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, struct nf_conn_acct *acct,
 215               enum ip_conntrack_dir dir, int type)
 216 {
 217         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 218         struct nf_conn_counter *counter = acct->counter;
 219         struct nlattr *nest_count;
 220         u64 pkts, bytes;
 221
 222         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 223                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 224                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 225         } else {
 226                 pkts = atomic64_read(&counter[dir].packets);
 227                 bytes = atomic64_read(&counter[dir].bytes);
 228         }
 229
 230         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 231         if (!nest_count)
 232                 goto nla_put_failure;
 233
 234         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 235             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 236                 goto nla_put_failure;
 237
 238         nla_nest_end(skb, nest_count);
 239
 240         return 0;
 241
 242 nla_put_failure:
 243         return -1;
 244 }
 245
 246 static int
 247 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 248 {
 249         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 250
 251         if (!acct)
 252                 return 0;
 253
 254         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 255                 return -1;
 256         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 257                 return -1;
 258
 259         return 0;
 260 }
 261
 262 static int
 263 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 264 {
 265         struct nlattr *nest_count;
 266         const struct nf_conn_tstamp *tstamp;
 267
 268         tstamp = nf_conn_tstamp_find(ct);
 269         if (!tstamp)
 270                 return 0;
 271
 272         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 273         if (!nest_count)
 274                 goto nla_put_failure;
 275
 276         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 277             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 278                                                cpu_to_be64(tstamp->stop))))
 279                 goto nla_put_failure;
 280         nla_nest_end(skb, nest_count);
 281
 282         return 0;
 283
 284 nla_put_failure:
 285         return -1;
 286 }
 287
 288 #ifdef CONFIG_NF_CONNTRACK_MARK
 289 static inline int
 290 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 291 {
 292         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 293                 goto nla_put_failure;
 294         return 0;
 295
 296 nla_put_failure:
 297         return -1;
 298 }
 299 #else
 300 #define ctnetlink_dump_mark(a, b) (0)
 301 #endif
 302
 303 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 304 static inline int
 305 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 306 {
 307         struct nlattr *nest_secctx;
 308         int len, ret;
 309         char *secctx;
 310
 311         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 312         if (ret)
 313                 return 0;
 314
 315         ret = -1;
 316         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 317         if (!nest_secctx)
 318                 goto nla_put_failure;
 319
 320         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 321                 goto nla_put_failure;
 322         nla_nest_end(skb, nest_secctx);
 323
 324         ret = 0;
 325 nla_put_failure:
 326         security_release_secctx(secctx, len);
 327         return ret;
 328 }
 329 #else
 330 #define ctnetlink_dump_secctx(a, b) (0)
 331 #endif
 332
 333 #ifdef CONFIG_NF_CONNTRACK_LABELS
 334 static int ctnetlink_label_size(const struct nf_conn *ct)
 335 {
 336         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 337
 338         if (!labels)
 339                 return 0;
 340         return nla_total_size(labels->words * sizeof(long));
 341 }
 342
 343 static int
 344 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 345 {
 346         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 347         unsigned int len, i;
 348
 349         if (!labels)
 350                 return 0;
 351
 352         len = labels->words * sizeof(long);
 353         i = 0;
 354         do {
 355                 if (labels->bits[i] != 0)
 356                         return nla_put(skb, CTA_LABELS, len, labels->bits);
 357                 i++;
 358         } while (i < labels->words);
 359
 360         return 0;
 361 }
 362 #else
 363 #define ctnetlink_dump_labels(a, b) (0)
 364 #define ctnetlink_label_size(a) (0)
 365 #endif
 366
 367 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 368
 369 static inline int
 370 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 371 {
 372         struct nlattr *nest_parms;
 373
 374         if (!(ct->status & IPS_EXPECTED))
 375                 return 0;
 376
 377         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 378         if (!nest_parms)
 379                 goto nla_put_failure;
 380         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 381                 goto nla_put_failure;
 382         nla_nest_end(skb, nest_parms);
 383
 384         return 0;
 385
 386 nla_put_failure:
 387         return -1;
 388 }
 389
 390 static int
 391 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 392 {
 393         struct nlattr *nest_parms;
 394
 395         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 396         if (!nest_parms)
 397                 goto nla_put_failure;
 398
 399         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 400                          htonl(seq->correction_pos)) ||
 401             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 402                          htonl(seq->offset_before)) ||
 403             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 404                          htonl(seq->offset_after)))
 405                 goto nla_put_failure;
 406
 407         nla_nest_end(skb, nest_parms);
 408
 409         return 0;
 410
 411 nla_put_failure:
 412         return -1;
 413 }
 414
 415 static inline int
 416 ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 417 {
 418         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 419         struct nf_ct_seqadj *seq;
 420
 421         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 422                 return 0;
 423
 424         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 425         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 426                 return -1;
 427
 428         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 429         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 430                 return -1;
 431
 432         return 0;
 433 }
 434
 435 static inline int
 436 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 437 {
 438         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 439                 goto nla_put_failure;
 440         return 0;
 441
 442 nla_put_failure:
 443         return -1;
 444 }
 445
 446 static inline int
 447 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 448 {
 449         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 450                 goto nla_put_failure;
 451         return 0;
 452
 453 nla_put_failure:
 454         return -1;
 455 }
 456
 457 static int
 458 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 459                     struct nf_conn *ct)
 460 {
 461         struct nlmsghdr *nlh;
 462         struct nfgenmsg *nfmsg;
 463         struct nlattr *nest_parms;
 464         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 465
 466         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 467         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 468         if (nlh == NULL)
 469                 goto nlmsg_failure;
 470
 471         nfmsg = nlmsg_data(nlh);
 472         nfmsg->nfgen_family = nf_ct_l3num(ct);
 473         nfmsg->version      = NFNETLINK_V0;
 474         nfmsg->res_id       = 0;
 475
 476         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 477         if (!nest_parms)
 478                 goto nla_put_failure;
 479         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 480                 goto nla_put_failure;
 481         nla_nest_end(skb, nest_parms);
 482
 483         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 484         if (!nest_parms)
 485                 goto nla_put_failure;
 486         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 487                 goto nla_put_failure;
 488         nla_nest_end(skb, nest_parms);
 489
 490         if (nf_ct_zone(ct) &&
 491             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 492                 goto nla_put_failure;
 493
 494         if (ctnetlink_dump_status(skb, ct) < 0 ||
 495             ctnetlink_dump_timeout(skb, ct) < 0 ||
 496             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 497             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 498             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 499             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 500             ctnetlink_dump_mark(skb, ct) < 0 ||
 501             ctnetlink_dump_secctx(skb, ct) < 0 ||
 502             ctnetlink_dump_labels(skb, ct) < 0 ||
 503             ctnetlink_dump_id(skb, ct) < 0 ||
 504             ctnetlink_dump_use(skb, ct) < 0 ||
 505             ctnetlink_dump_master(skb, ct) < 0 ||
 506             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 507                 goto nla_put_failure;
 508
 509         nlmsg_end(skb, nlh);
 510         return skb->len;
 511
 512 nlmsg_failure:
 513 nla_put_failure:
 514         nlmsg_cancel(skb, nlh);
 515         return -1;
 516 }
 517
 518 static inline size_t
 519 ctnetlink_proto_size(const struct nf_conn *ct)
 520 {
 521         struct nf_conntrack_l3proto *l3proto;
 522         struct nf_conntrack_l4proto *l4proto;
 523         size_t len = 0;
 524
 525         rcu_read_lock();
 526         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 527         len += l3proto->nla_size;
 528
 529         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 530         len += l4proto->nla_size;
 531         rcu_read_unlock();
 532
 533         return len;
 534 }
 535
 536 static inline size_t
 537 ctnetlink_acct_size(const struct nf_conn *ct)
 538 {
 539         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 540                 return 0;
 541         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 542                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 543                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 544                ;
 545 }
 546
 547 static inline int
 548 ctnetlink_secctx_size(const struct nf_conn *ct)
 549 {
 550 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 551         int len, ret;
 552
 553         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 554         if (ret)
 555                 return 0;
 556
 557         return nla_total_size(0) /* CTA_SECCTX */
 558                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 559 #else
 560         return 0;
 561 #endif
 562 }
 563
 564 static inline size_t
 565 ctnetlink_timestamp_size(const struct nf_conn *ct)
 566 {
 567 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 568         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 569                 return 0;
 570         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 571 #else
 572         return 0;
 573 #endif
 574 }
 575
 576 static inline size_t
 577 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 578 {
 579         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 580                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 581                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 582                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 583                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 584                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 585                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 586                + ctnetlink_acct_size(ct)
 587                + ctnetlink_timestamp_size(ct)
 588                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 589                + nla_total_size(0) /* CTA_PROTOINFO */
 590                + nla_total_size(0) /* CTA_HELP */
 591                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 592                + ctnetlink_secctx_size(ct)
 593 #ifdef CONFIG_NF_NAT_NEEDED
 594                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 595                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 596 #endif
 597 #ifdef CONFIG_NF_CONNTRACK_MARK
 598                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 599 #endif
 600                + ctnetlink_proto_size(ct)
 601                + ctnetlink_label_size(ct)
 602                ;
 603 }
 604
 605 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 606 static int
 607 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 608 {
 609         struct net *net;
 610         struct nlmsghdr *nlh;
 611         struct nfgenmsg *nfmsg;
 612         struct nlattr *nest_parms;
 613         struct nf_conn *ct = item->ct;
 614         struct sk_buff *skb;
 615         unsigned int type;
 616         unsigned int flags = 0, group;
 617         int err;
 618
 619         /* ignore our fake conntrack entry */
 620         if (nf_ct_is_untracked(ct))
 621                 return 0;
 622
 623         if (events & (1 << IPCT_DESTROY)) {
 624                 type = IPCTNL_MSG_CT_DELETE;
 625                 group = NFNLGRP_CONNTRACK_DESTROY;
 626         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 627                 type = IPCTNL_MSG_CT_NEW;
 628                 flags = NLM_F_CREATE|NLM_F_EXCL;
