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/jhash.h>
  25 #include <linux/siphash.h>
  26 #include <linux/err.h>
  27 #include <linux/percpu.h>
  28 #include <linux/moduleparam.h>
  29 #include <linux/notifier.h>
  30 #include <linux/kernel.h>
  31 #include <linux/netdevice.h>
  32 #include <linux/socket.h>
  33 #include <linux/mm.h>
  34 #include <linux/nsproxy.h>
  35 #include <linux/rculist_nulls.h>
  36
  37 #include <net/netfilter/nf_conntrack.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_seqadj.h>
  42 #include <net/netfilter/nf_conntrack_core.h>
  43 #include <net/netfilter/nf_conntrack_extend.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_ecache.h>
  46 #include <net/netfilter/nf_conntrack_zones.h>
  47 #include <net/netfilter/nf_conntrack_timestamp.h>
  48 #include <net/netfilter/nf_conntrack_timeout.h>
  49 #include <net/netfilter/nf_conntrack_labels.h>
  50 #include <net/netfilter/nf_conntrack_synproxy.h>
  51 #include <net/netfilter/nf_nat.h>
  52 #include <net/netfilter/nf_nat_helper.h>
  53 #include <net/netns/hash.h>
  54 #include <net/ip.h>
  55
  56 #include "nf_internals.h"
  57
  58 __cacheline_aligned_in_smp spinlock_t nf_conntrack_locks[CONNTRACK_LOCKS];
  59 EXPORT_SYMBOL_GPL(nf_conntrack_locks);
  60
  61 __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
  62 EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
  63
  64 struct hlist_nulls_head *nf_conntrack_hash __read_mostly;
  65 EXPORT_SYMBOL_GPL(nf_conntrack_hash);
  66
  67 struct conntrack_gc_work {
  68         struct delayed_work     dwork;
  69         u32                     last_bucket;
  70         bool                    exiting;
  71         bool                    early_drop;
  72         long                    next_gc_run;
  73 };
  74
  75 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
  76 static DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
  77 static __read_mostly bool nf_conntrack_locks_all;
  78
  79 /* every gc cycle scans at most 1/GC_MAX_BUCKETS_DIV part of table */
  80 #define GC_MAX_BUCKETS_DIV      128u
  81 /* upper bound of full table scan */
  82 #define GC_MAX_SCAN_JIFFIES     (16u * HZ)
  83 /* desired ratio of entries found to be expired */
  84 #define GC_EVICT_RATIO  50u
  85
  86 static struct conntrack_gc_work conntrack_gc_work;
  87
  88 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
  89 {
  90         /* 1) Acquire the lock */
  91         spin_lock(lock);
  92
  93         /* 2) read nf_conntrack_locks_all, with ACQUIRE semantics
  94          * It pairs with the smp_store_release() in nf_conntrack_all_unlock()
  95          */
  96         if (likely(smp_load_acquire(&nf_conntrack_locks_all) == false))
  97                 return;
  98
  99         /* fast path failed, unlock */
 100         spin_unlock(lock);
 101
 102         /* Slow path 1) get global lock */
 103         spin_lock(&nf_conntrack_locks_all_lock);
 104
 105         /* Slow path 2) get the lock we want */
 106         spin_lock(lock);
 107
 108         /* Slow path 3) release the global lock */
 109         spin_unlock(&nf_conntrack_locks_all_lock);
 110 }
 111 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
 112
 113 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 114 {
 115         h1 %= CONNTRACK_LOCKS;
 116         h2 %= CONNTRACK_LOCKS;
 117         spin_unlock(&nf_conntrack_locks[h1]);
 118         if (h1 != h2)
 119                 spin_unlock(&nf_conntrack_locks[h2]);
 120 }
 121
 122 /* return true if we need to recompute hashes (in case hash table was resized) */
 123 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
 124                                      unsigned int h2, unsigned int sequence)
 125 {
 126         h1 %= CONNTRACK_LOCKS;
 127         h2 %= CONNTRACK_LOCKS;
 128         if (h1 <= h2) {
 129                 nf_conntrack_lock(&nf_conntrack_locks[h1]);
 130                 if (h1 != h2)
 131                         spin_lock_nested(&nf_conntrack_locks[h2],
 132                                          SINGLE_DEPTH_NESTING);
 133         } else {
 134                 nf_conntrack_lock(&nf_conntrack_locks[h2]);
 135                 spin_lock_nested(&nf_conntrack_locks[h1],
 136                                  SINGLE_DEPTH_NESTING);
 137         }
 138         if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
 139                 nf_conntrack_double_unlock(h1, h2);
 140                 return true;
 141         }
 142         return false;
 143 }
 144
 145 static void nf_conntrack_all_lock(void)
 146         __acquires(&nf_conntrack_locks_all_lock)
 147 {
 148         int i;
 149
 150         spin_lock(&nf_conntrack_locks_all_lock);
 151
 152         nf_conntrack_locks_all = true;
 153
 154         for (i = 0; i < CONNTRACK_LOCKS; i++) {
 155                 spin_lock(&nf_conntrack_locks[i]);
 156
 157                 /* This spin_unlock provides the "release" to ensure that
 158                  * nf_conntrack_locks_all==true is visible to everyone that
 159                  * acquired spin_lock(&nf_conntrack_locks[]).
 160                  */
 161                 spin_unlock(&nf_conntrack_locks[i]);
 162         }
 163 }
 164
 165 static void nf_conntrack_all_unlock(void)
 166         __releases(&nf_conntrack_locks_all_lock)
 167 {
 168         /* All prior stores must be complete before we clear
 169          * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
 170          * might observe the false value but not the entire
 171          * critical section.
 172          * It pairs with the smp_load_acquire() in nf_conntrack_lock()
 173          */
 174         smp_store_release(&nf_conntrack_locks_all, false);
 175         spin_unlock(&nf_conntrack_locks_all_lock);
 176 }
 177
 178 unsigned int nf_conntrack_htable_size __read_mostly;
 179 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
 180
 181 unsigned int nf_conntrack_max __read_mostly;
 182 EXPORT_SYMBOL_GPL(nf_conntrack_max);
 183 seqcount_spinlock_t nf_conntrack_generation __read_mostly;
 184 static unsigned int nf_conntrack_hash_rnd __read_mostly;
 185
 186 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
 187                               const struct net *net)
 188 {
 189         unsigned int n;
 190         u32 seed;
 191
 192         get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
 193
 194         /* The direction must be ignored, so we hash everything up to the
 195          * destination ports (which is a multiple of 4) and treat the last
 196          * three bytes manually.
 197          */
 198         seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
 199         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
 200         return jhash2((u32 *)tuple, n, seed ^
 201                       (((__force __u16)tuple->dst.u.all << 16) |
 202                       tuple->dst.protonum));
 203 }
 204
 205 static u32 scale_hash(u32 hash)
 206 {
 207         return reciprocal_scale(hash, nf_conntrack_htable_size);
 208 }
 209
 210 static u32 __hash_conntrack(const struct net *net,
 211                             const struct nf_conntrack_tuple *tuple,
 212                             unsigned int size)
 213 {
 214         return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
 215 }
 216
 217 static u32 hash_conntrack(const struct net *net,
 218                           const struct nf_conntrack_tuple *tuple)
 219 {
 220         return scale_hash(hash_conntrack_raw(tuple, net));
 221 }
 222
 223 static bool nf_ct_get_tuple_ports(const struct sk_buff *skb,
 224                                   unsigned int dataoff,
 225                                   struct nf_conntrack_tuple *tuple)
 226 {       struct {
 227                 __be16 sport;
 228                 __be16 dport;
 229         } _inet_hdr, *inet_hdr;
 230
 231         /* Actually only need first 4 bytes to get ports. */
 232         inet_hdr = skb_header_pointer(skb, dataoff, sizeof(_inet_hdr), &_inet_hdr);
 233         if (!inet_hdr)
 234                 return false;
 235
 236         tuple->src.u.udp.port = inet_hdr->sport;
 237         tuple->dst.u.udp.port = inet_hdr->dport;
 238         return true;
 239 }
 240
 241 static bool
 242 nf_ct_get_tuple(const struct sk_buff *skb,
 243                 unsigned int nhoff,
 244                 unsigned int dataoff,
 245                 u_int16_t l3num,
 246                 u_int8_t protonum,
 247                 struct net *net,
 248                 struct nf_conntrack_tuple *tuple)
 249 {
 250         unsigned int size;
 251         const __be32 *ap;
 252         __be32 _addrs[8];
 253
 254         memset(tuple, 0, sizeof(*tuple));
 255
 256         tuple->src.l3num = l3num;
 257         switch (l3num) {
 258         case NFPROTO_IPV4:
 259                 nhoff += offsetof(struct iphdr, saddr);
 260                 size = 2 * sizeof(__be32);
 261                 break;
 262         case NFPROTO_IPV6:
 263                 nhoff += offsetof(struct ipv6hdr, saddr);
 264                 size = sizeof(_addrs);
 265                 break;
 266         default:
 267                 return true;
 268         }
 269
 270         ap = skb_header_pointer(skb, nhoff, size, _addrs);
 271         if (!ap)
 272                 return false;
 273
 274         switch (l3num) {
 275         case NFPROTO_IPV4:
 276                 tuple->src.u3.ip = ap[0];
 277                 tuple->dst.u3.ip = ap[1];
 278                 break;
 279         case NFPROTO_IPV6:
 280                 memcpy(tuple->src.u3.ip6, ap, sizeof(tuple->src.u3.ip6));
 281                 memcpy(tuple->dst.u3.ip6, ap + 4, sizeof(tuple->dst.u3.ip6));
 282                 break;
 283         }
 284
 285         tuple->dst.protonum = protonum;
 286         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 287
 288         switch (protonum) {
 289 #if IS_ENABLED(CONFIG_IPV6)
 290         case IPPROTO_ICMPV6:
 291                 return icmpv6_pkt_to_tuple(skb, dataoff, net, tuple);
 292 #endif
 293         case IPPROTO_ICMP:
 294                 return icmp_pkt_to_tuple(skb, dataoff, net, tuple);
 295 #ifdef CONFIG_NF_CT_PROTO_GRE
 296         case IPPROTO_GRE:
 297                 return gre_pkt_to_tuple(skb, dataoff, net, tuple);
 298 #endif
 299         case IPPROTO_TCP:
 300         case IPPROTO_UDP: /* fallthrough */
 301                 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
 302 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
 303         case IPPROTO_UDPLITE:
 304                 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
 305 #endif
 306 #ifdef CONFIG_NF_CT_PROTO_SCTP
 307         case IPPROTO_SCTP:
 308                 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
 309 #endif
 310 #ifdef CONFIG_NF_CT_PROTO_DCCP
 311         case IPPROTO_DCCP:
 312                 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
 313 #endif
 314         default:
 315                 break;
 316         }
 317
 318         return true;
 319 }
 320
 321 static int ipv4_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
 322                             u_int8_t *protonum)
 323 {
 324         int dataoff = -1;
 325         const struct iphdr *iph;
 326         struct iphdr _iph;
 327
 328         iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
 329         if (!iph)
 330                 return -1;
 331
 332         /* Conntrack defragments packets, we might still see fragments
 333          * inside ICMP packets though.
