net/netfilter/nf_conntrack_core.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Connection state tracking for netfilter.  This is separated from,
   3    but required by, the NAT layer; it can also be used by an iptables
   4    extension. */
   5
   6 /* (C) 1999-2001 Paul `Rusty' Russell
   7  * (C) 2002-2006 Netfilter Core Team <coreteam@netfilter.org>
   8  * (C) 2003,2004 USAGI/WIDE Project <http://www.linux-ipv6.org>
   9  * (C) 2005-2012 Patrick McHardy <kaber@trash.net>
  10  */
  11
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include <linux/types.h>
  15 #include <linux/netfilter.h>
  16 #include <linux/module.h>
  17 #include <linux/sched.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/proc_fs.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/stddef.h>
  22 #include <linux/slab.h>
  23 #include <linux/random.h>
  24 #include <linux/siphash.h>
  25 #include <linux/err.h>
  26 #include <linux/percpu.h>
  27 #include <linux/moduleparam.h>
  28 #include <linux/notifier.h>
  29 #include <linux/kernel.h>
  30 #include <linux/netdevice.h>
  31 #include <linux/socket.h>
  32 #include <linux/mm.h>
  33 #include <linux/nsproxy.h>
  34 #include <linux/rculist_nulls.h>
  35
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_bpf.h>
  38 #include <net/netfilter/nf_conntrack_l4proto.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_core.h>
  42 #include <net/netfilter/nf_conntrack_extend.h>
  43 #include <net/netfilter/nf_conntrack_acct.h>
  44 #include <net/netfilter/nf_conntrack_ecache.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_timeout.h>
  48 #include <net/netfilter/nf_conntrack_labels.h>
  49 #include <net/netfilter/nf_conntrack_synproxy.h>
  50 #include <net/netfilter/nf_nat.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #include <net/netns/hash.h>
  53 #include <net/ip.h>
  54
  55 #include "nf_internals.h"
  56
  57 __cacheline_aligned_in_smp spinlock_t nf_conntrack_locks[CONNTRACK_LOCKS];
  58 EXPORT_SYMBOL_GPL(nf_conntrack_locks);
  59
  60 __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
  61 EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
  62
  63 struct hlist_nulls_head *nf_conntrack_hash __read_mostly;
  64 EXPORT_SYMBOL_GPL(nf_conntrack_hash);
  65
  66 struct conntrack_gc_work {
  67         struct delayed_work     dwork;
  68         u32                     next_bucket;
  69         u32                     avg_timeout;
  70         u32                     count;
  71         u32                     start_time;
  72         bool                    exiting;
  73         bool                    early_drop;
  74 };
  75
  76 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
  77 static DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
  78 static __read_mostly bool nf_conntrack_locks_all;
  79
  80 /* serialize hash resizes and nf_ct_iterate_cleanup */
  81 static DEFINE_MUTEX(nf_conntrack_mutex);
  82
  83 #define GC_SCAN_INTERVAL_MAX    (60ul * HZ)
  84 #define GC_SCAN_INTERVAL_MIN    (1ul * HZ)
  85
  86 /* clamp timeouts to this value (TCP unacked) */
  87 #define GC_SCAN_INTERVAL_CLAMP  (300ul * HZ)
  88
  89 /* Initial bias pretending we have 100 entries at the upper bound so we don't
  90  * wakeup often just because we have three entries with a 1s timeout while still
  91  * allowing non-idle machines to wakeup more often when needed.
  92  */
  93 #define GC_SCAN_INITIAL_COUNT   100
  94 #define GC_SCAN_INTERVAL_INIT   GC_SCAN_INTERVAL_MAX
  95
  96 #define GC_SCAN_MAX_DURATION    msecs_to_jiffies(10)
  97 #define GC_SCAN_EXPIRED_MAX     (64000u / HZ)
  98
  99 #define MIN_CHAINLEN    50u
 100 #define MAX_CHAINLEN    (80u - MIN_CHAINLEN)
 101
 102 static struct conntrack_gc_work conntrack_gc_work;
 103
 104 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
 105 {
 106         /* 1) Acquire the lock */
 107         spin_lock(lock);
 108
 109         /* 2) read nf_conntrack_locks_all, with ACQUIRE semantics
 110          * It pairs with the smp_store_release() in nf_conntrack_all_unlock()
 111          */
 112         if (likely(smp_load_acquire(&nf_conntrack_locks_all) == false))
 113                 return;
 114
 115         /* fast path failed, unlock */
 116         spin_unlock(lock);
 117
 118         /* Slow path 1) get global lock */
 119         spin_lock(&nf_conntrack_locks_all_lock);
 120
 121         /* Slow path 2) get the lock we want */
 122         spin_lock(lock);
 123
 124         /* Slow path 3) release the global lock */
 125         spin_unlock(&nf_conntrack_locks_all_lock);
 126 }
 127 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
 128
 129 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 130 {
 131         h1 %= CONNTRACK_LOCKS;
 132         h2 %= CONNTRACK_LOCKS;
 133         spin_unlock(&nf_conntrack_locks[h1]);
 134         if (h1 != h2)
 135                 spin_unlock(&nf_conntrack_locks[h2]);
 136 }
 137
 138 /* return true if we need to recompute hashes (in case hash table was resized) */
 139 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
 140                                      unsigned int h2, unsigned int sequence)
 141 {
 142         h1 %= CONNTRACK_LOCKS;
 143         h2 %= CONNTRACK_LOCKS;
 144         if (h1 <= h2) {
 145                 nf_conntrack_lock(&nf_conntrack_locks[h1]);
 146                 if (h1 != h2)
 147                         spin_lock_nested(&nf_conntrack_locks[h2],
 148                                          SINGLE_DEPTH_NESTING);
 149         } else {
 150                 nf_conntrack_lock(&nf_conntrack_locks[h2]);
 151                 spin_lock_nested(&nf_conntrack_locks[h1],
 152                                  SINGLE_DEPTH_NESTING);
 153         }
 154         if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
 155                 nf_conntrack_double_unlock(h1, h2);
 156                 return true;
 157         }
 158         return false;
 159 }
 160
 161 static void nf_conntrack_all_lock(void)
 162         __acquires(&nf_conntrack_locks_all_lock)
 163 {
 164         int i;
 165
 166         spin_lock(&nf_conntrack_locks_all_lock);
 167
 168         /* For nf_contrack_locks_all, only the latest time when another
 169          * CPU will see an update is controlled, by the "release" of the
 170          * spin_lock below.
 171          * The earliest time is not controlled, an thus KCSAN could detect
 172          * a race when nf_conntract_lock() reads the variable.
 173          * WRITE_ONCE() is used to ensure the compiler will not
 174          * optimize the write.
 175          */
 176         WRITE_ONCE(nf_conntrack_locks_all, true);
 177
 178         for (i = 0; i < CONNTRACK_LOCKS; i++) {
 179                 spin_lock(&nf_conntrack_locks[i]);
 180
 181                 /* This spin_unlock provides the "release" to ensure that
 182                  * nf_conntrack_locks_all==true is visible to everyone that
 183                  * acquired spin_lock(&nf_conntrack_locks[]).
 184                  */
 185                 spin_unlock(&nf_conntrack_locks[i]);
 186         }
 187 }
 188
 189 static void nf_conntrack_all_unlock(void)
 190         __releases(&nf_conntrack_locks_all_lock)
 191 {
 192         /* All prior stores must be complete before we clear
 193          * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
 194          * might observe the false value but not the entire
 195          * critical section.
 196          * It pairs with the smp_load_acquire() in nf_conntrack_lock()
 197          */
 198         smp_store_release(&nf_conntrack_locks_all, false);
 199         spin_unlock(&nf_conntrack_locks_all_lock);
 200 }
 201
 202 unsigned int nf_conntrack_htable_size __read_mostly;
 203 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
 204
 205 unsigned int nf_conntrack_max __read_mostly;
 206 EXPORT_SYMBOL_GPL(nf_conntrack_max);
 207 seqcount_spinlock_t nf_conntrack_generation __read_mostly;
 208 static siphash_aligned_key_t nf_conntrack_hash_rnd;
 209
 210 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
 211                               unsigned int zoneid,
 212                               const struct net *net)
 213 {
 214         siphash_key_t key;
 215
 216         get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
 217
 218         key = nf_conntrack_hash_rnd;
 219
 220         key.key[0] ^= zoneid;
 221         key.key[1] ^= net_hash_mix(net);
 222
 223         return siphash((void *)tuple,
 224                         offsetofend(struct nf_conntrack_tuple, dst.__nfct_hash_offsetend),
 225                         &key);
 226 }
 227
 228 static u32 scale_hash(u32 hash)
 229 {
 230         return reciprocal_scale(hash, nf_conntrack_htable_size);
 231 }
 232
 233 static u32 __hash_conntrack(const struct net *net,
 234                             const struct nf_conntrack_tuple *tuple,
 235                             unsigned int zoneid,
 236                             unsigned int size)
 237 {
 238         return reciprocal_scale(hash_conntrack_raw(tuple, zoneid, net), size);
 239 }
 240
 241 static u32 hash_conntrack(const struct net *net,
 242                           const struct nf_conntrack_tuple *tuple,
 243                           unsigned int zoneid)
 244 {
 245         return scale_hash(hash_conntrack_raw(tuple, zoneid, net));
 246 }
 247
 248 static bool nf_ct_get_tuple_ports(const struct sk_buff *skb,
 249                                   unsigned int dataoff,
 250                                   struct nf_conntrack_tuple *tuple)
 251 {       struct {
 252                 __be16 sport;
 253                 __be16 dport;
 254         } _inet_hdr, *inet_hdr;
 255
 256         /* Actually only need first 4 bytes to get ports. */
 257         inet_hdr = skb_header_pointer(skb, dataoff, sizeof(_inet_hdr), &_inet_hdr);
 258         if (!inet_hdr)
 259                 return false;
 260
 261         tuple->src.u.udp.port = inet_hdr->sport;
 262         tuple->dst.u.udp.port = inet_hdr->dport;
 263         return true;
 264 }
 265
 266 static bool
 267 nf_ct_get_tuple(const struct sk_buff *skb,
 268                 unsigned int nhoff,
 269                 unsigned int dataoff,
 270                 u_int16_t l3num,
 271                 u_int8_t protonum,
 272                 struct net *net,
 273                 struct nf_conntrack_tuple *tuple)
 274 {
 275         unsigned int size;
 276         const __be32 *ap;
 277         __be32 _addrs[8];
 278
 279         memset(tuple, 0, sizeof(*tuple));
 280
 281         tuple->src.l3num = l3num;
 282         switch (l3num) {
 283         case NFPROTO_IPV4:
 284                 nhoff += offsetof(struct iphdr, saddr);
 285                 size = 2 * sizeof(__be32);
 286                 break;
 287         case NFPROTO_IPV6:
 288                 nhoff += offsetof(struct ipv6hdr, saddr);
 289                 size = sizeof(_addrs);
 290                 break;
 291         default:
 292                 return true;
 293         }
 294
 295         ap = skb_header_pointer(skb, nhoff, size, _addrs);
 296         if (!ap)
 297                 return false;
 298
 299         switch (l3num) {
 300         case NFPROTO_IPV4:
 301                 tuple->src.u3.ip = ap[0];
 302                 tuple->dst.u3.ip = ap[1];
 303                 break;
 304         case NFPROTO_IPV6:
 305                 memcpy(tuple->src.u3.ip6, ap, sizeof(tuple->src.u3.ip6));
 306                 memcpy(tuple->dst.u3.ip6, ap + 4, sizeof(tuple->dst.u3.ip6));
 307                 break;
 308         }
 309
 310         tuple->dst.protonum = protonum;
 311         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 312
 313         switch (protonum) {
 314 #if IS_ENABLED(CONFIG_IPV6)
 315         case IPPROTO_ICMPV6:
 316                 return icmpv6_pkt_to_tuple(skb, dataoff, net, tuple);
 317 #endif
 318         case IPPROTO_ICMP:
 319                 return icmp_pkt_to_tuple(skb, dataoff, net, tuple);
 320 #ifdef CONFIG_NF_CT_PROTO_GRE
 321         case IPPROTO_GRE:
 322                 return gre_pkt_to_tuple(skb, dataoff, net, tuple);
 323 #endif
 324         case IPPROTO_TCP:
 325         case IPPROTO_UDP:
 326 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
 327         case IPPROTO_UDPLITE:
 328 #endif
 329 #ifdef CONFIG_NF_CT_PROTO_SCTP
 330         case IPPROTO_SCTP:
 331 #endif
 332 #ifdef CONFIG_NF_CT_PROTO_DCCP
 333         case IPPROTO_DCCP:
 334 #endif
 335                 /* fallthrough */
 336                 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
 337         default:
 338                 break;
 339         }
 340
 341         return true;
 342 }
 343
 344 static int ipv4_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
 345                             u_int8_t *protonum)
 346 {
 347         int dataoff = -1;
 348         const struct iphdr *iph;
 349         struct iphdr _iph;
 350
 351         iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
 352         if (!iph)
 353                 return -1;
 354
 355         /* Conntrack defragments packets, we might still see fragments
 356          * inside ICMP packets though.
