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 /* nf_conn must be 8 bytes aligned, as the 3 LSB bits are used
 185  * for the nfctinfo. We cheat by (ab)using the PER CPU cache line
 186  * alignment to enforce this.
 187  */
 188 DEFINE_PER_CPU_ALIGNED(struct nf_conn, nf_conntrack_untracked);
 189 EXPORT_PER_CPU_SYMBOL(nf_conntrack_untracked);
 190
 191 static unsigned int nf_conntrack_hash_rnd __read_mostly;
 192
 193 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
 194                               const struct net *net)
 195 {
 196         unsigned int n;
 197         u32 seed;
 198
 199         get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
 200
 201         /* The direction must be ignored, so we hash everything up to the
 202          * destination ports (which is a multiple of 4) and treat the last
 203          * three bytes manually.
 204          */
 205         seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
 206         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
 207         return jhash2((u32 *)tuple, n, seed ^
 208                       (((__force __u16)tuple->dst.u.all << 16) |
 209                       tuple->dst.protonum));
 210 }
 211
 212 static u32 scale_hash(u32 hash)
 213 {
 214         return reciprocal_scale(hash, nf_conntrack_htable_size);
 215 }
 216
 217 static u32 __hash_conntrack(const struct net *net,
 218                             const struct nf_conntrack_tuple *tuple,
 219                             unsigned int size)
 220 {
 221         return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
 222 }
 223
 224 static u32 hash_conntrack(const struct net *net,
 225                           const struct nf_conntrack_tuple *tuple)
 226 {
 227         return scale_hash(hash_conntrack_raw(tuple, net));
 228 }
 229
 230 bool
 231 nf_ct_get_tuple(const struct sk_buff *skb,
 232                 unsigned int nhoff,
 233                 unsigned int dataoff,
 234                 u_int16_t l3num,
 235                 u_int8_t protonum,
 236                 struct net *net,
 237                 struct nf_conntrack_tuple *tuple,
 238                 const struct nf_conntrack_l3proto *l3proto,
 239                 const struct nf_conntrack_l4proto *l4proto)
 240 {
 241         memset(tuple, 0, sizeof(*tuple));
 242
 243         tuple->src.l3num = l3num;
 244         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 245                 return false;
 246
 247         tuple->dst.protonum = protonum;
 248         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 249
 250         return l4proto->pkt_to_tuple(skb, dataoff, net, tuple);
 251 }
 252 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 253
 254 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 255                        u_int16_t l3num,
 256                        struct net *net, struct nf_conntrack_tuple *tuple)
 257 {
 258         struct nf_conntrack_l3proto *l3proto;
 259         struct nf_conntrack_l4proto *l4proto;
 260         unsigned int protoff;
 261         u_int8_t protonum;
 262         int ret;
 263
 264         rcu_read_lock();
 265
 266         l3proto = __nf_ct_l3proto_find(l3num);
 267         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
 268         if (ret != NF_ACCEPT) {
 269                 rcu_read_unlock();
 270                 return false;
 271         }
 272
 273         l4proto = __nf_ct_l4proto_find(l3num, protonum);
 274
 275         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple,
 276                               l3proto, l4proto);
 277
 278         rcu_read_unlock();
 279         return ret;
 280 }
 281 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 282
 283 bool
 284 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 285                    const struct nf_conntrack_tuple *orig,
 286                    const struct nf_conntrack_l3proto *l3proto,
 287                    const struct nf_conntrack_l4proto *l4proto)
 288 {
 289         memset(inverse, 0, sizeof(*inverse));
 290
 291         inverse->src.l3num = orig->src.l3num;
 292         if (l3proto->invert_tuple(inverse, orig) == 0)
 293                 return false;
 294
 295         inverse->dst.dir = !orig->dst.dir;
 296
 297         inverse->dst.protonum = orig->dst.protonum;
 298         return l4proto->invert_tuple(inverse, orig);
 299 }
 300 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 301
 302 static void
 303 clean_from_lists(struct nf_conn *ct)
 304 {
 305         pr_debug("clean_from_lists(%p)\n", ct);
 306         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 307         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 308
 309         /* Destroy all pending expectations */
 310         nf_ct_remove_expectations(ct);
 311 }
 312
 313 /* must be called with local_bh_disable */
 314 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
 315 {
 316         struct ct_pcpu *pcpu;
 317
 318         /* add this conntrack to the (per cpu) dying list */
 319         ct->cpu = smp_processor_id();
 320         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 321
 322         spin_lock(&pcpu->lock);
 323         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 324                              &pcpu->dying);
 325         spin_unlock(&pcpu->lock);
 326 }
 327
 328 /* must be called with local_bh_disable */
 329 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
 330 {
 331         struct ct_pcpu *pcpu;
 332
 333         /* add this conntrack to the (per cpu) unconfirmed list */
 334         ct->cpu = smp_processor_id();
 335         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 336
 337         spin_lock(&pcpu->lock);
 338         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 339                              &pcpu->unconfirmed);
 340         spin_unlock(&pcpu->lock);
 341 }
 342
 343 /* must be called with local_bh_disable */
 344 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
 345 {
 346         struct ct_pcpu *pcpu;
 347
 348         /* We overload first tuple to link into unconfirmed or dying list.*/
 349         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 350
 351         spin_lock(&pcpu->lock);
 352         BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 353         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 354         spin_unlock(&pcpu->lock);
 355 }
 356
 357 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
 358
 359 /* Released via destroy_conntrack() */
 360 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
 361                                  const struct nf_conntrack_zone *zone,
 362                                  gfp_t flags)
 363 {
 364         struct nf_conn *tmpl, *p;
 365
 366         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
 367                 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
 368                 if (!tmpl)
 369                         return NULL;
 370
 371                 p = tmpl;
 372                 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 373                 if (tmpl != p) {
 374                         tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 375                         tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
 376                 }
 377         } else {
 378                 tmpl = kzalloc(sizeof(*tmpl), flags);
 379                 if (!tmpl)
 380                         return NULL;
 381         }
 382
 383         tmpl->status = IPS_TEMPLATE;
 384         write_pnet(&tmpl->ct_net, net);
 385         nf_ct_zone_add(tmpl, zone);
 386         atomic_set(&tmpl->ct_general.use, 0);
 387
 388         return tmpl;
 389 }
 390 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
 391
 392 void nf_ct_tmpl_free(struct nf_conn *tmpl)
 393 {
 394         nf_ct_ext_destroy(tmpl);
 395         nf_ct_ext_free(tmpl);
 396
 397         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
 398                 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
 399         else
 400                 kfree(tmpl);
 401 }
 402 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
 403
 404 static void
 405 destroy_conntrack(struct nf_conntrack *nfct)
 406 {
 407         struct nf_conn *ct = (struct nf_conn *)nfct;
 408         struct nf_conntrack_l4proto *l4proto;
 409
 410         pr_debug("destroy_conntrack(%p)\n", ct);
 411         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 412
 413         if (unlikely(nf_ct_is_template(ct))) {
 414                 nf_ct_tmpl_free(ct);
 415                 return;
 416         }
 417         rcu_read_lock();
 418         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 419         if (l4proto->destroy)
 420                 l4proto->destroy(ct);
 421
 422         rcu_read_unlock();
 423
 424         local_bh_disable();
 425         /* Expectations will have been removed in clean_from_lists,
 426          * except TFTP can create an expectation on the first packet,
 427          * before connection is in the list, so we need to clean here,
 428          * too.
