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) 2005-2011 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_l3proto.h>
  41 #include <net/netfilter/nf_conntrack_l4proto.h>
  42 #include <net/netfilter/nf_conntrack_tuple.h>
  43 #include <net/netfilter/nf_conntrack_acct.h>
  44 #include <net/netfilter/nf_conntrack_zones.h>
  45 #include <net/netfilter/nf_conntrack_timestamp.h>
  46 #ifdef CONFIG_NF_NAT_NEEDED
  47 #include <net/netfilter/nf_nat_core.h>
  48 #include <net/netfilter/nf_nat_protocol.h>
  49 #endif
  50
  51 #include <linux/netfilter/nfnetlink.h>
  52 #include <linux/netfilter/nfnetlink_conntrack.h>
  53
  54 MODULE_LICENSE("GPL");
  55
  56 static char __initdata version[] = "0.93";
  57
  58 static inline int
  59 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  60                             const struct nf_conntrack_tuple *tuple,
  61                             struct nf_conntrack_l4proto *l4proto)
  62 {
  63         int ret = 0;
  64         struct nlattr *nest_parms;
  65
  66         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  67         if (!nest_parms)
  68                 goto nla_put_failure;
  69         NLA_PUT_U8(skb, CTA_PROTO_NUM, tuple->dst.protonum);
  70
  71         if (likely(l4proto->tuple_to_nlattr))
  72                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  73
  74         nla_nest_end(skb, nest_parms);
  75
  76         return ret;
  77
  78 nla_put_failure:
  79         return -1;
  80 }
  81
  82 static inline int
  83 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  84                          const struct nf_conntrack_tuple *tuple,
  85                          struct nf_conntrack_l3proto *l3proto)
  86 {
  87         int ret = 0;
  88         struct nlattr *nest_parms;
  89
  90         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  91         if (!nest_parms)
  92                 goto nla_put_failure;
  93
  94         if (likely(l3proto->tuple_to_nlattr))
  95                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  96
  97         nla_nest_end(skb, nest_parms);
  98
  99         return ret;
 100
 101 nla_put_failure:
 102         return -1;
 103 }
 104
 105 static int
 106 ctnetlink_dump_tuples(struct sk_buff *skb,
 107                       const struct nf_conntrack_tuple *tuple)
 108 {
 109         int ret;
 110         struct nf_conntrack_l3proto *l3proto;
 111         struct nf_conntrack_l4proto *l4proto;
 112
 113         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 114         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 115
 116         if (unlikely(ret < 0))
 117                 return ret;
 118
 119         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 120         ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 121
 122         return ret;
 123 }
 124
 125 static inline int
 126 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 127 {
 128         NLA_PUT_BE32(skb, CTA_STATUS, htonl(ct->status));
 129         return 0;
 130
 131 nla_put_failure:
 132         return -1;
 133 }
 134
 135 static inline int
 136 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 137 {
 138         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 139
 140         if (timeout < 0)
 141                 timeout = 0;
 142
 143         NLA_PUT_BE32(skb, CTA_TIMEOUT, htonl(timeout));
 144         return 0;
 145
 146 nla_put_failure:
 147         return -1;
 148 }
 149
 150 static inline int
 151 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 152 {
 153         struct nf_conntrack_l4proto *l4proto;
 154         struct nlattr *nest_proto;
 155         int ret;
 156
 157         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 158         if (!l4proto->to_nlattr)
 159                 return 0;
 160
 161         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 162         if (!nest_proto)
 163                 goto nla_put_failure;
 164
 165         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 166
 167         nla_nest_end(skb, nest_proto);
 168
 169         return ret;
 170
 171 nla_put_failure:
 172         return -1;
 173 }
 174
 175 static inline int
 176 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 177 {
 178         struct nlattr *nest_helper;
 179         const struct nf_conn_help *help = nfct_help(ct);
 180         struct nf_conntrack_helper *helper;
 181
 182         if (!help)
 183                 return 0;
 184
 185         helper = rcu_dereference(help->helper);
 186         if (!helper)
 187                 goto out;
 188
 189         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 190         if (!nest_helper)
 191                 goto nla_put_failure;
 192         NLA_PUT_STRING(skb, CTA_HELP_NAME, helper->name);
 193
 194         if (helper->to_nlattr)
 195                 helper->to_nlattr(skb, ct);
 196
 197         nla_nest_end(skb, nest_helper);
 198 out:
 199         return 0;
 200
 201 nla_put_failure:
 202         return -1;
 203 }
 204
 205 static int
 206 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 207                         enum ip_conntrack_dir dir)
 208 {
 209         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 210         struct nlattr *nest_count;
 211         const struct nf_conn_counter *acct;
 212
 213         acct = nf_conn_acct_find(ct);
 214         if (!acct)
 215                 return 0;
 216
 217         nest_count = nla_nest_start(skb, type | NLA_F_NESTED);
 218         if (!nest_count)
 219                 goto nla_put_failure;
 220
 221         NLA_PUT_BE64(skb, CTA_COUNTERS_PACKETS,
 222                      cpu_to_be64(acct[dir].packets));
 223         NLA_PUT_BE64(skb, CTA_COUNTERS_BYTES,
 224                      cpu_to_be64(acct[dir].bytes));
 225
 226         nla_nest_end(skb, nest_count);
 227
 228         return 0;
 229
 230 nla_put_failure:
 231         return -1;
 232 }
 233
 234 static int
 235 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 236 {
 237         struct nlattr *nest_count;
 238         const struct nf_conn_tstamp *tstamp;
 239
 240         tstamp = nf_conn_tstamp_find(ct);
 241         if (!tstamp)
 242                 return 0;
 243
 244         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 245         if (!nest_count)
 246                 goto nla_put_failure;
 247
 248         NLA_PUT_BE64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start));
 249         if (tstamp->stop != 0) {
 250                 NLA_PUT_BE64(skb, CTA_TIMESTAMP_STOP,
 251                              cpu_to_be64(tstamp->stop));
 252         }
 253         nla_nest_end(skb, nest_count);
 254
 255         return 0;
 256
 257 nla_put_failure:
 258         return -1;
 259 }
 260
 261 #ifdef CONFIG_NF_CONNTRACK_MARK
 262 static inline int
 263 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 264 {
 265         NLA_PUT_BE32(skb, CTA_MARK, htonl(ct->mark));
 266         return 0;
 267
 268 nla_put_failure:
 269         return -1;
 270 }
 271 #else
 272 #define ctnetlink_dump_mark(a, b) (0)
 273 #endif
 274
 275 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 276 static inline int
 277 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 278 {
 279         struct nlattr *nest_secctx;
 280         int len, ret;
 281         char *secctx;
 282
 283         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 284         if (ret)
 285                 return 0;
 286
 287         ret = -1;
 288         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 289         if (!nest_secctx)
 290                 goto nla_put_failure;
 291
 292         NLA_PUT_STRING(skb, CTA_SECCTX_NAME, secctx);
 293         nla_nest_end(skb, nest_secctx);
 294
 295         ret = 0;
 296 nla_put_failure:
 297         security_release_secctx(secctx, len);
 298         return ret;
 299 }
 300 #else
 301 #define ctnetlink_dump_secctx(a, b) (0)
 302 #endif
 303
 304 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 305
 306 static inline int
 307 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 308 {
 309         struct nlattr *nest_parms;
 310
 311         if (!(ct->status & IPS_EXPECTED))
 312                 return 0;
 313
 314         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 315         if (!nest_parms)
 316                 goto nla_put_failure;
 317         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 318                 goto nla_put_failure;
 319         nla_nest_end(skb, nest_parms);
 320
 321         return 0;
 322
 323 nla_put_failure:
 324         return -1;
 325 }
 326
 327 #ifdef CONFIG_NF_NAT_NEEDED
 328 static int
 329 dump_nat_seq_adj(struct sk_buff *skb, const struct nf_nat_seq *natseq, int type)
 330 {
 331         struct nlattr *nest_parms;
 332
 333         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 334         if (!nest_parms)
 335                 goto nla_put_failure;
 336
 337         NLA_PUT_BE32(skb, CTA_NAT_SEQ_CORRECTION_POS,
 338                      htonl(natseq->correction_pos));
 339         NLA_PUT_BE32(skb, CTA_NAT_SEQ_OFFSET_BEFORE,
 340                      htonl(natseq->offset_before));
 341         NLA_PUT_BE32(skb, CTA_NAT_SEQ_OFFSET_AFTER,
 342                      htonl(natseq->offset_after));
 343
 344         nla_nest_end(skb, nest_parms);
 345
 346         return 0;
 347
 348 nla_put_failure:
 349         return -1;
 350 }
 351
 352 static inline int
 353 ctnetlink_dump_nat_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 354 {
 355         struct nf_nat_seq *natseq;
 356         struct nf_conn_nat *nat = nfct_nat(ct);
 357
 358         if (!(ct->status & IPS_SEQ_ADJUST) || !nat)
 359                 return 0;
 360
 361         natseq = &nat->seq[IP_CT_DIR_ORIGINAL];
 362         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_ORIG) == -1)
 363                 return -1;
 364
 365         natseq = &nat->seq[IP_CT_DIR_REPLY];
 366         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_REPLY) == -1)
