct hlist_nulls_node *n;
1538         int cpu;
1539         spinlock_t *lockp;
1540
1541         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1542                 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
1543                 local_bh_disable();
1544                 nf_conntrack_lock(lockp);
1545                 if (*bucket < nf_conntrack_htable_size) {
1546                         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
1547                                 if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1548                                         continue;
1549                                 ct = nf_ct_tuplehash_to_ctrack(h);
1550                                 if (net_eq(nf_ct_net(ct), net) &&
1551                                     iter(ct, data))
1552                                         goto found;
1553                         }
1554                 }
1555                 spin_unlock(lockp);
1556                 local_bh_enable();
1557                 cond_resched();
1558         }
1559
1560         for_each_possible_cpu(cpu) {
1561                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1562
1563                 spin_lock_bh(&pcpu->lock);
1564                 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
1565                         ct = nf_ct_tuplehash_to_ctrack(h);
1566                         if (iter(ct, data))
1567                                 set_bit(IPS_DYING_BIT, &ct->status);
1568                 }
1569                 spin_unlock_bh(&pcpu->lock);
1570                 cond_resched();
1571         }
1572         return NULL;
1573 found:
1574         atomic_inc(&ct->ct_general.use);
1575         spin_unlock(lockp);
1576         local_bh_enable();
1577         return ct;
1578 }
1579
1580 void nf_ct_iterate_cleanup(struct net *net,
1581                            int (*iter)(struct nf_conn *i, void *data),
1582                            void *data, u32 portid, int report)
1583 {
1584         struct nf_conn *ct;
1585         unsigned int bucket = 0;
1586
1587         might_sleep();
1588
1589         if (atomic_read(&net->ct.count) == 0)
1590                 return;
1591
1592         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1593                 /* Time to push up daises... */
1594
1595                 nf_ct_delete(ct, portid, report);
1596                 nf_ct_put(ct);
1597                 cond_resched();
1598         }
1599 }
1600 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1601
1602 static int kill_all(struct nf_conn *i, void *data)
1603 {
1604         return 1;
1605 }
1606
1607 void nf_ct_free_hashtable(void *hash, unsigned int size)
1608 {
1609         if (is_vmalloc_addr(hash))
1610                 vfree(hash);
1611         else
1612                 free_pages((unsigned long)hash,
1613                            get_order(sizeof(struct hlist_head) * size));
1614 }
1615 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1616
1617 static int untrack_refs(void)
1618 {
1619         int cnt = 0, cpu;
1620
1621         for_each_possible_cpu(cpu) {
1622                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1623
1624                 cnt += atomic_read(&ct->ct_general.use) - 1;
1625         }
1626         return cnt;
1627 }
1628
1629 void nf_conntrack_cleanup_start(void)
1630 {
1631         conntrack_gc_work.exiting = true;
1632         RCU_INIT_POINTER(ip_ct_attach, NULL);
1633 }
1634
1635 void nf_conntrack_cleanup_end(void)
1636 {
1637         RCU_INIT_POINTER(nf_ct_destroy, NULL);
1638         while (untrack_refs() > 0)
1639                 schedule();
1640
1641         cancel_delayed_work_sync(&conntrack_gc_work.dwork);
1642         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1643
1644         nf_conntrack_proto_fini();
1645         nf_conntrack_seqadj_fini();
1646         nf_conntrack_labels_fini();
1647         nf_conntrack_helper_fini();
1648         nf_conntrack_timeout_fini();
1649         nf_conntrack_ecache_fini();
1650         nf_conntrack_tstamp_fini();
1651         nf_conntrack_acct_fini();
1652         nf_conntrack_expect_fini();
1653
1654         kmem_cache_destroy(nf_conntrack_cachep);
1655 }
1656
1657 /*
1658  * Mishearing the voices in his head, our hero wonders how he's
1659  * supposed to kill the mall.