 429          */
 430         nf_ct_remove_expectations(ct);
 431
 432         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 433
 434         local_bh_enable();
 435
 436         if (ct->master)
 437                 nf_ct_put(ct->master);
 438
 439         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 440         nf_conntrack_free(ct);
 441 }
 442
 443 static void nf_ct_delete_from_lists(struct nf_conn *ct)
 444 {
 445         struct net *net = nf_ct_net(ct);
 446         unsigned int hash, reply_hash;
 447         unsigned int sequence;
 448
 449         nf_ct_helper_destroy(ct);
 450
 451         local_bh_disable();
 452         do {
 453                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 454                 hash = hash_conntrack(net,
 455                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 456                 reply_hash = hash_conntrack(net,
 457                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 458         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 459
 460         clean_from_lists(ct);
 461         nf_conntrack_double_unlock(hash, reply_hash);
 462
 463         nf_ct_add_to_dying_list(ct);
 464
 465         local_bh_enable();
 466 }
 467
 468 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
 469 {
 470         struct nf_conn_tstamp *tstamp;
 471
 472         if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
 473                 return false;
 474
 475         tstamp = nf_conn_tstamp_find(ct);
 476         if (tstamp && tstamp->stop == 0)
 477                 tstamp->stop = ktime_get_real_ns();
 478
 479         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 480                                     portid, report) < 0) {
 481                 /* destroy event was not delivered. nf_ct_put will
 482                  * be done by event cache worker on redelivery.
 483                  */
 484                 nf_ct_delete_from_lists(ct);
 485                 nf_conntrack_ecache_delayed_work(nf_ct_net(ct));
 486                 return false;
 487         }
 488
 489         nf_conntrack_ecache_work(nf_ct_net(ct));
 490         nf_ct_delete_from_lists(ct);
 491         nf_ct_put(ct);
 492         return true;
 493 }
 494 EXPORT_SYMBOL_GPL(nf_ct_delete);
 495
 496 static inline bool
 497 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
 498                 const struct nf_conntrack_tuple *tuple,
 499                 const struct nf_conntrack_zone *zone,
 500                 const struct net *net)
 501 {
 502         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 503
 504         /* A conntrack can be recreated with the equal tuple,
 505          * so we need to check that the conntrack is confirmed
 506          */
 507         return nf_ct_tuple_equal(tuple, &h->tuple) &&
 508                nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
 509                nf_ct_is_confirmed(ct) &&
 510                net_eq(net, nf_ct_net(ct));
 511 }
 512
 513 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
 514 static void nf_ct_gc_expired(struct nf_conn *ct)
 515 {
 516         if (!atomic_inc_not_zero(&ct->ct_general.use))
 517                 return;
 518
 519         if (nf_ct_should_gc(ct))
 520                 nf_ct_kill(ct);
 521
 522         nf_ct_put(ct);
 523 }
 524
 525 /*
 526  * Warning :
 527  * - Caller must take a reference on returned object
 528  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 529  */
 530 static struct nf_conntrack_tuple_hash *
 531 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
 532                       const struct nf_conntrack_tuple *tuple, u32 hash)
 533 {
 534         struct nf_conntrack_tuple_hash *h;
 535         struct hlist_nulls_head *ct_hash;
 536         struct hlist_nulls_node *n;
 537         unsigned int bucket, hsize;
 538
 539 begin:
 540         nf_conntrack_get_ht(&ct_hash, &hsize);
 541         bucket = reciprocal_scale(hash, hsize);
 542
 543         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
 544                 struct nf_conn *ct;
 545
 546                 ct = nf_ct_tuplehash_to_ctrack(h);
 547                 if (nf_ct_is_expired(ct)) {
 548                         nf_ct_gc_expired(ct);
 549                         continue;
 550                 }
 551
 552                 if (nf_ct_is_dying(ct))
 553                         continue;
 554
 555                 if (nf_ct_key_equal(h, tuple, zone, net))
 556                         return h;
 557         }
 558         /*
 559          * if the nulls value we got at the end of this lookup is
 560          * not the expected one, we must restart lookup.
 561          * We probably met an item that was moved to another chain.
 562          */
 563         if (get_nulls_value(n) != bucket) {
 564                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 565                 goto begin;
 566         }
 567
 568         return NULL;
 569 }
 570
 571 /* Find a connection corresponding to a tuple. */
 572 static struct nf_conntrack_tuple_hash *
 573 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 574                         const struct nf_conntrack_tuple *tuple, u32 hash)
 575 {
 576         struct nf_conntrack_tuple_hash *h;
 577         struct nf_conn *ct;
 578
 579         rcu_read_lock();
 580 begin:
 581         h = ____nf_conntrack_find(net, zone, tuple, hash);
 582         if (h) {
 583                 ct = nf_ct_tuplehash_to_ctrack(h);
 584                 if (unlikely(nf_ct_is_dying(ct) ||
 585                              !atomic_inc_not_zero(&ct->ct_general.use)))
 586                         h = NULL;
 587                 else {
 588                         if (unlikely(!nf_ct_key_equal(h, tuple, zone, net))) {
 589                                 nf_ct_put(ct);
 590                                 goto begin;
 591                         }
 592                 }
 593         }
 594         rcu_read_unlock();
 595
 596         return h;
 597 }
 598
 599 struct nf_conntrack_tuple_hash *
 600 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 601                       const struct nf_conntrack_tuple *tuple)
 602 {
 603         return __nf_conntrack_find_get(net, zone, tuple,
 604                                        hash_conntrack_raw(tuple, net));
 605 }
 606 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 607
 608 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 609                                        unsigned int hash,
 610                                        unsigned int reply_hash)
 611 {
 612         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 613                            &nf_conntrack_hash[hash]);
 614         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 615                            &nf_conntrack_hash[reply_hash]);
 616 }
 617
 618 int
 619 nf_conntrack_hash_check_insert(struct nf_conn *ct)
 620 {
 621         const struct nf_conntrack_zone *zone;
 622         struct net *net = nf_ct_net(ct);
 623         unsigned int hash, reply_hash;
 624         struct nf_conntrack_tuple_hash *h;
 625         struct hlist_nulls_node *n;
 626         unsigned int sequence;
 627
 628         zone = nf_ct_zone(ct);
 629
 630         local_bh_disable();
 631         do {
 632                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 633                 hash = hash_conntrack(net,
 634                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 635                 reply_hash = hash_conntrack(net,
 636                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 637         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 638
 639         /* See if there's one in the list already, including reverse */
 640         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 641                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 642                                     zone, net))
 643                         goto out;
 644
 645         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 646                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 647                                     zone, net))
 648                         goto out;
 649
 650         smp_wmb();
 651         /* The caller holds a reference to this object */
 652         atomic_set(&ct->ct_general.use, 2);
 653         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 654         nf_conntrack_double_unlock(hash, reply_hash);
 655         NF_CT_STAT_INC(net, insert);
 656         local_bh_enable();
 657         return 0;
 658
 659 out:
 660         nf_conntrack_double_unlock(hash, reply_hash);
 661         NF_CT_STAT_INC(net, insert_failed);
 662         local_bh_enable();
 663         return -EEXIST;
 664 }
 665 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 666
 667 static inline void nf_ct_acct_update(struct nf_conn *ct,
 668                                      enum ip_conntrack_info ctinfo,
 669                                      unsigned int len)
 670 {
 671         struct nf_conn_acct *acct;
 672
 673         acct = nf_conn_acct_find(ct);
 674         if (acct) {
 675                 struct nf_conn_counter *counter = acct->counter;
 676
 677                 atomic64_inc(&counter[CTINFO2DIR(ctinfo)].packets);
 678                 atomic64_add(len, &counter[CTINFO2DIR(ctinfo)].bytes);
 679         }
 680 }
 681
 682 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
 683                              const struct nf_conn *loser_ct)
 684 {
 685         struct nf_conn_acct *acct;
 686
