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