Pileus Git - ~andy/linux/blob - 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                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1043
1044         nf_ct_put(ct);
1045
1046         return 0;
1047 }
1048
1049 static int
1050 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1051                         const struct nlmsghdr *nlh,
1052                         const struct nlattr * const cda[])
1053 {
1054         struct net *net = sock_net(ctnl);
1055         struct nf_conntrack_tuple_hash *h;
1056         struct nf_conntrack_tuple tuple;
1057         struct nf_conn *ct;
1058         struct sk_buff *skb2 = NULL;
1059         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1060         u_int8_t u3 = nfmsg->nfgen_family;
1061         u16 zone;
1062         int err;
1063
1064         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1065                 struct netlink_dump_control c = {
1066                         .dump = ctnetlink_dump_table,
1067                         .done = ctnetlink_done,
1068                 };
1069 #ifdef CONFIG_NF_CONNTRACK_MARK
1070                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1071                         struct ctnetlink_dump_filter *filter;
1072
1073                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1074                                          GFP_ATOMIC);
1075                         if (filter == NULL)
1076                                 return -ENOMEM;
1077
1078                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1079                         filter->mark.mask =
1080                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1081                         c.data = filter;
1082                 }
1083 #endif
1084                 return netlink_dump_start(ctnl, skb, nlh, &c);
1085         }
1086
1087         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1088         if (err < 0)
1089                 return err;
1090
1091         if (cda[CTA_TUPLE_ORIG])
1092                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1093         else if (cda[CTA_TUPLE_REPLY])
1094                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1095         else
1096                 return -EINVAL;
1097
1098         if (err < 0)
1099                 return err;
1100
1101         h = nf_conntrack_find_get(net, zone, &tuple);
1102         if (!h)
1103                 return -ENOENT;
1104
1105         ct = nf_ct_tuplehash_to_ctrack(h);
1106
1107         err = -ENOMEM;
1108         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1109         if (skb2 == NULL) {
1110                 nf_ct_put(ct);
1111                 return -ENOMEM;
1112         }
1113
1114         rcu_read_lock();
1115         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1116                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1117         rcu_read_unlock();
1118         nf_ct_put(ct);
1119         if (err <= 0)
1120                 goto free;
1121
1122         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1123         if (err < 0)
1124                 goto out;
1125
1126         return 0;
1127
1128 free:
1129         kfree_skb(skb2);
1130 out:
1131         /* this avoids a loop in nfnetlink. */
1132         return err == -EAGAIN ? -ENOBUFS : err;
1133 }
1134
1135 static int ctnetlink_done_list(struct netlink_callback *cb)
1136 {
1137         if (cb->args[1])
1138                 nf_ct_put((struct nf_conn *)cb->args[1]);
1139         return 0;
1140 }
1141
1142 static int
1143 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb,
1144                     struct hlist_nulls_head *list)
1145 {
1146         struct nf_conn *ct, *last;
1147         struct nf_conntrack_tuple_hash *h;
1148         struct hlist_nulls_node *n;
1149         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1150         u_int8_t l3proto = nfmsg->nfgen_family;
1151         int res;
1152
1153         if (cb->args[2])
1154                 return 0;
1155
1156         spin_lock_bh(&nf_conntrack_lock);
1157         last = (struct nf_conn *)cb->args[1];
1158 restart:
1159         hlist_nulls_for_each_entry(h, n, list, hnnode) {
1160                 ct = nf_ct_tuplehash_to_ctrack(h);
1161                 if (l3proto && nf_ct_l3num(ct) != l3proto)
1162                         continue;
1163                 if (cb->args[1]) {
1164                         if (ct != last)
1165                                 continue;
1166                         cb->args[1] = 0;
1167                 }
1168                 rcu_read_lock();
1169                 res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1170                                           cb->nlh->nlmsg_seq,
1171                                           NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1172                                           ct);
1173                 rcu_read_unlock();
1174                 if (res < 0) {
1175                         nf_conntrack_get(&ct->ct_general);
1176                         cb->args[1] = (unsigned long)ct;
1177                         goto out;
1178                 }
1179         }
1180         if (cb->args[1]) {
1181                 cb->args[1] = 0;
1182                 goto restart;
1183         } else
1184                 cb->args[2] = 1;
1185 out:
1186         spin_unlock_bh(&nf_conntrack_lock);
1187         if (last)
1188                 nf_ct_put(last);
1189
1190         return skb->len;
1191 }
1192
1193 static int
1194 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1195 {
1196         struct net *net = sock_net(skb->sk);
1197
1198         return ctnetlink_dump_list(skb, cb, &net->ct.dying);
1199 }
1200
1201 static int
1202 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1203                        const struct nlmsghdr *nlh,
1204                        const struct nlattr * const cda[])
1205 {
1206         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1207                 struct netlink_dump_control c = {
1208                         .dump = ctnetlink_dump_dying,
1209                         .done = ctnetlink_done_list,
1210                 };
1211                 return netlink_dump_start(ctnl, skb, nlh, &c);
1212         }
1213
1214         return -EOPNOTSUPP;
1215 }
1216
1217 static int
1218 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1219 {
1220         struct net *net = sock_net(skb->sk);
1221
1222         return ctnetlink_dump_list(skb, cb, &net->ct.unconfirmed);
1223 }
1224
1225 static int
1226 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1227                              const struct nlmsghdr *nlh,
1228                              const struct nlattr * const cda[])
1229 {
1230         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1231                 struct netlink_dump_control c = {
1232                         .dump = ctnetlink_dump_unconfirmed,
1233                         .done = ctnetlink_done_list,
1234                 };
1235                 return netlink_dump_start(ctnl, skb, nlh, &c);
1236         }
1237
1238         return -EOPNOTSUPP;
1239 }
1240
1241 #ifdef CONFIG_NF_NAT_NEEDED
1242 static int
1243 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1244                           enum nf_nat_manip_type manip,
1245                           const struct nlattr *attr)
1246 {
1247         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1248         int err;
1249
1250         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1251         if (!parse_nat_setup) {
1252 #ifdef CONFIG_MODULES
1253                 rcu_read_unlock();
1254                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1255                 if (request_module("nf-nat") < 0) {
1256                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1257                         rcu_read_lock();
1258                         return -EOPNOTSUPP;
1259                 }
1260                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1261                 rcu_read_lock();
1262                 if (nfnetlink_parse_nat_setup_hook)
1263                         return -EAGAIN;
1264 #endif
1265                 return -EOPNOTSUPP;
1266         }
1267
1268         err = parse_nat_setup(ct, manip, attr);
1269         if (err == -EAGAIN) {
1270 #ifdef CONFIG_MODULES
1271                 rcu_read_unlock();
1272                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1273                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1274                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1275                         rcu_read_lock();
1276                         return -EOPNOTSUPP;
1277                 }
1278                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1279                 rcu_read_lock();
1280 #else
1281                 err = -EOPNOTSUPP;
1282 #endif
1283         }
1284         return err;
1285 }
1286 #endif
1287
1288 static int
1289 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1290 {
1291         unsigned long d;
1292         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1293         d = ct->status ^ status;
1294
1295         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1296                 /* unchangeable */
1297                 return -EBUSY;
1298
1299         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1300                 /* SEEN_REPLY bit can only be set */
1301                 return -EBUSY;
1302
1303         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1304                 /* ASSURED bit can only be set */
1305                 return -EBUSY;
1306
1307         /* Be careful here, modifying NAT bits can screw up things,
1308          * so don't let users modify them directly if they don't pass
1309          * nf_nat_range. */
1310         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1311         return 0;
1312 }
1313
1314 static int
1315 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1316 {
1317 #ifdef CONFIG_NF_NAT_NEEDED
1318         int ret;
1319
1320         if (cda[CTA_NAT_DST]) {
1321                 ret = ctnetlink_parse_nat_setup(ct,
1322                                                 NF_NAT_MANIP_DST,
1323                                                 cda[CTA_NAT_DST]);
1324                 if (ret < 0)
1325                         return ret;
1326         }
1327         if (cda[CTA_NAT_SRC]) {
