net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_l3proto.h>
  41 #include <net/netfilter/nf_conntrack_l4proto.h>
  42 #include <net/netfilter/nf_conntrack_tuple.h>
  43 #include <net/netfilter/nf_conntrack_acct.h>
  44 #include <net/netfilter/nf_conntrack_zones.h>
  45 #include <net/netfilter/nf_conntrack_timestamp.h>
  46 #include <net/netfilter/nf_conntrack_labels.h>
  47 #ifdef CONFIG_NF_NAT_NEEDED
  48 #include <net/netfilter/nf_nat_core.h>
  49 #include <net/netfilter/nf_nat_l4proto.h>
  50 #include <net/netfilter/nf_nat_helper.h>
  51 #endif
  52
  53 #include <linux/netfilter/nfnetlink.h>
  54 #include <linux/netfilter/nfnetlink_conntrack.h>
  55
  56 MODULE_LICENSE("GPL");
  57
  58 static char __initdata version[] = "0.93";
  59
  60 static inline int
  61 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  62                             const struct nf_conntrack_tuple *tuple,
  63                             struct nf_conntrack_l4proto *l4proto)
  64 {
  65         int ret = 0;
  66         struct nlattr *nest_parms;
  67
  68         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  69         if (!nest_parms)
  70                 goto nla_put_failure;
  71         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  72                 goto nla_put_failure;
  73
  74         if (likely(l4proto->tuple_to_nlattr))
  75                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  76
  77         nla_nest_end(skb, nest_parms);
  78
  79         return ret;
  80
  81 nla_put_failure:
  82         return -1;
  83 }
  84
  85 static inline int
  86 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  87                          const struct nf_conntrack_tuple *tuple,
  88                          struct nf_conntrack_l3proto *l3proto)
  89 {
  90         int ret = 0;
  91         struct nlattr *nest_parms;
  92
  93         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  94         if (!nest_parms)
  95                 goto nla_put_failure;
  96
  97         if (likely(l3proto->tuple_to_nlattr))
  98                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  99
 100         nla_nest_end(skb, nest_parms);
 101
 102         return ret;
 103
 104 nla_put_failure:
 105         return -1;
 106 }
 107
 108 static int
 109 ctnetlink_dump_tuples(struct sk_buff *skb,
 110                       const struct nf_conntrack_tuple *tuple)
 111 {
 112         int ret;
 113         struct nf_conntrack_l3proto *l3proto;
 114         struct nf_conntrack_l4proto *l4proto;
 115
 116         rcu_read_lock();
 117         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 118         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 119
 120         if (ret >= 0) {
 121                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 122                                                tuple->dst.protonum);
 123                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 124         }
 125         rcu_read_unlock();
 126         return ret;
 127 }
 128
 129 static inline int
 130 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 131 {
 132         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 133                 goto nla_put_failure;
 134         return 0;
 135
 136 nla_put_failure:
 137         return -1;
 138 }
 139
 140 static inline int
 141 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 142 {
 143         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 144
 145         if (timeout < 0)
 146                 timeout = 0;
 147
 148         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 149                 goto nla_put_failure;
 150         return 0;
 151
 152 nla_put_failure:
 153         return -1;
 154 }
 155
 156 static inline int
 157 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 158 {
 159         struct nf_conntrack_l4proto *l4proto;
 160         struct nlattr *nest_proto;
 161         int ret;
 162
 163         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 164         if (!l4proto->to_nlattr)
 165                 return 0;
 166
 167         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 168         if (!nest_proto)
 169                 goto nla_put_failure;
 170
 171         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 172
 173         nla_nest_end(skb, nest_proto);
 174
 175         return ret;
 176
 177 nla_put_failure:
 178         return -1;
 179 }
 180
 181 static inline int
 182 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 183 {
 184         struct nlattr *nest_helper;
 185         const struct nf_conn_help *help = nfct_help(ct);
 186         struct nf_conntrack_helper *helper;
 187
 188         if (!help)
 189                 return 0;
 190
 191         helper = rcu_dereference(help->helper);
 192         if (!helper)
 193                 goto out;
 194
 195         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 196         if (!nest_helper)
 197                 goto nla_put_failure;
 198         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 199                 goto nla_put_failure;
 200
 201         if (helper->to_nlattr)
 202                 helper->to_nlattr(skb, ct);
 203
 204         nla_nest_end(skb, nest_helper);
 205 out:
 206         return 0;
 207
 208 nla_put_failure:
 209         return -1;
 210 }
 211
 212 static int
 213 dump_counters(struct sk_buff *skb, u64 pkts, u64 bytes,
 214               enum ip_conntrack_dir dir)
 215 {
 216         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 217         struct nlattr *nest_count;
 218
 219         nest_count = nla_nest_start(skb, type | NLA_F_NESTED);
 220         if (!nest_count)
 221                 goto nla_put_failure;
 222
 223         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 224             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 225                 goto nla_put_failure;
 226
 227         nla_nest_end(skb, nest_count);
 228
 229         return 0;
 230
 231 nla_put_failure:
 232         return -1;
 233 }
 234
 235 static int
 236 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 237                         enum ip_conntrack_dir dir, int type)
 238 {
 239         struct nf_conn_counter *acct;
 240         u64 pkts, bytes;
 241
 242         acct = nf_conn_acct_find(ct);
 243         if (!acct)
 244                 return 0;
 245
 246         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 247                 pkts = atomic64_xchg(&acct[dir].packets, 0);
 248                 bytes = atomic64_xchg(&acct[dir].bytes, 0);
 249         } else {
 250                 pkts = atomic64_read(&acct[dir].packets);
 251                 bytes = atomic64_read(&acct[dir].bytes);
 252         }
 253         return dump_counters(skb, pkts, bytes, dir);
 254 }
 255
 256 static int
 257 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 258 {
 259         struct nlattr *nest_count;
 260         const struct nf_conn_tstamp *tstamp;
 261
 262         tstamp = nf_conn_tstamp_find(ct);
 263         if (!tstamp)
 264                 return 0;
 265
 266         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 267         if (!nest_count)
 268                 goto nla_put_failure;
 269
 270         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 271             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 272                                                cpu_to_be64(tstamp->stop))))
 273                 goto nla_put_failure;
 274         nla_nest_end(skb, nest_count);
 275
 276         return 0;
 277
 278 nla_put_failure:
 279         return -1;
 280 }
 281
 282 #ifdef CONFIG_NF_CONNTRACK_MARK
 283 static inline int
 284 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 285 {
 286         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 287                 goto nla_put_failure;
 288         return 0;
 289
 290 nla_put_failure:
 291         return -1;
 292 }
 293 #else
 294 #define ctnetlink_dump_mark(a, b) (0)
 295 #endif
 296
 297 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 298 static inline int
 299 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 300 {
 301         struct nlattr *nest_secctx;
 302         int len, ret;
 303         char *secctx;
 304
 305         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 306         if (ret)
 307                 return 0;
 308
 309         ret = -1;
 310         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 311         if (!nest_secctx)
 312                 goto nla_put_failure;
 313
 314         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 315                 goto nla_put_failure;
 316         nla_nest_end(skb, nest_secctx);
 317
 318         ret = 0;
 319 nla_put_failure:
 320         security_release_secctx(secctx, len);
 321         return ret;
 322 }
 323 #else
 324 #define ctnetlink_dump_secctx(a, b) (0)
 325 #endif
 326
 327 #ifdef CONFIG_NF_CONNTRACK_LABELS
 328 static int ctnetlink_label_size(const struct nf_conn *ct)
 329 {
 330         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 331
 332         if (!labels)
 333                 return 0;
 334         return nla_total_size(labels->words * sizeof(long));
 335 }
 336
 337 static int
 338 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 339 {
 340         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 341         unsigned int len, i;
 342
 343         if (!labels)
 344                 return 0;
 345
 346         len = labels->words * sizeof(long);
 347         i = 0;
 348         do {
 349                 if (labels->bits[i] != 0)
 350                         return nla_put(skb, CTA_LABELS, len, labels->bits);
 351                 i++;
 352         } while (i < labels->words);
 353
 354         return 0;
 355 }
 356 #else
 357 #define ctnetlink_dump_labels(a, b) (0)
 358 #define ctnetlink_label_size(a) (0)
 359 #endif
 360
 361 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 362
 363 static inline int
 364 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 365 {
 366         struct nlattr *nest_parms;
 367
 368         if (!(ct->status & IPS_EXPECTED))
 369                 return 0;
 370
 371         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 372         if (!nest_parms)
 373                 goto nla_put_failure;
 374         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 375                 goto nla_put_failure;
 376         nla_nest_end(skb, nest_parms);
 377
 378         return 0;
 379
 380 nla_put_failure:
 381         return -1;
 382 }
 383
 384 #ifdef CONFIG_NF_NAT_NEEDED
 385 static int
 386 dump_nat_seq_adj(struct sk_buff *skb, const struct nf_nat_seq *natseq, int type)
 387 {
 388         struct nlattr *nest_parms;
 389
 390         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 391         if (!nest_parms)
 392                 goto nla_put_failure;
 393
 394         if (nla_put_be32(skb, CTA_NAT_SEQ_CORRECTION_POS,
 395                          htonl(natseq->correction_pos)) ||
 396             nla_put_be32(skb, CTA_NAT_SEQ_OFFSET_BEFORE,
 397                          htonl(natseq->offset_before)) ||
 398             nla_put_be32(skb, CTA_NAT_SEQ_OFFSET_AFTER,
 399                          htonl(natseq->offset_after)))
 400                 goto nla_put_failure;
 401
 402         nla_nest_end(skb, nest_parms);
 403
 404         return 0;
 405
 406 nla_put_failure:
 407         return -1;
 408 }
 409
 410 static inline int
 411 ctnetlink_dump_nat_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 412 {
 413         struct nf_nat_seq *natseq;
 414         struct nf_conn_nat *nat = nfct_nat(ct);
 415
 416         if (!(ct->status & IPS_SEQ_ADJUST) || !nat)
 417                 return 0;
 418
 419         natseq = &nat->seq[IP_CT_DIR_ORIGINAL];
 420         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_ORIG) == -1)
 421                 return -1;
 422
 423         natseq = &nat->seq[IP_CT_DIR_REPLY];
 424         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_REPLY) == -1)
 425                 return -1;
 426
 427         return 0;
 428 }
 429 #else
 430 #define ctnetlink_dump_nat_seq_adj(a, b) (0)
 431 #endif
 432
 433 static inline int
 434 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 435 {
 436         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 437                 goto nla_put_failure;
 438         return 0;
 439
 440 nla_put_failure:
 441         return -1;
 442 }
 443
 444 static inline int
 445 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 446 {
 447         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 448                 goto nla_put_failure;
 449         return 0;
 450
 451 nla_put_failure:
 452         return -1;
 453 }