 357          */
 358         if (iph->frag_off & htons(IP_OFFSET))
 359                 return -1;
 360
 361         dataoff = nhoff + (iph->ihl << 2);
 362         *protonum = iph->protocol;
 363
 364         /* Check bogus IP headers */
 365         if (dataoff > skb->len) {
 366                 pr_debug("bogus IPv4 packet: nhoff %u, ihl %u, skblen %u\n",
 367                          nhoff, iph->ihl << 2, skb->len);
 368                 return -1;
 369         }
 370         return dataoff;
 371 }
 372
 373 #if IS_ENABLED(CONFIG_IPV6)
 374 static int ipv6_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
 375                             u8 *protonum)
 376 {
 377         int protoff = -1;
 378         unsigned int extoff = nhoff + sizeof(struct ipv6hdr);
 379         __be16 frag_off;
 380         u8 nexthdr;
 381
 382         if (skb_copy_bits(skb, nhoff + offsetof(struct ipv6hdr, nexthdr),
 383                           &nexthdr, sizeof(nexthdr)) != 0) {
 384                 pr_debug("can't get nexthdr\n");
 385                 return -1;
 386         }
 387         protoff = ipv6_skip_exthdr(skb, extoff, &nexthdr, &frag_off);
 388         /*
 389          * (protoff == skb->len) means the packet has not data, just
 390          * IPv6 and possibly extensions headers, but it is tracked anyway
 391          */
 392         if (protoff < 0 || (frag_off & htons(~0x7)) != 0) {
 393                 pr_debug("can't find proto in pkt\n");
 394                 return -1;
 395         }
 396
 397         *protonum = nexthdr;
 398         return protoff;
 399 }
 400 #endif
 401
 402 static int get_l4proto(const struct sk_buff *skb,
 403                        unsigned int nhoff, u8 pf, u8 *l4num)
 404 {
 405         switch (pf) {
 406         case NFPROTO_IPV4:
 407                 return ipv4_get_l4proto(skb, nhoff, l4num);
 408 #if IS_ENABLED(CONFIG_IPV6)
 409         case NFPROTO_IPV6:
 410                 return ipv6_get_l4proto(skb, nhoff, l4num);
 411 #endif
 412         default:
 413                 *l4num = 0;
 414                 break;
 415         }
 416         return -1;
 417 }
 418
 419 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 420                        u_int16_t l3num,
 421                        struct net *net, struct nf_conntrack_tuple *tuple)
 422 {
 423         u8 protonum;
 424         int protoff;
 425
 426         protoff = get_l4proto(skb, nhoff, l3num, &protonum);
 427         if (protoff <= 0)
 428                 return false;
 429
 430         return nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple);
 431 }
 432 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 433
 434 bool
 435 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 436                    const struct nf_conntrack_tuple *orig)
 437 {
 438         memset(inverse, 0, sizeof(*inverse));
 439
 440         inverse->src.l3num = orig->src.l3num;
 441
 442         switch (orig->src.l3num) {
 443         case NFPROTO_IPV4:
 444                 inverse->src.u3.ip = orig->dst.u3.ip;
 445                 inverse->dst.u3.ip = orig->src.u3.ip;
 446                 break;
 447         case NFPROTO_IPV6:
 448                 inverse->src.u3.in6 = orig->dst.u3.in6;
 449                 inverse->dst.u3.in6 = orig->src.u3.in6;
 450                 break;
 451         default:
 452                 break;
 453         }
 454
 455         inverse->dst.dir = !orig->dst.dir;
 456
 457         inverse->dst.protonum = orig->dst.protonum;
 458
 459         switch (orig->dst.protonum) {
 460         case IPPROTO_ICMP:
 461                 return nf_conntrack_invert_icmp_tuple(inverse, orig);
 462 #if IS_ENABLED(CONFIG_IPV6)
 463         case IPPROTO_ICMPV6:
 464                 return nf_conntrack_invert_icmpv6_tuple(inverse, orig);
 465 #endif
 466         }
 467
 468         inverse->src.u.all = orig->dst.u.all;
 469         inverse->dst.u.all = orig->src.u.all;
 470         return true;
 471 }
 472 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 473
 474 /* Generate a almost-unique pseudo-id for a given conntrack.
 475  *
 476  * intentionally doesn't re-use any of the seeds used for hash
 477  * table location, we assume id gets exposed to userspace.
 478  *
 479  * Following nf_conn items do not change throughout lifetime
 480  * of the nf_conn:
 481  *
 482  * 1. nf_conn address
 483  * 2. nf_conn->master address (normally NULL)
 484  * 3. the associated net namespace
 485  * 4. the original direction tuple
 486  */
 487 u32 nf_ct_get_id(const struct nf_conn *ct)
 488 {
 489         static siphash_aligned_key_t ct_id_seed;
 490         unsigned long a, b, c, d;
 491
 492         net_get_random_once(&ct_id_seed, sizeof(ct_id_seed));
 493
 494         a = (unsigned long)ct;
 495         b = (unsigned long)ct->master;
 496         c = (unsigned long)nf_ct_net(ct);
 497         d = (unsigned long)siphash(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 498                                    sizeof(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple),
 499                                    &ct_id_seed);
 500 #ifdef CONFIG_64BIT
 501         return siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &ct_id_seed);
 502 #else
 503         return siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &ct_id_seed);
 504 #endif
 505 }
 506 EXPORT_SYMBOL_GPL(nf_ct_get_id);
 507
 508 static void
 509 clean_from_lists(struct nf_conn *ct)
 510 {
 511         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 512         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 513
 514         /* Destroy all pending expectations */
 515         nf_ct_remove_expectations(ct);
 516 }
 517
 518 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
 519
 520 /* Released via nf_ct_destroy() */
 521 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
 522                                  const struct nf_conntrack_zone *zone,
 523                                  gfp_t flags)
 524 {
 525         struct nf_conn *tmpl, *p;
 526
 527         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
 528                 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
 529                 if (!tmpl)
 530                         return NULL;
 531
 532                 p = tmpl;
 533                 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 534                 if (tmpl != p) {
 535                         tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 536                         tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
 537                 }
 538         } else {
 539                 tmpl = kzalloc(sizeof(*tmpl), flags);
 540                 if (!tmpl)
 541                         return NULL;
 542         }
 543
 544         tmpl->status = IPS_TEMPLATE;
 545         write_pnet(&tmpl->ct_net, net);
 546         nf_ct_zone_add(tmpl, zone);
 547         refcount_set(&tmpl->ct_general.use, 1);
 548
 549         return tmpl;
 550 }
 551 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
 552
 553 void nf_ct_tmpl_free(struct nf_conn *tmpl)
 554 {
 555         kfree(tmpl->ext);
 556
 557         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
 558                 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
 559         else
 560                 kfree(tmpl);
 561 }
 562 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
 563
 564 static void destroy_gre_conntrack(struct nf_conn *ct)
 565 {
 566 #ifdef CONFIG_NF_CT_PROTO_GRE
 567         struct nf_conn *master = ct->master;
 568
 569         if (master)
 570                 nf_ct_gre_keymap_destroy(master);
 571 #endif
 572 }
 573
 574 void nf_ct_destroy(struct nf_conntrack *nfct)
 575 {
 576         struct nf_conn *ct = (struct nf_conn *)nfct;
 577
 578         WARN_ON(refcount_read(&nfct->use) != 0);
 579
 580         if (unlikely(nf_ct_is_template(ct))) {
 581                 nf_ct_tmpl_free(ct);
 582                 return;
 583         }
 584
 585         if (unlikely(nf_ct_protonum(ct) == IPPROTO_GRE))
 586                 destroy_gre_conntrack(ct);
 587
 588         /* Expectations will have been removed in clean_from_lists,
 589          * except TFTP can create an expectation on the first packet,
 590          * before connection is in the list, so we need to clean here,
 591          * too.
 592          */
 593         nf_ct_remove_expectations(ct);
 594
 595         if (ct->master)
 596                 nf_ct_put(ct->master);
 597
 598         nf_conntrack_free(ct);
 599 }
 600 EXPORT_SYMBOL(nf_ct_destroy);
 601
 602 static void __nf_ct_delete_from_lists(struct nf_conn *ct)
 603 {
 604         struct net *net = nf_ct_net(ct);
 605         unsigned int hash, reply_hash;
 606         unsigned int sequence;
 607
 608         do {
 609                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 610                 hash = hash_conntrack(net,
 611                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 612                                       nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_ORIGINAL));
 613                 reply_hash = hash_conntrack(net,
 614                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 615                                            nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
 616         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 617
 618         clean_from_lists(ct);
 619         nf_conntrack_double_unlock(hash, reply_hash);
 620 }
 621
 622 static void nf_ct_delete_from_lists(struct nf_conn *ct)
 623 {
 624         nf_ct_helper_destroy(ct);
 625         local_bh_disable();
 626
 627         __nf_ct_delete_from_lists(ct);
 628
 629         local_bh_enable();
 630 }
 631
 632 static void nf_ct_add_to_ecache_list(struct nf_conn *ct)
 633 {
 634 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 635         struct nf_conntrack_net *cnet = nf_ct_pernet(nf_ct_net(ct));
 636
 637         spin_lock(&cnet->ecache.dying_lock);
 638         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 639                                  &cnet->ecache.dying_list);
 640         spin_unlock(&cnet->ecache.dying_lock);
 641 #endif
 642 }
 643
 644 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
 645 {
 646         struct nf_conn_tstamp *tstamp;
 647         struct net *net;
 648
 649         if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
 650                 return false;
 651
 652         tstamp = nf_conn_tstamp_find(ct);
 653         if (tstamp) {
 654                 s32 timeout = READ_ONCE(ct->timeout) - nfct_time_stamp;
 655
 656                 tstamp->stop = ktime_get_real_ns();
 657                 if (timeout < 0)
 658                         tstamp->stop -= jiffies_to_nsecs(-timeout);
 659         }
 660
 661         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 662                                     portid, report) < 0) {
 663                 /* destroy event was not delivered. nf_ct_put will
 664                  * be done by event cache worker on redelivery.
 665                  */
 666                 nf_ct_helper_destroy(ct);
 667                 local_bh_disable();
 668                 __nf_ct_delete_from_lists(ct);
 669                 nf_ct_add_to_ecache_list(ct);
 670                 local_bh_enable();
 671
 672                 nf_conntrack_ecache_work(nf_ct_net(ct), NFCT_ECACHE_DESTROY_FAIL);
 673                 return false;
 674         }
 675
 676         net = nf_ct_net(ct);
 677         if (nf_conntrack_ecache_dwork_pending(net))
 678                 nf_conntrack_ecache_work(net, NFCT_ECACHE_DESTROY_SENT);
 679         nf_ct_delete_from_lists(ct);
 680         nf_ct_put(ct);
 681         return true;
 682 }
 683 EXPORT_SYMBOL_GPL(nf_ct_delete);
 684
 685 static inline bool
 686 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
 687                 const struct nf_conntrack_tuple *tuple,
 688                 const struct nf_conntrack_zone *zone,
 689                 const struct net *net)
 690 {
 691         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 692
 693         /* A conntrack can be recreated with the equal tuple,
 694          * so we need to check that the conntrack is confirmed
 695          */
 696         return nf_ct_tuple_equal(tuple, &h->tuple) &&
 697                nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
 698                nf_ct_is_confirmed(ct) &&
 699                net_eq(net, nf_ct_net(ct));
 700 }
 701
 702 static inline bool
 703 nf_ct_match(const struct nf_conn *ct1, const struct nf_conn *ct2)
 704 {
 705         return nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 706                                  &ct2->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
 707                nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_REPLY].tuple,
 708                                  &ct2->tuplehash[IP_CT_DIR_REPLY].tuple) &&
 709                nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_ORIGINAL) &&
 710                nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_REPLY) &&
 711                net_eq(nf_ct_net(ct1), nf_ct_net(ct2));
 712 }
 713
 714 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
 715 static void nf_ct_gc_expired(struct nf_conn *ct)
 716 {
 717         if (!refcount_inc_not_zero(&ct->ct_general.use))
 718                 return;
 719
 720         /* load ->status after refcount increase */
 721         smp_acquire__after_ctrl_dep();
 722
 723         if (nf_ct_should_gc(ct))
 724                 nf_ct_kill(ct);
 725
 726         nf_ct_put(ct);
 727 }
 728
 729 /*
 730  * Warning :
 731  * - Caller must take a reference on returned object
 732  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 733  */
 734 static struct nf_conntrack_tuple_hash *
 735 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
 736                       const struct nf_conntrack_tuple *tuple, u32 hash)
 737 {
 738         struct nf_conntrack_tuple_hash *h;
 739         struct hlist_nulls_head *ct_hash;
 740         struct hlist_nulls_node *n;
 741         unsigned int bucket, hsize;
 742
 743 begin:
 744         nf_conntrack_get_ht(&ct_hash, &hsize);
 745         bucket = reciprocal_scale(hash, hsize);
 746
 747         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
 748                 struct nf_conn *ct;
 749
 750                 ct = nf_ct_tuplehash_to_ctrack(h);
 751                 if (nf_ct_is_expired(ct)) {
 752                         nf_ct_gc_expired(ct);
 753                         continue;
 754                 }
 755
 756                 if (nf_ct_key_equal(h, tuple, zone, net))
 757                         return h;
 758         }
 759         /*
 760          * if the nulls value we got at the end of this lookup is
 761          * not the expected one, we must restart lookup.
