SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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) 2005-2006 by Pablo Neira Ayuso <pablo@eurodev.net>
   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/types.h>
  22 #include <linux/timer.h>
  23 #include <linux/skbuff.h>
  24 #include <linux/errno.h>
  25 #include <linux/netlink.h>
  26 #include <linux/spinlock.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/notifier.h>
  29
  30 #include <linux/netfilter.h>
  31 #include <net/netlink.h>
  32 #include <net/netfilter/nf_conntrack.h>
  33 #include <net/netfilter/nf_conntrack_core.h>
  34 #include <net/netfilter/nf_conntrack_expect.h>
  35 #include <net/netfilter/nf_conntrack_helper.h>
  36 #include <net/netfilter/nf_conntrack_l3proto.h>
  37 #include <net/netfilter/nf_conntrack_l4proto.h>
  38 #include <net/netfilter/nf_conntrack_tuple.h>
  39 #ifdef CONFIG_NF_NAT_NEEDED
  40 #include <net/netfilter/nf_nat_core.h>
  41 #include <net/netfilter/nf_nat_protocol.h>
  42 #endif
  43
  44 #include <linux/netfilter/nfnetlink.h>
  45 #include <linux/netfilter/nfnetlink_conntrack.h>
  46
  47 MODULE_LICENSE("GPL");
  48
  49 static char __initdata version[] = "0.93";
  50
  51 static inline int
  52 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  53                             const struct nf_conntrack_tuple *tuple,
  54                             struct nf_conntrack_l4proto *l4proto)
  55 {
  56         int ret = 0;
  57         struct nlattr *nest_parms;
  58
  59         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  60         if (!nest_parms)
  61                 goto nla_put_failure;
  62         NLA_PUT(skb, CTA_PROTO_NUM, sizeof(u_int8_t), &tuple->dst.protonum);
  63
  64         if (likely(l4proto->tuple_to_nfattr))
  65                 ret = l4proto->tuple_to_nfattr(skb, tuple);
  66
  67         nla_nest_end(skb, nest_parms);
  68
  69         return ret;
  70
  71 nla_put_failure:
  72         return -1;
  73 }
  74
  75 static inline int
  76 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  77                          const struct nf_conntrack_tuple *tuple,
  78                          struct nf_conntrack_l3proto *l3proto)
  79 {
  80         int ret = 0;
  81         struct nlattr *nest_parms;
  82
  83         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  84         if (!nest_parms)
  85                 goto nla_put_failure;
  86
  87         if (likely(l3proto->tuple_to_nfattr))
  88                 ret = l3proto->tuple_to_nfattr(skb, tuple);
  89
  90         nla_nest_end(skb, nest_parms);
  91
  92         return ret;
  93
  94 nla_put_failure:
  95         return -1;
  96 }
  97
  98 static inline int
  99 ctnetlink_dump_tuples(struct sk_buff *skb,
 100                       const struct nf_conntrack_tuple *tuple)
 101 {
 102         int ret;
 103         struct nf_conntrack_l3proto *l3proto;
 104         struct nf_conntrack_l4proto *l4proto;
 105
 106         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
 107         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 108         nf_ct_l3proto_put(l3proto);
 109
 110         if (unlikely(ret < 0))
 111                 return ret;
 112
 113         l4proto = nf_ct_l4proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 114         ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 115         nf_ct_l4proto_put(l4proto);
 116
 117         return ret;
 118 }
 119
 120 static inline int
 121 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 122 {
 123         __be32 status = htonl((u_int32_t) ct->status);
 124         NLA_PUT(skb, CTA_STATUS, sizeof(status), &status);
 125         return 0;
 126
 127 nla_put_failure:
 128         return -1;
 129 }
 130
 131 static inline int
 132 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 133 {
 134         long timeout_l = ct->timeout.expires - jiffies;
 135         __be32 timeout;
 136
 137         if (timeout_l < 0)
 138                 timeout = 0;
 139         else
 140                 timeout = htonl(timeout_l / HZ);
 141
 142         NLA_PUT(skb, CTA_TIMEOUT, sizeof(timeout), &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, const struct nf_conn *ct)
 151 {
 152         struct nf_conntrack_l4proto *l4proto = nf_ct_l4proto_find_get(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num, ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum);
 153         struct nlattr *nest_proto;
 154         int ret;
 155
 156         if (!l4proto->to_nfattr) {
 157                 nf_ct_l4proto_put(l4proto);
 158                 return 0;
 159         }
 160
 161         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 162         if (!nest_proto)
 163                 goto nla_put_failure;
 164
 165         ret = l4proto->to_nfattr(skb, nest_proto, ct);
 166
 167         nf_ct_l4proto_put(l4proto);
 168
 169         nla_nest_end(skb, nest_proto);
 170
 171         return ret;
 172
 173 nla_put_failure:
 174         nf_ct_l4proto_put(l4proto);
 175         return -1;
 176 }
 177
 178 static inline int
 179 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 180 {
 181         struct nlattr *nest_helper;
 182         const struct nf_conn_help *help = nfct_help(ct);
 183         struct nf_conntrack_helper *helper;
 184
 185         if (!help)
 186                 return 0;
 187
 188         rcu_read_lock();
 189         helper = rcu_dereference(help->helper);
 190         if (!helper)
 191                 goto out;
 192
 193         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 194         if (!nest_helper)
 195                 goto nla_put_failure;
 196         NLA_PUT(skb, CTA_HELP_NAME, strlen(helper->name), helper->name);
 197
 198         if (helper->to_nfattr)
 199                 helper->to_nfattr(skb, ct);
 200
 201         nla_nest_end(skb, nest_helper);
 202 out:
 203         rcu_read_unlock();
 204         return 0;
 205
 206 nla_put_failure:
 207         rcu_read_unlock();
 208         return -1;
 209 }
 210
 211 #ifdef CONFIG_NF_CT_ACCT
 212 static inline int
 213 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 214                         enum ip_conntrack_dir dir)
 215 {
 216         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 217         struct nlattr *nest_count;
 218         __be32 tmp;
 219
 220         nest_count = nla_nest_start(skb, type | NLA_F_NESTED);
 221         if (!nest_count)
 222                 goto nla_put_failure;
 223
 224         tmp = htonl(ct->counters[dir].packets);
 225         NLA_PUT(skb, CTA_COUNTERS32_PACKETS, sizeof(u_int32_t), &tmp);
 226
 227         tmp = htonl(ct->counters[dir].bytes);
 228         NLA_PUT(skb, CTA_COUNTERS32_BYTES, sizeof(u_int32_t), &tmp);
 229
 230         nla_nest_end(skb, nest_count);
 231
 232         return 0;
 233
 234 nla_put_failure:
 235         return -1;
 236 }
 237 #else
 238 #define ctnetlink_dump_counters(a, b, c) (0)
 239 #endif
 240
 241 #ifdef CONFIG_NF_CONNTRACK_MARK
 242 static inline int
 243 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 244 {
 245         __be32 mark = htonl(ct->mark);
 246
 247         NLA_PUT(skb, CTA_MARK, sizeof(u_int32_t), &mark);
 248         return 0;
 249
 250 nla_put_failure:
 251         return -1;
 252 }
 253 #else
 254 #define ctnetlink_dump_mark(a, b) (0)
 255 #endif
 256
 257 static inline int
 258 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 259 {
 260         __be32 id = htonl(ct->id);
 261         NLA_PUT(skb, CTA_ID, sizeof(u_int32_t), &id);
 262         return 0;
 263
 264 nla_put_failure:
 265         return -1;
 266 }
 267
 268 static inline int
 269 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 270 {
 271         __be32 use = htonl(atomic_read(&ct->ct_general.use));
 272
 273         NLA_PUT(skb, CTA_USE, sizeof(u_int32_t), &use);