 454
 455 static int
 456 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 457                     struct nf_conn *ct)
 458 {
 459         struct nlmsghdr *nlh;
 460         struct nfgenmsg *nfmsg;
 461         struct nlattr *nest_parms;
 462         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 463
 464         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 465         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 466         if (nlh == NULL)
 467                 goto nlmsg_failure;
 468
 469         nfmsg = nlmsg_data(nlh);
 470         nfmsg->nfgen_family = nf_ct_l3num(ct);
 471         nfmsg->version      = NFNETLINK_V0;
 472         nfmsg->res_id       = 0;
 473
 474         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 475         if (!nest_parms)
 476                 goto nla_put_failure;
 477         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 478                 goto nla_put_failure;
 479         nla_nest_end(skb, nest_parms);
 480
 481         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 482         if (!nest_parms)
 483                 goto nla_put_failure;
 484         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 485                 goto nla_put_failure;
 486         nla_nest_end(skb, nest_parms);
 487
 488         if (nf_ct_zone(ct) &&
 489             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 490                 goto nla_put_failure;
 491
 492         if (ctnetlink_dump_status(skb, ct) < 0 ||
 493             ctnetlink_dump_timeout(skb, ct) < 0 ||
 494             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL, type) < 0 ||
 495             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY, type) < 0 ||
 496             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 497             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 498             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 499             ctnetlink_dump_mark(skb, ct) < 0 ||
 500             ctnetlink_dump_secctx(skb, ct) < 0 ||
 501             ctnetlink_dump_labels(skb, ct) < 0 ||
 502             ctnetlink_dump_id(skb, ct) < 0 ||
 503             ctnetlink_dump_use(skb, ct) < 0 ||
 504             ctnetlink_dump_master(skb, ct) < 0 ||
 505             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 506                 goto nla_put_failure;
 507
 508         nlmsg_end(skb, nlh);
 509         return skb->len;
 510
 511 nlmsg_failure:
 512 nla_put_failure:
 513         nlmsg_cancel(skb, nlh);
 514         return -1;
 515 }
 516
 517 static inline size_t
 518 ctnetlink_proto_size(const struct nf_conn *ct)
 519 {
 520         struct nf_conntrack_l3proto *l3proto;
 521         struct nf_conntrack_l4proto *l4proto;
 522         size_t len = 0;
 523
 524         rcu_read_lock();
 525         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 526         len += l3proto->nla_size;
 527
 528         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 529         len += l4proto->nla_size;
 530         rcu_read_unlock();
 531
 532         return len;
 533 }
 534
 535 static inline size_t
 536 ctnetlink_counters_size(const struct nf_conn *ct)
 537 {
 538         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 539                 return 0;
 540         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 541                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 542                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 543                ;
 544 }
 545
 546 static inline int
 547 ctnetlink_secctx_size(const struct nf_conn *ct)
 548 {
 549 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 550         int len, ret;
 551
 552         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 553         if (ret)
 554                 return 0;
 555
 556         return nla_total_size(0) /* CTA_SECCTX */
 557                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 558 #else
 559         return 0;
 560 #endif
 561 }
 562
 563 static inline size_t
 564 ctnetlink_timestamp_size(const struct nf_conn *ct)
 565 {
 566 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 567         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 568                 return 0;
 569         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 570 #else
 571         return 0;
 572 #endif
 573 }
 574
 575 static inline size_t
 576 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 577 {
 578         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 579                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 580                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 581                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 582                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 583                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 584                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 585                + ctnetlink_counters_size(ct)
 586                + ctnetlink_timestamp_size(ct)
 587                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 588                + nla_total_size(0) /* CTA_PROTOINFO */
 589                + nla_total_size(0) /* CTA_HELP */
 590                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 591                + ctnetlink_secctx_size(ct)
 592 #ifdef CONFIG_NF_NAT_NEEDED
 593                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 594                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 595 #endif
 596 #ifdef CONFIG_NF_CONNTRACK_MARK
 597                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 598 #endif
 599                + ctnetlink_proto_size(ct)
 600                + ctnetlink_label_size(ct)
 601                ;
 602 }
 603
 604 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 605 static int
 606 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 607 {
 608         struct net *net;
 609         struct nlmsghdr *nlh;
 610         struct nfgenmsg *nfmsg;
 611         struct nlattr *nest_parms;
 612         struct nf_conn *ct = item->ct;
 613         struct sk_buff *skb;
 614         unsigned int type;
 615         unsigned int flags = 0, group;
 616         int err;
 617
 618         /* ignore our fake conntrack entry */
 619         if (nf_ct_is_untracked(ct))
 620                 return 0;
 621
 622         if (events & (1 << IPCT_DESTROY)) {
 623                 type = IPCTNL_MSG_CT_DELETE;
 624                 group = NFNLGRP_CONNTRACK_DESTROY;
 625         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 626                 type = IPCTNL_MSG_CT_NEW;
 627                 flags = NLM_F_CREATE|NLM_F_EXCL;
 628                 group = NFNLGRP_CONNTRACK_NEW;
 629         } else  if (events) {
 630                 type = IPCTNL_MSG_CT_NEW;
 631                 group = NFNLGRP_CONNTRACK_UPDATE;
 632         } else
 633                 return 0;
 634
 635         net = nf_ct_net(ct);
 636         if (!item->report && !nfnetlink_has_listeners(net, group))
 637                 return 0;
 638
 639         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 640         if (skb == NULL)
 641                 goto errout;
 642
 643         type |= NFNL_SUBSYS_CTNETLINK << 8;
 644         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 645         if (nlh == NULL)
 646                 goto nlmsg_failure;
 647
 648         nfmsg = nlmsg_data(nlh);
 649         nfmsg->nfgen_family = nf_ct_l3num(ct);
 650         nfmsg->version  = NFNETLINK_V0;
 651         nfmsg->res_id   = 0;
 652
 653         rcu_read_lock();
 654         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 655         if (!nest_parms)
 656                 goto nla_put_failure;
 657         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 658                 goto nla_put_failure;
 659         nla_nest_end(skb, nest_parms);
 660
 661         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 662         if (!nest_parms)
 663                 goto nla_put_failure;
 664         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 665                 goto nla_put_failure;
 666         nla_nest_end(skb, nest_parms);
 667
 668         if (nf_ct_zone(ct) &&
 669             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 670                 goto nla_put_failure;
 671
 672         if (ctnetlink_dump_id(skb, ct) < 0)
 673                 goto nla_put_failure;
 674
 675         if (ctnetlink_dump_status(skb, ct) < 0)
 676                 goto nla_put_failure;
 677
 678         if (events & (1 << IPCT_DESTROY)) {
 679                 if (ctnetlink_dump_counters(skb, ct,
 680                                             IP_CT_DIR_ORIGINAL, type) < 0 ||
 681                     ctnetlink_dump_counters(skb, ct,
 682                                             IP_CT_DIR_REPLY, type) < 0 ||
 683                     ctnetlink_dump_timestamp(skb, ct) < 0)
 684                         goto nla_put_failure;
 685         } else {
 686                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 687                         goto nla_put_failure;
 688
 689                 if (events & (1 << IPCT_PROTOINFO)
 690                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 691                         goto nla_put_failure;
 692
 693                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 694                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 695                         goto nla_put_failure;
 696
 697 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 698                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 699                     && ctnetlink_dump_secctx(skb, ct) < 0)
 700                         goto nla_put_failure;
 701 #endif
 702                 if (events & (1 << IPCT_LABEL) &&
 703                      ctnetlink_dump_labels(skb, ct) < 0)
 704                         goto nla_put_failure;
 705
 706                 if (events & (1 << IPCT_RELATED) &&
 707                     ctnetlink_dump_master(skb, ct) < 0)
 708                         goto nla_put_failure;
 709
 710                 if (events & (1 << IPCT_NATSEQADJ) &&
 711                     ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 712                         goto nla_put_failure;
 713         }
 714
 715 #ifdef CONFIG_NF_CONNTRACK_MARK
 716         if ((events & (1 << IPCT_MARK) || ct->mark)
 717             && ctnetlink_dump_mark(skb, ct) < 0)
 718                 goto nla_put_failure;
 719 #endif
 720         rcu_read_unlock();
 721
 722         nlmsg_end(skb, nlh);
 723         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 724                              GFP_ATOMIC);
 725         if (err == -ENOBUFS || err == -EAGAIN)
 726                 return -ENOBUFS;
 727
 728         return 0;
 729
 730 nla_put_failure:
 731         rcu_read_unlock();
 732         nlmsg_cancel(skb, nlh);
 733 nlmsg_failure:
 734         kfree_skb(skb);
 735 errout:
 736         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 737                 return -ENOBUFS;
 738
 739         return 0;
 740 }
 741 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 742
 743 static int ctnetlink_done(struct netlink_callback *cb)
 744 {
 745         if (cb->args[1])
 746                 nf_ct_put((struct nf_conn *)cb->args[1]);
 747         if (cb->data)
 748                 kfree(cb->data);
 749         return 0;
 750 }
 751
 752 struct ctnetlink_dump_filter {
 753         struct {
 754                 u_int32_t val;
 755                 u_int32_t mask;
 756         } mark;
 757 };
 758
 759 static int
 760 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 761 {
 762         struct net *net = sock_net(skb->sk);
 763         struct nf_conn *ct, *last;
 764         struct nf_conntrack_tuple_hash *h;
 765         struct hlist_nulls_node *n;
 766         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 767         u_int8_t l3proto = nfmsg->nfgen_family;
 768         int res;
 769 #ifdef CONFIG_NF_CONNTRACK_MARK
 770         const struct ctnetlink_dump_filter *filter = cb->data;
 771 #endif
 772
 773         spin_lock_bh(&nf_conntrack_lock);
 774         last = (struct nf_conn *)cb->args[1];
 775         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 776 restart:
 777                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 778                                          hnnode) {
 779                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 780                                 continue;
 781                         ct = nf_ct_tuplehash_to_ctrack(h);
 782                         /* Dump entries of a given L3 protocol number.