 687         acct = nf_conn_acct_find(loser_ct);
 688         if (acct) {
 689                 struct nf_conn_counter *counter = acct->counter;
 690                 unsigned int bytes;
 691
 692                 /* u32 should be fine since we must have seen one packet. */
 693                 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
 694                 nf_ct_acct_update(ct, ctinfo, bytes);
 695         }
 696 }
 697
 698 /* Resolve race on insertion if this protocol allows this. */
 699 static int nf_ct_resolve_clash(struct net *net, struct sk_buff *skb,
 700                                enum ip_conntrack_info ctinfo,
 701                                struct nf_conntrack_tuple_hash *h)
 702 {
 703         /* This is the conntrack entry already in hashes that won race. */
 704         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 705         struct nf_conntrack_l4proto *l4proto;
 706
 707         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 708         if (l4proto->allow_clash &&
 709             !nfct_nat(ct) &&
 710             !nf_ct_is_dying(ct) &&
 711             atomic_inc_not_zero(&ct->ct_general.use)) {
 712                 enum ip_conntrack_info oldinfo;
 713                 struct nf_conn *loser_ct = nf_ct_get(skb, &oldinfo);
 714
 715                 nf_ct_acct_merge(ct, ctinfo, loser_ct);
 716                 nf_conntrack_put(&loser_ct->ct_general);
 717                 nf_ct_set(skb, ct, oldinfo);
 718                 return NF_ACCEPT;
 719         }
 720         NF_CT_STAT_INC(net, drop);
 721         return NF_DROP;
 722 }
 723
 724 /* Confirm a connection given skb; places it in hash table */
 725 int
 726 __nf_conntrack_confirm(struct sk_buff *skb)
 727 {
 728         const struct nf_conntrack_zone *zone;
 729         unsigned int hash, reply_hash;
 730         struct nf_conntrack_tuple_hash *h;
 731         struct nf_conn *ct;
 732         struct nf_conn_help *help;
 733         struct nf_conn_tstamp *tstamp;
 734         struct hlist_nulls_node *n;
 735         enum ip_conntrack_info ctinfo;
 736         struct net *net;
 737         unsigned int sequence;
 738         int ret = NF_DROP;
 739
 740         ct = nf_ct_get(skb, &ctinfo);
 741         net = nf_ct_net(ct);
 742
 743         /* ipt_REJECT uses nf_conntrack_attach to attach related
 744            ICMP/TCP RST packets in other direction.  Actual packet
 745            which created connection will be IP_CT_NEW or for an
 746            expected connection, IP_CT_RELATED. */
 747         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 748                 return NF_ACCEPT;
 749
 750         zone = nf_ct_zone(ct);
 751         local_bh_disable();
 752
 753         do {
 754                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 755                 /* reuse the hash saved before */
 756                 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
 757                 hash = scale_hash(hash);
 758                 reply_hash = hash_conntrack(net,
 759                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 760
 761         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 762
 763         /* We're not in hash table, and we refuse to set up related
 764          * connections for unconfirmed conns.  But packet copies and
 765          * REJECT will give spurious warnings here.
 766          */
 767         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 768
 769         /* No external references means no one else could have
 770          * confirmed us.
 771          */
 772         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 773         pr_debug("Confirming conntrack %p\n", ct);
 774         /* We have to check the DYING flag after unlink to prevent
 775          * a race against nf_ct_get_next_corpse() possibly called from
 776          * user context, else we insert an already 'dead' hash, blocking
 777          * further use of that particular connection -JM.
 778          */
 779         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 780
 781         if (unlikely(nf_ct_is_dying(ct))) {
 782                 nf_ct_add_to_dying_list(ct);
 783                 goto dying;
 784         }
 785
 786         /* See if there's one in the list already, including reverse:
 787            NAT could have grabbed it without realizing, since we're
 788            not in the hash.  If there is, we lost race. */
 789         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 790                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 791                                     zone, net))
 792                         goto out;
 793
 794         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 795                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 796                                     zone, net))
 797                         goto out;
 798
 799         /* Timer relative to confirmation time, not original
 800            setting time, otherwise we'd get timer wrap in
 801            weird delay cases. */
 802         ct->timeout += nfct_time_stamp;
 803         atomic_inc(&ct->ct_general.use);
 804         ct->status |= IPS_CONFIRMED;
 805
 806         /* set conntrack timestamp, if enabled. */
 807         tstamp = nf_conn_tstamp_find(ct);
 808         if (tstamp) {
 809                 if (skb->tstamp == 0)
 810                         __net_timestamp(skb);
 811
 812                 tstamp->start = ktime_to_ns(skb->tstamp);
 813         }
 814         /* Since the lookup is lockless, hash insertion must be done after
 815          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 816          * guarantee that no other CPU can find the conntrack before the above
 817          * stores are visible.
 818          */
 819         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 820         nf_conntrack_double_unlock(hash, reply_hash);
 821         local_bh_enable();
 822
 823         help = nfct_help(ct);
 824         if (help && help->helper)
 825                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 826
 827         nf_conntrack_event_cache(master_ct(ct) ?
 828                                  IPCT_RELATED : IPCT_NEW, ct);
 829         return NF_ACCEPT;
 830
 831 out:
 832         nf_ct_add_to_dying_list(ct);
 833         ret = nf_ct_resolve_clash(net, skb, ctinfo, h);
 834 dying:
 835         nf_conntrack_double_unlock(hash, reply_hash);
 836         NF_CT_STAT_INC(net, insert_failed);
 837         local_bh_enable();
 838         return ret;
 839 }
 840 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 841
 842 /* Returns true if a connection correspondings to the tuple (required
 843    for NAT). */
 844 int
 845 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 846                          const struct nf_conn *ignored_conntrack)
 847 {
 848         struct net *net = nf_ct_net(ignored_conntrack);
 849         const struct nf_conntrack_zone *zone;
 850         struct nf_conntrack_tuple_hash *h;
 851         struct hlist_nulls_head *ct_hash;
 852         unsigned int hash, hsize;
 853         struct hlist_nulls_node *n;
 854         struct nf_conn *ct;
 855
 856         zone = nf_ct_zone(ignored_conntrack);
 857
 858         rcu_read_lock();
 859  begin:
 860         nf_conntrack_get_ht(&ct_hash, &hsize);
 861         hash = __hash_conntrack(net, tuple, hsize);
 862
 863         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
 864                 ct = nf_ct_tuplehash_to_ctrack(h);
 865
 866                 if (ct == ignored_conntrack)
 867                         continue;
 868
 869                 if (nf_ct_is_expired(ct)) {
 870                         nf_ct_gc_expired(ct);
 871                         continue;
 872                 }
 873
 874                 if (nf_ct_key_equal(h, tuple, zone, net)) {
 875                         NF_CT_STAT_INC_ATOMIC(net, found);
 876                         rcu_read_unlock();
 877                         return 1;
 878                 }
 879         }
 880
 881         if (get_nulls_value(n) != hash) {
 882                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 883                 goto begin;
 884         }
 885
 886         rcu_read_unlock();
 887
 888         return 0;
 889 }
 890 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 891
 892 #define NF_CT_EVICTION_RANGE    8
 893
 894 /* There's a small race here where we may free a just-assured
 895    connection.  Too bad: we're in trouble anyway. */
 896 static unsigned int early_drop_list(struct net *net,
 897                                     struct hlist_nulls_head *head)
 898 {
 899         struct nf_conntrack_tuple_hash *h;
 900         struct hlist_nulls_node *n;
 901         unsigned int drops = 0;
 902         struct nf_conn *tmp;
 903
 904         hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
 905                 tmp = nf_ct_tuplehash_to_ctrack(h);
 906
 907                 if (nf_ct_is_expired(tmp)) {
 908                         nf_ct_gc_expired(tmp);
 909                         continue;
 910                 }
 911
 912                 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
 913                     !net_eq(nf_ct_net(tmp), net) ||
 914                     nf_ct_is_dying(tmp))
 915                         continue;
 916
 917                 if (!atomic_inc_not_zero(&tmp->ct_general.use))
 918                         continue;
 919
 920                 /* kill only if still in same netns -- might have moved due to
 921                  * SLAB_TYPESAFE_BY_RCU rules.
