net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32 #include <linux/siphash.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netlink.h>
  36 #include <net/sock.h>
  37 #include <net/netfilter/nf_conntrack.h>
  38 #include <net/netfilter/nf_conntrack_core.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_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #include <net/netfilter/nf_conntrack_synproxy.h>
  49 #ifdef CONFIG_NF_NAT_NEEDED
  50 #include <net/netfilter/nf_nat_core.h>
  51 #include <net/netfilter/nf_nat_l4proto.h>
  52 #include <net/netfilter/nf_nat_helper.h>
  53 #endif
  54
  55 #include <linux/netfilter/nfnetlink.h>
  56 #include <linux/netfilter/nfnetlink_conntrack.h>
  57
  58 MODULE_LICENSE("GPL");
  59
  60 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  61                                 const struct nf_conntrack_tuple *tuple,
  62                                 const struct nf_conntrack_l4proto *l4proto)
  63 {
  64         int ret = 0;
  65         struct nlattr *nest_parms;
  66
  67         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  68         if (!nest_parms)
  69                 goto nla_put_failure;
  70         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  71                 goto nla_put_failure;
  72
  73         if (likely(l4proto->tuple_to_nlattr))
  74                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  75
  76         nla_nest_end(skb, nest_parms);
  77
  78         return ret;
  79
  80 nla_put_failure:
  81         return -1;
  82 }
  83
  84 static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  85                                 const struct nf_conntrack_tuple *tuple)
  86 {
  87         if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  88             nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  89                 return -EMSGSIZE;
  90         return 0;
  91 }
  92
  93 static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
  94                                 const struct nf_conntrack_tuple *tuple)
  95 {
  96         if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
  97             nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
  98                 return -EMSGSIZE;
  99         return 0;
 100 }
 101
 102 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 103                                     const struct nf_conntrack_tuple *tuple)
 104 {
 105         int ret = 0;
 106         struct nlattr *nest_parms;
 107
 108         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
 109         if (!nest_parms)
 110                 goto nla_put_failure;
 111
 112         switch (tuple->src.l3num) {
 113         case NFPROTO_IPV4:
 114                 ret = ipv4_tuple_to_nlattr(skb, tuple);
 115                 break;
 116         case NFPROTO_IPV6:
 117                 ret = ipv6_tuple_to_nlattr(skb, tuple);
 118                 break;
 119         }
 120
 121         nla_nest_end(skb, nest_parms);
 122
 123         return ret;
 124
 125 nla_put_failure:
 126         return -1;
 127 }
 128
 129 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 130                                  const struct nf_conntrack_tuple *tuple)
 131 {
 132         const struct nf_conntrack_l4proto *l4proto;
 133         int ret;
 134
 135         rcu_read_lock();
 136         ret = ctnetlink_dump_tuples_ip(skb, tuple);
 137
 138         if (ret >= 0) {
 139                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 140                                                tuple->dst.protonum);
 141                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 142         }
 143         rcu_read_unlock();
 144         return ret;
 145 }
 146
 147 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 148                                   const struct nf_conntrack_zone *zone, int dir)
 149 {
 150         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 151                 return 0;
 152         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 153                 goto nla_put_failure;
 154         return 0;
 155
 156 nla_put_failure:
 157         return -1;
 158 }
 159
 160 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 161 {
 162         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 163                 goto nla_put_failure;
 164         return 0;
 165
 166 nla_put_failure:
 167         return -1;
 168 }
 169
 170 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 171 {
 172         long timeout = nf_ct_expires(ct) / HZ;
 173
 174         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 175                 goto nla_put_failure;
 176         return 0;
 177
 178 nla_put_failure:
 179         return -1;
 180 }
 181
 182 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 183 {
 184         const struct nf_conntrack_l4proto *l4proto;
 185         struct nlattr *nest_proto;
 186         int ret;
 187
 188         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 189         if (!l4proto->to_nlattr)
 190                 return 0;
 191
 192         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 193         if (!nest_proto)
 194                 goto nla_put_failure;
 195
 196         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 197
 198         nla_nest_end(skb, nest_proto);
 199
 200         return ret;
 201
 202 nla_put_failure:
 203         return -1;
 204 }
 205
 206 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 207                                    const struct nf_conn *ct)
 208 {
 209         struct nlattr *nest_helper;
 210         const struct nf_conn_help *help = nfct_help(ct);
 211         struct nf_conntrack_helper *helper;
 212
 213         if (!help)
 214                 return 0;
 215
 216         rcu_read_lock();
 217         helper = rcu_dereference(help->helper);
 218         if (!helper)
 219                 goto out;
 220
 221         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 222         if (!nest_helper)
 223                 goto nla_put_failure;
 224         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 225                 goto nla_put_failure;
 226
 227         if (helper->to_nlattr)
 228                 helper->to_nlattr(skb, ct);
 229
 230         nla_nest_end(skb, nest_helper);
 231 out:
 232         rcu_read_unlock();
 233         return 0;
 234
 235 nla_put_failure:
 236         rcu_read_unlock();
 237         return -1;
 238 }
 239
 240 static int
 241 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 242               enum ip_conntrack_dir dir, int type)
 243 {
 244         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 245         struct nf_conn_counter *counter = acct->counter;
 246         struct nlattr *nest_count;
 247         u64 pkts, bytes;
 248
 249         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 250                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 251                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 252         } else {
 253                 pkts = atomic64_read(&counter[dir].packets);
 254                 bytes = atomic64_read(&counter[dir].bytes);
 255         }
 256
 257         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 258         if (!nest_count)
 259                 goto nla_put_failure;
 260
 261         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 262                          CTA_COUNTERS_PAD) ||
 263             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 264                          CTA_COUNTERS_PAD))
 265                 goto nla_put_failure;
 266
 267         nla_nest_end(skb, nest_count);
 268
 269         return 0;
 270
 271 nla_put_failure:
 272         return -1;
 273 }
 274
 275 static int
 276 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 277 {
 278         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 279
 280         if (!acct)
 281                 return 0;
 282
 283         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 284                 return -1;
 285         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 286                 return -1;
 287
 288         return 0;
 289 }
 290
 291 static int
 292 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 293 {
 294         struct nlattr *nest_count;
 295         const struct nf_conn_tstamp *tstamp;
 296
 297         tstamp = nf_conn_tstamp_find(ct);
 298         if (!tstamp)
 299                 return 0;
 300
 301         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 302         if (!nest_count)
 303                 goto nla_put_failure;
 304
 305         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 306                          CTA_TIMESTAMP_PAD) ||
 307             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 308                                                cpu_to_be64(tstamp->stop),
 309                                                CTA_TIMESTAMP_PAD)))
 310                 goto nla_put_failure;
 311         nla_nest_end(skb, nest_count);
 312
 313         return 0;
 314
 315 nla_put_failure:
 316         return -1;
 317 }
 318
 319 #ifdef CONFIG_NF_CONNTRACK_MARK
 320 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 321 {
 322         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 323                 goto nla_put_failure;
 324         return 0;
 325
 326 nla_put_failure:
 327         return -1;
 328 }
 329 #else
 330 #define ctnetlink_dump_mark(a, b) (0)
 331 #endif
 332
 333 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 334 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 335 {
 336         struct nlattr *nest_secctx;
 337         int len, ret;
 338         char *secctx;
 339
 340         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 341         if (ret)
 342                 return 0;
 343
 344         ret = -1;
 345         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 346         if (!nest_secctx)
 347                 goto nla_put_failure;
 348
 349         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 350                 goto nla_put_failure;
 351         nla_nest_end(skb, nest_secctx);
 352
 353         ret = 0;
 354 nla_put_failure:
 355         security_release_secctx(secctx, len);
 356         return ret;
 357 }
 358 #else
 359 #define ctnetlink_dump_secctx(a, b) (0)
 360 #endif
 361
 362 #ifdef CONFIG_NF_CONNTRACK_LABELS
 363 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 364 {
 365         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 366
 367         if (!labels)
 368                 return 0;
 369         return nla_total_size(sizeof(labels->bits));
 370 }
 371
 372 static int
 373 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 374 {
 375         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 376         unsigned int i;
 377
 378         if (!labels)
 379                 return 0;
 380
 381         i = 0;
 382         do {
 383                 if (labels->bits[i] != 0)
 384                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 385                                        labels->bits);
 386                 i++;
 387         } while (i < ARRAY_SIZE(labels->bits));
 388
 389         return 0;
 390 }
 391 #else
 392 #define ctnetlink_dump_labels(a, b) (0)
 393 #define ctnetlink_label_size(a) (0)
 394 #endif
 395
 396 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 397
 398 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 399 {
 400         struct nlattr *nest_parms;
 401
 402         if (!(ct->status & IPS_EXPECTED))
 403                 return 0;
 404
 405         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 406         if (!nest_parms)
 407                 goto nla_put_failure;
 408         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 409                 goto nla_put_failure;
 410         nla_nest_end(skb, nest_parms);
 411
 412         return 0;
 413
 414 nla_put_failure:
 415         return -1;
 416 }
 417
 418 static int
 419 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 420 {
 421         struct nlattr *nest_parms;
 422
 423         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 424         if (!nest_parms)
 425                 goto nla_put_failure;
 426
 427         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 428                          htonl(seq->correction_pos)) ||
 429             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 430                          htonl(seq->offset_before)) ||
 431             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 432                          htonl(seq->offset_after)))
 433                 goto nla_put_failure;
 434
 435         nla_nest_end(skb, nest_parms);
 436
 437         return 0;
 438
 439 nla_put_failure:
 440         return -1;
 441 }
 442
 443 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 444 {
 445         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 446         struct nf_ct_seqadj *seq;
 447
 448         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 449                 return 0;
 450
 451         spin_lock_bh(&ct->lock);
 452         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 453         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 454                 goto err;
 455
 456         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 457         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 458                 goto err;
 459
 460         spin_unlock_bh(&ct->lock);
 461         return 0;
 462 err:
 463         spin_unlock_bh(&ct->lock);
 464         return -1;
 465 }
 466
 467 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 468 {
 469         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 470         struct nlattr *nest_parms;
 471
 472         if (!synproxy)
 473                 return 0;
 474
 475         nest_parms = nla_nest_start(skb, CTA_SYNPROXY | NLA_F_NESTED);
 476         if (!nest_parms)
 477                 goto nla_put_failure;
 478
 479         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 480             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 481             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 482                 goto nla_put_failure;
 483
 484         nla_nest_end(skb, nest_parms);
 485
 486         return 0;
 487
 488 nla_put_failure:
 489         return -1;
 490 }
 491
 492 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 493 {
 494         __be32 id = (__force __be32)nf_ct_get_id(ct);
 495
 496         if (nla_put_be32(skb, CTA_ID, id))
 497                 goto nla_put_failure;
 498         return 0;
 499
 500 nla_put_failure:
 501         return -1;
 502 }
 503
 504 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 505 {
 506         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 507                 goto nla_put_failure;
 508         return 0;
 509
 510 nla_put_failure:
 511         return -1;
 512 }
 513
 514 static int
 515 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 516                     struct nf_conn *ct)
 517 {
 518         const struct nf_conntrack_zone *zone;
 519         struct nlmsghdr *nlh;
 520         struct nfgenmsg *nfmsg;
 521         struct nlattr *nest_parms;
 522         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 523
 524         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 525         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 526         if (nlh == NULL)
 527                 goto nlmsg_failure;
 528
 529         nfmsg = nlmsg_data(nlh);
 530         nfmsg->nfgen_family = nf_ct_l3num(ct);
 531         nfmsg->version      = NFNETLINK_V0;
 532         nfmsg->res_id       = 0;
 533
