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_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #ifdef CONFIG_NF_NAT_NEEDED
  49 #include <net/netfilter/nf_nat_core.h>
  50 #include <net/netfilter/nf_nat_l4proto.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 MODULE_LICENSE("GPL");
  58
  59 static char __initdata version[] = "0.93";
  60
  61 static inline int
  62 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  63                             const struct nf_conntrack_tuple *tuple,
  64                             struct nf_conntrack_l4proto *l4proto)
  65 {
  66         int ret = 0;
  67         struct nlattr *nest_parms;
  68
  69         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  70         if (!nest_parms)
  71                 goto nla_put_failure;
  72         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  73                 goto nla_put_failure;
  74
  75         if (likely(l4proto->tuple_to_nlattr))
  76                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  77
  78         nla_nest_end(skb, nest_parms);
  79
  80         return ret;
  81
  82 nla_put_failure:
  83         return -1;
  84 }
  85
  86 static inline int
  87 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  88                          const struct nf_conntrack_tuple *tuple,
  89                          struct nf_conntrack_l3proto *l3proto)
  90 {
  91         int ret = 0;
  92         struct nlattr *nest_parms;
  93
  94         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  95         if (!nest_parms)
  96                 goto nla_put_failure;
  97
  98         if (likely(l3proto->tuple_to_nlattr))
  99                 ret = l3proto->tuple_to_nlattr(skb, tuple);
 100
 101         nla_nest_end(skb, nest_parms);
 102
 103         return ret;
 104
 105 nla_put_failure:
 106         return -1;
 107 }
 108
 109 static int
 110 ctnetlink_dump_tuples(struct sk_buff *skb,
 111                       const struct nf_conntrack_tuple *tuple)
 112 {
 113         int ret;
 114         struct nf_conntrack_l3proto *l3proto;
 115         struct nf_conntrack_l4proto *l4proto;
 116
 117         rcu_read_lock();
 118         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 119         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 120
 121         if (ret >= 0) {
 122                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 123                                                tuple->dst.protonum);
 124                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 125         }
 126         rcu_read_unlock();
 127         return ret;
 128 }
 129
 130 static inline int
 131 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 132 {
 133         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 134                 goto nla_put_failure;
 135         return 0;
 136
 137 nla_put_failure:
 138         return -1;
 139 }
 140
 141 static inline int
 142 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 143 {
 144         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 145
 146         if (timeout < 0)
 147                 timeout = 0;
 148
 149         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 150                 goto nla_put_failure;
 151         return 0;
 152
 153 nla_put_failure:
 154         return -1;
 155 }
 156
 157 static inline int
 158 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 159 {
 160         struct nf_conntrack_l4proto *l4proto;
 161         struct nlattr *nest_proto;
 162         int ret;
 163
 164         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 165         if (!l4proto->to_nlattr)
 166                 return 0;
 167
 168         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 169         if (!nest_proto)
 170                 goto nla_put_failure;
 171
 172         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 173
 174         nla_nest_end(skb, nest_proto);
 175
 176         return ret;
 177
 178 nla_put_failure:
 179         return -1;
 180 }
 181
 182 static inline int
 183 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 184 {
 185         struct nlattr *nest_helper;
 186         const struct nf_conn_help *help = nfct_help(ct);
 187         struct nf_conntrack_helper *helper;
 188
 189         if (!help)
 190                 return 0;
 191
 192         helper = rcu_dereference(help->helper);
 193         if (!helper)
 194                 goto out;
 195
 196         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 197         if (!nest_helper)
 198                 goto nla_put_failure;
 199         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 200                 goto nla_put_failure;
 201
 202         if (helper->to_nlattr)
 203                 helper->to_nlattr(skb, ct);
 204
 205         nla_nest_end(skb, nest_helper);
 206 out:
 207         return 0;
 208
 209 nla_put_failure:
 210         return -1;
 211 }
 212
 213 static int
 214 dump_counters(struct sk_buff *skb, u64 pkts, u64 bytes,
 215               enum ip_conntrack_dir dir)
 216 {
 217         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 218         struct nlattr *nest_count;
 219
 220         nest_count = nla_nest_start(skb, type | NLA_F_NESTED);
 221         if (!nest_count)
 222                 goto nla_put_failure;
 223
 224         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 225             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 226                 goto nla_put_failure;
 227
 228         nla_nest_end(skb, nest_count);
 229
 230         return 0;
 231
 232 nla_put_failure:
 233         return -1;
 234 }
 235
 236 static int
 237 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 238                         enum ip_conntrack_dir dir, int type)
 239 {
 240         struct nf_conn_counter *acct;
 241         u64 pkts, bytes;
 242
 243         acct = nf_conn_acct_find(ct);
 244         if (!acct)
 245                 return 0;
 246
 247         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 248                 pkts = atomic64_xchg(&acct[dir].packets, 0);
 249                 bytes = atomic64_xchg(&acct[dir].bytes, 0);
 250         } else {
 251                 pkts = atomic64_read(&acct[dir].packets);
 252                 bytes = atomic64_read(&acct[dir].bytes);
 253         }
 254         return dump_counters(skb, pkts, bytes, dir);
 255 }
 256
 257 static int
 258 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 259 {
 260         struct nlattr *nest_count;
 261         const struct nf_conn_tstamp *tstamp;
 262
 263         tstamp = nf_conn_tstamp_find(ct);
 264         if (!tstamp)
 265                 return 0;
 266
 267         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 268         if (!nest_count)
 269                 goto nla_put_failure;
 270
 271         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 272             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 273                                                cpu_to_be64(tstamp->stop))))
 274                 goto nla_put_failure;
 275         nla_nest_end(skb, nest_count);
 276
 277         return 0;
 278
 279 nla_put_failure:
 280         return -1;
 281 }
 282
 283 #ifdef CONFIG_NF_CONNTRACK_MARK
 284 static inline int
 285 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 286 {
 287         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 288                 goto nla_put_failure;
 289         return 0;
 290
 291 nla_put_failure:
 292         return -1;
 293 }
 294 #else
 295 #define ctnetlink_dump_mark(a, b) (0)
 296 #endif
 297
 298 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 299 static inline int
 300 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 301 {
 302         struct nlattr *nest_secctx;
 303         int len, ret;
 304         char *secctx;
 305
 306         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 307         if (ret)
 308                 return 0;
 309
 310         ret = -1;
 311         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 312         if (!nest_secctx)
 313                 goto nla_put_failure;
 314
 315         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 316                 goto nla_put_failure;
 317         nla_nest_end(skb, nest_secctx);
 318
 319         ret = 0;
 320 nla_put_failure:
 321         security_release_secctx(secctx, len);
 322         return ret;
 323 }
 324 #else
 325 #define ctnetlink_dump_secctx(a, b) (0)
 326 #endif
 327
 328 #ifdef CONFIG_NF_CONNTRACK_LABELS
 329 static int ctnetlink_label_size(const struct nf_conn *ct)
 330 {
 331         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 332
 333         if (!labels)
 334                 return 0;
 335         return nla_total_size(labels->words * sizeof(long));
 336 }
 337
 338 static int
 339 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 340 {
 341         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 342         unsigned int len, i;
 343
 344         if (!labels)
 345                 return 0;
 346
 347         len = labels->words * sizeof(long);
 348         i = 0;
 349         do {
 350                 if (labels->bits[i] != 0)
 351                         return nla_put(skb, CTA_LABELS, len, labels->bits);
 352                 i++;
 353         } while (i < labels->words);
 354
 355         return 0;
 356 }
 357 #else
 358 #define ctnetlink_dump_labels(a, b) (0)
 359 #define ctnetlink_label_size(a) (0)
 360 #endif
 361
 362 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 363
 364 static inline int
 365 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 366 {
 367         struct nlattr *nest_parms;
 368
 369         if (!(ct->status & IPS_EXPECTED))
 370                 return 0;
 371
 372         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 373         if (!nest_parms)
 374                 goto nla_put_failure;
 375         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 376                 goto nla_put_failure;
 377         nla_nest_end(skb, nest_parms);
 378
 379         return 0;
 380
 381 nla_put_failure:
 382         return -1;
 383 }
 384
 385 static int
 386 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, 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_SEQADJ_CORRECTION_POS,
 395                          htonl(seq->correction_pos)) ||
 396             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 397                          htonl(seq->offset_before)) ||
 398             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 399                          htonl(seq->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_ct_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 412 {
 413         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 414         struct nf_ct_seqadj *seq;
 415
 416         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 417                 return 0;
 418
 419         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 420         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 421                 return -1;
 422
 423         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 424         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 425                 return -1;
 426
 427         return 0;
 428 }
 429
 430 static inline int
 431 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 432 {
 433         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 434                 goto nla_put_failure;
 435         return 0;
 436
 437 nla_put_failure:
 438         return -1;
 439 }
 440
 441 static inline int
 442 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 443 {
 444         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 445                 goto nla_put_failure;
 446         return 0;
 447
 448 nla_put_failure:
 449         return -1;
 450 }
 451
 452 static int
 453 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 454                     struct nf_conn *ct)
 455 {
 456         struct nlmsghdr *nlh;
 457         struct nfgenmsg *nfmsg;
 458         struct nlattr *nest_parms;
 459         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 460
 461         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 462         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 463         if (nlh == NULL)