 367                 return -1;
 368
 369         return 0;
 370 }
 371 #else
 372 #define ctnetlink_dump_nat_seq_adj(a, b) (0)
 373 #endif
 374
 375 static inline int
 376 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 377 {
 378         NLA_PUT_BE32(skb, CTA_ID, htonl((unsigned long)ct));
 379         return 0;
 380
 381 nla_put_failure:
 382         return -1;
 383 }
 384
 385 static inline int
 386 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 387 {
 388         NLA_PUT_BE32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use)));
 389         return 0;
 390
 391 nla_put_failure:
 392         return -1;
 393 }
 394
 395 static int
 396 ctnetlink_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
 397                     int event, struct nf_conn *ct)
 398 {
 399         struct nlmsghdr *nlh;
 400         struct nfgenmsg *nfmsg;
 401         struct nlattr *nest_parms;
 402         unsigned int flags = pid ? NLM_F_MULTI : 0;
 403
 404         event |= NFNL_SUBSYS_CTNETLINK << 8;
 405         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
 406         if (nlh == NULL)
 407                 goto nlmsg_failure;
 408
 409         nfmsg = nlmsg_data(nlh);
 410         nfmsg->nfgen_family = nf_ct_l3num(ct);
 411         nfmsg->version      = NFNETLINK_V0;
 412         nfmsg->res_id       = 0;
 413
 414         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 415         if (!nest_parms)
 416                 goto nla_put_failure;
 417         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 418                 goto nla_put_failure;
 419         nla_nest_end(skb, nest_parms);
 420
 421         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 422         if (!nest_parms)
 423                 goto nla_put_failure;
 424         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 425                 goto nla_put_failure;
 426         nla_nest_end(skb, nest_parms);
 427
 428         if (nf_ct_zone(ct))
 429                 NLA_PUT_BE16(skb, CTA_ZONE, htons(nf_ct_zone(ct)));
 430
 431         if (ctnetlink_dump_status(skb, ct) < 0 ||
 432             ctnetlink_dump_timeout(skb, ct) < 0 ||
 433             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 434             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 435             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 436             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 437             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 438             ctnetlink_dump_mark(skb, ct) < 0 ||
 439             ctnetlink_dump_secctx(skb, ct) < 0 ||
 440             ctnetlink_dump_id(skb, ct) < 0 ||
 441             ctnetlink_dump_use(skb, ct) < 0 ||
 442             ctnetlink_dump_master(skb, ct) < 0 ||
 443             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 444                 goto nla_put_failure;
 445
 446         nlmsg_end(skb, nlh);
 447         return skb->len;
 448
 449 nlmsg_failure:
 450 nla_put_failure:
 451         nlmsg_cancel(skb, nlh);
 452         return -1;
 453 }
 454
 455 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 456 static inline size_t
 457 ctnetlink_proto_size(const struct nf_conn *ct)
 458 {
 459         struct nf_conntrack_l3proto *l3proto;
 460         struct nf_conntrack_l4proto *l4proto;
 461         size_t len = 0;
 462
 463         rcu_read_lock();
 464         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 465         len += l3proto->nla_size;
 466
 467         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 468         len += l4proto->nla_size;
 469         rcu_read_unlock();
 470
 471         return len;
 472 }
 473
 474 static inline size_t
 475 ctnetlink_counters_size(const struct nf_conn *ct)
 476 {
 477         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 478                 return 0;
 479         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 480                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 481                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 482                ;
 483 }
 484
 485 static inline int
 486 ctnetlink_secctx_size(const struct nf_conn *ct)
 487 {
 488 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 489         int len, ret;
 490
 491         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 492         if (ret)
 493                 return 0;
 494
 495         return nla_total_size(0) /* CTA_SECCTX */
 496                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 497 #else
 498         return 0;
 499 #endif
 500 }
 501
 502 static inline size_t
 503 ctnetlink_timestamp_size(const struct nf_conn *ct)
 504 {
 505 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 506         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 507                 return 0;
 508         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 509 #else
 510         return 0;
 511 #endif
 512 }
 513
 514 static inline size_t
 515 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 516 {
 517         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 518                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 519                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 520                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 521                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 522                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 523                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 524                + ctnetlink_counters_size(ct)
 525                + ctnetlink_timestamp_size(ct)
 526                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 527                + nla_total_size(0) /* CTA_PROTOINFO */
 528                + nla_total_size(0) /* CTA_HELP */
 529                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 530                + ctnetlink_secctx_size(ct)
 531 #ifdef CONFIG_NF_NAT_NEEDED
 532                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 533                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 534 #endif
 535 #ifdef CONFIG_NF_CONNTRACK_MARK
 536                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 537 #endif
 538                + ctnetlink_proto_size(ct)
 539                ;
 540 }
 541
 542 static int
 543 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 544 {
 545         struct net *net;
 546         struct nlmsghdr *nlh;
 547         struct nfgenmsg *nfmsg;
 548         struct nlattr *nest_parms;
 549         struct nf_conn *ct = item->ct;
 550         struct sk_buff *skb;
 551         unsigned int type;
 552         unsigned int flags = 0, group;
 553         int err;
 554
 555         /* ignore our fake conntrack entry */
 556         if (nf_ct_is_untracked(ct))
 557                 return 0;
 558
 559         if (events & (1 << IPCT_DESTROY)) {
 560                 type = IPCTNL_MSG_CT_DELETE;
 561                 group = NFNLGRP_CONNTRACK_DESTROY;
 562         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 563                 type = IPCTNL_MSG_CT_NEW;
 564                 flags = NLM_F_CREATE|NLM_F_EXCL;
 565                 group = NFNLGRP_CONNTRACK_NEW;
 566         } else  if (events) {
 567                 type = IPCTNL_MSG_CT_NEW;
 568                 group = NFNLGRP_CONNTRACK_UPDATE;
 569         } else
 570                 return 0;
 571
 572         net = nf_ct_net(ct);
 573         if (!item->report && !nfnetlink_has_listeners(net, group))
 574                 return 0;
 575
 576         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 577         if (skb == NULL)
 578                 goto errout;
 579
 580         type |= NFNL_SUBSYS_CTNETLINK << 8;
 581         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
 582         if (nlh == NULL)
 583                 goto nlmsg_failure;
 584
 585         nfmsg = nlmsg_data(nlh);
 586         nfmsg->nfgen_family = nf_ct_l3num(ct);
 587         nfmsg->version  = NFNETLINK_V0;
 588         nfmsg->res_id   = 0;
 589
 590         rcu_read_lock();
 591         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 592         if (!nest_parms)
 593                 goto nla_put_failure;
 594         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 595                 goto nla_put_failure;
 596         nla_nest_end(skb, nest_parms);
 597
 598         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 599         if (!nest_parms)
 600                 goto nla_put_failure;
 601         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 602                 goto nla_put_failure;
 603         nla_nest_end(skb, nest_parms);
 604
 605         if (nf_ct_zone(ct))
 606                 NLA_PUT_BE16(skb, CTA_ZONE, htons(nf_ct_zone(ct)));
 607
 608         if (ctnetlink_dump_id(skb, ct) < 0)
 609                 goto nla_put_failure;
 610
 611         if (ctnetlink_dump_status(skb, ct) < 0)
 612                 goto nla_put_failure;
 613
 614         if (events & (1 << IPCT_DESTROY)) {
 615                 if (ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 616                     ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 617                     ctnetlink_dump_timestamp(skb, ct) < 0)
 618                         goto nla_put_failure;
 619         } else {
 620                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 621                         goto nla_put_failure;
 622
 623                 if (events & (1 << IPCT_PROTOINFO)
 624                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 625                         goto nla_put_failure;
 626
 627                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 628                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 629                         goto nla_put_failure;
 630
 631 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 632                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 633                     && ctnetlink_dump_secctx(skb, ct) < 0)