 762          * We probably met an item that was moved to another chain.
 763          */
 764         if (get_nulls_value(n) != bucket) {
 765                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 766                 goto begin;
 767         }
 768
 769         return NULL;
 770 }
 771
 772 /* Find a connection corresponding to a tuple. */
 773 static struct nf_conntrack_tuple_hash *
 774 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 775                         const struct nf_conntrack_tuple *tuple, u32 hash)
 776 {
 777         struct nf_conntrack_tuple_hash *h;
 778         struct nf_conn *ct;
 779
 780         h = ____nf_conntrack_find(net, zone, tuple, hash);
 781         if (h) {
 782                 /* We have a candidate that matches the tuple we're interested
 783                  * in, try to obtain a reference and re-check tuple
 784                  */
 785                 ct = nf_ct_tuplehash_to_ctrack(h);
 786                 if (likely(refcount_inc_not_zero(&ct->ct_general.use))) {
 787                         /* re-check key after refcount */
 788                         smp_acquire__after_ctrl_dep();
 789
 790                         if (likely(nf_ct_key_equal(h, tuple, zone, net)))
 791                                 return h;
 792
 793                         /* TYPESAFE_BY_RCU recycled the candidate */
 794                         nf_ct_put(ct);
 795                 }
 796
 797                 h = NULL;
 798         }
 799
 800         return h;
 801 }
 802
 803 struct nf_conntrack_tuple_hash *
 804 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 805                       const struct nf_conntrack_tuple *tuple)
 806 {
 807         unsigned int rid, zone_id = nf_ct_zone_id(zone, IP_CT_DIR_ORIGINAL);
 808         struct nf_conntrack_tuple_hash *thash;
 809
 810         rcu_read_lock();
 811
 812         thash = __nf_conntrack_find_get(net, zone, tuple,
 813                                         hash_conntrack_raw(tuple, zone_id, net));
 814
 815         if (thash)
 816                 goto out_unlock;
 817
 818         rid = nf_ct_zone_id(zone, IP_CT_DIR_REPLY);
 819         if (rid != zone_id)
 820                 thash = __nf_conntrack_find_get(net, zone, tuple,
 821                                                 hash_conntrack_raw(tuple, rid, net));
 822
 823 out_unlock:
 824         rcu_read_unlock();
 825         return thash;
 826 }
 827 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 828
 829 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 830                                        unsigned int hash,
 831                                        unsigned int reply_hash)
 832 {
 833         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 834                            &nf_conntrack_hash[hash]);
 835         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 836                            &nf_conntrack_hash[reply_hash]);
 837 }
 838
 839 static bool nf_ct_ext_valid_pre(const struct nf_ct_ext *ext)
 840 {
 841         /* if ext->gen_id is not equal to nf_conntrack_ext_genid, some extensions
 842          * may contain stale pointers to e.g. helper that has been removed.
 843          *
 844          * The helper can't clear this because the nf_conn object isn't in
 845          * any hash and synchronize_rcu() isn't enough because associated skb
 846          * might sit in a queue.
 847          */
 848         return !ext || ext->gen_id == atomic_read(&nf_conntrack_ext_genid);
 849 }
 850
 851 static bool nf_ct_ext_valid_post(struct nf_ct_ext *ext)
 852 {
 853         if (!ext)
 854                 return true;
 855
 856         if (ext->gen_id != atomic_read(&nf_conntrack_ext_genid))
 857                 return false;
 858
 859         /* inserted into conntrack table, nf_ct_iterate_cleanup()
 860          * will find it.  Disable nf_ct_ext_find() id check.
 861          */
 862         WRITE_ONCE(ext->gen_id, 0);
 863         return true;
 864 }
 865
 866 int
 867 nf_conntrack_hash_check_insert(struct nf_conn *ct)
 868 {
 869         const struct nf_conntrack_zone *zone;
 870         struct net *net = nf_ct_net(ct);
 871         unsigned int hash, reply_hash;
 872         struct nf_conntrack_tuple_hash *h;
 873         struct hlist_nulls_node *n;
 874         unsigned int max_chainlen;
 875         unsigned int chainlen = 0;
 876         unsigned int sequence;
 877         int err = -EEXIST;
 878
 879         zone = nf_ct_zone(ct);
 880
 881         if (!nf_ct_ext_valid_pre(ct->ext))
 882                 return -EAGAIN;
 883
 884         local_bh_disable();
 885         do {
 886                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 887                 hash = hash_conntrack(net,
 888                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 889                                       nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_ORIGINAL));
 890                 reply_hash = hash_conntrack(net,
 891                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 892                                            nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
 893         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 894
 895         max_chainlen = MIN_CHAINLEN + get_random_u32_below(MAX_CHAINLEN);
 896
 897         /* See if there's one in the list already, including reverse */
 898         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode) {
 899                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 900                                     zone, net))
 901                         goto out;
 902
 903                 if (chainlen++ > max_chainlen)
 904                         goto chaintoolong;
 905         }
 906
 907         chainlen = 0;
 908
 909         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode) {
 910                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 911                                     zone, net))
 912                         goto out;
 913                 if (chainlen++ > max_chainlen)
 914                         goto chaintoolong;
 915         }
 916
 917         /* If genid has changed, we can't insert anymore because ct
 918          * extensions could have stale pointers and nf_ct_iterate_destroy
 919          * might have completed its table scan already.
 920          *
 921          * Increment of the ext genid right after this check is fine:
 922          * nf_ct_iterate_destroy blocks until locks are released.
 923          */
 924         if (!nf_ct_ext_valid_post(ct->ext)) {
 925                 err = -EAGAIN;
 926                 goto out;
 927         }
 928
 929         smp_wmb();
 930         /* The caller holds a reference to this object */
 931         refcount_set(&ct->ct_general.use, 2);
 932         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 933         nf_conntrack_double_unlock(hash, reply_hash);
 934         NF_CT_STAT_INC(net, insert);
 935         local_bh_enable();
 936
 937         return 0;
 938 chaintoolong:
 939         NF_CT_STAT_INC(net, chaintoolong);
 940         err = -ENOSPC;
 941 out:
 942         nf_conntrack_double_unlock(hash, reply_hash);
 943         local_bh_enable();
 944         return err;
 945 }
 946 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 947
 948 void nf_ct_acct_add(struct nf_conn *ct, u32 dir, unsigned int packets,
 949                     unsigned int bytes)
 950 {
 951         struct nf_conn_acct *acct;
 952
 953         acct = nf_conn_acct_find(ct);
 954         if (acct) {
 955                 struct nf_conn_counter *counter = acct->counter;
 956
 957                 atomic64_add(packets, &counter[dir].packets);
 958                 atomic64_add(bytes, &counter[dir].bytes);
 959         }
 960 }
 961 EXPORT_SYMBOL_GPL(nf_ct_acct_add);
 962
 963 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
 964                              const struct nf_conn *loser_ct)
 965 {
 966         struct nf_conn_acct *acct;
 967
 968         acct = nf_conn_acct_find(loser_ct);
 969         if (acct) {
 970                 struct nf_conn_counter *counter = acct->counter;
 971                 unsigned int bytes;
 972
 973                 /* u32 should be fine since we must have seen one packet. */
 974                 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
 975                 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), bytes);
 976         }
 977 }
 978
 979 static void __nf_conntrack_insert_prepare(struct nf_conn *ct)
 980 {
 981         struct nf_conn_tstamp *tstamp;
 982
 983         refcount_inc(&ct->ct_general.use);
 984
 985         /* set conntrack timestamp, if enabled. */
 986         tstamp = nf_conn_tstamp_find(ct);
 987         if (tstamp)
 988                 tstamp->start = ktime_get_real_ns();
 989 }
 990
 991 /* caller must hold locks to prevent concurrent changes */
 992 static int __nf_ct_resolve_clash(struct sk_buff *skb,
 993                                  struct nf_conntrack_tuple_hash *h)
 994 {
 995         /* This is the conntrack entry already in hashes that won race. */
 996         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 997         enum ip_conntrack_info ctinfo;
 998         struct nf_conn *loser_ct;
 999
1000         loser_ct = nf_ct_get(skb, &ctinfo);
1001
1002         if (nf_ct_is_dying(ct))
1003                 return NF_DROP;
1004
1005         if (((ct->status & IPS_NAT_DONE_MASK) == 0) ||
1006             nf_ct_match(ct, loser_ct)) {
1007                 struct net *net = nf_ct_net(ct);
1008
1009                 nf_conntrack_get(&ct->ct_general);
1010
1011                 nf_ct_acct_merge(ct, ctinfo, loser_ct);
1012                 nf_ct_put(loser_ct);
1013                 nf_ct_set(skb, ct, ctinfo);
1014
1015                 NF_CT_STAT_INC(net, clash_resolve);
1016                 return NF_ACCEPT;
1017         }
1018
1019         return NF_DROP;
1020 }
1021
1022 /**
1023  * nf_ct_resolve_clash_harder - attempt to insert clashing conntrack entry
1024  *
1025  * @skb: skb that causes the collision
1026  * @repl_idx: hash slot for reply direction
1027  *
1028  * Called when origin or reply direction had a clash.
1029  * The skb can be handled without packet drop provided the reply direction
1030  * is unique or there the existing entry has the identical tuple in both
1031  * directions.
1032  *
1033  * Caller must hold conntrack table locks to prevent concurrent updates.
1034  *
1035  * Returns NF_DROP if the clash could not be handled.
1036  */
1037 static int nf_ct_resolve_clash_harder(struct sk_buff *skb, u32 repl_idx)
1038 {
1039         struct nf_conn *loser_ct = (struct nf_conn *)skb_nfct(skb);
1040         const struct nf_conntrack_zone *zone;
1041         struct nf_conntrack_tuple_hash *h;
1042         struct hlist_nulls_node *n;
1043         struct net *net;
1044
1045         zone = nf_ct_zone(loser_ct);
1046         net = nf_ct_net(loser_ct);
1047
1048         /* Reply direction must never result in a clash, unless both origin
1049          * and reply tuples are identical.
1050          */
1051         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[repl_idx], hnnode) {
1052                 if (nf_ct_key_equal(h,
1053                                     &loser_ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1054                                     zone, net))
1055                         return __nf_ct_resolve_clash(skb, h);
1056         }
1057
1058         /* We want the clashing entry to go away real soon: 1 second timeout. */
1059         WRITE_ONCE(loser_ct->timeout, nfct_time_stamp + HZ);
1060
1061         /* IPS_NAT_CLASH removes the entry automatically on the first
1062          * reply.  Also prevents UDP tracker from moving the entry to
1063          * ASSURED state, i.e. the entry can always be evicted under
1064          * pressure.
1065          */
1066         loser_ct->status |= IPS_FIXED_TIMEOUT | IPS_NAT_CLASH;
1067
1068         __nf_conntrack_insert_prepare(loser_ct);
1069
1070         /* fake add for ORIGINAL dir: we want lookups to only find the entry
1071          * already in the table.  This also hides the clashing entry from
1072          * ctnetlink iteration, i.e. conntrack -L won't show them.
