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 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 586                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 587                     && ctnetlink_dump_secctx(skb, ct) < 0)
 588                         goto nla_put_failure;
 589 #endif
 590
 591                 if (events & (1 << IPCT_RELATED) &&
 592                     ctnetlink_dump_master(skb, ct) < 0)
 593                         goto nla_put_failure;
 594
 595                 if (events & (1 << IPCT_NATSEQADJ) &&
 596                     ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 597                         goto nla_put_failure;
 598         }
 599
 600 #ifdef CONFIG_NF_CONNTRACK_MARK
 601         if ((events & (1 << IPCT_MARK) || ct->mark)
 602             && ctnetlink_dump_mark(skb, ct) < 0)
 603                 goto nla_put_failure;
 604 #endif
 605         rcu_read_unlock();
 606
 607         nlmsg_end(skb, nlh);
 608         err = nfnetlink_send(skb, net, item->pid, group, item->report,
 609                              GFP_ATOMIC);
 610         if (err == -ENOBUFS || err == -EAGAIN)
 611                 return -ENOBUFS;
 612
 613         return 0;
 614
 615 nla_put_failure:
 616         rcu_read_unlock();
 617         nlmsg_cancel(skb, nlh);
 618 nlmsg_failure:
 619         kfree_skb(skb);
 620 errout:
 621         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 622                 return -ENOBUFS;
 623
 624         return 0;
 625 }
 626 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 627
 628 static int ctnetlink_done(struct netlink_callback *cb)
 629 {
 630         if (cb->args[1])
 631                 nf_ct_put((struct nf_conn *)cb->args[1]);
 632         return 0;
 633 }
 634
 635 static int
 636 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 637 {
 638         struct net *net = sock_net(skb->sk);
 639         struct nf_conn *ct, *last;
 640         struct nf_conntrack_tuple_hash *h;
 641         struct hlist_nulls_node *n;
 642         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 643         u_int8_t l3proto = nfmsg->nfgen_family;
 644
 645         spin_lock_bh(&nf_conntrack_lock);
 646         last = (struct nf_conn *)cb->args[1];
 647         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 648 restart:
 649                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 650                                          hnnode) {
 651                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 652                                 continue;
 653                         ct = nf_ct_tuplehash_to_ctrack(h);
 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                                 continue;
 659                         if (cb->args[1]) {
 660                                 if (ct != last)
 661                                         continue;
 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                                 nf_conntrack_get(&ct->ct_general);
 668                                 cb->args[1] = (unsigned long)ct;
 669                                 goto out;
 670                         }
 671
 672                         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) ==
 673                                                 IPCTNL_MSG_CT_GET_CTRZERO) {
 674                                 struct nf_conn_counter *acct;
 675
 676                                 acct = nf_conn_acct_find(ct);
 677                                 if (acct)
 678                                         memset(acct, 0, sizeof(struct nf_conn_counter[IP_CT_DIR_MAX]));
 679                         }
 680                 }
 681                 if (cb->args[1]) {
 682                         cb->args[1] = 0;
 683                         goto restart;
 684                 }
 685         }
 686 out:
 687         spin_unlock_bh(&nf_conntrack_lock);
 688         if (last)
 689                 nf_ct_put(last);
 690
 691         return skb->len;
 692 }
 693
 694 static inline int
 695 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 696 {
 697         struct nlattr *tb[CTA_IP_MAX+1];
 698         struct nf_conntrack_l3proto *l3proto;
 699         int ret = 0;
 700
 701         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 702
 703         rcu_read_lock();
 704         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 705
 706         if (likely(l3proto->nlattr_to_tuple)) {
 707                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 708                                           l3proto->nla_policy);
 709                 if (ret == 0)
 710                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 711         }
 712
 713         rcu_read_unlock();
 714
 715         return ret;
 716 }
 717
 718 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 719         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 720 };
 721
 722 static inline int
 723 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 724                             struct nf_conntrack_tuple *tuple)
 725 {
 726         struct nlattr *tb[CTA_PROTO_MAX+1];
 727         struct nf_conntrack_l4proto *l4proto;
 728         int ret = 0;
 729
 730         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 731         if (ret < 0)
 732                 return ret;
 733
 734         if (!tb[CTA_PROTO_NUM])
 735                 return -EINVAL;
 736         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 737
 738         rcu_read_lock();
 739         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 740
 741         if (likely(l4proto->nlattr_to_tuple)) {
 742                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 743                                           l4proto->nla_policy);
 744                 if (ret == 0)
 745                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 746         }
 747
 748         rcu_read_unlock();
 749
 750         return ret;
 751 }
 752
 753 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 754         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 755         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 756 };
 757
 758 static int
 759 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 760                       struct nf_conntrack_tuple *tuple,
 761                       enum ctattr_tuple type, u_int8_t l3num)
 762 {
 763         struct nlattr *tb[CTA_TUPLE_MAX+1];
 764         int err;
 765
 766         memset(tuple, 0, sizeof(*tuple));
 767
 768         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 769
 770         if (!tb[CTA_TUPLE_IP])
 771                 return -EINVAL;
 772
 773         tuple->src.l3num = l3num;
 774
 775         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 776         if (err < 0)
 777                 return err;
 778
 779         if (!tb[CTA_TUPLE_PROTO])
 780                 return -EINVAL;
 781
 782         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 783         if (err < 0)
 784                 return err;
 785
 786         /* orig and expect tuples get DIR_ORIGINAL */
 787         if (type == CTA_TUPLE_REPLY)
 788                 tuple->dst.dir = IP_CT_DIR_REPLY;
 789         else
 790                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 791
 792         return 0;
 793 }
 794
 795 static int
 796 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 797 {
 798         if (attr)
 799 #ifdef CONFIG_NF_CONNTRACK_ZONES
 800                 *zone = ntohs(nla_get_be16(attr));