 464                 goto nlmsg_failure;
 465
 466         nfmsg = nlmsg_data(nlh);
 467         nfmsg->nfgen_family = nf_ct_l3num(ct);
 468         nfmsg->version      = NFNETLINK_V0;
 469         nfmsg->res_id       = 0;
 470
 471         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 472         if (!nest_parms)
 473                 goto nla_put_failure;
 474         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 475                 goto nla_put_failure;
 476         nla_nest_end(skb, nest_parms);
 477
 478         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 479         if (!nest_parms)
 480                 goto nla_put_failure;
 481         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 482                 goto nla_put_failure;
 483         nla_nest_end(skb, nest_parms);
 484
 485         if (nf_ct_zone(ct) &&
 486             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 487                 goto nla_put_failure;
 488
 489         if (ctnetlink_dump_status(skb, ct) < 0 ||
 490             ctnetlink_dump_timeout(skb, ct) < 0 ||
 491             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL, type) < 0 ||
 492             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY, type) < 0 ||
 493             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 494             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 495             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 496             ctnetlink_dump_mark(skb, ct) < 0 ||
 497             ctnetlink_dump_secctx(skb, ct) < 0 ||
 498             ctnetlink_dump_labels(skb, ct) < 0 ||
 499             ctnetlink_dump_id(skb, ct) < 0 ||
 500             ctnetlink_dump_use(skb, ct) < 0 ||
 501             ctnetlink_dump_master(skb, ct) < 0 ||
 502             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 503                 goto nla_put_failure;
 504
 505         nlmsg_end(skb, nlh);
 506         return skb->len;
 507
 508 nlmsg_failure:
 509 nla_put_failure:
 510         nlmsg_cancel(skb, nlh);
 511         return -1;
 512 }
 513
 514 static inline size_t
 515 ctnetlink_proto_size(const struct nf_conn *ct)
 516 {
 517         struct nf_conntrack_l3proto *l3proto;
 518         struct nf_conntrack_l4proto *l4proto;
 519         size_t len = 0;
 520
 521         rcu_read_lock();
 522         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 523         len += l3proto->nla_size;
 524
 525         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 526         len += l4proto->nla_size;
 527         rcu_read_unlock();
 528
 529         return len;
 530 }
 531
 532 static inline size_t
 533 ctnetlink_counters_size(const struct nf_conn *ct)
 534 {
 535         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 536                 return 0;
 537         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 538                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 539                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 540                ;
 541 }
 542
 543 static inline int
 544 ctnetlink_secctx_size(const struct nf_conn *ct)
 545 {
 546 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 547         int len, ret;
 548
 549         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 550         if (ret)
 551                 return 0;
 552
 553         return nla_total_size(0) /* CTA_SECCTX */
 554                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 555 #else
 556         return 0;
 557 #endif
 558 }
 559
 560 static inline size_t
 561 ctnetlink_timestamp_size(const struct nf_conn *ct)
 562 {
 563 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 564         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 565                 return 0;
 566         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 567 #else
 568         return 0;
 569 #endif
 570 }
 571
 572 static inline size_t
 573 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 574 {
 575         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 576                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 577                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 578                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 579                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 580                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 581                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 582                + ctnetlink_counters_size(ct)
 583                + ctnetlink_timestamp_size(ct)
 584                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 585                + nla_total_size(0) /* CTA_PROTOINFO */
 586                + nla_total_size(0) /* CTA_HELP */
 587                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 588                + ctnetlink_secctx_size(ct)
 589 #ifdef CONFIG_NF_NAT_NEEDED
 590                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 591                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 592 #endif
 593 #ifdef CONFIG_NF_CONNTRACK_MARK
 594                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 595 #endif
 596                + ctnetlink_proto_size(ct)
 597                + ctnetlink_label_size(ct)
 598                ;
 599 }
 600
 601 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 602 static int
 603 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 604 {
 605         struct net *net;
 606         struct nlmsghdr *nlh;
 607         struct nfgenmsg *nfmsg;
 608         struct nlattr *nest_parms;
 609         struct nf_conn *ct = item->ct;
 610         struct sk_buff *skb;
 611         unsigned int type;
 612         unsigned int flags = 0, group;
 613         int err;
 614
 615         /* ignore our fake conntrack entry */
 616         if (nf_ct_is_untracked(ct))
 617                 return 0;
 618
 619         if (events & (1 << IPCT_DESTROY)) {
 620                 type = IPCTNL_MSG_CT_DELETE;
 621                 group = NFNLGRP_CONNTRACK_DESTROY;
 622         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 623                 type = IPCTNL_MSG_CT_NEW;
 624                 flags = NLM_F_CREATE|NLM_F_EXCL;
 625                 group = NFNLGRP_CONNTRACK_NEW;
 626         } else  if (events) {
 627                 type = IPCTNL_MSG_CT_NEW;
 628                 group = NFNLGRP_CONNTRACK_UPDATE;
 629         } else
 630                 return 0;
 631
 632         net = nf_ct_net(ct);
 633         if (!item->report && !nfnetlink_has_listeners(net, group))
 634                 return 0;
 635
 636         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 637         if (skb == NULL)
 638                 goto errout;
 639
 640         type |= NFNL_SUBSYS_CTNETLINK << 8;
 641         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 642         if (nlh == NULL)
 643                 goto nlmsg_failure;
 644
 645         nfmsg = nlmsg_data(nlh);
 646         nfmsg->nfgen_family = nf_ct_l3num(ct);
 647         nfmsg->version  = NFNETLINK_V0;
 648         nfmsg->res_id   = 0;
 649
 650         rcu_read_lock();
 651         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 652         if (!nest_parms)
 653                 goto nla_put_failure;
 654         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 655                 goto nla_put_failure;
 656         nla_nest_end(skb, nest_parms);
 657
 658         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 659         if (!nest_parms)
 660                 goto nla_put_failure;
 661         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 662                 goto nla_put_failure;
 663         nla_nest_end(skb, nest_parms);
 664
 665         if (nf_ct_zone(ct) &&
 666             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 667                 goto nla_put_failure;
 668
 669         if (ctnetlink_dump_id(skb, ct) < 0)
 670                 goto nla_put_failure;
 671
 672         if (ctnetlink_dump_status(skb, ct) < 0)
 673                 goto nla_put_failure;
 674
 675         if (events & (1 << IPCT_DESTROY)) {
 676                 if (ctnetlink_dump_counters(skb, ct,
 677                                             IP_CT_DIR_ORIGINAL, type) < 0 ||
 678                     ctnetlink_dump_counters(skb, ct,
 679                                             IP_CT_DIR_REPLY, type) < 0 ||
 680                     ctnetlink_dump_timestamp(skb, ct) < 0)
 681                         goto nla_put_failure;
 682         } else {
 683                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 684                         goto nla_put_failure;
 685
 686                 if (events & (1 << IPCT_PROTOINFO)
 687                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 688                         goto nla_put_failure;
 689
 690                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 691                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 692                         goto nla_put_failure;
 693
 694 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 695                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 696                     && ctnetlink_dump_secctx(skb, ct) < 0)
 697                         goto nla_put_failure;
 698 #endif
 699                 if (events & (1 << IPCT_LABEL) &&
 700                      ctnetlink_dump_labels(skb, ct) < 0)
 701                         goto nla_put_failure;
 702
 703                 if (events & (1 << IPCT_RELATED) &&
 704                     ctnetlink_dump_master(skb, ct) < 0)
 705                         goto nla_put_failure;
 706
 707                 if (events & (1 << IPCT_SEQADJ) &&
 708                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 709                         goto nla_put_failure;
 710         }
 711
 712 #ifdef CONFIG_NF_CONNTRACK_MARK
 713         if ((events & (1 << IPCT_MARK) || ct->mark)
 714             && ctnetlink_dump_mark(skb, ct) < 0)
 715                 goto nla_put_failure;
 716 #endif
 717         rcu_read_unlock();
 718
 719         nlmsg_end(skb, nlh);
 720         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 721                              GFP_ATOMIC);
 722         if (err == -ENOBUFS || err == -EAGAIN)
 723                 return -ENOBUFS;
 724
 725         return 0;
 726
 727 nla_put_failure:
 728         rcu_read_unlock();
 729         nlmsg_cancel(skb, nlh);
 730 nlmsg_failure:
 731         kfree_skb(skb);
 732 errout:
 733         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 734                 return -ENOBUFS;
 735
 736         return 0;
 737 }
 738 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 739
 740 static int ctnetlink_done(struct netlink_callback *cb)
 741 {
 742         if (cb->args[1])
 743                 nf_ct_put((struct nf_conn *)cb->args[1]);
 744         if (cb->data)
 745                 kfree(cb->data);
 746         return 0;
 747 }
 748
 749 struct ctnetlink_dump_filter {
 750         struct {
 751                 u_int32_t val;
 752                 u_int32_t mask;
 753         } mark;
 754 };
 755
 756 static int
 757 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 758 {
 759         struct net *net = sock_net(skb->sk);
 760         struct nf_conn *ct, *last;
 761         struct nf_conntrack_tuple_hash *h;
 762         struct hlist_nulls_node *n;
 763         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 764         u_int8_t l3proto = nfmsg->nfgen_family;
 765         int res;
 766 #ifdef CONFIG_NF_CONNTRACK_MARK
 767         const struct ctnetlink_dump_filter *filter = cb->data;
 768 #endif
 769
 770         spin_lock_bh(&nf_conntrack_lock);
 771         last = (struct nf_conn *)cb->args[1];
 772         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 773 restart:
 774                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 775                                          hnnode) {
 776                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 777                                 continue;
 778                         ct = nf_ct_tuplehash_to_ctrack(h);
 779                         /* Dump entries of a given L3 protocol number.
