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  * I've reworked this stuff to use attributes instead of conntrack
  10  * structures. 5.44 am. I need more tea. --pablo 05/07/11.
  11  *
  12  * Initial connection tracking via netlink development funded and
  13  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  14  *
  15  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  16  *
  17  * This software may be used and distributed according to the terms
  18  * of the GNU General Public License, incorporated herein by reference.
  19  *
  20  * Derived from ip_conntrack_netlink.c: Port by Pablo Neira Ayuso (05/11/14)
  21  */
  22
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/kernel.h>
  26 #include <linux/types.h>
  27 #include <linux/timer.h>
  28 #include <linux/skbuff.h>
  29 #include <linux/errno.h>
  30 #include <linux/netlink.h>
  31 #include <linux/spinlock.h>
  32 #include <linux/interrupt.h>
  33 #include <linux/notifier.h>
  34
  35 #include <linux/netfilter.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_helper.h>
  39 #include <net/netfilter/nf_conntrack_l3proto.h>
  40 #include <net/netfilter/nf_conntrack_protocol.h>
  41 #include <linux/netfilter_ipv4/ip_nat_protocol.h>
  42
  43 #include <linux/netfilter/nfnetlink.h>
  44 #include <linux/netfilter/nfnetlink_conntrack.h>
  45
  46 MODULE_LICENSE("GPL");
  47
  48 static char __initdata version[] = "0.93";
  49
  50 static inline int
  51 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  52                             const struct nf_conntrack_tuple *tuple,
  53                             struct nf_conntrack_protocol *proto)
  54 {
  55         int ret = 0;
  56         struct nfattr *nest_parms = NFA_NEST(skb, CTA_TUPLE_PROTO);
  57
  58         NFA_PUT(skb, CTA_PROTO_NUM, sizeof(u_int8_t), &tuple->dst.protonum);
  59
  60         if (likely(proto->tuple_to_nfattr))
  61                 ret = proto->tuple_to_nfattr(skb, tuple);
  62
  63         NFA_NEST_END(skb, nest_parms);
  64
  65         return ret;
  66
  67 nfattr_failure:
  68         return -1;
  69 }
  70
  71 static inline int
  72 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  73                          const struct nf_conntrack_tuple *tuple,
  74                          struct nf_conntrack_l3proto *l3proto)
  75 {
  76         int ret = 0;
  77         struct nfattr *nest_parms = NFA_NEST(skb, CTA_TUPLE_IP);
  78
  79         if (likely(l3proto->tuple_to_nfattr))
  80                 ret = l3proto->tuple_to_nfattr(skb, tuple);
  81
  82         NFA_NEST_END(skb, nest_parms);
  83
  84         return ret;
  85
  86 nfattr_failure:
  87         return -1;
  88 }
  89
  90 static inline int
  91 ctnetlink_dump_tuples(struct sk_buff *skb,
  92                       const struct nf_conntrack_tuple *tuple)
  93 {
  94         int ret;
  95         struct nf_conntrack_l3proto *l3proto;
  96         struct nf_conntrack_protocol *proto;
  97
  98         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
  99         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 100         nf_ct_l3proto_put(l3proto);
 101
 102         if (unlikely(ret < 0))
 103                 return ret;
 104
 105         proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 106         ret = ctnetlink_dump_tuples_proto(skb, tuple, proto);
 107         nf_ct_proto_put(proto);
 108
 109         return ret;
 110 }
 111
 112 static inline int
 113 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 114 {
 115         u_int32_t status = htonl((u_int32_t) ct->status);
 116         NFA_PUT(skb, CTA_STATUS, sizeof(status), &status);
 117         return 0;
 118
 119 nfattr_failure:
 120         return -1;
 121 }
 122
 123 static inline int
 124 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 125 {
 126         long timeout_l = ct->timeout.expires - jiffies;
 127         u_int32_t timeout;
 128
 129         if (timeout_l < 0)
 130                 timeout = 0;
 131         else
 132                 timeout = htonl(timeout_l / HZ);
 133
 134         NFA_PUT(skb, CTA_TIMEOUT, sizeof(timeout), &timeout);
 135         return 0;
 136
 137 nfattr_failure:
 138         return -1;
 139 }
 140
 141 static inline int
 142 ctnetlink_dump_protoinfo(struct sk_buff *skb, const struct nf_conn *ct)
 143 {
 144         struct nf_conntrack_protocol *proto = nf_ct_proto_find_get(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num, ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum);
 145         struct nfattr *nest_proto;
 146         int ret;
 147
 148         if (!proto->to_nfattr) {
 149                 nf_ct_proto_put(proto);
 150                 return 0;
 151         }
 152
 153         nest_proto = NFA_NEST(skb, CTA_PROTOINFO);
 154
 155         ret = proto->to_nfattr(skb, nest_proto, ct);
 156
 157         nf_ct_proto_put(proto);
 158
 159         NFA_NEST_END(skb, nest_proto);
 160
 161         return ret;
 162
 163 nfattr_failure:
 164         return -1;
 165 }
 166
 167 static inline int
 168 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 169 {
 170         struct nfattr *nest_helper;
 171         const struct nf_conn_help *help = nfct_help(ct);
 172
 173         if (!help || !help->helper)
 174                 return 0;
 175
 176         nest_helper = NFA_NEST(skb, CTA_HELP);
 177         NFA_PUT(skb, CTA_HELP_NAME, strlen(help->helper->name), help->helper->name);
 178
 179         if (help->helper->to_nfattr)
 180                 help->helper->to_nfattr(skb, ct);
 181
 182         NFA_NEST_END(skb, nest_helper);
 183
 184         return 0;
 185
 186 nfattr_failure:
 187         return -1;
 188 }
 189
 190 #ifdef CONFIG_NF_CT_ACCT
 191 static inline int
 192 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 193                         enum ip_conntrack_dir dir)
 194 {
 195         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 196         struct nfattr *nest_count = NFA_NEST(skb, type);
 197         u_int32_t tmp;
 198
 199         tmp = htonl(ct->counters[dir].packets);
 200         NFA_PUT(skb, CTA_COUNTERS32_PACKETS, sizeof(u_int32_t), &tmp);
 201
 202         tmp = htonl(ct->counters[dir].bytes);
 203         NFA_PUT(skb, CTA_COUNTERS32_BYTES, sizeof(u_int32_t), &tmp);
 204
 205         NFA_NEST_END(skb, nest_count);
 206
 207         return 0;
 208
 209 nfattr_failure:
 210         return -1;
 211 }
 212 #else
 213 #define ctnetlink_dump_counters(a, b, c) (0)
 214 #endif
 215
 216 #ifdef CONFIG_NF_CONNTRACK_MARK
 217 static inline int
 218 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 219 {
 220         u_int32_t mark = htonl(ct->mark);
 221
 222         NFA_PUT(skb, CTA_MARK, sizeof(u_int32_t), &mark);
 223         return 0;
 224
 225 nfattr_failure:
 226         return -1;
 227 }
 228 #else
 229 #define ctnetlink_dump_mark(a, b) (0)
 230 #endif
 231
 232 static inline int
 233 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 234 {
 235         u_int32_t id = htonl(ct->id);
 236         NFA_PUT(skb, CTA_ID, sizeof(u_int32_t), &id);
 237         return 0;
 238
 239 nfattr_failure:
 240         return -1;
 241 }
 242
 243 static inline int
 244 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 245 {
