net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32 #include <linux/siphash.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netlink.h>
  36 #include <net/sock.h>
  37 #include <net/netfilter/nf_conntrack.h>
  38 #include <net/netfilter/nf_conntrack_core.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_seqadj.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #include <net/netfilter/nf_conntrack_synproxy.h>
  49 #if IS_ENABLED(CONFIG_NF_NAT)
  50 #include <net/netfilter/nf_nat.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 MODULE_LICENSE("GPL");
  58
  59 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  60                                 const struct nf_conntrack_tuple *tuple,
  61                                 const struct nf_conntrack_l4proto *l4proto)
  62 {
  63         int ret = 0;
  64         struct nlattr *nest_parms;
  65
  66         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  67         if (!nest_parms)
  68                 goto nla_put_failure;
  69         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  70                 goto nla_put_failure;
  71
  72         if (likely(l4proto->tuple_to_nlattr))
  73                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  74
  75         nla_nest_end(skb, nest_parms);
  76
  77         return ret;
  78
  79 nla_put_failure:
  80         return -1;
  81 }
  82
  83 static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  84                                 const struct nf_conntrack_tuple *tuple)
  85 {
  86         if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  87             nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  88                 return -EMSGSIZE;
  89         return 0;
  90 }
  91
  92 static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
  93                                 const struct nf_conntrack_tuple *tuple)
  94 {
  95         if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
  96             nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
  97                 return -EMSGSIZE;
  98         return 0;
  99 }
 100
 101 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 102                                     const struct nf_conntrack_tuple *tuple)
 103 {
 104         int ret = 0;
 105         struct nlattr *nest_parms;
 106
 107         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
 108         if (!nest_parms)
 109                 goto nla_put_failure;
 110
 111         switch (tuple->src.l3num) {
 112         case NFPROTO_IPV4:
 113                 ret = ipv4_tuple_to_nlattr(skb, tuple);
 114                 break;
 115         case NFPROTO_IPV6:
 116                 ret = ipv6_tuple_to_nlattr(skb, tuple);
 117                 break;
 118         }
 119
 120         nla_nest_end(skb, nest_parms);
 121
 122         return ret;
 123
 124 nla_put_failure:
 125         return -1;
 126 }
 127
 128 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 129                                  const struct nf_conntrack_tuple *tuple)
 130 {
 131         const struct nf_conntrack_l4proto *l4proto;
 132         int ret;
 133
 134         rcu_read_lock();
 135         ret = ctnetlink_dump_tuples_ip(skb, tuple);
 136
 137         if (ret >= 0) {
 138                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 139                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 140         }
 141         rcu_read_unlock();
 142         return ret;
 143 }
 144
 145 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 146                                   const struct nf_conntrack_zone *zone, int dir)
 147 {
 148         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 149                 return 0;
 150         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 151                 goto nla_put_failure;
 152         return 0;
 153
 154 nla_put_failure:
 155         return -1;
 156 }
 157
 158 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 159 {
 160         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 161                 goto nla_put_failure;
 162         return 0;
 163
 164 nla_put_failure:
 165         return -1;
 166 }
 167
 168 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 169 {
 170         long timeout = nf_ct_expires(ct) / HZ;
 171
 172         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 173                 goto nla_put_failure;
 174         return 0;
 175
 176 nla_put_failure:
 177         return -1;
 178 }
 179
 180 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 181 {
 182         const struct nf_conntrack_l4proto *l4proto;
 183         struct nlattr *nest_proto;
 184         int ret;
 185
 186         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 187         if (!l4proto->to_nlattr)
 188                 return 0;
 189
 190         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 191         if (!nest_proto)
 192                 goto nla_put_failure;
 193
 194         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 195
 196         nla_nest_end(skb, nest_proto);
 197
 198         return ret;
 199
 200 nla_put_failure:
 201         return -1;
 202 }
 203
 204 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 205                                    const struct nf_conn *ct)
 206 {
 207         struct nlattr *nest_helper;
 208         const struct nf_conn_help *help = nfct_help(ct);
 209         struct nf_conntrack_helper *helper;
 210
 211         if (!help)
 212                 return 0;
 213
 214         helper = rcu_dereference(help->helper);
 215         if (!helper)
 216                 goto out;
 217
 218         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 219         if (!nest_helper)
 220                 goto nla_put_failure;
 221         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 222                 goto nla_put_failure;
 223
 224         if (helper->to_nlattr)
 225                 helper->to_nlattr(skb, ct);
 226
 227         nla_nest_end(skb, nest_helper);
 228 out:
 229         return 0;
 230
 231 nla_put_failure:
 232         return -1;
 233 }
 234
 235 static int
 236 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 237               enum ip_conntrack_dir dir, int type)
 238 {
 239         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 240         struct nf_conn_counter *counter = acct->counter;
 241         struct nlattr *nest_count;
 242         u64 pkts, bytes;
 243
 244         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 245                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 246                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 247         } else {
 248                 pkts = atomic64_read(&counter[dir].packets);
 249                 bytes = atomic64_read(&counter[dir].bytes);
 250         }
 251
 252         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 253         if (!nest_count)
 254                 goto nla_put_failure;
 255
 256         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 257                          CTA_COUNTERS_PAD) ||
 258             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 259                          CTA_COUNTERS_PAD))
 260                 goto nla_put_failure;
 261
 262         nla_nest_end(skb, nest_count);
 263
 264         return 0;
 265
 266 nla_put_failure:
 267         return -1;
 268 }
 269
 270 static int
 271 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 272 {
 273         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 274
 275         if (!acct)
 276                 return 0;
 277
 278         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 279                 return -1;
 280         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 281                 return -1;
 282
 283         return 0;
 284 }
 285
 286 static int
 287 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 288 {
 289         struct nlattr *nest_count;
 290         const struct nf_conn_tstamp *tstamp;
 291
 292         tstamp = nf_conn_tstamp_find(ct);
 293         if (!tstamp)
 294                 return 0;
 295
 296         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 297         if (!nest_count)
 298                 goto nla_put_failure;
 299
 300         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 301                          CTA_TIMESTAMP_PAD) ||
 302             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 303                                                cpu_to_be64(tstamp->stop),
 304                                                CTA_TIMESTAMP_PAD)))
 305                 goto nla_put_failure;
 306         nla_nest_end(skb, nest_count);
 307
 308         return 0;
 309
 310 nla_put_failure:
 311         return -1;
 312 }
 313
 314 #ifdef CONFIG_NF_CONNTRACK_MARK
 315 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 316 {
 317         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 318                 goto nla_put_failure;
 319         return 0;
 320
 321 nla_put_failure:
 322         return -1;
 323 }
 324 #else
 325 #define ctnetlink_dump_mark(a, b) (0)
 326 #endif
 327
 328 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 329 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 330 {
 331         struct nlattr *nest_secctx;
 332         int len, ret;
 333         char *secctx;
 334
 335         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 336         if (ret)
 337                 return 0;
 338
 339         ret = -1;
 340         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 341         if (!nest_secctx)
 342                 goto nla_put_failure;
 343
 344         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 345                 goto nla_put_failure;
 346         nla_nest_end(skb, nest_secctx);
 347
 348         ret = 0;
 349 nla_put_failure:
 350         security_release_secctx(secctx, len);
 351         return ret;
 352 }
 353 #else
 354 #define ctnetlink_dump_secctx(a, b) (0)
 355 #endif
 356
 357 #ifdef CONFIG_NF_CONNTRACK_LABELS
 358 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 359 {
 360         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 361
 362         if (!labels)
 363                 return 0;
 364         return nla_total_size(sizeof(labels->bits));
 365 }
 366
 367 static int
 368 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 369 {
 370         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 371         unsigned int i;
 372
 373         if (!labels)
 374                 return 0;
 375
 376         i = 0;
 377         do {
 378                 if (labels->bits[i] != 0)
 379                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 380                                        labels->bits);
 381                 i++;
 382         } while (i < ARRAY_SIZE(labels->bits));
 383
 384         return 0;
 385 }
 386 #else
 387 #define ctnetlink_dump_labels(a, b) (0)
 388 #define ctnetlink_label_size(a) (0)
 389 #endif
 390
 391 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 392
 393 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 394 {
 395         struct nlattr *nest_parms;
 396
 397         if (!(ct->status & IPS_EXPECTED))
 398                 return 0;
 399
 400         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 401         if (!nest_parms)
 402                 goto nla_put_failure;
 403         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 404                 goto nla_put_failure;
 405         nla_nest_end(skb, nest_parms);
 406
 407         return 0;
 408
 409 nla_put_failure:
 410         return -1;
 411 }
 412
 413 static int
 414 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 415 {
 416         struct nlattr *nest_parms;
 417
 418         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 419         if (!nest_parms)
 420                 goto nla_put_failure;
 421
 422         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 423                          htonl(seq->correction_pos)) ||
 424             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 425                          htonl(seq->offset_before)) ||
 426             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 427                          htonl(seq->offset_after)))
 428                 goto nla_put_failure;
 429
 430         nla_nest_end(skb, nest_parms);
 431
 432         return 0;
 433
 434 nla_put_failure:
 435         return -1;
 436 }
 437
 438 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 439 {
 440         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 441         struct nf_ct_seqadj *seq;
 442
 443         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 444                 return 0;
 445
 446         spin_lock_bh(&ct->lock);
 447         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 448         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 449                 goto err;
 450
 451         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 452         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 453                 goto err;
 454
 455         spin_unlock_bh(&ct->lock);
 456         return 0;
 457 err:
 458         spin_unlock_bh(&ct->lock);
 459         return -1;
 460 }
 461
 462 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 463 {
 464         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 465         struct nlattr *nest_parms;
 466
 467         if (!synproxy)
 468                 return 0;
 469
 470         nest_parms = nla_nest_start(skb, CTA_SYNPROXY | NLA_F_NESTED);
 471         if (!nest_parms)
 472                 goto nla_put_failure;
 473
 474         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 475             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 476             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 477                 goto nla_put_failure;
 478
 479         nla_nest_end(skb, nest_parms);
 480
 481         return 0;
 482
 483 nla_put_failure:
 484         return -1;
 485 }
 486
 487 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 488 {
 489         __be32 id = (__force __be32)nf_ct_get_id(ct);
 490
 491         if (nla_put_be32(skb, CTA_ID, id))
 492                 goto nla_put_failure;
 493         return 0;
 494
 495 nla_put_failure:
 496         return -1;
 497 }
 498
 499 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 500 {
 501         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 502                 goto nla_put_failure;
 503         return 0;
 504
 505 nla_put_failure:
 506         return -1;
 507 }