 801 #else
 802                 return -EOPNOTSUPP;
 803 #endif
 804         else
 805                 *zone = 0;
 806
 807         return 0;
 808 }
 809
 810 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 811         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING },
 812 };
 813
 814 static inline int
 815 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name)
 816 {
 817         struct nlattr *tb[CTA_HELP_MAX+1];
 818
 819         nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 820
 821         if (!tb[CTA_HELP_NAME])
 822                 return -EINVAL;
 823
 824         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 825
 826         return 0;
 827 }
 828
 829 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 830         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 831         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 832         [CTA_STATUS]            = { .type = NLA_U32 },
 833         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 834         [CTA_HELP]              = { .type = NLA_NESTED },
 835         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 836         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 837         [CTA_MARK]              = { .type = NLA_U32 },
 838         [CTA_ID]                = { .type = NLA_U32 },
 839         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 840         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 841         [CTA_ZONE]              = { .type = NLA_U16 },
 842 };
 843
 844 static int
 845 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 846                         const struct nlmsghdr *nlh,
 847                         const struct nlattr * const cda[])
 848 {
 849         struct net *net = sock_net(ctnl);
 850         struct nf_conntrack_tuple_hash *h;
 851         struct nf_conntrack_tuple tuple;
 852         struct nf_conn *ct;
 853         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 854         u_int8_t u3 = nfmsg->nfgen_family;
 855         u16 zone;
 856         int err;
 857
 858         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 859         if (err < 0)
 860                 return err;
 861
 862         if (cda[CTA_TUPLE_ORIG])
 863                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 864         else if (cda[CTA_TUPLE_REPLY])
 865                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 866         else {
 867                 /* Flush the whole table */
 868                 nf_conntrack_flush_report(net,
 869                                          NETLINK_CB(skb).pid,
 870                                          nlmsg_report(nlh));
 871                 return 0;
 872         }
 873
 874         if (err < 0)
 875                 return err;
 876
 877         h = nf_conntrack_find_get(net, zone, &tuple);
 878         if (!h)
 879                 return -ENOENT;
 880
 881         ct = nf_ct_tuplehash_to_ctrack(h);
 882
 883         if (cda[CTA_ID]) {
 884                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
 885                 if (id != (u32)(unsigned long)ct) {
 886                         nf_ct_put(ct);
 887                         return -ENOENT;
 888                 }
 889         }
 890
 891         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 892                                       NETLINK_CB(skb).pid,
 893                                       nlmsg_report(nlh)) < 0) {
 894                 nf_ct_delete_from_lists(ct);
 895                 /* we failed to report the event, try later */
 896                 nf_ct_insert_dying_list(ct);
 897                 nf_ct_put(ct);
 898                 return 0;
 899         }
 900
 901         /* death_by_timeout would report the event again */
 902         set_bit(IPS_DYING_BIT, &ct->status);
 903
 904         nf_ct_kill(ct);
 905         nf_ct_put(ct);
 906
 907         return 0;
 908 }
 909
 910 static int
 911 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 912                         const struct nlmsghdr *nlh,
 913                         const struct nlattr * const cda[])
 914 {
 915         struct net *net = sock_net(ctnl);
 916         struct nf_conntrack_tuple_hash *h;
 917         struct nf_conntrack_tuple tuple;
 918         struct nf_conn *ct;
 919         struct sk_buff *skb2 = NULL;
 920         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 921         u_int8_t u3 = nfmsg->nfgen_family;
 922         u16 zone;
 923         int err;
 924
 925         if (nlh->nlmsg_flags & NLM_F_DUMP)
 926                 return netlink_dump_start(ctnl, skb, nlh, ctnetlink_dump_table,
 927                                           ctnetlink_done);
 928
 929         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 930         if (err < 0)
 931                 return err;
 932
 933         if (cda[CTA_TUPLE_ORIG])
 934                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 935         else if (cda[CTA_TUPLE_REPLY])
 936                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 937         else
 938                 return -EINVAL;
 939
 940         if (err < 0)
 941                 return err;
 942
 943         h = nf_conntrack_find_get(net, zone, &tuple);
 944         if (!h)
 945                 return -ENOENT;
 946
 947         ct = nf_ct_tuplehash_to_ctrack(h);
 948
 949         err = -ENOMEM;
 950         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 951         if (skb2 == NULL) {
 952                 nf_ct_put(ct);
 953                 return -ENOMEM;
 954         }
 955
 956         rcu_read_lock();
 957         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
 958                                   IPCTNL_MSG_CT_NEW, ct);
 959         rcu_read_unlock();
 960         nf_ct_put(ct);
 961         if (err <= 0)
 962                 goto free;
 963
 964         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
 965         if (err < 0)
 966                 goto out;
 967
 968         return 0;
 969
 970 free:
 971         kfree_skb(skb2);
 972 out:
 973         return err;
 974 }
 975
 976 #ifdef CONFIG_NF_NAT_NEEDED
 977 static int
 978 ctnetlink_parse_nat_setup(struct nf_conn *ct,
 979                           enum nf_nat_manip_type manip,
 980                           const struct nlattr *attr)
 981 {
 982         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
 983
 984         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
 985         if (!parse_nat_setup) {
 986 #ifdef CONFIG_MODULES
 987                 rcu_read_unlock();
 988                 spin_unlock_bh(&nf_conntrack_lock);
 989                 nfnl_unlock();
 990                 if (request_module("nf-nat-ipv4") < 0) {
 991                         nfnl_lock();
 992                         spin_lock_bh(&nf_conntrack_lock);
 993                         rcu_read_lock();
 994                         return -EOPNOTSUPP;
 995                 }
 996                 nfnl_lock();
 997                 spin_lock_bh(&nf_conntrack_lock);
 998                 rcu_read_lock();
 999                 if (nfnetlink_parse_nat_setup_hook)
1000                         return -EAGAIN;
1001 #endif
1002                 return -EOPNOTSUPP;
1003         }
1004
1005         return parse_nat_setup(ct, manip, attr);
1006 }
1007 #endif
1008
1009 static int
1010 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1011 {
1012         unsigned long d;
1013         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1014         d = ct->status ^ status;