 246         u_int32_t use = htonl(atomic_read(&ct->ct_general.use));
 247
 248         NFA_PUT(skb, CTA_USE, sizeof(u_int32_t), &use);
 249         return 0;
 250
 251 nfattr_failure:
 252         return -1;
 253 }
 254
 255 #define tuple(ct, dir) (&(ct)->tuplehash[dir].tuple)
 256
 257 static int
 258 ctnetlink_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
 259                     int event, int nowait,
 260                     const struct nf_conn *ct)
 261 {
 262         struct nlmsghdr *nlh;
 263         struct nfgenmsg *nfmsg;
 264         struct nfattr *nest_parms;
 265         unsigned char *b;
 266
 267         b = skb->tail;
 268
 269         event |= NFNL_SUBSYS_CTNETLINK << 8;
 270         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
 271         nfmsg  = NLMSG_DATA(nlh);
 272
 273         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
 274         nfmsg->nfgen_family =
 275                 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 276         nfmsg->version      = NFNETLINK_V0;
 277         nfmsg->res_id       = 0;
 278
 279         nest_parms = NFA_NEST(skb, CTA_TUPLE_ORIG);
 280         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 281                 goto nfattr_failure;
 282         NFA_NEST_END(skb, nest_parms);
 283
 284         nest_parms = NFA_NEST(skb, CTA_TUPLE_REPLY);
 285         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_REPLY)) < 0)
 286                 goto nfattr_failure;
 287         NFA_NEST_END(skb, nest_parms);
 288
 289         if (ctnetlink_dump_status(skb, ct) < 0 ||
 290             ctnetlink_dump_timeout(skb, ct) < 0 ||
 291             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 292             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 293             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 294             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 295             ctnetlink_dump_mark(skb, ct) < 0 ||
 296             ctnetlink_dump_id(skb, ct) < 0 ||
 297             ctnetlink_dump_use(skb, ct) < 0)
 298                 goto nfattr_failure;
 299
 300         nlh->nlmsg_len = skb->tail - b;
 301         return skb->len;
 302
 303 nlmsg_failure:
 304 nfattr_failure:
 305         skb_trim(skb, b - skb->data);
 306         return -1;
 307 }
 308
 309 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 310 static int ctnetlink_conntrack_event(struct notifier_block *this,
 311                                      unsigned long events, void *ptr)
 312 {
 313         struct nlmsghdr *nlh;
 314         struct nfgenmsg *nfmsg;
 315         struct nfattr *nest_parms;
 316         struct nf_conn *ct = (struct nf_conn *)ptr;
 317         struct sk_buff *skb;
 318         unsigned int type;
 319         unsigned char *b;
 320         unsigned int flags = 0, group;
 321
 322         /* ignore our fake conntrack entry */
 323         if (ct == &nf_conntrack_untracked)
 324                 return NOTIFY_DONE;
 325
 326         if (events & IPCT_DESTROY) {
 327                 type = IPCTNL_MSG_CT_DELETE;
 328                 group = NFNLGRP_CONNTRACK_DESTROY;
 329         } else  if (events & (IPCT_NEW | IPCT_RELATED)) {
 330                 type = IPCTNL_MSG_CT_NEW;
 331                 flags = NLM_F_CREATE|NLM_F_EXCL;
 332                 /* dump everything */
 333                 events = ~0UL;
 334                 group = NFNLGRP_CONNTRACK_NEW;
 335         } else  if (events & (IPCT_STATUS | IPCT_PROTOINFO)) {
 336                 type = IPCTNL_MSG_CT_NEW;
 337                 group = NFNLGRP_CONNTRACK_UPDATE;
 338         } else
 339                 return NOTIFY_DONE;
 340
 341         if (!nfnetlink_has_listeners(group))
 342                 return NOTIFY_DONE;
 343
 344         skb = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
 345         if (!skb)
 346                 return NOTIFY_DONE;
 347
 348         b = skb->tail;
 349
 350         type |= NFNL_SUBSYS_CTNETLINK << 8;
 351         nlh   = NLMSG_PUT(skb, 0, 0, type, sizeof(struct nfgenmsg));
 352         nfmsg = NLMSG_DATA(nlh);
 353
 354         nlh->nlmsg_flags    = flags;
 355         nfmsg->nfgen_family = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 356         nfmsg->version  = NFNETLINK_V0;
 357         nfmsg->res_id   = 0;
 358
 359         nest_parms = NFA_NEST(skb, CTA_TUPLE_ORIG);
 360         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 361                 goto nfattr_failure;
 362         NFA_NEST_END(skb, nest_parms);
 363
 364         nest_parms = NFA_NEST(skb, CTA_TUPLE_REPLY);
 365         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_REPLY)) < 0)
 366                 goto nfattr_failure;
 367         NFA_NEST_END(skb, nest_parms);
 368
 369         /* NAT stuff is now a status flag */
 370         if ((events & IPCT_STATUS || events & IPCT_NATINFO)
 371             && ctnetlink_dump_status(skb, ct) < 0)
 372                 goto nfattr_failure;
 373         if (events & IPCT_REFRESH
 374             && ctnetlink_dump_timeout(skb, ct) < 0)
 375                 goto nfattr_failure;
 376         if (events & IPCT_PROTOINFO
 377             && ctnetlink_dump_protoinfo(skb, ct) < 0)
 378                 goto nfattr_failure;
 379         if (events & IPCT_HELPINFO
 380             && ctnetlink_dump_helpinfo(skb, ct) < 0)
 381                 goto nfattr_failure;
 382
 383         if (ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 384             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0)
 385                 goto nfattr_failure;
 386
 387         if (events & IPCT_MARK
 388             && ctnetlink_dump_mark(skb, ct) < 0)
 389                 goto nfattr_failure;
 390
 391         nlh->nlmsg_len = skb->tail - b;
 392         nfnetlink_send(skb, 0, group, 0);
 393         return NOTIFY_DONE;
 394
 395 nlmsg_failure:
 396 nfattr_failure:
 397         kfree_skb(skb);
 398         return NOTIFY_DONE;
 399 }
 400 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 401
 402 static int ctnetlink_done(struct netlink_callback *cb)
 403 {
 404         if (cb->args[1])
 405                 nf_ct_put((struct nf_conn *)cb->args[1]);
 406         return 0;
 407 }
 408
 409 #define L3PROTO(ct) ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num
 410
 411 static int
 412 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 413 {
 414         struct nf_conn *ct, *last;
 415         struct nf_conntrack_tuple_hash *h;
 416         struct list_head *i;
 417         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
 418         u_int8_t l3proto = nfmsg->nfgen_family;
 419
 420         read_lock_bh(&nf_conntrack_lock);
 421         last = (struct nf_conn *)cb->args[1];
 422         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 423 restart:
 424                 list_for_each_prev(i, &nf_conntrack_hash[cb->args[0]]) {
 425                         h = (struct nf_conntrack_tuple_hash *) i;
 426                         if (DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 427                                 continue;
 428                         ct = nf_ct_tuplehash_to_ctrack(h);
 429                         /* Dump entries of a given L3 protocol number.
