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);
  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);
 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);
 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);
 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);
 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);
 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);
 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);
 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);
 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);
 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 /* all these functions access ct->ext. Caller must either hold a reference
 510  * on ct or prevent its deletion by holding either the bucket spinlock or
 511  * pcpu dying list lock.
 512  */
 513 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 514                                   struct nf_conn *ct, u32 type)
 515 {
 516         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 517             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 518             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 519             ctnetlink_dump_labels(skb, ct) < 0 ||
 520             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 521             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 522                 return -1;
 523
 524         return 0;
 525 }
 526
 527 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 528 {
 529         if (ctnetlink_dump_status(skb, ct) < 0 ||
 530             ctnetlink_dump_mark(skb, ct) < 0 ||
 531             ctnetlink_dump_secctx(skb, ct) < 0 ||
 532             ctnetlink_dump_id(skb, ct) < 0 ||
 533             ctnetlink_dump_use(skb, ct) < 0 ||
 534             ctnetlink_dump_master(skb, ct) < 0)
 535                 return -1;
 536
 537         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 538             (ctnetlink_dump_timeout(skb, ct) < 0 ||
 539              ctnetlink_dump_protoinfo(skb, ct) < 0))
 540                 return -1;
 541
 542         return 0;
 543 }
 544
 545 static int
 546 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 547                     struct nf_conn *ct, bool extinfo)
 548 {
 549         const struct nf_conntrack_zone *zone;
 550         struct nlmsghdr *nlh;
 551         struct nfgenmsg *nfmsg;
 552         struct nlattr *nest_parms;
 553         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 554
 555         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 556         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 557         if (nlh == NULL)
 558                 goto nlmsg_failure;
 559
 560         nfmsg = nlmsg_data(nlh);
 561         nfmsg->nfgen_family = nf_ct_l3num(ct);
 562         nfmsg->version      = NFNETLINK_V0;
 563         nfmsg->res_id       = 0;
 564
 565         zone = nf_ct_zone(ct);
 566
 567         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 568         if (!nest_parms)
 569                 goto nla_put_failure;
 570         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 571                 goto nla_put_failure;
 572         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 573                                    NF_CT_ZONE_DIR_ORIG) < 0)
 574                 goto nla_put_failure;
 575         nla_nest_end(skb, nest_parms);
 576
 577         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 578         if (!nest_parms)
 579                 goto nla_put_failure;
 580         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 581                 goto nla_put_failure;
 582         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 583                                    NF_CT_ZONE_DIR_REPL) < 0)
 584                 goto nla_put_failure;
 585         nla_nest_end(skb, nest_parms);
 586
 587         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 588                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 589                 goto nla_put_failure;
 590
 591         if (ctnetlink_dump_info(skb, ct) < 0)
 592                 goto nla_put_failure;
 593         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 594                 goto nla_put_failure;
 595
 596         nlmsg_end(skb, nlh);
 597         return skb->len;
 598
 599 nlmsg_failure:
 600 nla_put_failure:
 601         nlmsg_cancel(skb, nlh);
 602         return -1;
 603 }
 604
 605 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 606         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 607         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 608         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 609         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 610 };
 611
 612 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 613 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 614 {
 615         const struct nf_conntrack_l4proto *l4proto;
 616         size_t len, len4 = 0;
 617
 618         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 619         len *= 3u; /* ORIG, REPLY, MASTER */
 620
 621         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 622         len += l4proto->nlattr_size;
 623         if (l4proto->nlattr_tuple_size) {
 624                 len4 = l4proto->nlattr_tuple_size();
 625                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 626         }
 627
 628         return len + len4;
 629 }
 630 #endif
 631
 632 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 633 {
 634         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 635                 return 0;
 636         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 637                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 638                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 639                ;
 640 }
 641
 642 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 643 {
 644 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 645         int len, ret;
 646
 647         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 648         if (ret)
 649                 return 0;
 650
 651         return nla_total_size(0) /* CTA_SECCTX */
 652                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 653 #else
 654         return 0;
 655 #endif
 656 }
 657
 658 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 659 {
 660 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 661         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 662                 return 0;
 663         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 664 #else
 665         return 0;
 666 #endif
 667 }
 668
 669 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 670 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 671 {
 672         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 673                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 674                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 675                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 676                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 677                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 678                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 679                + ctnetlink_acct_size(ct)
 680                + ctnetlink_timestamp_size(ct)
 681                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 682                + nla_total_size(0) /* CTA_PROTOINFO */
 683                + nla_total_size(0) /* CTA_HELP */
 684                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 685                + ctnetlink_secctx_size(ct)
 686 #if IS_ENABLED(CONFIG_NF_NAT)
 687                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 688                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 689 #endif
 690 #ifdef CONFIG_NF_CONNTRACK_MARK
 691                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 692 #endif
 693 #ifdef CONFIG_NF_CONNTRACK_ZONES
 694                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 695 #endif
 696                + ctnetlink_proto_size(ct)
 697                + ctnetlink_label_size(ct)
 698                ;
 699 }
 700
 701 static int
 702 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 703 {
 704         const struct nf_conntrack_zone *zone;
 705         struct net *net;
 706         struct nlmsghdr *nlh;
 707         struct nfgenmsg *nfmsg;
 708         struct nlattr *nest_parms;
 709         struct nf_conn *ct = item->ct;
 710         struct sk_buff *skb;
 711         unsigned int type;
 712         unsigned int flags = 0, group;
 713         int err;
 714
 715         if (events & (1 << IPCT_DESTROY)) {
 716                 type = IPCTNL_MSG_CT_DELETE;
 717                 group = NFNLGRP_CONNTRACK_DESTROY;
 718         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 719                 type = IPCTNL_MSG_CT_NEW;
 720                 flags = NLM_F_CREATE|NLM_F_EXCL;
 721                 group = NFNLGRP_CONNTRACK_NEW;
 722         } else if (events) {
 723                 type = IPCTNL_MSG_CT_NEW;
 724                 group = NFNLGRP_CONNTRACK_UPDATE;
 725         } else
 726                 return 0;
 727
 728         net = nf_ct_net(ct);
 729         if (!item->report && !nfnetlink_has_listeners(net, group))
 730                 return 0;
 731
 732         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 733         if (skb == NULL)
 734                 goto errout;
 735
 736         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 737         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 738         if (nlh == NULL)
 739                 goto nlmsg_failure;
 740
 741         nfmsg = nlmsg_data(nlh);
 742         nfmsg->nfgen_family = nf_ct_l3num(ct);
 743         nfmsg->version  = NFNETLINK_V0;
 744         nfmsg->res_id   = 0;
 745
 746         zone = nf_ct_zone(ct);
 747
 748         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 749         if (!nest_parms)
 750                 goto nla_put_failure;
 751         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 752                 goto nla_put_failure;
 753         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 754                                    NF_CT_ZONE_DIR_ORIG) < 0)
 755                 goto nla_put_failure;
 756         nla_nest_end(skb, nest_parms);
 757
 758         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 759         if (!nest_parms)
 760                 goto nla_put_failure;
 761         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 762                 goto nla_put_failure;
 763         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 764                                    NF_CT_ZONE_DIR_REPL) < 0)
 765                 goto nla_put_failure;
 766         nla_nest_end(skb, nest_parms);
 767
 768         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 769                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 770                 goto nla_put_failure;
 771
 772         if (ctnetlink_dump_id(skb, ct) < 0)
 773                 goto nla_put_failure;
 774
 775         if (ctnetlink_dump_status(skb, ct) < 0)
 776                 goto nla_put_failure;
 777
 778         if (events & (1 << IPCT_DESTROY)) {
 779                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 780                     ctnetlink_dump_timestamp(skb, ct) < 0)
 781                         goto nla_put_failure;
 782         } else {
 783                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 784                         goto nla_put_failure;
 785
 786                 if (events & (1 << IPCT_PROTOINFO)
 787                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 788                         goto nla_put_failure;
 789
 790                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 791                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 792                         goto nla_put_failure;
 793
 794 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 795                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 796                     && ctnetlink_dump_secctx(skb, ct) < 0)
 797                         goto nla_put_failure;
 798 #endif
 799                 if (events & (1 << IPCT_LABEL) &&
 800                      ctnetlink_dump_labels(skb, ct) < 0)
 801                         goto nla_put_failure;
 802
 803                 if (events & (1 << IPCT_RELATED) &&
 804                     ctnetlink_dump_master(skb, ct) < 0)
 805                         goto nla_put_failure;
 806
 807                 if (events & (1 << IPCT_SEQADJ) &&
 808                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 809                         goto nla_put_failure;
 810
 811                 if (events & (1 << IPCT_SYNPROXY) &&
 812                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 813                         goto nla_put_failure;
 814         }
 815
 816 #ifdef CONFIG_NF_CONNTRACK_MARK
 817         if ((events & (1 << IPCT_MARK) || ct->mark)
 818             && ctnetlink_dump_mark(skb, ct) < 0)
 819                 goto nla_put_failure;
 820 #endif
 821         nlmsg_end(skb, nlh);
 822         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 823                              GFP_ATOMIC);
 824         if (err == -ENOBUFS || err == -EAGAIN)
 825                 return -ENOBUFS;
 826
 827         return 0;
 828
 829 nla_put_failure:
 830         nlmsg_cancel(skb, nlh);
 831 nlmsg_failure:
 832         kfree_skb(skb);
 833 errout:
 834         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 835                 return -ENOBUFS;
 836
 837         return 0;
 838 }
 839 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 840
 841 static int ctnetlink_done(struct netlink_callback *cb)
 842 {
 843         if (cb->args[1])
 844                 nf_ct_put((struct nf_conn *)cb->args[1]);
 845         kfree(cb->data);
 846         return 0;
 847 }
 848
 849 struct ctnetlink_filter {
 850         u8 family;
 851         struct {
 852                 u_int32_t val;
 853                 u_int32_t mask;
 854         } mark;
 855 };
 856
 857 static struct ctnetlink_filter *
 858 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 859 {
 860         struct ctnetlink_filter *filter;
 861
 862 #ifndef CONFIG_NF_CONNTRACK_MARK
 863         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 864                 return ERR_PTR(-EOPNOTSUPP);
 865 #endif
 866
 867         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 868         if (filter == NULL)
 869                 return ERR_PTR(-ENOMEM);
 870
 871         filter->family = family;
 872
 873 #ifdef CONFIG_NF_CONNTRACK_MARK
 874         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
 875                 filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 876                 filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 877         }
 878 #endif
 879         return filter;
 880 }
 881
 882 static int ctnetlink_start(struct netlink_callback *cb)
 883 {
 884         const struct nlattr * const *cda = cb->data;
 885         struct ctnetlink_filter *filter = NULL;
 886         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 887         u8 family = nfmsg->nfgen_family;
 888
 889         if (family || (cda[CTA_MARK] && cda[CTA_MARK_MASK])) {
 890                 filter = ctnetlink_alloc_filter(cda, family);
 891                 if (IS_ERR(filter))
 892                         return PTR_ERR(filter);
 893         }
 894
 895         cb->data = filter;
 896         return 0;
 897 }
 898
 899 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 900 {
 901         struct ctnetlink_filter *filter = data;
 902
 903         if (filter == NULL)
 904                 goto out;
 905
 906         /* Match entries of a given L3 protocol number.