1015
1016         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1017                 /* unchangeable */
1018                 return -EBUSY;
1019
1020         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1021                 /* SEEN_REPLY bit can only be set */
1022                 return -EBUSY;
1023
1024         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1025                 /* ASSURED bit can only be set */
1026                 return -EBUSY;
1027
1028         /* Be careful here, modifying NAT bits can screw up things,
1029          * so don't let users modify them directly if they don't pass
1030          * nf_nat_range. */
1031         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1032         return 0;
1033 }
1034
1035 static int
1036 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1037 {
1038 #ifdef CONFIG_NF_NAT_NEEDED
1039         int ret;
1040
1041         if (cda[CTA_NAT_DST]) {
1042                 ret = ctnetlink_parse_nat_setup(ct,
1043                                                 IP_NAT_MANIP_DST,
1044                                                 cda[CTA_NAT_DST]);
1045                 if (ret < 0)
1046                         return ret;
1047         }
1048         if (cda[CTA_NAT_SRC]) {
1049                 ret = ctnetlink_parse_nat_setup(ct,
1050                                                 IP_NAT_MANIP_SRC,
1051                                                 cda[CTA_NAT_SRC]);
1052                 if (ret < 0)
1053                         return ret;
1054         }
1055         return 0;
1056 #else
1057         return -EOPNOTSUPP;
1058 #endif
1059 }
1060
1061 static inline int
1062 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1063 {
1064         struct nf_conntrack_helper *helper;
1065         struct nf_conn_help *help = nfct_help(ct);
1066         char *helpname = NULL;
1067         int err;
1068
1069         /* don't change helper of sibling connections */
1070         if (ct->master)
1071                 return -EBUSY;
1072
1073         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1074         if (err < 0)
1075                 return err;
1076
1077         if (!strcmp(helpname, "")) {
1078                 if (help && help->helper) {
1079                         /* we had a helper before ... */
1080                         nf_ct_remove_expectations(ct);
1081                         rcu_assign_pointer(help->helper, NULL);
1082                 }
1083
1084                 return 0;
1085         }
1086
1087         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1088                                             nf_ct_protonum(ct));
1089         if (helper == NULL) {
1090 #ifdef CONFIG_MODULES
1091                 spin_unlock_bh(&nf_conntrack_lock);
1092
1093                 if (request_module("nfct-helper-%s", helpname) < 0) {
1094                         spin_lock_bh(&nf_conntrack_lock);
1095                         return -EOPNOTSUPP;
1096                 }
1097
1098                 spin_lock_bh(&nf_conntrack_lock);
1099                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1100                                                     nf_ct_protonum(ct));
1101                 if (helper)
1102                         return -EAGAIN;
1103 #endif
1104                 return -EOPNOTSUPP;
1105         }
1106
1107         if (help) {
1108                 if (help->helper == helper)
1109                         return 0;
1110                 if (help->helper)
1111                         return -EBUSY;
1112                 /* need to zero data of old helper */
1113                 memset(&help->help, 0, sizeof(help->help));
1114         } else {
1115                 /* we cannot set a helper for an existing conntrack */
1116                 return -EOPNOTSUPP;
1117         }
1118
1119         rcu_assign_pointer(help->helper, helper);
1120
1121         return 0;
1122 }
1123
1124 static inline int
1125 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1126 {
1127         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1128
1129         if (!del_timer(&ct->timeout))
1130                 return -ETIME;
1131
1132         ct->timeout.expires = jiffies + timeout * HZ;
1133         add_timer(&ct->timeout);
1134
1135         return 0;
1136 }
1137
1138 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1139         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1140         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1141         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1142 };
1143
1144 static inline int
1145 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1146 {
1147         const struct nlattr *attr = cda[CTA_PROTOINFO];
1148         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1149         struct nf_conntrack_l4proto *l4proto;
1150         int err = 0;
1151
1152         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1153
1154         rcu_read_lock();
1155         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1156         if (l4proto->from_nlattr)
1157                 err = l4proto->from_nlattr(tb, ct);
1158         rcu_read_unlock();
1159
1160         return err;
1161 }
1162
1163 #ifdef CONFIG_NF_NAT_NEEDED
1164 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1165         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1166         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1167         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1168 };
1169
1170 static inline int
1171 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1172 {
1173         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1174
1175         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1176
1177         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1178                 return -EINVAL;
1179
1180         natseq->correction_pos =
1181                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1182
1183         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1184                 return -EINVAL;
1185
1186         natseq->offset_before =
1187                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1188
1189         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1190                 return -EINVAL;
1191
1192         natseq->offset_after =
1193                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1194
1195         return 0;
1196 }
1197
1198 static int
1199 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1200                              const struct nlattr * const cda[])
1201 {
1202         int ret = 0;
1203         struct nf_conn_nat *nat = nfct_nat(ct);
1204
1205         if (!nat)
1206                 return 0;
1207
1208         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1209                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1210                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1211                 if (ret < 0)
1212                         return ret;
1213
1214                 ct->status |= IPS_SEQ_ADJUST;
1215         }
1216
1217         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1218                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1219                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1220                 if (ret < 0)
1221                         return ret;
1222
1223                 ct->status |= IPS_SEQ_ADJUST;
1224         }
1225
1226         return 0;
1227 }
1228 #endif
1229
1230 static int