 430                          * If it is not specified, ie. l3proto == 0,
 431                          * then dump everything. */
 432                         if (l3proto && L3PROTO(ct) != l3proto)
 433                                 continue;
 434                         if (cb->args[1]) {
 435                                 if (ct != last)
 436                                         continue;
 437                                 cb->args[1] = 0;
 438                         }
 439                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 440                                                 cb->nlh->nlmsg_seq,
 441                                                 IPCTNL_MSG_CT_NEW,
 442                                                 1, ct) < 0) {
 443                                 nf_conntrack_get(&ct->ct_general);
 444                                 cb->args[1] = (unsigned long)ct;
 445                                 goto out;
 446                         }
 447 #ifdef CONFIG_NF_CT_ACCT
 448                         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) ==
 449                                                 IPCTNL_MSG_CT_GET_CTRZERO)
 450                                 memset(&ct->counters, 0, sizeof(ct->counters));
 451 #endif
 452                 }
 453                 if (cb->args[1]) {
 454                         cb->args[1] = 0;
 455                         goto restart;
 456                 }
 457         }
 458 out:
 459         read_unlock_bh(&nf_conntrack_lock);
 460         if (last)
 461                 nf_ct_put(last);
 462
 463         return skb->len;
 464 }
 465
 466 static inline int
 467 ctnetlink_parse_tuple_ip(struct nfattr *attr, struct nf_conntrack_tuple *tuple)
 468 {
 469         struct nfattr *tb[CTA_IP_MAX];
 470         struct nf_conntrack_l3proto *l3proto;
 471         int ret = 0;
 472
 473         nfattr_parse_nested(tb, CTA_IP_MAX, attr);
 474
 475         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
 476
 477         if (likely(l3proto->nfattr_to_tuple))
 478                 ret = l3proto->nfattr_to_tuple(tb, tuple);
 479
 480         nf_ct_l3proto_put(l3proto);
 481
 482         return ret;
 483 }
 484
 485 static const size_t cta_min_proto[CTA_PROTO_MAX] = {
 486         [CTA_PROTO_NUM-1]       = sizeof(u_int8_t),
 487 };
 488
 489 static inline int
 490 ctnetlink_parse_tuple_proto(struct nfattr *attr,
 491                             struct nf_conntrack_tuple *tuple)
 492 {
 493         struct nfattr *tb[CTA_PROTO_MAX];
 494         struct nf_conntrack_protocol *proto;
 495         int ret = 0;
 496
 497         nfattr_parse_nested(tb, CTA_PROTO_MAX, attr);
 498
 499         if (nfattr_bad_size(tb, CTA_PROTO_MAX, cta_min_proto))
 500                 return -EINVAL;
 501
 502         if (!tb[CTA_PROTO_NUM-1])
 503                 return -EINVAL;
 504         tuple->dst.protonum = *(u_int8_t *)NFA_DATA(tb[CTA_PROTO_NUM-1]);
 505
 506         proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 507
 508         if (likely(proto->nfattr_to_tuple))
 509                 ret = proto->nfattr_to_tuple(tb, tuple);
 510
 511         nf_ct_proto_put(proto);
 512
 513         return ret;
 514 }
 515
 516 static inline int
 517 ctnetlink_parse_tuple(struct nfattr *cda[], struct nf_conntrack_tuple *tuple,
 518                       enum ctattr_tuple type, u_int8_t l3num)
 519 {
 520         struct nfattr *tb[CTA_TUPLE_MAX];
 521         int err;
 522
 523         memset(tuple, 0, sizeof(*tuple));
 524
 525         nfattr_parse_nested(tb, CTA_TUPLE_MAX, cda[type-1]);
 526
 527         if (!tb[CTA_TUPLE_IP-1])
 528                 return -EINVAL;
 529
 530         tuple->src.l3num = l3num;
 531
 532         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP-1], tuple);
 533         if (err < 0)
 534                 return err;
 535
 536         if (!tb[CTA_TUPLE_PROTO-1])
 537                 return -EINVAL;
 538
 539         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO-1], tuple);
 540         if (err < 0)
 541                 return err;
 542
 543         /* orig and expect tuples get DIR_ORIGINAL */
 544         if (type == CTA_TUPLE_REPLY)
 545                 tuple->dst.dir = IP_CT_DIR_REPLY;
 546         else
 547                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 548
 549         return 0;
 550 }
 551
 552 #ifdef CONFIG_IP_NF_NAT_NEEDED
 553 static const size_t cta_min_protonat[CTA_PROTONAT_MAX] = {
 554         [CTA_PROTONAT_PORT_MIN-1]       = sizeof(u_int16_t),
 555         [CTA_PROTONAT_PORT_MAX-1]       = sizeof(u_int16_t),
 556 };
 557
 558 static int ctnetlink_parse_nat_proto(struct nfattr *attr,
 559                                      const struct nf_conn *ct,
 560                                      struct ip_nat_range *range)
 561 {
 562         struct nfattr *tb[CTA_PROTONAT_MAX];
 563         struct ip_nat_protocol *npt;
 564
 565         nfattr_parse_nested(tb, CTA_PROTONAT_MAX, attr);
 566
 567         if (nfattr_bad_size(tb, CTA_PROTONAT_MAX, cta_min_protonat))
 568                 return -EINVAL;
 569
 570         npt = ip_nat_proto_find_get(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum);
 571
 572         if (!npt->nfattr_to_range) {
 573                 ip_nat_proto_put(npt);
 574                 return 0;
 575         }
 576
 577         /* nfattr_to_range returns 1 if it parsed, 0 if not, neg. on error */
 578         if (npt->nfattr_to_range(tb, range) > 0)
 579                 range->flags |= IP_NAT_RANGE_PROTO_SPECIFIED;
 580
 581         ip_nat_proto_put(npt);
 582
 583         return 0;
 584 }
 585
 586 static const size_t cta_min_nat[CTA_NAT_MAX] = {
 587         [CTA_NAT_MINIP-1]       = sizeof(u_int32_t),
 588         [CTA_NAT_MAXIP-1]       = sizeof(u_int32_t),
 589 };
 590
 591 static inline int
 592 ctnetlink_parse_nat(struct nfattr *nat,
 593                     const struct nf_conn *ct, struct ip_nat_range *range)
 594 {
 595         struct nfattr *tb[CTA_NAT_MAX];
 596         int err;
 597
 598         memset(range, 0, sizeof(*range));
 599
 600         nfattr_parse_nested(tb, CTA_NAT_MAX, nat);
 601
 602         if (nfattr_bad_size(tb, CTA_NAT_MAX, cta_min_nat))
 603                 return -EINVAL;
 604
 605         if (tb[CTA_NAT_MINIP-1])
 606                 range->min_ip = *(u_int32_t *)NFA_DATA(tb[CTA_NAT_MINIP-1]);
 607
 608         if (!tb[CTA_NAT_MAXIP-1])
 609                 range->max_ip = range->min_ip;
 610         else
 611                 range->max_ip = *(u_int32_t *)NFA_DATA(tb[CTA_NAT_MAXIP-1]);
