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 #include "nf_internals.h"
  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         /* 1) Acquire the lock */
 100         spin_lock(lock);
 101
 102         /* 2) read nf_conntrack_locks_all, with ACQUIRE semantics
 103          * It pairs with the smp_store_release() in nf_conntrack_all_unlock()
 104          */
 105         if (likely(smp_load_acquire(&nf_conntrack_locks_all) == false))
 106                 return;
 107
 108         /* fast path failed, unlock */
 109         spin_unlock(lock);
 110
 111         /* Slow path 1) get global lock */
 112         spin_lock(&nf_conntrack_locks_all_lock);
 113
 114         /* Slow path 2) get the lock we want */
 115         spin_lock(lock);
 116
 117         /* Slow path 3) release the global lock */
 118         spin_unlock(&nf_conntrack_locks_all_lock);
 119 }
 120 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
 121
 122 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 123 {
 124         h1 %= CONNTRACK_LOCKS;
 125         h2 %= CONNTRACK_LOCKS;
 126         spin_unlock(&nf_conntrack_locks[h1]);
 127         if (h1 != h2)
 128                 spin_unlock(&nf_conntrack_locks[h2]);
 129 }
 130
 131 /* return true if we need to recompute hashes (in case hash table was resized) */
 132 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
 133                                      unsigned int h2, unsigned int sequence)
 134 {
 135         h1 %= CONNTRACK_LOCKS;
 136         h2 %= CONNTRACK_LOCKS;
 137         if (h1 <= h2) {
 138                 nf_conntrack_lock(&nf_conntrack_locks[h1]);
 139                 if (h1 != h2)
 140                         spin_lock_nested(&nf_conntrack_locks[h2],
 141                                          SINGLE_DEPTH_NESTING);
 142         } else {
 143                 nf_conntrack_lock(&nf_conntrack_locks[h2]);
 144                 spin_lock_nested(&nf_conntrack_locks[h1],
 145                                  SINGLE_DEPTH_NESTING);
 146         }
 147         if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
 148                 nf_conntrack_double_unlock(h1, h2);
 149                 return true;
 150         }
 151         return false;
 152 }
 153
 154 static void nf_conntrack_all_lock(void)
 155 {
 156         int i;
 157
 158         spin_lock(&nf_conntrack_locks_all_lock);
 159
 160         nf_conntrack_locks_all = true;
 161
 162         for (i = 0; i < CONNTRACK_LOCKS; i++) {
 163                 spin_lock(&nf_conntrack_locks[i]);
 164
 165                 /* This spin_unlock provides the "release" to ensure that
 166                  * nf_conntrack_locks_all==true is visible to everyone that
 167                  * acquired spin_lock(&nf_conntrack_locks[]).
 168                  */
 169                 spin_unlock(&nf_conntrack_locks[i]);
 170         }
 171 }
 172
 173 static void nf_conntrack_all_unlock(void)
 174 {
 175         /* All prior stores must be complete before we clear
 176          * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
 177          * might observe the false value but not the entire
 178          * critical section.
 179          * It pairs with the smp_load_acquire() in nf_conntrack_lock()
 180          */
 181         smp_store_release(&nf_conntrack_locks_all, false);
 182         spin_unlock(&nf_conntrack_locks_all_lock);
 183 }
 184
 185 unsigned int nf_conntrack_htable_size __read_mostly;
 186 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
 187
 188 unsigned int nf_conntrack_max __read_mostly;
 189 seqcount_t nf_conntrack_generation __read_mostly;
 190 static unsigned int nf_conntrack_hash_rnd __read_mostly;
 191
 192 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
 193                               const struct net *net)
 194 {
 195         unsigned int n;
 196         u32 seed;
 197
 198         get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
 199
 200         /* The direction must be ignored, so we hash everything up to the
 201          * destination ports (which is a multiple of 4) and treat the last
 202          * three bytes manually.
 203          */
 204         seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
 205         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
 206         return jhash2((u32 *)tuple, n, seed ^
 207                       (((__force __u16)tuple->dst.u.all << 16) |
 208                       tuple->dst.protonum));
 209 }
 210
 211 static u32 scale_hash(u32 hash)
 212 {
 213         return reciprocal_scale(hash, nf_conntrack_htable_size);
 214 }
 215
 216 static u32 __hash_conntrack(const struct net *net,
 217                             const struct nf_conntrack_tuple *tuple,
 218                             unsigned int size)
 219 {
 220         return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
 221 }
 222
 223 static u32 hash_conntrack(const struct net *net,
 224                           const struct nf_conntrack_tuple *tuple)
 225 {
 226         return scale_hash(hash_conntrack_raw(tuple, net));
 227 }
 228
 229 bool
 230 nf_ct_get_tuple(const struct sk_buff *skb,
 231                 unsigned int nhoff,
 232                 unsigned int dataoff,
 233                 u_int16_t l3num,
 234                 u_int8_t protonum,
 235                 struct net *net,
 236                 struct nf_conntrack_tuple *tuple,
 237                 const struct nf_conntrack_l3proto *l3proto,
 238                 const struct nf_conntrack_l4proto *l4proto)
 239 {
 240         memset(tuple, 0, sizeof(*tuple));
 241
 242         tuple->src.l3num = l3num;
 243         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
 244                 return false;
 245
 246         tuple->dst.protonum = protonum;
 247         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 248
 249         return l4proto->pkt_to_tuple(skb, dataoff, net, tuple);
 250 }
 251 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
 252
 253 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 254                        u_int16_t l3num,
 255                        struct net *net, struct nf_conntrack_tuple *tuple)
 256 {
 257         const struct nf_conntrack_l3proto *l3proto;
 258         const struct nf_conntrack_l4proto *l4proto;
 259         unsigned int protoff;
 260         u_int8_t protonum;
 261         int ret;
 262
 263         rcu_read_lock();
 264
 265         l3proto = __nf_ct_l3proto_find(l3num);
 266         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
 267         if (ret != NF_ACCEPT) {
 268                 rcu_read_unlock();
 269                 return false;
 270         }
 271
 272         l4proto = __nf_ct_l4proto_find(l3num, protonum);
 273
 274         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple,
 275                               l3proto, l4proto);
 276
 277         rcu_read_unlock();
 278         return ret;
 279 }
 280 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 281
 282 bool
 283 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 284                    const struct nf_conntrack_tuple *orig,
 285                    const struct nf_conntrack_l3proto *l3proto,
 286                    const struct nf_conntrack_l4proto *l4proto)
 287 {
 288         memset(inverse, 0, sizeof(*inverse));
 289
 290         inverse->src.l3num = orig->src.l3num;
 291         if (l3proto->invert_tuple(inverse, orig) == 0)
 292                 return false;
 293
 294         inverse->dst.dir = !orig->dst.dir;
 295
 296         inverse->dst.protonum = orig->dst.protonum;
 297         return l4proto->invert_tuple(inverse, orig);
 298 }
 299 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 300
 301 static void
 302 clean_from_lists(struct nf_conn *ct)
 303 {
 304         pr_debug("clean_from_lists(%p)\n", ct);
 305         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 306         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 307
 308         /* Destroy all pending expectations */
 309         nf_ct_remove_expectations(ct);
 310 }
 311
 312 /* must be called with local_bh_disable */
 313 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
 314 {
 315         struct ct_pcpu *pcpu;
 316
 317         /* add this conntrack to the (per cpu) dying list */
 318         ct->cpu = smp_processor_id();
 319         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 320
 321         spin_lock(&pcpu->lock);
 322         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 323                              &pcpu->dying);
 324         spin_unlock(&pcpu->lock);
 325 }
 326
 327 /* must be called with local_bh_disable */
 328 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
 329 {
 330         struct ct_pcpu *pcpu;
 331
 332         /* add this conntrack to the (per cpu) unconfirmed list */
 333         ct->cpu = smp_processor_id();
 334         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 335
 336         spin_lock(&pcpu->lock);
 337         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 338                              &pcpu->unconfirmed);
 339         spin_unlock(&pcpu->lock);
 340 }
 341
 342 /* must be called with local_bh_disable */
 343 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
 344 {
 345         struct ct_pcpu *pcpu;
 346
 347         /* We overload first tuple to link into unconfirmed or dying list.*/
 348         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 349
 350         spin_lock(&pcpu->lock);
 351         BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 352         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 353         spin_unlock(&pcpu->lock);
 354 }
 355
 356 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
 357
 358 /* Released via destroy_conntrack() */
 359 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
 360                                  const struct nf_conntrack_zone *zone,
 361                                  gfp_t flags)
 362 {
 363         struct nf_conn *tmpl, *p;
 364
 365         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
 366                 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
 367                 if (!tmpl)
 368                         return NULL;
 369
 370                 p = tmpl;
 371                 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 372                 if (tmpl != p) {
 373                         tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 374                         tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
 375                 }
 376         } else {
 377                 tmpl = kzalloc(sizeof(*tmpl), flags);
 378                 if (!tmpl)
 379                         return NULL;
 380         }
 381
 382         tmpl->status = IPS_TEMPLATE;
 383         write_pnet(&tmpl->ct_net, net);
 384         nf_ct_zone_add(tmpl, zone);
 385         atomic_set(&tmpl->ct_general.use, 0);
 386
 387         return tmpl;
 388 }
 389 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
 390
 391 void nf_ct_tmpl_free(struct nf_conn *tmpl)
 392 {
 393         nf_ct_ext_destroy(tmpl);
 394         nf_ct_ext_free(tmpl);
 395
 396         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
 397                 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
 398         else
 399                 kfree(tmpl);
 400 }
 401 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
 402
 403 static void
 404 destroy_conntrack(struct nf_conntrack *nfct)
 405 {
 406         struct nf_conn *ct = (struct nf_conn *)nfct;
 407         const struct nf_conntrack_l4proto *l4proto;
 408
 409         pr_debug("destroy_conntrack(%p)\n", ct);
 410         WARN_ON(atomic_read(&nfct->use) != 0);
 411
 412         if (unlikely(nf_ct_is_template(ct))) {
 413                 nf_ct_tmpl_free(ct);
 414                 return;
 415         }
 416         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 417         if (l4proto->destroy)
 418                 l4proto->destroy(ct);
 419
 420         local_bh_disable();
 421         /* Expectations will have been removed in clean_from_lists,
 422          * except TFTP can create an expectation on the first packet,
 423          * before connection is in the list, so we need to clean here,
 424          * too.
