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) 2005-2011 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_l3proto.h>
  41 #include <net/netfilter/nf_conntrack_l4proto.h>
  42 #include <net/netfilter/nf_conntrack_tuple.h>
  43 #include <net/netfilter/nf_conntrack_acct.h>
  44 #include <net/netfilter/nf_conntrack_zones.h>
  45 #include <net/netfilter/nf_conntrack_timestamp.h>
  46 #ifdef CONFIG_NF_NAT_NEEDED
  47 #include <net/netfilter/nf_nat_core.h>
  48 #include <net/netfilter/nf_nat_protocol.h>
  49 #endif
  50
  51 #include <linux/netfilter/nfnetlink.h>
  52 #include <linux/netfilter/nfnetlink_conntrack.h>
  53
  54 MODULE_LICENSE("GPL");
  55
  56 static char __initdata version[] = "0.93";
  57
  58 static inline int
  59 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  60                             const struct nf_conntrack_tuple *tuple,
  61                             struct nf_conntrack_l4proto *l4proto)
  62 {
  63         int ret = 0;
  64         struct nlattr *nest_parms;
  65
  66         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  67         if (!nest_parms)
  68                 goto nla_put_failure;
  69         NLA_PUT_U8(skb, CTA_PROTO_NUM, tuple->dst.protonum);
  70
  71         if (likely(l4proto->tuple_to_nlattr))
  72                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  73
  74         nla_nest_end(skb, nest_parms);
  75
  76         return ret;
  77
  78 nla_put_failure:
  79         return -1;
  80 }
  81
  82 static inline int
  83 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  84                          const struct nf_conntrack_tuple *tuple,
  85                          struct nf_conntrack_l3proto *l3proto)
  86 {
  87         int ret = 0;
  88         struct nlattr *nest_parms;
  89
  90         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  91         if (!nest_parms)
  92                 goto nla_put_failure;
  93
  94         if (likely(l3proto->tuple_to_nlattr))
  95                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  96
  97         nla_nest_end(skb, nest_parms);
  98
  99         return ret;
 100
 101 nla_put_failure:
 102         return -1;
 103 }
 104
 105 static int
 106 ctnetlink_dump_tuples(struct sk_buff *skb,
 107                       const struct nf_conntrack_tuple *tuple)
 108 {
 109         int ret;
 110         struct nf_conntrack_l3proto *l3proto;
 111         struct nf_conntrack_l4proto *l4proto;
 112
 113         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 114         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 115
 116         if (unlikely(ret < 0))
 117                 return ret;
 118
 119         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 120         ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 121
 122         return ret;
 123 }
 124
 125 static inline int
 126 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 127 {
 128         NLA_PUT_BE32(skb, CTA_STATUS, htonl(ct->status));
 129         return 0;
 130
 131 nla_put_failure:
 132         return -1;
 133 }
 134
 135 static inline int
 136 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 137 {
 138         long timeout = (ct->timeout.expires - jiffies) / HZ;
 139
 140         if (timeout < 0)
 141                 timeout = 0;
 142
 143         NLA_PUT_BE32(skb, CTA_TIMEOUT, htonl(timeout));
 144         return 0;
 145
 146 nla_put_failure:
 147         return -1;
 148 }
 149
 150 static inline int
 151 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 152 {
 153         struct nf_conntrack_l4proto *l4proto;
 154         struct nlattr *nest_proto;
 155         int ret;
 156
 157         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 158         if (!l4proto->to_nlattr)
 159                 return 0;
 160
 161         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 162         if (!nest_proto)
 163                 goto nla_put_failure;
 164
 165         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 166
 167         nla_nest_end(skb, nest_proto);
 168
 169         return ret;
 170
 171 nla_put_failure:
 172         return -1;
 173 }
 174
 175 static inline int
 176 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 177 {
 178         struct nlattr *nest_helper;
 179         const struct nf_conn_help *help = nfct_help(ct);
 180         struct nf_conntrack_helper *helper;
 181
 182         if (!help)
 183                 return 0;
 184
 185         helper = rcu_dereference(help->helper);
 186         if (!helper)
 187                 goto out;
 188
 189         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 190         if (!nest_helper)
 191                 goto nla_put_failure;
 192         NLA_PUT_STRING(skb, CTA_HELP_NAME, helper->name);
 193
 194         if (helper->to_nlattr)
 195                 helper->to_nlattr(skb, ct);
 196
 197         nla_nest_end(skb, nest_helper);
 198 out:
 199         return 0;
 200
 201 nla_put_failure:
 202         return -1;
 203 }
 204
 205 static int
 206 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 207                         enum ip_conntrack_dir dir)
 208 {
 209         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 210         struct nlattr *nest_count;
 211         const struct nf_conn_counter *acct;
 212
 213         acct = nf_conn_acct_find(ct);
 214         if (!acct)
 215                 return 0;
 216
 217         nest_count = nla_nest_start(skb, type | NLA_F_NESTED);
 218         if (!nest_count)
 219                 goto nla_put_failure;
 220
 221         NLA_PUT_BE64(skb, CTA_COUNTERS_PACKETS,
 222                      cpu_to_be64(atomic64_read(&acct[dir].packets)));
 223         NLA_PUT_BE64(skb, CTA_COUNTERS_BYTES,
 224                      cpu_to_be64(atomic64_read(&acct[dir].bytes)));
 225
 226         nla_nest_end(skb, nest_count);
 227
 228         return 0;
 229
 230 nla_put_failure:
 231         return -1;
 232 }
 233
 234 static int
 235 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 236 {
 237         struct nlattr *nest_count;
 238         const struct nf_conn_tstamp *tstamp;
 239
 240         tstamp = nf_conn_tstamp_find(ct);
 241         if (!tstamp)
 242                 return 0;
 243
 244         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 245         if (!nest_count)
 246                 goto nla_put_failure;
 247
 248         NLA_PUT_BE64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start));
 249         if (tstamp->stop != 0) {
 250                 NLA_PUT_BE64(skb, CTA_TIMESTAMP_STOP,
 251                              cpu_to_be64(tstamp->stop));
 252         }
 253         nla_nest_end(skb, nest_count);
 254
 255         return 0;
 256
 257 nla_put_failure:
 258         return -1;
 259 }
 260
 261 #ifdef CONFIG_NF_CONNTRACK_MARK
 262 static inline int
 263 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 264 {
 265         NLA_PUT_BE32(skb, CTA_MARK, htonl(ct->mark));
 266         return 0;
 267
 268 nla_put_failure:
 269         return -1;
 270 }
 271 #else
 272 #define ctnetlink_dump_mark(a, b) (0)
 273 #endif
 274
 275 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 276 static inline int
 277 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 278 {
 279         struct nlattr *nest_secctx;
 280         int len, ret;
 281         char *secctx;
 282
 283         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 284         if (ret)
 285                 return 0;
 286
 287         ret = -1;