 334          */
 335         if (iph->frag_off & htons(IP_OFFSET))
 336                 return -1;
 337
 338         dataoff = nhoff + (iph->ihl << 2);
 339         *protonum = iph->protocol;
 340
 341         /* Check bogus IP headers */
 342         if (dataoff > skb->len) {
 343                 pr_debug("bogus IPv4 packet: nhoff %u, ihl %u, skblen %u\n",
 344                          nhoff, iph->ihl << 2, skb->len);
 345                 return -1;
 346         }
 347         return dataoff;
 348 }
 349
 350 #if IS_ENABLED(CONFIG_IPV6)
 351 static int ipv6_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
 352                             u8 *protonum)
 353 {
 354         int protoff = -1;
 355         unsigned int extoff = nhoff + sizeof(struct ipv6hdr);
 356         __be16 frag_off;
 357         u8 nexthdr;
 358
 359         if (skb_copy_bits(skb, nhoff + offsetof(struct ipv6hdr, nexthdr),
 360                           &nexthdr, sizeof(nexthdr)) != 0) {
 361                 pr_debug("can't get nexthdr\n");
 362                 return -1;
 363         }
 364         protoff = ipv6_skip_exthdr(skb, extoff, &nexthdr, &frag_off);
 365         /*
 366          * (protoff == skb->len) means the packet has not data, just
 367          * IPv6 and possibly extensions headers, but it is tracked anyway
 368          */
 369         if (protoff < 0 || (frag_off & htons(~0x7)) != 0) {
 370                 pr_debug("can't find proto in pkt\n");
 371                 return -1;
 372         }
 373
 374         *protonum = nexthdr;
 375         return protoff;
 376 }
 377 #endif
 378
 379 static int get_l4proto(const struct sk_buff *skb,
 380                        unsigned int nhoff, u8 pf, u8 *l4num)
 381 {
 382         switch (pf) {
 383         case NFPROTO_IPV4:
 384                 return ipv4_get_l4proto(skb, nhoff, l4num);
 385 #if IS_ENABLED(CONFIG_IPV6)
 386         case NFPROTO_IPV6:
 387                 return ipv6_get_l4proto(skb, nhoff, l4num);
 388 #endif
 389         default:
 390                 *l4num = 0;
 391                 break;
 392         }
 393         return -1;
 394 }
 395
 396 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
 397                        u_int16_t l3num,
 398                        struct net *net, struct nf_conntrack_tuple *tuple)
 399 {
 400         u8 protonum;
 401         int protoff;
 402
 403         protoff = get_l4proto(skb, nhoff, l3num, &protonum);
 404         if (protoff <= 0)
 405                 return false;
 406
 407         return nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple);
 408 }
 409 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
 410
 411 bool
 412 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
 413                    const struct nf_conntrack_tuple *orig)
 414 {
 415         memset(inverse, 0, sizeof(*inverse));
 416
 417         inverse->src.l3num = orig->src.l3num;
 418
 419         switch (orig->src.l3num) {
 420         case NFPROTO_IPV4:
 421                 inverse->src.u3.ip = orig->dst.u3.ip;
 422                 inverse->dst.u3.ip = orig->src.u3.ip;
 423                 break;
 424         case NFPROTO_IPV6:
 425                 inverse->src.u3.in6 = orig->dst.u3.in6;
 426                 inverse->dst.u3.in6 = orig->src.u3.in6;
 427                 break;
 428         default:
 429                 break;
 430         }
 431
 432         inverse->dst.dir = !orig->dst.dir;
 433
 434         inverse->dst.protonum = orig->dst.protonum;
 435
 436         switch (orig->dst.protonum) {
 437         case IPPROTO_ICMP:
 438                 return nf_conntrack_invert_icmp_tuple(inverse, orig);
 439 #if IS_ENABLED(CONFIG_IPV6)
 440         case IPPROTO_ICMPV6:
 441                 return nf_conntrack_invert_icmpv6_tuple(inverse, orig);
 442 #endif
 443         }
 444
 445         inverse->src.u.all = orig->dst.u.all;
 446         inverse->dst.u.all = orig->src.u.all;
 447         return true;
 448 }
 449 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
 450
 451 /* Generate a almost-unique pseudo-id for a given conntrack.
 452  *
 453  * intentionally doesn't re-use any of the seeds used for hash
 454  * table location, we assume id gets exposed to userspace.
 455  *
 456  * Following nf_conn items do not change throughout lifetime
 457  * of the nf_conn:
 458  *
 459  * 1. nf_conn address
 460  * 2. nf_conn->master address (normally NULL)
 461  * 3. the associated net namespace
 462  * 4. the original direction tuple
 463  */
 464 u32 nf_ct_get_id(const struct nf_conn *ct)
 465 {
 466         static __read_mostly siphash_key_t ct_id_seed;
 467         unsigned long a, b, c, d;
 468
 469         net_get_random_once(&ct_id_seed, sizeof(ct_id_seed));
 470
 471         a = (unsigned long)ct;
 472         b = (unsigned long)ct->master;
 473         c = (unsigned long)nf_ct_net(ct);
 474         d = (unsigned long)siphash(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 475                                    sizeof(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple),
 476                                    &ct_id_seed);
 477 #ifdef CONFIG_64BIT
 478         return siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &ct_id_seed);
 479 #else
 480         return siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &ct_id_seed);
 481 #endif
 482 }
 483 EXPORT_SYMBOL_GPL(nf_ct_get_id);
 484
 485 static void
 486 clean_from_lists(struct nf_conn *ct)
 487 {
 488         pr_debug("clean_from_lists(%p)\n", ct);
 489         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 490         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
 491
 492         /* Destroy all pending expectations */
 493         nf_ct_remove_expectations(ct);
 494 }
 495
 496 /* must be called with local_bh_disable */
 497 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
 498 {
 499         struct ct_pcpu *pcpu;
 500
 501         /* add this conntrack to the (per cpu) dying list */
 502         ct->cpu = smp_processor_id();
 503         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 504
 505         spin_lock(&pcpu->lock);
 506         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 507                              &pcpu->dying);
 508         spin_unlock(&pcpu->lock);
 509 }
 510
 511 /* must be called with local_bh_disable */
 512 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
 513 {
 514         struct ct_pcpu *pcpu;
 515
 516         /* add this conntrack to the (per cpu) unconfirmed list */
 517         ct->cpu = smp_processor_id();
 518         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 519
 520         spin_lock(&pcpu->lock);
 521         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 522                              &pcpu->unconfirmed);
 523         spin_unlock(&pcpu->lock);
 524 }
 525
 526 /* must be called with local_bh_disable */
 527 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
 528 {
 529         struct ct_pcpu *pcpu;
 530
 531         /* We overload first tuple to link into unconfirmed or dying list.*/
 532         pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
 533
 534         spin_lock(&pcpu->lock);
 535         BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
 536         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 537         spin_unlock(&pcpu->lock);
 538 }
 539
 540 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
 541
 542 /* Released via destroy_conntrack() */
 543 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
 544                                  const struct nf_conntrack_zone *zone,
 545                                  gfp_t flags)
 546 {
 547         struct nf_conn *tmpl, *p;
 548
 549         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
 550                 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
 551                 if (!tmpl)
 552                         return NULL;
 553
 554                 p = tmpl;
 555                 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 556                 if (tmpl != p) {
 557                         tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
 558                         tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
 559                 }
 560         } else {
 561                 tmpl = kzalloc(sizeof(*tmpl), flags);
 562                 if (!tmpl)
 563                         return NULL;
 564         }
 565
 566         tmpl->status = IPS_TEMPLATE;
 567         write_pnet(&tmpl->ct_net, net);
 568         nf_ct_zone_add(tmpl, zone);
 569         atomic_set(&tmpl->ct_general.use, 0);
 570
 571         return tmpl;
 572 }
 573 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
 574
 575 void nf_ct_tmpl_free(struct nf_conn *tmpl)
 576 {
 577         nf_ct_ext_destroy(tmpl);
 578
 579         if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
 580                 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
 581         else
 582                 kfree(tmpl);
 583 }
 584 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
 585
 586 static void destroy_gre_conntrack(struct nf_conn *ct)
 587 {
 588 #ifdef CONFIG_NF_CT_PROTO_GRE
 589         struct nf_conn *master = ct->master;
 590
 591         if (master)
 592                 nf_ct_gre_keymap_destroy(master);
 593 #endif
 594 }
 595
 596 static void
 597 destroy_conntrack(struct nf_conntrack *nfct)
 598 {
 599         struct nf_conn *ct = (struct nf_conn *)nfct;
 600
 601         pr_debug("destroy_conntrack(%p)\n", ct);
 602         WARN_ON(atomic_read(&nfct->use) != 0);
 603
 604         if (unlikely(nf_ct_is_template(ct))) {
 605                 nf_ct_tmpl_free(ct);
 606                 return;
 607         }
 608
 609         if (unlikely(nf_ct_protonum(ct) == IPPROTO_GRE))
 610                 destroy_gre_conntrack(ct);
 611
 612         local_bh_disable();
 613         /* Expectations will have been removed in clean_from_lists,
 614          * except TFTP can create an expectation on the first packet,
 615          * before connection is in the list, so we need to clean here,
 616          * too.
 617          */
 618         nf_ct_remove_expectations(ct);
 619
 620         nf_ct_del_from_dying_or_unconfirmed_list(ct);
 621
 622         local_bh_enable();
 623
 624         if (ct->master)
 625                 nf_ct_put(ct->master);
 626
 627         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
 628         nf_conntrack_free(ct);
 629 }
 630
 631 static void nf_ct_delete_from_lists(struct nf_conn *ct)
 632 {
 633         struct net *net = nf_ct_net(ct);
 634         unsigned int hash, reply_hash;
 635         unsigned int sequence;
 636
 637         nf_ct_helper_destroy(ct);
 638
 639         local_bh_disable();
 640         do {
 641                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 642                 hash = hash_conntrack(net,
 643                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 644                 reply_hash = hash_conntrack(net,
 645                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 646         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 647
 648         clean_from_lists(ct);
 649         nf_conntrack_double_unlock(hash, reply_hash);
 650
 651         nf_ct_add_to_dying_list(ct);
 652
 653         local_bh_enable();
 654 }
 655
 656 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
 657 {
 658         struct nf_conn_tstamp *tstamp;
 659
 660         if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
 661                 return false;
 662
 663         tstamp = nf_conn_tstamp_find(ct);
 664         if (tstamp && tstamp->stop == 0)
 665                 tstamp->stop = ktime_get_real_ns();
 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_delete_from_lists(ct);
 673                 nf_conntrack_ecache_delayed_work(nf_ct_net(ct));
 674                 return false;
 675         }
 676
 677         nf_conntrack_ecache_work(nf_ct_net(ct));
 678         nf_ct_delete_from_lists(ct);
 679         nf_ct_put(ct);
 680         return true;
 681 }
 682 EXPORT_SYMBOL_GPL(nf_ct_delete);
 683
 684 static inline bool
 685 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
 686                 const struct nf_conntrack_tuple *tuple,
 687                 const struct nf_conntrack_zone *zone,
 688                 const struct net *net)
 689 {
 690         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 691
 692         /* A conntrack can be recreated with the equal tuple,
 693          * so we need to check that the conntrack is confirmed
 694          */
 695         return nf_ct_tuple_equal(tuple, &h->tuple) &&
 696                nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
 697                nf_ct_is_confirmed(ct) &&
 698                net_eq(net, nf_ct_net(ct));
 699 }
 700
 701 static inline bool
 702 nf_ct_match(const struct nf_conn *ct1, const struct nf_conn *ct2)
 703 {
 704         return nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 705                                  &ct2->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
 706                nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_REPLY].tuple,
 707                                  &ct2->tuplehash[IP_CT_DIR_REPLY].tuple) &&
 708                nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_ORIGINAL) &&
 709                nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_REPLY) &&
 710                net_eq(nf_ct_net(ct1), nf_ct_net(ct2));
 711 }
 712
 713 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
 714 static void nf_ct_gc_expired(struct nf_conn *ct)
 715 {
 716         if (!atomic_inc_not_zero(&ct->ct_general.use))
 717                 return;
 718
 719         if (nf_ct_should_gc(ct))
 720                 nf_ct_kill(ct);
 721
 722         nf_ct_put(ct);
 723 }
 724
 725 /*
 726  * Warning :
 727  * - Caller must take a reference on returned object
 728  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
 729  */
 730 static struct nf_conntrack_tuple_hash *
 731 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
 732                       const struct nf_conntrack_tuple *tuple, u32 hash)
 733 {
 734         struct nf_conntrack_tuple_hash *h;
 735         struct hlist_nulls_head *ct_hash;
 736         struct hlist_nulls_node *n;
 737         unsigned int bucket, hsize;
 738
 739 begin:
 740         nf_conntrack_get_ht(&ct_hash, &hsize);
 741         bucket = reciprocal_scale(hash, hsize);
 742
 743         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
 744                 struct nf_conn *ct;
 745
 746                 ct = nf_ct_tuplehash_to_ctrack(h);
 747                 if (nf_ct_is_expired(ct)) {
 748                         nf_ct_gc_expired(ct);
 749                         continue;
 750                 }
 751
 752                 if (nf_ct_key_equal(h, tuple, zone, net))
 753                         return h;
 754         }
 755         /*
 756          * if the nulls value we got at the end of this lookup is
 757          * not the expected one, we must restart lookup.
