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 nfattr *nest_parms = NFA_NEST(skb, CTA_TUPLE_PROTO);
  58
  59         NFA_PUT(skb, CTA_PROTO_NUM, sizeof(u_int8_t), &tuple->dst.protonum);
  60
  61         if (likely(l4proto->tuple_to_nfattr))
  62                 ret = l4proto->tuple_to_nfattr(skb, tuple);
  63
  64         NFA_NEST_END(skb, nest_parms);
  65
  66         return ret;
  67
  68 nfattr_failure:
  69         return -1;
  70 }
  71
  72 static inline int
  73 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  74                          const struct nf_conntrack_tuple *tuple,
  75                          struct nf_conntrack_l3proto *l3proto)
  76 {
  77         int ret = 0;
  78         struct nfattr *nest_parms = NFA_NEST(skb, CTA_TUPLE_IP);
  79
  80         if (likely(l3proto->tuple_to_nfattr))
  81                 ret = l3proto->tuple_to_nfattr(skb, tuple);
  82
  83         NFA_NEST_END(skb, nest_parms);
  84
  85         return ret;
  86
  87 nfattr_failure:
  88         return -1;
  89 }
  90
  91 static inline int
  92 ctnetlink_dump_tuples(struct sk_buff *skb,
  93                       const struct nf_conntrack_tuple *tuple)
  94 {
  95         int ret;
  96         struct nf_conntrack_l3proto *l3proto;
  97         struct nf_conntrack_l4proto *l4proto;
  98
  99         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
 100         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 101         nf_ct_l3proto_put(l3proto);
 102
 103         if (unlikely(ret < 0))
 104                 return ret;
 105
 106         l4proto = nf_ct_l4proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 107         ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 108         nf_ct_l4proto_put(l4proto);
 109
 110         return ret;
 111 }
 112
 113 static inline int
 114 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 115 {
 116         __be32 status = htonl((u_int32_t) ct->status);
 117         NFA_PUT(skb, CTA_STATUS, sizeof(status), &status);
 118         return 0;
 119
 120 nfattr_failure:
 121         return -1;
 122 }
 123
 124 static inline int
 125 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 126 {
 127         long timeout_l = ct->timeout.expires - jiffies;
 128         __be32 timeout;
 129
 130         if (timeout_l < 0)
 131                 timeout = 0;
 132         else
 133                 timeout = htonl(timeout_l / HZ);
 134
 135         NFA_PUT(skb, CTA_TIMEOUT, sizeof(timeout), &timeout);
 136         return 0;
 137
 138 nfattr_failure:
 139         return -1;
 140 }
 141
 142 static inline int
 143 ctnetlink_dump_protoinfo(struct sk_buff *skb, const struct nf_conn *ct)
 144 {
 145         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);
 146         struct nfattr *nest_proto;
 147         int ret;
 148
 149         if (!l4proto->to_nfattr) {
 150                 nf_ct_l4proto_put(l4proto);
 151                 return 0;
 152         }
 153
 154         nest_proto = NFA_NEST(skb, CTA_PROTOINFO);
 155
 156         ret = l4proto->to_nfattr(skb, nest_proto, ct);
 157
 158         nf_ct_l4proto_put(l4proto);
 159
 160         NFA_NEST_END(skb, nest_proto);
 161
 162         return ret;
 163
 164 nfattr_failure:
 165         nf_ct_l4proto_put(l4proto);
 166         return -1;
 167 }
 168
 169 static inline int
 170 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 171 {
 172         struct nfattr *nest_helper;
 173         const struct nf_conn_help *help = nfct_help(ct);
 174         struct nf_conntrack_helper *helper;
 175
 176         if (!help)
 177                 return 0;
 178
 179         rcu_read_lock();
 180         helper = rcu_dereference(help->helper);
 181         if (!helper)
 182                 goto out;
 183
 184         nest_helper = NFA_NEST(skb, CTA_HELP);
 185         NFA_PUT(skb, CTA_HELP_NAME, strlen(helper->name), helper->name);
 186
 187         if (helper->to_nfattr)
 188                 helper->to_nfattr(skb, ct);
 189
 190         NFA_NEST_END(skb, nest_helper);
 191 out:
 192         rcu_read_unlock();
 193         return 0;
 194
 195 nfattr_failure:
 196         rcu_read_unlock();
 197         return -1;
 198 }
 199
 200 #ifdef CONFIG_NF_CT_ACCT
 201 static inline int
 202 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 203                         enum ip_conntrack_dir dir)
 204 {
 205         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 206         struct nfattr *nest_count = NFA_NEST(skb, type);
 207         __be32 tmp;
 208
 209         tmp = htonl(ct->counters[dir].packets);
 210         NFA_PUT(skb, CTA_COUNTERS32_PACKETS, sizeof(u_int32_t), &tmp);
 211
 212         tmp = htonl(ct->counters[dir].bytes);
 213         NFA_PUT(skb, CTA_COUNTERS32_BYTES, sizeof(u_int32_t), &tmp);
 214
 215         NFA_NEST_END(skb, nest_count);
 216
 217         return 0;
 218
 219 nfattr_failure:
 220         return -1;
 221 }
 222 #else
 223 #define ctnetlink_dump_counters(a, b, c) (0)
 224 #endif
 225
 226 #ifdef CONFIG_NF_CONNTRACK_MARK
 227 static inline int
 228 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 229 {
 230         __be32 mark = htonl(ct->mark);
 231
 232         NFA_PUT(skb, CTA_MARK, sizeof(u_int32_t), &mark);
 233         return 0;
 234
 235 nfattr_failure:
 236         return -1;
 237 }
 238 #else
 239 #define ctnetlink_dump_mark(a, b) (0)
 240 #endif
 241
 242 static inline int
 243 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 244 {
 245         __be32 id = htonl(ct->id);
 246         NFA_PUT(skb, CTA_ID, sizeof(u_int32_t), &id);
 247         return 0;
 248
 249 nfattr_failure:
 250         return -1;
 251 }
 252
 253 static inline int
 254 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 255 {
 256         __be32 use = htonl(atomic_read(&ct->ct_general.use));
 257
 258         NFA_PUT(skb, CTA_USE, sizeof(u_int32_t), &use);
 259         return 0;
 260
 261 nfattr_failure:
 262         return -1;
 263 }
 264
 265 #define tuple(ct, dir) (&(ct)->tuplehash[dir].tuple)
 266
 267 static int
 268 ctnetlink_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
 269                     int event, int nowait,
 270                     const struct nf_conn *ct)
 271 {
 272         struct nlmsghdr *nlh;
 273         struct nfgenmsg *nfmsg;
 274         struct nfattr *nest_parms;
 275         unsigned char *b = skb_tail_pointer(skb);
 276
 277         event |= NFNL_SUBSYS_CTNETLINK << 8;
 278         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
 279         nfmsg  = NLMSG_DATA(nlh);
 280
 281         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
 282         nfmsg->nfgen_family =