 612
 613         if (range->min_ip)
 614                 range->flags |= IP_NAT_RANGE_MAP_IPS;
 615
 616         if (!tb[CTA_NAT_PROTO-1])
 617                 return 0;
 618
 619         err = ctnetlink_parse_nat_proto(tb[CTA_NAT_PROTO-1], ct, range);
 620         if (err < 0)
 621                 return err;
 622
 623         return 0;
 624 }
 625 #endif
 626
 627 static inline int
 628 ctnetlink_parse_help(struct nfattr *attr, char **helper_name)
 629 {
 630         struct nfattr *tb[CTA_HELP_MAX];
 631
 632         nfattr_parse_nested(tb, CTA_HELP_MAX, attr);
 633
 634         if (!tb[CTA_HELP_NAME-1])
 635                 return -EINVAL;
 636
 637         *helper_name = NFA_DATA(tb[CTA_HELP_NAME-1]);
 638
 639         return 0;
 640 }
 641
 642 static const size_t cta_min[CTA_MAX] = {
 643         [CTA_STATUS-1]          = sizeof(u_int32_t),
 644         [CTA_TIMEOUT-1]         = sizeof(u_int32_t),
 645         [CTA_MARK-1]            = sizeof(u_int32_t),
 646         [CTA_USE-1]             = sizeof(u_int32_t),
 647         [CTA_ID-1]              = sizeof(u_int32_t)
 648 };
 649
 650 static int
 651 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 652                         struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
 653 {
 654         struct nf_conntrack_tuple_hash *h;
 655         struct nf_conntrack_tuple tuple;
 656         struct nf_conn *ct;
 657         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 658         u_int8_t u3 = nfmsg->nfgen_family;
 659         int err = 0;
 660
 661         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 662                 return -EINVAL;
 663
 664         if (cda[CTA_TUPLE_ORIG-1])
 665                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 666         else if (cda[CTA_TUPLE_REPLY-1])
 667                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 668         else {
 669                 /* Flush the whole table */
 670                 nf_conntrack_flush();
 671                 return 0;
 672         }
 673
 674         if (err < 0)
 675                 return err;
 676
 677         h = nf_conntrack_find_get(&tuple, NULL);
 678         if (!h)
 679                 return -ENOENT;
 680
 681         ct = nf_ct_tuplehash_to_ctrack(h);
 682
 683         if (cda[CTA_ID-1]) {
 684                 u_int32_t id = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_ID-1]));
 685                 if (ct->id != id) {
 686                         nf_ct_put(ct);
 687                         return -ENOENT;
 688                 }
 689         }
 690         if (del_timer(&ct->timeout))
 691                 ct->timeout.function((unsigned long)ct);
 692
 693         nf_ct_put(ct);
 694
 695         return 0;
 696 }
 697
 698 static int
 699 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 700                         struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
 701 {
 702         struct nf_conntrack_tuple_hash *h;
 703         struct nf_conntrack_tuple tuple;
 704         struct nf_conn *ct;
 705         struct sk_buff *skb2 = NULL;
 706         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 707         u_int8_t u3 = nfmsg->nfgen_family;
 708         int err = 0;
 709
 710         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 711                 u32 rlen;
 712
 713 #ifndef CONFIG_NF_CT_ACCT
 714                 if (NFNL_MSG_TYPE(nlh->nlmsg_type) == IPCTNL_MSG_CT_GET_CTRZERO)
 715                         return -ENOTSUPP;
 716 #endif
 717                 if ((*errp = netlink_dump_start(ctnl, skb, nlh,
 718                                                 ctnetlink_dump_table,
 719                                                 ctnetlink_done)) != 0)
 720                         return -EINVAL;
 721
 722                 rlen = NLMSG_ALIGN(nlh->nlmsg_len);
 723                 if (rlen > skb->len)
 724                         rlen = skb->len;
 725                 skb_pull(skb, rlen);
 726                 return 0;
 727         }
 728
 729         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 730                 return -EINVAL;
 731
 732         if (cda[CTA_TUPLE_ORIG-1])
 733                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 734         else if (cda[CTA_TUPLE_REPLY-1])
 735                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 736         else
 737                 return -EINVAL;
 738
 739         if (err < 0)
 740                 return err;
 741
 742         h = nf_conntrack_find_get(&tuple, NULL);
 743         if (!h)
 744                 return -ENOENT;
 745
 746         ct = nf_ct_tuplehash_to_ctrack(h);
 747
 748         err = -ENOMEM;
 749         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 750         if (!skb2) {
 751                 nf_ct_put(ct);
 752                 return -ENOMEM;
 753         }
 754         NETLINK_CB(skb2).dst_pid = NETLINK_CB(skb).pid;
 755
 756         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
 757                                   IPCTNL_MSG_CT_NEW, 1, ct);
 758         nf_ct_put(ct);
 759         if (err <= 0)
 760                 goto free;
 761
 762         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
 763         if (err < 0)
 764                 goto out;
 765
 766         return 0;
 767
 768 free:
 769         kfree_skb(skb2);
 770 out:
 771         return err;
 772 }
 773
 774 static inline int
 775 ctnetlink_change_status(struct nf_conn *ct, struct nfattr *cda[])
 776 {
 777         unsigned long d;
 778         unsigned status = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_STATUS-1]));
 779         d = ct->status ^ status;
 780
 781         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
 782                 /* unchangeable */
 783                 return -EINVAL;
 784
 785         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
 786                 /* SEEN_REPLY bit can only be set */
 787                 return -EINVAL;
 788
 789
 790         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
 791                 /* ASSURED bit can only be set */
 792                 return -EINVAL;
 793
 794         if (cda[CTA_NAT_SRC-1] || cda[CTA_NAT_DST-1]) {
 795 #ifndef CONFIG_IP_NF_NAT_NEEDED
 796                 return -EINVAL;
 797 #else
 798                 struct ip_nat_range range;
 799
 800                 if (cda[CTA_NAT_DST-1]) {
 801                         if (ctnetlink_parse_nat(cda[CTA_NAT_DST-1], ct,
 802                                                 &range) < 0)