 629                 group = NFNLGRP_CONNTRACK_NEW;
 630         } else  if (events) {
 631                 type = IPCTNL_MSG_CT_NEW;
 632                 group = NFNLGRP_CONNTRACK_UPDATE;
 633         } else
 634                 return 0;
 635
 636         net = nf_ct_net(ct);
 637         if (!item->report && !nfnetlink_has_listeners(net, group))
 638                 return 0;
 639
 640         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 641         if (skb == NULL)
 642                 goto errout;
 643
 644         type |= NFNL_SUBSYS_CTNETLINK << 8;
 645         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 646         if (nlh == NULL)
 647                 goto nlmsg_failure;
 648
 649         nfmsg = nlmsg_data(nlh);
 650         nfmsg->nfgen_family = nf_ct_l3num(ct);
 651         nfmsg->version  = NFNETLINK_V0;
 652         nfmsg->res_id   = 0;
 653
 654         rcu_read_lock();
 655         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 656         if (!nest_parms)
 657                 goto nla_put_failure;
 658         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 659                 goto nla_put_failure;
 660         nla_nest_end(skb, nest_parms);
 661
 662         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 663         if (!nest_parms)
 664                 goto nla_put_failure;
 665         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 666                 goto nla_put_failure;
 667         nla_nest_end(skb, nest_parms);
 668
 669         if (nf_ct_zone(ct) &&
 670             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 671                 goto nla_put_failure;
 672
 673         if (ctnetlink_dump_id(skb, ct) < 0)
 674                 goto nla_put_failure;
 675
 676         if (ctnetlink_dump_status(skb, ct) < 0)
 677                 goto nla_put_failure;
 678
 679         if (events & (1 << IPCT_DESTROY)) {
 680                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 681                     ctnetlink_dump_timestamp(skb, ct) < 0)
 682                         goto nla_put_failure;
 683         } else {
 684                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 685                         goto nla_put_failure;
 686
 687                 if (events & (1 << IPCT_PROTOINFO)
 688                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 689                         goto nla_put_failure;
 690
 691                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 692                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 693                         goto nla_put_failure;
 694
 695 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 696                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 697                     && ctnetlink_dump_secctx(skb, ct) < 0)
 698                         goto nla_put_failure;
 699 #endif
 700                 if (events & (1 << IPCT_LABEL) &&
 701                      ctnetlink_dump_labels(skb, ct) < 0)
 702                         goto nla_put_failure;
 703
 704                 if (events & (1 << IPCT_RELATED) &&
 705                     ctnetlink_dump_master(skb, ct) < 0)
 706                         goto nla_put_failure;
 707
 708                 if (events & (1 << IPCT_SEQADJ) &&
 709                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 710                         goto nla_put_failure;
 711         }
 712
 713 #ifdef CONFIG_NF_CONNTRACK_MARK
 714         if ((events & (1 << IPCT_MARK) || ct->mark)
 715             && ctnetlink_dump_mark(skb, ct) < 0)
 716                 goto nla_put_failure;
 717 #endif
 718         rcu_read_unlock();
 719
 720         nlmsg_end(skb, nlh);
 721         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 722                              GFP_ATOMIC);
 723         if (err == -ENOBUFS || err == -EAGAIN)
 724                 return -ENOBUFS;
 725
 726         return 0;
 727
 728 nla_put_failure:
 729         rcu_read_unlock();
 730         nlmsg_cancel(skb, nlh);
 731 nlmsg_failure:
 732         kfree_skb(skb);
 733 errout:
 734         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 735                 return -ENOBUFS;
 736
 737         return 0;
 738 }
 739 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 740
 741 static int ctnetlink_done(struct netlink_callback *cb)
 742 {
 743         if (cb->args[1])
 744                 nf_ct_put((struct nf_conn *)cb->args[1]);
 745         if (cb->data)
 746                 kfree(cb->data);
 747         return 0;
 748 }
 749
 750 struct ctnetlink_dump_filter {
 751         struct {
 752                 u_int32_t val;
 753                 u_int32_t mask;
 754         } mark;
 755 };
 756
 757 static int
 758 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 759 {
 760         struct net *net = sock_net(skb->sk);
 761         struct nf_conn *ct, *last;
 762         struct nf_conntrack_tuple_hash *h;
 763         struct hlist_nulls_node *n;
 764         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 765         u_int8_t l3proto = nfmsg->nfgen_family;
 766         int res;
 767 #ifdef CONFIG_NF_CONNTRACK_MARK
 768         const struct ctnetlink_dump_filter *filter = cb->data;
 769 #endif
 770
 771         spin_lock_bh(&nf_conntrack_lock);
 772         last = (struct nf_conn *)cb->args[1];
 773         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 774 restart:
 775                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 776                                          hnnode) {
 777                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 778                                 continue;
 779                         ct = nf_ct_tuplehash_to_ctrack(h);
 780                         /* Dump entries of a given L3 protocol number.
 781                          * If it is not specified, ie. l3proto == 0,
 782                          * then dump everything. */
 783                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 784                                 continue;
 785                         if (cb->args[1]) {
 786                                 if (ct != last)
 787                                         continue;
 788                                 cb->args[1] = 0;
 789                         }
 790 #ifdef CONFIG_NF_CONNTRACK_MARK
 791                         if (filter && !((ct->mark & filter->mark.mask) ==
 792                                         filter->mark.val)) {
 793                                 continue;
 794                         }
 795 #endif
 796                         rcu_read_lock();
 797                         res =
 798                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 799                                             cb->nlh->nlmsg_seq,
 800                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 801                                             ct);
 802                         rcu_read_unlock();
 803                         if (res < 0) {
 804                                 nf_conntrack_get(&ct->ct_general);
 805                                 cb->args[1] = (unsigned long)ct;
 806                                 goto out;
 807                         }
 808                 }
 809                 if (cb->args[1]) {
 810                         cb->args[1] = 0;
 811                         goto restart;
 812                 }
 813         }
 814 out:
 815         spin_unlock_bh(&nf_conntrack_lock);
 816         if (last)
 817                 nf_ct_put(last);
 818
 819         return skb->len;
 820 }
 821
 822 static inline int
 823 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 824 {
 825         struct nlattr *tb[CTA_IP_MAX+1];
 826         struct nf_conntrack_l3proto *l3proto;
 827         int ret = 0;
 828
 829         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 830         if (ret < 0)
 831                 return ret;
 832
 833         rcu_read_lock();
 834         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 835
 836         if (likely(l3proto->nlattr_to_tuple)) {
 837                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 838                                           l3proto->nla_policy);
 839                 if (ret == 0)
 840                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 841         }
 842
 843         rcu_read_unlock();
 844
 845         return ret;
 846 }
 847
 848 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 849         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 850 };
 851
 852 static inline int
 853 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 854                             struct nf_conntrack_tuple *tuple)
 855 {
 856         struct nlattr *tb[CTA_PROTO_MAX+1];
 857         struct nf_conntrack_l4proto *l4proto;
 858         int ret = 0;
 859
 860         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 861         if (ret < 0)
 862                 return ret;
 863
 864         if (!tb[CTA_PROTO_NUM])
 865                 return -EINVAL;
 866         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 867
 868         rcu_read_lock();
 869         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 870
 871         if (likely(l4proto->nlattr_to_tuple)) {
 872                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 873                                           l4proto->nla_policy);
 874                 if (ret == 0)
 875                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 876         }
 877
 878         rcu_read_unlock();
 879
 880         return ret;
 881 }
 882
 883 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 884         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 885         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 886 };
 887
 888 static int
 889 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 890                       struct nf_conntrack_tuple *tuple,
 891                       enum ctattr_type type, u_int8_t l3num)
 892 {
 893         struct nlattr *tb[CTA_TUPLE_MAX+1];
 894         int err;
 895
 896         memset(tuple, 0, sizeof(*tuple));
 897
 898         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 899         if (err < 0)
 900                 return err;
 901
 902         if (!tb[CTA_TUPLE_IP])
 903                 return -EINVAL;
 904
 905         tuple->src.l3num = l3num;
 906
 907         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 908         if (err < 0)
 909                 return err;
 910
 911         if (!tb[CTA_TUPLE_PROTO])
 912                 return -EINVAL;
 913
 914         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 915         if (err < 0)
 916                 return err;
 917
 918         /* orig and expect tuples get DIR_ORIGINAL */
 919         if (type == CTA_TUPLE_REPLY)
 920                 tuple->dst.dir = IP_CT_DIR_REPLY;
 921         else
 922                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 923
 924         return 0;
 925 }
 926
 927 static int
 928 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 929 {
 930         if (attr)
 931 #ifdef CONFIG_NF_CONNTRACK_ZONES
 932                 *zone = ntohs(nla_get_be16(attr));
 933 #else
 934                 return -EOPNOTSUPP;
 935 #endif
 936         else
 937                 *zone = 0;
 938
 939         return 0;
 940 }
 941
 942 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 943         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 944                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 945 };
 946
 947 static inline int
 948 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 949                      struct nlattr **helpinfo)
 950 {
 951         int err;
 952         struct nlattr *tb[CTA_HELP_MAX+1];
 953
 954         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 955         if (err < 0)
 956                 return err;
 957
 958         if (!tb[CTA_HELP_NAME])
 959                 return -EINVAL;
 960
 961         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 962
 963         if (tb[CTA_HELP_INFO])
 964                 *helpinfo = tb[CTA_HELP_INFO];
 965
 966         return 0;
 967 }
 968
 969 #define __CTA_LABELS_MAX_LENGTH ((XT_CONNLABEL_MAXBIT + 1) / BITS_PER_BYTE)
 970 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 971         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 972         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 973         [CTA_STATUS]            = { .type = NLA_U32 },
 974         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 975         [CTA_HELP]              = { .type = NLA_NESTED },
 976         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 977         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 978         [CTA_MARK]              = { .type = NLA_U32 },
 979         [CTA_ID]                = { .type = NLA_U32 },
 980         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 981         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 982         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
 983         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
 984         [CTA_ZONE]              = { .type = NLA_U16 },
 985         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 986         [CTA_LABELS]            = { .type = NLA_BINARY,
 987                                     .len = __CTA_LABELS_MAX_LENGTH },
 988         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
 989                                     .len = __CTA_LABELS_MAX_LENGTH },
 990 };
 991
 992 static int
 993 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 994                         const struct nlmsghdr *nlh,
 995                         const struct nlattr * const cda[])
 996 {
 997         struct net *net = sock_net(ctnl);
 998         struct nf_conntrack_tuple_hash *h;
 999         struct nf_conntrack_tuple tuple;
1000         struct nf_conn *ct;
1001         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1002         u_int8_t u3 = nfmsg->nfgen_family;
1003         u16 zone;