 508
 509 static int
 510 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 511                     struct nf_conn *ct)
 512 {
 513         const struct nf_conntrack_zone *zone;
 514         struct nlmsghdr *nlh;
 515         struct nfgenmsg *nfmsg;
 516         struct nlattr *nest_parms;
 517         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 518
 519         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 520         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 521         if (nlh == NULL)
 522                 goto nlmsg_failure;
 523
 524         nfmsg = nlmsg_data(nlh);
 525         nfmsg->nfgen_family = nf_ct_l3num(ct);
 526         nfmsg->version      = NFNETLINK_V0;
 527         nfmsg->res_id       = 0;
 528
 529         zone = nf_ct_zone(ct);
 530
 531         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 532         if (!nest_parms)
 533                 goto nla_put_failure;
 534         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 535                 goto nla_put_failure;
 536         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 537                                    NF_CT_ZONE_DIR_ORIG) < 0)
 538                 goto nla_put_failure;
 539         nla_nest_end(skb, nest_parms);
 540
 541         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 542         if (!nest_parms)
 543                 goto nla_put_failure;
 544         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 545                 goto nla_put_failure;
 546         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 547                                    NF_CT_ZONE_DIR_REPL) < 0)
 548                 goto nla_put_failure;
 549         nla_nest_end(skb, nest_parms);
 550
 551         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 552                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 553                 goto nla_put_failure;
 554
 555         if (ctnetlink_dump_status(skb, ct) < 0 ||
 556             ctnetlink_dump_timeout(skb, ct) < 0 ||
 557             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 558             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 559             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 560             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 561             ctnetlink_dump_mark(skb, ct) < 0 ||
 562             ctnetlink_dump_secctx(skb, ct) < 0 ||
 563             ctnetlink_dump_labels(skb, ct) < 0 ||
 564             ctnetlink_dump_id(skb, ct) < 0 ||
 565             ctnetlink_dump_use(skb, ct) < 0 ||
 566             ctnetlink_dump_master(skb, ct) < 0 ||
 567             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 568             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 569                 goto nla_put_failure;
 570
 571         nlmsg_end(skb, nlh);
 572         return skb->len;
 573
 574 nlmsg_failure:
 575 nla_put_failure:
 576         nlmsg_cancel(skb, nlh);
 577         return -1;
 578 }
 579
 580 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 581         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 582         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 583         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 584         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 585 };
 586
 587 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 588 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 589 {
 590         const struct nf_conntrack_l4proto *l4proto;
 591         size_t len, len4 = 0;
 592
 593         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 594         len *= 3u; /* ORIG, REPLY, MASTER */
 595
 596         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 597         len += l4proto->nlattr_size;
 598         if (l4proto->nlattr_tuple_size) {
 599                 len4 = l4proto->nlattr_tuple_size();
 600                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 601         }
 602
 603         return len + len4;
 604 }
 605 #endif
 606
 607 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 608 {
 609         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 610                 return 0;
 611         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 612                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 613                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 614                ;
 615 }
 616
 617 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 618 {
 619 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 620         int len, ret;
 621
 622         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 623         if (ret)
 624                 return 0;
 625
 626         return nla_total_size(0) /* CTA_SECCTX */
 627                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 628 #else
 629         return 0;
 630 #endif
 631 }
 632
 633 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 634 {
 635 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 636         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 637                 return 0;
 638         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 639 #else
 640         return 0;
 641 #endif
 642 }
 643
 644 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 645 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 646 {
 647         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 648                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 649                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 650                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 651                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 652                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 653                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 654                + ctnetlink_acct_size(ct)
 655                + ctnetlink_timestamp_size(ct)
 656                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 657                + nla_total_size(0) /* CTA_PROTOINFO */
 658                + nla_total_size(0) /* CTA_HELP */
 659                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 660                + ctnetlink_secctx_size(ct)
 661 #if IS_ENABLED(CONFIG_NF_NAT)
 662                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 663                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 664 #endif
 665 #ifdef CONFIG_NF_CONNTRACK_MARK
 666                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 667 #endif
 668 #ifdef CONFIG_NF_CONNTRACK_ZONES
 669                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 670 #endif
 671                + ctnetlink_proto_size(ct)
 672                + ctnetlink_label_size(ct)
 673                ;
 674 }
 675
 676 static int
 677 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 678 {
 679         const struct nf_conntrack_zone *zone;
 680         struct net *net;
 681         struct nlmsghdr *nlh;
 682         struct nfgenmsg *nfmsg;
 683         struct nlattr *nest_parms;
 684         struct nf_conn *ct = item->ct;
 685         struct sk_buff *skb;
 686         unsigned int type;
 687         unsigned int flags = 0, group;
 688         int err;
 689
 690         if (events & (1 << IPCT_DESTROY)) {
 691                 type = IPCTNL_MSG_CT_DELETE;
 692                 group = NFNLGRP_CONNTRACK_DESTROY;
 693         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 694                 type = IPCTNL_MSG_CT_NEW;
 695                 flags = NLM_F_CREATE|NLM_F_EXCL;
 696                 group = NFNLGRP_CONNTRACK_NEW;
 697         } else if (events) {
 698                 type = IPCTNL_MSG_CT_NEW;
 699                 group = NFNLGRP_CONNTRACK_UPDATE;
 700         } else
 701                 return 0;
 702
 703         net = nf_ct_net(ct);
 704         if (!item->report && !nfnetlink_has_listeners(net, group))
 705                 return 0;
 706
 707         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 708         if (skb == NULL)
 709                 goto errout;
 710
 711         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 712         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 713         if (nlh == NULL)
 714                 goto nlmsg_failure;
 715
 716         nfmsg = nlmsg_data(nlh);
 717         nfmsg->nfgen_family = nf_ct_l3num(ct);
 718         nfmsg->version  = NFNETLINK_V0;
 719         nfmsg->res_id   = 0;
 720
 721         zone = nf_ct_zone(ct);
 722
 723         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 724         if (!nest_parms)
 725                 goto nla_put_failure;
 726         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 727                 goto nla_put_failure;
 728         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 729                                    NF_CT_ZONE_DIR_ORIG) < 0)
 730                 goto nla_put_failure;
 731         nla_nest_end(skb, nest_parms);
 732
 733         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 734         if (!nest_parms)
 735                 goto nla_put_failure;
 736         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 737                 goto nla_put_failure;
 738         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 739                                    NF_CT_ZONE_DIR_REPL) < 0)
 740                 goto nla_put_failure;
 741         nla_nest_end(skb, nest_parms);
 742
 743         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 744                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 745                 goto nla_put_failure;
 746
 747         if (ctnetlink_dump_id(skb, ct) < 0)
 748                 goto nla_put_failure;
 749
 750         if (ctnetlink_dump_status(skb, ct) < 0)
 751                 goto nla_put_failure;
 752
 753         if (events & (1 << IPCT_DESTROY)) {
 754                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 755                     ctnetlink_dump_timestamp(skb, ct) < 0)
 756                         goto nla_put_failure;
 757         } else {
 758                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 759                         goto nla_put_failure;
 760
 761                 if (events & (1 << IPCT_PROTOINFO)
 762                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 763                         goto nla_put_failure;
 764
 765                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 766                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 767                         goto nla_put_failure;
 768
 769 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 770                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 771                     && ctnetlink_dump_secctx(skb, ct) < 0)
 772                         goto nla_put_failure;
 773 #endif
 774                 if (events & (1 << IPCT_LABEL) &&
 775                      ctnetlink_dump_labels(skb, ct) < 0)
 776                         goto nla_put_failure;
 777
 778                 if (events & (1 << IPCT_RELATED) &&
 779                     ctnetlink_dump_master(skb, ct) < 0)
 780                         goto nla_put_failure;
 781
 782                 if (events & (1 << IPCT_SEQADJ) &&
 783                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 784                         goto nla_put_failure;
 785
 786                 if (events & (1 << IPCT_SYNPROXY) &&
 787                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 788                         goto nla_put_failure;
 789         }
 790
 791 #ifdef CONFIG_NF_CONNTRACK_MARK
 792         if ((events & (1 << IPCT_MARK) || ct->mark)
 793             && ctnetlink_dump_mark(skb, ct) < 0)
 794                 goto nla_put_failure;
 795 #endif
 796         nlmsg_end(skb, nlh);
 797         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 798                              GFP_ATOMIC);
 799         if (err == -ENOBUFS || err == -EAGAIN)
 800                 return -ENOBUFS;
 801
 802         return 0;
 803
 804 nla_put_failure:
 805         nlmsg_cancel(skb, nlh);
 806 nlmsg_failure:
 807         kfree_skb(skb);
 808 errout:
 809         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 810                 return -ENOBUFS;
 811
 812         return 0;
 813 }
 814 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 815
 816 static int ctnetlink_done(struct netlink_callback *cb)
 817 {
 818         if (cb->args[1])
 819                 nf_ct_put((struct nf_conn *)cb->args[1]);
 820         kfree(cb->data);
 821         return 0;
 822 }
 823
 824 struct ctnetlink_filter {
 825         u8 family;
 826         struct {
 827                 u_int32_t val;
 828                 u_int32_t mask;
 829         } mark;
 830 };
 831
 832 static struct ctnetlink_filter *
 833 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 834 {
 835         struct ctnetlink_filter *filter;
 836
 837 #ifndef CONFIG_NF_CONNTRACK_MARK
 838         if (cda[CTA_MARK] && cda[CTA_MARK_MASK])
 839                 return ERR_PTR(-EOPNOTSUPP);
 840 #endif
 841
 842         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 843         if (filter == NULL)
 844                 return ERR_PTR(-ENOMEM);
 845
 846         filter->family = family;
 847
 848 #ifdef CONFIG_NF_CONNTRACK_MARK
 849         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
 850                 filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 851                 filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 852         }
 853 #endif
 854         return filter;
 855 }
 856
 857 static int ctnetlink_start(struct netlink_callback *cb)
 858 {
 859         const struct nlattr * const *cda = cb->data;
 860         struct ctnetlink_filter *filter = NULL;
 861         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 862         u8 family = nfmsg->nfgen_family;
 863
 864         if (family || (cda[CTA_MARK] && cda[CTA_MARK_MASK])) {
 865                 filter = ctnetlink_alloc_filter(cda, family);
 866                 if (IS_ERR(filter))
 867                         return PTR_ERR(filter);
 868         }
 869
 870         cb->data = filter;
 871         return 0;
 872 }
 873
 874 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 875 {
 876         struct ctnetlink_filter *filter = data;
 877
 878         if (filter == NULL)
 879                 goto out;
 880
 881         /* Match entries of a given L3 protocol number.
 882          * If it is not specified, ie. l3proto == 0,
 883          * then match everything.