 534         zone = nf_ct_zone(ct);
 535
 536         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 537         if (!nest_parms)
 538                 goto nla_put_failure;
 539         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 540                 goto nla_put_failure;
 541         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 542                                    NF_CT_ZONE_DIR_ORIG) < 0)
 543                 goto nla_put_failure;
 544         nla_nest_end(skb, nest_parms);
 545
 546         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 547         if (!nest_parms)
 548                 goto nla_put_failure;
 549         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 550                 goto nla_put_failure;
 551         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 552                                    NF_CT_ZONE_DIR_REPL) < 0)
 553                 goto nla_put_failure;
 554         nla_nest_end(skb, nest_parms);
 555
 556         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 557                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 558                 goto nla_put_failure;
 559
 560         if (ctnetlink_dump_status(skb, ct) < 0 ||
 561             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 562             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 563             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 564             ctnetlink_dump_mark(skb, ct) < 0 ||
 565             ctnetlink_dump_secctx(skb, ct) < 0 ||
 566             ctnetlink_dump_labels(skb, ct) < 0 ||
 567             ctnetlink_dump_id(skb, ct) < 0 ||
 568             ctnetlink_dump_use(skb, ct) < 0 ||
 569             ctnetlink_dump_master(skb, ct) < 0 ||
 570             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 571             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 572                 goto nla_put_failure;
 573
 574         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 575             (ctnetlink_dump_timeout(skb, ct) < 0 ||
 576              ctnetlink_dump_protoinfo(skb, ct) < 0))
 577                 goto nla_put_failure;
 578
 579         nlmsg_end(skb, nlh);
 580         return skb->len;
 581
 582 nlmsg_failure:
 583 nla_put_failure:
 584         nlmsg_cancel(skb, nlh);
 585         return -1;
 586 }
 587
 588 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 589         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 590         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 591         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 592         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 593 };
 594
 595 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 596 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 597 {
 598         const struct nf_conntrack_l4proto *l4proto;
 599         size_t len, len4 = 0;
 600
 601         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 602         len *= 3u; /* ORIG, REPLY, MASTER */
 603
 604         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 605         len += l4proto->nlattr_size;
 606         if (l4proto->nlattr_tuple_size) {
 607                 len4 = l4proto->nlattr_tuple_size();
 608                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 609         }
 610
 611         return len + len4;
 612 }
 613 #endif
 614
 615 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 616 {
 617         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 618                 return 0;
 619         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 620                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 621                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 622                ;
 623 }
 624
 625 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 626 {
 627 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 628         int len, ret;
 629
 630         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 631         if (ret)
 632                 return 0;
 633
 634         return nla_total_size(0) /* CTA_SECCTX */
 635                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 636 #else
 637         return 0;
 638 #endif
 639 }
 640
 641 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 642 {
 643 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 644         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 645                 return 0;
 646         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 647 #else
 648         return 0;
 649 #endif
 650 }
 651
 652 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 653 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 654 {
 655         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 656                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 657                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 658                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 659                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 660                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 661                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 662                + ctnetlink_acct_size(ct)
 663                + ctnetlink_timestamp_size(ct)
 664                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 665                + nla_total_size(0) /* CTA_PROTOINFO */
 666                + nla_total_size(0) /* CTA_HELP */
 667                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 668                + ctnetlink_secctx_size(ct)
 669 #ifdef CONFIG_NF_NAT_NEEDED
 670                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 671                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 672 #endif
 673 #ifdef CONFIG_NF_CONNTRACK_MARK
 674                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 675 #endif
 676 #ifdef CONFIG_NF_CONNTRACK_ZONES
 677                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 678 #endif
 679                + ctnetlink_proto_size(ct)
 680                + ctnetlink_label_size(ct)
 681                ;
 682 }
 683
 684 static int
 685 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 686 {
 687         const struct nf_conntrack_zone *zone;
 688         struct net *net;
 689         struct nlmsghdr *nlh;
 690         struct nfgenmsg *nfmsg;
 691         struct nlattr *nest_parms;
 692         struct nf_conn *ct = item->ct;
 693         struct sk_buff *skb;
 694         unsigned int type;
 695         unsigned int flags = 0, group;
 696         int err;
 697
 698         if (events & (1 << IPCT_DESTROY)) {
 699                 type = IPCTNL_MSG_CT_DELETE;
 700                 group = NFNLGRP_CONNTRACK_DESTROY;
 701         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 702                 type = IPCTNL_MSG_CT_NEW;
 703                 flags = NLM_F_CREATE|NLM_F_EXCL;
 704                 group = NFNLGRP_CONNTRACK_NEW;
 705         } else if (events) {
 706                 type = IPCTNL_MSG_CT_NEW;
 707                 group = NFNLGRP_CONNTRACK_UPDATE;
 708         } else
 709                 return 0;
 710
 711         net = nf_ct_net(ct);
 712         if (!item->report && !nfnetlink_has_listeners(net, group))
 713                 return 0;
 714
 715         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 716         if (skb == NULL)
 717                 goto errout;
 718
 719         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 720         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 721         if (nlh == NULL)
 722                 goto nlmsg_failure;
 723
 724         nfmsg = nlmsg_data(nlh);
 725         nfmsg->nfgen_family = nf_ct_l3num(ct);
 726         nfmsg->version  = NFNETLINK_V0;
 727         nfmsg->res_id   = 0;
 728
 729         zone = nf_ct_zone(ct);
 730
 731         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 732         if (!nest_parms)
 733                 goto nla_put_failure;
 734         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 735                 goto nla_put_failure;
 736         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 737                                    NF_CT_ZONE_DIR_ORIG) < 0)
 738                 goto nla_put_failure;
 739         nla_nest_end(skb, nest_parms);
 740
 741         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 742         if (!nest_parms)
 743                 goto nla_put_failure;
 744         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 745                 goto nla_put_failure;
 746         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 747                                    NF_CT_ZONE_DIR_REPL) < 0)
 748                 goto nla_put_failure;
 749         nla_nest_end(skb, nest_parms);
 750
 751         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 752                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 753                 goto nla_put_failure;
 754
 755         if (ctnetlink_dump_id(skb, ct) < 0)
 756                 goto nla_put_failure;
 757
 758         if (ctnetlink_dump_status(skb, ct) < 0)
 759                 goto nla_put_failure;
 760
 761         if (events & (1 << IPCT_DESTROY)) {
 762                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 763                     ctnetlink_dump_timestamp(skb, ct) < 0)
 764                         goto nla_put_failure;
 765         } else {
 766                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 767                         goto nla_put_failure;
 768
 769                 if (events & (1 << IPCT_PROTOINFO)
 770                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 771                         goto nla_put_failure;
 772
 773                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 774                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 775                         goto nla_put_failure;
 776
 777 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 778                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 779                     && ctnetlink_dump_secctx(skb, ct) < 0)
 780                         goto nla_put_failure;
 781 #endif
 782                 if (events & (1 << IPCT_LABEL) &&
 783                      ctnetlink_dump_labels(skb, ct) < 0)
 784                         goto nla_put_failure;
 785
 786                 if (events & (1 << IPCT_RELATED) &&
 787                     ctnetlink_dump_master(skb, ct) < 0)
 788                         goto nla_put_failure;
 789
 790                 if (events & (1 << IPCT_SEQADJ) &&
 791                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 792                         goto nla_put_failure;
 793
 794                 if (events & (1 << IPCT_SYNPROXY) &&
 795                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 796                         goto nla_put_failure;
 797         }
 798
 799 #ifdef CONFIG_NF_CONNTRACK_MARK
 800         if ((events & (1 << IPCT_MARK) || ct->mark)
 801             && ctnetlink_dump_mark(skb, ct) < 0)
 802                 goto nla_put_failure;
 803 #endif
 804         nlmsg_end(skb, nlh);
 805         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 806                              GFP_ATOMIC);
 807         if (err == -ENOBUFS || err == -EAGAIN)
 808                 return -ENOBUFS;
 809
 810         return 0;
 811
 812 nla_put_failure:
 813         nlmsg_cancel(skb, nlh);
 814 nlmsg_failure:
 815         kfree_skb(skb);
 816 errout:
 817         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 818                 return -ENOBUFS;
 819
 820         return 0;
 821 }
 822 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 823
 824 static int ctnetlink_done(struct netlink_callback *cb)
 825 {
 826         if (cb->args[1])
 827                 nf_ct_put((struct nf_conn *)cb->args[1]);
 828         kfree(cb->data);
 829         return 0;
 830 }
 831
 832 struct ctnetlink_filter {
 833         struct {
 834                 u_int32_t val;
 835                 u_int32_t mask;
 836         } mark;
 837 };
 838
 839 static struct ctnetlink_filter *
 840 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 841 {
 842 #ifdef CONFIG_NF_CONNTRACK_MARK
 843         struct ctnetlink_filter *filter;
 844
 845         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 846         if (filter == NULL)
 847                 return ERR_PTR(-ENOMEM);
 848
 849         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 850         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 851
 852         return filter;
 853 #else
 854         return ERR_PTR(-EOPNOTSUPP);
 855 #endif
 856 }
 857
 858 static int ctnetlink_start(struct netlink_callback *cb)
 859 {
 860         const struct nlattr * const *cda = cb->data;
 861         struct ctnetlink_filter *filter = NULL;
 862
 863         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
 864                 filter = ctnetlink_alloc_filter(cda);
 865                 if (IS_ERR(filter))
 866                         return PTR_ERR(filter);
 867         }
 868
 869         cb->data = filter;
 870         return 0;
 871 }
 872
 873 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 874 {
 875         struct ctnetlink_filter *filter = data;
 876
 877         if (filter == NULL)
 878                 return 1;
 879
 880 #ifdef CONFIG_NF_CONNTRACK_MARK
 881         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 882                 return 1;
 883 #endif
 884
 885         return 0;
 886 }
 887
 888 static int
 889 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 890 {
 891         struct net *net = sock_net(skb->sk);
 892         struct nf_conn *ct, *last;
 893         struct nf_conntrack_tuple_hash *h;
 894         struct hlist_nulls_node *n;
 895         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 896         u_int8_t l3proto = nfmsg->nfgen_family;
 897         struct nf_conn *nf_ct_evict[8];
 898         int res, i;
 899         spinlock_t *lockp;
 900
 901         last = (struct nf_conn *)cb->args[1];
 902         i = 0;
 903
 904         local_bh_disable();
 905         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 906 restart:
 907                 while (i) {
 908                         i--;
 909                         if (nf_ct_should_gc(nf_ct_evict[i]))
 910                                 nf_ct_kill(nf_ct_evict[i]);
 911                         nf_ct_put(nf_ct_evict[i]);
 912                 }
 913
 914                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 915                 nf_conntrack_lock(lockp);
 916                 if (cb->args[0] >= nf_conntrack_htable_size) {
 917                         spin_unlock(lockp);
 918                         goto out;
 919                 }
 920                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 921                                            hnnode) {
 922                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 923                                 continue;
 924                         ct = nf_ct_tuplehash_to_ctrack(h);
 925                         if (nf_ct_is_expired(ct)) {
 926                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 927                                     atomic_inc_not_zero(&ct->ct_general.use))
 928                                         nf_ct_evict[i++] = ct;
 929                                 continue;
 930                         }
 931
 932                         if (!net_eq(net, nf_ct_net(ct)))
 933                                 continue;
 934
 935                         /* Dump entries of a given L3 protocol number.
