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 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/notifier.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netfilter/nf_conntrack.h>
  36 #include <net/netfilter/nf_conntrack_core.h>
  37 #include <net/netfilter/nf_conntrack_helper.h>
  38 #include <net/netfilter/nf_conntrack_l3proto.h>
  39 #include <net/netfilter/nf_conntrack_protocol.h>
  40 #include <linux/netfilter_ipv4/ip_nat_protocol.h>
  41
  42 #include <linux/netfilter/nfnetlink.h>
  43 #include <linux/netfilter/nfnetlink_conntrack.h>
  44
  45 MODULE_LICENSE("GPL");
  46
  47 static char __initdata version[] = "0.93";
  48
  49 #if 0
  50 #define DEBUGP printk
  51 #else
  52 #define DEBUGP(format, args...)
  53 #endif
  54
  55
  56 static inline int
  57 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  58                             const struct nf_conntrack_tuple *tuple)
  59 {
  60         struct nf_conntrack_protocol *proto;
  61         int ret = 0;
  62
  63         NFA_PUT(skb, CTA_PROTO_NUM, sizeof(u_int8_t), &tuple->dst.protonum);
  64
  65         /* If no protocol helper is found, this function will return the
  66          * generic protocol helper, so proto won't *ever* be NULL */
  67         proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
  68         if (likely(proto->tuple_to_nfattr))
  69                 ret = proto->tuple_to_nfattr(skb, tuple);
  70
  71         nf_ct_proto_put(proto);
  72
  73         return ret;
  74
  75 nfattr_failure:
  76         return -1;
  77 }
  78
  79 static inline int
  80 ctnetlink_dump_tuples(struct sk_buff *skb,
  81                       const struct nf_conntrack_tuple *tuple)
  82 {
  83         struct nfattr *nest_parms;
  84         struct nf_conntrack_l3proto *l3proto;
  85         int ret = 0;
  86
  87         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
  88
  89         nest_parms = NFA_NEST(skb, CTA_TUPLE_IP);
  90         if (likely(l3proto->tuple_to_nfattr))
  91                 ret = l3proto->tuple_to_nfattr(skb, tuple);
  92         NFA_NEST_END(skb, nest_parms);
  93
  94         nf_ct_l3proto_put(l3proto);
  95
  96         if (unlikely(ret < 0))
  97                 return ret;
  98
  99         nest_parms = NFA_NEST(skb, CTA_TUPLE_PROTO);
 100         ret = ctnetlink_dump_tuples_proto(skb, tuple);
 101         NFA_NEST_END(skb, nest_parms);
 102
 103         return ret;
 104
 105 nfattr_failure:
 106         return -1;
 107 }
 108
 109 static inline int
 110 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 111 {
 112         u_int32_t status = htonl((u_int32_t) ct->status);
 113         NFA_PUT(skb, CTA_STATUS, sizeof(status), &status);
 114         return 0;
 115
 116 nfattr_failure:
 117         return -1;
 118 }
 119
 120 static inline int
 121 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 122 {
 123         long timeout_l = ct->timeout.expires - jiffies;
 124         u_int32_t timeout;
 125
 126         if (timeout_l < 0)
 127                 timeout = 0;
 128         else
 129                 timeout = htonl(timeout_l / HZ);
 130
 131         NFA_PUT(skb, CTA_TIMEOUT, sizeof(timeout), &timeout);
 132         return 0;
 133
 134 nfattr_failure:
 135         return -1;
 136 }
 137
 138 static inline int
 139 ctnetlink_dump_protoinfo(struct sk_buff *skb, const struct nf_conn *ct)
 140 {
 141         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);
 142         struct nfattr *nest_proto;
 143         int ret;
 144
 145         if (!proto->to_nfattr) {
 146                 nf_ct_proto_put(proto);
 147                 return 0;
 148         }
 149
 150         nest_proto = NFA_NEST(skb, CTA_PROTOINFO);
 151
 152         ret = proto->to_nfattr(skb, nest_proto, ct);
 153
 154         nf_ct_proto_put(proto);
 155
 156         NFA_NEST_END(skb, nest_proto);
 157
 158         return ret;
 159
 160 nfattr_failure:
 161         return -1;
 162 }
 163
 164 static inline int
 165 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 166 {
 167         struct nfattr *nest_helper;
 168         const struct nf_conn_help *help = nfct_help(ct);
 169
 170         if (!help || !help->helper)
 171                 return 0;
 172
 173         nest_helper = NFA_NEST(skb, CTA_HELP);
 174         NFA_PUT(skb, CTA_HELP_NAME, strlen(help->helper->name), help->helper->name);
 175
 176         if (help->helper->to_nfattr)
 177                 help->helper->to_nfattr(skb, ct);
 178
 179         NFA_NEST_END(skb, nest_helper);
 180
 181         return 0;
 182
 183 nfattr_failure:
 184         return -1;
 185 }
 186
 187 #ifdef CONFIG_NF_CT_ACCT
 188 static inline int
 189 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 190                         enum ip_conntrack_dir dir)
 191 {
 192         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 193         struct nfattr *nest_count = NFA_NEST(skb, type);
 194         u_int32_t tmp;
 195
 196         tmp = htonl(ct->counters[dir].packets);
 197         NFA_PUT(skb, CTA_COUNTERS32_PACKETS, sizeof(u_int32_t), &tmp);
 198
 199         tmp = htonl(ct->counters[dir].bytes);
 200         NFA_PUT(skb, CTA_COUNTERS32_BYTES, sizeof(u_int32_t), &tmp);
 201
 202         NFA_NEST_END(skb, nest_count);
 203
 204         return 0;
 205
 206 nfattr_failure:
 207         return -1;
 208 }
 209 #else
 210 #define ctnetlink_dump_counters(a, b, c) (0)
 211 #endif
 212
 213 #ifdef CONFIG_NF_CONNTRACK_MARK
 214 static inline int
 215 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 216 {
 217         u_int32_t mark = htonl(ct->mark);
 218
 219         NFA_PUT(skb, CTA_MARK, sizeof(u_int32_t), &mark);
 220         return 0;
 221
 222 nfattr_failure:
 223         return -1;
 224 }
 225 #else
 226 #define ctnetlink_dump_mark(a, b) (0)
 227 #endif
 228
 229 static inline int
 230 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 231 {
