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         bool                    early_drop;
  80         long                    next_gc_run;
  81 };
  82
  83 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
  84 static __read_mostly spinlock_t nf_conntrack_locks_all_lock;
  85 static __read_mostly DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
  86 static __read_mostly bool nf_conntrack_locks_all;
  87
  88 /* every gc cycle scans at most 1/GC_MAX_BUCKETS_DIV part of table */
  89 #define GC_MAX_BUCKETS_DIV      128u
  90 /* upper bound of full table scan */
  91 #define GC_MAX_SCAN_JIFFIES     (16u * HZ)
  92 /* desired ratio of entries found to be expired */
  93 #define GC_EVICT_RATIO  50u
  94
  95 static struct conntrack_gc_work conntrack_gc_work;
  96
  97 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
  98 {
  99         spin_lock(lock);
 100         while (unlikely(nf_conntrack_locks_all)) {
 101                 spin_unlock(lock);
 102
 103                 /*
 104                  * Order the 'nf_conntrack_locks_all' load vs. the
 105                  * spin_unlock_wait() loads below, to ensure
 106                  * that 'nf_conntrack_locks_all_lock' is indeed held:
 107                  */
 108                 smp_rmb(); /* spin_lock(&nf_conntrack_locks_all_lock) */
 109                 spin_unlock_wait(&nf_conntrack_locks_all_lock);
 110                 spin_lock(lock);
 111         }
 112 }
 113 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
 114
 115 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 116 {
 117         h1 %= CONNTRACK_LOCKS;
 118         h2 %= CONNTRACK_LOCKS;
 119         spin_unlock(&nf_conntrack_locks[h1]);
 120         if (h1 != h2)
 121                 spin_unlock(&nf_conntrack_locks[h2]);
 122 }
 123
 124 /* return true if we need to recompute hashes (in case hash table was resized) */
 125 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
 126                                      unsigned int h2, unsigned int sequence)
 127 {
 128         h1 %= CONNTRACK_LOCKS;
 129         h2 %= CONNTRACK_LOCKS;
 130         if (h1 <= h2) {
 131                 nf_conntrack_lock(&nf_conntrack_locks[h1]);
 132                 if (h1 != h2)
 133                         spin_lock_nested(&nf_conntrack_locks[h2],
 134                                          SINGLE_DEPTH_NESTING);
 135         } else {
 136                 nf_conntrack_lock(&nf_conntrack_locks[h2]);
 137                 spin_lock_nested(&nf_conntrack_locks[h1],
 138                                  SINGLE_DEPTH_NESTING);
 139         }
 140         if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
 141                 nf_conntrack_double_unlock(h1, h2);
 142                 return true;
 143         }
 144         return false;
 145 }
 146
 147 static void nf_conntrack_all_lock(void)
 148 {
 149         int i;
 150
 151         spin_lock(&nf_conntrack_locks_all_lock);
 152         nf_conntrack_locks_all = true;
 153
 154         /*
 155          * Order the above store of 'nf_conntrack_locks_all' against
 156          * the spin_unlock_wait() loads below, such that if
 157          * nf_conntrack_lock() observes 'nf_conntrack_locks_all'
 158          * we must observe nf_conntrack_locks[] held:
 159          */
 160         smp_mb(); /* spin_lock(&nf_conntrack_locks_all_lock) */
 161
 162         for (i = 0; i < CONNTRACK_LOCKS; i++) {
 163                 spin_unlock_wait(&nf_conntrack_locks[i]);
 164         }
 165 }
 166
 167 static void nf_conntrack_all_unlock(void)
 168 {
 169         /*
 170          * All prior stores must be complete before we clear
 171          * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
 172          * might observe the false value but not the entire
 173          * critical section:
 174          */
 175         smp_store_release(&nf_conntrack_locks_all, false);
 176         spin_unlock(&nf_conntrack_locks_all_lock);
 177 }
 178
 179 unsigned int nf_conntrack_htable_size __read_mostly;
 180 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
 181
 182 unsigned int nf_conntrack_max __read_mostly;
 183 seqcount_t nf_conntrack_generation __read_mostly;
 184 static unsigned int nf_conntrack_hash_rnd __read_mostly;
 185
 186 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
 187                               const struct net *net)
 188 {
 189         unsigned int n;
 190         u32 seed;
 191
 192         get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
 193
 194         /* The direction must be ignored, so we hash everything up to the
 195          * destination ports (which is a multiple of 4) and treat the last
 196          * three bytes manually.
 197          */
 198         seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
 199         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
 200         return jhash2((u32 *)tuple, n, seed ^
 201                       (((__force __u16)tuple->dst.u.all << 16) |
 202                       tuple->dst.protonum));
 203 }
 204
 205 static u32 scale_hash(u32 hash)
 206 {
 207         return reciprocal_scale(hash, nf_conntrack_htable_size);
 208 }
 209
 210 static u32 __hash_conntrack(const struct net *net,
 211                             const struct nf_conntrack_tuple *tuple,
 212                             unsigned int size)
 213 {
 214         return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
 215 }
 216
 217 static u32 hash_conntrack(const struct net *net,
 218                           const struct nf_conntrack_tuple *tuple)
 219 {
 220         return scale_hash(hash_conntrack_raw(tuple, net));
 221 }
 222
 223 bool
 224 nf_ct_get_tuple(const struct sk_buff *skb,
 225                 unsigned int nhoff,
 226                 unsigned int dataoff,
 227                 u_int16_t l3num,
 228                 u_int8_t protonum,
 229                 struct net *net,
 230                 struct nf_conntrack_tuple *tuple,
 231                 const struct nf_conntrack_l3proto *l3proto,
 232                 const struct nf_conntrack_l4proto *l4proto)
 233 {
 234         memset(tuple, 0, sizeof(*tuple));
 235
 236         tuple->src.l3num = l3num;
 237         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 238                 return false;
 239
 240         tuple->dst.protonum = protonum;
 241         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 242
 243         return l4proto->pkt_to_tuple(skb, dataoff, net, tuple);
 244 }
 245 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 246
 247 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 248                        u_int16_t l3num,
 249                        struct net *net, struct nf_conntrack_tuple *tuple)
 250 {
 251         struct nf_conntrack_l3proto *l3proto;
 252         struct nf_conntrack_l4proto *l4proto;
 253         unsigned int protoff;
 254         u_int8_t protonum;
 255         int ret;
 256
 257         rcu_read_lock();
 258
 259         l3proto = __nf_ct_l3proto_find(l3num);
 260         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
 261         if (ret != NF_ACCEPT) {
 262                 rcu_read_unlock();
 263                 return false;
 264         }
 265
 266         l4proto = __nf_ct_l4proto_find(l3num, protonum);
 267
 268         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple,
 269                               l3proto, l4proto);
 270
 271         rcu_read_unlock();
 272         return ret;
 273 }
 274 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 275
 276 bool
 277 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 278                    const struct nf_conntrack_tuple *orig,
 279                    const struct nf_conntrack_l3proto *l3proto,
 280                    const struct nf_conntrack_l4proto *l4proto)
 281 {
 282         memset(inverse, 0, sizeof(*inverse));
 283
 284         inverse->src.l3num = orig->src.l3num;
 285         if (l3proto->invert_tuple(inverse, orig) == 0)
 286                 return false;
 287
 288         inverse->dst.dir = !orig->dst.dir;
 289
 290         inverse->dst.protonum = orig->dst.protonum;
 291         return l4proto->invert_tuple(inverse, orig);
 292 }
 293 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 294
 295 static void
 296 clean_from_lists(struct nf_conn *ct)
 297 {
 298         pr_debug("clean_from_lists(%p)\n", ct);
 299         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 300         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 301
 302         /* Destroy all pending expectations */
 303         nf_ct_remove_expectations(ct);
 304 }
 305
 306 /* must be called with local_bh_disable */
 307 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
 308 {
 309         struct ct_pcpu *pcpu;
 310
 311         /* add this conntrack to the (per cpu) dying list */
 312         ct->cpu = smp_processor_id();
 313         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 314
 315         spin_lock(&pcpu->lock);
 316         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 317                              &pcpu->dying);
 318         spin_unlock(&pcpu->lock);
 319 }
 320
 321 /* must be called with local_bh_disable */
 322 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
 323 {
 324         struct ct_pcpu *pcpu;
 325
 326         /* add this conntrack to the (per cpu) unconfirmed list */
 327         ct->cpu = smp_processor_id();
 328         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 329
 330         spin_lock(&pcpu->lock);
 331         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 332                              &pcpu->unconfirmed);
 333         spin_unlock(&pcpu->lock);
 334 }
 335
 336 /* must be called with local_bh_disable */
 337 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
 338 {
 339         struct ct_pcpu *pcpu;
 340
 341         /* We overload first tuple to link into unconfirmed or dying list.*/
 342         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 343
 344         spin_lock(&pcpu->lock);
 345         BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 346         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 347         spin_unlock(&pcpu->lock);
 348 }
 349
 350 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
 351
 352 /* Released via destroy_conntrack() */
 353 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
 354                                  const struct nf_conntrack_zone *zone,
 355                                  gfp_t flags)
 356 {
 357         struct nf_conn *tmpl, *p;
 358
 359         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
 360                 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
 361                 if (!tmpl)
 362                         return NULL;
 363
 364                 p = tmpl;
 365                 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 366                 if (tmpl != p) {
 367                         tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 368                         tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
 369                 }
 370         } else {
 371                 tmpl = kzalloc(sizeof(*tmpl), flags);
 372                 if (!tmpl)
 373                         return NULL;
 374         }
 375
 376         tmpl->status = IPS_TEMPLATE;
 377         write_pnet(&tmpl->ct_net, net);
 378         nf_ct_zone_add(tmpl, zone);
 379         atomic_set(&tmpl->ct_general.use, 0);
 380
 381         return tmpl;
 382 }
 383 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
 384
 385 void nf_ct_tmpl_free(struct nf_conn *tmpl)
 386 {
 387         nf_ct_ext_destroy(tmpl);
 388         nf_ct_ext_free(tmpl);
 389
 390         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
 391                 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
 392         else
 393                 kfree(tmpl);
 394 }
 395 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
 396
 397 static void
 398 destroy_conntrack(struct nf_conntrack *nfct)
 399 {
 400         struct nf_conn *ct = (struct nf_conn *)nfct;
 401         struct nf_conntrack_l4proto *l4proto;
 402
 403         pr_debug("destroy_conntrack(%p)\n", ct);
 404         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
 405
 406         if (unlikely(nf_ct_is_template(ct))) {
 407                 nf_ct_tmpl_free(ct);
 408                 return;
 409         }
 410         rcu_read_lock();
 411         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 412         if (l4proto->destroy)
 413                 l4proto->destroy(ct);
 414
 415         rcu_read_unlock();
 416
 417         local_bh_disable();
 418         /* Expectations will have been removed in clean_from_lists,
 419          * except TFTP can create an expectation on the first packet,
 420          * before connection is in the list, so we need to clean here,
 421          * too.