1231 ctnetlink_change_conntrack(struct nf_conn *ct,
1232                            const struct nlattr * const cda[])
1233 {
1234         int err;
1235
1236         /* only allow NAT changes and master assignation for new conntracks */
1237         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1238                 return -EOPNOTSUPP;
1239
1240         if (cda[CTA_HELP]) {
1241                 err = ctnetlink_change_helper(ct, cda);
1242                 if (err < 0)
1243                         return err;
1244         }
1245
1246         if (cda[CTA_TIMEOUT]) {
1247                 err = ctnetlink_change_timeout(ct, cda);
1248                 if (err < 0)
1249                         return err;
1250         }
1251
1252         if (cda[CTA_STATUS]) {
1253                 err = ctnetlink_change_status(ct, cda);
1254                 if (err < 0)
1255                         return err;
1256         }
1257
1258         if (cda[CTA_PROTOINFO]) {
1259                 err = ctnetlink_change_protoinfo(ct, cda);
1260                 if (err < 0)
1261                         return err;
1262         }
1263
1264 #if defined(CONFIG_NF_CONNTRACK_MARK)
1265         if (cda[CTA_MARK])
1266                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1267 #endif
1268
1269 #ifdef CONFIG_NF_NAT_NEEDED
1270         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1271                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1272                 if (err < 0)
1273                         return err;
1274         }
1275 #endif
1276
1277         return 0;
1278 }
1279
1280 static struct nf_conn *
1281 ctnetlink_create_conntrack(struct net *net, u16 zone,
1282                            const struct nlattr * const cda[],
1283                            struct nf_conntrack_tuple *otuple,
1284                            struct nf_conntrack_tuple *rtuple,
1285                            u8 u3)
1286 {
1287         struct nf_conn *ct;
1288         int err = -EINVAL;
1289         struct nf_conntrack_helper *helper;
1290
1291         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1292         if (IS_ERR(ct))
1293                 return ERR_PTR(-ENOMEM);
1294
1295         if (!cda[CTA_TIMEOUT])
1296                 goto err1;
1297         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1298
1299         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1300
1301         rcu_read_lock();
1302         if (cda[CTA_HELP]) {
1303                 char *helpname = NULL;
1304
1305                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1306                 if (err < 0)
1307                         goto err2;
1308
1309                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1310                                                     nf_ct_protonum(ct));
1311                 if (helper == NULL) {
1312                         rcu_read_unlock();
1313 #ifdef CONFIG_MODULES
1314                         if (request_module("nfct-helper-%s", helpname) < 0) {
1315                                 err = -EOPNOTSUPP;
1316                                 goto err1;
1317                         }
1318
1319                         rcu_read_lock();
1320                         helper = __nf_conntrack_helper_find(helpname,
1321                                                             nf_ct_l3num(ct),
1322                                                             nf_ct_protonum(ct));
1323                         if (helper) {
1324                                 err = -EAGAIN;
1325                                 goto err2;
1326                         }
1327                         rcu_read_unlock();
1328 #endif
1329                         err = -EOPNOTSUPP;
1330                         goto err1;
1331                 } else {
1332                         struct nf_conn_help *help;
1333
1334                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1335                         if (help == NULL) {
1336                                 err = -ENOMEM;
1337                                 goto err2;
1338                         }
1339
1340                         /* not in hash table yet so not strictly necessary */
1341                         rcu_assign_pointer(help->helper, helper);
1342                 }
1343         } else {
1344                 /* try an implicit helper assignation */
1345                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1346                 if (err < 0)
1347                         goto err2;
1348         }
1349
1350         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1351                 err = ctnetlink_change_nat(ct, cda);
1352                 if (err < 0)
1353                         goto err2;
1354         }
1355
1356         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1357         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1358         /* we must add conntrack extensions before confirmation. */
1359         ct->status |= IPS_CONFIRMED;
1360
1361         if (cda[CTA_STATUS]) {
1362                 err = ctnetlink_change_status(ct, cda);
1363                 if (err < 0)
1364                         goto err2;
1365         }
1366
1367 #ifdef CONFIG_NF_NAT_NEEDED
1368         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1369                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1370                 if (err < 0)
1371                         goto err2;
1372         }
1373 #endif
1374
1375         if (cda[CTA_PROTOINFO]) {
1376                 err = ctnetlink_change_protoinfo(ct, cda);
1377                 if (err < 0)
1378                         goto err2;
1379         }
1380
1381 #if defined(CONFIG_NF_CONNTRACK_MARK)
1382         if (cda[CTA_MARK])
1383                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1384 #endif
1385
1386         /* setup master conntrack: this is a confirmed expectation */
1387         if (cda[CTA_TUPLE_MASTER]) {
1388                 struct nf_conntrack_tuple master;
1389                 struct nf_conntrack_tuple_hash *master_h;
1390                 struct nf_conn *master_ct;
1391
1392                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1393                 if (err < 0)
1394                         goto err2;
1395
1396                 master_h = nf_conntrack_find_get(net, zone, &master);
1397                 if (master_h == NULL) {
1398                         err = -ENOENT;
1399                         goto err2;
1400                 }
1401                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1402                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1403                 ct->master = master_ct;
1404         }
1405
1406         add_timer(&ct->timeout);
1407         nf_conntrack_hash_insert(ct);
1408         rcu_read_unlock();
1409
1410         return ct;
1411
1412 err2:
1413         rcu_read_unlock();
1414 err1:
1415         nf_conntrack_free(ct);
1416         return ERR_PTR(err);
1417 }
1418
1419 static int
1420 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1421                         const struct nlmsghdr *nlh,
1422                         const struct nlattr * const cda[])
1423 {
1424         struct net *net = sock_net(ctnl);
1425         struct nf_conntrack_tuple otuple, rtuple;
1426         struct nf_conntrack_tuple_hash *h = NULL;
1427         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1428         u_int8_t u3 = nfmsg->nfgen_family;
1429         u16 zone;
1430         int err;
1431