 425          */
 426         nf_ct_remove_expectations(ct);
 427
 428         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 429
 430         local_bh_enable();
 431
 432         if (ct->master)
 433                 nf_ct_put(ct->master);
 434
 435         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 436         nf_conntrack_free(ct);
 437 }
 438
 439 static void nf_ct_delete_from_lists(struct nf_conn *ct)
 440 {
 441         struct net *net = nf_ct_net(ct);
 442         unsigned int hash, reply_hash;
 443         unsigned int sequence;
 444
 445         nf_ct_helper_destroy(ct);
 446
 447         local_bh_disable();
 448         do {
 449                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 450                 hash = hash_conntrack(net,
 451                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 452                 reply_hash = hash_conntrack(net,
 453                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 454         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 455
 456         clean_from_lists(ct);
 457         nf_conntrack_double_unlock(hash, reply_hash);
 458
 459         nf_ct_add_to_dying_list(ct);
 460
 461         local_bh_enable();
 462 }
 463
 464 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
 465 {
 466         struct nf_conn_tstamp *tstamp;
 467
 468         if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
 469                 return false;
 470
 471         tstamp = nf_conn_tstamp_find(ct);
 472         if (tstamp && tstamp->stop == 0)
 473                 tstamp->stop = ktime_get_real_ns();
 474
 475         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 476                                     portid, report) < 0) {
 477                 /* destroy event was not delivered. nf_ct_put will
 478                  * be done by event cache worker on redelivery.
 479                  */
 480                 nf_ct_delete_from_lists(ct);
 481                 nf_conntrack_ecache_delayed_work(nf_ct_net(ct));
 482                 return false;
 483         }
 484
 485         nf_conntrack_ecache_work(nf_ct_net(ct));
 486         nf_ct_delete_from_lists(ct);
 487         nf_ct_put(ct);
 488         return true;
 489 }
 490 EXPORT_SYMBOL_GPL(nf_ct_delete);
 491
 492 static inline bool
 493 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
 494                 const struct nf_conntrack_tuple *tuple,
 495                 const struct nf_conntrack_zone *zone,
 496                 const struct net *net)
 497 {
 498         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 499
 500         /* A conntrack can be recreated with the equal tuple,
 501          * so we need to check that the conntrack is confirmed
 502          */
 503         return nf_ct_tuple_equal(tuple, &h->tuple) &&
 504                nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
 505                nf_ct_is_confirmed(ct) &&
 506                net_eq(net, nf_ct_net(ct));
 507 }
 508
 509 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
 510 static void nf_ct_gc_expired(struct nf_conn *ct)
 511 {
 512         if (!atomic_inc_not_zero(&ct->ct_general.use))
 513                 return;
 514
 515         if (nf_ct_should_gc(ct))
 516                 nf_ct_kill(ct);
 517
 518         nf_ct_put(ct);
 519 }
 520
 521 /*
 522  * Warning :
 523  * - Caller must take a reference on returned object
 524  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 525  */
 526 static struct nf_conntrack_tuple_hash *
 527 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
 528                       const struct nf_conntrack_tuple *tuple, u32 hash)
 529 {
 530         struct nf_conntrack_tuple_hash *h;
 531         struct hlist_nulls_head *ct_hash;
 532         struct hlist_nulls_node *n;
 533         unsigned int bucket, hsize;
 534
 535 begin:
 536         nf_conntrack_get_ht(&ct_hash, &hsize);
 537         bucket = reciprocal_scale(hash, hsize);
 538
 539         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
 540                 struct nf_conn *ct;
 541
 542                 ct = nf_ct_tuplehash_to_ctrack(h);
 543                 if (nf_ct_is_expired(ct)) {
 544                         nf_ct_gc_expired(ct);
 545                         continue;
 546                 }
 547
 548                 if (nf_ct_is_dying(ct))
 549                         continue;
 550
 551                 if (nf_ct_key_equal(h, tuple, zone, net))
 552                         return h;
 553         }
 554         /*
 555          * if the nulls value we got at the end of this lookup is
 556          * not the expected one, we must restart lookup.
 557          * We probably met an item that was moved to another chain.
 558          */
 559         if (get_nulls_value(n) != bucket) {
 560                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 561                 goto begin;
 562         }
 563
 564         return NULL;
 565 }
 566
 567 /* Find a connection corresponding to a tuple. */
 568 static struct nf_conntrack_tuple_hash *
 569 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 570                         const struct nf_conntrack_tuple *tuple, u32 hash)
 571 {
 572         struct nf_conntrack_tuple_hash *h;
 573         struct nf_conn *ct;
 574
 575         rcu_read_lock();
 576 begin:
 577         h = ____nf_conntrack_find(net, zone, tuple, hash);
 578         if (h) {
 579                 ct = nf_ct_tuplehash_to_ctrack(h);
 580                 if (unlikely(nf_ct_is_dying(ct) ||
 581                              !atomic_inc_not_zero(&ct->ct_general.use)))
 582                         h = NULL;
 583                 else {
 584                         if (unlikely(!nf_ct_key_equal(h, tuple, zone, net))) {
 585                                 nf_ct_put(ct);
 586                                 goto begin;
 587                         }
 588                 }
 589         }
 590         rcu_read_unlock();
 591
 592         return h;
 593 }
 594
 595 struct nf_conntrack_tuple_hash *
 596 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 597                       const struct nf_conntrack_tuple *tuple)
 598 {
 599         return __nf_conntrack_find_get(net, zone, tuple,
 600                                        hash_conntrack_raw(tuple, net));
 601 }
 602 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 603
 604 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 605                                        unsigned int hash,
 606                                        unsigned int reply_hash)
 607 {
 608         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 609                            &nf_conntrack_hash[hash]);
 610         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 611                            &nf_conntrack_hash[reply_hash]);
 612 }
 613
 614 int
 615 nf_conntrack_hash_check_insert(struct nf_conn *ct)
 616 {
 617         const struct nf_conntrack_zone *zone;
 618         struct net *net = nf_ct_net(ct);
 619         unsigned int hash, reply_hash;
 620         struct nf_conntrack_tuple_hash *h;
 621         struct hlist_nulls_node *n;
 622         unsigned int sequence;
 623
 624         zone = nf_ct_zone(ct);
 625
 626         local_bh_disable();
 627         do {
 628                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 629                 hash = hash_conntrack(net,
 630                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 631                 reply_hash = hash_conntrack(net,
 632                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 633         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 634
 635         /* See if there's one in the list already, including reverse */
 636         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 637                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 638                                     zone, net))
 639                         goto out;
 640
 641         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 642                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 643                                     zone, net))
 644                         goto out;
 645
 646         smp_wmb();
 647         /* The caller holds a reference to this object */
 648         atomic_set(&ct->ct_general.use, 2);
 649         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 650         nf_conntrack_double_unlock(hash, reply_hash);
 651         NF_CT_STAT_INC(net, insert);
 652         local_bh_enable();
 653         return 0;
 654
 655 out:
 656         nf_conntrack_double_unlock(hash, reply_hash);
 657         NF_CT_STAT_INC(net, insert_failed);
 658         local_bh_enable();
 659         return -EEXIST;
 660 }
 661 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 662
 663 static inline void nf_ct_acct_update(struct nf_conn *ct,
 664                                      enum ip_conntrack_info ctinfo,
 665                                      unsigned int len)
 666 {
 667         struct nf_conn_acct *acct;
 668
 669         acct = nf_conn_acct_find(ct);
 670         if (acct) {
 671                 struct nf_conn_counter *counter = acct->counter;
 672
 673                 atomic64_inc(&counter[CTINFO2DIR(ctinfo)].packets);
 674                 atomic64_add(len, &counter[CTINFO2DIR(ctinfo)].bytes);
 675         }
 676 }
 677
 678 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
 679                              const struct nf_conn *loser_ct)
 680 {
 681         struct nf_conn_acct *acct;
 682
 683         acct = nf_conn_acct_find(loser_ct);