1004         int err;
1005
1006         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1007         if (err < 0)
1008                 return err;
1009
1010         if (cda[CTA_TUPLE_ORIG])
1011                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1012         else if (cda[CTA_TUPLE_REPLY])
1013                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1014         else {
1015                 /* Flush the whole table */
1016                 nf_conntrack_flush_report(net,
1017                                          NETLINK_CB(skb).portid,
1018                                          nlmsg_report(nlh));
1019                 return 0;
1020         }
1021
1022         if (err < 0)
1023                 return err;
1024
1025         h = nf_conntrack_find_get(net, zone, &tuple);
1026         if (!h)
1027                 return -ENOENT;
1028
1029         ct = nf_ct_tuplehash_to_ctrack(h);
1030
1031         if (cda[CTA_ID]) {
1032                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1033                 if (id != (u32)(unsigned long)ct) {
1034                         nf_ct_put(ct);
1035                         return -ENOENT;
1036                 }
1037         }
1038
1039         if (del_timer(&ct->timeout))
1040                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1041
1042         nf_ct_put(ct);
1043
1044         return 0;
1045 }
1046
1047 static int
1048 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1049                         const struct nlmsghdr *nlh,
1050                         const struct nlattr * const cda[])
1051 {
1052         struct net *net = sock_net(ctnl);
1053         struct nf_conntrack_tuple_hash *h;
1054         struct nf_conntrack_tuple tuple;
1055         struct nf_conn *ct;
1056         struct sk_buff *skb2 = NULL;
1057         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1058         u_int8_t u3 = nfmsg->nfgen_family;
1059         u16 zone;
1060         int err;
1061
1062         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1063                 struct netlink_dump_control c = {
1064                         .dump = ctnetlink_dump_table,
1065                         .done = ctnetlink_done,
1066                 };
1067 #ifdef CONFIG_NF_CONNTRACK_MARK
1068                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1069                         struct ctnetlink_dump_filter *filter;
1070
1071                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1072                                          GFP_ATOMIC);
1073                         if (filter == NULL)
1074                                 return -ENOMEM;
1075
1076                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1077                         filter->mark.mask =
1078                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1079                         c.data = filter;
1080                 }
1081 #endif
1082                 return netlink_dump_start(ctnl, skb, nlh, &c);
1083         }
1084
1085         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1086         if (err < 0)
1087                 return err;
1088
1089         if (cda[CTA_TUPLE_ORIG])
1090                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1091         else if (cda[CTA_TUPLE_REPLY])
1092                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1093         else
1094                 return -EINVAL;
1095
1096         if (err < 0)
1097                 return err;
1098
1099         h = nf_conntrack_find_get(net, zone, &tuple);
1100         if (!h)
1101                 return -ENOENT;
1102
1103         ct = nf_ct_tuplehash_to_ctrack(h);
1104
1105         err = -ENOMEM;
1106         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1107         if (skb2 == NULL) {
1108                 nf_ct_put(ct);
1109                 return -ENOMEM;
1110         }
1111
1112         rcu_read_lock();
1113         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1114                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1115         rcu_read_unlock();
1116         nf_ct_put(ct);
1117         if (err <= 0)
1118                 goto free;
1119
1120         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1121         if (err < 0)
1122                 goto out;
1123
1124         return 0;
1125
1126 free:
1127         kfree_skb(skb2);
1128 out:
1129         /* this avoids a loop in nfnetlink. */
1130         return err == -EAGAIN ? -ENOBUFS : err;
1131 }
1132
1133 static int ctnetlink_done_list(struct netlink_callback *cb)
1134 {
1135         if (cb->args[1])
1136                 nf_ct_put((struct nf_conn *)cb->args[1]);
1137         return 0;
1138 }
1139
1140 static int
1141 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb,
1142                     struct hlist_nulls_head *list)
1143 {
1144         struct nf_conn *ct, *last;
1145         struct nf_conntrack_tuple_hash *h;
1146         struct hlist_nulls_node *n;
1147         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1148         u_int8_t l3proto = nfmsg->nfgen_family;
1149         int res;
1150
1151         if (cb->args[2])
1152                 return 0;
1153
1154         spin_lock_bh(&nf_conntrack_lock);
1155         last = (struct nf_conn *)cb->args[1];
1156 restart:
1157         hlist_nulls_for_each_entry(h, n, list, hnnode) {
1158                 ct = nf_ct_tuplehash_to_ctrack(h);
1159                 if (l3proto && nf_ct_l3num(ct) != l3proto)
1160                         continue;
1161                 if (cb->args[1]) {
1162                         if (ct != last)
1163                                 continue;
1164                         cb->args[1] = 0;
1165                 }
1166                 rcu_read_lock();
1167                 res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1168                                           cb->nlh->nlmsg_seq,
1169                                           NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1170                                           ct);
1171                 rcu_read_unlock();
1172                 if (res < 0) {
1173                         nf_conntrack_get(&ct->ct_general);
1174                         cb->args[1] = (unsigned long)ct;
1175                         goto out;
1176                 }
1177         }
1178         if (cb->args[1]) {
1179                 cb->args[1] = 0;
1180                 goto restart;
1181         } else
1182                 cb->args[2] = 1;
1183 out:
1184         spin_unlock_bh(&nf_conntrack_lock);
1185         if (last)
1186                 nf_ct_put(last);
1187
1188         return skb->len;
1189 }
1190
1191 static int
1192 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1193 {
1194         struct net *net = sock_net(skb->sk);
1195
1196         return ctnetlink_dump_list(skb, cb, &net->ct.dying);
1197 }
1198
1199 static int
1200 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1201                        const struct nlmsghdr *nlh,
1202                        const struct nlattr * const cda[])
1203 {
1204         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1205                 struct netlink_dump_control c = {
1206                         .dump = ctnetlink_dump_dying,
1207                         .done = ctnetlink_done_list,
1208                 };
1209                 return netlink_dump_start(ctnl, skb, nlh, &c);
1210         }
1211
1212         return -EOPNOTSUPP;
1213 }
1214
1215 static int
1216 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1217 {
1218         struct net *net = sock_net(skb->sk);
1219
1220         return ctnetlink_dump_list(skb, cb, &net->ct.unconfirmed);
1221 }
1222
1223 static int
1224 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1225                              const struct nlmsghdr *nlh,
1226                              const struct nlattr * const cda[])
1227 {
1228         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1229                 struct netlink_dump_control c = {
1230                         .dump = ctnetlink_dump_unconfirmed,
1231                         .done = ctnetlink_done_list,
1232                 };
1233                 return netlink_dump_start(ctnl, skb, nlh, &c);
1234         }
1235
1236         return -EOPNOTSUPP;
1237 }
1238
1239 #ifdef CONFIG_NF_NAT_NEEDED
1240 static int
1241 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1242                           enum nf_nat_manip_type manip,
1243                           const struct nlattr *attr)
1244 {
1245         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1246         int err;
1247
1248         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1249         if (!parse_nat_setup) {
1250 #ifdef CONFIG_MODULES
1251                 rcu_read_unlock();
1252                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1253                 if (request_module("nf-nat") < 0) {
1254                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1255                         rcu_read_lock();
1256                         return -EOPNOTSUPP;
1257                 }
1258                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1259                 rcu_read_lock();
1260                 if (nfnetlink_parse_nat_setup_hook)
1261                         return -EAGAIN;
1262 #endif
1263                 return -EOPNOTSUPP;
1264         }
1265
1266         err = parse_nat_setup(ct, manip, attr);
1267         if (err == -EAGAIN) {
1268 #ifdef CONFIG_MODULES
1269                 rcu_read_unlock();
1270                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1271                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1272                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1273                         rcu_read_lock();
1274                         return -EOPNOTSUPP;
1275                 }
1276                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1277                 rcu_read_lock();
1278 #else
1279                 err = -EOPNOTSUPP;
1280 #endif
1281         }
1282         return err;
1283 }
1284 #endif
1285
1286 static int
1287 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1288 {
1289         unsigned long d;
1290         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1291         d = ct->status ^ status;
1292
1293         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1294                 /* unchangeable */
1295                 return -EBUSY;
1296
1297         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1298                 /* SEEN_REPLY bit can only be set */
1299                 return -EBUSY;
1300
1301         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1302                 /* ASSURED bit can only be set */
1303                 return -EBUSY;
1304
1305         /* Be careful here, modifying NAT bits can screw up things,
1306          * so don't let users modify them directly if they don't pass
1307          * nf_nat_range. */
1308         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1309         return 0;
1310 }
1311
1312 static int
1313 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1314 {
1315 #ifdef CONFIG_NF_NAT_NEEDED
1316         int ret;
1317
1318         if (cda[CTA_NAT_DST]) {
1319                 ret = ctnetlink_parse_nat_setup(ct,
1320                                                 NF_NAT_MANIP_DST,
1321                                                 cda[CTA_NAT_DST]);
1322                 if (ret < 0)
1323                         return ret;
1324         }
1325         if (cda[CTA_NAT_SRC]) {
1326                 ret = ctnetlink_parse_nat_setup(ct,
1327                                                 NF_NAT_MANIP_SRC,
1328                                                 cda[CTA_NAT_SRC]);
1329                 if (ret < 0)
1330                         return ret;
1331         }
1332         return 0;
1333 #else
1334         return -EOPNOTSUPP;
1335 #endif
1336 }
1337
1338 static inline int
1339 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1340 {
1341         struct nf_conntrack_helper *helper;
1342         struct nf_conn_help *help = nfct_help(ct);
1343         char *helpname = NULL;
1344         struct nlattr *helpinfo = NULL;
1345         int err;
1346
1347         /* don't change helper of sibling connections */
1348         if (ct->master)
1349                 return -EBUSY;
1350
1351         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1352         if (err < 0)
1353                 return err;
1354
1355         if (!strcmp(helpname, "")) {
1356                 if (help && help->helper) {
1357                         /* we had a helper before ... */
1358                         nf_ct_remove_expectations(ct);
1359                         RCU_INIT_POINTER(help->helper, NULL);
1360                 }
1361
1362                 return 0;
1363         }
1364
1365         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1366                                             nf_ct_protonum(ct));
1367         if (helper == NULL) {
1368 #ifdef CONFIG_MODULES
1369                 spin_unlock_bh(&nf_conntrack_lock);
1370
1371                 if (request_module("nfct-helper-%s", helpname) < 0) {
1372                         spin_lock_bh(&nf_conntrack_lock);
1373                         return -EOPNOTSUPP;
1374                 }
1375
1376                 spin_lock_bh(&nf_conntrack_lock);
1377                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1378                                                     nf_ct_protonum(ct));
1379                 if (helper)
1380                         return -EAGAIN;
1381 #endif
1382                 return -EOPNOTSUPP;
1383         }
1384
1385         if (help) {
1386                 if (help->helper == helper) {
1387                         /* update private helper data if allowed. */
1388                         if (helper->from_nlattr)
1389                                 helper->from_nlattr(helpinfo, ct);
1390                         return 0;
1391                 } else