 232         u_int32_t id = htonl(ct->id);
 233         NFA_PUT(skb, CTA_ID, sizeof(u_int32_t), &id);
 234         return 0;
 235
 236 nfattr_failure:
 237         return -1;
 238 }
 239
 240 static inline int
 241 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 242 {
 243         u_int32_t use = htonl(atomic_read(&ct->ct_general.use));
 244
 245         NFA_PUT(skb, CTA_USE, sizeof(u_int32_t), &use);
 246         return 0;
 247
 248 nfattr_failure:
 249         return -1;
 250 }
 251
 252 #define tuple(ct, dir) (&(ct)->tuplehash[dir].tuple)
 253
 254 static int
 255 ctnetlink_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
 256                     int event, int nowait,
 257                     const struct nf_conn *ct)
 258 {
 259         struct nlmsghdr *nlh;
 260         struct nfgenmsg *nfmsg;
 261         struct nfattr *nest_parms;
 262         unsigned char *b;
 263
 264         b = skb->tail;
 265
 266         event |= NFNL_SUBSYS_CTNETLINK << 8;
 267         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
 268         nfmsg  = NLMSG_DATA(nlh);
 269
 270         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
 271         nfmsg->nfgen_family =
 272                 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 273         nfmsg->version      = NFNETLINK_V0;
 274         nfmsg->res_id       = 0;
 275
 276         nest_parms = NFA_NEST(skb, CTA_TUPLE_ORIG);
 277         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 278                 goto nfattr_failure;
 279         NFA_NEST_END(skb, nest_parms);
 280
 281         nest_parms = NFA_NEST(skb, CTA_TUPLE_REPLY);
 282         if (ctnetlink_dump_tuples(skb, tuple(ct, IP_CT_DIR_REPLY)) < 0)
 283                 goto nfattr_failure;
 284         NFA_NEST_END(skb, nest_parms);
 285
 286         if (ctnetlink_dump_status(skb, ct) < 0 ||
 287             ctnetlink_dump_timeout(skb, ct) < 0 ||
 288             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 289             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 290             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 291             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 292             ctnetlink_dump_mark(skb, ct) < 0 ||
 293             ctnetlink_dump_id(skb, ct) < 0 ||
 294             ctnetlink_dump_use(skb, ct) < 0)
 295                 goto nfattr_failure;
 296
 297         nlh->nlmsg_len = skb->tail - b;
 298         return skb->len;
 299
 300 nlmsg_failure:
 301 nfattr_failure:
 302         skb_trim(skb, b - skb->data);
 303         return -1;
 304 }
 305
 306 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 307 static int ctnetlink_conntrack_event(struct notifier_block *this,
 308                                      unsigned long events, void *ptr)
 309 {
 310         struct nlmsghdr *nlh;
 311         struct nfgenmsg *nfmsg;
 312         struct nfattr *nest_parms;
 313         struct nf_conn *ct = (struct nf_conn *)ptr;
 314         struct sk_buff *skb;
 315         unsigned int type;
 316         unsigned char *b;
 317         unsigned int flags = 0, group;
 318
 319         /* ignore our fake conntrack entry */
 320         if (ct == &nf_conntrack_untracked)
 321                 return NOTIFY_DONE;
 322
 323         if (events & IPCT_DESTROY) {
 324                 type = IPCTNL_MSG_CT_DELETE;
 325                 group = NFNLGRP_CONNTRACK_DESTROY;
 326         } else  if (events & (IPCT_NEW | IPCT_RELATED)) {
 327                 type = IPCTNL_MSG_CT_NEW;
 328                 flags = NLM_F_CREATE|NLM_F_EXCL;
 329                 /* dump everything */
 330                 events = ~0UL;
 331                 group = NFNLGRP_CONNTRACK_NEW;
 332         } else  if (events & (IPCT_STATUS |
 333                       IPCT_PROTOINFO |
 334                       IPCT_HELPER |
 335                       IPCT_HELPINFO |
 336                       IPCT_NATINFO)) {
 337                 type = IPCTNL_MSG_CT_NEW;
 338                 group = NFNLGRP_CONNTRACK_UPDATE;
 339         } else
 340                 return NOTIFY_DONE;
 341
 342   /* FIXME: Check if there are any listeners before, don't hurt performance */
 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         nlh->nlmsg_len = skb->tail - b;
 388         nfnetlink_send(skb, 0, group, 0);
 389         return NOTIFY_DONE;
 390
 391 nlmsg_failure:
 392 nfattr_failure:
 393         kfree_skb(skb);
 394         return NOTIFY_DONE;
 395 }
 396 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 397
 398 static int ctnetlink_done(struct netlink_callback *cb)
 399 {
 400         DEBUGP("entered %s\n", __FUNCTION__);
 401         return 0;
 402 }
 403
 404 #define L3PROTO(ct) ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num
 405
 406 static int
 407 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 408 {
 409         struct nf_conn *ct = NULL;
 410         struct nf_conntrack_tuple_hash *h;
 411         struct list_head *i;
 412         u_int32_t *id = (u_int32_t *) &cb->args[1];
 413         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
 414         u_int8_t l3proto = nfmsg->nfgen_family;
 415
 416         DEBUGP("entered %s, last bucket=%lu id=%u\n", __FUNCTION__,
 417                         cb->args[0], *id);
 418
 419         read_lock_bh(&nf_conntrack_lock);
 420         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++, *id = 0) {
 421                 list_for_each_prev(i, &nf_conntrack_hash[cb->args[0]]) {
 422                         h = (struct nf_conntrack_tuple_hash *) i;
 423                         if (DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 424                                 continue;
 425                         ct = nf_ct_tuplehash_to_ctrack(h);
 426                         /* Dump entries of a given L3 protocol number.