 803                                 return -EINVAL;
 804                         if (ip_nat_initialized(ct,
 805                                                HOOK2MANIP(NF_IP_PRE_ROUTING)))
 806                                 return -EEXIST;
 807                         ip_nat_setup_info(ct, &range, hooknum);
 808                 }
 809                 if (cda[CTA_NAT_SRC-1]) {
 810                         if (ctnetlink_parse_nat(cda[CTA_NAT_SRC-1], ct,
 811                                                 &range) < 0)
 812                                 return -EINVAL;
 813                         if (ip_nat_initialized(ct,
 814                                                HOOK2MANIP(NF_IP_POST_ROUTING)))
 815                                 return -EEXIST;
 816                         ip_nat_setup_info(ct, &range, hooknum);
 817                 }
 818 #endif
 819         }
 820
 821         /* Be careful here, modifying NAT bits can screw up things,
 822          * so don't let users modify them directly if they don't pass
 823          * ip_nat_range. */
 824         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
 825         return 0;
 826 }
 827
 828
 829 static inline int
 830 ctnetlink_change_helper(struct nf_conn *ct, struct nfattr *cda[])
 831 {
 832         struct nf_conntrack_helper *helper;
 833         struct nf_conn_help *help = nfct_help(ct);
 834         char *helpname;
 835         int err;
 836
 837         if (!help) {
 838                 /* FIXME: we need to reallocate and rehash */
 839                 return -EBUSY;
 840         }
 841
 842         /* don't change helper of sibling connections */
 843         if (ct->master)
 844                 return -EINVAL;
 845
 846         err = ctnetlink_parse_help(cda[CTA_HELP-1], &helpname);
 847         if (err < 0)
 848                 return err;
 849
 850         helper = __nf_conntrack_helper_find_byname(helpname);
 851         if (!helper) {
 852                 if (!strcmp(helpname, ""))
 853                         helper = NULL;
 854                 else
 855                         return -EINVAL;
 856         }
 857
 858         if (help->helper) {
 859                 if (!helper) {
 860                         /* we had a helper before ... */
 861                         nf_ct_remove_expectations(ct);
 862                         help->helper = NULL;
 863                 } else {
 864                         /* need to zero data of old helper */
 865                         memset(&help->help, 0, sizeof(help->help));
 866                 }
 867         }
 868
 869         help->helper = helper;
 870
 871         return 0;
 872 }
 873
 874 static inline int
 875 ctnetlink_change_timeout(struct nf_conn *ct, struct nfattr *cda[])
 876 {
 877         u_int32_t timeout = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_TIMEOUT-1]));
 878
 879         if (!del_timer(&ct->timeout))
 880                 return -ETIME;
 881
 882         ct->timeout.expires = jiffies + timeout * HZ;
 883         add_timer(&ct->timeout);
 884
 885         return 0;
 886 }
 887
 888 static inline int
 889 ctnetlink_change_protoinfo(struct nf_conn *ct, struct nfattr *cda[])
 890 {
 891         struct nfattr *tb[CTA_PROTOINFO_MAX], *attr = cda[CTA_PROTOINFO-1];
 892         struct nf_conntrack_protocol *proto;
 893         u_int16_t npt = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum;
 894         u_int16_t l3num = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 895         int err = 0;
 896
 897         nfattr_parse_nested(tb, CTA_PROTOINFO_MAX, attr);
 898
 899         proto = nf_ct_proto_find_get(l3num, npt);
 900
 901         if (proto->from_nfattr)
 902                 err = proto->from_nfattr(tb, ct);
 903         nf_ct_proto_put(proto);
 904
 905         return err;
 906 }
 907
 908 static int
 909 ctnetlink_change_conntrack(struct nf_conn *ct, struct nfattr *cda[])
 910 {
 911         int err;
 912
 913         if (cda[CTA_HELP-1]) {
 914                 err = ctnetlink_change_helper(ct, cda);
 915                 if (err < 0)
 916                         return err;
 917         }
 918
 919         if (cda[CTA_TIMEOUT-1]) {
 920                 err = ctnetlink_change_timeout(ct, cda);
 921                 if (err < 0)
 922                         return err;
 923         }
 924
 925         if (cda[CTA_STATUS-1]) {
 926                 err = ctnetlink_change_status(ct, cda);
 927                 if (err < 0)
 928                         return err;
 929         }
 930
 931         if (cda[CTA_PROTOINFO-1]) {
 932                 err = ctnetlink_change_protoinfo(ct, cda);
 933                 if (err < 0)
 934                         return err;
 935         }
 936
 937 #if defined(CONFIG_NF_CONNTRACK_MARK)
 938         if (cda[CTA_MARK-1])
 939                 ct->mark = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_MARK-1]));
 940 #endif
 941
 942         return 0;
 943 }
 944
 945 static int
 946 ctnetlink_create_conntrack(struct nfattr *cda[],
 947                            struct nf_conntrack_tuple *otuple,
 948                            struct nf_conntrack_tuple *rtuple)
 949 {
 950         struct nf_conn *ct;
 951         int err = -EINVAL;
 952
 953         ct = nf_conntrack_alloc(otuple, rtuple);
 954         if (ct == NULL || IS_ERR(ct))
 955                 return -ENOMEM;
 956
 957         if (!cda[CTA_TIMEOUT-1])
 958                 goto err;
 959         ct->timeout.expires = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_TIMEOUT-1]));
 960
 961         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
 962         ct->status |= IPS_CONFIRMED;
 963
 964         err = ctnetlink_change_status(ct, cda);
 965         if (err < 0)
 966                 goto err;
 967
 968         if (cda[CTA_PROTOINFO-1]) {
 969                 err = ctnetlink_change_protoinfo(ct, cda);
 970                 if (err < 0)
 971                         return err;
 972         }
 973
 974 #if defined(CONFIG_NF_CONNTRACK_MARK)
 975         if (cda[CTA_MARK-1])
 976                 ct->mark = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_MARK-1]));
 977 #endif
 978
 979         add_timer(&ct->timeout);
 980         nf_conntrack_hash_insert(ct);
 981
 982         return 0;
 983
 984 err:
 985         nf_conntrack_free(ct);
 986         return err;
 987 }
 988
 989 static int
 990 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
 991                         struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
 992 {
 993         struct nf_conntrack_tuple otuple, rtuple;