1073          */
1074         hlist_nulls_add_fake(&loser_ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
1075
1076         hlist_nulls_add_head_rcu(&loser_ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
1077                                  &nf_conntrack_hash[repl_idx]);
1078
1079         NF_CT_STAT_INC(net, clash_resolve);
1080         return NF_ACCEPT;
1081 }
1082
1083 /**
1084  * nf_ct_resolve_clash - attempt to handle clash without packet drop
1085  *
1086  * @skb: skb that causes the clash
1087  * @h: tuplehash of the clashing entry already in table
1088  * @reply_hash: hash slot for reply direction
1089  *
1090  * A conntrack entry can be inserted to the connection tracking table
1091  * if there is no existing entry with an identical tuple.
1092  *
1093  * If there is one, @skb (and the assocated, unconfirmed conntrack) has
1094  * to be dropped.  In case @skb is retransmitted, next conntrack lookup
1095  * will find the already-existing entry.
1096  *
1097  * The major problem with such packet drop is the extra delay added by
1098  * the packet loss -- it will take some time for a retransmit to occur
1099  * (or the sender to time out when waiting for a reply).
1100  *
1101  * This function attempts to handle the situation without packet drop.
1102  *
1103  * If @skb has no NAT transformation or if the colliding entries are
1104  * exactly the same, only the to-be-confirmed conntrack entry is discarded
1105  * and @skb is associated with the conntrack entry already in the table.
1106  *
1107  * Failing that, the new, unconfirmed conntrack is still added to the table
1108  * provided that the collision only occurs in the ORIGINAL direction.
1109  * The new entry will be added only in the non-clashing REPLY direction,
1110  * so packets in the ORIGINAL direction will continue to match the existing
1111  * entry.  The new entry will also have a fixed timeout so it expires --
1112  * due to the collision, it will only see reply traffic.
1113  *
1114  * Returns NF_DROP if the clash could not be resolved.
1115  */
1116 static __cold noinline int
1117 nf_ct_resolve_clash(struct sk_buff *skb, struct nf_conntrack_tuple_hash *h,
1118                     u32 reply_hash)
1119 {
1120         /* This is the conntrack entry already in hashes that won race. */
1121         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1122         const struct nf_conntrack_l4proto *l4proto;
1123         enum ip_conntrack_info ctinfo;
1124         struct nf_conn *loser_ct;
1125         struct net *net;
1126         int ret;
1127
1128         loser_ct = nf_ct_get(skb, &ctinfo);
1129         net = nf_ct_net(loser_ct);
1130
1131         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1132         if (!l4proto->allow_clash)
1133                 goto drop;
1134
1135         ret = __nf_ct_resolve_clash(skb, h);
1136         if (ret == NF_ACCEPT)
1137                 return ret;
1138
1139         ret = nf_ct_resolve_clash_harder(skb, reply_hash);
1140         if (ret == NF_ACCEPT)
1141                 return ret;
1142
1143 drop:
1144         NF_CT_STAT_INC(net, drop);
1145         NF_CT_STAT_INC(net, insert_failed);
1146         return NF_DROP;
1147 }
1148
1149 /* Confirm a connection given skb; places it in hash table */
1150 int
1151 __nf_conntrack_confirm(struct sk_buff *skb)
1152 {
1153         unsigned int chainlen = 0, sequence, max_chainlen;
1154         const struct nf_conntrack_zone *zone;
1155         unsigned int hash, reply_hash;
1156         struct nf_conntrack_tuple_hash *h;
1157         struct nf_conn *ct;
1158         struct nf_conn_help *help;
1159         struct hlist_nulls_node *n;
1160         enum ip_conntrack_info ctinfo;
1161         struct net *net;
1162         int ret = NF_DROP;
1163
1164         ct = nf_ct_get(skb, &ctinfo);
1165         net = nf_ct_net(ct);
1166
1167         /* ipt_REJECT uses nf_conntrack_attach to attach related
1168            ICMP/TCP RST packets in other direction.  Actual packet
1169            which created connection will be IP_CT_NEW or for an
1170            expected connection, IP_CT_RELATED. */
1171         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
1172                 return NF_ACCEPT;
1173
1174         zone = nf_ct_zone(ct);
1175         local_bh_disable();
1176
1177         do {
1178                 sequence = read_seqcount_begin(&nf_conntrack_generation);
1179                 /* reuse the hash saved before */
1180                 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
1181                 hash = scale_hash(hash);
1182                 reply_hash = hash_conntrack(net,
1183                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1184                                            nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
1185         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
1186
1187         /* We're not in hash table, and we refuse to set up related
1188          * connections for unconfirmed conns.  But packet copies and
1189          * REJECT will give spurious warnings here.
1190          */
1191
1192         /* Another skb with the same unconfirmed conntrack may
1193          * win the race. This may happen for bridge(br_flood)
1194          * or broadcast/multicast packets do skb_clone with
1195          * unconfirmed conntrack.
1196          */
1197         if (unlikely(nf_ct_is_confirmed(ct))) {
1198                 WARN_ON_ONCE(1);
1199                 nf_conntrack_double_unlock(hash, reply_hash);
1200                 local_bh_enable();
1201                 return NF_DROP;
1202         }
1203
1204         if (!nf_ct_ext_valid_pre(ct->ext)) {
1205                 NF_CT_STAT_INC(net, insert_failed);
1206                 goto dying;
1207         }
1208
1209         /* We have to check the DYING flag after unlink to prevent
1210          * a race against nf_ct_get_next_corpse() possibly called from
1211          * user context, else we insert an already 'dead' hash, blocking
1212          * further use of that particular connection -JM.
1213          */
1214         ct->status |= IPS_CONFIRMED;
1215
1216         if (unlikely(nf_ct_is_dying(ct))) {
1217                 NF_CT_STAT_INC(net, insert_failed);
1218                 goto dying;
1219         }
1220
1221         max_chainlen = MIN_CHAINLEN + get_random_u32_below(MAX_CHAINLEN);
1222         /* See if there's one in the list already, including reverse:
1223            NAT could have grabbed it without realizing, since we're
1224            not in the hash.  If there is, we lost race. */
1225         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode) {
1226                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1227                                     zone, net))
1228                         goto out;
1229                 if (chainlen++ > max_chainlen)
1230                         goto chaintoolong;
1231         }
1232
1233         chainlen = 0;
1234         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode) {
1235                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1236                                     zone, net))
1237                         goto out;
1238                 if (chainlen++ > max_chainlen) {
1239 chaintoolong:
1240                         NF_CT_STAT_INC(net, chaintoolong);
1241                         NF_CT_STAT_INC(net, insert_failed);
1242                         ret = NF_DROP;
1243                         goto dying;
1244                 }
1245         }
1246
1247         /* Timer relative to confirmation time, not original
1248            setting time, otherwise we'd get timer wrap in
1249            weird delay cases. */
1250         ct->timeout += nfct_time_stamp;
1251
1252         __nf_conntrack_insert_prepare(ct);
1253
1254         /* Since the lookup is lockless, hash insertion must be done after
1255          * starting the timer and setting the CONFIRMED bit. The RCU barriers
1256          * guarantee that no other CPU can find the conntrack before the above
1257          * stores are visible.
1258          */
1259         __nf_conntrack_hash_insert(ct, hash, reply_hash);
1260         nf_conntrack_double_unlock(hash, reply_hash);
1261         local_bh_enable();
1262
1263         /* ext area is still valid (rcu read lock is held,
1264          * but will go out of scope soon, we need to remove
1265          * this conntrack again.
1266          */
1267         if (!nf_ct_ext_valid_post(ct->ext)) {
1268                 nf_ct_kill(ct);
1269                 NF_CT_STAT_INC_ATOMIC(net, drop);
1270                 return NF_DROP;
1271         }
1272
1273         help = nfct_help(ct);
1274         if (help && help->helper)
1275                 nf_conntrack_event_cache(IPCT_HELPER, ct);
1276
1277         nf_conntrack_event_cache(master_ct(ct) ?
1278                                  IPCT_RELATED : IPCT_NEW, ct);
1279         return NF_ACCEPT;
1280
1281 out:
1282         ret = nf_ct_resolve_clash(skb, h, reply_hash);
1283 dying:
1284         nf_conntrack_double_unlock(hash, reply_hash);
1285         local_bh_enable();
1286         return ret;
1287 }
1288 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
1289
1290 /* Returns true if a connection corresponds to the tuple (required
1291    for NAT). */
1292 int
1293 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
1294                          const struct nf_conn *ignored_conntrack)
1295 {
1296         struct net *net = nf_ct_net(ignored_conntrack);
1297         const struct nf_conntrack_zone *zone;
1298         struct nf_conntrack_tuple_hash *h;
1299         struct hlist_nulls_head *ct_hash;
1300         unsigned int hash, hsize;
1301         struct hlist_nulls_node *n;
1302         struct nf_conn *ct;
1303
1304         zone = nf_ct_zone(ignored_conntrack);
1305
1306         rcu_read_lock();
1307  begin:
1308         nf_conntrack_get_ht(&ct_hash, &hsize);
1309         hash = __hash_conntrack(net, tuple, nf_ct_zone_id(zone, IP_CT_DIR_REPLY), hsize);
1310
1311         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
1312                 ct = nf_ct_tuplehash_to_ctrack(h);
1313
1314                 if (ct == ignored_conntrack)
1315                         continue;
1316
1317                 if (nf_ct_is_expired(ct)) {
1318                         nf_ct_gc_expired(ct);
1319                         continue;
1320                 }
1321
1322                 if (nf_ct_key_equal(h, tuple, zone, net)) {
1323                         /* Tuple is taken already, so caller will need to find
1324                          * a new source port to use.
1325                          *
1326                          * Only exception:
1327                          * If the *original tuples* are identical, then both
1328                          * conntracks refer to the same flow.
1329                          * This is a rare situation, it can occur e.g. when
1330                          * more than one UDP packet is sent from same socket
1331                          * in different threads.
1332                          *
1333                          * Let nf_ct_resolve_clash() deal with this later.
1334                          */
1335                         if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1336                                               &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
1337                                               nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL))
1338                                 continue;
1339
1340                         NF_CT_STAT_INC_ATOMIC(net, found);
1341                         rcu_read_unlock();
1342                         return 1;
1343                 }
1344         }
1345
1346         if (get_nulls_value(n) != hash) {
1347                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
1348                 goto begin;
1349         }
1350
1351         rcu_read_unlock();
1352
1353         return 0;
1354 }
1355 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
1356
1357 #define NF_CT_EVICTION_RANGE    8
1358
1359 /* There's a small race here where we may free a just-assured
1360    connection.  Too bad: we're in trouble anyway. */
1361 static unsigned int early_drop_list(struct net *net,
1362                                     struct hlist_nulls_head *head)
1363 {
1364         struct nf_conntrack_tuple_hash *h;
1365         struct hlist_nulls_node *n;
1366         unsigned int drops = 0;
1367         struct nf_conn *tmp;
1368
1369         hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
1370                 tmp = nf_ct_tuplehash_to_ctrack(h);
1371
1372                 if (nf_ct_is_expired(tmp)) {
1373                         nf_ct_gc_expired(tmp);
1374                         continue;
1375                 }
1376
1377                 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
1378                     !net_eq(nf_ct_net(tmp), net) ||
1379                     nf_ct_is_dying(tmp))
1380                         continue;
1381
1382                 if (!refcount_inc_not_zero(&tmp->ct_general.use))
1383                         continue;
1384
1385                 /* load ->ct_net and ->status after refcount increase */
1386                 smp_acquire__after_ctrl_dep();
1387
1388                 /* kill only if still in same netns -- might have moved due to
1389                  * SLAB_TYPESAFE_BY_RCU rules.
1390                  *
1391                  * We steal the timer reference.  If that fails timer has
1392                  * already fired or someone else deleted it. Just drop ref
1393                  * and move to next entry.