 283                 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 284         nfmsg->version      = NFNETLINK_V0;
 285         nfmsg->res_id       = 0;
 286
 287         nest_parms = NFA_NEST(skb, CTA_TUPLE_ORIG);
 288         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 289                 goto nfattr_failure;
 290         NFA_NEST_END(skb, nest_parms);
 291
 292         nest_parms = NFA_NEST(skb, CTA_TUPLE_REPLY);
 293         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_REPLY)) < 0)
 294                 goto nfattr_failure;
 295         NFA_NEST_END(skb, nest_parms);
 296
 297         if (ctnetlink_dump_status(skb, ct) < 0 ||
 298             ctnetlink_dump_timeout(skb, ct) < 0 ||
 299             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 300             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 301             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 302             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 303             ctnetlink_dump_mark(skb, ct) < 0 ||
 304             ctnetlink_dump_id(skb, ct) < 0 ||
 305             ctnetlink_dump_use(skb, ct) < 0)
 306                 goto nfattr_failure;
 307
 308         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 309         return skb->len;
 310
 311 nlmsg_failure:
 312 nfattr_failure:
 313         nlmsg_trim(skb, b);
 314         return -1;
 315 }
 316
 317 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 318 static int ctnetlink_conntrack_event(struct notifier_block *this,
 319                                      unsigned long events, void *ptr)
 320 {
 321         struct nlmsghdr *nlh;
 322         struct nfgenmsg *nfmsg;
 323         struct nfattr *nest_parms;
 324         struct nf_conn *ct = (struct nf_conn *)ptr;
 325         struct sk_buff *skb;
 326         unsigned int type;
 327         sk_buff_data_t b;
 328         unsigned int flags = 0, group;
 329
 330         /* ignore our fake conntrack entry */
 331         if (ct == &nf_conntrack_untracked)
 332                 return NOTIFY_DONE;
 333
 334         if (events & IPCT_DESTROY) {
 335                 type = IPCTNL_MSG_CT_DELETE;
 336                 group = NFNLGRP_CONNTRACK_DESTROY;
 337         } else  if (events & (IPCT_NEW | IPCT_RELATED)) {
 338                 type = IPCTNL_MSG_CT_NEW;
 339                 flags = NLM_F_CREATE|NLM_F_EXCL;
 340                 group = NFNLGRP_CONNTRACK_NEW;
 341         } else  if (events & (IPCT_STATUS | IPCT_PROTOINFO)) {
 342                 type = IPCTNL_MSG_CT_NEW;
 343                 group = NFNLGRP_CONNTRACK_UPDATE;
 344         } else
 345                 return NOTIFY_DONE;
 346
 347         if (!nfnetlink_has_listeners(group))
 348                 return NOTIFY_DONE;
 349
 350         skb = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
 351         if (!skb)
 352                 return NOTIFY_DONE;
 353
 354         b = skb->tail;
 355
 356         type |= NFNL_SUBSYS_CTNETLINK << 8;
 357         nlh   = NLMSG_PUT(skb, 0, 0, type, sizeof(struct nfgenmsg));
 358         nfmsg = NLMSG_DATA(nlh);
 359
 360         nlh->nlmsg_flags    = flags;
 361         nfmsg->nfgen_family = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 362         nfmsg->version  = NFNETLINK_V0;
 363         nfmsg->res_id   = 0;
 364
 365         nest_parms = NFA_NEST(skb, CTA_TUPLE_ORIG);
 366         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 367                 goto nfattr_failure;
 368         NFA_NEST_END(skb, nest_parms);
 369
 370         nest_parms = NFA_NEST(skb, CTA_TUPLE_REPLY);
 371         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_REPLY)) < 0)
 372                 goto nfattr_failure;
 373         NFA_NEST_END(skb, nest_parms);
 374
 375         if (events & IPCT_DESTROY) {
 376                 if (ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 377                     ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0)
 378                         goto nfattr_failure;
 379         } else {
 380                 if (ctnetlink_dump_status(skb, ct) < 0)
 381                         goto nfattr_failure;
 382
 383                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 384                         goto nfattr_failure;
 385
 386                 if (events & IPCT_PROTOINFO
 387                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 388                         goto nfattr_failure;
 389
 390                 if ((events & IPCT_HELPER || nfct_help(ct))
 391                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 392                         goto nfattr_failure;
 393
 394 #ifdef CONFIG_NF_CONNTRACK_MARK
 395                 if ((events & IPCT_MARK || ct->mark)
 396                     && ctnetlink_dump_mark(skb, ct) < 0)
 397                         goto nfattr_failure;
 398 #endif
 399
 400                 if (events & IPCT_COUNTER_FILLING &&
 401                     (ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 402                      ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0))
 403                         goto nfattr_failure;
 404         }
 405
 406         nlh->nlmsg_len = skb->tail - b;
 407         nfnetlink_send(skb, 0, group, 0);
 408         return NOTIFY_DONE;
 409
 410 nlmsg_failure:
 411 nfattr_failure:
 412         kfree_skb(skb);
 413         return NOTIFY_DONE;
 414 }
 415 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 416
 417 static int ctnetlink_done(struct netlink_callback *cb)
 418 {
 419         if (cb->args[1])
 420                 nf_ct_put((struct nf_conn *)cb->args[1]);
 421         return 0;
 422 }
 423
 424 #define L3PROTO(ct) ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num
 425
 426 static int
 427 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 428 {
 429         struct nf_conn *ct, *last;
 430         struct nf_conntrack_tuple_hash *h;
 431         struct hlist_node *n;
 432         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
 433         u_int8_t l3proto = nfmsg->nfgen_family;
 434
 435         read_lock_bh(&nf_conntrack_lock);
 436         last = (struct nf_conn *)cb->args[1];
 437         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 438 restart:
 439                 hlist_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 440                                      hnode) {
 441                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 442                                 continue;
 443                         ct = nf_ct_tuplehash_to_ctrack(h);
 444                         /* Dump entries of a given L3 protocol number.