 780                          * If it is not specified, ie. l3proto == 0,
 781                          * then dump everything. */
 782                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 783                                 continue;
 784                         if (cb->args[1]) {
 785                                 if (ct != last)
 786                                         continue;
 787                                 cb->args[1] = 0;
 788                         }
 789 #ifdef CONFIG_NF_CONNTRACK_MARK
 790                         if (filter && !((ct->mark & filter->mark.mask) ==
 791                                         filter->mark.val)) {
 792                                 continue;
 793                         }
 794 #endif
 795                         rcu_read_lock();
 796                         res =
 797                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 798                                             cb->nlh->nlmsg_seq,
 799                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 800                                             ct);
 801                         rcu_read_unlock();
 802                         if (res < 0) {
 803                                 nf_conntrack_get(&ct->ct_general);
 804                                 cb->args[1] = (unsigned long)ct;
 805                                 goto out;
 806                         }
 807                 }
 808                 if (cb->args[1]) {
 809                         cb->args[1] = 0;
 810                         goto restart;
 811                 }
 812         }
 813 out:
 814         spin_unlock_bh(&nf_conntrack_lock);
 815         if (last)
 816                 nf_ct_put(last);
 817
 818         return skb->len;
 819 }
 820
 821 static inline int
 822 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 823 {
 824         struct nlattr *tb[CTA_IP_MAX+1];
 825         struct nf_conntrack_l3proto *l3proto;
 826         int ret = 0;
 827
 828         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 829         if (ret < 0)
 830                 return ret;
 831
 832         rcu_read_lock();
 833         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 834
 835         if (likely(l3proto->nlattr_to_tuple)) {
 836                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 837                                           l3proto->nla_policy);
 838                 if (ret == 0)
 839                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 840         }
 841
 842         rcu_read_unlock();
 843
 844         return ret;
 845 }
 846
 847 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 848         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 849 };
 850
 851 static inline int
 852 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 853                             struct nf_conntrack_tuple *tuple)
 854 {
 855         struct nlattr *tb[CTA_PROTO_MAX+1];
 856         struct nf_conntrack_l4proto *l4proto;
 857         int ret = 0;
 858
 859         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 860         if (ret < 0)
 861                 return ret;
 862
 863         if (!tb[CTA_PROTO_NUM])
 864                 return -EINVAL;
 865         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 866
 867         rcu_read_lock();
 868         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 869
 870         if (likely(l4proto->nlattr_to_tuple)) {
 871                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 872                                           l4proto->nla_policy);
 873                 if (ret == 0)
 874                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 875         }
 876
 877         rcu_read_unlock();
 878
 879         return ret;
 880 }
 881
 882 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 883         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 884         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 885 };
 886
 887 static int
 888 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 889                       struct nf_conntrack_tuple *tuple,
 890                       enum ctattr_type type, u_int8_t l3num)
 891 {
 892         struct nlattr *tb[CTA_TUPLE_MAX+1];
 893         int err;
 894
 895         memset(tuple, 0, sizeof(*tuple));
 896
 897         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 898         if (err < 0)
 899                 return err;
 900
 901         if (!tb[CTA_TUPLE_IP])
 902                 return -EINVAL;
 903
 904         tuple->src.l3num = l3num;
 905
 906         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 907         if (err < 0)
 908                 return err;
 909
 910         if (!tb[CTA_TUPLE_PROTO])
 911                 return -EINVAL;
 912
 913         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 914         if (err < 0)
 915                 return err;
 916
 917         /* orig and expect tuples get DIR_ORIGINAL */
 918         if (type == CTA_TUPLE_REPLY)
 919                 tuple->dst.dir = IP_CT_DIR_REPLY;
 920         else
 921                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 922
 923         return 0;
 924 }
 925
 926 static int
 927 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 928 {
 929         if (attr)
 930 #ifdef CONFIG_NF_CONNTRACK_ZONES
 931                 *zone = ntohs(nla_get_be16(attr));
 932 #else
 933                 return -EOPNOTSUPP;
 934 #endif
 935         else
 936                 *zone = 0;
 937
 938         return 0;
 939 }
 940
 941 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 942         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 943                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 944 };
 945
 946 static inline int
 947 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 948                      struct nlattr **helpinfo)
 949 {
 950         int err;
 951         struct nlattr *tb[CTA_HELP_MAX+1];
 952
 953         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 954         if (err < 0)
 955                 return err;
 956
 957         if (!tb[CTA_HELP_NAME])
 958                 return -EINVAL;
 959
 960         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 961
 962         if (tb[CTA_HELP_INFO])
 963                 *helpinfo = tb[CTA_HELP_INFO];
 964
 965         return 0;
 966 }
 967
 968 #define __CTA_LABELS_MAX_LENGTH ((XT_CONNLABEL_MAXBIT + 1) / BITS_PER_BYTE)
 969 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 970         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 971         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 972         [CTA_STATUS]            = { .type = NLA_U32 },
 973         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 974         [CTA_HELP]              = { .type = NLA_NESTED },
 975         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 976         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 977         [CTA_MARK]              = { .type = NLA_U32 },
 978         [CTA_ID]                = { .type = NLA_U32 },
 979         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 980         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 981         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
 982         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
 983         [CTA_ZONE]              = { .type = NLA_U16 },
 984         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 985         [CTA_LABELS]            = { .type = NLA_BINARY,
 986                                     .len = __CTA_LABELS_MAX_LENGTH },
 987         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
 988                                     .len = __CTA_LABELS_MAX_LENGTH },
 989 };
 990
 991 static int
 992 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 993                         const struct nlmsghdr *nlh,
 994                         const struct nlattr * const cda[])
 995 {
 996         struct net *net = sock_net(ctnl);
 997         struct nf_conntrack_tuple_hash *h;
 998         struct nf_conntrack_tuple tuple;
 999         struct nf_conn *ct;
1000         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1001         u_int8_t u3 = nfmsg->nfgen_family;
1002         u16 zone;
1003         int err;
1004
1005         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1006         if (err < 0)
1007                 return err;
1008
1009         if (cda[CTA_TUPLE_ORIG])
1010                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1011         else if (cda[CTA_TUPLE_REPLY])
1012                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1013         else {
1014                 /* Flush the whole table */
1015                 nf_conntrack_flush_report(net,
1016                                          NETLINK_CB(skb).portid,
1017                                          nlmsg_report(nlh));
1018                 return 0;
1019         }
1020
1021         if (err < 0)
1022                 return err;
1023
1024         h = nf_conntrack_find_get(net, zone, &tuple);
1025         if (!h)
1026                 return -ENOENT;
1027
1028         ct = nf_ct_tuplehash_to_ctrack(h);
1029
1030         if (cda[CTA_ID]) {
1031                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1032                 if (id != (u32)(unsigned long)ct) {
1033                         nf_ct_put(ct);
1034                         return -ENOENT;
1035                 }
1036         }
1037
1038         if (del_timer(&ct->timeout))
1039                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1040
1041         nf_ct_put(ct);
1042
1043         return 0;
1044 }
1045
1046 static int
1047 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1048                         const struct nlmsghdr *nlh,
1049                         const struct nlattr * const cda[])
1050 {
1051         struct net *net = sock_net(ctnl);
1052         struct nf_conntrack_tuple_hash *h;
1053         struct nf_conntrack_tuple tuple;
1054         struct nf_conn *ct;
1055         struct sk_buff *skb2 = NULL;
1056         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1057         u_int8_t u3 = nfmsg->nfgen_family;
1058         u16 zone;
1059         int err;
1060
1061         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1062                 struct netlink_dump_control c = {
1063                         .dump = ctnetlink_dump_table,
1064                         .done = ctnetlink_done,
1065                 };
1066 #ifdef CONFIG_NF_CONNTRACK_MARK
1067                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1068                         struct ctnetlink_dump_filter *filter;
1069
1070                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1071                                          GFP_ATOMIC);
1072                         if (filter == NULL)
1073                                 return -ENOMEM;
1074
1075                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1076                         filter->mark.mask =
1077                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1078                         c.data = filter;
1079                 }
1080 #endif
1081                 return netlink_dump_start(ctnl, skb, nlh, &c);
1082         }
1083
1084         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1085         if (err < 0)
1086                 return err;
1087
1088         if (cda[CTA_TUPLE_ORIG])
1089                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1090         else if (cda[CTA_TUPLE_REPLY])
1091                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1092         else
1093                 return -EINVAL;
1094
1095         if (err < 0)
1096                 return err;
1097
1098         h = nf_conntrack_find_get(net, zone, &tuple);
1099         if (!h)
1100                 return -ENOENT;
1101
1102         ct = nf_ct_tuplehash_to_ctrack(h);
1103
1104         err = -ENOMEM;
1105         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1106         if (skb2 == NULL) {
1107                 nf_ct_put(ct);
1108                 return -ENOMEM;
1109         }
1110
1111         rcu_read_lock();
1112         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1113                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1114         rcu_read_unlock();
1115         nf_ct_put(ct);
1116         if (err <= 0)
1117                 goto free;
1118
1119         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1120         if (err < 0)
1121                 goto out;
1122
1123         return 0;
1124
1125 free:
1126         kfree_skb(skb2);
1127 out:
1128         /* this avoids a loop in nfnetlink. */
1129         return err == -EAGAIN ? -ENOBUFS : err;
1130 }
1131
1132 static int ctnetlink_done_list(struct netlink_callback *cb)
1133 {
1134         if (cb->args[1])
1135                 nf_ct_put((struct nf_conn *)cb->args[1]);
1136         return 0;
1137 }
1138
1139 static int
1140 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb,
1141                     struct hlist_nulls_head *list)
1142 {
1143         struct nf_conn *ct, *last;
1144         struct nf_conntrack_tuple_hash *h;
1145         struct hlist_nulls_node *n;
1146         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1147         u_int8_t l3proto = nfmsg->nfgen_family;
1148         int res;
1149
1150         if (cb->args[2])
1151                 return 0;
1152
1153         spin_lock_bh(&nf_conntrack_lock);
1154         last = (struct nf_conn *)cb->args[1];
1155 restart:
1156         hlist_nulls_for_each_entry(h, n, list, hnnode) {
1157                 ct = nf_ct_tuplehash_to_ctrack(h);
1158                 if (l3proto && nf_ct_l3num(ct) != l3proto)
1159                         continue;
1160                 if (cb->args[1]) {
1161                         if (ct != last)
1162                                 continue;
1163                         cb->args[1] = 0;
1164                 }
1165                 rcu_read_lock();
1166                 res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1167                                           cb->nlh->nlmsg_seq,
1168                                           NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1169                                           ct);
1170                 rcu_read_unlock();
1171                 if (res < 0) {
1172                         nf_conntrack_get(&ct->ct_general);
1173                         cb->args[1] = (unsigned long)ct;
1174                         goto out;
1175                 }
1176         }
1177         if (cb->args[1]) {
1178                 cb->args[1] = 0;
1179                 goto restart;
1180         } else
1181                 cb->args[2] = 1;
1182 out:
1183         spin_unlock_bh(&nf_conntrack_lock);
1184         if (last)
1185                 nf_ct_put(last);
1186
1187         return skb->len;
1188 }
1189
1190 static int
1191 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1192 {
1193         struct net *net = sock_net(skb->sk);
1194
1195         return ctnetlink_dump_list(skb, cb, &net->ct.dying);
1196 }
1197
1198 static int
1199 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1200                        const struct nlmsghdr *nlh,
1201                        const struct nlattr * const cda[])
1202 {