 907          * If it is not specified, ie. l3proto == 0,
 908          * then match everything.
 909          */
 910         if (filter->family && nf_ct_l3num(ct) != filter->family)
 911                 goto ignore_entry;
 912
 913 #ifdef CONFIG_NF_CONNTRACK_MARK
 914         if ((ct->mark & filter->mark.mask) != filter->mark.val)
 915                 goto ignore_entry;
 916 #endif
 917
 918 out:
 919         return 1;
 920
 921 ignore_entry:
 922         return 0;
 923 }
 924
 925 static int
 926 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 927 {
 928         struct net *net = sock_net(skb->sk);
 929         struct nf_conn *ct, *last;
 930         struct nf_conntrack_tuple_hash *h;
 931         struct hlist_nulls_node *n;
 932         struct nf_conn *nf_ct_evict[8];
 933         int res, i;
 934         spinlock_t *lockp;
 935
 936         last = (struct nf_conn *)cb->args[1];
 937         i = 0;
 938
 939         local_bh_disable();
 940         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 941 restart:
 942                 while (i) {
 943                         i--;
 944                         if (nf_ct_should_gc(nf_ct_evict[i]))
 945                                 nf_ct_kill(nf_ct_evict[i]);
 946                         nf_ct_put(nf_ct_evict[i]);
 947                 }
 948
 949                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 950                 nf_conntrack_lock(lockp);
 951                 if (cb->args[0] >= nf_conntrack_htable_size) {
 952                         spin_unlock(lockp);
 953                         goto out;
 954                 }
 955                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 956                                            hnnode) {
 957                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 958                                 continue;
 959                         ct = nf_ct_tuplehash_to_ctrack(h);
 960                         if (nf_ct_is_expired(ct)) {
 961                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 962                                     atomic_inc_not_zero(&ct->ct_general.use))
 963                                         nf_ct_evict[i++] = ct;
 964                                 continue;
 965                         }
 966
 967                         if (!net_eq(net, nf_ct_net(ct)))
 968                                 continue;
 969
 970                         if (cb->args[1]) {
 971                                 if (ct != last)
 972                                         continue;
 973                                 cb->args[1] = 0;
 974                         }
 975                         if (!ctnetlink_filter_match(ct, cb->data))
 976                                 continue;
 977
 978                         res =
 979                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 980                                             cb->nlh->nlmsg_seq,
 981                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 982                                             ct, true);
 983                         if (res < 0) {
 984                                 nf_conntrack_get(&ct->ct_general);
 985                                 cb->args[1] = (unsigned long)ct;
 986                                 spin_unlock(lockp);
 987                                 goto out;
 988                         }
 989                 }
 990                 spin_unlock(lockp);
 991                 if (cb->args[1]) {
 992                         cb->args[1] = 0;
 993                         goto restart;
 994                 }
 995         }
 996 out:
 997         local_bh_enable();
 998         if (last) {
 999                 /* nf ct hash resize happened, now clear the leftover. */
1000                 if ((struct nf_conn *)cb->args[1] == last)
1001                         cb->args[1] = 0;
1002
1003                 nf_ct_put(last);
1004         }
1005
1006         while (i) {
1007                 i--;
1008                 if (nf_ct_should_gc(nf_ct_evict[i]))
1009                         nf_ct_kill(nf_ct_evict[i]);
1010                 nf_ct_put(nf_ct_evict[i]);
1011         }
1012
1013         return skb->len;
1014 }
1015
1016 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1017                                 struct nf_conntrack_tuple *t)
1018 {
1019         if (!tb[CTA_IP_V4_SRC] || !tb[CTA_IP_V4_DST])
1020                 return -EINVAL;
1021
1022         t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1023         t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1024
1025         return 0;
1026 }
1027
1028 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1029                                 struct nf_conntrack_tuple *t)
1030 {
1031         if (!tb[CTA_IP_V6_SRC] || !tb[CTA_IP_V6_DST])
1032                 return -EINVAL;
1033
1034         t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1035         t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1036
1037         return 0;
1038 }
1039
1040 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1041                                     struct nf_conntrack_tuple *tuple)
1042 {
1043         struct nlattr *tb[CTA_IP_MAX+1];
1044         int ret = 0;
1045
1046         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1047         if (ret < 0)
1048                 return ret;
1049
1050         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1051                                              cta_ip_nla_policy, NULL);
1052         if (ret)
1053                 return ret;
1054
1055         switch (tuple->src.l3num) {
1056         case NFPROTO_IPV4:
1057                 ret = ipv4_nlattr_to_tuple(tb, tuple);
1058                 break;
1059         case NFPROTO_IPV6:
1060                 ret = ipv6_nlattr_to_tuple(tb, tuple);
1061                 break;
1062         }
1063
1064         return ret;
1065 }
1066
1067 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1068         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1069 };
1070
1071 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1072                                        struct nf_conntrack_tuple *tuple)
1073 {
1074         const struct nf_conntrack_l4proto *l4proto;
1075         struct nlattr *tb[CTA_PROTO_MAX+1];
1076         int ret = 0;
1077
1078         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1079                                           proto_nla_policy, NULL);
1080         if (ret < 0)
1081                 return ret;
1082
1083         if (!tb[CTA_PROTO_NUM])
1084                 return -EINVAL;
1085         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1086
1087         rcu_read_lock();
1088         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1089
1090         if (likely(l4proto->nlattr_to_tuple)) {
1091                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1092                                                      l4proto->nla_policy,
1093                                                      NULL);
1094                 if (ret == 0)
1095                         ret = l4proto->nlattr_to_tuple(tb, tuple);
1096         }
1097
1098         rcu_read_unlock();
1099
1100         return ret;
1101 }
1102
1103 static int
1104 ctnetlink_parse_zone(const struct nlattr *attr,
1105                      struct nf_conntrack_zone *zone)
1106 {
1107         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1108                         NF_CT_DEFAULT_ZONE_DIR, 0);
1109 #ifdef CONFIG_NF_CONNTRACK_ZONES
1110         if (attr)
1111                 zone->id = ntohs(nla_get_be16(attr));
1112 #else
1113         if (attr)
1114                 return -EOPNOTSUPP;
1115 #endif
1116         return 0;
1117 }
1118
1119 static int
1120 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1121                            struct nf_conntrack_zone *zone)
1122 {
1123         int ret;
1124
1125         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1126                 return -EINVAL;
1127
1128         ret = ctnetlink_parse_zone(attr, zone);
1129         if (ret < 0)
1130                 return ret;
1131
1132         if (type == CTA_TUPLE_REPLY)
1133                 zone->dir = NF_CT_ZONE_DIR_REPL;
1134         else
1135                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1136
1137         return 0;
1138 }
1139
1140 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1141         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1142         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1143         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1144 };
1145
1146 static int
1147 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1148                       struct nf_conntrack_tuple *tuple, u32 type,
1149                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1150 {
1151         struct nlattr *tb[CTA_TUPLE_MAX+1];
1152         int err;
1153
1154         memset(tuple, 0, sizeof(*tuple));
1155
1156         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1157                                           tuple_nla_policy, NULL);
1158         if (err < 0)
1159                 return err;
1160
1161         if (!tb[CTA_TUPLE_IP])
1162                 return -EINVAL;
1163
1164         tuple->src.l3num = l3num;
1165
1166         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1167         if (err < 0)
1168                 return err;
1169
1170         if (!tb[CTA_TUPLE_PROTO])
1171                 return -EINVAL;
1172
1173         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1174         if (err < 0)
1175                 return err;
1176
1177         if (tb[CTA_TUPLE_ZONE]) {
1178                 if (!zone)
1179                         return -EINVAL;
1180
1181                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1182                                                  type, zone);
1183                 if (err < 0)
1184                         return err;
1185         }
1186
1187         /* orig and expect tuples get DIR_ORIGINAL */
1188         if (type == CTA_TUPLE_REPLY)
1189                 tuple->dst.dir = IP_CT_DIR_REPLY;
1190         else
1191                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1192
1193         return 0;
1194 }
1195
1196 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1197         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1198                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1199 };
1200
1201 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1202                                 struct nlattr **helpinfo)
1203 {
1204         int err;
1205         struct nlattr *tb[CTA_HELP_MAX+1];
1206
1207         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1208                                           help_nla_policy, NULL);
1209         if (err < 0)
1210                 return err;
1211
1212         if (!tb[CTA_HELP_NAME])
1213                 return -EINVAL;
1214
1215         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1216
1217         if (tb[CTA_HELP_INFO])
1218                 *helpinfo = tb[CTA_HELP_INFO];
1219
1220         return 0;
1221 }
1222
1223 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1224         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1225         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1226         [CTA_STATUS]            = { .type = NLA_U32 },
1227         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1228         [CTA_HELP]              = { .type = NLA_NESTED },
1229         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1230         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1231         [CTA_MARK]              = { .type = NLA_U32 },
1232         [CTA_ID]                = { .type = NLA_U32 },
1233         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1234         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1235         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1236         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1237         [CTA_ZONE]              = { .type = NLA_U16 },
1238         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1239         [CTA_LABELS]            = { .type = NLA_BINARY,
1240                                     .len = NF_CT_LABELS_MAX_SIZE },
1241         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1242                                     .len = NF_CT_LABELS_MAX_SIZE },
1243 };
1244
1245 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1246 {
1247         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1248                 return 0;
1249
1250         return ctnetlink_filter_match(ct, data);
1251 }
1252
1253 static int ctnetlink_flush_conntrack(struct net *net,
1254                                      const struct nlattr * const cda[],
1255                                      u32 portid, int report, u8 family)
1256 {
1257         struct ctnetlink_filter *filter = NULL;
1258
1259         if (family || (cda[CTA_MARK] && cda[CTA_MARK_MASK])) {
1260                 filter = ctnetlink_alloc_filter(cda, family);
1261                 if (IS_ERR(filter))
1262                         return PTR_ERR(filter);
1263         }
1264
1265         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1266                                   portid, report);
1267         kfree(filter);
1268
1269         return 0;
1270 }
1271
1272 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1273                                    struct sk_buff *skb,
1274                                    const struct nlmsghdr *nlh,
1275                                    const struct nlattr * const cda[],
1276                                    struct netlink_ext_ack *extack)
1277 {
1278         struct nf_conntrack_tuple_hash *h;
1279         struct nf_conntrack_tuple tuple;
1280         struct nf_conn *ct;
1281         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1282         struct nf_conntrack_zone zone;
1283         int err;
1284
1285         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1286         if (err < 0)
