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         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 832         if (ret < 0)
 833                 return ret;
 834
 835         rcu_read_lock();
 836         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 837
 838         if (likely(l3proto->nlattr_to_tuple)) {
 839                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 840                                           l3proto->nla_policy);
 841                 if (ret == 0)
 842                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 843         }
 844
 845         rcu_read_unlock();
 846
 847         return ret;
 848 }
 849
 850 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 851         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 852 };
 853
 854 static inline int
 855 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 856                             struct nf_conntrack_tuple *tuple)
 857 {
 858         struct nlattr *tb[CTA_PROTO_MAX+1];
 859         struct nf_conntrack_l4proto *l4proto;
 860         int ret = 0;
 861
 862         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 863         if (ret < 0)
 864                 return ret;
 865
 866         if (!tb[CTA_PROTO_NUM])
 867                 return -EINVAL;
 868         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 869
 870         rcu_read_lock();
 871         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 872
 873         if (likely(l4proto->nlattr_to_tuple)) {
 874                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 875                                           l4proto->nla_policy);
 876                 if (ret == 0)
 877                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 878         }
 879
 880         rcu_read_unlock();
 881
 882         return ret;
 883 }
 884
 885 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 886         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 887         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 888 };
 889
 890 static int
 891 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 892                       struct nf_conntrack_tuple *tuple,
 893                       enum ctattr_type type, u_int8_t l3num)
 894 {
 895         struct nlattr *tb[CTA_TUPLE_MAX+1];
 896         int err;
 897
 898         memset(tuple, 0, sizeof(*tuple));
 899
 900         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 901         if (err < 0)
 902                 return err;
 903
 904         if (!tb[CTA_TUPLE_IP])
 905                 return -EINVAL;
 906
 907         tuple->src.l3num = l3num;
 908
 909         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 910         if (err < 0)
 911                 return err;
 912
 913         if (!tb[CTA_TUPLE_PROTO])
 914                 return -EINVAL;
 915
 916         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 917         if (err < 0)
 918                 return err;
 919
 920         /* orig and expect tuples get DIR_ORIGINAL */
 921         if (type == CTA_TUPLE_REPLY)
 922                 tuple->dst.dir = IP_CT_DIR_REPLY;
 923         else
 924                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 925
 926         return 0;
 927 }
 928
 929 static int
 930 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 931 {
 932         if (attr)
 933 #ifdef CONFIG_NF_CONNTRACK_ZONES
 934                 *zone = ntohs(nla_get_be16(attr));
 935 #else
 936                 return -EOPNOTSUPP;
 937 #endif
 938         else
 939                 *zone = 0;
 940
 941         return 0;
 942 }
 943
 944 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 945         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 946                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 947 };
 948
 949 static inline int
 950 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 951                      struct nlattr **helpinfo)
 952 {
 953         int err;
 954         struct nlattr *tb[CTA_HELP_MAX+1];
 955
 956         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 957         if (err < 0)
 958                 return err;
 959
 960         if (!tb[CTA_HELP_NAME])
 961                 return -EINVAL;
 962
 963         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 964
 965         if (tb[CTA_HELP_INFO])
 966                 *helpinfo = tb[CTA_HELP_INFO];
 967
 968         return 0;
 969 }
 970
 971 #define __CTA_LABELS_MAX_LENGTH ((XT_CONNLABEL_MAXBIT + 1) / BITS_PER_BYTE)
 972 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 973         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 974         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 975         [CTA_STATUS]            = { .type = NLA_U32 },
 976         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 977         [CTA_HELP]              = { .type = NLA_NESTED },
 978         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 979         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 980         [CTA_MARK]              = { .type = NLA_U32 },
 981         [CTA_ID]                = { .type = NLA_U32 },
 982         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 983         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 984         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
 985         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
 986         [CTA_ZONE]              = { .type = NLA_U16 },
 987         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 988         [CTA_LABELS]            = { .type = NLA_BINARY,
 989                                     .len = __CTA_LABELS_MAX_LENGTH },
 990         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
 991                                     .len = __CTA_LABELS_MAX_LENGTH },
 992 };
 993
 994 static int
 995 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 996                         const struct nlmsghdr *nlh,
 997                         const struct nlattr * const cda[])
 998 {
 999         struct net *net = sock_net(ctnl);
1000         struct nf_conntrack_tuple_hash *h;
1001         struct nf_conntrack_tuple tuple;
1002         struct nf_conn *ct;
1003         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1004         u_int8_t u3 = nfmsg->nfgen_family;
1005         u16 zone;
1006         int err;
1007
1008         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1009         if (err < 0)
1010                 return err;
1011
1012         if (cda[CTA_TUPLE_ORIG])
1013                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1014         else if (cda[CTA_TUPLE_REPLY])
1015                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1016         else {
1017                 /* Flush the whole table */
1018                 nf_conntrack_flush_report(net,
1019                                          NETLINK_CB(skb).portid,
1020                                          nlmsg_report(nlh));
1021                 return 0;
1022         }
1023
1024         if (err < 0)
1025                 return err;
1026
1027         h = nf_conntrack_find_get(net, zone, &tuple);
1028         if (!h)
1029                 return -ENOENT;
1030
1031         ct = nf_ct_tuplehash_to_ctrack(h);
1032
1033         if (cda[CTA_ID]) {
1034                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1035                 if (id != (u32)(unsigned long)ct) {
1036                         nf_ct_put(ct);
1037                         return -ENOENT;
1038                 }
1039         }
1040
1041         if (del_timer(&ct->timeout)) {
1042                 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
1043                                               NETLINK_CB(skb).portid,
1044                                               nlmsg_report(nlh)) < 0) {
1045                         nf_ct_delete_from_lists(ct);
1046                         /* we failed to report the event, try later */
1047                         nf_ct_dying_timeout(ct);
1048                         nf_ct_put(ct);
1049                         return 0;
1050                 }
1051                 /* death_by_timeout would report the event again */
1052                 set_bit(IPS_DYING_BIT, &ct->status);
1053                 nf_ct_delete_from_lists(ct);
1054                 nf_ct_put(ct);
1055         }
1056         nf_ct_put(ct);
1057
1058         return 0;
1059 }
1060
1061 static int
1062 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1063                         const struct nlmsghdr *nlh,
1064                         const struct nlattr * const cda[])
1065 {
1066         struct net *net = sock_net(ctnl);
1067         struct nf_conntrack_tuple_hash *h;
1068         struct nf_conntrack_tuple tuple;
1069         struct nf_conn *ct;
1070         struct sk_buff *skb2 = NULL;
1071         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1072         u_int8_t u3 = nfmsg->nfgen_family;
1073         u16 zone;
1074         int err;
1075
1076         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1077                 struct netlink_dump_control c = {
1078                         .dump = ctnetlink_dump_table,
1079                         .done = ctnetlink_done,
1080                 };
1081 #ifdef CONFIG_NF_CONNTRACK_MARK
1082                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1083                         struct ctnetlink_dump_filter *filter;
1084
1085                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1086                                          GFP_ATOMIC);
1087                         if (filter == NULL)
1088                                 return -ENOMEM;
1089
1090                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1091                         filter->mark.mask =
1092                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1093                         c.data = filter;
1094                 }
1095 #endif
1096                 return netlink_dump_start(ctnl, skb, nlh, &c);
1097         }
1098
1099         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1100         if (err < 0)
1101                 return err;
1102
1103         if (cda[CTA_TUPLE_ORIG])
1104                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1105         else if (cda[CTA_TUPLE_REPLY])
1106                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1107         else
1108                 return -EINVAL;
1109
1110         if (err < 0)
1111                 return err;
1112
1113         h = nf_conntrack_find_get(net, zone, &tuple);
1114         if (!h)
1115                 return -ENOENT;
1116
1117         ct = nf_ct_tuplehash_to_ctrack(h);
1118
1119         err = -ENOMEM;
1120         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1121         if (skb2 == NULL) {
1122                 nf_ct_put(ct);
1123                 return -ENOMEM;
1124         }
1125
1126         rcu_read_lock();
1127         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1128                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1129         rcu_read_unlock();
1130         nf_ct_put(ct);
1131         if (err <= 0)
1132                 goto free;
1133
1134         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1135         if (err < 0)
1136                 goto out;
1137
1138         return 0;