 288         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 289         if (!nest_secctx)
 290                 goto nla_put_failure;
 291
 292         NLA_PUT_STRING(skb, CTA_SECCTX_NAME, secctx);
 293         nla_nest_end(skb, nest_secctx);
 294
 295         ret = 0;
 296 nla_put_failure:
 297         security_release_secctx(secctx, len);
 298         return ret;
 299 }
 300 #else
 301 #define ctnetlink_dump_secctx(a, b) (0)
 302 #endif
 303
 304 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 305
 306 static inline int
 307 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 308 {
 309         struct nlattr *nest_parms;
 310
 311         if (!(ct->status & IPS_EXPECTED))
 312                 return 0;
 313
 314         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 315         if (!nest_parms)
 316                 goto nla_put_failure;
 317         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 318                 goto nla_put_failure;
 319         nla_nest_end(skb, nest_parms);
 320
 321         return 0;
 322
 323 nla_put_failure:
 324         return -1;
 325 }
 326
 327 #ifdef CONFIG_NF_NAT_NEEDED
 328 static int
 329 dump_nat_seq_adj(struct sk_buff *skb, const struct nf_nat_seq *natseq, int type)
 330 {
 331         struct nlattr *nest_parms;
 332
 333         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 334         if (!nest_parms)
 335                 goto nla_put_failure;
 336
 337         NLA_PUT_BE32(skb, CTA_NAT_SEQ_CORRECTION_POS,
 338                      htonl(natseq->correction_pos));
 339         NLA_PUT_BE32(skb, CTA_NAT_SEQ_OFFSET_BEFORE,
 340                      htonl(natseq->offset_before));
 341         NLA_PUT_BE32(skb, CTA_NAT_SEQ_OFFSET_AFTER,
 342                      htonl(natseq->offset_after));
 343
 344         nla_nest_end(skb, nest_parms);
 345
 346         return 0;
 347
 348 nla_put_failure:
 349         return -1;
 350 }
 351
 352 static inline int
 353 ctnetlink_dump_nat_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 354 {
 355         struct nf_nat_seq *natseq;
 356         struct nf_conn_nat *nat = nfct_nat(ct);
 357
 358         if (!(ct->status & IPS_SEQ_ADJUST) || !nat)
 359                 return 0;
 360
 361         natseq = &nat->seq[IP_CT_DIR_ORIGINAL];
 362         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_ORIG) == -1)
 363                 return -1;
 364
 365         natseq = &nat->seq[IP_CT_DIR_REPLY];
 366         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_REPLY) == -1)
 367                 return -1;
 368
 369         return 0;
 370 }
 371 #else
 372 #define ctnetlink_dump_nat_seq_adj(a, b) (0)
 373 #endif
 374
 375 static inline int
 376 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 377 {
 378         NLA_PUT_BE32(skb, CTA_ID, htonl((unsigned long)ct));
 379         return 0;
 380
 381 nla_put_failure:
 382         return -1;
 383 }
 384
 385 static inline int
 386 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 387 {
 388         NLA_PUT_BE32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use)));
 389         return 0;
 390
 391 nla_put_failure:
 392         return -1;
 393 }
 394
 395 static int
 396 ctnetlink_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
 397                     int event, struct nf_conn *ct)
 398 {
 399         struct nlmsghdr *nlh;
 400         struct nfgenmsg *nfmsg;
 401         struct nlattr *nest_parms;
 402         unsigned int flags = pid ? NLM_F_MULTI : 0;
 403
 404         event |= NFNL_SUBSYS_CTNETLINK << 8;
 405         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
 406         if (nlh == NULL)
 407                 goto nlmsg_failure;
 408
 409         nfmsg = nlmsg_data(nlh);
 410         nfmsg->nfgen_family = nf_ct_l3num(ct);
 411         nfmsg->version      = NFNETLINK_V0;
 412         nfmsg->res_id       = 0;
 413
 414         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 415         if (!nest_parms)
 416                 goto nla_put_failure;
 417         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 418                 goto nla_put_failure;
 419         nla_nest_end(skb, nest_parms);
 420
 421         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 422         if (!nest_parms)
 423                 goto nla_put_failure;
 424         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 425                 goto nla_put_failure;
 426         nla_nest_end(skb, nest_parms);
 427
 428         if (nf_ct_zone(ct))
 429                 NLA_PUT_BE16(skb, CTA_ZONE, htons(nf_ct_zone(ct)));
 430
 431         if (ctnetlink_dump_status(skb, ct) < 0 ||
 432             ctnetlink_dump_timeout(skb, ct) < 0 ||
 433             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 434             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 435             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 436             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 437             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 438             ctnetlink_dump_mark(skb, ct) < 0 ||
 439             ctnetlink_dump_secctx(skb, ct) < 0 ||
 440             ctnetlink_dump_id(skb, ct) < 0 ||
 441             ctnetlink_dump_use(skb, ct) < 0 ||
 442             ctnetlink_dump_master(skb, ct) < 0 ||
 443             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 444                 goto nla_put_failure;
 445
 446         nlmsg_end(skb, nlh);
 447         return skb->len;
 448
 449 nlmsg_failure:
 450 nla_put_failure:
 451         nlmsg_cancel(skb, nlh);
 452         return -1;
 453 }
 454
 455 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 456 static inline size_t
 457 ctnetlink_proto_size(const struct nf_conn *ct)
 458 {
 459         struct nf_conntrack_l3proto *l3proto;
 460         struct nf_conntrack_l4proto *l4proto;
 461         size_t len = 0;
 462
 463         rcu_read_lock();
 464         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 465         len += l3proto->nla_size;
 466
 467         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 468         len += l4proto->nla_size;
 469         rcu_read_unlock();
 470
 471         return len;
 472 }
 473
 474 static inline size_t
 475 ctnetlink_counters_size(const struct nf_conn *ct)
 476 {
 477         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 478                 return 0;
 479         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 480                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 481                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 482                ;
 483 }
 484
 485 static inline int
 486 ctnetlink_secctx_size(const struct nf_conn *ct)
 487 {
 488 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 489         int len, ret;
 490
 491         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 492         if (ret)
 493                 return 0;
 494
 495         return nla_total_size(0) /* CTA_SECCTX */
 496                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 497 #else
 498         return 0;
 499 #endif
 500 }
 501
 502 static inline size_t
 503 ctnetlink_timestamp_size(const struct nf_conn *ct)
 504 {
 505 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 506         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 507                 return 0;
 508         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 509 #else
 510         return 0;
 511 #endif
 512 }
 513
 514 static inline size_t
 515 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 516 {