 783                          * If it is not specified, ie. l3proto == 0,
 784                          * then dump everything. */
 785                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 786                                 continue;
 787                         if (cb->args[1]) {
 788                                 if (ct != last)
 789                                         continue;
 790                                 cb->args[1] = 0;
 791                         }
 792 #ifdef CONFIG_NF_CONNTRACK_MARK
 793                         if (filter && !((ct->mark & filter->mark.mask) ==
 794                                         filter->mark.val)) {
 795                                 continue;
 796                         }
 797 #endif
 798                         rcu_read_lock();
 799                         res =
 800                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 801                                             cb->nlh->nlmsg_seq,
 802                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 803                                             ct);
 804                         rcu_read_unlock();
 805                         if (res < 0) {
 806                                 nf_conntrack_get(&ct->ct_general);
 807                                 cb->args[1] = (unsigned long)ct;
 808                                 goto out;
 809                         }
 810                 }
 811                 if (cb->args[1]) {
 812                         cb->args[1] = 0;
 813                         goto restart;
 814                 }
 815         }
 816 out:
 817         spin_unlock_bh(&nf_conntrack_lock);
 818         if (last)
 819                 nf_ct_put(last);
 820
 821         return skb->len;
 822 }
 823
 824 static inline int
 825 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 826 {
 827         struct nlattr *tb[CTA_IP_MAX+1];
 828         struct nf_conntrack_l3proto *l3proto;
 829         int ret = 0;
 830
 831         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 832
 833         rcu_read_lock();
 834         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 835
 836         if (likely(l3proto->nlattr_to_tuple)) {
 837                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 838                                           l3proto->nla_policy);
 839                 if (ret == 0)
 840                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 841         }
 842
 843         rcu_read_unlock();
 844
 845         return ret;
 846 }
 847
 848 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 849         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 850 };
 851
 852 static inline int
 853 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 854                             struct nf_conntrack_tuple *tuple)
 855 {
 856         struct nlattr *tb[CTA_PROTO_MAX+1];
 857         struct nf_conntrack_l4proto *l4proto;
 858         int ret = 0;
 859
 860         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 861         if (ret < 0)
 862                 return ret;
 863
 864         if (!tb[CTA_PROTO_NUM])
 865                 return -EINVAL;
 866         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 867
 868         rcu_read_lock();
 869         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 870
 871         if (likely(l4proto->nlattr_to_tuple)) {
 872                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 873                                           l4proto->nla_policy);
 874                 if (ret == 0)
 875                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 876         }
 877
 878         rcu_read_unlock();
 879
 880         return ret;
 881 }
 882
 883 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 884         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 885         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 886 };
 887
 888 static int
 889 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 890                       struct nf_conntrack_tuple *tuple,
 891                       enum ctattr_type type, u_int8_t l3num)
 892 {
 893         struct nlattr *tb[CTA_TUPLE_MAX+1];
 894         int err;
 895
 896         memset(tuple, 0, sizeof(*tuple));
 897
 898         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 899
 900         if (!tb[CTA_TUPLE_IP])
 901                 return -EINVAL;
 902
 903         tuple->src.l3num = l3num;
 904
 905         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 906         if (err < 0)
 907                 return err;
 908
 909         if (!tb[CTA_TUPLE_PROTO])
 910                 return -EINVAL;
 911
 912         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 913         if (err < 0)
 914                 return err;
 915
 916         /* orig and expect tuples get DIR_ORIGINAL */
 917         if (type == CTA_TUPLE_REPLY)
 918                 tuple->dst.dir = IP_CT_DIR_REPLY;
 919         else
 920                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 921
 922         return 0;
 923 }
 924
 925 static int
 926 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 927 {
 928         if (attr)
 929 #ifdef CONFIG_NF_CONNTRACK_ZONES
 930                 *zone = ntohs(nla_get_be16(attr));
 931 #else
 932                 return -EOPNOTSUPP;
 933 #endif
 934         else
 935                 *zone = 0;
 936
 937         return 0;
 938 }
 939
 940 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 941         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 942                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 943 };
 944
 945 static inline int
 946 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 947                      struct nlattr **helpinfo)
 948 {
 949         struct nlattr *tb[CTA_HELP_MAX+1];
 950
 951         nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 952
 953         if (!tb[CTA_HELP_NAME])
 954                 return -EINVAL;
 955
 956         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 957
 958         if (tb[CTA_HELP_INFO])
 959                 *helpinfo = tb[CTA_HELP_INFO];
 960
 961         return 0;
 962 }
 963
 964 #define __CTA_LABELS_MAX_LENGTH ((XT_CONNLABEL_MAXBIT + 1) / BITS_PER_BYTE)
 965 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 966         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 967         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 968         [CTA_STATUS]            = { .type = NLA_U32 },
 969         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 970         [CTA_HELP]              = { .type = NLA_NESTED },
 971         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 972         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 973         [CTA_MARK]              = { .type = NLA_U32 },
 974         [CTA_ID]                = { .type = NLA_U32 },
 975         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 976         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 977         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
 978         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
 979         [CTA_ZONE]              = { .type = NLA_U16 },
 980         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 981         [CTA_LABELS]            = { .type = NLA_BINARY,
 982                                     .len = __CTA_LABELS_MAX_LENGTH },
 983         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
 984                                     .len = __CTA_LABELS_MAX_LENGTH },
 985 };
 986
 987 static int
 988 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 989                         const struct nlmsghdr *nlh,
 990                         const struct nlattr * const cda[])
 991 {
 992         struct net *net = sock_net(ctnl);
 993         struct nf_conntrack_tuple_hash *h;
 994         struct nf_conntrack_tuple tuple;
 995         struct nf_conn *ct;
 996         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 997         u_int8_t u3 = nfmsg->nfgen_family;
 998         u16 zone;
 999         int err;
1000
1001         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1002         if (err < 0)
1003                 return err;
1004
1005         if (cda[CTA_TUPLE_ORIG])
1006                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1007         else if (cda[CTA_TUPLE_REPLY])
1008                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1009         else {
1010                 /* Flush the whole table */
1011                 nf_conntrack_flush_report(net,
1012                                          NETLINK_CB(skb).portid,
1013                                          nlmsg_report(nlh));
1014                 return 0;
1015         }
1016
1017         if (err < 0)
1018                 return err;
1019
1020         h = nf_conntrack_find_get(net, zone, &tuple);
1021         if (!h)
1022                 return -ENOENT;
1023
1024         ct = nf_ct_tuplehash_to_ctrack(h);
1025
1026         if (cda[CTA_ID]) {
1027                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1028                 if (id != (u32)(unsigned long)ct) {
1029                         nf_ct_put(ct);
1030                         return -ENOENT;
1031                 }
1032         }
1033
1034         if (del_timer(&ct->timeout)) {
1035                 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
1036                                               NETLINK_CB(skb).portid,
1037                                               nlmsg_report(nlh)) < 0) {
1038                         nf_ct_delete_from_lists(ct);
1039                         /* we failed to report the event, try later */
1040                         nf_ct_dying_timeout(ct);
1041                         nf_ct_put(ct);
1042                         return 0;
1043                 }
1044                 /* death_by_timeout would report the event again */
1045                 set_bit(IPS_DYING_BIT, &ct->status);
1046                 nf_ct_delete_from_lists(ct);
1047                 nf_ct_put(ct);
1048         }
1049         nf_ct_put(ct);
1050
1051         return 0;
1052 }
1053
1054 static int
1055 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1056                         const struct nlmsghdr *nlh,
1057                         const struct nlattr * const cda[])
1058 {
1059         struct net *net = sock_net(ctnl);
1060         struct nf_conntrack_tuple_hash *h;
1061         struct nf_conntrack_tuple tuple;
1062         struct nf_conn *ct;
1063         struct sk_buff *skb2 = NULL;
1064         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1065         u_int8_t u3 = nfmsg->nfgen_family;
1066         u16 zone;
1067         int err;
1068
1069         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1070                 struct netlink_dump_control c = {
1071                         .dump = ctnetlink_dump_table,
1072                         .done = ctnetlink_done,
1073                 };
1074 #ifdef CONFIG_NF_CONNTRACK_MARK
1075                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1076                         struct ctnetlink_dump_filter *filter;
1077
1078                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1079                                          GFP_ATOMIC);
1080                         if (filter == NULL)
1081                                 return -ENOMEM;
1082
1083                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1084                         filter->mark.mask =
1085                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1086                         c.data = filter;
1087                 }
1088 #endif
1089                 return netlink_dump_start(ctnl, skb, nlh, &c);
1090         }
1091
1092         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1093         if (err < 0)
1094                 return err;
1095
1096         if (cda[CTA_TUPLE_ORIG])
1097                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1098         else if (cda[CTA_TUPLE_REPLY])
1099                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1100         else
1101                 return -EINVAL;
1102
1103         if (err < 0)
1104                 return err;
1105
1106         h = nf_conntrack_find_get(net, zone, &tuple);
1107         if (!h)
1108                 return -ENOENT;
1109
1110         ct = nf_ct_tuplehash_to_ctrack(h);
1111
1112         err = -ENOMEM;
1113         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1114         if (skb2 == NULL) {
1115                 nf_ct_put(ct);
1116                 return -ENOMEM;
1117         }
1118
1119         rcu_read_lock();
1120         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1121                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1122         rcu_read_unlock();
1123         nf_ct_put(ct);
1124         if (err <= 0)
1125                 goto free;
1126
1127         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1128         if (err < 0)
1129                 goto out;
1130
1131         return 0;
1132
1133 free:
1134         kfree_skb(skb2);
1135 out:
1136         /* this avoids a loop in nfnetlink. */
1137         return err == -EAGAIN ? -ENOBUFS : err;
1138 }
1139
1140 static int ctnetlink_done_list(struct netlink_callback *cb)
1141 {
1142         if (cb->args[1])
1143                 nf_ct_put((struct nf_conn *)cb->args[1]);
1144         return 0;
1145 }
1146
1147 static int
1148 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb,
1149                     struct hlist_nulls_head *list)
1150 {
1151         struct nf_conn *ct, *last;
1152         struct nf_conntrack_tuple_hash *h;
1153         struct hlist_nulls_node *n;
1154         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1155         u_int8_t l3proto = nfmsg->nfgen_family;
1156         int res;
1157
1158         if (cb->args[2])
1159                 return 0;
1160