1139
1140 free:
1141         kfree_skb(skb2);
1142 out:
1143         /* this avoids a loop in nfnetlink. */
1144         return err == -EAGAIN ? -ENOBUFS : err;
1145 }
1146
1147 static int ctnetlink_done_list(struct netlink_callback *cb)
1148 {
1149         if (cb->args[1])
1150                 nf_ct_put((struct nf_conn *)cb->args[1]);
1151         return 0;
1152 }
1153
1154 static int
1155 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb,
1156                     struct hlist_nulls_head *list)
1157 {
1158         struct nf_conn *ct, *last;
1159         struct nf_conntrack_tuple_hash *h;
1160         struct hlist_nulls_node *n;
1161         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1162         u_int8_t l3proto = nfmsg->nfgen_family;
1163         int res;
1164
1165         if (cb->args[2])
1166                 return 0;
1167
1168         spin_lock_bh(&nf_conntrack_lock);
1169         last = (struct nf_conn *)cb->args[1];
1170 restart:
1171         hlist_nulls_for_each_entry(h, n, list, hnnode) {
1172                 ct = nf_ct_tuplehash_to_ctrack(h);
1173                 if (l3proto && nf_ct_l3num(ct) != l3proto)
1174                         continue;
1175                 if (cb->args[1]) {
1176                         if (ct != last)
1177                                 continue;
1178                         cb->args[1] = 0;
1179                 }
1180                 rcu_read_lock();
1181                 res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1182                                           cb->nlh->nlmsg_seq,
1183                                           NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1184                                           ct);
1185                 rcu_read_unlock();
1186                 if (res < 0) {
1187                         nf_conntrack_get(&ct->ct_general);
1188                         cb->args[1] = (unsigned long)ct;
1189                         goto out;
1190                 }
1191         }
1192         if (cb->args[1]) {
1193                 cb->args[1] = 0;
1194                 goto restart;
1195         } else
1196                 cb->args[2] = 1;
1197 out:
1198         spin_unlock_bh(&nf_conntrack_lock);
1199         if (last)
1200                 nf_ct_put(last);
1201
1202         return skb->len;
1203 }
1204
1205 static int
1206 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1207 {
1208         struct net *net = sock_net(skb->sk);
1209
1210         return ctnetlink_dump_list(skb, cb, &net->ct.dying);
1211 }
1212
1213 static int
1214 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1215                        const struct nlmsghdr *nlh,
1216                        const struct nlattr * const cda[])
1217 {
1218         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1219                 struct netlink_dump_control c = {
1220                         .dump = ctnetlink_dump_dying,
1221                         .done = ctnetlink_done_list,
1222                 };
1223                 return netlink_dump_start(ctnl, skb, nlh, &c);
1224         }
1225
1226         return -EOPNOTSUPP;
1227 }
1228
1229 static int
1230 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1231 {
1232         struct net *net = sock_net(skb->sk);
1233
1234         return ctnetlink_dump_list(skb, cb, &net->ct.unconfirmed);
1235 }
1236
1237 static int
1238 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1239                              const struct nlmsghdr *nlh,
1240                              const struct nlattr * const cda[])
1241 {
1242         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1243                 struct netlink_dump_control c = {
1244                         .dump = ctnetlink_dump_unconfirmed,
1245                         .done = ctnetlink_done_list,
1246                 };
1247                 return netlink_dump_start(ctnl, skb, nlh, &c);
1248         }
1249
1250         return -EOPNOTSUPP;
1251 }
1252
1253 #ifdef CONFIG_NF_NAT_NEEDED
1254 static int
1255 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1256                           enum nf_nat_manip_type manip,
1257                           const struct nlattr *attr)
1258 {
1259         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1260         int err;
1261
1262         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1263         if (!parse_nat_setup) {
1264 #ifdef CONFIG_MODULES
1265                 rcu_read_unlock();
1266                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1267                 if (request_module("nf-nat") < 0) {
1268                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1269                         rcu_read_lock();
1270                         return -EOPNOTSUPP;
1271                 }
1272                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1273                 rcu_read_lock();
1274                 if (nfnetlink_parse_nat_setup_hook)
1275                         return -EAGAIN;
1276 #endif
1277                 return -EOPNOTSUPP;
1278         }
1279
1280         err = parse_nat_setup(ct, manip, attr);
1281         if (err == -EAGAIN) {
1282 #ifdef CONFIG_MODULES
1283                 rcu_read_unlock();
1284                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1285                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1286                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1287                         rcu_read_lock();
1288                         return -EOPNOTSUPP;
1289                 }
1290                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1291                 rcu_read_lock();
1292 #else
1293                 err = -EOPNOTSUPP;
1294 #endif
1295         }
1296         return err;
1297 }
1298 #endif
1299
1300 static int
1301 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1302 {
1303         unsigned long d;
1304         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1305         d = ct->status ^ status;
1306
1307         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1308                 /* unchangeable */
1309                 return -EBUSY;
1310
1311         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1312                 /* SEEN_REPLY bit can only be set */
1313                 return -EBUSY;
1314
1315         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1316                 /* ASSURED bit can only be set */
1317                 return -EBUSY;
1318
1319         /* Be careful here, modifying NAT bits can screw up things,
1320          * so don't let users modify them directly if they don't pass
1321          * nf_nat_range. */
1322         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1323         return 0;
1324 }
1325
1326 static int
1327 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1328 {
1329 #ifdef CONFIG_NF_NAT_NEEDED
1330         int ret;
1331
1332         if (cda[CTA_NAT_DST]) {
1333                 ret = ctnetlink_parse_nat_setup(ct,
1334                                                 NF_NAT_MANIP_DST,
1335                                                 cda[CTA_NAT_DST]);
1336                 if (ret < 0)
1337                         return ret;
1338         }
1339         if (cda[CTA_NAT_SRC]) {
1340                 ret = ctnetlink_parse_nat_setup(ct,
1341                                                 NF_NAT_MANIP_SRC,
1342                                                 cda[CTA_NAT_SRC]);
1343                 if (ret < 0)
1344                         return ret;
1345         }
1346         return 0;
1347 #else
1348         return -EOPNOTSUPP;
1349 #endif
1350 }
1351
1352 static inline int
1353 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1354 {
1355         struct nf_conntrack_helper *helper;
1356         struct nf_conn_help *help = nfct_help(ct);
1357         char *helpname = NULL;
1358         struct nlattr *helpinfo = NULL;
1359         int err;
1360
1361         /* don't change helper of sibling connections */
1362         if (ct->master)
1363                 return -EBUSY;
1364
1365         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1366         if (err < 0)
1367                 return err;
1368
1369         if (!strcmp(helpname, "")) {
1370                 if (help && help->helper) {
1371                         /* we had a helper before ... */
1372                         nf_ct_remove_expectations(ct);
1373                         RCU_INIT_POINTER(help->helper, NULL);
1374                 }
1375
1376                 return 0;
1377         }
1378
1379         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1380                                             nf_ct_protonum(ct));
1381         if (helper == NULL) {
1382 #ifdef CONFIG_MODULES
1383                 spin_unlock_bh(&nf_conntrack_lock);
1384
1385                 if (request_module("nfct-helper-%s", helpname) < 0) {
1386                         spin_lock_bh(&nf_conntrack_lock);
1387                         return -EOPNOTSUPP;
1388                 }
1389
1390                 spin_lock_bh(&nf_conntrack_lock);
1391                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1392                                                     nf_ct_protonum(ct));
1393                 if (helper)
1394                         return -EAGAIN;
1395 #endif
1396                 return -EOPNOTSUPP;
1397         }
1398
1399         if (help) {
1400                 if (help->helper == helper) {
1401                         /* update private helper data if allowed. */
1402                         if (helper->from_nlattr)
1403                                 helper->from_nlattr(helpinfo, ct);
1404                         return 0;
1405                 } else
1406                         return -EBUSY;
1407         }
1408
1409         /* we cannot set a helper for an existing conntrack */
1410         return -EOPNOTSUPP;
1411 }
1412
1413 static inline int
1414 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1415 {
1416         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1417
1418         if (!del_timer(&ct->timeout))
1419                 return -ETIME;
1420
1421         ct->timeout.expires = jiffies + timeout * HZ;
1422         add_timer(&ct->timeout);
1423
1424         return 0;
1425 }
1426
1427 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1428         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1429         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1430         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1431 };
1432
1433 static inline int
1434 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1435 {
1436         const struct nlattr *attr = cda[CTA_PROTOINFO];
1437         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1438         struct nf_conntrack_l4proto *l4proto;
1439         int err = 0;
1440
1441         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1442         if (err < 0)
1443                 return err;
1444
1445         rcu_read_lock();
1446         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1447         if (l4proto->from_nlattr)
1448                 err = l4proto->from_nlattr(tb, ct);
1449         rcu_read_unlock();
1450
1451         return err;
1452 }
1453
1454 #ifdef CONFIG_NF_NAT_NEEDED