1432         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1433         if (err < 0)
1434                 return err;
1435
1436         if (cda[CTA_TUPLE_ORIG]) {
1437                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1438                 if (err < 0)
1439                         return err;
1440         }
1441
1442         if (cda[CTA_TUPLE_REPLY]) {
1443                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1444                 if (err < 0)
1445                         return err;
1446         }
1447
1448         spin_lock_bh(&nf_conntrack_lock);
1449         if (cda[CTA_TUPLE_ORIG])
1450                 h = __nf_conntrack_find(net, zone, &otuple);
1451         else if (cda[CTA_TUPLE_REPLY])
1452                 h = __nf_conntrack_find(net, zone, &rtuple);
1453
1454         if (h == NULL) {
1455                 err = -ENOENT;
1456                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1457                         struct nf_conn *ct;
1458                         enum ip_conntrack_events events;
1459
1460                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1461                                                         &rtuple, u3);
1462                         if (IS_ERR(ct)) {
1463                                 err = PTR_ERR(ct);
1464                                 goto out_unlock;
1465                         }
1466                         err = 0;
1467                         nf_conntrack_get(&ct->ct_general);
1468                         spin_unlock_bh(&nf_conntrack_lock);
1469                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1470                                 events = IPCT_RELATED;
1471                         else
1472                                 events = IPCT_NEW;
1473
1474                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1475                                                       (1 << IPCT_ASSURED) |
1476                                                       (1 << IPCT_HELPER) |
1477                                                       (1 << IPCT_PROTOINFO) |
1478                                                       (1 << IPCT_NATSEQADJ) |
1479                                                       (1 << IPCT_MARK) | events,
1480                                                       ct, NETLINK_CB(skb).pid,
1481                                                       nlmsg_report(nlh));
1482                         nf_ct_put(ct);
1483                 } else
1484                         spin_unlock_bh(&nf_conntrack_lock);
1485
1486                 return err;
1487         }
1488         /* implicit 'else' */
1489
1490         /* We manipulate the conntrack inside the global conntrack table lock,
1491          * so there's no need to increase the refcount */
1492         err = -EEXIST;
1493         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1494                 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1495
1496                 err = ctnetlink_change_conntrack(ct, cda);
1497                 if (err == 0) {
1498                         nf_conntrack_get(&ct->ct_general);
1499                         spin_unlock_bh(&nf_conntrack_lock);
1500                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1501                                                       (1 << IPCT_ASSURED) |
1502                                                       (1 << IPCT_HELPER) |
1503                                                       (1 << IPCT_PROTOINFO) |
1504                                                       (1 << IPCT_NATSEQADJ) |
1505                                                       (1 << IPCT_MARK),
1506                                                       ct, NETLINK_CB(skb).pid,
1507                                                       nlmsg_report(nlh));
1508                         nf_ct_put(ct);
1509                 } else
1510                         spin_unlock_bh(&nf_conntrack_lock);
1511
1512                 return err;
1513         }
1514
1515 out_unlock:
1516         spin_unlock_bh(&nf_conntrack_lock);
1517         return err;
1518 }
1519
1520 /***********************************************************************
1521  * EXPECT
1522  ***********************************************************************/
1523
1524 static inline int
1525 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1526                          const struct nf_conntrack_tuple *tuple,
1527                          enum ctattr_expect type)
1528 {
1529         struct nlattr *nest_parms;
1530
1531         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
1532         if (!nest_parms)
1533                 goto nla_put_failure;
1534         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1535                 goto nla_put_failure;
1536         nla_nest_end(skb, nest_parms);
1537
1538         return 0;
1539
1540 nla_put_failure:
1541         return -1;
1542 }
1543
1544 static inline int
1545 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1546                         const struct nf_conntrack_tuple *tuple,
1547                         const struct nf_conntrack_tuple_mask *mask)
1548 {
1549         int ret;
1550         struct nf_conntrack_l3proto *l3proto;
1551         struct nf_conntrack_l4proto *l4proto;
1552         struct nf_conntrack_tuple m;
1553         struct nlattr *nest_parms;
1554
1555         memset(&m, 0xFF, sizeof(m));
1556         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1557         m.src.u.all = mask->src.u.all;
1558         m.dst.protonum = tuple->dst.protonum;
1559
1560         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
1561         if (!nest_parms)
1562                 goto nla_put_failure;
1563
1564         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
1565         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1566
1567         if (unlikely(ret < 0))
1568                 goto nla_put_failure;
1569
1570         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
1571         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1572         if (unlikely(ret < 0))
1573                 goto nla_put_failure;
1574
1575         nla_nest_end(skb, nest_parms);
1576
1577         return 0;
1578
1579 nla_put_failure:
1580         return -1;
1581 }
1582
1583 static int
1584 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1585                           const struct nf_conntrack_expect *exp)
1586 {
1587         struct nf_conn *master = exp->master;
1588         long timeout = (exp->timeout.expires - jiffies) / HZ;
1589         struct nf_conn_help *help;
1590
1591         if (timeout < 0)
1592                 timeout = 0;
1593
1594         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1595                 goto nla_put_failure;
1596         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1597                 goto nla_put_failure;
1598         if (ctnetlink_exp_dump_tuple(skb,
1599                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1600                                  CTA_EXPECT_MASTER) < 0)
1601                 goto nla_put_failure;
1602
1603         NLA_PUT_BE32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout));
1604         NLA_PUT_BE32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp));
1605         NLA_PUT_BE32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags));
1606         help = nfct_help(master);
1607         if (help) {
1608                 struct nf_conntrack_helper *helper;
1609