1161         spin_lock_bh(&nf_conntrack_lock);
1162         last = (struct nf_conn *)cb->args[1];
1163 restart:
1164         hlist_nulls_for_each_entry(h, n, list, hnnode) {
1165                 ct = nf_ct_tuplehash_to_ctrack(h);
1166                 if (l3proto && nf_ct_l3num(ct) != l3proto)
1167                         continue;
1168                 if (cb->args[1]) {
1169                         if (ct != last)
1170                                 continue;
1171                         cb->args[1] = 0;
1172                 }
1173                 rcu_read_lock();
1174                 res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1175                                           cb->nlh->nlmsg_seq,
1176                                           NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1177                                           ct);
1178                 rcu_read_unlock();
1179                 if (res < 0) {
1180                         nf_conntrack_get(&ct->ct_general);
1181                         cb->args[1] = (unsigned long)ct;
1182                         goto out;
1183                 }
1184         }
1185         if (cb->args[1]) {
1186                 cb->args[1] = 0;
1187                 goto restart;
1188         } else
1189                 cb->args[2] = 1;
1190 out:
1191         spin_unlock_bh(&nf_conntrack_lock);
1192         if (last)
1193                 nf_ct_put(last);
1194
1195         return skb->len;
1196 }
1197
1198 static int
1199 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1200 {
1201         struct net *net = sock_net(skb->sk);
1202
1203         return ctnetlink_dump_list(skb, cb, &net->ct.dying);
1204 }
1205
1206 static int
1207 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1208                        const struct nlmsghdr *nlh,
1209                        const struct nlattr * const cda[])
1210 {
1211         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1212                 struct netlink_dump_control c = {
1213                         .dump = ctnetlink_dump_dying,
1214                         .done = ctnetlink_done_list,
1215                 };
1216                 return netlink_dump_start(ctnl, skb, nlh, &c);
1217         }
1218
1219         return -EOPNOTSUPP;
1220 }
1221
1222 static int
1223 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1224 {
1225         struct net *net = sock_net(skb->sk);
1226
1227         return ctnetlink_dump_list(skb, cb, &net->ct.unconfirmed);
1228 }
1229
1230 static int
1231 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1232                              const struct nlmsghdr *nlh,
1233                              const struct nlattr * const cda[])
1234 {
1235         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1236                 struct netlink_dump_control c = {
1237                         .dump = ctnetlink_dump_unconfirmed,
1238                         .done = ctnetlink_done_list,
1239                 };
1240                 return netlink_dump_start(ctnl, skb, nlh, &c);
1241         }
1242
1243         return -EOPNOTSUPP;
1244 }
1245
1246 #ifdef CONFIG_NF_NAT_NEEDED
1247 static int
1248 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1249                           enum nf_nat_manip_type manip,
1250                           const struct nlattr *attr)
1251 {
1252         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1253         int err;
1254
1255         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1256         if (!parse_nat_setup) {
1257 #ifdef CONFIG_MODULES
1258                 rcu_read_unlock();
1259                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1260                 if (request_module("nf-nat") < 0) {
1261                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1262                         rcu_read_lock();
1263                         return -EOPNOTSUPP;
1264                 }
1265                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1266                 rcu_read_lock();
1267                 if (nfnetlink_parse_nat_setup_hook)
1268                         return -EAGAIN;
1269 #endif
1270                 return -EOPNOTSUPP;
1271         }
1272
1273         err = parse_nat_setup(ct, manip, attr);
1274         if (err == -EAGAIN) {
1275 #ifdef CONFIG_MODULES
1276                 rcu_read_unlock();
1277                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1278                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1279                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1280                         rcu_read_lock();
1281                         return -EOPNOTSUPP;
1282                 }
1283                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1284                 rcu_read_lock();
1285 #else
1286                 err = -EOPNOTSUPP;
1287 #endif
1288         }
1289         return err;
1290 }
1291 #endif
1292
1293 static int
1294 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1295 {
1296         unsigned long d;
1297         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1298         d = ct->status ^ status;
1299
1300         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1301                 /* unchangeable */
1302                 return -EBUSY;
1303
1304         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1305                 /* SEEN_REPLY bit can only be set */
1306                 return -EBUSY;
1307
1308         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1309                 /* ASSURED bit can only be set */
1310                 return -EBUSY;
1311
1312         /* Be careful here, modifying NAT bits can screw up things,
1313          * so don't let users modify them directly if they don't pass
1314          * nf_nat_range. */
1315         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1316         return 0;
1317 }
1318
1319 static int
1320 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1321 {
1322 #ifdef CONFIG_NF_NAT_NEEDED
1323         int ret;
1324
1325         if (cda[CTA_NAT_DST]) {
1326                 ret = ctnetlink_parse_nat_setup(ct,
1327                                                 NF_NAT_MANIP_DST,
1328                                                 cda[CTA_NAT_DST]);
1329                 if (ret < 0)
1330                         return ret;
1331         }
1332         if (cda[CTA_NAT_SRC]) {
1333                 ret = ctnetlink_parse_nat_setup(ct,
1334                                                 NF_NAT_MANIP_SRC,
1335                                                 cda[CTA_NAT_SRC]);
1336                 if (ret < 0)
1337                         return ret;
1338         }
1339         return 0;
1340 #else
1341         return -EOPNOTSUPP;
1342 #endif
1343 }
1344
1345 static inline int
1346 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1347 {
1348         struct nf_conntrack_helper *helper;
1349         struct nf_conn_help *help = nfct_help(ct);
1350         char *helpname = NULL;
1351         struct nlattr *helpinfo = NULL;
1352         int err;
1353
1354         /* don't change helper of sibling connections */
1355         if (ct->master)
1356                 return -EBUSY;
1357
1358         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1359         if (err < 0)
1360                 return err;
1361
1362         if (!strcmp(helpname, "")) {
1363                 if (help && help->helper) {
1364                         /* we had a helper before ... */
1365                         nf_ct_remove_expectations(ct);
1366                         RCU_INIT_POINTER(help->helper, NULL);
1367                 }
1368
1369                 return 0;
1370         }
1371
1372         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1373                                             nf_ct_protonum(ct));
1374         if (helper == NULL) {
1375 #ifdef CONFIG_MODULES
1376                 spin_unlock_bh(&nf_conntrack_lock);
1377
1378                 if (request_module("nfct-helper-%s", helpname) < 0) {
1379                         spin_lock_bh(&nf_conntrack_lock);
1380                         return -EOPNOTSUPP;
1381                 }
1382
1383                 spin_lock_bh(&nf_conntrack_lock);
1384                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1385                                                     nf_ct_protonum(ct));
1386                 if (helper)
1387                         return -EAGAIN;
1388 #endif
1389                 return -EOPNOTSUPP;
1390         }
1391
1392         if (help) {
1393                 if (help->helper == helper) {
1394                         /* update private helper data if allowed. */
1395                         if (helper->from_nlattr)
1396                                 helper->from_nlattr(helpinfo, ct);
1397                         return 0;
1398                 } else
1399                         return -EBUSY;
1400         }
1401
1402         /* we cannot set a helper for an existing conntrack */
1403         return -EOPNOTSUPP;
1404 }
1405
1406 static inline int
1407 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1408 {
1409         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1410
1411         if (!del_timer(&ct->timeout))
1412                 return -ETIME;
1413
1414         ct->timeout.expires = jiffies + timeout * HZ;
1415         add_timer(&ct->timeout);
1416
1417         return 0;
1418 }
1419
1420 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1421         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1422         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1423         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1424 };
1425
1426 static inline int
1427 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1428 {
1429         const struct nlattr *attr = cda[CTA_PROTOINFO];
1430         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1431         struct nf_conntrack_l4proto *l4proto;
1432         int err = 0;
1433
1434         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1435
1436         rcu_read_lock();
1437         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1438         if (l4proto->from_nlattr)
1439                 err = l4proto->from_nlattr(tb, ct);
1440         rcu_read_unlock();
1441
1442         return err;
1443 }
1444
1445 #ifdef CONFIG_NF_NAT_NEEDED
1446 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1447         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1448         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1449         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1450 };
1451
1452 static inline int
1453 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1454 {
1455         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1456
1457         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1458
1459         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1460                 return -EINVAL;
1461
1462         natseq->correction_pos =
1463                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1464
1465         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1466                 return -EINVAL;
1467
1468         natseq->offset_before =
1469                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1470
1471         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1472                 return -EINVAL;
1473
1474         natseq->offset_after =
1475                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1476
1477         return 0;
1478 }
1479
1480 static int
1481 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1482                              const struct nlattr * const cda[])
1483 {
1484         int ret = 0;
1485         struct nf_conn_nat *nat = nfct_nat(ct);
1486
1487         if (!nat)
1488                 return 0;
1489
1490         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1491                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1492                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1493                 if (ret < 0)
1494                         return ret;
1495
1496                 ct->status |= IPS_SEQ_ADJUST;
1497         }
1498
1499         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1500                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1501                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1502                 if (ret < 0)
1503                         return ret;
1504
1505                 ct->status |= IPS_SEQ_ADJUST;
1506         }
1507
1508         return 0;
1509 }
1510 #endif
1511
1512 static int
1513 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1514 {
1515 #ifdef CONFIG_NF_CONNTRACK_LABELS
1516         size_t len = nla_len(cda[CTA_LABELS]);
1517         const void *mask = cda[CTA_LABELS_MASK];
1518
1519         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1520                 return -EINVAL;
1521
1522         if (mask) {
1523                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1524                     nla_len(cda[CTA_LABELS_MASK]) != len)
1525                         return -EINVAL;
1526                 mask = nla_data(cda[CTA_LABELS_MASK]);
1527         }
1528
1529         len /= sizeof(u32);
1530
1531         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1532 #else
1533         return -EOPNOTSUPP;
1534 #endif
1535 }
1536
1537 static int
1538 ctnetlink_change_conntrack(struct nf_conn *ct,
1539                            const struct nlattr * const cda[])
1540 {
1541         int err;
1542
1543         /* only allow NAT changes and master assignation for new conntracks */
1544         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1545                 return -EOPNOTSUPP;
1546
1547         if (cda[CTA_HELP]) {
1548                 err = ctnetlink_change_helper(ct, cda);
1549                 if (err < 0)
1550                         return err;
1551         }
1552