 445                          * If it is not specified, ie. l3proto == 0,
 446                          * then dump everything. */
 447                         if (l3proto && L3PROTO(ct) != l3proto)
 448                                 continue;
 449                         if (cb->args[1]) {
 450                                 if (ct != last)
 451                                         continue;
 452                                 cb->args[1] = 0;
 453                         }
 454                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 455                                                 cb->nlh->nlmsg_seq,
 456                                                 IPCTNL_MSG_CT_NEW,
 457                                                 1, ct) < 0) {
 458                                 nf_conntrack_get(&ct->ct_general);
 459                                 cb->args[1] = (unsigned long)ct;
 460                                 goto out;
 461                         }
 462 #ifdef CONFIG_NF_CT_ACCT
 463                         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) ==
 464                                                 IPCTNL_MSG_CT_GET_CTRZERO)
 465                                 memset(&ct->counters, 0, sizeof(ct->counters));
 466 #endif
 467                 }
 468                 if (cb->args[1]) {
 469                         cb->args[1] = 0;
 470                         goto restart;
 471                 }
 472         }
 473 out:
 474         read_unlock_bh(&nf_conntrack_lock);
 475         if (last)
 476                 nf_ct_put(last);
 477
 478         return skb->len;
 479 }
 480
 481 static inline int
 482 ctnetlink_parse_tuple_ip(struct nfattr *attr, struct nf_conntrack_tuple *tuple)
 483 {
 484         struct nfattr *tb[CTA_IP_MAX];
 485         struct nf_conntrack_l3proto *l3proto;
 486         int ret = 0;
 487
 488         nfattr_parse_nested(tb, CTA_IP_MAX, attr);
 489
 490         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
 491
 492         if (likely(l3proto->nfattr_to_tuple))
 493                 ret = l3proto->nfattr_to_tuple(tb, tuple);
 494
 495         nf_ct_l3proto_put(l3proto);
 496
 497         return ret;
 498 }
 499
 500 static const size_t cta_min_proto[CTA_PROTO_MAX] = {
 501         [CTA_PROTO_NUM-1]       = sizeof(u_int8_t),
 502 };
 503
 504 static inline int
 505 ctnetlink_parse_tuple_proto(struct nfattr *attr,
 506                             struct nf_conntrack_tuple *tuple)
 507 {
 508         struct nfattr *tb[CTA_PROTO_MAX];
 509         struct nf_conntrack_l4proto *l4proto;
 510         int ret = 0;
 511
 512         nfattr_parse_nested(tb, CTA_PROTO_MAX, attr);
 513
 514         if (nfattr_bad_size(tb, CTA_PROTO_MAX, cta_min_proto))
 515                 return -EINVAL;
 516
 517         if (!tb[CTA_PROTO_NUM-1])
 518                 return -EINVAL;
 519         tuple->dst.protonum = *(u_int8_t *)NFA_DATA(tb[CTA_PROTO_NUM-1]);
 520
 521         l4proto = nf_ct_l4proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 522
 523         if (likely(l4proto->nfattr_to_tuple))
 524                 ret = l4proto->nfattr_to_tuple(tb, tuple);
 525
 526         nf_ct_l4proto_put(l4proto);
 527
 528         return ret;
 529 }
 530
 531 static inline int
 532 ctnetlink_parse_tuple(struct nfattr *cda[], struct nf_conntrack_tuple *tuple,
 533                       enum ctattr_tuple type, u_int8_t l3num)
 534 {
 535         struct nfattr *tb[CTA_TUPLE_MAX];
 536         int err;
 537
 538         memset(tuple, 0, sizeof(*tuple));
 539
 540         nfattr_parse_nested(tb, CTA_TUPLE_MAX, cda[type-1]);
 541
 542         if (!tb[CTA_TUPLE_IP-1])
 543                 return -EINVAL;
 544
 545         tuple->src.l3num = l3num;
 546
 547         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP-1], tuple);
 548         if (err < 0)
 549                 return err;
 550
 551         if (!tb[CTA_TUPLE_PROTO-1])
 552                 return -EINVAL;
 553
 554         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO-1], tuple);
 555         if (err < 0)
 556                 return err;
 557
 558         /* orig and expect tuples get DIR_ORIGINAL */
 559         if (type == CTA_TUPLE_REPLY)
 560                 tuple->dst.dir = IP_CT_DIR_REPLY;
 561         else
 562                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 563
 564         return 0;
 565 }
 566
 567 #ifdef CONFIG_NF_NAT_NEEDED
 568 static const size_t cta_min_protonat[CTA_PROTONAT_MAX] = {
 569         [CTA_PROTONAT_PORT_MIN-1]       = sizeof(u_int16_t),
 570         [CTA_PROTONAT_PORT_MAX-1]       = sizeof(u_int16_t),
 571 };
 572
 573 static int nfnetlink_parse_nat_proto(struct nfattr *attr,
 574                                      const struct nf_conn *ct,
 575                                      struct nf_nat_range *range)
 576 {
 577         struct nfattr *tb[CTA_PROTONAT_MAX];
 578         struct nf_nat_protocol *npt;
 579
 580         nfattr_parse_nested(tb, CTA_PROTONAT_MAX, attr);
 581
 582         if (nfattr_bad_size(tb, CTA_PROTONAT_MAX, cta_min_protonat))
 583                 return -EINVAL;
 584
 585         npt = nf_nat_proto_find_get(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum);
 586
 587         if (!npt->nfattr_to_range) {
 588                 nf_nat_proto_put(npt);
 589                 return 0;
 590         }
 591
 592         /* nfattr_to_range returns 1 if it parsed, 0 if not, neg. on error */
 593         if (npt->nfattr_to_range(tb, range) > 0)
 594                 range->flags |= IP_NAT_RANGE_PROTO_SPECIFIED;
 595
 596         nf_nat_proto_put(npt);
 597
 598         return 0;
 599 }
 600
 601 static const size_t cta_min_nat[CTA_NAT_MAX] = {
 602         [CTA_NAT_MINIP-1]       = sizeof(u_int32_t),
 603         [CTA_NAT_MAXIP-1]       = sizeof(u_int32_t),
 604 };
 605
 606 static inline int
 607 nfnetlink_parse_nat(struct nfattr *nat,
 608                     const struct nf_conn *ct, struct nf_nat_range *range)
 609 {
 610         struct nfattr *tb[CTA_NAT_MAX];
 611         int err;
 612
 613         memset(range, 0, sizeof(*range));
 614
 615         nfattr_parse_nested(tb, CTA_NAT_MAX, nat);
 616
 617         if (nfattr_bad_size(tb, CTA_NAT_MAX, cta_min_nat))
 618                 return -EINVAL;
 619
 620         if (tb[CTA_NAT_MINIP-1])
 621                 range->min_ip = *(__be32 *)NFA_DATA(tb[CTA_NAT_MINIP-1]);
 622
 623         if (!tb[CTA_NAT_MAXIP-1])
 624                 range->max_ip = range->min_ip;
 625         else
 626                 range->max_ip = *(__be32 *)NFA_DATA(tb[CTA_NAT_MAXIP-1]);
 627
 628         if (range->min_ip)
 629                 range->flags |= IP_NAT_RANGE_MAP_IPS;
 630
 631         if (!tb[CTA_NAT_PROTO-1])
 632                 return 0;
 633