1392                         return -EBUSY;
1393         }
1394
1395         /* we cannot set a helper for an existing conntrack */
1396         return -EOPNOTSUPP;
1397 }
1398
1399 static inline int
1400 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1401 {
1402         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1403
1404         if (!del_timer(&ct->timeout))
1405                 return -ETIME;
1406
1407         ct->timeout.expires = jiffies + timeout * HZ;
1408         add_timer(&ct->timeout);
1409
1410         return 0;
1411 }
1412
1413 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1414         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1415         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1416         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1417 };
1418
1419 static inline int
1420 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1421 {
1422         const struct nlattr *attr = cda[CTA_PROTOINFO];
1423         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1424         struct nf_conntrack_l4proto *l4proto;
1425         int err = 0;
1426
1427         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1428         if (err < 0)
1429                 return err;
1430
1431         rcu_read_lock();
1432         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1433         if (l4proto->from_nlattr)
1434                 err = l4proto->from_nlattr(tb, ct);
1435         rcu_read_unlock();
1436
1437         return err;
1438 }
1439
1440 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1441         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1442         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1443         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1444 };
1445
1446 static inline int
1447 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1448 {
1449         int err;
1450         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1451
1452         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1453         if (err < 0)
1454                 return err;
1455
1456         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1457                 return -EINVAL;
1458
1459         seq->correction_pos =
1460                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1461
1462         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1463                 return -EINVAL;
1464
1465         seq->offset_before =
1466                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1467
1468         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1469                 return -EINVAL;
1470
1471         seq->offset_after =
1472                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1473
1474         return 0;
1475 }
1476
1477 static int
1478 ctnetlink_change_seq_adj(struct nf_conn *ct,
1479                          const struct nlattr * const cda[])
1480 {
1481         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1482         int ret = 0;
1483
1484         if (!seqadj)
1485                 return 0;
1486
1487         if (cda[CTA_SEQ_ADJ_ORIG]) {
1488                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1489                                      cda[CTA_SEQ_ADJ_ORIG]);
1490                 if (ret < 0)
1491                         return ret;
1492
1493                 ct->status |= IPS_SEQ_ADJUST;
1494         }
1495
1496         if (cda[CTA_SEQ_ADJ_REPLY]) {
1497                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1498                                      cda[CTA_SEQ_ADJ_REPLY]);
1499                 if (ret < 0)
1500                         return ret;
1501
1502                 ct->status |= IPS_SEQ_ADJUST;
1503         }
1504
1505         return 0;
1506 }
1507
1508 static int
1509 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1510 {
1511 #ifdef CONFIG_NF_CONNTRACK_LABELS
1512         size_t len = nla_len(cda[CTA_LABELS]);
1513         const void *mask = cda[CTA_LABELS_MASK];
1514
1515         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1516                 return -EINVAL;
1517
1518         if (mask) {
1519                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1520                     nla_len(cda[CTA_LABELS_MASK]) != len)
1521                         return -EINVAL;
1522                 mask = nla_data(cda[CTA_LABELS_MASK]);
1523         }
1524
1525         len /= sizeof(u32);
1526
1527         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1528 #else
1529         return -EOPNOTSUPP;
1530 #endif
1531 }
1532
1533 static int
1534 ctnetlink_change_conntrack(struct nf_conn *ct,
1535                            const struct nlattr * const cda[])
1536 {
1537         int err;
1538
1539         /* only allow NAT changes and master assignation for new conntracks */
1540         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1541                 return -EOPNOTSUPP;
1542
1543         if (cda[CTA_HELP]) {
1544                 err = ctnetlink_change_helper(ct, cda);
1545                 if (err < 0)
1546                         return err;
1547         }
1548
1549         if (cda[CTA_TIMEOUT]) {
1550                 err = ctnetlink_change_timeout(ct, cda);
1551                 if (err < 0)
1552                         return err;
1553         }
1554
1555         if (cda[CTA_STATUS]) {
1556                 err = ctnetlink_change_status(ct, cda);
1557                 if (err < 0)
1558                         return err;
1559         }
1560
1561         if (cda[CTA_PROTOINFO]) {
1562                 err = ctnetlink_change_protoinfo(ct, cda);
1563                 if (err < 0)
1564                         return err;
1565         }
1566
1567 #if defined(CONFIG_NF_CONNTRACK_MARK)
1568         if (cda[CTA_MARK])
1569                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1570 #endif
1571
1572         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1573                 err = ctnetlink_change_seq_adj(ct, cda);
1574                 if (err < 0)
1575                         return err;
1576         }
1577
1578         if (cda[CTA_LABELS]) {
1579                 err = ctnetlink_attach_labels(ct, cda);
1580                 if (err < 0)
1581                         return err;
1582         }
1583
1584         return 0;
1585 }
1586
1587 static struct nf_conn *
1588 ctnetlink_create_conntrack(struct net *net, u16 zone,
1589                            const struct nlattr * const cda[],
1590                            struct nf_conntrack_tuple *otuple,
1591                            struct nf_conntrack_tuple *rtuple,
1592                            u8 u3)
1593 {
1594         struct nf_conn *ct;
1595         int err = -EINVAL;
1596         struct nf_conntrack_helper *helper;
1597         struct nf_conn_tstamp *tstamp;
1598
1599         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1600         if (IS_ERR(ct))
1601                 return ERR_PTR(-ENOMEM);
1602
1603         if (!cda[CTA_TIMEOUT])
1604                 goto err1;
1605         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1606
1607         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1608
1609         rcu_read_lock();
1610         if (cda[CTA_HELP]) {
1611                 char *helpname = NULL;
1612                 struct nlattr *helpinfo = NULL;
1613
1614                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1615                 if (err < 0)
1616                         goto err2;
1617
1618                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1619                                                     nf_ct_protonum(ct));
1620                 if (helper == NULL) {
1621                         rcu_read_unlock();
1622 #ifdef CONFIG_MODULES
1623                         if (request_module("nfct-helper-%s", helpname) < 0) {
1624                                 err = -EOPNOTSUPP;
1625                                 goto err1;
1626                         }
1627
1628                         rcu_read_lock();
1629                         helper = __nf_conntrack_helper_find(helpname,
1630                                                             nf_ct_l3num(ct),
1631                                                             nf_ct_protonum(ct));
1632                         if (helper) {
1633                                 err = -EAGAIN;
1634                                 goto err2;
1635                         }
1636                         rcu_read_unlock();
1637 #endif
1638                         err = -EOPNOTSUPP;
1639                         goto err1;
1640                 } else {
1641                         struct nf_conn_help *help;
1642
1643                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1644                         if (help == NULL) {
1645                                 err = -ENOMEM;
1646                                 goto err2;
1647                         }
1648                         /* set private helper data if allowed. */
1649                         if (helper->from_nlattr)
1650                                 helper->from_nlattr(helpinfo, ct);
1651
1652                         /* not in hash table yet so not strictly necessary */
1653                         RCU_INIT_POINTER(help->helper, helper);
1654                 }
1655         } else {
1656                 /* try an implicit helper assignation */
1657                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1658                 if (err < 0)
1659                         goto err2;
1660         }
1661
1662         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1663                 err = ctnetlink_change_nat(ct, cda);
1664                 if (err < 0)
1665                         goto err2;
1666         }
1667
1668         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1669         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1670         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1671         nf_ct_labels_ext_add(ct);
1672
1673         /* we must add conntrack extensions before confirmation. */
1674         ct->status |= IPS_CONFIRMED;
1675
1676         if (cda[CTA_STATUS]) {
1677                 err = ctnetlink_change_status(ct, cda);
1678                 if (err < 0)
1679                         goto err2;
1680         }
1681
1682         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1683                 err = ctnetlink_change_seq_adj(ct, cda);
1684                 if (err < 0)
1685                         goto err2;
1686         }
1687
1688         memset(&ct->proto, 0, sizeof(ct->proto));
1689         if (cda[CTA_PROTOINFO]) {
1690                 err = ctnetlink_change_protoinfo(ct, cda);
1691                 if (err < 0)
1692                         goto err2;
1693         }
1694
1695 #if defined(CONFIG_NF_CONNTRACK_MARK)
1696         if (cda[CTA_MARK])
1697                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1698 #endif
1699
1700         /* setup master conntrack: this is a confirmed expectation */
1701         if (cda[CTA_TUPLE_MASTER]) {
1702                 struct nf_conntrack_tuple master;
1703                 struct nf_conntrack_tuple_hash *master_h;
1704                 struct nf_conn *master_ct;
1705
1706                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1707                 if (err < 0)
1708                         goto err2;
1709
1710                 master_h = nf_conntrack_find_get(net, zone, &master);
1711                 if (master_h == NULL) {
1712                         err = -ENOENT;
1713                         goto err2;
1714                 }
1715                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1716                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1717                 ct->master = master_ct;
1718         }
1719         tstamp = nf_conn_tstamp_find(ct);
1720         if (tstamp)
1721                 tstamp->start = ktime_to_ns(ktime_get_real());
1722
1723         err = nf_conntrack_hash_check_insert(ct);
1724         if (err < 0)
1725                 goto err2;
1726
1727         rcu_read_unlock();
1728
1729         return ct;
1730
1731 err2:
1732         rcu_read_unlock();
1733 err1:
1734         nf_conntrack_free(ct);
1735         return ERR_PTR(err);
1736 }
1737
1738 static int
1739 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1740                         const struct nlmsghdr *nlh,
1741                         const struct nlattr * const cda[])
1742 {
1743         struct net *net = sock_net(ctnl);
1744         struct nf_conntrack_tuple otuple, rtuple;
1745         struct nf_conntrack_tuple_hash *h = NULL;
1746         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1747         struct nf_conn *ct;
1748         u_int8_t u3 = nfmsg->nfgen_family;
1749         u16 zone;
1750         int err;
1751
1752         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1753         if (err < 0)
1754                 return err;
1755
1756         if (cda[CTA_TUPLE_ORIG]) {
1757                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1758                 if (err < 0)
1759                         return err;
1760         }
1761
1762         if (cda[CTA_TUPLE_REPLY]) {
1763                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1764                 if (err < 0)
1765                         return err;
1766         }
1767
1768         if (cda[CTA_TUPLE_ORIG])
1769                 h = nf_conntrack_find_get(net, zone, &otuple);
1770         else if (cda[CTA_TUPLE_REPLY])
1771                 h = nf_conntrack_find_get(net, zone, &rtuple);
1772
1773         if (h == NULL) {
1774                 err = -ENOENT;
1775                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1776                         enum ip_conntrack_events events;
1777
1778                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1779                                 return -EINVAL;
1780
1781                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1782                                                         &rtuple, u3);