 274         return 0;
 275
 276 nla_put_failure:
 277         return -1;
 278 }
 279
 280 #define tuple(ct, dir) (&(ct)->tuplehash[dir].tuple)
 281
 282 static int
 283 ctnetlink_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
 284                     int event, int nowait,
 285                     const struct nf_conn *ct)
 286 {
 287         struct nlmsghdr *nlh;
 288         struct nfgenmsg *nfmsg;
 289         struct nlattr *nest_parms;
 290         unsigned char *b = skb_tail_pointer(skb);
 291
 292         event |= NFNL_SUBSYS_CTNETLINK << 8;
 293         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
 294         nfmsg  = NLMSG_DATA(nlh);
 295
 296         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
 297         nfmsg->nfgen_family =
 298                 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 299         nfmsg->version      = NFNETLINK_V0;
 300         nfmsg->res_id       = 0;
 301
 302         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 303         if (!nest_parms)
 304                 goto nla_put_failure;
 305         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 306                 goto nla_put_failure;
 307         nla_nest_end(skb, nest_parms);
 308
 309         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 310         if (!nest_parms)
 311                 goto nla_put_failure;
 312         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_REPLY)) < 0)
 313                 goto nla_put_failure;
 314         nla_nest_end(skb, nest_parms);
 315
 316         if (ctnetlink_dump_status(skb, ct) < 0 ||
 317             ctnetlink_dump_timeout(skb, ct) < 0 ||
 318             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 319             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 320             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 321             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 322             ctnetlink_dump_mark(skb, ct) < 0 ||
 323             ctnetlink_dump_id(skb, ct) < 0 ||
 324             ctnetlink_dump_use(skb, ct) < 0)
 325                 goto nla_put_failure;
 326
 327         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 328         return skb->len;
 329
 330 nlmsg_failure:
 331 nla_put_failure:
 332         nlmsg_trim(skb, b);
 333         return -1;
 334 }
 335
 336 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 337 static int ctnetlink_conntrack_event(struct notifier_block *this,
 338                                      unsigned long events, void *ptr)
 339 {
 340         struct nlmsghdr *nlh;
 341         struct nfgenmsg *nfmsg;
 342         struct nlattr *nest_parms;
 343         struct nf_conn *ct = (struct nf_conn *)ptr;
 344         struct sk_buff *skb;
 345         unsigned int type;
 346         sk_buff_data_t b;
 347         unsigned int flags = 0, group;
 348
 349         /* ignore our fake conntrack entry */
 350         if (ct == &nf_conntrack_untracked)
 351                 return NOTIFY_DONE;
 352
 353         if (events & IPCT_DESTROY) {
 354                 type = IPCTNL_MSG_CT_DELETE;
 355                 group = NFNLGRP_CONNTRACK_DESTROY;
 356         } else  if (events & (IPCT_NEW | IPCT_RELATED)) {
 357                 type = IPCTNL_MSG_CT_NEW;
 358                 flags = NLM_F_CREATE|NLM_F_EXCL;
 359                 group = NFNLGRP_CONNTRACK_NEW;
 360         } else  if (events & (IPCT_STATUS | IPCT_PROTOINFO)) {
 361                 type = IPCTNL_MSG_CT_NEW;
 362                 group = NFNLGRP_CONNTRACK_UPDATE;
 363         } else
 364                 return NOTIFY_DONE;
 365
 366         if (!nfnetlink_has_listeners(group))
 367                 return NOTIFY_DONE;
 368
 369         skb = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
 370         if (!skb)
 371                 return NOTIFY_DONE;
 372
 373         b = skb->tail;
 374
 375         type |= NFNL_SUBSYS_CTNETLINK << 8;
 376         nlh   = NLMSG_PUT(skb, 0, 0, type, sizeof(struct nfgenmsg));
 377         nfmsg = NLMSG_DATA(nlh);
 378
 379         nlh->nlmsg_flags    = flags;
 380         nfmsg->nfgen_family = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 381         nfmsg->version  = NFNETLINK_V0;
 382         nfmsg->res_id   = 0;
 383
 384         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 385         if (!nest_parms)
 386                 goto nla_put_failure;
 387         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 388                 goto nla_put_failure;
 389         nla_nest_end(skb, nest_parms);
 390
 391         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 392         if (!nest_parms)
 393                 goto nla_put_failure;
 394         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_REPLY)) < 0)
 395                 goto nla_put_failure;
 396         nla_nest_end(skb, nest_parms);
 397
 398         if (events & IPCT_DESTROY) {
 399                 if (ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 400                     ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0)
 401                         goto nla_put_failure;
 402         } else {
 403                 if (ctnetlink_dump_status(skb, ct) < 0)
 404                         goto nla_put_failure;
 405
 406                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 407                         goto nla_put_failure;
 408
 409                 if (events & IPCT_PROTOINFO
 410                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 411                         goto nla_put_failure;
 412
 413                 if ((events & IPCT_HELPER || nfct_help(ct))
 414                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 415                         goto nla_put_failure;
 416
 417 #ifdef CONFIG_NF_CONNTRACK_MARK
 418                 if ((events & IPCT_MARK || ct->mark)
 419                     && ctnetlink_dump_mark(skb, ct) < 0)
 420                         goto nla_put_failure;
 421 #endif
 422
 423                 if (events & IPCT_COUNTER_FILLING &&
 424                     (ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 425                      ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0))
 426                         goto nla_put_failure;
 427         }
 428
 429         nlh->nlmsg_len = skb->tail - b;
 430         nfnetlink_send(skb, 0, group, 0);
 431         return NOTIFY_DONE;
 432
 433 nlmsg_failure:
 434 nla_put_failure:
 435         kfree_skb(skb);
 436         return NOTIFY_DONE;
 437 }
 438 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 439
 440 static int ctnetlink_done(struct netlink_callback *cb)
 441 {
 442         if (cb->args[1])
 443                 nf_ct_put((struct nf_conn *)cb->args[1]);
 444         return 0;
 445 }
 446
 447 #define L3PROTO(ct) ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num
 448
 449 static int
 450 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 451 {
 452         struct nf_conn *ct, *last;
 453         struct nf_conntrack_tuple_hash *h;
 454         struct hlist_node *n;
 455         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
 456         u_int8_t l3proto = nfmsg->nfgen_family;
 457
 458         read_lock_bh(&nf_conntrack_lock);
 459         last = (struct nf_conn *)cb->args[1];
 460         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 461 restart:
 462                 hlist_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 463                                      hnode) {
 464                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 465                                 continue;
 466                         ct = nf_ct_tuplehash_to_ctrack(h);
 467                         /* Dump entries of a given L3 protocol number.