 884          */
 885         if (filter->family && nf_ct_l3num(ct) != filter->family)
 886                 goto ignore_entry;
 887
 888 #ifdef CONFIG_NF_CONNTRACK_MARK
 889         if ((ct->mark & filter->mark.mask) != filter->mark.val)
 890                 goto ignore_entry;
 891 #endif
 892
 893 out:
 894         return 1;
 895
 896 ignore_entry:
 897         return 0;
 898 }
 899
 900 static int
 901 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 902 {
 903         struct net *net = sock_net(skb->sk);
 904         struct nf_conn *ct, *last;
 905         struct nf_conntrack_tuple_hash *h;
 906         struct hlist_nulls_node *n;
 907         struct nf_conn *nf_ct_evict[8];
 908         int res, i;
 909         spinlock_t *lockp;
 910
 911         last = (struct nf_conn *)cb->args[1];
 912         i = 0;
 913
 914         local_bh_disable();
 915         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 916 restart:
 917                 while (i) {
 918                         i--;
 919                         if (nf_ct_should_gc(nf_ct_evict[i]))
 920                                 nf_ct_kill(nf_ct_evict[i]);
 921                         nf_ct_put(nf_ct_evict[i]);
 922                 }
 923
 924                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 925                 nf_conntrack_lock(lockp);
 926                 if (cb->args[0] >= nf_conntrack_htable_size) {
 927                         spin_unlock(lockp);
 928                         goto out;
 929                 }
 930                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 931                                            hnnode) {
 932                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 933                                 continue;
 934                         ct = nf_ct_tuplehash_to_ctrack(h);
 935                         if (nf_ct_is_expired(ct)) {
 936                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 937                                     atomic_inc_not_zero(&ct->ct_general.use))
 938                                         nf_ct_evict[i++] = ct;
 939                                 continue;
 940                         }
 941
 942                         if (!net_eq(net, nf_ct_net(ct)))
 943                                 continue;
 944
 945                         if (cb->args[1]) {
 946                                 if (ct != last)
 947                                         continue;
 948                                 cb->args[1] = 0;
 949                         }
 950                         if (!ctnetlink_filter_match(ct, cb->data))
 951                                 continue;
 952
 953                         rcu_read_lock();
 954                         res =
 955                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 956                                             cb->nlh->nlmsg_seq,
 957                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 958                                             ct);
 959                         rcu_read_unlock();
 960                         if (res < 0) {
 961                                 nf_conntrack_get(&ct->ct_general);
 962                                 cb->args[1] = (unsigned long)ct;
 963                                 spin_unlock(lockp);
 964                                 goto out;
 965                         }
 966                 }
 967                 spin_unlock(lockp);
 968                 if (cb->args[1]) {
 969                         cb->args[1] = 0;
 970                         goto restart;
 971                 }
 972         }
 973 out:
 974         local_bh_enable();
 975         if (last) {
 976                 /* nf ct hash resize happened, now clear the leftover. */
 977                 if ((struct nf_conn *)cb->args[1] == last)
 978                         cb->args[1] = 0;
 979
 980                 nf_ct_put(last);
 981         }
 982
 983         while (i) {
 984                 i--;
 985                 if (nf_ct_should_gc(nf_ct_evict[i]))
 986                         nf_ct_kill(nf_ct_evict[i]);
 987                 nf_ct_put(nf_ct_evict[i]);
 988         }
 989
 990         return skb->len;
 991 }
 992
 993 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
 994                                 struct nf_conntrack_tuple *t)
 995 {
 996         if (!tb[CTA_IP_V4_SRC] || !tb[CTA_IP_V4_DST])
 997                 return -EINVAL;
 998
 999         t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1000         t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1001
1002         return 0;
1003 }
1004
1005 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1006                                 struct nf_conntrack_tuple *t)
1007 {
1008         if (!tb[CTA_IP_V6_SRC] || !tb[CTA_IP_V6_DST])
1009                 return -EINVAL;
1010
1011         t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1012         t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1013
1014         return 0;
1015 }
1016
1017 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1018                                     struct nf_conntrack_tuple *tuple)
1019 {
1020         struct nlattr *tb[CTA_IP_MAX+1];
1021         int ret = 0;
1022
1023         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL, NULL);
1024         if (ret < 0)
1025                 return ret;
1026
1027         ret = nla_validate_nested(attr, CTA_IP_MAX,
1028                                   cta_ip_nla_policy, NULL);
1029         if (ret)
1030                 return ret;
1031
1032         switch (tuple->src.l3num) {
1033         case NFPROTO_IPV4:
1034                 ret = ipv4_nlattr_to_tuple(tb, tuple);
1035                 break;
1036         case NFPROTO_IPV6:
1037                 ret = ipv6_nlattr_to_tuple(tb, tuple);
1038                 break;
1039         }
1040
1041         return ret;
1042 }
1043
1044 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1045         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1046 };
1047
1048 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1049                                        struct nf_conntrack_tuple *tuple)
1050 {
1051         const struct nf_conntrack_l4proto *l4proto;
1052         struct nlattr *tb[CTA_PROTO_MAX+1];
1053         int ret = 0;
1054
1055         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy,
1056                                NULL);
1057         if (ret < 0)
1058                 return ret;
1059
1060         if (!tb[CTA_PROTO_NUM])
1061                 return -EINVAL;
1062         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1063
1064         rcu_read_lock();
1065         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1066
1067         if (likely(l4proto->nlattr_to_tuple)) {
1068                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
1069                                           l4proto->nla_policy, NULL);
1070                 if (ret == 0)
1071                         ret = l4proto->nlattr_to_tuple(tb, tuple);
1072         }
1073
1074         rcu_read_unlock();
1075
1076         return ret;
1077 }
1078
1079 static int
1080 ctnetlink_parse_zone(const struct nlattr *attr,
1081                      struct nf_conntrack_zone *zone)
1082 {
1083         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1084                         NF_CT_DEFAULT_ZONE_DIR, 0);
1085 #ifdef CONFIG_NF_CONNTRACK_ZONES
1086         if (attr)
1087                 zone->id = ntohs(nla_get_be16(attr));
1088 #else
1089         if (attr)
1090                 return -EOPNOTSUPP;
1091 #endif
1092         return 0;
1093 }
1094
1095 static int
1096 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1097                            struct nf_conntrack_zone *zone)
1098 {
1099         int ret;
1100
1101         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1102                 return -EINVAL;
1103
1104         ret = ctnetlink_parse_zone(attr, zone);
1105         if (ret < 0)
1106                 return ret;
1107
1108         if (type == CTA_TUPLE_REPLY)
1109                 zone->dir = NF_CT_ZONE_DIR_REPL;
1110         else
1111                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1112
1113         return 0;
1114 }
1115
1116 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1117         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1118         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1119         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1120 };
1121
1122 static int
1123 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1124                       struct nf_conntrack_tuple *tuple, u32 type,
1125                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1126 {
1127         struct nlattr *tb[CTA_TUPLE_MAX+1];
1128         int err;
1129
1130         memset(tuple, 0, sizeof(*tuple));
1131
1132         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1133                                NULL);
1134         if (err < 0)
1135                 return err;
1136
1137         if (!tb[CTA_TUPLE_IP])
1138                 return -EINVAL;
1139
1140         tuple->src.l3num = l3num;
1141
1142         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1143         if (err < 0)
1144                 return err;
1145
1146         if (!tb[CTA_TUPLE_PROTO])
1147                 return -EINVAL;
1148
1149         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1150         if (err < 0)
1151                 return err;
1152
1153         if (tb[CTA_TUPLE_ZONE]) {
1154                 if (!zone)
1155                         return -EINVAL;
1156
1157                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1158                                                  type, zone);
1159                 if (err < 0)
1160                         return err;
1161         }
1162
1163         /* orig and expect tuples get DIR_ORIGINAL */
1164         if (type == CTA_TUPLE_REPLY)
1165                 tuple->dst.dir = IP_CT_DIR_REPLY;
1166         else
1167                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1168
1169         return 0;
1170 }
1171
1172 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1173         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1174                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1175 };
1176
1177 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1178                                 struct nlattr **helpinfo)
1179 {
1180         int err;
1181         struct nlattr *tb[CTA_HELP_MAX+1];
1182
1183         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1184         if (err < 0)
1185                 return err;
1186
1187         if (!tb[CTA_HELP_NAME])
1188                 return -EINVAL;
1189
1190         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1191
1192         if (tb[CTA_HELP_INFO])
1193                 *helpinfo = tb[CTA_HELP_INFO];
1194
1195         return 0;
1196 }
1197
1198 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1199         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1200         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1201         [CTA_STATUS]            = { .type = NLA_U32 },
1202         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1203         [CTA_HELP]              = { .type = NLA_NESTED },
1204         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1205         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1206         [CTA_MARK]              = { .type = NLA_U32 },
1207         [CTA_ID]                = { .type = NLA_U32 },
1208         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1209         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1210         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1211         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1212         [CTA_ZONE]              = { .type = NLA_U16 },
1213         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1214         [CTA_LABELS]            = { .type = NLA_BINARY,
1215                                     .len = NF_CT_LABELS_MAX_SIZE },
1216         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1217                                     .len = NF_CT_LABELS_MAX_SIZE },
1218 };
1219
1220 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1221 {
1222         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1223                 return 0;
1224
1225         return ctnetlink_filter_match(ct, data);
1226 }
1227
1228 static int ctnetlink_flush_conntrack(struct net *net,
1229                                      const struct nlattr * const cda[],
1230                                      u32 portid, int report, u8 family)
1231 {
1232         struct ctnetlink_filter *filter = NULL;
1233
1234         if (family || (cda[CTA_MARK] && cda[CTA_MARK_MASK])) {
1235                 filter = ctnetlink_alloc_filter(cda, family);
1236                 if (IS_ERR(filter))
1237                         return PTR_ERR(filter);
1238         }
1239
1240         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1241                                   portid, report);
1242         kfree(filter);
1243
1244         return 0;
1245 }
1246
1247 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1248                                    struct sk_buff *skb,
1249                                    const struct nlmsghdr *nlh,
1250                                    const struct nlattr * const cda[],
1251                                    struct netlink_ext_ack *extack)
1252 {
1253         struct nf_conntrack_tuple_hash *h;
1254         struct nf_conntrack_tuple tuple;
1255         struct nf_conn *ct;
1256         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1257         u_int8_t u3 = nfmsg->nfgen_family;
1258         struct nf_conntrack_zone zone;
1259         int err;
1260
1261         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1262         if (err < 0)
1263                 return err;
1264
1265         if (cda[CTA_TUPLE_ORIG])
1266                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1267                                             u3, &zone);
1268         else if (cda[CTA_TUPLE_REPLY])
1269                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1270                                             u3, &zone);
1271         else {
1272                 return ctnetlink_flush_conntrack(net, cda,
1273                                                  NETLINK_CB(skb).portid,
1274                                                  nlmsg_report(nlh), u3);
1275         }
1276
1277         if (err < 0)
1278                 return err;
1279
1280         h = nf_conntrack_find_get(net, &zone, &tuple);
1281         if (!h)
1282                 return -ENOENT;
1283
1284         ct = nf_ct_tuplehash_to_ctrack(h);
1285
1286         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1287                 nf_ct_put(ct);
1288                 return -EBUSY;
1289         }
1290
1291         if (cda[CTA_ID]) {
1292                 __be32 id = nla_get_be32(cda[CTA_ID]);
1293
1294                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1295                         nf_ct_put(ct);
1296                         return -ENOENT;
1297                 }
1298         }
1299
1300         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1301         nf_ct_put(ct);
1302
1303         return 0;
1304 }
1305
1306 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1307                                    struct sk_buff *skb,
1308                                    const struct nlmsghdr *nlh,
1309                                    const struct nlattr * const cda[],
1310                                    struct netlink_ext_ack *extack)
1311 {
1312         struct nf_conntrack_tuple_hash *h;
1313         struct nf_conntrack_tuple tuple;
1314         struct nf_conn *ct;
1315         struct sk_buff *skb2 = NULL;
1316         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1317         u_int8_t u3 = nfmsg->nfgen_family;
1318         struct nf_conntrack_zone zone;
1319         int err;
1320
1321         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1322                 struct netlink_dump_control c = {
1323                         .start = ctnetlink_start,
1324                         .dump = ctnetlink_dump_table,
1325                         .done = ctnetlink_done,
1326                         .data = (void *)cda,
1327                 };
1328
1329                 return netlink_dump_start(ctnl, skb, nlh, &c);
1330         }
1331
1332         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1333         if (err < 0)
1334                 return err;
1335
1336         if (cda[CTA_TUPLE_ORIG])
1337                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1338                                             u3, &zone);
1339         else if (cda[CTA_TUPLE_REPLY])
1340                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1341                                             u3, &zone);
1342         else
1343                 return -EINVAL;
1344
1345         if (err < 0)
1346                 return err;
1347
1348         h = nf_conntrack_find_get(net, &zone, &tuple);
