net/netfilter/nft_ct.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2008-2009 Patrick McHardy <[email protected]>
   4  * Copyright (c) 2016 Pablo Neira Ayuso <[email protected]>
   5  *
   6  * Development of this code funded by Astaro AG (http://www.astaro.com/)
   7  */
   8
   9 #include <linux/kernel.h>
  10 #include <linux/init.h>
  11 #include <linux/module.h>
  12 #include <linux/netlink.h>
  13 #include <linux/netfilter.h>
  14 #include <linux/netfilter/nf_tables.h>
  15 #include <net/netfilter/nf_tables.h>
  16 #include <net/netfilter/nf_conntrack.h>
  17 #include <net/netfilter/nf_conntrack_acct.h>
  18 #include <net/netfilter/nf_conntrack_tuple.h>
  19 #include <net/netfilter/nf_conntrack_helper.h>
  20 #include <net/netfilter/nf_conntrack_ecache.h>
  21 #include <net/netfilter/nf_conntrack_labels.h>
  22 #include <net/netfilter/nf_conntrack_timeout.h>
  23 #include <net/netfilter/nf_conntrack_l4proto.h>
  24 #include <net/netfilter/nf_conntrack_expect.h>
  25
  26 struct nft_ct {
  27         enum nft_ct_keys        key:8;
  28         enum ip_conntrack_dir   dir:8;
  29         u8                      len;
  30         union {
  31                 u8              dreg;
  32                 u8              sreg;
  33         };
  34 };
  35
  36 struct nft_ct_helper_obj  {
  37         struct nf_conntrack_helper *helper4;
  38         struct nf_conntrack_helper *helper6;
  39         u8 l4proto;
  40 };
  41
  42 #ifdef CONFIG_NF_CONNTRACK_ZONES
  43 static DEFINE_PER_CPU(struct nf_conn *, nft_ct_pcpu_template);
  44 static unsigned int nft_ct_pcpu_template_refcnt __read_mostly;
  45 static DEFINE_MUTEX(nft_ct_pcpu_mutex);
  46 #endif
  47
  48 static u64 nft_ct_get_eval_counter(const struct nf_conn_counter *c,
  49                                    enum nft_ct_keys k,
  50                                    enum ip_conntrack_dir d)
  51 {
  52         if (d < IP_CT_DIR_MAX)
  53                 return k == NFT_CT_BYTES ? atomic64_read(&c[d].bytes) :
  54                                            atomic64_read(&c[d].packets);
  55
  56         return nft_ct_get_eval_counter(c, k, IP_CT_DIR_ORIGINAL) +
  57                nft_ct_get_eval_counter(c, k, IP_CT_DIR_REPLY);
  58 }
  59
  60 static void nft_ct_get_eval(const struct nft_expr *expr,
  61                             struct nft_regs *regs,
  62                             const struct nft_pktinfo *pkt)
  63 {
  64         const struct nft_ct *priv = nft_expr_priv(expr);
  65         u32 *dest = &regs->data[priv->dreg];
  66         enum ip_conntrack_info ctinfo;
  67         const struct nf_conn *ct;
  68         const struct nf_conn_help *help;
  69         const struct nf_conntrack_tuple *tuple;
  70         const struct nf_conntrack_helper *helper;
  71         unsigned int state;
  72
  73         ct = nf_ct_get(pkt->skb, &ctinfo);
  74
  75         switch (priv->key) {
  76         case NFT_CT_STATE:
  77                 if (ct)
  78                         state = NF_CT_STATE_BIT(ctinfo);
  79                 else if (ctinfo == IP_CT_UNTRACKED)
  80                         state = NF_CT_STATE_UNTRACKED_BIT;
  81                 else
  82                         state = NF_CT_STATE_INVALID_BIT;
  83                 *dest = state;
  84                 return;
  85         default:
  86                 break;
  87         }
  88
  89         if (ct == NULL)
  90                 goto err;
  91
  92         switch (priv->key) {
  93         case NFT_CT_DIRECTION:
  94                 nft_reg_store8(dest, CTINFO2DIR(ctinfo));
  95                 return;
  96         case NFT_CT_STATUS:
  97                 *dest = ct->status;
  98                 return;
  99 #ifdef CONFIG_NF_CONNTRACK_MARK
 100         case NFT_CT_MARK:
 101                 *dest = READ_ONCE(ct->mark);
 102                 return;
 103 #endif
 104 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 105         case NFT_CT_SECMARK:
 106                 *dest = ct->secmark;
 107                 return;
 108 #endif
 109         case NFT_CT_EXPIRATION:
 110                 *dest = jiffies_to_msecs(nf_ct_expires(ct));
 111                 return;
 112         case NFT_CT_HELPER:
 113                 if (ct->master == NULL)
 114                         goto err;
 115                 help = nfct_help(ct->master);
 116                 if (help == NULL)
 117                         goto err;
 118                 helper = rcu_dereference(help->helper);
 119                 if (helper == NULL)
 120                         goto err;
 121                 strncpy((char *)dest, helper->name, NF_CT_HELPER_NAME_LEN);
 122                 return;
 123 #ifdef CONFIG_NF_CONNTRACK_LABELS
 124         case NFT_CT_LABELS: {
 125                 struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 126
 127                 if (labels)
 128                         memcpy(dest, labels->bits, NF_CT_LABELS_MAX_SIZE);
 129                 else
 130                         memset(dest, 0, NF_CT_LABELS_MAX_SIZE);
 131                 return;
 132         }
 133 #endif
 134         case NFT_CT_BYTES:
 135         case NFT_CT_PKTS: {
 136                 const struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 137                 u64 count = 0;
 138
 139                 if (acct)
 140                         count = nft_ct_get_eval_counter(acct->counter,
 141                                                         priv->key, priv->dir);
 142                 memcpy(dest, &count, sizeof(count));
 143                 return;
 144         }
 145         case NFT_CT_AVGPKT: {
 146                 const struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 147                 u64 avgcnt = 0, bcnt = 0, pcnt = 0;
 148
 149                 if (acct) {
 150                         pcnt = nft_ct_get_eval_counter(acct->counter,
 151                                                        NFT_CT_PKTS, priv->dir);
 152                         bcnt = nft_ct_get_eval_counter(acct->counter,
 153                                                        NFT_CT_BYTES, priv->dir);
 154                         if (pcnt != 0)
 155                                 avgcnt = div64_u64(bcnt, pcnt);
 156                 }
 157
 158                 memcpy(dest, &avgcnt, sizeof(avgcnt));
 159                 return;
 160         }
 161         case NFT_CT_L3PROTOCOL:
 162                 nft_reg_store8(dest, nf_ct_l3num(ct));
 163                 return;
 164         case NFT_CT_PROTOCOL:
 165                 nft_reg_store8(dest, nf_ct_protonum(ct));
 166                 return;
 167 #ifdef CONFIG_NF_CONNTRACK_ZONES
 168         case NFT_CT_ZONE: {
 169                 const struct nf_conntrack_zone *zone = nf_ct_zone(ct);