 468                          * If it is not specified, ie. l3proto == 0,
 469                          * then dump everything. */
 470                         if (l3proto && L3PROTO(ct) != l3proto)
 471                                 continue;
 472                         if (cb->args[1]) {
 473                                 if (ct != last)
 474                                         continue;
 475                                 cb->args[1] = 0;
 476                         }
 477                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 478                                                 cb->nlh->nlmsg_seq,
 479                                                 IPCTNL_MSG_CT_NEW,
 480                                                 1, ct) < 0) {
 481                                 nf_conntrack_get(&ct->ct_general);
 482                                 cb->args[1] = (unsigned long)ct;
 483                                 goto out;
 484                         }
 485 #ifdef CONFIG_NF_CT_ACCT
 486                         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) ==
 487                                                 IPCTNL_MSG_CT_GET_CTRZERO)
 488                                 memset(&ct->counters, 0, sizeof(ct->counters));
 489 #endif
 490                 }
 491                 if (cb->args[1]) {
 492                         cb->args[1] = 0;
 493                         goto restart;
 494                 }
 495         }
 496 out:
 497         read_unlock_bh(&nf_conntrack_lock);
 498         if (last)
 499                 nf_ct_put(last);
 500
 501         return skb->len;
 502 }
 503
 504 static inline int
 505 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 506 {
 507         struct nlattr *tb[CTA_IP_MAX+1];
 508         struct nf_conntrack_l3proto *l3proto;
 509         int ret = 0;
 510
 511         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 512
 513         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
 514
 515         if (likely(l3proto->nfattr_to_tuple))
 516                 ret = l3proto->nfattr_to_tuple(tb, tuple);
 517
 518         nf_ct_l3proto_put(l3proto);
 519
 520         return ret;
 521 }
 522
 523 static const size_t cta_min_proto[CTA_PROTO_MAX+1] = {
 524         [CTA_PROTO_NUM] = sizeof(u_int8_t),
 525 };
 526
 527 static inline int
 528 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 529                             struct nf_conntrack_tuple *tuple)
 530 {
 531         struct nlattr *tb[CTA_PROTO_MAX+1];
 532         struct nf_conntrack_l4proto *l4proto;
 533         int ret = 0;
 534
 535         nla_parse_nested(tb, CTA_PROTO_MAX, attr, NULL);
 536
 537         if (nfattr_bad_size(tb, CTA_PROTO_MAX, cta_min_proto))
 538                 return -EINVAL;
 539
 540         if (!tb[CTA_PROTO_NUM])
 541                 return -EINVAL;
 542         tuple->dst.protonum = *(u_int8_t *)nla_data(tb[CTA_PROTO_NUM]);
 543
 544         l4proto = nf_ct_l4proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 545
 546         if (likely(l4proto->nfattr_to_tuple))
 547                 ret = l4proto->nfattr_to_tuple(tb, tuple);
 548
 549         nf_ct_l4proto_put(l4proto);
 550
 551         return ret;
 552 }
 553
 554 static inline int
 555 ctnetlink_parse_tuple(struct nlattr *cda[], struct nf_conntrack_tuple *tuple,
 556                       enum ctattr_tuple type, u_int8_t l3num)
 557 {
 558         struct nlattr *tb[CTA_TUPLE_MAX+1];
 559         int err;
 560
 561         memset(tuple, 0, sizeof(*tuple));
 562
 563         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], NULL);
 564
 565         if (!tb[CTA_TUPLE_IP])
 566                 return -EINVAL;
 567
 568         tuple->src.l3num = l3num;
 569
 570         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 571         if (err < 0)
 572                 return err;
 573
 574         if (!tb[CTA_TUPLE_PROTO])
 575                 return -EINVAL;
 576
 577         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 578         if (err < 0)
 579                 return err;
 580
 581         /* orig and expect tuples get DIR_ORIGINAL */
 582         if (type == CTA_TUPLE_REPLY)
 583                 tuple->dst.dir = IP_CT_DIR_REPLY;
 584         else
 585                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 586
 587         return 0;
 588 }
 589
 590 #ifdef CONFIG_NF_NAT_NEEDED
 591 static const size_t cta_min_protonat[CTA_PROTONAT_MAX+1] = {
 592         [CTA_PROTONAT_PORT_MIN] = sizeof(u_int16_t),
 593         [CTA_PROTONAT_PORT_MAX] = sizeof(u_int16_t),
 594 };
 595
 596 static int nfnetlink_parse_nat_proto(struct nlattr *attr,
 597                                      const struct nf_conn *ct,
 598                                      struct nf_nat_range *range)
 599 {
 600         struct nlattr *tb[CTA_PROTONAT_MAX+1];
 601         struct nf_nat_protocol *npt;
 602
 603         nla_parse_nested(tb, CTA_PROTONAT_MAX, attr, NULL);
 604
 605         if (nfattr_bad_size(tb, CTA_PROTONAT_MAX, cta_min_protonat))
 606                 return -EINVAL;
 607
 608         npt = nf_nat_proto_find_get(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum);
 609
 610         if (!npt->nfattr_to_range) {
 611                 nf_nat_proto_put(npt);
 612                 return 0;
 613         }
 614
 615         /* nfattr_to_range returns 1 if it parsed, 0 if not, neg. on error */
 616         if (npt->nfattr_to_range(tb, range) > 0)
 617                 range->flags |= IP_NAT_RANGE_PROTO_SPECIFIED;
 618
 619         nf_nat_proto_put(npt);
 620
 621         return 0;
 622 }
 623
 624 static const size_t cta_min_nat[CTA_NAT_MAX+1] = {
 625         [CTA_NAT_MINIP]         = sizeof(u_int32_t),
 626         [CTA_NAT_MAXIP]         = sizeof(u_int32_t),
 627 };
 628
 629 static inline int
 630 nfnetlink_parse_nat(struct nlattr *nat,
 631                     const struct nf_conn *ct, struct nf_nat_range *range)
 632 {
 633         struct nlattr *tb[CTA_NAT_MAX+1];
 634         int err;
 635
 636         memset(range, 0, sizeof(*range));
 637
 638         nla_parse_nested(tb, CTA_NAT_MAX, nat, NULL);
 639
 640         if (nfattr_bad_size(tb, CTA_NAT_MAX, cta_min_nat))
 641                 return -EINVAL;
 642
 643         if (tb[CTA_NAT_MINIP])
 644                 range->min_ip = *(__be32 *)nla_data(tb[CTA_NAT_MINIP]);
 645
 646         if (!tb[CTA_NAT_MAXIP])
 647                 range->max_ip = range->min_ip;
 648         else
 649                 range->max_ip = *(__be32 *)nla_data(tb[CTA_NAT_MAXIP]);
 650
 651         if (range->min_ip)
 652                 range->flags |= IP_NAT_RANGE_MAP_IPS;
 653
 654         if (!tb[CTA_NAT_PROTO])
 655                 return 0;
 656
 657         err = nfnetlink_parse_nat_proto(tb[CTA_NAT_PROTO], ct, range);
 658         if (err < 0)