 684         if (acct) {
 685                 struct nf_conn_counter *counter = acct->counter;
 686                 unsigned int bytes;
 687
 688                 /* u32 should be fine since we must have seen one packet. */
 689                 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
 690                 nf_ct_acct_update(ct, ctinfo, bytes);
 691         }
 692 }
 693
 694 /* Resolve race on insertion if this protocol allows this. */
 695 static int nf_ct_resolve_clash(struct net *net, struct sk_buff *skb,
 696                                enum ip_conntrack_info ctinfo,
 697                                struct nf_conntrack_tuple_hash *h)
 698 {
 699         /* This is the conntrack entry already in hashes that won race. */
 700         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 701         const struct nf_conntrack_l4proto *l4proto;
 702
 703         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 704         if (l4proto->allow_clash &&
 705             ((ct->status & IPS_NAT_DONE_MASK) == 0) &&
 706             !nf_ct_is_dying(ct) &&
 707             atomic_inc_not_zero(&ct->ct_general.use)) {
 708                 enum ip_conntrack_info oldinfo;
 709                 struct nf_conn *loser_ct = nf_ct_get(skb, &oldinfo);
 710
 711                 nf_ct_acct_merge(ct, ctinfo, loser_ct);
 712                 nf_conntrack_put(&loser_ct->ct_general);
 713                 nf_ct_set(skb, ct, oldinfo);
 714                 return NF_ACCEPT;
 715         }
 716         NF_CT_STAT_INC(net, drop);
 717         return NF_DROP;
 718 }
 719
 720 /* Confirm a connection given skb; places it in hash table */
 721 int
 722 __nf_conntrack_confirm(struct sk_buff *skb)
 723 {
 724         const struct nf_conntrack_zone *zone;
 725         unsigned int hash, reply_hash;
 726         struct nf_conntrack_tuple_hash *h;
 727         struct nf_conn *ct;
 728         struct nf_conn_help *help;
 729         struct nf_conn_tstamp *tstamp;
 730         struct hlist_nulls_node *n;
 731         enum ip_conntrack_info ctinfo;
 732         struct net *net;
 733         unsigned int sequence;
 734         int ret = NF_DROP;
 735
 736         ct = nf_ct_get(skb, &ctinfo);
 737         net = nf_ct_net(ct);
 738
 739         /* ipt_REJECT uses nf_conntrack_attach to attach related
 740            ICMP/TCP RST packets in other direction.  Actual packet
 741            which created connection will be IP_CT_NEW or for an
 742            expected connection, IP_CT_RELATED. */
 743         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
 744                 return NF_ACCEPT;
 745
 746         zone = nf_ct_zone(ct);
 747         local_bh_disable();
 748
 749         do {
 750                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 751                 /* reuse the hash saved before */
 752                 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
 753                 hash = scale_hash(hash);
 754                 reply_hash = hash_conntrack(net,
 755                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 756
 757         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 758
 759         /* We're not in hash table, and we refuse to set up related
 760          * connections for unconfirmed conns.  But packet copies and
 761          * REJECT will give spurious warnings here.
 762          */
 763
 764         /* No external references means no one else could have
 765          * confirmed us.
 766          */
 767         WARN_ON(nf_ct_is_confirmed(ct));
 768         pr_debug("Confirming conntrack %p\n", ct);
 769         /* We have to check the DYING flag after unlink to prevent
 770          * a race against nf_ct_get_next_corpse() possibly called from
 771          * user context, else we insert an already 'dead' hash, blocking
 772          * further use of that particular connection -JM.
 773          */
 774         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 775
 776         if (unlikely(nf_ct_is_dying(ct))) {
 777                 nf_ct_add_to_dying_list(ct);
 778                 goto dying;
 779         }
 780
 781         /* See if there's one in the list already, including reverse:
 782            NAT could have grabbed it without realizing, since we're
 783            not in the hash.  If there is, we lost race. */
 784         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 785                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 786                                     zone, net))
 787                         goto out;
 788
 789         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 790                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 791                                     zone, net))
 792                         goto out;
 793
 794         /* Timer relative to confirmation time, not original
 795            setting time, otherwise we'd get timer wrap in
 796            weird delay cases. */
 797         ct->timeout += nfct_time_stamp;
 798         atomic_inc(&ct->ct_general.use);
 799         ct->status |= IPS_CONFIRMED;
 800
 801         /* set conntrack timestamp, if enabled. */
 802         tstamp = nf_conn_tstamp_find(ct);
 803         if (tstamp) {
 804                 if (skb->tstamp == 0)
 805                         __net_timestamp(skb);
 806
 807                 tstamp->start = ktime_to_ns(skb->tstamp);
 808         }
 809         /* Since the lookup is lockless, hash insertion must be done after
 810          * starting the timer and setting the CONFIRMED bit. The RCU barriers
 811          * guarantee that no other CPU can find the conntrack before the above
 812          * stores are visible.
 813          */
 814         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 815         nf_conntrack_double_unlock(hash, reply_hash);
 816         local_bh_enable();
 817
 818         help = nfct_help(ct);
 819         if (help && help->helper)
 820                 nf_conntrack_event_cache(IPCT_HELPER, ct);
 821
 822         nf_conntrack_event_cache(master_ct(ct) ?
 823                                  IPCT_RELATED : IPCT_NEW, ct);
 824         return NF_ACCEPT;
 825
 826 out:
 827         nf_ct_add_to_dying_list(ct);
 828         ret = nf_ct_resolve_clash(net, skb, ctinfo, h);
 829 dying:
 830         nf_conntrack_double_unlock(hash, reply_hash);
 831         NF_CT_STAT_INC(net, insert_failed);
 832         local_bh_enable();
 833         return ret;
 834 }
 835 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
 836
 837 /* Returns true if a connection correspondings to the tuple (required
 838    for NAT). */
 839 int
 840 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 841                          const struct nf_conn *ignored_conntrack)
 842 {
 843         struct net *net = nf_ct_net(ignored_conntrack);
 844         const struct nf_conntrack_zone *zone;
 845         struct nf_conntrack_tuple_hash *h;
 846         struct hlist_nulls_head *ct_hash;
 847         unsigned int hash, hsize;
 848         struct hlist_nulls_node *n;
 849         struct nf_conn *ct;
 850
 851         zone = nf_ct_zone(ignored_conntrack);
 852
 853         rcu_read_lock();
 854  begin:
 855         nf_conntrack_get_ht(&ct_hash, &hsize);
 856         hash = __hash_conntrack(net, tuple, hsize);
 857
 858         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
 859                 ct = nf_ct_tuplehash_to_ctrack(h);
 860
 861                 if (ct == ignored_conntrack)
 862                         continue;
 863
 864                 if (nf_ct_is_expired(ct)) {
 865                         nf_ct_gc_expired(ct);
 866                         continue;
 867                 }
 868
 869                 if (nf_ct_key_equal(h, tuple, zone, net)) {
 870                         NF_CT_STAT_INC_ATOMIC(net, found);
 871                         rcu_read_unlock();
 872                         return 1;
 873                 }
 874         }
 875
 876         if (get_nulls_value(n) != hash) {
 877                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 878                 goto begin;
 879         }
 880
 881         rcu_read_unlock();
 882
 883         return 0;
 884 }
 885 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
 886
 887 #define NF_CT_EVICTION_RANGE    8
 888
 889 /* There's a small race here where we may free a just-assured
 890    connection.  Too bad: we're in trouble anyway. */
 891 static unsigned int early_drop_list(struct net *net,
 892                                     struct hlist_nulls_head *head)
 893 {
 894         struct nf_conntrack_tuple_hash *h;
 895         struct hlist_nulls_node *n;
 896         unsigned int drops = 0;
 897         struct nf_conn *tmp;
 898
 899         hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
 900                 tmp = nf_ct_tuplehash_to_ctrack(h);
 901
 902                 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status))
 903                         continue;
 904
 905                 if (nf_ct_is_expired(tmp)) {
 906                         nf_ct_gc_expired(tmp);
 907                         continue;
 908                 }
 909
 910                 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
 911                     !net_eq(nf_ct_net(tmp), net) ||
 912                     nf_ct_is_dying(tmp))
 913                         continue;
 914
 915                 if (!atomic_inc_not_zero(&tmp->ct_general.use))
 916                         continue;
 917
 918                 /* kill only if still in same netns -- might have moved due to
 919                  * SLAB_TYPESAFE_BY_RCU rules.