 517         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 518                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 519                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 520                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 521                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 522                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 523                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 524                + ctnetlink_counters_size(ct)
 525                + ctnetlink_timestamp_size(ct)
 526                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 527                + nla_total_size(0) /* CTA_PROTOINFO */
 528                + nla_total_size(0) /* CTA_HELP */
 529                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 530                + ctnetlink_secctx_size(ct)
 531 #ifdef CONFIG_NF_NAT_NEEDED
 532                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 533                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 534 #endif
 535 #ifdef CONFIG_NF_CONNTRACK_MARK
 536                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 537 #endif
 538                + ctnetlink_proto_size(ct)
 539                ;
 540 }
 541
 542 static int
 543 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 544 {
 545         struct net *net;
 546         struct nlmsghdr *nlh;
 547         struct nfgenmsg *nfmsg;
 548         struct nlattr *nest_parms;
 549         struct nf_conn *ct = item->ct;
 550         struct sk_buff *skb;
 551         unsigned int type;
 552         unsigned int flags = 0, group;
 553         int err;
 554
 555         /* ignore our fake conntrack entry */
 556         if (nf_ct_is_untracked(ct))
 557                 return 0;
 558
 559         if (events & (1 << IPCT_DESTROY)) {
 560                 type = IPCTNL_MSG_CT_DELETE;
 561                 group = NFNLGRP_CONNTRACK_DESTROY;
 562         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 563                 type = IPCTNL_MSG_CT_NEW;
 564                 flags = NLM_F_CREATE|NLM_F_EXCL;
 565                 group = NFNLGRP_CONNTRACK_NEW;
 566         } else  if (events) {
 567                 type = IPCTNL_MSG_CT_NEW;
 568                 group = NFNLGRP_CONNTRACK_UPDATE;
 569         } else
 570                 return 0;
 571
 572         net = nf_ct_net(ct);
 573         if (!item->report && !nfnetlink_has_listeners(net, group))
 574                 return 0;
 575
 576         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 577         if (skb == NULL)
 578                 goto errout;
 579
 580         type |= NFNL_SUBSYS_CTNETLINK << 8;
 581         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
 582         if (nlh == NULL)
 583                 goto nlmsg_failure;
 584
 585         nfmsg = nlmsg_data(nlh);
 586         nfmsg->nfgen_family = nf_ct_l3num(ct);
 587         nfmsg->version  = NFNETLINK_V0;
 588         nfmsg->res_id   = 0;
 589
 590         rcu_read_lock();
 591         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 592         if (!nest_parms)
 593                 goto nla_put_failure;
 594         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 595                 goto nla_put_failure;
 596         nla_nest_end(skb, nest_parms);
 597
 598         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 599         if (!nest_parms)
 600                 goto nla_put_failure;
 601         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 602                 goto nla_put_failure;
 603         nla_nest_end(skb, nest_parms);
 604
 605         if (nf_ct_zone(ct))
 606                 NLA_PUT_BE16(skb, CTA_ZONE, htons(nf_ct_zone(ct)));
 607
 608         if (ctnetlink_dump_id(skb, ct) < 0)
 609                 goto nla_put_failure;
 610
 611         if (ctnetlink_dump_status(skb, ct) < 0)
 612                 goto nla_put_failure;
 613
 614         if (events & (1 << IPCT_DESTROY)) {
 615                 if (ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL) < 0 ||
 616                     ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY) < 0 ||
 617                     ctnetlink_dump_timestamp(skb, ct) < 0)
 618                         goto nla_put_failure;
 619         } else {
 620                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 621                         goto nla_put_failure;
 622
 623                 if (events & (1 << IPCT_PROTOINFO)
 624                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 625                         goto nla_put_failure;
 626
 627                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 628                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 629                         goto nla_put_failure;
 630
 631 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 632                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 633                     && ctnetlink_dump_secctx(skb, ct) < 0)
 634                         goto nla_put_failure;
 635 #endif
 636
 637                 if (events & (1 << IPCT_RELATED) &&
 638                     ctnetlink_dump_master(skb, ct) < 0)
 639                         goto nla_put_failure;
 640
 641                 if (events & (1 << IPCT_NATSEQADJ) &&
 642                     ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 643                         goto nla_put_failure;
 644         }
 645
 646 #ifdef CONFIG_NF_CONNTRACK_MARK
 647         if ((events & (1 << IPCT_MARK) || ct->mark)
 648             && ctnetlink_dump_mark(skb, ct) < 0)
 649                 goto nla_put_failure;
 650 #endif
 651         rcu_read_unlock();
 652
 653         nlmsg_end(skb, nlh);
 654         err = nfnetlink_send(skb, net, item->pid, group, item->report,
 655                              GFP_ATOMIC);
 656         if (err == -ENOBUFS || err == -EAGAIN)
 657                 return -ENOBUFS;
 658
 659         return 0;
 660
 661 nla_put_failure:
 662         rcu_read_unlock();
 663         nlmsg_cancel(skb, nlh);
 664 nlmsg_failure:
 665         kfree_skb(skb);
 666 errout:
 667         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 668                 return -ENOBUFS;
 669
 670         return 0;
 671 }
 672 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 673
 674 static int ctnetlink_done(struct netlink_callback *cb)
 675 {
 676         if (cb->args[1])
 677                 nf_ct_put((struct nf_conn *)cb->args[1]);
 678         return 0;
 679 }
 680
 681 static int
 682 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 683 {
 684         struct net *net = sock_net(skb->sk);
 685         struct nf_conn *ct, *last;
 686         struct nf_conntrack_tuple_hash *h;
 687         struct hlist_nulls_node *n;
 688         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 689         u_int8_t l3proto = nfmsg->nfgen_family;
 690
 691         spin_lock_bh(&nf_conntrack_lock);
 692         last = (struct nf_conn *)cb->args[1];
 693         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 694 restart:
 695                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 696                                          hnnode) {
 697                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 698                                 continue;
 699                         ct = nf_ct_tuplehash_to_ctrack(h);
 700                         /* Dump entries of a given L3 protocol number.