 758          * We probably met an item that was moved to another chain.
 759          */
 760         if (get_nulls_value(n) != bucket) {
 761                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
 762                 goto begin;
 763         }
 764
 765         return NULL;
 766 }
 767
 768 /* Find a connection corresponding to a tuple. */
 769 static struct nf_conntrack_tuple_hash *
 770 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 771                         const struct nf_conntrack_tuple *tuple, u32 hash)
 772 {
 773         struct nf_conntrack_tuple_hash *h;
 774         struct nf_conn *ct;
 775
 776         rcu_read_lock();
 777
 778         h = ____nf_conntrack_find(net, zone, tuple, hash);
 779         if (h) {
 780                 /* We have a candidate that matches the tuple we're interested
 781                  * in, try to obtain a reference and re-check tuple
 782                  */
 783                 ct = nf_ct_tuplehash_to_ctrack(h);
 784                 if (likely(atomic_inc_not_zero(&ct->ct_general.use))) {
 785                         if (likely(nf_ct_key_equal(h, tuple, zone, net)))
 786                                 goto found;
 787
 788                         /* TYPESAFE_BY_RCU recycled the candidate */
 789                         nf_ct_put(ct);
 790                 }
 791
 792                 h = NULL;
 793         }
 794 found:
 795         rcu_read_unlock();
 796
 797         return h;
 798 }
 799
 800 struct nf_conntrack_tuple_hash *
 801 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
 802                       const struct nf_conntrack_tuple *tuple)
 803 {
 804         return __nf_conntrack_find_get(net, zone, tuple,
 805                                        hash_conntrack_raw(tuple, net));
 806 }
 807 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
 808
 809 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
 810                                        unsigned int hash,
 811                                        unsigned int reply_hash)
 812 {
 813         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
 814                            &nf_conntrack_hash[hash]);
 815         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 816                            &nf_conntrack_hash[reply_hash]);
 817 }
 818
 819 int
 820 nf_conntrack_hash_check_insert(struct nf_conn *ct)
 821 {
 822         const struct nf_conntrack_zone *zone;
 823         struct net *net = nf_ct_net(ct);
 824         unsigned int hash, reply_hash;
 825         struct nf_conntrack_tuple_hash *h;
 826         struct hlist_nulls_node *n;
 827         unsigned int sequence;
 828
 829         zone = nf_ct_zone(ct);
 830
 831         local_bh_disable();
 832         do {
 833                 sequence = read_seqcount_begin(&nf_conntrack_generation);
 834                 hash = hash_conntrack(net,
 835                                       &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
 836                 reply_hash = hash_conntrack(net,
 837                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
 838         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
 839
 840         /* See if there's one in the list already, including reverse */
 841         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
 842                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
 843                                     zone, net))
 844                         goto out;
 845
 846         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
 847                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 848                                     zone, net))
 849                         goto out;
 850
 851         smp_wmb();
 852         /* The caller holds a reference to this object */
 853         atomic_set(&ct->ct_general.use, 2);
 854         __nf_conntrack_hash_insert(ct, hash, reply_hash);
 855         nf_conntrack_double_unlock(hash, reply_hash);
 856         NF_CT_STAT_INC(net, insert);
 857         local_bh_enable();
 858         return 0;
 859
 860 out:
 861         nf_conntrack_double_unlock(hash, reply_hash);
 862         local_bh_enable();
 863         return -EEXIST;
 864 }
 865 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
 866
 867 void nf_ct_acct_add(struct nf_conn *ct, u32 dir, unsigned int packets,
 868                     unsigned int bytes)
 869 {
 870         struct nf_conn_acct *acct;
 871
 872         acct = nf_conn_acct_find(ct);
 873         if (acct) {
 874                 struct nf_conn_counter *counter = acct->counter;
 875
 876                 atomic64_add(packets, &counter[dir].packets);
 877                 atomic64_add(bytes, &counter[dir].bytes);
 878         }
 879 }
 880 EXPORT_SYMBOL_GPL(nf_ct_acct_add);
 881
 882 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
 883                              const struct nf_conn *loser_ct)
 884 {
 885         struct nf_conn_acct *acct;
 886
 887         acct = nf_conn_acct_find(loser_ct);
 888         if (acct) {
 889                 struct nf_conn_counter *counter = acct->counter;
 890                 unsigned int bytes;
 891
 892                 /* u32 should be fine since we must have seen one packet. */
 893                 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
 894                 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), bytes);
 895         }
 896 }
 897
 898 static void __nf_conntrack_insert_prepare(struct nf_conn *ct)
 899 {
 900         struct nf_conn_tstamp *tstamp;
 901
 902         atomic_inc(&ct->ct_general.use);
 903         ct->status |= IPS_CONFIRMED;
 904
 905         /* set conntrack timestamp, if enabled. */
 906         tstamp = nf_conn_tstamp_find(ct);
 907         if (tstamp)
 908                 tstamp->start = ktime_get_real_ns();
 909 }
 910
 911 /* caller must hold locks to prevent concurrent changes */
 912 static int __nf_ct_resolve_clash(struct sk_buff *skb,
 913                                  struct nf_conntrack_tuple_hash *h)
 914 {
 915         /* This is the conntrack entry already in hashes that won race. */
 916         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
 917         enum ip_conntrack_info ctinfo;
 918         struct nf_conn *loser_ct;
 919
 920         loser_ct = nf_ct_get(skb, &ctinfo);
 921
 922         if (nf_ct_is_dying(ct))
 923                 return NF_DROP;
 924
 925         if (((ct->status & IPS_NAT_DONE_MASK) == 0) ||
 926             nf_ct_match(ct, loser_ct)) {
 927                 struct net *net = nf_ct_net(ct);
 928
 929                 nf_conntrack_get(&ct->ct_general);
 930
 931                 nf_ct_acct_merge(ct, ctinfo, loser_ct);
 932                 nf_ct_add_to_dying_list(loser_ct);
 933                 nf_conntrack_put(&loser_ct->ct_general);
 934                 nf_ct_set(skb, ct, ctinfo);
 935
 936                 NF_CT_STAT_INC(net, clash_resolve);
 937                 return NF_ACCEPT;
 938         }
 939
 940         return NF_DROP;
 941 }
 942
 943 /**
 944  * nf_ct_resolve_clash_harder - attempt to insert clashing conntrack entry
 945  *
 946  * @skb: skb that causes the collision
 947  * @repl_idx: hash slot for reply direction
 948  *
 949  * Called when origin or reply direction had a clash.
 950  * The skb can be handled without packet drop provided the reply direction
 951  * is unique or there the existing entry has the identical tuple in both
 952  * directions.
 953  *
 954  * Caller must hold conntrack table locks to prevent concurrent updates.
 955  *
 956  * Returns NF_DROP if the clash could not be handled.
 957  */
 958 static int nf_ct_resolve_clash_harder(struct sk_buff *skb, u32 repl_idx)
 959 {
 960         struct nf_conn *loser_ct = (struct nf_conn *)skb_nfct(skb);
 961         const struct nf_conntrack_zone *zone;
 962         struct nf_conntrack_tuple_hash *h;
 963         struct hlist_nulls_node *n;
 964         struct net *net;
 965
 966         zone = nf_ct_zone(loser_ct);
 967         net = nf_ct_net(loser_ct);
 968
 969         /* Reply direction must never result in a clash, unless both origin
 970          * and reply tuples are identical.
 971          */
 972         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[repl_idx], hnnode) {
 973                 if (nf_ct_key_equal(h,
 974                                     &loser_ct->tuplehash[IP_CT_DIR_REPLY].tuple,
 975                                     zone, net))
 976                         return __nf_ct_resolve_clash(skb, h);
 977         }
 978
 979         /* We want the clashing entry to go away real soon: 1 second timeout. */
 980         loser_ct->timeout = nfct_time_stamp + HZ;
 981
 982         /* IPS_NAT_CLASH removes the entry automatically on the first
 983          * reply.  Also prevents UDP tracker from moving the entry to
 984          * ASSURED state, i.e. the entry can always be evicted under
 985          * pressure.
 986          */
 987         loser_ct->status |= IPS_FIXED_TIMEOUT | IPS_NAT_CLASH;
 988
 989         __nf_conntrack_insert_prepare(loser_ct);
 990
 991         /* fake add for ORIGINAL dir: we want lookups to only find the entry
 992          * already in the table.  This also hides the clashing entry from
 993          * ctnetlink iteration, i.e. conntrack -L won't show them.
 994          */
 995         hlist_nulls_add_fake(&loser_ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
 996
 997         hlist_nulls_add_head_rcu(&loser_ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
 998                                  &nf_conntrack_hash[repl_idx]);
 999
1000         NF_CT_STAT_INC(net, clash_resolve);
1001         return NF_ACCEPT;
1002 }
1003
1004 /**
1005  * nf_ct_resolve_clash - attempt to handle clash without packet drop
1006  *
1007  * @skb: skb that causes the clash
1008  * @h: tuplehash of the clashing entry already in table
1009  * @reply_hash: hash slot for reply direction
1010  *
1011  * A conntrack entry can be inserted to the connection tracking table
1012  * if there is no existing entry with an identical tuple.
1013  *
1014  * If there is one, @skb (and the assocated, unconfirmed conntrack) has
1015  * to be dropped.  In case @skb is retransmitted, next conntrack lookup
1016  * will find the already-existing entry.
1017  *
1018  * The major problem with such packet drop is the extra delay added by
1019  * the packet loss -- it will take some time for a retransmit to occur
1020  * (or the sender to time out when waiting for a reply).
1021  *
1022  * This function attempts to handle the situation without packet drop.
1023  *
1024  * If @skb has no NAT transformation or if the colliding entries are
1025  * exactly the same, only the to-be-confirmed conntrack entry is discarded
1026  * and @skb is associated with the conntrack entry already in the table.
1027  *
1028  * Failing that, the new, unconfirmed conntrack is still added to the table
1029  * provided that the collision only occurs in the ORIGINAL direction.
1030  * The new entry will be added only in the non-clashing REPLY direction,
1031  * so packets in the ORIGINAL direction will continue to match the existing
1032  * entry.  The new entry will also have a fixed timeout so it expires --
1033  * due to the collision, it will only see reply traffic.
1034  *
1035  * Returns NF_DROP if the clash could not be resolved.