1455 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1456         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1457         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1458         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1459 };
1460
1461 static inline int
1462 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1463 {
1464         int err;
1465         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1466
1467         err = nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1468         if (err < 0)
1469                 return err;
1470
1471         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1472                 return -EINVAL;
1473
1474         natseq->correction_pos =
1475                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1476
1477         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1478                 return -EINVAL;
1479
1480         natseq->offset_before =
1481                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1482
1483         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1484                 return -EINVAL;
1485
1486         natseq->offset_after =
1487                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1488
1489         return 0;
1490 }
1491
1492 static int
1493 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1494                              const struct nlattr * const cda[])
1495 {
1496         int ret = 0;
1497         struct nf_conn_nat *nat = nfct_nat(ct);
1498
1499         if (!nat)
1500                 return 0;
1501
1502         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1503                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1504                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1505                 if (ret < 0)
1506                         return ret;
1507
1508                 ct->status |= IPS_SEQ_ADJUST;
1509         }
1510
1511         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1512                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1513                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1514                 if (ret < 0)
1515                         return ret;
1516
1517                 ct->status |= IPS_SEQ_ADJUST;
1518         }
1519
1520         return 0;
1521 }
1522 #endif
1523
1524 static int
1525 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1526 {
1527 #ifdef CONFIG_NF_CONNTRACK_LABELS
1528         size_t len = nla_len(cda[CTA_LABELS]);
1529         const void *mask = cda[CTA_LABELS_MASK];
1530
1531         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1532                 return -EINVAL;
1533
1534         if (mask) {
1535                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1536                     nla_len(cda[CTA_LABELS_MASK]) != len)
1537                         return -EINVAL;
1538                 mask = nla_data(cda[CTA_LABELS_MASK]);
1539         }
1540
1541         len /= sizeof(u32);
1542
1543         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1544 #else
1545         return -EOPNOTSUPP;
1546 #endif
1547 }
1548
1549 static int
1550 ctnetlink_change_conntrack(struct nf_conn *ct,
1551                            const struct nlattr * const cda[])
1552 {
1553         int err;
1554
1555         /* only allow NAT changes and master assignation for new conntracks */
1556         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1557                 return -EOPNOTSUPP;
1558
1559         if (cda[CTA_HELP]) {
1560                 err = ctnetlink_change_helper(ct, cda);
1561                 if (err < 0)
1562                         return err;
1563         }
1564
1565         if (cda[CTA_TIMEOUT]) {
1566                 err = ctnetlink_change_timeout(ct, cda);
1567                 if (err < 0)
1568                         return err;
1569         }
1570
1571         if (cda[CTA_STATUS]) {
1572                 err = ctnetlink_change_status(ct, cda);
1573                 if (err < 0)
1574                         return err;
1575         }
1576
1577         if (cda[CTA_PROTOINFO]) {
1578                 err = ctnetlink_change_protoinfo(ct, cda);
1579                 if (err < 0)
1580                         return err;
1581         }
1582
1583 #if defined(CONFIG_NF_CONNTRACK_MARK)
1584         if (cda[CTA_MARK])
1585                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1586 #endif
1587
1588 #ifdef CONFIG_NF_NAT_NEEDED
1589         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1590                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1591                 if (err < 0)
1592                         return err;
1593         }
1594 #endif
1595         if (cda[CTA_LABELS]) {
1596                 err = ctnetlink_attach_labels(ct, cda);
1597                 if (err < 0)
1598                         return err;
1599         }
1600
1601         return 0;
1602 }
1603
1604 static struct nf_conn *
1605 ctnetlink_create_conntrack(struct net *net, u16 zone,
1606                            const struct nlattr * const cda[],
1607                            struct nf_conntrack_tuple *otuple,
1608                            struct nf_conntrack_tuple *rtuple,
1609                            u8 u3)
1610 {
1611         struct nf_conn *ct;
1612         int err = -EINVAL;
1613         struct nf_conntrack_helper *helper;
1614         struct nf_conn_tstamp *tstamp;
1615
1616         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1617         if (IS_ERR(ct))
1618                 return ERR_PTR(-ENOMEM);
1619
1620         if (!cda[CTA_TIMEOUT])
1621                 goto err1;
1622         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1623
1624         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1625
1626         rcu_read_lock();
1627         if (cda[CTA_HELP]) {
1628                 char *helpname = NULL;
1629                 struct nlattr *helpinfo = NULL;
1630
1631                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1632                 if (err < 0)
1633                         goto err2;
1634
1635                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1636                                                     nf_ct_protonum(ct));
1637                 if (helper == NULL) {
1638                         rcu_read_unlock();
1639 #ifdef CONFIG_MODULES
1640                         if (request_module("nfct-helper-%s", helpname) < 0) {
1641                                 err = -EOPNOTSUPP;
1642                                 goto err1;
1643                         }
1644
1645                         rcu_read_lock();
1646                         helper = __nf_conntrack_helper_find(helpname,
1647                                                             nf_ct_l3num(ct),
1648                                                             nf_ct_protonum(ct));
1649                         if (helper) {
1650                                 err = -EAGAIN;
1651                                 goto err2;
1652                         }
1653                         rcu_read_unlock();
1654 #endif
1655                         err = -EOPNOTSUPP;
1656                         goto err1;
1657                 } else {
1658                         struct nf_conn_help *help;
1659
1660                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1661                         if (help == NULL) {
1662                                 err = -ENOMEM;
1663                                 goto err2;
1664                         }
1665                         /* set private helper data if allowed. */
1666                         if (helper->from_nlattr)
1667                                 helper->from_nlattr(helpinfo, ct);
1668
1669                         /* not in hash table yet so not strictly necessary */
1670                         RCU_INIT_POINTER(help->helper, helper);
1671                 }
1672         } else {
1673                 /* try an implicit helper assignation */
1674                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1675                 if (err < 0)
1676                         goto err2;
1677         }
1678
1679         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1680                 err = ctnetlink_change_nat(ct, cda);
1681                 if (err < 0)
1682                         goto err2;
1683         }
1684
1685         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1686         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1687         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1688         nf_ct_labels_ext_add(ct);
1689
1690         /* we must add conntrack extensions before confirmation. */
1691         ct->status |= IPS_CONFIRMED;
1692
1693         if (cda[CTA_STATUS]) {
1694                 err = ctnetlink_change_status(ct, cda);
1695                 if (err < 0)
1696                         goto err2;
1697         }
1698
1699 #ifdef CONFIG_NF_NAT_NEEDED
1700         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1701                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1702                 if (err < 0)
1703                         goto err2;
1704         }
1705 #endif
1706
1707         memset(&ct->proto, 0, sizeof(ct->proto));
1708         if (cda[CTA_PROTOINFO]) {
1709                 err = ctnetlink_change_protoinfo(ct, cda);
1710                 if (err < 0)
1711                         goto err2;
1712         }
1713
1714 #if defined(CONFIG_NF_CONNTRACK_MARK)
1715         if (cda[CTA_MARK])
1716                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1717 #endif
1718
1719         /* setup master conntrack: this is a confirmed expectation */
1720         if (cda[CTA_TUPLE_MASTER]) {
1721                 struct nf_conntrack_tuple master;
1722                 struct nf_conntrack_tuple_hash *master_h;
1723                 struct nf_conn *master_ct;
1724
1725                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1726                 if (err < 0)
1727                         goto err2;
1728
1729                 master_h = nf_conntrack_find_get(net, zone, &master);
1730                 if (master_h == NULL) {
1731                         err = -ENOENT;
1732                         goto err2;
1733                 }
1734                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1735                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1736                 ct->master = master_ct;
1737         }
1738         tstamp = nf_conn_tstamp_find(ct);
1739         if (tstamp)
1740                 tstamp->start = ktime_to_ns(ktime_get_real());
1741
1742         err = nf_conntrack_hash_check_insert(ct);
1743         if (err < 0)
1744                 goto err2;
1745
1746         rcu_read_unlock();
1747
1748         return ct;
1749
1750 err2:
1751         rcu_read_unlock();
1752 err1:
1753         nf_conntrack_free(ct);
1754         return ERR_PTR(err);
1755 }
1756
1757 static int
1758 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1759                         const struct nlmsghdr *nlh,
1760                         const struct nlattr * const cda[])
1761 {
1762         struct net *net = sock_net(ctnl);
1763         struct nf_conntrack_tuple otuple, rtuple;
1764         struct nf_conntrack_tuple_hash *h = NULL;
1765         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1766         struct nf_conn *ct;
1767         u_int8_t u3 = nfmsg->nfgen_family;