1287                 return err;
1288
1289         if (cda[CTA_TUPLE_ORIG])
1290                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1291                                             nfmsg->nfgen_family, &zone);
1292         else if (cda[CTA_TUPLE_REPLY])
1293                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1294                                             nfmsg->nfgen_family, &zone);
1295         else {
1296                 u_int8_t u3 = nfmsg->version ? nfmsg->nfgen_family : AF_UNSPEC;
1297
1298                 return ctnetlink_flush_conntrack(net, cda,
1299                                                  NETLINK_CB(skb).portid,
1300                                                  nlmsg_report(nlh), u3);
1301         }
1302
1303         if (err < 0)
1304                 return err;
1305
1306         h = nf_conntrack_find_get(net, &zone, &tuple);
1307         if (!h)
1308                 return -ENOENT;
1309
1310         ct = nf_ct_tuplehash_to_ctrack(h);
1311
1312         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1313                 nf_ct_put(ct);
1314                 return -EBUSY;
1315         }
1316
1317         if (cda[CTA_ID]) {
1318                 __be32 id = nla_get_be32(cda[CTA_ID]);
1319
1320                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1321                         nf_ct_put(ct);
1322                         return -ENOENT;
1323                 }
1324         }
1325
1326         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1327         nf_ct_put(ct);
1328
1329         return 0;
1330 }
1331
1332 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1333                                    struct sk_buff *skb,
1334                                    const struct nlmsghdr *nlh,
1335                                    const struct nlattr * const cda[],
1336                                    struct netlink_ext_ack *extack)
1337 {
1338         struct nf_conntrack_tuple_hash *h;
1339         struct nf_conntrack_tuple tuple;
1340         struct nf_conn *ct;
1341         struct sk_buff *skb2 = NULL;
1342         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1343         u_int8_t u3 = nfmsg->nfgen_family;
1344         struct nf_conntrack_zone zone;
1345         int err;
1346
1347         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1348                 struct netlink_dump_control c = {
1349                         .start = ctnetlink_start,
1350                         .dump = ctnetlink_dump_table,
1351                         .done = ctnetlink_done,
1352                         .data = (void *)cda,
1353                 };
1354
1355                 return netlink_dump_start(ctnl, skb, nlh, &c);
1356         }
1357
1358         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1359         if (err < 0)
1360                 return err;
1361
1362         if (cda[CTA_TUPLE_ORIG])
1363                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1364                                             u3, &zone);
1365         else if (cda[CTA_TUPLE_REPLY])
1366                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1367                                             u3, &zone);
1368         else
1369                 return -EINVAL;
1370
1371         if (err < 0)
1372                 return err;
1373
1374         h = nf_conntrack_find_get(net, &zone, &tuple);
1375         if (!h)
1376                 return -ENOENT;
1377
1378         ct = nf_ct_tuplehash_to_ctrack(h);
1379
1380         err = -ENOMEM;
1381         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1382         if (skb2 == NULL) {
1383                 nf_ct_put(ct);
1384                 return -ENOMEM;
1385         }
1386
1387         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1388                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct, true);
1389         nf_ct_put(ct);
1390         if (err <= 0)
1391                 goto free;
1392
1393         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1394         if (err < 0)
1395                 goto out;
1396
1397         return 0;
1398
1399 free:
1400         kfree_skb(skb2);
1401 out:
1402         /* this avoids a loop in nfnetlink. */
1403         return err == -EAGAIN ? -ENOBUFS : err;
1404 }
1405
1406 static int ctnetlink_done_list(struct netlink_callback *cb)
1407 {
1408         if (cb->args[1])
1409                 nf_ct_put((struct nf_conn *)cb->args[1]);
1410         return 0;
1411 }
1412
1413 static int
1414 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1415 {
1416         struct nf_conn *ct, *last;
1417         struct nf_conntrack_tuple_hash *h;
1418         struct hlist_nulls_node *n;
1419         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1420         u_int8_t l3proto = nfmsg->nfgen_family;
1421         int res;
1422         int cpu;
1423         struct hlist_nulls_head *list;
1424         struct net *net = sock_net(skb->sk);
1425
1426         if (cb->args[2])
1427                 return 0;
1428
1429         last = (struct nf_conn *)cb->args[1];
1430
1431         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1432                 struct ct_pcpu *pcpu;
1433
1434                 if (!cpu_possible(cpu))
1435                         continue;
1436
1437                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1438                 spin_lock_bh(&pcpu->lock);
1439                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1440 restart:
1441                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1442                         ct = nf_ct_tuplehash_to_ctrack(h);
1443                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1444                                 continue;
1445                         if (cb->args[1]) {
1446                                 if (ct != last)
1447                                         continue;
1448                                 cb->args[1] = 0;
1449                         }
1450
1451                         /* We can't dump extension info for the unconfirmed
1452                          * list because unconfirmed conntracks can have
1453                          * ct->ext reallocated (and thus freed).
1454                          *
1455                          * In the dying list case ct->ext can't be free'd
1456                          * until after we drop pcpu->lock.
1457                          */
1458                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1459                                                   cb->nlh->nlmsg_seq,
1460                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1461                                                   ct, dying ? true : false);
1462                         if (res < 0) {
1463                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1464                                         continue;
1465                                 cb->args[0] = cpu;
1466                                 cb->args[1] = (unsigned long)ct;
1467                                 spin_unlock_bh(&pcpu->lock);
1468                                 goto out;
1469                         }
1470                 }
1471                 if (cb->args[1]) {
1472                         cb->args[1] = 0;
1473                         goto restart;
1474                 }
1475                 spin_unlock_bh(&pcpu->lock);
1476         }
1477         cb->args[2] = 1;
1478 out:
1479         if (last)
1480                 nf_ct_put(last);
1481
1482         return skb->len;
1483 }
1484
1485 static int
1486 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1487 {
1488         return ctnetlink_dump_list(skb, cb, true);
1489 }
1490
1491 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1492                                   struct sk_buff *skb,
1493                                   const struct nlmsghdr *nlh,
1494                                   const struct nlattr * const cda[],
1495                                   struct netlink_ext_ack *extack)
1496 {
1497         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1498                 struct netlink_dump_control c = {
1499                         .dump = ctnetlink_dump_dying,
1500                         .done = ctnetlink_done_list,
1501                 };
1502                 return netlink_dump_start(ctnl, skb, nlh, &c);
1503         }
1504
1505         return -EOPNOTSUPP;
1506 }
1507
1508 static int
1509 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1510 {
1511         return ctnetlink_dump_list(skb, cb, false);
1512 }
1513
1514 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1515                                         struct sk_buff *skb,
1516                                         const struct nlmsghdr *nlh,
1517                                         const struct nlattr * const cda[],
1518                                         struct netlink_ext_ack *extack)
1519 {
1520         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1521                 struct netlink_dump_control c = {
1522                         .dump = ctnetlink_dump_unconfirmed,
1523                         .done = ctnetlink_done_list,
1524                 };
1525                 return netlink_dump_start(ctnl, skb, nlh, &c);
1526         }
1527
1528         return -EOPNOTSUPP;
1529 }
1530
1531 #if IS_ENABLED(CONFIG_NF_NAT)
1532 static int
1533 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1534                           enum nf_nat_manip_type manip,
1535                           const struct nlattr *attr)
1536         __must_hold(RCU)
1537 {
1538         struct nf_nat_hook *nat_hook;
1539         int err;
1540
1541         nat_hook = rcu_dereference(nf_nat_hook);
1542         if (!nat_hook) {
1543 #ifdef CONFIG_MODULES
1544                 rcu_read_unlock();
1545                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1546                 if (request_module("nf-nat") < 0) {
1547                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1548                         rcu_read_lock();
1549                         return -EOPNOTSUPP;
1550                 }
1551                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1552                 rcu_read_lock();
1553                 nat_hook = rcu_dereference(nf_nat_hook);
1554                 if (nat_hook)
1555                         return -EAGAIN;
1556 #endif
1557                 return -EOPNOTSUPP;
1558         }
1559
1560         err = nat_hook->parse_nat_setup(ct, manip, attr);
1561         if (err == -EAGAIN) {
1562 #ifdef CONFIG_MODULES
1563                 rcu_read_unlock();
1564                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1565                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1566                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1567                         rcu_read_lock();
1568                         return -EOPNOTSUPP;
1569                 }
1570                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1571                 rcu_read_lock();
1572 #else
1573                 err = -EOPNOTSUPP;
1574 #endif
1575         }
1576         return err;
1577 }
1578 #endif
1579
1580 static void
1581 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1582                           unsigned long off)
1583 {
1584         unsigned int bit;
1585
1586         /* Ignore these unchangable bits */
1587         on &= ~IPS_UNCHANGEABLE_MASK;
1588         off &= ~IPS_UNCHANGEABLE_MASK;
1589
1590         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1591                 if (on & (1 << bit))
1592                         set_bit(bit, &ct->status);
1593                 else if (off & (1 << bit))
1594                         clear_bit(bit, &ct->status);
1595         }
1596 }
1597
1598 static int
1599 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1600 {
1601         unsigned long d;
1602         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1603         d = ct->status ^ status;
1604
1605         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1606                 /* unchangeable */
1607                 return -EBUSY;
1608
1609         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1610                 /* SEEN_REPLY bit can only be set */
1611                 return -EBUSY;
1612
1613         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1614                 /* ASSURED bit can only be set */
1615                 return -EBUSY;
1616
1617         __ctnetlink_change_status(ct, status, 0);
1618         return 0;
1619 }
1620
1621 static int
1622 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1623 {
1624 #if IS_ENABLED(CONFIG_NF_NAT)
1625         int ret;
1626
1627         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1628                 return 0;
1629
1630         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1631                                         cda[CTA_NAT_DST]);
1632         if (ret < 0)
1633                 return ret;
1634
1635         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1636                                          cda[CTA_NAT_SRC]);
1637 #else
1638         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1639                 return 0;
1640         return -EOPNOTSUPP;
1641 #endif
1642 }
1643
1644 static int ctnetlink_change_helper(struct nf_conn *ct,
1645                                    const struct nlattr * const cda[])
1646 {
1647         struct nf_conntrack_helper *helper;
1648         struct nf_conn_help *help = nfct_help(ct);
1649         char *helpname = NULL;
1650         struct nlattr *helpinfo = NULL;
1651         int err;
1652
1653         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1654         if (err < 0)
1655                 return err;
1656
1657         /* don't change helper of sibling connections */
1658         if (ct->master) {
1659                 /* If we try to change the helper to the same thing twice,
1660                  * treat the second attempt as a no-op instead of returning
1661                  * an error.