1349         if (!h)
1350                 return -ENOENT;
1351
1352         ct = nf_ct_tuplehash_to_ctrack(h);
1353
1354         err = -ENOMEM;
1355         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1356         if (skb2 == NULL) {
1357                 nf_ct_put(ct);
1358                 return -ENOMEM;
1359         }
1360
1361         rcu_read_lock();
1362         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1363                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1364         rcu_read_unlock();
1365         nf_ct_put(ct);
1366         if (err <= 0)
1367                 goto free;
1368
1369         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1370         if (err < 0)
1371                 goto out;
1372
1373         return 0;
1374
1375 free:
1376         kfree_skb(skb2);
1377 out:
1378         /* this avoids a loop in nfnetlink. */
1379         return err == -EAGAIN ? -ENOBUFS : err;
1380 }
1381
1382 static int ctnetlink_done_list(struct netlink_callback *cb)
1383 {
1384         if (cb->args[1])
1385                 nf_ct_put((struct nf_conn *)cb->args[1]);
1386         return 0;
1387 }
1388
1389 static int
1390 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1391 {
1392         struct nf_conn *ct, *last;
1393         struct nf_conntrack_tuple_hash *h;
1394         struct hlist_nulls_node *n;
1395         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1396         u_int8_t l3proto = nfmsg->nfgen_family;
1397         int res;
1398         int cpu;
1399         struct hlist_nulls_head *list;
1400         struct net *net = sock_net(skb->sk);
1401
1402         if (cb->args[2])
1403                 return 0;
1404
1405         last = (struct nf_conn *)cb->args[1];
1406
1407         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1408                 struct ct_pcpu *pcpu;
1409
1410                 if (!cpu_possible(cpu))
1411                         continue;
1412
1413                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1414                 spin_lock_bh(&pcpu->lock);
1415                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1416 restart:
1417                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1418                         ct = nf_ct_tuplehash_to_ctrack(h);
1419                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1420                                 continue;
1421                         if (cb->args[1]) {
1422                                 if (ct != last)
1423                                         continue;
1424                                 cb->args[1] = 0;
1425                         }
1426                         rcu_read_lock();
1427                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1428                                                   cb->nlh->nlmsg_seq,
1429                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1430                                                   ct);
1431                         rcu_read_unlock();
1432                         if (res < 0) {
1433                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1434                                         continue;
1435                                 cb->args[0] = cpu;
1436                                 cb->args[1] = (unsigned long)ct;
1437                                 spin_unlock_bh(&pcpu->lock);
1438                                 goto out;
1439                         }
1440                 }
1441                 if (cb->args[1]) {
1442                         cb->args[1] = 0;
1443                         goto restart;
1444                 }
1445                 spin_unlock_bh(&pcpu->lock);
1446         }
1447         cb->args[2] = 1;
1448 out:
1449         if (last)
1450                 nf_ct_put(last);
1451
1452         return skb->len;
1453 }
1454
1455 static int
1456 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1457 {
1458         return ctnetlink_dump_list(skb, cb, true);
1459 }
1460
1461 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1462                                   struct sk_buff *skb,
1463                                   const struct nlmsghdr *nlh,
1464                                   const struct nlattr * const cda[],
1465                                   struct netlink_ext_ack *extack)
1466 {
1467         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1468                 struct netlink_dump_control c = {
1469                         .dump = ctnetlink_dump_dying,
1470                         .done = ctnetlink_done_list,
1471                 };
1472                 return netlink_dump_start(ctnl, skb, nlh, &c);
1473         }
1474
1475         return -EOPNOTSUPP;
1476 }
1477
1478 static int
1479 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1480 {
1481         return ctnetlink_dump_list(skb, cb, false);
1482 }
1483
1484 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1485                                         struct sk_buff *skb,
1486                                         const struct nlmsghdr *nlh,
1487                                         const struct nlattr * const cda[],
1488                                         struct netlink_ext_ack *extack)
1489 {
1490         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1491                 struct netlink_dump_control c = {
1492                         .dump = ctnetlink_dump_unconfirmed,
1493                         .done = ctnetlink_done_list,
1494                 };
1495                 return netlink_dump_start(ctnl, skb, nlh, &c);
1496         }
1497
1498         return -EOPNOTSUPP;
1499 }
1500
1501 #if IS_ENABLED(CONFIG_NF_NAT)
1502 static int
1503 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1504                           enum nf_nat_manip_type manip,
1505                           const struct nlattr *attr)
1506 {
1507         struct nf_nat_hook *nat_hook;
1508         int err;
1509
1510         nat_hook = rcu_dereference(nf_nat_hook);
1511         if (!nat_hook) {
1512 #ifdef CONFIG_MODULES
1513                 rcu_read_unlock();
1514                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1515                 if (request_module("nf-nat") < 0) {
1516                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1517                         rcu_read_lock();
1518                         return -EOPNOTSUPP;
1519                 }
1520                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1521                 rcu_read_lock();
1522                 nat_hook = rcu_dereference(nf_nat_hook);
1523                 if (nat_hook)
1524                         return -EAGAIN;
1525 #endif
1526                 return -EOPNOTSUPP;
1527         }
1528
1529         err = nat_hook->parse_nat_setup(ct, manip, attr);
1530         if (err == -EAGAIN) {
1531 #ifdef CONFIG_MODULES
1532                 rcu_read_unlock();
1533                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1534                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1535                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1536                         rcu_read_lock();
1537                         return -EOPNOTSUPP;
1538                 }
1539                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1540                 rcu_read_lock();
1541 #else
1542                 err = -EOPNOTSUPP;
1543 #endif
1544         }
1545         return err;
1546 }
1547 #endif
1548
1549 static void
1550 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1551                           unsigned long off)
1552 {
1553         unsigned int bit;
1554
1555         /* Ignore these unchangable bits */
1556         on &= ~IPS_UNCHANGEABLE_MASK;
1557         off &= ~IPS_UNCHANGEABLE_MASK;
1558
1559         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1560                 if (on & (1 << bit))
1561                         set_bit(bit, &ct->status);
1562                 else if (off & (1 << bit))
1563                         clear_bit(bit, &ct->status);
1564         }
1565 }
1566
1567 static int
1568 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1569 {
1570         unsigned long d;
1571         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1572         d = ct->status ^ status;
1573
1574         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1575                 /* unchangeable */
1576                 return -EBUSY;
1577
1578         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1579                 /* SEEN_REPLY bit can only be set */
1580                 return -EBUSY;
1581
1582         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1583                 /* ASSURED bit can only be set */
1584                 return -EBUSY;
1585
1586         __ctnetlink_change_status(ct, status, 0);
1587         return 0;
1588 }
1589
1590 static int
1591 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1592 {
1593 #if IS_ENABLED(CONFIG_NF_NAT)
1594         int ret;
1595
1596         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1597                 return 0;
1598
1599         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1600                                         cda[CTA_NAT_DST]);
1601         if (ret < 0)
1602                 return ret;
1603
1604         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1605                                          cda[CTA_NAT_SRC]);
1606 #else
1607         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1608                 return 0;
1609         return -EOPNOTSUPP;
1610 #endif
1611 }
1612
1613 static int ctnetlink_change_helper(struct nf_conn *ct,
1614                                    const struct nlattr * const cda[])
1615 {
1616         struct nf_conntrack_helper *helper;
1617         struct nf_conn_help *help = nfct_help(ct);
1618         char *helpname = NULL;
1619         struct nlattr *helpinfo = NULL;
1620         int err;
1621
1622         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1623         if (err < 0)
1624                 return err;
1625
1626         /* don't change helper of sibling connections */
1627         if (ct->master) {
1628                 /* If we try to change the helper to the same thing twice,
1629                  * treat the second attempt as a no-op instead of returning
1630                  * an error.
1631                  */
1632                 err = -EBUSY;
1633                 if (help) {
1634                         rcu_read_lock();
1635                         helper = rcu_dereference(help->helper);
1636                         if (helper && !strcmp(helper->name, helpname))
1637                                 err = 0;
1638                         rcu_read_unlock();
1639                 }
1640
1641                 return err;
1642         }
1643
1644         if (!strcmp(helpname, "")) {
1645                 if (help && help->helper) {
1646                         /* we had a helper before ... */
1647                         nf_ct_remove_expectations(ct);
1648                         RCU_INIT_POINTER(help->helper, NULL);
1649                 }
1650
1651                 return 0;
1652         }
1653
1654         rcu_read_lock();
1655         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1656                                             nf_ct_protonum(ct));
1657         if (helper == NULL) {
1658                 rcu_read_unlock();
1659                 return -EOPNOTSUPP;
1660         }
1661
1662         if (help) {
1663                 if (help->helper == helper) {
1664                         /* update private helper data if allowed. */
1665                         if (helper->from_nlattr)
1666                                 helper->from_nlattr(helpinfo, ct);
1667                         err = 0;
1668                 } else
1669                         err = -EBUSY;
1670         } else {
1671                 /* we cannot set a helper for an existing conntrack */
1672                 err = -EOPNOTSUPP;
1673         }
1674
1675         rcu_read_unlock();
1676         return err;
1677 }
1678
1679 static int ctnetlink_change_timeout(struct nf_conn *ct,
1680                                     const struct nlattr * const cda[])
1681 {
1682         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1683
1684         if (timeout > INT_MAX)
1685                 timeout = INT_MAX;
1686         ct->timeout = nfct_time_stamp + (u32)timeout;
1687
1688         if (test_bit(IPS_DYING_BIT, &ct->status))
1689                 return -ETIME;
1690
1691         return 0;
1692 }
1693
1694 #if defined(CONFIG_NF_CONNTRACK_MARK)
1695 static void ctnetlink_change_mark(struct nf_conn *ct,
1696                                     const struct nlattr * const cda[])
1697 {
1698         u32 mark, newmark, mask = 0;
1699
1700         if (cda[CTA_MARK_MASK])
1701                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1702
1703         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1704         newmark = (ct->mark & mask) ^ mark;
1705         if (newmark != ct->mark)
1706                 ct->mark = newmark;
1707 }
1708 #endif
1709
1710 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1711         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1712         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1713         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1714 };
1715
1716 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1717                                       const struct nlattr * const cda[])
1718 {
1719         const struct nlattr *attr = cda[CTA_PROTOINFO];
1720         const struct nf_conntrack_l4proto *l4proto;
1721         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1722         int err = 0;
1723
1724         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1725                                NULL);
1726         if (err < 0)
1727                 return err;
1728
1729         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1730         if (l4proto->from_nlattr)
1731                 err = l4proto->from_nlattr(tb, ct);
1732
1733         return err;
1734 }
1735
1736 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1737         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1738         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1739         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1740 };
1741
1742 static int change_seq_adj(struct nf_ct_seqadj *seq,
1743                           const struct nlattr * const attr)
1744 {
1745         int err;
1746         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1747
1748         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1749         if (err < 0)
1750                 return err;
1751
1752         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1753                 return -EINVAL;
1754
1755         seq->correction_pos =
1756                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1757
1758         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1759                 return -EINVAL;
1760
1761         seq->offset_before =
1762                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1763
1764         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1765                 return -EINVAL;
1766
1767         seq->offset_after =
1768                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1769
1770         return 0;
1771 }
1772
1773 static int
1774 ctnetlink_change_seq_adj(struct nf_conn *ct,
1775                          const struct nlattr * const cda[])
1776 {
1777         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1778         int ret = 0;
1779
1780         if (!seqadj)
1781                 return 0;
1782
1783         spin_lock_bh(&ct->lock);
1784         if (cda[CTA_SEQ_ADJ_ORIG]) {
1785                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1786                                      cda[CTA_SEQ_ADJ_ORIG]);
1787                 if (ret < 0)
1788                         goto err;
1789
1790                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1791         }
1792
1793         if (cda[CTA_SEQ_ADJ_REPLY]) {
1794                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1795                                      cda[CTA_SEQ_ADJ_REPLY]);
1796                 if (ret < 0)
1797                         goto err;
1798
1799                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1800         }
1801
1802         spin_unlock_bh(&ct->lock);
1803         return 0;
1804 err:
1805         spin_unlock_bh(&ct->lock);
1806         return ret;
1807 }
1808
1809 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
1810         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
1811         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
1812         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
1813 };
1814
1815 static int ctnetlink_change_synproxy(struct nf_conn *ct,
1816                                      const struct nlattr * const cda[])
1817 {
1818         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
1819         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
1820         int err;
1821
1822         if (!synproxy)