1768         u16 zone;
1769         int err;
1770
1771         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1772         if (err < 0)
1773                 return err;
1774
1775         if (cda[CTA_TUPLE_ORIG]) {
1776                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1777                 if (err < 0)
1778                         return err;
1779         }
1780
1781         if (cda[CTA_TUPLE_REPLY]) {
1782                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1783                 if (err < 0)
1784                         return err;
1785         }
1786
1787         if (cda[CTA_TUPLE_ORIG])
1788                 h = nf_conntrack_find_get(net, zone, &otuple);
1789         else if (cda[CTA_TUPLE_REPLY])
1790                 h = nf_conntrack_find_get(net, zone, &rtuple);
1791
1792         if (h == NULL) {
1793                 err = -ENOENT;
1794                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1795                         enum ip_conntrack_events events;
1796
1797                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1798                                 return -EINVAL;
1799
1800                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1801                                                         &rtuple, u3);
1802                         if (IS_ERR(ct))
1803                                 return PTR_ERR(ct);
1804
1805                         err = 0;
1806                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1807                                 events = IPCT_RELATED;
1808                         else
1809                                 events = IPCT_NEW;
1810
1811                         if (cda[CTA_LABELS] &&
1812                             ctnetlink_attach_labels(ct, cda) == 0)
1813                                 events |= (1 << IPCT_LABEL);
1814
1815                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1816                                                       (1 << IPCT_ASSURED) |
1817                                                       (1 << IPCT_HELPER) |
1818                                                       (1 << IPCT_PROTOINFO) |
1819                                                       (1 << IPCT_NATSEQADJ) |
1820                                                       (1 << IPCT_MARK) | events,
1821                                                       ct, NETLINK_CB(skb).portid,
1822                                                       nlmsg_report(nlh));
1823                         nf_ct_put(ct);
1824                 }
1825
1826                 return err;
1827         }
1828         /* implicit 'else' */
1829
1830         err = -EEXIST;
1831         ct = nf_ct_tuplehash_to_ctrack(h);
1832         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1833                 spin_lock_bh(&nf_conntrack_lock);
1834                 err = ctnetlink_change_conntrack(ct, cda);
1835                 spin_unlock_bh(&nf_conntrack_lock);
1836                 if (err == 0) {
1837                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1838                                                       (1 << IPCT_ASSURED) |
1839                                                       (1 << IPCT_HELPER) |
1840                                                       (1 << IPCT_PROTOINFO) |
1841                                                       (1 << IPCT_NATSEQADJ) |
1842                                                       (1 << IPCT_MARK),
1843                                                       ct, NETLINK_CB(skb).portid,
1844                                                       nlmsg_report(nlh));
1845                 }
1846         }
1847
1848         nf_ct_put(ct);
1849         return err;
1850 }
1851
1852 static int
1853 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1854                                 __u16 cpu, const struct ip_conntrack_stat *st)
1855 {
1856         struct nlmsghdr *nlh;
1857         struct nfgenmsg *nfmsg;
1858         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1859
1860         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1861         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1862         if (nlh == NULL)
1863                 goto nlmsg_failure;
1864
1865         nfmsg = nlmsg_data(nlh);
1866         nfmsg->nfgen_family = AF_UNSPEC;
1867         nfmsg->version      = NFNETLINK_V0;
1868         nfmsg->res_id       = htons(cpu);
1869
1870         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1871             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1872             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1873             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1874             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1875             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1876             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1877             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1878             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1879                                 htonl(st->insert_failed)) ||
1880             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1881             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1882             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1883             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1884                                 htonl(st->search_restart)))
1885                 goto nla_put_failure;
1886
1887         nlmsg_end(skb, nlh);
1888         return skb->len;
1889
1890 nla_put_failure:
1891 nlmsg_failure:
1892         nlmsg_cancel(skb, nlh);
1893         return -1;
1894 }
1895
1896 static int
1897 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1898 {
1899         int cpu;
1900         struct net *net = sock_net(skb->sk);
1901
1902         if (cb->args[0] == nr_cpu_ids)
1903                 return 0;
1904
1905         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1906                 const struct ip_conntrack_stat *st;
1907
1908                 if (!cpu_possible(cpu))
1909                         continue;
1910
1911                 st = per_cpu_ptr(net->ct.stat, cpu);
1912                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1913                                                     NETLINK_CB(cb->skb).portid,
1914                                                     cb->nlh->nlmsg_seq,
1915                                                     cpu, st) < 0)
1916                                 break;
1917         }
1918         cb->args[0] = cpu;
1919
1920         return skb->len;
1921 }
1922
1923 static int
1924 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1925                       const struct nlmsghdr *nlh,
1926                       const struct nlattr * const cda[])
1927 {
1928         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1929                 struct netlink_dump_control c = {
1930                         .dump = ctnetlink_ct_stat_cpu_dump,
1931                 };
1932                 return netlink_dump_start(ctnl, skb, nlh, &c);
1933         }
1934
1935         return 0;
1936 }
1937
1938 static int
1939 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1940                             struct net *net)
1941 {
1942         struct nlmsghdr *nlh;
1943         struct nfgenmsg *nfmsg;
1944         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1945         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1946
1947         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1948         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1949         if (nlh == NULL)
1950                 goto nlmsg_failure;
1951
1952         nfmsg = nlmsg_data(nlh);
1953         nfmsg->nfgen_family = AF_UNSPEC;
1954         nfmsg->version      = NFNETLINK_V0;
1955         nfmsg->res_id       = 0;
1956
1957         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1958                 goto nla_put_failure;
1959
1960         nlmsg_end(skb, nlh);
1961         return skb->len;
1962
1963 nla_put_failure:
1964 nlmsg_failure:
1965         nlmsg_cancel(skb, nlh);
1966         return -1;
1967 }
1968
1969 static int
1970 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1971                   const struct nlmsghdr *nlh,
1972                   const struct nlattr * const cda[])
1973 {
1974         struct sk_buff *skb2;
1975         int err;
1976
1977         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1978         if (skb2 == NULL)
1979                 return -ENOMEM;
1980
1981         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1982                                           nlh->nlmsg_seq,
1983                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1984                                           sock_net(skb->sk));
1985         if (err <= 0)
1986                 goto free;
1987
1988         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1989         if (err < 0)
1990                 goto out;
1991
1992         return 0;
1993
1994 free:
1995         kfree_skb(skb2);
1996 out:
1997         /* this avoids a loop in nfnetlink. */
1998         return err == -EAGAIN ? -ENOBUFS : err;
1999 }
2000
2001 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2002 static size_t
2003 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2004 {
2005         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2006                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2007                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2008                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2009                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2010                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2011                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2012                + nla_total_size(0) /* CTA_PROTOINFO */
2013                + nla_total_size(0) /* CTA_HELP */
2014                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2015                + ctnetlink_secctx_size(ct)
2016 #ifdef CONFIG_NF_NAT_NEEDED
2017                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2018                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2019 #endif
2020 #ifdef CONFIG_NF_CONNTRACK_MARK
2021                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2022 #endif
2023                + ctnetlink_proto_size(ct)
2024                ;
2025 }
2026
2027 static int
2028 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2029 {
2030         struct nlattr *nest_parms;
2031
2032         rcu_read_lock();
2033         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2034         if (!nest_parms)
2035                 goto nla_put_failure;
2036         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2037                 goto nla_put_failure;
2038         nla_nest_end(skb, nest_parms);
2039
2040         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2041         if (!nest_parms)
2042                 goto nla_put_failure;
2043         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2044                 goto nla_put_failure;
2045         nla_nest_end(skb, nest_parms);
2046
2047         if (nf_ct_zone(ct)) {