 634                         goto nla_put_failure;
 635 #endif
 636
 637                 if (events & (1 << IPCT_RELATED) &&
 638                     ctnetlink_dump_master(skb, ct) < 0)
 639                         goto nla_put_failure;
 640
 641                 if (events & (1 << IPCT_NATSEQADJ) &&
 642                     ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 643                         goto nla_put_failure;
 644         }
 645
 646 #ifdef CONFIG_NF_CONNTRACK_MARK
 647         if ((events & (1 << IPCT_MARK) || ct->mark)
 648             && ctnetlink_dump_mark(skb, ct) < 0)
 649                 goto nla_put_failure;
 650 #endif
 651         rcu_read_unlock();
 652
 653         nlmsg_end(skb, nlh);
 654         err = nfnetlink_send(skb, net, item->pid, group, item->report,
 655                              GFP_ATOMIC);
 656         if (err == -ENOBUFS || err == -EAGAIN)
 657                 return -ENOBUFS;
 658
 659         return 0;
 660
 661 nla_put_failure:
 662         rcu_read_unlock();
 663         nlmsg_cancel(skb, nlh);
 664 nlmsg_failure:
 665         kfree_skb(skb);
 666 errout:
 667         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 668                 return -ENOBUFS;
 669
 670         return 0;
 671 }
 672 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 673
 674 static int ctnetlink_done(struct netlink_callback *cb)
 675 {
 676         if (cb->args[1])
 677                 nf_ct_put((struct nf_conn *)cb->args[1]);
 678         return 0;
 679 }
 680
 681 static int
 682 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 683 {
 684         struct net *net = sock_net(skb->sk);
 685         struct nf_conn *ct, *last;
 686         struct nf_conntrack_tuple_hash *h;
 687         struct hlist_nulls_node *n;
 688         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 689         u_int8_t l3proto = nfmsg->nfgen_family;
 690
 691         spin_lock_bh(&nf_conntrack_lock);
 692         last = (struct nf_conn *)cb->args[1];
 693         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 694 restart:
 695                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 696                                          hnnode) {
 697                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 698                                 continue;
 699                         ct = nf_ct_tuplehash_to_ctrack(h);
 700                         /* Dump entries of a given L3 protocol number.
 701                          * If it is not specified, ie. l3proto == 0,
 702                          * then dump everything. */
 703                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 704                                 continue;
 705                         if (cb->args[1]) {
 706                                 if (ct != last)
 707                                         continue;
 708                                 cb->args[1] = 0;
 709                         }
 710                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 711                                                 cb->nlh->nlmsg_seq,
 712                                                 IPCTNL_MSG_CT_NEW, ct) < 0) {
 713                                 nf_conntrack_get(&ct->ct_general);
 714                                 cb->args[1] = (unsigned long)ct;
 715                                 goto out;
 716                         }
 717
 718                         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) ==
 719                                                 IPCTNL_MSG_CT_GET_CTRZERO) {
 720                                 struct nf_conn_counter *acct;
 721
 722                                 acct = nf_conn_acct_find(ct);
 723                                 if (acct)
 724                                         memset(acct, 0, sizeof(struct nf_conn_counter[IP_CT_DIR_MAX]));
 725                         }
 726                 }
 727                 if (cb->args[1]) {
 728                         cb->args[1] = 0;
 729                         goto restart;
 730                 }
 731         }
 732 out:
 733         spin_unlock_bh(&nf_conntrack_lock);
 734         if (last) {
 735                 /* nf ct hash resize happened, now clear the leftover. */
 736                 if ((struct nf_conn *)cb->args[1] == last)
 737                         cb->args[1] = 0;
 738
 739                 nf_ct_put(last);
 740         }
 741
 742         return skb->len;
 743 }
 744
 745 static inline int
 746 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 747 {
 748         struct nlattr *tb[CTA_IP_MAX+1];
 749         struct nf_conntrack_l3proto *l3proto;
 750         int ret = 0;
 751
 752         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 753
 754         rcu_read_lock();
 755         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 756
 757         if (likely(l3proto->nlattr_to_tuple)) {
 758                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 759                                           l3proto->nla_policy);
 760                 if (ret == 0)
 761                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 762         }
 763
 764         rcu_read_unlock();
 765
 766         return ret;
 767 }
 768
 769 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 770         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 771 };
 772
 773 static inline int
 774 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 775                             struct nf_conntrack_tuple *tuple)
 776 {
 777         struct nlattr *tb[CTA_PROTO_MAX+1];
 778         struct nf_conntrack_l4proto *l4proto;
 779         int ret = 0;
 780
 781         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 782         if (ret < 0)
 783                 return ret;
 784
 785         if (!tb[CTA_PROTO_NUM])
 786                 return -EINVAL;
 787         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 788
 789         rcu_read_lock();
 790         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 791
 792         if (likely(l4proto->nlattr_to_tuple)) {
 793                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 794                                           l4proto->nla_policy);
 795                 if (ret == 0)
 796                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 797         }
 798
 799         rcu_read_unlock();
 800
 801         return ret;
 802 }
 803
 804 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 805         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 806         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 807 };
 808
 809 static int
 810 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 811                       struct nf_conntrack_tuple *tuple,
 812                       enum ctattr_type type, u_int8_t l3num)
 813 {
 814         struct nlattr *tb[CTA_TUPLE_MAX+1];
 815         int err;
 816
 817         memset(tuple, 0, sizeof(*tuple));
 818
 819         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 820
 821         if (!tb[CTA_TUPLE_IP])
 822                 return -EINVAL;
 823
 824         tuple->src.l3num = l3num;
 825
 826         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 827         if (err < 0)
 828                 return err;
 829
 830         if (!tb[CTA_TUPLE_PROTO])
 831                 return -EINVAL;
 832
 833         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 834         if (err < 0)
 835                 return err;
 836
 837         /* orig and expect tuples get DIR_ORIGINAL */
 838         if (type == CTA_TUPLE_REPLY)
 839                 tuple->dst.dir = IP_CT_DIR_REPLY;
 840         else
 841                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 842
 843         return 0;
 844 }
 845
 846 static int
 847 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 848 {
 849         if (attr)
 850 #ifdef CONFIG_NF_CONNTRACK_ZONES
 851                 *zone = ntohs(nla_get_be16(attr));
 852 #else
 853                 return -EOPNOTSUPP;
 854 #endif
 855         else
 856                 *zone = 0;
 857
 858         return 0;
 859 }
 860
 861 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 862         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING },
 863 };
 864
 865 static inline int
 866 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name)
 867 {
 868         struct nlattr *tb[CTA_HELP_MAX+1];
 869
 870         nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 871
 872         if (!tb[CTA_HELP_NAME])
 873                 return -EINVAL;
 874
 875         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 876
 877         return 0;
 878 }
 879
 880 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 881         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 882         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 883         [CTA_STATUS]            = { .type = NLA_U32 },
 884         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 885         [CTA_HELP]              = { .type = NLA_NESTED },
 886         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 887         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 888         [CTA_MARK]              = { .type = NLA_U32 },
 889         [CTA_ID]                = { .type = NLA_U32 },
 890         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 891         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 892         [CTA_ZONE]              = { .type = NLA_U16 },
 893 };
 894
 895 static int
 896 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 897                         const struct nlmsghdr *nlh,
 898                         const struct nlattr * const cda[])
 899 {
 900         struct net *net = sock_net(ctnl);
 901         struct nf_conntrack_tuple_hash *h;
 902         struct nf_conntrack_tuple tuple;
 903         struct nf_conn *ct;