 170                 u16 zoneid;
 171
 172                 if (priv->dir < IP_CT_DIR_MAX)
 173                         zoneid = nf_ct_zone_id(zone, priv->dir);
 174                 else
 175                         zoneid = zone->id;
 176
 177                 nft_reg_store16(dest, zoneid);
 178                 return;
 179         }
 180 #endif
 181         case NFT_CT_ID:
 182                 *dest = nf_ct_get_id(ct);
 183                 return;
 184         default:
 185                 break;
 186         }
 187
 188         tuple = &ct->tuplehash[priv->dir].tuple;
 189         switch (priv->key) {
 190         case NFT_CT_SRC:
 191                 memcpy(dest, tuple->src.u3.all,
 192                        nf_ct_l3num(ct) == NFPROTO_IPV4 ? 4 : 16);
 193                 return;
 194         case NFT_CT_DST:
 195                 memcpy(dest, tuple->dst.u3.all,
 196                        nf_ct_l3num(ct) == NFPROTO_IPV4 ? 4 : 16);
 197                 return;
 198         case NFT_CT_PROTO_SRC:
 199                 nft_reg_store16(dest, (__force u16)tuple->src.u.all);
 200                 return;
 201         case NFT_CT_PROTO_DST:
 202                 nft_reg_store16(dest, (__force u16)tuple->dst.u.all);
 203                 return;
 204         case NFT_CT_SRC_IP:
 205                 if (nf_ct_l3num(ct) != NFPROTO_IPV4)
 206                         goto err;
 207                 *dest = (__force __u32)tuple->src.u3.ip;
 208                 return;
 209         case NFT_CT_DST_IP:
 210                 if (nf_ct_l3num(ct) != NFPROTO_IPV4)
 211                         goto err;
 212                 *dest = (__force __u32)tuple->dst.u3.ip;
 213                 return;
 214         case NFT_CT_SRC_IP6:
 215                 if (nf_ct_l3num(ct) != NFPROTO_IPV6)
 216                         goto err;
 217                 memcpy(dest, tuple->src.u3.ip6, sizeof(struct in6_addr));
 218                 return;
 219         case NFT_CT_DST_IP6:
 220                 if (nf_ct_l3num(ct) != NFPROTO_IPV6)
 221                         goto err;
 222                 memcpy(dest, tuple->dst.u3.ip6, sizeof(struct in6_addr));
 223                 return;
 224         default:
 225                 break;
 226         }
 227         return;
 228 err:
 229         regs->verdict.code = NFT_BREAK;
 230 }
 231
 232 #ifdef CONFIG_NF_CONNTRACK_ZONES
 233 static void nft_ct_set_zone_eval(const struct nft_expr *expr,
 234                                  struct nft_regs *regs,
 235                                  const struct nft_pktinfo *pkt)
 236 {
 237         struct nf_conntrack_zone zone = { .dir = NF_CT_DEFAULT_ZONE_DIR };
 238         const struct nft_ct *priv = nft_expr_priv(expr);
 239         struct sk_buff *skb = pkt->skb;
 240         enum ip_conntrack_info ctinfo;
 241         u16 value = nft_reg_load16(&regs->data[priv->sreg]);
 242         struct nf_conn *ct;
 243
 244         ct = nf_ct_get(skb, &ctinfo);
 245         if (ct) /* already tracked */
 246                 return;
 247
 248         zone.id = value;
 249
 250         switch (priv->dir) {
 251         case IP_CT_DIR_ORIGINAL:
 252                 zone.dir = NF_CT_ZONE_DIR_ORIG;
 253                 break;
 254         case IP_CT_DIR_REPLY:
 255                 zone.dir = NF_CT_ZONE_DIR_REPL;
 256                 break;
 257         default:
 258                 break;
 259         }
 260
 261         ct = this_cpu_read(nft_ct_pcpu_template);
 262
 263         if (likely(refcount_read(&ct->ct_general.use) == 1)) {
 264                 refcount_inc(&ct->ct_general.use);
 265                 nf_ct_zone_add(ct, &zone);
 266         } else {
 267                 /* previous skb got queued to userspace, allocate temporary
 268                  * one until percpu template can be reused.
 269                  */
 270                 ct = nf_ct_tmpl_alloc(nft_net(pkt), &zone, GFP_ATOMIC);
 271                 if (!ct) {
 272                         regs->verdict.code = NF_DROP;
 273                         return;
 274                 }
 275         }
 276
 277         nf_ct_set(skb, ct, IP_CT_NEW);
 278 }
 279 #endif
 280
 281 static void nft_ct_set_eval(const struct nft_expr *expr,
 282                             struct nft_regs *regs,
 283                             const struct nft_pktinfo *pkt)
 284 {
 285         const struct nft_ct *priv = nft_expr_priv(expr);
 286         struct sk_buff *skb = pkt->skb;
 287 #if defined(CONFIG_NF_CONNTRACK_MARK) || defined(CONFIG_NF_CONNTRACK_SECMARK)
 288         u32 value = regs->data[priv->sreg];
 289 #endif
 290         enum ip_conntrack_info ctinfo;
 291         struct nf_conn *ct;
 292
 293         ct = nf_ct_get(skb, &ctinfo);
 294         if (ct == NULL || nf_ct_is_template(ct))
 295                 return;
 296
 297         switch (priv->key) {
 298 #ifdef CONFIG_NF_CONNTRACK_MARK
 299         case NFT_CT_MARK:
 300                 if (READ_ONCE(ct->mark) != value) {
 301                         WRITE_ONCE(ct->mark, value);
 302                         nf_conntrack_event_cache(IPCT_MARK, ct);
 303                 }
 304                 break;
 305 #endif
 306 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 307         case NFT_CT_SECMARK:
 308                 if (ct->secmark != value) {
 309                         ct->secmark = value;
 310                         nf_conntrack_event_cache(IPCT_SECMARK, ct);
 311                 }
 312                 break;
 313 #endif
 314 #ifdef CONFIG_NF_CONNTRACK_LABELS
 315         case NFT_CT_LABELS:
 316                 nf_connlabels_replace(ct,
 317                                       &regs->data[priv->sreg],
 318                                       &regs->data[priv->sreg],
 319                                       NF_CT_LABELS_MAX_SIZE / sizeof(u32));
 320                 break;
 321 #endif
 322 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 323         case NFT_CT_EVENTMASK: {
 324                 struct nf_conntrack_ecache *e = nf_ct_ecache_find(ct);
 325                 u32 ctmask = regs->data[priv->sreg];
 326
 327                 if (e) {
 328                         if (e->ctmask != ctmask)
 329                                 e->ctmask = ctmask;
 330                         break;
 331                 }
 332
 333                 if (ctmask && !nf_ct_is_confirmed(ct))
 334                         nf_ct_ecache_ext_add(ct, ctmask, 0, GFP_ATOMIC);
 335                 break;
 336         }
 337 #endif
 338         default:
 339                 break;
 340         }
 341 }
 342
 343 static const struct nla_policy nft_ct_policy[NFTA_CT_MAX + 1] = {
 344         [NFTA_CT_DREG]          = { .type = NLA_U32 },