2048                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2049                         goto nla_put_failure;
2050         }
2051
2052         if (ctnetlink_dump_id(skb, ct) < 0)
2053                 goto nla_put_failure;
2054
2055         if (ctnetlink_dump_status(skb, ct) < 0)
2056                 goto nla_put_failure;
2057
2058         if (ctnetlink_dump_timeout(skb, ct) < 0)
2059                 goto nla_put_failure;
2060
2061         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2062                 goto nla_put_failure;
2063
2064         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2065                 goto nla_put_failure;
2066
2067 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2068         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2069                 goto nla_put_failure;
2070 #endif
2071         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2072                 goto nla_put_failure;
2073
2074         if ((ct->status & IPS_SEQ_ADJUST) &&
2075             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
2076                 goto nla_put_failure;
2077
2078 #ifdef CONFIG_NF_CONNTRACK_MARK
2079         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2080                 goto nla_put_failure;
2081 #endif
2082         if (ctnetlink_dump_labels(skb, ct) < 0)
2083                 goto nla_put_failure;
2084         rcu_read_unlock();
2085         return 0;
2086
2087 nla_put_failure:
2088         rcu_read_unlock();
2089         return -ENOSPC;
2090 }
2091
2092 static int
2093 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2094 {
2095         int err;
2096
2097         if (cda[CTA_TIMEOUT]) {
2098                 err = ctnetlink_change_timeout(ct, cda);
2099                 if (err < 0)
2100                         return err;
2101         }
2102         if (cda[CTA_STATUS]) {
2103                 err = ctnetlink_change_status(ct, cda);
2104                 if (err < 0)
2105                         return err;
2106         }
2107         if (cda[CTA_HELP]) {
2108                 err = ctnetlink_change_helper(ct, cda);
2109                 if (err < 0)
2110                         return err;
2111         }
2112         if (cda[CTA_LABELS]) {
2113                 err = ctnetlink_attach_labels(ct, cda);
2114                 if (err < 0)
2115                         return err;
2116         }
2117 #if defined(CONFIG_NF_CONNTRACK_MARK)
2118         if (cda[CTA_MARK])
2119                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2120 #endif
2121         return 0;
2122 }
2123
2124 static int
2125 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2126 {
2127         struct nlattr *cda[CTA_MAX+1];
2128         int ret;
2129
2130         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2131         if (ret < 0)
2132                 return ret;
2133
2134         spin_lock_bh(&nf_conntrack_lock);
2135         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2136         spin_unlock_bh(&nf_conntrack_lock);
2137
2138         return ret;
2139 }
2140
2141 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2142         .build_size     = ctnetlink_nfqueue_build_size,
2143         .build          = ctnetlink_nfqueue_build,
2144         .parse          = ctnetlink_nfqueue_parse,
2145 };
2146 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2147
2148 /***********************************************************************
2149  * EXPECT
2150  ***********************************************************************/
2151
2152 static inline int
2153 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2154                          const struct nf_conntrack_tuple *tuple,
2155                          enum ctattr_expect type)
2156 {
2157         struct nlattr *nest_parms;
2158
2159         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2160         if (!nest_parms)
2161                 goto nla_put_failure;
2162         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2163                 goto nla_put_failure;
2164         nla_nest_end(skb, nest_parms);
2165
2166         return 0;
2167
2168 nla_put_failure:
2169         return -1;
2170 }
2171
2172 static inline int
2173 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2174                         const struct nf_conntrack_tuple *tuple,
2175                         const struct nf_conntrack_tuple_mask *mask)
2176 {
2177         int ret;
2178         struct nf_conntrack_l3proto *l3proto;
2179         struct nf_conntrack_l4proto *l4proto;
2180         struct nf_conntrack_tuple m;
2181         struct nlattr *nest_parms;
2182
2183         memset(&m, 0xFF, sizeof(m));
2184         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2185         m.src.u.all = mask->src.u.all;
2186         m.dst.protonum = tuple->dst.protonum;
2187
2188         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2189         if (!nest_parms)
2190                 goto nla_put_failure;
2191
2192         rcu_read_lock();
2193         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2194         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2195         if (ret >= 0) {
2196                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2197                                                tuple->dst.protonum);
2198         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2199         }
2200         rcu_read_unlock();
2201
2202         if (unlikely(ret < 0))
2203                 goto nla_put_failure;
2204
2205         nla_nest_end(skb, nest_parms);
2206
2207         return 0;
2208
2209 nla_put_failure:
2210         return -1;
2211 }
2212
2213 static const union nf_inet_addr any_addr;
2214
2215 static int
2216 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2217                           const struct nf_conntrack_expect *exp)
2218 {
2219         struct nf_conn *master = exp->master;
2220         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2221         struct nf_conn_help *help;
2222 #ifdef CONFIG_NF_NAT_NEEDED
2223         struct nlattr *nest_parms;
2224         struct nf_conntrack_tuple nat_tuple = {};
2225 #endif
2226         struct nf_ct_helper_expectfn *expfn;
2227
2228         if (timeout < 0)
2229                 timeout = 0;
2230
2231         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2232                 goto nla_put_failure;
2233         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2234                 goto nla_put_failure;
2235         if (ctnetlink_exp_dump_tuple(skb,
2236                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2237                                  CTA_EXPECT_MASTER) < 0)
2238                 goto nla_put_failure;
2239
2240 #ifdef CONFIG_NF_NAT_NEEDED
2241         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2242             exp->saved_proto.all) {
2243                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2244                 if (!nest_parms)
2245                         goto nla_put_failure;
2246
2247                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2248                         goto nla_put_failure;
2249
2250                 nat_tuple.src.l3num = nf_ct_l3num(master);
2251                 nat_tuple.src.u3 = exp->saved_addr;
2252                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2253                 nat_tuple.src.u = exp->saved_proto;
2254
2255                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2256                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2257                         goto nla_put_failure;
2258                 nla_nest_end(skb, nest_parms);
2259         }
2260 #endif
2261         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2262             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2263             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2264             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2265                 goto nla_put_failure;
2266         help = nfct_help(master);
2267         if (help) {
2268                 struct nf_conntrack_helper *helper;
2269
2270                 helper = rcu_dereference(help->helper);
2271                 if (helper &&
2272                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2273                         goto nla_put_failure;
2274         }
2275         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2276         if (expfn != NULL &&
2277             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2278                 goto nla_put_failure;
2279
2280         return 0;
2281
2282 nla_put_failure:
2283         return -1;
2284 }
2285
2286 static int
2287 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2288                         int event, const struct nf_conntrack_expect *exp)
2289 {
2290         struct nlmsghdr *nlh;
2291         struct nfgenmsg *nfmsg;
2292         unsigned int flags = portid ? NLM_F_MULTI : 0;
2293
2294         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2295         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2296         if (nlh == NULL)
2297                 goto nlmsg_failure;
2298
2299         nfmsg = nlmsg_data(nlh);
2300         nfmsg->nfgen_family = exp->tuple.src.l3num;
2301         nfmsg->version      = NFNETLINK_V0;
2302         nfmsg->res_id       = 0;
2303
2304         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2305                 goto nla_put_failure;
2306
2307         nlmsg_end(skb, nlh);
2308         return skb->len;
2309
2310 nlmsg_failure:
2311 nla_put_failure:
2312         nlmsg_cancel(skb, nlh);
2313         return -1;
2314 }
2315
2316 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2317 static int
2318 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2319 {
2320         struct nf_conntrack_expect *exp = item->exp;
2321         struct net *net = nf_ct_exp_net(exp);
2322         struct nlmsghdr *nlh;
2323         struct nfgenmsg *nfmsg;
2324         struct sk_buff *skb;
2325         unsigned int type, group;
2326         int flags = 0;
2327
2328         if (events & (1 << IPEXP_DESTROY)) {
2329                 type = IPCTNL_MSG_EXP_DELETE;
2330                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2331         } else if (events & (1 << IPEXP_NEW)) {
2332                 type = IPCTNL_MSG_EXP_NEW;
2333                 flags = NLM_F_CREATE|NLM_F_EXCL;
2334                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2335         } else
2336                 return 0;
2337
2338         if (!item->report && !nfnetlink_has_listeners(net, group))
2339                 return 0;
2340
2341         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2342         if (skb == NULL)
2343                 goto errout;
2344
2345         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2346         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2347         if (nlh == NULL)
2348                 goto nlmsg_failure;
2349
2350         nfmsg = nlmsg_data(nlh);
2351         nfmsg->nfgen_family = exp->tuple.src.l3num;
2352         nfmsg->version      = NFNETLINK_V0;
2353         nfmsg->res_id       = 0;
2354
2355         rcu_read_lock();
2356         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2357                 goto nla_put_failure;
2358         rcu_read_unlock();
2359
2360         nlmsg_end(skb, nlh);