1328                 ret = ctnetlink_parse_nat_setup(ct,
1329                                                 NF_NAT_MANIP_SRC,
1330                                                 cda[CTA_NAT_SRC]);
1331                 if (ret < 0)
1332                         return ret;
1333         }
1334         return 0;
1335 #else
1336         return -EOPNOTSUPP;
1337 #endif
1338 }
1339
1340 static inline int
1341 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1342 {
1343         struct nf_conntrack_helper *helper;
1344         struct nf_conn_help *help = nfct_help(ct);
1345         char *helpname = NULL;
1346         struct nlattr *helpinfo = NULL;
1347         int err;
1348
1349         /* don't change helper of sibling connections */
1350         if (ct->master)
1351                 return -EBUSY;
1352
1353         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1354         if (err < 0)
1355                 return err;
1356
1357         if (!strcmp(helpname, "")) {
1358                 if (help && help->helper) {
1359                         /* we had a helper before ... */
1360                         nf_ct_remove_expectations(ct);
1361                         RCU_INIT_POINTER(help->helper, NULL);
1362                 }
1363
1364                 return 0;
1365         }
1366
1367         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1368                                             nf_ct_protonum(ct));
1369         if (helper == NULL) {
1370 #ifdef CONFIG_MODULES
1371                 spin_unlock_bh(&nf_conntrack_lock);
1372
1373                 if (request_module("nfct-helper-%s", helpname) < 0) {
1374                         spin_lock_bh(&nf_conntrack_lock);
1375                         return -EOPNOTSUPP;
1376                 }
1377
1378                 spin_lock_bh(&nf_conntrack_lock);
1379                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1380                                                     nf_ct_protonum(ct));
1381                 if (helper)
1382                         return -EAGAIN;
1383 #endif
1384                 return -EOPNOTSUPP;
1385         }
1386
1387         if (help) {
1388                 if (help->helper == helper) {
1389                         /* update private helper data if allowed. */
1390                         if (helper->from_nlattr)
1391                                 helper->from_nlattr(helpinfo, ct);
1392                         return 0;
1393                 } else
1394                         return -EBUSY;
1395         }
1396
1397         /* we cannot set a helper for an existing conntrack */
1398         return -EOPNOTSUPP;
1399 }
1400
1401 static inline int
1402 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1403 {
1404         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1405
1406         if (!del_timer(&ct->timeout))
1407                 return -ETIME;
1408
1409         ct->timeout.expires = jiffies + timeout * HZ;
1410         add_timer(&ct->timeout);
1411
1412         return 0;
1413 }
1414
1415 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1416         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1417         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1418         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1419 };
1420
1421 static inline int
1422 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1423 {
1424         const struct nlattr *attr = cda[CTA_PROTOINFO];
1425         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1426         struct nf_conntrack_l4proto *l4proto;
1427         int err = 0;
1428
1429         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1430         if (err < 0)
1431                 return err;
1432
1433         rcu_read_lock();
1434         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1435         if (l4proto->from_nlattr)
1436                 err = l4proto->from_nlattr(tb, ct);
1437         rcu_read_unlock();
1438
1439         return err;
1440 }
1441
1442 #ifdef CONFIG_NF_NAT_NEEDED
1443 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1444         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1445         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1446         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1447 };
1448
1449 static inline int
1450 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1451 {
1452         int err;
1453         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1454
1455         err = nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1456         if (err < 0)
1457                 return err;
1458
1459         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1460                 return -EINVAL;
1461
1462         natseq->correction_pos =
1463                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1464
1465         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1466                 return -EINVAL;
1467
1468         natseq->offset_before =
1469                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1470
1471         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1472                 return -EINVAL;
1473
1474         natseq->offset_after =
1475                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1476
1477         return 0;
1478 }
1479
1480 static int
1481 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1482                              const struct nlattr * const cda[])
1483 {
1484         int ret = 0;
1485         struct nf_conn_nat *nat = nfct_nat(ct);
1486
1487         if (!nat)
1488                 return 0;
1489
1490         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1491                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1492                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1493                 if (ret < 0)
1494                         return ret;
1495
1496                 ct->status |= IPS_SEQ_ADJUST;
1497         }
1498
1499         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1500                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1501                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1502                 if (ret < 0)
1503                         return ret;
1504
1505                 ct->status |= IPS_SEQ_ADJUST;
1506         }
1507
1508         return 0;
1509 }
1510 #endif
1511
1512 static int
1513 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1514 {
1515 #ifdef CONFIG_NF_CONNTRACK_LABELS
1516         size_t len = nla_len(cda[CTA_LABELS]);
1517         const void *mask = cda[CTA_LABELS_MASK];
1518
1519         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1520                 return -EINVAL;
1521
1522         if (mask) {
1523                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1524                     nla_len(cda[CTA_LABELS_MASK]) != len)
1525                         return -EINVAL;
1526                 mask = nla_data(cda[CTA_LABELS_MASK]);
1527         }
1528
1529         len /= sizeof(u32);
1530
1531         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1532 #else
1533         return -EOPNOTSUPP;
1534 #endif
1535 }
1536
1537 static int
1538 ctnetlink_change_conntrack(struct nf_conn *ct,
1539                            const struct nlattr * const cda[])
1540 {
1541         int err;
1542
1543         /* only allow NAT changes and master assignation for new conntracks */
1544         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1545                 return -EOPNOTSUPP;
1546
1547         if (cda[CTA_HELP]) {
1548                 err = ctnetlink_change_helper(ct, cda);
1549                 if (err < 0)
1550                         return err;
1551         }
1552
1553         if (cda[CTA_TIMEOUT]) {
1554                 err = ctnetlink_change_timeout(ct, cda);
1555                 if (err < 0)
1556                         return err;
1557         }
1558
1559         if (cda[CTA_STATUS]) {
1560                 err = ctnetlink_change_status(ct, cda);
1561                 if (err < 0)
1562                         return err;
1563         }
1564
1565         if (cda[CTA_PROTOINFO]) {
1566                 err = ctnetlink_change_protoinfo(ct, cda);
1567                 if (err < 0)
1568                         return err;
1569         }
1570
1571 #if defined(CONFIG_NF_CONNTRACK_MARK)
1572         if (cda[CTA_MARK])
1573                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1574 #endif
1575
1576 #ifdef CONFIG_NF_NAT_NEEDED
1577         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1578                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1579                 if (err < 0)
1580                         return err;
1581         }
1582 #endif
1583         if (cda[CTA_LABELS]) {
1584                 err = ctnetlink_attach_labels(ct, cda);
1585                 if (err < 0)
1586                         return err;
1587         }
1588
1589         return 0;
1590 }
1591
1592 static struct nf_conn *
1593 ctnetlink_create_conntrack(struct net *net, u16 zone,
1594                            const struct nlattr * const cda[],
1595                            struct nf_conntrack_tuple *otuple,
1596                            struct nf_conntrack_tuple *rtuple,
1597                            u8 u3)
1598 {
1599         struct nf_conn *ct;
1600         int err = -EINVAL;
1601         struct nf_conntrack_helper *helper;
1602         struct nf_conn_tstamp *tstamp;
1603
1604         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1605         if (IS_ERR(ct))
1606                 return ERR_PTR(-ENOMEM);
1607
1608         if (!cda[CTA_TIMEOUT])
1609                 goto err1;
1610         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1611
1612         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1613
1614         rcu_read_lock();
1615         if (cda[CTA_HELP]) {
1616                 char *helpname = NULL;
1617                 struct nlattr *helpinfo = NULL;
1618
1619                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1620                 if (err < 0)
1621                         goto err2;
1622
1623                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1624                                                     nf_ct_protonum(ct));
1625                 if (helper == NULL) {
1626                         rcu_read_unlock();
1627 #ifdef CONFIG_MODULES
1628                         if (request_module("nfct-helper-%s", helpname) < 0) {