 345         [NFTA_CT_KEY]           = { .type = NLA_U32 },
 346         [NFTA_CT_DIRECTION]     = { .type = NLA_U8 },
 347         [NFTA_CT_SREG]          = { .type = NLA_U32 },
 348 };
 349
 350 #ifdef CONFIG_NF_CONNTRACK_ZONES
 351 static void nft_ct_tmpl_put_pcpu(void)
 352 {
 353         struct nf_conn *ct;
 354         int cpu;
 355
 356         for_each_possible_cpu(cpu) {
 357                 ct = per_cpu(nft_ct_pcpu_template, cpu);
 358                 if (!ct)
 359                         break;
 360                 nf_ct_put(ct);
 361                 per_cpu(nft_ct_pcpu_template, cpu) = NULL;
 362         }
 363 }
 364
 365 static bool nft_ct_tmpl_alloc_pcpu(void)
 366 {
 367         struct nf_conntrack_zone zone = { .id = 0 };
 368         struct nf_conn *tmp;
 369         int cpu;
 370
 371         if (nft_ct_pcpu_template_refcnt)
 372                 return true;
 373
 374         for_each_possible_cpu(cpu) {
 375                 tmp = nf_ct_tmpl_alloc(&init_net, &zone, GFP_KERNEL);
 376                 if (!tmp) {
 377                         nft_ct_tmpl_put_pcpu();
 378                         return false;
 379                 }
 380
 381                 per_cpu(nft_ct_pcpu_template, cpu) = tmp;
 382         }
 383
 384         return true;
 385 }
 386 #endif
 387
 388 static int nft_ct_get_init(const struct nft_ctx *ctx,
 389                            const struct nft_expr *expr,
 390                            const struct nlattr * const tb[])
 391 {
 392         struct nft_ct *priv = nft_expr_priv(expr);
 393         unsigned int len;
 394         int err;
 395
 396         priv->key = ntohl(nla_get_be32(tb[NFTA_CT_KEY]));
 397         priv->dir = IP_CT_DIR_MAX;
 398         switch (priv->key) {
 399         case NFT_CT_DIRECTION:
 400                 if (tb[NFTA_CT_DIRECTION] != NULL)
 401                         return -EINVAL;
 402                 len = sizeof(u8);
 403                 break;
 404         case NFT_CT_STATE:
 405         case NFT_CT_STATUS:
 406 #ifdef CONFIG_NF_CONNTRACK_MARK
 407         case NFT_CT_MARK:
 408 #endif
 409 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 410         case NFT_CT_SECMARK:
 411 #endif
 412         case NFT_CT_EXPIRATION:
 413                 if (tb[NFTA_CT_DIRECTION] != NULL)
 414                         return -EINVAL;
 415                 len = sizeof(u32);
 416                 break;
 417 #ifdef CONFIG_NF_CONNTRACK_LABELS
 418         case NFT_CT_LABELS:
 419                 if (tb[NFTA_CT_DIRECTION] != NULL)
 420                         return -EINVAL;
 421                 len = NF_CT_LABELS_MAX_SIZE;
 422                 break;
 423 #endif
 424         case NFT_CT_HELPER:
 425                 if (tb[NFTA_CT_DIRECTION] != NULL)
 426                         return -EINVAL;
 427                 len = NF_CT_HELPER_NAME_LEN;
 428                 break;
 429
 430         case NFT_CT_L3PROTOCOL:
 431         case NFT_CT_PROTOCOL:
 432                 /* For compatibility, do not report error if NFTA_CT_DIRECTION
 433                  * attribute is specified.
 434                  */
 435                 len = sizeof(u8);
 436                 break;
 437         case NFT_CT_SRC:
 438         case NFT_CT_DST:
 439                 if (tb[NFTA_CT_DIRECTION] == NULL)
 440                         return -EINVAL;
 441
 442                 switch (ctx->family) {
 443                 case NFPROTO_IPV4:
 444                         len = sizeof_field(struct nf_conntrack_tuple,
 445                                            src.u3.ip);
 446                         break;
 447                 case NFPROTO_IPV6:
 448                 case NFPROTO_INET:
 449                         len = sizeof_field(struct nf_conntrack_tuple,
 450                                            src.u3.ip6);
 451                         break;
 452                 default:
 453                         return -EAFNOSUPPORT;
 454                 }
 455                 break;
 456         case NFT_CT_SRC_IP:
 457         case NFT_CT_DST_IP:
 458                 if (tb[NFTA_CT_DIRECTION] == NULL)
 459                         return -EINVAL;
 460
 461                 len = sizeof_field(struct nf_conntrack_tuple, src.u3.ip);
 462                 break;
 463         case NFT_CT_SRC_IP6:
 464         case NFT_CT_DST_IP6:
 465                 if (tb[NFTA_CT_DIRECTION] == NULL)
 466                         return -EINVAL;
 467
 468                 len = sizeof_field(struct nf_conntrack_tuple, src.u3.ip6);
 469                 break;
 470         case NFT_CT_PROTO_SRC:
 471         case NFT_CT_PROTO_DST:
 472                 if (tb[NFTA_CT_DIRECTION] == NULL)
 473                         return -EINVAL;
 474                 len = sizeof_field(struct nf_conntrack_tuple, src.u.all);
 475                 break;
 476         case NFT_CT_BYTES:
 477         case NFT_CT_PKTS:
 478         case NFT_CT_AVGPKT:
 479                 len = sizeof(u64);
 480                 break;
 481 #ifdef CONFIG_NF_CONNTRACK_ZONES
 482         case NFT_CT_ZONE:
 483                 len = sizeof(u16);
 484                 break;
 485 #endif
 486         case NFT_CT_ID:
 487                 len = sizeof(u32);
 488                 break;
 489         default:
 490                 return -EOPNOTSUPP;
 491         }
 492
 493         if (tb[NFTA_CT_DIRECTION] != NULL) {
 494                 priv->dir = nla_get_u8(tb[NFTA_CT_DIRECTION]);
 495                 switch (priv->dir) {
 496                 case IP_CT_DIR_ORIGINAL:
 497                 case IP_CT_DIR_REPLY:
 498                         break;
 499                 default:
 500                         return -EINVAL;
 501                 }
 502         }
 503
 504         priv->len = len;
 505         err = nft_parse_register_store(ctx, tb[NFTA_CT_DREG], &priv->dreg, NULL,
 506                                        NFT_DATA_VALUE, len);
 507         if (err < 0)
 508                 return err;
 509
 510         err = nf_ct_netns_get(ctx->net, ctx->family);
 511         if (err < 0)
 512                 return err;
 513
 514         if (priv->key == NFT_CT_BYTES ||
 515             priv->key == NFT_CT_PKTS  ||
 516             priv->key == NFT_CT_AVGPKT)
 517                 nf_ct_set_acct(ctx->net, true);
 518
 519         return 0;
 520 }
 521
 522 static void __nft_ct_set_destroy(const struct nft_ctx *ctx, struct nft_ct *priv)
 523 {
 524         switch (priv->key) {
 525 #ifdef CONFIG_NF_CONNTRACK_LABELS
 526         case NFT_CT_LABELS:
 527                 nf_connlabels_put(ctx->net);
 528                 break;
 529 #endif
 530 #ifdef CONFIG_NF_CONNTRACK_ZONES
 531         case NFT_CT_ZONE:
 532                 mutex_lock(&nft_ct_pcpu_mutex);