2361         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2362         return 0;
2363
2364 nla_put_failure:
2365         rcu_read_unlock();
2366         nlmsg_cancel(skb, nlh);
2367 nlmsg_failure:
2368         kfree_skb(skb);
2369 errout:
2370         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2371         return 0;
2372 }
2373 #endif
2374 static int ctnetlink_exp_done(struct netlink_callback *cb)
2375 {
2376         if (cb->args[1])
2377                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2378         return 0;
2379 }
2380
2381 static int
2382 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2383 {
2384         struct net *net = sock_net(skb->sk);
2385         struct nf_conntrack_expect *exp, *last;
2386         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2387         u_int8_t l3proto = nfmsg->nfgen_family;
2388
2389         rcu_read_lock();
2390         last = (struct nf_conntrack_expect *)cb->args[1];
2391         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2392 restart:
2393                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2394                                      hnode) {
2395                         if (l3proto && exp->tuple.src.l3num != l3proto)
2396                                 continue;
2397                         if (cb->args[1]) {
2398                                 if (exp != last)
2399                                         continue;
2400                                 cb->args[1] = 0;
2401                         }
2402                         if (ctnetlink_exp_fill_info(skb,
2403                                                     NETLINK_CB(cb->skb).portid,
2404                                                     cb->nlh->nlmsg_seq,
2405                                                     IPCTNL_MSG_EXP_NEW,
2406                                                     exp) < 0) {
2407                                 if (!atomic_inc_not_zero(&exp->use))
2408                                         continue;
2409                                 cb->args[1] = (unsigned long)exp;
2410                                 goto out;
2411                         }
2412                 }
2413                 if (cb->args[1]) {
2414                         cb->args[1] = 0;
2415                         goto restart;
2416                 }
2417         }
2418 out:
2419         rcu_read_unlock();
2420         if (last)
2421                 nf_ct_expect_put(last);
2422
2423         return skb->len;
2424 }
2425
2426 static int
2427 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2428 {
2429         struct nf_conntrack_expect *exp, *last;
2430         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2431         struct nf_conn *ct = cb->data;
2432         struct nf_conn_help *help = nfct_help(ct);
2433         u_int8_t l3proto = nfmsg->nfgen_family;
2434
2435         if (cb->args[0])
2436                 return 0;
2437
2438         rcu_read_lock();
2439         last = (struct nf_conntrack_expect *)cb->args[1];
2440 restart:
2441         hlist_for_each_entry(exp, &help->expectations, lnode) {
2442                 if (l3proto && exp->tuple.src.l3num != l3proto)
2443                         continue;
2444                 if (cb->args[1]) {
2445                         if (exp != last)
2446                                 continue;
2447                         cb->args[1] = 0;
2448                 }
2449                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2450                                             cb->nlh->nlmsg_seq,
2451                                             IPCTNL_MSG_EXP_NEW,
2452                                             exp) < 0) {
2453                         if (!atomic_inc_not_zero(&exp->use))
2454                                 continue;
2455                         cb->args[1] = (unsigned long)exp;
2456                         goto out;
2457                 }
2458         }
2459         if (cb->args[1]) {
2460                 cb->args[1] = 0;
2461                 goto restart;
2462         }
2463         cb->args[0] = 1;
2464 out:
2465         rcu_read_unlock();
2466         if (last)
2467                 nf_ct_expect_put(last);
2468
2469         return skb->len;
2470 }
2471
2472 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2473                                  const struct nlmsghdr *nlh,
2474                                  const struct nlattr * const cda[])
2475 {
2476         int err;
2477         struct net *net = sock_net(ctnl);
2478         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2479         u_int8_t u3 = nfmsg->nfgen_family;
2480         struct nf_conntrack_tuple tuple;
2481         struct nf_conntrack_tuple_hash *h;
2482         struct nf_conn *ct;
2483         u16 zone = 0;
2484         struct netlink_dump_control c = {
2485                 .dump = ctnetlink_exp_ct_dump_table,
2486                 .done = ctnetlink_exp_done,
2487         };
2488
2489         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2490         if (err < 0)
2491                 return err;
2492
2493         if (cda[CTA_EXPECT_ZONE]) {
2494                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2495                 if (err < 0)
2496                         return err;
2497         }
2498
2499         h = nf_conntrack_find_get(net, zone, &tuple);
2500         if (!h)
2501                 return -ENOENT;
2502
2503         ct = nf_ct_tuplehash_to_ctrack(h);
2504         c.data = ct;
2505
2506         err = netlink_dump_start(ctnl, skb, nlh, &c);
2507         nf_ct_put(ct);
2508
2509         return err;
2510 }
2511
2512 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2513         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2514         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2515         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2516         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2517         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2518         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2519                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2520         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2521         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2522         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2523         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2524         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2525 };
2526
2527 static int
2528 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2529                      const struct nlmsghdr *nlh,
2530                      const struct nlattr * const cda[])
2531 {
2532         struct net *net = sock_net(ctnl);
2533         struct nf_conntrack_tuple tuple;
2534         struct nf_conntrack_expect *exp;
2535         struct sk_buff *skb2;
2536         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2537         u_int8_t u3 = nfmsg->nfgen_family;
2538         u16 zone;
2539         int err;
2540
2541         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2542                 if (cda[CTA_EXPECT_MASTER])
2543                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2544                 else {
2545                         struct netlink_dump_control c = {
2546                                 .dump = ctnetlink_exp_dump_table,
2547                                 .done = ctnetlink_exp_done,
2548                         };
2549                         return netlink_dump_start(ctnl, skb, nlh, &c);
2550                 }
2551         }
2552
2553         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2554         if (err < 0)
2555                 return err;
2556
2557         if (cda[CTA_EXPECT_TUPLE])
2558                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2559         else if (cda[CTA_EXPECT_MASTER])
2560                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2561         else
2562                 return -EINVAL;
2563
2564         if (err < 0)
2565                 return err;
2566
2567         exp = nf_ct_expect_find_get(net, zone, &tuple);
2568         if (!exp)
2569                 return -ENOENT;
2570
2571         if (cda[CTA_EXPECT_ID]) {
2572                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2573                 if (ntohl(id) != (u32)(unsigned long)exp) {
2574                         nf_ct_expect_put(exp);
2575                         return -ENOENT;
2576                 }
2577         }
2578
2579         err = -ENOMEM;
2580         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2581         if (skb2 == NULL) {
2582                 nf_ct_expect_put(exp);
2583                 goto out;
2584         }
2585
2586         rcu_read_lock();
2587         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2588                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2589         rcu_read_unlock();
2590         nf_ct_expect_put(exp);
2591         if (err <= 0)
2592                 goto free;
2593
2594         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2595         if (err < 0)
2596                 goto out;
2597
2598         return 0;
2599
2600 free:
2601         kfree_skb(skb2);
2602 out:
2603         /* this avoids a loop in nfnetlink. */
2604         return err == -EAGAIN ? -ENOBUFS : err;
2605 }
2606
2607 static int
2608 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2609                      const struct nlmsghdr *nlh,
2610                      const struct nlattr * const cda[])
2611 {
2612         struct net *net = sock_net(ctnl);
2613         struct nf_conntrack_expect *exp;
2614         struct nf_conntrack_tuple tuple;
2615         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2616         struct hlist_node *next;
2617         u_int8_t u3 = nfmsg->nfgen_family;
2618         unsigned int i;
2619         u16 zone;
2620         int err;
2621
2622         if (cda[CTA_EXPECT_TUPLE]) {
2623                 /* delete a single expect by tuple */
2624                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2625                 if (err < 0)
2626                         return err;
2627
2628                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2629                 if (err < 0)
2630                         return err;
2631
2632                 /* bump usage count to 2 */
2633                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2634                 if (!exp)
2635                         return -ENOENT;
2636
2637                 if (cda[CTA_EXPECT_ID]) {
2638                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2639                         if (ntohl(id) != (u32)(unsigned long)exp) {
2640                                 nf_ct_expect_put(exp);
2641                                 return -ENOENT;
2642                         }
2643                 }
2644
2645                 /* after list removal, usage count == 1 */
2646                 spin_lock_bh(&nf_conntrack_lock);
2647                 if (del_timer(&exp->timeout)) {
2648                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2649                                                    nlmsg_report(nlh));
2650                         nf_ct_expect_put(exp);
2651                 }
2652                 spin_unlock_bh(&nf_conntrack_lock);
2653                 /* have to put what we 'get' above.