1203         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1204                 struct netlink_dump_control c = {
1205                         .dump = ctnetlink_dump_dying,
1206                         .done = ctnetlink_done_list,
1207                 };
1208                 return netlink_dump_start(ctnl, skb, nlh, &c);
1209         }
1210
1211         return -EOPNOTSUPP;
1212 }
1213
1214 static int
1215 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1216 {
1217         struct net *net = sock_net(skb->sk);
1218
1219         return ctnetlink_dump_list(skb, cb, &net->ct.unconfirmed);
1220 }
1221
1222 static int
1223 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1224                              const struct nlmsghdr *nlh,
1225                              const struct nlattr * const cda[])
1226 {
1227         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1228                 struct netlink_dump_control c = {
1229                         .dump = ctnetlink_dump_unconfirmed,
1230                         .done = ctnetlink_done_list,
1231                 };
1232                 return netlink_dump_start(ctnl, skb, nlh, &c);
1233         }
1234
1235         return -EOPNOTSUPP;
1236 }
1237
1238 #ifdef CONFIG_NF_NAT_NEEDED
1239 static int
1240 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1241                           enum nf_nat_manip_type manip,
1242                           const struct nlattr *attr)
1243 {
1244         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1245         int err;
1246
1247         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1248         if (!parse_nat_setup) {
1249 #ifdef CONFIG_MODULES
1250                 rcu_read_unlock();
1251                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1252                 if (request_module("nf-nat") < 0) {
1253                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1254                         rcu_read_lock();
1255                         return -EOPNOTSUPP;
1256                 }
1257                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1258                 rcu_read_lock();
1259                 if (nfnetlink_parse_nat_setup_hook)
1260                         return -EAGAIN;
1261 #endif
1262                 return -EOPNOTSUPP;
1263         }
1264
1265         err = parse_nat_setup(ct, manip, attr);
1266         if (err == -EAGAIN) {
1267 #ifdef CONFIG_MODULES
1268                 rcu_read_unlock();
1269                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1270                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1271                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1272                         rcu_read_lock();
1273                         return -EOPNOTSUPP;
1274                 }
1275                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1276                 rcu_read_lock();
1277 #else
1278                 err = -EOPNOTSUPP;
1279 #endif
1280         }
1281         return err;
1282 }
1283 #endif
1284
1285 static int
1286 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1287 {
1288         unsigned long d;
1289         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1290         d = ct->status ^ status;
1291
1292         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1293                 /* unchangeable */
1294                 return -EBUSY;
1295
1296         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1297                 /* SEEN_REPLY bit can only be set */
1298                 return -EBUSY;
1299
1300         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1301                 /* ASSURED bit can only be set */
1302                 return -EBUSY;
1303
1304         /* Be careful here, modifying NAT bits can screw up things,
1305          * so don't let users modify them directly if they don't pass
1306          * nf_nat_range. */
1307         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1308         return 0;
1309 }
1310
1311 static int
1312 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1313 {
1314 #ifdef CONFIG_NF_NAT_NEEDED
1315         int ret;
1316
1317         if (cda[CTA_NAT_DST]) {
1318                 ret = ctnetlink_parse_nat_setup(ct,
1319                                                 NF_NAT_MANIP_DST,
1320                                                 cda[CTA_NAT_DST]);
1321                 if (ret < 0)
1322                         return ret;
1323         }
1324         if (cda[CTA_NAT_SRC]) {
1325                 ret = ctnetlink_parse_nat_setup(ct,
1326                                                 NF_NAT_MANIP_SRC,
1327                                                 cda[CTA_NAT_SRC]);
1328                 if (ret < 0)
1329                         return ret;
1330         }
1331         return 0;
1332 #else
1333         return -EOPNOTSUPP;
1334 #endif
1335 }
1336
1337 static inline int
1338 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1339 {
1340         struct nf_conntrack_helper *helper;
1341         struct nf_conn_help *help = nfct_help(ct);
1342         char *helpname = NULL;
1343         struct nlattr *helpinfo = NULL;
1344         int err;
1345
1346         /* don't change helper of sibling connections */
1347         if (ct->master)
1348                 return -EBUSY;
1349
1350         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1351         if (err < 0)
1352                 return err;
1353
1354         if (!strcmp(helpname, "")) {
1355                 if (help && help->helper) {
1356                         /* we had a helper before ... */
1357                         nf_ct_remove_expectations(ct);
1358                         RCU_INIT_POINTER(help->helper, NULL);
1359                 }
1360
1361                 return 0;
1362         }
1363
1364         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1365                                             nf_ct_protonum(ct));
1366         if (helper == NULL) {
1367 #ifdef CONFIG_MODULES
1368                 spin_unlock_bh(&nf_conntrack_lock);
1369
1370                 if (request_module("nfct-helper-%s", helpname) < 0) {
1371                         spin_lock_bh(&nf_conntrack_lock);
1372                         return -EOPNOTSUPP;
1373                 }
1374
1375                 spin_lock_bh(&nf_conntrack_lock);
1376                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1377                                                     nf_ct_protonum(ct));
1378                 if (helper)
1379                         return -EAGAIN;
1380 #endif
1381                 return -EOPNOTSUPP;
1382         }
1383
1384         if (help) {
1385                 if (help->helper == helper) {
1386                         /* update private helper data if allowed. */
1387                         if (helper->from_nlattr)
1388                                 helper->from_nlattr(helpinfo, ct);
1389                         return 0;
1390                 } else
1391                         return -EBUSY;
1392         }
1393
1394         /* we cannot set a helper for an existing conntrack */
1395         return -EOPNOTSUPP;
1396 }
1397
1398 static inline int
1399 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1400 {
1401         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1402
1403         if (!del_timer(&ct->timeout))
1404                 return -ETIME;
1405
1406         ct->timeout.expires = jiffies + timeout * HZ;
1407         add_timer(&ct->timeout);
1408
1409         return 0;
1410 }
1411
1412 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1413         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1414         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1415         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1416 };
1417
1418 static inline int
1419 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1420 {
1421         const struct nlattr *attr = cda[CTA_PROTOINFO];
1422         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1423         struct nf_conntrack_l4proto *l4proto;
1424         int err = 0;
1425
1426         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1427         if (err < 0)
1428                 return err;
1429
1430         rcu_read_lock();
1431         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1432         if (l4proto->from_nlattr)
1433                 err = l4proto->from_nlattr(tb, ct);
1434         rcu_read_unlock();
1435
1436         return err;
1437 }
1438
1439 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1440         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1441         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1442         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1443 };
1444
1445 static inline int
1446 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1447 {
1448         int err;
1449         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1450
1451         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1452         if (err < 0)
1453                 return err;
1454
1455         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1456                 return -EINVAL;
1457
1458         seq->correction_pos =
1459                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1460
1461         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1462                 return -EINVAL;
1463
1464         seq->offset_before =
1465                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1466
1467         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1468                 return -EINVAL;
1469
1470         seq->offset_after =
1471                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1472
1473         return 0;
1474 }
1475
1476 static int
1477 ctnetlink_change_seq_adj(struct nf_conn *ct,
1478                          const struct nlattr * const cda[])
1479 {
1480         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1481         int ret = 0;
1482
1483         if (!seqadj)
1484                 return 0;
1485
1486         if (cda[CTA_SEQ_ADJ_ORIG]) {
1487                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1488                                      cda[CTA_SEQ_ADJ_ORIG]);
1489                 if (ret < 0)
1490                         return ret;
1491
1492                 ct->status |= IPS_SEQ_ADJUST;
1493         }
1494
1495         if (cda[CTA_SEQ_ADJ_REPLY]) {
1496                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1497                                      cda[CTA_SEQ_ADJ_REPLY]);
1498                 if (ret < 0)
1499                         return ret;
1500
1501                 ct->status |= IPS_SEQ_ADJUST;
1502         }
1503
1504         return 0;
1505 }
1506
1507 static int
1508 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1509 {
1510 #ifdef CONFIG_NF_CONNTRACK_LABELS
1511         size_t len = nla_len(cda[CTA_LABELS]);
1512         const void *mask = cda[CTA_LABELS_MASK];
1513
1514         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1515                 return -EINVAL;
1516
1517         if (mask) {
1518                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1519                     nla_len(cda[CTA_LABELS_MASK]) != len)
1520                         return -EINVAL;
1521                 mask = nla_data(cda[CTA_LABELS_MASK]);
1522         }
1523
1524         len /= sizeof(u32);
1525
1526         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1527 #else
1528         return -EOPNOTSUPP;
1529 #endif
1530 }
1531
1532 static int
1533 ctnetlink_change_conntrack(struct nf_conn *ct,
1534                            const struct nlattr * const cda[])
1535 {
1536         int err;
1537
1538         /* only allow NAT changes and master assignation for new conntracks */
1539         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1540                 return -EOPNOTSUPP;
1541
1542         if (cda[CTA_HELP]) {
1543                 err = ctnetlink_change_helper(ct, cda);
1544                 if (err < 0)
1545                         return err;
1546         }
1547
1548         if (cda[CTA_TIMEOUT]) {
1549                 err = ctnetlink_change_timeout(ct, cda);
1550                 if (err < 0)
1551                         return err;
1552         }
1553
1554         if (cda[CTA_STATUS]) {
1555                 err = ctnetlink_change_status(ct, cda);
1556                 if (err < 0)
1557                         return err;
1558         }
1559
1560         if (cda[CTA_PROTOINFO]) {
1561                 err = ctnetlink_change_protoinfo(ct, cda);
1562                 if (err < 0)
1563                         return err;
1564         }
1565
1566 #if defined(CONFIG_NF_CONNTRACK_MARK)
1567         if (cda[CTA_MARK])
1568                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1569 #endif
1570
1571         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1572                 err = ctnetlink_change_seq_adj(ct, cda);
1573                 if (err < 0)
1574                         return err;
1575         }
1576
1577         if (cda[CTA_LABELS]) {
1578                 err = ctnetlink_attach_labels(ct, cda);
1579                 if (err < 0)
1580                         return err;
1581         }
1582
1583         return 0;
1584 }
1585
1586 static struct nf_conn *
1587 ctnetlink_create_conntrack(struct net *net, u16 zone,
1588                            const struct nlattr * const cda[],
1589                            struct nf_conntrack_tuple *otuple,
1590                            struct nf_conntrack_tuple *rtuple,
1591                            u8 u3)
1592 {
1593         struct nf_conn *ct;
1594         int err = -EINVAL;
1595         struct nf_conntrack_helper *helper;
1596         struct nf_conn_tstamp *tstamp;
1597
1598         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1599         if (IS_ERR(ct))
1600                 return ERR_PTR(-ENOMEM);
1601
1602         if (!cda[CTA_TIMEOUT])
1603                 goto err1;
1604         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1605
1606         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1607
1608         rcu_read_lock();
1609         if (cda[CTA_HELP]) {
1610                 char *helpname = NULL;
1611                 struct nlattr *helpinfo = NULL;
1612
1613                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1614                 if (err < 0)
1615                         goto err2;
1616
1617                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1618                                                     nf_ct_protonum(ct));
1619                 if (helper == NULL) {
1620                         rcu_read_unlock();
1621 #ifdef CONFIG_MODULES
1622                         if (request_module("nfct-helper-%s", helpname) < 0) {
1623                                 err = -EOPNOTSUPP;
1624                                 goto err1;
1625                         }
1626
1627                         rcu_read_lock();
1628                         helper = __nf_conntrack_helper_find(helpname,
1629                                                             nf_ct_l3num(ct),
1630                                                             nf_ct_protonum(ct));
1631                         if (helper) {
1632                                 err = -EAGAIN;
1633                                 goto err2;
1634                         }