 904         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 905         u_int8_t u3 = nfmsg->nfgen_family;
 906         u16 zone;
 907         int err;
 908
 909         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 910         if (err < 0)
 911                 return err;
 912
 913         if (cda[CTA_TUPLE_ORIG])
 914                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 915         else if (cda[CTA_TUPLE_REPLY])
 916                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 917         else {
 918                 /* Flush the whole table */
 919                 nf_conntrack_flush_report(net,
 920                                          NETLINK_CB(skb).pid,
 921                                          nlmsg_report(nlh));
 922                 return 0;
 923         }
 924
 925         if (err < 0)
 926                 return err;
 927
 928         h = nf_conntrack_find_get(net, zone, &tuple);
 929         if (!h)
 930                 return -ENOENT;
 931
 932         ct = nf_ct_tuplehash_to_ctrack(h);
 933
 934         if (cda[CTA_ID]) {
 935                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
 936                 if (id != (u32)(unsigned long)ct) {
 937                         nf_ct_put(ct);
 938                         return -ENOENT;
 939                 }
 940         }
 941
 942         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 943                                       NETLINK_CB(skb).pid,
 944                                       nlmsg_report(nlh)) < 0) {
 945                 nf_ct_delete_from_lists(ct);
 946                 /* we failed to report the event, try later */
 947                 nf_ct_insert_dying_list(ct);
 948                 nf_ct_put(ct);
 949                 return 0;
 950         }
 951
 952         /* death_by_timeout would report the event again */
 953         set_bit(IPS_DYING_BIT, &ct->status);
 954
 955         nf_ct_kill(ct);
 956         nf_ct_put(ct);
 957
 958         return 0;
 959 }
 960
 961 static int
 962 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 963                         const struct nlmsghdr *nlh,
 964                         const struct nlattr * const cda[])
 965 {
 966         struct net *net = sock_net(ctnl);
 967         struct nf_conntrack_tuple_hash *h;
 968         struct nf_conntrack_tuple tuple;
 969         struct nf_conn *ct;
 970         struct sk_buff *skb2 = NULL;
 971         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 972         u_int8_t u3 = nfmsg->nfgen_family;
 973         u16 zone;
 974         int err;
 975
 976         if (nlh->nlmsg_flags & NLM_F_DUMP)
 977                 return netlink_dump_start(ctnl, skb, nlh, ctnetlink_dump_table,
 978                                           ctnetlink_done, 0);
 979
 980         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 981         if (err < 0)
 982                 return err;
 983
 984         if (cda[CTA_TUPLE_ORIG])
 985                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 986         else if (cda[CTA_TUPLE_REPLY])
 987                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 988         else
 989                 return -EINVAL;
 990
 991         if (err < 0)
 992                 return err;
 993
 994         h = nf_conntrack_find_get(net, zone, &tuple);
 995         if (!h)
 996                 return -ENOENT;
 997
 998         ct = nf_ct_tuplehash_to_ctrack(h);
 999
1000         err = -ENOMEM;
1001         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1002         if (skb2 == NULL) {
1003                 nf_ct_put(ct);
1004                 return -ENOMEM;
1005         }
1006
1007         rcu_read_lock();
1008         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
1009                                   IPCTNL_MSG_CT_NEW, ct);
1010         rcu_read_unlock();
1011         nf_ct_put(ct);
1012         if (err <= 0)
1013                 goto free;
1014
1015         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1016         if (err < 0)
1017                 goto out;
1018
1019         return 0;
1020
1021 free:
1022         kfree_skb(skb2);
1023 out:
1024         /* this avoids a loop in nfnetlink. */
1025         return err == -EAGAIN ? -ENOBUFS : err;
1026 }
1027
1028 #ifdef CONFIG_NF_NAT_NEEDED
1029 static int
1030 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1031                           enum nf_nat_manip_type manip,
1032                           const struct nlattr *attr)
1033 {
1034         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1035
1036         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1037         if (!parse_nat_setup) {
1038 #ifdef CONFIG_MODULES
1039                 rcu_read_unlock();
1040                 spin_unlock_bh(&nf_conntrack_lock);
1041                 nfnl_unlock();
1042                 if (request_module("nf-nat-ipv4") < 0) {
1043                         nfnl_lock();
1044                         spin_lock_bh(&nf_conntrack_lock);
1045                         rcu_read_lock();
1046                         return -EOPNOTSUPP;
1047                 }
1048                 nfnl_lock();
1049                 spin_lock_bh(&nf_conntrack_lock);
1050                 rcu_read_lock();
1051                 if (nfnetlink_parse_nat_setup_hook)
1052                         return -EAGAIN;
1053 #endif
1054                 return -EOPNOTSUPP;
1055         }
1056
1057         return parse_nat_setup(ct, manip, attr);
1058 }
1059 #endif
1060
1061 static void
1062 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1063                           unsigned long off)
1064 {
1065         unsigned int bit;
1066
1067         /* Ignore these unchangable bits */
1068         on &= ~IPS_UNCHANGEABLE_MASK;
1069         off &= ~IPS_UNCHANGEABLE_MASK;
1070
1071         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1072                 if (on & (1 << bit))
1073                         set_bit(bit, &ct->status);
1074                 else if (off & (1 << bit))
1075                         clear_bit(bit, &ct->status);
1076         }
1077 }
1078
1079 static int
1080 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1081 {
1082         unsigned long d;
1083         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1084         d = ct->status ^ status;
1085
1086         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1087                 /* unchangeable */
1088                 return -EBUSY;
1089
1090         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1091                 /* SEEN_REPLY bit can only be set */
1092                 return -EBUSY;
1093
1094         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1095                 /* ASSURED bit can only be set */
1096                 return -EBUSY;
1097
1098         __ctnetlink_change_status(ct, status, 0);
1099         return 0;
1100 }
1101
1102 static int
1103 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1104 {
1105 #ifdef CONFIG_NF_NAT_NEEDED
1106         int ret;
1107
1108         if (cda[CTA_NAT_DST]) {
1109                 ret = ctnetlink_parse_nat_setup(ct,
1110                                                 IP_NAT_MANIP_DST,
1111                                                 cda[CTA_NAT_DST]);
1112                 if (ret < 0)
1113                         return ret;
1114         }
1115         if (cda[CTA_NAT_SRC]) {
1116                 ret = ctnetlink_parse_nat_setup(ct,
1117                                                 IP_NAT_MANIP_SRC,
1118                                                 cda[CTA_NAT_SRC]);
1119                 if (ret < 0)
1120                         return ret;
1121         }
1122         return 0;
1123 #else
1124         return -EOPNOTSUPP;
1125 #endif
1126 }
1127
1128 static inline int
1129 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1130 {
1131         struct nf_conntrack_helper *helper;
1132         struct nf_conn_help *help = nfct_help(ct);
1133         char *helpname = NULL;
1134         int err;
1135
1136         /* don't change helper of sibling connections */
1137         if (ct->master)
1138                 return -EBUSY;
1139
1140         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1141         if (err < 0)
1142                 return err;
1143
1144         if (!strcmp(helpname, "")) {
1145                 if (help && help->helper) {
1146                         /* we had a helper before ... */
1147                         nf_ct_remove_expectations(ct);
1148                         RCU_INIT_POINTER(help->helper, NULL);
1149                 }
1150
1151                 return 0;
1152         }
1153
1154         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1155                                             nf_ct_protonum(ct));
1156         if (helper == NULL) {
1157 #ifdef CONFIG_MODULES
1158                 spin_unlock_bh(&nf_conntrack_lock);
1159
1160                 if (request_module("nfct-helper-%s", helpname) < 0) {
1161                         spin_lock_bh(&nf_conntrack_lock);
1162                         return -EOPNOTSUPP;
1163                 }
1164
1165                 spin_lock_bh(&nf_conntrack_lock);
1166                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1167                                                     nf_ct_protonum(ct));
1168                 if (helper)
1169                         return -EAGAIN;
1170 #endif
1171                 return -EOPNOTSUPP;
1172         }
1173
1174         if (help) {
1175                 if (help->helper == helper)
1176                         return 0;
1177                 if (help->helper)
1178                         return -EBUSY;
1179                 /* need to zero data of old helper */
1180                 memset(&help->help, 0, sizeof(help->help));
1181         } else {
1182                 /* we cannot set a helper for an existing conntrack */
1183                 return -EOPNOTSUPP;
1184         }
1185
1186         rcu_assign_pointer(help->helper, helper);