1660  */
1661 void nf_conntrack_cleanup_net(struct net *net)
1662 {
1663         LIST_HEAD(single);
1664
1665         list_add(&net->exit_list, &single);
1666         nf_conntrack_cleanup_net_list(&single);
1667 }
1668
1669 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
1670 {
1671         int busy;
1672         struct net *net;
1673
1674         /*
1675          * This makes sure all current packets have passed through
1676          *  netfilter framework.  Roll on, two-stage module
1677          *  delete...
1678          */
1679         synchronize_net();
1680 i_see_dead_people:
1681         busy = 0;
1682         list_for_each_entry(net, net_exit_list, exit_list) {
1683                 nf_ct_iterate_cleanup(net, kill_all, NULL, 0, 0);
1684                 if (atomic_read(&net->ct.count) != 0)
1685                         busy = 1;
1686         }
1687         if (busy) {
1688                 schedule();
1689                 goto i_see_dead_people;
1690         }
1691
1692         list_for_each_entry(net, net_exit_list, exit_list) {
1693                 nf_conntrack_proto_pernet_fini(net);
1694                 nf_conntrack_helper_pernet_fini(net);
1695                 nf_conntrack_ecache_pernet_fini(net);
1696                 nf_conntrack_tstamp_pernet_fini(net);
1697                 nf_conntrack_acct_pernet_fini(net);
1698                 nf_conntrack_expect_pernet_fini(net);
1699                 free_percpu(net->ct.stat);
1700                 free_percpu(net->ct.pcpu_lists);
1701         }
1702 }
1703
1704 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
1705 {
1706         struct hlist_nulls_head *hash;
1707         unsigned int nr_slots, i;
1708         size_t sz;
1709
1710         if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1711                 return NULL;
1712
1713         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1714         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1715
1716         if (nr_slots > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1717                 return NULL;
1718
1719         sz = nr_slots * sizeof(struct hlist_nulls_head);
1720         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1721                                         get_order(sz));
1722         if (!hash)
1723                 hash = vzalloc(sz);
1724
1725         if (hash && nulls)
1726                 for (i = 0; i < nr_slots; i++)
1727                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1728
1729         return hash;
1730 }
1731 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1732
1733 int nf_conntrack_hash_resize(unsigned int hashsize)
1734 {
1735         int i, bucket;
1736         unsigned int old_size;
1737         struct hlist_nulls_head *hash, *old_hash;
1738         struct nf_conntrack_tuple_hash *h;
1739         struct nf_conn *ct;
1740
1741         if (!hashsize)
1742                 return -EINVAL;
1743
1744         hash = nf_ct_alloc_hashtable(&hashsize, 1);
1745         if (!hash)
1746                 return -ENOMEM;
1747
1748         old_size = nf_conntrack_htable_size;
1749         if (old_size == hashsize) {
1750                 nf_ct_free_hashtable(hash, hashsize);
1751                 return 0;
1752         }
1753
1754         local_bh_disable();
1755         nf_conntrack_all_lock();
1756         write_seqcount_begin(&nf_conntrack_generation);
1757
1758         /* Lookups in the old hash might happen in parallel, which means we
1759          * might get false negatives during connection lookup. New connections
1760          * created because of a false negative won't make it into the hash
1761          * though since that required taking the locks.