2654                  * after this line usage count == 0 */
2655                 nf_ct_expect_put(exp);
2656         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2657                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2658                 struct nf_conn_help *m_help;
2659
2660                 /* delete all expectations for this helper */
2661                 spin_lock_bh(&nf_conntrack_lock);
2662                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2663                         hlist_for_each_entry_safe(exp, next,
2664                                                   &net->ct.expect_hash[i],
2665                                                   hnode) {
2666                                 m_help = nfct_help(exp->master);
2667                                 if (!strcmp(m_help->helper->name, name) &&
2668                                     del_timer(&exp->timeout)) {
2669                                         nf_ct_unlink_expect_report(exp,
2670                                                         NETLINK_CB(skb).portid,
2671                                                         nlmsg_report(nlh));
2672                                         nf_ct_expect_put(exp);
2673                                 }
2674                         }
2675                 }
2676                 spin_unlock_bh(&nf_conntrack_lock);
2677         } else {
2678                 /* This basically means we have to flush everything*/
2679                 spin_lock_bh(&nf_conntrack_lock);
2680                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2681                         hlist_for_each_entry_safe(exp, next,
2682                                                   &net->ct.expect_hash[i],
2683                                                   hnode) {
2684                                 if (del_timer(&exp->timeout)) {
2685                                         nf_ct_unlink_expect_report(exp,
2686                                                         NETLINK_CB(skb).portid,
2687                                                         nlmsg_report(nlh));
2688                                         nf_ct_expect_put(exp);
2689                                 }
2690                         }
2691                 }
2692                 spin_unlock_bh(&nf_conntrack_lock);
2693         }
2694
2695         return 0;
2696 }
2697 static int
2698 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2699                         const struct nlattr * const cda[])
2700 {
2701         if (cda[CTA_EXPECT_TIMEOUT]) {
2702                 if (!del_timer(&x->timeout))
2703                         return -ETIME;
2704
2705                 x->timeout.expires = jiffies +
2706                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2707                 add_timer(&x->timeout);
2708         }
2709         return 0;
2710 }
2711
2712 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2713         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2714         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2715 };
2716
2717 static int
2718 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2719                            struct nf_conntrack_expect *exp,
2720                            u_int8_t u3)
2721 {
2722 #ifdef CONFIG_NF_NAT_NEEDED
2723         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2724         struct nf_conntrack_tuple nat_tuple = {};
2725         int err;
2726
2727         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2728         if (err < 0)
2729                 return err;
2730
2731         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2732                 return -EINVAL;
2733
2734         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2735                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2736         if (err < 0)
2737                 return err;
2738
2739         exp->saved_addr = nat_tuple.src.u3;
2740         exp->saved_proto = nat_tuple.src.u;
2741         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2742
2743         return 0;
2744 #else
2745         return -EOPNOTSUPP;
2746 #endif
2747 }
2748
2749 static int
2750 ctnetlink_create_expect(struct net *net, u16 zone,
2751                         const struct nlattr * const cda[],
2752                         u_int8_t u3,
2753                         u32 portid, int report)
2754 {
2755         struct nf_conntrack_tuple tuple, mask, master_tuple;
2756         struct nf_conntrack_tuple_hash *h = NULL;
2757         struct nf_conntrack_expect *exp;
2758         struct nf_conn *ct;
2759         struct nf_conn_help *help;
2760         struct nf_conntrack_helper *helper = NULL;
2761         u_int32_t class = 0;
2762         int err = 0;
2763
2764         /* caller guarantees that those three CTA_EXPECT_* exist */
2765         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2766         if (err < 0)
2767                 return err;
2768         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2769         if (err < 0)
2770                 return err;
2771         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2772         if (err < 0)
2773                 return err;
2774
2775         /* Look for master conntrack of this expectation */
2776         h = nf_conntrack_find_get(net, zone, &master_tuple);
2777         if (!h)
2778                 return -ENOENT;
2779         ct = nf_ct_tuplehash_to_ctrack(h);
2780
2781         /* Look for helper of this expectation */
2782         if (cda[CTA_EXPECT_HELP_NAME]) {
2783                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2784
2785                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2786                                                     nf_ct_protonum(ct));
2787                 if (helper == NULL) {
2788 #ifdef CONFIG_MODULES
2789                         if (request_module("nfct-helper-%s", helpname) < 0) {
2790                                 err = -EOPNOTSUPP;
2791                                 goto out;
2792                         }
2793
2794                         helper = __nf_conntrack_helper_find(helpname,
2795                                                             nf_ct_l3num(ct),
2796                                                             nf_ct_protonum(ct));
2797                         if (helper) {
2798                                 err = -EAGAIN;
2799                                 goto out;
2800                         }
2801 #endif
2802                         err = -EOPNOTSUPP;
2803                         goto out;
2804                 }
2805         }
2806
2807         if (cda[CTA_EXPECT_CLASS] && helper) {
2808                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2809                 if (class > helper->expect_class_max) {
2810                         err = -EINVAL;
2811                         goto out;
2812                 }
2813         }
2814         exp = nf_ct_expect_alloc(ct);
2815         if (!exp) {
2816                 err = -ENOMEM;
2817                 goto out;
2818         }
2819         help = nfct_help(ct);
2820         if (!help) {
2821                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2822                         err = -EINVAL;
2823                         goto err_out;
2824                 }
2825                 exp->timeout.expires =
2826                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2827
2828                 exp->flags = NF_CT_EXPECT_USERSPACE;
2829                 if (cda[CTA_EXPECT_FLAGS]) {
2830                         exp->flags |=
2831                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2832                 }
2833         } else {
2834                 if (cda[CTA_EXPECT_FLAGS]) {
2835                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2836                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2837                 } else
2838                         exp->flags = 0;
2839         }
2840         if (cda[CTA_EXPECT_FN]) {
2841                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2842                 struct nf_ct_helper_expectfn *expfn;
2843
2844                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2845                 if (expfn == NULL) {
2846                         err = -EINVAL;
2847                         goto err_out;
2848                 }
2849                 exp->expectfn = expfn->expectfn;
2850         } else
2851                 exp->expectfn = NULL;
2852
2853         exp->class = class;
2854         exp->master = ct;
2855         exp->helper = helper;
2856         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2857         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2858         exp->mask.src.u.all = mask.src.u.all;
2859
2860         if (cda[CTA_EXPECT_NAT]) {
2861                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2862                                                  exp, u3);
2863                 if (err < 0)
2864                         goto err_out;
2865         }
2866         err = nf_ct_expect_related_report(exp, portid, report);
2867 err_out:
2868         nf_ct_expect_put(exp);
2869 out:
2870         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2871         return err;
2872 }
2873
2874 static int
2875 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2876                      const struct nlmsghdr *nlh,
2877                      const struct nlattr * const cda[])
2878 {
2879         struct net *net = sock_net(ctnl);
2880         struct nf_conntrack_tuple tuple;
2881         struct nf_conntrack_expect *exp;
2882         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2883         u_int8_t u3 = nfmsg->nfgen_family;
2884         u16 zone;
2885         int err;
2886
2887         if (!cda[CTA_EXPECT_TUPLE]
2888             || !cda[CTA_EXPECT_MASK]
2889             || !cda[CTA_EXPECT_MASTER])
2890                 return -EINVAL;
2891
2892         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2893         if (err < 0)