1610                 helper = rcu_dereference(help->helper);
1611                 if (helper)
1612                         NLA_PUT_STRING(skb, CTA_EXPECT_HELP_NAME, helper->name);
1613         }
1614
1615         return 0;
1616
1617 nla_put_failure:
1618         return -1;
1619 }
1620
1621 static int
1622 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1623                         int event, const struct nf_conntrack_expect *exp)
1624 {
1625         struct nlmsghdr *nlh;
1626         struct nfgenmsg *nfmsg;
1627         unsigned int flags = pid ? NLM_F_MULTI : 0;
1628
1629         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1630         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
1631         if (nlh == NULL)
1632                 goto nlmsg_failure;
1633
1634         nfmsg = nlmsg_data(nlh);
1635         nfmsg->nfgen_family = exp->tuple.src.l3num;
1636         nfmsg->version      = NFNETLINK_V0;
1637         nfmsg->res_id       = 0;
1638
1639         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1640                 goto nla_put_failure;
1641
1642         nlmsg_end(skb, nlh);
1643         return skb->len;
1644
1645 nlmsg_failure:
1646 nla_put_failure:
1647         nlmsg_cancel(skb, nlh);
1648         return -1;
1649 }
1650
1651 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1652 static int
1653 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
1654 {
1655         struct nf_conntrack_expect *exp = item->exp;
1656         struct net *net = nf_ct_exp_net(exp);
1657         struct nlmsghdr *nlh;
1658         struct nfgenmsg *nfmsg;
1659         struct sk_buff *skb;
1660         unsigned int type, group;
1661         int flags = 0;
1662
1663         if (events & (1 << IPEXP_DESTROY)) {
1664                 type = IPCTNL_MSG_EXP_DELETE;
1665                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
1666         } else if (events & (1 << IPEXP_NEW)) {
1667                 type = IPCTNL_MSG_EXP_NEW;
1668                 flags = NLM_F_CREATE|NLM_F_EXCL;
1669                 group = NFNLGRP_CONNTRACK_EXP_NEW;
1670         } else
1671                 return 0;
1672
1673         if (!item->report && !nfnetlink_has_listeners(net, group))
1674                 return 0;
1675
1676         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
1677         if (skb == NULL)
1678                 goto errout;
1679
1680         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1681         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
1682         if (nlh == NULL)
1683                 goto nlmsg_failure;
1684
1685         nfmsg = nlmsg_data(nlh);
1686         nfmsg->nfgen_family = exp->tuple.src.l3num;
1687         nfmsg->version      = NFNETLINK_V0;
1688         nfmsg->res_id       = 0;
1689
1690         rcu_read_lock();
1691         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1692                 goto nla_put_failure;
1693         rcu_read_unlock();
1694
1695         nlmsg_end(skb, nlh);
1696         nfnetlink_send(skb, net, item->pid, group, item->report, GFP_ATOMIC);
1697         return 0;
1698
1699 nla_put_failure:
1700         rcu_read_unlock();
1701         nlmsg_cancel(skb, nlh);
1702 nlmsg_failure:
1703         kfree_skb(skb);
1704 errout:
1705         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
1706         return 0;
1707 }
1708 #endif
1709 static int ctnetlink_exp_done(struct netlink_callback *cb)
1710 {
1711         if (cb->args[1])
1712                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1713         return 0;
1714 }
1715
1716 static int
1717 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1718 {
1719         struct net *net = sock_net(skb->sk);
1720         struct nf_conntrack_expect *exp, *last;
1721         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1722         struct hlist_node *n;
1723         u_int8_t l3proto = nfmsg->nfgen_family;
1724
1725         rcu_read_lock();
1726         last = (struct nf_conntrack_expect *)cb->args[1];
1727         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1728 restart:
1729                 hlist_for_each_entry(exp, n, &net->ct.expect_hash[cb->args[0]],
1730                                      hnode) {
1731                         if (l3proto && exp->tuple.src.l3num != l3proto)
1732                                 continue;
1733                         if (cb->args[1]) {
1734                                 if (exp != last)
1735                                         continue;
1736                                 cb->args[1] = 0;
1737                         }
1738                         if (ctnetlink_exp_fill_info(skb,
1739                                                     NETLINK_CB(cb->skb).pid,
1740                                                     cb->nlh->nlmsg_seq,
1741                                                     IPCTNL_MSG_EXP_NEW,
1742                                                     exp) < 0) {
1743                                 if (!atomic_inc_not_zero(&exp->use))
1744                                         continue;
1745                                 cb->args[1] = (unsigned long)exp;
1746                                 goto out;
1747                         }
1748                 }
1749                 if (cb->args[1]) {
1750                         cb->args[1] = 0;
1751                         goto restart;
1752                 }
1753         }
1754 out:
1755         rcu_read_unlock();
1756         if (last)
1757                 nf_ct_expect_put(last);
1758
1759         return skb->len;
1760 }
1761
1762 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1763         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1764         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1765         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1766         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1767         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1768         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING },
1769         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1770         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1771 };
1772
1773 static int
1774 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1775                      const struct nlmsghdr *nlh,
1776                      const struct nlattr * const cda[])
1777 {
1778         struct net *net = sock_net(ctnl);
1779         struct nf_conntrack_tuple tuple;
1780         struct nf_conntrack_expect *exp;
1781         struct sk_buff *skb2;
1782         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1783         u_int8_t u3 = nfmsg->nfgen_family;
1784         u16 zone;
1785         int err;
1786
1787         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1788                 return netlink_dump_start(ctnl, skb, nlh,
1789                                           ctnetlink_exp_dump_table,
1790                                           ctnetlink_exp_done);
1791         }
1792
1793         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1794         if (err < 0)
1795                 return err;
1796
1797         if (cda[CTA_EXPECT_MASTER])
1798                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1799         else
1800                 return -EINVAL;
1801
1802         if (err < 0)
1803                 return err;
1804
1805         exp = nf_ct_expect_find_get(net, zone, &tuple);
1806         if (!exp)
1807                 return -ENOENT;
1808
1809         if (cda[CTA_EXPECT_ID]) {
