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