1762          */
1763
1764         for (i = 0; i < nf_conntrack_htable_size; i++) {
1765                 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
1766                         h = hlist_nulls_entry(nf_conntrack_hash[i].first,
1767                                               struct nf_conntrack_tuple_hash, hnnode);
1768                         ct = nf_ct_tuplehash_to_ctrack(h);
1769                         hlist_nulls_del_rcu(&h->hnnode);
1770                         bucket = __hash_conntrack(nf_ct_net(ct),
1771                                                   &h->tuple, hashsize);
1772                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1773                 }
1774         }
1775         old_size = nf_conntrack_htable_size;
1776         old_hash = nf_conntrack_hash;
1777
1778         nf_conntrack_hash = hash;
1779         nf_conntrack_htable_size = hashsize;
1780
1781         write_seqcount_end(&nf_conntrack_generation);
1782         nf_conntrack_all_unlock();
1783         local_bh_enable();
1784
1785         synchronize_net();
1786         nf_ct_free_hashtable(old_hash, old_size);
1787         return 0;
1788 }
1789
1790 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1791 {
1792         unsigned int hashsize;
1793         int rc;
1794
1795         if (current->nsproxy->net_ns != &init_net)
1796                 return -EOPNOTSUPP;
1797
1798         /* On boot, we can set this without any fancy locking. */
1799         if (!nf_conntrack_htable_size)
1800                 return param_set_uint(val, kp);
1801
1802         rc = kstrtouint(val, 0, &hashsize);
1803         if (rc)
1804                 return rc;
1805
1806         return nf_conntrack_hash_resize(hashsize);
1807 }
1808 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1809
1810 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1811                   &nf_conntrack_htable_size, 0600);
1812
1813 void nf_ct_untracked_status_or(unsigned long bits)
1814 {
1815         int cpu;
1816
1817         for_each_possible_cpu(cpu)
1818                 per_cpu(nf_conntrack_untracked, cpu).status |= bits;
1819 }
1820 EXPORT_SYMBOL_GPL(nf_ct_untracked_status_or);
1821
1822 int nf_conntrack_init_start(void)
1823 {
1824         int max_factor = 8;
1825         int ret = -ENOMEM;
1826         int i, cpu;
1827
1828         seqcount_init(&nf_conntrack_generation);
1829
1830         for (i = 0; i < CONNTRACK_LOCKS; i++)
1831                 spin_lock_init(&nf_conntrack_locks[i]);
1832
1833         if (!nf_conntrack_htable_size) {
1834                 /* Idea from tcp.c: use 1/16384 of memory.
1835                  * On i386: 32MB machine has 512 buckets.
1836                  * >= 1GB machines have 16384 buckets.
1837                  * >= 4GB machines have 65536 buckets.
1838                  */
1839                 nf_conntrack_htable_size
1840                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1841                            / sizeof(struct hlist_head));
1842                 if (totalram_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
1843                         nf_conntrack_htable_size = 65536;
1844                 else if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1845                         nf_conntrack_htable_size = 16384;
1846                 if (nf_conntrack_htable_size < 32)
1847                         nf_conntrack_htable_size = 32;
1848
1849                 /* Use a max. factor of four by default to get the same max as
1850                  * with the old struct list_heads. When a table size is given
1851                  * we use the old value of 8 to avoid reducing the max.
1852                  * entries. */
1853                 max_factor = 4;
1854         }
1855
1856         nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
1857         if (!nf_conntrack_hash)
1858                 return -ENOMEM;
1859
1860         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1861
1862         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
1863                                                 sizeof(struct nf_conn), 0,
1864                                                 SLAB_DESTROY_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
1865         if (!nf_conntrack_cachep)
1866                 goto err_cachep;
1867
1868         printk(KERN_INFO "nf_conntrack version %s (%u buckets, %d max)\n",
1869                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1870                nf_conntrack_max);
1871
1872         ret = nf_conntrack_expect_init();
1873         if (ret < 0)
1874                 goto err_expect;
1875
1876         ret = nf_conntrack_acct_init();