1783                         if (IS_ERR(ct))
1784                                 return PTR_ERR(ct);
1785
1786                         err = 0;
1787                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1788                                 events = IPCT_RELATED;
1789                         else
1790                                 events = IPCT_NEW;
1791
1792                         if (cda[CTA_LABELS] &&
1793                             ctnetlink_attach_labels(ct, cda) == 0)
1794                                 events |= (1 << IPCT_LABEL);
1795
1796                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1797                                                       (1 << IPCT_ASSURED) |
1798                                                       (1 << IPCT_HELPER) |
1799                                                       (1 << IPCT_PROTOINFO) |
1800                                                       (1 << IPCT_SEQADJ) |
1801                                                       (1 << IPCT_MARK) | events,
1802                                                       ct, NETLINK_CB(skb).portid,
1803                                                       nlmsg_report(nlh));
1804                         nf_ct_put(ct);
1805                 }
1806
1807                 return err;
1808         }
1809         /* implicit 'else' */
1810
1811         err = -EEXIST;
1812         ct = nf_ct_tuplehash_to_ctrack(h);
1813         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1814                 spin_lock_bh(&nf_conntrack_lock);
1815                 err = ctnetlink_change_conntrack(ct, cda);
1816                 spin_unlock_bh(&nf_conntrack_lock);
1817                 if (err == 0) {
1818                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1819                                                       (1 << IPCT_ASSURED) |
1820                                                       (1 << IPCT_HELPER) |
1821                                                       (1 << IPCT_LABEL) |
1822                                                       (1 << IPCT_PROTOINFO) |
1823                                                       (1 << IPCT_SEQADJ) |
1824                                                       (1 << IPCT_MARK),
1825                                                       ct, NETLINK_CB(skb).portid,
1826                                                       nlmsg_report(nlh));
1827                 }
1828         }
1829
1830         nf_ct_put(ct);
1831         return err;
1832 }
1833
1834 static int
1835 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1836                                 __u16 cpu, const struct ip_conntrack_stat *st)
1837 {
1838         struct nlmsghdr *nlh;
1839         struct nfgenmsg *nfmsg;
1840         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1841
1842         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1843         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1844         if (nlh == NULL)
1845                 goto nlmsg_failure;
1846
1847         nfmsg = nlmsg_data(nlh);
1848         nfmsg->nfgen_family = AF_UNSPEC;
1849         nfmsg->version      = NFNETLINK_V0;
1850         nfmsg->res_id       = htons(cpu);
1851
1852         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1853             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1854             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1855             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1856             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1857             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1858             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1859             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1860             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1861                                 htonl(st->insert_failed)) ||
1862             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1863             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1864             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1865             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1866                                 htonl(st->search_restart)))
1867                 goto nla_put_failure;
1868
1869         nlmsg_end(skb, nlh);
1870         return skb->len;
1871
1872 nla_put_failure:
1873 nlmsg_failure:
1874         nlmsg_cancel(skb, nlh);
1875         return -1;
1876 }
1877
1878 static int
1879 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1880 {
1881         int cpu;
1882         struct net *net = sock_net(skb->sk);
1883
1884         if (cb->args[0] == nr_cpu_ids)
1885                 return 0;
1886
1887         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1888                 const struct ip_conntrack_stat *st;
1889
1890                 if (!cpu_possible(cpu))
1891                         continue;
1892
1893                 st = per_cpu_ptr(net->ct.stat, cpu);
1894                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1895                                                     NETLINK_CB(cb->skb).portid,
1896                                                     cb->nlh->nlmsg_seq,
1897                                                     cpu, st) < 0)
1898                                 break;
1899         }
1900         cb->args[0] = cpu;
1901
1902         return skb->len;
1903 }
1904
1905 static int
1906 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1907                       const struct nlmsghdr *nlh,
1908                       const struct nlattr * const cda[])
1909 {
1910         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1911                 struct netlink_dump_control c = {
1912                         .dump = ctnetlink_ct_stat_cpu_dump,
1913                 };
1914                 return netlink_dump_start(ctnl, skb, nlh, &c);
1915         }
1916
1917         return 0;
1918 }
1919
1920 static int
1921 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1922                             struct net *net)
1923 {
1924         struct nlmsghdr *nlh;
1925         struct nfgenmsg *nfmsg;
1926         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1927         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1928
1929         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1930         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1931         if (nlh == NULL)
1932                 goto nlmsg_failure;
1933
1934         nfmsg = nlmsg_data(nlh);
1935         nfmsg->nfgen_family = AF_UNSPEC;
1936         nfmsg->version      = NFNETLINK_V0;
1937         nfmsg->res_id       = 0;
1938
1939         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1940                 goto nla_put_failure;
1941
1942         nlmsg_end(skb, nlh);
1943         return skb->len;
1944
1945 nla_put_failure:
1946 nlmsg_failure:
1947         nlmsg_cancel(skb, nlh);
1948         return -1;
1949 }
1950
1951 static int
1952 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1953                   const struct nlmsghdr *nlh,
1954                   const struct nlattr * const cda[])
1955 {
1956         struct sk_buff *skb2;
1957         int err;
1958
1959         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1960         if (skb2 == NULL)
1961                 return -ENOMEM;
1962
1963         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1964                                           nlh->nlmsg_seq,
1965                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1966                                           sock_net(skb->sk));
1967         if (err <= 0)
1968                 goto free;
1969
1970         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1971         if (err < 0)
1972                 goto out;
1973
1974         return 0;
1975
1976 free:
1977         kfree_skb(skb2);
1978 out:
1979         /* this avoids a loop in nfnetlink. */
1980         return err == -EAGAIN ? -ENOBUFS : err;
1981 }
1982
1983 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1984         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1985         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1986         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1987         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1988         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1989         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
1990                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1991         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1992         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1993         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
1994         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
1995         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
1996 };
1997
1998 static struct nf_conntrack_expect *
1999 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2000                        struct nf_conntrack_helper *helper,
2001                        struct nf_conntrack_tuple *tuple,
2002                        struct nf_conntrack_tuple *mask);
2003
2004 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2005 static size_t
2006 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2007 {
2008         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2009                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2010                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2011                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2012                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2013                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2014                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2015                + nla_total_size(0) /* CTA_PROTOINFO */
2016                + nla_total_size(0) /* CTA_HELP */
2017                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2018                + ctnetlink_secctx_size(ct)
2019 #ifdef CONFIG_NF_NAT_NEEDED
2020                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2021                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2022 #endif
2023 #ifdef CONFIG_NF_CONNTRACK_MARK
2024                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2025 #endif
2026                + ctnetlink_proto_size(ct)
2027                ;
2028 }
2029
2030 static int
2031 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2032 {
2033         struct nlattr *nest_parms;
2034
2035         rcu_read_lock();
2036         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2037         if (!nest_parms)
2038                 goto nla_put_failure;
2039         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2040                 goto nla_put_failure;
2041         nla_nest_end(skb, nest_parms);
2042
2043         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2044         if (!nest_parms)
2045                 goto nla_put_failure;
2046         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2047                 goto nla_put_failure;
2048         nla_nest_end(skb, nest_parms);
2049
2050         if (nf_ct_zone(ct)) {
2051                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2052                         goto nla_put_failure;
2053         }
2054
2055         if (ctnetlink_dump_id(skb, ct) < 0)
2056                 goto nla_put_failure;
2057
2058         if (ctnetlink_dump_status(skb, ct) < 0)
2059                 goto nla_put_failure;
2060
2061         if (ctnetlink_dump_timeout(skb, ct) < 0)
2062                 goto nla_put_failure;
2063
2064         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2065                 goto nla_put_failure;
2066
2067         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2068                 goto nla_put_failure;
2069
2070 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2071         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2072                 goto nla_put_failure;
2073 #endif
2074         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2075                 goto nla_put_failure;
2076
2077         if ((ct->status & IPS_SEQ_ADJUST) &&
2078             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2079                 goto nla_put_failure;
2080
2081 #ifdef CONFIG_NF_CONNTRACK_MARK
2082         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2083                 goto nla_put_failure;
2084 #endif
2085         if (ctnetlink_dump_labels(skb, ct) < 0)
2086                 goto nla_put_failure;
2087         rcu_read_unlock();
2088         return 0;
2089
2090 nla_put_failure:
2091         rcu_read_unlock();
2092         return -ENOSPC;
2093 }
2094
2095 static int
2096 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2097 {
2098         int err;
2099
2100         if (cda[CTA_TIMEOUT]) {
2101                 err = ctnetlink_change_timeout(ct, cda);
2102                 if (err < 0)
2103                         return err;
2104         }
2105         if (cda[CTA_STATUS]) {
2106                 err = ctnetlink_change_status(ct, cda);
2107                 if (err < 0)
2108                         return err;
2109         }
2110         if (cda[CTA_HELP]) {
2111                 err = ctnetlink_change_helper(ct, cda);
2112                 if (err < 0)
2113                         return err;
2114         }
2115         if (cda[CTA_LABELS]) {
2116                 err = ctnetlink_attach_labels(ct, cda);
2117                 if (err < 0)
2118                         return err;
2119         }
2120 #if defined(CONFIG_NF_CONNTRACK_MARK)
2121         if (cda[CTA_MARK]) {
2122                 u32 mask = 0, mark, newmark;
2123                 if (cda[CTA_MARK_MASK])
2124                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2125
2126                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2127                 newmark = (ct->mark & mask) ^ mark;
2128                 if (newmark != ct->mark)