1823                 return 0;
1824
1825         err = nla_parse_nested(tb, CTA_SYNPROXY_MAX, cda[CTA_SYNPROXY],
1826                                synproxy_policy, NULL);
1827         if (err < 0)
1828                 return err;
1829
1830         if (!tb[CTA_SYNPROXY_ISN] ||
1831             !tb[CTA_SYNPROXY_ITS] ||
1832             !tb[CTA_SYNPROXY_TSOFF])
1833                 return -EINVAL;
1834
1835         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
1836         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
1837         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
1838
1839         return 0;
1840 }
1841
1842 static int
1843 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1844 {
1845 #ifdef CONFIG_NF_CONNTRACK_LABELS
1846         size_t len = nla_len(cda[CTA_LABELS]);
1847         const void *mask = cda[CTA_LABELS_MASK];
1848
1849         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1850                 return -EINVAL;
1851
1852         if (mask) {
1853                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1854                     nla_len(cda[CTA_LABELS_MASK]) != len)
1855                         return -EINVAL;
1856                 mask = nla_data(cda[CTA_LABELS_MASK]);
1857         }
1858
1859         len /= sizeof(u32);
1860
1861         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1862 #else
1863         return -EOPNOTSUPP;
1864 #endif
1865 }
1866
1867 static int
1868 ctnetlink_change_conntrack(struct nf_conn *ct,
1869                            const struct nlattr * const cda[])
1870 {
1871         int err;
1872
1873         /* only allow NAT changes and master assignation for new conntracks */
1874         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1875                 return -EOPNOTSUPP;
1876
1877         if (cda[CTA_HELP]) {
1878                 err = ctnetlink_change_helper(ct, cda);
1879                 if (err < 0)
1880                         return err;
1881         }
1882
1883         if (cda[CTA_TIMEOUT]) {
1884                 err = ctnetlink_change_timeout(ct, cda);
1885                 if (err < 0)
1886                         return err;
1887         }
1888
1889         if (cda[CTA_STATUS]) {
1890                 err = ctnetlink_change_status(ct, cda);
1891                 if (err < 0)
1892                         return err;
1893         }
1894
1895         if (cda[CTA_PROTOINFO]) {
1896                 err = ctnetlink_change_protoinfo(ct, cda);
1897                 if (err < 0)
1898                         return err;
1899         }
1900
1901 #if defined(CONFIG_NF_CONNTRACK_MARK)
1902         if (cda[CTA_MARK])
1903                 ctnetlink_change_mark(ct, cda);
1904 #endif
1905
1906         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1907                 err = ctnetlink_change_seq_adj(ct, cda);
1908                 if (err < 0)
1909                         return err;
1910         }
1911
1912         if (cda[CTA_SYNPROXY]) {
1913                 err = ctnetlink_change_synproxy(ct, cda);
1914                 if (err < 0)
1915                         return err;
1916         }
1917
1918         if (cda[CTA_LABELS]) {
1919                 err = ctnetlink_attach_labels(ct, cda);
1920                 if (err < 0)
1921                         return err;
1922         }
1923
1924         return 0;
1925 }
1926
1927 static struct nf_conn *
1928 ctnetlink_create_conntrack(struct net *net,
1929                            const struct nf_conntrack_zone *zone,
1930                            const struct nlattr * const cda[],
1931                            struct nf_conntrack_tuple *otuple,
1932                            struct nf_conntrack_tuple *rtuple,
1933                            u8 u3)
1934 {
1935         struct nf_conn *ct;
1936         int err = -EINVAL;
1937         struct nf_conntrack_helper *helper;
1938         struct nf_conn_tstamp *tstamp;
1939         u64 timeout;
1940
1941         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1942         if (IS_ERR(ct))
1943                 return ERR_PTR(-ENOMEM);
1944
1945         if (!cda[CTA_TIMEOUT])
1946                 goto err1;
1947
1948         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1949         if (timeout > INT_MAX)
1950                 timeout = INT_MAX;
1951         ct->timeout = (u32)timeout + nfct_time_stamp;
1952
1953         rcu_read_lock();
1954         if (cda[CTA_HELP]) {
1955                 char *helpname = NULL;
1956                 struct nlattr *helpinfo = NULL;
1957
1958                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1959                 if (err < 0)
1960                         goto err2;
1961
1962                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1963                                                     nf_ct_protonum(ct));
1964                 if (helper == NULL) {
1965                         rcu_read_unlock();
1966 #ifdef CONFIG_MODULES
1967                         if (request_module("nfct-helper-%s", helpname) < 0) {
1968                                 err = -EOPNOTSUPP;
1969                                 goto err1;
1970                         }
1971
1972                         rcu_read_lock();
1973                         helper = __nf_conntrack_helper_find(helpname,
1974                                                             nf_ct_l3num(ct),
1975                                                             nf_ct_protonum(ct));
1976                         if (helper) {
1977                                 err = -EAGAIN;
1978                                 goto err2;
1979                         }
1980                         rcu_read_unlock();
1981 #endif
1982                         err = -EOPNOTSUPP;
1983                         goto err1;
1984                 } else {
1985                         struct nf_conn_help *help;
1986
1987                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1988                         if (help == NULL) {
1989                                 err = -ENOMEM;
1990                                 goto err2;
1991                         }
1992                         /* set private helper data if allowed. */
1993                         if (helper->from_nlattr)
1994                                 helper->from_nlattr(helpinfo, ct);
1995
1996                         /* not in hash table yet so not strictly necessary */
1997                         RCU_INIT_POINTER(help->helper, helper);
1998                 }
1999         } else {
2000                 /* try an implicit helper assignation */
2001                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2002                 if (err < 0)
2003                         goto err2;
2004         }
2005
2006         err = ctnetlink_setup_nat(ct, cda);
2007         if (err < 0)
2008                 goto err2;
2009
2010         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2011         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2012         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2013         nf_ct_labels_ext_add(ct);
2014         nfct_seqadj_ext_add(ct);
2015         nfct_synproxy_ext_add(ct);
2016
2017         /* we must add conntrack extensions before confirmation. */
2018         ct->status |= IPS_CONFIRMED;
2019
2020         if (cda[CTA_STATUS]) {
2021                 err = ctnetlink_change_status(ct, cda);
2022                 if (err < 0)
2023                         goto err2;
2024         }
2025
2026         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2027                 err = ctnetlink_change_seq_adj(ct, cda);
2028                 if (err < 0)
2029                         goto err2;
2030         }
2031
2032         memset(&ct->proto, 0, sizeof(ct->proto));
2033         if (cda[CTA_PROTOINFO]) {
2034                 err = ctnetlink_change_protoinfo(ct, cda);
2035                 if (err < 0)
2036                         goto err2;
2037         }
2038
2039         if (cda[CTA_SYNPROXY]) {
2040                 err = ctnetlink_change_synproxy(ct, cda);
2041                 if (err < 0)
2042                         goto err2;
2043         }
2044
2045 #if defined(CONFIG_NF_CONNTRACK_MARK)
2046         if (cda[CTA_MARK])
2047                 ctnetlink_change_mark(ct, cda);
2048 #endif
2049
2050         /* setup master conntrack: this is a confirmed expectation */
2051         if (cda[CTA_TUPLE_MASTER]) {
2052                 struct nf_conntrack_tuple master;
2053                 struct nf_conntrack_tuple_hash *master_h;
2054                 struct nf_conn *master_ct;
2055
2056                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2057                                             u3, NULL);
2058                 if (err < 0)
2059                         goto err2;
2060
2061                 master_h = nf_conntrack_find_get(net, zone, &master);
2062                 if (master_h == NULL) {
2063                         err = -ENOENT;
2064                         goto err2;
2065                 }
2066                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2067                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2068                 ct->master = master_ct;
2069         }
2070         tstamp = nf_conn_tstamp_find(ct);
2071         if (tstamp)
2072                 tstamp->start = ktime_get_real_ns();
2073
2074         err = nf_conntrack_hash_check_insert(ct);
2075         if (err < 0)
2076                 goto err2;
2077
2078         rcu_read_unlock();
2079
2080         return ct;
2081
2082 err2:
2083         rcu_read_unlock();
2084 err1:
2085         nf_conntrack_free(ct);
2086         return ERR_PTR(err);
2087 }
2088
2089 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2090                                    struct sk_buff *skb,
2091                                    const struct nlmsghdr *nlh,
2092                                    const struct nlattr * const cda[],
2093                                    struct netlink_ext_ack *extack)
2094 {
2095         struct nf_conntrack_tuple otuple, rtuple;
2096         struct nf_conntrack_tuple_hash *h = NULL;
2097         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2098         struct nf_conn *ct;
2099         u_int8_t u3 = nfmsg->nfgen_family;
2100         struct nf_conntrack_zone zone;
2101         int err;
2102
2103         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2104         if (err < 0)
2105                 return err;
2106
2107         if (cda[CTA_TUPLE_ORIG]) {
2108                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2109                                             u3, &zone);
2110                 if (err < 0)
2111                         return err;
2112         }
2113
2114         if (cda[CTA_TUPLE_REPLY]) {
2115                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2116                                             u3, &zone);
2117                 if (err < 0)
2118                         return err;
2119         }
2120
2121         if (cda[CTA_TUPLE_ORIG])
2122                 h = nf_conntrack_find_get(net, &zone, &otuple);
2123         else if (cda[CTA_TUPLE_REPLY])
2124                 h = nf_conntrack_find_get(net, &zone, &rtuple);
2125
2126         if (h == NULL) {
2127                 err = -ENOENT;
2128                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2129                         enum ip_conntrack_events events;
2130
2131                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2132                                 return -EINVAL;
2133                         if (otuple.dst.protonum != rtuple.dst.protonum)
2134                                 return -EINVAL;
2135
2136                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2137                                                         &rtuple, u3);
2138                         if (IS_ERR(ct))
2139                                 return PTR_ERR(ct);
2140
2141                         err = 0;
2142                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2143                                 events = 1 << IPCT_RELATED;
2144                         else
2145                                 events = 1 << IPCT_NEW;
2146
2147                         if (cda[CTA_LABELS] &&
2148                             ctnetlink_attach_labels(ct, cda) == 0)
2149                                 events |= (1 << IPCT_LABEL);
2150
2151                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2152                                                       (1 << IPCT_ASSURED) |
2153                                                       (1 << IPCT_HELPER) |
2154                                                       (1 << IPCT_PROTOINFO) |
2155                                                       (1 << IPCT_SEQADJ) |
2156                                                       (1 << IPCT_MARK) |
2157                                                       (1 << IPCT_SYNPROXY) |
2158                                                       events,
2159                                                       ct, NETLINK_CB(skb).portid,
2160                                                       nlmsg_report(nlh));
2161                         nf_ct_put(ct);
2162                 }
2163
2164                 return err;
2165         }
2166         /* implicit 'else' */
2167
2168         err = -EEXIST;
2169         ct = nf_ct_tuplehash_to_ctrack(h);
2170         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2171                 err = ctnetlink_change_conntrack(ct, cda);
2172                 if (err == 0) {
2173                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2174                                                       (1 << IPCT_ASSURED) |
2175                                                       (1 << IPCT_HELPER) |
2176                                                       (1 << IPCT_LABEL) |
2177                                                       (1 << IPCT_PROTOINFO) |
2178                                                       (1 << IPCT_SEQADJ) |
2179                                                       (1 << IPCT_MARK) |
2180                                                       (1 << IPCT_SYNPROXY),
2181                                                       ct, NETLINK_CB(skb).portid,
2182                                                       nlmsg_report(nlh));
2183                 }
2184         }
2185
2186         nf_ct_put(ct);
2187         return err;
2188 }
2189
2190 static int
2191 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2192                                 __u16 cpu, const struct ip_conntrack_stat *st)
2193 {
2194         struct nlmsghdr *nlh;
2195         struct nfgenmsg *nfmsg;
2196         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2197
2198         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2199                               IPCTNL_MSG_CT_GET_STATS_CPU);
2200         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2201         if (nlh == NULL)
2202                 goto nlmsg_failure;
2203
2204         nfmsg = nlmsg_data(nlh);
2205         nfmsg->nfgen_family = AF_UNSPEC;
2206         nfmsg->version      = NFNETLINK_V0;
2207         nfmsg->res_id       = htons(cpu);
2208
2209         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2210             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2211             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2212             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2213             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2214                                 htonl(st->insert_failed)) ||
2215             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2216             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2217             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2218             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2219                                 htonl(st->search_restart)))
2220                 goto nla_put_failure;
2221
2222         nlmsg_end(skb, nlh);
2223         return skb->len;
2224
2225 nla_put_failure:
2226 nlmsg_failure:
2227         nlmsg_cancel(skb, nlh);
2228         return -1;
2229 }
2230
2231 static int
2232 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2233 {
2234         int cpu;
2235         struct net *net = sock_net(skb->sk);
2236
2237         if (cb->args[0] == nr_cpu_ids)
2238                 return 0;
2239
2240         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2241                 const struct ip_conntrack_stat *st;
2242
2243                 if (!cpu_possible(cpu))
2244                         continue;
2245
2246                 st = per_cpu_ptr(net->ct.stat, cpu);
2247                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2248                                                     NETLINK_CB(cb->skb).portid,
2249                                                     cb->nlh->nlmsg_seq,
2250                                                     cpu, st) < 0)
2251                                 break;
2252         }
2253         cb->args[0] = cpu;
2254
2255         return skb->len;
2256 }