1394                  */
1395                 if (net_eq(nf_ct_net(tmp), net) &&
1396                     nf_ct_is_confirmed(tmp) &&
1397                     nf_ct_delete(tmp, 0, 0))
1398                         drops++;
1399
1400                 nf_ct_put(tmp);
1401         }
1402
1403         return drops;
1404 }
1405
1406 static noinline int early_drop(struct net *net, unsigned int hash)
1407 {
1408         unsigned int i, bucket;
1409
1410         for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
1411                 struct hlist_nulls_head *ct_hash;
1412                 unsigned int hsize, drops;
1413
1414                 rcu_read_lock();
1415                 nf_conntrack_get_ht(&ct_hash, &hsize);
1416                 if (!i)
1417                         bucket = reciprocal_scale(hash, hsize);
1418                 else
1419                         bucket = (bucket + 1) % hsize;
1420
1421                 drops = early_drop_list(net, &ct_hash[bucket]);
1422                 rcu_read_unlock();
1423
1424                 if (drops) {
1425                         NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
1426                         return true;
1427                 }
1428         }
1429
1430         return false;
1431 }
1432
1433 static bool gc_worker_skip_ct(const struct nf_conn *ct)
1434 {
1435         return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
1436 }
1437
1438 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
1439 {
1440         const struct nf_conntrack_l4proto *l4proto;
1441         u8 protonum = nf_ct_protonum(ct);
1442
1443         if (test_bit(IPS_OFFLOAD_BIT, &ct->status) && protonum != IPPROTO_UDP)
1444                 return false;
1445         if (!test_bit(IPS_ASSURED_BIT, &ct->status))
1446                 return true;
1447
1448         l4proto = nf_ct_l4proto_find(protonum);
1449         if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
1450                 return true;
1451
1452         return false;
1453 }
1454
1455 static void gc_worker(struct work_struct *work)
1456 {
1457         unsigned int i, hashsz, nf_conntrack_max95 = 0;
1458         u32 end_time, start_time = nfct_time_stamp;
1459         struct conntrack_gc_work *gc_work;
1460         unsigned int expired_count = 0;
1461         unsigned long next_run;
1462         s32 delta_time;
1463         long count;
1464
1465         gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1466
1467         i = gc_work->next_bucket;
1468         if (gc_work->early_drop)
1469                 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1470
1471         if (i == 0) {
1472                 gc_work->avg_timeout = GC_SCAN_INTERVAL_INIT;
1473                 gc_work->count = GC_SCAN_INITIAL_COUNT;
1474                 gc_work->start_time = start_time;
1475         }
1476
1477         next_run = gc_work->avg_timeout;
1478         count = gc_work->count;
1479
1480         end_time = start_time + GC_SCAN_MAX_DURATION;
1481
1482         do {
1483                 struct nf_conntrack_tuple_hash *h;
1484                 struct hlist_nulls_head *ct_hash;
1485                 struct hlist_nulls_node *n;
1486                 struct nf_conn *tmp;
1487
1488                 rcu_read_lock();
1489
1490                 nf_conntrack_get_ht(&ct_hash, &hashsz);
1491                 if (i >= hashsz) {
1492                         rcu_read_unlock();
1493                         break;
1494                 }
1495
1496                 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1497                         struct nf_conntrack_net *cnet;
1498                         struct net *net;
1499                         long expires;
1500
1501                         tmp = nf_ct_tuplehash_to_ctrack(h);
1502
1503                         if (test_bit(IPS_OFFLOAD_BIT, &tmp->status)) {
1504                                 nf_ct_offload_timeout(tmp);
1505                                 if (!nf_conntrack_max95)
1506                                         continue;
1507                         }
1508
1509                         if (expired_count > GC_SCAN_EXPIRED_MAX) {
1510                                 rcu_read_unlock();
1511
1512                                 gc_work->next_bucket = i;
1513                                 gc_work->avg_timeout = next_run;
1514                                 gc_work->count = count;
1515
1516                                 delta_time = nfct_time_stamp - gc_work->start_time;
1517
1518                                 /* re-sched immediately if total cycle time is exceeded */
1519                                 next_run = delta_time < (s32)GC_SCAN_INTERVAL_MAX;
1520                                 goto early_exit;
1521                         }
1522
1523                         if (nf_ct_is_expired(tmp)) {
1524                                 nf_ct_gc_expired(tmp);
1525                                 expired_count++;
1526                                 continue;
1527                         }
1528
1529                         expires = clamp(nf_ct_expires(tmp), GC_SCAN_INTERVAL_MIN, GC_SCAN_INTERVAL_CLAMP);
1530                         expires = (expires - (long)next_run) / ++count;
1531                         next_run += expires;
1532
1533                         if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1534                                 continue;
1535
1536                         net = nf_ct_net(tmp);
1537                         cnet = nf_ct_pernet(net);
1538                         if (atomic_read(&cnet->count) < nf_conntrack_max95)
1539                                 continue;
1540
1541                         /* need to take reference to avoid possible races */
1542                         if (!refcount_inc_not_zero(&tmp->ct_general.use))
1543                                 continue;
1544
1545                         /* load ->status after refcount increase */
1546                         smp_acquire__after_ctrl_dep();
1547
1548                         if (gc_worker_skip_ct(tmp)) {
1549                                 nf_ct_put(tmp);
1550                                 continue;
1551                         }
1552
1553                         if (gc_worker_can_early_drop(tmp)) {
1554                                 nf_ct_kill(tmp);
1555                                 expired_count++;
1556                         }
1557
1558                         nf_ct_put(tmp);
1559                 }
1560
1561                 /* could check get_nulls_value() here and restart if ct
1562                  * was moved to another chain.  But given gc is best-effort
1563                  * we will just continue with next hash slot.
1564                  */
1565                 rcu_read_unlock();
1566                 cond_resched();
1567                 i++;
1568
1569                 delta_time = nfct_time_stamp - end_time;
1570                 if (delta_time > 0 && i < hashsz) {
1571                         gc_work->avg_timeout = next_run;
1572                         gc_work->count = count;
1573                         gc_work->next_bucket = i;
1574                         next_run = 0;
1575                         goto early_exit;
1576                 }
1577         } while (i < hashsz);
1578
1579         gc_work->next_bucket = 0;
1580
1581         next_run = clamp(next_run, GC_SCAN_INTERVAL_MIN, GC_SCAN_INTERVAL_MAX);
1582
1583         delta_time = max_t(s32, nfct_time_stamp - gc_work->start_time, 1);
1584         if (next_run > (unsigned long)delta_time)
1585                 next_run -= delta_time;
1586         else
1587                 next_run = 1;
1588
1589 early_exit:
1590         if (gc_work->exiting)
1591                 return;
1592
1593         if (next_run)
1594                 gc_work->early_drop = false;
1595
1596         queue_delayed_work(system_power_efficient_wq, &gc_work->dwork, next_run);
1597 }
1598
1599 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1600 {
1601         INIT_DELAYED_WORK(&gc_work->dwork, gc_worker);
1602         gc_work->exiting = false;
1603 }
1604
1605 static struct nf_conn *
1606 __nf_conntrack_alloc(struct net *net,
1607                      const struct nf_conntrack_zone *zone,
1608                      const struct nf_conntrack_tuple *orig,
1609                      const struct nf_conntrack_tuple *repl,
1610                      gfp_t gfp, u32 hash)
1611 {
1612         struct nf_conntrack_net *cnet = nf_ct_pernet(net);
1613         unsigned int ct_count;
1614         struct nf_conn *ct;
1615
1616         /* We don't want any race condition at early drop stage */
1617         ct_count = atomic_inc_return(&cnet->count);
1618
1619         if (nf_conntrack_max && unlikely(ct_count > nf_conntrack_max)) {
1620                 if (!early_drop(net, hash)) {
1621                         if (!conntrack_gc_work.early_drop)
1622                                 conntrack_gc_work.early_drop = true;
1623                         atomic_dec(&cnet->count);
1624                         net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1625                         return ERR_PTR(-ENOMEM);
1626                 }
1627         }
1628
1629         /*
1630          * Do not use kmem_cache_zalloc(), as this cache uses
1631          * SLAB_TYPESAFE_BY_RCU.
1632          */
1633         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1634         if (ct == NULL)
1635                 goto out;
1636
1637         spin_lock_init(&ct->lock);
1638         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1639         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1640         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1641         /* save hash for reusing when confirming */
1642         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1643         ct->status = 0;
1644         WRITE_ONCE(ct->timeout, 0);
1645         write_pnet(&ct->ct_net, net);
1646         memset_after(ct, 0, __nfct_init_offset);
1647
1648         nf_ct_zone_add(ct, zone);
1649
1650         /* Because we use RCU lookups, we set ct_general.use to zero before
1651          * this is inserted in any list.
1652          */
1653         refcount_set(&ct->ct_general.use, 0);
1654         return ct;
1655 out:
1656         atomic_dec(&cnet->count);
1657         return ERR_PTR(-ENOMEM);
1658 }
1659
1660 struct nf_conn *nf_conntrack_alloc(struct net *net,
1661                                    const struct nf_conntrack_zone *zone,
1662                                    const struct nf_conntrack_tuple *orig,
1663                                    const struct nf_conntrack_tuple *repl,
1664                                    gfp_t gfp)
1665 {
1666         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1667 }
1668 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1669
1670 void nf_conntrack_free(struct nf_conn *ct)
1671 {
1672         struct net *net = nf_ct_net(ct);
1673         struct nf_conntrack_net *cnet;
1674
1675         /* A freed object has refcnt == 0, that's
1676          * the golden rule for SLAB_TYPESAFE_BY_RCU
1677          */
1678         WARN_ON(refcount_read(&ct->ct_general.use) != 0);
1679
1680         if (ct->status & IPS_SRC_NAT_DONE) {
1681                 const struct nf_nat_hook *nat_hook;
1682
1683                 rcu_read_lock();
1684                 nat_hook = rcu_dereference(nf_nat_hook);
1685                 if (nat_hook)
1686                         nat_hook->remove_nat_bysrc(ct);
1687                 rcu_read_unlock();
1688         }
1689
1690         kfree(ct->ext);
1691         kmem_cache_free(nf_conntrack_cachep, ct);
1692         cnet = nf_ct_pernet(net);
1693
1694         smp_mb__before_atomic();
1695         atomic_dec(&cnet->count);
1696 }
1697 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1698
1699
1700 /* Allocate a new conntrack: we return -ENOMEM if classification
1701    failed due to stress.  Otherwise it really is unclassifiable. */
1702 static noinline struct nf_conntrack_tuple_hash *
1703 init_conntrack(struct net *net, struct nf_conn *tmpl,
1704                const struct nf_conntrack_tuple *tuple,
1705                struct sk_buff *skb,
1706                unsigned int dataoff, u32 hash)
1707 {
1708         struct nf_conn *ct;
1709         struct nf_conn_help *help;
1710         struct nf_conntrack_tuple repl_tuple;
1711 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1712         struct nf_conntrack_ecache *ecache;
1713 #endif
1714         struct nf_conntrack_expect *exp = NULL;
1715         const struct nf_conntrack_zone *zone;
1716         struct nf_conn_timeout *timeout_ext;
1717         struct nf_conntrack_zone tmp;
1718         struct nf_conntrack_net *cnet;
1719
1720         if (!nf_ct_invert_tuple(&repl_tuple, tuple))
1721                 return NULL;
1722
1723         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1724         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1725                                   hash);
1726         if (IS_ERR(ct))
1727                 return (struct nf_conntrack_tuple_hash *)ct;
1728
1729         if (!nf_ct_add_synproxy(ct, tmpl)) {
1730                 nf_conntrack_free(ct);
1731                 return ERR_PTR(-ENOMEM);
1732         }
1733
1734         timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1735
1736         if (timeout_ext)
1737                 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1738                                       GFP_ATOMIC);
1739
1740         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1741         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1742         nf_ct_labels_ext_add(ct);
1743
1744 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1745         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1746
1747         if ((ecache || net->ct.sysctl_events) &&
1748             !nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1749                                   ecache ? ecache->expmask : 0,
1750                                   GFP_ATOMIC)) {
1751                 nf_conntrack_free(ct);
1752                 return ERR_PTR(-ENOMEM);
1753         }
1754 #endif
1755
1756         cnet = nf_ct_pernet(net);
1757         if (cnet->expect_count) {
1758                 spin_lock_bh(&nf_conntrack_expect_lock);
1759                 exp = nf_ct_find_expectation(net, zone, tuple, !tmpl || nf_ct_is_confirmed(tmpl));
1760                 if (exp) {
1761                         /* Welcome, Mr. Bond.  We've been expecting you... */
1762                         __set_bit(IPS_EXPECTED_BIT, &ct->status);
1763                         /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1764                         ct->master = exp->master;
1765                         if (exp->helper) {
1766                                 help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1767                                 if (help)
1768                                         rcu_assign_pointer(help->helper, exp->helper);
1769                         }
1770
1771 #ifdef CONFIG_NF_CONNTRACK_MARK
1772                         ct->mark = READ_ONCE(exp->master->mark);
1773 #endif
1774 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1775                         ct->secmark = exp->master->secmark;
1776 #endif
1777                         NF_CT_STAT_INC(net, expect_new);
1778                 }
1779                 spin_unlock_bh(&nf_conntrack_expect_lock);
1780         }
1781         if (!exp && tmpl)
1782                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1783
1784         /* Other CPU might have obtained a pointer to this object before it was
1785          * released.  Because refcount is 0, refcount_inc_not_zero() will fail.