2129                         ct->mark = newmark;
2130         }
2131 #endif
2132         return 0;
2133 }
2134
2135 static int
2136 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2137 {
2138         struct nlattr *cda[CTA_MAX+1];
2139         int ret;
2140
2141         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2142         if (ret < 0)
2143                 return ret;
2144
2145         spin_lock_bh(&nf_conntrack_lock);
2146         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2147         spin_unlock_bh(&nf_conntrack_lock);
2148
2149         return ret;
2150 }
2151
2152 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2153                                        const struct nf_conn *ct,
2154                                        struct nf_conntrack_tuple *tuple,
2155                                        struct nf_conntrack_tuple *mask)
2156 {
2157         int err;
2158
2159         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2160                                     nf_ct_l3num(ct));
2161         if (err < 0)
2162                 return err;
2163
2164         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2165                                      nf_ct_l3num(ct));
2166 }
2167
2168 static int
2169 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2170                                 u32 portid, u32 report)
2171 {
2172         struct nlattr *cda[CTA_EXPECT_MAX+1];
2173         struct nf_conntrack_tuple tuple, mask;
2174         struct nf_conntrack_helper *helper = NULL;
2175         struct nf_conntrack_expect *exp;
2176         int err;
2177
2178         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2179         if (err < 0)
2180                 return err;
2181
2182         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2183                                           ct, &tuple, &mask);
2184         if (err < 0)
2185                 return err;
2186
2187         if (cda[CTA_EXPECT_HELP_NAME]) {
2188                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2189
2190                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2191                                                     nf_ct_protonum(ct));
2192                 if (helper == NULL)
2193                         return -EOPNOTSUPP;
2194         }
2195
2196         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2197                                      helper, &tuple, &mask);
2198         if (IS_ERR(exp))
2199                 return PTR_ERR(exp);
2200
2201         err = nf_ct_expect_related_report(exp, portid, report);
2202         if (err < 0) {
2203                 nf_ct_expect_put(exp);
2204                 return err;
2205         }
2206
2207         return 0;
2208 }
2209
2210 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2211         .build_size     = ctnetlink_nfqueue_build_size,
2212         .build          = ctnetlink_nfqueue_build,
2213         .parse          = ctnetlink_nfqueue_parse,
2214         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2215         .seq_adjust     = nf_ct_tcp_seqadj_set,
2216 };
2217 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2218
2219 /***********************************************************************
2220  * EXPECT
2221  ***********************************************************************/
2222
2223 static inline int
2224 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2225                          const struct nf_conntrack_tuple *tuple,
2226                          enum ctattr_expect type)
2227 {
2228         struct nlattr *nest_parms;
2229
2230         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2231         if (!nest_parms)
2232                 goto nla_put_failure;
2233         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2234                 goto nla_put_failure;
2235         nla_nest_end(skb, nest_parms);
2236
2237         return 0;
2238
2239 nla_put_failure:
2240         return -1;
2241 }
2242
2243 static inline int
2244 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2245                         const struct nf_conntrack_tuple *tuple,
2246                         const struct nf_conntrack_tuple_mask *mask)
2247 {
2248         int ret;
2249         struct nf_conntrack_l3proto *l3proto;
2250         struct nf_conntrack_l4proto *l4proto;
2251         struct nf_conntrack_tuple m;
2252         struct nlattr *nest_parms;
2253
2254         memset(&m, 0xFF, sizeof(m));
2255         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2256         m.src.u.all = mask->src.u.all;
2257         m.dst.protonum = tuple->dst.protonum;
2258
2259         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2260         if (!nest_parms)
2261                 goto nla_put_failure;
2262
2263         rcu_read_lock();
2264         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2265         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2266         if (ret >= 0) {
2267                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2268                                                tuple->dst.protonum);
2269         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2270         }
2271         rcu_read_unlock();
2272
2273         if (unlikely(ret < 0))
2274                 goto nla_put_failure;
2275
2276         nla_nest_end(skb, nest_parms);
2277
2278         return 0;
2279
2280 nla_put_failure:
2281         return -1;
2282 }
2283
2284 static const union nf_inet_addr any_addr;
2285
2286 static int
2287 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2288                           const struct nf_conntrack_expect *exp)
2289 {
2290         struct nf_conn *master = exp->master;
2291         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2292         struct nf_conn_help *help;
2293 #ifdef CONFIG_NF_NAT_NEEDED
2294         struct nlattr *nest_parms;
2295         struct nf_conntrack_tuple nat_tuple = {};
2296 #endif
2297         struct nf_ct_helper_expectfn *expfn;
2298
2299         if (timeout < 0)
2300                 timeout = 0;
2301
2302         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2303                 goto nla_put_failure;
2304         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2305                 goto nla_put_failure;
2306         if (ctnetlink_exp_dump_tuple(skb,
2307                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2308                                  CTA_EXPECT_MASTER) < 0)
2309                 goto nla_put_failure;
2310
2311 #ifdef CONFIG_NF_NAT_NEEDED
2312         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2313             exp->saved_proto.all) {
2314                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2315                 if (!nest_parms)
2316                         goto nla_put_failure;
2317
2318                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2319                         goto nla_put_failure;
2320
2321                 nat_tuple.src.l3num = nf_ct_l3num(master);
2322                 nat_tuple.src.u3 = exp->saved_addr;
2323                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2324                 nat_tuple.src.u = exp->saved_proto;
2325
2326                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2327                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2328                         goto nla_put_failure;
2329                 nla_nest_end(skb, nest_parms);
2330         }
2331 #endif
2332         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2333             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2334             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2335             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2336                 goto nla_put_failure;
2337         help = nfct_help(master);
2338         if (help) {
2339                 struct nf_conntrack_helper *helper;
2340
2341                 helper = rcu_dereference(help->helper);
2342                 if (helper &&
2343                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2344                         goto nla_put_failure;
2345         }
2346         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2347         if (expfn != NULL &&
2348             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2349                 goto nla_put_failure;
2350
2351         return 0;
2352
2353 nla_put_failure:
2354         return -1;
2355 }
2356
2357 static int
2358 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2359                         int event, const struct nf_conntrack_expect *exp)
2360 {
2361         struct nlmsghdr *nlh;
2362         struct nfgenmsg *nfmsg;
2363         unsigned int flags = portid ? NLM_F_MULTI : 0;
2364
2365         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2366         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2367         if (nlh == NULL)
2368                 goto nlmsg_failure;
2369
2370         nfmsg = nlmsg_data(nlh);
2371         nfmsg->nfgen_family = exp->tuple.src.l3num;
2372         nfmsg->version      = NFNETLINK_V0;
2373         nfmsg->res_id       = 0;
2374
2375         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2376                 goto nla_put_failure;
2377
2378         nlmsg_end(skb, nlh);
2379         return skb->len;
2380
2381 nlmsg_failure:
2382 nla_put_failure:
2383         nlmsg_cancel(skb, nlh);
2384         return -1;
2385 }
2386
2387 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2388 static int
2389 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2390 {
2391         struct nf_conntrack_expect *exp = item->exp;
2392         struct net *net = nf_ct_exp_net(exp);
2393         struct nlmsghdr *nlh;
2394         struct nfgenmsg *nfmsg;
2395         struct sk_buff *skb;
2396         unsigned int type, group;
2397         int flags = 0;
2398
2399         if (events & (1 << IPEXP_DESTROY)) {
2400                 type = IPCTNL_MSG_EXP_DELETE;
2401                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2402         } else if (events & (1 << IPEXP_NEW)) {
2403                 type = IPCTNL_MSG_EXP_NEW;
2404                 flags = NLM_F_CREATE|NLM_F_EXCL;
2405                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2406         } else
2407                 return 0;
2408
2409         if (!item->report && !nfnetlink_has_listeners(net, group))
2410                 return 0;
2411
2412         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2413         if (skb == NULL)
2414                 goto errout;
2415
2416         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2417         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2418         if (nlh == NULL)
2419                 goto nlmsg_failure;
2420
2421         nfmsg = nlmsg_data(nlh);
2422         nfmsg->nfgen_family = exp->tuple.src.l3num;
2423         nfmsg->version      = NFNETLINK_V0;
2424         nfmsg->res_id       = 0;
2425
2426         rcu_read_lock();
2427         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2428                 goto nla_put_failure;
2429         rcu_read_unlock();
2430
2431         nlmsg_end(skb, nlh);
2432         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2433         return 0;
2434
2435 nla_put_failure:
2436         rcu_read_unlock();
2437         nlmsg_cancel(skb, nlh);
2438 nlmsg_failure:
2439         kfree_skb(skb);
2440 errout:
2441         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2442         return 0;
2443 }
2444 #endif
2445 static int ctnetlink_exp_done(struct netlink_callback *cb)
2446 {
2447         if (cb->args[1])
2448                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2449         return 0;
2450 }
2451
2452 static int
2453 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2454 {
2455         struct net *net = sock_net(skb->sk);
2456         struct nf_conntrack_expect *exp, *last;
2457         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2458         u_int8_t l3proto = nfmsg->nfgen_family;
2459
2460         rcu_read_lock();
2461         last = (struct nf_conntrack_expect *)cb->args[1];
2462         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2463 restart:
2464                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2465                                      hnode) {
2466                         if (l3proto && exp->tuple.src.l3num != l3proto)
2467                                 continue;
2468                         if (cb->args[1]) {
2469                                 if (exp != last)
2470                                         continue;
2471                                 cb->args[1] = 0;
2472                         }
2473                         if (ctnetlink_exp_fill_info(skb,
2474                                                     NETLINK_CB(cb->skb).portid,
2475                                                     cb->nlh->nlmsg_seq,
2476                                                     IPCTNL_MSG_EXP_NEW,
2477                                                     exp) < 0) {
2478                                 if (!atomic_inc_not_zero(&exp->use))
2479                                         continue;
2480                                 cb->args[1] = (unsigned long)exp;
2481                                 goto out;
2482                         }
2483                 }
2484                 if (cb->args[1]) {
2485                         cb->args[1] = 0;
2486                         goto restart;
2487                 }
2488         }
2489 out:
2490         rcu_read_unlock();
2491         if (last)
2492                 nf_ct_expect_put(last);
2493
2494         return skb->len;
2495 }
2496
2497 static int
2498 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2499 {
2500         struct nf_conntrack_expect *exp, *last;
2501         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2502         struct nf_conn *ct = cb->data;
2503         struct nf_conn_help *help = nfct_help(ct);
2504         u_int8_t l3proto = nfmsg->nfgen_family;
2505
2506         if (cb->args[0])