1036  */
1037 static __cold noinline int
1038 nf_ct_resolve_clash(struct sk_buff *skb, struct nf_conntrack_tuple_hash *h,
1039                     u32 reply_hash)
1040 {
1041         /* This is the conntrack entry already in hashes that won race. */
1042         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1043         const struct nf_conntrack_l4proto *l4proto;
1044         enum ip_conntrack_info ctinfo;
1045         struct nf_conn *loser_ct;
1046         struct net *net;
1047         int ret;
1048
1049         loser_ct = nf_ct_get(skb, &ctinfo);
1050         net = nf_ct_net(loser_ct);
1051
1052         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1053         if (!l4proto->allow_clash)
1054                 goto drop;
1055
1056         ret = __nf_ct_resolve_clash(skb, h);
1057         if (ret == NF_ACCEPT)
1058                 return ret;
1059
1060         ret = nf_ct_resolve_clash_harder(skb, reply_hash);
1061         if (ret == NF_ACCEPT)
1062                 return ret;
1063
1064 drop:
1065         nf_ct_add_to_dying_list(loser_ct);
1066         NF_CT_STAT_INC(net, drop);
1067         NF_CT_STAT_INC(net, insert_failed);
1068         return NF_DROP;
1069 }
1070
1071 /* Confirm a connection given skb; places it in hash table */
1072 int
1073 __nf_conntrack_confirm(struct sk_buff *skb)
1074 {
1075         const struct nf_conntrack_zone *zone;
1076         unsigned int hash, reply_hash;
1077         struct nf_conntrack_tuple_hash *h;
1078         struct nf_conn *ct;
1079         struct nf_conn_help *help;
1080         struct hlist_nulls_node *n;
1081         enum ip_conntrack_info ctinfo;
1082         struct net *net;
1083         unsigned int sequence;
1084         int ret = NF_DROP;
1085
1086         ct = nf_ct_get(skb, &ctinfo);
1087         net = nf_ct_net(ct);
1088
1089         /* ipt_REJECT uses nf_conntrack_attach to attach related
1090            ICMP/TCP RST packets in other direction.  Actual packet
1091            which created connection will be IP_CT_NEW or for an
1092            expected connection, IP_CT_RELATED. */
1093         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
1094                 return NF_ACCEPT;
1095
1096         zone = nf_ct_zone(ct);
1097         local_bh_disable();
1098
1099         do {
1100                 sequence = read_seqcount_begin(&nf_conntrack_generation);
1101                 /* reuse the hash saved before */
1102                 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
1103                 hash = scale_hash(hash);
1104                 reply_hash = hash_conntrack(net,
1105                                            &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
1106
1107         } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
1108
1109         /* We're not in hash table, and we refuse to set up related
1110          * connections for unconfirmed conns.  But packet copies and
1111          * REJECT will give spurious warnings here.
1112          */
1113
1114         /* Another skb with the same unconfirmed conntrack may
1115          * win the race. This may happen for bridge(br_flood)
1116          * or broadcast/multicast packets do skb_clone with
1117          * unconfirmed conntrack.
1118          */
1119         if (unlikely(nf_ct_is_confirmed(ct))) {
1120                 WARN_ON_ONCE(1);
1121                 nf_conntrack_double_unlock(hash, reply_hash);
1122                 local_bh_enable();
1123                 return NF_DROP;
1124         }
1125
1126         pr_debug("Confirming conntrack %p\n", ct);
1127         /* We have to check the DYING flag after unlink to prevent
1128          * a race against nf_ct_get_next_corpse() possibly called from
1129          * user context, else we insert an already 'dead' hash, blocking
1130          * further use of that particular connection -JM.
1131          */
1132         nf_ct_del_from_dying_or_unconfirmed_list(ct);
1133
1134         if (unlikely(nf_ct_is_dying(ct))) {
1135                 nf_ct_add_to_dying_list(ct);
1136                 NF_CT_STAT_INC(net, insert_failed);
1137                 goto dying;
1138         }
1139
1140         /* See if there's one in the list already, including reverse:
1141            NAT could have grabbed it without realizing, since we're
1142            not in the hash.  If there is, we lost race. */
1143         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
1144                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1145                                     zone, net))
1146                         goto out;
1147
1148         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
1149                 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1150                                     zone, net))
1151                         goto out;
1152
1153         /* Timer relative to confirmation time, not original
1154            setting time, otherwise we'd get timer wrap in
1155            weird delay cases. */
1156         ct->timeout += nfct_time_stamp;
1157
1158         __nf_conntrack_insert_prepare(ct);
1159
1160         /* Since the lookup is lockless, hash insertion must be done after
1161          * starting the timer and setting the CONFIRMED bit. The RCU barriers
1162          * guarantee that no other CPU can find the conntrack before the above
1163          * stores are visible.
1164          */
1165         __nf_conntrack_hash_insert(ct, hash, reply_hash);
1166         nf_conntrack_double_unlock(hash, reply_hash);
1167         local_bh_enable();
1168
1169         help = nfct_help(ct);
1170         if (help && help->helper)
1171                 nf_conntrack_event_cache(IPCT_HELPER, ct);
1172
1173         nf_conntrack_event_cache(master_ct(ct) ?
1174                                  IPCT_RELATED : IPCT_NEW, ct);
1175         return NF_ACCEPT;
1176
1177 out:
1178         ret = nf_ct_resolve_clash(skb, h, reply_hash);
1179 dying:
1180         nf_conntrack_double_unlock(hash, reply_hash);
1181         local_bh_enable();
1182         return ret;
1183 }
1184 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
1185
1186 /* Returns true if a connection correspondings to the tuple (required
1187    for NAT). */
1188 int
1189 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
1190                          const struct nf_conn *ignored_conntrack)
1191 {
1192         struct net *net = nf_ct_net(ignored_conntrack);
1193         const struct nf_conntrack_zone *zone;
1194         struct nf_conntrack_tuple_hash *h;
1195         struct hlist_nulls_head *ct_hash;
1196         unsigned int hash, hsize;
1197         struct hlist_nulls_node *n;
1198         struct nf_conn *ct;
1199
1200         zone = nf_ct_zone(ignored_conntrack);
1201
1202         rcu_read_lock();
1203  begin:
1204         nf_conntrack_get_ht(&ct_hash, &hsize);
1205         hash = __hash_conntrack(net, tuple, hsize);
1206
1207         hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
1208                 ct = nf_ct_tuplehash_to_ctrack(h);
1209
1210                 if (ct == ignored_conntrack)
1211                         continue;
1212
1213                 if (nf_ct_is_expired(ct)) {
1214                         nf_ct_gc_expired(ct);
1215                         continue;
1216                 }
1217
1218                 if (nf_ct_key_equal(h, tuple, zone, net)) {
1219                         /* Tuple is taken already, so caller will need to find
1220                          * a new source port to use.
1221                          *
1222                          * Only exception:
1223                          * If the *original tuples* are identical, then both
1224                          * conntracks refer to the same flow.
1225                          * This is a rare situation, it can occur e.g. when
1226                          * more than one UDP packet is sent from same socket
1227                          * in different threads.
1228                          *
1229                          * Let nf_ct_resolve_clash() deal with this later.
1230                          */
1231                         if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1232                                               &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
1233                                               nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL))
1234                                 continue;
1235
1236                         NF_CT_STAT_INC_ATOMIC(net, found);
1237                         rcu_read_unlock();
1238                         return 1;
1239                 }
1240         }
1241
1242         if (get_nulls_value(n) != hash) {
1243                 NF_CT_STAT_INC_ATOMIC(net, search_restart);
1244                 goto begin;
1245         }
1246
1247         rcu_read_unlock();
1248
1249         return 0;
1250 }
1251 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
1252
1253 #define NF_CT_EVICTION_RANGE    8
1254
1255 /* There's a small race here where we may free a just-assured
1256    connection.  Too bad: we're in trouble anyway. */
1257 static unsigned int early_drop_list(struct net *net,
1258                                     struct hlist_nulls_head *head)
1259 {
1260         struct nf_conntrack_tuple_hash *h;
1261         struct hlist_nulls_node *n;
1262         unsigned int drops = 0;
1263         struct nf_conn *tmp;
1264
1265         hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
1266                 tmp = nf_ct_tuplehash_to_ctrack(h);
1267
1268                 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status))
1269                         continue;
1270
1271                 if (nf_ct_is_expired(tmp)) {
1272                         nf_ct_gc_expired(tmp);
1273                         continue;
1274                 }
1275
1276                 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
1277                     !net_eq(nf_ct_net(tmp), net) ||
1278                     nf_ct_is_dying(tmp))
1279                         continue;
1280
1281                 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1282                         continue;
1283
1284                 /* kill only if still in same netns -- might have moved due to
1285                  * SLAB_TYPESAFE_BY_RCU rules.
1286                  *
1287                  * We steal the timer reference.  If that fails timer has
1288                  * already fired or someone else deleted it. Just drop ref
1289                  * and move to next entry.
1290                  */
1291                 if (net_eq(nf_ct_net(tmp), net) &&
1292                     nf_ct_is_confirmed(tmp) &&
1293                     nf_ct_delete(tmp, 0, 0))
1294                         drops++;
1295
1296                 nf_ct_put(tmp);
1297         }
1298
1299         return drops;
1300 }
1301
1302 static noinline int early_drop(struct net *net, unsigned int hash)
1303 {
1304         unsigned int i, bucket;
1305
1306         for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
1307                 struct hlist_nulls_head *ct_hash;
1308                 unsigned int hsize, drops;
1309
1310                 rcu_read_lock();
1311                 nf_conntrack_get_ht(&ct_hash, &hsize);
1312                 if (!i)
1313                         bucket = reciprocal_scale(hash, hsize);
1314                 else
1315                         bucket = (bucket + 1) % hsize;
1316
1317                 drops = early_drop_list(net, &ct_hash[bucket]);
1318                 rcu_read_unlock();
1319
1320                 if (drops) {
1321                         NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
1322                         return true;
1323                 }
1324         }
1325
1326         return false;
1327 }
1328
1329 static bool gc_worker_skip_ct(const struct nf_conn *ct)
1330 {
1331         return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
1332 }
1333
1334 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
1335 {
1336         const struct nf_conntrack_l4proto *l4proto;
1337
1338         if (!test_bit(IPS_ASSURED_BIT, &ct->status))
1339                 return true;
1340
1341         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1342         if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
1343                 return true;
1344
1345         return false;
1346 }
1347
1348 static void gc_worker(struct work_struct *work)
1349 {
1350         unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
1351         unsigned int i, goal, buckets = 0, expired_count = 0;
1352         unsigned int nf_conntrack_max95 = 0;
1353         struct conntrack_gc_work *gc_work;
1354         unsigned int ratio, scanned = 0;
1355         unsigned long next_run;
1356
1357         gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1358
1359         goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
1360         i = gc_work->last_bucket;
1361         if (gc_work->early_drop)
1362                 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1363
1364         do {
1365                 struct nf_conntrack_tuple_hash *h;
1366                 struct hlist_nulls_head *ct_hash;
1367                 struct hlist_nulls_node *n;
1368                 unsigned int hashsz;
1369                 struct nf_conn *tmp;
1370
1371                 i++;
1372                 rcu_read_lock();
1373
1374                 nf_conntrack_get_ht(&ct_hash, &hashsz);
1375                 if (i >= hashsz)
1376                         i = 0;
1377
1378                 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1379                         struct net *net;
1380
1381                         tmp = nf_ct_tuplehash_to_ctrack(h);
1382
1383                         scanned++;
1384                         if (test_bit(IPS_OFFLOAD_BIT, &tmp->status)) {
1385                                 nf_ct_offload_timeout(tmp);
1386                                 continue;
1387                         }
1388
1389                         if (nf_ct_is_expired(tmp)) {
1390                                 nf_ct_gc_expired(tmp);
1391                                 expired_count++;
1392                                 continue;
1393                         }
1394
1395                         if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1396                                 continue;
1397
1398                         net = nf_ct_net(tmp);
1399                         if (atomic_read(&net->ct.count) < nf_conntrack_max95)
1400                                 continue;
1401
1402                         /* need to take reference to avoid possible races */
1403                         if (!atomic_inc_not_zero(&tmp->ct_general.use))
1404                                 continue;
1405
1406                         if (gc_worker_skip_ct(tmp)) {
1407                                 nf_ct_put(tmp);
1408                                 continue;
1409                         }
1410
1411                         if (gc_worker_can_early_drop(tmp))
1412                                 nf_ct_kill(tmp);
1413
1414                         nf_ct_put(tmp);
1415                 }
1416
1417                 /* could check get_nulls_value() here and restart if ct
1418                  * was moved to another chain.  But given gc is best-effort
1419                  * we will just continue with next hash slot.