 920                  *
 921                  * We steal the timer reference.  If that fails timer has
 922                  * already fired or someone else deleted it. Just drop ref
 923                  * and move to next entry.
 924                  */
 925                 if (net_eq(nf_ct_net(tmp), net) &&
 926                     nf_ct_is_confirmed(tmp) &&
 927                     nf_ct_delete(tmp, 0, 0))
 928                         drops++;
 929
 930                 nf_ct_put(tmp);
 931         }
 932
 933         return drops;
 934 }
 935
 936 static noinline int early_drop(struct net *net, unsigned int _hash)
 937 {
 938         unsigned int i;
 939
 940         for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
 941                 struct hlist_nulls_head *ct_hash;
 942                 unsigned int hash, hsize, drops;
 943
 944                 rcu_read_lock();
 945                 nf_conntrack_get_ht(&ct_hash, &hsize);
 946                 hash = reciprocal_scale(_hash++, hsize);
 947
 948                 drops = early_drop_list(net, &ct_hash[hash]);
 949                 rcu_read_unlock();
 950
 951                 if (drops) {
 952                         NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
 953                         return true;
 954                 }
 955         }
 956
 957         return false;
 958 }
 959
 960 static bool gc_worker_skip_ct(const struct nf_conn *ct)
 961 {
 962         return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
 963 }
 964
 965 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
 966 {
 967         const struct nf_conntrack_l4proto *l4proto;
 968
 969         if (!test_bit(IPS_ASSURED_BIT, &ct->status))
 970                 return true;
 971
 972         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 973         if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
 974                 return true;
 975
 976         return false;
 977 }
 978
 979 #define DAY     (86400 * HZ)
 980
 981 /* Set an arbitrary timeout large enough not to ever expire, this save
 982  * us a check for the IPS_OFFLOAD_BIT from the packet path via
 983  * nf_ct_is_expired().
 984  */
 985 static void nf_ct_offload_timeout(struct nf_conn *ct)
 986 {
 987         if (nf_ct_expires(ct) < DAY / 2)
 988                 ct->timeout = nfct_time_stamp + DAY;
 989 }
 990
 991 static void gc_worker(struct work_struct *work)
 992 {
 993         unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
 994         unsigned int i, goal, buckets = 0, expired_count = 0;
 995         unsigned int nf_conntrack_max95 = 0;
 996         struct conntrack_gc_work *gc_work;
 997         unsigned int ratio, scanned = 0;
 998         unsigned long next_run;
 999
1000         gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1001
1002         goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
1003         i = gc_work->last_bucket;
1004         if (gc_work->early_drop)
1005                 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1006
1007         do {
1008                 struct nf_conntrack_tuple_hash *h;
1009                 struct hlist_nulls_head *ct_hash;
1010                 struct hlist_nulls_node *n;
1011                 unsigned int hashsz;
1012                 struct nf_conn *tmp;
1013
1014                 i++;
1015                 rcu_read_lock();
1016
1017                 nf_conntrack_get_ht(&ct_hash, &hashsz);
1018                 if (i >= hashsz)
1019                         i = 0;
1020
1021                 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1022                         struct net *net;
1023
1024                         tmp = nf_ct_tuplehash_to_ctrack(h);
1025
1026                         scanned++;
1027                         if (test_bit(IPS_OFFLOAD_BIT, &tmp->status)) {
1028                                 nf_ct_offload_timeout(tmp);
1029                                 continue;
1030                         }
1031
1032                         if (nf_ct_is_expired(tmp)) {
1033                                 nf_ct_gc_expired(tmp);
1034                                 expired_count++;
1035                                 continue;
1036                         }
1037
1038                         if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1039                                 continue;
1040
1041                         net = nf_ct_net(tmp);
1042                         if (atomic_read(&net->ct.count) < nf_conntrack_max95)
1043                                 continue;
1044
1045                         /* need to take reference to avoid possible races */
1046                         if (!atomic_inc_not_zero(&tmp->ct_general.use))
1047                                 continue;
1048
1049                         if (gc_worker_skip_ct(tmp)) {
1050                                 nf_ct_put(tmp);
1051                                 continue;
1052                         }
1053
1054                         if (gc_worker_can_early_drop(tmp))
1055                                 nf_ct_kill(tmp);
1056
1057                         nf_ct_put(tmp);
1058                 }
1059
1060                 /* could check get_nulls_value() here and restart if ct
1061                  * was moved to another chain.  But given gc is best-effort
1062                  * we will just continue with next hash slot.
1063                  */
1064                 rcu_read_unlock();
1065                 cond_resched();
1066         } while (++buckets < goal);
1067
1068         if (gc_work->exiting)
1069                 return;
1070
1071         /*
1072          * Eviction will normally happen from the packet path, and not
1073          * from this gc worker.
1074          *
1075          * This worker is only here to reap expired entries when system went
1076          * idle after a busy period.
1077          *
1078          * The heuristics below are supposed to balance conflicting goals:
1079          *
1080          * 1. Minimize time until we notice a stale entry
1081          * 2. Maximize scan intervals to not waste cycles
1082          *
1083          * Normally, expire ratio will be close to 0.
1084          *
1085          * As soon as a sizeable fraction of the entries have expired
1086          * increase scan frequency.
1087          */
1088         ratio = scanned ? expired_count * 100 / scanned : 0;
1089         if (ratio > GC_EVICT_RATIO) {
1090                 gc_work->next_gc_run = min_interval;
1091         } else {
1092                 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1093
1094                 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1095
1096                 gc_work->next_gc_run += min_interval;
1097                 if (gc_work->next_gc_run > max)
1098                         gc_work->next_gc_run = max;
1099         }
1100
1101         next_run = gc_work->next_gc_run;
1102         gc_work->last_bucket = i;
1103         gc_work->early_drop = false;
1104         queue_delayed_work(system_power_efficient_wq, &gc_work->dwork, next_run);
1105 }
1106
1107 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1108 {
1109         INIT_DEFERRABLE_WORK(&gc_work->dwork, gc_worker);
1110         gc_work->next_gc_run = HZ;
1111         gc_work->exiting = false;
1112 }
1113
1114 static struct nf_conn *
1115 __nf_conntrack_alloc(struct net *net,
1116                      const struct nf_conntrack_zone *zone,
1117                      const struct nf_conntrack_tuple *orig,
1118                      const struct nf_conntrack_tuple *repl,
1119                      gfp_t gfp, u32 hash)
1120 {
1121         struct nf_conn *ct;
1122
1123         /* We don't want any race condition at early drop stage */
1124         atomic_inc(&net->ct.count);
1125
1126         if (nf_conntrack_max &&
1127             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
1128                 if (!early_drop(net, hash)) {
1129                         if (!conntrack_gc_work.early_drop)
1130                                 conntrack_gc_work.early_drop = true;
1131                         atomic_dec(&net->ct.count);
1132                         net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1133                         return ERR_PTR(-ENOMEM);
1134                 }
1135         }
1136
1137         /*
1138          * Do not use kmem_cache_zalloc(), as this cache uses
1139          * SLAB_TYPESAFE_BY_RCU.
1140          */
1141         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1142         if (ct == NULL)
1143                 goto out;
1144
1145         spin_lock_init(&ct->lock);
1146         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1147         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1148         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1149         /* save hash for reusing when confirming */
1150         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1151         ct->status = 0;
1152         write_pnet(&ct->ct_net, net);
1153         memset(&ct->__nfct_init_offset[0], 0,
1154                offsetof(struct nf_conn, proto) -
1155                offsetof(struct nf_conn, __nfct_init_offset[0]));
1156
1157         nf_ct_zone_add(ct, zone);
1158
1159         /* Because we use RCU lookups, we set ct_general.use to zero before
1160          * this is inserted in any list.