1629                                 err = -EOPNOTSUPP;
1630                                 goto err1;
1631                         }
1632
1633                         rcu_read_lock();
1634                         helper = __nf_conntrack_helper_find(helpname,
1635                                                             nf_ct_l3num(ct),
1636                                                             nf_ct_protonum(ct));
1637                         if (helper) {
1638                                 err = -EAGAIN;
1639                                 goto err2;
1640                         }
1641                         rcu_read_unlock();
1642 #endif
1643                         err = -EOPNOTSUPP;
1644                         goto err1;
1645                 } else {
1646                         struct nf_conn_help *help;
1647
1648                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1649                         if (help == NULL) {
1650                                 err = -ENOMEM;
1651                                 goto err2;
1652                         }
1653                         /* set private helper data if allowed. */
1654                         if (helper->from_nlattr)
1655                                 helper->from_nlattr(helpinfo, ct);
1656
1657                         /* not in hash table yet so not strictly necessary */
1658                         RCU_INIT_POINTER(help->helper, helper);
1659                 }
1660         } else {
1661                 /* try an implicit helper assignation */
1662                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1663                 if (err < 0)
1664                         goto err2;
1665         }
1666
1667         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1668                 err = ctnetlink_change_nat(ct, cda);
1669                 if (err < 0)
1670                         goto err2;
1671         }
1672
1673         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1674         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1675         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1676         nf_ct_labels_ext_add(ct);
1677
1678         /* we must add conntrack extensions before confirmation. */
1679         ct->status |= IPS_CONFIRMED;
1680
1681         if (cda[CTA_STATUS]) {
1682                 err = ctnetlink_change_status(ct, cda);
1683                 if (err < 0)
1684                         goto err2;
1685         }
1686
1687 #ifdef CONFIG_NF_NAT_NEEDED
1688         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1689                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1690                 if (err < 0)
1691                         goto err2;
1692         }
1693 #endif
1694
1695         memset(&ct->proto, 0, sizeof(ct->proto));
1696         if (cda[CTA_PROTOINFO]) {
1697                 err = ctnetlink_change_protoinfo(ct, cda);
1698                 if (err < 0)
1699                         goto err2;
1700         }
1701
1702 #if defined(CONFIG_NF_CONNTRACK_MARK)
1703         if (cda[CTA_MARK])
1704                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1705 #endif
1706
1707         /* setup master conntrack: this is a confirmed expectation */
1708         if (cda[CTA_TUPLE_MASTER]) {
1709                 struct nf_conntrack_tuple master;
1710                 struct nf_conntrack_tuple_hash *master_h;
1711                 struct nf_conn *master_ct;
1712
1713                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1714                 if (err < 0)
1715                         goto err2;
1716
1717                 master_h = nf_conntrack_find_get(net, zone, &master);
1718                 if (master_h == NULL) {
1719                         err = -ENOENT;
1720                         goto err2;
1721                 }
1722                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1723                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1724                 ct->master = master_ct;
1725         }
1726         tstamp = nf_conn_tstamp_find(ct);
1727         if (tstamp)
1728                 tstamp->start = ktime_to_ns(ktime_get_real());
1729
1730         err = nf_conntrack_hash_check_insert(ct);
1731         if (err < 0)
1732                 goto err2;
1733
1734         rcu_read_unlock();
1735
1736         return ct;
1737
1738 err2:
1739         rcu_read_unlock();
1740 err1:
1741         nf_conntrack_free(ct);
1742         return ERR_PTR(err);
1743 }
1744
1745 static int
1746 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1747                         const struct nlmsghdr *nlh,
1748                         const struct nlattr * const cda[])
1749 {
1750         struct net *net = sock_net(ctnl);
1751         struct nf_conntrack_tuple otuple, rtuple;
1752         struct nf_conntrack_tuple_hash *h = NULL;
1753         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1754         struct nf_conn *ct;
1755         u_int8_t u3 = nfmsg->nfgen_family;
1756         u16 zone;
1757         int err;
1758
1759         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1760         if (err < 0)
1761                 return err;
1762
1763         if (cda[CTA_TUPLE_ORIG]) {
1764                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1765                 if (err < 0)
1766                         return err;
1767         }
1768
1769         if (cda[CTA_TUPLE_REPLY]) {
1770                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1771                 if (err < 0)
1772                         return err;
1773         }
1774
1775         if (cda[CTA_TUPLE_ORIG])
1776                 h = nf_conntrack_find_get(net, zone, &otuple);
1777         else if (cda[CTA_TUPLE_REPLY])
1778                 h = nf_conntrack_find_get(net, zone, &rtuple);
1779
1780         if (h == NULL) {
1781                 err = -ENOENT;
1782                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1783                         enum ip_conntrack_events events;
1784
1785                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1786                                 return -EINVAL;
1787
1788                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1789                                                         &rtuple, u3);
1790                         if (IS_ERR(ct))
1791                                 return PTR_ERR(ct);
1792
1793                         err = 0;
1794                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1795                                 events = IPCT_RELATED;
1796                         else
1797                                 events = IPCT_NEW;
1798
1799                         if (cda[CTA_LABELS] &&
1800                             ctnetlink_attach_labels(ct, cda) == 0)
1801                                 events |= (1 << IPCT_LABEL);
1802
1803                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1804                                                       (1 << IPCT_ASSURED) |
1805                                                       (1 << IPCT_HELPER) |
1806                                                       (1 << IPCT_PROTOINFO) |
1807                                                       (1 << IPCT_NATSEQADJ) |
1808                                                       (1 << IPCT_MARK) | events,
1809                                                       ct, NETLINK_CB(skb).portid,
1810                                                       nlmsg_report(nlh));
1811                         nf_ct_put(ct);
1812                 }
1813
1814                 return err;
1815         }
1816         /* implicit 'else' */
1817
1818         err = -EEXIST;
1819         ct = nf_ct_tuplehash_to_ctrack(h);
1820         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1821                 spin_lock_bh(&nf_conntrack_lock);
1822                 err = ctnetlink_change_conntrack(ct, cda);
1823                 spin_unlock_bh(&nf_conntrack_lock);
1824                 if (err == 0) {
1825                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1826                                                       (1 << IPCT_ASSURED) |
1827                                                       (1 << IPCT_HELPER) |
1828                                                       (1 << IPCT_LABEL) |
1829                                                       (1 << IPCT_PROTOINFO) |
1830                                                       (1 << IPCT_NATSEQADJ) |
1831                                                       (1 << IPCT_MARK),
1832                                                       ct, NETLINK_CB(skb).portid,
1833                                                       nlmsg_report(nlh));
1834                 }
1835         }
1836
1837         nf_ct_put(ct);
1838         return err;
1839 }
1840
1841 static int
1842 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1843                                 __u16 cpu, const struct ip_conntrack_stat *st)
1844 {
1845         struct nlmsghdr *nlh;
1846         struct nfgenmsg *nfmsg;
1847         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1848
1849         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1850         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1851         if (nlh == NULL)
1852                 goto nlmsg_failure;
1853
1854         nfmsg = nlmsg_data(nlh);
1855         nfmsg->nfgen_family = AF_UNSPEC;
1856         nfmsg->version      = NFNETLINK_V0;
1857         nfmsg->res_id       = htons(cpu);
1858
1859         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1860             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1861             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1862             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1863             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1864             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1865             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1866             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1867             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1868                                 htonl(st->insert_failed)) ||