1420                  */
1421                 rcu_read_unlock();
1422                 cond_resched();
1423         } while (++buckets < goal);
1424
1425         if (gc_work->exiting)
1426                 return;
1427
1428         /*
1429          * Eviction will normally happen from the packet path, and not
1430          * from this gc worker.
1431          *
1432          * This worker is only here to reap expired entries when system went
1433          * idle after a busy period.
1434          *
1435          * The heuristics below are supposed to balance conflicting goals:
1436          *
1437          * 1. Minimize time until we notice a stale entry
1438          * 2. Maximize scan intervals to not waste cycles
1439          *
1440          * Normally, expire ratio will be close to 0.
1441          *
1442          * As soon as a sizeable fraction of the entries have expired
1443          * increase scan frequency.
1444          */
1445         ratio = scanned ? expired_count * 100 / scanned : 0;
1446         if (ratio > GC_EVICT_RATIO) {
1447                 gc_work->next_gc_run = min_interval;
1448         } else {
1449                 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1450
1451                 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1452
1453                 gc_work->next_gc_run += min_interval;
1454                 if (gc_work->next_gc_run > max)
1455                         gc_work->next_gc_run = max;
1456         }
1457
1458         next_run = gc_work->next_gc_run;
1459         gc_work->last_bucket = i;
1460         gc_work->early_drop = false;
1461         queue_delayed_work(system_power_efficient_wq, &gc_work->dwork, next_run);
1462 }
1463
1464 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1465 {
1466         INIT_DEFERRABLE_WORK(&gc_work->dwork, gc_worker);
1467         gc_work->next_gc_run = HZ;
1468         gc_work->exiting = false;
1469 }
1470
1471 static struct nf_conn *
1472 __nf_conntrack_alloc(struct net *net,
1473                      const struct nf_conntrack_zone *zone,
1474                      const struct nf_conntrack_tuple *orig,
1475                      const struct nf_conntrack_tuple *repl,
1476                      gfp_t gfp, u32 hash)
1477 {
1478         struct nf_conn *ct;
1479
1480         /* We don't want any race condition at early drop stage */
1481         atomic_inc(&net->ct.count);
1482
1483         if (nf_conntrack_max &&
1484             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
1485                 if (!early_drop(net, hash)) {
1486                         if (!conntrack_gc_work.early_drop)
1487                                 conntrack_gc_work.early_drop = true;
1488                         atomic_dec(&net->ct.count);
1489                         net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1490                         return ERR_PTR(-ENOMEM);
1491                 }
1492         }
1493
1494         /*
1495          * Do not use kmem_cache_zalloc(), as this cache uses
1496          * SLAB_TYPESAFE_BY_RCU.
1497          */
1498         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1499         if (ct == NULL)
1500                 goto out;
1501
1502         spin_lock_init(&ct->lock);
1503         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1504         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1505         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1506         /* save hash for reusing when confirming */
1507         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1508         ct->status = 0;
1509         ct->timeout = 0;
1510         write_pnet(&ct->ct_net, net);
1511         memset(&ct->__nfct_init_offset, 0,
1512                offsetof(struct nf_conn, proto) -
1513                offsetof(struct nf_conn, __nfct_init_offset));
1514
1515         nf_ct_zone_add(ct, zone);
1516
1517         /* Because we use RCU lookups, we set ct_general.use to zero before
1518          * this is inserted in any list.
1519          */
1520         atomic_set(&ct->ct_general.use, 0);
1521         return ct;
1522 out:
1523         atomic_dec(&net->ct.count);
1524         return ERR_PTR(-ENOMEM);
1525 }
1526
1527 struct nf_conn *nf_conntrack_alloc(struct net *net,
1528                                    const struct nf_conntrack_zone *zone,
1529                                    const struct nf_conntrack_tuple *orig,
1530                                    const struct nf_conntrack_tuple *repl,
1531                                    gfp_t gfp)
1532 {
1533         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1534 }
1535 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1536
1537 void nf_conntrack_free(struct nf_conn *ct)
1538 {
1539         struct net *net = nf_ct_net(ct);
1540
1541         /* A freed object has refcnt == 0, that's
1542          * the golden rule for SLAB_TYPESAFE_BY_RCU
1543          */
1544         WARN_ON(atomic_read(&ct->ct_general.use) != 0);
1545
1546         nf_ct_ext_destroy(ct);
1547         kmem_cache_free(nf_conntrack_cachep, ct);
1548         smp_mb__before_atomic();
1549         atomic_dec(&net->ct.count);
1550 }
1551 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1552
1553
1554 /* Allocate a new conntrack: we return -ENOMEM if classification
1555    failed due to stress.  Otherwise it really is unclassifiable. */
1556 static noinline struct nf_conntrack_tuple_hash *
1557 init_conntrack(struct net *net, struct nf_conn *tmpl,
1558                const struct nf_conntrack_tuple *tuple,
1559                struct sk_buff *skb,
1560                unsigned int dataoff, u32 hash)
1561 {
1562         struct nf_conn *ct;
1563         struct nf_conn_help *help;
1564         struct nf_conntrack_tuple repl_tuple;
1565         struct nf_conntrack_ecache *ecache;
1566         struct nf_conntrack_expect *exp = NULL;
1567         const struct nf_conntrack_zone *zone;
1568         struct nf_conn_timeout *timeout_ext;
1569         struct nf_conntrack_zone tmp;
1570
1571         if (!nf_ct_invert_tuple(&repl_tuple, tuple)) {
1572                 pr_debug("Can't invert tuple.\n");
1573                 return NULL;
1574         }
1575
1576         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1577         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1578                                   hash);
1579         if (IS_ERR(ct))
1580                 return (struct nf_conntrack_tuple_hash *)ct;
1581
1582         if (!nf_ct_add_synproxy(ct, tmpl)) {
1583                 nf_conntrack_free(ct);
1584                 return ERR_PTR(-ENOMEM);
1585         }
1586
1587         timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1588
1589         if (timeout_ext)
1590                 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1591                                       GFP_ATOMIC);
1592
1593         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1594         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1595         nf_ct_labels_ext_add(ct);
1596
1597         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1598         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1599                                  ecache ? ecache->expmask : 0,
1600                              GFP_ATOMIC);
1601
1602         local_bh_disable();
1603         if (net->ct.expect_count) {
1604                 spin_lock(&nf_conntrack_expect_lock);
1605                 exp = nf_ct_find_expectation(net, zone, tuple);
1606                 if (exp) {
1607                         pr_debug("expectation arrives ct=%p exp=%p\n",
1608                                  ct, exp);
1609                         /* Welcome, Mr. Bond.  We've been expecting you... */
1610                         __set_bit(IPS_EXPECTED_BIT, &ct->status);
1611                         /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1612                         ct->master = exp->master;
1613                         if (exp->helper) {
1614                                 help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1615                                 if (help)
1616                                         rcu_assign_pointer(help->helper, exp->helper);
1617                         }
1618
1619 #ifdef CONFIG_NF_CONNTRACK_MARK
1620                         ct->mark = exp->master->mark;
1621 #endif
1622 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1623                         ct->secmark = exp->master->secmark;
1624 #endif
1625                         NF_CT_STAT_INC(net, expect_new);
1626                 }
1627                 spin_unlock(&nf_conntrack_expect_lock);
1628         }
1629         if (!exp)
1630                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1631
1632         /* Now it is inserted into the unconfirmed list, bump refcount */
1633         nf_conntrack_get(&ct->ct_general);
1634         nf_ct_add_to_unconfirmed_list(ct);
1635
1636         local_bh_enable();
1637
1638         if (exp) {
1639                 if (exp->expectfn)
1640                         exp->expectfn(ct, exp);
1641                 nf_ct_expect_put(exp);
1642         }
1643
1644         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1645 }
1646
1647 /* On success, returns 0, sets skb->_nfct | ctinfo */
1648 static int
1649 resolve_normal_ct(struct nf_conn *tmpl,
1650                   struct sk_buff *skb,
1651                   unsigned int dataoff,
1652                   u_int8_t protonum,
1653                   const struct nf_hook_state *state)
1654 {
1655         const struct nf_conntrack_zone *zone;
1656         struct nf_conntrack_tuple tuple;
1657         struct nf_conntrack_tuple_hash *h;
1658         enum ip_conntrack_info ctinfo;
1659         struct nf_conntrack_zone tmp;
1660         struct nf_conn *ct;
1661         u32 hash;
1662
1663         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1664                              dataoff, state->pf, protonum, state->net,
1665                              &tuple)) {
1666                 pr_debug("Can't get tuple\n");
1667                 return 0;
1668         }
1669
1670         /* look for tuple match */
1671         zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1672         hash = hash_conntrack_raw(&tuple, state->net);
1673         h = __nf_conntrack_find_get(state->net, zone, &tuple, hash);
1674         if (!h) {
1675                 h = init_conntrack(state->net, tmpl, &tuple,
1676                                    skb, dataoff, hash);
1677                 if (!h)
1678                         return 0;
1679                 if (IS_ERR(h))
1680                         return PTR_ERR(h);
1681         }
1682         ct = nf_ct_tuplehash_to_ctrack(h);
1683
1684         /* It exists; we have (non-exclusive) reference. */
1685         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1686                 ctinfo = IP_CT_ESTABLISHED_REPLY;
1687         } else {
1688                 /* Once we've had two way comms, always ESTABLISHED. */
1689                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1690                         pr_debug("normal packet for %p\n", ct);
1691                         ctinfo = IP_CT_ESTABLISHED;
1692                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1693                         pr_debug("related packet for %p\n", ct);
1694                         ctinfo = IP_CT_RELATED;
1695                 } else {
1696                         pr_debug("new packet for %p\n", ct);
1697                         ctinfo = IP_CT_NEW;
1698                 }
1699         }
1700         nf_ct_set(skb, ct, ctinfo);
1701         return 0;
1702 }
1703
1704 /*
1705  * icmp packets need special treatment to handle error messages that are
1706  * related to a connection.
1707  *
1708  * Callers need to check if skb has a conntrack assigned when this
1709  * helper returns; in such case skb belongs to an already known connection.