1662                  */
1663                 err = -EBUSY;
1664                 if (help) {
1665                         rcu_read_lock();
1666                         helper = rcu_dereference(help->helper);
1667                         if (helper && !strcmp(helper->name, helpname))
1668                                 err = 0;
1669                         rcu_read_unlock();
1670                 }
1671
1672                 return err;
1673         }
1674
1675         if (!strcmp(helpname, "")) {
1676                 if (help && help->helper) {
1677                         /* we had a helper before ... */
1678                         nf_ct_remove_expectations(ct);
1679                         RCU_INIT_POINTER(help->helper, NULL);
1680                 }
1681
1682                 return 0;
1683         }
1684
1685         rcu_read_lock();
1686         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1687                                             nf_ct_protonum(ct));
1688         if (helper == NULL) {
1689                 rcu_read_unlock();
1690                 return -EOPNOTSUPP;
1691         }
1692
1693         if (help) {
1694                 if (help->helper == helper) {
1695                         /* update private helper data if allowed. */
1696                         if (helper->from_nlattr)
1697                                 helper->from_nlattr(helpinfo, ct);
1698                         err = 0;
1699                 } else
1700                         err = -EBUSY;
1701         } else {
1702                 /* we cannot set a helper for an existing conntrack */
1703                 err = -EOPNOTSUPP;
1704         }
1705
1706         rcu_read_unlock();
1707         return err;
1708 }
1709
1710 static int ctnetlink_change_timeout(struct nf_conn *ct,
1711                                     const struct nlattr * const cda[])
1712 {
1713         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1714
1715         if (timeout > INT_MAX)
1716                 timeout = INT_MAX;
1717         ct->timeout = nfct_time_stamp + (u32)timeout;
1718
1719         if (test_bit(IPS_DYING_BIT, &ct->status))
1720                 return -ETIME;
1721
1722         return 0;
1723 }
1724
1725 #if defined(CONFIG_NF_CONNTRACK_MARK)
1726 static void ctnetlink_change_mark(struct nf_conn *ct,
1727                                     const struct nlattr * const cda[])
1728 {
1729         u32 mark, newmark, mask = 0;
1730
1731         if (cda[CTA_MARK_MASK])
1732                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1733
1734         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1735         newmark = (ct->mark & mask) ^ mark;
1736         if (newmark != ct->mark)
1737                 ct->mark = newmark;
1738 }
1739 #endif
1740
1741 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1742         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1743         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1744         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1745 };
1746
1747 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1748                                       const struct nlattr * const cda[])
1749 {
1750         const struct nlattr *attr = cda[CTA_PROTOINFO];
1751         const struct nf_conntrack_l4proto *l4proto;
1752         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1753         int err = 0;
1754
1755         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
1756                                           protoinfo_policy, NULL);
1757         if (err < 0)
1758                 return err;
1759
1760         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1761         if (l4proto->from_nlattr)
1762                 err = l4proto->from_nlattr(tb, ct);
1763
1764         return err;
1765 }
1766
1767 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1768         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1769         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1770         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1771 };
1772
1773 static int change_seq_adj(struct nf_ct_seqadj *seq,
1774                           const struct nlattr * const attr)
1775 {
1776         int err;
1777         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1778
1779         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
1780                                           seqadj_policy, NULL);
1781         if (err < 0)
1782                 return err;
1783
1784         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1785                 return -EINVAL;
1786
1787         seq->correction_pos =
1788                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1789
1790         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1791                 return -EINVAL;
1792
1793         seq->offset_before =
1794                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1795
1796         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1797                 return -EINVAL;
1798
1799         seq->offset_after =
1800                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1801
1802         return 0;
1803 }
1804
1805 static int
1806 ctnetlink_change_seq_adj(struct nf_conn *ct,
1807                          const struct nlattr * const cda[])
1808 {
1809         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1810         int ret = 0;
1811
1812         if (!seqadj)
1813                 return 0;
1814
1815         spin_lock_bh(&ct->lock);
1816         if (cda[CTA_SEQ_ADJ_ORIG]) {
1817                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1818                                      cda[CTA_SEQ_ADJ_ORIG]);
1819                 if (ret < 0)
1820                         goto err;
1821
1822                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1823         }
1824
1825         if (cda[CTA_SEQ_ADJ_REPLY]) {
1826                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1827                                      cda[CTA_SEQ_ADJ_REPLY]);
1828                 if (ret < 0)
1829                         goto err;
1830
1831                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1832         }
1833
1834         spin_unlock_bh(&ct->lock);
1835         return 0;
1836 err:
1837         spin_unlock_bh(&ct->lock);
1838         return ret;
1839 }
1840
1841 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
1842         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
1843         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
1844         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
1845 };
1846
1847 static int ctnetlink_change_synproxy(struct nf_conn *ct,
1848                                      const struct nlattr * const cda[])
1849 {
1850         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
1851         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
1852         int err;
1853
1854         if (!synproxy)
1855                 return 0;
1856
1857         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
1858                                           cda[CTA_SYNPROXY], synproxy_policy,
1859                                           NULL);
1860         if (err < 0)
1861                 return err;
1862
1863         if (!tb[CTA_SYNPROXY_ISN] ||
1864             !tb[CTA_SYNPROXY_ITS] ||
1865             !tb[CTA_SYNPROXY_TSOFF])
1866                 return -EINVAL;
1867
1868         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
1869         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
1870         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
1871
1872         return 0;
1873 }
1874
1875 static int
1876 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1877 {
1878 #ifdef CONFIG_NF_CONNTRACK_LABELS
1879         size_t len = nla_len(cda[CTA_LABELS]);
1880         const void *mask = cda[CTA_LABELS_MASK];
1881
1882         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1883                 return -EINVAL;
1884
1885         if (mask) {
1886                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1887                     nla_len(cda[CTA_LABELS_MASK]) != len)
1888                         return -EINVAL;
1889                 mask = nla_data(cda[CTA_LABELS_MASK]);
1890         }
1891
1892         len /= sizeof(u32);
1893
1894         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1895 #else
1896         return -EOPNOTSUPP;
1897 #endif
1898 }
1899
1900 static int
1901 ctnetlink_change_conntrack(struct nf_conn *ct,
1902                            const struct nlattr * const cda[])
1903 {
1904         int err;
1905
1906         /* only allow NAT changes and master assignation for new conntracks */
1907         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1908                 return -EOPNOTSUPP;
1909
1910         if (cda[CTA_HELP]) {
1911                 err = ctnetlink_change_helper(ct, cda);
1912                 if (err < 0)
1913                         return err;
1914         }
1915
1916         if (cda[CTA_TIMEOUT]) {
1917                 err = ctnetlink_change_timeout(ct, cda);
1918                 if (err < 0)
1919                         return err;
1920         }
1921
1922         if (cda[CTA_STATUS]) {
1923                 err = ctnetlink_change_status(ct, cda);
1924                 if (err < 0)
1925                         return err;
1926         }
1927
1928         if (cda[CTA_PROTOINFO]) {
1929                 err = ctnetlink_change_protoinfo(ct, cda);
1930                 if (err < 0)
1931                         return err;
1932         }
1933
1934 #if defined(CONFIG_NF_CONNTRACK_MARK)
1935         if (cda[CTA_MARK])
1936                 ctnetlink_change_mark(ct, cda);
1937 #endif
1938
1939         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1940                 err = ctnetlink_change_seq_adj(ct, cda);
1941                 if (err < 0)
1942                         return err;
1943         }
1944
1945         if (cda[CTA_SYNPROXY]) {
1946                 err = ctnetlink_change_synproxy(ct, cda);
1947                 if (err < 0)
1948                         return err;
1949         }
1950
1951         if (cda[CTA_LABELS]) {
1952                 err = ctnetlink_attach_labels(ct, cda);
1953                 if (err < 0)
1954                         return err;
1955         }
1956
1957         return 0;
1958 }
1959
1960 static struct nf_conn *
1961 ctnetlink_create_conntrack(struct net *net,
1962                            const struct nf_conntrack_zone *zone,
1963                            const struct nlattr * const cda[],
1964                            struct nf_conntrack_tuple *otuple,
1965                            struct nf_conntrack_tuple *rtuple,
1966                            u8 u3)
1967 {
1968         struct nf_conn *ct;
1969         int err = -EINVAL;
1970         struct nf_conntrack_helper *helper;
1971         struct nf_conn_tstamp *tstamp;
1972         u64 timeout;
1973
1974         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1975         if (IS_ERR(ct))
1976                 return ERR_PTR(-ENOMEM);
1977
1978         if (!cda[CTA_TIMEOUT])
1979                 goto err1;
1980
1981         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1982         if (timeout > INT_MAX)
1983                 timeout = INT_MAX;
1984         ct->timeout = (u32)timeout + nfct_time_stamp;
1985
1986         rcu_read_lock();
1987         if (cda[CTA_HELP]) {
1988                 char *helpname = NULL;
1989                 struct nlattr *helpinfo = NULL;
1990
1991                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1992                 if (err < 0)
1993                         goto err2;
1994
1995                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1996                                                     nf_ct_protonum(ct));
1997                 if (helper == NULL) {
1998                         rcu_read_unlock();
1999 #ifdef CONFIG_MODULES
2000                         if (request_module("nfct-helper-%s", helpname) < 0) {
2001                                 err = -EOPNOTSUPP;
2002                                 goto err1;
2003                         }
2004
2005                         rcu_read_lock();
2006                         helper = __nf_conntrack_helper_find(helpname,
2007                                                             nf_ct_l3num(ct),
2008                                                             nf_ct_protonum(ct));
2009                         if (helper) {
2010                                 err = -EAGAIN;
2011                                 goto err2;
2012                         }
2013                         rcu_read_unlock();
2014 #endif
2015                         err = -EOPNOTSUPP;
2016                         goto err1;
2017                 } else {
2018                         struct nf_conn_help *help;
2019
2020                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2021                         if (help == NULL) {
2022                                 err = -ENOMEM;
2023                                 goto err2;
2024                         }
2025                         /* set private helper data if allowed. */
2026                         if (helper->from_nlattr)
2027                                 helper->from_nlattr(helpinfo, ct);
2028
2029                         /* not in hash table yet so not strictly necessary */
2030                         RCU_INIT_POINTER(help->helper, helper);
2031                 }
2032         } else {
2033                 /* try an implicit helper assignation */
2034                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2035                 if (err < 0)
2036                         goto err2;
2037         }
2038
2039         err = ctnetlink_setup_nat(ct, cda);
2040         if (err < 0)
2041                 goto err2;
2042
2043         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2044         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2045         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2046         nf_ct_labels_ext_add(ct);
2047         nfct_seqadj_ext_add(ct);
2048         nfct_synproxy_ext_add(ct);
2049
2050         /* we must add conntrack extensions before confirmation. */
2051         ct->status |= IPS_CONFIRMED;
2052
2053         if (cda[CTA_STATUS]) {
2054                 err = ctnetlink_change_status(ct, cda);
2055                 if (err < 0)
2056                         goto err2;
2057         }
2058
2059         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2060                 err = ctnetlink_change_seq_adj(ct, cda);
2061                 if (err < 0)
2062                         goto err2;
2063         }
2064
2065         memset(&ct->proto, 0, sizeof(ct->proto));
2066         if (cda[CTA_PROTOINFO]) {
2067                 err = ctnetlink_change_protoinfo(ct, cda);
2068                 if (err < 0)
2069                         goto err2;
2070         }
2071
2072         if (cda[CTA_SYNPROXY]) {
2073                 err = ctnetlink_change_synproxy(ct, cda);
2074                 if (err < 0)
2075                         goto err2;
2076         }
2077
2078 #if defined(CONFIG_NF_CONNTRACK_MARK)
2079         if (cda[CTA_MARK])
2080                 ctnetlink_change_mark(ct, cda);
2081 #endif
2082
2083         /* setup master conntrack: this is a confirmed expectation */
2084         if (cda[CTA_TUPLE_MASTER]) {
2085                 struct nf_conntrack_tuple master;
2086                 struct nf_conntrack_tuple_hash *master_h;
2087                 struct nf_conn *master_ct;
2088
2089                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2090                                             u3, NULL);
2091                 if (err < 0)
2092                         goto err2;
2093
2094                 master_h = nf_conntrack_find_get(net, zone, &master);
2095                 if (master_h == NULL) {
2096                         err = -ENOENT;
2097                         goto err2;
2098                 }
2099                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2100                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2101                 ct->master = master_ct;
2102         }
2103         tstamp = nf_conn_tstamp_find(ct);
2104         if (tstamp)
2105                 tstamp->start = ktime_get_real_ns();
2106
2107         err = nf_conntrack_hash_check_insert(ct);
2108         if (err < 0)
2109                 goto err2;
2110
2111         rcu_read_unlock();
2112
2113         return ct;
2114
2115 err2:
2116         rcu_read_unlock();
2117 err1:
2118         nf_conntrack_free(ct);
2119         return ERR_PTR(err);
2120 }
2121
2122 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2123                                    struct sk_buff *skb,
2124                                    const struct nlmsghdr *nlh,
2125                                    const struct nlattr * const cda[],
2126                                    struct netlink_ext_ack *extack)
2127 {
2128         struct nf_conntrack_tuple otuple, rtuple;
2129         struct nf_conntrack_tuple_hash *h = NULL;