 936                          * If it is not specified, ie. l3proto == 0,
 937                          * then dump everything. */
 938                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 939                                 continue;
 940                         if (cb->args[1]) {
 941                                 if (ct != last)
 942                                         continue;
 943                                 cb->args[1] = 0;
 944                         }
 945                         if (!ctnetlink_filter_match(ct, cb->data))
 946                                 continue;
 947
 948                         rcu_read_lock();
 949                         res =
 950                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 951                                             cb->nlh->nlmsg_seq,
 952                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 953                                             ct);
 954                         rcu_read_unlock();
 955                         if (res < 0) {
 956                                 nf_conntrack_get(&ct->ct_general);
 957                                 cb->args[1] = (unsigned long)ct;
 958                                 spin_unlock(lockp);
 959                                 goto out;
 960                         }
 961                 }
 962                 spin_unlock(lockp);
 963                 if (cb->args[1]) {
 964                         cb->args[1] = 0;
 965                         goto restart;
 966                 }
 967         }
 968 out:
 969         local_bh_enable();
 970         if (last) {
 971                 /* nf ct hash resize happened, now clear the leftover. */
 972                 if ((struct nf_conn *)cb->args[1] == last)
 973                         cb->args[1] = 0;
 974
 975                 nf_ct_put(last);
 976         }
 977
 978         while (i) {
 979                 i--;
 980                 if (nf_ct_should_gc(nf_ct_evict[i]))
 981                         nf_ct_kill(nf_ct_evict[i]);
 982                 nf_ct_put(nf_ct_evict[i]);
 983         }
 984
 985         return skb->len;
 986 }
 987
 988 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
 989                                 struct nf_conntrack_tuple *t)
 990 {
 991         if (!tb[CTA_IP_V4_SRC] || !tb[CTA_IP_V4_DST])
 992                 return -EINVAL;
 993
 994         t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
 995         t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
 996
 997         return 0;
 998 }
 999
1000 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1001                                 struct nf_conntrack_tuple *t)
1002 {
1003         if (!tb[CTA_IP_V6_SRC] || !tb[CTA_IP_V6_DST])
1004                 return -EINVAL;
1005
1006         t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1007         t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1008
1009         return 0;
1010 }
1011
1012 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1013                                     struct nf_conntrack_tuple *tuple)
1014 {
1015         struct nlattr *tb[CTA_IP_MAX+1];
1016         int ret = 0;
1017
1018         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL, NULL);
1019         if (ret < 0)
1020                 return ret;
1021
1022         ret = nla_validate_nested(attr, CTA_IP_MAX,
1023                                   cta_ip_nla_policy, NULL);
1024         if (ret)
1025                 return ret;
1026
1027         switch (tuple->src.l3num) {
1028         case NFPROTO_IPV4:
1029                 ret = ipv4_nlattr_to_tuple(tb, tuple);
1030                 break;
1031         case NFPROTO_IPV6:
1032                 ret = ipv6_nlattr_to_tuple(tb, tuple);
1033                 break;
1034         }
1035
1036         return ret;
1037 }
1038
1039 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1040         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1041 };
1042
1043 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1044                                        struct nf_conntrack_tuple *tuple)
1045 {
1046         const struct nf_conntrack_l4proto *l4proto;
1047         struct nlattr *tb[CTA_PROTO_MAX+1];
1048         int ret = 0;
1049
1050         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy,
1051                                NULL);
1052         if (ret < 0)
1053                 return ret;
1054
1055         if (!tb[CTA_PROTO_NUM])
1056                 return -EINVAL;
1057         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1058
1059         rcu_read_lock();
1060         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
1061
1062         if (likely(l4proto->nlattr_to_tuple)) {
1063                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
1064                                           l4proto->nla_policy, NULL);
1065                 if (ret == 0)
1066                         ret = l4proto->nlattr_to_tuple(tb, tuple);
1067         }
1068
1069         rcu_read_unlock();
1070
1071         return ret;
1072 }
1073
1074 static int
1075 ctnetlink_parse_zone(const struct nlattr *attr,
1076                      struct nf_conntrack_zone *zone)
1077 {
1078         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1079                         NF_CT_DEFAULT_ZONE_DIR, 0);
1080 #ifdef CONFIG_NF_CONNTRACK_ZONES
1081         if (attr)
1082                 zone->id = ntohs(nla_get_be16(attr));
1083 #else
1084         if (attr)
1085                 return -EOPNOTSUPP;
1086 #endif
1087         return 0;
1088 }
1089
1090 static int
1091 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1092                            struct nf_conntrack_zone *zone)
1093 {
1094         int ret;
1095
1096         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1097                 return -EINVAL;
1098
1099         ret = ctnetlink_parse_zone(attr, zone);
1100         if (ret < 0)
1101                 return ret;
1102
1103         if (type == CTA_TUPLE_REPLY)
1104                 zone->dir = NF_CT_ZONE_DIR_REPL;
1105         else
1106                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1107
1108         return 0;
1109 }
1110
1111 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1112         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1113         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1114         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1115 };
1116
1117 static int
1118 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1119                       struct nf_conntrack_tuple *tuple, u32 type,
1120                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1121 {
1122         struct nlattr *tb[CTA_TUPLE_MAX+1];
1123         int err;
1124
1125         memset(tuple, 0, sizeof(*tuple));
1126
1127         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1128                                NULL);
1129         if (err < 0)
1130                 return err;
1131
1132         if (!tb[CTA_TUPLE_IP])
1133                 return -EINVAL;
1134
1135         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1136                 return -EOPNOTSUPP;
1137         tuple->src.l3num = l3num;
1138
1139         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1140         if (err < 0)
1141                 return err;
1142
1143         if (!tb[CTA_TUPLE_PROTO])
1144                 return -EINVAL;
1145
1146         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1147         if (err < 0)
1148                 return err;
1149
1150         if (tb[CTA_TUPLE_ZONE]) {
1151                 if (!zone)
1152                         return -EINVAL;
1153
1154                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1155                                                  type, zone);
1156                 if (err < 0)
1157                         return err;
1158         }
1159
1160         /* orig and expect tuples get DIR_ORIGINAL */
1161         if (type == CTA_TUPLE_REPLY)
1162                 tuple->dst.dir = IP_CT_DIR_REPLY;
1163         else
1164                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1165
1166         return 0;
1167 }
1168
1169 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1170         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1171                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1172 };
1173
1174 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1175                                 struct nlattr **helpinfo)
1176 {
1177         int err;
1178         struct nlattr *tb[CTA_HELP_MAX+1];
1179
1180         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1181         if (err < 0)
1182                 return err;
1183
1184         if (!tb[CTA_HELP_NAME])
1185                 return -EINVAL;
1186
1187         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1188
1189         if (tb[CTA_HELP_INFO])
1190                 *helpinfo = tb[CTA_HELP_INFO];
1191
1192         return 0;
1193 }
1194
1195 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1196         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1197         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1198         [CTA_STATUS]            = { .type = NLA_U32 },
1199         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1200         [CTA_HELP]              = { .type = NLA_NESTED },
1201         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1202         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1203         [CTA_MARK]              = { .type = NLA_U32 },
1204         [CTA_ID]                = { .type = NLA_U32 },
1205         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1206         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1207         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1208         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1209         [CTA_ZONE]              = { .type = NLA_U16 },
1210         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1211         [CTA_LABELS]            = { .type = NLA_BINARY,
1212                                     .len = NF_CT_LABELS_MAX_SIZE },
1213         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1214                                     .len = NF_CT_LABELS_MAX_SIZE },
1215 };
1216
1217 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1218 {
1219         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1220                 return 0;
1221
1222         return ctnetlink_filter_match(ct, data);
1223 }
1224
1225 static int ctnetlink_flush_conntrack(struct net *net,
1226                                      const struct nlattr * const cda[],
1227                                      u32 portid, int report)
1228 {
1229         struct ctnetlink_filter *filter = NULL;
1230
1231         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1232                 filter = ctnetlink_alloc_filter(cda);
1233                 if (IS_ERR(filter))
1234                         return PTR_ERR(filter);
1235         }
1236
1237         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1238                                   portid, report);
1239         kfree(filter);
1240
1241         return 0;
1242 }
1243
1244 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1245                                    struct sk_buff *skb,
1246                                    const struct nlmsghdr *nlh,
1247                                    const struct nlattr * const cda[],
1248                                    struct netlink_ext_ack *extack)
1249 {
1250         struct nf_conntrack_tuple_hash *h;
1251         struct nf_conntrack_tuple tuple;
1252         struct nf_conn *ct;
1253         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1254         u_int8_t u3 = nfmsg->nfgen_family;
1255         struct nf_conntrack_zone zone;
1256         int err;
1257
1258         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1259         if (err < 0)
1260                 return err;
1261
1262         if (cda[CTA_TUPLE_ORIG])
1263                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1264                                             u3, &zone);
1265         else if (cda[CTA_TUPLE_REPLY])
1266                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1267                                             u3, &zone);
1268         else {
1269                 return ctnetlink_flush_conntrack(net, cda,
1270                                                  NETLINK_CB(skb).portid,
1271                                                  nlmsg_report(nlh));
1272         }
1273
1274         if (err < 0)
1275                 return err;
1276
1277         h = nf_conntrack_find_get(net, &zone, &tuple);
1278         if (!h)
1279                 return -ENOENT;
1280
1281         ct = nf_ct_tuplehash_to_ctrack(h);
1282
1283         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1284                 nf_ct_put(ct);
1285                 return -EBUSY;
1286         }
1287
1288         if (cda[CTA_ID]) {
1289                 __be32 id = nla_get_be32(cda[CTA_ID]);
1290
1291                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1292                         nf_ct_put(ct);
1293                         return -ENOENT;
1294                 }
1295         }
1296
1297         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1298         nf_ct_put(ct);
1299
1300         return 0;
1301 }
1302
1303 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1304                                    struct sk_buff *skb,
1305                                    const struct nlmsghdr *nlh,
1306                                    const struct nlattr * const cda[],
1307                                    struct netlink_ext_ack *extack)
1308 {
1309         struct nf_conntrack_tuple_hash *h;
1310         struct nf_conntrack_tuple tuple;
1311         struct nf_conn *ct;
1312         struct sk_buff *skb2 = NULL;
1313         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1314         u_int8_t u3 = nfmsg->nfgen_family;
1315         struct nf_conntrack_zone zone;
1316         int err;
1317
1318         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1319                 struct netlink_dump_control c = {
1320                         .start = ctnetlink_start,
1321                         .dump = ctnetlink_dump_table,
1322                         .done = ctnetlink_done,
1323                         .data = (void *)cda,
1324                 };
1325
1326                 return netlink_dump_start(ctnl, skb, nlh, &c);
1327         }
1328
1329         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1330         if (err < 0)
1331                 return err;
1332
1333         if (cda[CTA_TUPLE_ORIG])
1334                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1335                                             u3, &zone);
1336         else if (cda[CTA_TUPLE_REPLY])
1337                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1338                                             u3, &zone);
1339         else
1340                 return -EINVAL;
1341
1342         if (err < 0)
1343                 return err;
1344
1345         h = nf_conntrack_find_get(net, &zone, &tuple);
1346         if (!h)
1347                 return -ENOENT;
1348
1349         ct = nf_ct_tuplehash_to_ctrack(h);
1350
1351         err = -ENOMEM;
1352         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1353         if (skb2 == NULL) {
1354                 nf_ct_put(ct);
1355                 return -ENOMEM;
1356         }
1357
1358         rcu_read_lock();
1359         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1360                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1361         rcu_read_unlock();
1362         nf_ct_put(ct);
1363         if (err <= 0)
1364                 goto free;
1365
1366         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1367         if (err < 0)
1368                 goto out;
1369
1370         return 0;
1371
1372 free:
1373         kfree_skb(skb2);
1374 out:
1375         /* this avoids a loop in nfnetlink. */
1376         return err == -EAGAIN ? -ENOBUFS : err;
1377 }
1378
1379 static int ctnetlink_done_list(struct netlink_callback *cb)
1380 {
1381         if (cb->args[1])
1382                 nf_ct_put((struct nf_conn *)cb->args[1]);
1383         return 0;
1384 }
1385
1386 static int
1387 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1388 {
1389         struct nf_conn *ct, *last;
1390         struct nf_conntrack_tuple_hash *h;
1391         struct hlist_nulls_node *n;
1392         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1393         u_int8_t l3proto = nfmsg->nfgen_family;
1394         int res;
1395         int cpu;
1396         struct hlist_nulls_head *list;