 422          */
 423         nf_ct_remove_expectations(ct);
 424
 425         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 426
 427         local_bh_enable();
 428
 429         if (ct->master)
 430                 nf_ct_put(ct->master);
 431
 432         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 433         nf_conntrack_free(ct);
 434 }
 435
 436 static void nf_ct_delete_from_lists(struct nf_conn *ct)
 437 {
 438         struct net *net = nf_ct_net(ct);
 439         unsigned int hash, reply_hash;
 440         unsigned int sequence;
 441
 442         nf_ct_helper_destroy(ct);
 443
 444         local_bh_disable();
 445         do {
 446                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 447                 hash = hash_conntrack(net,
 448                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 449                 reply_hash = hash_conntrack(net,
 450                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 451         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 452
 453         clean_from_lists(ct);
 454         nf_conntrack_double_unlock(hash, reply_hash);
 455
 456         nf_ct_add_to_dying_list(ct);
 457
 458         local_bh_enable();
 459 }
 460
 461 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
 462 {
 463         struct nf_conn_tstamp *tstamp;
 464
 465         if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
 466                 return false;
 467
 468         tstamp = nf_conn_tstamp_find(ct);
 469         if (tstamp && tstamp->stop == 0)
 470                 tstamp->stop = ktime_get_real_ns();
 471
 472         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 473                                     portid, report) < 0) {
 474                 /* destroy event was not delivered. nf_ct_put will
 475                  * be done by event cache worker on redelivery.
 476                  */
 477                 nf_ct_delete_from_lists(ct);
 478                 nf_conntrack_ecache_delayed_work(nf_ct_net(ct));
 479                 return false;
 480         }
 481
 482         nf_conntrack_ecache_work(nf_ct_net(ct));
 483         nf_ct_delete_from_lists(ct);
 484         nf_ct_put(ct);
 485         return true;
 486 }
 487 EXPORT_SYMBOL_GPL(nf_ct_delete);
 488
 489 static inline bool
 490 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
 491                 const struct nf_conntrack_tuple *tuple,
 492                 const struct nf_conntrack_zone *zone,
 493                 const struct net *net)
 494 {
 495         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 496
 497         /* A conntrack can be recreated with the equal tuple,
 498          * so we need to check that the conntrack is confirmed
 499          */
 500         return nf_ct_tuple_equal(tuple, &h->tuple) &&
 501                nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
 502                nf_ct_is_confirmed(ct) &&
 503                net_eq(net, nf_ct_net(ct));
 504 }
 505
 506 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
 507 static void nf_ct_gc_expired(struct nf_conn *ct)
 508 {
 509         if (!atomic_inc_not_zero(&ct->ct_general.use))
 510                 return;
 511
 512         if (nf_ct_should_gc(ct))
 513                 nf_ct_kill(ct);
 514
 515         nf_ct_put(ct);
 516 }
 517
 518 /*
 519  * Warning :
 520  * - Caller must take a reference on returned object
 521  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 522  */
 523 static struct nf_conntrack_tuple_hash *
 524 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
 525                       const struct nf_conntrack_tuple *tuple, u32 hash)
 526 {
 527         struct nf_conntrack_tuple_hash *h;
 528         struct hlist_nulls_head *ct_hash;
 529         struct hlist_nulls_node *n;
 530         unsigned int bucket, hsize;
 531
 532 begin:
 533         nf_conntrack_get_ht(&ct_hash, &hsize);
 534         bucket = reciprocal_scale(hash, hsize);
 535
 536         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
 537                 struct nf_conn *ct;
 538
 539                 ct = nf_ct_tuplehash_to_ctrack(h);
 540                 if (nf_ct_is_expired(ct)) {
 541                         nf_ct_gc_expired(ct);
 542                         continue;
 543                 }
 544
 545                 if (nf_ct_is_dying(ct))
 546                         continue;
 547
 548                 if (nf_ct_key_equal(h, tuple, zone, net))
 549                         return h;
 550         }
 551         /*
 552          * if the nulls value we got at the end of this lookup is
 553          * not the expected one, we must restart lookup.
 554          * We probably met an item that was moved to another chain.