1553         if (cda[CTA_TIMEOUT]) {
1554                 err = ctnetlink_change_timeout(ct, cda);
1555                 if (err < 0)
1556                         return err;
1557         }
1558
1559         if (cda[CTA_STATUS]) {
1560                 err = ctnetlink_change_status(ct, cda);
1561                 if (err < 0)
1562                         return err;
1563         }
1564
1565         if (cda[CTA_PROTOINFO]) {
1566                 err = ctnetlink_change_protoinfo(ct, cda);
1567                 if (err < 0)
1568                         return err;
1569         }
1570
1571 #if defined(CONFIG_NF_CONNTRACK_MARK)
1572         if (cda[CTA_MARK])
1573                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1574 #endif
1575
1576 #ifdef CONFIG_NF_NAT_NEEDED
1577         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1578                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1579                 if (err < 0)
1580                         return err;
1581         }
1582 #endif
1583         if (cda[CTA_LABELS]) {
1584                 err = ctnetlink_attach_labels(ct, cda);
1585                 if (err < 0)
1586                         return err;
1587         }
1588
1589         return 0;
1590 }
1591
1592 static struct nf_conn *
1593 ctnetlink_create_conntrack(struct net *net, u16 zone,
1594                            const struct nlattr * const cda[],
1595                            struct nf_conntrack_tuple *otuple,
1596                            struct nf_conntrack_tuple *rtuple,
1597                            u8 u3)
1598 {
1599         struct nf_conn *ct;
1600         int err = -EINVAL;
1601         struct nf_conntrack_helper *helper;
1602         struct nf_conn_tstamp *tstamp;
1603
1604         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1605         if (IS_ERR(ct))
1606                 return ERR_PTR(-ENOMEM);
1607
1608         if (!cda[CTA_TIMEOUT])
1609                 goto err1;
1610         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1611
1612         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1613
1614         rcu_read_lock();
1615         if (cda[CTA_HELP]) {
1616                 char *helpname = NULL;
1617                 struct nlattr *helpinfo = NULL;
1618
1619                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1620                 if (err < 0)
1621                         goto err2;
1622
1623                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1624                                                     nf_ct_protonum(ct));
1625                 if (helper == NULL) {
1626                         rcu_read_unlock();
1627 #ifdef CONFIG_MODULES
1628                         if (request_module("nfct-helper-%s", helpname) < 0) {
1629                                 err = -EOPNOTSUPP;
1630                                 goto err1;
1631                         }
1632
1633                         rcu_read_lock();
1634                         helper = __nf_conntrack_helper_find(helpname,
1635                                                             nf_ct_l3num(ct),
1636                                                             nf_ct_protonum(ct));
1637                         if (helper) {
1638                                 err = -EAGAIN;
1639                                 goto err2;
1640                         }
1641                         rcu_read_unlock();
1642 #endif
1643                         err = -EOPNOTSUPP;
1644                         goto err1;
1645                 } else {
1646                         struct nf_conn_help *help;
1647
1648                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1649                         if (help == NULL) {
1650                                 err = -ENOMEM;
1651                                 goto err2;
1652                         }
1653                         /* set private helper data if allowed. */
1654                         if (helper->from_nlattr)
1655                                 helper->from_nlattr(helpinfo, ct);
1656
1657                         /* not in hash table yet so not strictly necessary */
1658                         RCU_INIT_POINTER(help->helper, helper);
1659                 }
1660         } else {
1661                 /* try an implicit helper assignation */
1662                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1663                 if (err < 0)
1664                         goto err2;
1665         }
1666
1667         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1668                 err = ctnetlink_change_nat(ct, cda);
1669                 if (err < 0)
1670                         goto err2;
1671         }
1672
1673         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1674         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1675         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1676         nf_ct_labels_ext_add(ct);
1677
1678         /* we must add conntrack extensions before confirmation. */
1679         ct->status |= IPS_CONFIRMED;
1680
1681         if (cda[CTA_STATUS]) {
1682                 err = ctnetlink_change_status(ct, cda);
1683                 if (err < 0)
1684                         goto err2;
1685         }
1686
1687 #ifdef CONFIG_NF_NAT_NEEDED
1688         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1689                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1690                 if (err < 0)
1691                         goto err2;
1692         }
1693 #endif
1694
1695         memset(&ct->proto, 0, sizeof(ct->proto));
1696         if (cda[CTA_PROTOINFO]) {
1697                 err = ctnetlink_change_protoinfo(ct, cda);
1698                 if (err < 0)
1699                         goto err2;
1700         }
1701
1702 #if defined(CONFIG_NF_CONNTRACK_MARK)
1703         if (cda[CTA_MARK])
1704                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1705 #endif
1706
1707         /* setup master conntrack: this is a confirmed expectation */
1708         if (cda[CTA_TUPLE_MASTER]) {
1709                 struct nf_conntrack_tuple master;
1710                 struct nf_conntrack_tuple_hash *master_h;
1711                 struct nf_conn *master_ct;
1712
1713                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1714                 if (err < 0)
1715                         goto err2;
1716
1717                 master_h = nf_conntrack_find_get(net, zone, &master);
1718                 if (master_h == NULL) {
1719                         err = -ENOENT;
1720                         goto err2;
1721                 }
1722                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1723                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1724                 ct->master = master_ct;
1725         }
1726         tstamp = nf_conn_tstamp_find(ct);
1727         if (tstamp)
1728                 tstamp->start = ktime_to_ns(ktime_get_real());
1729
1730         err = nf_conntrack_hash_check_insert(ct);
1731         if (err < 0)
1732                 goto err2;
1733
1734         rcu_read_unlock();
1735
1736         return ct;
1737
1738 err2:
1739         rcu_read_unlock();
1740 err1:
1741         nf_conntrack_free(ct);
1742         return ERR_PTR(err);
1743 }
1744
1745 static int
1746 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1747                         const struct nlmsghdr *nlh,
1748                         const struct nlattr * const cda[])
1749 {
1750         struct net *net = sock_net(ctnl);
1751         struct nf_conntrack_tuple otuple, rtuple;
1752         struct nf_conntrack_tuple_hash *h = NULL;
1753         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1754         struct nf_conn *ct;
1755         u_int8_t u3 = nfmsg->nfgen_family;
1756         u16 zone;
1757         int err;
1758
1759         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1760         if (err < 0)
1761                 return err;
1762
1763         if (cda[CTA_TUPLE_ORIG]) {
1764                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1765                 if (err < 0)
1766                         return err;
1767         }
1768
1769         if (cda[CTA_TUPLE_REPLY]) {
1770                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1771                 if (err < 0)
1772                         return err;
1773         }
1774
1775         if (cda[CTA_TUPLE_ORIG])
1776                 h = nf_conntrack_find_get(net, zone, &otuple);
1777         else if (cda[CTA_TUPLE_REPLY])
1778                 h = nf_conntrack_find_get(net, zone, &rtuple);
1779
1780         if (h == NULL) {
1781                 err = -ENOENT;
1782                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1783                         enum ip_conntrack_events events;
1784
1785                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1786                                 return -EINVAL;
1787
1788                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1789                                                         &rtuple, u3);
1790                         if (IS_ERR(ct))
1791                                 return PTR_ERR(ct);
1792
1793                         err = 0;
1794                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1795                                 events = IPCT_RELATED;
1796                         else
1797                                 events = IPCT_NEW;
1798
1799                         if (cda[CTA_LABELS] &&
1800                             ctnetlink_attach_labels(ct, cda) == 0)
1801                                 events |= (1 << IPCT_LABEL);
1802
1803                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1804                                                       (1 << IPCT_ASSURED) |
1805                                                       (1 << IPCT_HELPER) |
1806                                                       (1 << IPCT_PROTOINFO) |
1807                                                       (1 << IPCT_NATSEQADJ) |
1808                                                       (1 << IPCT_MARK) | events,
1809                                                       ct, NETLINK_CB(skb).portid,
1810                                                       nlmsg_report(nlh));
1811                         nf_ct_put(ct);
1812                 }
1813
1814                 return err;
1815         }
1816         /* implicit 'else' */
1817
1818         err = -EEXIST;
1819         ct = nf_ct_tuplehash_to_ctrack(h);
1820         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1821                 spin_lock_bh(&nf_conntrack_lock);
1822                 err = ctnetlink_change_conntrack(ct, cda);
1823                 spin_unlock_bh(&nf_conntrack_lock);
1824                 if (err == 0) {
1825                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1826                                                       (1 << IPCT_ASSURED) |
1827                                                       (1 << IPCT_HELPER) |
1828                                                       (1 << IPCT_LABEL) |
1829                                                       (1 << IPCT_PROTOINFO) |
1830                                                       (1 << IPCT_NATSEQADJ) |
1831                                                       (1 << IPCT_MARK),
1832                                                       ct, NETLINK_CB(skb).portid,
1833                                                       nlmsg_report(nlh));
1834                 }
1835         }
1836
1837         nf_ct_put(ct);
1838         return err;
1839 }
1840
1841 static int
1842 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1843                                 __u16 cpu, const struct ip_conntrack_stat *st)
1844 {
1845         struct nlmsghdr *nlh;
1846         struct nfgenmsg *nfmsg;
1847         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1848
1849         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1850         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1851         if (nlh == NULL)
1852                 goto nlmsg_failure;
1853
1854         nfmsg = nlmsg_data(nlh);
1855         nfmsg->nfgen_family = AF_UNSPEC;
1856         nfmsg->version      = NFNETLINK_V0;
1857         nfmsg->res_id       = htons(cpu);
1858
1859         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1860             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1861             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1862             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1863             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1864             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1865             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1866             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1867             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1868                                 htonl(st->insert_failed)) ||
1869             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1870             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1871             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1872             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1873                                 htonl(st->search_restart)))
1874                 goto nla_put_failure;
1875
1876         nlmsg_end(skb, nlh);
1877         return skb->len;
1878
1879 nla_put_failure:
1880 nlmsg_failure:
1881         nlmsg_cancel(skb, nlh);
1882         return -1;
1883 }
1884
1885 static int
1886 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1887 {
1888         int cpu;
1889         struct net *net = sock_net(skb->sk);
1890
1891         if (cb->args[0] == nr_cpu_ids)
1892                 return 0;
1893
1894         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1895                 const struct ip_conntrack_stat *st;
1896