 634         err = nfnetlink_parse_nat_proto(tb[CTA_NAT_PROTO-1], ct, range);
 635         if (err < 0)
 636                 return err;
 637
 638         return 0;
 639 }
 640 #endif
 641
 642 static inline int
 643 ctnetlink_parse_help(struct nfattr *attr, char **helper_name)
 644 {
 645         struct nfattr *tb[CTA_HELP_MAX];
 646
 647         nfattr_parse_nested(tb, CTA_HELP_MAX, attr);
 648
 649         if (!tb[CTA_HELP_NAME-1])
 650                 return -EINVAL;
 651
 652         *helper_name = NFA_DATA(tb[CTA_HELP_NAME-1]);
 653
 654         return 0;
 655 }
 656
 657 static const size_t cta_min[CTA_MAX] = {
 658         [CTA_STATUS-1]          = sizeof(u_int32_t),
 659         [CTA_TIMEOUT-1]         = sizeof(u_int32_t),
 660         [CTA_MARK-1]            = sizeof(u_int32_t),
 661         [CTA_USE-1]             = sizeof(u_int32_t),
 662         [CTA_ID-1]              = sizeof(u_int32_t)
 663 };
 664
 665 static int
 666 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 667                         struct nlmsghdr *nlh, struct nfattr *cda[])
 668 {
 669         struct nf_conntrack_tuple_hash *h;
 670         struct nf_conntrack_tuple tuple;
 671         struct nf_conn *ct;
 672         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 673         u_int8_t u3 = nfmsg->nfgen_family;
 674         int err = 0;
 675
 676         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 677                 return -EINVAL;
 678
 679         if (cda[CTA_TUPLE_ORIG-1])
 680                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 681         else if (cda[CTA_TUPLE_REPLY-1])
 682                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 683         else {
 684                 /* Flush the whole table */
 685                 nf_conntrack_flush();
 686                 return 0;
 687         }
 688
 689         if (err < 0)
 690                 return err;
 691
 692         h = nf_conntrack_find_get(&tuple);
 693         if (!h)
 694                 return -ENOENT;
 695
 696         ct = nf_ct_tuplehash_to_ctrack(h);
 697
 698         if (cda[CTA_ID-1]) {
 699                 u_int32_t id = ntohl(*(__be32 *)NFA_DATA(cda[CTA_ID-1]));
 700                 if (ct->id != id) {
 701                         nf_ct_put(ct);
 702                         return -ENOENT;
 703                 }
 704         }
 705         if (del_timer(&ct->timeout))
 706                 ct->timeout.function((unsigned long)ct);
 707
 708         nf_ct_put(ct);
 709
 710         return 0;
 711 }
 712
 713 static int
 714 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 715                         struct nlmsghdr *nlh, struct nfattr *cda[])
 716 {
 717         struct nf_conntrack_tuple_hash *h;
 718         struct nf_conntrack_tuple tuple;
 719         struct nf_conn *ct;
 720         struct sk_buff *skb2 = NULL;
 721         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 722         u_int8_t u3 = nfmsg->nfgen_family;
 723         int err = 0;
 724
 725         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 726 #ifndef CONFIG_NF_CT_ACCT
 727                 if (NFNL_MSG_TYPE(nlh->nlmsg_type) == IPCTNL_MSG_CT_GET_CTRZERO)
 728                         return -ENOTSUPP;
 729 #endif
 730                 return netlink_dump_start(ctnl, skb, nlh, ctnetlink_dump_table,
 731                                           ctnetlink_done);
 732         }
 733
 734         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 735                 return -EINVAL;
 736
 737         if (cda[CTA_TUPLE_ORIG-1])
 738                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 739         else if (cda[CTA_TUPLE_REPLY-1])
 740                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 741         else
 742                 return -EINVAL;
 743
 744         if (err < 0)
 745                 return err;
 746
 747         h = nf_conntrack_find_get(&tuple);
 748         if (!h)
 749                 return -ENOENT;
 750
 751         ct = nf_ct_tuplehash_to_ctrack(h);
 752
 753         err = -ENOMEM;
 754         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 755         if (!skb2) {
 756                 nf_ct_put(ct);
 757                 return -ENOMEM;
 758         }
 759
 760         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
 761                                   IPCTNL_MSG_CT_NEW, 1, ct);
 762         nf_ct_put(ct);
 763         if (err <= 0)
 764                 goto free;
 765
 766         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
 767         if (err < 0)
 768                 goto out;
 769
 770         return 0;
 771
 772 free:
 773         kfree_skb(skb2);
 774 out:
 775         return err;
 776 }
 777
 778 static inline int
 779 ctnetlink_change_status(struct nf_conn *ct, struct nfattr *cda[])
 780 {
 781         unsigned long d;
 782         unsigned int status = ntohl(*(__be32 *)NFA_DATA(cda[CTA_STATUS-1]));
 783         d = ct->status ^ status;
 784
 785         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
 786                 /* unchangeable */
 787                 return -EINVAL;
 788
 789         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
 790                 /* SEEN_REPLY bit can only be set */
 791                 return -EINVAL;
 792
 793
 794         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
 795                 /* ASSURED bit can only be set */
 796                 return -EINVAL;
 797
 798         if (cda[CTA_NAT_SRC-1] || cda[CTA_NAT_DST-1]) {
 799 #ifndef CONFIG_NF_NAT_NEEDED
 800                 return -EINVAL;
 801 #else
 802                 struct nf_nat_range range;
 803
 804                 if (cda[CTA_NAT_DST-1]) {
 805                         if (nfnetlink_parse_nat(cda[CTA_NAT_DST-1], ct,
 806                                                 &range) < 0)
 807                                 return -EINVAL;
 808                         if (nf_nat_initialized(ct,
 809                                                HOOK2MANIP(NF_IP_PRE_ROUTING)))
 810                                 return -EEXIST;
 811                         nf_nat_setup_info(ct, &range, NF_IP_PRE_ROUTING);
 812                 }
 813                 if (cda[CTA_NAT_SRC-1]) {
 814                         if (nfnetlink_parse_nat(cda[CTA_NAT_SRC-1], ct,
 815                                                 &range) < 0)
 816                                 return -EINVAL;
 817                         if (nf_nat_initialized(ct,
 818                                                HOOK2MANIP(NF_IP_POST_ROUTING)))
 819                                 return -EEXIST;
 820                         nf_nat_setup_info(ct, &range, NF_IP_POST_ROUTING);
 821                 }
 822 #endif
 823         }
 824