1710  */
1711 static unsigned int __cold
1712 nf_conntrack_handle_icmp(struct nf_conn *tmpl,
1713                          struct sk_buff *skb,
1714                          unsigned int dataoff,
1715                          u8 protonum,
1716                          const struct nf_hook_state *state)
1717 {
1718         int ret;
1719
1720         if (state->pf == NFPROTO_IPV4 && protonum == IPPROTO_ICMP)
1721                 ret = nf_conntrack_icmpv4_error(tmpl, skb, dataoff, state);
1722 #if IS_ENABLED(CONFIG_IPV6)
1723         else if (state->pf == NFPROTO_IPV6 && protonum == IPPROTO_ICMPV6)
1724                 ret = nf_conntrack_icmpv6_error(tmpl, skb, dataoff, state);
1725 #endif
1726         else
1727                 return NF_ACCEPT;
1728
1729         if (ret <= 0)
1730                 NF_CT_STAT_INC_ATOMIC(state->net, error);
1731
1732         return ret;
1733 }
1734
1735 static int generic_packet(struct nf_conn *ct, struct sk_buff *skb,
1736                           enum ip_conntrack_info ctinfo)
1737 {
1738         const unsigned int *timeout = nf_ct_timeout_lookup(ct);
1739
1740         if (!timeout)
1741                 timeout = &nf_generic_pernet(nf_ct_net(ct))->timeout;
1742
1743         nf_ct_refresh_acct(ct, ctinfo, skb, *timeout);
1744         return NF_ACCEPT;
1745 }
1746
1747 /* Returns verdict for packet, or -1 for invalid. */
1748 static int nf_conntrack_handle_packet(struct nf_conn *ct,
1749                                       struct sk_buff *skb,
1750                                       unsigned int dataoff,
1751                                       enum ip_conntrack_info ctinfo,
1752                                       const struct nf_hook_state *state)
1753 {
1754         switch (nf_ct_protonum(ct)) {
1755         case IPPROTO_TCP:
1756                 return nf_conntrack_tcp_packet(ct, skb, dataoff,
1757                                                ctinfo, state);
1758         case IPPROTO_UDP:
1759                 return nf_conntrack_udp_packet(ct, skb, dataoff,
1760                                                ctinfo, state);
1761         case IPPROTO_ICMP:
1762                 return nf_conntrack_icmp_packet(ct, skb, ctinfo, state);
1763 #if IS_ENABLED(CONFIG_IPV6)
1764         case IPPROTO_ICMPV6:
1765                 return nf_conntrack_icmpv6_packet(ct, skb, ctinfo, state);
1766 #endif
1767 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
1768         case IPPROTO_UDPLITE:
1769                 return nf_conntrack_udplite_packet(ct, skb, dataoff,
1770                                                    ctinfo, state);
1771 #endif
1772 #ifdef CONFIG_NF_CT_PROTO_SCTP
1773         case IPPROTO_SCTP:
1774                 return nf_conntrack_sctp_packet(ct, skb, dataoff,
1775                                                 ctinfo, state);
1776 #endif
1777 #ifdef CONFIG_NF_CT_PROTO_DCCP
1778         case IPPROTO_DCCP:
1779                 return nf_conntrack_dccp_packet(ct, skb, dataoff,
1780                                                 ctinfo, state);
1781 #endif
1782 #ifdef CONFIG_NF_CT_PROTO_GRE
1783         case IPPROTO_GRE:
1784                 return nf_conntrack_gre_packet(ct, skb, dataoff,
1785                                                ctinfo, state);
1786 #endif
1787         }
1788
1789         return generic_packet(ct, skb, ctinfo);
1790 }
1791
1792 unsigned int
1793 nf_conntrack_in(struct sk_buff *skb, const struct nf_hook_state *state)
1794 {
1795         enum ip_conntrack_info ctinfo;
1796         struct nf_conn *ct, *tmpl;
1797         u_int8_t protonum;
1798         int dataoff, ret;
1799
1800         tmpl = nf_ct_get(skb, &ctinfo);
1801         if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1802                 /* Previously seen (loopback or untracked)?  Ignore. */
1803                 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1804                      ctinfo == IP_CT_UNTRACKED)
1805                         return NF_ACCEPT;
1806                 skb->_nfct = 0;
1807         }
1808
1809         /* rcu_read_lock()ed by nf_hook_thresh */
1810         dataoff = get_l4proto(skb, skb_network_offset(skb), state->pf, &protonum);
1811         if (dataoff <= 0) {
1812                 pr_debug("not prepared to track yet or error occurred\n");
1813                 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1814                 ret = NF_ACCEPT;
1815                 goto out;
1816         }
1817
1818         if (protonum == IPPROTO_ICMP || protonum == IPPROTO_ICMPV6) {
1819                 ret = nf_conntrack_handle_icmp(tmpl, skb, dataoff,
1820                                                protonum, state);
1821                 if (ret <= 0) {
1822                         ret = -ret;
1823                         goto out;
1824                 }
1825                 /* ICMP[v6] protocol trackers may assign one conntrack. */
1826                 if (skb->_nfct)
1827                         goto out;
1828         }
1829 repeat:
1830         ret = resolve_normal_ct(tmpl, skb, dataoff,
1831                                 protonum, state);
1832         if (ret < 0) {
1833                 /* Too stressed to deal. */
1834                 NF_CT_STAT_INC_ATOMIC(state->net, drop);
1835                 ret = NF_DROP;
1836                 goto out;
1837         }
1838
1839         ct = nf_ct_get(skb, &ctinfo);
1840         if (!ct) {
1841                 /* Not valid part of a connection */
1842                 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1843                 ret = NF_ACCEPT;
1844                 goto out;
1845         }
1846
1847         ret = nf_conntrack_handle_packet(ct, skb, dataoff, ctinfo, state);
1848         if (ret <= 0) {
1849                 /* Invalid: inverse of the return code tells
1850                  * the netfilter core what to do */
1851                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1852                 nf_conntrack_put(&ct->ct_general);
1853                 skb->_nfct = 0;
1854                 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1855                 if (ret == -NF_DROP)
1856                         NF_CT_STAT_INC_ATOMIC(state->net, drop);
1857                 /* Special case: TCP tracker reports an attempt to reopen a
1858                  * closed/aborted connection. We have to go back and create a
1859                  * fresh conntrack.
1860                  */
1861                 if (ret == -NF_REPEAT)
1862                         goto repeat;
1863                 ret = -ret;
1864                 goto out;
1865         }
1866
1867         if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
1868             !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1869                 nf_conntrack_event_cache(IPCT_REPLY, ct);
1870 out:
1871         if (tmpl)
1872                 nf_ct_put(tmpl);
1873
1874         return ret;
1875 }
1876 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1877
1878 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
1879    implicitly racy: see __nf_conntrack_confirm */
1880 void nf_conntrack_alter_reply(struct nf_conn *ct,
1881                               const struct nf_conntrack_tuple *newreply)
1882 {
1883         struct nf_conn_help *help = nfct_help(ct);
1884
1885         /* Should be unconfirmed, so not in hash table yet */
1886         WARN_ON(nf_ct_is_confirmed(ct));
1887
1888         pr_debug("Altering reply tuple of %p to ", ct);
1889         nf_ct_dump_tuple(newreply);
1890
1891         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1892         if (ct->master || (help && !hlist_empty(&help->expectations)))
1893                 return;
1894
1895         rcu_read_lock();
1896         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1897         rcu_read_unlock();
1898 }
1899 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1900
1901 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1902 void __nf_ct_refresh_acct(struct nf_conn *ct,
1903                           enum ip_conntrack_info ctinfo,
1904                           const struct sk_buff *skb,
1905                           u32 extra_jiffies,
1906                           bool do_acct)
1907 {
1908         /* Only update if this is not a fixed timeout */
1909         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1910                 goto acct;
1911
1912         /* If not in hash table, timer will not be active yet */
1913         if (nf_ct_is_confirmed(ct))
1914                 extra_jiffies += nfct_time_stamp;
1915
1916         if (READ_ONCE(ct->timeout) != extra_jiffies)
1917                 WRITE_ONCE(ct->timeout, extra_jiffies);
1918 acct:
1919         if (do_acct)
1920                 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
1921 }
1922 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1923
1924 bool nf_ct_kill_acct(struct nf_conn *ct,
1925                      enum ip_conntrack_info ctinfo,
1926                      const struct sk_buff *skb)
1927 {
1928         nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
1929
1930         return nf_ct_delete(ct, 0, 0);
1931 }
1932 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1933
1934 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1935
1936 #include <linux/netfilter/nfnetlink.h>
1937 #include <linux/netfilter/nfnetlink_conntrack.h>
1938 #include <linux/mutex.h>
1939
1940 /* Generic function for tcp/udp/sctp/dccp and alike. */
1941 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1942                                const struct nf_conntrack_tuple *tuple)
1943 {
1944         if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1945             nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1946                 goto nla_put_failure;
1947         return 0;
1948
1949 nla_put_failure:
1950         return -1;
1951 }
1952 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1953
1954 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1955         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
1956         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
1957 };
1958 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1959
1960 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1961                                struct nf_conntrack_tuple *t,
1962                                u_int32_t flags)
1963 {
1964         if (flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) {
1965                 if (!tb[CTA_PROTO_SRC_PORT])
1966                         return -EINVAL;
1967
1968                 t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1969         }
1970
1971         if (flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) {
1972                 if (!tb[CTA_PROTO_DST_PORT])
1973                         return -EINVAL;
1974
1975                 t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1976         }
1977
1978         return 0;
1979 }
1980 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1981
1982 unsigned int nf_ct_port_nlattr_tuple_size(void)
1983 {
1984         static unsigned int size __read_mostly;
1985
1986         if (!size)
1987                 size = nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1988
1989         return size;
1990 }
1991 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1992 #endif
1993
1994 /* Used by ipt_REJECT and ip6t_REJECT. */
1995 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
1996 {
1997         struct nf_conn *ct;
1998         enum ip_conntrack_info ctinfo;
1999
2000         /* This ICMP is in reverse direction to the packet which caused it */
2001         ct = nf_ct_get(skb, &ctinfo);
2002         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
2003                 ctinfo = IP_CT_RELATED_REPLY;
2004         else
2005                 ctinfo = IP_CT_RELATED;
2006
2007         /* Attach to new skbuff, and increment count */
2008         nf_ct_set(nskb, ct, ctinfo);
2009         nf_conntrack_get(skb_nfct(nskb));
2010 }
2011
2012 static int __nf_conntrack_update(struct net *net, struct sk_buff *skb,
2013                                  struct nf_conn *ct,
2014                                  enum ip_conntrack_info ctinfo)
2015 {
2016         struct nf_conntrack_tuple_hash *h;
2017         struct nf_conntrack_tuple tuple;
2018         struct nf_nat_hook *nat_hook;
2019         unsigned int status;
2020         int dataoff;
2021         u16 l3num;
2022         u8 l4num;
2023
2024         l3num = nf_ct_l3num(ct);
2025
2026         dataoff = get_l4proto(skb, skb_network_offset(skb), l3num, &l4num);
2027         if (dataoff <= 0)
2028                 return -1;
2029
2030         if (!nf_ct_get_tuple(skb, skb_network_offset(skb), dataoff, l3num,
2031                              l4num, net, &tuple))
2032                 return -1;
2033
2034         if (ct->status & IPS_SRC_NAT) {
2035                 memcpy(tuple.src.u3.all,
2036                        ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.all,
2037                        sizeof(tuple.src.u3.all));
2038                 tuple.src.u.all =
2039                         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u.all;
2040         }
2041
2042         if (ct->status & IPS_DST_NAT) {
2043                 memcpy(tuple.dst.u3.all,
2044                        ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.all,
2045                        sizeof(tuple.dst.u3.all));
2046                 tuple.dst.u.all =
2047                         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u.all;
2048         }
2049
2050         h = nf_conntrack_find_get(net, nf_ct_zone(ct), &tuple);
2051         if (!h)
2052                 return 0;
2053
2054         /* Store status bits of the conntrack that is clashing to re-do NAT
2055          * mangling according to what it has been done already to this packet.