 427                          * If it is not specified, ie. l3proto == 0,
 428                          * then dump everything. */
 429                         if (l3proto && L3PROTO(ct) != l3proto)
 430                                 continue;
 431                         if (ct->id <= *id)
 432                                 continue;
 433                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 434                                                 cb->nlh->nlmsg_seq,
 435                                                 IPCTNL_MSG_CT_NEW,
 436                                                 1, ct) < 0)
 437                                 goto out;
 438                         *id = ct->id;
 439                 }
 440         }
 441 out:
 442         read_unlock_bh(&nf_conntrack_lock);
 443
 444         DEBUGP("leaving, last bucket=%lu id=%u\n", cb->args[0], *id);
 445
 446         return skb->len;
 447 }
 448
 449 #ifdef CONFIG_NF_CT_ACCT
 450 static int
 451 ctnetlink_dump_table_w(struct sk_buff *skb, struct netlink_callback *cb)
 452 {
 453         struct nf_conn *ct = NULL;
 454         struct nf_conntrack_tuple_hash *h;
 455         struct list_head *i;
 456         u_int32_t *id = (u_int32_t *) &cb->args[1];
 457         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
 458         u_int8_t l3proto = nfmsg->nfgen_family;
 459
 460         DEBUGP("entered %s, last bucket=%u id=%u\n", __FUNCTION__,
 461                         cb->args[0], *id);
 462
 463         write_lock_bh(&nf_conntrack_lock);
 464         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++, *id = 0) {
 465                 list_for_each_prev(i, &nf_conntrack_hash[cb->args[0]]) {
 466                         h = (struct nf_conntrack_tuple_hash *) i;
 467                         if (DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 468                                 continue;
 469                         ct = nf_ct_tuplehash_to_ctrack(h);
 470                         if (l3proto && L3PROTO(ct) != l3proto)
 471                                 continue;
 472                         if (ct->id <= *id)
 473                                 continue;
 474                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 475                                                 cb->nlh->nlmsg_seq,
 476                                                 IPCTNL_MSG_CT_NEW,
 477                                                 1, ct) < 0)
 478                                 goto out;
 479                         *id = ct->id;
 480
 481                         memset(&ct->counters, 0, sizeof(ct->counters));
 482                 }
 483         }
 484 out:
 485         write_unlock_bh(&nf_conntrack_lock);
 486
 487         DEBUGP("leaving, last bucket=%lu id=%u\n", cb->args[0], *id);
 488
 489         return skb->len;
 490 }
 491 #endif
 492
 493 static inline int
 494 ctnetlink_parse_tuple_ip(struct nfattr *attr, struct nf_conntrack_tuple *tuple)
 495 {
 496         struct nfattr *tb[CTA_IP_MAX];
 497         struct nf_conntrack_l3proto *l3proto;
 498         int ret = 0;
 499
 500         DEBUGP("entered %s\n", __FUNCTION__);
 501
 502         nfattr_parse_nested(tb, CTA_IP_MAX, attr);
 503
 504         l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
 505
 506         if (likely(l3proto->nfattr_to_tuple))
 507                 ret = l3proto->nfattr_to_tuple(tb, tuple);
 508
 509         nf_ct_l3proto_put(l3proto);
 510
 511         DEBUGP("leaving\n");
 512
 513         return ret;
 514 }
 515
 516 static const size_t cta_min_proto[CTA_PROTO_MAX] = {
 517         [CTA_PROTO_NUM-1]       = sizeof(u_int8_t),
 518 };
 519
 520 static inline int
 521 ctnetlink_parse_tuple_proto(struct nfattr *attr,
 522                             struct nf_conntrack_tuple *tuple)
 523 {
 524         struct nfattr *tb[CTA_PROTO_MAX];
 525         struct nf_conntrack_protocol *proto;
 526         int ret = 0;
 527
 528         DEBUGP("entered %s\n", __FUNCTION__);
 529
 530         nfattr_parse_nested(tb, CTA_PROTO_MAX, attr);
 531
 532         if (nfattr_bad_size(tb, CTA_PROTO_MAX, cta_min_proto))
 533                 return -EINVAL;
 534
 535         if (!tb[CTA_PROTO_NUM-1])
 536                 return -EINVAL;
 537         tuple->dst.protonum = *(u_int8_t *)NFA_DATA(tb[CTA_PROTO_NUM-1]);
 538
 539         proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 540
 541         if (likely(proto->nfattr_to_tuple))
 542                 ret = proto->nfattr_to_tuple(tb, tuple);
 543
 544         nf_ct_proto_put(proto);
 545
 546         return ret;
 547 }
 548
 549 static inline int
 550 ctnetlink_parse_tuple(struct nfattr *cda[], struct nf_conntrack_tuple *tuple,
 551                       enum ctattr_tuple type, u_int8_t l3num)
 552 {
 553         struct nfattr *tb[CTA_TUPLE_MAX];
 554         int err;
 555
 556         DEBUGP("entered %s\n", __FUNCTION__);
 557
 558         memset(tuple, 0, sizeof(*tuple));
 559
 560         nfattr_parse_nested(tb, CTA_TUPLE_MAX, cda[type-1]);
 561
 562         if (!tb[CTA_TUPLE_IP-1])
 563                 return -EINVAL;
 564
 565         tuple->src.l3num = l3num;
 566
 567         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP-1], tuple);
 568         if (err < 0)
 569                 return err;
 570
 571         if (!tb[CTA_TUPLE_PROTO-1])
 572                 return -EINVAL;
 573
 574         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO-1], tuple);
 575         if (err < 0)
 576                 return err;
 577
 578         /* orig and expect tuples get DIR_ORIGINAL */
 579         if (type == CTA_TUPLE_REPLY)
 580                 tuple->dst.dir = IP_CT_DIR_REPLY;
 581         else
 582                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 583
 584         NF_CT_DUMP_TUPLE(tuple);
 585
 586         DEBUGP("leaving\n");
 587
 588         return 0;
 589 }
 590
 591 #ifdef CONFIG_IP_NF_NAT_NEEDED
 592 static const size_t cta_min_protonat[CTA_PROTONAT_MAX] = {
 593         [CTA_PROTONAT_PORT_MIN-1]       = sizeof(u_int16_t),
 594         [CTA_PROTONAT_PORT_MAX-1]       = sizeof(u_int16_t),
 595 };
 596
 597 static int ctnetlink_parse_nat_proto(struct nfattr *attr,
 598                                      const struct nf_conn *ct,
 599                                      struct ip_nat_range *range)
 600 {
 601         struct nfattr *tb[CTA_PROTONAT_MAX];
 602         struct ip_nat_protocol *npt;
 603
 604         DEBUGP("entered %s\n", __FUNCTION__);
 605
 606         nfattr_parse_nested(tb, CTA_PROTONAT_MAX, attr);
 607
 608         if (nfattr_bad_size(tb, CTA_PROTONAT_MAX, cta_min_protonat))
 609                 return -EINVAL;
 610
 611         npt = ip_nat_proto_find_get(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum);
 612
 613         if (!npt->nfattr_to_range) {
 614                 ip_nat_proto_put(npt);
 615                 return 0;
 616         }
 617
 618         /* nfattr_to_range returns 1 if it parsed, 0 if not, neg. on error */