1786          *
1787          * After refcount_set(1) it will succeed; ensure that zeroing of
1788          * ct->status and the correct ct->net pointer are visible; else other
1789          * core might observe CONFIRMED bit which means the entry is valid and
1790          * in the hash table, but its not (anymore).
1791          */
1792         smp_wmb();
1793
1794         /* Now it is going to be associated with an sk_buff, set refcount to 1. */
1795         refcount_set(&ct->ct_general.use, 1);
1796
1797         if (exp) {
1798                 if (exp->expectfn)
1799                         exp->expectfn(ct, exp);
1800                 nf_ct_expect_put(exp);
1801         }
1802
1803         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1804 }
1805
1806 /* On success, returns 0, sets skb->_nfct | ctinfo */
1807 static int
1808 resolve_normal_ct(struct nf_conn *tmpl,
1809                   struct sk_buff *skb,
1810                   unsigned int dataoff,
1811                   u_int8_t protonum,
1812                   const struct nf_hook_state *state)
1813 {
1814         const struct nf_conntrack_zone *zone;
1815         struct nf_conntrack_tuple tuple;
1816         struct nf_conntrack_tuple_hash *h;
1817         enum ip_conntrack_info ctinfo;
1818         struct nf_conntrack_zone tmp;
1819         u32 hash, zone_id, rid;
1820         struct nf_conn *ct;
1821
1822         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1823                              dataoff, state->pf, protonum, state->net,
1824                              &tuple))
1825                 return 0;
1826
1827         /* look for tuple match */
1828         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1829
1830         zone_id = nf_ct_zone_id(zone, IP_CT_DIR_ORIGINAL);
1831         hash = hash_conntrack_raw(&tuple, zone_id, state->net);
1832         h = __nf_conntrack_find_get(state->net, zone, &tuple, hash);
1833
1834         if (!h) {
1835                 rid = nf_ct_zone_id(zone, IP_CT_DIR_REPLY);
1836                 if (zone_id != rid) {
1837                         u32 tmp = hash_conntrack_raw(&tuple, rid, state->net);
1838
1839                         h = __nf_conntrack_find_get(state->net, zone, &tuple, tmp);
1840                 }
1841         }
1842
1843         if (!h) {
1844                 h = init_conntrack(state->net, tmpl, &tuple,
1845                                    skb, dataoff, hash);
1846                 if (!h)
1847                         return 0;
1848                 if (IS_ERR(h))
1849                         return PTR_ERR(h);
1850         }
1851         ct = nf_ct_tuplehash_to_ctrack(h);
1852
1853         /* It exists; we have (non-exclusive) reference. */
1854         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1855                 ctinfo = IP_CT_ESTABLISHED_REPLY;
1856         } else {
1857                 unsigned long status = READ_ONCE(ct->status);
1858
1859                 /* Once we've had two way comms, always ESTABLISHED. */
1860                 if (likely(status & IPS_SEEN_REPLY))
1861                         ctinfo = IP_CT_ESTABLISHED;
1862                 else if (status & IPS_EXPECTED)
1863                         ctinfo = IP_CT_RELATED;
1864                 else
1865                         ctinfo = IP_CT_NEW;
1866         }
1867         nf_ct_set(skb, ct, ctinfo);
1868         return 0;
1869 }
1870
1871 /*
1872  * icmp packets need special treatment to handle error messages that are
1873  * related to a connection.
1874  *
1875  * Callers need to check if skb has a conntrack assigned when this
1876  * helper returns; in such case skb belongs to an already known connection.
1877  */
1878 static unsigned int __cold
1879 nf_conntrack_handle_icmp(struct nf_conn *tmpl,
1880                          struct sk_buff *skb,
1881                          unsigned int dataoff,
1882                          u8 protonum,
1883                          const struct nf_hook_state *state)
1884 {
1885         int ret;
1886
1887         if (state->pf == NFPROTO_IPV4 && protonum == IPPROTO_ICMP)
1888                 ret = nf_conntrack_icmpv4_error(tmpl, skb, dataoff, state);
1889 #if IS_ENABLED(CONFIG_IPV6)
1890         else if (state->pf == NFPROTO_IPV6 && protonum == IPPROTO_ICMPV6)
1891                 ret = nf_conntrack_icmpv6_error(tmpl, skb, dataoff, state);
1892 #endif
1893         else
1894                 return NF_ACCEPT;
1895
1896         if (ret <= 0)
1897                 NF_CT_STAT_INC_ATOMIC(state->net, error);
1898
1899         return ret;
1900 }
1901
1902 static int generic_packet(struct nf_conn *ct, struct sk_buff *skb,
1903                           enum ip_conntrack_info ctinfo)
1904 {
1905         const unsigned int *timeout = nf_ct_timeout_lookup(ct);
1906
1907         if (!timeout)
1908                 timeout = &nf_generic_pernet(nf_ct_net(ct))->timeout;
1909
1910         nf_ct_refresh_acct(ct, ctinfo, skb, *timeout);
1911         return NF_ACCEPT;
1912 }
1913
1914 /* Returns verdict for packet, or -1 for invalid. */
1915 static int nf_conntrack_handle_packet(struct nf_conn *ct,
1916                                       struct sk_buff *skb,
1917                                       unsigned int dataoff,
1918                                       enum ip_conntrack_info ctinfo,
1919                                       const struct nf_hook_state *state)
1920 {
1921         switch (nf_ct_protonum(ct)) {
1922         case IPPROTO_TCP:
1923                 return nf_conntrack_tcp_packet(ct, skb, dataoff,
1924                                                ctinfo, state);
1925         case IPPROTO_UDP:
1926                 return nf_conntrack_udp_packet(ct, skb, dataoff,
1927                                                ctinfo, state);
1928         case IPPROTO_ICMP:
1929                 return nf_conntrack_icmp_packet(ct, skb, ctinfo, state);
1930 #if IS_ENABLED(CONFIG_IPV6)
1931         case IPPROTO_ICMPV6:
1932                 return nf_conntrack_icmpv6_packet(ct, skb, ctinfo, state);
1933 #endif
1934 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
1935         case IPPROTO_UDPLITE:
1936                 return nf_conntrack_udplite_packet(ct, skb, dataoff,
1937                                                    ctinfo, state);
1938 #endif
1939 #ifdef CONFIG_NF_CT_PROTO_SCTP
1940         case IPPROTO_SCTP:
1941                 return nf_conntrack_sctp_packet(ct, skb, dataoff,
1942                                                 ctinfo, state);
1943 #endif
1944 #ifdef CONFIG_NF_CT_PROTO_DCCP
1945         case IPPROTO_DCCP:
1946                 return nf_conntrack_dccp_packet(ct, skb, dataoff,
1947                                                 ctinfo, state);
1948 #endif
1949 #ifdef CONFIG_NF_CT_PROTO_GRE
1950         case IPPROTO_GRE:
1951                 return nf_conntrack_gre_packet(ct, skb, dataoff,
1952                                                ctinfo, state);
1953 #endif
1954         }
1955
1956         return generic_packet(ct, skb, ctinfo);
1957 }
1958
1959 unsigned int
1960 nf_conntrack_in(struct sk_buff *skb, const struct nf_hook_state *state)
1961 {
1962         enum ip_conntrack_info ctinfo;
1963         struct nf_conn *ct, *tmpl;
1964         u_int8_t protonum;
1965         int dataoff, ret;
1966
1967         tmpl = nf_ct_get(skb, &ctinfo);
1968         if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1969                 /* Previously seen (loopback or untracked)?  Ignore. */
1970                 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1971                      ctinfo == IP_CT_UNTRACKED)
1972                         return NF_ACCEPT;
1973                 skb->_nfct = 0;
1974         }
1975
1976         /* rcu_read_lock()ed by nf_hook_thresh */
1977         dataoff = get_l4proto(skb, skb_network_offset(skb), state->pf, &protonum);
1978         if (dataoff <= 0) {
1979                 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1980                 ret = NF_ACCEPT;
1981                 goto out;
1982         }
1983
1984         if (protonum == IPPROTO_ICMP || protonum == IPPROTO_ICMPV6) {
1985                 ret = nf_conntrack_handle_icmp(tmpl, skb, dataoff,
1986                                                protonum, state);
1987                 if (ret <= 0) {
1988                         ret = -ret;
1989                         goto out;
1990                 }
1991                 /* ICMP[v6] protocol trackers may assign one conntrack. */
1992                 if (skb->_nfct)
1993                         goto out;
1994         }
1995 repeat:
1996         ret = resolve_normal_ct(tmpl, skb, dataoff,
1997                                 protonum, state);
1998         if (ret < 0) {
1999                 /* Too stressed to deal. */
2000                 NF_CT_STAT_INC_ATOMIC(state->net, drop);
2001                 ret = NF_DROP;
2002                 goto out;
2003         }
2004
2005         ct = nf_ct_get(skb, &ctinfo);
2006         if (!ct) {
2007                 /* Not valid part of a connection */
2008                 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
2009                 ret = NF_ACCEPT;
2010                 goto out;
2011         }
2012
2013         ret = nf_conntrack_handle_packet(ct, skb, dataoff, ctinfo, state);
2014         if (ret <= 0) {
2015                 /* Invalid: inverse of the return code tells
2016                  * the netfilter core what to do */
2017                 nf_ct_put(ct);
2018                 skb->_nfct = 0;
2019                 /* Special case: TCP tracker reports an attempt to reopen a
2020                  * closed/aborted connection. We have to go back and create a
2021                  * fresh conntrack.