1810                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1811                 if (ntohl(id) != (u32)(unsigned long)exp) {
1812                         nf_ct_expect_put(exp);
1813                         return -ENOENT;
1814                 }
1815         }
1816
1817         err = -ENOMEM;
1818         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1819         if (skb2 == NULL)
1820                 goto out;
1821
1822         rcu_read_lock();
1823         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1824                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
1825         rcu_read_unlock();
1826         if (err <= 0)
1827                 goto free;
1828
1829         nf_ct_expect_put(exp);
1830
1831         return netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1832
1833 free:
1834         kfree_skb(skb2);
1835 out:
1836         nf_ct_expect_put(exp);
1837         return err;
1838 }
1839
1840 static int
1841 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1842                      const struct nlmsghdr *nlh,
1843                      const struct nlattr * const cda[])
1844 {
1845         struct net *net = sock_net(ctnl);
1846         struct nf_conntrack_expect *exp;
1847         struct nf_conntrack_tuple tuple;
1848         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1849         struct hlist_node *n, *next;
1850         u_int8_t u3 = nfmsg->nfgen_family;
1851         unsigned int i;
1852         u16 zone;
1853         int err;
1854
1855         if (cda[CTA_EXPECT_TUPLE]) {
1856                 /* delete a single expect by tuple */
1857                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1858                 if (err < 0)
1859                         return err;
1860
1861                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1862                 if (err < 0)
1863                         return err;
1864
1865                 /* bump usage count to 2 */
1866                 exp = nf_ct_expect_find_get(net, zone, &tuple);
1867                 if (!exp)
1868                         return -ENOENT;
1869
1870                 if (cda[CTA_EXPECT_ID]) {
1871                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1872                         if (ntohl(id) != (u32)(unsigned long)exp) {
1873                                 nf_ct_expect_put(exp);
1874                                 return -ENOENT;
1875                         }
1876                 }
1877
1878                 /* after list removal, usage count == 1 */
1879                 spin_lock_bh(&nf_conntrack_lock);
1880                 if (del_timer(&exp->timeout)) {
1881                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).pid,
1882                                                    nlmsg_report(nlh));
1883                         nf_ct_expect_put(exp);
1884                 }
1885                 spin_unlock_bh(&nf_conntrack_lock);
1886                 /* have to put what we 'get' above.
1887                  * after this line usage count == 0 */
1888                 nf_ct_expect_put(exp);
1889         } else if (cda[CTA_EXPECT_HELP_NAME]) {
1890                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
1891                 struct nf_conn_help *m_help;
1892
1893                 /* delete all expectations for this helper */
1894                 spin_lock_bh(&nf_conntrack_lock);
1895                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1896                         hlist_for_each_entry_safe(exp, n, next,
1897                                                   &net->ct.expect_hash[i],
1898                                                   hnode) {
1899                                 m_help = nfct_help(exp->master);
1900                                 if (!strcmp(m_help->helper->name, name) &&
1901                                     del_timer(&exp->timeout)) {
1902                                         nf_ct_unlink_expect_report(exp,
1903                                                         NETLINK_CB(skb).pid,
1904                                                         nlmsg_report(nlh));
1905                                         nf_ct_expect_put(exp);
1906                                 }
1907                         }
1908                 }
1909                 spin_unlock_bh(&nf_conntrack_lock);
1910         } else {
1911                 /* This basically means we have to flush everything*/
1912                 spin_lock_bh(&nf_conntrack_lock);
1913                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1914                         hlist_for_each_entry_safe(exp, n, next,
1915                                                   &net->ct.expect_hash[i],
1916                                                   hnode) {
1917                                 if (del_timer(&exp->timeout)) {
1918                                         nf_ct_unlink_expect_report(exp,
1919                                                         NETLINK_CB(skb).pid,
1920                                                         nlmsg_report(nlh));
1921                                         nf_ct_expect_put(exp);
1922                                 }
1923                         }
1924                 }
1925                 spin_unlock_bh(&nf_conntrack_lock);
1926         }
1927
1928         return 0;
1929 }
1930 static int
1931 ctnetlink_change_expect(struct nf_conntrack_expect *x,
1932                         const struct nlattr * const cda[])
1933 {
1934         return -EOPNOTSUPP;
1935 }
1936
1937 static int
1938 ctnetlink_create_expect(struct net *net, u16 zone,
1939                         const struct nlattr * const cda[],
1940                         u_int8_t u3,
1941                         u32 pid, int report)
1942 {
1943         struct nf_conntrack_tuple tuple, mask, master_tuple;
1944         struct nf_conntrack_tuple_hash *h = NULL;
1945         struct nf_conntrack_expect *exp;
1946         struct nf_conn *ct;
1947         struct nf_conn_help *help;
1948         int err = 0;
1949
1950         /* caller guarantees that those three CTA_EXPECT_* exist */
1951         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1952         if (err < 0)
1953                 return err;
1954         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
1955         if (err < 0)
1956                 return err;
1957         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
1958         if (err < 0)
1959                 return err;
1960
1961         /* Look for master conntrack of this expectation */
1962         h = nf_conntrack_find_get(net, zone, &master_tuple);
1963         if (!h)
1964                 return -ENOENT;
1965         ct = nf_ct_tuplehash_to_ctrack(h);
1966         exp = nf_ct_expect_alloc(ct);
1967         if (!exp) {
1968                 err = -ENOMEM;
1969                 goto out;
1970         }
1971         help = nfct_help(ct);
1972         if (!help) {
1973                 if (!cda[CTA_EXPECT_TIMEOUT]) {
1974                         err = -EINVAL;
1975                         goto out;
1976                 }
1977                 exp->timeout.expires =
1978                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
1979
1980                 exp->flags = NF_CT_EXPECT_USERSPACE;
1981                 if (cda[CTA_EXPECT_FLAGS]) {
1982                         exp->flags |=
1983                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
1984                 }
1985         } else {
1986                 if (cda[CTA_EXPECT_FLAGS]) {