 659                 return err;
 660
 661         return 0;
 662 }
 663 #endif
 664
 665 static inline int
 666 ctnetlink_parse_help(struct nlattr *attr, char **helper_name)
 667 {
 668         struct nlattr *tb[CTA_HELP_MAX+1];
 669
 670         nla_parse_nested(tb, CTA_HELP_MAX, attr, NULL);
 671
 672         if (!tb[CTA_HELP_NAME])
 673                 return -EINVAL;
 674
 675         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 676
 677         return 0;
 678 }
 679
 680 static const size_t cta_min[CTA_MAX+1] = {
 681         [CTA_STATUS]            = sizeof(u_int32_t),
 682         [CTA_TIMEOUT]           = sizeof(u_int32_t),
 683         [CTA_MARK]              = sizeof(u_int32_t),
 684         [CTA_USE]               = sizeof(u_int32_t),
 685         [CTA_ID]                = sizeof(u_int32_t)
 686 };
 687
 688 static int
 689 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 690                         struct nlmsghdr *nlh, struct nlattr *cda[])
 691 {
 692         struct nf_conntrack_tuple_hash *h;
 693         struct nf_conntrack_tuple tuple;
 694         struct nf_conn *ct;
 695         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 696         u_int8_t u3 = nfmsg->nfgen_family;
 697         int err = 0;
 698
 699         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 700                 return -EINVAL;
 701
 702         if (cda[CTA_TUPLE_ORIG])
 703                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 704         else if (cda[CTA_TUPLE_REPLY])
 705                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 706         else {
 707                 /* Flush the whole table */
 708                 nf_conntrack_flush();
 709                 return 0;
 710         }
 711
 712         if (err < 0)
 713                 return err;
 714
 715         h = nf_conntrack_find_get(&tuple);
 716         if (!h)
 717                 return -ENOENT;
 718
 719         ct = nf_ct_tuplehash_to_ctrack(h);
 720
 721         if (cda[CTA_ID]) {
 722                 u_int32_t id = ntohl(*(__be32 *)nla_data(cda[CTA_ID]));
 723                 if (ct->id != id) {
 724                         nf_ct_put(ct);
 725                         return -ENOENT;
 726                 }
 727         }
 728         if (del_timer(&ct->timeout))
 729                 ct->timeout.function((unsigned long)ct);
 730
 731         nf_ct_put(ct);
 732
 733         return 0;
 734 }
 735
 736 static int
 737 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 738                         struct nlmsghdr *nlh, struct nlattr *cda[])
 739 {
 740         struct nf_conntrack_tuple_hash *h;
 741         struct nf_conntrack_tuple tuple;
 742         struct nf_conn *ct;
 743         struct sk_buff *skb2 = NULL;
 744         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 745         u_int8_t u3 = nfmsg->nfgen_family;
 746         int err = 0;
 747
 748         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 749 #ifndef CONFIG_NF_CT_ACCT
 750                 if (NFNL_MSG_TYPE(nlh->nlmsg_type) == IPCTNL_MSG_CT_GET_CTRZERO)
 751                         return -ENOTSUPP;
 752 #endif
 753                 return netlink_dump_start(ctnl, skb, nlh, ctnetlink_dump_table,
 754                                           ctnetlink_done);
 755         }
 756
 757         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 758                 return -EINVAL;
 759
 760         if (cda[CTA_TUPLE_ORIG])
 761                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 762         else if (cda[CTA_TUPLE_REPLY])
 763                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 764         else
 765                 return -EINVAL;
 766
 767         if (err < 0)
 768                 return err;
 769
 770         h = nf_conntrack_find_get(&tuple);
 771         if (!h)
 772                 return -ENOENT;
 773
 774         ct = nf_ct_tuplehash_to_ctrack(h);
 775
 776         err = -ENOMEM;
 777         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 778         if (!skb2) {
 779                 nf_ct_put(ct);
 780                 return -ENOMEM;
 781         }
 782
 783         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
 784                                   IPCTNL_MSG_CT_NEW, 1, ct);
 785         nf_ct_put(ct);
 786         if (err <= 0)
 787                 goto free;
 788
 789         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
 790         if (err < 0)
 791                 goto out;
 792
 793         return 0;
 794
 795 free:
 796         kfree_skb(skb2);
 797 out:
 798         return err;
 799 }
 800
 801 static inline int
 802 ctnetlink_change_status(struct nf_conn *ct, struct nlattr *cda[])
 803 {
 804         unsigned long d;
 805         unsigned int status = ntohl(*(__be32 *)nla_data(cda[CTA_STATUS]));
 806         d = ct->status ^ status;
 807
 808         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
 809                 /* unchangeable */
 810                 return -EINVAL;
 811
 812         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
 813                 /* SEEN_REPLY bit can only be set */
 814                 return -EINVAL;
 815
 816
 817         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
 818                 /* ASSURED bit can only be set */
 819                 return -EINVAL;
 820
 821         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
 822 #ifndef CONFIG_NF_NAT_NEEDED
 823                 return -EINVAL;
 824 #else
 825                 struct nf_nat_range range;
 826
 827                 if (cda[CTA_NAT_DST]) {
 828                         if (nfnetlink_parse_nat(cda[CTA_NAT_DST], ct,
 829                                                 &range) < 0)
 830                                 return -EINVAL;
 831                         if (nf_nat_initialized(ct,
 832                                                HOOK2MANIP(NF_IP_PRE_ROUTING)))
 833                                 return -EEXIST;
 834                         nf_nat_setup_info(ct, &range, NF_IP_PRE_ROUTING);
 835                 }
 836                 if (cda[CTA_NAT_SRC]) {
 837                         if (nfnetlink_parse_nat(cda[CTA_NAT_SRC], ct,
 838                                                 &range) < 0)
 839                                 return -EINVAL;
 840                         if (nf_nat_initialized(ct,
 841                                                HOOK2MANIP(NF_IP_POST_ROUTING)))
 842                                 return -EEXIST;
 843                         nf_nat_setup_info(ct, &range, NF_IP_POST_ROUTING);
 844                 }
 845 #endif
 846         }
 847
 848         /* Be careful here, modifying NAT bits can screw up things,
 849          * so don't let users modify them directly if they don't pass
