net/netfilter/nf_conntrack_netlink.c

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