 555          */
 556         if (get_nulls_value(n) != bucket) {
 557                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 558                 goto begin;
 559         }
 560
 561         return NULL;
 562 }
 563
 564 /* Find a connection corresponding to a tuple. */
 565 static struct nf_conntrack_tuple_hash *
 566 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 567                         const struct nf_conntrack_tuple *tuple, u32 hash)
 568 {
 569         struct nf_conntrack_tuple_hash *h;
 570         struct nf_conn *ct;
 571
 572         rcu_read_lock();
 573 begin:
 574         h = ____nf_conntrack_find(net, zone, tuple, hash);
 575         if (h) {
 576                 ct = nf_ct_tuplehash_to_ctrack(h);
 577                 if (unlikely(nf_ct_is_dying(ct) ||
 578                              !atomic_inc_not_zero(&ct->ct_general.use)))
 579                         h = NULL;
 580                 else {
 581                         if (unlikely(!nf_ct_key_equal(h, tuple, zone, net))) {
 582                                 nf_ct_put(ct);
 583                                 goto begin;
 584                         }
 585                 }
 586         }
 587         rcu_read_unlock();
 588
 589         return h;
 590 }
 591
 592 struct nf_conntrack_tuple_hash *
 593 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 594                       const struct nf_conntrack_tuple *tuple)
 595 {
 596         return __nf_conntrack_find_get(net, zone, tuple,
 597                                        hash_conntrack_raw(tuple, net));
 598 }
 599 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 600
 601 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 602                                        unsigned int hash,
 603                                        unsigned int reply_hash)
 604 {
 605         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 606                            &nf_conntrack_hash[hash]);
 607         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 608                            &nf_conntrack_hash[reply_hash]);
 609 }
 610
 611 int
 612 nf_conntrack_hash_check_insert(struct nf_conn *ct)
 613 {
 614         const struct nf_conntrack_zone *zone;
 615         struct net *net = nf_ct_net(ct);
 616         unsigned int hash, reply_hash;
 617         struct nf_conntrack_tuple_hash *h;
 618         struct hlist_nulls_node *n;
 619         unsigned int sequence;
 620
 621         zone = nf_ct_zone(ct);
 622
 623         local_bh_disable();
 624         do {
 625                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 626                 hash = hash_conntrack(net,
 627                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 628                 reply_hash = hash_conntrack(net,
 629                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 630         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 631
 632         /* See if there's one in the list already, including reverse */
 633         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 634                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 635                                     zone, net))
 636                         goto out;
 637
 638         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 639                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 640                                     zone, net))
 641                         goto out;
 642
 643         smp_wmb();
 644         /* The caller holds a reference to this object */
 645         atomic_set(&ct->ct_general.use, 2);
 646         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 647         nf_conntrack_double_unlock(hash, reply_hash);
 648         NF_CT_STAT_INC(net, insert);
 649         local_bh_enable();
 650         return 0;
 651
 652 out:
 653         nf_conntrack_double_unlock(hash, reply_hash);
 654         NF_CT_STAT_INC(net, insert_failed);
 655         local_bh_enable();
 656         return -EEXIST;
 657 }
 658 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 659
 660 static inline void nf_ct_acct_update(struct nf_conn *ct,
 661                                      enum ip_conntrack_info ctinfo,
 662                                      unsigned int len)
 663 {
 664         struct nf_conn_acct *acct;
 665
 666         acct = nf_conn_acct_find(ct);
 667         if (acct) {
 668                 struct nf_conn_counter *counter = acct->counter;
 669
 670                 atomic64_inc(&counter[CTINFO2DIR(ctinfo)].packets);
 671                 atomic64_add(len, &counter[CTINFO2DIR(ctinfo)].bytes);
 672         }
 673 }
 674
 675 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
 676                              const struct nf_conn *loser_ct)
 677 {
 678         struct nf_conn_acct *acct;
 679
 680         acct = nf_conn_acct_find(loser_ct);
 681         if (acct) {
 682                 struct nf_conn_counter *counter = acct->counter;
 683                 unsigned int bytes;
 684
 685                 /* u32 should be fine since we must have seen one packet. */
 686                 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
 687                 nf_ct_acct_update(ct, ctinfo, bytes);
 688         }
 689 }
 690
 691 /* Resolve race on insertion if this protocol allows this. */
 692 static int nf_ct_resolve_clash(struct net *net, struct sk_buff *skb,
 693                                enum ip_conntrack_info ctinfo,
 694                                struct nf_conntrack_tuple_hash *h)
 695 {
 696         /* This is the conntrack entry already in hashes that won race. */
 697         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 698         struct nf_conntrack_l4proto *l4proto;
 699
 700         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 701         if (l4proto->allow_clash &&
 702             ((ct->status & IPS_NAT_DONE_MASK) == 0) &&
 703             !nf_ct_is_dying(ct) &&
 704             atomic_inc_not_zero(&ct->ct_general.use)) {
 705                 enum ip_conntrack_info oldinfo;
 706                 struct nf_conn *loser_ct = nf_ct_get(skb, &oldinfo);
 707
 708                 nf_ct_acct_merge(ct, ctinfo, loser_ct);
 709                 nf_conntrack_put(&loser_ct->ct_general);
 710                 nf_ct_set(skb, ct, oldinfo);
 711                 return NF_ACCEPT;
 712         }
 713         NF_CT_STAT_INC(net, drop);
 714         return NF_DROP;
 715 }
 716
 717 /* Confirm a connection given skb; places it in hash table */
 718 int
 719 __nf_conntrack_confirm(struct sk_buff *skb)
 720 {
 721         const struct nf_conntrack_zone *zone;
 722         unsigned int hash, reply_hash;
 723         struct nf_conntrack_tuple_hash *h;
 724         struct nf_conn *ct;
 725         struct nf_conn_help *help;
 726         struct nf_conn_tstamp *tstamp;
 727         struct hlist_nulls_node *n;
 728         enum ip_conntrack_info ctinfo;
 729         struct net *net;
 730         unsigned int sequence;
 731         int ret = NF_DROP;
 732
 733         ct = nf_ct_get(skb, &ctinfo);
 734         net = nf_ct_net(ct);
 735
 736         /* ipt_REJECT uses nf_conntrack_attach to attach related
 737            ICMP/TCP RST packets in other direction.  Actual packet
 738            which created connection will be IP_CT_NEW or for an
 739            expected connection, IP_CT_RELATED. */
 740         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 741                 return NF_ACCEPT;
 742
 743         zone = nf_ct_zone(ct);
 744         local_bh_disable();
 745
 746         do {
 747                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 748                 /* reuse the hash saved before */
 749                 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
 750                 hash = scale_hash(hash);
 751                 reply_hash = hash_conntrack(net,
 752                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 753
 754         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 755
 756         /* We're not in hash table, and we refuse to set up related
 757          * connections for unconfirmed conns.  But packet copies and
 758          * REJECT will give spurious warnings here.
 759          */
 760         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
 761
 762         /* No external references means no one else could have
 763          * confirmed us.
 764          */
 765         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
 766         pr_debug("Confirming conntrack %p\n", ct);
 767         /* We have to check the DYING flag after unlink to prevent
 768          * a race against nf_ct_get_next_corpse() possibly called from
 769          * user context, else we insert an already 'dead' hash, blocking
 770          * further use of that particular connection -JM.
 771          */
 772         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 773
 774         if (unlikely(nf_ct_is_dying(ct))) {
 775                 nf_ct_add_to_dying_list(ct);
 776                 goto dying;
 777         }
 778
 779         /* See if there's one in the list already, including reverse:
 780            NAT could have grabbed it without realizing, since we're
 781            not in the hash.  If there is, we lost race. */
 782         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 783                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 784                                     zone, net))
 785                         goto out;
 786
 787         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 788                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 789                                     zone, net))
 790                         goto out;
 791
 792         /* Timer relative to confirmation time, not original
 793            setting time, otherwise we'd get timer wrap in
 794            weird delay cases. */
 795         ct->timeout += nfct_time_stamp;
 796         atomic_inc(&ct->ct_general.use);
 797         ct->status |= IPS_CONFIRMED;
 798
 799         /* set conntrack timestamp, if enabled. */
 800         tstamp = nf_conn_tstamp_find(ct);
 801         if (tstamp) {
 802                 if (skb->tstamp == 0)
 803                         __net_timestamp(skb);
 804
 805                 tstamp->start = ktime_to_ns(skb->tstamp);
 806         }
 807         /* Since the lookup is lockless, hash insertion must be done after
 808          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 809          * guarantee that no other CPU can find the conntrack before the above
 810          * stores are visible.
 811          */
 812         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 813         nf_conntrack_double_unlock(hash, reply_hash);
 814         local_bh_enable();
 815
 816         help = nfct_help(ct);
 817         if (help && help->helper)
 818                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 819
 820         nf_conntrack_event_cache(master_ct(ct) ?
 821                                  IPCT_RELATED : IPCT_NEW, ct);
 822         return NF_ACCEPT;
 823
 824 out:
 825         nf_ct_add_to_dying_list(ct);
 826         ret = nf_ct_resolve_clash(net, skb, ctinfo, h);
 827 dying:
 828         nf_conntrack_double_unlock(hash, reply_hash);
 829         NF_CT_STAT_INC(net, insert_failed);
 830         local_bh_enable();
 831         return ret;
 832 }
 833 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 834
 835 /* Returns true if a connection correspondings to the tuple (required
 836    for NAT). */
 837 int
 838 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 839                          const struct nf_conn *ignored_conntrack)
 840 {
 841         struct net *net = nf_ct_net(ignored_conntrack);
 842         const struct nf_conntrack_zone *zone;
 843         struct nf_conntrack_tuple_hash *h;
 844         struct hlist_nulls_head *ct_hash;
 845         unsigned int hash, hsize;
 846         struct hlist_nulls_node *n;
 847         struct nf_conn *ct;
 848
 849         zone = nf_ct_zone(ignored_conntrack);
 850
 851         rcu_read_lock();
 852  begin:
 853         nf_conntrack_get_ht(&ct_hash, &hsize);
 854         hash = __hash_conntrack(net, tuple, hsize);
 855
 856         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
 857                 ct = nf_ct_tuplehash_to_ctrack(h);
 858
 859                 if (ct == ignored_conntrack)
 860                         continue;
 861
 862                 if (nf_ct_is_expired(ct)) {
 863                         nf_ct_gc_expired(ct);
 864                         continue;
 865                 }
 866
 867                 if (nf_ct_key_equal(h, tuple, zone, net)) {
 868                         NF_CT_STAT_INC_ATOMIC(net, found);
 869                         rcu_read_unlock();
 870                         return 1;
 871                 }
 872         }
 873
 874         if (get_nulls_value(n) != hash) {
 875                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 876                 goto begin;
 877         }
 878
 879         rcu_read_unlock();
 880
 881         return 0;
 882 }
 883 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 884
 885 #define NF_CT_EVICTION_RANGE    8
 886
 887 /* There's a small race here where we may free a just-assured
 888    connection.  Too bad: we're in trouble anyway. */
 889 static unsigned int early_drop_list(struct net *net,
 890                                     struct hlist_nulls_head *head)
 891 {
 892         struct nf_conntrack_tuple_hash *h;
 893         struct hlist_nulls_node *n;
 894         unsigned int drops = 0;
 895         struct nf_conn *tmp;
 896
 897         hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
 898                 tmp = nf_ct_tuplehash_to_ctrack(h);
 899
 900                 if (nf_ct_is_expired(tmp)) {
 901                         nf_ct_gc_expired(tmp);
 902                         continue;
 903                 }
 904
 905                 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
 906                     !net_eq(nf_ct_net(tmp), net) ||
 907                     nf_ct_is_dying(tmp))
 908                         continue;
 909
 910                 if (!atomic_inc_not_zero(&tmp->ct_general.use))
 911                         continue;
 912
 913                 /* kill only if still in same netns -- might have moved due to
 914                  * SLAB_TYPESAFE_BY_RCU rules.