 701                          * If it is not specified, ie. l3proto == 0,
 702                          * then dump everything. */
 703                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 704                                 continue;
 705                         if (cb->args[1]) {
 706                                 if (ct != last)
 707                                         continue;
 708                                 cb->args[1] = 0;
 709                         }
 710                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 711                                                 cb->nlh->nlmsg_seq,
 712                                                 IPCTNL_MSG_CT_NEW, ct) < 0) {
 713                                 nf_conntrack_get(&ct->ct_general);
 714                                 cb->args[1] = (unsigned long)ct;
 715                                 goto out;
 716                         }
 717
 718                         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) ==
 719                                                 IPCTNL_MSG_CT_GET_CTRZERO) {
 720                                 struct nf_conn_counter *acct;
 721
 722                                 acct = nf_conn_acct_find(ct);
 723                                 if (acct) {
 724                                         atomic64_set(&acct[IP_CT_DIR_ORIGINAL].bytes, 0);
 725                                         atomic64_set(&acct[IP_CT_DIR_ORIGINAL].packets, 0);
 726                                         atomic64_set(&acct[IP_CT_DIR_REPLY].bytes, 0);
 727                                         atomic64_set(&acct[IP_CT_DIR_REPLY].packets, 0);
 728                                         }
 729                         }
 730                 }
 731                 if (cb->args[1]) {
 732                         cb->args[1] = 0;
 733                         goto restart;
 734                 }
 735         }
 736 out:
 737         spin_unlock_bh(&nf_conntrack_lock);
 738         if (last)
 739                 nf_ct_put(last);
 740
 741         return skb->len;
 742 }
 743
 744 static inline int
 745 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 746 {
 747         struct nlattr *tb[CTA_IP_MAX+1];
 748         struct nf_conntrack_l3proto *l3proto;
 749         int ret = 0;
 750
 751         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 752
 753         rcu_read_lock();
 754         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 755
 756         if (likely(l3proto->nlattr_to_tuple)) {
 757                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 758                                           l3proto->nla_policy);
 759                 if (ret == 0)
 760                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 761         }
 762
 763         rcu_read_unlock();
 764
 765         return ret;
 766 }
 767
 768 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 769         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 770 };
 771
 772 static inline int
 773 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 774                             struct nf_conntrack_tuple *tuple)
 775 {
 776         struct nlattr *tb[CTA_PROTO_MAX+1];
 777         struct nf_conntrack_l4proto *l4proto;
 778         int ret = 0;
 779
 780         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 781         if (ret < 0)
 782                 return ret;
 783
 784         if (!tb[CTA_PROTO_NUM])
 785                 return -EINVAL;
 786         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 787
 788         rcu_read_lock();
 789         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 790
 791         if (likely(l4proto->nlattr_to_tuple)) {
 792                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 793                                           l4proto->nla_policy);
 794                 if (ret == 0)
 795                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 796         }
 797
 798         rcu_read_unlock();
 799
 800         return ret;
 801 }
 802
 803 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 804         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 805         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 806 };
 807
 808 static int
 809 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 810                       struct nf_conntrack_tuple *tuple,
 811                       enum ctattr_type type, u_int8_t l3num)
 812 {
 813         struct nlattr *tb[CTA_TUPLE_MAX+1];
 814         int err;
 815
 816         memset(tuple, 0, sizeof(*tuple));
 817
 818         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 819
 820         if (!tb[CTA_TUPLE_IP])
 821                 return -EINVAL;
 822
 823         tuple->src.l3num = l3num;
 824
 825         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 826         if (err < 0)
 827                 return err;
 828
 829         if (!tb[CTA_TUPLE_PROTO])
 830                 return -EINVAL;
 831
 832         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 833         if (err < 0)
 834                 return err;
 835
 836         /* orig and expect tuples get DIR_ORIGINAL */
 837         if (type == CTA_TUPLE_REPLY)
 838                 tuple->dst.dir = IP_CT_DIR_REPLY;
 839         else
 840                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 841
 842         return 0;
 843 }
 844
 845 static int
 846 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 847 {
 848         if (attr)
 849 #ifdef CONFIG_NF_CONNTRACK_ZONES
 850                 *zone = ntohs(nla_get_be16(attr));
 851 #else
 852                 return -EOPNOTSUPP;
 853 #endif
 854         else
 855                 *zone = 0;
 856
 857         return 0;
 858 }
 859
 860 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 861         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING },
 862 };
 863
 864 static inline int
 865 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name)
 866 {
 867         struct nlattr *tb[CTA_HELP_MAX+1];
 868
 869         nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 870
 871         if (!tb[CTA_HELP_NAME])
 872                 return -EINVAL;
 873
 874         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 875
 876         return 0;
 877 }
 878
 879 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 880         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 881         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 882         [CTA_STATUS]            = { .type = NLA_U32 },
 883         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 884         [CTA_HELP]              = { .type = NLA_NESTED },
 885         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 886         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 887         [CTA_MARK]              = { .type = NLA_U32 },
 888         [CTA_ID]                = { .type = NLA_U32 },
 889         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 890         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 891         [CTA_ZONE]              = { .type = NLA_U16 },
 892 };
 893
 894 static int
 895 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 896                         const struct nlmsghdr *nlh,
 897                         const struct nlattr * const cda[])
 898 {
 899         struct net *net = sock_net(ctnl);
 900         struct nf_conntrack_tuple_hash *h;
 901         struct nf_conntrack_tuple tuple;
 902         struct nf_conn *ct;
 903         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 904         u_int8_t u3 = nfmsg->nfgen_family;
 905         u16 zone;
 906         int err;
 907
 908         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 909         if (err < 0)
 910                 return err;
 911
 912         if (cda[CTA_TUPLE_ORIG])
 913                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 914         else if (cda[CTA_TUPLE_REPLY])