 533                 if (--nft_ct_pcpu_template_refcnt == 0)
 534                         nft_ct_tmpl_put_pcpu();
 535                 mutex_unlock(&nft_ct_pcpu_mutex);
 536                 break;
 537 #endif
 538         default:
 539                 break;
 540         }
 541 }
 542
 543 static int nft_ct_set_init(const struct nft_ctx *ctx,
 544                            const struct nft_expr *expr,
 545                            const struct nlattr * const tb[])
 546 {
 547         struct nft_ct *priv = nft_expr_priv(expr);
 548         unsigned int len;
 549         int err;
 550
 551         priv->dir = IP_CT_DIR_MAX;
 552         priv->key = ntohl(nla_get_be32(tb[NFTA_CT_KEY]));
 553         switch (priv->key) {
 554 #ifdef CONFIG_NF_CONNTRACK_MARK
 555         case NFT_CT_MARK:
 556                 if (tb[NFTA_CT_DIRECTION])
 557                         return -EINVAL;
 558                 len = sizeof_field(struct nf_conn, mark);
 559                 break;
 560 #endif
 561 #ifdef CONFIG_NF_CONNTRACK_LABELS
 562         case NFT_CT_LABELS:
 563                 if (tb[NFTA_CT_DIRECTION])
 564                         return -EINVAL;
 565                 len = NF_CT_LABELS_MAX_SIZE;
 566                 err = nf_connlabels_get(ctx->net, (len * BITS_PER_BYTE) - 1);
 567                 if (err)
 568                         return err;
 569                 break;
 570 #endif
 571 #ifdef CONFIG_NF_CONNTRACK_ZONES
 572         case NFT_CT_ZONE:
 573                 mutex_lock(&nft_ct_pcpu_mutex);
 574                 if (!nft_ct_tmpl_alloc_pcpu()) {
 575                         mutex_unlock(&nft_ct_pcpu_mutex);
 576                         return -ENOMEM;
 577                 }
 578                 nft_ct_pcpu_template_refcnt++;
 579                 mutex_unlock(&nft_ct_pcpu_mutex);
 580                 len = sizeof(u16);
 581                 break;
 582 #endif
 583 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 584         case NFT_CT_EVENTMASK:
 585                 if (tb[NFTA_CT_DIRECTION])
 586                         return -EINVAL;
 587                 len = sizeof(u32);
 588                 break;
 589 #endif
 590 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 591         case NFT_CT_SECMARK:
 592                 if (tb[NFTA_CT_DIRECTION])
 593                         return -EINVAL;
 594                 len = sizeof(u32);
 595                 break;
 596 #endif
 597         default:
 598                 return -EOPNOTSUPP;
 599         }
 600
 601         if (tb[NFTA_CT_DIRECTION]) {
 602                 priv->dir = nla_get_u8(tb[NFTA_CT_DIRECTION]);
 603                 switch (priv->dir) {
 604                 case IP_CT_DIR_ORIGINAL:
 605                 case IP_CT_DIR_REPLY:
 606                         break;
 607                 default:
 608                         err = -EINVAL;
 609                         goto err1;
 610                 }
 611         }
 612
 613         priv->len = len;
 614         err = nft_parse_register_load(tb[NFTA_CT_SREG], &priv->sreg, len);
 615         if (err < 0)
 616                 goto err1;
 617
 618         err = nf_ct_netns_get(ctx->net, ctx->family);
 619         if (err < 0)
 620                 goto err1;
 621
 622         return 0;
 623
 624 err1:
 625         __nft_ct_set_destroy(ctx, priv);
 626         return err;
 627 }
 628
 629 static void nft_ct_get_destroy(const struct nft_ctx *ctx,
 630                                const struct nft_expr *expr)
 631 {
 632         nf_ct_netns_put(ctx->net, ctx->family);
 633 }
 634
 635 static void nft_ct_set_destroy(const struct nft_ctx *ctx,
 636                                const struct nft_expr *expr)
 637 {
 638         struct nft_ct *priv = nft_expr_priv(expr);
 639
 640         __nft_ct_set_destroy(ctx, priv);
 641         nf_ct_netns_put(ctx->net, ctx->family);
 642 }
 643
 644 static int nft_ct_get_dump(struct sk_buff *skb,
 645                            const struct nft_expr *expr, bool reset)
 646 {
 647         const struct nft_ct *priv = nft_expr_priv(expr);
 648
 649         if (nft_dump_register(skb, NFTA_CT_DREG, priv->dreg))
 650                 goto nla_put_failure;
 651         if (nla_put_be32(skb, NFTA_CT_KEY, htonl(priv->key)))
 652                 goto nla_put_failure;
 653
 654         switch (priv->key) {
 655         case NFT_CT_SRC:
 656         case NFT_CT_DST:
 657         case NFT_CT_SRC_IP:
 658         case NFT_CT_DST_IP:
 659         case NFT_CT_SRC_IP6:
 660         case NFT_CT_DST_IP6:
 661         case NFT_CT_PROTO_SRC:
 662         case NFT_CT_PROTO_DST:
 663                 if (nla_put_u8(skb, NFTA_CT_DIRECTION, priv->dir))
 664                         goto nla_put_failure;
 665                 break;
 666         case NFT_CT_BYTES:
 667         case NFT_CT_PKTS:
 668         case NFT_CT_AVGPKT:
 669         case NFT_CT_ZONE:
 670                 if (priv->dir < IP_CT_DIR_MAX &&
 671                     nla_put_u8(skb, NFTA_CT_DIRECTION, priv->dir))
 672                         goto nla_put_failure;
 673                 break;
 674         default:
 675                 break;
 676         }
 677
 678         return 0;
 679
 680 nla_put_failure:
 681         return -1;
 682 }
 683
 684 static bool nft_ct_get_reduce(struct nft_regs_track *track,
 685                               const struct nft_expr *expr)
 686 {
 687         const struct nft_ct *priv = nft_expr_priv(expr);
 688         const struct nft_ct *ct;
 689
 690         if (!nft_reg_track_cmp(track, expr, priv->dreg)) {
 691                 nft_reg_track_update(track, expr, priv->dreg, priv->len);
 692                 return false;
 693         }
 694
 695         ct = nft_expr_priv(track->regs[priv->dreg].selector);
 696         if (priv->key != ct->key) {
 697                 nft_reg_track_update(track, expr, priv->dreg, priv->len);
 698                 return false;
 699         }
 700
 701         if (!track->regs[priv->dreg].bitwise)
 702                 return true;
 703
 704         return nft_expr_reduce_bitwise(track, expr);
 705 }
 706
 707 static int nft_ct_set_dump(struct sk_buff *skb,
 708                            const struct nft_expr *expr, bool reset)
 709 {
 710         const struct nft_ct *priv = nft_expr_priv(expr);
 711
 712         if (nft_dump_register(skb, NFTA_CT_SREG, priv->sreg))
 713                 goto nla_put_failure;
 714         if (nla_put_be32(skb, NFTA_CT_KEY, htonl(priv->key)))
 715                 goto nla_put_failure;
 716
 717         switch (priv->key) {
 718         case NFT_CT_ZONE:
 719                 if (priv->dir < IP_CT_DIR_MAX &&
 720                     nla_put_u8(skb, NFTA_CT_DIRECTION, priv->dir))