 915                  *
 916                  * We steal the timer reference.  If that fails timer has
 917                  * already fired or someone else deleted it. Just drop ref
 918                  * and move to next entry.
 919                  */
 920                 if (net_eq(nf_ct_net(tmp), net) &&
 921                     nf_ct_is_confirmed(tmp) &&
 922                     nf_ct_delete(tmp, 0, 0))
 923                         drops++;
 924
 925                 nf_ct_put(tmp);
 926         }
 927
 928         return drops;
 929 }
 930
 931 static noinline int early_drop(struct net *net, unsigned int _hash)
 932 {
 933         unsigned int i;
 934
 935         for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
 936                 struct hlist_nulls_head *ct_hash;
 937                 unsigned int hash, hsize, drops;
 938
 939                 rcu_read_lock();
 940                 nf_conntrack_get_ht(&ct_hash, &hsize);
 941                 hash = reciprocal_scale(_hash++, hsize);
 942
 943                 drops = early_drop_list(net, &ct_hash[hash]);
 944                 rcu_read_unlock();
 945
 946                 if (drops) {
 947                         NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
 948                         return true;
 949                 }
 950         }
 951
 952         return false;
 953 }
 954
 955 static bool gc_worker_skip_ct(const struct nf_conn *ct)
 956 {
 957         return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
 958 }
 959
 960 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
 961 {
 962         const struct nf_conntrack_l4proto *l4proto;
 963
 964         if (!test_bit(IPS_ASSURED_BIT, &ct->status))
 965                 return true;
 966
 967         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 968         if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
 969                 return true;
 970
 971         return false;
 972 }
 973
 974 static void gc_worker(struct work_struct *work)
 975 {
 976         unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
 977         unsigned int i, goal, buckets = 0, expired_count = 0;
 978         unsigned int nf_conntrack_max95 = 0;
 979         struct conntrack_gc_work *gc_work;
 980         unsigned int ratio, scanned = 0;
 981         unsigned long next_run;
 982
 983         gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
 984
 985         goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
 986         i = gc_work->last_bucket;
 987         if (gc_work->early_drop)
 988                 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
 989
 990         do {
 991                 struct nf_conntrack_tuple_hash *h;
 992                 struct hlist_nulls_head *ct_hash;
 993                 struct hlist_nulls_node *n;
 994                 unsigned int hashsz;
 995                 struct nf_conn *tmp;
 996
 997                 i++;
 998                 rcu_read_lock();
 999
1000                 nf_conntrack_get_ht(&ct_hash, &hashsz);
1001                 if (i >= hashsz)
1002                         i = 0;
1003
1004                 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1005                         struct net *net;
1006
1007                         tmp = nf_ct_tuplehash_to_ctrack(h);
1008
1009                         scanned++;
1010                         if (nf_ct_is_expired(tmp)) {
1011                                 nf_ct_gc_expired(tmp);
1012                                 expired_count++;
1013                                 continue;
1014                         }
1015
1016                         if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1017                                 continue;
1018
1019                         net = nf_ct_net(tmp);
1020                         if (atomic_read(&net->ct.count) < nf_conntrack_max95)
1021                                 continue;
1022
1023                         /* need to take reference to avoid possible races */
1024                         if (!atomic_inc_not_zero(&tmp->ct_general.use))
1025                                 continue;
1026
1027                         if (gc_worker_skip_ct(tmp)) {
1028                                 nf_ct_put(tmp);
1029                                 continue;
1030                         }
1031
1032                         if (gc_worker_can_early_drop(tmp))
1033                                 nf_ct_kill(tmp);
1034
1035                         nf_ct_put(tmp);
1036                 }
1037
1038                 /* could check get_nulls_value() here and restart if ct
1039                  * was moved to another chain.  But given gc is best-effort
1040                  * we will just continue with next hash slot.
1041                  */
1042                 rcu_read_unlock();
1043                 cond_resched_rcu_qs();
1044         } while (++buckets < goal);
1045
1046         if (gc_work->exiting)
1047                 return;
1048
1049         /*
1050          * Eviction will normally happen from the packet path, and not
1051          * from this gc worker.
1052          *
1053          * This worker is only here to reap expired entries when system went
1054          * idle after a busy period.
1055          *
1056          * The heuristics below are supposed to balance conflicting goals:
1057          *
1058          * 1. Minimize time until we notice a stale entry
1059          * 2. Maximize scan intervals to not waste cycles
1060          *
1061          * Normally, expire ratio will be close to 0.
1062          *
1063          * As soon as a sizeable fraction of the entries have expired
1064          * increase scan frequency.
1065          */
1066         ratio = scanned ? expired_count * 100 / scanned : 0;
1067         if (ratio > GC_EVICT_RATIO) {
1068                 gc_work->next_gc_run = min_interval;
1069         } else {
1070                 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1071
1072                 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1073
1074                 gc_work->next_gc_run += min_interval;
1075                 if (gc_work->next_gc_run > max)
1076                         gc_work->next_gc_run = max;
1077         }
1078
1079         next_run = gc_work->next_gc_run;
1080         gc_work->last_bucket = i;
1081         gc_work->early_drop = false;
1082         queue_delayed_work(system_long_wq, &gc_work->dwork, next_run);
1083 }
1084
1085 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1086 {
1087         INIT_DELAYED_WORK(&gc_work->dwork, gc_worker);
1088         gc_work->next_gc_run = HZ;
1089         gc_work->exiting = false;
1090 }
1091
1092 static struct nf_conn *
1093 __nf_conntrack_alloc(struct net *net,
1094                      const struct nf_conntrack_zone *zone,
1095                      const struct nf_conntrack_tuple *orig,
1096                      const struct nf_conntrack_tuple *repl,
1097                      gfp_t gfp, u32 hash)
1098 {
1099         struct nf_conn *ct;
1100
1101         /* We don't want any race condition at early drop stage */
1102         atomic_inc(&net->ct.count);
1103
1104         if (nf_conntrack_max &&
1105             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
1106                 if (!early_drop(net, hash)) {
1107                         if (!conntrack_gc_work.early_drop)
1108                                 conntrack_gc_work.early_drop = true;
1109                         atomic_dec(&net->ct.count);
1110                         net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1111                         return ERR_PTR(-ENOMEM);
1112                 }
1113         }
1114
1115         /*
1116          * Do not use kmem_cache_zalloc(), as this cache uses
1117          * SLAB_TYPESAFE_BY_RCU.
1118          */
1119         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1120         if (ct == NULL)
1121                 goto out;
1122
1123         spin_lock_init(&ct->lock);
1124         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1125         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1126         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1127         /* save hash for reusing when confirming */
1128         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1129         ct->status = 0;
1130         write_pnet(&ct->ct_net, net);
1131         memset(&ct->__nfct_init_offset[0], 0,
1132                offsetof(struct nf_conn, proto) -
1133                offsetof(struct nf_conn, __nfct_init_offset[0]));
1134
1135         nf_ct_zone_add(ct, zone);
1136
1137         /* Because we use RCU lookups, we set ct_general.use to zero before
1138          * this is inserted in any list.
