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