2056          */
2057         status = ct->status;
2058
2059         nf_ct_put(ct);
2060         ct = nf_ct_tuplehash_to_ctrack(h);
2061         nf_ct_set(skb, ct, ctinfo);
2062
2063         nat_hook = rcu_dereference(nf_nat_hook);
2064         if (!nat_hook)
2065                 return 0;
2066
2067         if (status & IPS_SRC_NAT &&
2068             nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_SRC,
2069                                 IP_CT_DIR_ORIGINAL) == NF_DROP)
2070                 return -1;
2071
2072         if (status & IPS_DST_NAT &&
2073             nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_DST,
2074                                 IP_CT_DIR_ORIGINAL) == NF_DROP)
2075                 return -1;
2076
2077         return 0;
2078 }
2079
2080 /* This packet is coming from userspace via nf_queue, complete the packet
2081  * processing after the helper invocation in nf_confirm().
2082  */
2083 static int nf_confirm_cthelper(struct sk_buff *skb, struct nf_conn *ct,
2084                                enum ip_conntrack_info ctinfo)
2085 {
2086         const struct nf_conntrack_helper *helper;
2087         const struct nf_conn_help *help;
2088         int protoff;
2089
2090         help = nfct_help(ct);
2091         if (!help)
2092                 return 0;
2093
2094         helper = rcu_dereference(help->helper);
2095         if (!(helper->flags & NF_CT_HELPER_F_USERSPACE))
2096                 return 0;
2097
2098         switch (nf_ct_l3num(ct)) {
2099         case NFPROTO_IPV4:
2100                 protoff = skb_network_offset(skb) + ip_hdrlen(skb);
2101                 break;
2102 #if IS_ENABLED(CONFIG_IPV6)
2103         case NFPROTO_IPV6: {
2104                 __be16 frag_off;
2105                 u8 pnum;
2106
2107                 pnum = ipv6_hdr(skb)->nexthdr;
2108                 protoff = ipv6_skip_exthdr(skb, sizeof(struct ipv6hdr), &pnum,
2109                                            &frag_off);
2110                 if (protoff < 0 || (frag_off & htons(~0x7)) != 0)
2111                         return 0;
2112                 break;
2113         }
2114 #endif
2115         default:
2116                 return 0;
2117         }
2118
2119         if (test_bit(IPS_SEQ_ADJUST_BIT, &ct->status) &&
2120             !nf_is_loopback_packet(skb)) {
2121                 if (!nf_ct_seq_adjust(skb, ct, ctinfo, protoff)) {
2122                         NF_CT_STAT_INC_ATOMIC(nf_ct_net(ct), drop);
2123                         return -1;
2124                 }
2125         }
2126
2127         /* We've seen it coming out the other side: confirm it */
2128         return nf_conntrack_confirm(skb) == NF_DROP ? - 1 : 0;
2129 }
2130
2131 static int nf_conntrack_update(struct net *net, struct sk_buff *skb)
2132 {
2133         enum ip_conntrack_info ctinfo;
2134         struct nf_conn *ct;
2135         int err;
2136
2137         ct = nf_ct_get(skb, &ctinfo);
2138         if (!ct)
2139                 return 0;
2140
2141         if (!nf_ct_is_confirmed(ct)) {
2142                 err = __nf_conntrack_update(net, skb, ct, ctinfo);
2143                 if (err < 0)
2144                         return err;
2145
2146                 ct = nf_ct_get(skb, &ctinfo);
2147         }
2148
2149         return nf_confirm_cthelper(skb, ct, ctinfo);
2150 }
2151
2152 static bool nf_conntrack_get_tuple_skb(struct nf_conntrack_tuple *dst_tuple,
2153                                        const struct sk_buff *skb)
2154 {
2155         const struct nf_conntrack_tuple *src_tuple;
2156         const struct nf_conntrack_tuple_hash *hash;
2157         struct nf_conntrack_tuple srctuple;
2158         enum ip_conntrack_info ctinfo;
2159         struct nf_conn *ct;
2160
2161         ct = nf_ct_get(skb, &ctinfo);
2162         if (ct) {
2163                 src_tuple = nf_ct_tuple(ct, CTINFO2DIR(ctinfo));
2164                 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2165                 return true;
2166         }
2167
2168         if (!nf_ct_get_tuplepr(skb, skb_network_offset(skb),
2169                                NFPROTO_IPV4, dev_net(skb->dev),
2170                                &srctuple))
2171                 return false;
2172
2173         hash = nf_conntrack_find_get(dev_net(skb->dev),
2174                                      &nf_ct_zone_dflt,
2175                                      &srctuple);
2176         if (!hash)
2177                 return false;
2178
2179         ct = nf_ct_tuplehash_to_ctrack(hash);
2180         src_tuple = nf_ct_tuple(ct, !hash->tuple.dst.dir);
2181         memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2182         nf_ct_put(ct);
2183
2184         return true;
2185 }
2186
2187 /* Bring out ya dead! */
2188 static struct nf_conn *
2189 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
2190                 void *data, unsigned int *bucket)
2191 {
2192         struct nf_conntrack_tuple_hash *h;
2193         struct nf_conn *ct;
2194         struct hlist_nulls_node *n;
2195         spinlock_t *lockp;
2196
2197         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
2198                 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
2199                 local_bh_disable();
2200                 nf_conntrack_lock(lockp);
2201                 if (*bucket < nf_conntrack_htable_size) {
2202                         hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
2203                                 if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
2204                                         continue;
2205                                 /* All nf_conn objects are added to hash table twice, one
2206                                  * for original direction tuple, once for the reply tuple.
2207                                  *
2208                                  * Exception: In the IPS_NAT_CLASH case, only the reply
2209                                  * tuple is added (the original tuple already existed for
2210                                  * a different object).
2211                                  *
2212                                  * We only need to call the iterator once for each
2213                                  * conntrack, so we just use the 'reply' direction
2214                                  * tuple while iterating.
2215                                  */
2216                                 ct = nf_ct_tuplehash_to_ctrack(h);
2217                                 if (iter(ct, data))
2218                                         goto found;
2219                         }
2220                 }
2221                 spin_unlock(lockp);
2222                 local_bh_enable();
2223                 cond_resched();
2224         }
2225
2226         return NULL;
2227 found:
2228         atomic_inc(&ct->ct_general.use);
2229         spin_unlock(lockp);
2230         local_bh_enable();
2231         return ct;
2232 }
2233
2234 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
2235                                   void *data, u32 portid, int report)
2236 {
2237         unsigned int bucket = 0, sequence;
2238         struct nf_conn *ct;
2239
2240         might_sleep();
2241
2242         for (;;) {
2243                 sequence = read_seqcount_begin(&nf_conntrack_generation);
2244
2245                 while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
2246                         /* Time to push up daises... */
2247
2248                         nf_ct_delete(ct, portid, report);
2249                         nf_ct_put(ct);
2250                         cond_resched();
2251                 }
2252
2253                 if (!read_seqcount_retry(&nf_conntrack_generation, sequence))
2254                         break;
2255                 bucket = 0;
2256         }
2257 }
2258
2259 struct iter_data {
2260         int (*iter)(struct nf_conn *i, void *data);
2261         void *data;
2262         struct net *net;
2263 };
2264
2265 static int iter_net_only(struct nf_conn *i, void *data)
2266 {
2267         struct iter_data *d = data;
2268
2269         if (!net_eq(d->net, nf_ct_net(i)))
2270                 return 0;
2271
2272         return d->iter(i, d->data);
2273 }
2274
2275 static void
2276 __nf_ct_unconfirmed_destroy(struct net *net)
2277 {
2278         int cpu;
2279
2280         for_each_possible_cpu(cpu) {
2281                 struct nf_conntrack_tuple_hash *h;
2282                 struct hlist_nulls_node *n;
2283                 struct ct_pcpu *pcpu;
2284
2285                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2286
2287                 spin_lock_bh(&pcpu->lock);
2288                 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
2289                         struct nf_conn *ct;
2290
2291                         ct = nf_ct_tuplehash_to_ctrack(h);
2292
2293                         /* we cannot call iter() on unconfirmed list, the
2294                          * owning cpu can reallocate ct->ext at any time.
2295                          */
2296                         set_bit(IPS_DYING_BIT, &ct->status);
2297                 }
2298                 spin_unlock_bh(&pcpu->lock);
2299                 cond_resched();
2300         }
2301 }
2302
2303 void nf_ct_unconfirmed_destroy(struct net *net)
2304 {
2305         might_sleep();
2306
2307         if (atomic_read(&net->ct.count) > 0) {
2308                 __nf_ct_unconfirmed_destroy(net);
2309                 nf_queue_nf_hook_drop(net);
2310                 synchronize_net();
2311         }
2312 }
2313 EXPORT_SYMBOL_GPL(nf_ct_unconfirmed_destroy);
2314
2315 void nf_ct_iterate_cleanup_net(struct net *net,
2316                                int (*iter)(struct nf_conn *i, void *data),
2317                                void *data, u32 portid, int report)
2318 {
2319         struct iter_data d;
2320
2321         might_sleep();
2322
2323         if (atomic_read(&net->ct.count) == 0)
2324                 return;
2325
2326         d.iter = iter;
2327         d.data = data;
2328         d.net = net;
2329
2330         nf_ct_iterate_cleanup(iter_net_only, &d, portid, report);
2331 }
2332 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
2333
2334 /**
2335  * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
2336  * @iter: callback to invoke for each conntrack
2337  * @data: data to pass to @iter
2338  *
2339  * Like nf_ct_iterate_cleanup, but first marks conntracks on the
2340  * unconfirmed list as dying (so they will not be inserted into
2341  * main table).
2342  *
2343  * Can only be called in module exit path.
2344  */
2345 void
2346 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
2347 {
2348         struct net *net;
2349
2350         down_read(&net_rwsem);
2351         for_each_net(net) {
2352                 if (atomic_read(&net->ct.count) == 0)
2353                         continue;
2354                 __nf_ct_unconfirmed_destroy(net);
2355                 nf_queue_nf_hook_drop(net);
2356         }
2357         up_read(&net_rwsem);
2358
2359         /* Need to wait for netns cleanup worker to finish, if its
2360          * running -- it might have deleted a net namespace from
2361          * the global list, so our __nf_ct_unconfirmed_destroy() might
2362          * not have affected all namespaces.
2363          */
2364         net_ns_barrier();
2365
2366         /* a conntrack could have been unlinked from unconfirmed list
2367          * before we grabbed pcpu lock in __nf_ct_unconfirmed_destroy().
2368          * This makes sure its inserted into conntrack table.
2369          */
2370         synchronize_net();
2371
2372         nf_ct_iterate_cleanup(iter, data, 0, 0);
2373 }
2374 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
2375
2376 static int kill_all(struct nf_conn *i, void *data)
2377 {
2378         return net_eq(nf_ct_net(i), data);
2379 }
2380
2381 void nf_conntrack_cleanup_start(void)
2382 {
2383         conntrack_gc_work.exiting = true;
2384         RCU_INIT_POINTER(ip_ct_attach, NULL);
2385 }
2386
2387 void nf_conntrack_cleanup_end(void)
2388 {
2389         RCU_INIT_POINTER(nf_ct_hook, NULL);
2390         cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2391         kvfree(nf_conntrack_hash);
2392
2393         nf_conntrack_proto_fini();
2394         nf_conntrack_seqadj_fini();
2395         nf_conntrack_labels_fini();
2396         nf_conntrack_helper_fini();
2397         nf_conntrack_timeout_fini();
2398         nf_conntrack_ecache_fini();
2399         nf_conntrack_tstamp_fini();
2400         nf_conntrack_acct_fini();
2401         nf_conntrack_expect_fini();
2402
2403         kmem_cache_destroy(nf_conntrack_cachep);
2404 }
2405
2406 /*
2407  * Mishearing the voices in his head, our hero wonders how he's
2408  * supposed to kill the mall.