1869             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1870             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1871             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1872             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1873                                 htonl(st->search_restart)))
1874                 goto nla_put_failure;
1875
1876         nlmsg_end(skb, nlh);
1877         return skb->len;
1878
1879 nla_put_failure:
1880 nlmsg_failure:
1881         nlmsg_cancel(skb, nlh);
1882         return -1;
1883 }
1884
1885 static int
1886 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1887 {
1888         int cpu;
1889         struct net *net = sock_net(skb->sk);
1890
1891         if (cb->args[0] == nr_cpu_ids)
1892                 return 0;
1893
1894         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1895                 const struct ip_conntrack_stat *st;
1896
1897                 if (!cpu_possible(cpu))
1898                         continue;
1899
1900                 st = per_cpu_ptr(net->ct.stat, cpu);
1901                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1902                                                     NETLINK_CB(cb->skb).portid,
1903                                                     cb->nlh->nlmsg_seq,
1904                                                     cpu, st) < 0)
1905                                 break;
1906         }
1907         cb->args[0] = cpu;
1908
1909         return skb->len;
1910 }
1911
1912 static int
1913 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1914                       const struct nlmsghdr *nlh,
1915                       const struct nlattr * const cda[])
1916 {
1917         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1918                 struct netlink_dump_control c = {
1919                         .dump = ctnetlink_ct_stat_cpu_dump,
1920                 };
1921                 return netlink_dump_start(ctnl, skb, nlh, &c);
1922         }
1923
1924         return 0;
1925 }
1926
1927 static int
1928 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1929                             struct net *net)
1930 {
1931         struct nlmsghdr *nlh;
1932         struct nfgenmsg *nfmsg;
1933         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1934         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1935
1936         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1937         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1938         if (nlh == NULL)
1939                 goto nlmsg_failure;
1940
1941         nfmsg = nlmsg_data(nlh);
1942         nfmsg->nfgen_family = AF_UNSPEC;
1943         nfmsg->version      = NFNETLINK_V0;
1944         nfmsg->res_id       = 0;
1945
1946         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1947                 goto nla_put_failure;
1948
1949         nlmsg_end(skb, nlh);
1950         return skb->len;
1951
1952 nla_put_failure:
1953 nlmsg_failure:
1954         nlmsg_cancel(skb, nlh);
1955         return -1;
1956 }
1957
1958 static int
1959 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1960                   const struct nlmsghdr *nlh,
1961                   const struct nlattr * const cda[])
1962 {
1963         struct sk_buff *skb2;
1964         int err;
1965
1966         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1967         if (skb2 == NULL)
1968                 return -ENOMEM;
1969
1970         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1971                                           nlh->nlmsg_seq,
1972                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1973                                           sock_net(skb->sk));
1974         if (err <= 0)
1975                 goto free;
1976
1977         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1978         if (err < 0)
1979                 goto out;
1980
1981         return 0;
1982
1983 free:
1984         kfree_skb(skb2);
1985 out:
1986         /* this avoids a loop in nfnetlink. */
1987         return err == -EAGAIN ? -ENOBUFS : err;
1988 }
1989
1990 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1991         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1992         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1993         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1994         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1995         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1996         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
1997                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1998         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1999         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2000         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2001         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2002         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2003 };
2004
2005 static struct nf_conntrack_expect *
2006 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2007                        struct nf_conntrack_helper *helper,
2008                        struct nf_conntrack_tuple *tuple,
2009                        struct nf_conntrack_tuple *mask);
2010
2011 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2012 static size_t
2013 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2014 {
2015         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2016                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2017                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2018                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2019                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2020                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2021                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2022                + nla_total_size(0) /* CTA_PROTOINFO */
2023                + nla_total_size(0) /* CTA_HELP */
2024                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2025                + ctnetlink_secctx_size(ct)
2026 #ifdef CONFIG_NF_NAT_NEEDED
2027                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2028                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2029 #endif
2030 #ifdef CONFIG_NF_CONNTRACK_MARK
2031                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2032 #endif
2033                + ctnetlink_proto_size(ct)
2034                ;
2035 }
2036
2037 static int
2038 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2039 {
2040         struct nlattr *nest_parms;
2041
2042         rcu_read_lock();
2043         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2044         if (!nest_parms)
2045                 goto nla_put_failure;
2046         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2047                 goto nla_put_failure;
2048         nla_nest_end(skb, nest_parms);
2049
2050         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2051         if (!nest_parms)
2052                 goto nla_put_failure;
2053         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2054                 goto nla_put_failure;
2055         nla_nest_end(skb, nest_parms);
2056
2057         if (nf_ct_zone(ct)) {
2058                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2059                         goto nla_put_failure;
2060         }
2061
2062         if (ctnetlink_dump_id(skb, ct) < 0)
2063                 goto nla_put_failure;
2064
2065         if (ctnetlink_dump_status(skb, ct) < 0)
2066                 goto nla_put_failure;
2067
2068         if (ctnetlink_dump_timeout(skb, ct) < 0)
2069                 goto nla_put_failure;
2070
2071         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2072                 goto nla_put_failure;
2073
2074         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2075                 goto nla_put_failure;
2076
2077 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2078         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2079                 goto nla_put_failure;
2080 #endif
2081         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2082                 goto nla_put_failure;
2083
2084         if ((ct->status & IPS_SEQ_ADJUST) &&
2085             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
2086                 goto nla_put_failure;
2087
2088 #ifdef CONFIG_NF_CONNTRACK_MARK
2089         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2090                 goto nla_put_failure;
2091 #endif
2092         if (ctnetlink_dump_labels(skb, ct) < 0)
2093                 goto nla_put_failure;
2094         rcu_read_unlock();
2095         return 0;
2096
2097 nla_put_failure:
2098         rcu_read_unlock();
2099         return -ENOSPC;
2100 }
2101
2102 static int
2103 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2104 {
2105         int err;
2106
2107         if (cda[CTA_TIMEOUT]) {
2108                 err = ctnetlink_change_timeout(ct, cda);
2109                 if (err < 0)
2110                         return err;
2111         }
2112         if (cda[CTA_STATUS]) {
2113                 err = ctnetlink_change_status(ct, cda);
2114                 if (err < 0)
2115                         return err;
2116         }
2117         if (cda[CTA_HELP]) {
2118                 err = ctnetlink_change_helper(ct, cda);
2119                 if (err < 0)
2120                         return err;
2121         }
2122         if (cda[CTA_LABELS]) {
2123                 err = ctnetlink_attach_labels(ct, cda);
2124                 if (err < 0)
2125                         return err;
2126         }
2127 #if defined(CONFIG_NF_CONNTRACK_MARK)
2128         if (cda[CTA_MARK])
2129                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2130 #endif
2131         return 0;
2132 }
2133
2134 static int
2135 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2136 {
2137         struct nlattr *cda[CTA_MAX+1];
2138         int ret;
2139
2140         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2141         if (ret < 0)
2142                 return ret;
2143
2144         spin_lock_bh(&nf_conntrack_lock);
2145         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2146         spin_unlock_bh(&nf_conntrack_lock);
2147
2148         return ret;
2149 }
2150
2151 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2152                                        const struct nf_conn *ct,
2153                                        struct nf_conntrack_tuple *tuple,