 721                         goto nla_put_failure;
 722                 break;
 723         default:
 724                 break;
 725         }
 726
 727         return 0;
 728
 729 nla_put_failure:
 730         return -1;
 731 }
 732
 733 static struct nft_expr_type nft_ct_type;
 734 static const struct nft_expr_ops nft_ct_get_ops = {
 735         .type           = &nft_ct_type,
 736         .size           = NFT_EXPR_SIZE(sizeof(struct nft_ct)),
 737         .eval           = nft_ct_get_eval,
 738         .init           = nft_ct_get_init,
 739         .destroy        = nft_ct_get_destroy,
 740         .dump           = nft_ct_get_dump,
 741         .reduce         = nft_ct_get_reduce,
 742 };
 743
 744 static bool nft_ct_set_reduce(struct nft_regs_track *track,
 745                               const struct nft_expr *expr)
 746 {
 747         int i;
 748
 749         for (i = 0; i < NFT_REG32_NUM; i++) {
 750                 if (!track->regs[i].selector)
 751                         continue;
 752
 753                 if (track->regs[i].selector->ops != &nft_ct_get_ops)
 754                         continue;
 755
 756                 __nft_reg_track_cancel(track, i);
 757         }
 758
 759         return false;
 760 }
 761
 762 static const struct nft_expr_ops nft_ct_set_ops = {
 763         .type           = &nft_ct_type,
 764         .size           = NFT_EXPR_SIZE(sizeof(struct nft_ct)),
 765         .eval           = nft_ct_set_eval,
 766         .init           = nft_ct_set_init,
 767         .destroy        = nft_ct_set_destroy,
 768         .dump           = nft_ct_set_dump,
 769         .reduce         = nft_ct_set_reduce,
 770 };
 771
 772 #ifdef CONFIG_NF_CONNTRACK_ZONES
 773 static const struct nft_expr_ops nft_ct_set_zone_ops = {
 774         .type           = &nft_ct_type,
 775         .size           = NFT_EXPR_SIZE(sizeof(struct nft_ct)),
 776         .eval           = nft_ct_set_zone_eval,
 777         .init           = nft_ct_set_init,
 778         .destroy        = nft_ct_set_destroy,
 779         .dump           = nft_ct_set_dump,
 780         .reduce         = nft_ct_set_reduce,
 781 };
 782 #endif
 783
 784 static const struct nft_expr_ops *
 785 nft_ct_select_ops(const struct nft_ctx *ctx,
 786                     const struct nlattr * const tb[])
 787 {
 788         if (tb[NFTA_CT_KEY] == NULL)
 789                 return ERR_PTR(-EINVAL);
 790
 791         if (tb[NFTA_CT_DREG] && tb[NFTA_CT_SREG])
 792                 return ERR_PTR(-EINVAL);
 793
 794         if (tb[NFTA_CT_DREG])
 795                 return &nft_ct_get_ops;
 796
 797         if (tb[NFTA_CT_SREG]) {
 798 #ifdef CONFIG_NF_CONNTRACK_ZONES
 799                 if (nla_get_be32(tb[NFTA_CT_KEY]) == htonl(NFT_CT_ZONE))
 800                         return &nft_ct_set_zone_ops;
 801 #endif
 802                 return &nft_ct_set_ops;
 803         }
 804
 805         return ERR_PTR(-EINVAL);
 806 }
 807
 808 static struct nft_expr_type nft_ct_type __read_mostly = {
 809         .name           = "ct",
 810         .select_ops     = nft_ct_select_ops,
 811         .policy         = nft_ct_policy,
 812         .maxattr        = NFTA_CT_MAX,
 813         .owner          = THIS_MODULE,
 814 };
 815
 816 static void nft_notrack_eval(const struct nft_expr *expr,
 817                              struct nft_regs *regs,
 818                              const struct nft_pktinfo *pkt)
 819 {
 820         struct sk_buff *skb = pkt->skb;
 821         enum ip_conntrack_info ctinfo;
 822         struct nf_conn *ct;
 823
 824         ct = nf_ct_get(pkt->skb, &ctinfo);
 825         /* Previously seen (loopback or untracked)?  Ignore. */
 826         if (ct || ctinfo == IP_CT_UNTRACKED)
 827                 return;
 828
 829         nf_ct_set(skb, ct, IP_CT_UNTRACKED);
 830 }
 831
 832 static struct nft_expr_type nft_notrack_type;
 833 static const struct nft_expr_ops nft_notrack_ops = {
 834         .type           = &nft_notrack_type,
 835         .size           = NFT_EXPR_SIZE(0),
 836         .eval           = nft_notrack_eval,
 837         .reduce         = NFT_REDUCE_READONLY,
 838 };
 839
 840 static struct nft_expr_type nft_notrack_type __read_mostly = {
 841         .name           = "notrack",
 842         .ops            = &nft_notrack_ops,
 843         .owner          = THIS_MODULE,
 844 };
 845
 846 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
 847 static int
 848 nft_ct_timeout_parse_policy(void *timeouts,
 849                             const struct nf_conntrack_l4proto *l4proto,
 850                             struct net *net, const struct nlattr *attr)
 851 {
 852         struct nlattr **tb;
 853         int ret = 0;
 854
 855         tb = kcalloc(l4proto->ctnl_timeout.nlattr_max + 1, sizeof(*tb),
 856                      GFP_KERNEL);
 857
 858         if (!tb)
 859                 return -ENOMEM;
 860
 861         ret = nla_parse_nested_deprecated(tb,
 862                                           l4proto->ctnl_timeout.nlattr_max,
 863                                           attr,
 864                                           l4proto->ctnl_timeout.nla_policy,
 865                                           NULL);
 866         if (ret < 0)
 867                 goto err;
 868
 869         ret = l4proto->ctnl_timeout.nlattr_to_obj(tb, net, timeouts);
 870
 871 err:
 872         kfree(tb);
 873         return ret;
 874 }
 875
 876 struct nft_ct_timeout_obj {
 877         struct nf_ct_timeout    *timeout;
 878         u8                      l4proto;
 879 };
 880
 881 static void nft_ct_timeout_obj_eval(struct nft_object *obj,
 882                                     struct nft_regs *regs,
 883                                     const struct nft_pktinfo *pkt)
 884 {
 885         const struct nft_ct_timeout_obj *priv = nft_obj_data(obj);
 886         struct nf_conn *ct = (struct nf_conn *)skb_nfct(pkt->skb);
 887         struct nf_conn_timeout *timeout;
 888         const unsigned int *values;
 889
 890         if (priv->l4proto != pkt->tprot)
 891                 return;
 892
 893         if (!ct || nf_ct_is_template(ct) || nf_ct_is_confirmed(ct))
 894                 return;
 895
 896         timeout = nf_ct_timeout_find(ct);
 897         if (!timeout) {
 898                 timeout = nf_ct_timeout_ext_add(ct, priv->timeout, GFP_ATOMIC);
 899                 if (!timeout) {
 900                         regs->verdict.code = NF_DROP;
 901                         return;
 902                 }
 903         }
 904
 905         rcu_assign_pointer(timeout->timeout, priv->timeout);
 906
 907         /* adjust the timeout as per 'new' state. ct is unconfirmed,
 908          * so the current timestamp must not be added.