 915                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 916         else {
 917                 /* Flush the whole table */
 918                 nf_conntrack_flush_report(net,
 919                                          NETLINK_CB(skb).pid,
 920                                          nlmsg_report(nlh));
 921                 return 0;
 922         }
 923
 924         if (err < 0)
 925                 return err;
 926
 927         h = nf_conntrack_find_get(net, zone, &tuple);
 928         if (!h)
 929                 return -ENOENT;
 930
 931         ct = nf_ct_tuplehash_to_ctrack(h);
 932
 933         if (cda[CTA_ID]) {
 934                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
 935                 if (id != (u32)(unsigned long)ct) {
 936                         nf_ct_put(ct);
 937                         return -ENOENT;
 938                 }
 939         }
 940
 941         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 942                                       NETLINK_CB(skb).pid,
 943                                       nlmsg_report(nlh)) < 0) {
 944                 nf_ct_delete_from_lists(ct);
 945                 /* we failed to report the event, try later */
 946                 nf_ct_insert_dying_list(ct);
 947                 nf_ct_put(ct);
 948                 return 0;
 949         }
 950
 951         /* death_by_timeout would report the event again */
 952         set_bit(IPS_DYING_BIT, &ct->status);
 953
 954         nf_ct_kill(ct);
 955         nf_ct_put(ct);
 956
 957         return 0;
 958 }
 959
 960 static int
 961 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 962                         const struct nlmsghdr *nlh,
 963                         const struct nlattr * const cda[])
 964 {
 965         struct net *net = sock_net(ctnl);
 966         struct nf_conntrack_tuple_hash *h;
 967         struct nf_conntrack_tuple tuple;
 968         struct nf_conn *ct;
 969         struct sk_buff *skb2 = NULL;
 970         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 971         u_int8_t u3 = nfmsg->nfgen_family;
 972         u16 zone;
 973         int err;
 974
 975         if (nlh->nlmsg_flags & NLM_F_DUMP)
 976                 return netlink_dump_start(ctnl, skb, nlh, ctnetlink_dump_table,
 977                                           ctnetlink_done, 0);
 978
 979         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 980         if (err < 0)
 981                 return err;
 982
 983         if (cda[CTA_TUPLE_ORIG])
 984                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 985         else if (cda[CTA_TUPLE_REPLY])
 986                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 987         else
 988                 return -EINVAL;
 989
 990         if (err < 0)
 991                 return err;
 992
 993         h = nf_conntrack_find_get(net, zone, &tuple);
 994         if (!h)
 995                 return -ENOENT;
 996
 997         ct = nf_ct_tuplehash_to_ctrack(h);
 998
 999         err = -ENOMEM;
1000         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1001         if (skb2 == NULL) {
1002                 nf_ct_put(ct);
1003                 return -ENOMEM;
1004         }
1005
1006         rcu_read_lock();
1007         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
1008                                   IPCTNL_MSG_CT_NEW, ct);
1009         rcu_read_unlock();
1010         nf_ct_put(ct);
1011         if (err <= 0)
1012                 goto free;
1013
1014         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1015         if (err < 0)
1016                 goto out;
1017
1018         if (NFNL_MSG_TYPE(nlh->nlmsg_type) == IPCTNL_MSG_CT_GET_CTRZERO) {
1019                 struct nf_conn_counter *acct;
1020
1021                 acct = nf_conn_acct_find(ct);
1022                 if (acct) {
1023                         atomic64_set(&acct[IP_CT_DIR_ORIGINAL].bytes, 0);
1024                         atomic64_set(&acct[IP_CT_DIR_ORIGINAL].packets, 0);
1025                         atomic64_set(&acct[IP_CT_DIR_REPLY].bytes, 0);
1026                         atomic64_set(&acct[IP_CT_DIR_REPLY].packets, 0);
1027                 }
1028         }
1029         return 0;
1030
1031 free:
1032         kfree_skb(skb2);
1033 out:
1034         /* this avoids a loop in nfnetlink. */
1035         return err == -EAGAIN ? -ENOBUFS : err;
1036 }
1037
1038 #ifdef CONFIG_NF_NAT_NEEDED
1039 static int
1040 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1041                           enum nf_nat_manip_type manip,
1042                           const struct nlattr *attr)
1043 {
1044         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1045
1046         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1047         if (!parse_nat_setup) {
1048 #ifdef CONFIG_MODULES
1049                 rcu_read_unlock();
1050                 spin_unlock_bh(&nf_conntrack_lock);
1051                 nfnl_unlock();
1052                 if (request_module("nf-nat-ipv4") < 0) {
1053                         nfnl_lock();
1054                         spin_lock_bh(&nf_conntrack_lock);
1055                         rcu_read_lock();
1056                         return -EOPNOTSUPP;
1057                 }
1058                 nfnl_lock();
1059                 spin_lock_bh(&nf_conntrack_lock);
1060                 rcu_read_lock();
1061                 if (nfnetlink_parse_nat_setup_hook)
1062                         return -EAGAIN;
1063 #endif
1064                 return -EOPNOTSUPP;
1065         }
1066
1067         return parse_nat_setup(ct, manip, attr);
1068 }
1069 #endif
1070
1071 static int
1072 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1073 {
1074         unsigned long d;
1075         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1076         d = ct->status ^ status;
1077
1078         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1079                 /* unchangeable */
1080                 return -EBUSY;
1081
1082         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1083                 /* SEEN_REPLY bit can only be set */
1084                 return -EBUSY;
1085
1086         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1087                 /* ASSURED bit can only be set */
1088                 return -EBUSY;
1089
1090         /* Be careful here, modifying NAT bits can screw up things,
1091          * so don't let users modify them directly if they don't pass
1092          * nf_nat_range. */
1093         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1094         return 0;
1095 }
1096
1097 static int
1098 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1099 {
1100 #ifdef CONFIG_NF_NAT_NEEDED
1101         int ret;
1102
1103         if (cda[CTA_NAT_DST]) {
1104                 ret = ctnetlink_parse_nat_setup(ct,
1105                                                 NF_NAT_MANIP_DST,
1106                                                 cda[CTA_NAT_DST]);
1107                 if (ret < 0)
1108                         return ret;
1109         }
1110         if (cda[CTA_NAT_SRC]) {
1111                 ret = ctnetlink_parse_nat_setup(ct,
1112                                                 NF_NAT_MANIP_SRC,
1113                                                 cda[CTA_NAT_SRC]);
1114                 if (ret < 0)
1115                         return ret;
1116         }
1117         return 0;
1118 #else
1119         return -EOPNOTSUPP;
1120 #endif
1121 }
1122
1123 static inline int
1124 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1125 {
1126         struct nf_conntrack_helper *helper;
1127         struct nf_conn_help *help = nfct_help(ct);
1128         char *helpname = NULL;
1129         int err;
1130
1131         /* don't change helper of sibling connections */
1132         if (ct->master)
1133                 return -EBUSY;
1134
1135         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1136         if (err < 0)