2154                                        struct nf_conntrack_tuple *mask)
2155 {
2156         int err;
2157
2158         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2159                                     nf_ct_l3num(ct));
2160         if (err < 0)
2161                 return err;
2162
2163         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2164                                      nf_ct_l3num(ct));
2165 }
2166
2167 static int
2168 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2169                                 u32 portid, u32 report)
2170 {
2171         struct nlattr *cda[CTA_EXPECT_MAX+1];
2172         struct nf_conntrack_tuple tuple, mask;
2173         struct nf_conntrack_helper *helper;
2174         struct nf_conntrack_expect *exp;
2175         int err;
2176
2177         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2178         if (err < 0)
2179                 return err;
2180
2181         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2182                                           ct, &tuple, &mask);
2183         if (err < 0)
2184                 return err;
2185
2186         if (cda[CTA_EXPECT_HELP_NAME]) {
2187                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2188
2189                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2190                                                     nf_ct_protonum(ct));
2191                 if (helper == NULL)
2192                         return -EOPNOTSUPP;
2193         }
2194
2195         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2196                                      helper, &tuple, &mask);
2197         if (IS_ERR(exp))
2198                 return PTR_ERR(exp);
2199
2200         err = nf_ct_expect_related_report(exp, portid, report);
2201         if (err < 0) {
2202                 nf_ct_expect_put(exp);
2203                 return err;
2204         }
2205
2206         return 0;
2207 }
2208
2209 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2210         .build_size     = ctnetlink_nfqueue_build_size,
2211         .build          = ctnetlink_nfqueue_build,
2212         .parse          = ctnetlink_nfqueue_parse,
2213         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2214 };
2215 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2216
2217 /***********************************************************************
2218  * EXPECT
2219  ***********************************************************************/
2220
2221 static inline int
2222 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2223                          const struct nf_conntrack_tuple *tuple,
2224                          enum ctattr_expect type)
2225 {
2226         struct nlattr *nest_parms;
2227
2228         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2229         if (!nest_parms)
2230                 goto nla_put_failure;
2231         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2232                 goto nla_put_failure;
2233         nla_nest_end(skb, nest_parms);
2234
2235         return 0;
2236
2237 nla_put_failure:
2238         return -1;
2239 }
2240
2241 static inline int
2242 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2243                         const struct nf_conntrack_tuple *tuple,
2244                         const struct nf_conntrack_tuple_mask *mask)
2245 {
2246         int ret;
2247         struct nf_conntrack_l3proto *l3proto;
2248         struct nf_conntrack_l4proto *l4proto;
2249         struct nf_conntrack_tuple m;
2250         struct nlattr *nest_parms;
2251
2252         memset(&m, 0xFF, sizeof(m));
2253         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2254         m.src.u.all = mask->src.u.all;
2255         m.dst.protonum = tuple->dst.protonum;
2256
2257         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2258         if (!nest_parms)
2259                 goto nla_put_failure;
2260
2261         rcu_read_lock();
2262         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2263         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2264         if (ret >= 0) {
2265                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2266                                                tuple->dst.protonum);
2267         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2268         }
2269         rcu_read_unlock();
2270
2271         if (unlikely(ret < 0))
2272                 goto nla_put_failure;
2273
2274         nla_nest_end(skb, nest_parms);
2275
2276         return 0;
2277
2278 nla_put_failure:
2279         return -1;
2280 }
2281
2282 static const union nf_inet_addr any_addr;
2283
2284 static int
2285 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2286                           const struct nf_conntrack_expect *exp)
2287 {
2288         struct nf_conn *master = exp->master;
2289         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2290         struct nf_conn_help *help;
2291 #ifdef CONFIG_NF_NAT_NEEDED
2292         struct nlattr *nest_parms;
2293         struct nf_conntrack_tuple nat_tuple = {};
2294 #endif
2295         struct nf_ct_helper_expectfn *expfn;
2296
2297         if (timeout < 0)
2298                 timeout = 0;
2299
2300         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2301                 goto nla_put_failure;
2302         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2303                 goto nla_put_failure;
2304         if (ctnetlink_exp_dump_tuple(skb,
2305                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2306                                  CTA_EXPECT_MASTER) < 0)
2307                 goto nla_put_failure;
2308
2309 #ifdef CONFIG_NF_NAT_NEEDED
2310         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2311             exp->saved_proto.all) {
2312                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2313                 if (!nest_parms)
2314                         goto nla_put_failure;
2315
2316                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2317                         goto nla_put_failure;
2318
2319                 nat_tuple.src.l3num = nf_ct_l3num(master);
2320                 nat_tuple.src.u3 = exp->saved_addr;
2321                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2322                 nat_tuple.src.u = exp->saved_proto;
2323
2324                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2325                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2326                         goto nla_put_failure;
2327                 nla_nest_end(skb, nest_parms);
2328         }
2329 #endif
2330         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2331             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2332             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2333             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2334                 goto nla_put_failure;
2335         help = nfct_help(master);
2336         if (help) {
2337                 struct nf_conntrack_helper *helper;
2338
2339                 helper = rcu_dereference(help->helper);
2340                 if (helper &&
2341                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2342                         goto nla_put_failure;
2343         }
2344         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2345         if (expfn != NULL &&
2346             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2347                 goto nla_put_failure;
2348
2349         return 0;
2350
2351 nla_put_failure:
2352         return -1;
2353 }
2354
2355 static int
2356 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2357                         int event, const struct nf_conntrack_expect *exp)
2358 {
2359         struct nlmsghdr *nlh;
2360         struct nfgenmsg *nfmsg;
2361         unsigned int flags = portid ? NLM_F_MULTI : 0;
2362
2363         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2364         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2365         if (nlh == NULL)
2366                 goto nlmsg_failure;
2367
2368         nfmsg = nlmsg_data(nlh);
2369         nfmsg->nfgen_family = exp->tuple.src.l3num;
2370         nfmsg->version      = NFNETLINK_V0;
2371         nfmsg->res_id       = 0;
2372
2373         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2374                 goto nla_put_failure;
2375
2376         nlmsg_end(skb, nlh);
2377         return skb->len;
2378
2379 nlmsg_failure:
2380 nla_put_failure:
2381         nlmsg_cancel(skb, nlh);
2382         return -1;
2383 }
2384
2385 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2386 static int
2387 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2388 {
2389         struct nf_conntrack_expect *exp = item->exp;
2390         struct net *net = nf_ct_exp_net(exp);
2391         struct nlmsghdr *nlh;
2392         struct nfgenmsg *nfmsg;
2393         struct sk_buff *skb;
2394         unsigned int type, group;
2395         int flags = 0;
2396
2397         if (events & (1 << IPEXP_DESTROY)) {
2398                 type = IPCTNL_MSG_EXP_DELETE;
2399                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2400         } else if (events & (1 << IPEXP_NEW)) {
2401                 type = IPCTNL_MSG_EXP_NEW;
2402                 flags = NLM_F_CREATE|NLM_F_EXCL;
2403                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2404         } else
2405                 return 0;
2406
2407         if (!item->report && !nfnetlink_has_listeners(net, group))
2408                 return 0;
2409
2410         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2411         if (skb == NULL)
2412                 goto errout;
2413
2414         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2415         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2416         if (nlh == NULL)
2417                 goto nlmsg_failure;
2418
2419         nfmsg = nlmsg_data(nlh);
2420         nfmsg->nfgen_family = exp->tuple.src.l3num;
2421         nfmsg->version      = NFNETLINK_V0;
2422         nfmsg->res_id       = 0;
2423
2424         rcu_read_lock();
2425         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2426                 goto nla_put_failure;
2427         rcu_read_unlock();
2428
2429         nlmsg_end(skb, nlh);
2430         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2431         return 0;
2432
2433 nla_put_failure:
2434         rcu_read_unlock();
2435         nlmsg_cancel(skb, nlh);
2436 nlmsg_failure:
2437         kfree_skb(skb);
2438 errout:
2439         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2440         return 0;
2441 }
2442 #endif
2443 static int ctnetlink_exp_done(struct netlink_callback *cb)
2444 {
2445         if (cb->args[1])
2446                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2447         return 0;
2448 }
2449
2450 static int
2451 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2452 {
2453         struct net *net = sock_net(skb->sk);
2454         struct nf_conntrack_expect *exp, *last;
2455         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2456         u_int8_t l3proto = nfmsg->nfgen_family;
2457
2458         rcu_read_lock();
2459         last = (struct nf_conntrack_expect *)cb->args[1];
2460         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2461 restart:
2462                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2463                                      hnode) {
2464                         if (l3proto && exp->tuple.src.l3num != l3proto)
2465                                 continue;
2466                         if (cb->args[1]) {
2467                                 if (exp != last)
2468                                         continue;
2469                                 cb->args[1] = 0;
2470                         }
2471                         if (ctnetlink_exp_fill_info(skb,
2472                                                     NETLINK_CB(cb->skb).portid,
2473                                                     cb->nlh->nlmsg_seq,
2474                                                     IPCTNL_MSG_EXP_NEW,
2475                                                     exp) < 0) {
2476                                 if (!atomic_inc_not_zero(&exp->use))
2477                                         continue;
2478                                 cb->args[1] = (unsigned long)exp;
2479                                 goto out;
2480                         }
2481                 }
2482                 if (cb->args[1]) {
2483                         cb->args[1] = 0;
2484                         goto restart;
2485                 }
2486         }
2487 out:
2488         rcu_read_unlock();
2489         if (last)
2490                 nf_ct_expect_put(last);
2491
2492         return skb->len;
2493 }
2494
2495 static int
2496 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2497 {
2498         struct nf_conntrack_expect *exp, *last;
2499         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2500         struct nf_conn *ct = cb->data;
2501         struct nf_conn_help *help = nfct_help(ct);
2502         u_int8_t l3proto = nfmsg->nfgen_family;
2503
2504         if (cb->args[0])
2505                 return 0;
2506
2507         rcu_read_lock();
2508         last = (struct nf_conntrack_expect *)cb->args[1];
2509 restart:
2510         hlist_for_each_entry(exp, &help->expectations, lnode) {
2511                 if (l3proto && exp->tuple.src.l3num != l3proto)
2512                         continue;
2513                 if (cb->args[1]) {
2514                         if (exp != last)
2515                                 continue;
2516                         cb->args[1] = 0;
2517                 }
2518                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2519                                             cb->nlh->nlmsg_seq,
2520                                             IPCTNL_MSG_EXP_NEW,
2521                                             exp) < 0) {
2522                         if (!atomic_inc_not_zero(&exp->use))
2523                                 continue;
2524                         cb->args[1] = (unsigned long)exp;
2525                         goto out;
2526                 }
2527         }
2528         if (cb->args[1]) {
2529                 cb->args[1] = 0;
2530                 goto restart;
2531         }
2532         cb->args[0] = 1;
2533 out:
2534         rcu_read_unlock();
2535         if (last)
2536                 nf_ct_expect_put(last);
2537
2538         return skb->len;
2539 }
2540
2541 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2542                                  const struct nlmsghdr *nlh,
2543                                  const struct nlattr * const cda[])
2544 {
2545         int err;
2546         struct net *net = sock_net(ctnl);
2547         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2548         u_int8_t u3 = nfmsg->nfgen_family;
2549         struct nf_conntrack_tuple tuple;
2550         struct nf_conntrack_tuple_hash *h;
2551         struct nf_conn *ct;
2552         u16 zone = 0;
2553         struct netlink_dump_control c = {
2554                 .dump = ctnetlink_exp_ct_dump_table,
2555                 .done = ctnetlink_exp_done,
2556         };
2557
2558         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2559         if (err < 0)
2560                 return err;
2561
2562         if (cda[CTA_EXPECT_ZONE]) {
2563                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2564                 if (err < 0)
2565                         return err;
2566         }
2567
2568         h = nf_conntrack_find_get(net, zone, &tuple);
2569         if (!h)
2570                 return -ENOENT;
2571
2572         ct = nf_ct_tuplehash_to_ctrack(h);
2573         c.data = ct;
2574
2575         err = netlink_dump_start(ctnl, skb, nlh, &c);
2576         nf_ct_put(ct);
2577
2578         return err;
2579 }
2580
2581 static int
2582 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2583                      const struct nlmsghdr *nlh,
2584                      const struct nlattr * const cda[])
2585 {
2586         struct net *net = sock_net(ctnl);
2587         struct nf_conntrack_tuple tuple;
2588         struct nf_conntrack_expect *exp;
2589         struct sk_buff *skb2;
2590         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2591         u_int8_t u3 = nfmsg->nfgen_family;
2592         u16 zone;
2593         int err;
2594
2595         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2596                 if (cda[CTA_EXPECT_MASTER])
2597                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2598                 else {
2599                         struct netlink_dump_control c = {
2600                                 .dump = ctnetlink_exp_dump_table,
2601                                 .done = ctnetlink_exp_done,
2602                         };
2603                         return netlink_dump_start(ctnl, skb, nlh, &c);
2604                 }
2605         }
2606
2607         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2608         if (err < 0)
2609                 return err;
2610
2611         if (cda[CTA_EXPECT_TUPLE])
2612                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2613         else if (cda[CTA_EXPECT_MASTER])
2614                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2615         else
2616                 return -EINVAL;
2617
2618         if (err < 0)
2619                 return err;
2620
2621         exp = nf_ct_expect_find_get(net, zone, &tuple);
2622         if (!exp)
2623                 return -ENOENT;
2624
2625         if (cda[CTA_EXPECT_ID]) {
2626                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2627                 if (ntohl(id) != (u32)(unsigned long)exp) {
2628                         nf_ct_expect_put(exp);
2629                         return -ENOENT;
2630                 }
2631         }
2632
2633         err = -ENOMEM;
2634         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2635         if (skb2 == NULL) {
2636                 nf_ct_expect_put(exp);
2637                 goto out;
2638         }
2639
2640         rcu_read_lock();
2641         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2642                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2643         rcu_read_unlock();
2644         nf_ct_expect_put(exp);
2645         if (err <= 0)
2646                 goto free;
2647
2648         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2649         if (err < 0)
2650                 goto out;
2651
2652         return 0;
2653
2654 free:
2655         kfree_skb(skb2);
2656 out:
2657         /* this avoids a loop in nfnetlink. */
2658         return err == -EAGAIN ? -ENOBUFS : err;
2659 }
2660
2661 static int
2662 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2663                      const struct nlmsghdr *nlh,
2664                      const struct nlattr * const cda[])
2665 {
2666         struct net *net = sock_net(ctnl);
2667         struct nf_conntrack_expect *exp;
2668         struct nf_conntrack_tuple tuple;
2669         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2670         struct hlist_node *next;
2671         u_int8_t u3 = nfmsg->nfgen_family;
2672         unsigned int i;
2673         u16 zone;
2674         int err;
2675
2676         if (cda[CTA_EXPECT_TUPLE]) {
2677                 /* delete a single expect by tuple */
2678                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2679                 if (err < 0)
2680                         return err;
2681
2682                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2683                 if (err < 0)
2684                         return err;
2685
2686                 /* bump usage count to 2 */
2687                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2688                 if (!exp)
2689                         return -ENOENT;
2690
2691                 if (cda[CTA_EXPECT_ID]) {
2692                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2693                         if (ntohl(id) != (u32)(unsigned long)exp) {
2694                                 nf_ct_expect_put(exp);
2695                                 return -ENOENT;
2696                         }
2697                 }
2698
2699                 /* after list removal, usage count == 1 */
2700                 spin_lock_bh(&nf_conntrack_lock);
2701                 if (del_timer(&exp->timeout)) {
2702                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2703                                                    nlmsg_report(nlh));
2704                         nf_ct_expect_put(exp);
2705                 }
2706                 spin_unlock_bh(&nf_conntrack_lock);
2707                 /* have to put what we 'get' above.