 909          */
 910         values = nf_ct_timeout_data(timeout);
 911         if (values)
 912                 nf_ct_refresh(ct, pkt->skb, values[0]);
 913 }
 914
 915 static int nft_ct_timeout_obj_init(const struct nft_ctx *ctx,
 916                                    const struct nlattr * const tb[],
 917                                    struct nft_object *obj)
 918 {
 919         struct nft_ct_timeout_obj *priv = nft_obj_data(obj);
 920         const struct nf_conntrack_l4proto *l4proto;
 921         struct nf_ct_timeout *timeout;
 922         int l3num = ctx->family;
 923         __u8 l4num;
 924         int ret;
 925
 926         if (!tb[NFTA_CT_TIMEOUT_L4PROTO] ||
 927             !tb[NFTA_CT_TIMEOUT_DATA])
 928                 return -EINVAL;
 929
 930         if (tb[NFTA_CT_TIMEOUT_L3PROTO])
 931                 l3num = ntohs(nla_get_be16(tb[NFTA_CT_TIMEOUT_L3PROTO]));
 932
 933         l4num = nla_get_u8(tb[NFTA_CT_TIMEOUT_L4PROTO]);
 934         priv->l4proto = l4num;
 935
 936         l4proto = nf_ct_l4proto_find(l4num);
 937
 938         if (l4proto->l4proto != l4num) {
 939                 ret = -EOPNOTSUPP;
 940                 goto err_proto_put;
 941         }
 942
 943         timeout = kzalloc(sizeof(struct nf_ct_timeout) +
 944                           l4proto->ctnl_timeout.obj_size, GFP_KERNEL);
 945         if (timeout == NULL) {
 946                 ret = -ENOMEM;
 947                 goto err_proto_put;
 948         }
 949
 950         ret = nft_ct_timeout_parse_policy(&timeout->data, l4proto, ctx->net,
 951                                           tb[NFTA_CT_TIMEOUT_DATA]);
 952         if (ret < 0)
 953                 goto err_free_timeout;
 954
 955         timeout->l3num = l3num;
 956         timeout->l4proto = l4proto;
 957
 958         ret = nf_ct_netns_get(ctx->net, ctx->family);
 959         if (ret < 0)
 960                 goto err_free_timeout;
 961
 962         priv->timeout = timeout;
 963         return 0;
 964
 965 err_free_timeout:
 966         kfree(timeout);
 967 err_proto_put:
 968         return ret;
 969 }
 970
 971 static void nft_ct_timeout_obj_destroy(const struct nft_ctx *ctx,
 972                                        struct nft_object *obj)
 973 {
 974         struct nft_ct_timeout_obj *priv = nft_obj_data(obj);
 975         struct nf_ct_timeout *timeout = priv->timeout;
 976
 977         nf_ct_untimeout(ctx->net, timeout);
 978         nf_ct_netns_put(ctx->net, ctx->family);
 979         kfree(priv->timeout);
 980 }
 981
 982 static int nft_ct_timeout_obj_dump(struct sk_buff *skb,
 983                                    struct nft_object *obj, bool reset)
 984 {
 985         const struct nft_ct_timeout_obj *priv = nft_obj_data(obj);
 986         const struct nf_ct_timeout *timeout = priv->timeout;
 987         struct nlattr *nest_params;
 988         int ret;
 989
 990         if (nla_put_u8(skb, NFTA_CT_TIMEOUT_L4PROTO, timeout->l4proto->l4proto) ||
 991             nla_put_be16(skb, NFTA_CT_TIMEOUT_L3PROTO, htons(timeout->l3num)))
 992                 return -1;
 993
 994         nest_params = nla_nest_start(skb, NFTA_CT_TIMEOUT_DATA);
 995         if (!nest_params)
 996                 return -1;
 997
 998         ret = timeout->l4proto->ctnl_timeout.obj_to_nlattr(skb, &timeout->data);
 999         if (ret < 0)
1000                 return -1;
1001         nla_nest_end(skb, nest_params);
1002         return 0;
1003 }
1004
1005 static const struct nla_policy nft_ct_timeout_policy[NFTA_CT_TIMEOUT_MAX + 1] = {
1006         [NFTA_CT_TIMEOUT_L3PROTO] = {.type = NLA_U16 },
1007         [NFTA_CT_TIMEOUT_L4PROTO] = {.type = NLA_U8 },
1008         [NFTA_CT_TIMEOUT_DATA]    = {.type = NLA_NESTED },
1009 };
1010
1011 static struct nft_object_type nft_ct_timeout_obj_type;
1012
1013 static const struct nft_object_ops nft_ct_timeout_obj_ops = {
1014         .type           = &nft_ct_timeout_obj_type,
1015         .size           = sizeof(struct nft_ct_timeout_obj),
1016         .eval           = nft_ct_timeout_obj_eval,
1017         .init           = nft_ct_timeout_obj_init,
1018         .destroy        = nft_ct_timeout_obj_destroy,
1019         .dump           = nft_ct_timeout_obj_dump,
1020 };
1021
1022 static struct nft_object_type nft_ct_timeout_obj_type __read_mostly = {
1023         .type           = NFT_OBJECT_CT_TIMEOUT,
1024         .ops            = &nft_ct_timeout_obj_ops,
1025         .maxattr        = NFTA_CT_TIMEOUT_MAX,
1026         .policy         = nft_ct_timeout_policy,
1027         .owner          = THIS_MODULE,
1028 };
1029 #endif /* CONFIG_NF_CONNTRACK_TIMEOUT */
1030
1031 static int nft_ct_helper_obj_init(const struct nft_ctx *ctx,
1032                                   const struct nlattr * const tb[],
1033                                   struct nft_object *obj)
1034 {
1035         struct nft_ct_helper_obj *priv = nft_obj_data(obj);
1036         struct nf_conntrack_helper *help4, *help6;
1037         char name[NF_CT_HELPER_NAME_LEN];
1038         int family = ctx->family;
1039         int err;
1040
1041         if (!tb[NFTA_CT_HELPER_NAME] || !tb[NFTA_CT_HELPER_L4PROTO])
1042                 return -EINVAL;
1043
1044         priv->l4proto = nla_get_u8(tb[NFTA_CT_HELPER_L4PROTO]);
1045         if (!priv->l4proto)
1046                 return -ENOENT;
1047
1048         nla_strscpy(name, tb[NFTA_CT_HELPER_NAME], sizeof(name));
1049
1050         if (tb[NFTA_CT_HELPER_L3PROTO])
1051                 family = ntohs(nla_get_be16(tb[NFTA_CT_HELPER_L3PROTO]));
1052
1053         help4 = NULL;
1054         help6 = NULL;
1055
1056         switch (family) {
1057         case NFPROTO_IPV4:
1058                 if (ctx->family == NFPROTO_IPV6)
1059                         return -EINVAL;
1060
1061                 help4 = nf_conntrack_helper_try_module_get(name, family,
1062                                                            priv->l4proto);
1063                 break;
1064         case NFPROTO_IPV6:
1065                 if (ctx->family == NFPROTO_IPV4)
1066                         return -EINVAL;
1067
1068                 help6 = nf_conntrack_helper_try_module_get(name, family,
1069                                                            priv->l4proto);
1070                 break;
1071         case NFPROTO_NETDEV:
1072         case NFPROTO_BRIDGE:
1073         case NFPROTO_INET:
1074                 help4 = nf_conntrack_helper_try_module_get(name, NFPROTO_IPV4,
1075                                                            priv->l4proto);
1076                 help6 = nf_conntrack_helper_try_module_get(name, NFPROTO_IPV6,
1077                                                            priv->l4proto);
1078                 break;
1079         default:
1080                 return -EAFNOSUPPORT;
1081         }
1082
1083         /* && is intentional; only error if INET found neither ipv4 or ipv6 */
1084         if (!help4 && !help6)
1085                 return -ENOENT;
1086
1087         priv->helper4 = help4;
1088         priv->helper6 = help6;
1089
1090         err = nf_ct_netns_get(ctx->net, ctx->family);
1091         if (err < 0)
1092                 goto err_put_helper;
1093
1094         return 0;
1095
1096 err_put_helper:
1097         if (priv->helper4)
1098                 nf_conntrack_helper_put(priv->helper4);
1099         if (priv->helper6)
1100                 nf_conntrack_helper_put(priv->helper6);
1101         return err;
1102 }
1103
1104 static void nft_ct_helper_obj_destroy(const struct nft_ctx *ctx,
1105                                       struct nft_object *obj)
1106 {
1107         struct nft_ct_helper_obj *priv = nft_obj_data(obj);
1108
1109         if (priv->helper4)
1110                 nf_conntrack_helper_put(priv->helper4);
1111         if (priv->helper6)
1112                 nf_conntrack_helper_put(priv->helper6);
1113
1114         nf_ct_netns_put(ctx->net, ctx->family);
1115 }
1116
1117 static void nft_ct_helper_obj_eval(struct nft_object *obj,
1118                                    struct nft_regs *regs,
1119                                    const struct nft_pktinfo *pkt)
1120 {
1121         const struct nft_ct_helper_obj *priv = nft_obj_data(obj);
1122         struct nf_conn *ct = (struct nf_conn *)skb_nfct(pkt->skb);
1123         struct nf_conntrack_helper *to_assign = NULL;
1124         struct nf_conn_help *help;
1125
1126         if (!ct ||
1127             nf_ct_is_confirmed(ct) ||
1128             nf_ct_is_template(ct) ||
1129             priv->l4proto != nf_ct_protonum(ct))
1130                 return;
1131
1132         switch (nf_ct_l3num(ct)) {
1133         case NFPROTO_IPV4:
1134                 to_assign = priv->helper4;
1135                 break;
1136         case NFPROTO_IPV6:
1137                 to_assign = priv->helper6;
1138                 break;
1139         default:
1140                 WARN_ON_ONCE(1);
1141                 return;
1142         }
1143
1144         if (!to_assign)
1145                 return;
1146
1147         if (test_bit(IPS_HELPER_BIT, &ct->status))
1148                 return;
1149
1150         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1151         if (help) {
1152                 rcu_assign_pointer(help->helper, to_assign);
1153                 set_bit(IPS_HELPER_BIT, &ct->status);
1154         }
1155 }
1156
1157 static int nft_ct_helper_obj_dump(struct sk_buff *skb,
1158                                   struct nft_object *obj, bool reset)
1159 {
1160         const struct nft_ct_helper_obj *priv = nft_obj_data(obj);
1161         const struct nf_conntrack_helper *helper;
1162         u16 family;
1163
1164         if (priv->helper4 && priv->helper6) {
1165                 family = NFPROTO_INET;
1166                 helper = priv->helper4;
1167         } else if (priv->helper6) {
1168                 family = NFPROTO_IPV6;
1169                 helper = priv->helper6;
1170         } else {
1171                 family = NFPROTO_IPV4;
1172                 helper = priv->helper4;
1173         }
1174
1175         if (nla_put_string(skb, NFTA_CT_HELPER_NAME, helper->name))
1176                 return -1;
1177
1178         if (nla_put_u8(skb, NFTA_CT_HELPER_L4PROTO, priv->l4proto))
1179                 return -1;
1180
1181         if (nla_put_be16(skb, NFTA_CT_HELPER_L3PROTO, htons(family)))
1182                 return -1;
1183
1184         return 0;
1185 }
1186
1187 static const struct nla_policy nft_ct_helper_policy[NFTA_CT_HELPER_MAX + 1] = {
1188         [NFTA_CT_HELPER_NAME] = { .type = NLA_STRING,
1189                                   .len = NF_CT_HELPER_NAME_LEN - 1 },
1190         [NFTA_CT_HELPER_L3PROTO] = { .type = NLA_U16 },
1191         [NFTA_CT_HELPER_L4PROTO] = { .type = NLA_U8 },
1192 };
1193
1194 static struct nft_object_type nft_ct_helper_obj_type;
1195 static const struct nft_object_ops nft_ct_helper_obj_ops = {
1196         .type           = &nft_ct_helper_obj_type,
1197         .size           = sizeof(struct nft_ct_helper_obj),
1198         .eval           = nft_ct_helper_obj_eval,
1199         .init           = nft_ct_helper_obj_init,
1200         .destroy        = nft_ct_helper_obj_destroy,
1201         .dump           = nft_ct_helper_obj_dump,
1202 };
1203
1204 static struct nft_object_type nft_ct_helper_obj_type __read_mostly = {
1205         .type           = NFT_OBJECT_CT_HELPER,
1206         .ops            = &nft_ct_helper_obj_ops,
1207         .maxattr        = NFTA_CT_HELPER_MAX,
1208         .policy         = nft_ct_helper_policy,
1209         .owner          = THIS_MODULE,
1210 };
1211
1212 struct nft_ct_expect_obj {
1213         u16             l3num;
1214         __be16          dport;
1215         u8              l4proto;
1216         u8              size;
1217         u32             timeout;
1218 };
1219
1220 static int nft_ct_expect_obj_init(const struct nft_ctx *ctx,
1221                                   const struct nlattr * const tb[],
1222                                   struct nft_object *obj)
1223 {
1224         struct nft_ct_expect_obj *priv = nft_obj_data(obj);
1225
1226         if (!tb[NFTA_CT_EXPECT_L4PROTO] ||
1227             !tb[NFTA_CT_EXPECT_DPORT] ||
1228             !tb[NFTA_CT_EXPECT_TIMEOUT] ||
1229             !tb[NFTA_CT_EXPECT_SIZE])
1230                 return -EINVAL;
1231
1232         priv->l3num = ctx->family;
1233         if (tb[NFTA_CT_EXPECT_L3PROTO])
1234                 priv->l3num = ntohs(nla_get_be16(tb[NFTA_CT_EXPECT_L3PROTO]));
1235
1236         priv->l4proto = nla_get_u8(tb[NFTA_CT_EXPECT_L4PROTO]);
1237         priv->dport = nla_get_be16(tb[NFTA_CT_EXPECT_DPORT]);
1238         priv->timeout = nla_get_u32(tb[NFTA_CT_EXPECT_TIMEOUT]);