 994         struct nf_conntrack_tuple_hash *h = NULL;
 995         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 996         u_int8_t u3 = nfmsg->nfgen_family;
 997         int err = 0;
 998
 999         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
1000                 return -EINVAL;
1001
1002         if (cda[CTA_TUPLE_ORIG-1]) {
1003                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1004                 if (err < 0)
1005                         return err;
1006         }
1007
1008         if (cda[CTA_TUPLE_REPLY-1]) {
1009                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1010                 if (err < 0)
1011                         return err;
1012         }
1013
1014         write_lock_bh(&nf_conntrack_lock);
1015         if (cda[CTA_TUPLE_ORIG-1])
1016                 h = __nf_conntrack_find(&otuple, NULL);
1017         else if (cda[CTA_TUPLE_REPLY-1])
1018                 h = __nf_conntrack_find(&rtuple, NULL);
1019
1020         if (h == NULL) {
1021                 write_unlock_bh(&nf_conntrack_lock);
1022                 err = -ENOENT;
1023                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1024                         err = ctnetlink_create_conntrack(cda, &otuple, &rtuple);
1025                 return err;
1026         }
1027         /* implicit 'else' */
1028
1029         /* we only allow nat config for new conntracks */
1030         if (cda[CTA_NAT_SRC-1] || cda[CTA_NAT_DST-1]) {
1031                 err = -EINVAL;
1032                 goto out_unlock;
1033         }
1034
1035         /* We manipulate the conntrack inside the global conntrack table lock,
1036          * so there's no need to increase the refcount */
1037         err = -EEXIST;
1038         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
1039                 err = ctnetlink_change_conntrack(nf_ct_tuplehash_to_ctrack(h), cda);
1040
1041 out_unlock:
1042         write_unlock_bh(&nf_conntrack_lock);
1043         return err;
1044 }
1045
1046 /***********************************************************************
1047  * EXPECT
1048  ***********************************************************************/
1049
1050 static inline int
1051 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1052                          const struct nf_conntrack_tuple *tuple,
1053                          enum ctattr_expect type)
1054 {
1055         struct nfattr *nest_parms = NFA_NEST(skb, type);
1056
1057         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1058                 goto nfattr_failure;
1059
1060         NFA_NEST_END(skb, nest_parms);
1061
1062         return 0;
1063
1064 nfattr_failure:
1065         return -1;
1066 }
1067
1068 static inline int
1069 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1070                         const struct nf_conntrack_tuple *tuple,
1071                         const struct nf_conntrack_tuple *mask)
1072 {
1073         int ret;
1074         struct nf_conntrack_l3proto *l3proto;
1075         struct nf_conntrack_protocol *proto;
1076         struct nfattr *nest_parms = NFA_NEST(skb, CTA_EXPECT_MASK);
1077
1078         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
1079         ret = ctnetlink_dump_tuples_ip(skb, mask, l3proto);
1080         nf_ct_l3proto_put(l3proto);
1081
1082         if (unlikely(ret < 0))
1083                 goto nfattr_failure;
1084
1085         proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
1086         ret = ctnetlink_dump_tuples_proto(skb, mask, proto);
1087         nf_ct_proto_put(proto);
1088         if (unlikely(ret < 0))
1089                 goto nfattr_failure;
1090
1091         NFA_NEST_END(skb, nest_parms);
1092
1093         return 0;
1094
1095 nfattr_failure:
1096         return -1;
1097 }
1098
1099 static inline int
1100 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1101                           const struct nf_conntrack_expect *exp)
1102 {
1103         struct nf_conn *master = exp->master;
1104         u_int32_t timeout = htonl((exp->timeout.expires - jiffies) / HZ);
1105         u_int32_t id = htonl(exp->id);
1106
1107         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1108                 goto nfattr_failure;
1109         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1110                 goto nfattr_failure;
1111         if (ctnetlink_exp_dump_tuple(skb,
1112                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1113                                  CTA_EXPECT_MASTER) < 0)
1114                 goto nfattr_failure;
1115
1116         NFA_PUT(skb, CTA_EXPECT_TIMEOUT, sizeof(timeout), &timeout);
1117         NFA_PUT(skb, CTA_EXPECT_ID, sizeof(u_int32_t), &id);
1118
1119         return 0;
1120
1121 nfattr_failure:
1122         return -1;
1123 }
1124
1125 static int
1126 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1127                     int event,
1128                     int nowait,
1129                     const struct nf_conntrack_expect *exp)
1130 {
1131         struct nlmsghdr *nlh;
1132         struct nfgenmsg *nfmsg;
1133         unsigned char *b;
1134
1135         b = skb->tail;
1136
1137         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1138         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
1139         nfmsg  = NLMSG_DATA(nlh);
1140
1141         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
1142         nfmsg->nfgen_family = exp->tuple.src.l3num;
1143         nfmsg->version      = NFNETLINK_V0;
1144         nfmsg->res_id       = 0;
1145
1146         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1147                 goto nfattr_failure;
1148
1149         nlh->nlmsg_len = skb->tail - b;
1150         return skb->len;
1151
1152 nlmsg_failure:
1153 nfattr_failure:
1154         skb_trim(skb, b - skb->data);
1155         return -1;
1156 }
1157
1158 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1159 static int ctnetlink_expect_event(struct notifier_block *this,
1160                                   unsigned long events, void *ptr)
1161 {
1162         struct nlmsghdr *nlh;
1163         struct nfgenmsg *nfmsg;
1164         struct nf_conntrack_expect *exp = (struct nf_conntrack_expect *)ptr;
1165         struct sk_buff *skb;
1166         unsigned int type;
1167         unsigned char *b;
1168         int flags = 0;
1169
1170         if (events & IPEXP_NEW) {
1171                 type = IPCTNL_MSG_EXP_NEW;
1172                 flags = NLM_F_CREATE|NLM_F_EXCL;
1173         } else
1174                 return NOTIFY_DONE;
1175