2257
2258 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2259                                  struct sk_buff *skb,
2260                                  const struct nlmsghdr *nlh,
2261                                  const struct nlattr * const cda[],
2262                                  struct netlink_ext_ack *extack)
2263 {
2264         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2265                 struct netlink_dump_control c = {
2266                         .dump = ctnetlink_ct_stat_cpu_dump,
2267                 };
2268                 return netlink_dump_start(ctnl, skb, nlh, &c);
2269         }
2270
2271         return 0;
2272 }
2273
2274 static int
2275 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2276                             struct net *net)
2277 {
2278         struct nlmsghdr *nlh;
2279         struct nfgenmsg *nfmsg;
2280         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2281         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2282
2283         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2284         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2285         if (nlh == NULL)
2286                 goto nlmsg_failure;
2287
2288         nfmsg = nlmsg_data(nlh);
2289         nfmsg->nfgen_family = AF_UNSPEC;
2290         nfmsg->version      = NFNETLINK_V0;
2291         nfmsg->res_id       = 0;
2292
2293         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2294                 goto nla_put_failure;
2295
2296         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2297                 goto nla_put_failure;
2298
2299         nlmsg_end(skb, nlh);
2300         return skb->len;
2301
2302 nla_put_failure:
2303 nlmsg_failure:
2304         nlmsg_cancel(skb, nlh);
2305         return -1;
2306 }
2307
2308 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2309                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2310                              const struct nlattr * const cda[],
2311                              struct netlink_ext_ack *extack)
2312 {
2313         struct sk_buff *skb2;
2314         int err;
2315
2316         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2317         if (skb2 == NULL)
2318                 return -ENOMEM;
2319
2320         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2321                                           nlh->nlmsg_seq,
2322                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2323                                           sock_net(skb->sk));
2324         if (err <= 0)
2325                 goto free;
2326
2327         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2328         if (err < 0)
2329                 goto out;
2330
2331         return 0;
2332
2333 free:
2334         kfree_skb(skb2);
2335 out:
2336         /* this avoids a loop in nfnetlink. */
2337         return err == -EAGAIN ? -ENOBUFS : err;
2338 }
2339
2340 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2341         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2342         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2343         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2344         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2345         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2346         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2347                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2348         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2349         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2350         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2351         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2352         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2353 };
2354
2355 static struct nf_conntrack_expect *
2356 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2357                        struct nf_conntrack_helper *helper,
2358                        struct nf_conntrack_tuple *tuple,
2359                        struct nf_conntrack_tuple *mask);
2360
2361 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2362 static size_t
2363 ctnetlink_glue_build_size(const struct nf_conn *ct)
2364 {
2365         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2366                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2367                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2368                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2369                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2370                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2371                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2372                + nla_total_size(0) /* CTA_PROTOINFO */
2373                + nla_total_size(0) /* CTA_HELP */
2374                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2375                + ctnetlink_secctx_size(ct)
2376 #if IS_ENABLED(CONFIG_NF_NAT)
2377                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2378                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2379 #endif
2380 #ifdef CONFIG_NF_CONNTRACK_MARK
2381                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2382 #endif
2383 #ifdef CONFIG_NF_CONNTRACK_ZONES
2384                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2385 #endif
2386                + ctnetlink_proto_size(ct)
2387                ;
2388 }
2389
2390 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2391                                              enum ip_conntrack_info *ctinfo)
2392 {
2393         return nf_ct_get(skb, ctinfo);
2394 }
2395
2396 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2397 {
2398         const struct nf_conntrack_zone *zone;
2399         struct nlattr *nest_parms;
2400
2401         zone = nf_ct_zone(ct);
2402
2403         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2404         if (!nest_parms)
2405                 goto nla_put_failure;
2406         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2407                 goto nla_put_failure;
2408         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2409                                    NF_CT_ZONE_DIR_ORIG) < 0)
2410                 goto nla_put_failure;
2411         nla_nest_end(skb, nest_parms);
2412
2413         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2414         if (!nest_parms)
2415                 goto nla_put_failure;
2416         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2417                 goto nla_put_failure;
2418         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2419                                    NF_CT_ZONE_DIR_REPL) < 0)
2420                 goto nla_put_failure;
2421         nla_nest_end(skb, nest_parms);
2422
2423         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2424                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2425                 goto nla_put_failure;
2426
2427         if (ctnetlink_dump_id(skb, ct) < 0)
2428                 goto nla_put_failure;
2429
2430         if (ctnetlink_dump_status(skb, ct) < 0)
2431                 goto nla_put_failure;
2432
2433         if (ctnetlink_dump_timeout(skb, ct) < 0)
2434                 goto nla_put_failure;
2435
2436         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2437                 goto nla_put_failure;
2438
2439         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2440                 goto nla_put_failure;
2441
2442 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2443         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2444                 goto nla_put_failure;
2445 #endif
2446         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2447                 goto nla_put_failure;
2448
2449         if ((ct->status & IPS_SEQ_ADJUST) &&
2450             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2451                 goto nla_put_failure;
2452
2453         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2454                 goto nla_put_failure;
2455
2456 #ifdef CONFIG_NF_CONNTRACK_MARK
2457         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2458                 goto nla_put_failure;
2459 #endif
2460         if (ctnetlink_dump_labels(skb, ct) < 0)
2461                 goto nla_put_failure;
2462         return 0;
2463
2464 nla_put_failure:
2465         return -ENOSPC;
2466 }
2467
2468 static int
2469 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2470                      enum ip_conntrack_info ctinfo,
2471                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2472 {
2473         struct nlattr *nest_parms;
2474
2475         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2476         if (!nest_parms)
2477                 goto nla_put_failure;
2478
2479         if (__ctnetlink_glue_build(skb, ct) < 0)
2480                 goto nla_put_failure;
2481
2482         nla_nest_end(skb, nest_parms);
2483
2484         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2485                 goto nla_put_failure;
2486
2487         return 0;
2488
2489 nla_put_failure:
2490         return -ENOSPC;
2491 }
2492
2493 static int
2494 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2495 {
2496         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2497         unsigned long d = ct->status ^ status;
2498
2499         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2500                 /* SEEN_REPLY bit can only be set */
2501                 return -EBUSY;
2502
2503         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2504                 /* ASSURED bit can only be set */
2505                 return -EBUSY;
2506
2507         /* This check is less strict than ctnetlink_change_status()
2508          * because callers often flip IPS_EXPECTED bits when sending
2509          * an NFQA_CT attribute to the kernel.  So ignore the
2510          * unchangeable bits but do not error out. Also user programs
2511          * are allowed to clear the bits that they are allowed to change.
2512          */
2513         __ctnetlink_change_status(ct, status, ~status);
2514         return 0;
2515 }
2516
2517 static int
2518 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2519 {
2520         int err;
2521
2522         if (cda[CTA_TIMEOUT]) {
2523                 err = ctnetlink_change_timeout(ct, cda);
2524                 if (err < 0)
2525                         return err;
2526         }
2527         if (cda[CTA_STATUS]) {
2528                 err = ctnetlink_update_status(ct, cda);
2529                 if (err < 0)
2530                         return err;
2531         }
2532         if (cda[CTA_HELP]) {
2533                 err = ctnetlink_change_helper(ct, cda);
2534                 if (err < 0)
2535                         return err;
2536         }
2537         if (cda[CTA_LABELS]) {
2538                 err = ctnetlink_attach_labels(ct, cda);
2539                 if (err < 0)
2540                         return err;
2541         }
2542 #if defined(CONFIG_NF_CONNTRACK_MARK)
2543         if (cda[CTA_MARK]) {
2544                 ctnetlink_change_mark(ct, cda);
2545         }
2546 #endif
2547         return 0;
2548 }
2549
2550 static int
2551 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2552 {
2553         struct nlattr *cda[CTA_MAX+1];
2554         int ret;
2555
2556         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2557         if (ret < 0)
2558                 return ret;
2559
2560         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2561 }
2562
2563 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2564                                     const struct nf_conn *ct,
2565                                     struct nf_conntrack_tuple *tuple,
2566                                     struct nf_conntrack_tuple *mask)
2567 {
2568         int err;
2569
2570         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2571                                     nf_ct_l3num(ct), NULL);
2572         if (err < 0)
2573                 return err;
2574
2575         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2576                                      nf_ct_l3num(ct), NULL);
2577 }
2578
2579 static int
2580 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2581                              u32 portid, u32 report)
2582 {
2583         struct nlattr *cda[CTA_EXPECT_MAX+1];
2584         struct nf_conntrack_tuple tuple, mask;
2585         struct nf_conntrack_helper *helper = NULL;
2586         struct nf_conntrack_expect *exp;
2587         int err;
2588
2589         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2590                                NULL);
2591         if (err < 0)
2592                 return err;
2593
2594         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2595                                        ct, &tuple, &mask);
2596         if (err < 0)
2597                 return err;
2598
2599         if (cda[CTA_EXPECT_HELP_NAME]) {
2600                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2601
2602                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2603                                                     nf_ct_protonum(ct));
2604                 if (helper == NULL)
2605                         return -EOPNOTSUPP;
2606         }
2607
2608         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2609                                      helper, &tuple, &mask);
2610         if (IS_ERR(exp))
2611                 return PTR_ERR(exp);
2612
2613         err = nf_ct_expect_related_report(exp, portid, report);
2614         nf_ct_expect_put(exp);
2615         return err;
2616 }
2617
2618 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2619                                   enum ip_conntrack_info ctinfo, int diff)
2620 {
2621         if (!(ct->status & IPS_NAT_MASK))
2622                 return;
2623
2624         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2625 }
2626
2627 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2628         .get_ct         = ctnetlink_glue_get_ct,
2629         .build_size     = ctnetlink_glue_build_size,
2630         .build          = ctnetlink_glue_build,
2631         .parse          = ctnetlink_glue_parse,
2632         .attach_expect  = ctnetlink_glue_attach_expect,
2633         .seq_adjust     = ctnetlink_glue_seqadj,
2634 };
2635 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2636
2637 /***********************************************************************
2638  * EXPECT
2639  ***********************************************************************/
2640
2641 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2642                                     const struct nf_conntrack_tuple *tuple,
2643                                     u32 type)
2644 {
2645         struct nlattr *nest_parms;
2646
2647         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2648         if (!nest_parms)
2649                 goto nla_put_failure;
2650         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2651                 goto nla_put_failure;
2652         nla_nest_end(skb, nest_parms);
2653
2654         return 0;
2655
2656 nla_put_failure:
2657         return -1;
2658 }
2659
2660 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2661                                    const struct nf_conntrack_tuple *tuple,
2662                                    const struct nf_conntrack_tuple_mask *mask)
2663 {
2664         const struct nf_conntrack_l4proto *l4proto;
2665         struct nf_conntrack_tuple m;
2666         struct nlattr *nest_parms;
2667         int ret;
2668
2669         memset(&m, 0xFF, sizeof(m));
2670         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2671         m.src.u.all = mask->src.u.all;
2672         m.dst.protonum = tuple->dst.protonum;
2673
2674         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2675         if (!nest_parms)
2676                 goto nla_put_failure;
2677
2678         rcu_read_lock();
2679         ret = ctnetlink_dump_tuples_ip(skb, &m);
2680         if (ret >= 0) {
2681                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2682                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2683         }
2684         rcu_read_unlock();
2685
2686         if (unlikely(ret < 0))
2687                 goto nla_put_failure;
2688
2689         nla_nest_end(skb, nest_parms);
2690
2691         return 0;
2692
2693 nla_put_failure:
2694         return -1;
2695 }
2696
2697 static const union nf_inet_addr any_addr;
2698
2699 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2700 {
2701         static __read_mostly siphash_key_t exp_id_seed;
2702         unsigned long a, b, c, d;
2703
2704         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2705
2706         a = (unsigned long)exp;
2707         b = (unsigned long)exp->helper;
2708         c = (unsigned long)exp->master;
2709         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2710
2711 #ifdef CONFIG_64BIT
2712         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2713 #else
2714         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2715 #endif