 850          * nf_nat_range. */
 851         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
 852         return 0;
 853 }
 854
 855
 856 static inline int
 857 ctnetlink_change_helper(struct nf_conn *ct, struct nlattr *cda[])
 858 {
 859         struct nf_conntrack_helper *helper;
 860         struct nf_conn_help *help = nfct_help(ct);
 861         char *helpname;
 862         int err;
 863
 864         /* don't change helper of sibling connections */
 865         if (ct->master)
 866                 return -EINVAL;
 867
 868         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
 869         if (err < 0)
 870                 return err;
 871
 872         if (!strcmp(helpname, "")) {
 873                 if (help && help->helper) {
 874                         /* we had a helper before ... */
 875                         nf_ct_remove_expectations(ct);
 876                         rcu_assign_pointer(help->helper, NULL);
 877                 }
 878
 879                 return 0;
 880         }
 881
 882         helper = __nf_conntrack_helper_find_byname(helpname);
 883         if (helper == NULL)
 884                 return -EINVAL;
 885
 886         if (help) {
 887                 if (help->helper == helper)
 888                         return 0;
 889                 if (help->helper)
 890                         return -EBUSY;
 891                 /* need to zero data of old helper */
 892                 memset(&help->help, 0, sizeof(help->help));
 893         } else {
 894                 help = nf_ct_helper_ext_add(ct, GFP_KERNEL);
 895                 if (help == NULL)
 896                         return -ENOMEM;
 897         }
 898
 899         rcu_assign_pointer(help->helper, helper);
 900
 901         return 0;
 902 }
 903
 904 static inline int
 905 ctnetlink_change_timeout(struct nf_conn *ct, struct nlattr *cda[])
 906 {
 907         u_int32_t timeout = ntohl(*(__be32 *)nla_data(cda[CTA_TIMEOUT]));
 908
 909         if (!del_timer(&ct->timeout))
 910                 return -ETIME;
 911
 912         ct->timeout.expires = jiffies + timeout * HZ;
 913         add_timer(&ct->timeout);
 914
 915         return 0;
 916 }
 917
 918 static inline int
 919 ctnetlink_change_protoinfo(struct nf_conn *ct, struct nlattr *cda[])
 920 {
 921         struct nlattr *tb[CTA_PROTOINFO_MAX+1], *attr = cda[CTA_PROTOINFO];
 922         struct nf_conntrack_l4proto *l4proto;
 923         u_int16_t npt = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum;
 924         u_int16_t l3num = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 925         int err = 0;
 926
 927         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, NULL);
 928
 929         l4proto = nf_ct_l4proto_find_get(l3num, npt);
 930
 931         if (l4proto->from_nfattr)
 932                 err = l4proto->from_nfattr(tb, ct);
 933         nf_ct_l4proto_put(l4proto);
 934
 935         return err;
 936 }
 937
 938 static int
 939 ctnetlink_change_conntrack(struct nf_conn *ct, struct nlattr *cda[])
 940 {
 941         int err;
 942
 943         if (cda[CTA_HELP]) {
 944                 err = ctnetlink_change_helper(ct, cda);
 945                 if (err < 0)
 946                         return err;
 947         }
 948
 949         if (cda[CTA_TIMEOUT]) {
 950                 err = ctnetlink_change_timeout(ct, cda);
 951                 if (err < 0)
 952                         return err;
 953         }
 954
 955         if (cda[CTA_STATUS]) {
 956                 err = ctnetlink_change_status(ct, cda);
 957                 if (err < 0)
 958                         return err;
 959         }
 960
 961         if (cda[CTA_PROTOINFO]) {
 962                 err = ctnetlink_change_protoinfo(ct, cda);
 963                 if (err < 0)
 964                         return err;
 965         }
 966
 967 #if defined(CONFIG_NF_CONNTRACK_MARK)
 968         if (cda[CTA_MARK])
 969                 ct->mark = ntohl(*(__be32 *)nla_data(cda[CTA_MARK]));
 970 #endif
 971
 972         return 0;
 973 }
 974
 975 static int
 976 ctnetlink_create_conntrack(struct nlattr *cda[],
 977                            struct nf_conntrack_tuple *otuple,
 978                            struct nf_conntrack_tuple *rtuple)
 979 {
 980         struct nf_conn *ct;
 981         int err = -EINVAL;
 982         struct nf_conn_help *help;
 983         struct nf_conntrack_helper *helper;
 984
 985         ct = nf_conntrack_alloc(otuple, rtuple);
 986         if (ct == NULL || IS_ERR(ct))
 987                 return -ENOMEM;
 988
 989         if (!cda[CTA_TIMEOUT])
 990                 goto err;
 991         ct->timeout.expires = ntohl(*(__be32 *)nla_data(cda[CTA_TIMEOUT]));
 992
 993         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
 994         ct->status |= IPS_CONFIRMED;
 995
 996         if (cda[CTA_STATUS]) {
 997                 err = ctnetlink_change_status(ct, cda);
 998                 if (err < 0)
 999                         goto err;
1000         }
1001
1002         if (cda[CTA_PROTOINFO]) {
1003                 err = ctnetlink_change_protoinfo(ct, cda);
1004                 if (err < 0)
1005                         goto err;
1006         }
1007
1008 #if defined(CONFIG_NF_CONNTRACK_MARK)
1009         if (cda[CTA_MARK])
1010                 ct->mark = ntohl(*(__be32 *)nla_data(cda[CTA_MARK]));
1011 #endif
1012
1013         helper = nf_ct_helper_find_get(rtuple);
1014         if (helper) {
1015                 help = nf_ct_helper_ext_add(ct, GFP_KERNEL);
1016                 if (help == NULL) {
1017                         nf_ct_helper_put(helper);
1018                         err = -ENOMEM;
1019                         goto err;
1020                 }
1021                 /* not in hash table yet so not strictly necessary */
1022                 rcu_assign_pointer(help->helper, helper);
1023         }
1024
1025         add_timer(&ct->timeout);
1026         nf_conntrack_hash_insert(ct);
1027
1028         if (helper)
1029                 nf_ct_helper_put(helper);
1030
1031         return 0;
1032
1033 err:
1034         nf_conntrack_free(ct);
1035         return err;
1036 }
1037
1038 static int
1039 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1040                         struct nlmsghdr *nlh, struct nlattr *cda[])
1041 {
1042         struct nf_conntrack_tuple otuple, rtuple;
1043         struct nf_conntrack_tuple_hash *h = NULL;
1044         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1045         u_int8_t u3 = nfmsg->nfgen_family;
1046         int err = 0;
1047
1048         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
1049                 return -EINVAL;
1050
1051         if (cda[CTA_TUPLE_ORIG]) {
1052                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1053                 if (err < 0)
1054                         return err;