1187
1188         return 0;
1189 }
1190
1191 static inline int
1192 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1193 {
1194         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1195
1196         if (!del_timer(&ct->timeout))
1197                 return -ETIME;
1198
1199         ct->timeout.expires = jiffies + timeout * HZ;
1200         add_timer(&ct->timeout);
1201
1202         return 0;
1203 }
1204
1205 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1206         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1207         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1208         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1209 };
1210
1211 static inline int
1212 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1213 {
1214         const struct nlattr *attr = cda[CTA_PROTOINFO];
1215         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1216         struct nf_conntrack_l4proto *l4proto;
1217         int err = 0;
1218
1219         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1220
1221         rcu_read_lock();
1222         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1223         if (l4proto->from_nlattr)
1224                 err = l4proto->from_nlattr(tb, ct);
1225         rcu_read_unlock();
1226
1227         return err;
1228 }
1229
1230 #ifdef CONFIG_NF_NAT_NEEDED
1231 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1232         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1233         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1234         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1235 };
1236
1237 static inline int
1238 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1239 {
1240         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1241
1242         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1243
1244         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1245                 return -EINVAL;
1246
1247         natseq->correction_pos =
1248                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1249
1250         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1251                 return -EINVAL;
1252
1253         natseq->offset_before =
1254                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1255
1256         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1257                 return -EINVAL;
1258
1259         natseq->offset_after =
1260                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1261
1262         return 0;
1263 }
1264
1265 static int
1266 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1267                              const struct nlattr * const cda[])
1268 {
1269         int ret = 0;
1270         struct nf_conn_nat *nat = nfct_nat(ct);
1271
1272         if (!nat)
1273                 return 0;
1274
1275         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1276                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1277                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1278                 if (ret < 0)
1279                         return ret;
1280
1281                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1282         }
1283
1284         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1285                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1286                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1287                 if (ret < 0)
1288                         return ret;
1289
1290                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1291         }
1292
1293         return 0;
1294 }
1295 #endif
1296
1297 static int
1298 ctnetlink_change_conntrack(struct nf_conn *ct,
1299                            const struct nlattr * const cda[])
1300 {
1301         int err;
1302
1303         /* only allow NAT changes and master assignation for new conntracks */
1304         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1305                 return -EOPNOTSUPP;
1306
1307         if (cda[CTA_HELP]) {
1308                 err = ctnetlink_change_helper(ct, cda);
1309                 if (err < 0)
1310                         return err;
1311         }
1312
1313         if (cda[CTA_TIMEOUT]) {
1314                 err = ctnetlink_change_timeout(ct, cda);
1315                 if (err < 0)
1316                         return err;
1317         }
1318
1319         if (cda[CTA_STATUS]) {
1320                 err = ctnetlink_change_status(ct, cda);
1321                 if (err < 0)
1322                         return err;
1323         }
1324
1325         if (cda[CTA_PROTOINFO]) {
1326                 err = ctnetlink_change_protoinfo(ct, cda);
1327                 if (err < 0)
1328                         return err;
1329         }
1330
1331 #if defined(CONFIG_NF_CONNTRACK_MARK)
1332         if (cda[CTA_MARK])
1333                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1334 #endif
1335
1336 #ifdef CONFIG_NF_NAT_NEEDED
1337         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1338                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1339                 if (err < 0)
1340                         return err;
1341         }
1342 #endif
1343
1344         return 0;
1345 }
1346
1347 static struct nf_conn *
1348 ctnetlink_create_conntrack(struct net *net, u16 zone,
1349                            const struct nlattr * const cda[],
1350                            struct nf_conntrack_tuple *otuple,
1351                            struct nf_conntrack_tuple *rtuple,
1352                            u8 u3)
1353 {
1354         struct nf_conn *ct;
1355         int err = -EINVAL;
1356         struct nf_conntrack_helper *helper;
1357         struct nf_conn_tstamp *tstamp;
1358
1359         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1360         if (IS_ERR(ct))
1361                 return ERR_PTR(-ENOMEM);
1362
1363         if (!cda[CTA_TIMEOUT])
1364                 goto err1;
1365         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1366
1367         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1368
1369         rcu_read_lock();
1370         if (cda[CTA_HELP]) {
1371                 char *helpname = NULL;
1372
1373                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1374                 if (err < 0)
1375                         goto err2;
1376
1377                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1378                                                     nf_ct_protonum(ct));
1379                 if (helper == NULL) {
1380                         rcu_read_unlock();
1381                         spin_unlock_bh(&nf_conntrack_lock);
1382 #ifdef CONFIG_MODULES
1383                         if (request_module("nfct-helper-%s", helpname) < 0) {
1384                                 spin_lock_bh(&nf_conntrack_lock);
1385                                 err = -EOPNOTSUPP;
1386                                 goto err1;
1387                         }
1388
1389                         spin_lock_bh(&nf_conntrack_lock);
1390                         rcu_read_lock();
1391                         helper = __nf_conntrack_helper_find(helpname,
1392                                                             nf_ct_l3num(ct),
1393                                                             nf_ct_protonum(ct));
1394                         if (helper) {
1395                                 err = -EAGAIN;
1396                                 goto err2;
1397                         }
1398                         rcu_read_unlock();
1399 #endif
1400                         err = -EOPNOTSUPP;
1401                         goto err1;
1402                 } else {
1403                         struct nf_conn_help *help;
1404
1405                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1406                         if (help == NULL) {
1407                                 err = -ENOMEM;
1408                                 goto err2;
1409                         }
1410
1411                         /* not in hash table yet so not strictly necessary */
1412                         RCU_INIT_POINTER(help->helper, helper);
1413                 }
1414         } else {
1415                 /* try an implicit helper assignation */
1416                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1417                 if (err < 0)
1418                         goto err2;
1419         }
1420
1421         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1422                 err = ctnetlink_change_nat(ct, cda);
1423                 if (err < 0)
1424                         goto err2;
1425         }
1426
1427         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1428         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1429         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1430         /* we must add conntrack extensions before confirmation. */
1431         ct->status |= IPS_CONFIRMED;
1432
1433         if (cda[CTA_STATUS]) {
1434                 err = ctnetlink_change_status(ct, cda);
1435                 if (err < 0)
1436                         goto err2;
1437         }
1438
1439 #ifdef CONFIG_NF_NAT_NEEDED
1440         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1441                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1442                 if (err < 0)
1443                         goto err2;
1444         }
1445 #endif
1446
1447         memset(&ct->proto, 0, sizeof(ct->proto));
1448         if (cda[CTA_PROTOINFO]) {
1449                 err = ctnetlink_change_protoinfo(ct, cda);
1450                 if (err < 0)
1451                         goto err2;
1452         }
1453
1454 #if defined(CONFIG_NF_CONNTRACK_MARK)
1455         if (cda[CTA_MARK])
1456                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1457 #endif
1458
1459         /* setup master conntrack: this is a confirmed expectation */