2130         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2131         struct nf_conn *ct;
2132         u_int8_t u3 = nfmsg->nfgen_family;
2133         struct nf_conntrack_zone zone;
2134         int err;
2135
2136         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2137         if (err < 0)
2138                 return err;
2139
2140         if (cda[CTA_TUPLE_ORIG]) {
2141                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2142                                             u3, &zone);
2143                 if (err < 0)
2144                         return err;
2145         }
2146
2147         if (cda[CTA_TUPLE_REPLY]) {
2148                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2149                                             u3, &zone);
2150                 if (err < 0)
2151                         return err;
2152         }
2153
2154         if (cda[CTA_TUPLE_ORIG])
2155                 h = nf_conntrack_find_get(net, &zone, &otuple);
2156         else if (cda[CTA_TUPLE_REPLY])
2157                 h = nf_conntrack_find_get(net, &zone, &rtuple);
2158
2159         if (h == NULL) {
2160                 err = -ENOENT;
2161                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2162                         enum ip_conntrack_events events;
2163
2164                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2165                                 return -EINVAL;
2166                         if (otuple.dst.protonum != rtuple.dst.protonum)
2167                                 return -EINVAL;
2168
2169                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2170                                                         &rtuple, u3);
2171                         if (IS_ERR(ct))
2172                                 return PTR_ERR(ct);
2173
2174                         err = 0;
2175                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2176                                 events = 1 << IPCT_RELATED;
2177                         else
2178                                 events = 1 << IPCT_NEW;
2179
2180                         if (cda[CTA_LABELS] &&
2181                             ctnetlink_attach_labels(ct, cda) == 0)
2182                                 events |= (1 << IPCT_LABEL);
2183
2184                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2185                                                       (1 << IPCT_ASSURED) |
2186                                                       (1 << IPCT_HELPER) |
2187                                                       (1 << IPCT_PROTOINFO) |
2188                                                       (1 << IPCT_SEQADJ) |
2189                                                       (1 << IPCT_MARK) |
2190                                                       (1 << IPCT_SYNPROXY) |
2191                                                       events,
2192                                                       ct, NETLINK_CB(skb).portid,
2193                                                       nlmsg_report(nlh));
2194                         nf_ct_put(ct);
2195                 }
2196
2197                 return err;
2198         }
2199         /* implicit 'else' */
2200
2201         err = -EEXIST;
2202         ct = nf_ct_tuplehash_to_ctrack(h);
2203         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2204                 err = ctnetlink_change_conntrack(ct, cda);
2205                 if (err == 0) {
2206                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2207                                                       (1 << IPCT_ASSURED) |
2208                                                       (1 << IPCT_HELPER) |
2209                                                       (1 << IPCT_LABEL) |
2210                                                       (1 << IPCT_PROTOINFO) |
2211                                                       (1 << IPCT_SEQADJ) |
2212                                                       (1 << IPCT_MARK) |
2213                                                       (1 << IPCT_SYNPROXY),
2214                                                       ct, NETLINK_CB(skb).portid,
2215                                                       nlmsg_report(nlh));
2216                 }
2217         }
2218
2219         nf_ct_put(ct);
2220         return err;
2221 }
2222
2223 static int
2224 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2225                                 __u16 cpu, const struct ip_conntrack_stat *st)
2226 {
2227         struct nlmsghdr *nlh;
2228         struct nfgenmsg *nfmsg;
2229         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2230
2231         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2232                               IPCTNL_MSG_CT_GET_STATS_CPU);
2233         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2234         if (nlh == NULL)
2235                 goto nlmsg_failure;
2236
2237         nfmsg = nlmsg_data(nlh);
2238         nfmsg->nfgen_family = AF_UNSPEC;
2239         nfmsg->version      = NFNETLINK_V0;
2240         nfmsg->res_id       = htons(cpu);
2241
2242         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2243             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2244             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2245             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2246             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2247                                 htonl(st->insert_failed)) ||
2248             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2249             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2250             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2251             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2252                                 htonl(st->search_restart)))
2253                 goto nla_put_failure;
2254
2255         nlmsg_end(skb, nlh);
2256         return skb->len;
2257
2258 nla_put_failure:
2259 nlmsg_failure:
2260         nlmsg_cancel(skb, nlh);
2261         return -1;
2262 }
2263
2264 static int
2265 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2266 {
2267         int cpu;
2268         struct net *net = sock_net(skb->sk);
2269
2270         if (cb->args[0] == nr_cpu_ids)
2271                 return 0;
2272
2273         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2274                 const struct ip_conntrack_stat *st;
2275
2276                 if (!cpu_possible(cpu))
2277                         continue;
2278
2279                 st = per_cpu_ptr(net->ct.stat, cpu);
2280                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2281                                                     NETLINK_CB(cb->skb).portid,
2282                                                     cb->nlh->nlmsg_seq,
2283                                                     cpu, st) < 0)
2284                                 break;
2285         }
2286         cb->args[0] = cpu;
2287
2288         return skb->len;
2289 }
2290
2291 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2292                                  struct sk_buff *skb,
2293                                  const struct nlmsghdr *nlh,
2294                                  const struct nlattr * const cda[],
2295                                  struct netlink_ext_ack *extack)
2296 {
2297         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2298                 struct netlink_dump_control c = {
2299                         .dump = ctnetlink_ct_stat_cpu_dump,
2300                 };
2301                 return netlink_dump_start(ctnl, skb, nlh, &c);
2302         }
2303
2304         return 0;
2305 }
2306
2307 static int
2308 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2309                             struct net *net)
2310 {
2311         struct nlmsghdr *nlh;
2312         struct nfgenmsg *nfmsg;
2313         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2314         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2315
2316         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2317         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2318         if (nlh == NULL)
2319                 goto nlmsg_failure;
2320
2321         nfmsg = nlmsg_data(nlh);
2322         nfmsg->nfgen_family = AF_UNSPEC;
2323         nfmsg->version      = NFNETLINK_V0;
2324         nfmsg->res_id       = 0;
2325
2326         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2327                 goto nla_put_failure;
2328
2329         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2330                 goto nla_put_failure;
2331
2332         nlmsg_end(skb, nlh);
2333         return skb->len;
2334
2335 nla_put_failure:
2336 nlmsg_failure:
2337         nlmsg_cancel(skb, nlh);
2338         return -1;
2339 }
2340
2341 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2342                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2343                              const struct nlattr * const cda[],
2344                              struct netlink_ext_ack *extack)
2345 {
2346         struct sk_buff *skb2;
2347         int err;
2348
2349         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2350         if (skb2 == NULL)
2351                 return -ENOMEM;
2352
2353         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2354                                           nlh->nlmsg_seq,
2355                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2356                                           sock_net(skb->sk));
2357         if (err <= 0)
2358                 goto free;
2359
2360         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2361         if (err < 0)
2362                 goto out;
2363
2364         return 0;
2365
2366 free:
2367         kfree_skb(skb2);
2368 out:
2369         /* this avoids a loop in nfnetlink. */
2370         return err == -EAGAIN ? -ENOBUFS : err;
2371 }
2372
2373 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2374         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2375         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2376         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2377         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2378         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2379         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2380                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2381         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2382         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2383         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2384         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2385         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2386 };
2387
2388 static struct nf_conntrack_expect *
2389 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2390                        struct nf_conntrack_helper *helper,
2391                        struct nf_conntrack_tuple *tuple,
2392                        struct nf_conntrack_tuple *mask);
2393
2394 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2395 static size_t
2396 ctnetlink_glue_build_size(const struct nf_conn *ct)
2397 {
2398         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2399                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2400                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2401                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2402                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2403                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2404                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2405                + nla_total_size(0) /* CTA_PROTOINFO */
2406                + nla_total_size(0) /* CTA_HELP */
2407                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2408                + ctnetlink_secctx_size(ct)
2409 #if IS_ENABLED(CONFIG_NF_NAT)
2410                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2411                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2412 #endif
2413 #ifdef CONFIG_NF_CONNTRACK_MARK
2414                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2415 #endif
2416 #ifdef CONFIG_NF_CONNTRACK_ZONES
2417                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2418 #endif
2419                + ctnetlink_proto_size(ct)
2420                ;
2421 }
2422
2423 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2424                                              enum ip_conntrack_info *ctinfo)
2425 {
2426         return nf_ct_get(skb, ctinfo);
2427 }
2428
2429 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2430 {
2431         const struct nf_conntrack_zone *zone;
2432         struct nlattr *nest_parms;
2433
2434         zone = nf_ct_zone(ct);
2435
2436         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2437         if (!nest_parms)
2438                 goto nla_put_failure;
2439         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2440                 goto nla_put_failure;
2441         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2442                                    NF_CT_ZONE_DIR_ORIG) < 0)
2443                 goto nla_put_failure;
2444         nla_nest_end(skb, nest_parms);
2445
2446         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2447         if (!nest_parms)
2448                 goto nla_put_failure;
2449         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2450                 goto nla_put_failure;
2451         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2452                                    NF_CT_ZONE_DIR_REPL) < 0)
2453                 goto nla_put_failure;
2454         nla_nest_end(skb, nest_parms);
2455
2456         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2457                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2458                 goto nla_put_failure;
2459
2460         if (ctnetlink_dump_id(skb, ct) < 0)
2461                 goto nla_put_failure;
2462
2463         if (ctnetlink_dump_status(skb, ct) < 0)
2464                 goto nla_put_failure;
2465
2466         if (ctnetlink_dump_timeout(skb, ct) < 0)
2467                 goto nla_put_failure;
2468
2469         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2470                 goto nla_put_failure;
2471
2472         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2473                 goto nla_put_failure;
2474
2475 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2476         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2477                 goto nla_put_failure;
2478 #endif
2479         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2480                 goto nla_put_failure;
2481
2482         if ((ct->status & IPS_SEQ_ADJUST) &&
2483             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2484                 goto nla_put_failure;
2485
2486         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2487                 goto nla_put_failure;
2488
2489 #ifdef CONFIG_NF_CONNTRACK_MARK
2490         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2491                 goto nla_put_failure;
2492 #endif
2493         if (ctnetlink_dump_labels(skb, ct) < 0)
2494                 goto nla_put_failure;
2495         return 0;
2496
2497 nla_put_failure:
2498         return -ENOSPC;
2499 }
2500
2501 static int
2502 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2503                      enum ip_conntrack_info ctinfo,
2504                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2505 {
2506         struct nlattr *nest_parms;
2507
2508         nest_parms = nla_nest_start(skb, ct_attr);
2509         if (!nest_parms)
2510                 goto nla_put_failure;
2511
2512         if (__ctnetlink_glue_build(skb, ct) < 0)
2513                 goto nla_put_failure;
2514
2515         nla_nest_end(skb, nest_parms);
2516
2517         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2518                 goto nla_put_failure;
2519
2520         return 0;
2521
2522 nla_put_failure:
2523         return -ENOSPC;
2524 }
2525
2526 static int
2527 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2528 {
2529         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2530         unsigned long d = ct->status ^ status;
2531
2532         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2533                 /* SEEN_REPLY bit can only be set */
2534                 return -EBUSY;
2535
2536         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2537                 /* ASSURED bit can only be set */
2538                 return -EBUSY;
2539
2540         /* This check is less strict than ctnetlink_change_status()
2541          * because callers often flip IPS_EXPECTED bits when sending
2542          * an NFQA_CT attribute to the kernel.  So ignore the
2543          * unchangeable bits but do not error out. Also user programs
2544          * are allowed to clear the bits that they are allowed to change.
