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