1460         if (cda[CTA_TUPLE_MASTER]) {
1461                 struct nf_conntrack_tuple master;
1462                 struct nf_conntrack_tuple_hash *master_h;
1463                 struct nf_conn *master_ct;
1464
1465                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1466                 if (err < 0)
1467                         goto err2;
1468
1469                 master_h = nf_conntrack_find_get(net, zone, &master);
1470                 if (master_h == NULL) {
1471                         err = -ENOENT;
1472                         goto err2;
1473                 }
1474                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1475                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1476                 ct->master = master_ct;
1477         }
1478         tstamp = nf_conn_tstamp_find(ct);
1479         if (tstamp)
1480                 tstamp->start = ktime_to_ns(ktime_get_real());
1481
1482         add_timer(&ct->timeout);
1483         nf_conntrack_hash_insert(ct);
1484         rcu_read_unlock();
1485
1486         return ct;
1487
1488 err2:
1489         rcu_read_unlock();
1490 err1:
1491         nf_conntrack_free(ct);
1492         return ERR_PTR(err);
1493 }
1494
1495 static int
1496 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1497                         const struct nlmsghdr *nlh,
1498                         const struct nlattr * const cda[])
1499 {
1500         struct net *net = sock_net(ctnl);
1501         struct nf_conntrack_tuple otuple, rtuple;
1502         struct nf_conntrack_tuple_hash *h = NULL;
1503         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1504         u_int8_t u3 = nfmsg->nfgen_family;
1505         u16 zone;
1506         int err;
1507
1508         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1509         if (err < 0)
1510                 return err;
1511
1512         if (cda[CTA_TUPLE_ORIG]) {
1513                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1514                 if (err < 0)
1515                         return err;
1516         }
1517
1518         if (cda[CTA_TUPLE_REPLY]) {
1519                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1520                 if (err < 0)
1521                         return err;
1522         }
1523
1524         spin_lock_bh(&nf_conntrack_lock);
1525         if (cda[CTA_TUPLE_ORIG])
1526                 h = __nf_conntrack_find(net, zone, &otuple);
1527         else if (cda[CTA_TUPLE_REPLY])
1528                 h = __nf_conntrack_find(net, zone, &rtuple);
1529
1530         if (h == NULL) {
1531                 err = -ENOENT;
1532                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1533                         struct nf_conn *ct;
1534                         enum ip_conntrack_events events;
1535
1536                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1537                                                         &rtuple, u3);
1538                         if (IS_ERR(ct)) {
1539                                 err = PTR_ERR(ct);
1540                                 goto out_unlock;
1541                         }
1542                         err = 0;
1543                         nf_conntrack_get(&ct->ct_general);
1544                         spin_unlock_bh(&nf_conntrack_lock);
1545                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1546                                 events = IPCT_RELATED;
1547                         else
1548                                 events = IPCT_NEW;
1549
1550                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1551                                                       (1 << IPCT_ASSURED) |
1552                                                       (1 << IPCT_HELPER) |
1553                                                       (1 << IPCT_PROTOINFO) |
1554                                                       (1 << IPCT_NATSEQADJ) |
1555                                                       (1 << IPCT_MARK) | events,
1556                                                       ct, NETLINK_CB(skb).pid,
1557                                                       nlmsg_report(nlh));
1558                         nf_ct_put(ct);
1559                 } else
1560                         spin_unlock_bh(&nf_conntrack_lock);
1561
1562                 return err;
1563         }
1564         /* implicit 'else' */
1565
1566         /* We manipulate the conntrack inside the global conntrack table lock,
1567          * so there's no need to increase the refcount */
1568         err = -EEXIST;
1569         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1570                 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1571
1572                 err = ctnetlink_change_conntrack(ct, cda);
1573                 if (err == 0) {
1574                         nf_conntrack_get(&ct->ct_general);
1575                         spin_unlock_bh(&nf_conntrack_lock);
1576                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1577                                                       (1 << IPCT_ASSURED) |
1578                                                       (1 << IPCT_HELPER) |
1579                                                       (1 << IPCT_PROTOINFO) |
1580                                                       (1 << IPCT_NATSEQADJ) |
1581                                                       (1 << IPCT_MARK),
1582                                                       ct, NETLINK_CB(skb).pid,
1583                                                       nlmsg_report(nlh));
1584                         nf_ct_put(ct);
1585                 } else
1586                         spin_unlock_bh(&nf_conntrack_lock);
1587
1588                 return err;
1589         }
1590
1591 out_unlock:
1592         spin_unlock_bh(&nf_conntrack_lock);
1593         return err;
1594 }
1595
1596 /***********************************************************************
1597  * EXPECT
1598  ***********************************************************************/
1599
1600 static inline int
1601 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1602                          const struct nf_conntrack_tuple *tuple,
1603                          enum ctattr_expect type)
1604 {
1605         struct nlattr *nest_parms;
1606
1607         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
1608         if (!nest_parms)
1609                 goto nla_put_failure;
1610         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1611                 goto nla_put_failure;
1612         nla_nest_end(skb, nest_parms);
1613
1614         return 0;
1615
1616 nla_put_failure:
1617         return -1;
1618 }
1619
1620 static inline int
1621 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1622                         const struct nf_conntrack_tuple *tuple,
1623                         const struct nf_conntrack_tuple_mask *mask)
1624 {
1625         int ret;
1626         struct nf_conntrack_l3proto *l3proto;
1627         struct nf_conntrack_l4proto *l4proto;
1628         struct nf_conntrack_tuple m;
1629         struct nlattr *nest_parms;
1630
1631         memset(&m, 0xFF, sizeof(m));
1632         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1633         m.src.u.all = mask->src.u.all;
1634         m.dst.protonum = tuple->dst.protonum;
1635
1636         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
1637         if (!nest_parms)
1638                 goto nla_put_failure;
1639
1640         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
1641         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1642
1643         if (unlikely(ret < 0))
1644                 goto nla_put_failure;
1645
1646         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
1647         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1648         if (unlikely(ret < 0))
1649                 goto nla_put_failure;
1650
1651         nla_nest_end(skb, nest_parms);
1652
1653         return 0;
1654
1655 nla_put_failure:
1656         return -1;
1657 }
1658
1659 static int
1660 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1661                           const struct nf_conntrack_expect *exp)
1662 {
1663         struct nf_conn *master = exp->master;
1664         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
1665         struct nf_conn_help *help;
1666
1667         if (timeout < 0)
1668                 timeout = 0;
1669
1670         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1671                 goto nla_put_failure;
1672         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1673                 goto nla_put_failure;
1674         if (ctnetlink_exp_dump_tuple(skb,
1675                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1676                                  CTA_EXPECT_MASTER) < 0)
1677                 goto nla_put_failure;
1678
1679         NLA_PUT_BE32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout));
1680         NLA_PUT_BE32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp));
1681         NLA_PUT_BE32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags));
1682         help = nfct_help(master);
1683         if (help) {
1684                 struct nf_conntrack_helper *helper;
1685
1686                 helper = rcu_dereference(help->helper);
1687                 if (helper)
1688                         NLA_PUT_STRING(skb, CTA_EXPECT_HELP_NAME, helper->name);
1689         }
1690
1691         return 0;
1692
1693 nla_put_failure:
1694         return -1;
1695 }
1696
1697 static int
1698 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1699                         int event, const struct nf_conntrack_expect *exp)
1700 {
1701         struct nlmsghdr *nlh;
1702         struct nfgenmsg *nfmsg;
1703         unsigned int flags = pid ? NLM_F_MULTI : 0;
1704
1705         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1706         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