2545          */
2546         __ctnetlink_change_status(ct, status, ~status);
2547         return 0;
2548 }
2549
2550 static int
2551 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2552 {
2553         int err;
2554
2555         if (cda[CTA_TIMEOUT]) {
2556                 err = ctnetlink_change_timeout(ct, cda);
2557                 if (err < 0)
2558                         return err;
2559         }
2560         if (cda[CTA_STATUS]) {
2561                 err = ctnetlink_update_status(ct, cda);
2562                 if (err < 0)
2563                         return err;
2564         }
2565         if (cda[CTA_HELP]) {
2566                 err = ctnetlink_change_helper(ct, cda);
2567                 if (err < 0)
2568                         return err;
2569         }
2570         if (cda[CTA_LABELS]) {
2571                 err = ctnetlink_attach_labels(ct, cda);
2572                 if (err < 0)
2573                         return err;
2574         }
2575 #if defined(CONFIG_NF_CONNTRACK_MARK)
2576         if (cda[CTA_MARK]) {
2577                 ctnetlink_change_mark(ct, cda);
2578         }
2579 #endif
2580         return 0;
2581 }
2582
2583 static int
2584 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2585 {
2586         struct nlattr *cda[CTA_MAX+1];
2587         int ret;
2588
2589         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2590                                           NULL);
2591         if (ret < 0)
2592                 return ret;
2593
2594         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2595 }
2596
2597 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2598                                     const struct nf_conn *ct,
2599                                     struct nf_conntrack_tuple *tuple,
2600                                     struct nf_conntrack_tuple *mask)
2601 {
2602         int err;
2603
2604         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2605                                     nf_ct_l3num(ct), NULL);
2606         if (err < 0)
2607                 return err;
2608
2609         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2610                                      nf_ct_l3num(ct), NULL);
2611 }
2612
2613 static int
2614 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2615                              u32 portid, u32 report)
2616 {
2617         struct nlattr *cda[CTA_EXPECT_MAX+1];
2618         struct nf_conntrack_tuple tuple, mask;
2619         struct nf_conntrack_helper *helper = NULL;
2620         struct nf_conntrack_expect *exp;
2621         int err;
2622
2623         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2624                                           exp_nla_policy, NULL);
2625         if (err < 0)
2626                 return err;
2627
2628         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2629                                        ct, &tuple, &mask);
2630         if (err < 0)
2631                 return err;
2632
2633         if (cda[CTA_EXPECT_HELP_NAME]) {
2634                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2635
2636                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2637                                                     nf_ct_protonum(ct));
2638                 if (helper == NULL)
2639                         return -EOPNOTSUPP;
2640         }
2641
2642         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2643                                      helper, &tuple, &mask);
2644         if (IS_ERR(exp))
2645                 return PTR_ERR(exp);
2646
2647         err = nf_ct_expect_related_report(exp, portid, report, 0);
2648         nf_ct_expect_put(exp);
2649         return err;
2650 }
2651
2652 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2653                                   enum ip_conntrack_info ctinfo, int diff)
2654 {
2655         if (!(ct->status & IPS_NAT_MASK))
2656                 return;
2657
2658         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2659 }
2660
2661 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2662         .get_ct         = ctnetlink_glue_get_ct,
2663         .build_size     = ctnetlink_glue_build_size,
2664         .build          = ctnetlink_glue_build,
2665         .parse          = ctnetlink_glue_parse,
2666         .attach_expect  = ctnetlink_glue_attach_expect,
2667         .seq_adjust     = ctnetlink_glue_seqadj,
2668 };
2669 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2670
2671 /***********************************************************************
2672  * EXPECT
2673  ***********************************************************************/
2674
2675 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2676                                     const struct nf_conntrack_tuple *tuple,
2677                                     u32 type)
2678 {
2679         struct nlattr *nest_parms;
2680
2681         nest_parms = nla_nest_start(skb, type);
2682         if (!nest_parms)
2683                 goto nla_put_failure;
2684         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2685                 goto nla_put_failure;
2686         nla_nest_end(skb, nest_parms);
2687
2688         return 0;
2689
2690 nla_put_failure:
2691         return -1;
2692 }
2693
2694 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2695                                    const struct nf_conntrack_tuple *tuple,
2696                                    const struct nf_conntrack_tuple_mask *mask)
2697 {
2698         const struct nf_conntrack_l4proto *l4proto;
2699         struct nf_conntrack_tuple m;
2700         struct nlattr *nest_parms;
2701         int ret;
2702
2703         memset(&m, 0xFF, sizeof(m));
2704         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2705         m.src.u.all = mask->src.u.all;
2706         m.dst.protonum = tuple->dst.protonum;
2707
2708         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2709         if (!nest_parms)
2710                 goto nla_put_failure;
2711
2712         rcu_read_lock();
2713         ret = ctnetlink_dump_tuples_ip(skb, &m);
2714         if (ret >= 0) {
2715                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2716                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2717         }
2718         rcu_read_unlock();
2719
2720         if (unlikely(ret < 0))
2721                 goto nla_put_failure;
2722
2723         nla_nest_end(skb, nest_parms);
2724
2725         return 0;
2726
2727 nla_put_failure:
2728         return -1;
2729 }
2730
2731 static const union nf_inet_addr any_addr;
2732
2733 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2734 {
2735         static __read_mostly siphash_key_t exp_id_seed;
2736         unsigned long a, b, c, d;
2737
2738         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2739
2740         a = (unsigned long)exp;
2741         b = (unsigned long)exp->helper;
2742         c = (unsigned long)exp->master;
2743         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2744
2745 #ifdef CONFIG_64BIT
2746         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2747 #else
2748         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2749 #endif
2750 }
2751
2752 static int
2753 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2754                           const struct nf_conntrack_expect *exp)
2755 {
2756         struct nf_conn *master = exp->master;
2757         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2758         struct nf_conn_help *help;
2759 #if IS_ENABLED(CONFIG_NF_NAT)
2760         struct nlattr *nest_parms;
2761         struct nf_conntrack_tuple nat_tuple = {};
2762 #endif
2763         struct nf_ct_helper_expectfn *expfn;
2764
2765         if (timeout < 0)
2766                 timeout = 0;
2767
2768         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2769                 goto nla_put_failure;
2770         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2771                 goto nla_put_failure;
2772         if (ctnetlink_exp_dump_tuple(skb,
2773                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2774                                  CTA_EXPECT_MASTER) < 0)
2775                 goto nla_put_failure;
2776
2777 #if IS_ENABLED(CONFIG_NF_NAT)
2778         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2779             exp->saved_proto.all) {
2780                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
2781                 if (!nest_parms)
2782                         goto nla_put_failure;
2783
2784                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2785                         goto nla_put_failure;
2786
2787                 nat_tuple.src.l3num = nf_ct_l3num(master);
2788                 nat_tuple.src.u3 = exp->saved_addr;
2789                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2790                 nat_tuple.src.u = exp->saved_proto;
2791
2792                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2793                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2794                         goto nla_put_failure;
2795                 nla_nest_end(skb, nest_parms);
2796         }
2797 #endif
2798         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2799             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
2800             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2801             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2802                 goto nla_put_failure;
2803         help = nfct_help(master);
2804         if (help) {
2805                 struct nf_conntrack_helper *helper;
2806
2807                 helper = rcu_dereference(help->helper);
2808                 if (helper &&
2809                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2810                         goto nla_put_failure;
2811         }
2812         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2813         if (expfn != NULL &&
2814             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2815                 goto nla_put_failure;
2816
2817         return 0;
2818
2819 nla_put_failure:
2820         return -1;
2821 }
2822
2823 static int
2824 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2825                         int event, const struct nf_conntrack_expect *exp)
2826 {
2827         struct nlmsghdr *nlh;
2828         struct nfgenmsg *nfmsg;
2829         unsigned int flags = portid ? NLM_F_MULTI : 0;
2830
2831         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2832         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2833         if (nlh == NULL)
2834                 goto nlmsg_failure;
2835
2836         nfmsg = nlmsg_data(nlh);
2837         nfmsg->nfgen_family = exp->tuple.src.l3num;
2838         nfmsg->version      = NFNETLINK_V0;
2839         nfmsg->res_id       = 0;
2840
2841         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2842                 goto nla_put_failure;
2843
2844         nlmsg_end(skb, nlh);
2845         return skb->len;
2846
2847 nlmsg_failure:
2848 nla_put_failure:
2849         nlmsg_cancel(skb, nlh);
2850         return -1;
2851 }
2852
2853 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2854 static int
2855 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2856 {
2857         struct nf_conntrack_expect *exp = item->exp;
2858         struct net *net = nf_ct_exp_net(exp);
2859         struct nlmsghdr *nlh;
2860         struct nfgenmsg *nfmsg;
2861         struct sk_buff *skb;
2862         unsigned int type, group;
2863         int flags = 0;
2864
2865         if (events & (1 << IPEXP_DESTROY)) {
2866                 type = IPCTNL_MSG_EXP_DELETE;
2867                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2868         } else if (events & (1 << IPEXP_NEW)) {
2869                 type = IPCTNL_MSG_EXP_NEW;
2870                 flags = NLM_F_CREATE|NLM_F_EXCL;
2871                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2872         } else
2873                 return 0;
2874
2875         if (!item->report && !nfnetlink_has_listeners(net, group))
2876                 return 0;
2877
2878         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2879         if (skb == NULL)