1176         if (!nfnetlink_has_listeners(NFNLGRP_CONNTRACK_EXP_NEW))
1177                 return NOTIFY_DONE;
1178
1179         skb = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1180         if (!skb)
1181                 return NOTIFY_DONE;
1182
1183         b = skb->tail;
1184
1185         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1186         nlh   = NLMSG_PUT(skb, 0, 0, type, sizeof(struct nfgenmsg));
1187         nfmsg = NLMSG_DATA(nlh);
1188
1189         nlh->nlmsg_flags    = flags;
1190         nfmsg->nfgen_family = exp->tuple.src.l3num;
1191         nfmsg->version      = NFNETLINK_V0;
1192         nfmsg->res_id       = 0;
1193
1194         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1195                 goto nfattr_failure;
1196
1197         nlh->nlmsg_len = skb->tail - b;
1198         nfnetlink_send(skb, 0, NFNLGRP_CONNTRACK_EXP_NEW, 0);
1199         return NOTIFY_DONE;
1200
1201 nlmsg_failure:
1202 nfattr_failure:
1203         kfree_skb(skb);
1204         return NOTIFY_DONE;
1205 }
1206 #endif
1207
1208 static int
1209 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1210 {
1211         struct nf_conntrack_expect *exp = NULL;
1212         struct list_head *i;
1213         u_int32_t *id = (u_int32_t *) &cb->args[0];
1214         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
1215         u_int8_t l3proto = nfmsg->nfgen_family;
1216
1217         read_lock_bh(&nf_conntrack_lock);
1218         list_for_each_prev(i, &nf_conntrack_expect_list) {
1219                 exp = (struct nf_conntrack_expect *) i;
1220                 if (l3proto && exp->tuple.src.l3num != l3proto)
1221                         continue;
1222                 if (exp->id <= *id)
1223                         continue;
1224                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).pid,
1225                                             cb->nlh->nlmsg_seq,
1226                                             IPCTNL_MSG_EXP_NEW,
1227                                             1, exp) < 0)
1228                         goto out;
1229                 *id = exp->id;
1230         }
1231 out:
1232         read_unlock_bh(&nf_conntrack_lock);
1233
1234         return skb->len;
1235 }
1236
1237 static const size_t cta_min_exp[CTA_EXPECT_MAX] = {
1238         [CTA_EXPECT_TIMEOUT-1]          = sizeof(u_int32_t),
1239         [CTA_EXPECT_ID-1]               = sizeof(u_int32_t)
1240 };
1241
1242 static int
1243 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1244                      struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
1245 {
1246         struct nf_conntrack_tuple tuple;
1247         struct nf_conntrack_expect *exp;
1248         struct sk_buff *skb2;
1249         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1250         u_int8_t u3 = nfmsg->nfgen_family;
1251         int err = 0;
1252
1253         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1254                 return -EINVAL;
1255
1256         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1257                 u32 rlen;
1258
1259                 if ((*errp = netlink_dump_start(ctnl, skb, nlh,
1260                                                 ctnetlink_exp_dump_table,
1261                                                 ctnetlink_done)) != 0)
1262                         return -EINVAL;
1263                 rlen = NLMSG_ALIGN(nlh->nlmsg_len);
1264                 if (rlen > skb->len)
1265                         rlen = skb->len;
1266                 skb_pull(skb, rlen);
1267                 return 0;
1268         }
1269
1270         if (cda[CTA_EXPECT_MASTER-1])
1271                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1272         else
1273                 return -EINVAL;
1274
1275         if (err < 0)
1276                 return err;
1277
1278         exp = nf_conntrack_expect_find(&tuple);
1279         if (!exp)
1280                 return -ENOENT;
1281
1282         if (cda[CTA_EXPECT_ID-1]) {
1283                 u_int32_t id = *(u_int32_t *)NFA_DATA(cda[CTA_EXPECT_ID-1]);
1284                 if (exp->id != ntohl(id)) {
1285                         nf_conntrack_expect_put(exp);
1286                         return -ENOENT;
1287                 }
1288         }
1289
1290         err = -ENOMEM;
1291         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
1292         if (!skb2)
1293                 goto out;
1294         NETLINK_CB(skb2).dst_pid = NETLINK_CB(skb).pid;
1295
1296         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1297                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
1298                                       1, exp);
1299         if (err <= 0)
1300                 goto free;
1301
1302         nf_conntrack_expect_put(exp);
1303
1304         return netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1305
1306 free:
1307         kfree_skb(skb2);
1308 out:
1309         nf_conntrack_expect_put(exp);
1310         return err;
1311 }
1312
1313 static int
1314 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1315                      struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
1316 {
1317         struct nf_conntrack_expect *exp, *tmp;
1318         struct nf_conntrack_tuple tuple;
1319         struct nf_conntrack_helper *h;
1320         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1321         u_int8_t u3 = nfmsg->nfgen_family;
1322         int err;
1323
1324         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1325                 return -EINVAL;
1326
1327         if (cda[CTA_EXPECT_TUPLE-1]) {
1328                 /* delete a single expect by tuple */
1329                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1330                 if (err < 0)
1331                         return err;
1332
1333                 /* bump usage count to 2 */
1334                 exp = nf_conntrack_expect_find(&tuple);
1335                 if (!exp)
1336                         return -ENOENT;
1337
1338                 if (cda[CTA_EXPECT_ID-1]) {
1339                         u_int32_t id =
1340                                 *(u_int32_t *)NFA_DATA(cda[CTA_EXPECT_ID-1]);
1341                         if (exp->id != ntohl(id)) {
1342                                 nf_conntrack_expect_put(exp);
1343                                 return -ENOENT;
1344                         }
1345                 }
1346
1347                 /* after list removal, usage count == 1 */
1348                 nf_conntrack_unexpect_related(exp);
1349                 /* have to put what we 'get' above.