1707         if (nlh == NULL)
1708                 goto nlmsg_failure;
1709
1710         nfmsg = nlmsg_data(nlh);
1711         nfmsg->nfgen_family = exp->tuple.src.l3num;
1712         nfmsg->version      = NFNETLINK_V0;
1713         nfmsg->res_id       = 0;
1714
1715         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1716                 goto nla_put_failure;
1717
1718         nlmsg_end(skb, nlh);
1719         return skb->len;
1720
1721 nlmsg_failure:
1722 nla_put_failure:
1723         nlmsg_cancel(skb, nlh);
1724         return -1;
1725 }
1726
1727 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1728 static int
1729 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
1730 {
1731         struct nf_conntrack_expect *exp = item->exp;
1732         struct net *net = nf_ct_exp_net(exp);
1733         struct nlmsghdr *nlh;
1734         struct nfgenmsg *nfmsg;
1735         struct sk_buff *skb;
1736         unsigned int type, group;
1737         int flags = 0;
1738
1739         if (events & (1 << IPEXP_DESTROY)) {
1740                 type = IPCTNL_MSG_EXP_DELETE;
1741                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
1742         } else if (events & (1 << IPEXP_NEW)) {
1743                 type = IPCTNL_MSG_EXP_NEW;
1744                 flags = NLM_F_CREATE|NLM_F_EXCL;
1745                 group = NFNLGRP_CONNTRACK_EXP_NEW;
1746         } else
1747                 return 0;
1748
1749         if (!item->report && !nfnetlink_has_listeners(net, group))
1750                 return 0;
1751
1752         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
1753         if (skb == NULL)
1754                 goto errout;
1755
1756         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1757         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
1758         if (nlh == NULL)
1759                 goto nlmsg_failure;
1760
1761         nfmsg = nlmsg_data(nlh);
1762         nfmsg->nfgen_family = exp->tuple.src.l3num;
1763         nfmsg->version      = NFNETLINK_V0;
1764         nfmsg->res_id       = 0;
1765
1766         rcu_read_lock();
1767         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1768                 goto nla_put_failure;
1769         rcu_read_unlock();
1770
1771         nlmsg_end(skb, nlh);
1772         nfnetlink_send(skb, net, item->pid, group, item->report, GFP_ATOMIC);
1773         return 0;
1774
1775 nla_put_failure:
1776         rcu_read_unlock();
1777         nlmsg_cancel(skb, nlh);
1778 nlmsg_failure:
1779         kfree_skb(skb);
1780 errout:
1781         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
1782         return 0;
1783 }
1784 #endif
1785 static int ctnetlink_exp_done(struct netlink_callback *cb)
1786 {
1787         if (cb->args[1])
1788                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1789         return 0;
1790 }
1791
1792 static int
1793 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1794 {
1795         struct net *net = sock_net(skb->sk);
1796         struct nf_conntrack_expect *exp, *last;
1797         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1798         struct hlist_node *n;
1799         u_int8_t l3proto = nfmsg->nfgen_family;
1800
1801         rcu_read_lock();
1802         last = (struct nf_conntrack_expect *)cb->args[1];
1803         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1804 restart:
1805                 hlist_for_each_entry(exp, n, &net->ct.expect_hash[cb->args[0]],
1806                                      hnode) {
1807                         if (l3proto && exp->tuple.src.l3num != l3proto)
1808                                 continue;
1809                         if (cb->args[1]) {
1810                                 if (exp != last)
1811                                         continue;
1812                                 cb->args[1] = 0;
1813                         }
1814                         if (ctnetlink_exp_fill_info(skb,
1815                                                     NETLINK_CB(cb->skb).pid,
1816                                                     cb->nlh->nlmsg_seq,
1817                                                     IPCTNL_MSG_EXP_NEW,
1818                                                     exp) < 0) {
1819                                 if (!atomic_inc_not_zero(&exp->use))
1820                                         continue;
1821                                 cb->args[1] = (unsigned long)exp;
1822                                 goto out;
1823                         }
1824                 }
1825                 if (cb->args[1]) {
1826                         cb->args[1] = 0;
1827                         goto restart;
1828                 }
1829         }
1830 out:
1831         rcu_read_unlock();
1832         if (last)
1833                 nf_ct_expect_put(last);
1834
1835         return skb->len;
1836 }
1837
1838 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1839         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1840         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1841         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1842         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1843         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1844         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING },
1845         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1846         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1847 };
1848
1849 static int
1850 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1851                      const struct nlmsghdr *nlh,
1852                      const struct nlattr * const cda[])
1853 {
1854         struct net *net = sock_net(ctnl);
1855         struct nf_conntrack_tuple tuple;
1856         struct nf_conntrack_expect *exp;
1857         struct sk_buff *skb2;
1858         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1859         u_int8_t u3 = nfmsg->nfgen_family;
1860         u16 zone;
1861         int err;
1862
1863         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1864                 return netlink_dump_start(ctnl, skb, nlh,
1865                                           ctnetlink_exp_dump_table,
1866                                           ctnetlink_exp_done, 0);
1867         }
1868
1869         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1870         if (err < 0)
1871                 return err;
1872
1873         if (cda[CTA_EXPECT_MASTER])
1874                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1875         else
1876                 return -EINVAL;
1877
1878         if (err < 0)
1879                 return err;
1880
1881         exp = nf_ct_expect_find_get(net, zone, &tuple);
1882         if (!exp)
1883                 return -ENOENT;
1884
1885         if (cda[CTA_EXPECT_ID]) {
1886                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1887                 if (ntohl(id) != (u32)(unsigned long)exp) {
1888                         nf_ct_expect_put(exp);
1889                         return -ENOENT;
1890                 }
1891         }
1892
1893         err = -ENOMEM;
1894         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1895         if (skb2 == NULL) {
1896                 nf_ct_expect_put(exp);
1897                 goto out;
1898         }
1899
1900         rcu_read_lock();
1901         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1902                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
1903         rcu_read_unlock();
1904         nf_ct_expect_put(exp);
1905         if (err <= 0)
1906                 goto free;
1907
1908         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1909         if (err < 0)
1910                 goto out;
1911
1912         return 0;
1913
1914 free:
1915         kfree_skb(skb2);
1916 out:
1917         /* this avoids a loop in nfnetlink. */
1918         return err == -EAGAIN ? -ENOBUFS : err;
1919 }
1920
1921 static int
1922 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1923                      const struct nlmsghdr *nlh,
1924                      const struct nlattr * const cda[])
1925 {
1926         struct net *net = sock_net(ctnl);
1927         struct nf_conntrack_expect *exp;
1928         struct nf_conntrack_tuple tuple;
1929         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1930         struct hlist_node *n, *next;
1931         u_int8_t u3 = nfmsg->nfgen_family;
1932         unsigned int i;
1933         u16 zone;
1934         int err;
1935
1936         if (cda[CTA_EXPECT_TUPLE]) {
1937                 /* delete a single expect by tuple */
1938                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1939                 if (err < 0)
1940                         return err;
1941
1942                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1943                 if (err < 0)
1944                         return err;
1945
1946                 /* bump usage count to 2 */
1947                 exp = nf_ct_expect_find_get(net, zone, &tuple);
1948                 if (!exp)
1949                         return -ENOENT;
1950
1951                 if (cda[CTA_EXPECT_ID]) {
1952                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1953                         if (ntohl(id) != (u32)(unsigned long)exp) {
1954                                 nf_ct_expect_put(exp);
1955                                 return -ENOENT;
1956                         }
1957                 }
1958
1959                 /* after list removal, usage count == 1 */
1960                 spin_lock_bh(&nf_conntrack_lock);
1961                 if (del_timer(&exp->timeout)) {
1962                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).pid,
1963                                                    nlmsg_report(nlh));
1964                         nf_ct_expect_put(exp);
1965                 }
1966                 spin_unlock_bh(&nf_conntrack_lock);
1967                 /* have to put what we 'get' above.