2716 }
2717
2718 static int
2719 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2720                           const struct nf_conntrack_expect *exp)
2721 {
2722         struct nf_conn *master = exp->master;
2723         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2724         struct nf_conn_help *help;
2725 #if IS_ENABLED(CONFIG_NF_NAT)
2726         struct nlattr *nest_parms;
2727         struct nf_conntrack_tuple nat_tuple = {};
2728 #endif
2729         struct nf_ct_helper_expectfn *expfn;
2730
2731         if (timeout < 0)
2732                 timeout = 0;
2733
2734         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2735                 goto nla_put_failure;
2736         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2737                 goto nla_put_failure;
2738         if (ctnetlink_exp_dump_tuple(skb,
2739                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2740                                  CTA_EXPECT_MASTER) < 0)
2741                 goto nla_put_failure;
2742
2743 #if IS_ENABLED(CONFIG_NF_NAT)
2744         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2745             exp->saved_proto.all) {
2746                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2747                 if (!nest_parms)
2748                         goto nla_put_failure;
2749
2750                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2751                         goto nla_put_failure;
2752
2753                 nat_tuple.src.l3num = nf_ct_l3num(master);
2754                 nat_tuple.src.u3 = exp->saved_addr;
2755                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2756                 nat_tuple.src.u = exp->saved_proto;
2757
2758                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2759                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2760                         goto nla_put_failure;
2761                 nla_nest_end(skb, nest_parms);
2762         }
2763 #endif
2764         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2765             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
2766             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2767             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2768                 goto nla_put_failure;
2769         help = nfct_help(master);
2770         if (help) {
2771                 struct nf_conntrack_helper *helper;
2772
2773                 helper = rcu_dereference(help->helper);
2774                 if (helper &&
2775                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2776                         goto nla_put_failure;
2777         }
2778         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2779         if (expfn != NULL &&
2780             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2781                 goto nla_put_failure;
2782
2783         return 0;
2784
2785 nla_put_failure:
2786         return -1;
2787 }
2788
2789 static int
2790 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2791                         int event, const struct nf_conntrack_expect *exp)
2792 {
2793         struct nlmsghdr *nlh;
2794         struct nfgenmsg *nfmsg;
2795         unsigned int flags = portid ? NLM_F_MULTI : 0;
2796
2797         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2798         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2799         if (nlh == NULL)
2800                 goto nlmsg_failure;
2801
2802         nfmsg = nlmsg_data(nlh);
2803         nfmsg->nfgen_family = exp->tuple.src.l3num;
2804         nfmsg->version      = NFNETLINK_V0;
2805         nfmsg->res_id       = 0;
2806
2807         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2808                 goto nla_put_failure;
2809
2810         nlmsg_end(skb, nlh);
2811         return skb->len;
2812
2813 nlmsg_failure:
2814 nla_put_failure:
2815         nlmsg_cancel(skb, nlh);
2816         return -1;
2817 }
2818
2819 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2820 static int
2821 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2822 {
2823         struct nf_conntrack_expect *exp = item->exp;
2824         struct net *net = nf_ct_exp_net(exp);
2825         struct nlmsghdr *nlh;
2826         struct nfgenmsg *nfmsg;
2827         struct sk_buff *skb;
2828         unsigned int type, group;
2829         int flags = 0;
2830
2831         if (events & (1 << IPEXP_DESTROY)) {
2832                 type = IPCTNL_MSG_EXP_DELETE;
2833                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2834         } else if (events & (1 << IPEXP_NEW)) {
2835                 type = IPCTNL_MSG_EXP_NEW;
2836                 flags = NLM_F_CREATE|NLM_F_EXCL;
2837                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2838         } else
2839                 return 0;
2840
2841         if (!item->report && !nfnetlink_has_listeners(net, group))
2842                 return 0;
2843
2844         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2845         if (skb == NULL)
2846                 goto errout;
2847
2848         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2849         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2850         if (nlh == NULL)
2851                 goto nlmsg_failure;
2852
2853         nfmsg = nlmsg_data(nlh);
2854         nfmsg->nfgen_family = exp->tuple.src.l3num;
2855         nfmsg->version      = NFNETLINK_V0;
2856         nfmsg->res_id       = 0;
2857
2858         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2859                 goto nla_put_failure;
2860
2861         nlmsg_end(skb, nlh);
2862         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2863         return 0;
2864
2865 nla_put_failure:
2866         nlmsg_cancel(skb, nlh);
2867 nlmsg_failure:
2868         kfree_skb(skb);
2869 errout:
2870         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2871         return 0;
2872 }
2873 #endif
2874 static int ctnetlink_exp_done(struct netlink_callback *cb)
2875 {
2876         if (cb->args[1])
2877                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2878         return 0;
2879 }
2880
2881 static int
2882 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2883 {
2884         struct net *net = sock_net(skb->sk);
2885         struct nf_conntrack_expect *exp, *last;
2886         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2887         u_int8_t l3proto = nfmsg->nfgen_family;
2888
2889         rcu_read_lock();
2890         last = (struct nf_conntrack_expect *)cb->args[1];
2891         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2892 restart:
2893                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2894                                          hnode) {
2895                         if (l3proto && exp->tuple.src.l3num != l3proto)
2896                                 continue;
2897
2898                         if (!net_eq(nf_ct_net(exp->master), net))
2899                                 continue;
2900
2901                         if (cb->args[1]) {
2902                                 if (exp != last)
2903                                         continue;
2904                                 cb->args[1] = 0;
2905                         }
2906                         if (ctnetlink_exp_fill_info(skb,
2907                                                     NETLINK_CB(cb->skb).portid,
2908                                                     cb->nlh->nlmsg_seq,
2909                                                     IPCTNL_MSG_EXP_NEW,
2910                                                     exp) < 0) {
2911                                 if (!refcount_inc_not_zero(&exp->use))
2912                                         continue;
2913                                 cb->args[1] = (unsigned long)exp;
2914                                 goto out;
2915                         }
2916                 }
2917                 if (cb->args[1]) {
2918                         cb->args[1] = 0;
2919                         goto restart;
2920                 }
2921         }
2922 out:
2923         rcu_read_unlock();
2924         if (last)
2925                 nf_ct_expect_put(last);
2926
2927         return skb->len;
2928 }
2929
2930 static int
2931 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2932 {
2933         struct nf_conntrack_expect *exp, *last;
2934         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2935         struct nf_conn *ct = cb->data;
2936         struct nf_conn_help *help = nfct_help(ct);
2937         u_int8_t l3proto = nfmsg->nfgen_family;
2938
2939         if (cb->args[0])
2940                 return 0;
2941
2942         rcu_read_lock();
2943         last = (struct nf_conntrack_expect *)cb->args[1];
2944 restart:
2945         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2946                 if (l3proto && exp->tuple.src.l3num != l3proto)
2947                         continue;
2948                 if (cb->args[1]) {
2949                         if (exp != last)
2950                                 continue;
2951                         cb->args[1] = 0;
2952                 }
2953                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2954                                             cb->nlh->nlmsg_seq,
2955                                             IPCTNL_MSG_EXP_NEW,
2956                                             exp) < 0) {
2957                         if (!refcount_inc_not_zero(&exp->use))
2958                                 continue;
2959                         cb->args[1] = (unsigned long)exp;
2960                         goto out;
2961                 }
2962         }
2963         if (cb->args[1]) {
2964                 cb->args[1] = 0;
2965                 goto restart;
2966         }
2967         cb->args[0] = 1;
2968 out:
2969         rcu_read_unlock();
2970         if (last)
2971                 nf_ct_expect_put(last);
2972
2973         return skb->len;
2974 }
2975
2976 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2977                                  struct sk_buff *skb,
2978                                  const struct nlmsghdr *nlh,
2979                                  const struct nlattr * const cda[],
2980                                  struct netlink_ext_ack *extack)
2981 {
2982         int err;
2983         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2984         u_int8_t u3 = nfmsg->nfgen_family;
2985         struct nf_conntrack_tuple tuple;
2986         struct nf_conntrack_tuple_hash *h;
2987         struct nf_conn *ct;
2988         struct nf_conntrack_zone zone;
2989         struct netlink_dump_control c = {
2990                 .dump = ctnetlink_exp_ct_dump_table,
2991                 .done = ctnetlink_exp_done,
2992         };
2993
2994         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2995                                     u3, NULL);
2996         if (err < 0)
2997                 return err;
2998
2999         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3000         if (err < 0)
3001                 return err;
3002
3003         h = nf_conntrack_find_get(net, &zone, &tuple);
3004         if (!h)
3005                 return -ENOENT;
3006
3007         ct = nf_ct_tuplehash_to_ctrack(h);
3008         /* No expectation linked to this connection tracking. */
3009         if (!nfct_help(ct)) {
3010                 nf_ct_put(ct);
3011                 return 0;
3012         }
3013
3014         c.data = ct;
3015
3016         err = netlink_dump_start(ctnl, skb, nlh, &c);
3017         nf_ct_put(ct);
3018
3019         return err;
3020 }
3021
3022 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
3023                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3024                                 const struct nlattr * const cda[],
3025                                 struct netlink_ext_ack *extack)
3026 {
3027         struct nf_conntrack_tuple tuple;
3028         struct nf_conntrack_expect *exp;
3029         struct sk_buff *skb2;
3030         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3031         u_int8_t u3 = nfmsg->nfgen_family;
3032         struct nf_conntrack_zone zone;
3033         int err;
3034
3035         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3036                 if (cda[CTA_EXPECT_MASTER])
3037                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
3038                                                      extack);
3039                 else {
3040                         struct netlink_dump_control c = {
3041                                 .dump = ctnetlink_exp_dump_table,
3042                                 .done = ctnetlink_exp_done,
3043                         };
3044                         return netlink_dump_start(ctnl, skb, nlh, &c);
3045                 }
3046         }
3047
3048         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3049         if (err < 0)
3050                 return err;
3051
3052         if (cda[CTA_EXPECT_TUPLE])
3053                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3054                                             u3, NULL);
3055         else if (cda[CTA_EXPECT_MASTER])
3056                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3057                                             u3, NULL);
3058         else
3059                 return -EINVAL;
3060
3061         if (err < 0)
3062                 return err;
3063
3064         exp = nf_ct_expect_find_get(net, &zone, &tuple);
3065         if (!exp)
3066                 return -ENOENT;
3067
3068         if (cda[CTA_EXPECT_ID]) {
3069                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3070
3071                 if (id != nf_expect_get_id(exp)) {
3072                         nf_ct_expect_put(exp);
3073                         return -ENOENT;
3074                 }
3075         }
3076
3077         err = -ENOMEM;
3078         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3079         if (skb2 == NULL) {
3080                 nf_ct_expect_put(exp);
3081                 goto out;
3082         }
3083
3084         rcu_read_lock();
3085         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3086                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
3087         rcu_read_unlock();
3088         nf_ct_expect_put(exp);
3089         if (err <= 0)
3090                 goto free;
3091
3092         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
3093         if (err < 0)
3094                 goto out;
3095
3096         return 0;
3097
3098 free:
3099         kfree_skb(skb2);
3100 out:
3101         /* this avoids a loop in nfnetlink. */
3102         return err == -EAGAIN ? -ENOBUFS : err;
3103 }
3104
3105 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3106 {
3107         const struct nf_conn_help *m_help;
3108         const char *name = data;
3109
3110         m_help = nfct_help(exp->master);
3111
3112         return strcmp(m_help->helper->name, name) == 0;
3113 }
3114
3115 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3116 {
3117         return true;
3118 }
3119
3120 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3121                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3122                                 const struct nlattr * const cda[],
3123                                 struct netlink_ext_ack *extack)
3124 {
3125         struct nf_conntrack_expect *exp;
3126         struct nf_conntrack_tuple tuple;
3127         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3128         u_int8_t u3 = nfmsg->nfgen_family;
3129         struct nf_conntrack_zone zone;
3130         int err;
3131
3132         if (cda[CTA_EXPECT_TUPLE]) {
3133                 /* delete a single expect by tuple */
3134                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3135                 if (err < 0)
3136                         return err;
3137
3138                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3139                                             u3, NULL);
3140                 if (err < 0)
3141                         return err;
3142
3143                 /* bump usage count to 2 */
3144                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
3145                 if (!exp)
3146                         return -ENOENT;
3147
3148                 if (cda[CTA_EXPECT_ID]) {
3149                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3150                         if (ntohl(id) != (u32)(unsigned long)exp) {
3151                                 nf_ct_expect_put(exp);
3152                                 return -ENOENT;
3153                         }
3154                 }
3155
3156                 /* after list removal, usage count == 1 */
3157                 spin_lock_bh(&nf_conntrack_expect_lock);
3158                 if (del_timer(&exp->timeout)) {
3159                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3160                                                    nlmsg_report(nlh));
3161                         nf_ct_expect_put(exp);
3162                 }
3163                 spin_unlock_bh(&nf_conntrack_expect_lock);
3164                 /* have to put what we 'get' above.