1139          */
1140         atomic_set(&ct->ct_general.use, 0);
1141         return ct;
1142 out:
1143         atomic_dec(&net->ct.count);
1144         return ERR_PTR(-ENOMEM);
1145 }
1146
1147 struct nf_conn *nf_conntrack_alloc(struct net *net,
1148                                    const struct nf_conntrack_zone *zone,
1149                                    const struct nf_conntrack_tuple *orig,
1150                                    const struct nf_conntrack_tuple *repl,
1151                                    gfp_t gfp)
1152 {
1153         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1154 }
1155 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1156
1157 void nf_conntrack_free(struct nf_conn *ct)
1158 {
1159         struct net *net = nf_ct_net(ct);
1160
1161         /* A freed object has refcnt == 0, that's
1162          * the golden rule for SLAB_TYPESAFE_BY_RCU
1163          */
1164         NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 0);
1165
1166         nf_ct_ext_destroy(ct);
1167         nf_ct_ext_free(ct);
1168         kmem_cache_free(nf_conntrack_cachep, ct);
1169         smp_mb__before_atomic();
1170         atomic_dec(&net->ct.count);
1171 }
1172 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1173
1174
1175 /* Allocate a new conntrack: we return -ENOMEM if classification
1176    failed due to stress.  Otherwise it really is unclassifiable. */
1177 static noinline struct nf_conntrack_tuple_hash *
1178 init_conntrack(struct net *net, struct nf_conn *tmpl,
1179                const struct nf_conntrack_tuple *tuple,
1180                struct nf_conntrack_l3proto *l3proto,
1181                struct nf_conntrack_l4proto *l4proto,
1182                struct sk_buff *skb,
1183                unsigned int dataoff, u32 hash)
1184 {
1185         struct nf_conn *ct;
1186         struct nf_conn_help *help;
1187         struct nf_conntrack_tuple repl_tuple;
1188         struct nf_conntrack_ecache *ecache;
1189         struct nf_conntrack_expect *exp = NULL;
1190         const struct nf_conntrack_zone *zone;
1191         struct nf_conn_timeout *timeout_ext;
1192         struct nf_conntrack_zone tmp;
1193         unsigned int *timeouts;
1194
1195         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
1196                 pr_debug("Can't invert tuple.\n");
1197                 return NULL;
1198         }
1199
1200         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1201         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1202                                   hash);
1203         if (IS_ERR(ct))
1204                 return (struct nf_conntrack_tuple_hash *)ct;
1205
1206         if (!nf_ct_add_synproxy(ct, tmpl)) {
1207                 nf_conntrack_free(ct);
1208                 return ERR_PTR(-ENOMEM);
1209         }
1210
1211         timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1212         if (timeout_ext) {
1213                 timeouts = nf_ct_timeout_data(timeout_ext);
1214                 if (unlikely(!timeouts))
1215                         timeouts = l4proto->get_timeouts(net);
1216         } else {
1217                 timeouts = l4proto->get_timeouts(net);
1218         }
1219
1220         if (!l4proto->new(ct, skb, dataoff, timeouts)) {
1221                 nf_conntrack_free(ct);
1222                 pr_debug("can't track with proto module\n");
1223                 return NULL;
1224         }
1225
1226         if (timeout_ext)
1227                 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1228                                       GFP_ATOMIC);
1229
1230         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1231         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1232         nf_ct_labels_ext_add(ct);
1233
1234         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1235         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1236                                  ecache ? ecache->expmask : 0,
1237                              GFP_ATOMIC);
1238
1239         local_bh_disable();
1240         if (net->ct.expect_count) {
1241                 spin_lock(&nf_conntrack_expect_lock);
1242                 exp = nf_ct_find_expectation(net, zone, tuple);
1243                 if (exp) {
1244                         pr_debug("expectation arrives ct=%p exp=%p\n",
1245                                  ct, exp);
1246                         /* Welcome, Mr. Bond.  We've been expecting you... */
1247                         __set_bit(IPS_EXPECTED_BIT, &ct->status);
1248                         /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1249                         ct->master = exp->master;
1250                         if (exp->helper) {
1251                                 help = nf_ct_helper_ext_add(ct, exp->helper,
1252                                                             GFP_ATOMIC);
1253                                 if (help)
1254                                         rcu_assign_pointer(help->helper, exp->helper);
1255                         }
1256
1257 #ifdef CONFIG_NF_CONNTRACK_MARK
1258                         ct->mark = exp->master->mark;
1259 #endif
1260 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1261                         ct->secmark = exp->master->secmark;
1262 #endif
1263                         NF_CT_STAT_INC(net, expect_new);
1264                 }
1265                 spin_unlock(&nf_conntrack_expect_lock);
1266         }
1267         if (!exp)
1268                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1269
1270         /* Now it is inserted into the unconfirmed list, bump refcount */
1271         nf_conntrack_get(&ct->ct_general);
1272         nf_ct_add_to_unconfirmed_list(ct);
1273
1274         local_bh_enable();
1275
1276         if (exp) {
1277                 if (exp->expectfn)
1278                         exp->expectfn(ct, exp);
1279                 nf_ct_expect_put(exp);
1280         }
1281
1282         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1283 }
1284
1285 /* On success, returns 0, sets skb->_nfct | ctinfo */
1286 static int
1287 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
1288                   struct sk_buff *skb,
1289                   unsigned int dataoff,
1290                   u_int16_t l3num,
1291                   u_int8_t protonum,
1292                   struct nf_conntrack_l3proto *l3proto,
1293                   struct nf_conntrack_l4proto *l4proto)
1294 {
1295         const struct nf_conntrack_zone *zone;
1296         struct nf_conntrack_tuple tuple;
1297         struct nf_conntrack_tuple_hash *h;
1298         enum ip_conntrack_info ctinfo;
1299         struct nf_conntrack_zone tmp;
1300         struct nf_conn *ct;
1301         u32 hash;
1302
1303         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1304                              dataoff, l3num, protonum, net, &tuple, l3proto,
1305                              l4proto)) {
1306                 pr_debug("Can't get tuple\n");
1307                 return 0;
1308         }
1309
1310         /* look for tuple match */
1311         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1312         hash = hash_conntrack_raw(&tuple, net);
1313         h = __nf_conntrack_find_get(net, zone, &tuple, hash);
1314         if (!h) {
1315                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
1316                                    skb, dataoff, hash);
1317                 if (!h)
1318                         return 0;
1319                 if (IS_ERR(h))
1320                         return PTR_ERR(h);
1321         }
1322         ct = nf_ct_tuplehash_to_ctrack(h);
1323
1324         /* It exists; we have (non-exclusive) reference. */
1325         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1326                 ctinfo = IP_CT_ESTABLISHED_REPLY;
1327         } else {
1328                 /* Once we've had two way comms, always ESTABLISHED. */
1329                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1330                         pr_debug("normal packet for %p\n", ct);
1331                         ctinfo = IP_CT_ESTABLISHED;
1332                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1333                         pr_debug("related packet for %p\n", ct);
1334                         ctinfo = IP_CT_RELATED;
1335                 } else {
1336                         pr_debug("new packet for %p\n", ct);
1337                         ctinfo = IP_CT_NEW;
1338                 }
1339         }
1340         nf_ct_set(skb, ct, ctinfo);
1341         return 0;
1342 }
1343
1344 unsigned int
1345 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
1346                 struct sk_buff *skb)
1347 {
1348         struct nf_conn *ct, *tmpl;
1349         enum ip_conntrack_info ctinfo;
1350         struct nf_conntrack_l3proto *l3proto;
1351         struct nf_conntrack_l4proto *l4proto;
1352         unsigned int *timeouts;
1353         unsigned int dataoff;
1354         u_int8_t protonum;
1355         int ret;
1356
1357         tmpl = nf_ct_get(skb, &ctinfo);
1358         if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1359                 /* Previously seen (loopback or untracked)?  Ignore. */
1360                 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1361                      ctinfo == IP_CT_UNTRACKED) {
1362                         NF_CT_STAT_INC_ATOMIC(net, ignore);
1363                         return NF_ACCEPT;
1364                 }
1365                 skb->_nfct = 0;
1366         }
1367
1368         /* rcu_read_lock()ed by nf_hook_thresh */
1369         l3proto = __nf_ct_l3proto_find(pf);
1370         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
1371                                    &dataoff, &protonum);
1372         if (ret <= 0) {
1373                 pr_debug("not prepared to track yet or error occurred\n");
1374                 NF_CT_STAT_INC_ATOMIC(net, error);
1375                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1376                 ret = -ret;
1377                 goto out;
1378         }
1379
1380         l4proto = __nf_ct_l4proto_find(pf, protonum);
1381
1382         /* It may be an special packet, error, unclean...