1897                 if (!cpu_possible(cpu))
1898                         continue;
1899
1900                 st = per_cpu_ptr(net->ct.stat, cpu);
1901                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1902                                                     NETLINK_CB(cb->skb).portid,
1903                                                     cb->nlh->nlmsg_seq,
1904                                                     cpu, st) < 0)
1905                                 break;
1906         }
1907         cb->args[0] = cpu;
1908
1909         return skb->len;
1910 }
1911
1912 static int
1913 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1914                       const struct nlmsghdr *nlh,
1915                       const struct nlattr * const cda[])
1916 {
1917         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1918                 struct netlink_dump_control c = {
1919                         .dump = ctnetlink_ct_stat_cpu_dump,
1920                 };
1921                 return netlink_dump_start(ctnl, skb, nlh, &c);
1922         }
1923
1924         return 0;
1925 }
1926
1927 static int
1928 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1929                             struct net *net)
1930 {
1931         struct nlmsghdr *nlh;
1932         struct nfgenmsg *nfmsg;
1933         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1934         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1935
1936         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1937         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1938         if (nlh == NULL)
1939                 goto nlmsg_failure;
1940
1941         nfmsg = nlmsg_data(nlh);
1942         nfmsg->nfgen_family = AF_UNSPEC;
1943         nfmsg->version      = NFNETLINK_V0;
1944         nfmsg->res_id       = 0;
1945
1946         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1947                 goto nla_put_failure;
1948
1949         nlmsg_end(skb, nlh);
1950         return skb->len;
1951
1952 nla_put_failure:
1953 nlmsg_failure:
1954         nlmsg_cancel(skb, nlh);
1955         return -1;
1956 }
1957
1958 static int
1959 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1960                   const struct nlmsghdr *nlh,
1961                   const struct nlattr * const cda[])
1962 {
1963         struct sk_buff *skb2;
1964         int err;
1965
1966         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1967         if (skb2 == NULL)
1968                 return -ENOMEM;
1969
1970         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1971                                           nlh->nlmsg_seq,
1972                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1973                                           sock_net(skb->sk));
1974         if (err <= 0)
1975                 goto free;
1976
1977         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1978         if (err < 0)
1979                 goto out;
1980
1981         return 0;
1982
1983 free:
1984         kfree_skb(skb2);
1985 out:
1986         /* this avoids a loop in nfnetlink. */
1987         return err == -EAGAIN ? -ENOBUFS : err;
1988 }
1989
1990 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
1991 static size_t
1992 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
1993 {
1994         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
1995                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
1996                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
1997                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
1998                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
1999                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2000                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2001                + nla_total_size(0) /* CTA_PROTOINFO */
2002                + nla_total_size(0) /* CTA_HELP */
2003                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2004                + ctnetlink_secctx_size(ct)
2005 #ifdef CONFIG_NF_NAT_NEEDED
2006                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2007                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2008 #endif
2009 #ifdef CONFIG_NF_CONNTRACK_MARK
2010                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2011 #endif
2012                + ctnetlink_proto_size(ct)
2013                ;
2014 }
2015
2016 static int
2017 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2018 {
2019         struct nlattr *nest_parms;
2020
2021         rcu_read_lock();
2022         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2023         if (!nest_parms)
2024                 goto nla_put_failure;
2025         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2026                 goto nla_put_failure;
2027         nla_nest_end(skb, nest_parms);
2028
2029         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2030         if (!nest_parms)
2031                 goto nla_put_failure;
2032         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2033                 goto nla_put_failure;
2034         nla_nest_end(skb, nest_parms);
2035
2036         if (nf_ct_zone(ct)) {
2037                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2038                         goto nla_put_failure;
2039         }
2040
2041         if (ctnetlink_dump_id(skb, ct) < 0)
2042                 goto nla_put_failure;
2043
2044         if (ctnetlink_dump_status(skb, ct) < 0)
2045                 goto nla_put_failure;
2046
2047         if (ctnetlink_dump_timeout(skb, ct) < 0)
2048                 goto nla_put_failure;
2049
2050         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2051                 goto nla_put_failure;
2052
2053         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2054                 goto nla_put_failure;
2055
2056 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2057         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2058                 goto nla_put_failure;
2059 #endif
2060         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2061                 goto nla_put_failure;
2062
2063         if ((ct->status & IPS_SEQ_ADJUST) &&
2064             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
2065                 goto nla_put_failure;
2066
2067 #ifdef CONFIG_NF_CONNTRACK_MARK
2068         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2069                 goto nla_put_failure;
2070 #endif
2071         if (ctnetlink_dump_labels(skb, ct) < 0)
2072                 goto nla_put_failure;
2073         rcu_read_unlock();
2074         return 0;
2075
2076 nla_put_failure:
2077         rcu_read_unlock();
2078         return -ENOSPC;
2079 }
2080
2081 static int
2082 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2083 {
2084         int err;
2085
2086         if (cda[CTA_TIMEOUT]) {
2087                 err = ctnetlink_change_timeout(ct, cda);
2088                 if (err < 0)
2089                         return err;
2090         }
2091         if (cda[CTA_STATUS]) {
2092                 err = ctnetlink_change_status(ct, cda);
2093                 if (err < 0)
2094                         return err;
2095         }
2096         if (cda[CTA_HELP]) {
2097                 err = ctnetlink_change_helper(ct, cda);
2098                 if (err < 0)
2099                         return err;
2100         }
2101         if (cda[CTA_LABELS]) {
2102                 err = ctnetlink_attach_labels(ct, cda);
2103                 if (err < 0)
2104                         return err;
2105         }
2106 #if defined(CONFIG_NF_CONNTRACK_MARK)
2107         if (cda[CTA_MARK])
2108                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2109 #endif
2110         return 0;
2111 }
2112
2113 static int
2114 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2115 {
2116         struct nlattr *cda[CTA_MAX+1];
2117         int ret;
2118
2119         nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2120
2121         spin_lock_bh(&nf_conntrack_lock);
2122         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2123         spin_unlock_bh(&nf_conntrack_lock);
2124
2125         return ret;
2126 }
2127
2128 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2129         .build_size     = ctnetlink_nfqueue_build_size,
2130         .build          = ctnetlink_nfqueue_build,
2131         .parse          = ctnetlink_nfqueue_parse,
2132 };
2133 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2134
2135 /***********************************************************************
2136  * EXPECT
2137  ***********************************************************************/
2138
2139 static inline int
2140 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2141                          const struct nf_conntrack_tuple *tuple,
2142                          enum ctattr_expect type)
2143 {
2144         struct nlattr *nest_parms;
2145
2146         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2147         if (!nest_parms)
2148                 goto nla_put_failure;
2149         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2150                 goto nla_put_failure;
2151         nla_nest_end(skb, nest_parms);
2152
2153         return 0;
2154
2155 nla_put_failure:
2156         return -1;
2157 }
2158
2159 static inline int
2160 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2161                         const struct nf_conntrack_tuple *tuple,
2162                         const struct nf_conntrack_tuple_mask *mask)
2163 {
2164         int ret;
2165         struct nf_conntrack_l3proto *l3proto;
2166         struct nf_conntrack_l4proto *l4proto;
2167         struct nf_conntrack_tuple m;
2168         struct nlattr *nest_parms;
2169
2170         memset(&m, 0xFF, sizeof(m));
2171         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2172         m.src.u.all = mask->src.u.all;
2173         m.dst.protonum = tuple->dst.protonum;
2174
2175         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2176         if (!nest_parms)
2177                 goto nla_put_failure;
2178
2179         rcu_read_lock();
2180         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2181         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2182         if (ret >= 0) {
2183                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2184                                                tuple->dst.protonum);
2185         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2186         }
2187         rcu_read_unlock();
2188
2189         if (unlikely(ret < 0))
2190                 goto nla_put_failure;
2191
2192         nla_nest_end(skb, nest_parms);
2193
2194         return 0;
2195
2196 nla_put_failure:
2197         return -1;
2198 }
2199
2200 static const union nf_inet_addr any_addr;
2201
2202 static int
2203 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2204                           const struct nf_conntrack_expect *exp)
2205 {
2206         struct nf_conn *master = exp->master;
2207         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2208         struct nf_conn_help *help;
2209 #ifdef CONFIG_NF_NAT_NEEDED
2210         struct nlattr *nest_parms;
2211         struct nf_conntrack_tuple nat_tuple = {};
2212 #endif
2213         struct nf_ct_helper_expectfn *expfn;
2214
2215         if (timeout < 0)
2216                 timeout = 0;
2217
2218         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2219                 goto nla_put_failure;
2220         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2221                 goto nla_put_failure;
2222         if (ctnetlink_exp_dump_tuple(skb,
2223                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2224                                  CTA_EXPECT_MASTER) < 0)
2225                 goto nla_put_failure;
2226
2227 #ifdef CONFIG_NF_NAT_NEEDED
2228         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2229             exp->saved_proto.all) {
2230                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2231                 if (!nest_parms)
2232                         goto nla_put_failure;
2233
2234                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2235                         goto nla_put_failure;
2236
2237                 nat_tuple.src.l3num = nf_ct_l3num(master);
2238                 nat_tuple.src.u3 = exp->saved_addr;
2239                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2240                 nat_tuple.src.u = exp->saved_proto;
2241
2242                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2243                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2244                         goto nla_put_failure;
2245                 nla_nest_end(skb, nest_parms);
2246         }
2247 #endif
2248         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2249             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2250             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2251             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2252                 goto nla_put_failure;
2253         help = nfct_help(master);
2254         if (help) {
2255                 struct nf_conntrack_helper *helper;
2256
2257                 helper = rcu_dereference(help->helper);
2258                 if (helper &&
2259                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2260                         goto nla_put_failure;
2261         }
2262         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2263         if (expfn != NULL &&
2264             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2265                 goto nla_put_failure;
2266
2267         return 0;
2268
2269 nla_put_failure:
2270         return -1;
2271 }
2272
2273 static int
2274 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2275                         int event, const struct nf_conntrack_expect *exp)