 922                  *
 923                  * We steal the timer reference.  If that fails timer has
 924                  * already fired or someone else deleted it. Just drop ref
 925                  * and move to next entry.
 926                  */
 927                 if (net_eq(nf_ct_net(tmp), net) &&
 928                     nf_ct_is_confirmed(tmp) &&
 929                     nf_ct_delete(tmp, 0, 0))
 930                         drops++;
 931
 932                 nf_ct_put(tmp);
 933         }
 934
 935         return drops;
 936 }
 937
 938 static noinline int early_drop(struct net *net, unsigned int _hash)
 939 {
 940         unsigned int i;
 941
 942         for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
 943                 struct hlist_nulls_head *ct_hash;
 944                 unsigned int hash, hsize, drops;
 945
 946                 rcu_read_lock();
 947                 nf_conntrack_get_ht(&ct_hash, &hsize);
 948                 hash = reciprocal_scale(_hash++, hsize);
 949
 950                 drops = early_drop_list(net, &ct_hash[hash]);
 951                 rcu_read_unlock();
 952
 953                 if (drops) {
 954                         NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
 955                         return true;
 956                 }
 957         }
 958
 959         return false;
 960 }
 961
 962 static void gc_worker(struct work_struct *work)
 963 {
 964         unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
 965         unsigned int i, goal, buckets = 0, expired_count = 0;
 966         struct conntrack_gc_work *gc_work;
 967         unsigned int ratio, scanned = 0;
 968         unsigned long next_run;
 969
 970         gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
 971
 972         goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
 973         i = gc_work->last_bucket;
 974
 975         do {
 976                 struct nf_conntrack_tuple_hash *h;
 977                 struct hlist_nulls_head *ct_hash;
 978                 struct hlist_nulls_node *n;
 979                 unsigned int hashsz;
 980                 struct nf_conn *tmp;
 981
 982                 i++;
 983                 rcu_read_lock();
 984
 985                 nf_conntrack_get_ht(&ct_hash, &hashsz);
 986                 if (i >= hashsz)
 987                         i = 0;
 988
 989                 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
 990                         tmp = nf_ct_tuplehash_to_ctrack(h);
 991
 992                         scanned++;
 993                         if (nf_ct_is_expired(tmp)) {
 994                                 nf_ct_gc_expired(tmp);
 995                                 expired_count++;
 996                                 continue;
 997                         }
 998                 }
 999
1000                 /* could check get_nulls_value() here and restart if ct
1001                  * was moved to another chain.  But given gc is best-effort
1002                  * we will just continue with next hash slot.
1003                  */
1004                 rcu_read_unlock();
1005                 cond_resched_rcu_qs();
1006         } while (++buckets < goal);
1007
1008         if (gc_work->exiting)
1009                 return;
1010
1011         /*
1012          * Eviction will normally happen from the packet path, and not
1013          * from this gc worker.
1014          *
1015          * This worker is only here to reap expired entries when system went
1016          * idle after a busy period.
1017          *
1018          * The heuristics below are supposed to balance conflicting goals:
1019          *
1020          * 1. Minimize time until we notice a stale entry
1021          * 2. Maximize scan intervals to not waste cycles
1022          *
1023          * Normally, expire ratio will be close to 0.
1024          *
1025          * As soon as a sizeable fraction of the entries have expired
1026          * increase scan frequency.
1027          */
1028         ratio = scanned ? expired_count * 100 / scanned : 0;
1029         if (ratio > GC_EVICT_RATIO) {
1030                 gc_work->next_gc_run = min_interval;
1031         } else {
1032                 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1033
1034                 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1035
1036                 gc_work->next_gc_run += min_interval;
1037                 if (gc_work->next_gc_run > max)
1038                         gc_work->next_gc_run = max;
1039         }
1040
1041         next_run = gc_work->next_gc_run;
1042         gc_work->last_bucket = i;
1043         queue_delayed_work(system_long_wq, &gc_work->dwork, next_run);
1044 }
1045
1046 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1047 {
1048         INIT_DELAYED_WORK(&gc_work->dwork, gc_worker);
1049         gc_work->next_gc_run = HZ;
1050         gc_work->exiting = false;
1051 }
1052
1053 static struct nf_conn *
1054 __nf_conntrack_alloc(struct net *net,
1055                      const struct nf_conntrack_zone *zone,
1056                      const struct nf_conntrack_tuple *orig,
1057                      const struct nf_conntrack_tuple *repl,
1058                      gfp_t gfp, u32 hash)
1059 {
1060         struct nf_conn *ct;
1061
1062         /* We don't want any race condition at early drop stage */
1063         atomic_inc(&net->ct.count);
1064
1065         if (nf_conntrack_max &&
1066             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
1067                 if (!early_drop(net, hash)) {
1068                         atomic_dec(&net->ct.count);
1069                         net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1070                         return ERR_PTR(-ENOMEM);
1071                 }
1072         }
1073
1074         /*
1075          * Do not use kmem_cache_zalloc(), as this cache uses
1076          * SLAB_TYPESAFE_BY_RCU.
1077          */
1078         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1079         if (ct == NULL)
1080                 goto out;
1081
1082         spin_lock_init(&ct->lock);
1083         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1084         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1085         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1086         /* save hash for reusing when confirming */
1087         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1088         ct->status = 0;
1089         write_pnet(&ct->ct_net, net);
1090         memset(&ct->__nfct_init_offset[0], 0,
1091                offsetof(struct nf_conn, proto) -
1092                offsetof(struct nf_conn, __nfct_init_offset[0]));
1093
1094         nf_ct_zone_add(ct, zone);
1095
1096         /* Because we use RCU lookups, we set ct_general.use to zero before
1097          * this is inserted in any list.