 619         if (npt->nfattr_to_range(tb, range) > 0)
 620                 range->flags |= IP_NAT_RANGE_PROTO_SPECIFIED;
 621
 622         ip_nat_proto_put(npt);
 623
 624         DEBUGP("leaving\n");
 625         return 0;
 626 }
 627
 628 static const size_t cta_min_nat[CTA_NAT_MAX] = {
 629         [CTA_NAT_MINIP-1]       = sizeof(u_int32_t),
 630         [CTA_NAT_MAXIP-1]       = sizeof(u_int32_t),
 631 };
 632
 633 static inline int
 634 ctnetlink_parse_nat(struct nfattr *cda[],
 635                     const struct nf_conn *ct, struct ip_nat_range *range)
 636 {
 637         struct nfattr *tb[CTA_NAT_MAX];
 638         int err;
 639
 640         DEBUGP("entered %s\n", __FUNCTION__);
 641
 642         memset(range, 0, sizeof(*range));
 643
 644         nfattr_parse_nested(tb, CTA_NAT_MAX, cda[CTA_NAT-1]);
 645
 646         if (nfattr_bad_size(tb, CTA_NAT_MAX, cta_min_nat))
 647                 return -EINVAL;
 648
 649         if (tb[CTA_NAT_MINIP-1])
 650                 range->min_ip = *(u_int32_t *)NFA_DATA(tb[CTA_NAT_MINIP-1]);
 651
 652         if (!tb[CTA_NAT_MAXIP-1])
 653                 range->max_ip = range->min_ip;
 654         else
 655                 range->max_ip = *(u_int32_t *)NFA_DATA(tb[CTA_NAT_MAXIP-1]);
 656
 657         if (range->min_ip)
 658                 range->flags |= IP_NAT_RANGE_MAP_IPS;
 659
 660         if (!tb[CTA_NAT_PROTO-1])
 661                 return 0;
 662
 663         err = ctnetlink_parse_nat_proto(tb[CTA_NAT_PROTO-1], ct, range);
 664         if (err < 0)
 665                 return err;
 666
 667         DEBUGP("leaving\n");
 668         return 0;
 669 }
 670 #endif
 671
 672 static inline int
 673 ctnetlink_parse_help(struct nfattr *attr, char **helper_name)
 674 {
 675         struct nfattr *tb[CTA_HELP_MAX];
 676
 677         DEBUGP("entered %s\n", __FUNCTION__);
 678
 679         nfattr_parse_nested(tb, CTA_HELP_MAX, attr);
 680
 681         if (!tb[CTA_HELP_NAME-1])
 682                 return -EINVAL;
 683
 684         *helper_name = NFA_DATA(tb[CTA_HELP_NAME-1]);
 685
 686         return 0;
 687 }
 688
 689 static const size_t cta_min[CTA_MAX] = {
 690         [CTA_STATUS-1]          = sizeof(u_int32_t),
 691         [CTA_TIMEOUT-1]         = sizeof(u_int32_t),
 692         [CTA_MARK-1]            = sizeof(u_int32_t),
 693         [CTA_USE-1]             = sizeof(u_int32_t),
 694         [CTA_ID-1]              = sizeof(u_int32_t)
 695 };
 696
 697 static int
 698 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 699                         struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
 700 {
 701         struct nf_conntrack_tuple_hash *h;
 702         struct nf_conntrack_tuple tuple;
 703         struct nf_conn *ct;
 704         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 705         u_int8_t u3 = nfmsg->nfgen_family;
 706         int err = 0;
 707
 708         DEBUGP("entered %s\n", __FUNCTION__);
 709
 710         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 711                 return -EINVAL;
 712
 713         if (cda[CTA_TUPLE_ORIG-1])
 714                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 715         else if (cda[CTA_TUPLE_REPLY-1])
 716                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 717         else {
 718                 /* Flush the whole table */
 719                 nf_conntrack_flush();
 720                 return 0;
 721         }
 722
 723         if (err < 0)
 724                 return err;
 725
 726         h = nf_conntrack_find_get(&tuple, NULL);
 727         if (!h) {
 728                 DEBUGP("tuple not found in conntrack hash\n");
 729                 return -ENOENT;
 730         }
 731
 732         ct = nf_ct_tuplehash_to_ctrack(h);
 733
 734         if (cda[CTA_ID-1]) {
 735                 u_int32_t id = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_ID-1]));
 736                 if (ct->id != id) {
 737                         nf_ct_put(ct);
 738                         return -ENOENT;
 739                 }
 740         }
 741         if (del_timer(&ct->timeout))
 742                 ct->timeout.function((unsigned long)ct);
 743
 744         nf_ct_put(ct);
 745         DEBUGP("leaving\n");
 746
 747         return 0;
 748 }
 749
 750 static int
 751 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 752                         struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
 753 {
 754         struct nf_conntrack_tuple_hash *h;
 755         struct nf_conntrack_tuple tuple;
 756         struct nf_conn *ct;
 757         struct sk_buff *skb2 = NULL;
 758         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
 759         u_int8_t u3 = nfmsg->nfgen_family;
 760         int err = 0;
 761
 762         DEBUGP("entered %s\n", __FUNCTION__);
 763
 764         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 765                 u32 rlen;
 766
 767                 if (NFNL_MSG_TYPE(nlh->nlmsg_type) ==
 768                                         IPCTNL_MSG_CT_GET_CTRZERO) {
 769 #ifdef CONFIG_NF_CT_ACCT
 770                         if ((*errp = netlink_dump_start(ctnl, skb, nlh,
 771                                                 ctnetlink_dump_table_w,
 772                                                 ctnetlink_done)) != 0)
 773                                 return -EINVAL;
 774 #else
 775                         return -ENOTSUPP;
 776 #endif
 777                 } else {
 778                         if ((*errp = netlink_dump_start(ctnl, skb, nlh,
 779                                                         ctnetlink_dump_table,
 780                                                         ctnetlink_done)) != 0)
 781                         return -EINVAL;
 782                 }
 783
 784                 rlen = NLMSG_ALIGN(nlh->nlmsg_len);
 785                 if (rlen > skb->len)
 786                         rlen = skb->len;
 787                 skb_pull(skb, rlen);
 788                 return 0;
 789         }
 790
 791         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
 792                 return -EINVAL;
 793
 794         if (cda[CTA_TUPLE_ORIG-1])
 795                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 796         else if (cda[CTA_TUPLE_REPLY-1])
 797                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 798         else
 799                 return -EINVAL;
 800
 801         if (err < 0)
 802                 return err;
 803
 804         h = nf_conntrack_find_get(&tuple, NULL);
 805         if (!h) {
 806                 DEBUGP("tuple not found in conntrack hash");
 807                 return -ENOENT;
 808         }
 809         DEBUGP("tuple found\n");
 810         ct = nf_ct_tuplehash_to_ctrack(h);
 811
 812         err = -ENOMEM;
 813         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 814         if (!skb2) {
 815                 nf_ct_put(ct);
 816                 return -ENOMEM;
 817         }
 818         NETLINK_CB(skb2).dst_pid = NETLINK_CB(skb).pid;
 819