2507                 return 0;
2508
2509         rcu_read_lock();
2510         last = (struct nf_conntrack_expect *)cb->args[1];
2511 restart:
2512         hlist_for_each_entry(exp, &help->expectations, lnode) {
2513                 if (l3proto && exp->tuple.src.l3num != l3proto)
2514                         continue;
2515                 if (cb->args[1]) {
2516                         if (exp != last)
2517                                 continue;
2518                         cb->args[1] = 0;
2519                 }
2520                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2521                                             cb->nlh->nlmsg_seq,
2522                                             IPCTNL_MSG_EXP_NEW,
2523                                             exp) < 0) {
2524                         if (!atomic_inc_not_zero(&exp->use))
2525                                 continue;
2526                         cb->args[1] = (unsigned long)exp;
2527                         goto out;
2528                 }
2529         }
2530         if (cb->args[1]) {
2531                 cb->args[1] = 0;
2532                 goto restart;
2533         }
2534         cb->args[0] = 1;
2535 out:
2536         rcu_read_unlock();
2537         if (last)
2538                 nf_ct_expect_put(last);
2539
2540         return skb->len;
2541 }
2542
2543 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2544                                  const struct nlmsghdr *nlh,
2545                                  const struct nlattr * const cda[])
2546 {
2547         int err;
2548         struct net *net = sock_net(ctnl);
2549         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2550         u_int8_t u3 = nfmsg->nfgen_family;
2551         struct nf_conntrack_tuple tuple;
2552         struct nf_conntrack_tuple_hash *h;
2553         struct nf_conn *ct;
2554         u16 zone = 0;
2555         struct netlink_dump_control c = {
2556                 .dump = ctnetlink_exp_ct_dump_table,
2557                 .done = ctnetlink_exp_done,
2558         };
2559
2560         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2561         if (err < 0)
2562                 return err;
2563
2564         if (cda[CTA_EXPECT_ZONE]) {
2565                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2566                 if (err < 0)
2567                         return err;
2568         }
2569
2570         h = nf_conntrack_find_get(net, zone, &tuple);
2571         if (!h)
2572                 return -ENOENT;
2573
2574         ct = nf_ct_tuplehash_to_ctrack(h);
2575         c.data = ct;
2576
2577         err = netlink_dump_start(ctnl, skb, nlh, &c);
2578         nf_ct_put(ct);
2579
2580         return err;
2581 }
2582
2583 static int
2584 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2585                      const struct nlmsghdr *nlh,
2586                      const struct nlattr * const cda[])
2587 {
2588         struct net *net = sock_net(ctnl);
2589         struct nf_conntrack_tuple tuple;
2590         struct nf_conntrack_expect *exp;
2591         struct sk_buff *skb2;
2592         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2593         u_int8_t u3 = nfmsg->nfgen_family;
2594         u16 zone;
2595         int err;
2596
2597         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2598                 if (cda[CTA_EXPECT_MASTER])
2599                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2600                 else {
2601                         struct netlink_dump_control c = {
2602                                 .dump = ctnetlink_exp_dump_table,
2603                                 .done = ctnetlink_exp_done,
2604                         };
2605                         return netlink_dump_start(ctnl, skb, nlh, &c);
2606                 }
2607         }
2608
2609         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2610         if (err < 0)
2611                 return err;
2612
2613         if (cda[CTA_EXPECT_TUPLE])
2614                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2615         else if (cda[CTA_EXPECT_MASTER])
2616                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2617         else
2618                 return -EINVAL;
2619
2620         if (err < 0)
2621                 return err;
2622
2623         exp = nf_ct_expect_find_get(net, zone, &tuple);
2624         if (!exp)
2625                 return -ENOENT;
2626
2627         if (cda[CTA_EXPECT_ID]) {
2628                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2629                 if (ntohl(id) != (u32)(unsigned long)exp) {
2630                         nf_ct_expect_put(exp);
2631                         return -ENOENT;
2632                 }
2633         }
2634
2635         err = -ENOMEM;
2636         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2637         if (skb2 == NULL) {
2638                 nf_ct_expect_put(exp);
2639                 goto out;
2640         }
2641
2642         rcu_read_lock();
2643         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2644                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2645         rcu_read_unlock();
2646         nf_ct_expect_put(exp);
2647         if (err <= 0)
2648                 goto free;
2649
2650         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2651         if (err < 0)
2652                 goto out;
2653
2654         return 0;
2655
2656 free:
2657         kfree_skb(skb2);
2658 out:
2659         /* this avoids a loop in nfnetlink. */
2660         return err == -EAGAIN ? -ENOBUFS : err;
2661 }
2662
2663 static int
2664 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2665                      const struct nlmsghdr *nlh,
2666                      const struct nlattr * const cda[])
2667 {
2668         struct net *net = sock_net(ctnl);
2669         struct nf_conntrack_expect *exp;
2670         struct nf_conntrack_tuple tuple;
2671         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2672         struct hlist_node *next;
2673         u_int8_t u3 = nfmsg->nfgen_family;
2674         unsigned int i;
2675         u16 zone;
2676         int err;
2677
2678         if (cda[CTA_EXPECT_TUPLE]) {
2679                 /* delete a single expect by tuple */
2680                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2681                 if (err < 0)
2682                         return err;
2683
2684                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2685                 if (err < 0)
2686                         return err;
2687
2688                 /* bump usage count to 2 */
2689                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2690                 if (!exp)
2691                         return -ENOENT;
2692
2693                 if (cda[CTA_EXPECT_ID]) {
2694                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2695                         if (ntohl(id) != (u32)(unsigned long)exp) {
2696                                 nf_ct_expect_put(exp);
2697                                 return -ENOENT;
2698                         }
2699                 }
2700
2701                 /* after list removal, usage count == 1 */
2702                 spin_lock_bh(&nf_conntrack_lock);
2703                 if (del_timer(&exp->timeout)) {
2704                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2705                                                    nlmsg_report(nlh));
2706                         nf_ct_expect_put(exp);
2707                 }
2708                 spin_unlock_bh(&nf_conntrack_lock);
2709                 /* have to put what we 'get' above.
2710                  * after this line usage count == 0 */
2711                 nf_ct_expect_put(exp);
2712         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2713                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2714                 struct nf_conn_help *m_help;
2715
2716                 /* delete all expectations for this helper */
2717                 spin_lock_bh(&nf_conntrack_lock);
2718                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2719                         hlist_for_each_entry_safe(exp, next,
2720                                                   &net->ct.expect_hash[i],
2721                                                   hnode) {
2722                                 m_help = nfct_help(exp->master);
2723                                 if (!strcmp(m_help->helper->name, name) &&
2724                                     del_timer(&exp->timeout)) {
2725                                         nf_ct_unlink_expect_report(exp,
2726                                                         NETLINK_CB(skb).portid,
2727                                                         nlmsg_report(nlh));
2728                                         nf_ct_expect_put(exp);
2729                                 }
2730                         }
2731                 }
2732                 spin_unlock_bh(&nf_conntrack_lock);
2733         } else {
2734                 /* This basically means we have to flush everything*/
2735                 spin_lock_bh(&nf_conntrack_lock);
2736                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2737                         hlist_for_each_entry_safe(exp, next,
2738                                                   &net->ct.expect_hash[i],
2739                                                   hnode) {
2740                                 if (del_timer(&exp->timeout)) {
2741                                         nf_ct_unlink_expect_report(exp,
2742                                                         NETLINK_CB(skb).portid,
2743                                                         nlmsg_report(nlh));
2744                                         nf_ct_expect_put(exp);
2745                                 }
2746                         }
2747                 }
2748                 spin_unlock_bh(&nf_conntrack_lock);
2749         }
2750
2751         return 0;
2752 }
2753 static int
2754 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2755                         const struct nlattr * const cda[])
2756 {
2757         if (cda[CTA_EXPECT_TIMEOUT]) {
2758                 if (!del_timer(&x->timeout))
2759                         return -ETIME;
2760
2761                 x->timeout.expires = jiffies +
2762                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2763                 add_timer(&x->timeout);
2764         }
2765         return 0;
2766 }
2767
2768 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2769         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2770         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2771 };
2772
2773 static int
2774 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2775                            struct nf_conntrack_expect *exp,
2776                            u_int8_t u3)
2777 {
2778 #ifdef CONFIG_NF_NAT_NEEDED
2779         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2780         struct nf_conntrack_tuple nat_tuple = {};
2781         int err;
2782
2783         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2784         if (err < 0)
2785                 return err;
2786
2787         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2788                 return -EINVAL;
2789
2790         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2791                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2792         if (err < 0)
2793                 return err;
2794
2795         exp->saved_addr = nat_tuple.src.u3;
2796         exp->saved_proto = nat_tuple.src.u;
2797         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2798
2799         return 0;
2800 #else
2801         return -EOPNOTSUPP;
2802 #endif
2803 }
2804
2805 static struct nf_conntrack_expect *
2806 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2807                        struct nf_conntrack_helper *helper,
2808                        struct nf_conntrack_tuple *tuple,
2809                        struct nf_conntrack_tuple *mask)
2810 {
2811         u_int32_t class = 0;
2812         struct nf_conntrack_expect *exp;
2813         struct nf_conn_help *help;
2814         int err;
2815
2816         if (cda[CTA_EXPECT_CLASS] && helper) {
2817                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2818                 if (class > helper->expect_class_max)
2819                         return ERR_PTR(-EINVAL);
2820         }
2821         exp = nf_ct_expect_alloc(ct);
2822         if (!exp)
2823                 return ERR_PTR(-ENOMEM);
2824
2825         help = nfct_help(ct);
2826         if (!help) {
2827                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2828                         err = -EINVAL;
2829                         goto err_out;
2830                 }
2831                 exp->timeout.expires =
2832                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2833
2834                 exp->flags = NF_CT_EXPECT_USERSPACE;
2835                 if (cda[CTA_EXPECT_FLAGS]) {
2836                         exp->flags |=
2837                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2838                 }
2839         } else {
2840                 if (cda[CTA_EXPECT_FLAGS]) {
2841                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2842                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2843                 } else
2844                         exp->flags = 0;
2845         }
2846         if (cda[CTA_EXPECT_FN]) {
2847                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2848                 struct nf_ct_helper_expectfn *expfn;
2849
2850                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2851                 if (expfn == NULL) {
2852                         err = -EINVAL;
2853                         goto err_out;
2854                 }
2855                 exp->expectfn = expfn->expectfn;
2856         } else
2857                 exp->expectfn = NULL;
2858
2859         exp->class = class;