2276 {
2277         struct nlmsghdr *nlh;
2278         struct nfgenmsg *nfmsg;
2279         unsigned int flags = portid ? NLM_F_MULTI : 0;
2280
2281         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2282         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2283         if (nlh == NULL)
2284                 goto nlmsg_failure;
2285
2286         nfmsg = nlmsg_data(nlh);
2287         nfmsg->nfgen_family = exp->tuple.src.l3num;
2288         nfmsg->version      = NFNETLINK_V0;
2289         nfmsg->res_id       = 0;
2290
2291         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2292                 goto nla_put_failure;
2293
2294         nlmsg_end(skb, nlh);
2295         return skb->len;
2296
2297 nlmsg_failure:
2298 nla_put_failure:
2299         nlmsg_cancel(skb, nlh);
2300         return -1;
2301 }
2302
2303 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2304 static int
2305 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2306 {
2307         struct nf_conntrack_expect *exp = item->exp;
2308         struct net *net = nf_ct_exp_net(exp);
2309         struct nlmsghdr *nlh;
2310         struct nfgenmsg *nfmsg;
2311         struct sk_buff *skb;
2312         unsigned int type, group;
2313         int flags = 0;
2314
2315         if (events & (1 << IPEXP_DESTROY)) {
2316                 type = IPCTNL_MSG_EXP_DELETE;
2317                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2318         } else if (events & (1 << IPEXP_NEW)) {
2319                 type = IPCTNL_MSG_EXP_NEW;
2320                 flags = NLM_F_CREATE|NLM_F_EXCL;
2321                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2322         } else
2323                 return 0;
2324
2325         if (!item->report && !nfnetlink_has_listeners(net, group))
2326                 return 0;
2327
2328         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2329         if (skb == NULL)
2330                 goto errout;
2331
2332         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2333         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2334         if (nlh == NULL)
2335                 goto nlmsg_failure;
2336
2337         nfmsg = nlmsg_data(nlh);
2338         nfmsg->nfgen_family = exp->tuple.src.l3num;
2339         nfmsg->version      = NFNETLINK_V0;
2340         nfmsg->res_id       = 0;
2341
2342         rcu_read_lock();
2343         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2344                 goto nla_put_failure;
2345         rcu_read_unlock();
2346
2347         nlmsg_end(skb, nlh);
2348         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2349         return 0;
2350
2351 nla_put_failure:
2352         rcu_read_unlock();
2353         nlmsg_cancel(skb, nlh);
2354 nlmsg_failure:
2355         kfree_skb(skb);
2356 errout:
2357         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2358         return 0;
2359 }
2360 #endif
2361 static int ctnetlink_exp_done(struct netlink_callback *cb)
2362 {
2363         if (cb->args[1])
2364                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2365         return 0;
2366 }
2367
2368 static int
2369 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2370 {
2371         struct net *net = sock_net(skb->sk);
2372         struct nf_conntrack_expect *exp, *last;
2373         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2374         u_int8_t l3proto = nfmsg->nfgen_family;
2375
2376         rcu_read_lock();
2377         last = (struct nf_conntrack_expect *)cb->args[1];
2378         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2379 restart:
2380                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2381                                      hnode) {
2382                         if (l3proto && exp->tuple.src.l3num != l3proto)
2383                                 continue;
2384                         if (cb->args[1]) {
2385                                 if (exp != last)
2386                                         continue;
2387                                 cb->args[1] = 0;
2388                         }
2389                         if (ctnetlink_exp_fill_info(skb,
2390                                                     NETLINK_CB(cb->skb).portid,
2391                                                     cb->nlh->nlmsg_seq,
2392                                                     IPCTNL_MSG_EXP_NEW,
2393                                                     exp) < 0) {
2394                                 if (!atomic_inc_not_zero(&exp->use))
2395                                         continue;
2396                                 cb->args[1] = (unsigned long)exp;
2397                                 goto out;
2398                         }
2399                 }
2400                 if (cb->args[1]) {
2401                         cb->args[1] = 0;
2402                         goto restart;
2403                 }
2404         }
2405 out:
2406         rcu_read_unlock();
2407         if (last)
2408                 nf_ct_expect_put(last);
2409
2410         return skb->len;
2411 }
2412
2413 static int
2414 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2415 {
2416         struct nf_conntrack_expect *exp, *last;
2417         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2418         struct nf_conn *ct = cb->data;
2419         struct nf_conn_help *help = nfct_help(ct);
2420         u_int8_t l3proto = nfmsg->nfgen_family;
2421
2422         if (cb->args[0])
2423                 return 0;
2424
2425         rcu_read_lock();
2426         last = (struct nf_conntrack_expect *)cb->args[1];
2427 restart:
2428         hlist_for_each_entry(exp, &help->expectations, lnode) {
2429                 if (l3proto && exp->tuple.src.l3num != l3proto)
2430                         continue;
2431                 if (cb->args[1]) {
2432                         if (exp != last)
2433                                 continue;
2434                         cb->args[1] = 0;
2435                 }
2436                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2437                                             cb->nlh->nlmsg_seq,
2438                                             IPCTNL_MSG_EXP_NEW,
2439                                             exp) < 0) {
2440                         if (!atomic_inc_not_zero(&exp->use))
2441                                 continue;
2442                         cb->args[1] = (unsigned long)exp;
2443                         goto out;
2444                 }
2445         }
2446         if (cb->args[1]) {
2447                 cb->args[1] = 0;
2448                 goto restart;
2449         }
2450         cb->args[0] = 1;
2451 out:
2452         rcu_read_unlock();
2453         if (last)
2454                 nf_ct_expect_put(last);
2455
2456         return skb->len;
2457 }
2458
2459 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2460                                  const struct nlmsghdr *nlh,
2461                                  const struct nlattr * const cda[])
2462 {
2463         int err;
2464         struct net *net = sock_net(ctnl);
2465         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2466         u_int8_t u3 = nfmsg->nfgen_family;
2467         struct nf_conntrack_tuple tuple;
2468         struct nf_conntrack_tuple_hash *h;
2469         struct nf_conn *ct;
2470         u16 zone = 0;
2471         struct netlink_dump_control c = {
2472                 .dump = ctnetlink_exp_ct_dump_table,
2473                 .done = ctnetlink_exp_done,
2474         };
2475
2476         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2477         if (err < 0)
2478                 return err;
2479
2480         if (cda[CTA_EXPECT_ZONE]) {
2481                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2482                 if (err < 0)
2483                         return err;
2484         }
2485
2486         h = nf_conntrack_find_get(net, zone, &tuple);
2487         if (!h)
2488                 return -ENOENT;
2489
2490         ct = nf_ct_tuplehash_to_ctrack(h);
2491         c.data = ct;
2492
2493         err = netlink_dump_start(ctnl, skb, nlh, &c);
2494         nf_ct_put(ct);
2495
2496         return err;
2497 }
2498
2499 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2500         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2501         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2502         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2503         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2504         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2505         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2506                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2507         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2508         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2509         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2510         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2511         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2512 };
2513
2514 static int
2515 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2516                      const struct nlmsghdr *nlh,
2517                      const struct nlattr * const cda[])
2518 {
2519         struct net *net = sock_net(ctnl);
2520         struct nf_conntrack_tuple tuple;
2521         struct nf_conntrack_expect *exp;
2522         struct sk_buff *skb2;
2523         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2524         u_int8_t u3 = nfmsg->nfgen_family;
2525         u16 zone;
2526         int err;
2527
2528         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2529                 if (cda[CTA_EXPECT_MASTER])
2530                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2531                 else {
2532                         struct netlink_dump_control c = {
2533                                 .dump = ctnetlink_exp_dump_table,
2534                                 .done = ctnetlink_exp_done,
2535                         };
2536                         return netlink_dump_start(ctnl, skb, nlh, &c);
2537                 }
2538         }
2539
2540         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2541         if (err < 0)
2542                 return err;
2543
2544         if (cda[CTA_EXPECT_TUPLE])
2545                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2546         else if (cda[CTA_EXPECT_MASTER])
2547                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2548         else
2549                 return -EINVAL;
2550
2551         if (err < 0)
2552                 return err;
2553
2554         exp = nf_ct_expect_find_get(net, zone, &tuple);
2555         if (!exp)
2556                 return -ENOENT;
2557
2558         if (cda[CTA_EXPECT_ID]) {
2559                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2560                 if (ntohl(id) != (u32)(unsigned long)exp) {
2561                         nf_ct_expect_put(exp);
2562                         return -ENOENT;
2563                 }
2564         }
2565
2566         err = -ENOMEM;
2567         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2568         if (skb2 == NULL) {
2569                 nf_ct_expect_put(exp);
2570                 goto out;
2571         }
2572
2573         rcu_read_lock();
2574         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2575                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2576         rcu_read_unlock();
2577         nf_ct_expect_put(exp);
2578         if (err <= 0)
2579                 goto free;
2580
2581         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2582         if (err < 0)
2583                 goto out;
2584
2585         return 0;
2586
2587 free:
2588         kfree_skb(skb2);
2589 out:
2590         /* this avoids a loop in nfnetlink. */
2591         return err == -EAGAIN ? -ENOBUFS : err;
2592 }
2593
2594 static int
2595 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2596                      const struct nlmsghdr *nlh,
2597                      const struct nlattr * const cda[])
2598 {
2599         struct net *net = sock_net(ctnl);
2600         struct nf_conntrack_expect *exp;
2601         struct nf_conntrack_tuple tuple;
2602         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2603         struct hlist_node *next;
2604         u_int8_t u3 = nfmsg->nfgen_family;
2605         unsigned int i;
2606         u16 zone;
2607         int err;
2608
2609         if (cda[CTA_EXPECT_TUPLE]) {
2610                 /* delete a single expect by tuple */
2611                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2612                 if (err < 0)
2613                         return err;
2614
2615                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2616                 if (err < 0)
2617                         return err;
2618
2619                 /* bump usage count to 2 */
2620                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2621                 if (!exp)
2622                         return -ENOENT;
2623
2624                 if (cda[CTA_EXPECT_ID]) {
2625                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2626                         if (ntohl(id) != (u32)(unsigned long)exp) {
2627                                 nf_ct_expect_put(exp);
2628                                 return -ENOENT;
2629                         }
2630                 }
2631
2632                 /* after list removal, usage count == 1 */
2633                 spin_lock_bh(&nf_conntrack_lock);
2634                 if (del_timer(&exp->timeout)) {
2635                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2636                                                    nlmsg_report(nlh));
2637                         nf_ct_expect_put(exp);
2638                 }
2639                 spin_unlock_bh(&nf_conntrack_lock);
2640                 /* have to put what we 'get' above.