1383          * inverse of the return code tells to the netfilter
1384          * core what to do with the packet. */
1385         if (l4proto->error != NULL) {
1386                 ret = l4proto->error(net, tmpl, skb, dataoff, pf, hooknum);
1387                 if (ret <= 0) {
1388                         NF_CT_STAT_INC_ATOMIC(net, error);
1389                         NF_CT_STAT_INC_ATOMIC(net, invalid);
1390                         ret = -ret;
1391                         goto out;
1392                 }
1393                 /* ICMP[v6] protocol trackers may assign one conntrack. */
1394                 if (skb->_nfct)
1395                         goto out;
1396         }
1397 repeat:
1398         ret = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
1399                                 l3proto, l4proto);
1400         if (ret < 0) {
1401                 /* Too stressed to deal. */
1402                 NF_CT_STAT_INC_ATOMIC(net, drop);
1403                 ret = NF_DROP;
1404                 goto out;
1405         }
1406
1407         ct = nf_ct_get(skb, &ctinfo);
1408         if (!ct) {
1409                 /* Not valid part of a connection */
1410                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1411                 ret = NF_ACCEPT;
1412                 goto out;
1413         }
1414
1415         /* Decide what timeout policy we want to apply to this flow. */
1416         timeouts = nf_ct_timeout_lookup(net, ct, l4proto);
1417
1418         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum, timeouts);
1419         if (ret <= 0) {
1420                 /* Invalid: inverse of the return code tells
1421                  * the netfilter core what to do */
1422                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1423                 nf_conntrack_put(&ct->ct_general);
1424                 skb->_nfct = 0;
1425                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1426                 if (ret == -NF_DROP)
1427                         NF_CT_STAT_INC_ATOMIC(net, drop);
1428                 /* Special case: TCP tracker reports an attempt to reopen a
1429                  * closed/aborted connection. We have to go back and create a
1430                  * fresh conntrack.
1431                  */
1432                 if (ret == -NF_REPEAT)
1433                         goto repeat;
1434                 ret = -ret;
1435                 goto out;
1436         }
1437
1438         if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
1439             !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1440                 nf_conntrack_event_cache(IPCT_REPLY, ct);
1441 out:
1442         if (tmpl)
1443                 nf_ct_put(tmpl);
1444
1445         return ret;
1446 }
1447 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1448
1449 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
1450                           const struct nf_conntrack_tuple *orig)
1451 {
1452         bool ret;
1453
1454         rcu_read_lock();
1455         ret = nf_ct_invert_tuple(inverse, orig,
1456                                  __nf_ct_l3proto_find(orig->src.l3num),
1457                                  __nf_ct_l4proto_find(orig->src.l3num,
1458                                                       orig->dst.protonum));
1459         rcu_read_unlock();
1460         return ret;
1461 }
1462 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
1463
1464 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
1465    implicitly racy: see __nf_conntrack_confirm */
1466 void nf_conntrack_alter_reply(struct nf_conn *ct,
1467                               const struct nf_conntrack_tuple *newreply)
1468 {
1469         struct nf_conn_help *help = nfct_help(ct);
1470
1471         /* Should be unconfirmed, so not in hash table yet */
1472         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
1473
1474         pr_debug("Altering reply tuple of %p to ", ct);
1475         nf_ct_dump_tuple(newreply);
1476
1477         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1478         if (ct->master || (help && !hlist_empty(&help->expectations)))
1479                 return;
1480
1481         rcu_read_lock();
1482         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1483         rcu_read_unlock();
1484 }
1485 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1486
1487 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1488 void __nf_ct_refresh_acct(struct nf_conn *ct,
1489                           enum ip_conntrack_info ctinfo,
1490                           const struct sk_buff *skb,
1491                           unsigned long extra_jiffies,
1492                           int do_acct)
1493 {
1494         NF_CT_ASSERT(skb);
1495
1496         /* Only update if this is not a fixed timeout */
1497         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1498                 goto acct;
1499
1500         /* If not in hash table, timer will not be active yet */
1501         if (nf_ct_is_confirmed(ct))
1502                 extra_jiffies += nfct_time_stamp;
1503
1504         ct->timeout = extra_jiffies;
1505 acct:
1506         if (do_acct)
1507                 nf_ct_acct_update(ct, ctinfo, skb->len);
1508 }
1509 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1510
1511 bool nf_ct_kill_acct(struct nf_conn *ct,
1512                      enum ip_conntrack_info ctinfo,
1513                      const struct sk_buff *skb)
1514 {
1515         nf_ct_acct_update(ct, ctinfo, skb->len);
1516
1517         return nf_ct_delete(ct, 0, 0);
1518 }
1519 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1520
1521 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1522
1523 #include <linux/netfilter/nfnetlink.h>
1524 #include <linux/netfilter/nfnetlink_conntrack.h>
1525 #include <linux/mutex.h>
1526
1527 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
1528  * in ip_conntrack_core, since we don't want the protocols to autoload
1529  * or depend on ctnetlink */
1530 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1531                                const struct nf_conntrack_tuple *tuple)
1532 {
1533         if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1534             nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1535                 goto nla_put_failure;
1536         return 0;
1537
1538 nla_put_failure:
1539         return -1;
1540 }
1541 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1542
1543 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1544         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
1545         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
1546 };
1547 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1548
1549 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1550                                struct nf_conntrack_tuple *t)
1551 {
1552         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
1553                 return -EINVAL;
1554
1555         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1556         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1557
1558         return 0;
1559 }
1560 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1561
1562 int nf_ct_port_nlattr_tuple_size(void)
1563 {
1564         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1565 }
1566 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1567 #endif
1568
1569 /* Used by ipt_REJECT and ip6t_REJECT. */
1570 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
1571 {
1572         struct nf_conn *ct;
1573         enum ip_conntrack_info ctinfo;
1574
1575         /* This ICMP is in reverse direction to the packet which caused it */
1576         ct = nf_ct_get(skb, &ctinfo);
1577         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1578                 ctinfo = IP_CT_RELATED_REPLY;
1579         else
1580                 ctinfo = IP_CT_RELATED;
1581
1582         /* Attach to new skbuff, and increment count */
1583         nf_ct_set(nskb, ct, ctinfo);
1584         nf_conntrack_get(skb_nfct(nskb));
1585 }
1586
1587 /* Bring out ya dead! */
1588 static struct nf_conn *
1589 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
1590                 void *data, unsigned int *bucket)
1591 {
1592         struct nf_conntrack_tuple_hash *h;
1593         struct nf_conn *ct;
1594         struct hlist_nulls_node *n;
1595         spinlock_t *lockp;
1596
1597         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1598                 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
1599                 local_bh_disable();
1600                 nf_conntrack_lock(lockp);
1601                 if (*bucket < nf_conntrack_htable_size) {
1602                         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
1603                                 if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1604                                         continue;
1605                                 ct = nf_ct_tuplehash_to_ctrack(h);
1606                                 if (iter(ct, data))
1607                                         goto found;
1608                         }
1609                 }
1610                 spin_unlock(lockp);
1611                 local_bh_enable();
1612                 cond_resched();
1613         }
1614
1615         return NULL;
1616 found:
1617         atomic_inc(&ct->ct_general.use);
1618         spin_unlock(lockp);
1619         local_bh_enable();
1620         return ct;
1621 }
1622
1623 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
1624                                   void *data, u32 portid, int report)
1625 {
1626         unsigned int bucket = 0, sequence;
1627         struct nf_conn *ct;
1628
1629         might_sleep();
1630
1631         for (;;) {
1632                 sequence = read_seqcount_begin(&nf_conntrack_generation);
1633
1634                 while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
1635                         /* Time to push up daises... */
1636
1637                         nf_ct_delete(ct, portid, report);
1638                         nf_ct_put(ct);
1639                         cond_resched();
1640                 }
1641
1642                 if (!read_seqcount_retry(&nf_conntrack_generation, sequence))
1643                         break;
1644                 bucket = 0;
1645         }
1646 }
1647
1648 struct iter_data {
1649         int (*iter)(struct nf_conn *i, void *data);
1650         void *data;
1651         struct net *net;
1652 };
1653
1654 static int iter_net_only(struct nf_conn *i, void *data)
1655 {
1656         struct iter_data *d = data;
1657
1658         if (!net_eq(d->net, nf_ct_net(i)))
1659                 return 0;
1660
1661         return d->iter(i, d->data);
1662 }
1663
1664 static void
1665 __nf_ct_unconfirmed_destroy(struct net *net)
1666 {
1667         int cpu;
1668
1669         for_each_possible_cpu(cpu) {
1670                 struct nf_conntrack_tuple_hash *h;
1671                 struct hlist_nulls_node *n;
1672                 struct ct_pcpu *pcpu;
1673
1674                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1675
1676                 spin_lock_bh(&pcpu->lock);
1677                 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
1678                         struct nf_conn *ct;
1679
1680                         ct = nf_ct_tuplehash_to_ctrack(h);
1681
1682                         /* we cannot call iter() on unconfirmed list, the
1683                          * owning cpu can reallocate ct->ext at any time.