3165                  * after this line usage count == 0 */
3166                 nf_ct_expect_put(exp);
3167         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3168                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3169
3170                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
3171                                          NETLINK_CB(skb).portid,
3172                                          nlmsg_report(nlh));
3173         } else {
3174                 /* This basically means we have to flush everything*/
3175                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3176                                          NETLINK_CB(skb).portid,
3177                                          nlmsg_report(nlh));
3178         }
3179
3180         return 0;
3181 }
3182 static int
3183 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3184                         const struct nlattr * const cda[])
3185 {
3186         if (cda[CTA_EXPECT_TIMEOUT]) {
3187                 if (!del_timer(&x->timeout))
3188                         return -ETIME;
3189
3190                 x->timeout.expires = jiffies +
3191                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3192                 add_timer(&x->timeout);
3193         }
3194         return 0;
3195 }
3196
3197 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3198         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3199         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3200 };
3201
3202 static int
3203 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3204                            struct nf_conntrack_expect *exp,
3205                            u_int8_t u3)
3206 {
3207 #if IS_ENABLED(CONFIG_NF_NAT)
3208         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3209         struct nf_conntrack_tuple nat_tuple = {};
3210         int err;
3211
3212         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3213                                exp_nat_nla_policy, NULL);
3214         if (err < 0)
3215                 return err;
3216
3217         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3218                 return -EINVAL;
3219
3220         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3221                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3222                                     u3, NULL);
3223         if (err < 0)
3224                 return err;
3225
3226         exp->saved_addr = nat_tuple.src.u3;
3227         exp->saved_proto = nat_tuple.src.u;
3228         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3229
3230         return 0;
3231 #else
3232         return -EOPNOTSUPP;
3233 #endif
3234 }
3235
3236 static struct nf_conntrack_expect *
3237 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3238                        struct nf_conntrack_helper *helper,
3239                        struct nf_conntrack_tuple *tuple,
3240                        struct nf_conntrack_tuple *mask)
3241 {
3242         u_int32_t class = 0;
3243         struct nf_conntrack_expect *exp;
3244         struct nf_conn_help *help;
3245         int err;
3246
3247         help = nfct_help(ct);
3248         if (!help)
3249                 return ERR_PTR(-EOPNOTSUPP);
3250
3251         if (cda[CTA_EXPECT_CLASS] && helper) {
3252                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3253                 if (class > helper->expect_class_max)
3254                         return ERR_PTR(-EINVAL);
3255         }
3256         exp = nf_ct_expect_alloc(ct);
3257         if (!exp)
3258                 return ERR_PTR(-ENOMEM);
3259
3260         if (cda[CTA_EXPECT_FLAGS]) {
3261                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3262                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3263         } else {
3264                 exp->flags = 0;
3265         }
3266         if (cda[CTA_EXPECT_FN]) {
3267                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3268                 struct nf_ct_helper_expectfn *expfn;
3269
3270                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3271                 if (expfn == NULL) {
3272                         err = -EINVAL;
3273                         goto err_out;
3274                 }
3275                 exp->expectfn = expfn->expectfn;
3276         } else
3277                 exp->expectfn = NULL;
3278
3279         exp->class = class;
3280         exp->master = ct;
3281         exp->helper = helper;
3282         exp->tuple = *tuple;
3283         exp->mask.src.u3 = mask->src.u3;
3284         exp->mask.src.u.all = mask->src.u.all;
3285
3286         if (cda[CTA_EXPECT_NAT]) {
3287                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3288                                                  exp, nf_ct_l3num(ct));
3289                 if (err < 0)
3290                         goto err_out;
3291         }
3292         return exp;
3293 err_out:
3294         nf_ct_expect_put(exp);
3295         return ERR_PTR(err);
3296 }
3297
3298 static int
3299 ctnetlink_create_expect(struct net *net,
3300                         const struct nf_conntrack_zone *zone,
3301                         const struct nlattr * const cda[],
3302                         u_int8_t u3, u32 portid, int report)
3303 {
3304         struct nf_conntrack_tuple tuple, mask, master_tuple;
3305         struct nf_conntrack_tuple_hash *h = NULL;
3306         struct nf_conntrack_helper *helper = NULL;
3307         struct nf_conntrack_expect *exp;
3308         struct nf_conn *ct;
3309         int err;
3310
3311         /* caller guarantees that those three CTA_EXPECT_* exist */
3312         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3313                                     u3, NULL);
3314         if (err < 0)
3315                 return err;
3316         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3317                                     u3, NULL);
3318         if (err < 0)
3319                 return err;
3320         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3321                                     u3, NULL);
3322         if (err < 0)
3323                 return err;
3324
3325         /* Look for master conntrack of this expectation */
3326         h = nf_conntrack_find_get(net, zone, &master_tuple);
3327         if (!h)
3328                 return -ENOENT;
3329         ct = nf_ct_tuplehash_to_ctrack(h);
3330
3331         rcu_read_lock();
3332         if (cda[CTA_EXPECT_HELP_NAME]) {
3333                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3334
3335                 helper = __nf_conntrack_helper_find(helpname, u3,
3336                                                     nf_ct_protonum(ct));
3337                 if (helper == NULL) {
3338                         rcu_read_unlock();
3339 #ifdef CONFIG_MODULES
3340                         if (request_module("nfct-helper-%s", helpname) < 0) {
3341                                 err = -EOPNOTSUPP;
3342                                 goto err_ct;
3343                         }
3344                         rcu_read_lock();
3345                         helper = __nf_conntrack_helper_find(helpname, u3,
3346                                                             nf_ct_protonum(ct));
3347                         if (helper) {
3348                                 err = -EAGAIN;
3349                                 goto err_rcu;
3350                         }
3351                         rcu_read_unlock();
3352 #endif
3353                         err = -EOPNOTSUPP;
3354                         goto err_ct;
3355                 }
3356         }
3357
3358         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3359         if (IS_ERR(exp)) {
3360                 err = PTR_ERR(exp);
3361                 goto err_rcu;
3362         }
3363
3364         err = nf_ct_expect_related_report(exp, portid, report);
3365         nf_ct_expect_put(exp);
3366 err_rcu:
3367         rcu_read_unlock();
3368 err_ct:
3369         nf_ct_put(ct);
3370         return err;
3371 }
3372
3373 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3374                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3375                                 const struct nlattr * const cda[],
3376                                 struct netlink_ext_ack *extack)
3377 {
3378         struct nf_conntrack_tuple tuple;
3379         struct nf_conntrack_expect *exp;
3380         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3381         u_int8_t u3 = nfmsg->nfgen_family;
3382         struct nf_conntrack_zone zone;
3383         int err;
3384
3385         if (!cda[CTA_EXPECT_TUPLE]
3386             || !cda[CTA_EXPECT_MASK]
3387             || !cda[CTA_EXPECT_MASTER])
3388                 return -EINVAL;
3389
3390         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3391         if (err < 0)
3392                 return err;
3393
3394         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3395                                     u3, NULL);
3396         if (err < 0)
3397                 return err;
3398
3399         spin_lock_bh(&nf_conntrack_expect_lock);
3400         exp = __nf_ct_expect_find(net, &zone, &tuple);
3401         if (!exp) {
3402                 spin_unlock_bh(&nf_conntrack_expect_lock);
3403                 err = -ENOENT;
3404                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3405                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3406                                                       NETLINK_CB(skb).portid,
3407                                                       nlmsg_report(nlh));
3408                 }
3409                 return err;
3410         }
3411
3412         err = -EEXIST;
3413         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3414                 err = ctnetlink_change_expect(exp, cda);
3415         spin_unlock_bh(&nf_conntrack_expect_lock);
3416
3417         return err;
3418 }
3419
3420 static int
3421 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3422                              const struct ip_conntrack_stat *st)
3423 {
3424         struct nlmsghdr *nlh;
3425         struct nfgenmsg *nfmsg;
3426         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3427
3428         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3429                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3430         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3431         if (nlh == NULL)
3432                 goto nlmsg_failure;
3433
3434         nfmsg = nlmsg_data(nlh);
3435         nfmsg->nfgen_family = AF_UNSPEC;
3436         nfmsg->version      = NFNETLINK_V0;
3437         nfmsg->res_id       = htons(cpu);
3438
3439         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3440             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3441             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3442                 goto nla_put_failure;
3443
3444         nlmsg_end(skb, nlh);
3445         return skb->len;
3446
3447 nla_put_failure:
3448 nlmsg_failure:
3449         nlmsg_cancel(skb, nlh);
3450         return -1;
3451 }
3452
3453 static int
3454 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3455 {
3456         int cpu;
3457         struct net *net = sock_net(skb->sk);
3458
3459         if (cb->args[0] == nr_cpu_ids)
3460                 return 0;
3461
3462         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3463                 const struct ip_conntrack_stat *st;
3464
3465                 if (!cpu_possible(cpu))
3466                         continue;
3467
3468                 st = per_cpu_ptr(net->ct.stat, cpu);
3469                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3470                                                  cb->nlh->nlmsg_seq,
3471                                                  cpu, st) < 0)
3472                         break;
3473         }
3474         cb->args[0] = cpu;
3475
3476         return skb->len;
3477 }
3478
3479 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3480                                   struct sk_buff *skb,
3481                                   const struct nlmsghdr *nlh,
3482                                   const struct nlattr * const cda[],
3483                                   struct netlink_ext_ack *extack)
3484 {
3485         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3486                 struct netlink_dump_control c = {
3487                         .dump = ctnetlink_exp_stat_cpu_dump,
3488                 };
3489                 return netlink_dump_start(ctnl, skb, nlh, &c);
3490         }
3491
3492         return 0;
3493 }
3494
3495 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3496 static struct nf_ct_event_notifier ctnl_notifier = {
3497         .fcn = ctnetlink_conntrack_event,
3498 };
3499
3500 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3501         .fcn = ctnetlink_expect_event,
3502 };
3503 #endif
3504
3505 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3506         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3507                                             .attr_count = CTA_MAX,
3508                                             .policy = ct_nla_policy },
3509         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3510                                             .attr_count = CTA_MAX,
3511                                             .policy = ct_nla_policy },
3512         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3513                                             .attr_count = CTA_MAX,
3514                                             .policy = ct_nla_policy },
3515         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3516                                             .attr_count = CTA_MAX,
3517                                             .policy = ct_nla_policy },
3518         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3519         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3520         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3521         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3522 };
3523
3524 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3525         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3526                                             .attr_count = CTA_EXPECT_MAX,
3527                                             .policy = exp_nla_policy },
3528         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3529                                             .attr_count = CTA_EXPECT_MAX,
3530                                             .policy = exp_nla_policy },
3531         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3532                                             .attr_count = CTA_EXPECT_MAX,
3533                                             .policy = exp_nla_policy },
3534         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3535 };
3536
3537 static const struct nfnetlink_subsystem ctnl_subsys = {
3538         .name                           = "conntrack",
3539         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3540         .cb_count                       = IPCTNL_MSG_MAX,
3541         .cb                             = ctnl_cb,
3542 };
3543
3544 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3545         .name                           = "conntrack_expect",
3546         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3547         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3548         .cb                             = ctnl_exp_cb,
3549 };
3550
3551 MODULE_ALIAS("ip_conntrack_netlink");
3552 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3553 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3554
3555 static int __net_init ctnetlink_net_init(struct net *net)
3556 {
3557 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3558         int ret;
3559
3560         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3561         if (ret < 0) {
3562                 pr_err("ctnetlink_init: cannot register notifier.\n");
3563                 goto err_out;
3564         }
3565
3566         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3567         if (ret < 0) {
3568                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3569                 goto err_unreg_notifier;
3570         }
3571 #endif
3572         return 0;
3573
3574 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3575 err_unreg_notifier:
3576         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3577 err_out:
3578         return ret;
3579 #endif
3580 }
3581
3582 static void ctnetlink_net_exit(struct net *net)
3583 {
3584 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3585         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3586         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3587 #endif
3588 }
3589
3590 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3591 {
3592         struct net *net;
3593
3594         list_for_each_entry(net, net_exit_list, exit_list)
3595                 ctnetlink_net_exit(net);
3596 }
3597
3598 static struct pernet_operations ctnetlink_net_ops = {
3599         .init           = ctnetlink_net_init,
3600         .exit_batch     = ctnetlink_net_exit_batch,
3601 };
3602
3603 static int __init ctnetlink_init(void)
3604 {
3605         int ret;
3606
3607         ret = nfnetlink_subsys_register(&ctnl_subsys);
3608         if (ret < 0) {
3609                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3610                 goto err_out;
3611         }
3612
3613         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3614         if (ret < 0) {
3615                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3616                 goto err_unreg_subsys;
3617         }
3618
3619         ret = register_pernet_subsys(&ctnetlink_net_ops);
3620         if (ret < 0) {
3621                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3622                 goto err_unreg_exp_subsys;
3623         }
3624 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3625         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3626         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3627 #endif
3628         return 0;
3629
3630 err_unreg_exp_subsys:
3631         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3632 err_unreg_subsys:
3633         nfnetlink_subsys_unregister(&ctnl_subsys);
3634 err_out:
3635         return ret;
3636 }
3637
3638 static void __exit ctnetlink_exit(void)
3639 {
3640         unregister_pernet_subsys(&ctnetlink_net_ops);
3641         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3642         nfnetlink_subsys_unregister(&ctnl_subsys);
3643 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3644         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3645 #endif
3646         synchronize_rcu();
3647 }
3648
3649 module_init(ctnetlink_init);
3650 module_exit(ctnetlink_exit);