1098          */
1099         atomic_set(&ct->ct_general.use, 0);
1100         return ct;
1101 out:
1102         atomic_dec(&net->ct.count);
1103         return ERR_PTR(-ENOMEM);
1104 }
1105
1106 struct nf_conn *nf_conntrack_alloc(struct net *net,
1107                                    const struct nf_conntrack_zone *zone,
1108                                    const struct nf_conntrack_tuple *orig,
1109                                    const struct nf_conntrack_tuple *repl,
1110                                    gfp_t gfp)
1111 {
1112         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1113 }
1114 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1115
1116 void nf_conntrack_free(struct nf_conn *ct)
1117 {
1118         struct net *net = nf_ct_net(ct);
1119
1120         /* A freed object has refcnt == 0, that's
1121          * the golden rule for SLAB_TYPESAFE_BY_RCU
1122          */
1123         NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 0);
1124
1125         nf_ct_ext_destroy(ct);
1126         nf_ct_ext_free(ct);
1127         kmem_cache_free(nf_conntrack_cachep, ct);
1128         smp_mb__before_atomic();
1129         atomic_dec(&net->ct.count);
1130 }
1131 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1132
1133
1134 /* Allocate a new conntrack: we return -ENOMEM if classification
1135    failed due to stress.  Otherwise it really is unclassifiable. */
1136 static struct nf_conntrack_tuple_hash *
1137 init_conntrack(struct net *net, struct nf_conn *tmpl,
1138                const struct nf_conntrack_tuple *tuple,
1139                struct nf_conntrack_l3proto *l3proto,
1140                struct nf_conntrack_l4proto *l4proto,
1141                struct sk_buff *skb,
1142                unsigned int dataoff, u32 hash)
1143 {
1144         struct nf_conn *ct;
1145         struct nf_conn_help *help;
1146         struct nf_conntrack_tuple repl_tuple;
1147         struct nf_conntrack_ecache *ecache;
1148         struct nf_conntrack_expect *exp = NULL;
1149         const struct nf_conntrack_zone *zone;
1150         struct nf_conn_timeout *timeout_ext;
1151         struct nf_conntrack_zone tmp;
1152         unsigned int *timeouts;
1153
1154         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
1155                 pr_debug("Can't invert tuple.\n");
1156                 return NULL;
1157         }
1158
1159         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1160         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1161                                   hash);
1162         if (IS_ERR(ct))
1163                 return (struct nf_conntrack_tuple_hash *)ct;
1164
1165         if (!nf_ct_add_synproxy(ct, tmpl)) {
1166                 nf_conntrack_free(ct);
1167                 return ERR_PTR(-ENOMEM);
1168         }
1169
1170         timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1171         if (timeout_ext) {
1172                 timeouts = nf_ct_timeout_data(timeout_ext);
1173                 if (unlikely(!timeouts))
1174                         timeouts = l4proto->get_timeouts(net);
1175         } else {
1176                 timeouts = l4proto->get_timeouts(net);
1177         }
1178
1179         if (!l4proto->new(ct, skb, dataoff, timeouts)) {
1180                 nf_conntrack_free(ct);
1181                 pr_debug("can't track with proto module\n");
1182                 return NULL;
1183         }
1184
1185         if (timeout_ext)
1186                 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1187                                       GFP_ATOMIC);
1188
1189         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1190         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1191         nf_ct_labels_ext_add(ct);
1192
1193         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1194         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1195                                  ecache ? ecache->expmask : 0,
1196                              GFP_ATOMIC);
1197
1198         local_bh_disable();
1199         if (net->ct.expect_count) {
1200                 spin_lock(&nf_conntrack_expect_lock);
1201                 exp = nf_ct_find_expectation(net, zone, tuple);
1202                 if (exp) {
1203                         pr_debug("expectation arrives ct=%p exp=%p\n",
1204                                  ct, exp);
1205                         /* Welcome, Mr. Bond.  We've been expecting you... */
1206                         __set_bit(IPS_EXPECTED_BIT, &ct->status);
1207                         /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1208                         ct->master = exp->master;
1209                         if (exp->helper) {
1210                                 help = nf_ct_helper_ext_add(ct, exp->helper,
1211                                                             GFP_ATOMIC);
1212                                 if (help)
1213                                         rcu_assign_pointer(help->helper, exp->helper);
1214                         }
1215
1216 #ifdef CONFIG_NF_CONNTRACK_MARK
1217                         ct->mark = exp->master->mark;
1218 #endif
1219 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1220                         ct->secmark = exp->master->secmark;
1221 #endif
1222                         NF_CT_STAT_INC(net, expect_new);
1223                 }
1224                 spin_unlock(&nf_conntrack_expect_lock);
1225         }
1226         if (!exp)
1227                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1228
1229         /* Now it is inserted into the unconfirmed list, bump refcount */
1230         nf_conntrack_get(&ct->ct_general);
1231         nf_ct_add_to_unconfirmed_list(ct);
1232
1233         local_bh_enable();
1234
1235         if (exp) {
1236                 if (exp->expectfn)
1237                         exp->expectfn(ct, exp);
1238                 nf_ct_expect_put(exp);
1239         }
1240
1241         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1242 }
1243
1244 /* On success, returns conntrack ptr, sets skb->_nfct | ctinfo */
1245 static inline struct nf_conn *
1246 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
1247                   struct sk_buff *skb,
1248                   unsigned int dataoff,
1249                   u_int16_t l3num,
1250                   u_int8_t protonum,
1251                   struct nf_conntrack_l3proto *l3proto,
1252                   struct nf_conntrack_l4proto *l4proto,
1253                   int *set_reply,
1254                   enum ip_conntrack_info *ctinfo)
1255 {
1256         const struct nf_conntrack_zone *zone;
1257         struct nf_conntrack_tuple tuple;
1258         struct nf_conntrack_tuple_hash *h;
1259         struct nf_conntrack_zone tmp;
1260         struct nf_conn *ct;
1261         u32 hash;
1262
1263         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1264                              dataoff, l3num, protonum, net, &tuple, l3proto,
1265                              l4proto)) {
1266                 pr_debug("Can't get tuple\n");
1267                 return NULL;
1268         }
1269
1270         /* look for tuple match */
1271         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1272         hash = hash_conntrack_raw(&tuple, net);
1273         h = __nf_conntrack_find_get(net, zone, &tuple, hash);
1274         if (!h) {
1275                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
1276                                    skb, dataoff, hash);
1277                 if (!h)
1278                         return NULL;
1279                 if (IS_ERR(h))
1280                         return (void *)h;
1281         }
1282         ct = nf_ct_tuplehash_to_ctrack(h);
1283
1284         /* It exists; we have (non-exclusive) reference. */
1285         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1286                 *ctinfo = IP_CT_ESTABLISHED_REPLY;
1287                 /* Please set reply bit if this packet OK */
1288                 *set_reply = 1;
1289         } else {
1290                 /* Once we've had two way comms, always ESTABLISHED. */
1291                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1292                         pr_debug("normal packet for %p\n", ct);
1293                         *ctinfo = IP_CT_ESTABLISHED;
1294                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1295                         pr_debug("related packet for %p\n", ct);
1296                         *ctinfo = IP_CT_RELATED;
1297                 } else {
1298                         pr_debug("new packet for %p\n", ct);
1299                         *ctinfo = IP_CT_NEW;
1300                 }
1301                 *set_reply = 0;
1302         }
1303         nf_ct_set(skb, ct, *ctinfo);
1304         return ct;
1305 }
1306
1307 unsigned int
1308 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
1309                 struct sk_buff *skb)
1310 {
1311         struct nf_conn *ct, *tmpl;
1312         enum ip_conntrack_info ctinfo;
1313         struct nf_conntrack_l3proto *l3proto;
1314         struct nf_conntrack_l4proto *l4proto;
1315         unsigned int *timeouts;
1316         unsigned int dataoff;
1317         u_int8_t protonum;
1318         int set_reply = 0;
1319         int ret;
1320
1321         tmpl = nf_ct_get(skb, &ctinfo);
1322         if (tmpl) {
1323                 /* Previously seen (loopback or untracked)?  Ignore. */
1324                 if (!nf_ct_is_template(tmpl)) {
1325                         NF_CT_STAT_INC_ATOMIC(net, ignore);
1326                         return NF_ACCEPT;
1327                 }
1328                 skb->_nfct = 0;
1329         }
1330
1331         /* rcu_read_lock()ed by nf_hook_thresh */
1332         l3proto = __nf_ct_l3proto_find(pf);
1333         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
1334                                    &dataoff, &protonum);
1335         if (ret <= 0) {
1336                 pr_debug("not prepared to track yet or error occurred\n");
1337                 NF_CT_STAT_INC_ATOMIC(net, error);
1338                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1339                 ret = -ret;
1340                 goto out;
1341         }
1342
1343         l4proto = __nf_ct_l4proto_find(pf, protonum);
1344
1345         /* It may be an special packet, error, unclean...