2860         exp->master = ct;
2861         exp->helper = helper;
2862         exp->tuple = *tuple;
2863         exp->mask.src.u3 = mask->src.u3;
2864         exp->mask.src.u.all = mask->src.u.all;
2865
2866         if (cda[CTA_EXPECT_NAT]) {
2867                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2868                                                  exp, nf_ct_l3num(ct));
2869                 if (err < 0)
2870                         goto err_out;
2871         }
2872         return exp;
2873 err_out:
2874         nf_ct_expect_put(exp);
2875         return ERR_PTR(err);
2876 }
2877
2878 static int
2879 ctnetlink_create_expect(struct net *net, u16 zone,
2880                         const struct nlattr * const cda[],
2881                         u_int8_t u3, u32 portid, int report)
2882 {
2883         struct nf_conntrack_tuple tuple, mask, master_tuple;
2884         struct nf_conntrack_tuple_hash *h = NULL;
2885         struct nf_conntrack_helper *helper = NULL;
2886         struct nf_conntrack_expect *exp;
2887         struct nf_conn *ct;
2888         int err;
2889
2890         /* caller guarantees that those three CTA_EXPECT_* exist */
2891         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2892         if (err < 0)
2893                 return err;
2894         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2895         if (err < 0)
2896                 return err;
2897         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2898         if (err < 0)
2899                 return err;
2900
2901         /* Look for master conntrack of this expectation */
2902         h = nf_conntrack_find_get(net, zone, &master_tuple);
2903         if (!h)
2904                 return -ENOENT;
2905         ct = nf_ct_tuplehash_to_ctrack(h);
2906
2907         if (cda[CTA_EXPECT_HELP_NAME]) {
2908                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2909
2910                 helper = __nf_conntrack_helper_find(helpname, u3,
2911                                                     nf_ct_protonum(ct));
2912                 if (helper == NULL) {
2913 #ifdef CONFIG_MODULES
2914                         if (request_module("nfct-helper-%s", helpname) < 0) {
2915                                 err = -EOPNOTSUPP;
2916                                 goto err_ct;
2917                         }
2918                         helper = __nf_conntrack_helper_find(helpname, u3,
2919                                                             nf_ct_protonum(ct));
2920                         if (helper) {
2921                                 err = -EAGAIN;
2922                                 goto err_ct;
2923                         }
2924 #endif
2925                         err = -EOPNOTSUPP;
2926                         goto err_ct;
2927                 }
2928         }
2929
2930         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
2931         if (IS_ERR(exp)) {
2932                 err = PTR_ERR(exp);
2933                 goto err_ct;
2934         }
2935
2936         err = nf_ct_expect_related_report(exp, portid, report);
2937         if (err < 0)
2938                 goto err_exp;
2939
2940         return 0;
2941 err_exp:
2942         nf_ct_expect_put(exp);
2943 err_ct:
2944         nf_ct_put(ct);
2945         return err;
2946 }
2947
2948 static int
2949 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2950                      const struct nlmsghdr *nlh,
2951                      const struct nlattr * const cda[])
2952 {
2953         struct net *net = sock_net(ctnl);
2954         struct nf_conntrack_tuple tuple;
2955         struct nf_conntrack_expect *exp;
2956         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2957         u_int8_t u3 = nfmsg->nfgen_family;
2958         u16 zone;
2959         int err;
2960
2961         if (!cda[CTA_EXPECT_TUPLE]
2962             || !cda[CTA_EXPECT_MASK]
2963             || !cda[CTA_EXPECT_MASTER])
2964                 return -EINVAL;
2965
2966         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2967         if (err < 0)
2968                 return err;
2969
2970         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2971         if (err < 0)
2972                 return err;
2973
2974         spin_lock_bh(&nf_conntrack_lock);
2975         exp = __nf_ct_expect_find(net, zone, &tuple);
2976
2977         if (!exp) {
2978                 spin_unlock_bh(&nf_conntrack_lock);
2979                 err = -ENOENT;
2980                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2981                         err = ctnetlink_create_expect(net, zone, cda,
2982                                                       u3,
2983                                                       NETLINK_CB(skb).portid,
2984                                                       nlmsg_report(nlh));
2985                 }
2986                 return err;
2987         }
2988
2989         err = -EEXIST;
2990         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2991                 err = ctnetlink_change_expect(exp, cda);
2992         spin_unlock_bh(&nf_conntrack_lock);
2993
2994         return err;
2995 }
2996
2997 static int
2998 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
2999                              const struct ip_conntrack_stat *st)
3000 {
3001         struct nlmsghdr *nlh;
3002         struct nfgenmsg *nfmsg;
3003         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3004
3005         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3006         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3007         if (nlh == NULL)
3008                 goto nlmsg_failure;
3009
3010         nfmsg = nlmsg_data(nlh);
3011         nfmsg->nfgen_family = AF_UNSPEC;
3012         nfmsg->version      = NFNETLINK_V0;
3013         nfmsg->res_id       = htons(cpu);
3014
3015         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3016             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3017             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3018                 goto nla_put_failure;
3019
3020         nlmsg_end(skb, nlh);
3021         return skb->len;
3022
3023 nla_put_failure:
3024 nlmsg_failure:
3025         nlmsg_cancel(skb, nlh);
3026         return -1;
3027 }
3028
3029 static int
3030 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3031 {
3032         int cpu;
3033         struct net *net = sock_net(skb->sk);
3034
3035         if (cb->args[0] == nr_cpu_ids)
3036                 return 0;
3037
3038         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3039                 const struct ip_conntrack_stat *st;
3040
3041                 if (!cpu_possible(cpu))
3042                         continue;
3043
3044                 st = per_cpu_ptr(net->ct.stat, cpu);
3045                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3046                                                  cb->nlh->nlmsg_seq,
3047                                                  cpu, st) < 0)
3048                         break;
3049         }
3050         cb->args[0] = cpu;
3051
3052         return skb->len;
3053 }
3054
3055 static int
3056 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3057                        const struct nlmsghdr *nlh,
3058                        const struct nlattr * const cda[])
3059 {
3060         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3061                 struct netlink_dump_control c = {
3062                         .dump = ctnetlink_exp_stat_cpu_dump,
3063                 };
3064                 return netlink_dump_start(ctnl, skb, nlh, &c);
3065         }
3066
3067         return 0;
3068 }
3069
3070 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3071 static struct nf_ct_event_notifier ctnl_notifier = {
3072         .fcn = ctnetlink_conntrack_event,
3073 };
3074
3075 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3076         .fcn = ctnetlink_expect_event,
3077 };
3078 #endif
3079
3080 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3081         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3082                                             .attr_count = CTA_MAX,
3083                                             .policy = ct_nla_policy },
3084         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3085                                             .attr_count = CTA_MAX,
3086                                             .policy = ct_nla_policy },
3087         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3088                                             .attr_count = CTA_MAX,
3089                                             .policy = ct_nla_policy },
3090         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3091                                             .attr_count = CTA_MAX,
3092                                             .policy = ct_nla_policy },
3093         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3094         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3095         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3096         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3097 };
3098
3099 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3100         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3101                                             .attr_count = CTA_EXPECT_MAX,
3102                                             .policy = exp_nla_policy },
3103         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3104                                             .attr_count = CTA_EXPECT_MAX,
3105                                             .policy = exp_nla_policy },
3106         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3107                                             .attr_count = CTA_EXPECT_MAX,
3108                                             .policy = exp_nla_policy },
3109         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3110 };
3111
3112 static const struct nfnetlink_subsystem ctnl_subsys = {
3113         .name                           = "conntrack",
3114         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3115         .cb_count                       = IPCTNL_MSG_MAX,
3116         .cb                             = ctnl_cb,
3117 };
3118
3119 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3120         .name                           = "conntrack_expect",
3121         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3122         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3123         .cb                             = ctnl_exp_cb,
3124 };
3125
3126 MODULE_ALIAS("ip_conntrack_netlink");
3127 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3128 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3129
3130 static int __net_init ctnetlink_net_init(struct net *net)
3131 {
3132 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3133         int ret;
3134
3135         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3136         if (ret < 0) {
3137                 pr_err("ctnetlink_init: cannot register notifier.\n");
3138                 goto err_out;
3139         }
3140
3141         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3142         if (ret < 0) {
3143                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3144                 goto err_unreg_notifier;
3145         }
3146 #endif
3147         return 0;
3148
3149 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3150 err_unreg_notifier:
3151         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3152 err_out:
3153         return ret;
3154 #endif
3155 }
3156
3157 static void ctnetlink_net_exit(struct net *net)
3158 {
3159 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3160         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3161         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3162 #endif
3163 }
3164
3165 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3166 {
3167         struct net *net;
3168
3169         list_for_each_entry(net, net_exit_list, exit_list)
3170                 ctnetlink_net_exit(net);
3171 }
3172
3173 static struct pernet_operations ctnetlink_net_ops = {
3174         .init           = ctnetlink_net_init,
3175         .exit_batch     = ctnetlink_net_exit_batch,
3176 };
3177
3178 static int __init ctnetlink_init(void)
3179 {
3180         int ret;
3181
3182         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3183         ret = nfnetlink_subsys_register(&ctnl_subsys);
3184         if (ret < 0) {
3185                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3186                 goto err_out;
3187         }
3188
3189         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3190         if (ret < 0) {
3191                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3192                 goto err_unreg_subsys;
3193         }
3194
3195         ret = register_pernet_subsys(&ctnetlink_net_ops);
3196         if (ret < 0) {
3197                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3198                 goto err_unreg_exp_subsys;
3199         }
3200 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3201         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3202         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3203 #endif
3204         return 0;
3205
3206 err_unreg_exp_subsys:
3207         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3208 err_unreg_subsys:
3209         nfnetlink_subsys_unregister(&ctnl_subsys);
3210 err_out:
3211         return ret;
3212 }
3213
3214 static void __exit ctnetlink_exit(void)
3215 {
3216         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3217
3218         unregister_pernet_subsys(&ctnetlink_net_ops);
3219         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3220         nfnetlink_subsys_unregister(&ctnl_subsys);
3221 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3222         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3223 #endif
3224 }
3225
3226 module_init(ctnetlink_init);
3227 module_exit(ctnetlink_exit);