1635                         rcu_read_unlock();
1636 #endif
1637                         err = -EOPNOTSUPP;
1638                         goto err1;
1639                 } else {
1640                         struct nf_conn_help *help;
1641
1642                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1643                         if (help == NULL) {
1644                                 err = -ENOMEM;
1645                                 goto err2;
1646                         }
1647                         /* set private helper data if allowed. */
1648                         if (helper->from_nlattr)
1649                                 helper->from_nlattr(helpinfo, ct);
1650
1651                         /* not in hash table yet so not strictly necessary */
1652                         RCU_INIT_POINTER(help->helper, helper);
1653                 }
1654         } else {
1655                 /* try an implicit helper assignation */
1656                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1657                 if (err < 0)
1658                         goto err2;
1659         }
1660
1661         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1662                 err = ctnetlink_change_nat(ct, cda);
1663                 if (err < 0)
1664                         goto err2;
1665         }
1666
1667         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1668         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1669         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1670         nf_ct_labels_ext_add(ct);
1671
1672         /* we must add conntrack extensions before confirmation. */
1673         ct->status |= IPS_CONFIRMED;
1674
1675         if (cda[CTA_STATUS]) {
1676                 err = ctnetlink_change_status(ct, cda);
1677                 if (err < 0)
1678                         goto err2;
1679         }
1680
1681         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1682                 err = ctnetlink_change_seq_adj(ct, cda);
1683                 if (err < 0)
1684                         goto err2;
1685         }
1686
1687         memset(&ct->proto, 0, sizeof(ct->proto));
1688         if (cda[CTA_PROTOINFO]) {
1689                 err = ctnetlink_change_protoinfo(ct, cda);
1690                 if (err < 0)
1691                         goto err2;
1692         }
1693
1694 #if defined(CONFIG_NF_CONNTRACK_MARK)
1695         if (cda[CTA_MARK])
1696                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1697 #endif
1698
1699         /* setup master conntrack: this is a confirmed expectation */
1700         if (cda[CTA_TUPLE_MASTER]) {
1701                 struct nf_conntrack_tuple master;
1702                 struct nf_conntrack_tuple_hash *master_h;
1703                 struct nf_conn *master_ct;
1704
1705                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1706                 if (err < 0)
1707                         goto err2;
1708
1709                 master_h = nf_conntrack_find_get(net, zone, &master);
1710                 if (master_h == NULL) {
1711                         err = -ENOENT;
1712                         goto err2;
1713                 }
1714                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1715                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1716                 ct->master = master_ct;
1717         }
1718         tstamp = nf_conn_tstamp_find(ct);
1719         if (tstamp)
1720                 tstamp->start = ktime_to_ns(ktime_get_real());
1721
1722         err = nf_conntrack_hash_check_insert(ct);
1723         if (err < 0)
1724                 goto err2;
1725
1726         rcu_read_unlock();
1727
1728         return ct;
1729
1730 err2:
1731         rcu_read_unlock();
1732 err1:
1733         nf_conntrack_free(ct);
1734         return ERR_PTR(err);
1735 }
1736
1737 static int
1738 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1739                         const struct nlmsghdr *nlh,
1740                         const struct nlattr * const cda[])
1741 {
1742         struct net *net = sock_net(ctnl);
1743         struct nf_conntrack_tuple otuple, rtuple;
1744         struct nf_conntrack_tuple_hash *h = NULL;
1745         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1746         struct nf_conn *ct;
1747         u_int8_t u3 = nfmsg->nfgen_family;
1748         u16 zone;
1749         int err;
1750
1751         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1752         if (err < 0)
1753                 return err;
1754
1755         if (cda[CTA_TUPLE_ORIG]) {
1756                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1757                 if (err < 0)
1758                         return err;
1759         }
1760
1761         if (cda[CTA_TUPLE_REPLY]) {
1762                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1763                 if (err < 0)
1764                         return err;
1765         }
1766
1767         if (cda[CTA_TUPLE_ORIG])
1768                 h = nf_conntrack_find_get(net, zone, &otuple);
1769         else if (cda[CTA_TUPLE_REPLY])
1770                 h = nf_conntrack_find_get(net, zone, &rtuple);
1771
1772         if (h == NULL) {
1773                 err = -ENOENT;
1774                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1775                         enum ip_conntrack_events events;
1776
1777                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1778                                 return -EINVAL;
1779
1780                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1781                                                         &rtuple, u3);
1782                         if (IS_ERR(ct))
1783                                 return PTR_ERR(ct);
1784
1785                         err = 0;
1786                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1787                                 events = IPCT_RELATED;
1788                         else
1789                                 events = IPCT_NEW;
1790
1791                         if (cda[CTA_LABELS] &&
1792                             ctnetlink_attach_labels(ct, cda) == 0)
1793                                 events |= (1 << IPCT_LABEL);
1794
1795                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1796                                                       (1 << IPCT_ASSURED) |
1797                                                       (1 << IPCT_HELPER) |
1798                                                       (1 << IPCT_PROTOINFO) |
1799                                                       (1 << IPCT_SEQADJ) |
1800                                                       (1 << IPCT_MARK) | events,
1801                                                       ct, NETLINK_CB(skb).portid,
1802                                                       nlmsg_report(nlh));
1803                         nf_ct_put(ct);
1804                 }
1805
1806                 return err;
1807         }
1808         /* implicit 'else' */
1809
1810         err = -EEXIST;
1811         ct = nf_ct_tuplehash_to_ctrack(h);
1812         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1813                 spin_lock_bh(&nf_conntrack_lock);
1814                 err = ctnetlink_change_conntrack(ct, cda);
1815                 spin_unlock_bh(&nf_conntrack_lock);
1816                 if (err == 0) {
1817                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1818                                                       (1 << IPCT_ASSURED) |
1819                                                       (1 << IPCT_HELPER) |
1820                                                       (1 << IPCT_LABEL) |
1821                                                       (1 << IPCT_PROTOINFO) |
1822                                                       (1 << IPCT_SEQADJ) |
1823                                                       (1 << IPCT_MARK),
1824                                                       ct, NETLINK_CB(skb).portid,
1825                                                       nlmsg_report(nlh));
1826                 }
1827         }
1828
1829         nf_ct_put(ct);
1830         return err;
1831 }
1832
1833 static int
1834 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1835                                 __u16 cpu, const struct ip_conntrack_stat *st)
1836 {
1837         struct nlmsghdr *nlh;
1838         struct nfgenmsg *nfmsg;
1839         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1840
1841         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1842         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1843         if (nlh == NULL)
1844                 goto nlmsg_failure;
1845
1846         nfmsg = nlmsg_data(nlh);
1847         nfmsg->nfgen_family = AF_UNSPEC;
1848         nfmsg->version      = NFNETLINK_V0;
1849         nfmsg->res_id       = htons(cpu);
1850
1851         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1852             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1853             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1854             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1855             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1856             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1857             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1858             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1859             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1860                                 htonl(st->insert_failed)) ||
1861             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1862             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1863             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1864             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1865                                 htonl(st->search_restart)))
1866                 goto nla_put_failure;
1867
1868         nlmsg_end(skb, nlh);
1869         return skb->len;
1870
1871 nla_put_failure:
1872 nlmsg_failure:
1873         nlmsg_cancel(skb, nlh);
1874         return -1;
1875 }
1876
1877 static int
1878 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1879 {
1880         int cpu;
1881         struct net *net = sock_net(skb->sk);
1882
1883         if (cb->args[0] == nr_cpu_ids)
1884                 return 0;
1885
1886         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1887                 const struct ip_conntrack_stat *st;
1888
1889                 if (!cpu_possible(cpu))
1890                         continue;
1891
1892                 st = per_cpu_ptr(net->ct.stat, cpu);
1893                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1894                                                     NETLINK_CB(cb->skb).portid,
1895                                                     cb->nlh->nlmsg_seq,
1896                                                     cpu, st) < 0)
1897                                 break;
1898         }
1899         cb->args[0] = cpu;
1900
1901         return skb->len;
1902 }
1903
1904 static int
1905 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1906                       const struct nlmsghdr *nlh,
1907                       const struct nlattr * const cda[])
1908 {
1909         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1910                 struct netlink_dump_control c = {
1911                         .dump = ctnetlink_ct_stat_cpu_dump,
1912                 };
1913                 return netlink_dump_start(ctnl, skb, nlh, &c);
1914         }
1915
1916         return 0;
1917 }
1918
1919 static int
1920 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1921                             struct net *net)
1922 {
1923         struct nlmsghdr *nlh;
1924         struct nfgenmsg *nfmsg;
1925         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1926         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1927
1928         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1929         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1930         if (nlh == NULL)
1931                 goto nlmsg_failure;
1932
1933         nfmsg = nlmsg_data(nlh);
1934         nfmsg->nfgen_family = AF_UNSPEC;
1935         nfmsg->version      = NFNETLINK_V0;
1936         nfmsg->res_id       = 0;
1937
1938         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1939                 goto nla_put_failure;
1940
1941         nlmsg_end(skb, nlh);
1942         return skb->len;
1943
1944 nla_put_failure:
1945 nlmsg_failure:
1946         nlmsg_cancel(skb, nlh);
1947         return -1;
1948 }
1949
1950 static int
1951 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1952                   const struct nlmsghdr *nlh,
1953                   const struct nlattr * const cda[])
1954 {
1955         struct sk_buff *skb2;
1956         int err;
1957
1958         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1959         if (skb2 == NULL)
1960                 return -ENOMEM;
1961
1962         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1963                                           nlh->nlmsg_seq,
1964                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1965                                           sock_net(skb->sk));
1966         if (err <= 0)
1967                 goto free;
1968
1969         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1970         if (err < 0)
1971                 goto out;
1972
1973         return 0;
1974
1975 free:
1976         kfree_skb(skb2);
1977 out:
1978         /* this avoids a loop in nfnetlink. */
1979         return err == -EAGAIN ? -ENOBUFS : err;
1980 }
1981
1982 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1983         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1984         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1985         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1986         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1987         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1988         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
1989                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1990         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1991         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1992         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
1993         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
1994         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
1995 };
1996
1997 static struct nf_conntrack_expect *
1998 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
1999                        struct nf_conntrack_helper *helper,
2000                        struct nf_conntrack_tuple *tuple,
2001                        struct nf_conntrack_tuple *mask);
2002
2003 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2004 static size_t
2005 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2006 {
2007         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2008                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2009                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2010                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2011                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2012                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2013                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2014                + nla_total_size(0) /* CTA_PROTOINFO */