1055         }
1056
1057         if (cda[CTA_TUPLE_REPLY]) {
1058                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1059                 if (err < 0)
1060                         return err;
1061         }
1062
1063         write_lock_bh(&nf_conntrack_lock);
1064         if (cda[CTA_TUPLE_ORIG])
1065                 h = __nf_conntrack_find(&otuple, NULL);
1066         else if (cda[CTA_TUPLE_REPLY])
1067                 h = __nf_conntrack_find(&rtuple, NULL);
1068
1069         if (h == NULL) {
1070                 write_unlock_bh(&nf_conntrack_lock);
1071                 err = -ENOENT;
1072                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1073                         err = ctnetlink_create_conntrack(cda, &otuple, &rtuple);
1074                 return err;
1075         }
1076         /* implicit 'else' */
1077
1078         /* We manipulate the conntrack inside the global conntrack table lock,
1079          * so there's no need to increase the refcount */
1080         err = -EEXIST;
1081         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1082                 /* we only allow nat config for new conntracks */
1083                 if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1084                         err = -EINVAL;
1085                         goto out_unlock;
1086                 }
1087                 err = ctnetlink_change_conntrack(nf_ct_tuplehash_to_ctrack(h),
1088                                                  cda);
1089         }
1090
1091 out_unlock:
1092         write_unlock_bh(&nf_conntrack_lock);
1093         return err;
1094 }
1095
1096 /***********************************************************************
1097  * EXPECT
1098  ***********************************************************************/
1099
1100 static inline int
1101 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1102                          const struct nf_conntrack_tuple *tuple,
1103                          enum ctattr_expect type)
1104 {
1105         struct nlattr *nest_parms;
1106
1107         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
1108         if (!nest_parms)
1109                 goto nla_put_failure;
1110         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1111                 goto nla_put_failure;
1112         nla_nest_end(skb, nest_parms);
1113
1114         return 0;
1115
1116 nla_put_failure:
1117         return -1;
1118 }
1119
1120 static inline int
1121 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1122                         const struct nf_conntrack_tuple *tuple,
1123                         const struct nf_conntrack_tuple_mask *mask)
1124 {
1125         int ret;
1126         struct nf_conntrack_l3proto *l3proto;
1127         struct nf_conntrack_l4proto *l4proto;
1128         struct nf_conntrack_tuple m;
1129         struct nlattr *nest_parms;
1130
1131         memset(&m, 0xFF, sizeof(m));
1132         m.src.u.all = mask->src.u.all;
1133         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1134
1135         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
1136         if (!nest_parms)
1137                 goto nla_put_failure;
1138
1139         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
1140         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1141         nf_ct_l3proto_put(l3proto);
1142
1143         if (unlikely(ret < 0))
1144                 goto nla_put_failure;
1145
1146         l4proto = nf_ct_l4proto_find_get(tuple->src.l3num, tuple->dst.protonum);
1147         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1148         nf_ct_l4proto_put(l4proto);
1149         if (unlikely(ret < 0))
1150                 goto nla_put_failure;
1151
1152         nla_nest_end(skb, nest_parms);
1153
1154         return 0;
1155
1156 nla_put_failure:
1157         return -1;
1158 }
1159
1160 static inline int
1161 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1162                           const struct nf_conntrack_expect *exp)
1163 {
1164         struct nf_conn *master = exp->master;
1165         __be32 timeout = htonl((exp->timeout.expires - jiffies) / HZ);
1166         __be32 id = htonl(exp->id);
1167
1168         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1169                 goto nla_put_failure;
1170         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1171                 goto nla_put_failure;
1172         if (ctnetlink_exp_dump_tuple(skb,
1173                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1174                                  CTA_EXPECT_MASTER) < 0)
1175                 goto nla_put_failure;
1176
1177         NLA_PUT(skb, CTA_EXPECT_TIMEOUT, sizeof(timeout), &timeout);
1178         NLA_PUT(skb, CTA_EXPECT_ID, sizeof(u_int32_t), &id);
1179
1180         return 0;
1181
1182 nla_put_failure:
1183         return -1;
1184 }
1185
1186 static int
1187 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1188                     int event,
1189                     int nowait,
1190                     const struct nf_conntrack_expect *exp)
1191 {
1192         struct nlmsghdr *nlh;
1193         struct nfgenmsg *nfmsg;
1194         unsigned char *b = skb_tail_pointer(skb);
1195
1196         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1197         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
1198         nfmsg  = NLMSG_DATA(nlh);
1199
1200         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
1201         nfmsg->nfgen_family = exp->tuple.src.l3num;
1202         nfmsg->version      = NFNETLINK_V0;
1203         nfmsg->res_id       = 0;
1204
1205         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1206                 goto nla_put_failure;
1207
1208         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
1209         return skb->len;
1210
1211 nlmsg_failure:
1212 nla_put_failure:
1213         nlmsg_trim(skb, b);
1214         return -1;
1215 }
1216
1217 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1218 static int ctnetlink_expect_event(struct notifier_block *this,
1219                                   unsigned long events, void *ptr)
1220 {
1221         struct nlmsghdr *nlh;
1222         struct nfgenmsg *nfmsg;
1223         struct nf_conntrack_expect *exp = (struct nf_conntrack_expect *)ptr;
1224         struct sk_buff *skb;
1225         unsigned int type;
1226         sk_buff_data_t b;
1227         int flags = 0;
1228
1229         if (events & IPEXP_NEW) {
1230                 type = IPCTNL_MSG_EXP_NEW;
1231                 flags = NLM_F_CREATE|NLM_F_EXCL;
1232         } else
1233                 return NOTIFY_DONE;
1234
1235         if (!nfnetlink_has_listeners(NFNLGRP_CONNTRACK_EXP_NEW))
1236                 return NOTIFY_DONE;
1237
1238         skb = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1239         if (!skb)