2880                 goto errout;
2881
2882         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2883         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2884         if (nlh == NULL)
2885                 goto nlmsg_failure;
2886
2887         nfmsg = nlmsg_data(nlh);
2888         nfmsg->nfgen_family = exp->tuple.src.l3num;
2889         nfmsg->version      = NFNETLINK_V0;
2890         nfmsg->res_id       = 0;
2891
2892         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2893                 goto nla_put_failure;
2894
2895         nlmsg_end(skb, nlh);
2896         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2897         return 0;
2898
2899 nla_put_failure:
2900         nlmsg_cancel(skb, nlh);
2901 nlmsg_failure:
2902         kfree_skb(skb);
2903 errout:
2904         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2905         return 0;
2906 }
2907 #endif
2908 static int ctnetlink_exp_done(struct netlink_callback *cb)
2909 {
2910         if (cb->args[1])
2911                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2912         return 0;
2913 }
2914
2915 static int
2916 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2917 {
2918         struct net *net = sock_net(skb->sk);
2919         struct nf_conntrack_expect *exp, *last;
2920         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2921         u_int8_t l3proto = nfmsg->nfgen_family;
2922
2923         rcu_read_lock();
2924         last = (struct nf_conntrack_expect *)cb->args[1];
2925         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2926 restart:
2927                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2928                                          hnode) {
2929                         if (l3proto && exp->tuple.src.l3num != l3proto)
2930                                 continue;
2931
2932                         if (!net_eq(nf_ct_net(exp->master), net))
2933                                 continue;
2934
2935                         if (cb->args[1]) {
2936                                 if (exp != last)
2937                                         continue;
2938                                 cb->args[1] = 0;
2939                         }
2940                         if (ctnetlink_exp_fill_info(skb,
2941                                                     NETLINK_CB(cb->skb).portid,
2942                                                     cb->nlh->nlmsg_seq,
2943                                                     IPCTNL_MSG_EXP_NEW,
2944                                                     exp) < 0) {
2945                                 if (!refcount_inc_not_zero(&exp->use))
2946                                         continue;
2947                                 cb->args[1] = (unsigned long)exp;
2948                                 goto out;
2949                         }
2950                 }
2951                 if (cb->args[1]) {
2952                         cb->args[1] = 0;
2953                         goto restart;
2954                 }
2955         }
2956 out:
2957         rcu_read_unlock();
2958         if (last)
2959                 nf_ct_expect_put(last);
2960
2961         return skb->len;
2962 }
2963
2964 static int
2965 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2966 {
2967         struct nf_conntrack_expect *exp, *last;
2968         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2969         struct nf_conn *ct = cb->data;
2970         struct nf_conn_help *help = nfct_help(ct);
2971         u_int8_t l3proto = nfmsg->nfgen_family;
2972
2973         if (cb->args[0])
2974                 return 0;
2975
2976         rcu_read_lock();
2977         last = (struct nf_conntrack_expect *)cb->args[1];
2978 restart:
2979         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2980                 if (l3proto && exp->tuple.src.l3num != l3proto)
2981                         continue;
2982                 if (cb->args[1]) {
2983                         if (exp != last)
2984                                 continue;
2985                         cb->args[1] = 0;
2986                 }
2987                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2988                                             cb->nlh->nlmsg_seq,
2989                                             IPCTNL_MSG_EXP_NEW,
2990                                             exp) < 0) {
2991                         if (!refcount_inc_not_zero(&exp->use))
2992                                 continue;
2993                         cb->args[1] = (unsigned long)exp;
2994                         goto out;
2995                 }
2996         }
2997         if (cb->args[1]) {
2998                 cb->args[1] = 0;
2999                 goto restart;
3000         }
3001         cb->args[0] = 1;
3002 out:
3003         rcu_read_unlock();
3004         if (last)
3005                 nf_ct_expect_put(last);
3006
3007         return skb->len;
3008 }
3009
3010 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3011                                  struct sk_buff *skb,
3012                                  const struct nlmsghdr *nlh,
3013                                  const struct nlattr * const cda[],
3014                                  struct netlink_ext_ack *extack)
3015 {
3016         int err;
3017         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3018         u_int8_t u3 = nfmsg->nfgen_family;
3019         struct nf_conntrack_tuple tuple;
3020         struct nf_conntrack_tuple_hash *h;
3021         struct nf_conn *ct;
3022         struct nf_conntrack_zone zone;
3023         struct netlink_dump_control c = {
3024                 .dump = ctnetlink_exp_ct_dump_table,
3025                 .done = ctnetlink_exp_done,
3026         };
3027
3028         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3029                                     u3, NULL);
3030         if (err < 0)
3031                 return err;
3032
3033         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3034         if (err < 0)
3035                 return err;
3036
3037         h = nf_conntrack_find_get(net, &zone, &tuple);
3038         if (!h)
3039                 return -ENOENT;
3040
3041         ct = nf_ct_tuplehash_to_ctrack(h);
3042         /* No expectation linked to this connection tracking. */
3043         if (!nfct_help(ct)) {
3044                 nf_ct_put(ct);
3045                 return 0;
3046         }
3047
3048         c.data = ct;
3049
3050         err = netlink_dump_start(ctnl, skb, nlh, &c);
3051         nf_ct_put(ct);
3052
3053         return err;
3054 }
3055
3056 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
3057                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3058                                 const struct nlattr * const cda[],
3059                                 struct netlink_ext_ack *extack)
3060 {
3061         struct nf_conntrack_tuple tuple;
3062         struct nf_conntrack_expect *exp;
3063         struct sk_buff *skb2;
3064         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3065         u_int8_t u3 = nfmsg->nfgen_family;
3066         struct nf_conntrack_zone zone;
3067         int err;
3068
3069         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3070                 if (cda[CTA_EXPECT_MASTER])
3071                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
3072                                                      extack);
3073                 else {
3074                         struct netlink_dump_control c = {
3075                                 .dump = ctnetlink_exp_dump_table,
3076                                 .done = ctnetlink_exp_done,
3077                         };
3078                         return netlink_dump_start(ctnl, skb, nlh, &c);
3079                 }
3080         }
3081
3082         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3083         if (err < 0)
3084                 return err;
3085
3086         if (cda[CTA_EXPECT_TUPLE])
3087                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3088                                             u3, NULL);
3089         else if (cda[CTA_EXPECT_MASTER])
3090                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3091                                             u3, NULL);
3092         else
3093                 return -EINVAL;
3094
3095         if (err < 0)
3096                 return err;
3097
3098         exp = nf_ct_expect_find_get(net, &zone, &tuple);
3099         if (!exp)
3100                 return -ENOENT;
3101
3102         if (cda[CTA_EXPECT_ID]) {
3103                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3104
3105                 if (id != nf_expect_get_id(exp)) {
3106                         nf_ct_expect_put(exp);
3107                         return -ENOENT;
3108                 }
3109         }
3110
3111         err = -ENOMEM;
3112         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3113         if (skb2 == NULL) {
3114                 nf_ct_expect_put(exp);
3115                 goto out;
3116         }
3117
3118         rcu_read_lock();
3119         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3120                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
3121         rcu_read_unlock();
3122         nf_ct_expect_put(exp);
3123         if (err <= 0)
3124                 goto free;
3125
3126         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
3127         if (err < 0)
3128                 goto out;
3129
3130         return 0;
3131
3132 free:
3133         kfree_skb(skb2);
3134 out:
3135         /* this avoids a loop in nfnetlink. */
3136         return err == -EAGAIN ? -ENOBUFS : err;
3137 }
3138
3139 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3140 {
3141         const struct nf_conn_help *m_help;
3142         const char *name = data;
3143
3144         m_help = nfct_help(exp->master);
3145
3146         return strcmp(m_help->helper->name, name) == 0;
3147 }
3148
3149 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3150 {
3151         return true;
3152 }
3153
3154 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3155                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3156                                 const struct nlattr * const cda[],
3157                                 struct netlink_ext_ack *extack)
3158 {
3159         struct nf_conntrack_expect *exp;
3160         struct nf_conntrack_tuple tuple;
3161         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3162         u_int8_t u3 = nfmsg->nfgen_family;
3163         struct nf_conntrack_zone zone;
3164         int err;
3165
3166         if (cda[CTA_EXPECT_TUPLE]) {
3167                 /* delete a single expect by tuple */
3168                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3169                 if (err < 0)
3170                         return err;
3171
3172                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3173                                             u3, NULL);
3174                 if (err < 0)
3175                         return err;
3176
3177                 /* bump usage count to 2 */
3178                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
3179                 if (!exp)
3180                         return -ENOENT;
3181
3182                 if (cda[CTA_EXPECT_ID]) {
3183                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3184                         if (ntohl(id) != (u32)(unsigned long)exp) {
3185                                 nf_ct_expect_put(exp);
3186                                 return -ENOENT;
3187                         }
3188                 }
3189
3190                 /* after list removal, usage count == 1 */
3191                 spin_lock_bh(&nf_conntrack_expect_lock);
3192                 if (del_timer(&exp->timeout)) {
3193                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3194                                                    nlmsg_report(nlh));
3195                         nf_ct_expect_put(exp);
3196                 }
3197                 spin_unlock_bh(&nf_conntrack_expect_lock);
3198                 /* have to put what we 'get' above.