1397         struct net *net = sock_net(skb->sk);
1398
1399         if (cb->args[2])
1400                 return 0;
1401
1402         last = (struct nf_conn *)cb->args[1];
1403
1404         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1405                 struct ct_pcpu *pcpu;
1406
1407                 if (!cpu_possible(cpu))
1408                         continue;
1409
1410                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1411                 spin_lock_bh(&pcpu->lock);
1412                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1413 restart:
1414                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1415                         ct = nf_ct_tuplehash_to_ctrack(h);
1416                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1417                                 continue;
1418                         if (cb->args[1]) {
1419                                 if (ct != last)
1420                                         continue;
1421                                 cb->args[1] = 0;
1422                         }
1423                         rcu_read_lock();
1424                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1425                                                   cb->nlh->nlmsg_seq,
1426                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1427                                                   ct);
1428                         rcu_read_unlock();
1429                         if (res < 0) {
1430                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1431                                         continue;
1432                                 cb->args[0] = cpu;
1433                                 cb->args[1] = (unsigned long)ct;
1434                                 spin_unlock_bh(&pcpu->lock);
1435                                 goto out;
1436                         }
1437                 }
1438                 if (cb->args[1]) {
1439                         cb->args[1] = 0;
1440                         goto restart;
1441                 }
1442                 spin_unlock_bh(&pcpu->lock);
1443         }
1444         cb->args[2] = 1;
1445 out:
1446         if (last)
1447                 nf_ct_put(last);
1448
1449         return skb->len;
1450 }
1451
1452 static int
1453 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1454 {
1455         return ctnetlink_dump_list(skb, cb, true);
1456 }
1457
1458 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1459                                   struct sk_buff *skb,
1460                                   const struct nlmsghdr *nlh,
1461                                   const struct nlattr * const cda[],
1462                                   struct netlink_ext_ack *extack)
1463 {
1464         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1465                 struct netlink_dump_control c = {
1466                         .dump = ctnetlink_dump_dying,
1467                         .done = ctnetlink_done_list,
1468                 };
1469                 return netlink_dump_start(ctnl, skb, nlh, &c);
1470         }
1471
1472         return -EOPNOTSUPP;
1473 }
1474
1475 static int
1476 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1477 {
1478         return ctnetlink_dump_list(skb, cb, false);
1479 }
1480
1481 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1482                                         struct sk_buff *skb,
1483                                         const struct nlmsghdr *nlh,
1484                                         const struct nlattr * const cda[],
1485                                         struct netlink_ext_ack *extack)
1486 {
1487         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1488                 struct netlink_dump_control c = {
1489                         .dump = ctnetlink_dump_unconfirmed,
1490                         .done = ctnetlink_done_list,
1491                 };
1492                 return netlink_dump_start(ctnl, skb, nlh, &c);
1493         }
1494
1495         return -EOPNOTSUPP;
1496 }
1497
1498 #ifdef CONFIG_NF_NAT_NEEDED
1499 static int
1500 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1501                           enum nf_nat_manip_type manip,
1502                           const struct nlattr *attr)
1503 {
1504         struct nf_nat_hook *nat_hook;
1505         int err;
1506
1507         nat_hook = rcu_dereference(nf_nat_hook);
1508         if (!nat_hook) {
1509 #ifdef CONFIG_MODULES
1510                 rcu_read_unlock();
1511                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1512                 if (request_module("nf-nat") < 0) {
1513                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1514                         rcu_read_lock();
1515                         return -EOPNOTSUPP;
1516                 }
1517                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1518                 rcu_read_lock();
1519                 nat_hook = rcu_dereference(nf_nat_hook);
1520                 if (nat_hook)
1521                         return -EAGAIN;
1522 #endif
1523                 return -EOPNOTSUPP;
1524         }
1525
1526         err = nat_hook->parse_nat_setup(ct, manip, attr);
1527         if (err == -EAGAIN) {
1528 #ifdef CONFIG_MODULES
1529                 rcu_read_unlock();
1530                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1531                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1532                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1533                         rcu_read_lock();
1534                         return -EOPNOTSUPP;
1535                 }
1536                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1537                 rcu_read_lock();
1538 #else
1539                 err = -EOPNOTSUPP;
1540 #endif
1541         }
1542         return err;
1543 }
1544 #endif
1545
1546 static void
1547 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1548                           unsigned long off)
1549 {
1550         unsigned int bit;
1551
1552         /* Ignore these unchangable bits */
1553         on &= ~IPS_UNCHANGEABLE_MASK;
1554         off &= ~IPS_UNCHANGEABLE_MASK;
1555
1556         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1557                 if (on & (1 << bit))
1558                         set_bit(bit, &ct->status);
1559                 else if (off & (1 << bit))
1560                         clear_bit(bit, &ct->status);
1561         }
1562 }
1563
1564 static int
1565 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1566 {
1567         unsigned long d;
1568         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1569         d = ct->status ^ status;
1570
1571         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1572                 /* unchangeable */
1573                 return -EBUSY;
1574
1575         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1576                 /* SEEN_REPLY bit can only be set */
1577                 return -EBUSY;
1578
1579         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1580                 /* ASSURED bit can only be set */
1581                 return -EBUSY;
1582
1583         __ctnetlink_change_status(ct, status, 0);
1584         return 0;
1585 }
1586
1587 static int
1588 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1589 {
1590 #ifdef CONFIG_NF_NAT_NEEDED
1591         int ret;
1592
1593         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1594                 return 0;
1595
1596         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1597                                         cda[CTA_NAT_DST]);
1598         if (ret < 0)
1599                 return ret;
1600
1601         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1602                                          cda[CTA_NAT_SRC]);
1603 #else
1604         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1605                 return 0;
1606         return -EOPNOTSUPP;
1607 #endif
1608 }
1609
1610 static int ctnetlink_change_helper(struct nf_conn *ct,
1611                                    const struct nlattr * const cda[])
1612 {
1613         struct nf_conntrack_helper *helper;
1614         struct nf_conn_help *help = nfct_help(ct);
1615         char *helpname = NULL;
1616         struct nlattr *helpinfo = NULL;
1617         int err;
1618
1619         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1620         if (err < 0)
1621                 return err;
1622
1623         /* don't change helper of sibling connections */
1624         if (ct->master) {
1625                 /* If we try to change the helper to the same thing twice,
1626                  * treat the second attempt as a no-op instead of returning
1627                  * an error.
1628                  */
1629                 err = -EBUSY;
1630                 if (help) {
1631                         rcu_read_lock();
1632                         helper = rcu_dereference(help->helper);
1633                         if (helper && !strcmp(helper->name, helpname))
1634                                 err = 0;
1635                         rcu_read_unlock();
1636                 }
1637
1638                 return err;
1639         }
1640
1641         if (!strcmp(helpname, "")) {
1642                 if (help && help->helper) {
1643                         /* we had a helper before ... */
1644                         nf_ct_remove_expectations(ct);
1645                         RCU_INIT_POINTER(help->helper, NULL);
1646                 }
1647
1648                 return 0;
1649         }
1650
1651         rcu_read_lock();
1652         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1653                                             nf_ct_protonum(ct));
1654         if (helper == NULL) {
1655                 rcu_read_unlock();
1656                 return -EOPNOTSUPP;
1657         }
1658
1659         if (help) {
1660                 if (help->helper == helper) {
1661                         /* update private helper data if allowed. */
1662                         if (helper->from_nlattr)
1663                                 helper->from_nlattr(helpinfo, ct);
1664                         err = 0;
1665                 } else
1666                         err = -EBUSY;
1667         } else {
1668                 /* we cannot set a helper for an existing conntrack */
1669                 err = -EOPNOTSUPP;
1670         }
1671
1672         rcu_read_unlock();
1673         return err;
1674 }
1675
1676 static int ctnetlink_change_timeout(struct nf_conn *ct,
1677                                     const struct nlattr * const cda[])
1678 {
1679         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1680
1681         if (timeout > INT_MAX)
1682                 timeout = INT_MAX;
1683         ct->timeout = nfct_time_stamp + (u32)timeout;
1684
1685         if (test_bit(IPS_DYING_BIT, &ct->status))
1686                 return -ETIME;
1687
1688         return 0;
1689 }
1690
1691 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1692         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1693         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1694         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1695 };
1696
1697 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1698                                       const struct nlattr * const cda[])
1699 {
1700         const struct nlattr *attr = cda[CTA_PROTOINFO];
1701         const struct nf_conntrack_l4proto *l4proto;
1702         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1703         int err = 0;
1704
1705         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1706                                NULL);
1707         if (err < 0)
1708                 return err;
1709
1710         rcu_read_lock();
1711         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1712         if (l4proto->from_nlattr)
1713                 err = l4proto->from_nlattr(tb, ct);
1714         rcu_read_unlock();
1715
1716         return err;
1717 }
1718
1719 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1720         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1721         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1722         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1723 };
1724
1725 static int change_seq_adj(struct nf_ct_seqadj *seq,
1726                           const struct nlattr * const attr)
1727 {
1728         int err;
1729         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1730
1731         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1732         if (err < 0)
1733                 return err;
1734
1735         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1736                 return -EINVAL;
1737
1738         seq->correction_pos =
1739                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1740
1741         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1742                 return -EINVAL;
1743
1744         seq->offset_before =
1745                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1746
1747         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1748                 return -EINVAL;
1749
1750         seq->offset_after =
1751                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1752
1753         return 0;
1754 }
1755
1756 static int
1757 ctnetlink_change_seq_adj(struct nf_conn *ct,
1758                          const struct nlattr * const cda[])
1759 {
1760         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1761         int ret = 0;
1762
1763         if (!seqadj)
1764                 return 0;
1765
1766         spin_lock_bh(&ct->lock);
1767         if (cda[CTA_SEQ_ADJ_ORIG]) {
1768                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1769                                      cda[CTA_SEQ_ADJ_ORIG]);
1770                 if (ret < 0)
1771                         goto err;
1772
1773                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1774         }
1775
1776         if (cda[CTA_SEQ_ADJ_REPLY]) {
1777                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1778                                      cda[CTA_SEQ_ADJ_REPLY]);
1779                 if (ret < 0)
1780                         goto err;
1781
1782                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1783         }
1784
1785         spin_unlock_bh(&ct->lock);
1786         return 0;
1787 err:
1788         spin_unlock_bh(&ct->lock);
1789         return ret;
1790 }
1791
1792 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
1793         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
1794         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
1795         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
1796 };
1797
1798 static int ctnetlink_change_synproxy(struct nf_conn *ct,
1799                                      const struct nlattr * const cda[])
1800 {
1801         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
1802         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
1803         int err;
1804
1805         if (!synproxy)
1806                 return 0;
1807
1808         err = nla_parse_nested(tb, CTA_SYNPROXY_MAX, cda[CTA_SYNPROXY],
1809                                synproxy_policy, NULL);
1810         if (err < 0)
1811                 return err;
1812
1813         if (!tb[CTA_SYNPROXY_ISN] ||
1814             !tb[CTA_SYNPROXY_ITS] ||
1815             !tb[CTA_SYNPROXY_TSOFF])
1816                 return -EINVAL;
1817
1818         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
1819         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
1820         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
1821
1822         return 0;
1823 }
1824
1825 static int
1826 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1827 {
1828 #ifdef CONFIG_NF_CONNTRACK_LABELS
1829         size_t len = nla_len(cda[CTA_LABELS]);
1830         const void *mask = cda[CTA_LABELS_MASK];
1831
1832         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1833                 return -EINVAL;
1834
1835         if (mask) {
1836                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1837                     nla_len(cda[CTA_LABELS_MASK]) != len)
1838                         return -EINVAL;
1839                 mask = nla_data(cda[CTA_LABELS_MASK]);
1840         }
1841
1842         len /= sizeof(u32);
1843
1844         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1845 #else
1846         return -EOPNOTSUPP;
1847 #endif
1848 }
1849
1850 static int
1851 ctnetlink_change_conntrack(struct nf_conn *ct,
1852                            const struct nlattr * const cda[])
1853 {
1854         int err;
1855
1856         /* only allow NAT changes and master assignation for new conntracks */