2409  */
2410 void nf_conntrack_cleanup_net(struct net *net)
2411 {
2412         LIST_HEAD(single);
2413
2414         list_add(&net->exit_list, &single);
2415         nf_conntrack_cleanup_net_list(&single);
2416 }
2417
2418 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
2419 {
2420         int busy;
2421         struct net *net;
2422
2423         /*
2424          * This makes sure all current packets have passed through
2425          *  netfilter framework.  Roll on, two-stage module
2426          *  delete...
2427          */
2428         synchronize_net();
2429 i_see_dead_people:
2430         busy = 0;
2431         list_for_each_entry(net, net_exit_list, exit_list) {
2432                 nf_ct_iterate_cleanup(kill_all, net, 0, 0);
2433                 if (atomic_read(&net->ct.count) != 0)
2434                         busy = 1;
2435         }
2436         if (busy) {
2437                 schedule();
2438                 goto i_see_dead_people;
2439         }
2440
2441         list_for_each_entry(net, net_exit_list, exit_list) {
2442                 nf_conntrack_proto_pernet_fini(net);
2443                 nf_conntrack_ecache_pernet_fini(net);
2444                 nf_conntrack_expect_pernet_fini(net);
2445                 free_percpu(net->ct.stat);
2446                 free_percpu(net->ct.pcpu_lists);
2447         }
2448 }
2449
2450 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
2451 {
2452         struct hlist_nulls_head *hash;
2453         unsigned int nr_slots, i;
2454
2455         if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
2456                 return NULL;
2457
2458         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
2459         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
2460
2461         hash = kvcalloc(nr_slots, sizeof(struct hlist_nulls_head), GFP_KERNEL);
2462
2463         if (hash && nulls)
2464                 for (i = 0; i < nr_slots; i++)
2465                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
2466
2467         return hash;
2468 }
2469 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
2470
2471 int nf_conntrack_hash_resize(unsigned int hashsize)
2472 {
2473         int i, bucket;
2474         unsigned int old_size;
2475         struct hlist_nulls_head *hash, *old_hash;
2476         struct nf_conntrack_tuple_hash *h;
2477         struct nf_conn *ct;
2478
2479         if (!hashsize)
2480                 return -EINVAL;
2481
2482         hash = nf_ct_alloc_hashtable(&hashsize, 1);
2483         if (!hash)
2484                 return -ENOMEM;
2485
2486         old_size = nf_conntrack_htable_size;
2487         if (old_size == hashsize) {
2488                 kvfree(hash);
2489                 return 0;
2490         }
2491
2492         local_bh_disable();
2493         nf_conntrack_all_lock();
2494         write_seqcount_begin(&nf_conntrack_generation);
2495
2496         /* Lookups in the old hash might happen in parallel, which means we
2497          * might get false negatives during connection lookup. New connections
2498          * created because of a false negative won't make it into the hash
2499          * though since that required taking the locks.
2500          */
2501
2502         for (i = 0; i < nf_conntrack_htable_size; i++) {
2503                 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
2504                         h = hlist_nulls_entry(nf_conntrack_hash[i].first,
2505                                               struct nf_conntrack_tuple_hash, hnnode);
2506                         ct = nf_ct_tuplehash_to_ctrack(h);
2507                         hlist_nulls_del_rcu(&h->hnnode);
2508                         bucket = __hash_conntrack(nf_ct_net(ct),
2509                                                   &h->tuple, hashsize);
2510                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
2511                 }
2512         }
2513         old_size = nf_conntrack_htable_size;
2514         old_hash = nf_conntrack_hash;
2515
2516         nf_conntrack_hash = hash;
2517         nf_conntrack_htable_size = hashsize;
2518
2519         write_seqcount_end(&nf_conntrack_generation);
2520         nf_conntrack_all_unlock();
2521         local_bh_enable();
2522
2523         synchronize_net();
2524         kvfree(old_hash);
2525         return 0;
2526 }
2527
2528 int nf_conntrack_set_hashsize(const char *val, const struct kernel_param *kp)
2529 {
2530         unsigned int hashsize;
2531         int rc;
2532
2533         if (current->nsproxy->net_ns != &init_net)
2534                 return -EOPNOTSUPP;
2535
2536         /* On boot, we can set this without any fancy locking. */
2537         if (!nf_conntrack_hash)
2538                 return param_set_uint(val, kp);
2539
2540         rc = kstrtouint(val, 0, &hashsize);
2541         if (rc)
2542                 return rc;
2543
2544         return nf_conntrack_hash_resize(hashsize);
2545 }
2546
2547 static __always_inline unsigned int total_extension_size(void)
2548 {
2549         /* remember to add new extensions below */
2550         BUILD_BUG_ON(NF_CT_EXT_NUM > 9);
2551
2552         return sizeof(struct nf_ct_ext) +
2553                sizeof(struct nf_conn_help)
2554 #if IS_ENABLED(CONFIG_NF_NAT)
2555                 + sizeof(struct nf_conn_nat)
2556 #endif
2557                 + sizeof(struct nf_conn_seqadj)
2558                 + sizeof(struct nf_conn_acct)
2559 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2560                 + sizeof(struct nf_conntrack_ecache)
2561 #endif
2562 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
2563                 + sizeof(struct nf_conn_tstamp)
2564 #endif
2565 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
2566                 + sizeof(struct nf_conn_timeout)
2567 #endif
2568 #ifdef CONFIG_NF_CONNTRACK_LABELS
2569                 + sizeof(struct nf_conn_labels)
2570 #endif
2571 #if IS_ENABLED(CONFIG_NETFILTER_SYNPROXY)
2572                 + sizeof(struct nf_conn_synproxy)
2573 #endif
2574         ;
2575 };
2576
2577 int nf_conntrack_init_start(void)
2578 {
2579         unsigned long nr_pages = totalram_pages();
2580         int max_factor = 8;
2581         int ret = -ENOMEM;
2582         int i;
2583
2584         /* struct nf_ct_ext uses u8 to store offsets/size */
2585         BUILD_BUG_ON(total_extension_size() > 255u);
2586
2587         seqcount_spinlock_init(&nf_conntrack_generation,
2588                                &nf_conntrack_locks_all_lock);
2589
2590         for (i = 0; i < CONNTRACK_LOCKS; i++)
2591                 spin_lock_init(&nf_conntrack_locks[i]);
2592
2593         if (!nf_conntrack_htable_size) {
2594                 /* Idea from tcp.c: use 1/16384 of memory.
2595                  * On i386: 32MB machine has 512 buckets.
2596                  * >= 1GB machines have 16384 buckets.
2597                  * >= 4GB machines have 65536 buckets.
2598                  */
2599                 nf_conntrack_htable_size
2600                         = (((nr_pages << PAGE_SHIFT) / 16384)
2601                            / sizeof(struct hlist_head));
2602                 if (nr_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2603                         nf_conntrack_htable_size = 65536;
2604                 else if (nr_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2605                         nf_conntrack_htable_size = 16384;
2606                 if (nf_conntrack_htable_size < 32)
2607                         nf_conntrack_htable_size = 32;
2608
2609                 /* Use a max. factor of four by default to get the same max as
2610                  * with the old struct list_heads. When a table size is given
2611                  * we use the old value of 8 to avoid reducing the max.
2612                  * entries. */
2613                 max_factor = 4;
2614         }
2615
2616         nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2617         if (!nf_conntrack_hash)
2618                 return -ENOMEM;
2619
2620         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2621
2622         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2623                                                 sizeof(struct nf_conn),
2624                                                 NFCT_INFOMASK + 1,
2625                                                 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2626         if (!nf_conntrack_cachep)
2627                 goto err_cachep;
2628
2629         ret = nf_conntrack_expect_init();
2630         if (ret < 0)
2631                 goto err_expect;
2632
2633         ret = nf_conntrack_acct_init();
2634         if (ret < 0)
2635                 goto err_acct;
2636
2637         ret = nf_conntrack_tstamp_init();
2638         if (ret < 0)
2639                 goto err_tstamp;
2640
2641         ret = nf_conntrack_ecache_init();
2642         if (ret < 0)
2643                 goto err_ecache;
2644
2645         ret = nf_conntrack_timeout_init();
2646         if (ret < 0)
2647                 goto err_timeout;
2648
2649         ret = nf_conntrack_helper_init();
2650         if (ret < 0)
2651                 goto err_helper;
2652
2653         ret = nf_conntrack_labels_init();
2654         if (ret < 0)
2655                 goto err_labels;
2656
2657         ret = nf_conntrack_seqadj_init();
2658         if (ret < 0)
2659                 goto err_seqadj;
2660
2661         ret = nf_conntrack_proto_init();
2662         if (ret < 0)
2663                 goto err_proto;
2664
2665         conntrack_gc_work_init(&conntrack_gc_work);
2666         queue_delayed_work(system_power_efficient_wq, &conntrack_gc_work.dwork, HZ);
2667
2668         return 0;
2669
2670 err_proto:
2671         nf_conntrack_seqadj_fini();
2672 err_seqadj:
2673         nf_conntrack_labels_fini();
2674 err_labels:
2675         nf_conntrack_helper_fini();
2676 err_helper:
2677         nf_conntrack_timeout_fini();
2678 err_timeout:
2679         nf_conntrack_ecache_fini();
2680 err_ecache:
2681         nf_conntrack_tstamp_fini();
2682 err_tstamp:
2683         nf_conntrack_acct_fini();
2684 err_acct:
2685         nf_conntrack_expect_fini();
2686 err_expect:
2687         kmem_cache_destroy(nf_conntrack_cachep);
2688 err_cachep:
2689         kvfree(nf_conntrack_hash);
2690         return ret;
2691 }
2692
2693 static struct nf_ct_hook nf_conntrack_hook = {
2694         .update         = nf_conntrack_update,
2695         .destroy        = destroy_conntrack,
2696         .get_tuple_skb  = nf_conntrack_get_tuple_skb,
2697 };
2698
2699 void nf_conntrack_init_end(void)
2700 {
2701         /* For use by REJECT target */
2702         RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
2703         RCU_INIT_POINTER(nf_ct_hook, &nf_conntrack_hook);
2704 }
2705
2706 /*
2707  * We need to use special "null" values, not used in hash table
2708  */
2709 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
2710 #define DYING_NULLS_VAL         ((1<<30)+1)
2711
2712 int nf_conntrack_init_net(struct net *net)
2713 {
2714         int ret = -ENOMEM;
2715         int cpu;
2716
2717         BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2718         BUILD_BUG_ON_NOT_POWER_OF_2(CONNTRACK_LOCKS);
2719         atomic_set(&net->ct.count, 0);
2720
2721         net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
2722         if (!net->ct.pcpu_lists)
2723                 goto err_stat;
2724
2725         for_each_possible_cpu(cpu) {
2726                 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2727
2728                 spin_lock_init(&pcpu->lock);
2729                 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
2730                 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
2731         }
2732
2733         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2734         if (!net->ct.stat)
2735                 goto err_pcpu_lists;
2736
2737         ret = nf_conntrack_expect_pernet_init(net);
2738         if (ret < 0)
2739                 goto err_expect;
2740
2741         nf_conntrack_acct_pernet_init(net);
2742         nf_conntrack_tstamp_pernet_init(net);
2743         nf_conntrack_ecache_pernet_init(net);
2744         nf_conntrack_helper_pernet_init(net);
2745         nf_conntrack_proto_pernet_init(net);
2746
2747         return 0;
2748
2749 err_expect:
2750         free_percpu(net->ct.stat);
2751 err_pcpu_lists:
2752         free_percpu(net->ct.pcpu_lists);
2753 err_stat:
2754         return ret;
2755 }