1346          * inverse of the return code tells to the netfilter
1347          * core what to do with the packet. */
1348         if (l4proto->error != NULL) {
1349                 ret = l4proto->error(net, tmpl, skb, dataoff, pf, hooknum);
1350                 if (ret <= 0) {
1351                         NF_CT_STAT_INC_ATOMIC(net, error);
1352                         NF_CT_STAT_INC_ATOMIC(net, invalid);
1353                         ret = -ret;
1354                         goto out;
1355                 }
1356                 /* ICMP[v6] protocol trackers may assign one conntrack. */
1357                 if (skb->_nfct)
1358                         goto out;
1359         }
1360 repeat:
1361         ct = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
1362                                l3proto, l4proto, &set_reply, &ctinfo);
1363         if (!ct) {
1364                 /* Not valid part of a connection */
1365                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1366                 ret = NF_ACCEPT;
1367                 goto out;
1368         }
1369
1370         if (IS_ERR(ct)) {
1371                 /* Too stressed to deal. */
1372                 NF_CT_STAT_INC_ATOMIC(net, drop);
1373                 ret = NF_DROP;
1374                 goto out;
1375         }
1376
1377         NF_CT_ASSERT(skb_nfct(skb));
1378
1379         /* Decide what timeout policy we want to apply to this flow. */
1380         timeouts = nf_ct_timeout_lookup(net, ct, l4proto);
1381
1382         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum, timeouts);
1383         if (ret <= 0) {
1384                 /* Invalid: inverse of the return code tells
1385                  * the netfilter core what to do */
1386                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1387                 nf_conntrack_put(&ct->ct_general);
1388                 skb->_nfct = 0;
1389                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1390                 if (ret == -NF_DROP)
1391                         NF_CT_STAT_INC_ATOMIC(net, drop);
1392                 /* Special case: TCP tracker reports an attempt to reopen a
1393                  * closed/aborted connection. We have to go back and create a
1394                  * fresh conntrack.
1395                  */
1396                 if (ret == -NF_REPEAT)
1397                         goto repeat;
1398                 ret = -ret;
1399                 goto out;
1400         }
1401
1402         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1403                 nf_conntrack_event_cache(IPCT_REPLY, ct);
1404 out:
1405         if (tmpl)
1406                 nf_ct_put(tmpl);
1407
1408         return ret;
1409 }
1410 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1411
1412 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
1413                           const struct nf_conntrack_tuple *orig)
1414 {
1415         bool ret;
1416
1417         rcu_read_lock();
1418         ret = nf_ct_invert_tuple(inverse, orig,
1419                                  __nf_ct_l3proto_find(orig->src.l3num),
1420                                  __nf_ct_l4proto_find(orig->src.l3num,
1421                                                       orig->dst.protonum));
1422         rcu_read_unlock();
1423         return ret;
1424 }
1425 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
1426
1427 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
1428    implicitly racy: see __nf_conntrack_confirm */
1429 void nf_conntrack_alter_reply(struct nf_conn *ct,
1430                               const struct nf_conntrack_tuple *newreply)
1431 {
1432         struct nf_conn_help *help = nfct_help(ct);
1433
1434         /* Should be unconfirmed, so not in hash table yet */
1435         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
1436
1437         pr_debug("Altering reply tuple of %p to ", ct);
1438         nf_ct_dump_tuple(newreply);
1439
1440         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1441         if (ct->master || (help && !hlist_empty(&help->expectations)))
1442                 return;
1443
1444         rcu_read_lock();
1445         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1446         rcu_read_unlock();
1447 }
1448 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1449
1450 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1451 void __nf_ct_refresh_acct(struct nf_conn *ct,
1452                           enum ip_conntrack_info ctinfo,
1453                           const struct sk_buff *skb,
1454                           unsigned long extra_jiffies,
1455                           int do_acct)
1456 {
1457         NF_CT_ASSERT(skb);
1458
1459         /* Only update if this is not a fixed timeout */
1460         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1461                 goto acct;
1462
1463         /* If not in hash table, timer will not be active yet */
1464         if (nf_ct_is_confirmed(ct))
1465                 extra_jiffies += nfct_time_stamp;
1466
1467         ct->timeout = extra_jiffies;
1468 acct:
1469         if (do_acct)
1470                 nf_ct_acct_update(ct, ctinfo, skb->len);
1471 }
1472 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1473
1474 bool nf_ct_kill_acct(struct nf_conn *ct,
1475                      enum ip_conntrack_info ctinfo,
1476                      const struct sk_buff *skb)
1477 {
1478         nf_ct_acct_update(ct, ctinfo, skb->len);
1479
1480         return nf_ct_delete(ct, 0, 0);
1481 }
1482 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1483
1484 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1485
1486 #include <linux/netfilter/nfnetlink.h>
1487 #include <linux/netfilter/nfnetlink_conntrack.h>
1488 #include <linux/mutex.h>
1489
1490 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
1491  * in ip_conntrack_core, since we don't want the protocols to autoload
1492  * or depend on ctnetlink */
1493 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1494                                const struct nf_conntrack_tuple *tuple)
1495 {
1496         if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1497             nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1498                 goto nla_put_failure;
1499         return 0;
1500
1501 nla_put_failure:
1502         return -1;
1503 }
1504 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1505
1506 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1507         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
1508         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
1509 };
1510 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1511
1512 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1513                                struct nf_conntrack_tuple *t)
1514 {
1515         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
1516                 return -EINVAL;
1517
1518         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1519         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1520
1521         return 0;
1522 }
1523 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1524
1525 int nf_ct_port_nlattr_tuple_size(void)
1526 {
1527         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1528 }
1529 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1530 #endif
1531
1532 /* Used by ipt_REJECT and ip6t_REJECT. */
1533 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
1534 {
1535         struct nf_conn *ct;
1536         enum ip_conntrack_info ctinfo;
1537
1538         /* This ICMP is in reverse direction to the packet which caused it */
1539         ct = nf_ct_get(skb, &ctinfo);
1540         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1541                 ctinfo = IP_CT_RELATED_REPLY;
1542         else
1543                 ctinfo = IP_CT_RELATED;
1544
1545         /* Attach to new skbuff, and increment count */
1546         nf_ct_set(nskb, ct, ctinfo);
1547         nf_conntrack_get(skb_nfct(nskb));
1548 }
1549
1550 /* Bring out ya dead! */
1551 static struct nf_conn *
1552 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1553                 void *data, unsigned int *bucket)