 825         /* Be careful here, modifying NAT bits can screw up things,
 826          * so don't let users modify them directly if they don't pass
 827          * nf_nat_range. */
 828         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
 829         return 0;
 830 }
 831
 832
 833 static inline int
 834 ctnetlink_change_helper(struct nf_conn *ct, struct nfattr *cda[])
 835 {
 836         struct nf_conntrack_helper *helper;
 837         struct nf_conn_help *help = nfct_help(ct);
 838         char *helpname;
 839         int err;
 840
 841         /* don't change helper of sibling connections */
 842         if (ct->master)
 843                 return -EINVAL;
 844
 845         err = ctnetlink_parse_help(cda[CTA_HELP-1], &helpname);
 846         if (err < 0)
 847                 return err;
 848
 849         if (!strcmp(helpname, "")) {
 850                 if (help && help->helper) {
 851                         /* we had a helper before ... */
 852                         nf_ct_remove_expectations(ct);
 853                         rcu_assign_pointer(help->helper, NULL);
 854                 }
 855
 856                 return 0;
 857         }
 858
 859         helper = __nf_conntrack_helper_find_byname(helpname);
 860         if (helper == NULL)
 861                 return -EINVAL;
 862
 863         if (help) {
 864                 if (help->helper == helper)
 865                         return 0;
 866                 if (help->helper)
 867                         return -EBUSY;
 868                 /* need to zero data of old helper */
 869                 memset(&help->help, 0, sizeof(help->help));
 870         } else {
 871                 help = nf_ct_helper_ext_add(ct, GFP_KERNEL);
 872                 if (help == NULL)
 873                         return -ENOMEM;
 874         }
 875
 876         rcu_assign_pointer(help->helper, helper);
 877
 878         return 0;
 879 }
 880
 881 static inline int
 882 ctnetlink_change_timeout(struct nf_conn *ct, struct nfattr *cda[])
 883 {
 884         u_int32_t timeout = ntohl(*(__be32 *)NFA_DATA(cda[CTA_TIMEOUT-1]));
 885
 886         if (!del_timer(&ct->timeout))
 887                 return -ETIME;
 888
 889         ct->timeout.expires = jiffies + timeout * HZ;
 890         add_timer(&ct->timeout);
 891
 892         return 0;
 893 }
 894
 895 static inline int
 896 ctnetlink_change_protoinfo(struct nf_conn *ct, struct nfattr *cda[])
 897 {
 898         struct nfattr *tb[CTA_PROTOINFO_MAX], *attr = cda[CTA_PROTOINFO-1];
 899         struct nf_conntrack_l4proto *l4proto;
 900         u_int16_t npt = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum;
 901         u_int16_t l3num = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 902         int err = 0;
 903
 904         nfattr_parse_nested(tb, CTA_PROTOINFO_MAX, attr);
 905
 906         l4proto = nf_ct_l4proto_find_get(l3num, npt);
 907
 908         if (l4proto->from_nfattr)
 909                 err = l4proto->from_nfattr(tb, ct);
 910         nf_ct_l4proto_put(l4proto);
 911
 912         return err;
 913 }
 914
 915 static int
 916 ctnetlink_change_conntrack(struct nf_conn *ct, struct nfattr *cda[])
 917 {
 918         int err;
 919
 920         if (cda[CTA_HELP-1]) {
 921                 err = ctnetlink_change_helper(ct, cda);
 922                 if (err < 0)
 923                         return err;
 924         }
 925
 926         if (cda[CTA_TIMEOUT-1]) {
 927                 err = ctnetlink_change_timeout(ct, cda);
 928                 if (err < 0)
 929                         return err;
 930         }
 931
 932         if (cda[CTA_STATUS-1]) {
 933                 err = ctnetlink_change_status(ct, cda);
 934                 if (err < 0)
 935                         return err;
 936         }
 937
 938         if (cda[CTA_PROTOINFO-1]) {
 939                 err = ctnetlink_change_protoinfo(ct, cda);
 940                 if (err < 0)
 941                         return err;
 942         }
 943
 944 #if defined(CONFIG_NF_CONNTRACK_MARK)
 945         if (cda[CTA_MARK-1])
 946                 ct->mark = ntohl(*(__be32 *)NFA_DATA(cda[CTA_MARK-1]));
 947 #endif
 948
 949         return 0;
 950 }
 951
 952 static int
 953 ctnetlink_create_conntrack(struct nfattr *cda[],
 954                            struct nf_conntrack_tuple *otuple,
 955                            struct nf_conntrack_tuple *rtuple)
 956 {
 957         struct nf_conn *ct;
 958         int err = -EINVAL;
 959         struct nf_conn_help *help;
 960         struct nf_conntrack_helper *helper;
 961
 962         ct = nf_conntrack_alloc(otuple, rtuple);
 963         if (ct == NULL || IS_ERR(ct))
 964                 return -ENOMEM;
 965
 966         if (!cda[CTA_TIMEOUT-1])
 967                 goto err;
 968         ct->timeout.expires = ntohl(*(__be32 *)NFA_DATA(cda[CTA_TIMEOUT-1]));
 969
 970         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
 971         ct->status |= IPS_CONFIRMED;
 972
 973         if (cda[CTA_STATUS-1]) {
 974                 err = ctnetlink_change_status(ct, cda);
 975                 if (err < 0)
 976                         goto err;
 977         }
 978
 979         if (cda[CTA_PROTOINFO-1]) {
 980                 err = ctnetlink_change_protoinfo(ct, cda);
 981                 if (err < 0)
 982                         goto err;
 983         }
 984
 985 #if defined(CONFIG_NF_CONNTRACK_MARK)
 986         if (cda[CTA_MARK-1])
 987                 ct->mark = ntohl(*(__be32 *)NFA_DATA(cda[CTA_MARK-1]));
 988 #endif
 989
 990         helper = nf_ct_helper_find_get(rtuple);
 991         if (helper) {
 992                 help = nf_ct_helper_ext_add(ct, GFP_KERNEL);
 993                 if (help == NULL) {
 994                         nf_ct_helper_put(helper);
 995                         err = -ENOMEM;
 996                         goto err;
 997                 }
 998                 /* not in hash table yet so not strictly necessary */
 999                 rcu_assign_pointer(help->helper, helper);
1000         }
1001
1002         add_timer(&ct->timeout);
1003         nf_conntrack_hash_insert(ct);
1004
1005         if (helper)
1006                 nf_ct_helper_put(helper);
1007
1008         return 0;
1009
1010 err:
1011         nf_conntrack_free(ct);
1012         return err;
1013 }
1014
1015 static int
1016 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1017                         struct nlmsghdr *nlh, struct nfattr *cda[])
1018 {
1019         struct nf_conntrack_tuple otuple, rtuple;
1020         struct nf_conntrack_tuple_hash *h = NULL;
1021         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1022         u_int8_t u3 = nfmsg->nfgen_family;
1023         int err = 0;
1024
1025         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
1026                 return -EINVAL;