2015                + nla_total_size(0) /* CTA_HELP */
2016                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2017                + ctnetlink_secctx_size(ct)
2018 #ifdef CONFIG_NF_NAT_NEEDED
2019                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2020                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2021 #endif
2022 #ifdef CONFIG_NF_CONNTRACK_MARK
2023                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2024 #endif
2025                + ctnetlink_proto_size(ct)
2026                ;
2027 }
2028
2029 static int
2030 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2031 {
2032         struct nlattr *nest_parms;
2033
2034         rcu_read_lock();
2035         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2036         if (!nest_parms)
2037                 goto nla_put_failure;
2038         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2039                 goto nla_put_failure;
2040         nla_nest_end(skb, nest_parms);
2041
2042         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2043         if (!nest_parms)
2044                 goto nla_put_failure;
2045         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2046                 goto nla_put_failure;
2047         nla_nest_end(skb, nest_parms);
2048
2049         if (nf_ct_zone(ct)) {
2050                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2051                         goto nla_put_failure;
2052         }
2053
2054         if (ctnetlink_dump_id(skb, ct) < 0)
2055                 goto nla_put_failure;
2056
2057         if (ctnetlink_dump_status(skb, ct) < 0)
2058                 goto nla_put_failure;
2059
2060         if (ctnetlink_dump_timeout(skb, ct) < 0)
2061                 goto nla_put_failure;
2062
2063         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2064                 goto nla_put_failure;
2065
2066         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2067                 goto nla_put_failure;
2068
2069 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2070         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2071                 goto nla_put_failure;
2072 #endif
2073         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2074                 goto nla_put_failure;
2075
2076         if ((ct->status & IPS_SEQ_ADJUST) &&
2077             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2078                 goto nla_put_failure;
2079
2080 #ifdef CONFIG_NF_CONNTRACK_MARK
2081         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2082                 goto nla_put_failure;
2083 #endif
2084         if (ctnetlink_dump_labels(skb, ct) < 0)
2085                 goto nla_put_failure;
2086         rcu_read_unlock();
2087         return 0;
2088
2089 nla_put_failure:
2090         rcu_read_unlock();
2091         return -ENOSPC;
2092 }
2093
2094 static int
2095 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2096 {
2097         int err;
2098
2099         if (cda[CTA_TIMEOUT]) {
2100                 err = ctnetlink_change_timeout(ct, cda);
2101                 if (err < 0)
2102                         return err;
2103         }
2104         if (cda[CTA_STATUS]) {
2105                 err = ctnetlink_change_status(ct, cda);
2106                 if (err < 0)
2107                         return err;
2108         }
2109         if (cda[CTA_HELP]) {
2110                 err = ctnetlink_change_helper(ct, cda);
2111                 if (err < 0)
2112                         return err;
2113         }
2114         if (cda[CTA_LABELS]) {
2115                 err = ctnetlink_attach_labels(ct, cda);
2116                 if (err < 0)
2117                         return err;
2118         }
2119 #if defined(CONFIG_NF_CONNTRACK_MARK)
2120         if (cda[CTA_MARK])
2121                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2122 #endif
2123         return 0;
2124 }
2125
2126 static int
2127 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2128 {
2129         struct nlattr *cda[CTA_MAX+1];
2130         int ret;
2131
2132         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2133         if (ret < 0)
2134                 return ret;
2135
2136         spin_lock_bh(&nf_conntrack_lock);
2137         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2138         spin_unlock_bh(&nf_conntrack_lock);
2139
2140         return ret;
2141 }
2142
2143 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2144                                        const struct nf_conn *ct,
2145                                        struct nf_conntrack_tuple *tuple,
2146                                        struct nf_conntrack_tuple *mask)
2147 {
2148         int err;
2149
2150         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2151                                     nf_ct_l3num(ct));
2152         if (err < 0)
2153                 return err;
2154
2155         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2156                                      nf_ct_l3num(ct));
2157 }
2158
2159 static int
2160 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2161                                 u32 portid, u32 report)
2162 {
2163         struct nlattr *cda[CTA_EXPECT_MAX+1];
2164         struct nf_conntrack_tuple tuple, mask;
2165         struct nf_conntrack_helper *helper = NULL;
2166         struct nf_conntrack_expect *exp;
2167         int err;
2168
2169         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2170         if (err < 0)
2171                 return err;
2172
2173         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2174                                           ct, &tuple, &mask);
2175         if (err < 0)
2176                 return err;
2177
2178         if (cda[CTA_EXPECT_HELP_NAME]) {
2179                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2180
2181                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2182                                                     nf_ct_protonum(ct));
2183                 if (helper == NULL)
2184                         return -EOPNOTSUPP;
2185         }
2186
2187         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2188                                      helper, &tuple, &mask);
2189         if (IS_ERR(exp))
2190                 return PTR_ERR(exp);
2191
2192         err = nf_ct_expect_related_report(exp, portid, report);
2193         if (err < 0) {
2194                 nf_ct_expect_put(exp);
2195                 return err;
2196         }
2197
2198         return 0;
2199 }
2200
2201 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2202         .build_size     = ctnetlink_nfqueue_build_size,
2203         .build          = ctnetlink_nfqueue_build,
2204         .parse          = ctnetlink_nfqueue_parse,
2205         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2206         .seq_adjust     = nf_ct_tcp_seqadj_set,
2207 };
2208 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2209
2210 /***********************************************************************
2211  * EXPECT
2212  ***********************************************************************/
2213
2214 static inline int
2215 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2216                          const struct nf_conntrack_tuple *tuple,
2217                          enum ctattr_expect type)
2218 {
2219         struct nlattr *nest_parms;
2220
2221         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2222         if (!nest_parms)
2223                 goto nla_put_failure;
2224         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2225                 goto nla_put_failure;
2226         nla_nest_end(skb, nest_parms);
2227
2228         return 0;
2229
2230 nla_put_failure:
2231         return -1;
2232 }
2233
2234 static inline int
2235 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2236                         const struct nf_conntrack_tuple *tuple,
2237                         const struct nf_conntrack_tuple_mask *mask)
2238 {
2239         int ret;
2240         struct nf_conntrack_l3proto *l3proto;
2241         struct nf_conntrack_l4proto *l4proto;
2242         struct nf_conntrack_tuple m;
2243         struct nlattr *nest_parms;
2244
2245         memset(&m, 0xFF, sizeof(m));
2246         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2247         m.src.u.all = mask->src.u.all;
2248         m.dst.protonum = tuple->dst.protonum;
2249
2250         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2251         if (!nest_parms)
2252                 goto nla_put_failure;
2253
2254         rcu_read_lock();
2255         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2256         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2257         if (ret >= 0) {
2258                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2259                                                tuple->dst.protonum);
2260         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2261         }
2262         rcu_read_unlock();
2263
2264         if (unlikely(ret < 0))
2265                 goto nla_put_failure;
2266
2267         nla_nest_end(skb, nest_parms);
2268
2269         return 0;
2270
2271 nla_put_failure:
2272         return -1;
2273 }
2274
2275 static const union nf_inet_addr any_addr;
2276
2277 static int
2278 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2279                           const struct nf_conntrack_expect *exp)
2280 {
2281         struct nf_conn *master = exp->master;
2282         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2283         struct nf_conn_help *help;
2284 #ifdef CONFIG_NF_NAT_NEEDED
2285         struct nlattr *nest_parms;
2286         struct nf_conntrack_tuple nat_tuple = {};
2287 #endif
2288         struct nf_ct_helper_expectfn *expfn;
2289
2290         if (timeout < 0)
2291                 timeout = 0;
2292
2293         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2294                 goto nla_put_failure;
2295         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2296                 goto nla_put_failure;
2297         if (ctnetlink_exp_dump_tuple(skb,
2298                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2299                                  CTA_EXPECT_MASTER) < 0)
2300                 goto nla_put_failure;
2301
2302 #ifdef CONFIG_NF_NAT_NEEDED
2303         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2304             exp->saved_proto.all) {
2305                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2306                 if (!nest_parms)
2307                         goto nla_put_failure;
2308
2309                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2310                         goto nla_put_failure;
2311
2312                 nat_tuple.src.l3num = nf_ct_l3num(master);
2313                 nat_tuple.src.u3 = exp->saved_addr;
2314                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2315                 nat_tuple.src.u = exp->saved_proto;
2316
2317                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2318                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2319                         goto nla_put_failure;
2320                 nla_nest_end(skb, nest_parms);
2321         }
2322 #endif
2323         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2324             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2325             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2326             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2327                 goto nla_put_failure;
2328         help = nfct_help(master);
2329         if (help) {
2330                 struct nf_conntrack_helper *helper;
2331
2332                 helper = rcu_dereference(help->helper);
2333                 if (helper &&
2334                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2335                         goto nla_put_failure;
2336         }
2337         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2338         if (expfn != NULL &&
2339             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2340                 goto nla_put_failure;
2341
2342         return 0;
2343
2344 nla_put_failure:
2345         return -1;
2346 }
2347
2348 static int
2349 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2350                         int event, const struct nf_conntrack_expect *exp)
2351 {
2352         struct nlmsghdr *nlh;
2353         struct nfgenmsg *nfmsg;
2354         unsigned int flags = portid ? NLM_F_MULTI : 0;
2355
2356         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2357         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2358         if (nlh == NULL)
2359                 goto nlmsg_failure;
2360
2361         nfmsg = nlmsg_data(nlh);
2362         nfmsg->nfgen_family = exp->tuple.src.l3num;
2363         nfmsg->version      = NFNETLINK_V0;
2364         nfmsg->res_id       = 0;
2365
2366         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2367                 goto nla_put_failure;
2368
2369         nlmsg_end(skb, nlh);
2370         return skb->len;
2371
2372 nlmsg_failure:
2373 nla_put_failure:
2374         nlmsg_cancel(skb, nlh);
2375         return -1;
2376 }
2377
2378 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2379 static int
2380 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2381 {
2382         struct nf_conntrack_expect *exp = item->exp;
2383         struct net *net = nf_ct_exp_net(exp);
2384         struct nlmsghdr *nlh;
2385         struct nfgenmsg *nfmsg;
2386         struct sk_buff *skb;
2387         unsigned int type, group;
2388         int flags = 0;
2389
2390         if (events & (1 << IPEXP_DESTROY)) {
2391                 type = IPCTNL_MSG_EXP_DELETE;
2392                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2393         } else if (events & (1 << IPEXP_NEW)) {
2394                 type = IPCTNL_MSG_EXP_NEW;
2395                 flags = NLM_F_CREATE|NLM_F_EXCL;
2396                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2397         } else
2398                 return 0;
2399
2400         if (!item->report && !nfnetlink_has_listeners(net, group))
2401                 return 0;
2402
2403         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2404         if (skb == NULL)
2405                 goto errout;
2406
2407         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2408         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2409         if (nlh == NULL)
2410                 goto nlmsg_failure;
2411
2412         nfmsg = nlmsg_data(nlh);
2413         nfmsg->nfgen_family = exp->tuple.src.l3num;
2414         nfmsg->version      = NFNETLINK_V0;
2415         nfmsg->res_id       = 0;
2416
2417         rcu_read_lock();
2418         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2419                 goto nla_put_failure;
2420         rcu_read_unlock();
2421
2422         nlmsg_end(skb, nlh);
2423         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2424         return 0;
2425
2426 nla_put_failure:
2427         rcu_read_unlock();
2428         nlmsg_cancel(skb, nlh);
2429 nlmsg_failure:
2430         kfree_skb(skb);
2431 errout:
2432         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2433         return 0;
2434 }
2435 #endif
2436 static int ctnetlink_exp_done(struct netlink_callback *cb)
2437 {
2438         if (cb->args[1])
2439                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2440         return 0;
2441 }
2442
2443 static int
2444 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2445 {
2446         struct net *net = sock_net(skb->sk);
2447         struct nf_conntrack_expect *exp, *last;
2448         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2449         u_int8_t l3proto = nfmsg->nfgen_family;
2450
2451         rcu_read_lock();
2452         last = (struct nf_conntrack_expect *)cb->args[1];
2453         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2454 restart:
2455                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2456                                      hnode) {
2457                         if (l3proto && exp->tuple.src.l3num != l3proto)
2458                                 continue;
2459                         if (cb->args[1]) {
2460                                 if (exp != last)
2461                                         continue;
2462                                 cb->args[1] = 0;
2463                         }
2464                         if (ctnetlink_exp_fill_info(skb,
2465                                                     NETLINK_CB(cb->skb).portid,
2466                                                     cb->nlh->nlmsg_seq,
2467                                                     IPCTNL_MSG_EXP_NEW,
2468                                                     exp) < 0) {
2469                                 if (!atomic_inc_not_zero(&exp->use))
2470                                         continue;
2471                                 cb->args[1] = (unsigned long)exp;
2472                                 goto out;
2473                         }
2474                 }
2475                 if (cb->args[1]) {
2476                         cb->args[1] = 0;
2477                         goto restart;
2478                 }
2479         }
2480 out:
2481         rcu_read_unlock();
2482         if (last)
2483                 nf_ct_expect_put(last);
2484
2485         return skb->len;
2486 }
2487
2488 static int
2489 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2490 {
2491         struct nf_conntrack_expect *exp, *last;
2492         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2493         struct nf_conn *ct = cb->data;
2494         struct nf_conn_help *help = nfct_help(ct);
2495         u_int8_t l3proto = nfmsg->nfgen_family;
2496
2497         if (cb->args[0])
2498                 return 0;
2499
2500         rcu_read_lock();
2501         last = (struct nf_conntrack_expect *)cb->args[1];
2502 restart:
2503         hlist_for_each_entry(exp, &help->expectations, lnode) {
2504                 if (l3proto && exp->tuple.src.l3num != l3proto)
2505                         continue;
2506                 if (cb->args[1]) {
2507                         if (exp != last)
2508                                 continue;
2509                         cb->args[1] = 0;
2510                 }
2511                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2512                                             cb->nlh->nlmsg_seq,
2513                                             IPCTNL_MSG_EXP_NEW,
2514                                             exp) < 0) {
2515                         if (!atomic_inc_not_zero(&exp->use))
2516                                 continue;
2517                         cb->args[1] = (unsigned long)exp;
2518                         goto out;
2519                 }
2520         }
2521         if (cb->args[1]) {
2522                 cb->args[1] = 0;
2523                 goto restart;
2524         }
2525         cb->args[0] = 1;
2526 out:
2527         rcu_read_unlock();
2528         if (last)
2529                 nf_ct_expect_put(last);
2530
2531         return skb->len;
2532 }
2533
2534 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2535                                  const struct nlmsghdr *nlh,
2536                                  const struct nlattr * const cda[])
2537 {
2538         int err;
2539         struct net *net = sock_net(ctnl);
2540         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2541         u_int8_t u3 = nfmsg->nfgen_family;
2542         struct nf_conntrack_tuple tuple;
2543         struct nf_conntrack_tuple_hash *h;
2544         struct nf_conn *ct;
2545         u16 zone = 0;
2546         struct netlink_dump_control c = {
2547                 .dump = ctnetlink_exp_ct_dump_table,
2548                 .done = ctnetlink_exp_done,
2549         };
2550
2551         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2552         if (err < 0)
2553                 return err;
2554
2555         if (cda[CTA_EXPECT_ZONE]) {
2556                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2557                 if (err < 0)
2558                         return err;
2559         }
2560
2561         h = nf_conntrack_find_get(net, zone, &tuple);
2562         if (!h)
2563                 return -ENOENT;
2564
2565         ct = nf_ct_tuplehash_to_ctrack(h);
2566         c.data = ct;
2567
2568         err = netlink_dump_start(ctnl, skb, nlh, &c);
2569         nf_ct_put(ct);
2570
2571         return err;
2572 }
2573
2574 static int
2575 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2576                      const struct nlmsghdr *nlh,
2577                      const struct nlattr * const cda[])
2578 {
2579         struct net *net = sock_net(ctnl);
2580         struct nf_conntrack_tuple tuple;
2581         struct nf_conntrack_expect *exp;
2582         struct sk_buff *skb2;
2583         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2584         u_int8_t u3 = nfmsg->nfgen_family;
2585         u16 zone;
2586         int err;
2587
2588         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2589                 if (cda[CTA_EXPECT_MASTER])
2590                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2591                 else {
2592                         struct netlink_dump_control c = {
2593                                 .dump = ctnetlink_exp_dump_table,
2594                                 .done = ctnetlink_exp_done,
2595                         };
2596                         return netlink_dump_start(ctnl, skb, nlh, &c);
2597                 }
2598         }
2599
2600         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2601         if (err < 0)
2602                 return err;
2603
2604         if (cda[CTA_EXPECT_TUPLE])
2605                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2606         else if (cda[CTA_EXPECT_MASTER])
2607                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2608         else
2609                 return -EINVAL;
2610
2611         if (err < 0)
2612                 return err;
2613
2614         exp = nf_ct_expect_find_get(net, zone, &tuple);
2615         if (!exp)
2616                 return -ENOENT;
2617
2618         if (cda[CTA_EXPECT_ID]) {
2619                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2620                 if (ntohl(id) != (u32)(unsigned long)exp) {
2621                         nf_ct_expect_put(exp);
2622                         return -ENOENT;
2623                 }
2624         }
2625
2626         err = -ENOMEM;
2627         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2628         if (skb2 == NULL) {
2629                 nf_ct_expect_put(exp);
2630                 goto out;
2631         }
2632
2633         rcu_read_lock();
2634         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2635                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2636         rcu_read_unlock();
2637         nf_ct_expect_put(exp);
2638         if (err <= 0)
2639                 goto free;
2640
2641         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2642         if (err < 0)
2643                 goto out;
2644
2645         return 0;
2646
2647 free:
2648         kfree_skb(skb2);
2649 out:
2650         /* this avoids a loop in nfnetlink. */
2651         return err == -EAGAIN ? -ENOBUFS : err;
2652 }
2653
2654 static int
2655 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2656                      const struct nlmsghdr *nlh,
2657                      const struct nlattr * const cda[])
2658 {
2659         struct net *net = sock_net(ctnl);
2660         struct nf_conntrack_expect *exp;
2661         struct nf_conntrack_tuple tuple;
2662         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2663         struct hlist_node *next;
2664         u_int8_t u3 = nfmsg->nfgen_family;
2665         unsigned int i;
2666         u16 zone;
2667         int err;
2668
2669         if (cda[CTA_EXPECT_TUPLE]) {
2670                 /* delete a single expect by tuple */
2671                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2672                 if (err < 0)
2673                         return err;
2674
2675                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2676                 if (err < 0)
2677                         return err;
2678
2679                 /* bump usage count to 2 */
2680                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2681                 if (!exp)
2682                         return -ENOENT;
2683
2684                 if (cda[CTA_EXPECT_ID]) {
2685                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2686                         if (ntohl(id) != (u32)(unsigned long)exp) {
2687                                 nf_ct_expect_put(exp);
2688                                 return -ENOENT;
2689                         }
2690                 }
2691
2692                 /* after list removal, usage count == 1 */
2693                 spin_lock_bh(&nf_conntrack_lock);
2694                 if (del_timer(&exp->timeout)) {
2695                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2696                                                    nlmsg_report(nlh));
2697                         nf_ct_expect_put(exp);
2698                 }
2699                 spin_unlock_bh(&nf_conntrack_lock);
2700                 /* have to put what we 'get' above.
2701                  * after this line usage count == 0 */
2702                 nf_ct_expect_put(exp);
2703         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2704                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2705                 struct nf_conn_help *m_help;
2706
2707                 /* delete all expectations for this helper */
2708                 spin_lock_bh(&nf_conntrack_lock);
2709                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2710                         hlist_for_each_entry_safe(exp, next,
2711                                                   &net->ct.expect_hash[i],
2712                                                   hnode) {
2713                                 m_help = nfct_help(exp->master);
2714                                 if (!strcmp(m_help->helper->name, name) &&
2715                                     del_timer(&exp->timeout)) {
2716                                         nf_ct_unlink_expect_report(exp,
2717                                                         NETLINK_CB(skb).portid,
2718                                                         nlmsg_report(nlh));
2719                                         nf_ct_expect_put(exp);
2720                                 }
2721                         }
2722                 }
2723                 spin_unlock_bh(&nf_conntrack_lock);
2724         } else {
2725                 /* This basically means we have to flush everything*/
2726                 spin_lock_bh(&nf_conntrack_lock);
2727                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2728                         hlist_for_each_entry_safe(exp, next,
2729                                                   &net->ct.expect_hash[i],
2730                                                   hnode) {
2731                                 if (del_timer(&exp->timeout)) {
2732                                         nf_ct_unlink_expect_report(exp,
2733                                                         NETLINK_CB(skb).portid,
2734                                                         nlmsg_report(nlh));
2735                                         nf_ct_expect_put(exp);
2736                                 }
2737                         }
2738                 }
2739                 spin_unlock_bh(&nf_conntrack_lock);
2740         }
2741
2742         return 0;
2743 }
2744 static int
2745 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2746                         const struct nlattr * const cda[])
2747 {
2748         if (cda[CTA_EXPECT_TIMEOUT]) {
2749                 if (!del_timer(&x->timeout))
2750                         return -ETIME;
2751
2752                 x->timeout.expires = jiffies +
2753                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2754                 add_timer(&x->timeout);
2755         }
2756         return 0;
2757 }
2758
2759 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2760         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2761         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2762 };
2763
2764 static int
2765 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2766                            struct nf_conntrack_expect *exp,
2767                            u_int8_t u3)
2768 {
2769 #ifdef CONFIG_NF_NAT_NEEDED
2770         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2771         struct nf_conntrack_tuple nat_tuple = {};
2772         int err;
2773
2774         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2775         if (err < 0)
2776                 return err;
2777
2778         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2779                 return -EINVAL;
2780
2781         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2782                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2783         if (err < 0)
2784                 return err;
2785
2786         exp->saved_addr = nat_tuple.src.u3;
2787         exp->saved_proto = nat_tuple.src.u;
2788         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2789
2790         return 0;
2791 #else
2792         return -EOPNOTSUPP;
2793 #endif
2794 }
2795
2796 static struct nf_conntrack_expect *
2797 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2798                        struct nf_conntrack_helper *helper,
2799                        struct nf_conntrack_tuple *tuple,
2800                        struct nf_conntrack_tuple *mask)
2801 {
2802         u_int32_t class = 0;
2803         struct nf_conntrack_expect *exp;
2804         struct nf_conn_help *help;
2805         int err;
2806
2807         if (cda[CTA_EXPECT_CLASS] && helper) {
2808                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2809                 if (class > helper->expect_class_max)
2810                         return ERR_PTR(-EINVAL);
2811         }
2812         exp = nf_ct_expect_alloc(ct);
2813         if (!exp)
2814                 return ERR_PTR(-ENOMEM);
2815
2816         help = nfct_help(ct);
2817         if (!help) {
2818                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2819                         err = -EINVAL;