3199                  * after this line usage count == 0 */
3200                 nf_ct_expect_put(exp);
3201         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3202                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3203
3204                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
3205                                          NETLINK_CB(skb).portid,
3206                                          nlmsg_report(nlh));
3207         } else {
3208                 /* This basically means we have to flush everything*/
3209                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3210                                          NETLINK_CB(skb).portid,
3211                                          nlmsg_report(nlh));
3212         }
3213
3214         return 0;
3215 }
3216 static int
3217 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3218                         const struct nlattr * const cda[])
3219 {
3220         if (cda[CTA_EXPECT_TIMEOUT]) {
3221                 if (!del_timer(&x->timeout))
3222                         return -ETIME;
3223
3224                 x->timeout.expires = jiffies +
3225                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3226                 add_timer(&x->timeout);
3227         }
3228         return 0;
3229 }
3230
3231 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3232         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3233         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3234 };
3235
3236 static int
3237 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3238                            struct nf_conntrack_expect *exp,
3239                            u_int8_t u3)
3240 {
3241 #if IS_ENABLED(CONFIG_NF_NAT)
3242         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3243         struct nf_conntrack_tuple nat_tuple = {};
3244         int err;
3245
3246         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3247                                           exp_nat_nla_policy, NULL);
3248         if (err < 0)
3249                 return err;
3250
3251         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3252                 return -EINVAL;
3253
3254         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3255                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3256                                     u3, NULL);
3257         if (err < 0)
3258                 return err;
3259
3260         exp->saved_addr = nat_tuple.src.u3;
3261         exp->saved_proto = nat_tuple.src.u;
3262         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3263
3264         return 0;
3265 #else
3266         return -EOPNOTSUPP;
3267 #endif
3268 }
3269
3270 static struct nf_conntrack_expect *
3271 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3272                        struct nf_conntrack_helper *helper,
3273                        struct nf_conntrack_tuple *tuple,
3274                        struct nf_conntrack_tuple *mask)
3275 {
3276         u_int32_t class = 0;
3277         struct nf_conntrack_expect *exp;
3278         struct nf_conn_help *help;
3279         int err;
3280
3281         help = nfct_help(ct);
3282         if (!help)
3283                 return ERR_PTR(-EOPNOTSUPP);
3284
3285         if (cda[CTA_EXPECT_CLASS] && helper) {
3286                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3287                 if (class > helper->expect_class_max)
3288                         return ERR_PTR(-EINVAL);
3289         }
3290         exp = nf_ct_expect_alloc(ct);
3291         if (!exp)
3292                 return ERR_PTR(-ENOMEM);
3293
3294         if (cda[CTA_EXPECT_FLAGS]) {
3295                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3296                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3297         } else {
3298                 exp->flags = 0;
3299         }
3300         if (cda[CTA_EXPECT_FN]) {
3301                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3302                 struct nf_ct_helper_expectfn *expfn;
3303
3304                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3305                 if (expfn == NULL) {
3306                         err = -EINVAL;
3307                         goto err_out;
3308                 }
3309                 exp->expectfn = expfn->expectfn;
3310         } else
3311                 exp->expectfn = NULL;
3312
3313         exp->class = class;
3314         exp->master = ct;
3315         exp->helper = helper;
3316         exp->tuple = *tuple;
3317         exp->mask.src.u3 = mask->src.u3;
3318         exp->mask.src.u.all = mask->src.u.all;
3319
3320         if (cda[CTA_EXPECT_NAT]) {
3321                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3322                                                  exp, nf_ct_l3num(ct));
3323                 if (err < 0)
3324                         goto err_out;
3325         }
3326         return exp;
3327 err_out:
3328         nf_ct_expect_put(exp);
3329         return ERR_PTR(err);
3330 }
3331
3332 static int
3333 ctnetlink_create_expect(struct net *net,
3334                         const struct nf_conntrack_zone *zone,
3335                         const struct nlattr * const cda[],
3336                         u_int8_t u3, u32 portid, int report)
3337 {
3338         struct nf_conntrack_tuple tuple, mask, master_tuple;
3339         struct nf_conntrack_tuple_hash *h = NULL;
3340         struct nf_conntrack_helper *helper = NULL;
3341         struct nf_conntrack_expect *exp;
3342         struct nf_conn *ct;
3343         int err;
3344
3345         /* caller guarantees that those three CTA_EXPECT_* exist */
3346         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3347                                     u3, NULL);
3348         if (err < 0)
3349                 return err;
3350         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3351                                     u3, NULL);
3352         if (err < 0)
3353                 return err;
3354         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3355                                     u3, NULL);
3356         if (err < 0)
3357                 return err;
3358
3359         /* Look for master conntrack of this expectation */
3360         h = nf_conntrack_find_get(net, zone, &master_tuple);
3361         if (!h)
3362                 return -ENOENT;
3363         ct = nf_ct_tuplehash_to_ctrack(h);
3364
3365         rcu_read_lock();
3366         if (cda[CTA_EXPECT_HELP_NAME]) {
3367                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3368
3369                 helper = __nf_conntrack_helper_find(helpname, u3,
3370                                                     nf_ct_protonum(ct));
3371                 if (helper == NULL) {
3372                         rcu_read_unlock();
3373 #ifdef CONFIG_MODULES
3374                         if (request_module("nfct-helper-%s", helpname) < 0) {
3375                                 err = -EOPNOTSUPP;
3376                                 goto err_ct;
3377                         }
3378                         rcu_read_lock();
3379                         helper = __nf_conntrack_helper_find(helpname, u3,
3380                                                             nf_ct_protonum(ct));
3381                         if (helper) {
3382                                 err = -EAGAIN;
3383                                 goto err_rcu;
3384                         }
3385                         rcu_read_unlock();
3386 #endif
3387                         err = -EOPNOTSUPP;
3388                         goto err_ct;
3389                 }
3390         }
3391
3392         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3393         if (IS_ERR(exp)) {
3394                 err = PTR_ERR(exp);
3395                 goto err_rcu;
3396         }
3397
3398         err = nf_ct_expect_related_report(exp, portid, report, 0);
3399         nf_ct_expect_put(exp);
3400 err_rcu:
3401         rcu_read_unlock();
3402 err_ct:
3403         nf_ct_put(ct);
3404         return err;
3405 }
3406
3407 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3408                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3409                                 const struct nlattr * const cda[],
3410                                 struct netlink_ext_ack *extack)
3411 {
3412         struct nf_conntrack_tuple tuple;
3413         struct nf_conntrack_expect *exp;
3414         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3415         u_int8_t u3 = nfmsg->nfgen_family;
3416         struct nf_conntrack_zone zone;
3417         int err;
3418
3419         if (!cda[CTA_EXPECT_TUPLE]
3420             || !cda[CTA_EXPECT_MASK]
3421             || !cda[CTA_EXPECT_MASTER])
3422                 return -EINVAL;
3423
3424         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3425         if (err < 0)
3426                 return err;
3427
3428         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3429                                     u3, NULL);
3430         if (err < 0)
3431                 return err;
3432
3433         spin_lock_bh(&nf_conntrack_expect_lock);
3434         exp = __nf_ct_expect_find(net, &zone, &tuple);
3435         if (!exp) {
3436                 spin_unlock_bh(&nf_conntrack_expect_lock);
3437                 err = -ENOENT;
3438                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3439                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3440                                                       NETLINK_CB(skb).portid,
3441                                                       nlmsg_report(nlh));
3442                 }
3443                 return err;
3444         }
3445
3446         err = -EEXIST;
3447         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3448                 err = ctnetlink_change_expect(exp, cda);
3449         spin_unlock_bh(&nf_conntrack_expect_lock);
3450
3451         return err;
3452 }
3453
3454 static int
3455 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3456                              const struct ip_conntrack_stat *st)
3457 {
3458         struct nlmsghdr *nlh;
3459         struct nfgenmsg *nfmsg;
3460         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3461
3462         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3463                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3464         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3465         if (nlh == NULL)
3466                 goto nlmsg_failure;
3467
3468         nfmsg = nlmsg_data(nlh);
3469         nfmsg->nfgen_family = AF_UNSPEC;
3470         nfmsg->version      = NFNETLINK_V0;
3471         nfmsg->res_id       = htons(cpu);
3472
3473         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3474             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3475             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3476                 goto nla_put_failure;
3477
3478         nlmsg_end(skb, nlh);
3479         return skb->len;
3480
3481 nla_put_failure:
3482 nlmsg_failure:
3483         nlmsg_cancel(skb, nlh);
3484         return -1;
3485 }
3486
3487 static int
3488 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3489 {
3490         int cpu;
3491         struct net *net = sock_net(skb->sk);
3492
3493         if (cb->args[0] == nr_cpu_ids)
3494                 return 0;
3495
3496         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3497                 const struct ip_conntrack_stat *st;
3498
3499                 if (!cpu_possible(cpu))
3500                         continue;
3501
3502                 st = per_cpu_ptr(net->ct.stat, cpu);
3503                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3504                                                  cb->nlh->nlmsg_seq,
3505                                                  cpu, st) < 0)
3506                         break;
3507         }
3508         cb->args[0] = cpu;
3509
3510         return skb->len;
3511 }
3512
3513 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3514                                   struct sk_buff *skb,
3515                                   const struct nlmsghdr *nlh,
3516                                   const struct nlattr * const cda[],
3517                                   struct netlink_ext_ack *extack)
3518 {
3519         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3520                 struct netlink_dump_control c = {
3521                         .dump = ctnetlink_exp_stat_cpu_dump,
3522                 };
3523                 return netlink_dump_start(ctnl, skb, nlh, &c);
3524         }
3525
3526         return 0;
3527 }
3528
3529 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3530 static struct nf_ct_event_notifier ctnl_notifier = {
3531         .fcn = ctnetlink_conntrack_event,
3532 };
3533
3534 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3535         .fcn = ctnetlink_expect_event,
3536 };
3537 #endif
3538
3539 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3540         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3541                                             .attr_count = CTA_MAX,
3542                                             .policy = ct_nla_policy },
3543         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3544                                             .attr_count = CTA_MAX,
3545                                             .policy = ct_nla_policy },
3546         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3547                                             .attr_count = CTA_MAX,
3548                                             .policy = ct_nla_policy },
3549         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3550                                             .attr_count = CTA_MAX,
3551                                             .policy = ct_nla_policy },
3552         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3553         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3554         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3555         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3556 };
3557
3558 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3559         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3560                                             .attr_count = CTA_EXPECT_MAX,
3561                                             .policy = exp_nla_policy },
3562         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3563                                             .attr_count = CTA_EXPECT_MAX,
3564                                             .policy = exp_nla_policy },
3565         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3566                                             .attr_count = CTA_EXPECT_MAX,
3567                                             .policy = exp_nla_policy },
3568         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3569 };
3570
3571 static const struct nfnetlink_subsystem ctnl_subsys = {
3572         .name                           = "conntrack",
3573         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3574         .cb_count                       = IPCTNL_MSG_MAX,
3575         .cb                             = ctnl_cb,
3576 };
3577
3578 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3579         .name                           = "conntrack_expect",
3580         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3581         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3582         .cb                             = ctnl_exp_cb,
3583 };
3584
3585 MODULE_ALIAS("ip_conntrack_netlink");
3586 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3587 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3588
3589 static int __net_init ctnetlink_net_init(struct net *net)
3590 {
3591 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3592         int ret;
3593
3594         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3595         if (ret < 0) {
3596                 pr_err("ctnetlink_init: cannot register notifier.\n");
3597                 goto err_out;
3598         }
3599
3600         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3601         if (ret < 0) {
3602                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3603                 goto err_unreg_notifier;
3604         }
3605 #endif
3606         return 0;
3607
3608 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3609 err_unreg_notifier:
3610         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3611 err_out:
3612         return ret;
3613 #endif
3614 }
3615
3616 static void ctnetlink_net_exit(struct net *net)
3617 {
3618 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3619         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3620         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3621 #endif
3622 }
3623
3624 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3625 {
3626         struct net *net;
3627
3628         list_for_each_entry(net, net_exit_list, exit_list)
3629                 ctnetlink_net_exit(net);
3630
3631         /* wait for other cpus until they are done with ctnl_notifiers */
3632         synchronize_rcu();
3633 }
3634
3635 static struct pernet_operations ctnetlink_net_ops = {
3636         .init           = ctnetlink_net_init,
3637         .exit_batch     = ctnetlink_net_exit_batch,
3638 };
3639
3640 static int __init ctnetlink_init(void)
3641 {
3642         int ret;
3643
3644         ret = nfnetlink_subsys_register(&ctnl_subsys);
3645         if (ret < 0) {
3646                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3647                 goto err_out;
3648         }
3649
3650         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3651         if (ret < 0) {
3652                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3653                 goto err_unreg_subsys;
3654         }
3655
3656         ret = register_pernet_subsys(&ctnetlink_net_ops);
3657         if (ret < 0) {
3658                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3659                 goto err_unreg_exp_subsys;
3660         }
3661 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3662         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3663         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3664 #endif
3665         return 0;
3666
3667 err_unreg_exp_subsys:
3668         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3669 err_unreg_subsys:
3670         nfnetlink_subsys_unregister(&ctnl_subsys);
3671 err_out:
3672         return ret;
3673 }
3674
3675 static void __exit ctnetlink_exit(void)
3676 {
3677         unregister_pernet_subsys(&ctnetlink_net_ops);
3678         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3679         nfnetlink_subsys_unregister(&ctnl_subsys);
3680 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3681         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3682 #endif
3683         synchronize_rcu();
3684 }
3685
3686 module_init(ctnetlink_init);
3687 module_exit(ctnetlink_exit);