1161          */
1162         atomic_set(&ct->ct_general.use, 0);
1163         return ct;
1164 out:
1165         atomic_dec(&net->ct.count);
1166         return ERR_PTR(-ENOMEM);
1167 }
1168
1169 struct nf_conn *nf_conntrack_alloc(struct net *net,
1170                                    const struct nf_conntrack_zone *zone,
1171                                    const struct nf_conntrack_tuple *orig,
1172                                    const struct nf_conntrack_tuple *repl,
1173                                    gfp_t gfp)
1174 {
1175         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1176 }
1177 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1178
1179 void nf_conntrack_free(struct nf_conn *ct)
1180 {
1181         struct net *net = nf_ct_net(ct);
1182
1183         /* A freed object has refcnt == 0, that's
1184          * the golden rule for SLAB_TYPESAFE_BY_RCU
1185          */
1186         WARN_ON(atomic_read(&ct->ct_general.use) != 0);
1187
1188         nf_ct_ext_destroy(ct);
1189         nf_ct_ext_free(ct);
1190         kmem_cache_free(nf_conntrack_cachep, ct);
1191         smp_mb__before_atomic();
1192         atomic_dec(&net->ct.count);
1193 }
1194 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1195
1196
1197 /* Allocate a new conntrack: we return -ENOMEM if classification
1198    failed due to stress.  Otherwise it really is unclassifiable. */
1199 static noinline struct nf_conntrack_tuple_hash *
1200 init_conntrack(struct net *net, struct nf_conn *tmpl,
1201                const struct nf_conntrack_tuple *tuple,
1202                const struct nf_conntrack_l3proto *l3proto,
1203                const struct nf_conntrack_l4proto *l4proto,
1204                struct sk_buff *skb,
1205                unsigned int dataoff, u32 hash)
1206 {
1207         struct nf_conn *ct;
1208         struct nf_conn_help *help;
1209         struct nf_conntrack_tuple repl_tuple;
1210         struct nf_conntrack_ecache *ecache;
1211         struct nf_conntrack_expect *exp = NULL;
1212         const struct nf_conntrack_zone *zone;
1213         struct nf_conn_timeout *timeout_ext;
1214         struct nf_conntrack_zone tmp;
1215         unsigned int *timeouts;
1216
1217         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
1218                 pr_debug("Can't invert tuple.\n");
1219                 return NULL;
1220         }
1221
1222         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1223         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1224                                   hash);
1225         if (IS_ERR(ct))
1226                 return (struct nf_conntrack_tuple_hash *)ct;
1227
1228         if (!nf_ct_add_synproxy(ct, tmpl)) {
1229                 nf_conntrack_free(ct);
1230                 return ERR_PTR(-ENOMEM);
1231         }
1232
1233         timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1234         if (timeout_ext) {
1235                 timeouts = nf_ct_timeout_data(timeout_ext);
1236                 if (unlikely(!timeouts))
1237                         timeouts = l4proto->get_timeouts(net);
1238         } else {
1239                 timeouts = l4proto->get_timeouts(net);
1240         }
1241
1242         if (!l4proto->new(ct, skb, dataoff, timeouts)) {
1243                 nf_conntrack_free(ct);
1244                 pr_debug("can't track with proto module\n");
1245                 return NULL;
1246         }
1247
1248         if (timeout_ext)
1249                 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1250                                       GFP_ATOMIC);
1251
1252         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1253         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1254         nf_ct_labels_ext_add(ct);
1255
1256         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1257         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1258                                  ecache ? ecache->expmask : 0,
1259                              GFP_ATOMIC);
1260
1261         local_bh_disable();
1262         if (net->ct.expect_count) {
1263                 spin_lock(&nf_conntrack_expect_lock);
1264                 exp = nf_ct_find_expectation(net, zone, tuple);
1265                 if (exp) {
1266                         pr_debug("expectation arrives ct=%p exp=%p\n",
1267                                  ct, exp);
1268                         /* Welcome, Mr. Bond.  We've been expecting you... */
1269                         __set_bit(IPS_EXPECTED_BIT, &ct->status);
1270                         /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1271                         ct->master = exp->master;
1272                         if (exp->helper) {
1273                                 help = nf_ct_helper_ext_add(ct, exp->helper,
1274                                                             GFP_ATOMIC);
1275                                 if (help)
1276                                         rcu_assign_pointer(help->helper, exp->helper);
1277                         }
1278
1279 #ifdef CONFIG_NF_CONNTRACK_MARK
1280                         ct->mark = exp->master->mark;
1281 #endif
1282 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1283                         ct->secmark = exp->master->secmark;
1284 #endif
1285                         NF_CT_STAT_INC(net, expect_new);
1286                 }
1287                 spin_unlock(&nf_conntrack_expect_lock);
1288         }
1289         if (!exp)
1290                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1291
1292         /* Now it is inserted into the unconfirmed list, bump refcount */
1293         nf_conntrack_get(&ct->ct_general);
1294         nf_ct_add_to_unconfirmed_list(ct);
1295
1296         local_bh_enable();
1297
1298         if (exp) {
1299                 if (exp->expectfn)
1300                         exp->expectfn(ct, exp);
1301                 nf_ct_expect_put(exp);
1302         }
1303
1304         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1305 }
1306
1307 /* On success, returns 0, sets skb->_nfct | ctinfo */
1308 static int
1309 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
1310                   struct sk_buff *skb,
1311                   unsigned int dataoff,
1312                   u_int16_t l3num,
1313                   u_int8_t protonum,
1314                   const struct nf_conntrack_l3proto *l3proto,
1315                   const struct nf_conntrack_l4proto *l4proto)
1316 {
1317         const struct nf_conntrack_zone *zone;
1318         struct nf_conntrack_tuple tuple;
1319         struct nf_conntrack_tuple_hash *h;
1320         enum ip_conntrack_info ctinfo;
1321         struct nf_conntrack_zone tmp;
1322         struct nf_conn *ct;
1323         u32 hash;
1324
1325         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1326                              dataoff, l3num, protonum, net, &tuple, l3proto,
1327                              l4proto)) {
1328                 pr_debug("Can't get tuple\n");
1329                 return 0;
1330         }
1331
1332         /* look for tuple match */
1333         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1334         hash = hash_conntrack_raw(&tuple, net);
1335         h = __nf_conntrack_find_get(net, zone, &tuple, hash);
1336         if (!h) {
1337                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
1338                                    skb, dataoff, hash);
1339                 if (!h)
1340                         return 0;
1341                 if (IS_ERR(h))
1342                         return PTR_ERR(h);
1343         }
1344         ct = nf_ct_tuplehash_to_ctrack(h);
1345
1346         /* It exists; we have (non-exclusive) reference. */
1347         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1348                 ctinfo = IP_CT_ESTABLISHED_REPLY;
1349         } else {
1350                 /* Once we've had two way comms, always ESTABLISHED. */
1351                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1352                         pr_debug("normal packet for %p\n", ct);
1353                         ctinfo = IP_CT_ESTABLISHED;
1354                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1355                         pr_debug("related packet for %p\n", ct);
1356                         ctinfo = IP_CT_RELATED;
1357                 } else {
1358                         pr_debug("new packet for %p\n", ct);
1359                         ctinfo = IP_CT_NEW;
1360                 }
1361         }
1362         nf_ct_set(skb, ct, ctinfo);
1363         return 0;
1364 }
1365
1366 unsigned int
1367 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
1368                 struct sk_buff *skb)
1369 {
1370         const struct nf_conntrack_l3proto *l3proto;
1371         const struct nf_conntrack_l4proto *l4proto;
1372         struct nf_conn *ct, *tmpl;
1373         enum ip_conntrack_info ctinfo;
1374         unsigned int *timeouts;
1375         unsigned int dataoff;
1376         u_int8_t protonum;
1377         int ret;
1378
1379         tmpl = nf_ct_get(skb, &ctinfo);
1380         if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1381                 /* Previously seen (loopback or untracked)?  Ignore. */
1382                 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1383                      ctinfo == IP_CT_UNTRACKED) {
1384                         NF_CT_STAT_INC_ATOMIC(net, ignore);
1385                         return NF_ACCEPT;
1386                 }
1387                 skb->_nfct = 0;
1388         }
1389
1390         /* rcu_read_lock()ed by nf_hook_thresh */
1391         l3proto = __nf_ct_l3proto_find(pf);
1392         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
1393                                    &dataoff, &protonum);
1394         if (ret <= 0) {
1395                 pr_debug("not prepared to track yet or error occurred\n");
1396                 NF_CT_STAT_INC_ATOMIC(net, error);
1397                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1398                 ret = -ret;
1399                 goto out;
1400         }
1401
1402         l4proto = __nf_ct_l4proto_find(pf, protonum);
1403
1404         /* It may be an special packet, error, unclean...