2022                  */
2023                 if (ret == -NF_REPEAT)
2024                         goto repeat;
2025
2026                 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
2027                 if (ret == -NF_DROP)
2028                         NF_CT_STAT_INC_ATOMIC(state->net, drop);
2029
2030                 ret = -ret;
2031                 goto out;
2032         }
2033
2034         if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
2035             !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
2036                 nf_conntrack_event_cache(IPCT_REPLY, ct);
2037 out:
2038         if (tmpl)
2039                 nf_ct_put(tmpl);
2040
2041         return ret;
2042 }
2043 EXPORT_SYMBOL_GPL(nf_conntrack_in);
2044
2045 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
2046 void __nf_ct_refresh_acct(struct nf_conn *ct,
2047                           enum ip_conntrack_info ctinfo,
2048                           const struct sk_buff *skb,
2049                           u32 extra_jiffies,
2050                           bool do_acct)
2051 {
2052         /* Only update if this is not a fixed timeout */
2053         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
2054                 goto acct;
2055
2056         /* If not in hash table, timer will not be active yet */
2057         if (nf_ct_is_confirmed(ct))
2058                 extra_jiffies += nfct_time_stamp;
2059
2060         if (READ_ONCE(ct->timeout) != extra_jiffies)
2061                 WRITE_ONCE(ct->timeout, extra_jiffies);
2062 acct:
2063         if (do_acct)
2064                 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
2065 }
2066 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
2067
2068 bool nf_ct_kill_acct(struct nf_conn *ct,
2069                      enum ip_conntrack_info ctinfo,
2070                      const struct sk_buff *skb)
2071 {
2072         nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
2073
2074         return nf_ct_delete(ct, 0, 0);
2075 }
2076 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
2077
2078 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
2079
2080 #include <linux/netfilter/nfnetlink.h>
2081 #include <linux/netfilter/nfnetlink_conntrack.h>
2082 #include <linux/mutex.h>
2083
2084 /* Generic function for tcp/udp/sctp/dccp and alike. */
2085 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
2086                                const struct nf_conntrack_tuple *tuple)
2087 {
2088         if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
2089             nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
2090                 goto nla_put_failure;
2091         return 0;
2092
2093 nla_put_failure:
2094         return -1;
2095 }
2096 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
2097
2098 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
2099         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
2100         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
2101 };
2102 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
2103
2104 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
2105                                struct nf_conntrack_tuple *t,
2106                                u_int32_t flags)
2107 {
2108         if (flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) {
2109                 if (!tb[CTA_PROTO_SRC_PORT])
2110                         return -EINVAL;
2111
2112                 t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
2113         }
2114
2115         if (flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) {
2116                 if (!tb[CTA_PROTO_DST_PORT])
2117                         return -EINVAL;
2118
2119                 t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
2120         }
2121
2122         return 0;
2123 }
2124 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
2125
2126 unsigned int nf_ct_port_nlattr_tuple_size(void)
2127 {
2128         static unsigned int size __read_mostly;
2129
2130         if (!size)
2131                 size = nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
2132
2133         return size;
2134 }
2135 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
2136 #endif
2137
2138 /* Used by ipt_REJECT and ip6t_REJECT. */
2139 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
2140 {
2141         struct nf_conn *ct;
2142         enum ip_conntrack_info ctinfo;
2143
2144         /* This ICMP is in reverse direction to the packet which caused it */
2145         ct = nf_ct_get(skb, &ctinfo);
2146         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
2147                 ctinfo = IP_CT_RELATED_REPLY;
2148         else
2149                 ctinfo = IP_CT_RELATED;
2150
2151         /* Attach to new skbuff, and increment count */
2152         nf_ct_set(nskb, ct, ctinfo);
2153         nf_conntrack_get(skb_nfct(nskb));
2154 }
2155
2156 static int __nf_conntrack_update(struct net *net, struct sk_buff *skb,
2157                                  struct nf_conn *ct,
2158                                  enum ip_conntrack_info ctinfo)
2159 {
2160         const struct nf_nat_hook *nat_hook;
2161         struct nf_conntrack_tuple_hash *h;
2162         struct nf_conntrack_tuple tuple;
2163         unsigned int status;
2164         int dataoff;
2165         u16 l3num;
2166         u8 l4num;
2167
2168         l3num = nf_ct_l3num(ct);
2169
2170         dataoff = get_l4proto(skb, skb_network_offset(skb), l3num, &l4num);
2171         if (dataoff <= 0)
2172                 return NF_DROP;
2173
2174         if (!nf_ct_get_tuple(skb, skb_network_offset(skb), dataoff, l3num,
2175                              l4num, net, &tuple))
2176                 return NF_DROP;
2177
2178         if (ct->status & IPS_SRC_NAT) {
2179                 memcpy(tuple.src.u3.all,
2180                        ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.all,
2181                        sizeof(tuple.src.u3.all));
2182                 tuple.src.u.all =
2183                         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u.all;
2184         }
2185
2186         if (ct->status & IPS_DST_NAT) {
2187                 memcpy(tuple.dst.u3.all,
2188                        ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.all,
2189                        sizeof(tuple.dst.u3.all));
2190                 tuple.dst.u.all =
2191                         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u.all;
2192         }
2193
2194         h = nf_conntrack_find_get(net, nf_ct_zone(ct), &tuple);
2195         if (!h)
2196                 return NF_ACCEPT;
2197
2198         /* Store status bits of the conntrack that is clashing to re-do NAT
2199          * mangling according to what it has been done already to this packet.
2200          */
2201         status = ct->status;
2202
2203         nf_ct_put(ct);
2204         ct = nf_ct_tuplehash_to_ctrack(h);
2205         nf_ct_set(skb, ct, ctinfo);
2206
2207         nat_hook = rcu_dereference(nf_nat_hook);
2208         if (!nat_hook)
2209                 return NF_ACCEPT;
2210
2211         if (status & IPS_SRC_NAT) {
2212                 unsigned int verdict = nat_hook->manip_pkt(skb, ct,
2213                                                            NF_NAT_MANIP_SRC,
2214                                                            IP_CT_DIR_ORIGINAL);
2215                 if (verdict != NF_ACCEPT)
2216                         return verdict;
2217         }
2218
2219         if (status & IPS_DST_NAT) {
2220                 unsigned int verdict = nat_hook->manip_pkt(skb, ct,
2221                                                            NF_NAT_MANIP_DST,
2222                                                            IP_CT_DIR_ORIGINAL);
2223                 if (verdict != NF_ACCEPT)
2224                         return verdict;
2225         }
2226
2227         return NF_ACCEPT;
2228 }
2229
2230 /* This packet is coming from userspace via nf_queue, complete the packet
2231  * processing after the helper invocation in nf_confirm().
2232  */
2233 static int nf_confirm_cthelper(struct sk_buff *skb, struct nf_conn *ct,
2234                                enum ip_conntrack_info ctinfo)
2235 {
2236         const struct nf_conntrack_helper *helper;
2237         const struct nf_conn_help *help;
2238         int protoff;
2239
2240         help = nfct_help(ct);
2241         if (!help)
2242                 return NF_ACCEPT;
2243
2244         helper = rcu_dereference(help->helper);
2245         if (!helper)
2246                 return NF_ACCEPT;
2247
2248         if (!(helper->flags & NF_CT_HELPER_F_USERSPACE))
2249                 return NF_ACCEPT;
2250
2251         switch (nf_ct_l3num(ct)) {
2252         case NFPROTO_IPV4:
2253                 protoff = skb_network_offset(skb) + ip_hdrlen(skb);
2254                 break;
2255 #if IS_ENABLED(CONFIG_IPV6)
2256         case NFPROTO_IPV6: {
2257                 __be16 frag_off;
2258                 u8 pnum;
2259
2260                 pnum = ipv6_hdr(skb)->nexthdr;
2261                 protoff = ipv6_skip_exthdr(skb, sizeof(struct ipv6hdr), &pnum,
2262                                            &frag_off);
2263                 if (protoff < 0 || (frag_off & htons(~0x7)) != 0)
2264                         return NF_ACCEPT;
2265                 break;
2266         }
2267 #endif
2268         default:
2269                 return NF_ACCEPT;
2270         }
2271
2272         if (test_bit(IPS_SEQ_ADJUST_BIT, &ct->status) &&
2273             !nf_is_loopback_packet(skb)) {
2274                 if (!nf_ct_seq_adjust(skb, ct, ctinfo, protoff)) {
2275                         NF_CT_STAT_INC_ATOMIC(nf_ct_net(ct), drop);
2276                         return NF_DROP;
2277                 }
2278         }
2279
2280         /* We've seen it coming out the other side: confirm it */
2281         return nf_conntrack_confirm(skb);
2282 }
2283
2284 static int nf_conntrack_update(struct net *net, struct sk_buff *skb)
2285 {
2286         enum ip_conntrack_info ctinfo;
2287         struct nf_conn *ct;
2288
2289         ct = nf_ct_get(skb, &ctinfo);
2290         if (!ct)
2291                 return NF_ACCEPT;
2292
2293         if (!nf_ct_is_confirmed(ct)) {
2294                 int ret = __nf_conntrack_update(net, skb, ct, ctinfo);
2295
2296                 if (ret != NF_ACCEPT)
2297                         return ret;
2298
2299                 ct = nf_ct_get(skb, &ctinfo);
2300                 if (!ct)
2301                         return NF_ACCEPT;
2302         }
2303
2304         return nf_confirm_cthelper(skb, ct, ctinfo);
2305 }
2306
2307 static bool nf_conntrack_get_tuple_skb(struct nf_conntrack_tuple *dst_tuple,
2308                                        const struct sk_buff *skb)
2309 {
2310         const struct nf_conntrack_tuple *src_tuple;
2311         const struct nf_conntrack_tuple_hash *hash;
2312         struct nf_conntrack_tuple srctuple;
2313         enum ip_conntrack_info ctinfo;
2314         struct nf_conn *ct;
2315
2316         ct = nf_ct_get(skb, &ctinfo);
2317         if (ct) {
2318                 src_tuple = nf_ct_tuple(ct, CTINFO2DIR(ctinfo));
2319                 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2320                 return true;
2321         }
2322
2323         if (!nf_ct_get_tuplepr(skb, skb_network_offset(skb),
2324                                NFPROTO_IPV4, dev_net(skb->dev),
2325                                &srctuple))
2326                 return false;
2327
2328         hash = nf_conntrack_find_get(dev_net(skb->dev),
2329                                      &nf_ct_zone_dflt,
2330                                      &srctuple);
2331         if (!hash)
2332                 return false;
2333
2334         ct = nf_ct_tuplehash_to_ctrack(hash);
2335         src_tuple = nf_ct_tuple(ct, !hash->tuple.dst.dir);
2336         memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2337         nf_ct_put(ct);
2338
2339         return true;
2340 }
2341
2342 /* Bring out ya dead! */
2343 static struct nf_conn *
2344 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
2345                 const struct nf_ct_iter_data *iter_data, unsigned int *bucket)
2346 {
2347         struct nf_conntrack_tuple_hash *h;
2348         struct nf_conn *ct;
2349         struct hlist_nulls_node *n;
2350         spinlock_t *lockp;
2351
2352         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
2353                 struct hlist_nulls_head *hslot = &nf_conntrack_hash[*bucket];
2354
2355                 if (hlist_nulls_empty(hslot))
2356                         continue;
2357
2358                 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
2359                 local_bh_disable();
2360                 nf_conntrack_lock(lockp);
2361                 hlist_nulls_for_each_entry(h, n, hslot, hnnode) {
2362                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
2363                                 continue;
2364                         /* All nf_conn objects are added to hash table twice, one
2365                          * for original direction tuple, once for the reply tuple.
2366                          *
2367                          * Exception: In the IPS_NAT_CLASH case, only the reply
2368                          * tuple is added (the original tuple already existed for
2369                          * a different object).
2370                          *
2371                          * We only need to call the iterator once for each
2372                          * conntrack, so we just use the 'reply' direction
2373                          * tuple while iterating.
2374                          */
2375                         ct = nf_ct_tuplehash_to_ctrack(h);
2376
2377                         if (iter_data->net &&
2378                             !net_eq(iter_data->net, nf_ct_net(ct)))
2379                                 continue;
2380
2381                         if (iter(ct, iter_data->data))
2382                                 goto found;
2383                 }
2384                 spin_unlock(lockp);
2385                 local_bh_enable();
2386                 cond_resched();
2387         }
2388
2389         return NULL;
2390 found:
2391         refcount_inc(&ct->ct_general.use);
2392         spin_unlock(lockp);
2393         local_bh_enable();
2394         return ct;
2395 }
2396
2397 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
2398                                   const struct nf_ct_iter_data *iter_data)
2399 {
2400         unsigned int bucket = 0;
2401         struct nf_conn *ct;
2402
2403         might_sleep();
2404
2405         mutex_lock(&nf_conntrack_mutex);
2406         while ((ct = get_next_corpse(iter, iter_data, &bucket)) != NULL) {
2407                 /* Time to push up daises... */
2408
2409                 nf_ct_delete(ct, iter_data->portid, iter_data->report);
2410                 nf_ct_put(ct);
2411                 cond_resched();
2412         }
2413         mutex_unlock(&nf_conntrack_mutex);
2414 }
2415
2416 void nf_ct_iterate_cleanup_net(int (*iter)(struct nf_conn *i, void *data),
2417                                const struct nf_ct_iter_data *iter_data)
2418 {
2419         struct net *net = iter_data->net;
2420         struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2421
2422         might_sleep();
2423
2424         if (atomic_read(&cnet->count) == 0)
2425                 return;
2426
2427         nf_ct_iterate_cleanup(iter, iter_data);
2428 }
2429 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
2430
2431 /**
2432  * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
2433  * @iter: callback to invoke for each conntrack
2434  * @data: data to pass to @iter
2435  *
2436  * Like nf_ct_iterate_cleanup, but first marks conntracks on the
2437  * unconfirmed list as dying (so they will not be inserted into
2438  * main table).
2439  *
2440  * Can only be called in module exit path.
2441  */
2442 void
2443 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
2444 {
2445         struct nf_ct_iter_data iter_data = {};
2446         struct net *net;
2447
2448         down_read(&net_rwsem);
2449         for_each_net(net) {
2450                 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2451
2452                 if (atomic_read(&cnet->count) == 0)
2453                         continue;
2454                 nf_queue_nf_hook_drop(net);
2455         }
2456         up_read(&net_rwsem);
2457
2458         /* Need to wait for netns cleanup worker to finish, if its
2459          * running -- it might have deleted a net namespace from
2460          * the global list, so hook drop above might not have
2461          * affected all namespaces.
2462          */
2463         net_ns_barrier();
2464
2465         /* a skb w. unconfirmed conntrack could have been reinjected just
2466          * before we called nf_queue_nf_hook_drop().
2467          *
2468          * This makes sure its inserted into conntrack table.
2469          */
2470         synchronize_net();
2471
2472         nf_ct_ext_bump_genid();
2473         iter_data.data = data;
2474         nf_ct_iterate_cleanup(iter, &iter_data);
2475
2476         /* Another cpu might be in a rcu read section with
2477          * rcu protected pointer cleared in iter callback
2478          * or hidden via nf_ct_ext_bump_genid() above.