1877         if (ret < 0)
1878                 goto err_acct;
1879
1880         ret = nf_conntrack_tstamp_init();
1881         if (ret < 0)
1882                 goto err_tstamp;
1883
1884         ret = nf_conntrack_ecache_init();
1885         if (ret < 0)
1886                 goto err_ecache;
1887
1888         ret = nf_conntrack_timeout_init();
1889         if (ret < 0)
1890                 goto err_timeout;
1891
1892         ret = nf_conntrack_helper_init();
1893         if (ret < 0)
1894                 goto err_helper;
1895
1896         ret = nf_conntrack_labels_init();
1897         if (ret < 0)
1898                 goto err_labels;
1899
1900         ret = nf_conntrack_seqadj_init();
1901         if (ret < 0)
1902                 goto err_seqadj;
1903
1904         ret = nf_conntrack_proto_init();
1905         if (ret < 0)
1906                 goto err_proto;
1907
1908         /* Set up fake conntrack: to never be deleted, not in any hashes */
1909         for_each_possible_cpu(cpu) {
1910                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1911                 write_pnet(&ct->ct_net, &init_net);
1912                 atomic_set(&ct->ct_general.use, 1);
1913         }
1914         /*  - and look it like as a confirmed connection */
1915         nf_ct_untracked_status_or(IPS_CONFIRMED | IPS_UNTRACKED);
1916
1917         conntrack_gc_work_init(&conntrack_gc_work);
1918         queue_delayed_work(system_long_wq, &conntrack_gc_work.dwork, HZ);
1919
1920         return 0;
1921
1922 err_proto:
1923         nf_conntrack_seqadj_fini();
1924 err_seqadj:
1925         nf_conntrack_labels_fini();
1926 err_labels:
1927         nf_conntrack_helper_fini();
1928 err_helper:
1929         nf_conntrack_timeout_fini();
1930 err_timeout:
1931         nf_conntrack_ecache_fini();
1932 err_ecache:
1933         nf_conntrack_tstamp_fini();
1934 err_tstamp:
1935         nf_conntrack_acct_fini();
1936 err_acct:
1937         nf_conntrack_expect_fini();
1938 err_expect:
1939         kmem_cache_destroy(nf_conntrack_cachep);
1940 err_cachep:
1941         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1942         return ret;
1943 }
1944
1945 void nf_conntrack_init_end(void)
1946 {
1947         /* For use by REJECT target */
1948         RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
1949         RCU_INIT_POINTER(nf_ct_destroy, destroy_conntrack);
1950 }
1951
1952 /*
1953  * We need to use special "null" values, not used in hash table
1954  */
1955 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1956 #define DYING_NULLS_VAL         ((1<<30)+1)
1957 #define TEMPLATE_NULLS_VAL      ((1<<30)+2)
1958
1959 int nf_conntrack_init_net(struct net *net)
1960 {
1961         int ret = -ENOMEM;
1962         int cpu;
1963
1964         atomic_set(&net->ct.count, 0);
1965
1966         net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
1967         if (!net->ct.pcpu_lists)
1968                 goto err_stat;
1969
1970         for_each_possible_cpu(cpu) {
1971                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1972
1973                 spin_lock_init(&pcpu->lock);
1974                 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
1975                 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
1976         }
1977
1978         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
1979         if (!net->ct.stat)
1980                 goto err_pcpu_lists;
1981
1982         ret = nf_conntrack_expect_pernet_init(net);
1983         if (ret < 0)
1984                 goto err_expect;
1985         ret = nf_conntrack_acct_pernet_init(net);
1986         if (ret < 0)
1987                 goto err_acct;
1988         ret = nf_conntrack_tstamp_pernet_init(net);
1989         if (ret < 0)
1990                 goto err_tstamp;
1991         ret = nf_conntrack_ecache_pernet_init(net);
1992         if (ret < 0)
1993                 goto err_ecache;
1994         ret = nf_conntrack_helper_pernet_init(net);
1995         if (ret < 0)
1996                 goto err_helper;
1997         ret = nf_conntrack_proto_pernet_init(net);
1998         if (ret < 0)
1999                 goto err_proto;
2000         return 0;
2001
2002 err_proto:
2003         nf_conntrack_helper_pernet_fini(net);
2004 err_helper:
2005         nf_conntrack_ecache_pernet_fini(net);
2006 err_ecache:
2007         nf_conntrack_tstamp_pernet_fini(net);
2008 err_tstamp:
2009         nf_conntrack_acct_pernet_fini(net);
2010 err_acct:
2011         nf_conntrack_expect_pernet_fini(net);
2012 err_expect:
2013         free_percpu(net->ct.stat);
2014 err_pcpu_lists:
2015         free_percpu(net->ct.pcpu_lists);
2016 err_stat:
2017         return ret;
2018 }