 820         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
 821                                   IPCTNL_MSG_CT_NEW, 1, ct);
 822         nf_ct_put(ct);
 823         if (err <= 0)
 824                 goto free;
 825
 826         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
 827         if (err < 0)
 828                 goto out;
 829
 830         DEBUGP("leaving\n");
 831         return 0;
 832
 833 free:
 834         kfree_skb(skb2);
 835 out:
 836         return err;
 837 }
 838
 839 static inline int
 840 ctnetlink_change_status(struct nf_conn *ct, struct nfattr *cda[])
 841 {
 842         unsigned long d;
 843         unsigned status = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_STATUS-1]));
 844         d = ct->status ^ status;
 845
 846         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
 847                 /* unchangeable */
 848                 return -EINVAL;
 849
 850         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
 851                 /* SEEN_REPLY bit can only be set */
 852                 return -EINVAL;
 853
 854
 855         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
 856                 /* ASSURED bit can only be set */
 857                 return -EINVAL;
 858
 859         if (cda[CTA_NAT-1]) {
 860 #ifndef CONFIG_IP_NF_NAT_NEEDED
 861                 return -EINVAL;
 862 #else
 863                 unsigned int hooknum;
 864                 struct ip_nat_range range;
 865
 866                 if (ctnetlink_parse_nat(cda, ct, &range) < 0)
 867                         return -EINVAL;
 868
 869                 DEBUGP("NAT: %u.%u.%u.%u-%u.%u.%u.%u:%u-%u\n",
 870                        NIPQUAD(range.min_ip), NIPQUAD(range.max_ip),
 871                        htons(range.min.all), htons(range.max.all));
 872
 873                 /* This is tricky but it works. ip_nat_setup_info needs the
 874                  * hook number as parameter, so let's do the correct
 875                  * conversion and run away */
 876                 if (status & IPS_SRC_NAT_DONE)
 877                         hooknum = NF_IP_POST_ROUTING; /* IP_NAT_MANIP_SRC */
 878                 else if (status & IPS_DST_NAT_DONE)
 879                         hooknum = NF_IP_PRE_ROUTING;  /* IP_NAT_MANIP_DST */
 880                 else
 881                         return -EINVAL; /* Missing NAT flags */
 882
 883                 DEBUGP("NAT status: %lu\n",
 884                        status & (IPS_NAT_MASK | IPS_NAT_DONE_MASK));
 885
 886                 if (ip_nat_initialized(ct, HOOK2MANIP(hooknum)))
 887                         return -EEXIST;
 888                 ip_nat_setup_info(ct, &range, hooknum);
 889
 890                 DEBUGP("NAT status after setup_info: %lu\n",
 891                        ct->status & (IPS_NAT_MASK | IPS_NAT_DONE_MASK));
 892 #endif
 893         }
 894
 895         /* Be careful here, modifying NAT bits can screw up things,
 896          * so don't let users modify them directly if they don't pass
 897          * ip_nat_range. */
 898         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
 899         return 0;
 900 }
 901
 902
 903 static inline int
 904 ctnetlink_change_helper(struct nf_conn *ct, struct nfattr *cda[])
 905 {
 906         struct nf_conntrack_helper *helper;
 907         struct nf_conn_help *help = nfct_help(ct);
 908         char *helpname;
 909         int err;
 910
 911         DEBUGP("entered %s\n", __FUNCTION__);
 912
 913         if (!help) {
 914                 /* FIXME: we need to reallocate and rehash */
 915                 return -EBUSY;
 916         }
 917
 918         /* don't change helper of sibling connections */
 919         if (ct->master)
 920                 return -EINVAL;
 921
 922         err = ctnetlink_parse_help(cda[CTA_HELP-1], &helpname);
 923         if (err < 0)
 924                 return err;
 925
 926         helper = __nf_conntrack_helper_find_byname(helpname);
 927         if (!helper) {
 928                 if (!strcmp(helpname, ""))
 929                         helper = NULL;
 930                 else
 931                         return -EINVAL;
 932         }
 933
 934         if (help->helper) {
 935                 if (!helper) {
 936                         /* we had a helper before ... */
 937                         nf_ct_remove_expectations(ct);
 938                         help->helper = NULL;
 939                 } else {
 940                         /* need to zero data of old helper */
 941                         memset(&help->help, 0, sizeof(help->help));
 942                 }
 943         }
 944
 945         help->helper = helper;
 946
 947         return 0;
 948 }
 949
 950 static inline int
 951 ctnetlink_change_timeout(struct nf_conn *ct, struct nfattr *cda[])
 952 {
 953         u_int32_t timeout = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_TIMEOUT-1]));
 954
 955         if (!del_timer(&ct->timeout))
 956                 return -ETIME;
 957
 958         ct->timeout.expires = jiffies + timeout * HZ;
 959         add_timer(&ct->timeout);
 960
 961         return 0;
 962 }
 963
 964 static inline int
 965 ctnetlink_change_protoinfo(struct nf_conn *ct, struct nfattr *cda[])
 966 {
 967         struct nfattr *tb[CTA_PROTOINFO_MAX], *attr = cda[CTA_PROTOINFO-1];
 968         struct nf_conntrack_protocol *proto;
 969         u_int16_t npt = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.protonum;
 970         u_int16_t l3num = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.l3num;
 971         int err = 0;
 972
 973         nfattr_parse_nested(tb, CTA_PROTOINFO_MAX, attr);
 974
 975         proto = nf_ct_proto_find_get(l3num, npt);
 976
 977         if (proto->from_nfattr)
 978                 err = proto->from_nfattr(tb, ct);
 979         nf_ct_proto_put(proto);
 980
 981         return err;
 982 }
 983
 984 static int
 985 ctnetlink_change_conntrack(struct nf_conn *ct, struct nfattr *cda[])
 986 {
 987         int err;
 988
 989         DEBUGP("entered %s\n", __FUNCTION__);
 990
 991         if (cda[CTA_HELP-1]) {
 992                 err = ctnetlink_change_helper(ct, cda);
 993                 if (err < 0)
 994                         return err;
 995         }
 996
 997         if (cda[CTA_TIMEOUT-1]) {
 998                 err = ctnetlink_change_timeout(ct, cda);
 999                 if (err < 0)
1000                         return err;
1001         }
1002
1003         if (cda[CTA_STATUS-1]) {
1004                 err = ctnetlink_change_status(ct, cda);
1005                 if (err < 0)
1006                         return err;
1007         }
1008
1009         if (cda[CTA_PROTOINFO-1]) {
1010                 err = ctnetlink_change_protoinfo(ct, cda);
1011                 if (err < 0)
1012                         return err;
1013         }
1014
1015 #if defined(CONFIG_IP_NF_CONNTRACK_MARK)
1016         if (cda[CTA_MARK-1])
1017                 ct->mark = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_MARK-1]));
1018 #endif
1019
1020         DEBUGP("all done\n");