1027
1028         if (cda[CTA_TUPLE_ORIG-1]) {
1029                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1030                 if (err < 0)
1031                         return err;
1032         }
1033
1034         if (cda[CTA_TUPLE_REPLY-1]) {
1035                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1036                 if (err < 0)
1037                         return err;
1038         }
1039
1040         write_lock_bh(&nf_conntrack_lock);
1041         if (cda[CTA_TUPLE_ORIG-1])
1042                 h = __nf_conntrack_find(&otuple, NULL);
1043         else if (cda[CTA_TUPLE_REPLY-1])
1044                 h = __nf_conntrack_find(&rtuple, NULL);
1045
1046         if (h == NULL) {
1047                 write_unlock_bh(&nf_conntrack_lock);
1048                 err = -ENOENT;
1049                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1050                         err = ctnetlink_create_conntrack(cda, &otuple, &rtuple);
1051                 return err;
1052         }
1053         /* implicit 'else' */
1054
1055         /* we only allow nat config for new conntracks */
1056         if (cda[CTA_NAT_SRC-1] || cda[CTA_NAT_DST-1]) {
1057                 err = -EINVAL;
1058                 goto out_unlock;
1059         }
1060
1061         /* We manipulate the conntrack inside the global conntrack table lock,
1062          * so there's no need to increase the refcount */
1063         err = -EEXIST;
1064         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
1065                 err = ctnetlink_change_conntrack(nf_ct_tuplehash_to_ctrack(h), cda);
1066
1067 out_unlock:
1068         write_unlock_bh(&nf_conntrack_lock);
1069         return err;
1070 }
1071
1072 /***********************************************************************
1073  * EXPECT
1074  ***********************************************************************/
1075
1076 static inline int
1077 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1078                          const struct nf_conntrack_tuple *tuple,
1079                          enum ctattr_expect type)
1080 {
1081         struct nfattr *nest_parms = NFA_NEST(skb, type);
1082
1083         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1084                 goto nfattr_failure;
1085
1086         NFA_NEST_END(skb, nest_parms);
1087
1088         return 0;
1089
1090 nfattr_failure:
1091         return -1;
1092 }
1093
1094 static inline int
1095 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1096                         const struct nf_conntrack_tuple *tuple,
1097                         const struct nf_conntrack_tuple_mask *mask)
1098 {
1099         int ret;
1100         struct nf_conntrack_l3proto *l3proto;
1101         struct nf_conntrack_l4proto *l4proto;
1102         struct nf_conntrack_tuple m;
1103         struct nfattr *nest_parms;
1104
1105         memset(&m, 0xFF, sizeof(m));
1106         m.src.u.all = mask->src.u.all;
1107         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1108
1109         nest_parms = NFA_NEST(skb, CTA_EXPECT_MASK);
1110
1111         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
1112         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1113         nf_ct_l3proto_put(l3proto);
1114
1115         if (unlikely(ret < 0))
1116                 goto nfattr_failure;
1117
1118         l4proto = nf_ct_l4proto_find_get(tuple->src.l3num, tuple->dst.protonum);
1119         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1120         nf_ct_l4proto_put(l4proto);
1121         if (unlikely(ret < 0))
1122                 goto nfattr_failure;
1123
1124         NFA_NEST_END(skb, nest_parms);
1125
1126         return 0;
1127
1128 nfattr_failure:
1129         return -1;
1130 }
1131
1132 static inline int
1133 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1134                           const struct nf_conntrack_expect *exp)
1135 {
1136         struct nf_conn *master = exp->master;
1137         __be32 timeout = htonl((exp->timeout.expires - jiffies) / HZ);
1138         __be32 id = htonl(exp->id);
1139
1140         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1141                 goto nfattr_failure;
1142         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1143                 goto nfattr_failure;
1144         if (ctnetlink_exp_dump_tuple(skb,
1145                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1146                                  CTA_EXPECT_MASTER) < 0)
1147                 goto nfattr_failure;
1148
1149         NFA_PUT(skb, CTA_EXPECT_TIMEOUT, sizeof(timeout), &timeout);
1150         NFA_PUT(skb, CTA_EXPECT_ID, sizeof(u_int32_t), &id);
1151
1152         return 0;
1153
1154 nfattr_failure:
1155         return -1;
1156 }
1157
1158 static int
1159 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1160                     int event,
1161                     int nowait,
1162                     const struct nf_conntrack_expect *exp)
1163 {
1164         struct nlmsghdr *nlh;
1165         struct nfgenmsg *nfmsg;
1166         unsigned char *b = skb_tail_pointer(skb);
1167
1168         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1169         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
1170         nfmsg  = NLMSG_DATA(nlh);
1171
1172         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
1173         nfmsg->nfgen_family = exp->tuple.src.l3num;
1174         nfmsg->version      = NFNETLINK_V0;
1175         nfmsg->res_id       = 0;
1176
1177         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1178                 goto nfattr_failure;
1179
1180         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
1181         return skb->len;
1182
1183 nlmsg_failure:
1184 nfattr_failure:
1185         nlmsg_trim(skb, b);
1186         return -1;
1187 }
1188
1189 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1190 static int ctnetlink_expect_event(struct notifier_block *this,
1191                                   unsigned long events, void *ptr)
1192 {
1193         struct nlmsghdr *nlh;
1194         struct nfgenmsg *nfmsg;
1195         struct nf_conntrack_expect *exp = (struct nf_conntrack_expect *)ptr;
1196         struct sk_buff *skb;
1197         unsigned int type;
1198         sk_buff_data_t b;
1199         int flags = 0;
1200
1201         if (events & IPEXP_NEW) {
1202                 type = IPCTNL_MSG_EXP_NEW;
1203                 flags = NLM_F_CREATE|NLM_F_EXCL;
1204         } else
1205                 return NOTIFY_DONE;
1206
1207         if (!nfnetlink_has_listeners(NFNLGRP_CONNTRACK_EXP_NEW))
1208                 return NOTIFY_DONE;
1209
1210         skb = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1211         if (!skb)
1212                 return NOTIFY_DONE;
1213