2894                 return err;
2895
2896         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2897         if (err < 0)
2898                 return err;
2899
2900         spin_lock_bh(&nf_conntrack_lock);
2901         exp = __nf_ct_expect_find(net, zone, &tuple);
2902
2903         if (!exp) {
2904                 spin_unlock_bh(&nf_conntrack_lock);
2905                 err = -ENOENT;
2906                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2907                         err = ctnetlink_create_expect(net, zone, cda,
2908                                                       u3,
2909                                                       NETLINK_CB(skb).portid,
2910                                                       nlmsg_report(nlh));
2911                 }
2912                 return err;
2913         }
2914
2915         err = -EEXIST;
2916         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2917                 err = ctnetlink_change_expect(exp, cda);
2918         spin_unlock_bh(&nf_conntrack_lock);
2919
2920         return err;
2921 }
2922
2923 static int
2924 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
2925                              const struct ip_conntrack_stat *st)
2926 {
2927         struct nlmsghdr *nlh;
2928         struct nfgenmsg *nfmsg;
2929         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2930
2931         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
2932         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2933         if (nlh == NULL)
2934                 goto nlmsg_failure;
2935
2936         nfmsg = nlmsg_data(nlh);
2937         nfmsg->nfgen_family = AF_UNSPEC;
2938         nfmsg->version      = NFNETLINK_V0;
2939         nfmsg->res_id       = htons(cpu);
2940
2941         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
2942             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
2943             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
2944                 goto nla_put_failure;
2945
2946         nlmsg_end(skb, nlh);
2947         return skb->len;
2948
2949 nla_put_failure:
2950 nlmsg_failure:
2951         nlmsg_cancel(skb, nlh);
2952         return -1;
2953 }
2954
2955 static int
2956 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2957 {
2958         int cpu;
2959         struct net *net = sock_net(skb->sk);
2960
2961         if (cb->args[0] == nr_cpu_ids)
2962                 return 0;
2963
2964         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2965                 const struct ip_conntrack_stat *st;
2966
2967                 if (!cpu_possible(cpu))
2968                         continue;
2969
2970                 st = per_cpu_ptr(net->ct.stat, cpu);
2971                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
2972                                                  cb->nlh->nlmsg_seq,
2973                                                  cpu, st) < 0)
2974                         break;
2975         }
2976         cb->args[0] = cpu;
2977
2978         return skb->len;
2979 }
2980
2981 static int
2982 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
2983                        const struct nlmsghdr *nlh,
2984                        const struct nlattr * const cda[])
2985 {
2986         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2987                 struct netlink_dump_control c = {
2988                         .dump = ctnetlink_exp_stat_cpu_dump,
2989                 };
2990                 return netlink_dump_start(ctnl, skb, nlh, &c);
2991         }
2992
2993         return 0;
2994 }
2995
2996 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2997 static struct nf_ct_event_notifier ctnl_notifier = {
2998         .fcn = ctnetlink_conntrack_event,
2999 };
3000
3001 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3002         .fcn = ctnetlink_expect_event,
3003 };
3004 #endif
3005
3006 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3007         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3008                                             .attr_count = CTA_MAX,
3009                                             .policy = ct_nla_policy },
3010         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3011                                             .attr_count = CTA_MAX,
3012                                             .policy = ct_nla_policy },
3013         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3014                                             .attr_count = CTA_MAX,
3015                                             .policy = ct_nla_policy },
3016         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3017                                             .attr_count = CTA_MAX,
3018                                             .policy = ct_nla_policy },
3019         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3020         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3021         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3022         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3023 };
3024
3025 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3026         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3027                                             .attr_count = CTA_EXPECT_MAX,
3028                                             .policy = exp_nla_policy },
3029         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3030                                             .attr_count = CTA_EXPECT_MAX,
3031                                             .policy = exp_nla_policy },
3032         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3033                                             .attr_count = CTA_EXPECT_MAX,
3034                                             .policy = exp_nla_policy },
3035         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3036 };
3037
3038 static const struct nfnetlink_subsystem ctnl_subsys = {
3039         .name                           = "conntrack",
3040         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3041         .cb_count                       = IPCTNL_MSG_MAX,
3042         .cb                             = ctnl_cb,
3043 };
3044
3045 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3046         .name                           = "conntrack_expect",
3047         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3048         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3049         .cb                             = ctnl_exp_cb,
3050 };
3051
3052 MODULE_ALIAS("ip_conntrack_netlink");
3053 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3054 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3055
3056 static int __net_init ctnetlink_net_init(struct net *net)
3057 {
3058 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3059         int ret;
3060
3061         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3062         if (ret < 0) {
3063                 pr_err("ctnetlink_init: cannot register notifier.\n");
3064                 goto err_out;
3065         }
3066
3067         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3068         if (ret < 0) {
3069                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3070                 goto err_unreg_notifier;
3071         }
3072 #endif
3073         return 0;
3074
3075 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3076 err_unreg_notifier:
3077         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3078 err_out:
3079         return ret;
3080 #endif
3081 }
3082
3083 static void ctnetlink_net_exit(struct net *net)
3084 {
3085 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3086         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3087         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3088 #endif
3089 }
3090
3091 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3092 {
3093         struct net *net;
3094
3095         list_for_each_entry(net, net_exit_list, exit_list)
3096                 ctnetlink_net_exit(net);
3097 }
3098
3099 static struct pernet_operations ctnetlink_net_ops = {
3100         .init           = ctnetlink_net_init,
3101         .exit_batch     = ctnetlink_net_exit_batch,
3102 };
3103
3104 static int __init ctnetlink_init(void)
3105 {
3106         int ret;
3107
3108         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3109         ret = nfnetlink_subsys_register(&ctnl_subsys);
3110         if (ret < 0) {
3111                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3112                 goto err_out;
3113         }
3114
3115         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3116         if (ret < 0) {
3117                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3118                 goto err_unreg_subsys;
3119         }
3120
3121         ret = register_pernet_subsys(&ctnetlink_net_ops);
3122         if (ret < 0) {
3123                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3124                 goto err_unreg_exp_subsys;
3125         }
3126 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3127         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3128         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3129 #endif
3130         return 0;
3131
3132 err_unreg_exp_subsys:
3133         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3134 err_unreg_subsys:
3135         nfnetlink_subsys_unregister(&ctnl_subsys);
3136 err_out:
3137         return ret;
3138 }
3139
3140 static void __exit ctnetlink_exit(void)
3141 {
3142         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3143
3144         unregister_pernet_subsys(&ctnetlink_net_ops);
3145         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3146         nfnetlink_subsys_unregister(&ctnl_subsys);
3147 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3148         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3149 #endif
3150 }
3151
3152 module_init(ctnetlink_init);
3153 module_exit(ctnetlink_exit);