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