1021         return 0;
1022 }
1023
1024 static int
1025 ctnetlink_create_conntrack(struct nfattr *cda[],
1026                            struct nf_conntrack_tuple *otuple,
1027                            struct nf_conntrack_tuple *rtuple)
1028 {
1029         struct nf_conn *ct;
1030         int err = -EINVAL;
1031
1032         DEBUGP("entered %s\n", __FUNCTION__);
1033
1034         ct = nf_conntrack_alloc(otuple, rtuple);
1035         if (ct == NULL || IS_ERR(ct))
1036                 return -ENOMEM;
1037
1038         if (!cda[CTA_TIMEOUT-1])
1039                 goto err;
1040         ct->timeout.expires = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_TIMEOUT-1]));
1041
1042         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1043         ct->status |= IPS_CONFIRMED;
1044
1045         err = ctnetlink_change_status(ct, cda);
1046         if (err < 0)
1047                 goto err;
1048
1049         if (cda[CTA_PROTOINFO-1]) {
1050                 err = ctnetlink_change_protoinfo(ct, cda);
1051                 if (err < 0)
1052                         return err;
1053         }
1054
1055 #if defined(CONFIG_IP_NF_CONNTRACK_MARK)
1056         if (cda[CTA_MARK-1])
1057                 ct->mark = ntohl(*(u_int32_t *)NFA_DATA(cda[CTA_MARK-1]));
1058 #endif
1059
1060         add_timer(&ct->timeout);
1061         nf_conntrack_hash_insert(ct);
1062
1063         DEBUGP("conntrack with id %u inserted\n", ct->id);
1064         return 0;
1065
1066 err:
1067         nf_conntrack_free(ct);
1068         return err;
1069 }
1070
1071 static int
1072 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1073                         struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
1074 {
1075         struct nf_conntrack_tuple otuple, rtuple;
1076         struct nf_conntrack_tuple_hash *h = NULL;
1077         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1078         u_int8_t u3 = nfmsg->nfgen_family;
1079         int err = 0;
1080
1081         DEBUGP("entered %s\n", __FUNCTION__);
1082
1083         if (nfattr_bad_size(cda, CTA_MAX, cta_min))
1084                 return -EINVAL;
1085
1086         if (cda[CTA_TUPLE_ORIG-1]) {
1087                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1088                 if (err < 0)
1089                         return err;
1090         }
1091
1092         if (cda[CTA_TUPLE_REPLY-1]) {
1093                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1094                 if (err < 0)
1095                         return err;
1096         }
1097
1098         write_lock_bh(&nf_conntrack_lock);
1099         if (cda[CTA_TUPLE_ORIG-1])
1100                 h = __nf_conntrack_find(&otuple, NULL);
1101         else if (cda[CTA_TUPLE_REPLY-1])
1102                 h = __nf_conntrack_find(&rtuple, NULL);
1103
1104         if (h == NULL) {
1105                 write_unlock_bh(&nf_conntrack_lock);
1106                 DEBUGP("no such conntrack, create new\n");
1107                 err = -ENOENT;
1108                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1109                         err = ctnetlink_create_conntrack(cda, &otuple, &rtuple);
1110                 return err;
1111         }
1112         /* implicit 'else' */
1113
1114         /* we only allow nat config for new conntracks */
1115         if (cda[CTA_NAT-1]) {
1116                 err = -EINVAL;
1117                 goto out_unlock;
1118         }
1119
1120         /* We manipulate the conntrack inside the global conntrack table lock,
1121          * so there's no need to increase the refcount */
1122         DEBUGP("conntrack found\n");
1123         err = -EEXIST;
1124         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
1125                 err = ctnetlink_change_conntrack(nf_ct_tuplehash_to_ctrack(h), cda);
1126
1127 out_unlock:
1128         write_unlock_bh(&nf_conntrack_lock);
1129         return err;
1130 }
1131
1132 /***********************************************************************
1133  * EXPECT
1134  ***********************************************************************/
1135
1136 static inline int
1137 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1138                          const struct nf_conntrack_tuple *tuple,
1139                          enum ctattr_expect type)
1140 {
1141         struct nfattr *nest_parms = NFA_NEST(skb, type);
1142
1143         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1144                 goto nfattr_failure;
1145
1146         NFA_NEST_END(skb, nest_parms);
1147
1148         return 0;
1149
1150 nfattr_failure:
1151         return -1;
1152 }
1153
1154 static inline int
1155 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1156                           const struct nf_conntrack_expect *exp)
1157 {
1158         struct nf_conn *master = exp->master;
1159         u_int32_t timeout = htonl((exp->timeout.expires - jiffies) / HZ);
1160         u_int32_t id = htonl(exp->id);
1161
1162         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1163                 goto nfattr_failure;
1164         if (ctnetlink_exp_dump_tuple(skb, &exp->mask, CTA_EXPECT_MASK) < 0)
1165                 goto nfattr_failure;
1166         if (ctnetlink_exp_dump_tuple(skb,
1167                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1168                                  CTA_EXPECT_MASTER) < 0)
1169                 goto nfattr_failure;
1170
1171         NFA_PUT(skb, CTA_EXPECT_TIMEOUT, sizeof(timeout), &timeout);
1172         NFA_PUT(skb, CTA_EXPECT_ID, sizeof(u_int32_t), &id);
1173
1174         return 0;
1175
1176 nfattr_failure:
1177         return -1;
1178 }
1179
1180 static int
1181 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1182                     int event,
1183                     int nowait,
1184                     const struct nf_conntrack_expect *exp)
1185 {
1186         struct nlmsghdr *nlh;
1187         struct nfgenmsg *nfmsg;
1188         unsigned char *b;
1189
1190         b = skb->tail;
1191
1192         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1193         nlh    = NLMSG_PUT(skb, pid, seq, event, sizeof(struct nfgenmsg));
1194         nfmsg  = NLMSG_DATA(nlh);
1195
1196         nlh->nlmsg_flags    = (nowait && pid) ? NLM_F_MULTI : 0;
1197         nfmsg->nfgen_family = exp->tuple.src.l3num;
1198         nfmsg->version      = NFNETLINK_V0;
1199         nfmsg->res_id       = 0;
1200
1201         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1202                 goto nfattr_failure;
1203
1204         nlh->nlmsg_len = skb->tail - b;
1205         return skb->len;
1206
1207 nlmsg_failure:
1208 nfattr_failure:
1209         skb_trim(skb, b - skb->data);
1210         return -1;
1211 }
1212
1213 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1214 static int ctnetlink_expect_event(struct notifier_block *this,
1215                                   unsigned long events, void *ptr)
1216 {
1217         struct nlmsghdr *nlh;
1218         struct nfgenmsg *nfmsg;