2708                  * after this line usage count == 0 */
2709                 nf_ct_expect_put(exp);
2710         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2711                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2712                 struct nf_conn_help *m_help;
2713
2714                 /* delete all expectations for this helper */
2715                 spin_lock_bh(&nf_conntrack_lock);
2716                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2717                         hlist_for_each_entry_safe(exp, next,
2718                                                   &net->ct.expect_hash[i],
2719                                                   hnode) {
2720                                 m_help = nfct_help(exp->master);
2721                                 if (!strcmp(m_help->helper->name, name) &&
2722                                     del_timer(&exp->timeout)) {
2723                                         nf_ct_unlink_expect_report(exp,
2724                                                         NETLINK_CB(skb).portid,
2725                                                         nlmsg_report(nlh));
2726                                         nf_ct_expect_put(exp);
2727                                 }
2728                         }
2729                 }
2730                 spin_unlock_bh(&nf_conntrack_lock);
2731         } else {
2732                 /* This basically means we have to flush everything*/
2733                 spin_lock_bh(&nf_conntrack_lock);
2734                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2735                         hlist_for_each_entry_safe(exp, next,
2736                                                   &net->ct.expect_hash[i],
2737                                                   hnode) {
2738                                 if (del_timer(&exp->timeout)) {
2739                                         nf_ct_unlink_expect_report(exp,
2740                                                         NETLINK_CB(skb).portid,
2741                                                         nlmsg_report(nlh));
2742                                         nf_ct_expect_put(exp);
2743                                 }
2744                         }
2745                 }
2746                 spin_unlock_bh(&nf_conntrack_lock);
2747         }
2748
2749         return 0;
2750 }
2751 static int
2752 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2753                         const struct nlattr * const cda[])
2754 {
2755         if (cda[CTA_EXPECT_TIMEOUT]) {
2756                 if (!del_timer(&x->timeout))
2757                         return -ETIME;
2758
2759                 x->timeout.expires = jiffies +
2760                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2761                 add_timer(&x->timeout);
2762         }
2763         return 0;
2764 }
2765
2766 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2767         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2768         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2769 };
2770
2771 static int
2772 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2773                            struct nf_conntrack_expect *exp,
2774                            u_int8_t u3)
2775 {
2776 #ifdef CONFIG_NF_NAT_NEEDED
2777         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2778         struct nf_conntrack_tuple nat_tuple = {};
2779         int err;
2780
2781         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2782         if (err < 0)
2783                 return err;
2784
2785         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2786                 return -EINVAL;
2787
2788         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2789                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2790         if (err < 0)
2791                 return err;
2792
2793         exp->saved_addr = nat_tuple.src.u3;
2794         exp->saved_proto = nat_tuple.src.u;
2795         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2796
2797         return 0;
2798 #else
2799         return -EOPNOTSUPP;
2800 #endif
2801 }
2802
2803 static struct nf_conntrack_expect *
2804 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2805                        struct nf_conntrack_helper *helper,
2806                        struct nf_conntrack_tuple *tuple,
2807                        struct nf_conntrack_tuple *mask)
2808 {
2809         u_int32_t class = 0;
2810         struct nf_conntrack_expect *exp;
2811         struct nf_conn_help *help;
2812         int err;
2813
2814         if (cda[CTA_EXPECT_CLASS] && helper) {
2815                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2816                 if (class > helper->expect_class_max)
2817                         return ERR_PTR(-EINVAL);
2818         }
2819         exp = nf_ct_expect_alloc(ct);
2820         if (!exp)
2821                 return ERR_PTR(-ENOMEM);
2822
2823         help = nfct_help(ct);
2824         if (!help) {
2825                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2826                         err = -EINVAL;
2827                         goto err_out;
2828                 }
2829                 exp->timeout.expires =
2830                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2831
2832                 exp->flags = NF_CT_EXPECT_USERSPACE;
2833                 if (cda[CTA_EXPECT_FLAGS]) {
2834                         exp->flags |=
2835                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2836                 }
2837         } else {
2838                 if (cda[CTA_EXPECT_FLAGS]) {
2839                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2840                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2841                 } else
2842                         exp->flags = 0;
2843         }
2844         if (cda[CTA_EXPECT_FN]) {
2845                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2846                 struct nf_ct_helper_expectfn *expfn;
2847
2848                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2849                 if (expfn == NULL) {
2850                         err = -EINVAL;
2851                         goto err_out;
2852                 }
2853                 exp->expectfn = expfn->expectfn;
2854         } else
2855                 exp->expectfn = NULL;
2856
2857         exp->class = class;
2858         exp->master = ct;
2859         exp->helper = helper;
2860         exp->tuple = *tuple;
2861         exp->mask.src.u3 = mask->src.u3;
2862         exp->mask.src.u.all = mask->src.u.all;
2863
2864         if (cda[CTA_EXPECT_NAT]) {
2865                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2866                                                  exp, nf_ct_l3num(ct));
2867                 if (err < 0)
2868                         goto err_out;
2869         }
2870         return exp;
2871 err_out:
2872         nf_ct_expect_put(exp);
2873         return ERR_PTR(err);
2874 }
2875
2876 static int
2877 ctnetlink_create_expect(struct net *net, u16 zone,
2878                         const struct nlattr * const cda[],
2879                         u_int8_t u3, u32 portid, int report)
2880 {
2881         struct nf_conntrack_tuple tuple, mask, master_tuple;
2882         struct nf_conntrack_tuple_hash *h = NULL;
2883         struct nf_conntrack_helper *helper = NULL;
2884         struct nf_conntrack_expect *exp;
2885         struct nf_conn *ct;
2886         int err;
2887
2888         /* caller guarantees that those three CTA_EXPECT_* exist */
2889         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2890         if (err < 0)
2891                 return err;
2892         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2893         if (err < 0)
2894                 return err;
2895         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2896         if (err < 0)
2897                 return err;
2898
2899         /* Look for master conntrack of this expectation */
2900         h = nf_conntrack_find_get(net, zone, &master_tuple);
2901         if (!h)
2902                 return -ENOENT;
2903         ct = nf_ct_tuplehash_to_ctrack(h);
2904
2905         if (cda[CTA_EXPECT_HELP_NAME]) {
2906                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2907
2908                 helper = __nf_conntrack_helper_find(helpname, u3,
2909                                                     nf_ct_protonum(ct));
2910                 if (helper == NULL) {
2911 #ifdef CONFIG_MODULES
2912                         if (request_module("nfct-helper-%s", helpname) < 0) {
2913                                 err = -EOPNOTSUPP;
2914                                 goto err_ct;
2915                         }
2916                         helper = __nf_conntrack_helper_find(helpname, u3,
2917                                                             nf_ct_protonum(ct));
2918                         if (helper) {
2919                                 err = -EAGAIN;
2920                                 goto err_ct;
2921                         }
2922 #endif
2923                         err = -EOPNOTSUPP;
2924                         goto err_ct;
2925                 }
2926         }
2927
2928         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
2929         if (IS_ERR(exp)) {
2930                 err = PTR_ERR(exp);
2931                 goto err_ct;
2932         }
2933
2934         err = nf_ct_expect_related_report(exp, portid, report);
2935         if (err < 0)
2936                 goto err_exp;
2937
2938         return 0;
2939 err_exp:
2940         nf_ct_expect_put(exp);
2941 err_ct:
2942         nf_ct_put(ct);
2943         return err;
2944 }
2945
2946 static int
2947 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2948                      const struct nlmsghdr *nlh,
2949                      const struct nlattr * const cda[])
2950 {
2951         struct net *net = sock_net(ctnl);
2952         struct nf_conntrack_tuple tuple;
2953         struct nf_conntrack_expect *exp;
2954         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2955         u_int8_t u3 = nfmsg->nfgen_family;
2956         u16 zone;
2957         int err;
2958