1554 {
1555         struct nf_conntrack_tuple_hash *h;
1556         struct nf_conn *ct;
1557         struct hlist_nulls_node *n;
1558         int cpu;
1559         spinlock_t *lockp;
1560
1561         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1562                 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
1563                 local_bh_disable();
1564                 nf_conntrack_lock(lockp);
1565                 if (*bucket < nf_conntrack_htable_size) {
1566                         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
1567                                 if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1568                                         continue;
1569                                 ct = nf_ct_tuplehash_to_ctrack(h);
1570                                 if (net_eq(nf_ct_net(ct), net) &&
1571                                     iter(ct, data))
1572                                         goto found;
1573                         }
1574                 }
1575                 spin_unlock(lockp);
1576                 local_bh_enable();
1577                 cond_resched();
1578         }
1579
1580         for_each_possible_cpu(cpu) {
1581                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1582
1583                 spin_lock_bh(&pcpu->lock);
1584                 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
1585                         ct = nf_ct_tuplehash_to_ctrack(h);
1586                         if (iter(ct, data))
1587                                 set_bit(IPS_DYING_BIT, &ct->status);
1588                 }
1589                 spin_unlock_bh(&pcpu->lock);
1590                 cond_resched();
1591         }
1592         return NULL;
1593 found:
1594         atomic_inc(&ct->ct_general.use);
1595         spin_unlock(lockp);
1596         local_bh_enable();
1597         return ct;
1598 }
1599
1600 void nf_ct_iterate_cleanup(struct net *net,
1601                            int (*iter)(struct nf_conn *i, void *data),
1602                            void *data, u32 portid, int report)
1603 {
1604         struct nf_conn *ct;
1605         unsigned int bucket = 0;
1606
1607         might_sleep();
1608
1609         if (atomic_read(&net->ct.count) == 0)
1610                 return;
1611
1612         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1613                 /* Time to push up daises... */
1614
1615                 nf_ct_delete(ct, portid, report);
1616                 nf_ct_put(ct);
1617                 cond_resched();
1618         }
1619 }
1620 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1621
1622 static int kill_all(struct nf_conn *i, void *data)
1623 {
1624         return 1;
1625 }
1626
1627 void nf_ct_free_hashtable(void *hash, unsigned int size)
1628 {
1629         if (is_vmalloc_addr(hash))
1630                 vfree(hash);
1631         else
1632                 free_pages((unsigned long)hash,
1633                            get_order(sizeof(struct hlist_head) * size));
1634 }
1635 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1636
1637 static int untrack_refs(void)
1638 {
1639         int cnt = 0, cpu;
1640
1641         for_each_possible_cpu(cpu) {
1642                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1643
1644                 cnt += atomic_read(&ct->ct_general.use) - 1;
1645         }
1646         return cnt;
1647 }
1648
1649 void nf_conntrack_cleanup_start(void)
1650 {
1651         conntrack_gc_work.exiting = true;
1652         RCU_INIT_POINTER(ip_ct_attach, NULL);
1653 }
1654
1655 void nf_conntrack_cleanup_end(void)
1656 {
1657         RCU_INIT_POINTER(nf_ct_destroy, NULL);
1658         while (untrack_refs() > 0)
1659                 schedule();
1660
1661         cancel_delayed_work_sync(&conntrack_gc_work.dwork);
1662         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1663
1664         nf_conntrack_proto_fini();
1665         nf_conntrack_seqadj_fini();
1666         nf_conntrack_labels_fini();
1667         nf_conntrack_helper_fini();
1668         nf_conntrack_timeout_fini();
1669         nf_conntrack_ecache_fini();
1670         nf_conntrack_tstamp_fini();
1671         nf_conntrack_acct_fini();
1672         nf_conntrack_expect_fini();
1673
1674         kmem_cache_destroy(nf_conntrack_cachep);
1675 }
1676
1677 /*
1678  * Mishearing the voices in his head, our hero wonders how he's
1679  * supposed to kill the mall.
1680  */
1681 void nf_conntrack_cleanup_net(struct net *net)
1682 {
1683         LIST_HEAD(single);
1684
1685         list_add(&net->exit_list, &single);
1686         nf_conntrack_cleanup_net_list(&single);
1687 }
1688
1689 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
1690 {
1691         int busy;
1692         struct net *net;
1693
1694         /*
1695          * This makes sure all current packets have passed through
1696          *  netfilter framework.  Roll on, two-stage module
1697          *  delete...
1698          */
1699         synchronize_net();
1700 i_see_dead_people:
1701         busy = 0;
1702         list_for_each_entry(net, net_exit_list, exit_list) {
1703                 nf_ct_iterate_cleanup(net, kill_all, NULL, 0, 0);
1704                 if (atomic_read(&net->ct.count) != 0)
1705                         busy = 1;
1706         }
1707         if (busy) {
1708                 schedule();
1709                 goto i_see_dead_people;
1710         }
1711
1712         list_for_each_entry(net, net_exit_list, exit_list) {
1713                 nf_conntrack_proto_pernet_fini(net);
1714                 nf_conntrack_helper_pernet_fini(net);
1715                 nf_conntrack_ecache_pernet_fini(net);
1716                 nf_conntrack_tstamp_pernet_fini(net);
1717                 nf_conntrack_acct_pernet_fini(net);
1718                 nf_conntrack_expect_pernet_fini(net);
1719                 free_percpu(net->ct.stat);
1720                 free_percpu(net->ct.pcpu_lists);
1721         }
1722 }
1723
1724 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
1725 {
1726         struct hlist_nulls_head *hash;
1727         unsigned int nr_slots, i;
1728         size_t sz;
1729
1730         if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1731                 return NULL;
1732
1733         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1734         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1735
1736         if (nr_slots > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1737                 return NULL;
1738
1739         sz = nr_slots * sizeof(struct hlist_nulls_head);
1740         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1741                                         get_order(sz));
1742         if (!hash)
1743                 hash = vzalloc(sz);
1744
1745         if (hash && nulls)
1746                 for (i = 0; i < nr_slots; i++)
1747                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1748
1749         return hash;
1750 }
1751 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1752
1753 int nf_conntrack_hash_resize(unsigned int hashsize)
1754 {
1755         int i, bucket;
1756         unsigned int old_size;
1757         struct hlist_nulls_head *hash, *old_hash;
1758         struct nf_conntrack_tuple_hash *h;
1759         struct nf_conn *ct;
1760
1761         if (!hashsize)
1762                 return -EINVAL;
1763
1764         hash = nf_ct_alloc_hashtable(&hashsize, 1);
1765         if (!hash)
1766                 return -ENOMEM;
1767
1768         old_size = nf_conntrack_htable_size;
1769         if (old_size == hashsize) {
1770                 nf_ct_free_hashtable(hash, hashsize);
1771                 return 0;
1772         }
1773
1774         local_bh_disable();
1775         nf_conntrack_all_lock();
1776         write_seqcount_begin(&nf_conntrack_generation);
1777
1778         /* Lookups in the old hash might happen in parallel, which means we
1779          * might get false negatives during connection lookup. New connections
1780          * created because of a false negative won't make it into the hash
1781          * though since that required taking the locks.