1137                 return err;
1138
1139         if (!strcmp(helpname, "")) {
1140                 if (help && help->helper) {
1141                         /* we had a helper before ... */
1142                         nf_ct_remove_expectations(ct);
1143                         RCU_INIT_POINTER(help->helper, NULL);
1144                 }
1145
1146                 return 0;
1147         }
1148
1149         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1150                                             nf_ct_protonum(ct));
1151         if (helper == NULL) {
1152 #ifdef CONFIG_MODULES
1153                 spin_unlock_bh(&nf_conntrack_lock);
1154
1155                 if (request_module("nfct-helper-%s", helpname) < 0) {
1156                         spin_lock_bh(&nf_conntrack_lock);
1157                         return -EOPNOTSUPP;
1158                 }
1159
1160                 spin_lock_bh(&nf_conntrack_lock);
1161                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1162                                                     nf_ct_protonum(ct));
1163                 if (helper)
1164                         return -EAGAIN;
1165 #endif
1166                 return -EOPNOTSUPP;
1167         }
1168
1169         if (help) {
1170                 if (help->helper == helper)
1171                         return 0;
1172                 if (help->helper)
1173                         return -EBUSY;
1174                 /* need to zero data of old helper */
1175                 memset(&help->help, 0, sizeof(help->help));
1176         } else {
1177                 /* we cannot set a helper for an existing conntrack */
1178                 return -EOPNOTSUPP;
1179         }
1180
1181         RCU_INIT_POINTER(help->helper, helper);
1182
1183         return 0;
1184 }
1185
1186 static inline int
1187 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1188 {
1189         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1190
1191         if (!del_timer(&ct->timeout))
1192                 return -ETIME;
1193
1194         ct->timeout.expires = jiffies + timeout * HZ;
1195         add_timer(&ct->timeout);
1196
1197         return 0;
1198 }
1199
1200 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1201         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1202         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1203         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1204 };
1205
1206 static inline int
1207 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1208 {
1209         const struct nlattr *attr = cda[CTA_PROTOINFO];
1210         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1211         struct nf_conntrack_l4proto *l4proto;
1212         int err = 0;
1213
1214         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1215
1216         rcu_read_lock();
1217         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1218         if (l4proto->from_nlattr)
1219                 err = l4proto->from_nlattr(tb, ct);
1220         rcu_read_unlock();
1221
1222         return err;
1223 }
1224
1225 #ifdef CONFIG_NF_NAT_NEEDED
1226 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1227         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1228         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1229         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1230 };
1231
1232 static inline int
1233 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1234 {
1235         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1236
1237         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1238
1239         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1240                 return -EINVAL;
1241
1242         natseq->correction_pos =
1243                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1244
1245         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1246                 return -EINVAL;
1247
1248         natseq->offset_before =
1249                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1250
1251         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1252                 return -EINVAL;
1253
1254         natseq->offset_after =
1255                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1256
1257         return 0;
1258 }
1259
1260 static int
1261 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1262                              const struct nlattr * const cda[])
1263 {
1264         int ret = 0;
1265         struct nf_conn_nat *nat = nfct_nat(ct);
1266
1267         if (!nat)
1268                 return 0;
1269
1270         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1271                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1272                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1273                 if (ret < 0)
1274                         return ret;
1275
1276                 ct->status |= IPS_SEQ_ADJUST;
1277         }
1278
1279         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1280                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1281                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1282                 if (ret < 0)
1283                         return ret;
1284
1285                 ct->status |= IPS_SEQ_ADJUST;
1286         }
1287
1288         return 0;
1289 }
1290 #endif
1291
1292 static int
1293 ctnetlink_change_conntrack(struct nf_conn *ct,
1294                            const struct nlattr * const cda[])
1295 {
1296         int err;
1297
1298         /* only allow NAT changes and master assignation for new conntracks */
1299         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1300                 return -EOPNOTSUPP;
1301
1302         if (cda[CTA_HELP]) {
1303                 err = ctnetlink_change_helper(ct, cda);
1304                 if (err < 0)
1305                         return err;
1306         }
1307
1308         if (cda[CTA_TIMEOUT]) {
1309                 err = ctnetlink_change_timeout(ct, cda);
1310                 if (err < 0)
1311                         return err;
1312         }
1313
1314         if (cda[CTA_STATUS]) {
1315                 err = ctnetlink_change_status(ct, cda);
1316                 if (err < 0)
1317                         return err;
1318         }
1319
1320         if (cda[CTA_PROTOINFO]) {
1321                 err = ctnetlink_change_protoinfo(ct, cda);
1322                 if (err < 0)
1323                         return err;
1324         }
1325
1326 #if defined(CONFIG_NF_CONNTRACK_MARK)
1327         if (cda[CTA_MARK])
1328                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1329 #endif
1330
1331 #ifdef CONFIG_NF_NAT_NEEDED
1332         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1333                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1334                 if (err < 0)
1335                         return err;
1336         }
1337 #endif
1338
1339         return 0;
1340 }
1341
1342 static struct nf_conn *
1343 ctnetlink_create_conntrack(struct net *net, u16 zone,
1344                            const struct nlattr * const cda[],
1345                            struct nf_conntrack_tuple *otuple,
1346                            struct nf_conntrack_tuple *rtuple,
1347                            u8 u3)
1348 {
1349         struct nf_conn *ct;
1350         int err = -EINVAL;
1351         struct nf_conntrack_helper *helper;
1352         struct nf_conn_tstamp *tstamp;
1353
1354         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1355         if (IS_ERR(ct))
1356                 return ERR_PTR(-ENOMEM);
1357
1358         if (!cda[CTA_TIMEOUT])
1359                 goto err1;
1360         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1361
1362         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1363
1364         rcu_read_lock();
1365         if (cda[CTA_HELP]) {
1366                 char *helpname = NULL;
1367