1405          * inverse of the return code tells to the netfilter
1406          * core what to do with the packet. */
1407         if (l4proto->error != NULL) {
1408                 ret = l4proto->error(net, tmpl, skb, dataoff, pf, hooknum);
1409                 if (ret <= 0) {
1410                         NF_CT_STAT_INC_ATOMIC(net, error);
1411                         NF_CT_STAT_INC_ATOMIC(net, invalid);
1412                         ret = -ret;
1413                         goto out;
1414                 }
1415                 /* ICMP[v6] protocol trackers may assign one conntrack. */
1416                 if (skb->_nfct)
1417                         goto out;
1418         }
1419 repeat:
1420         ret = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
1421                                 l3proto, l4proto);
1422         if (ret < 0) {
1423                 /* Too stressed to deal. */
1424                 NF_CT_STAT_INC_ATOMIC(net, drop);
1425                 ret = NF_DROP;
1426                 goto out;
1427         }
1428
1429         ct = nf_ct_get(skb, &ctinfo);
1430         if (!ct) {
1431                 /* Not valid part of a connection */
1432                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1433                 ret = NF_ACCEPT;
1434                 goto out;
1435         }
1436
1437         /* Decide what timeout policy we want to apply to this flow. */
1438         timeouts = nf_ct_timeout_lookup(net, ct, l4proto);
1439
1440         ret = l4proto->packet(ct, skb, dataoff, ctinfo, timeouts);
1441         if (ret <= 0) {
1442                 /* Invalid: inverse of the return code tells
1443                  * the netfilter core what to do */
1444                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1445                 nf_conntrack_put(&ct->ct_general);
1446                 skb->_nfct = 0;
1447                 NF_CT_STAT_INC_ATOMIC(net, invalid);
1448                 if (ret == -NF_DROP)
1449                         NF_CT_STAT_INC_ATOMIC(net, drop);
1450                 /* Special case: TCP tracker reports an attempt to reopen a
1451                  * closed/aborted connection. We have to go back and create a
1452                  * fresh conntrack.
1453                  */
1454                 if (ret == -NF_REPEAT)
1455                         goto repeat;
1456                 ret = -ret;
1457                 goto out;
1458         }
1459
1460         if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
1461             !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1462                 nf_conntrack_event_cache(IPCT_REPLY, ct);
1463 out:
1464         if (tmpl)
1465                 nf_ct_put(tmpl);
1466
1467         return ret;
1468 }
1469 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1470
1471 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
1472                           const struct nf_conntrack_tuple *orig)
1473 {
1474         bool ret;
1475
1476         rcu_read_lock();
1477         ret = nf_ct_invert_tuple(inverse, orig,
1478                                  __nf_ct_l3proto_find(orig->src.l3num),
1479                                  __nf_ct_l4proto_find(orig->src.l3num,
1480                                                       orig->dst.protonum));
1481         rcu_read_unlock();
1482         return ret;
1483 }
1484 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
1485
1486 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
1487    implicitly racy: see __nf_conntrack_confirm */
1488 void nf_conntrack_alter_reply(struct nf_conn *ct,
1489                               const struct nf_conntrack_tuple *newreply)
1490 {
1491         struct nf_conn_help *help = nfct_help(ct);
1492
1493         /* Should be unconfirmed, so not in hash table yet */
1494         WARN_ON(nf_ct_is_confirmed(ct));
1495
1496         pr_debug("Altering reply tuple of %p to ", ct);
1497         nf_ct_dump_tuple(newreply);
1498
1499         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1500         if (ct->master || (help && !hlist_empty(&help->expectations)))
1501                 return;
1502
1503         rcu_read_lock();
1504         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1505         rcu_read_unlock();
1506 }
1507 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1508
1509 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1510 void __nf_ct_refresh_acct(struct nf_conn *ct,
1511                           enum ip_conntrack_info ctinfo,
1512                           const struct sk_buff *skb,
1513                           unsigned long extra_jiffies,
1514                           int do_acct)
1515 {
1516         WARN_ON(!skb);
1517
1518         /* Only update if this is not a fixed timeout */
1519         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1520                 goto acct;
1521
1522         /* If not in hash table, timer will not be active yet */
1523         if (nf_ct_is_confirmed(ct))
1524                 extra_jiffies += nfct_time_stamp;
1525
1526         ct->timeout = extra_jiffies;
1527 acct:
1528         if (do_acct)
1529                 nf_ct_acct_update(ct, ctinfo, skb->len);
1530 }
1531 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1532
1533 bool nf_ct_kill_acct(struct nf_conn *ct,
1534                      enum ip_conntrack_info ctinfo,
1535                      const struct sk_buff *skb)
1536 {
1537         nf_ct_acct_update(ct, ctinfo, skb->len);
1538
1539         return nf_ct_delete(ct, 0, 0);
1540 }
1541 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1542
1543 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1544
1545 #include <linux/netfilter/nfnetlink.h>
1546 #include <linux/netfilter/nfnetlink_conntrack.h>
1547 #include <linux/mutex.h>
1548
1549 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
1550  * in ip_conntrack_core, since we don't want the protocols to autoload
1551  * or depend on ctnetlink */
1552 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1553                                const struct nf_conntrack_tuple *tuple)
1554 {
1555         if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1556             nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1557                 goto nla_put_failure;
1558         return 0;
1559
1560 nla_put_failure:
1561         return -1;
1562 }
1563 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1564
1565 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1566         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
1567         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
1568 };
1569 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1570
1571 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1572                                struct nf_conntrack_tuple *t)
1573 {
1574         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
1575                 return -EINVAL;
1576
1577         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1578         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1579
1580         return 0;
1581 }
1582 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1583
1584 unsigned int nf_ct_port_nlattr_tuple_size(void)
1585 {
1586         static unsigned int size __read_mostly;
1587
1588         if (!size)
1589                 size = nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1590
1591         return size;
1592 }
1593 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1594 #endif
1595
1596 /* Used by ipt_REJECT and ip6t_REJECT. */
1597 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
1598 {
1599         struct nf_conn *ct;
1600         enum ip_conntrack_info ctinfo;
1601
1602         /* This ICMP is in reverse direction to the packet which caused it */
1603         ct = nf_ct_get(skb, &ctinfo);
1604         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1605                 ctinfo = IP_CT_RELATED_REPLY;
1606         else
1607                 ctinfo = IP_CT_RELATED;
1608
1609         /* Attach to new skbuff, and increment count */
1610         nf_ct_set(nskb, ct, ctinfo);
1611         nf_conntrack_get(skb_nfct(nskb));
1612 }
1613
1614 /* Bring out ya dead! */
1615 static struct nf_conn *
1616 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
1617                 void *data, unsigned int *bucket)
1618 {
1619         struct nf_conntrack_tuple_hash *h;
1620         struct nf_conn *ct;
1621         struct hlist_nulls_node *n;
1622         spinlock_t *lockp;
1623
1624         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1625                 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
1626                 local_bh_disable();
1627                 nf_conntrack_lock(lockp);
1628                 if (*bucket < nf_conntrack_htable_size) {
1629                         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
1630                                 if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1631                                         continue;
1632                                 ct = nf_ct_tuplehash_to_ctrack(h);
1633                                 if (iter(ct, data))
1634                                         goto found;
1635                         }
1636                 }
1637                 spin_unlock(lockp);
1638                 local_bh_enable();
1639                 cond_resched();
1640         }
1641
1642         return NULL;
1643 found:
1644         atomic_inc(&ct->ct_general.use);
1645         spin_unlock(lockp);
1646         local_bh_enable();
1647         return ct;
1648 }
1649
1650 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
1651                                   void *data, u32 portid, int report)
1652 {
1653         unsigned int bucket = 0, sequence;
1654         struct nf_conn *ct;
1655
1656         might_sleep();
1657
1658         for (;;) {
1659                 sequence = read_seqcount_begin(&nf_conntrack_generation);
1660
1661                 while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
1662                         /* Time to push up daises... */
1663
1664                         nf_ct_delete(ct, portid, report);
1665                         nf_ct_put(ct);
1666                         cond_resched();
1667                 }
1668
1669                 if (!read_seqcount_retry(&nf_conntrack_generation, sequence))
1670                         break;
1671                 bucket = 0;
1672         }
1673 }
1674
1675 struct iter_data {
1676         int (*iter)(struct nf_conn *i, void *data);
1677         void *data;
1678         struct net *net;
1679 };
1680
1681 static int iter_net_only(struct nf_conn *i, void *data)
1682 {
1683         struct iter_data *d = data;
1684
1685         if (!net_eq(d->net, nf_ct_net(i)))
1686                 return 0;
1687
1688         return d->iter(i, d->data);
1689 }
1690
1691 static void
1692 __nf_ct_unconfirmed_destroy(struct net *net)
1693 {
1694         int cpu;
1695
1696         for_each_possible_cpu(cpu) {
1697                 struct nf_conntrack_tuple_hash *h;
1698                 struct hlist_nulls_node *n;
1699                 struct ct_pcpu *pcpu;
1700
1701                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1702
1703                 spin_lock_bh(&pcpu->lock);
1704                 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
1705                         struct nf_conn *ct;
1706
1707                         ct = nf_ct_tuplehash_to_ctrack(h);
1708
1709                         /* we cannot call iter() on unconfirmed list, the
1710                          * owning cpu can reallocate ct->ext at any time.