1968                  * after this line usage count == 0 */
1969                 nf_ct_expect_put(exp);
1970         } else if (cda[CTA_EXPECT_HELP_NAME]) {
1971                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
1972                 struct nf_conn_help *m_help;
1973
1974                 /* delete all expectations for this helper */
1975                 spin_lock_bh(&nf_conntrack_lock);
1976                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1977                         hlist_for_each_entry_safe(exp, n, next,
1978                                                   &net->ct.expect_hash[i],
1979                                                   hnode) {
1980                                 m_help = nfct_help(exp->master);
1981                                 if (!strcmp(m_help->helper->name, name) &&
1982                                     del_timer(&exp->timeout)) {
1983                                         nf_ct_unlink_expect_report(exp,
1984                                                         NETLINK_CB(skb).pid,
1985                                                         nlmsg_report(nlh));
1986                                         nf_ct_expect_put(exp);
1987                                 }
1988                         }
1989                 }
1990                 spin_unlock_bh(&nf_conntrack_lock);
1991         } else {
1992                 /* This basically means we have to flush everything*/
1993                 spin_lock_bh(&nf_conntrack_lock);
1994                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1995                         hlist_for_each_entry_safe(exp, n, next,
1996                                                   &net->ct.expect_hash[i],
1997                                                   hnode) {
1998                                 if (del_timer(&exp->timeout)) {
1999                                         nf_ct_unlink_expect_report(exp,
2000                                                         NETLINK_CB(skb).pid,
2001                                                         nlmsg_report(nlh));
2002                                         nf_ct_expect_put(exp);
2003                                 }
2004                         }
2005                 }
2006                 spin_unlock_bh(&nf_conntrack_lock);
2007         }
2008
2009         return 0;
2010 }
2011 static int
2012 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2013                         const struct nlattr * const cda[])
2014 {
2015         return -EOPNOTSUPP;
2016 }
2017
2018 static int
2019 ctnetlink_create_expect(struct net *net, u16 zone,
2020                         const struct nlattr * const cda[],
2021                         u_int8_t u3,
2022                         u32 pid, int report)
2023 {
2024         struct nf_conntrack_tuple tuple, mask, master_tuple;
2025         struct nf_conntrack_tuple_hash *h = NULL;
2026         struct nf_conntrack_expect *exp;
2027         struct nf_conn *ct;
2028         struct nf_conn_help *help;
2029         int err = 0;
2030
2031         /* caller guarantees that those three CTA_EXPECT_* exist */
2032         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2033         if (err < 0)
2034                 return err;
2035         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2036         if (err < 0)
2037                 return err;
2038         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2039         if (err < 0)
2040                 return err;
2041
2042         /* Look for master conntrack of this expectation */
2043         h = nf_conntrack_find_get(net, zone, &master_tuple);
2044         if (!h)
2045                 return -ENOENT;
2046         ct = nf_ct_tuplehash_to_ctrack(h);
2047         exp = nf_ct_expect_alloc(ct);
2048         if (!exp) {
2049                 err = -ENOMEM;
2050                 goto out;
2051         }
2052         help = nfct_help(ct);
2053         if (!help) {
2054                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2055                         err = -EINVAL;
2056                         goto out;
2057                 }
2058                 exp->timeout.expires =
2059                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2060
2061                 exp->flags = NF_CT_EXPECT_USERSPACE;
2062                 if (cda[CTA_EXPECT_FLAGS]) {
2063                         exp->flags |=
2064                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2065                 }
2066         } else {
2067                 if (cda[CTA_EXPECT_FLAGS]) {
2068                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2069                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2070                 } else
2071                         exp->flags = 0;
2072         }
2073
2074         exp->class = 0;
2075         exp->expectfn = NULL;
2076         exp->master = ct;
2077         exp->helper = NULL;
2078         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2079         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2080         exp->mask.src.u.all = mask.src.u.all;
2081
2082         err = nf_ct_expect_related_report(exp, pid, report);
2083         nf_ct_expect_put(exp);
2084
2085 out:
2086         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2087         return err;
2088 }
2089
2090 static int
2091 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2092                      const struct nlmsghdr *nlh,
2093                      const struct nlattr * const cda[])
2094 {
2095         struct net *net = sock_net(ctnl);
2096         struct nf_conntrack_tuple tuple;
2097         struct nf_conntrack_expect *exp;
2098         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2099         u_int8_t u3 = nfmsg->nfgen_family;
2100         u16 zone;
2101         int err;
2102
2103         if (!cda[CTA_EXPECT_TUPLE]
2104             || !cda[CTA_EXPECT_MASK]
2105             || !cda[CTA_EXPECT_MASTER])
2106                 return -EINVAL;
2107
2108         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2109         if (err < 0)
2110                 return err;
2111
2112         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2113         if (err < 0)
2114                 return err;
2115
2116         spin_lock_bh(&nf_conntrack_lock);
2117         exp = __nf_ct_expect_find(net, zone, &tuple);
2118
2119         if (!exp) {
2120                 spin_unlock_bh(&nf_conntrack_lock);
2121                 err = -ENOENT;
2122                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2123                         err = ctnetlink_create_expect(net, zone, cda,
2124                                                       u3,
2125                                                       NETLINK_CB(skb).pid,
2126                                                       nlmsg_report(nlh));
2127                 }
2128                 return err;
2129         }
2130
2131         err = -EEXIST;
2132         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2133                 err = ctnetlink_change_expect(exp, cda);
2134         spin_unlock_bh(&nf_conntrack_lock);
2135
2136         return err;
2137 }
2138
2139 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2140 static struct nf_ct_event_notifier ctnl_notifier = {
2141         .fcn = ctnetlink_conntrack_event,
2142 };
2143
2144 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2145         .fcn = ctnetlink_expect_event,
2146 };
2147 #endif
2148
2149 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2150         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2151                                             .attr_count = CTA_MAX,
2152                                             .policy = ct_nla_policy },
2153         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2154                                             .attr_count = CTA_MAX,
2155                                             .policy = ct_nla_policy },
2156         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2157                                             .attr_count = CTA_MAX,
2158                                             .policy = ct_nla_policy },
2159         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
2160                                             .attr_count = CTA_MAX,
2161                                             .policy = ct_nla_policy },
2162 };
2163
2164 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
2165         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
2166                                             .attr_count = CTA_EXPECT_MAX,
2167                                             .policy = exp_nla_policy },
2168         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
2169                                             .attr_count = CTA_EXPECT_MAX,
2170                                             .policy = exp_nla_policy },
2171         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
2172                                             .attr_count = CTA_EXPECT_MAX,
2173                                             .policy = exp_nla_policy },
2174 };
2175
2176 static const struct nfnetlink_subsystem ctnl_subsys = {
2177         .name                           = "conntrack",
2178         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
2179         .cb_count                       = IPCTNL_MSG_MAX,
2180         .cb                             = ctnl_cb,
2181 };
2182
2183 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
2184         .name                           = "conntrack_expect",
2185         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
2186         .cb_count                       = IPCTNL_MSG_EXP_MAX,
2187         .cb                             = ctnl_exp_cb,
2188 };
2189
2190 MODULE_ALIAS("ip_conntrack_netlink");
2191 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
2192 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
2193
2194 static int __net_init ctnetlink_net_init(struct net *net)
2195 {
2196 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2197         int ret;
2198
2199         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
2200         if (ret < 0) {
2201                 pr_err("ctnetlink_init: cannot register notifier.\n");
2202                 goto err_out;
2203         }
2204
2205         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
2206         if (ret < 0) {
2207                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
2208                 goto err_unreg_notifier;
2209         }
2210 #endif
2211         return 0;
2212
2213 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2214 err_unreg_notifier:
2215         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2216 err_out:
2217         return ret;
2218 #endif
2219 }
2220
2221 static void ctnetlink_net_exit(struct net *net)
2222 {
2223 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2224         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
2225         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2226 #endif
2227 }
2228
2229 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
2230 {
2231         struct net *net;
2232
2233         list_for_each_entry(net, net_exit_list, exit_list)
2234                 ctnetlink_net_exit(net);
2235 }
2236
2237 static struct pernet_operations ctnetlink_net_ops = {
2238         .init           = ctnetlink_net_init,
2239         .exit_batch     = ctnetlink_net_exit_batch,
2240 };
2241
2242 static int __init ctnetlink_init(void)
2243 {
2244         int ret;
2245
2246         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
2247         ret = nfnetlink_subsys_register(&ctnl_subsys);
2248         if (ret < 0) {
2249                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
2250                 goto err_out;
2251         }
2252
2253         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
2254         if (ret < 0) {
2255                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
2256                 goto err_unreg_subsys;
2257         }
2258
2259         if (register_pernet_subsys(&ctnetlink_net_ops)) {
2260                 pr_err("ctnetlink_init: cannot register pernet operations\n");
2261                 goto err_unreg_exp_subsys;
2262         }
2263
2264         return 0;
2265
2266 err_unreg_exp_subsys:
2267         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2268 err_unreg_subsys:
2269         nfnetlink_subsys_unregister(&ctnl_subsys);
2270 err_out:
2271         return ret;
2272 }
2273
2274 static void __exit ctnetlink_exit(void)
2275 {
2276         pr_info("ctnetlink: unregistering from nfnetlink.\n");
2277
2278         nf_ct_remove_userspace_expectations();
2279         unregister_pernet_subsys(&ctnetlink_net_ops);
2280         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2281         nfnetlink_subsys_unregister(&ctnl_subsys);
2282 }
2283
2284 module_init(ctnetlink_init);
2285 module_exit(ctnetlink_exit);