1368                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1369                 if (err < 0)
1370                         goto err2;
1371
1372                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1373                                                     nf_ct_protonum(ct));
1374                 if (helper == NULL) {
1375                         rcu_read_unlock();
1376 #ifdef CONFIG_MODULES
1377                         if (request_module("nfct-helper-%s", helpname) < 0) {
1378                                 err = -EOPNOTSUPP;
1379                                 goto err1;
1380                         }
1381
1382                         rcu_read_lock();
1383                         helper = __nf_conntrack_helper_find(helpname,
1384                                                             nf_ct_l3num(ct),
1385                                                             nf_ct_protonum(ct));
1386                         if (helper) {
1387                                 err = -EAGAIN;
1388                                 goto err2;
1389                         }
1390                         rcu_read_unlock();
1391 #endif
1392                         err = -EOPNOTSUPP;
1393                         goto err1;
1394                 } else {
1395                         struct nf_conn_help *help;
1396
1397                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1398                         if (help == NULL) {
1399                                 err = -ENOMEM;
1400                                 goto err2;
1401                         }
1402
1403                         /* not in hash table yet so not strictly necessary */
1404                         RCU_INIT_POINTER(help->helper, helper);
1405                 }
1406         } else {
1407                 /* try an implicit helper assignation */
1408                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1409                 if (err < 0)
1410                         goto err2;
1411         }
1412
1413         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1414                 err = ctnetlink_change_nat(ct, cda);
1415                 if (err < 0)
1416                         goto err2;
1417         }
1418
1419         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1420         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1421         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1422         /* we must add conntrack extensions before confirmation. */
1423         ct->status |= IPS_CONFIRMED;
1424
1425         if (cda[CTA_STATUS]) {
1426                 err = ctnetlink_change_status(ct, cda);
1427                 if (err < 0)
1428                         goto err2;
1429         }
1430
1431 #ifdef CONFIG_NF_NAT_NEEDED
1432         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1433                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1434                 if (err < 0)
1435                         goto err2;
1436         }
1437 #endif
1438
1439         memset(&ct->proto, 0, sizeof(ct->proto));
1440         if (cda[CTA_PROTOINFO]) {
1441                 err = ctnetlink_change_protoinfo(ct, cda);
1442                 if (err < 0)
1443                         goto err2;
1444         }
1445
1446 #if defined(CONFIG_NF_CONNTRACK_MARK)
1447         if (cda[CTA_MARK])
1448                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1449 #endif
1450
1451         /* setup master conntrack: this is a confirmed expectation */
1452         if (cda[CTA_TUPLE_MASTER]) {
1453                 struct nf_conntrack_tuple master;
1454                 struct nf_conntrack_tuple_hash *master_h;
1455                 struct nf_conn *master_ct;
1456
1457                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1458                 if (err < 0)
1459                         goto err2;
1460
1461                 master_h = nf_conntrack_find_get(net, zone, &master);
1462                 if (master_h == NULL) {
1463                         err = -ENOENT;
1464                         goto err2;
1465                 }
1466                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1467                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1468                 ct->master = master_ct;
1469         }
1470         tstamp = nf_conn_tstamp_find(ct);
1471         if (tstamp)
1472                 tstamp->start = ktime_to_ns(ktime_get_real());
1473
1474         add_timer(&ct->timeout);
1475         nf_conntrack_hash_insert(ct);
1476         rcu_read_unlock();
1477
1478         return ct;
1479
1480 err2:
1481         rcu_read_unlock();
1482 err1:
1483         nf_conntrack_free(ct);
1484         return ERR_PTR(err);
1485 }
1486
1487 static int
1488 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1489                         const struct nlmsghdr *nlh,
1490                         const struct nlattr * const cda[])
1491 {
1492         struct net *net = sock_net(ctnl);
1493         struct nf_conntrack_tuple otuple, rtuple;
1494         struct nf_conntrack_tuple_hash *h = NULL;
1495         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1496         u_int8_t u3 = nfmsg->nfgen_family;
1497         u16 zone;
1498         int err;
1499
1500         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1501         if (err < 0)
1502                 return err;
1503
1504         if (cda[CTA_TUPLE_ORIG]) {
1505                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1506                 if (err < 0)
1507                         return err;
1508         }
1509
1510         if (cda[CTA_TUPLE_REPLY]) {
1511                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1512                 if (err < 0)
1513                         return err;
1514         }
1515
1516         spin_lock_bh(&nf_conntrack_lock);
1517         if (cda[CTA_TUPLE_ORIG])
1518                 h = __nf_conntrack_find(net, zone, &otuple);
1519         else if (cda[CTA_TUPLE_REPLY])
1520                 h = __nf_conntrack_find(net, zone, &rtuple);
1521
1522         if (h == NULL) {
1523                 err = -ENOENT;
1524                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1525                         struct nf_conn *ct;
1526                         enum ip_conntrack_events events;
1527
1528                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1529                                                         &rtuple, u3);
1530                         if (IS_ERR(ct)) {
1531                                 err = PTR_ERR(ct);
1532                                 goto out_unlock;
1533                         }
1534                         err = 0;
1535                         nf_conntrack_get(&ct->ct_general);
1536                         spin_unlock_bh(&nf_conntrack_lock);
1537                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1538                                 events = IPCT_RELATED;
1539                         else
1540                                 events = IPCT_NEW;
1541
1542                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1543                                                       (1 << IPCT_ASSURED) |
1544                                                       (1 << IPCT_HELPER) |
1545                                                       (1 << IPCT_PROTOINFO) |
1546                                                       (1 << IPCT_NATSEQADJ) |
1547                                                       (1 << IPCT_MARK) | events,
1548                                                       ct, NETLINK_CB(skb).pid,
1549                                                       nlmsg_report(nlh));
1550                         nf_ct_put(ct);
1551                 } else
1552                         spin_unlock_bh(&nf_conntrack_lock);
1553
1554                 return err;
1555         }
1556         /* implicit 'else' */
1557
1558         /* We manipulate the conntrack inside the global conntrack table lock,