1350                  * after this line usage count == 0 */
1351                 nf_conntrack_expect_put(exp);
1352         } else if (cda[CTA_EXPECT_HELP_NAME-1]) {
1353                 char *name = NFA_DATA(cda[CTA_EXPECT_HELP_NAME-1]);
1354
1355                 /* delete all expectations for this helper */
1356                 write_lock_bh(&nf_conntrack_lock);
1357                 h = __nf_conntrack_helper_find_byname(name);
1358                 if (!h) {
1359                         write_unlock_bh(&nf_conntrack_lock);
1360                         return -EINVAL;
1361                 }
1362                 list_for_each_entry_safe(exp, tmp, &nf_conntrack_expect_list,
1363                                          list) {
1364                         struct nf_conn_help *m_help = nfct_help(exp->master);
1365                         if (m_help->helper == h
1366                             && del_timer(&exp->timeout)) {
1367                                 nf_ct_unlink_expect(exp);
1368                                 nf_conntrack_expect_put(exp);
1369                         }
1370                 }
1371                 write_unlock_bh(&nf_conntrack_lock);
1372         } else {
1373                 /* This basically means we have to flush everything*/
1374                 write_lock_bh(&nf_conntrack_lock);
1375                 list_for_each_entry_safe(exp, tmp, &nf_conntrack_expect_list,
1376                                          list) {
1377                         if (del_timer(&exp->timeout)) {
1378                                 nf_ct_unlink_expect(exp);
1379                                 nf_conntrack_expect_put(exp);
1380                         }
1381                 }
1382                 write_unlock_bh(&nf_conntrack_lock);
1383         }
1384
1385         return 0;
1386 }
1387 static int
1388 ctnetlink_change_expect(struct nf_conntrack_expect *x, struct nfattr *cda[])
1389 {
1390         return -EOPNOTSUPP;
1391 }
1392
1393 static int
1394 ctnetlink_create_expect(struct nfattr *cda[], u_int8_t u3)
1395 {
1396         struct nf_conntrack_tuple tuple, mask, master_tuple;
1397         struct nf_conntrack_tuple_hash *h = NULL;
1398         struct nf_conntrack_expect *exp;
1399         struct nf_conn *ct;
1400         struct nf_conn_help *help;
1401         int err = 0;
1402
1403         /* caller guarantees that those three CTA_EXPECT_* exist */
1404         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1405         if (err < 0)
1406                 return err;
1407         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
1408         if (err < 0)
1409                 return err;
1410         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
1411         if (err < 0)
1412                 return err;
1413
1414         /* Look for master conntrack of this expectation */
1415         h = nf_conntrack_find_get(&master_tuple, NULL);
1416         if (!h)
1417                 return -ENOENT;
1418         ct = nf_ct_tuplehash_to_ctrack(h);
1419         help = nfct_help(ct);
1420
1421         if (!help || !help->helper) {
1422                 /* such conntrack hasn't got any helper, abort */
1423                 err = -EINVAL;
1424                 goto out;
1425         }
1426
1427         exp = nf_conntrack_expect_alloc(ct);
1428         if (!exp) {
1429                 err = -ENOMEM;
1430                 goto out;
1431         }
1432
1433         exp->expectfn = NULL;
1434         exp->flags = 0;
1435         exp->master = ct;
1436         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
1437         memcpy(&exp->mask, &mask, sizeof(struct nf_conntrack_tuple));
1438
1439         err = nf_conntrack_expect_related(exp);
1440         nf_conntrack_expect_put(exp);
1441
1442 out:
1443         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
1444         return err;
1445 }
1446
1447 static int
1448 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
1449                      struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
1450 {
1451         struct nf_conntrack_tuple tuple;
1452         struct nf_conntrack_expect *exp;
1453         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1454         u_int8_t u3 = nfmsg->nfgen_family;
1455         int err = 0;
1456
1457         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1458                 return -EINVAL;
1459
1460         if (!cda[CTA_EXPECT_TUPLE-1]
1461             || !cda[CTA_EXPECT_MASK-1]
1462             || !cda[CTA_EXPECT_MASTER-1])
1463                 return -EINVAL;
1464
1465         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1466         if (err < 0)
1467                 return err;
1468
1469         write_lock_bh(&nf_conntrack_lock);
1470         exp = __nf_conntrack_expect_find(&tuple);
1471
1472         if (!exp) {
1473                 write_unlock_bh(&nf_conntrack_lock);
1474                 err = -ENOENT;
1475                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1476                         err = ctnetlink_create_expect(cda, u3);
1477                 return err;
1478         }
1479
1480         err = -EEXIST;
1481         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
1482                 err = ctnetlink_change_expect(exp, cda);
1483         write_unlock_bh(&nf_conntrack_lock);
1484
1485         return err;
1486 }
1487
1488 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1489 static struct notifier_block ctnl_notifier = {
1490         .notifier_call  = ctnetlink_conntrack_event,
1491 };
1492
1493 static struct notifier_block ctnl_notifier_exp = {
1494         .notifier_call  = ctnetlink_expect_event,
1495 };
1496 #endif
1497
1498 static struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
1499         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
1500                                             .attr_count = CTA_MAX, },
1501         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
1502                                             .attr_count = CTA_MAX, },
1503         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
1504                                             .attr_count = CTA_MAX, },
1505         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
1506                                             .attr_count = CTA_MAX, },
1507 };
1508
1509 static struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
1510         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
1511                                             .attr_count = CTA_EXPECT_MAX, },
1512         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
1513                                             .attr_count = CTA_EXPECT_MAX, },
1514         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
1515                                             .attr_count = CTA_EXPECT_MAX, },
1516 };
1517
1518 static struct nfnetlink_subsystem ctnl_subsys = {
1519         .name                           = "conntrack",
1520         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
1521         .cb_count                       = IPCTNL_MSG_MAX,
1522         .cb                             = ctnl_cb,
1523 };
1524
1525 static struct nfnetlink_subsystem ctnl_exp_subsys = {
1526         .name                           = "conntrack_expect",
1527         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
1528         .cb_count                       = IPCTNL_MSG_EXP_MAX,
1529         .cb                             = ctnl_exp_cb,
1530 };
1531
1532 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
1533 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
1534
1535 static int __init ctnetlink_init(void)
1536 {
1537         int ret;
1538
1539         printk("ctnetlink v%s: registering with nfnetlink.\n", version);
1540         ret = nfnetlink_subsys_register(&ctnl_subsys);
1541         if (ret < 0) {
1542                 printk("ctnetlink_init: cannot register with nfnetlink.\n");
1543                 goto err_out;
1544         }
1545
1546         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
1547         if (ret < 0) {
1548                 printk("ctnetlink_init: cannot register exp with nfnetlink.\n");
1549                 goto err_unreg_subsys;
1550         }
1551
1552 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1553         ret = nf_conntrack_register_notifier(&ctnl_notifier);
1554         if (ret < 0) {
1555                 printk("ctnetlink_init: cannot register notifier.\n");
1556                 goto err_unreg_exp_subsys;
1557         }
1558
1559         ret = nf_conntrack_expect_register_notifier(&ctnl_notifier_exp);
1560         if (ret < 0) {
1561                 printk("ctnetlink_init: cannot expect register notifier.\n");
1562                 goto err_unreg_notifier;
1563         }
1564 #endif
1565
1566         return 0;
1567
1568 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1569 err_unreg_notifier:
1570         nf_conntrack_unregister_notifier(&ctnl_notifier);
1571 err_unreg_exp_subsys:
1572         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1573 #endif
1574 err_unreg_subsys:
1575         nfnetlink_subsys_unregister(&ctnl_subsys);
1576 err_out:
1577         return ret;
1578 }
1579
1580 static void __exit ctnetlink_exit(void)
1581 {
1582         printk("ctnetlink: unregistering from nfnetlink.\n");
1583
1584 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1585         nf_conntrack_expect_unregister_notifier(&ctnl_notifier_exp);
1586         nf_conntrack_unregister_notifier(&ctnl_notifier);
1587 #endif
1588
1589         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1590         nfnetlink_subsys_unregister(&ctnl_subsys);
1591         return;
1592 }
1593
1594 module_init(ctnetlink_init);
1595 module_exit(ctnetlink_exit);