2641                  * after this line usage count == 0 */
2642                 nf_ct_expect_put(exp);
2643         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2644                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2645                 struct nf_conn_help *m_help;
2646
2647                 /* delete all expectations for this helper */
2648                 spin_lock_bh(&nf_conntrack_lock);
2649                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2650                         hlist_for_each_entry_safe(exp, next,
2651                                                   &net->ct.expect_hash[i],
2652                                                   hnode) {
2653                                 m_help = nfct_help(exp->master);
2654                                 if (!strcmp(m_help->helper->name, name) &&
2655                                     del_timer(&exp->timeout)) {
2656                                         nf_ct_unlink_expect_report(exp,
2657                                                         NETLINK_CB(skb).portid,
2658                                                         nlmsg_report(nlh));
2659                                         nf_ct_expect_put(exp);
2660                                 }
2661                         }
2662                 }
2663                 spin_unlock_bh(&nf_conntrack_lock);
2664         } else {
2665                 /* This basically means we have to flush everything*/
2666                 spin_lock_bh(&nf_conntrack_lock);
2667                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2668                         hlist_for_each_entry_safe(exp, next,
2669                                                   &net->ct.expect_hash[i],
2670                                                   hnode) {
2671                                 if (del_timer(&exp->timeout)) {
2672                                         nf_ct_unlink_expect_report(exp,
2673                                                         NETLINK_CB(skb).portid,
2674                                                         nlmsg_report(nlh));
2675                                         nf_ct_expect_put(exp);
2676                                 }
2677                         }
2678                 }
2679                 spin_unlock_bh(&nf_conntrack_lock);
2680         }
2681
2682         return 0;
2683 }
2684 static int
2685 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2686                         const struct nlattr * const cda[])
2687 {
2688         if (cda[CTA_EXPECT_TIMEOUT]) {
2689                 if (!del_timer(&x->timeout))
2690                         return -ETIME;
2691
2692                 x->timeout.expires = jiffies +
2693                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2694                 add_timer(&x->timeout);
2695         }
2696         return 0;
2697 }
2698
2699 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2700         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2701         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2702 };
2703
2704 static int
2705 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2706                            struct nf_conntrack_expect *exp,
2707                            u_int8_t u3)
2708 {
2709 #ifdef CONFIG_NF_NAT_NEEDED
2710         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2711         struct nf_conntrack_tuple nat_tuple = {};
2712         int err;
2713
2714         nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2715
2716         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2717                 return -EINVAL;
2718
2719         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2720                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2721         if (err < 0)
2722                 return err;
2723
2724         exp->saved_addr = nat_tuple.src.u3;
2725         exp->saved_proto = nat_tuple.src.u;
2726         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2727
2728         return 0;
2729 #else
2730         return -EOPNOTSUPP;
2731 #endif
2732 }
2733
2734 static int
2735 ctnetlink_create_expect(struct net *net, u16 zone,
2736                         const struct nlattr * const cda[],
2737                         u_int8_t u3,
2738                         u32 portid, int report)
2739 {
2740         struct nf_conntrack_tuple tuple, mask, master_tuple;
2741         struct nf_conntrack_tuple_hash *h = NULL;
2742         struct nf_conntrack_expect *exp;
2743         struct nf_conn *ct;
2744         struct nf_conn_help *help;
2745         struct nf_conntrack_helper *helper = NULL;
2746         u_int32_t class = 0;
2747         int err = 0;
2748
2749         /* caller guarantees that those three CTA_EXPECT_* exist */
2750         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2751         if (err < 0)
2752                 return err;
2753         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2754         if (err < 0)
2755                 return err;
2756         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2757         if (err < 0)
2758                 return err;
2759
2760         /* Look for master conntrack of this expectation */
2761         h = nf_conntrack_find_get(net, zone, &master_tuple);
2762         if (!h)
2763                 return -ENOENT;
2764         ct = nf_ct_tuplehash_to_ctrack(h);
2765
2766         /* Look for helper of this expectation */
2767         if (cda[CTA_EXPECT_HELP_NAME]) {
2768                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2769
2770                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2771                                                     nf_ct_protonum(ct));
2772                 if (helper == NULL) {
2773 #ifdef CONFIG_MODULES
2774                         if (request_module("nfct-helper-%s", helpname) < 0) {
2775                                 err = -EOPNOTSUPP;
2776                                 goto out;
2777                         }
2778
2779                         helper = __nf_conntrack_helper_find(helpname,
2780                                                             nf_ct_l3num(ct),
2781                                                             nf_ct_protonum(ct));
2782                         if (helper) {
2783                                 err = -EAGAIN;
2784                                 goto out;
2785                         }
2786 #endif
2787                         err = -EOPNOTSUPP;
2788                         goto out;
2789                 }
2790         }
2791
2792         if (cda[CTA_EXPECT_CLASS] && helper) {
2793                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2794                 if (class > helper->expect_class_max) {
2795                         err = -EINVAL;
2796                         goto out;
2797                 }
2798         }
2799         exp = nf_ct_expect_alloc(ct);
2800         if (!exp) {
2801                 err = -ENOMEM;
2802                 goto out;
2803         }
2804         help = nfct_help(ct);
2805         if (!help) {
2806                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2807                         err = -EINVAL;
2808                         goto err_out;
2809                 }
2810                 exp->timeout.expires =
2811                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2812
2813                 exp->flags = NF_CT_EXPECT_USERSPACE;
2814                 if (cda[CTA_EXPECT_FLAGS]) {
2815                         exp->flags |=
2816                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2817                 }
2818         } else {
2819                 if (cda[CTA_EXPECT_FLAGS]) {
2820                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2821                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2822                 } else
2823                         exp->flags = 0;
2824         }
2825         if (cda[CTA_EXPECT_FN]) {
2826                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2827                 struct nf_ct_helper_expectfn *expfn;
2828
2829                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2830                 if (expfn == NULL) {
2831                         err = -EINVAL;
2832                         goto err_out;
2833                 }
2834                 exp->expectfn = expfn->expectfn;
2835         } else
2836                 exp->expectfn = NULL;
2837
2838         exp->class = class;
2839         exp->master = ct;
2840         exp->helper = helper;
2841         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2842         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2843         exp->mask.src.u.all = mask.src.u.all;
2844
2845         if (cda[CTA_EXPECT_NAT]) {
2846                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2847                                                  exp, u3);
2848                 if (err < 0)
2849                         goto err_out;
2850         }
2851         err = nf_ct_expect_related_report(exp, portid, report);
2852 err_out:
2853         nf_ct_expect_put(exp);
2854 out:
2855         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2856         return err;
2857 }
2858
2859 static int
2860 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2861                      const struct nlmsghdr *nlh,
2862                      const struct nlattr * const cda[])
2863 {
2864         struct net *net = sock_net(ctnl);
2865         struct nf_conntrack_tuple tuple;
2866         struct nf_conntrack_expect *exp;
2867         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2868         u_int8_t u3 = nfmsg->nfgen_family;
2869         u16 zone;
2870         int err;
2871
2872         if (!cda[CTA_EXPECT_TUPLE]
2873             || !cda[CTA_EXPECT_MASK]
2874             || !cda[CTA_EXPECT_MASTER])
2875                 return -EINVAL;
2876
2877         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2878         if (err < 0)
2879                 return err;
2880
2881         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2882         if (err < 0)
2883                 return err;
2884
2885         spin_lock_bh(&nf_conntrack_lock);
2886         exp = __nf_ct_expect_find(net, zone, &tuple);
2887
2888         if (!exp) {
2889                 spin_unlock_bh(&nf_conntrack_lock);
2890                 err = -ENOENT;
2891                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2892                         err = ctnetlink_create_expect(net, zone, cda,
2893                                                       u3,
2894                                                       NETLINK_CB(skb).portid,
2895                                                       nlmsg_report(nlh));
2896                 }
2897                 return err;
2898         }
2899
2900         err = -EEXIST;
2901         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2902                 err = ctnetlink_change_expect(exp, cda);
2903         spin_unlock_bh(&nf_conntrack_lock);
2904
2905         return err;
2906 }
2907
2908 static int
2909 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
2910                              const struct ip_conntrack_stat *st)
2911 {
2912         struct nlmsghdr *nlh;
2913         struct nfgenmsg *nfmsg;
2914         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2915
2916         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
2917         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2918         if (nlh == NULL)
2919                 goto nlmsg_failure;
2920
2921         nfmsg = nlmsg_data(nlh);
2922         nfmsg->nfgen_family = AF_UNSPEC;
2923         nfmsg->version      = NFNETLINK_V0;
2924         nfmsg->res_id       = htons(cpu);
2925
2926         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
2927             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
2928             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
2929                 goto nla_put_failure;
2930
2931         nlmsg_end(skb, nlh);
2932         return skb->len;
2933
2934 nla_put_failure:
2935 nlmsg_failure:
2936         nlmsg_cancel(skb, nlh);
2937         return -1;
2938 }
2939
2940 static int
2941 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2942 {
2943         int cpu;
2944         struct net *net = sock_net(skb->sk);
2945
2946         if (cb->args[0] == nr_cpu_ids)
2947                 return 0;
2948
2949         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2950                 const struct ip_conntrack_stat *st;
2951
2952                 if (!cpu_possible(cpu))
2953                         continue;
2954
2955                 st = per_cpu_ptr(net->ct.stat, cpu);
2956                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
2957                                                  cb->nlh->nlmsg_seq,
2958                                                  cpu, st) < 0)
2959                         break;
2960         }
2961         cb->args[0] = cpu;
2962
2963         return skb->len;
2964 }
2965
2966 static int
2967 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
2968                        const struct nlmsghdr *nlh,
2969                        const struct nlattr * const cda[])
2970 {
2971         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2972                 struct netlink_dump_control c = {
2973                         .dump = ctnetlink_exp_stat_cpu_dump,
2974                 };
2975                 return netlink_dump_start(ctnl, skb, nlh, &c);
2976         }
2977
2978         return 0;
2979 }
2980
2981 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2982 static struct nf_ct_event_notifier ctnl_notifier = {
2983         .fcn = ctnetlink_conntrack_event,
2984 };
2985
2986 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2987         .fcn = ctnetlink_expect_event,
2988 };
2989 #endif
2990
2991 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2992         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2993                                             .attr_count = CTA_MAX,
2994                                             .policy = ct_nla_policy },
2995         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2996                                             .attr_count = CTA_MAX,
2997                                             .policy = ct_nla_policy },
2998         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2999                                             .attr_count = CTA_MAX,
3000                                             .policy = ct_nla_policy },
3001         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3002                                             .attr_count = CTA_MAX,
3003                                             .policy = ct_nla_policy },
3004         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3005         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3006         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3007         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3008 };
3009
3010 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3011         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3012                                             .attr_count = CTA_EXPECT_MAX,
3013                                             .policy = exp_nla_policy },
3014         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3015                                             .attr_count = CTA_EXPECT_MAX,
3016                                             .policy = exp_nla_policy },
3017         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3018                                             .attr_count = CTA_EXPECT_MAX,
3019                                             .policy = exp_nla_policy },
3020         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3021 };
3022
3023 static const struct nfnetlink_subsystem ctnl_subsys = {
3024         .name                           = "conntrack",
3025         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3026         .cb_count                       = IPCTNL_MSG_MAX,
3027         .cb                             = ctnl_cb,
3028 };
3029
3030 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3031         .name                           = "conntrack_expect",
3032         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3033         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3034         .cb                             = ctnl_exp_cb,
3035 };
3036
3037 MODULE_ALIAS("ip_conntrack_netlink");
3038 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3039 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3040
3041 static int __net_init ctnetlink_net_init(struct net *net)
3042 {
3043 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3044         int ret;
3045
3046         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3047         if (ret < 0) {
3048                 pr_err("ctnetlink_init: cannot register notifier.\n");
3049                 goto err_out;
3050         }
3051
3052         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3053         if (ret < 0) {
3054                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3055                 goto err_unreg_notifier;
3056         }
3057 #endif
3058         return 0;
3059
3060 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3061 err_unreg_notifier:
3062         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3063 err_out:
3064         return ret;
3065 #endif
3066 }
3067
3068 static void ctnetlink_net_exit(struct net *net)
3069 {
3070 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3071         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3072         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3073 #endif
3074 }
3075
3076 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3077 {
3078         struct net *net;
3079
3080         list_for_each_entry(net, net_exit_list, exit_list)
3081                 ctnetlink_net_exit(net);
3082 }
3083
3084 static struct pernet_operations ctnetlink_net_ops = {
3085         .init           = ctnetlink_net_init,
3086         .exit_batch     = ctnetlink_net_exit_batch,
3087 };
3088
3089 static int __init ctnetlink_init(void)
3090 {
3091         int ret;
3092
3093         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3094         ret = nfnetlink_subsys_register(&ctnl_subsys);
3095         if (ret < 0) {
3096                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3097                 goto err_out;
3098         }
3099
3100         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3101         if (ret < 0) {
3102                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3103                 goto err_unreg_subsys;
3104         }
3105
3106         ret = register_pernet_subsys(&ctnetlink_net_ops);
3107         if (ret < 0) {
3108                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3109                 goto err_unreg_exp_subsys;
3110         }
3111 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3112         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3113         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3114 #endif
3115         return 0;
3116
3117 err_unreg_exp_subsys:
3118         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3119 err_unreg_subsys:
3120         nfnetlink_subsys_unregister(&ctnl_subsys);
3121 err_out:
3122         return ret;
3123 }
3124
3125 static void __exit ctnetlink_exit(void)
3126 {
3127         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3128
3129         unregister_pernet_subsys(&ctnetlink_net_ops);
3130         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3131         nfnetlink_subsys_unregister(&ctnl_subsys);
3132 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3133         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3134 #endif
3135 }
3136
3137 module_init(ctnetlink_init);
3138 module_exit(ctnetlink_exit);