1711                          */
1712                         set_bit(IPS_DYING_BIT, &ct->status);
1713                 }
1714                 spin_unlock_bh(&pcpu->lock);
1715                 cond_resched();
1716         }
1717 }
1718
1719 void nf_ct_unconfirmed_destroy(struct net *net)
1720 {
1721         might_sleep();
1722
1723         if (atomic_read(&net->ct.count) > 0) {
1724                 __nf_ct_unconfirmed_destroy(net);
1725                 nf_queue_nf_hook_drop(net);
1726                 synchronize_net();
1727         }
1728 }
1729 EXPORT_SYMBOL_GPL(nf_ct_unconfirmed_destroy);
1730
1731 void nf_ct_iterate_cleanup_net(struct net *net,
1732                                int (*iter)(struct nf_conn *i, void *data),
1733                                void *data, u32 portid, int report)
1734 {
1735         struct iter_data d;
1736
1737         might_sleep();
1738
1739         if (atomic_read(&net->ct.count) == 0)
1740                 return;
1741
1742         d.iter = iter;
1743         d.data = data;
1744         d.net = net;
1745
1746         nf_ct_iterate_cleanup(iter_net_only, &d, portid, report);
1747 }
1748 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
1749
1750 /**
1751  * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
1752  * @iter: callback to invoke for each conntrack
1753  * @data: data to pass to @iter
1754  *
1755  * Like nf_ct_iterate_cleanup, but first marks conntracks on the
1756  * unconfirmed list as dying (so they will not be inserted into
1757  * main table).
1758  *
1759  * Can only be called in module exit path.
1760  */
1761 void
1762 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
1763 {
1764         struct net *net;
1765
1766         rtnl_lock();
1767         for_each_net(net) {
1768                 if (atomic_read(&net->ct.count) == 0)
1769                         continue;
1770                 __nf_ct_unconfirmed_destroy(net);
1771                 nf_queue_nf_hook_drop(net);
1772         }
1773         rtnl_unlock();
1774
1775         /* Need to wait for netns cleanup worker to finish, if its
1776          * running -- it might have deleted a net namespace from
1777          * the global list, so our __nf_ct_unconfirmed_destroy() might
1778          * not have affected all namespaces.
1779          */
1780         net_ns_barrier();
1781
1782         /* a conntrack could have been unlinked from unconfirmed list
1783          * before we grabbed pcpu lock in __nf_ct_unconfirmed_destroy().
1784          * This makes sure its inserted into conntrack table.
1785          */
1786         synchronize_net();
1787
1788         nf_ct_iterate_cleanup(iter, data, 0, 0);
1789 }
1790 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
1791
1792 static int kill_all(struct nf_conn *i, void *data)
1793 {
1794         return net_eq(nf_ct_net(i), data);
1795 }
1796
1797 void nf_ct_free_hashtable(void *hash, unsigned int size)
1798 {
1799         if (is_vmalloc_addr(hash))
1800                 vfree(hash);
1801         else
1802                 free_pages((unsigned long)hash,
1803                            get_order(sizeof(struct hlist_head) * size));
1804 }
1805 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1806
1807 void nf_conntrack_cleanup_start(void)
1808 {
1809         conntrack_gc_work.exiting = true;
1810         RCU_INIT_POINTER(ip_ct_attach, NULL);
1811 }
1812
1813 void nf_conntrack_cleanup_end(void)
1814 {
1815         RCU_INIT_POINTER(nf_ct_destroy, NULL);
1816
1817         cancel_delayed_work_sync(&conntrack_gc_work.dwork);
1818         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1819
1820         nf_conntrack_proto_fini();
1821         nf_conntrack_seqadj_fini();
1822         nf_conntrack_labels_fini();
1823         nf_conntrack_helper_fini();
1824         nf_conntrack_timeout_fini();
1825         nf_conntrack_ecache_fini();
1826         nf_conntrack_tstamp_fini();
1827         nf_conntrack_acct_fini();
1828         nf_conntrack_expect_fini();
1829
1830         kmem_cache_destroy(nf_conntrack_cachep);
1831 }
1832
1833 /*
1834  * Mishearing the voices in his head, our hero wonders how he's
1835  * supposed to kill the mall.
1836  */
1837 void nf_conntrack_cleanup_net(struct net *net)
1838 {
1839         LIST_HEAD(single);
1840
1841         list_add(&net->exit_list, &single);
1842         nf_conntrack_cleanup_net_list(&single);
1843 }
1844
1845 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
1846 {
1847         int busy;
1848         struct net *net;
1849
1850         /*
1851          * This makes sure all current packets have passed through
1852          *  netfilter framework.  Roll on, two-stage module
1853          *  delete...
1854          */
1855         synchronize_net();
1856 i_see_dead_people:
1857         busy = 0;
1858         list_for_each_entry(net, net_exit_list, exit_list) {
1859                 nf_ct_iterate_cleanup(kill_all, net, 0, 0);
1860                 if (atomic_read(&net->ct.count) != 0)
1861                         busy = 1;
1862         }
1863         if (busy) {
1864                 schedule();
1865                 goto i_see_dead_people;
1866         }
1867
1868         list_for_each_entry(net, net_exit_list, exit_list) {
1869                 nf_conntrack_proto_pernet_fini(net);
1870                 nf_conntrack_helper_pernet_fini(net);
1871                 nf_conntrack_ecache_pernet_fini(net);
1872                 nf_conntrack_tstamp_pernet_fini(net);
1873                 nf_conntrack_acct_pernet_fini(net);
1874                 nf_conntrack_expect_pernet_fini(net);
1875                 free_percpu(net->ct.stat);
1876                 free_percpu(net->ct.pcpu_lists);
1877         }
1878 }
1879
1880 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
1881 {
1882         struct hlist_nulls_head *hash;
1883         unsigned int nr_slots, i;
1884         size_t sz;
1885
1886         if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1887                 return NULL;
1888
1889         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1890         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1891
1892         if (nr_slots > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1893                 return NULL;
1894
1895         sz = nr_slots * sizeof(struct hlist_nulls_head);
1896         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1897                                         get_order(sz));
1898         if (!hash)
1899                 hash = vzalloc(sz);
1900
1901         if (hash && nulls)
1902                 for (i = 0; i < nr_slots; i++)
1903                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1904
1905         return hash;
1906 }
1907 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1908
1909 int nf_conntrack_hash_resize(unsigned int hashsize)
1910 {
1911         int i, bucket;
1912         unsigned int old_size;
1913         struct hlist_nulls_head *hash, *old_hash;
1914         struct nf_conntrack_tuple_hash *h;
1915         struct nf_conn *ct;
1916
1917         if (!hashsize)
1918                 return -EINVAL;
1919
1920         hash = nf_ct_alloc_hashtable(&hashsize, 1);
1921         if (!hash)
1922                 return -ENOMEM;
1923
1924         old_size = nf_conntrack_htable_size;
1925         if (old_size == hashsize) {
1926                 nf_ct_free_hashtable(hash, hashsize);
1927                 return 0;
1928         }
1929
1930         local_bh_disable();
1931         nf_conntrack_all_lock();
1932         write_seqcount_begin(&nf_conntrack_generation);
1933
1934         /* Lookups in the old hash might happen in parallel, which means we
1935          * might get false negatives during connection lookup. New connections
1936          * created because of a false negative won't make it into the hash
1937          * though since that required taking the locks.