2479          *
2480          * Wait until those are done.
2481          */
2482         synchronize_rcu();
2483 }
2484 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
2485
2486 static int kill_all(struct nf_conn *i, void *data)
2487 {
2488         return 1;
2489 }
2490
2491 void nf_conntrack_cleanup_start(void)
2492 {
2493         cleanup_nf_conntrack_bpf();
2494         conntrack_gc_work.exiting = true;
2495 }
2496
2497 void nf_conntrack_cleanup_end(void)
2498 {
2499         RCU_INIT_POINTER(nf_ct_hook, NULL);
2500         cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2501         kvfree(nf_conntrack_hash);
2502
2503         nf_conntrack_proto_fini();
2504         nf_conntrack_helper_fini();
2505         nf_conntrack_expect_fini();
2506
2507         kmem_cache_destroy(nf_conntrack_cachep);
2508 }
2509
2510 /*
2511  * Mishearing the voices in his head, our hero wonders how he's
2512  * supposed to kill the mall.
2513  */
2514 void nf_conntrack_cleanup_net(struct net *net)
2515 {
2516         LIST_HEAD(single);
2517
2518         list_add(&net->exit_list, &single);
2519         nf_conntrack_cleanup_net_list(&single);
2520 }
2521
2522 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
2523 {
2524         struct nf_ct_iter_data iter_data = {};
2525         struct net *net;
2526         int busy;
2527
2528         /*
2529          * This makes sure all current packets have passed through
2530          *  netfilter framework.  Roll on, two-stage module
2531          *  delete...
2532          */
2533         synchronize_rcu_expedited();
2534 i_see_dead_people:
2535         busy = 0;
2536         list_for_each_entry(net, net_exit_list, exit_list) {
2537                 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2538
2539                 iter_data.net = net;
2540                 nf_ct_iterate_cleanup_net(kill_all, &iter_data);
2541                 if (atomic_read(&cnet->count) != 0)
2542                         busy = 1;
2543         }
2544         if (busy) {
2545                 schedule();
2546                 goto i_see_dead_people;
2547         }
2548
2549         list_for_each_entry(net, net_exit_list, exit_list) {
2550                 nf_conntrack_ecache_pernet_fini(net);
2551                 nf_conntrack_expect_pernet_fini(net);
2552                 free_percpu(net->ct.stat);
2553         }
2554 }
2555
2556 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
2557 {
2558         struct hlist_nulls_head *hash;
2559         unsigned int nr_slots, i;
2560
2561         if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
2562                 return NULL;
2563
2564         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
2565         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
2566
2567         hash = kvcalloc(nr_slots, sizeof(struct hlist_nulls_head), GFP_KERNEL);
2568
2569         if (hash && nulls)
2570                 for (i = 0; i < nr_slots; i++)
2571                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
2572
2573         return hash;
2574 }
2575 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
2576
2577 int nf_conntrack_hash_resize(unsigned int hashsize)
2578 {
2579         int i, bucket;
2580         unsigned int old_size;
2581         struct hlist_nulls_head *hash, *old_hash;
2582         struct nf_conntrack_tuple_hash *h;
2583         struct nf_conn *ct;
2584
2585         if (!hashsize)
2586                 return -EINVAL;
2587
2588         hash = nf_ct_alloc_hashtable(&hashsize, 1);
2589         if (!hash)
2590                 return -ENOMEM;
2591
2592         mutex_lock(&nf_conntrack_mutex);
2593         old_size = nf_conntrack_htable_size;
2594         if (old_size == hashsize) {
2595                 mutex_unlock(&nf_conntrack_mutex);
2596                 kvfree(hash);
2597                 return 0;
2598         }
2599
2600         local_bh_disable();
2601         nf_conntrack_all_lock();
2602         write_seqcount_begin(&nf_conntrack_generation);
2603
2604         /* Lookups in the old hash might happen in parallel, which means we
2605          * might get false negatives during connection lookup. New connections
2606          * created because of a false negative won't make it into the hash
2607          * though since that required taking the locks.
2608          */
2609
2610         for (i = 0; i < nf_conntrack_htable_size; i++) {
2611                 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
2612                         unsigned int zone_id;
2613
2614                         h = hlist_nulls_entry(nf_conntrack_hash[i].first,
2615                                               struct nf_conntrack_tuple_hash, hnnode);
2616                         ct = nf_ct_tuplehash_to_ctrack(h);
2617                         hlist_nulls_del_rcu(&h->hnnode);
2618
2619                         zone_id = nf_ct_zone_id(nf_ct_zone(ct), NF_CT_DIRECTION(h));
2620                         bucket = __hash_conntrack(nf_ct_net(ct),
2621                                                   &h->tuple, zone_id, hashsize);
2622                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
2623                 }
2624         }
2625         old_hash = nf_conntrack_hash;
2626
2627         nf_conntrack_hash = hash;
2628         nf_conntrack_htable_size = hashsize;
2629
2630         write_seqcount_end(&nf_conntrack_generation);
2631         nf_conntrack_all_unlock();
2632         local_bh_enable();
2633
2634         mutex_unlock(&nf_conntrack_mutex);
2635
2636         synchronize_net();
2637         kvfree(old_hash);
2638         return 0;
2639 }
2640
2641 int nf_conntrack_set_hashsize(const char *val, const struct kernel_param *kp)
2642 {
2643         unsigned int hashsize;
2644         int rc;
2645
2646         if (current->nsproxy->net_ns != &init_net)
2647                 return -EOPNOTSUPP;
2648
2649         /* On boot, we can set this without any fancy locking. */
2650         if (!nf_conntrack_hash)
2651                 return param_set_uint(val, kp);
2652
2653         rc = kstrtouint(val, 0, &hashsize);
2654         if (rc)
2655                 return rc;
2656
2657         return nf_conntrack_hash_resize(hashsize);
2658 }
2659
2660 int nf_conntrack_init_start(void)
2661 {
2662         unsigned long nr_pages = totalram_pages();
2663         int max_factor = 8;
2664         int ret = -ENOMEM;
2665         int i;
2666
2667         seqcount_spinlock_init(&nf_conntrack_generation,
2668                                &nf_conntrack_locks_all_lock);
2669
2670         for (i = 0; i < CONNTRACK_LOCKS; i++)
2671                 spin_lock_init(&nf_conntrack_locks[i]);
2672
2673         if (!nf_conntrack_htable_size) {
2674                 nf_conntrack_htable_size
2675                         = (((nr_pages << PAGE_SHIFT) / 16384)
2676                            / sizeof(struct hlist_head));
2677                 if (BITS_PER_LONG >= 64 &&
2678                     nr_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2679                         nf_conntrack_htable_size = 262144;
2680                 else if (nr_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2681                         nf_conntrack_htable_size = 65536;
2682
2683                 if (nf_conntrack_htable_size < 1024)
2684                         nf_conntrack_htable_size = 1024;
2685                 /* Use a max. factor of one by default to keep the average
2686                  * hash chain length at 2 entries.  Each entry has to be added
2687                  * twice (once for original direction, once for reply).
2688                  * When a table size is given we use the old value of 8 to
2689                  * avoid implicit reduction of the max entries setting.
2690                  */
2691                 max_factor = 1;
2692         }
2693
2694         nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2695         if (!nf_conntrack_hash)
2696                 return -ENOMEM;
2697
2698         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2699
2700         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2701                                                 sizeof(struct nf_conn),
2702                                                 NFCT_INFOMASK + 1,
2703                                                 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2704         if (!nf_conntrack_cachep)
2705                 goto err_cachep;
2706
2707         ret = nf_conntrack_expect_init();
2708         if (ret < 0)
2709                 goto err_expect;
2710
2711         ret = nf_conntrack_helper_init();
2712         if (ret < 0)
2713                 goto err_helper;
2714
2715         ret = nf_conntrack_proto_init();
2716         if (ret < 0)
2717                 goto err_proto;
2718
2719         conntrack_gc_work_init(&conntrack_gc_work);
2720         queue_delayed_work(system_power_efficient_wq, &conntrack_gc_work.dwork, HZ);
2721
2722         ret = register_nf_conntrack_bpf();
2723         if (ret < 0)
2724                 goto err_kfunc;
2725
2726         return 0;
2727
2728 err_kfunc:
2729         cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2730         nf_conntrack_proto_fini();
2731 err_proto:
2732         nf_conntrack_helper_fini();
2733 err_helper:
2734         nf_conntrack_expect_fini();
2735 err_expect:
2736         kmem_cache_destroy(nf_conntrack_cachep);
2737 err_cachep:
2738         kvfree(nf_conntrack_hash);
2739         return ret;
2740 }
2741
2742 static void nf_conntrack_set_closing(struct nf_conntrack *nfct)
2743 {
2744         struct nf_conn *ct = nf_ct_to_nf_conn(nfct);
2745
2746         switch (nf_ct_protonum(ct)) {
2747         case IPPROTO_TCP:
2748                 nf_conntrack_tcp_set_closing(ct);
2749                 break;
2750         }
2751 }
2752
2753 static const struct nf_ct_hook nf_conntrack_hook = {
2754         .update         = nf_conntrack_update,
2755         .destroy        = nf_ct_destroy,
2756         .get_tuple_skb  = nf_conntrack_get_tuple_skb,
2757         .attach         = nf_conntrack_attach,
2758         .set_closing    = nf_conntrack_set_closing,
2759         .confirm        = __nf_conntrack_confirm,
2760 };
2761
2762 void nf_conntrack_init_end(void)
2763 {
2764         RCU_INIT_POINTER(nf_ct_hook, &nf_conntrack_hook);
2765 }
2766
2767 /*
2768  * We need to use special "null" values, not used in hash table
2769  */
2770 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
2771
2772 int nf_conntrack_init_net(struct net *net)
2773 {
2774         struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2775         int ret = -ENOMEM;
2776
2777         BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2778         BUILD_BUG_ON_NOT_POWER_OF_2(CONNTRACK_LOCKS);
2779         atomic_set(&cnet->count, 0);
2780
2781         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2782         if (!net->ct.stat)
2783                 return ret;
2784
2785         ret = nf_conntrack_expect_pernet_init(net);
2786         if (ret < 0)
2787                 goto err_expect;
2788
2789         nf_conntrack_acct_pernet_init(net);
2790         nf_conntrack_tstamp_pernet_init(net);
2791         nf_conntrack_ecache_pernet_init(net);
2792         nf_conntrack_proto_pernet_init(net);
2793
2794         return 0;
2795
2796 err_expect:
2797         free_percpu(net->ct.stat);
2798         return ret;
2799 }
2800
2801 /* ctnetlink code shared by both ctnetlink and nf_conntrack_bpf */
2802
2803 int __nf_ct_change_timeout(struct nf_conn *ct, u64 timeout)
2804 {
2805         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
2806                 return -EPERM;
2807
2808         __nf_ct_set_timeout(ct, timeout);
2809
2810         if (test_bit(IPS_DYING_BIT, &ct->status))
2811                 return -ETIME;
2812
2813         return 0;
2814 }
2815 EXPORT_SYMBOL_GPL(__nf_ct_change_timeout);
2816
2817 void __nf_ct_change_status(struct nf_conn *ct, unsigned long on, unsigned long off)
2818 {
2819         unsigned int bit;
2820
2821         /* Ignore these unchangable bits */
2822         on &= ~IPS_UNCHANGEABLE_MASK;
2823         off &= ~IPS_UNCHANGEABLE_MASK;
2824
2825         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
2826                 if (on & (1 << bit))
2827                         set_bit(bit, &ct->status);
2828                 else if (off & (1 << bit))
2829                         clear_bit(bit, &ct->status);
2830         }
2831 }
2832 EXPORT_SYMBOL_GPL(__nf_ct_change_status);
2833
2834 int nf_ct_change_status_common(struct nf_conn *ct, unsigned int status)
2835 {
2836         unsigned long d;
2837
2838         d = ct->status ^ status;
2839
2840         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
2841                 /* unchangeable */
2842                 return -EBUSY;
2843
2844         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2845                 /* SEEN_REPLY bit can only be set */
2846                 return -EBUSY;
2847
2848         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2849                 /* ASSURED bit can only be set */
2850                 return -EBUSY;
2851
2852         __nf_ct_change_status(ct, status, 0);
2853         return 0;
2854 }
2855 EXPORT_SYMBOL_GPL(nf_ct_change_status_common);