1857         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1858                 return -EOPNOTSUPP;
1859
1860         if (cda[CTA_HELP]) {
1861                 err = ctnetlink_change_helper(ct, cda);
1862                 if (err < 0)
1863                         return err;
1864         }
1865
1866         if (cda[CTA_TIMEOUT]) {
1867                 err = ctnetlink_change_timeout(ct, cda);
1868                 if (err < 0)
1869                         return err;
1870         }
1871
1872         if (cda[CTA_STATUS]) {
1873                 err = ctnetlink_change_status(ct, cda);
1874                 if (err < 0)
1875                         return err;
1876         }
1877
1878         if (cda[CTA_PROTOINFO]) {
1879                 err = ctnetlink_change_protoinfo(ct, cda);
1880                 if (err < 0)
1881                         return err;
1882         }
1883
1884 #if defined(CONFIG_NF_CONNTRACK_MARK)
1885         if (cda[CTA_MARK])
1886                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1887 #endif
1888
1889         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1890                 err = ctnetlink_change_seq_adj(ct, cda);
1891                 if (err < 0)
1892                         return err;
1893         }
1894
1895         if (cda[CTA_SYNPROXY]) {
1896                 err = ctnetlink_change_synproxy(ct, cda);
1897                 if (err < 0)
1898                         return err;
1899         }
1900
1901         if (cda[CTA_LABELS]) {
1902                 err = ctnetlink_attach_labels(ct, cda);
1903                 if (err < 0)
1904                         return err;
1905         }
1906
1907         return 0;
1908 }
1909
1910 static struct nf_conn *
1911 ctnetlink_create_conntrack(struct net *net,
1912                            const struct nf_conntrack_zone *zone,
1913                            const struct nlattr * const cda[],
1914                            struct nf_conntrack_tuple *otuple,
1915                            struct nf_conntrack_tuple *rtuple,
1916                            u8 u3)
1917 {
1918         struct nf_conn *ct;
1919         int err = -EINVAL;
1920         struct nf_conntrack_helper *helper;
1921         struct nf_conn_tstamp *tstamp;
1922         u64 timeout;
1923
1924         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1925         if (IS_ERR(ct))
1926                 return ERR_PTR(-ENOMEM);
1927
1928         if (!cda[CTA_TIMEOUT])
1929                 goto err1;
1930
1931         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1932         if (timeout > INT_MAX)
1933                 timeout = INT_MAX;
1934         ct->timeout = (u32)timeout + nfct_time_stamp;
1935
1936         rcu_read_lock();
1937         if (cda[CTA_HELP]) {
1938                 char *helpname = NULL;
1939                 struct nlattr *helpinfo = NULL;
1940
1941                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1942                 if (err < 0)
1943                         goto err2;
1944
1945                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1946                                                     nf_ct_protonum(ct));
1947                 if (helper == NULL) {
1948                         rcu_read_unlock();
1949 #ifdef CONFIG_MODULES
1950                         if (request_module("nfct-helper-%s", helpname) < 0) {
1951                                 err = -EOPNOTSUPP;
1952                                 goto err1;
1953                         }
1954
1955                         rcu_read_lock();
1956                         helper = __nf_conntrack_helper_find(helpname,
1957                                                             nf_ct_l3num(ct),
1958                                                             nf_ct_protonum(ct));
1959                         if (helper) {
1960                                 err = -EAGAIN;
1961                                 goto err2;
1962                         }
1963                         rcu_read_unlock();
1964 #endif
1965                         err = -EOPNOTSUPP;
1966                         goto err1;
1967                 } else {
1968                         struct nf_conn_help *help;
1969
1970                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1971                         if (help == NULL) {
1972                                 err = -ENOMEM;
1973                                 goto err2;
1974                         }
1975                         /* set private helper data if allowed. */
1976                         if (helper->from_nlattr)
1977                                 helper->from_nlattr(helpinfo, ct);
1978
1979                         /* not in hash table yet so not strictly necessary */
1980                         RCU_INIT_POINTER(help->helper, helper);
1981                 }
1982         } else {
1983                 /* try an implicit helper assignation */
1984                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1985                 if (err < 0)
1986                         goto err2;
1987         }
1988
1989         err = ctnetlink_setup_nat(ct, cda);
1990         if (err < 0)
1991                 goto err2;
1992
1993         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1994         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1995         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1996         nf_ct_labels_ext_add(ct);
1997         nfct_seqadj_ext_add(ct);
1998         nfct_synproxy_ext_add(ct);
1999
2000         /* we must add conntrack extensions before confirmation. */
2001         ct->status |= IPS_CONFIRMED;
2002
2003         if (cda[CTA_STATUS]) {
2004                 err = ctnetlink_change_status(ct, cda);
2005                 if (err < 0)
2006                         goto err2;
2007         }
2008
2009         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2010                 err = ctnetlink_change_seq_adj(ct, cda);
2011                 if (err < 0)
2012                         goto err2;
2013         }
2014
2015         memset(&ct->proto, 0, sizeof(ct->proto));
2016         if (cda[CTA_PROTOINFO]) {
2017                 err = ctnetlink_change_protoinfo(ct, cda);
2018                 if (err < 0)
2019                         goto err2;
2020         }
2021
2022         if (cda[CTA_SYNPROXY]) {
2023                 err = ctnetlink_change_synproxy(ct, cda);
2024                 if (err < 0)
2025                         goto err2;
2026         }
2027
2028 #if defined(CONFIG_NF_CONNTRACK_MARK)
2029         if (cda[CTA_MARK])
2030                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2031 #endif
2032
2033         /* setup master conntrack: this is a confirmed expectation */
2034         if (cda[CTA_TUPLE_MASTER]) {
2035                 struct nf_conntrack_tuple master;
2036                 struct nf_conntrack_tuple_hash *master_h;
2037                 struct nf_conn *master_ct;
2038
2039                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2040                                             u3, NULL);
2041                 if (err < 0)
2042                         goto err2;
2043
2044                 master_h = nf_conntrack_find_get(net, zone, &master);
2045                 if (master_h == NULL) {
2046                         err = -ENOENT;
2047                         goto err2;
2048                 }
2049                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2050                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2051                 ct->master = master_ct;
2052         }
2053         tstamp = nf_conn_tstamp_find(ct);
2054         if (tstamp)
2055                 tstamp->start = ktime_get_real_ns();
2056
2057         err = nf_conntrack_hash_check_insert(ct);
2058         if (err < 0)
2059                 goto err3;
2060
2061         rcu_read_unlock();
2062
2063         return ct;
2064
2065 err3:
2066         if (ct->master)
2067                 nf_ct_put(ct->master);
2068 err2:
2069         rcu_read_unlock();
2070 err1:
2071         nf_conntrack_free(ct);
2072         return ERR_PTR(err);
2073 }
2074
2075 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2076                                    struct sk_buff *skb,
2077                                    const struct nlmsghdr *nlh,
2078                                    const struct nlattr * const cda[],
2079                                    struct netlink_ext_ack *extack)
2080 {
2081         struct nf_conntrack_tuple otuple, rtuple;
2082         struct nf_conntrack_tuple_hash *h = NULL;
2083         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2084         struct nf_conn *ct;
2085         u_int8_t u3 = nfmsg->nfgen_family;
2086         struct nf_conntrack_zone zone;
2087         int err;
2088
2089         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2090         if (err < 0)
2091                 return err;
2092
2093         if (cda[CTA_TUPLE_ORIG]) {
2094                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2095                                             u3, &zone);
2096                 if (err < 0)
2097                         return err;
2098         }
2099
2100         if (cda[CTA_TUPLE_REPLY]) {
2101                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2102                                             u3, &zone);
2103                 if (err < 0)
2104                         return err;
2105         }
2106
2107         if (cda[CTA_TUPLE_ORIG])
2108                 h = nf_conntrack_find_get(net, &zone, &otuple);
2109         else if (cda[CTA_TUPLE_REPLY])
2110                 h = nf_conntrack_find_get(net, &zone, &rtuple);
2111
2112         if (h == NULL) {
2113                 err = -ENOENT;
2114                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2115                         enum ip_conntrack_events events;
2116
2117                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2118                                 return -EINVAL;
2119                         if (otuple.dst.protonum != rtuple.dst.protonum)
2120                                 return -EINVAL;
2121
2122                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2123                                                         &rtuple, u3);
2124                         if (IS_ERR(ct))
2125                                 return PTR_ERR(ct);
2126
2127                         err = 0;
2128                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2129                                 events = 1 << IPCT_RELATED;
2130                         else
2131                                 events = 1 << IPCT_NEW;
2132
2133                         if (cda[CTA_LABELS] &&
2134                             ctnetlink_attach_labels(ct, cda) == 0)
2135                                 events |= (1 << IPCT_LABEL);
2136
2137                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2138                                                       (1 << IPCT_ASSURED) |
2139                                                       (1 << IPCT_HELPER) |
2140                                                       (1 << IPCT_PROTOINFO) |
2141                                                       (1 << IPCT_SEQADJ) |
2142                                                       (1 << IPCT_MARK) |
2143                                                       (1 << IPCT_SYNPROXY) |
2144                                                       events,
2145                                                       ct, NETLINK_CB(skb).portid,
2146                                                       nlmsg_report(nlh));
2147                         nf_ct_put(ct);
2148                 }
2149
2150                 return err;
2151         }
2152         /* implicit 'else' */
2153
2154         err = -EEXIST;
2155         ct = nf_ct_tuplehash_to_ctrack(h);
2156         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2157                 err = ctnetlink_change_conntrack(ct, cda);
2158                 if (err == 0) {
2159                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2160                                                       (1 << IPCT_ASSURED) |
2161                                                       (1 << IPCT_HELPER) |
2162                                                       (1 << IPCT_LABEL) |
2163                                                       (1 << IPCT_PROTOINFO) |
2164                                                       (1 << IPCT_SEQADJ) |
2165                                                       (1 << IPCT_MARK) |
2166                                                       (1 << IPCT_SYNPROXY),
2167                                                       ct, NETLINK_CB(skb).portid,
2168                                                       nlmsg_report(nlh));
2169                 }
2170         }
2171
2172         nf_ct_put(ct);
2173         return err;
2174 }
2175
2176 static int
2177 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2178                                 __u16 cpu, const struct ip_conntrack_stat *st)
2179 {
2180         struct nlmsghdr *nlh;
2181         struct nfgenmsg *nfmsg;
2182         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2183
2184         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2185                               IPCTNL_MSG_CT_GET_STATS_CPU);
2186         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2187         if (nlh == NULL)
2188                 goto nlmsg_failure;
2189
2190         nfmsg = nlmsg_data(nlh);
2191         nfmsg->nfgen_family = AF_UNSPEC;
2192         nfmsg->version      = NFNETLINK_V0;
2193         nfmsg->res_id       = htons(cpu);
2194
2195         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2196             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2197             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2198             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2199             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2200                                 htonl(st->insert_failed)) ||
2201             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2202             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2203             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2204             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2205                                 htonl(st->search_restart)))
2206                 goto nla_put_failure;
2207
2208         nlmsg_end(skb, nlh);
2209         return skb->len;
2210
2211 nla_put_failure:
2212 nlmsg_failure:
2213         nlmsg_cancel(skb, nlh);
2214         return -1;
2215 }
2216
2217 static int
2218 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2219 {
2220         int cpu;
2221         struct net *net = sock_net(skb->sk);
2222
2223         if (cb->args[0] == nr_cpu_ids)
2224                 return 0;
2225
2226         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2227                 const struct ip_conntrack_stat *st;
2228
2229                 if (!cpu_possible(cpu))
2230                         continue;
2231
2232                 st = per_cpu_ptr(net->ct.stat, cpu);
2233                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2234                                                     NETLINK_CB(cb->skb).portid,
2235                                                     cb->nlh->nlmsg_seq,
2236                                                     cpu, st) < 0)
2237                                 break;
2238         }
2239         cb->args[0] = cpu;
2240
2241         return skb->len;
2242 }
2243
2244 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2245                                  struct sk_buff *skb,
2246                                  const struct nlmsghdr *nlh,
2247                                  const struct nlattr * const cda[],
2248                                  struct netlink_ext_ack *extack)
2249 {
2250         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2251                 struct netlink_dump_control c = {
2252                         .dump = ctnetlink_ct_stat_cpu_dump,
2253                 };
2254                 return netlink_dump_start(ctnl, skb, nlh, &c);
2255         }
2256
2257         return 0;
2258 }
2259
2260 static int
2261 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2262                             struct net *net)
2263 {
2264         struct nlmsghdr *nlh;
2265         struct nfgenmsg *nfmsg;
2266         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2267         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2268
2269         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2270         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2271         if (nlh == NULL)
2272                 goto nlmsg_failure;