2959         if (!cda[CTA_EXPECT_TUPLE]
2960             || !cda[CTA_EXPECT_MASK]
2961             || !cda[CTA_EXPECT_MASTER])
2962                 return -EINVAL;
2963
2964         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2965         if (err < 0)
2966                 return err;
2967
2968         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2969         if (err < 0)
2970                 return err;
2971
2972         spin_lock_bh(&nf_conntrack_lock);
2973         exp = __nf_ct_expect_find(net, zone, &tuple);
2974
2975         if (!exp) {
2976                 spin_unlock_bh(&nf_conntrack_lock);
2977                 err = -ENOENT;
2978                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2979                         err = ctnetlink_create_expect(net, zone, cda,
2980                                                       u3,
2981                                                       NETLINK_CB(skb).portid,
2982                                                       nlmsg_report(nlh));
2983                 }
2984                 return err;
2985         }
2986
2987         err = -EEXIST;
2988         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2989                 err = ctnetlink_change_expect(exp, cda);
2990         spin_unlock_bh(&nf_conntrack_lock);
2991
2992         return err;
2993 }
2994
2995 static int
2996 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
2997                              const struct ip_conntrack_stat *st)
2998 {
2999         struct nlmsghdr *nlh;
3000         struct nfgenmsg *nfmsg;
3001         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3002
3003         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3004         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3005         if (nlh == NULL)
3006                 goto nlmsg_failure;
3007
3008         nfmsg = nlmsg_data(nlh);
3009         nfmsg->nfgen_family = AF_UNSPEC;
3010         nfmsg->version      = NFNETLINK_V0;
3011         nfmsg->res_id       = htons(cpu);
3012
3013         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3014             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3015             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3016                 goto nla_put_failure;
3017
3018         nlmsg_end(skb, nlh);
3019         return skb->len;
3020
3021 nla_put_failure:
3022 nlmsg_failure:
3023         nlmsg_cancel(skb, nlh);
3024         return -1;
3025 }
3026
3027 static int
3028 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3029 {
3030         int cpu;
3031         struct net *net = sock_net(skb->sk);
3032
3033         if (cb->args[0] == nr_cpu_ids)
3034                 return 0;
3035
3036         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3037                 const struct ip_conntrack_stat *st;
3038
3039                 if (!cpu_possible(cpu))
3040                         continue;
3041
3042                 st = per_cpu_ptr(net->ct.stat, cpu);
3043                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3044                                                  cb->nlh->nlmsg_seq,
3045                                                  cpu, st) < 0)
3046                         break;
3047         }
3048         cb->args[0] = cpu;
3049
3050         return skb->len;
3051 }
3052
3053 static int
3054 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3055                        const struct nlmsghdr *nlh,
3056                        const struct nlattr * const cda[])
3057 {
3058         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3059                 struct netlink_dump_control c = {
3060                         .dump = ctnetlink_exp_stat_cpu_dump,
3061                 };
3062                 return netlink_dump_start(ctnl, skb, nlh, &c);
3063         }
3064
3065         return 0;
3066 }
3067
3068 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3069 static struct nf_ct_event_notifier ctnl_notifier = {
3070         .fcn = ctnetlink_conntrack_event,
3071 };
3072
3073 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3074         .fcn = ctnetlink_expect_event,
3075 };
3076 #endif
3077
3078 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3079         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3080                                             .attr_count = CTA_MAX,
3081                                             .policy = ct_nla_policy },
3082         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3083                                             .attr_count = CTA_MAX,
3084                                             .policy = ct_nla_policy },
3085         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3086                                             .attr_count = CTA_MAX,
3087                                             .policy = ct_nla_policy },
3088         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3089                                             .attr_count = CTA_MAX,
3090                                             .policy = ct_nla_policy },
3091         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3092         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3093         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3094         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3095 };
3096
3097 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3098         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3099                                             .attr_count = CTA_EXPECT_MAX,
3100                                             .policy = exp_nla_policy },
3101         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3102                                             .attr_count = CTA_EXPECT_MAX,
3103                                             .policy = exp_nla_policy },
3104         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3105                                             .attr_count = CTA_EXPECT_MAX,
3106                                             .policy = exp_nla_policy },
3107         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3108 };
3109
3110 static const struct nfnetlink_subsystem ctnl_subsys = {
3111         .name                           = "conntrack",
3112         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3113         .cb_count                       = IPCTNL_MSG_MAX,
3114         .cb                             = ctnl_cb,
3115 };
3116
3117 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3118         .name                           = "conntrack_expect",
3119         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3120         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3121         .cb                             = ctnl_exp_cb,
3122 };
3123
3124 MODULE_ALIAS("ip_conntrack_netlink");
3125 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3126 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3127
3128 static int __net_init ctnetlink_net_init(struct net *net)
3129 {
3130 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3131         int ret;
3132
3133         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3134         if (ret < 0) {
3135                 pr_err("ctnetlink_init: cannot register notifier.\n");
3136                 goto err_out;
3137         }
3138
3139         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3140         if (ret < 0) {
3141                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3142                 goto err_unreg_notifier;
3143         }
3144 #endif
3145         return 0;
3146
3147 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3148 err_unreg_notifier:
3149         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3150 err_out:
3151         return ret;
3152 #endif
3153 }
3154
3155 static void ctnetlink_net_exit(struct net *net)
3156 {
3157 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3158         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3159         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3160 #endif
3161 }
3162
3163 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3164 {
3165         struct net *net;
3166
3167         list_for_each_entry(net, net_exit_list, exit_list)
3168                 ctnetlink_net_exit(net);
3169 }
3170
3171 static struct pernet_operations ctnetlink_net_ops = {
3172         .init           = ctnetlink_net_init,
3173         .exit_batch     = ctnetlink_net_exit_batch,
3174 };
3175
3176 static int __init ctnetlink_init(void)
3177 {
3178         int ret;
3179
3180         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3181         ret = nfnetlink_subsys_register(&ctnl_subsys);
3182         if (ret < 0) {
3183                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3184                 goto err_out;
3185         }
3186
3187         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3188         if (ret < 0) {
3189                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3190                 goto err_unreg_subsys;
3191         }
3192
3193         ret = register_pernet_subsys(&ctnetlink_net_ops);
3194         if (ret < 0) {
3195                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3196                 goto err_unreg_exp_subsys;
3197         }
3198 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3199         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3200         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3201 #endif
3202         return 0;
3203
3204 err_unreg_exp_subsys:
3205         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3206 err_unreg_subsys:
3207         nfnetlink_subsys_unregister(&ctnl_subsys);
3208 err_out:
3209         return ret;
3210 }
3211
3212 static void __exit ctnetlink_exit(void)
3213 {
3214         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3215
3216         unregister_pernet_subsys(&ctnetlink_net_ops);
3217         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3218         nfnetlink_subsys_unregister(&ctnl_subsys);
3219 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3220         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3221 #endif
3222 }
3223
3224 module_init(ctnetlink_init);
3225 module_exit(ctnetlink_exit);