1240                 return NOTIFY_DONE;
1241
1242         b = skb->tail;
1243
1244         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1245         nlh   = NLMSG_PUT(skb, 0, 0, type, sizeof(struct nfgenmsg));
1246         nfmsg = NLMSG_DATA(nlh);
1247
1248         nlh->nlmsg_flags    = flags;
1249         nfmsg->nfgen_family = exp->tuple.src.l3num;
1250         nfmsg->version      = NFNETLINK_V0;
1251         nfmsg->res_id       = 0;
1252
1253         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1254                 goto nla_put_failure;
1255
1256         nlh->nlmsg_len = skb->tail - b;
1257         nfnetlink_send(skb, 0, NFNLGRP_CONNTRACK_EXP_NEW, 0);
1258         return NOTIFY_DONE;
1259
1260 nlmsg_failure:
1261 nla_put_failure:
1262         kfree_skb(skb);
1263         return NOTIFY_DONE;
1264 }
1265 #endif
1266 static int ctnetlink_exp_done(struct netlink_callback *cb)
1267 {
1268         if (cb->args[1])
1269                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1270         return 0;
1271 }
1272
1273 static int
1274 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1275 {
1276         struct nf_conntrack_expect *exp, *last;
1277         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
1278         struct hlist_node *n;
1279         u_int8_t l3proto = nfmsg->nfgen_family;
1280
1281         read_lock_bh(&nf_conntrack_lock);
1282         last = (struct nf_conntrack_expect *)cb->args[1];
1283         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1284 restart:
1285                 hlist_for_each_entry(exp, n, &nf_ct_expect_hash[cb->args[0]],
1286                                      hnode) {
1287                         if (l3proto && exp->tuple.src.l3num != l3proto)
1288                                 continue;
1289                         if (cb->args[1]) {
1290                                 if (exp != last)
1291                                         continue;
1292                                 cb->args[1] = 0;
1293                         }
1294                         if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).pid,
1295                                                     cb->nlh->nlmsg_seq,
1296                                                     IPCTNL_MSG_EXP_NEW,
1297                                                     1, exp) < 0) {
1298                                 atomic_inc(&exp->use);
1299                                 cb->args[1] = (unsigned long)exp;
1300                                 goto out;
1301                         }
1302                 }
1303                 if (cb->args[1]) {
1304                         cb->args[1] = 0;
1305                         goto restart;
1306                 }
1307         }
1308 out:
1309         read_unlock_bh(&nf_conntrack_lock);
1310         if (last)
1311                 nf_ct_expect_put(last);
1312
1313         return skb->len;
1314 }
1315
1316 static const size_t cta_min_exp[CTA_EXPECT_MAX+1] = {
1317         [CTA_EXPECT_TIMEOUT]    = sizeof(u_int32_t),
1318         [CTA_EXPECT_ID]         = sizeof(u_int32_t)
1319 };
1320
1321 static int
1322 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1323                      struct nlmsghdr *nlh, struct nlattr *cda[])
1324 {
1325         struct nf_conntrack_tuple tuple;
1326         struct nf_conntrack_expect *exp;
1327         struct sk_buff *skb2;
1328         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1329         u_int8_t u3 = nfmsg->nfgen_family;
1330         int err = 0;
1331
1332         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1333                 return -EINVAL;
1334
1335         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1336                 return netlink_dump_start(ctnl, skb, nlh,
1337                                           ctnetlink_exp_dump_table,
1338                                           ctnetlink_exp_done);
1339         }
1340
1341         if (cda[CTA_EXPECT_MASTER])
1342                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1343         else
1344                 return -EINVAL;
1345
1346         if (err < 0)
1347                 return err;
1348
1349         exp = nf_ct_expect_find_get(&tuple);
1350         if (!exp)
1351                 return -ENOENT;
1352
1353         if (cda[CTA_EXPECT_ID]) {
1354                 __be32 id = *(__be32 *)nla_data(cda[CTA_EXPECT_ID]);
1355                 if (exp->id != ntohl(id)) {
1356                         nf_ct_expect_put(exp);
1357                         return -ENOENT;
1358                 }
1359         }
1360
1361         err = -ENOMEM;
1362         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
1363         if (!skb2)
1364                 goto out;
1365
1366         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1367                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
1368                                       1, exp);
1369         if (err <= 0)
1370                 goto free;
1371
1372         nf_ct_expect_put(exp);
1373
1374         return netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1375
1376 free:
1377         kfree_skb(skb2);
1378 out:
1379         nf_ct_expect_put(exp);
1380         return err;
1381 }
1382
1383 static int
1384 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1385                      struct nlmsghdr *nlh, struct nlattr *cda[])
1386 {
1387         struct nf_conntrack_expect *exp;
1388         struct nf_conntrack_tuple tuple;
1389         struct nf_conntrack_helper *h;
1390         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1391         struct hlist_node *n, *next;
1392         u_int8_t u3 = nfmsg->nfgen_family;
1393         unsigned int i;
1394         int err;
1395
1396         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1397                 return -EINVAL;
1398
1399         if (cda[CTA_EXPECT_TUPLE]) {
1400                 /* delete a single expect by tuple */
1401                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1402                 if (err < 0)
1403                         return err;
1404
1405                 /* bump usage count to 2 */
1406                 exp = nf_ct_expect_find_get(&tuple);
1407                 if (!exp)
1408                         return -ENOENT;
1409
1410                 if (cda[CTA_EXPECT_ID]) {
1411                         __be32 id = *(__be32 *)nla_data(cda[CTA_EXPECT_ID]);
1412                         if (exp->id != ntohl(id)) {
1413                                 nf_ct_expect_put(exp);
1414                                 return -ENOENT;
1415                         }
1416                 }
1417
1418                 /* after list removal, usage count == 1 */
1419                 nf_ct_unexpect_related(exp);
1420                 /* have to put what we 'get' above.