1214         b = skb->tail;
1215
1216         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1217         nlh   = NLMSG_PUT(skb, 0, 0, type, sizeof(struct nfgenmsg));
1218         nfmsg = NLMSG_DATA(nlh);
1219
1220         nlh->nlmsg_flags    = flags;
1221         nfmsg->nfgen_family = exp->tuple.src.l3num;
1222         nfmsg->version      = NFNETLINK_V0;
1223         nfmsg->res_id       = 0;
1224
1225         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1226                 goto nfattr_failure;
1227
1228         nlh->nlmsg_len = skb->tail - b;
1229         nfnetlink_send(skb, 0, NFNLGRP_CONNTRACK_EXP_NEW, 0);
1230         return NOTIFY_DONE;
1231
1232 nlmsg_failure:
1233 nfattr_failure:
1234         kfree_skb(skb);
1235         return NOTIFY_DONE;
1236 }
1237 #endif
1238 static int ctnetlink_exp_done(struct netlink_callback *cb)
1239 {
1240         if (cb->args[1])
1241                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1242         return 0;
1243 }
1244
1245 static int
1246 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1247 {
1248         struct nf_conntrack_expect *exp, *last;
1249         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
1250         struct hlist_node *n;
1251         u_int8_t l3proto = nfmsg->nfgen_family;
1252
1253         read_lock_bh(&nf_conntrack_lock);
1254         last = (struct nf_conntrack_expect *)cb->args[1];
1255         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1256 restart:
1257                 hlist_for_each_entry(exp, n, &nf_ct_expect_hash[cb->args[0]],
1258                                      hnode) {
1259                         if (l3proto && exp->tuple.src.l3num != l3proto)
1260                                 continue;
1261                         if (cb->args[1]) {
1262                                 if (exp != last)
1263                                         continue;
1264                                 cb->args[1] = 0;
1265                         }
1266                         if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).pid,
1267                                                     cb->nlh->nlmsg_seq,
1268                                                     IPCTNL_MSG_EXP_NEW,
1269                                                     1, exp) < 0) {
1270                                 atomic_inc(&exp->use);
1271                                 cb->args[1] = (unsigned long)exp;
1272                                 goto out;
1273                         }
1274                 }
1275                 if (cb->args[1]) {
1276                         cb->args[1] = 0;
1277                         goto restart;
1278                 }
1279         }
1280 out:
1281         read_unlock_bh(&nf_conntrack_lock);
1282         if (last)
1283                 nf_ct_expect_put(last);
1284
1285         return skb->len;
1286 }
1287
1288 static const size_t cta_min_exp[CTA_EXPECT_MAX] = {
1289         [CTA_EXPECT_TIMEOUT-1]          = sizeof(u_int32_t),
1290         [CTA_EXPECT_ID-1]               = sizeof(u_int32_t)
1291 };
1292
1293 static int
1294 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1295                      struct nlmsghdr *nlh, struct nfattr *cda[])
1296 {
1297         struct nf_conntrack_tuple tuple;
1298         struct nf_conntrack_expect *exp;
1299         struct sk_buff *skb2;
1300         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1301         u_int8_t u3 = nfmsg->nfgen_family;
1302         int err = 0;
1303
1304         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1305                 return -EINVAL;
1306
1307         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1308                 return netlink_dump_start(ctnl, skb, nlh,
1309                                           ctnetlink_exp_dump_table,
1310                                           ctnetlink_exp_done);
1311         }
1312
1313         if (cda[CTA_EXPECT_MASTER-1])
1314                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1315         else
1316                 return -EINVAL;
1317
1318         if (err < 0)
1319                 return err;
1320
1321         exp = nf_ct_expect_find_get(&tuple);
1322         if (!exp)
1323                 return -ENOENT;
1324
1325         if (cda[CTA_EXPECT_ID-1]) {
1326                 __be32 id = *(__be32 *)NFA_DATA(cda[CTA_EXPECT_ID-1]);
1327                 if (exp->id != ntohl(id)) {
1328                         nf_ct_expect_put(exp);
1329                         return -ENOENT;
1330                 }
1331         }
1332
1333         err = -ENOMEM;
1334         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
1335         if (!skb2)
1336                 goto out;
1337
1338         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1339                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
1340                                       1, exp);
1341         if (err <= 0)
1342                 goto free;
1343
1344         nf_ct_expect_put(exp);
1345
1346         return netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1347
1348 free:
1349         kfree_skb(skb2);
1350 out:
1351         nf_ct_expect_put(exp);
1352         return err;
1353 }
1354
1355 static int
1356 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1357                      struct nlmsghdr *nlh, struct nfattr *cda[])
1358 {
1359         struct nf_conntrack_expect *exp;
1360         struct nf_conntrack_tuple tuple;
1361         struct nf_conntrack_helper *h;
1362         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1363         struct hlist_node *n, *next;
1364         u_int8_t u3 = nfmsg->nfgen_family;
1365         unsigned int i;
1366         int err;
1367
1368         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1369                 return -EINVAL;
1370
1371         if (cda[CTA_EXPECT_TUPLE-1]) {
1372                 /* delete a single expect by tuple */
1373                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1374                 if (err < 0)
1375                         return err;
1376
1377                 /* bump usage count to 2 */
1378                 exp = nf_ct_expect_find_get(&tuple);
1379                 if (!exp)
1380                         return -ENOENT;
1381
1382                 if (cda[CTA_EXPECT_ID-1]) {
1383                         __be32 id = *(__be32 *)NFA_DATA(cda[CTA_EXPECT_ID-1]);
1384                         if (exp->id != ntohl(id)) {
1385                                 nf_ct_expect_put(exp);
1386                                 return -ENOENT;
1387                         }
1388                 }
1389
1390                 /* after list removal, usage count == 1 */
1391                 nf_ct_unexpect_related(exp);
1392                 /* have to put what we 'get' above.