1987                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
1988                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
1989                 } else
1990                         exp->flags = 0;
1991         }
1992
1993         exp->class = 0;
1994         exp->expectfn = NULL;
1995         exp->master = ct;
1996         exp->helper = NULL;
1997         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
1998         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
1999         exp->mask.src.u.all = mask.src.u.all;
2000
2001         err = nf_ct_expect_related_report(exp, pid, report);
2002         nf_ct_expect_put(exp);
2003
2004 out:
2005         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2006         return err;
2007 }
2008
2009 static int
2010 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2011                      const struct nlmsghdr *nlh,
2012                      const struct nlattr * const cda[])
2013 {
2014         struct net *net = sock_net(ctnl);
2015         struct nf_conntrack_tuple tuple;
2016         struct nf_conntrack_expect *exp;
2017         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2018         u_int8_t u3 = nfmsg->nfgen_family;
2019         u16 zone;
2020         int err;
2021
2022         if (!cda[CTA_EXPECT_TUPLE]
2023             || !cda[CTA_EXPECT_MASK]
2024             || !cda[CTA_EXPECT_MASTER])
2025                 return -EINVAL;
2026
2027         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2028         if (err < 0)
2029                 return err;
2030
2031         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2032         if (err < 0)
2033                 return err;
2034
2035         spin_lock_bh(&nf_conntrack_lock);
2036         exp = __nf_ct_expect_find(net, zone, &tuple);
2037
2038         if (!exp) {
2039                 spin_unlock_bh(&nf_conntrack_lock);
2040                 err = -ENOENT;
2041                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2042                         err = ctnetlink_create_expect(net, zone, cda,
2043                                                       u3,
2044                                                       NETLINK_CB(skb).pid,
2045                                                       nlmsg_report(nlh));
2046                 }
2047                 return err;
2048         }
2049
2050         err = -EEXIST;
2051         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2052                 err = ctnetlink_change_expect(exp, cda);
2053         spin_unlock_bh(&nf_conntrack_lock);
2054
2055         return err;
2056 }
2057
2058 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2059 static struct nf_ct_event_notifier ctnl_notifier = {
2060         .fcn = ctnetlink_conntrack_event,
2061 };
2062
2063 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2064         .fcn = ctnetlink_expect_event,
2065 };
2066 #endif
2067
2068 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2069         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2070                                             .attr_count = CTA_MAX,
2071                                             .policy = ct_nla_policy },
2072         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2073                                             .attr_count = CTA_MAX,
2074                                             .policy = ct_nla_policy },
2075         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2076                                             .attr_count = CTA_MAX,
2077                                             .policy = ct_nla_policy },
2078         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
2079                                             .attr_count = CTA_MAX,
2080                                             .policy = ct_nla_policy },
2081 };
2082
2083 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
2084         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
2085                                             .attr_count = CTA_EXPECT_MAX,
2086                                             .policy = exp_nla_policy },
2087         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
2088                                             .attr_count = CTA_EXPECT_MAX,
2089                                             .policy = exp_nla_policy },
2090         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
2091                                             .attr_count = CTA_EXPECT_MAX,
2092                                             .policy = exp_nla_policy },
2093 };
2094
2095 static const struct nfnetlink_subsystem ctnl_subsys = {
2096         .name                           = "conntrack",
2097         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
2098         .cb_count                       = IPCTNL_MSG_MAX,
2099         .cb                             = ctnl_cb,
2100 };
2101
2102 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
2103         .name                           = "conntrack_expect",
2104         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
2105         .cb_count                       = IPCTNL_MSG_EXP_MAX,
2106         .cb                             = ctnl_exp_cb,
2107 };
2108
2109 MODULE_ALIAS("ip_conntrack_netlink");
2110 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
2111 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
2112
2113 static int __init ctnetlink_init(void)
2114 {
2115         int ret;
2116
2117         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
2118         ret = nfnetlink_subsys_register(&ctnl_subsys);
2119         if (ret < 0) {
2120                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
2121                 goto err_out;
2122         }
2123
2124         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
2125         if (ret < 0) {
2126                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
2127                 goto err_unreg_subsys;
2128         }
2129
2130 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2131         ret = nf_conntrack_register_notifier(&ctnl_notifier);
2132         if (ret < 0) {
2133                 pr_err("ctnetlink_init: cannot register notifier.\n");
2134                 goto err_unreg_exp_subsys;
2135         }
2136
2137         ret = nf_ct_expect_register_notifier(&ctnl_notifier_exp);
2138         if (ret < 0) {
2139                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
2140                 goto err_unreg_notifier;
2141         }
2142 #endif
2143
2144         return 0;
2145
2146 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2147 err_unreg_notifier:
2148         nf_conntrack_unregister_notifier(&ctnl_notifier);
2149 err_unreg_exp_subsys:
2150         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2151 #endif
2152 err_unreg_subsys:
2153         nfnetlink_subsys_unregister(&ctnl_subsys);
2154 err_out:
2155         return ret;
2156 }
2157
2158 static void __exit ctnetlink_exit(void)
2159 {
2160         pr_info("ctnetlink: unregistering from nfnetlink.\n");
2161
2162         nf_ct_remove_userspace_expectations();
2163 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2164         nf_ct_expect_unregister_notifier(&ctnl_notifier_exp);
2165         nf_conntrack_unregister_notifier(&ctnl_notifier);
2166 #endif
2167
2168         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2169         nfnetlink_subsys_unregister(&ctnl_subsys);
2170 }
2171
2172 module_init(ctnetlink_init);
2173 module_exit(ctnetlink_exit);