1421                  * after this line usage count == 0 */
1422                 nf_ct_expect_put(exp);
1423         } else if (cda[CTA_EXPECT_HELP_NAME]) {
1424                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
1425                 struct nf_conn_help *m_help;
1426
1427                 /* delete all expectations for this helper */
1428                 write_lock_bh(&nf_conntrack_lock);
1429                 h = __nf_conntrack_helper_find_byname(name);
1430                 if (!h) {
1431                         write_unlock_bh(&nf_conntrack_lock);
1432                         return -EINVAL;
1433                 }
1434                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1435                         hlist_for_each_entry_safe(exp, n, next,
1436                                                   &nf_ct_expect_hash[i],
1437                                                   hnode) {
1438                                 m_help = nfct_help(exp->master);
1439                                 if (m_help->helper == h
1440                                     && del_timer(&exp->timeout)) {
1441                                         nf_ct_unlink_expect(exp);
1442                                         nf_ct_expect_put(exp);
1443                                 }
1444                         }
1445                 }
1446                 write_unlock_bh(&nf_conntrack_lock);
1447         } else {
1448                 /* This basically means we have to flush everything*/
1449                 write_lock_bh(&nf_conntrack_lock);
1450                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1451                         hlist_for_each_entry_safe(exp, n, next,
1452                                                   &nf_ct_expect_hash[i],
1453                                                   hnode) {
1454                                 if (del_timer(&exp->timeout)) {
1455                                         nf_ct_unlink_expect(exp);
1456                                         nf_ct_expect_put(exp);
1457                                 }
1458                         }
1459                 }
1460                 write_unlock_bh(&nf_conntrack_lock);
1461         }
1462
1463         return 0;
1464 }
1465 static int
1466 ctnetlink_change_expect(struct nf_conntrack_expect *x, struct nlattr *cda[])
1467 {
1468         return -EOPNOTSUPP;
1469 }
1470
1471 static int
1472 ctnetlink_create_expect(struct nlattr *cda[], u_int8_t u3)
1473 {
1474         struct nf_conntrack_tuple tuple, mask, master_tuple;
1475         struct nf_conntrack_tuple_hash *h = NULL;
1476         struct nf_conntrack_expect *exp;
1477         struct nf_conn *ct;
1478         struct nf_conn_help *help;
1479         int err = 0;
1480
1481         /* caller guarantees that those three CTA_EXPECT_* exist */
1482         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1483         if (err < 0)
1484                 return err;
1485         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
1486         if (err < 0)
1487                 return err;
1488         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
1489         if (err < 0)
1490                 return err;
1491
1492         /* Look for master conntrack of this expectation */
1493         h = nf_conntrack_find_get(&master_tuple);
1494         if (!h)
1495                 return -ENOENT;
1496         ct = nf_ct_tuplehash_to_ctrack(h);
1497         help = nfct_help(ct);
1498
1499         if (!help || !help->helper) {
1500                 /* such conntrack hasn't got any helper, abort */
1501                 err = -EINVAL;
1502                 goto out;
1503         }
1504
1505         exp = nf_ct_expect_alloc(ct);
1506         if (!exp) {
1507                 err = -ENOMEM;
1508                 goto out;
1509         }
1510
1511         exp->expectfn = NULL;
1512         exp->flags = 0;
1513         exp->master = ct;
1514         exp->helper = NULL;
1515         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
1516         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
1517         exp->mask.src.u.all = mask.src.u.all;
1518
1519         err = nf_ct_expect_related(exp);
1520         nf_ct_expect_put(exp);
1521
1522 out:
1523         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
1524         return err;
1525 }
1526
1527 static int
1528 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
1529                      struct nlmsghdr *nlh, struct nlattr *cda[])
1530 {
1531         struct nf_conntrack_tuple tuple;
1532         struct nf_conntrack_expect *exp;
1533         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1534         u_int8_t u3 = nfmsg->nfgen_family;
1535         int err = 0;
1536
1537         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1538                 return -EINVAL;
1539
1540         if (!cda[CTA_EXPECT_TUPLE]
1541             || !cda[CTA_EXPECT_MASK]
1542             || !cda[CTA_EXPECT_MASTER])
1543                 return -EINVAL;
1544
1545         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1546         if (err < 0)
1547                 return err;
1548
1549         write_lock_bh(&nf_conntrack_lock);
1550         exp = __nf_ct_expect_find(&tuple);
1551
1552         if (!exp) {
1553                 write_unlock_bh(&nf_conntrack_lock);
1554                 err = -ENOENT;
1555                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1556                         err = ctnetlink_create_expect(cda, u3);
1557                 return err;
1558         }
1559
1560         err = -EEXIST;
1561         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
1562                 err = ctnetlink_change_expect(exp, cda);
1563         write_unlock_bh(&nf_conntrack_lock);
1564
1565         return err;
1566 }
1567
1568 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1569 static struct notifier_block ctnl_notifier = {
1570         .notifier_call  = ctnetlink_conntrack_event,
1571 };
1572
1573 static struct notifier_block ctnl_notifier_exp = {
1574         .notifier_call  = ctnetlink_expect_event,
1575 };
1576 #endif
1577
1578 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
1579         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
1580                                             .attr_count = CTA_MAX, },
1581         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
1582                                             .attr_count = CTA_MAX, },
1583         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
1584                                             .attr_count = CTA_MAX, },
1585         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
1586                                             .attr_count = CTA_MAX, },
1587 };
1588
1589 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
1590         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
1591                                             .attr_count = CTA_EXPECT_MAX, },
1592         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
1593                                             .attr_count = CTA_EXPECT_MAX, },
1594         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
1595                                             .attr_count = CTA_EXPECT_MAX, },
1596 };
1597
1598 static const struct nfnetlink_subsystem ctnl_subsys = {
1599         .name                           = "conntrack",
1600         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
1601         .cb_count                       = IPCTNL_MSG_MAX,
1602         .cb                             = ctnl_cb,
1603 };
1604
1605 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
1606         .name                           = "conntrack_expect",
1607         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
1608         .cb_count                       = IPCTNL_MSG_EXP_MAX,
1609         .cb                             = ctnl_exp_cb,
1610 };
1611
1612 MODULE_ALIAS("ip_conntrack_netlink");
1613 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
1614 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
1615
1616 static int __init ctnetlink_init(void)
1617 {
1618         int ret;
1619
1620         printk("ctnetlink v%s: registering with nfnetlink.\n", version);
1621         ret = nfnetlink_subsys_register(&ctnl_subsys);
1622         if (ret < 0) {
1623                 printk("ctnetlink_init: cannot register with nfnetlink.\n");
1624                 goto err_out;
1625         }
1626
1627         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
1628         if (ret < 0) {
1629                 printk("ctnetlink_init: cannot register exp with nfnetlink.\n");
1630                 goto err_unreg_subsys;
1631         }
1632
1633 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1634         ret = nf_conntrack_register_notifier(&ctnl_notifier);
1635         if (ret < 0) {
1636                 printk("ctnetlink_init: cannot register notifier.\n");
1637                 goto err_unreg_exp_subsys;
1638         }
1639
1640         ret = nf_ct_expect_register_notifier(&ctnl_notifier_exp);
1641         if (ret < 0) {
1642                 printk("ctnetlink_init: cannot expect register notifier.\n");
1643                 goto err_unreg_notifier;
1644         }
1645 #endif
1646
1647         return 0;
1648
1649 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1650 err_unreg_notifier:
1651         nf_conntrack_unregister_notifier(&ctnl_notifier);
1652 err_unreg_exp_subsys:
1653         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1654 #endif
1655 err_unreg_subsys:
1656         nfnetlink_subsys_unregister(&ctnl_subsys);
1657 err_out:
1658         return ret;
1659 }
1660
1661 static void __exit ctnetlink_exit(void)
1662 {
1663         printk("ctnetlink: unregistering from nfnetlink.\n");
1664
1665 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1666         nf_ct_expect_unregister_notifier(&ctnl_notifier_exp);
1667         nf_conntrack_unregister_notifier(&ctnl_notifier);
1668 #endif
1669
1670         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1671         nfnetlink_subsys_unregister(&ctnl_subsys);
1672         return;
1673 }
1674
1675 module_init(ctnetlink_init);
1676 module_exit(ctnetlink_exit);