2273
2274         nfmsg = nlmsg_data(nlh);
2275         nfmsg->nfgen_family = AF_UNSPEC;
2276         nfmsg->version      = NFNETLINK_V0;
2277         nfmsg->res_id       = 0;
2278
2279         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2280                 goto nla_put_failure;
2281
2282         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2283                 goto nla_put_failure;
2284
2285         nlmsg_end(skb, nlh);
2286         return skb->len;
2287
2288 nla_put_failure:
2289 nlmsg_failure:
2290         nlmsg_cancel(skb, nlh);
2291         return -1;
2292 }
2293
2294 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2295                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2296                              const struct nlattr * const cda[],
2297                              struct netlink_ext_ack *extack)
2298 {
2299         struct sk_buff *skb2;
2300         int err;
2301
2302         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2303         if (skb2 == NULL)
2304                 return -ENOMEM;
2305
2306         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2307                                           nlh->nlmsg_seq,
2308                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2309                                           sock_net(skb->sk));
2310         if (err <= 0)
2311                 goto free;
2312
2313         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2314         if (err < 0)
2315                 goto out;
2316
2317         return 0;
2318
2319 free:
2320         kfree_skb(skb2);
2321 out:
2322         /* this avoids a loop in nfnetlink. */
2323         return err == -EAGAIN ? -ENOBUFS : err;
2324 }
2325
2326 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2327         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2328         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2329         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2330         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2331         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2332         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2333                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2334         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2335         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2336         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2337         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2338         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2339 };
2340
2341 static struct nf_conntrack_expect *
2342 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2343                        struct nf_conntrack_helper *helper,
2344                        struct nf_conntrack_tuple *tuple,
2345                        struct nf_conntrack_tuple *mask);
2346
2347 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2348 static size_t
2349 ctnetlink_glue_build_size(const struct nf_conn *ct)
2350 {
2351         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2352                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2353                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2354                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2355                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2356                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2357                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2358                + nla_total_size(0) /* CTA_PROTOINFO */
2359                + nla_total_size(0) /* CTA_HELP */
2360                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2361                + ctnetlink_secctx_size(ct)
2362 #ifdef CONFIG_NF_NAT_NEEDED
2363                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2364                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2365 #endif
2366 #ifdef CONFIG_NF_CONNTRACK_MARK
2367                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2368 #endif
2369 #ifdef CONFIG_NF_CONNTRACK_ZONES
2370                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2371 #endif
2372                + ctnetlink_proto_size(ct)
2373                ;
2374 }
2375
2376 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2377                                              enum ip_conntrack_info *ctinfo)
2378 {
2379         return nf_ct_get(skb, ctinfo);
2380 }
2381
2382 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2383 {
2384         const struct nf_conntrack_zone *zone;
2385         struct nlattr *nest_parms;
2386
2387         zone = nf_ct_zone(ct);
2388
2389         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2390         if (!nest_parms)
2391                 goto nla_put_failure;
2392         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2393                 goto nla_put_failure;
2394         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2395                                    NF_CT_ZONE_DIR_ORIG) < 0)
2396                 goto nla_put_failure;
2397         nla_nest_end(skb, nest_parms);
2398
2399         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2400         if (!nest_parms)
2401                 goto nla_put_failure;
2402         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2403                 goto nla_put_failure;
2404         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2405                                    NF_CT_ZONE_DIR_REPL) < 0)
2406                 goto nla_put_failure;
2407         nla_nest_end(skb, nest_parms);
2408
2409         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2410                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2411                 goto nla_put_failure;
2412
2413         if (ctnetlink_dump_id(skb, ct) < 0)
2414                 goto nla_put_failure;
2415
2416         if (ctnetlink_dump_status(skb, ct) < 0)
2417                 goto nla_put_failure;
2418
2419         if (ctnetlink_dump_timeout(skb, ct) < 0)
2420                 goto nla_put_failure;
2421
2422         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2423                 goto nla_put_failure;
2424
2425         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2426                 goto nla_put_failure;
2427
2428 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2429         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2430                 goto nla_put_failure;
2431 #endif
2432         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2433                 goto nla_put_failure;
2434
2435         if ((ct->status & IPS_SEQ_ADJUST) &&
2436             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2437                 goto nla_put_failure;
2438
2439         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2440                 goto nla_put_failure;
2441
2442 #ifdef CONFIG_NF_CONNTRACK_MARK
2443         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2444                 goto nla_put_failure;
2445 #endif
2446         if (ctnetlink_dump_labels(skb, ct) < 0)
2447                 goto nla_put_failure;
2448         return 0;
2449
2450 nla_put_failure:
2451         return -ENOSPC;
2452 }
2453
2454 static int
2455 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2456                      enum ip_conntrack_info ctinfo,
2457                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2458 {
2459         struct nlattr *nest_parms;
2460
2461         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2462         if (!nest_parms)
2463                 goto nla_put_failure;
2464
2465         if (__ctnetlink_glue_build(skb, ct) < 0)
2466                 goto nla_put_failure;
2467
2468         nla_nest_end(skb, nest_parms);
2469
2470         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2471                 goto nla_put_failure;
2472
2473         return 0;
2474
2475 nla_put_failure:
2476         return -ENOSPC;
2477 }
2478
2479 static int
2480 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2481 {
2482         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2483         unsigned long d = ct->status ^ status;
2484
2485         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2486                 /* SEEN_REPLY bit can only be set */
2487                 return -EBUSY;
2488
2489         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2490                 /* ASSURED bit can only be set */
2491                 return -EBUSY;
2492
2493         /* This check is less strict than ctnetlink_change_status()
2494          * because callers often flip IPS_EXPECTED bits when sending
2495          * an NFQA_CT attribute to the kernel.  So ignore the
2496          * unchangeable bits but do not error out. Also user programs
2497          * are allowed to clear the bits that they are allowed to change.
2498          */
2499         __ctnetlink_change_status(ct, status, ~status);
2500         return 0;
2501 }
2502
2503 static int
2504 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2505 {
2506         int err;
2507
2508         if (cda[CTA_TIMEOUT]) {
2509                 err = ctnetlink_change_timeout(ct, cda);
2510                 if (err < 0)
2511                         return err;
2512         }
2513         if (cda[CTA_STATUS]) {
2514                 err = ctnetlink_update_status(ct, cda);
2515                 if (err < 0)
2516                         return err;
2517         }
2518         if (cda[CTA_HELP]) {
2519                 err = ctnetlink_change_helper(ct, cda);
2520                 if (err < 0)
2521                         return err;
2522         }
2523         if (cda[CTA_LABELS]) {
2524                 err = ctnetlink_attach_labels(ct, cda);
2525                 if (err < 0)
2526                         return err;
2527         }
2528 #if defined(CONFIG_NF_CONNTRACK_MARK)
2529         if (cda[CTA_MARK]) {
2530                 u32 mask = 0, mark, newmark;
2531                 if (cda[CTA_MARK_MASK])
2532                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2533
2534                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2535                 newmark = (ct->mark & mask) ^ mark;
2536                 if (newmark != ct->mark)
2537                         ct->mark = newmark;
2538         }
2539 #endif
2540         return 0;
2541 }
2542
2543 static int
2544 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2545 {
2546         struct nlattr *cda[CTA_MAX+1];
2547         int ret;
2548
2549         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2550         if (ret < 0)
2551                 return ret;
2552
2553         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2554 }
2555
2556 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2557                                     const struct nf_conn *ct,
2558                                     struct nf_conntrack_tuple *tuple,
2559                                     struct nf_conntrack_tuple *mask)
2560 {
2561         int err;
2562
2563         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2564                                     nf_ct_l3num(ct), NULL);
2565         if (err < 0)
2566                 return err;
2567
2568         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2569                                      nf_ct_l3num(ct), NULL);
2570 }
2571
2572 static int
2573 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2574                              u32 portid, u32 report)
2575 {
2576         struct nlattr *cda[CTA_EXPECT_MAX+1];
2577         struct nf_conntrack_tuple tuple, mask;
2578         struct nf_conntrack_helper *helper = NULL;
2579         struct nf_conntrack_expect *exp;
2580         int err;
2581
2582         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2583                                NULL);
2584         if (err < 0)
2585                 return err;
2586
2587         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2588                                        ct, &tuple, &mask);
2589         if (err < 0)
2590                 return err;
2591
2592         if (cda[CTA_EXPECT_HELP_NAME]) {
2593                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2594
2595                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2596                                                     nf_ct_protonum(ct));
2597                 if (helper == NULL)
2598                         return -EOPNOTSUPP;
2599         }
2600
2601         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2602                                      helper, &tuple, &mask);
2603         if (IS_ERR(exp))
2604                 return PTR_ERR(exp);
2605
2606         err = nf_ct_expect_related_report(exp, portid, report);
2607         nf_ct_expect_put(exp);
2608         return err;
2609 }
2610
2611 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2612                                   enum ip_conntrack_info ctinfo, int diff)
2613 {
2614         if (!(ct->status & IPS_NAT_MASK))
2615                 return;
2616
2617         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2618 }
2619
2620 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2621         .get_ct         = ctnetlink_glue_get_ct,
2622         .build_size     = ctnetlink_glue_build_size,
2623         .build          = ctnetlink_glue_build,
2624         .parse          = ctnetlink_glue_parse,
2625         .attach_expect  = ctnetlink_glue_attach_expect,
2626         .seq_adjust     = ctnetlink_glue_seqadj,
2627 };
2628 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2629
2630 /***********************************************************************
2631  * EXPECT
2632  ***********************************************************************/
2633
2634 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2635                                     const struct nf_conntrack_tuple *tuple,
2636                                     u32 type)
2637 {
2638         struct nlattr *nest_parms;
2639
2640         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2641         if (!nest_parms)
2642                 goto nla_put_failure;
2643         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2644                 goto nla_put_failure;
2645         nla_nest_end(skb, nest_parms);
2646
2647         return 0;
2648
2649 nla_put_failure:
2650         return -1;
2651 }
2652
2653 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2654                                    const struct nf_conntrack_tuple *tuple,
2655                                    const struct nf_conntrack_tuple_mask *mask)
2656 {
2657         const struct nf_conntrack_l4proto *l4proto;
2658         struct nf_conntrack_tuple m;
2659         struct nlattr *nest_parms;
2660         int ret;
2661
2662         memset(&m, 0xFF, sizeof(m));
2663         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2664         m.src.u.all = mask->src.u.all;
2665         m.src.l3num = tuple->src.l3num;
2666         m.dst.protonum = tuple->dst.protonum;
2667
2668         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2669         if (!nest_parms)
2670                 goto nla_put_failure;
2671
2672         rcu_read_lock();
2673         ret = ctnetlink_dump_tuples_ip(skb, &m);
2674         if (ret >= 0) {
2675                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2676                                                tuple->dst.protonum);
2677         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2678         }
2679         rcu_read_unlock();
2680
2681         if (unlikely(ret < 0))
2682                 goto nla_put_failure;
2683
2684         nla_nest_end(skb, nest_parms);
2685
2686         return 0;
2687
2688 nla_put_failure:
2689         return -1;
2690 }
2691
2692 static const union nf_inet_addr any_addr;
2693
2694 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2695 {
2696         static __read_mostly siphash_key_t exp_id_seed;
2697         unsigned long a, b, c, d;
2698
2699         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2700
2701         a = (unsigned long)exp;
2702         b = (unsigned long)exp->helper;
2703         c = (unsigned long)exp->master;
2704         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2705
2706 #ifdef CONFIG_64BIT
2707         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2708 #else
2709         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2710 #endif
2711 }
2712
2713 static int
2714 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2715                           const struct nf_conntrack_expect *exp)
2716 {
2717         struct nf_conn *master = exp->master;
2718         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2719         struct nf_conn_help *help;
2720 #ifdef CONFIG_NF_NAT_NEEDED
2721         struct nlattr *nest_parms;