1938          */
1939
1940         for (i = 0; i < nf_conntrack_htable_size; i++) {
1941                 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
1942                         h = hlist_nulls_entry(nf_conntrack_hash[i].first,
1943                                               struct nf_conntrack_tuple_hash, hnnode);
1944                         ct = nf_ct_tuplehash_to_ctrack(h);
1945                         hlist_nulls_del_rcu(&h->hnnode);
1946                         bucket = __hash_conntrack(nf_ct_net(ct),
1947                                                   &h->tuple, hashsize);
1948                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1949                 }
1950         }
1951         old_size = nf_conntrack_htable_size;
1952         old_hash = nf_conntrack_hash;
1953
1954         nf_conntrack_hash = hash;
1955         nf_conntrack_htable_size = hashsize;
1956
1957         write_seqcount_end(&nf_conntrack_generation);
1958         nf_conntrack_all_unlock();
1959         local_bh_enable();
1960
1961         synchronize_net();
1962         nf_ct_free_hashtable(old_hash, old_size);
1963         return 0;
1964 }
1965
1966 int nf_conntrack_set_hashsize(const char *val, const struct kernel_param *kp)
1967 {
1968         unsigned int hashsize;
1969         int rc;
1970
1971         if (current->nsproxy->net_ns != &init_net)
1972                 return -EOPNOTSUPP;
1973
1974         /* On boot, we can set this without any fancy locking. */
1975         if (!nf_conntrack_htable_size)
1976                 return param_set_uint(val, kp);
1977
1978         rc = kstrtouint(val, 0, &hashsize);
1979         if (rc)
1980                 return rc;
1981
1982         return nf_conntrack_hash_resize(hashsize);
1983 }
1984 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1985
1986 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1987                   &nf_conntrack_htable_size, 0600);
1988
1989 static __always_inline unsigned int total_extension_size(void)
1990 {
1991         /* remember to add new extensions below */
1992         BUILD_BUG_ON(NF_CT_EXT_NUM > 9);
1993
1994         return sizeof(struct nf_ct_ext) +
1995                sizeof(struct nf_conn_help)
1996 #if IS_ENABLED(CONFIG_NF_NAT)
1997                 + sizeof(struct nf_conn_nat)
1998 #endif
1999                 + sizeof(struct nf_conn_seqadj)
2000                 + sizeof(struct nf_conn_acct)
2001 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2002                 + sizeof(struct nf_conntrack_ecache)
2003 #endif
2004 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
2005                 + sizeof(struct nf_conn_tstamp)
2006 #endif
2007 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
2008                 + sizeof(struct nf_conn_timeout)
2009 #endif
2010 #ifdef CONFIG_NF_CONNTRACK_LABELS
2011                 + sizeof(struct nf_conn_labels)
2012 #endif
2013 #if IS_ENABLED(CONFIG_NETFILTER_SYNPROXY)
2014                 + sizeof(struct nf_conn_synproxy)
2015 #endif
2016         ;
2017 };
2018
2019 int nf_conntrack_init_start(void)
2020 {
2021         int max_factor = 8;
2022         int ret = -ENOMEM;
2023         int i;
2024
2025         /* struct nf_ct_ext uses u8 to store offsets/size */
2026         BUILD_BUG_ON(total_extension_size() > 255u);
2027
2028         seqcount_init(&nf_conntrack_generation);
2029
2030         for (i = 0; i < CONNTRACK_LOCKS; i++)
2031                 spin_lock_init(&nf_conntrack_locks[i]);
2032
2033         if (!nf_conntrack_htable_size) {
2034                 /* Idea from tcp.c: use 1/16384 of memory.
2035                  * On i386: 32MB machine has 512 buckets.
2036                  * >= 1GB machines have 16384 buckets.
2037                  * >= 4GB machines have 65536 buckets.
2038                  */
2039                 nf_conntrack_htable_size
2040                         = (((totalram_pages << PAGE_SHIFT) / 16384)
2041                            / sizeof(struct hlist_head));
2042                 if (totalram_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2043                         nf_conntrack_htable_size = 65536;
2044                 else if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2045                         nf_conntrack_htable_size = 16384;
2046                 if (nf_conntrack_htable_size < 32)
2047                         nf_conntrack_htable_size = 32;
2048
2049                 /* Use a max. factor of four by default to get the same max as
2050                  * with the old struct list_heads. When a table size is given
2051                  * we use the old value of 8 to avoid reducing the max.
2052                  * entries. */
2053                 max_factor = 4;
2054         }
2055
2056         nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2057         if (!nf_conntrack_hash)
2058                 return -ENOMEM;
2059
2060         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2061
2062         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2063                                                 sizeof(struct nf_conn),
2064                                                 NFCT_INFOMASK + 1,
2065                                                 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2066         if (!nf_conntrack_cachep)
2067                 goto err_cachep;
2068
2069         ret = nf_conntrack_expect_init();
2070         if (ret < 0)
2071                 goto err_expect;
2072
2073         ret = nf_conntrack_acct_init();
2074         if (ret < 0)
2075                 goto err_acct;
2076
2077         ret = nf_conntrack_tstamp_init();
2078         if (ret < 0)
2079                 goto err_tstamp;
2080
2081         ret = nf_conntrack_ecache_init();
2082         if (ret < 0)
2083                 goto err_ecache;
2084
2085         ret = nf_conntrack_timeout_init();
2086         if (ret < 0)
2087                 goto err_timeout;
2088
2089         ret = nf_conntrack_helper_init();
2090         if (ret < 0)
2091                 goto err_helper;
2092
2093         ret = nf_conntrack_labels_init();
2094         if (ret < 0)
2095                 goto err_labels;
2096
2097         ret = nf_conntrack_seqadj_init();
2098         if (ret < 0)
2099                 goto err_seqadj;
2100
2101         ret = nf_conntrack_proto_init();
2102         if (ret < 0)
2103                 goto err_proto;
2104
2105         conntrack_gc_work_init(&conntrack_gc_work);
2106         queue_delayed_work(system_power_efficient_wq, &conntrack_gc_work.dwork, HZ);
2107
2108         return 0;
2109
2110 err_proto:
2111         nf_conntrack_seqadj_fini();
2112 err_seqadj:
2113         nf_conntrack_labels_fini();
2114 err_labels:
2115         nf_conntrack_helper_fini();
2116 err_helper:
2117         nf_conntrack_timeout_fini();
2118 err_timeout:
2119         nf_conntrack_ecache_fini();
2120 err_ecache:
2121         nf_conntrack_tstamp_fini();
2122 err_tstamp:
2123         nf_conntrack_acct_fini();
2124 err_acct:
2125         nf_conntrack_expect_fini();
2126 err_expect:
2127         kmem_cache_destroy(nf_conntrack_cachep);
2128 err_cachep:
2129         nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
2130         return ret;
2131 }
2132
2133 void nf_conntrack_init_end(void)
2134 {
2135         /* For use by REJECT target */
2136         RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
2137         RCU_INIT_POINTER(nf_ct_destroy, destroy_conntrack);
2138 }
2139
2140 /*
2141  * We need to use special "null" values, not used in hash table
2142  */
2143 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
2144 #define DYING_NULLS_VAL         ((1<<30)+1)
2145 #define TEMPLATE_NULLS_VAL      ((1<<30)+2)
2146
2147 int nf_conntrack_init_net(struct net *net)
2148 {
2149         int ret = -ENOMEM;
2150         int cpu;
2151
2152         BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2153         atomic_set(&net->ct.count, 0);
2154
2155         net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
2156         if (!net->ct.pcpu_lists)
2157                 goto err_stat;
2158
2159         for_each_possible_cpu(cpu) {
2160                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2161
2162                 spin_lock_init(&pcpu->lock);
2163                 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
2164                 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
2165         }
2166
2167         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2168         if (!net->ct.stat)
2169                 goto err_pcpu_lists;
2170
2171         ret = nf_conntrack_expect_pernet_init(net);
2172         if (ret < 0)
2173                 goto err_expect;
2174         ret = nf_conntrack_acct_pernet_init(net);
2175         if (ret < 0)
2176                 goto err_acct;
2177         ret = nf_conntrack_tstamp_pernet_init(net);
2178         if (ret < 0)
2179                 goto err_tstamp;
2180         ret = nf_conntrack_ecache_pernet_init(net);
2181         if (ret < 0)
2182                 goto err_ecache;
2183         ret = nf_conntrack_helper_pernet_init(net);
2184         if (ret < 0)
2185                 goto err_helper;
2186         ret = nf_conntrack_proto_pernet_init(net);
2187         if (ret < 0)
2188                 goto err_proto;
2189         return 0;
2190
2191 err_proto:
2192         nf_conntrack_helper_pernet_fini(net);
2193 err_helper:
2194         nf_conntrack_ecache_pernet_fini(net);
2195 err_ecache:
2196         nf_conntrack_tstamp_pernet_fini(net);
2197 err_tstamp:
2198         nf_conntrack_acct_pernet_fini(net);
2199 err_acct:
2200         nf_conntrack_expect_pernet_fini(net);
2201 err_expect:
2202         free_percpu(net->ct.stat);
2203 err_pcpu_lists:
2204         free_percpu(net->ct.pcpu_lists);
2205 err_stat:
2206         return ret;
2207 }