1393                  * after this line usage count == 0 */
1394                 nf_ct_expect_put(exp);
1395         } else if (cda[CTA_EXPECT_HELP_NAME-1]) {
1396                 char *name = NFA_DATA(cda[CTA_EXPECT_HELP_NAME-1]);
1397                 struct nf_conn_help *m_help;
1398
1399                 /* delete all expectations for this helper */
1400                 write_lock_bh(&nf_conntrack_lock);
1401                 h = __nf_conntrack_helper_find_byname(name);
1402                 if (!h) {
1403                         write_unlock_bh(&nf_conntrack_lock);
1404                         return -EINVAL;
1405                 }
1406                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1407                         hlist_for_each_entry_safe(exp, n, next,
1408                                                   &nf_ct_expect_hash[i],
1409                                                   hnode) {
1410                                 m_help = nfct_help(exp->master);
1411                                 if (m_help->helper == h
1412                                     && del_timer(&exp->timeout)) {
1413                                         nf_ct_unlink_expect(exp);
1414                                         nf_ct_expect_put(exp);
1415                                 }
1416                         }
1417                 }
1418                 write_unlock_bh(&nf_conntrack_lock);
1419         } else {
1420                 /* This basically means we have to flush everything*/
1421                 write_lock_bh(&nf_conntrack_lock);
1422                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1423                         hlist_for_each_entry_safe(exp, n, next,
1424                                                   &nf_ct_expect_hash[i],
1425                                                   hnode) {
1426                                 if (del_timer(&exp->timeout)) {
1427                                         nf_ct_unlink_expect(exp);
1428                                         nf_ct_expect_put(exp);
1429                                 }
1430                         }
1431                 }
1432                 write_unlock_bh(&nf_conntrack_lock);
1433         }
1434
1435         return 0;
1436 }
1437 static int
1438 ctnetlink_change_expect(struct nf_conntrack_expect *x, struct nfattr *cda[])
1439 {
1440         return -EOPNOTSUPP;
1441 }
1442
1443 static int
1444 ctnetlink_create_expect(struct nfattr *cda[], u_int8_t u3)
1445 {
1446         struct nf_conntrack_tuple tuple, mask, master_tuple;
1447         struct nf_conntrack_tuple_hash *h = NULL;
1448         struct nf_conntrack_expect *exp;
1449         struct nf_conn *ct;
1450         struct nf_conn_help *help;
1451         int err = 0;
1452
1453         /* caller guarantees that those three CTA_EXPECT_* exist */
1454         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1455         if (err < 0)
1456                 return err;
1457         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
1458         if (err < 0)
1459                 return err;
1460         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
1461         if (err < 0)
1462                 return err;
1463
1464         /* Look for master conntrack of this expectation */
1465         h = nf_conntrack_find_get(&master_tuple);
1466         if (!h)
1467                 return -ENOENT;
1468         ct = nf_ct_tuplehash_to_ctrack(h);
1469         help = nfct_help(ct);
1470
1471         if (!help || !help->helper) {
1472                 /* such conntrack hasn't got any helper, abort */
1473                 err = -EINVAL;
1474                 goto out;
1475         }
1476
1477         exp = nf_ct_expect_alloc(ct);
1478         if (!exp) {
1479                 err = -ENOMEM;
1480                 goto out;
1481         }
1482
1483         exp->expectfn = NULL;
1484         exp->flags = 0;
1485         exp->master = ct;
1486         exp->helper = NULL;
1487         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
1488         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
1489         exp->mask.src.u.all = mask.src.u.all;
1490
1491         err = nf_ct_expect_related(exp);
1492         nf_ct_expect_put(exp);
1493
1494 out:
1495         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
1496         return err;
1497 }
1498
1499 static int
1500 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
1501                      struct nlmsghdr *nlh, struct nfattr *cda[])
1502 {
1503         struct nf_conntrack_tuple tuple;
1504         struct nf_conntrack_expect *exp;
1505         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1506         u_int8_t u3 = nfmsg->nfgen_family;
1507         int err = 0;
1508
1509         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1510                 return -EINVAL;
1511
1512         if (!cda[CTA_EXPECT_TUPLE-1]
1513             || !cda[CTA_EXPECT_MASK-1]
1514             || !cda[CTA_EXPECT_MASTER-1])
1515                 return -EINVAL;
1516
1517         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1518         if (err < 0)
1519                 return err;
1520
1521         write_lock_bh(&nf_conntrack_lock);
1522         exp = __nf_ct_expect_find(&tuple);
1523
1524         if (!exp) {
1525                 write_unlock_bh(&nf_conntrack_lock);
1526                 err = -ENOENT;
1527                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1528                         err = ctnetlink_create_expect(cda, u3);
1529                 return err;
1530         }
1531
1532         err = -EEXIST;
1533         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
1534                 err = ctnetlink_change_expect(exp, cda);
1535         write_unlock_bh(&nf_conntrack_lock);
1536
1537         return err;
1538 }
1539
1540 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1541 static struct notifier_block ctnl_notifier = {
1542         .notifier_call  = ctnetlink_conntrack_event,
1543 };
1544
1545 static struct notifier_block ctnl_notifier_exp = {
1546         .notifier_call  = ctnetlink_expect_event,
1547 };
1548 #endif
1549
1550 static struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
1551         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
1552                                             .attr_count = CTA_MAX, },
1553         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
1554                                             .attr_count = CTA_MAX, },
1555         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
1556                                             .attr_count = CTA_MAX, },
1557         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
1558                                             .attr_count = CTA_MAX, },
1559 };
1560
1561 static struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
1562         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
1563                                             .attr_count = CTA_EXPECT_MAX, },
1564         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
1565                                             .attr_count = CTA_EXPECT_MAX, },
1566         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
1567                                             .attr_count = CTA_EXPECT_MAX, },
1568 };
1569
1570 static struct nfnetlink_subsystem ctnl_subsys = {
1571         .name                           = "conntrack",
1572         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
1573         .cb_count                       = IPCTNL_MSG_MAX,
1574         .cb                             = ctnl_cb,
1575 };
1576
1577 static struct nfnetlink_subsystem ctnl_exp_subsys = {
1578         .name                           = "conntrack_expect",
1579         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
1580         .cb_count                       = IPCTNL_MSG_EXP_MAX,
1581         .cb                             = ctnl_exp_cb,
1582 };
1583
1584 MODULE_ALIAS("ip_conntrack_netlink");
1585 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
1586 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
1587
1588 static int __init ctnetlink_init(void)
1589 {
1590         int ret;
1591
1592         printk("ctnetlink v%s: registering with nfnetlink.\n", version);
1593         ret = nfnetlink_subsys_register(&ctnl_subsys);
1594         if (ret < 0) {
1595                 printk("ctnetlink_init: cannot register with nfnetlink.\n");
1596                 goto err_out;
1597         }
1598
1599         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
1600         if (ret < 0) {
1601                 printk("ctnetlink_init: cannot register exp with nfnetlink.\n");
1602                 goto err_unreg_subsys;
1603         }
1604
1605 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1606         ret = nf_conntrack_register_notifier(&ctnl_notifier);
1607         if (ret < 0) {
1608                 printk("ctnetlink_init: cannot register notifier.\n");
1609                 goto err_unreg_exp_subsys;
1610         }
1611
1612         ret = nf_ct_expect_register_notifier(&ctnl_notifier_exp);
1613         if (ret < 0) {
1614                 printk("ctnetlink_init: cannot expect register notifier.\n");
1615                 goto err_unreg_notifier;
1616         }
1617 #endif
1618
1619         return 0;
1620
1621 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1622 err_unreg_notifier:
1623         nf_conntrack_unregister_notifier(&ctnl_notifier);
1624 err_unreg_exp_subsys:
1625         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1626 #endif
1627 err_unreg_subsys:
1628         nfnetlink_subsys_unregister(&ctnl_subsys);
1629 err_out:
1630         return ret;
1631 }
1632
1633 static void __exit ctnetlink_exit(void)
1634 {
1635         printk("ctnetlink: unregistering from nfnetlink.\n");
1636
1637 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1638         nf_ct_expect_unregister_notifier(&ctnl_notifier_exp);
1639         nf_conntrack_unregister_notifier(&ctnl_notifier);
1640 #endif
1641
1642         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1643         nfnetlink_subsys_unregister(&ctnl_subsys);
1644         return;
1645 }
1646
1647 module_init(ctnetlink_init);
1648 module_exit(ctnetlink_exit);