2722         struct nf_conntrack_tuple nat_tuple = {};
2723 #endif
2724         struct nf_ct_helper_expectfn *expfn;
2725
2726         if (timeout < 0)
2727                 timeout = 0;
2728
2729         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2730                 goto nla_put_failure;
2731         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2732                 goto nla_put_failure;
2733         if (ctnetlink_exp_dump_tuple(skb,
2734                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2735                                  CTA_EXPECT_MASTER) < 0)
2736                 goto nla_put_failure;
2737
2738 #ifdef CONFIG_NF_NAT_NEEDED
2739         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2740             exp->saved_proto.all) {
2741                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2742                 if (!nest_parms)
2743                         goto nla_put_failure;
2744
2745                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2746                         goto nla_put_failure;
2747
2748                 nat_tuple.src.l3num = nf_ct_l3num(master);
2749                 nat_tuple.src.u3 = exp->saved_addr;
2750                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2751                 nat_tuple.src.u = exp->saved_proto;
2752
2753                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2754                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2755                         goto nla_put_failure;
2756                 nla_nest_end(skb, nest_parms);
2757         }
2758 #endif
2759         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2760             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
2761             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2762             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2763                 goto nla_put_failure;
2764         help = nfct_help(master);
2765         if (help) {
2766                 struct nf_conntrack_helper *helper;
2767
2768                 helper = rcu_dereference(help->helper);
2769                 if (helper &&
2770                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2771                         goto nla_put_failure;
2772         }
2773         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2774         if (expfn != NULL &&
2775             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2776                 goto nla_put_failure;
2777
2778         return 0;
2779
2780 nla_put_failure:
2781         return -1;
2782 }
2783
2784 static int
2785 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2786                         int event, const struct nf_conntrack_expect *exp)
2787 {
2788         struct nlmsghdr *nlh;
2789         struct nfgenmsg *nfmsg;
2790         unsigned int flags = portid ? NLM_F_MULTI : 0;
2791
2792         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2793         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2794         if (nlh == NULL)
2795                 goto nlmsg_failure;
2796
2797         nfmsg = nlmsg_data(nlh);
2798         nfmsg->nfgen_family = exp->tuple.src.l3num;
2799         nfmsg->version      = NFNETLINK_V0;
2800         nfmsg->res_id       = 0;
2801
2802         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2803                 goto nla_put_failure;
2804
2805         nlmsg_end(skb, nlh);
2806         return skb->len;
2807
2808 nlmsg_failure:
2809 nla_put_failure:
2810         nlmsg_cancel(skb, nlh);
2811         return -1;
2812 }
2813
2814 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2815 static int
2816 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2817 {
2818         struct nf_conntrack_expect *exp = item->exp;
2819         struct net *net = nf_ct_exp_net(exp);
2820         struct nlmsghdr *nlh;
2821         struct nfgenmsg *nfmsg;
2822         struct sk_buff *skb;
2823         unsigned int type, group;
2824         int flags = 0;
2825
2826         if (events & (1 << IPEXP_DESTROY)) {
2827                 type = IPCTNL_MSG_EXP_DELETE;
2828                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2829         } else if (events & (1 << IPEXP_NEW)) {
2830                 type = IPCTNL_MSG_EXP_NEW;
2831                 flags = NLM_F_CREATE|NLM_F_EXCL;
2832                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2833         } else
2834                 return 0;
2835
2836         if (!item->report && !nfnetlink_has_listeners(net, group))
2837                 return 0;
2838
2839         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2840         if (skb == NULL)
2841                 goto errout;
2842
2843         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2844         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2845         if (nlh == NULL)
2846                 goto nlmsg_failure;
2847
2848         nfmsg = nlmsg_data(nlh);
2849         nfmsg->nfgen_family = exp->tuple.src.l3num;
2850         nfmsg->version      = NFNETLINK_V0;
2851         nfmsg->res_id       = 0;
2852
2853         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2854                 goto nla_put_failure;
2855
2856         nlmsg_end(skb, nlh);
2857         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2858         return 0;
2859
2860 nla_put_failure:
2861         nlmsg_cancel(skb, nlh);
2862 nlmsg_failure:
2863         kfree_skb(skb);
2864 errout:
2865         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2866         return 0;
2867 }
2868 #endif
2869 static int ctnetlink_exp_done(struct netlink_callback *cb)
2870 {
2871         if (cb->args[1])
2872                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2873         return 0;
2874 }
2875
2876 static int
2877 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2878 {
2879         struct net *net = sock_net(skb->sk);
2880         struct nf_conntrack_expect *exp, *last;
2881         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2882         u_int8_t l3proto = nfmsg->nfgen_family;
2883
2884         rcu_read_lock();
2885         last = (struct nf_conntrack_expect *)cb->args[1];
2886         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2887 restart:
2888                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2889                                          hnode) {
2890                         if (l3proto && exp->tuple.src.l3num != l3proto)
2891                                 continue;
2892
2893                         if (!net_eq(nf_ct_net(exp->master), net))
2894                                 continue;
2895
2896                         if (cb->args[1]) {
2897                                 if (exp != last)
2898                                         continue;
2899                                 cb->args[1] = 0;
2900                         }
2901                         if (ctnetlink_exp_fill_info(skb,
2902                                                     NETLINK_CB(cb->skb).portid,
2903                                                     cb->nlh->nlmsg_seq,
2904                                                     IPCTNL_MSG_EXP_NEW,
2905                                                     exp) < 0) {
2906                                 if (!refcount_inc_not_zero(&exp->use))
2907                                         continue;
2908                                 cb->args[1] = (unsigned long)exp;
2909                                 goto out;
2910                         }
2911                 }
2912                 if (cb->args[1]) {
2913                         cb->args[1] = 0;
2914                         goto restart;
2915                 }
2916         }
2917 out:
2918         rcu_read_unlock();
2919         if (last)
2920                 nf_ct_expect_put(last);
2921
2922         return skb->len;
2923 }
2924
2925 static int
2926 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2927 {
2928         struct nf_conntrack_expect *exp, *last;
2929         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2930         struct nf_conn *ct = cb->data;
2931         struct nf_conn_help *help = nfct_help(ct);
2932         u_int8_t l3proto = nfmsg->nfgen_family;
2933
2934         if (cb->args[0])
2935                 return 0;
2936
2937         rcu_read_lock();
2938         last = (struct nf_conntrack_expect *)cb->args[1];
2939 restart:
2940         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2941                 if (l3proto && exp->tuple.src.l3num != l3proto)
2942                         continue;
2943                 if (cb->args[1]) {
2944                         if (exp != last)
2945                                 continue;
2946                         cb->args[1] = 0;
2947                 }
2948                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2949                                             cb->nlh->nlmsg_seq,
2950                                             IPCTNL_MSG_EXP_NEW,
2951                                             exp) < 0) {
2952                         if (!refcount_inc_not_zero(&exp->use))
2953                                 continue;
2954                         cb->args[1] = (unsigned long)exp;
2955                         goto out;
2956                 }
2957         }
2958         if (cb->args[1]) {
2959                 cb->args[1] = 0;
2960                 goto restart;
2961         }
2962         cb->args[0] = 1;
2963 out:
2964         rcu_read_unlock();
2965         if (last)
2966                 nf_ct_expect_put(last);
2967
2968         return skb->len;
2969 }
2970
2971 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2972                                  struct sk_buff *skb,
2973                                  const struct nlmsghdr *nlh,
2974                                  const struct nlattr * const cda[],
2975                                  struct netlink_ext_ack *extack)
2976 {
2977         int err;
2978         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2979         u_int8_t u3 = nfmsg->nfgen_family;
2980         struct nf_conntrack_tuple tuple;
2981         struct nf_conntrack_tuple_hash *h;
2982         struct nf_conn *ct;
2983         struct nf_conntrack_zone zone;
2984         struct netlink_dump_control c = {
2985                 .dump = ctnetlink_exp_ct_dump_table,
2986                 .done = ctnetlink_exp_done,
2987         };
2988
2989         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2990                                     u3, NULL);
2991         if (err < 0)
2992                 return err;
2993
2994         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2995         if (err < 0)
2996                 return err;
2997
2998         h = nf_conntrack_find_get(net, &zone, &tuple);
2999         if (!h)
3000                 return -ENOENT;
3001
3002         ct = nf_ct_tuplehash_to_ctrack(h);
3003         /* No expectation linked to this connection tracking. */
3004         if (!nfct_help(ct)) {
3005                 nf_ct_put(ct);
3006                 return 0;
3007         }
3008
3009         c.data = ct;
3010
3011         err = netlink_dump_start(ctnl, skb, nlh, &c);
3012         nf_ct_put(ct);
3013
3014         return err;
3015 }
3016
3017 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
3018                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3019                                 const struct nlattr * const cda[],
3020                                 struct netlink_ext_ack *extack)
3021 {
3022         struct nf_conntrack_tuple tuple;
3023         struct nf_conntrack_expect *exp;
3024         struct sk_buff *skb2;
3025         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3026         u_int8_t u3 = nfmsg->nfgen_family;
3027         struct nf_conntrack_zone zone;
3028         int err;
3029
3030         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3031                 if (cda[CTA_EXPECT_MASTER])
3032                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
3033                                                      extack);
3034                 else {
3035                         struct netlink_dump_control c = {
3036                                 .dump = ctnetlink_exp_dump_table,
3037                                 .done = ctnetlink_exp_done,
3038                         };
3039                         return netlink_dump_start(ctnl, skb, nlh, &c);
3040                 }
3041         }
3042
3043         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3044         if (err < 0)
3045                 return err;
3046
3047         if (cda[CTA_EXPECT_TUPLE])
3048                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3049                                             u3, NULL);
3050         else if (cda[CTA_EXPECT_MASTER])
3051                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3052                                             u3, NULL);
3053         else
3054                 return -EINVAL;
3055
3056         if (err < 0)
3057                 return err;
3058
3059         exp = nf_ct_expect_find_get(net, &zone, &tuple);
3060         if (!exp)
3061                 return -ENOENT;
3062
3063         if (cda[CTA_EXPECT_ID]) {
3064                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3065
3066                 if (id != nf_expect_get_id(exp)) {
3067                         nf_ct_expect_put(exp);
3068                         return -ENOENT;
3069                 }
3070         }
3071
3072         err = -ENOMEM;
3073         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3074         if (skb2 == NULL) {
3075                 nf_ct_expect_put(exp);
3076                 goto out;
3077         }
3078
3079         rcu_read_lock();
3080         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3081                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
3082         rcu_read_unlock();
3083         nf_ct_expect_put(exp);
3084         if (err <= 0)
3085                 goto free;
3086
3087         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
3088         if (err < 0)
3089                 goto out;
3090
3091         return 0;
3092
3093 free:
3094         kfree_skb(skb2);
3095 out:
3096         /* this avoids a loop in nfnetlink. */
3097         return err == -EAGAIN ? -ENOBUFS : err;
3098 }
3099
3100 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3101 {
3102         const struct nf_conn_help *m_help;
3103         const char *name = data;
3104
3105         m_help = nfct_help(exp->master);
3106
3107         return strcmp(m_help->helper->name, name) == 0;
3108 }
3109
3110 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3111 {
3112         return true;
3113 }
3114
3115 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3116                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3117                                 const struct nlattr * const cda[],
3118                                 struct netlink_ext_ack *extack)
3119 {
3120         struct nf_conntrack_expect *exp;
3121         struct nf_conntrack_tuple tuple;
3122         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3123         u_int8_t u3 = nfmsg->nfgen_family;
3124         struct nf_conntrack_zone zone;
3125         int err;
3126
3127         if (cda[CTA_EXPECT_TUPLE]) {
3128                 /* delete a single expect by tuple */
3129                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3130                 if (err < 0)
3131                         return err;
3132
3133                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3134                                             u3, NULL);
3135                 if (err < 0)
3136                         return err;
3137
3138                 /* bump usage count to 2 */
3139                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
3140                 if (!exp)
3141                         return -ENOENT;
3142
3143                 if (cda[CTA_EXPECT_ID]) {
3144                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3145                         if (ntohl(id) != (u32)(unsigned long)exp) {
3146                                 nf_ct_expect_put(exp);
3147                                 return -ENOENT;
3148                         }
3149                 }
3150
3151                 /* after list removal, usage count == 1 */
3152                 spin_lock_bh(&nf_conntrack_expect_lock);
3153                 if (del_timer(&exp->timeout)) {
3154                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3155                                                    nlmsg_report(nlh));
3156                         nf_ct_expect_put(exp);
3157                 }
3158                 spin_unlock_bh(&nf_conntrack_expect_lock);
3159                 /* have to put what we 'get' above.