1782          */
1783
1784         for (i = 0; i < nf_conntrack_htable_size; i++) {
1785                 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
1786                         h = hlist_nulls_entry(nf_conntrack_hash[i].first,
1787                                               struct nf_conntrack_tuple_hash, hnnode);
1788                         ct = nf_ct_tuplehash_to_ctrack(h);
1789                         hlist_nulls_del_rcu(&h->hnnode);
1790                         bucket = __hash_conntrack(nf_ct_net(ct),
1791                                                   &h->tuple, hashsize);
1792                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1793                 }
1794         }
1795         old_size = nf_conntrack_htable_size;
1796         old_hash = nf_conntrack_hash;
1797
1798         nf_conntrack_hash = hash;
1799         nf_conntrack_htable_size = hashsize;
1800
1801         write_seqcount_end(&nf_conntrack_generation);
1802         nf_conntrack_all_unlock();
1803         local_bh_enable();
1804
1805         synchronize_net();
1806         nf_ct_free_hashtable(old_hash, old_size);
1807         return 0;
1808 }
1809
1810 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1811 {
1812         unsigned int hashsize;
1813         int rc;
1814
1815         if (current->nsproxy->net_ns != &init_net)
1816                 return -EOPNOTSUPP;
1817
1818         /* On boot, we can set this without any fancy locking. */
1819         if (!nf_conntrack_htable_size)
1820                 return param_set_uint(val, kp);
1821
1822         rc = kstrtouint(val, 0, &hashsize);
1823         if (rc)
1824                 return rc;
1825
1826         return nf_conntrack_hash_resize(hashsize);
1827 }
1828 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1829
1830 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1831                   &nf_conntrack_htable_size, 0600);
1832
1833 void nf_ct_untracked_status_or(unsigned long bits)
1834 {
1835         int cpu;
1836
1837         for_each_possible_cpu(cpu)
1838                 per_cpu(nf_conntrack_untracked, cpu).status |= bits;
1839 }
1840 EXPORT_SYMBOL_GPL(nf_ct_untracked_status_or);
1841
1842 int nf_conntrack_init_start(void)
1843 {
1844         int max_factor = 8;
1845         int ret = -ENOMEM;
1846         int i, cpu;
1847
1848         seqcount_init(&nf_conntrack_generation);
1849
1850         for (i = 0; i < CONNTRACK_LOCKS; i++)
1851                 spin_lock_init(&nf_conntrack_locks[i]);
1852
1853         if (!nf_conntrack_htable_size) {
1854                 /* Idea from tcp.c: use 1/16384 of memory.
1855                  * On i386: 32MB machine has 512 buckets.
1856                  * >= 1GB machines have 16384 buckets.
1857                  * >= 4GB machines have 65536 buckets.
1858                  */
1859                 nf_conntrack_htable_size
1860                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1861                            / sizeof(struct hlist_head));
1862                 if (totalram_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
1863                         nf_conntrack_htable_size = 65536;
1864                 else if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1865                         nf_conntrack_htable_size = 16384;
1866                 if (nf_conntrack_htable_size < 32)
1867                         nf_conntrack_htable_size = 32;
1868
1869                 /* Use a max. factor of four by default to get the same max as
1870                  * with the old struct list_heads. When a table size is given
1871                  * we use the old value of 8 to avoid reducing the max.
1872                  * entries. */
1873                 max_factor = 4;
1874         }
1875
1876         nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
1877         if (!nf_conntrack_hash)
1878                 return -ENOMEM;
1879
1880         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1881
1882         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
1883                                                 sizeof(struct nf_conn),
1884                                                 NFCT_INFOMASK + 1,
1885                                                 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
1886         if (!nf_conntrack_cachep)
1887                 goto err_cachep;
1888
1889         printk(KERN_INFO "nf_conntrack version %s (%u buckets, %d max)\n",
1890                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1891                nf_conntrack_max);
1892
1893         ret = nf_conntrack_expect_init();
1894         if (ret < 0)
1895                 goto err_expect;
1896
1897         ret = nf_conntrack_acct_init();
1898         if (ret < 0)
1899                 goto err_acct;
1900
1901         ret = nf_conntrack_tstamp_init();
1902         if (ret < 0)
1903                 goto err_tstamp;
1904
1905         ret = nf_conntrack_ecache_init();
1906         if (ret < 0)
1907                 goto err_ecache;
1908
1909         ret = nf_conntrack_timeout_init();
1910         if (ret < 0)
1911                 goto err_timeout;
1912
1913         ret = nf_conntrack_helper_init();
1914         if (ret < 0)
1915                 goto err_helper;
1916
1917         ret = nf_conntrack_labels_init();
1918         if (ret < 0)
1919                 goto err_labels;
1920
1921         ret = nf_conntrack_seqadj_init();
1922         if (ret < 0)
1923                 goto err_seqadj;
1924
1925         ret = nf_conntrack_proto_init();
1926         if (ret < 0)
1927                 goto err_proto;
1928
1929         /* Set up fake conntrack: to never be deleted, not in any hashes */
1930         for_each_possible_cpu(cpu) {
1931                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1932                 write_pnet(&ct->ct_net, &init_net);
1933                 atomic_set(&ct->ct_general.use, 1);
1934         }
1935         /*  - and look it like as a confirmed connection */
1936         nf_ct_untracked_status_or(IPS_CONFIRMED | IPS_UNTRACKED);
1937
1938         conntrack_gc_work_init(&conntrack_gc_work);
1939         queue_delayed_work(system_long_wq, &conntrack_gc_work.dwork, HZ);
1940
1941         return 0;
1942
1943 err_proto:
1944         nf_conntrack_seqadj_fini();
1945 err_seqadj:
1946         nf_conntrack_labels_fini();
1947 err_labels:
1948         nf_conntrack_helper_fini();
1949 err_helper:
1950         nf_conntrack_timeout_fini();
1951 err_timeout:
1952         nf_conntrack_ecache_fini();
1953 err_ecache:
1954         nf_conntrack_tstamp_fini();
1955 err_tstamp:
1956         nf_conntrack_acct_fini();
1957 err_acct:
1958         nf_conntrack_expect_fini();
1959 err_expect:
1960         kmem_cache_destroy(nf_conntrack_cachep);
1961 err_cachep:
1962         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1963         return ret;
1964 }
1965
1966 void nf_conntrack_init_end(void)
1967 {
1968         /* For use by REJECT target */
1969         RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
1970         RCU_INIT_POINTER(nf_ct_destroy, destroy_conntrack);
1971 }
1972
1973 /*
1974  * We need to use special "null" values, not used in hash table
1975  */
1976 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1977 #define DYING_NULLS_VAL         ((1<<30)+1)
1978 #define TEMPLATE_NULLS_VAL      ((1<<30)+2)
1979
1980 int nf_conntrack_init_net(struct net *net)
1981 {
1982         int ret = -ENOMEM;
1983         int cpu;
1984
1985         atomic_set(&net->ct.count, 0);
1986
1987         net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
1988         if (!net->ct.pcpu_lists)
1989                 goto err_stat;
1990
1991         for_each_possible_cpu(cpu) {
1992                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1993
1994                 spin_lock_init(&pcpu->lock);
1995                 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
1996                 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
1997         }
1998
1999         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2000         if (!net->ct.stat)
2001                 goto err_pcpu_lists;
2002
2003         ret = nf_conntrack_expect_pernet_init(net);
2004         if (ret < 0)
2005                 goto err_expect;
2006         ret = nf_conntrack_acct_pernet_init(net);
2007         if (ret < 0)
2008                 goto err_acct;
2009         ret = nf_conntrack_tstamp_pernet_init(net);
2010         if (ret < 0)
2011                 goto err_tstamp;
2012         ret = nf_conntrack_ecache_pernet_init(net);
2013         if (ret < 0)
2014                 goto err_ecache;
2015         ret = nf_conntrack_helper_pernet_init(net);
2016         if (ret < 0)
2017                 goto err_helper;
2018         ret = nf_conntrack_proto_pernet_init(net);
2019         if (ret < 0)
2020                 goto err_proto;
2021         return 0;
2022
2023 err_proto:
2024         nf_conntrack_helper_pernet_fini(net);
2025 err_helper:
2026         nf_conntrack_ecache_pernet_fini(net);
2027 err_ecache:
2028         nf_conntrack_tstamp_pernet_fini(net);
2029 err_tstamp:
2030         nf_conntrack_acct_pernet_fini(net);
2031 err_acct:
2032         nf_conntrack_expect_pernet_fini(net);
2033 err_expect:
2034         free_percpu(net->ct.stat);
2035 err_pcpu_lists:
2036         free_percpu(net->ct.pcpu_lists);
2037 err_stat:
2038         return ret;
2039 }