1684                          */
1685                         set_bit(IPS_DYING_BIT, &ct->status);
1686                 }
1687                 spin_unlock_bh(&pcpu->lock);
1688                 cond_resched();
1689         }
1690 }
1691
1692 void nf_ct_iterate_cleanup_net(struct net *net,
1693                                int (*iter)(struct nf_conn *i, void *data),
1694                                void *data, u32 portid, int report)
1695 {
1696         struct iter_data d;
1697
1698         might_sleep();
1699
1700         if (atomic_read(&net->ct.count) == 0)
1701                 return;
1702
1703         __nf_ct_unconfirmed_destroy(net);
1704
1705         d.iter = iter;
1706         d.data = data;
1707         d.net = net;
1708
1709         synchronize_net();
1710
1711         nf_ct_iterate_cleanup(iter_net_only, &d, portid, report);
1712 }
1713 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
1714
1715 /**
1716  * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
1717  * @iter: callback to invoke for each conntrack
1718  * @data: data to pass to @iter
1719  *
1720  * Like nf_ct_iterate_cleanup, but first marks conntracks on the
1721  * unconfirmed list as dying (so they will not be inserted into
1722  * main table).
1723  *
1724  * Can only be called in module exit path.
1725  */
1726 void
1727 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
1728 {
1729         struct net *net;
1730
1731         rtnl_lock();
1732         for_each_net(net) {
1733                 if (atomic_read(&net->ct.count) == 0)
1734                         continue;
1735                 __nf_ct_unconfirmed_destroy(net);
1736         }
1737         rtnl_unlock();
1738
1739         /* Need to wait for netns cleanup worker to finish, if its
1740          * running -- it might have deleted a net namespace from
1741          * the global list, so our __nf_ct_unconfirmed_destroy() might
1742          * not have affected all namespaces.
1743          */
1744         net_ns_barrier();
1745
1746         /* a conntrack could have been unlinked from unconfirmed list
1747          * before we grabbed pcpu lock in __nf_ct_unconfirmed_destroy().
1748          * This makes sure its inserted into conntrack table.
1749          */
1750         synchronize_net();
1751
1752         nf_ct_iterate_cleanup(iter, data, 0, 0);
1753 }
1754 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
1755
1756 static int kill_all(struct nf_conn *i, void *data)
1757 {
1758         return net_eq(nf_ct_net(i), data);
1759 }
1760
1761 void nf_ct_free_hashtable(void *hash, unsigned int size)
1762 {
1763         if (is_vmalloc_addr(hash))
1764                 vfree(hash);
1765         else
1766                 free_pages((unsigned long)hash,
1767                            get_order(sizeof(struct hlist_head) * size));
1768 }
1769 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1770
1771 void nf_conntrack_cleanup_start(void)
1772 {
1773         conntrack_gc_work.exiting = true;
1774         RCU_INIT_POINTER(ip_ct_attach, NULL);
1775 }
1776
1777 void nf_conntrack_cleanup_end(void)
1778 {
1779         RCU_INIT_POINTER(nf_ct_destroy, NULL);
1780
1781         cancel_delayed_work_sync(&conntrack_gc_work.dwork);
1782         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1783
1784         nf_conntrack_proto_fini();
1785         nf_conntrack_seqadj_fini();
1786         nf_conntrack_labels_fini();
1787         nf_conntrack_helper_fini();
1788         nf_conntrack_timeout_fini();
1789         nf_conntrack_ecache_fini();
1790         nf_conntrack_tstamp_fini();
1791         nf_conntrack_acct_fini();
1792         nf_conntrack_expect_fini();
1793
1794         kmem_cache_destroy(nf_conntrack_cachep);
1795 }
1796
1797 /*
1798  * Mishearing the voices in his head, our hero wonders how he's
1799  * supposed to kill the mall.
1800  */
1801 void nf_conntrack_cleanup_net(struct net *net)
1802 {
1803         LIST_HEAD(single);
1804
1805         list_add(&net->exit_list, &single);
1806         nf_conntrack_cleanup_net_list(&single);
1807 }
1808
1809 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
1810 {
1811         int busy;
1812         struct net *net;
1813
1814         /*
1815          * This makes sure all current packets have passed through
1816          *  netfilter framework.  Roll on, two-stage module
1817          *  delete...
1818          */
1819         synchronize_net();
1820 i_see_dead_people:
1821         busy = 0;
1822         list_for_each_entry(net, net_exit_list, exit_list) {
1823                 nf_ct_iterate_cleanup(kill_all, net, 0, 0);
1824                 if (atomic_read(&net->ct.count) != 0)
1825                         busy = 1;
1826         }
1827         if (busy) {
1828                 schedule();
1829                 goto i_see_dead_people;
1830         }
1831
1832         list_for_each_entry(net, net_exit_list, exit_list) {
1833                 nf_conntrack_proto_pernet_fini(net);
1834                 nf_conntrack_helper_pernet_fini(net);
1835                 nf_conntrack_ecache_pernet_fini(net);
1836                 nf_conntrack_tstamp_pernet_fini(net);
1837                 nf_conntrack_acct_pernet_fini(net);
1838                 nf_conntrack_expect_pernet_fini(net);
1839                 free_percpu(net->ct.stat);
1840                 free_percpu(net->ct.pcpu_lists);
1841         }
1842 }
1843
1844 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
1845 {
1846         struct hlist_nulls_head *hash;
1847         unsigned int nr_slots, i;
1848         size_t sz;
1849
1850         if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1851                 return NULL;
1852
1853         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1854         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1855
1856         if (nr_slots > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1857                 return NULL;
1858
1859         sz = nr_slots * sizeof(struct hlist_nulls_head);
1860         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1861                                         get_order(sz));
1862         if (!hash)
1863                 hash = vzalloc(sz);
1864
1865         if (hash && nulls)
1866                 for (i = 0; i < nr_slots; i++)
1867                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1868
1869         return hash;
1870 }
1871 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1872
1873 int nf_conntrack_hash_resize(unsigned int hashsize)
1874 {
1875         int i, bucket;
1876         unsigned int old_size;
1877         struct hlist_nulls_head *hash, *old_hash;
1878         struct nf_conntrack_tuple_hash *h;
1879         struct nf_conn *ct;
1880
1881         if (!hashsize)
1882                 return -EINVAL;
1883
1884         hash = nf_ct_alloc_hashtable(&hashsize, 1);
1885         if (!hash)
1886                 return -ENOMEM;
1887
1888         old_size = nf_conntrack_htable_size;
1889         if (old_size == hashsize) {
1890                 nf_ct_free_hashtable(hash, hashsize);
1891                 return 0;
1892         }
1893
1894         local_bh_disable();
1895         nf_conntrack_all_lock();
1896         write_seqcount_begin(&nf_conntrack_generation);
1897
1898         /* Lookups in the old hash might happen in parallel, which means we
1899          * might get false negatives during connection lookup. New connections
1900          * created because of a false negative won't make it into the hash
1901          * though since that required taking the locks.