1239         priv->size = nla_get_u8(tb[NFTA_CT_EXPECT_SIZE]);
1240
1241         return nf_ct_netns_get(ctx->net, ctx->family);
1242 }
1243
1244 static void nft_ct_expect_obj_destroy(const struct nft_ctx *ctx,
1245                                        struct nft_object *obj)
1246 {
1247         nf_ct_netns_put(ctx->net, ctx->family);
1248 }
1249
1250 static int nft_ct_expect_obj_dump(struct sk_buff *skb,
1251                                   struct nft_object *obj, bool reset)
1252 {
1253         const struct nft_ct_expect_obj *priv = nft_obj_data(obj);
1254
1255         if (nla_put_be16(skb, NFTA_CT_EXPECT_L3PROTO, htons(priv->l3num)) ||
1256             nla_put_u8(skb, NFTA_CT_EXPECT_L4PROTO, priv->l4proto) ||
1257             nla_put_be16(skb, NFTA_CT_EXPECT_DPORT, priv->dport) ||
1258             nla_put_u32(skb, NFTA_CT_EXPECT_TIMEOUT, priv->timeout) ||
1259             nla_put_u8(skb, NFTA_CT_EXPECT_SIZE, priv->size))
1260                 return -1;
1261
1262         return 0;
1263 }
1264
1265 static void nft_ct_expect_obj_eval(struct nft_object *obj,
1266                                    struct nft_regs *regs,
1267                                    const struct nft_pktinfo *pkt)
1268 {
1269         const struct nft_ct_expect_obj *priv = nft_obj_data(obj);
1270         struct nf_conntrack_expect *exp;
1271         enum ip_conntrack_info ctinfo;
1272         struct nf_conn_help *help;
1273         enum ip_conntrack_dir dir;
1274         u16 l3num = priv->l3num;
1275         struct nf_conn *ct;
1276
1277         ct = nf_ct_get(pkt->skb, &ctinfo);
1278         if (!ct || nf_ct_is_confirmed(ct) || nf_ct_is_template(ct)) {
1279                 regs->verdict.code = NFT_BREAK;
1280                 return;
1281         }
1282         dir = CTINFO2DIR(ctinfo);
1283
1284         help = nfct_help(ct);
1285         if (!help)
1286                 help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1287         if (!help) {
1288                 regs->verdict.code = NF_DROP;
1289                 return;
1290         }
1291
1292         if (help->expecting[NF_CT_EXPECT_CLASS_DEFAULT] >= priv->size) {
1293                 regs->verdict.code = NFT_BREAK;
1294                 return;
1295         }
1296         if (l3num == NFPROTO_INET)
1297                 l3num = nf_ct_l3num(ct);
1298
1299         exp = nf_ct_expect_alloc(ct);
1300         if (exp == NULL) {
1301                 regs->verdict.code = NF_DROP;
1302                 return;
1303         }
1304         nf_ct_expect_init(exp, NF_CT_EXPECT_CLASS_DEFAULT, l3num,
1305                           &ct->tuplehash[!dir].tuple.src.u3,
1306                           &ct->tuplehash[!dir].tuple.dst.u3,
1307                           priv->l4proto, NULL, &priv->dport);
1308         exp->timeout.expires = jiffies + priv->timeout * HZ;
1309
1310         if (nf_ct_expect_related(exp, 0) != 0)
1311                 regs->verdict.code = NF_DROP;
1312 }
1313
1314 static const struct nla_policy nft_ct_expect_policy[NFTA_CT_EXPECT_MAX + 1] = {
1315         [NFTA_CT_EXPECT_L3PROTO]        = { .type = NLA_U16 },
1316         [NFTA_CT_EXPECT_L4PROTO]        = { .type = NLA_U8 },
1317         [NFTA_CT_EXPECT_DPORT]          = { .type = NLA_U16 },
1318         [NFTA_CT_EXPECT_TIMEOUT]        = { .type = NLA_U32 },
1319         [NFTA_CT_EXPECT_SIZE]           = { .type = NLA_U8 },
1320 };
1321
1322 static struct nft_object_type nft_ct_expect_obj_type;
1323
1324 static const struct nft_object_ops nft_ct_expect_obj_ops = {
1325         .type           = &nft_ct_expect_obj_type,
1326         .size           = sizeof(struct nft_ct_expect_obj),
1327         .eval           = nft_ct_expect_obj_eval,
1328         .init           = nft_ct_expect_obj_init,
1329         .destroy        = nft_ct_expect_obj_destroy,
1330         .dump           = nft_ct_expect_obj_dump,
1331 };
1332
1333 static struct nft_object_type nft_ct_expect_obj_type __read_mostly = {
1334         .type           = NFT_OBJECT_CT_EXPECT,
1335         .ops            = &nft_ct_expect_obj_ops,
1336         .maxattr        = NFTA_CT_EXPECT_MAX,
1337         .policy         = nft_ct_expect_policy,
1338         .owner          = THIS_MODULE,
1339 };
1340
1341 static int __init nft_ct_module_init(void)
1342 {
1343         int err;
1344
1345         BUILD_BUG_ON(NF_CT_LABELS_MAX_SIZE > NFT_REG_SIZE);
1346
1347         err = nft_register_expr(&nft_ct_type);
1348         if (err < 0)
1349                 return err;
1350
1351         err = nft_register_expr(&nft_notrack_type);
1352         if (err < 0)
1353                 goto err1;
1354
1355         err = nft_register_obj(&nft_ct_helper_obj_type);
1356         if (err < 0)
1357                 goto err2;
1358
1359         err = nft_register_obj(&nft_ct_expect_obj_type);
1360         if (err < 0)
1361                 goto err3;
1362 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
1363         err = nft_register_obj(&nft_ct_timeout_obj_type);
1364         if (err < 0)
1365                 goto err4;
1366 #endif
1367         return 0;
1368
1369 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
1370 err4:
1371         nft_unregister_obj(&nft_ct_expect_obj_type);
1372 #endif
1373 err3:
1374         nft_unregister_obj(&nft_ct_helper_obj_type);
1375 err2:
1376         nft_unregister_expr(&nft_notrack_type);
1377 err1:
1378         nft_unregister_expr(&nft_ct_type);
1379         return err;
1380 }
1381
1382 static void __exit nft_ct_module_exit(void)
1383 {
1384 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
1385         nft_unregister_obj(&nft_ct_timeout_obj_type);
1386 #endif
1387         nft_unregister_obj(&nft_ct_expect_obj_type);
1388         nft_unregister_obj(&nft_ct_helper_obj_type);
1389         nft_unregister_expr(&nft_notrack_type);
1390         nft_unregister_expr(&nft_ct_type);
1391 }
1392
1393 module_init(nft_ct_module_init);
1394 module_exit(nft_ct_module_exit);
1395
1396 MODULE_LICENSE("GPL");
1397 MODULE_AUTHOR("Patrick McHardy <[email protected]>");
1398 MODULE_ALIAS_NFT_EXPR("ct");
1399 MODULE_ALIAS_NFT_EXPR("notrack");
1400 MODULE_ALIAS_NFT_OBJ(NFT_OBJECT_CT_HELPER);
1401 MODULE_ALIAS_NFT_OBJ(NFT_OBJECT_CT_TIMEOUT);
1402 MODULE_ALIAS_NFT_OBJ(NFT_OBJECT_CT_EXPECT);
1403 MODULE_DESCRIPTION("Netfilter nf_tables conntrack module");