1219         struct nf_conntrack_expect *exp = (struct nf_conntrack_expect *)ptr;
1220         struct sk_buff *skb;
1221         unsigned int type;
1222         unsigned char *b;
1223         int flags = 0;
1224
1225         if (events & IPEXP_NEW) {
1226                 type = IPCTNL_MSG_EXP_NEW;
1227                 flags = NLM_F_CREATE|NLM_F_EXCL;
1228         } else
1229                 return NOTIFY_DONE;
1230
1231         skb = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1232         if (!skb)
1233                 return NOTIFY_DONE;
1234
1235         b = skb->tail;
1236
1237         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1238         nlh   = NLMSG_PUT(skb, 0, 0, type, sizeof(struct nfgenmsg));
1239         nfmsg = NLMSG_DATA(nlh);
1240
1241         nlh->nlmsg_flags    = flags;
1242         nfmsg->nfgen_family = exp->tuple.src.l3num;
1243         nfmsg->version      = NFNETLINK_V0;
1244         nfmsg->res_id       = 0;
1245
1246         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1247                 goto nfattr_failure;
1248
1249         nlh->nlmsg_len = skb->tail - b;
1250         nfnetlink_send(skb, 0, NFNLGRP_CONNTRACK_EXP_NEW, 0);
1251         return NOTIFY_DONE;
1252
1253 nlmsg_failure:
1254 nfattr_failure:
1255         kfree_skb(skb);
1256         return NOTIFY_DONE;
1257 }
1258 #endif
1259
1260 static int
1261 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1262 {
1263         struct nf_conntrack_expect *exp = NULL;
1264         struct list_head *i;
1265         u_int32_t *id = (u_int32_t *) &cb->args[0];
1266         struct nfgenmsg *nfmsg = NLMSG_DATA(cb->nlh);
1267         u_int8_t l3proto = nfmsg->nfgen_family;
1268
1269         DEBUGP("entered %s, last id=%llu\n", __FUNCTION__, *id);
1270
1271         read_lock_bh(&nf_conntrack_lock);
1272         list_for_each_prev(i, &nf_conntrack_expect_list) {
1273                 exp = (struct nf_conntrack_expect *) i;
1274                 if (l3proto && exp->tuple.src.l3num != l3proto)
1275                         continue;
1276                 if (exp->id <= *id)
1277                         continue;
1278                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).pid,
1279                                             cb->nlh->nlmsg_seq,
1280                                             IPCTNL_MSG_EXP_NEW,
1281                                             1, exp) < 0)
1282                         goto out;
1283                 *id = exp->id;
1284         }
1285 out:
1286         read_unlock_bh(&nf_conntrack_lock);
1287
1288         DEBUGP("leaving, last id=%llu\n", *id);
1289
1290         return skb->len;
1291 }
1292
1293 static const size_t cta_min_exp[CTA_EXPECT_MAX] = {
1294         [CTA_EXPECT_TIMEOUT-1]          = sizeof(u_int32_t),
1295         [CTA_EXPECT_ID-1]               = sizeof(u_int32_t)
1296 };
1297
1298 static int
1299 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1300                      struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
1301 {
1302         struct nf_conntrack_tuple tuple;
1303         struct nf_conntrack_expect *exp;
1304         struct sk_buff *skb2;
1305         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1306         u_int8_t u3 = nfmsg->nfgen_family;
1307         int err = 0;
1308
1309         DEBUGP("entered %s\n", __FUNCTION__);
1310
1311         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1312                 return -EINVAL;
1313
1314         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1315                 u32 rlen;
1316
1317                 if ((*errp = netlink_dump_start(ctnl, skb, nlh,
1318                                                 ctnetlink_exp_dump_table,
1319                                                 ctnetlink_done)) != 0)
1320                         return -EINVAL;
1321                 rlen = NLMSG_ALIGN(nlh->nlmsg_len);
1322                 if (rlen > skb->len)
1323                         rlen = skb->len;
1324                 skb_pull(skb, rlen);
1325                 return 0;
1326         }
1327
1328         if (cda[CTA_EXPECT_MASTER-1])
1329                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1330         else
1331                 return -EINVAL;
1332
1333         if (err < 0)
1334                 return err;
1335
1336         exp = nf_conntrack_expect_find(&tuple);
1337         if (!exp)
1338                 return -ENOENT;
1339
1340         if (cda[CTA_EXPECT_ID-1]) {
1341                 u_int32_t id = *(u_int32_t *)NFA_DATA(cda[CTA_EXPECT_ID-1]);
1342                 if (exp->id != ntohl(id)) {
1343                         nf_conntrack_expect_put(exp);
1344                         return -ENOENT;
1345                 }
1346         }
1347
1348         err = -ENOMEM;
1349         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
1350         if (!skb2)
1351                 goto out;
1352         NETLINK_CB(skb2).dst_pid = NETLINK_CB(skb).pid;
1353
1354         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1355                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
1356                                       1, exp);
1357         if (err <= 0)
1358                 goto free;
1359
1360         nf_conntrack_expect_put(exp);
1361
1362         return netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1363
1364 free:
1365         kfree_skb(skb2);
1366 out:
1367         nf_conntrack_expect_put(exp);
1368         return err;
1369 }
1370
1371 static int
1372 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1373                      struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
1374 {
1375         struct nf_conntrack_expect *exp, *tmp;
1376         struct nf_conntrack_tuple tuple;
1377         struct nf_conntrack_helper *h;
1378         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1379         u_int8_t u3 = nfmsg->nfgen_family;
1380         int err;
1381
1382         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1383                 return -EINVAL;
1384
1385         if (cda[CTA_EXPECT_TUPLE-1]) {
1386                 /* delete a single expect by tuple */
1387                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1388                 if (err < 0)
1389                         return err;
1390
1391                 /* bump usage count to 2 */
1392                 exp = nf_conntrack_expect_find(&tuple);
1393                 if (!exp)
1394                         return -ENOENT;
1395
1396                 if (cda[CTA_EXPECT_ID-1]) {
1397                         u_int32_t id =
1398                                 *(u_int32_t *)NFA_DATA(cda[CTA_EXPECT_ID-1]);
1399                         if (exp->id != ntohl(id)) {
1400                                 nf_conntrack_expect_put(exp);
1401                                 return -ENOENT;
1402                         }
1403                 }
1404
1405                 /* after list removal, usage count == 1 */
1406                 nf_conntrack_unexpect_related(exp);
1407                 /* have to put what we 'get' above.