2820                         goto err_out;
2821                 }
2822                 exp->timeout.expires =
2823                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2824
2825                 exp->flags = NF_CT_EXPECT_USERSPACE;
2826                 if (cda[CTA_EXPECT_FLAGS]) {
2827                         exp->flags |=
2828                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2829                 }
2830         } else {
2831                 if (cda[CTA_EXPECT_FLAGS]) {
2832                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2833                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2834                 } else
2835                         exp->flags = 0;
2836         }
2837         if (cda[CTA_EXPECT_FN]) {
2838                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2839                 struct nf_ct_helper_expectfn *expfn;
2840
2841                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2842                 if (expfn == NULL) {
2843                         err = -EINVAL;
2844                         goto err_out;
2845                 }
2846                 exp->expectfn = expfn->expectfn;
2847         } else
2848                 exp->expectfn = NULL;
2849
2850         exp->class = class;
2851         exp->master = ct;
2852         exp->helper = helper;
2853         exp->tuple = *tuple;
2854         exp->mask.src.u3 = mask->src.u3;
2855         exp->mask.src.u.all = mask->src.u.all;
2856
2857         if (cda[CTA_EXPECT_NAT]) {
2858                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2859                                                  exp, nf_ct_l3num(ct));
2860                 if (err < 0)
2861                         goto err_out;
2862         }
2863         return exp;
2864 err_out:
2865         nf_ct_expect_put(exp);
2866         return ERR_PTR(err);
2867 }
2868
2869 static int
2870 ctnetlink_create_expect(struct net *net, u16 zone,
2871                         const struct nlattr * const cda[],
2872                         u_int8_t u3, u32 portid, int report)
2873 {
2874         struct nf_conntrack_tuple tuple, mask, master_tuple;
2875         struct nf_conntrack_tuple_hash *h = NULL;
2876         struct nf_conntrack_helper *helper = NULL;
2877         struct nf_conntrack_expect *exp;
2878         struct nf_conn *ct;
2879         int err;
2880
2881         /* caller guarantees that those three CTA_EXPECT_* exist */
2882         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2883         if (err < 0)
2884                 return err;
2885         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2886         if (err < 0)
2887                 return err;
2888         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2889         if (err < 0)
2890                 return err;
2891
2892         /* Look for master conntrack of this expectation */
2893         h = nf_conntrack_find_get(net, zone, &master_tuple);
2894         if (!h)
2895                 return -ENOENT;
2896         ct = nf_ct_tuplehash_to_ctrack(h);
2897
2898         if (cda[CTA_EXPECT_HELP_NAME]) {
2899                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2900
2901                 helper = __nf_conntrack_helper_find(helpname, u3,
2902                                                     nf_ct_protonum(ct));
2903                 if (helper == NULL) {
2904 #ifdef CONFIG_MODULES
2905                         if (request_module("nfct-helper-%s", helpname) < 0) {
2906                                 err = -EOPNOTSUPP;
2907                                 goto err_ct;
2908                         }
2909                         helper = __nf_conntrack_helper_find(helpname, u3,
2910                                                             nf_ct_protonum(ct));
2911                         if (helper) {
2912                                 err = -EAGAIN;
2913                                 goto err_ct;
2914                         }
2915 #endif
2916                         err = -EOPNOTSUPP;
2917                         goto err_ct;
2918                 }
2919         }
2920
2921         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
2922         if (IS_ERR(exp)) {
2923                 err = PTR_ERR(exp);
2924                 goto err_ct;
2925         }
2926
2927         err = nf_ct_expect_related_report(exp, portid, report);
2928         if (err < 0)
2929                 goto err_exp;
2930
2931         return 0;
2932 err_exp:
2933         nf_ct_expect_put(exp);
2934 err_ct:
2935         nf_ct_put(ct);
2936         return err;
2937 }
2938
2939 static int
2940 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2941                      const struct nlmsghdr *nlh,
2942                      const struct nlattr * const cda[])
2943 {
2944         struct net *net = sock_net(ctnl);
2945         struct nf_conntrack_tuple tuple;
2946         struct nf_conntrack_expect *exp;
2947         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2948         u_int8_t u3 = nfmsg->nfgen_family;
2949         u16 zone;
2950         int err;
2951
2952         if (!cda[CTA_EXPECT_TUPLE]
2953             || !cda[CTA_EXPECT_MASK]
2954             || !cda[CTA_EXPECT_MASTER])
2955                 return -EINVAL;
2956
2957         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2958         if (err < 0)
2959                 return err;
2960
2961         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2962         if (err < 0)
2963                 return err;
2964
2965         spin_lock_bh(&nf_conntrack_lock);
2966         exp = __nf_ct_expect_find(net, zone, &tuple);
2967
2968         if (!exp) {
2969                 spin_unlock_bh(&nf_conntrack_lock);
2970                 err = -ENOENT;
2971                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2972                         err = ctnetlink_create_expect(net, zone, cda,
2973                                                       u3,
2974                                                       NETLINK_CB(skb).portid,
2975                                                       nlmsg_report(nlh));
2976                 }
2977                 return err;
2978         }
2979
2980         err = -EEXIST;
2981         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2982                 err = ctnetlink_change_expect(exp, cda);
2983         spin_unlock_bh(&nf_conntrack_lock);
2984
2985         return err;
2986 }
2987
2988 static int
2989 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
2990                              const struct ip_conntrack_stat *st)
2991 {
2992         struct nlmsghdr *nlh;
2993         struct nfgenmsg *nfmsg;
2994         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2995
2996         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
2997         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2998         if (nlh == NULL)
2999                 goto nlmsg_failure;
3000
3001         nfmsg = nlmsg_data(nlh);
3002         nfmsg->nfgen_family = AF_UNSPEC;
3003         nfmsg->version      = NFNETLINK_V0;
3004         nfmsg->res_id       = htons(cpu);
3005
3006         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3007             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3008             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3009                 goto nla_put_failure;
3010
3011         nlmsg_end(skb, nlh);
3012         return skb->len;
3013
3014 nla_put_failure:
3015 nlmsg_failure:
3016         nlmsg_cancel(skb, nlh);
3017         return -1;
3018 }
3019
3020 static int
3021 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3022 {
3023         int cpu;
3024         struct net *net = sock_net(skb->sk);
3025
3026         if (cb->args[0] == nr_cpu_ids)
3027                 return 0;
3028
3029         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3030                 const struct ip_conntrack_stat *st;
3031
3032                 if (!cpu_possible(cpu))
3033                         continue;
3034
3035                 st = per_cpu_ptr(net->ct.stat, cpu);
3036                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3037                                                  cb->nlh->nlmsg_seq,
3038                                                  cpu, st) < 0)
3039                         break;
3040         }
3041         cb->args[0] = cpu;
3042
3043         return skb->len;
3044 }
3045
3046 static int
3047 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3048                        const struct nlmsghdr *nlh,
3049                        const struct nlattr * const cda[])
3050 {
3051         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3052                 struct netlink_dump_control c = {
3053                         .dump = ctnetlink_exp_stat_cpu_dump,
3054                 };
3055                 return netlink_dump_start(ctnl, skb, nlh, &c);
3056         }
3057
3058         return 0;
3059 }
3060
3061 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3062 static struct nf_ct_event_notifier ctnl_notifier = {
3063         .fcn = ctnetlink_conntrack_event,
3064 };
3065
3066 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3067         .fcn = ctnetlink_expect_event,
3068 };
3069 #endif
3070
3071 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3072         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3073                                             .attr_count = CTA_MAX,
3074                                             .policy = ct_nla_policy },
3075         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3076                                             .attr_count = CTA_MAX,
3077                                             .policy = ct_nla_policy },
3078         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3079                                             .attr_count = CTA_MAX,
3080                                             .policy = ct_nla_policy },
3081         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3082                                             .attr_count = CTA_MAX,
3083                                             .policy = ct_nla_policy },
3084         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3085         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3086         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3087         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3088 };
3089
3090 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3091         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3092                                             .attr_count = CTA_EXPECT_MAX,
3093                                             .policy = exp_nla_policy },
3094         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3095                                             .attr_count = CTA_EXPECT_MAX,
3096                                             .policy = exp_nla_policy },
3097         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3098                                             .attr_count = CTA_EXPECT_MAX,
3099                                             .policy = exp_nla_policy },
3100         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3101 };
3102
3103 static const struct nfnetlink_subsystem ctnl_subsys = {
3104         .name                           = "conntrack",
3105         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3106         .cb_count                       = IPCTNL_MSG_MAX,
3107         .cb                             = ctnl_cb,
3108 };
3109
3110 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3111         .name                           = "conntrack_expect",
3112         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3113         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3114         .cb                             = ctnl_exp_cb,
3115 };
3116
3117 MODULE_ALIAS("ip_conntrack_netlink");
3118 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3119 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3120
3121 static int __net_init ctnetlink_net_init(struct net *net)
3122 {
3123 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3124         int ret;
3125
3126         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3127         if (ret < 0) {
3128                 pr_err("ctnetlink_init: cannot register notifier.\n");
3129                 goto err_out;
3130         }
3131
3132         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3133         if (ret < 0) {
3134                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3135                 goto err_unreg_notifier;
3136         }
3137 #endif
3138         return 0;
3139
3140 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3141 err_unreg_notifier:
3142         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3143 err_out:
3144         return ret;
3145 #endif
3146 }
3147
3148 static void ctnetlink_net_exit(struct net *net)
3149 {
3150 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3151         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3152         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3153 #endif
3154 }
3155
3156 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3157 {
3158         struct net *net;
3159
3160         list_for_each_entry(net, net_exit_list, exit_list)
3161                 ctnetlink_net_exit(net);
3162 }
3163
3164 static struct pernet_operations ctnetlink_net_ops = {
3165         .init           = ctnetlink_net_init,
3166         .exit_batch     = ctnetlink_net_exit_batch,
3167 };
3168
3169 static int __init ctnetlink_init(void)
3170 {
3171         int ret;
3172
3173         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3174         ret = nfnetlink_subsys_register(&ctnl_subsys);
3175         if (ret < 0) {
3176                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3177                 goto err_out;
3178         }
3179
3180         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3181         if (ret < 0) {
3182                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3183                 goto err_unreg_subsys;
3184         }
3185
3186         ret = register_pernet_subsys(&ctnetlink_net_ops);
3187         if (ret < 0) {
3188                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3189                 goto err_unreg_exp_subsys;
3190         }
3191 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3192         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3193         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3194 #endif
3195         return 0;
3196
3197 err_unreg_exp_subsys:
3198         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3199 err_unreg_subsys:
3200         nfnetlink_subsys_unregister(&ctnl_subsys);
3201 err_out:
3202         return ret;
3203 }
3204
3205 static void __exit ctnetlink_exit(void)
3206 {
3207         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3208
3209         unregister_pernet_subsys(&ctnetlink_net_ops);
3210         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3211         nfnetlink_subsys_unregister(&ctnl_subsys);
3212 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3213         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3214 #endif
3215 }
3216
3217 module_init(ctnetlink_init);
3218 module_exit(ctnetlink_exit);