1902          */
1903
1904         for (i = 0; i < nf_conntrack_htable_size; i++) {
1905                 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
1906                         h = hlist_nulls_entry(nf_conntrack_hash[i].first,
1907                                               struct nf_conntrack_tuple_hash, hnnode);
1908                         ct = nf_ct_tuplehash_to_ctrack(h);
1909                         hlist_nulls_del_rcu(&h->hnnode);
1910                         bucket = __hash_conntrack(nf_ct_net(ct),
1911                                                   &h->tuple, hashsize);
1912                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1913                 }
1914         }
1915         old_size = nf_conntrack_htable_size;
1916         old_hash = nf_conntrack_hash;
1917
1918         nf_conntrack_hash = hash;
1919         nf_conntrack_htable_size = hashsize;
1920
1921         write_seqcount_end(&nf_conntrack_generation);
1922         nf_conntrack_all_unlock();
1923         local_bh_enable();
1924
1925         synchronize_net();
1926         nf_ct_free_hashtable(old_hash, old_size);
1927         return 0;
1928 }
1929
1930 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1931 {
1932         unsigned int hashsize;
1933         int rc;
1934
1935         if (current->nsproxy->net_ns != &init_net)
1936                 return -EOPNOTSUPP;
1937
1938         /* On boot, we can set this without any fancy locking. */
1939         if (!nf_conntrack_htable_size)
1940                 return param_set_uint(val, kp);
1941
1942         rc = kstrtouint(val, 0, &hashsize);
1943         if (rc)
1944                 return rc;
1945
1946         return nf_conntrack_hash_resize(hashsize);
1947 }
1948 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1949
1950 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1951                   &nf_conntrack_htable_size, 0600);
1952
1953 static __always_inline unsigned int total_extension_size(void)
1954 {
1955         /* remember to add new extensions below */
1956         BUILD_BUG_ON(NF_CT_EXT_NUM > 9);
1957
1958         return sizeof(struct nf_ct_ext) +
1959                sizeof(struct nf_conn_help)
1960 #if IS_ENABLED(CONFIG_NF_NAT)
1961                 + sizeof(struct nf_conn_nat)
1962 #endif
1963                 + sizeof(struct nf_conn_seqadj)
1964                 + sizeof(struct nf_conn_acct)
1965 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1966                 + sizeof(struct nf_conntrack_ecache)
1967 #endif
1968 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
1969                 + sizeof(struct nf_conn_tstamp)
1970 #endif
1971 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
1972                 + sizeof(struct nf_conn_timeout)
1973 #endif
1974 #ifdef CONFIG_NF_CONNTRACK_LABELS
1975                 + sizeof(struct nf_conn_labels)
1976 #endif
1977 #if IS_ENABLED(CONFIG_NETFILTER_SYNPROXY)
1978                 + sizeof(struct nf_conn_synproxy)
1979 #endif
1980         ;
1981 };
1982
1983 int nf_conntrack_init_start(void)
1984 {
1985         int max_factor = 8;
1986         int ret = -ENOMEM;
1987         int i;
1988
1989         /* struct nf_ct_ext uses u8 to store offsets/size */
1990         BUILD_BUG_ON(total_extension_size() > 255u);
1991
1992         seqcount_init(&nf_conntrack_generation);
1993
1994         for (i = 0; i < CONNTRACK_LOCKS; i++)
1995                 spin_lock_init(&nf_conntrack_locks[i]);
1996
1997         if (!nf_conntrack_htable_size) {
1998                 /* Idea from tcp.c: use 1/16384 of memory.
1999                  * On i386: 32MB machine has 512 buckets.
2000                  * >= 1GB machines have 16384 buckets.
2001                  * >= 4GB machines have 65536 buckets.
2002                  */
2003                 nf_conntrack_htable_size
2004                         = (((totalram_pages << PAGE_SHIFT) / 16384)
2005                            / sizeof(struct hlist_head));
2006                 if (totalram_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2007                         nf_conntrack_htable_size = 65536;
2008                 else if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2009                         nf_conntrack_htable_size = 16384;
2010                 if (nf_conntrack_htable_size < 32)
2011                         nf_conntrack_htable_size = 32;
2012
2013                 /* Use a max. factor of four by default to get the same max as
2014                  * with the old struct list_heads. When a table size is given
2015                  * we use the old value of 8 to avoid reducing the max.
2016                  * entries. */
2017                 max_factor = 4;
2018         }
2019
2020         nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2021         if (!nf_conntrack_hash)
2022                 return -ENOMEM;
2023
2024         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2025
2026         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2027                                                 sizeof(struct nf_conn),
2028                                                 NFCT_INFOMASK + 1,
2029                                                 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2030         if (!nf_conntrack_cachep)
2031                 goto err_cachep;
2032
2033         printk(KERN_INFO "nf_conntrack version %s (%u buckets, %d max)\n",
2034                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
2035                nf_conntrack_max);
2036
2037         ret = nf_conntrack_expect_init();
2038         if (ret < 0)
2039                 goto err_expect;
2040
2041         ret = nf_conntrack_acct_init();
2042         if (ret < 0)
2043                 goto err_acct;
2044
2045         ret = nf_conntrack_tstamp_init();
2046         if (ret < 0)
2047                 goto err_tstamp;
2048
2049         ret = nf_conntrack_ecache_init();
2050         if (ret < 0)
2051                 goto err_ecache;
2052
2053         ret = nf_conntrack_timeout_init();
2054         if (ret < 0)
2055                 goto err_timeout;
2056
2057         ret = nf_conntrack_helper_init();
2058         if (ret < 0)
2059                 goto err_helper;
2060
2061         ret = nf_conntrack_labels_init();
2062         if (ret < 0)
2063                 goto err_labels;
2064
2065         ret = nf_conntrack_seqadj_init();
2066         if (ret < 0)
2067                 goto err_seqadj;
2068
2069         ret = nf_conntrack_proto_init();
2070         if (ret < 0)
2071                 goto err_proto;
2072
2073         conntrack_gc_work_init(&conntrack_gc_work);
2074         queue_delayed_work(system_long_wq, &conntrack_gc_work.dwork, HZ);
2075
2076         return 0;
2077
2078 err_proto:
2079         nf_conntrack_seqadj_fini();
2080 err_seqadj:
2081         nf_conntrack_labels_fini();
2082 err_labels:
2083         nf_conntrack_helper_fini();
2084 err_helper:
2085         nf_conntrack_timeout_fini();
2086 err_timeout:
2087         nf_conntrack_ecache_fini();
2088 err_ecache:
2089         nf_conntrack_tstamp_fini();
2090 err_tstamp:
2091         nf_conntrack_acct_fini();
2092 err_acct:
2093         nf_conntrack_expect_fini();
2094 err_expect:
2095         kmem_cache_destroy(nf_conntrack_cachep);
2096 err_cachep:
2097         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
2098         return ret;
2099 }
2100
2101 void nf_conntrack_init_end(void)
2102 {
2103         /* For use by REJECT target */
2104         RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
2105         RCU_INIT_POINTER(nf_ct_destroy, destroy_conntrack);
2106 }
2107
2108 /*
2109  * We need to use special "null" values, not used in hash table
2110  */
2111 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
2112 #define DYING_NULLS_VAL         ((1<<30)+1)
2113 #define TEMPLATE_NULLS_VAL      ((1<<30)+2)
2114
2115 int nf_conntrack_init_net(struct net *net)
2116 {
2117         int ret = -ENOMEM;
2118         int cpu;
2119
2120         BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2121         atomic_set(&net->ct.count, 0);
2122
2123         net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
2124         if (!net->ct.pcpu_lists)
2125                 goto err_stat;
2126
2127         for_each_possible_cpu(cpu) {
2128                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2129
2130                 spin_lock_init(&pcpu->lock);
2131                 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
2132                 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
2133         }
2134
2135         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2136         if (!net->ct.stat)
2137                 goto err_pcpu_lists;
2138
2139         ret = nf_conntrack_expect_pernet_init(net);
2140         if (ret < 0)
2141                 goto err_expect;
2142         ret = nf_conntrack_acct_pernet_init(net);
2143         if (ret < 0)
2144                 goto err_acct;
2145         ret = nf_conntrack_tstamp_pernet_init(net);
2146         if (ret < 0)
2147                 goto err_tstamp;
2148         ret = nf_conntrack_ecache_pernet_init(net);
2149         if (ret < 0)
2150                 goto err_ecache;
2151         ret = nf_conntrack_helper_pernet_init(net);
2152         if (ret < 0)
2153                 goto err_helper;
2154         ret = nf_conntrack_proto_pernet_init(net);
2155         if (ret < 0)
2156                 goto err_proto;
2157         return 0;
2158
2159 err_proto:
2160         nf_conntrack_helper_pernet_fini(net);
2161 err_helper:
2162         nf_conntrack_ecache_pernet_fini(net);
2163 err_ecache:
2164         nf_conntrack_tstamp_pernet_fini(net);
2165 err_tstamp:
2166         nf_conntrack_acct_pernet_fini(net);
2167 err_acct:
2168         nf_conntrack_expect_pernet_fini(net);
2169 err_expect:
2170         free_percpu(net->ct.stat);
2171 err_pcpu_lists:
2172         free_percpu(net->ct.pcpu_lists);
2173 err_stat:
2174         return ret;
2175 }