3160                  * after this line usage count == 0 */
3161                 nf_ct_expect_put(exp);
3162         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3163                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3164
3165                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
3166                                          NETLINK_CB(skb).portid,
3167                                          nlmsg_report(nlh));
3168         } else {
3169                 /* This basically means we have to flush everything*/
3170                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3171                                          NETLINK_CB(skb).portid,
3172                                          nlmsg_report(nlh));
3173         }
3174
3175         return 0;
3176 }
3177 static int
3178 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3179                         const struct nlattr * const cda[])
3180 {
3181         if (cda[CTA_EXPECT_TIMEOUT]) {
3182                 if (!del_timer(&x->timeout))
3183                         return -ETIME;
3184
3185                 x->timeout.expires = jiffies +
3186                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3187                 add_timer(&x->timeout);
3188         }
3189         return 0;
3190 }
3191
3192 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3193         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3194         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3195 };
3196
3197 static int
3198 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3199                            struct nf_conntrack_expect *exp,
3200                            u_int8_t u3)
3201 {
3202 #ifdef CONFIG_NF_NAT_NEEDED
3203         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3204         struct nf_conntrack_tuple nat_tuple = {};
3205         int err;
3206
3207         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3208                                exp_nat_nla_policy, NULL);
3209         if (err < 0)
3210                 return err;
3211
3212         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3213                 return -EINVAL;
3214
3215         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3216                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3217                                     u3, NULL);
3218         if (err < 0)
3219                 return err;
3220
3221         exp->saved_addr = nat_tuple.src.u3;
3222         exp->saved_proto = nat_tuple.src.u;
3223         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3224
3225         return 0;
3226 #else
3227         return -EOPNOTSUPP;
3228 #endif
3229 }
3230
3231 static struct nf_conntrack_expect *
3232 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3233                        struct nf_conntrack_helper *helper,
3234                        struct nf_conntrack_tuple *tuple,
3235                        struct nf_conntrack_tuple *mask)
3236 {
3237         u_int32_t class = 0;
3238         struct nf_conntrack_expect *exp;
3239         struct nf_conn_help *help;
3240         int err;
3241
3242         help = nfct_help(ct);
3243         if (!help)
3244                 return ERR_PTR(-EOPNOTSUPP);
3245
3246         if (cda[CTA_EXPECT_CLASS] && helper) {
3247                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3248                 if (class > helper->expect_class_max)
3249                         return ERR_PTR(-EINVAL);
3250         }
3251         exp = nf_ct_expect_alloc(ct);
3252         if (!exp)
3253                 return ERR_PTR(-ENOMEM);
3254
3255         if (cda[CTA_EXPECT_FLAGS]) {
3256                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3257                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3258         } else {
3259                 exp->flags = 0;
3260         }
3261         if (cda[CTA_EXPECT_FN]) {
3262                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3263                 struct nf_ct_helper_expectfn *expfn;
3264
3265                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3266                 if (expfn == NULL) {
3267                         err = -EINVAL;
3268                         goto err_out;
3269                 }
3270                 exp->expectfn = expfn->expectfn;
3271         } else
3272                 exp->expectfn = NULL;
3273
3274         exp->class = class;
3275         exp->master = ct;
3276         exp->helper = helper;
3277         exp->tuple = *tuple;
3278         exp->mask.src.u3 = mask->src.u3;
3279         exp->mask.src.u.all = mask->src.u.all;
3280
3281         if (cda[CTA_EXPECT_NAT]) {
3282                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3283                                                  exp, nf_ct_l3num(ct));
3284                 if (err < 0)
3285                         goto err_out;
3286         }
3287         return exp;
3288 err_out:
3289         nf_ct_expect_put(exp);
3290         return ERR_PTR(err);
3291 }
3292
3293 static int
3294 ctnetlink_create_expect(struct net *net,
3295                         const struct nf_conntrack_zone *zone,
3296                         const struct nlattr * const cda[],
3297                         u_int8_t u3, u32 portid, int report)
3298 {
3299         struct nf_conntrack_tuple tuple, mask, master_tuple;
3300         struct nf_conntrack_tuple_hash *h = NULL;
3301         struct nf_conntrack_helper *helper = NULL;
3302         struct nf_conntrack_expect *exp;
3303         struct nf_conn *ct;
3304         int err;
3305
3306         /* caller guarantees that those three CTA_EXPECT_* exist */
3307         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3308                                     u3, NULL);
3309         if (err < 0)
3310                 return err;
3311         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3312                                     u3, NULL);
3313         if (err < 0)
3314                 return err;
3315         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3316                                     u3, NULL);
3317         if (err < 0)
3318                 return err;
3319
3320         /* Look for master conntrack of this expectation */
3321         h = nf_conntrack_find_get(net, zone, &master_tuple);
3322         if (!h)
3323                 return -ENOENT;
3324         ct = nf_ct_tuplehash_to_ctrack(h);
3325
3326         rcu_read_lock();
3327         if (cda[CTA_EXPECT_HELP_NAME]) {
3328                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3329
3330                 helper = __nf_conntrack_helper_find(helpname, u3,
3331                                                     nf_ct_protonum(ct));
3332                 if (helper == NULL) {
3333                         rcu_read_unlock();
3334 #ifdef CONFIG_MODULES
3335                         if (request_module("nfct-helper-%s", helpname) < 0) {
3336                                 err = -EOPNOTSUPP;
3337                                 goto err_ct;
3338                         }
3339                         rcu_read_lock();
3340                         helper = __nf_conntrack_helper_find(helpname, u3,
3341                                                             nf_ct_protonum(ct));
3342                         if (helper) {
3343                                 err = -EAGAIN;
3344                                 goto err_rcu;
3345                         }
3346                         rcu_read_unlock();
3347 #endif
3348                         err = -EOPNOTSUPP;
3349                         goto err_ct;
3350                 }
3351         }
3352
3353         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3354         if (IS_ERR(exp)) {
3355                 err = PTR_ERR(exp);
3356                 goto err_rcu;
3357         }
3358
3359         err = nf_ct_expect_related_report(exp, portid, report);
3360         nf_ct_expect_put(exp);
3361 err_rcu:
3362         rcu_read_unlock();
3363 err_ct:
3364         nf_ct_put(ct);
3365         return err;
3366 }
3367
3368 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3369                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3370                                 const struct nlattr * const cda[],
3371                                 struct netlink_ext_ack *extack)
3372 {
3373         struct nf_conntrack_tuple tuple;
3374         struct nf_conntrack_expect *exp;
3375         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3376         u_int8_t u3 = nfmsg->nfgen_family;
3377         struct nf_conntrack_zone zone;
3378         int err;
3379
3380         if (!cda[CTA_EXPECT_TUPLE]
3381             || !cda[CTA_EXPECT_MASK]
3382             || !cda[CTA_EXPECT_MASTER])
3383                 return -EINVAL;
3384
3385         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3386         if (err < 0)
3387                 return err;
3388
3389         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3390                                     u3, NULL);
3391         if (err < 0)
3392                 return err;
3393
3394         spin_lock_bh(&nf_conntrack_expect_lock);
3395         exp = __nf_ct_expect_find(net, &zone, &tuple);
3396         if (!exp) {
3397                 spin_unlock_bh(&nf_conntrack_expect_lock);
3398                 err = -ENOENT;
3399                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3400                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3401                                                       NETLINK_CB(skb).portid,
3402                                                       nlmsg_report(nlh));
3403                 }
3404                 return err;
3405         }
3406
3407         err = -EEXIST;
3408         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3409                 err = ctnetlink_change_expect(exp, cda);
3410         spin_unlock_bh(&nf_conntrack_expect_lock);
3411
3412         return err;
3413 }
3414
3415 static int
3416 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3417                              const struct ip_conntrack_stat *st)
3418 {
3419         struct nlmsghdr *nlh;
3420         struct nfgenmsg *nfmsg;
3421         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3422
3423         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3424                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3425         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3426         if (nlh == NULL)
3427                 goto nlmsg_failure;
3428
3429         nfmsg = nlmsg_data(nlh);
3430         nfmsg->nfgen_family = AF_UNSPEC;
3431         nfmsg->version      = NFNETLINK_V0;
3432         nfmsg->res_id       = htons(cpu);
3433
3434         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3435             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3436             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3437                 goto nla_put_failure;
3438
3439         nlmsg_end(skb, nlh);
3440         return skb->len;
3441
3442 nla_put_failure:
3443 nlmsg_failure:
3444         nlmsg_cancel(skb, nlh);
3445         return -1;
3446 }
3447
3448 static int
3449 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3450 {
3451         int cpu;
3452         struct net *net = sock_net(skb->sk);
3453
3454         if (cb->args[0] == nr_cpu_ids)
3455                 return 0;
3456
3457         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3458                 const struct ip_conntrack_stat *st;
3459
3460                 if (!cpu_possible(cpu))
3461                         continue;
3462
3463                 st = per_cpu_ptr(net->ct.stat, cpu);
3464                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3465                                                  cb->nlh->nlmsg_seq,
3466                                                  cpu, st) < 0)
3467                         break;
3468         }
3469         cb->args[0] = cpu;
3470
3471         return skb->len;
3472 }
3473
3474 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3475                                   struct sk_buff *skb,
3476                                   const struct nlmsghdr *nlh,
3477                                   const struct nlattr * const cda[],
3478                                   struct netlink_ext_ack *extack)
3479 {
3480         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3481                 struct netlink_dump_control c = {
3482                         .dump = ctnetlink_exp_stat_cpu_dump,
3483                 };
3484                 return netlink_dump_start(ctnl, skb, nlh, &c);
3485         }
3486
3487         return 0;
3488 }
3489
3490 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3491 static struct nf_ct_event_notifier ctnl_notifier = {
3492         .fcn = ctnetlink_conntrack_event,
3493 };
3494
3495 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3496         .fcn = ctnetlink_expect_event,
3497 };
3498 #endif
3499
3500 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3501         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3502                                             .attr_count = CTA_MAX,
3503                                             .policy = ct_nla_policy },
3504         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3505                                             .attr_count = CTA_MAX,
3506                                             .policy = ct_nla_policy },
3507         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3508                                             .attr_count = CTA_MAX,
3509                                             .policy = ct_nla_policy },
3510         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3511                                             .attr_count = CTA_MAX,
3512                                             .policy = ct_nla_policy },
3513         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3514         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3515         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3516         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3517 };
3518
3519 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3520         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3521                                             .attr_count = CTA_EXPECT_MAX,
3522                                             .policy = exp_nla_policy },
3523         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3524                                             .attr_count = CTA_EXPECT_MAX,
3525                                             .policy = exp_nla_policy },
3526         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3527                                             .attr_count = CTA_EXPECT_MAX,
3528                                             .policy = exp_nla_policy },
3529         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3530 };
3531
3532 static const struct nfnetlink_subsystem ctnl_subsys = {
3533         .name                           = "conntrack",
3534         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3535         .cb_count                       = IPCTNL_MSG_MAX,
3536         .cb                             = ctnl_cb,
3537 };
3538
3539 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3540         .name                           = "conntrack_expect",
3541         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3542         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3543         .cb                             = ctnl_exp_cb,
3544 };
3545
3546 MODULE_ALIAS("ip_conntrack_netlink");
3547 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3548 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3549
3550 static int __net_init ctnetlink_net_init(struct net *net)
3551 {
3552 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3553         int ret;
3554
3555         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3556         if (ret < 0) {
3557                 pr_err("ctnetlink_init: cannot register notifier.\n");
3558                 goto err_out;
3559         }
3560
3561         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3562         if (ret < 0) {
3563                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3564                 goto err_unreg_notifier;
3565         }
3566 #endif
3567         return 0;
3568
3569 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3570 err_unreg_notifier:
3571         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3572 err_out:
3573         return ret;
3574 #endif
3575 }
3576
3577 static void ctnetlink_net_exit(struct net *net)
3578 {
3579 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3580         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3581         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3582 #endif
3583 }
3584
3585 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3586 {
3587         struct net *net;
3588
3589         list_for_each_entry(net, net_exit_list, exit_list)
3590                 ctnetlink_net_exit(net);
3591
3592         /* wait for other cpus until they are done with ctnl_notifiers */
3593         synchronize_rcu();
3594 }
3595
3596 static struct pernet_operations ctnetlink_net_ops = {
3597         .init           = ctnetlink_net_init,
3598         .exit_batch     = ctnetlink_net_exit_batch,
3599 };
3600
3601 static int __init ctnetlink_init(void)
3602 {
3603         int ret;
3604
3605         ret = nfnetlink_subsys_register(&ctnl_subsys);
3606         if (ret < 0) {
3607                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3608                 goto err_out;
3609         }
3610
3611         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3612         if (ret < 0) {
3613                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3614                 goto err_unreg_subsys;
3615         }
3616
3617         ret = register_pernet_subsys(&ctnetlink_net_ops);
3618         if (ret < 0) {
3619                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3620                 goto err_unreg_exp_subsys;
3621         }
3622 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3623         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3624         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3625 #endif
3626         return 0;
3627
3628 err_unreg_exp_subsys:
3629         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3630 err_unreg_subsys:
3631         nfnetlink_subsys_unregister(&ctnl_subsys);
3632 err_out:
3633         return ret;
3634 }
3635
3636 static void __exit ctnetlink_exit(void)
3637 {
3638         unregister_pernet_subsys(&ctnetlink_net_ops);
3639         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3640         nfnetlink_subsys_unregister(&ctnl_subsys);
3641 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3642         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3643 #endif
3644         synchronize_rcu();
3645 }
3646
3647 module_init(ctnetlink_init);
3648 module_exit(ctnetlink_exit);