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