1408                  * after this line usage count == 0 */
1409                 nf_conntrack_expect_put(exp);
1410         } else if (cda[CTA_EXPECT_HELP_NAME-1]) {
1411                 char *name = NFA_DATA(cda[CTA_EXPECT_HELP_NAME-1]);
1412
1413                 /* delete all expectations for this helper */
1414                 write_lock_bh(&nf_conntrack_lock);
1415                 h = __nf_conntrack_helper_find_byname(name);
1416                 if (!h) {
1417                         write_unlock_bh(&nf_conntrack_lock);
1418                         return -EINVAL;
1419                 }
1420                 list_for_each_entry_safe(exp, tmp, &nf_conntrack_expect_list,
1421                                          list) {
1422                         struct nf_conn_help *m_help = nfct_help(exp->master);
1423                         if (m_help->helper == h
1424                             && del_timer(&exp->timeout)) {
1425                                 nf_ct_unlink_expect(exp);
1426                                 nf_conntrack_expect_put(exp);
1427                         }
1428                 }
1429                 write_unlock_bh(&nf_conntrack_lock);
1430         } else {
1431                 /* This basically means we have to flush everything*/
1432                 write_lock_bh(&nf_conntrack_lock);
1433                 list_for_each_entry_safe(exp, tmp, &nf_conntrack_expect_list,
1434                                          list) {
1435                         if (del_timer(&exp->timeout)) {
1436                                 nf_ct_unlink_expect(exp);
1437                                 nf_conntrack_expect_put(exp);
1438                         }
1439                 }
1440                 write_unlock_bh(&nf_conntrack_lock);
1441         }
1442
1443         return 0;
1444 }
1445 static int
1446 ctnetlink_change_expect(struct nf_conntrack_expect *x, struct nfattr *cda[])
1447 {
1448         return -EOPNOTSUPP;
1449 }
1450
1451 static int
1452 ctnetlink_create_expect(struct nfattr *cda[], u_int8_t u3)
1453 {
1454         struct nf_conntrack_tuple tuple, mask, master_tuple;
1455         struct nf_conntrack_tuple_hash *h = NULL;
1456         struct nf_conntrack_expect *exp;
1457         struct nf_conn *ct;
1458         struct nf_conn_help *help;
1459         int err = 0;
1460
1461         DEBUGP("entered %s\n", __FUNCTION__);
1462
1463         /* caller guarantees that those three CTA_EXPECT_* exist */
1464         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1465         if (err < 0)
1466                 return err;
1467         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
1468         if (err < 0)
1469                 return err;
1470         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
1471         if (err < 0)
1472                 return err;
1473
1474         /* Look for master conntrack of this expectation */
1475         h = nf_conntrack_find_get(&master_tuple, NULL);
1476         if (!h)
1477                 return -ENOENT;
1478         ct = nf_ct_tuplehash_to_ctrack(h);
1479         help = nfct_help(ct);
1480
1481         if (!help || !help->helper) {
1482                 /* such conntrack hasn't got any helper, abort */
1483                 err = -EINVAL;
1484                 goto out;
1485         }
1486
1487         exp = nf_conntrack_expect_alloc(ct);
1488         if (!exp) {
1489                 err = -ENOMEM;
1490                 goto out;
1491         }
1492
1493         exp->expectfn = NULL;
1494         exp->flags = 0;
1495         exp->master = ct;
1496         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
1497         memcpy(&exp->mask, &mask, sizeof(struct nf_conntrack_tuple));
1498
1499         err = nf_conntrack_expect_related(exp);
1500         nf_conntrack_expect_put(exp);
1501
1502 out:
1503         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
1504         return err;
1505 }
1506
1507 static int
1508 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
1509                      struct nlmsghdr *nlh, struct nfattr *cda[], int *errp)
1510 {
1511         struct nf_conntrack_tuple tuple;
1512         struct nf_conntrack_expect *exp;
1513         struct nfgenmsg *nfmsg = NLMSG_DATA(nlh);
1514         u_int8_t u3 = nfmsg->nfgen_family;
1515         int err = 0;
1516
1517         DEBUGP("entered %s\n", __FUNCTION__);
1518
1519         if (nfattr_bad_size(cda, CTA_EXPECT_MAX, cta_min_exp))
1520                 return -EINVAL;
1521
1522         if (!cda[CTA_EXPECT_TUPLE-1]
1523             || !cda[CTA_EXPECT_MASK-1]
1524             || !cda[CTA_EXPECT_MASTER-1])
1525                 return -EINVAL;
1526
1527         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1528         if (err < 0)
1529                 return err;
1530
1531         write_lock_bh(&nf_conntrack_lock);
1532         exp = __nf_conntrack_expect_find(&tuple);
1533
1534         if (!exp) {
1535                 write_unlock_bh(&nf_conntrack_lock);
1536                 err = -ENOENT;
1537                 if (nlh->nlmsg_flags & NLM_F_CREATE)
1538                         err = ctnetlink_create_expect(cda, u3);
1539                 return err;
1540         }
1541
1542         err = -EEXIST;
1543         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
1544                 err = ctnetlink_change_expect(exp, cda);
1545         write_unlock_bh(&nf_conntrack_lock);
1546
1547         DEBUGP("leaving\n");
1548
1549         return err;
1550 }
1551
1552 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1553 static struct notifier_block ctnl_notifier = {
1554         .notifier_call  = ctnetlink_conntrack_event,
1555 };
1556
1557 static struct notifier_block ctnl_notifier_exp = {
1558         .notifier_call  = ctnetlink_expect_event,
1559 };
1560 #endif
1561
1562 static struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
1563         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
1564                                             .attr_count = CTA_MAX, },
1565         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
1566                                             .attr_count = CTA_MAX, },
1567         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
1568                                             .attr_count = CTA_MAX, },
1569         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
1570                                             .attr_count = CTA_MAX, },
1571 };
1572
1573 static struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
1574         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
1575                                             .attr_count = CTA_EXPECT_MAX, },
1576         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
1577                                             .attr_count = CTA_EXPECT_MAX, },
1578         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
1579                                             .attr_count = CTA_EXPECT_MAX, },
1580 };
1581
1582 static struct nfnetlink_subsystem ctnl_subsys = {
1583         .name                           = "conntrack",
1584         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
1585         .cb_count                       = IPCTNL_MSG_MAX,
1586         .cb                             = ctnl_cb,
1587 };
1588
1589 static struct nfnetlink_subsystem ctnl_exp_subsys = {
1590         .name                           = "conntrack_expect",
1591         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
1592         .cb_count                       = IPCTNL_MSG_EXP_MAX,
1593         .cb                             = ctnl_exp_cb,
1594 };
1595
1596 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
1597 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
1598
1599 static int __init ctnetlink_init(void)
1600 {
1601         int ret;
1602
1603         printk("ctnetlink v%s: registering with nfnetlink.\n", version);
1604         ret = nfnetlink_subsys_register(&ctnl_subsys);
1605         if (ret < 0) {
1606                 printk("ctnetlink_init: cannot register with nfnetlink.\n");
1607                 goto err_out;
1608         }
1609
1610         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
1611         if (ret < 0) {
1612                 printk("ctnetlink_init: cannot register exp with nfnetlink.\n");
1613                 goto err_unreg_subsys;
1614         }
1615
1616 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1617         ret = nf_conntrack_register_notifier(&ctnl_notifier);
1618         if (ret < 0) {
1619                 printk("ctnetlink_init: cannot register notifier.\n");
1620                 goto err_unreg_exp_subsys;
1621         }
1622
1623         ret = nf_conntrack_expect_register_notifier(&ctnl_notifier_exp);
1624         if (ret < 0) {
1625                 printk("ctnetlink_init: cannot expect register notifier.\n");
1626                 goto err_unreg_notifier;
1627         }
1628 #endif
1629
1630         return 0;
1631
1632 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1633 err_unreg_notifier:
1634         nf_conntrack_unregister_notifier(&ctnl_notifier);
1635 err_unreg_exp_subsys:
1636         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1637 #endif
1638 err_unreg_subsys:
1639         nfnetlink_subsys_unregister(&ctnl_subsys);
1640 err_out:
1641         return ret;
1642 }
1643
1644 static void __exit ctnetlink_exit(void)
1645 {
1646         printk("ctnetlink: unregistering from nfnetlink.\n");
1647
1648 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1649         nf_conntrack_unregister_notifier(&ctnl_notifier_exp);
1650         nf_conntrack_unregister_notifier(&ctnl_notifier);
1651 #endif
1652
1653         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
1654         nfnetlink_subsys_unregister(&ctnl_subsys);
1655         return;
1656 }
1657
1658 module_init(ctnetlink_init);
1659 module_exit(ctnetlink_exit);