1559          * so there's no need to increase the refcount */
1560         err = -EEXIST;
1561         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1562                 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1563
1564                 err = ctnetlink_change_conntrack(ct, cda);
1565                 if (err == 0) {
1566                         nf_conntrack_get(&ct->ct_general);
1567                         spin_unlock_bh(&nf_conntrack_lock);
1568                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1569                                                       (1 << IPCT_ASSURED) |
1570                                                       (1 << IPCT_HELPER) |
1571                                                       (1 << IPCT_PROTOINFO) |
1572                                                       (1 << IPCT_NATSEQADJ) |
1573                                                       (1 << IPCT_MARK),
1574                                                       ct, NETLINK_CB(skb).pid,
1575                                                       nlmsg_report(nlh));
1576                         nf_ct_put(ct);
1577                 } else
1578                         spin_unlock_bh(&nf_conntrack_lock);
1579
1580                 return err;
1581         }
1582
1583 out_unlock:
1584         spin_unlock_bh(&nf_conntrack_lock);
1585         return err;
1586 }
1587
1588 /***********************************************************************
1589  * EXPECT
1590  ***********************************************************************/
1591
1592 static inline int
1593 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1594                          const struct nf_conntrack_tuple *tuple,
1595                          enum ctattr_expect type)
1596 {
1597         struct nlattr *nest_parms;
1598
1599         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
1600         if (!nest_parms)
1601                 goto nla_put_failure;
1602         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1603                 goto nla_put_failure;
1604         nla_nest_end(skb, nest_parms);
1605
1606         return 0;
1607
1608 nla_put_failure:
1609         return -1;
1610 }
1611
1612 static inline int
1613 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1614                         const struct nf_conntrack_tuple *tuple,
1615                         const struct nf_conntrack_tuple_mask *mask)
1616 {
1617         int ret;
1618         struct nf_conntrack_l3proto *l3proto;
1619         struct nf_conntrack_l4proto *l4proto;
1620         struct nf_conntrack_tuple m;
1621         struct nlattr *nest_parms;
1622
1623         memset(&m, 0xFF, sizeof(m));
1624         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1625         m.src.u.all = mask->src.u.all;
1626         m.dst.protonum = tuple->dst.protonum;
1627
1628         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
1629         if (!nest_parms)
1630                 goto nla_put_failure;
1631
1632         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
1633         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1634
1635         if (unlikely(ret < 0))
1636                 goto nla_put_failure;
1637
1638         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
1639         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1640         if (unlikely(ret < 0))
1641                 goto nla_put_failure;
1642
1643         nla_nest_end(skb, nest_parms);
1644
1645         return 0;
1646
1647 nla_put_failure:
1648         return -1;
1649 }
1650
1651 static int
1652 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1653                           const struct nf_conntrack_expect *exp)
1654 {
1655         struct nf_conn *master = exp->master;
1656         long timeout = (exp->timeout.expires - jiffies) / HZ;
1657         struct nf_conn_help *help;
1658
1659         if (timeout < 0)
1660                 timeout = 0;
1661
1662         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1663                 goto nla_put_failure;
1664         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1665                 goto nla_put_failure;
1666         if (ctnetlink_exp_dump_tuple(skb,
1667                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1668                                  CTA_EXPECT_MASTER) < 0)
1669                 goto nla_put_failure;
1670
1671         NLA_PUT_BE32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout));
1672         NLA_PUT_BE32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp));
1673         NLA_PUT_BE32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags));
1674         help = nfct_help(master);
1675         if (help) {
1676                 struct nf_conntrack_helper *helper;
1677
1678                 helper = rcu_dereference(help->helper);
1679                 if (helper)
1680                         NLA_PUT_STRING(skb, CTA_EXPECT_HELP_NAME, helper->name);
1681         }
1682
1683         return 0;
1684
1685 nla_put_failure:
1686         return -1;
1687 }
1688
1689 static int
1690 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1691                         int event, const struct nf_conntrack_expect *exp)
1692 {
1693         struct nlmsghdr *nlh;
1694         struct nfgenmsg *nfmsg;
1695         unsigned int flags = pid ? NLM_F_MULTI : 0;
1696
1697         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1698         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
1699         if (nlh == NULL)
1700                 goto nlmsg_failure;
1701
1702         nfmsg = nlmsg_data(nlh);
1703         nfmsg->nfgen_family = exp->tuple.src.l3num;
1704         nfmsg->version      = NFNETLINK_V0;
1705         nfmsg->res_id       = 0;
1706
1707         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1708                 goto nla_put_failure;
1709
1710         nlmsg_end(skb, nlh);
1711         return skb->len;
1712
1713 nlmsg_failure:
1714 nla_put_failure:
1715         nlmsg_cancel(skb, nlh);
1716         return -1;
1717 }
1718
1719 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1720 static int
1721 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
1722 {
1723         struct nf_conntrack_expect *exp = item->exp;
1724         struct net *net = nf_ct_exp_net(exp);
1725         struct nlmsghdr *nlh;
1726         struct nfgenmsg *nfmsg;
1727         struct sk_buff *skb;
1728         unsigned int type, group;
1729         int flags = 0;
1730
1731         if (events & (1 << IPEXP_DESTROY)) {
1732                 type = IPCTNL_MSG_EXP_DELETE;
1733                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
1734         } else if (events & (1 << IPEXP_NEW)) {
1735                 type = IPCTNL_MSG_EXP_NEW;
1736                 flags = NLM_F_CREATE|NLM_F_EXCL;
1737                 group = NFNLGRP_CONNTRACK_EXP_NEW;
1738         } else
1739                 return 0;
1740
1741         if (!item->report && !nfnetlink_has_listeners(net, group))
1742                 return 0;
1743
1744         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
1745         if (skb == NULL)
1746                 goto errout;
1747
1748         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1749         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
1750         if (nlh == NULL)
1751                 goto nlmsg_failure;
1752
1753         nfmsg = nlmsg_data(nlh);
1754         nfmsg->nfgen_family = exp->tuple.src.l3num;
1755         nfmsg->version      = NFNETLINK_V0;
1756         nfmsg->res_id       = 0;
1757
1758         rcu_read_lock();
1759         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1760                 goto nla_put_failure;
1761         rcu_read_unlock();
1762
1763         nlmsg_end(skb, nlh);
1764         nfnetlink_send(skb, net, item->pid, group, item->report, GFP_ATOMIC);
1765         return 0;
1766
1767 nla_put_failure:
1768         rcu_read_unlock();
1769         nlmsg_cancel(skb, nlh);
1770 nlmsg_failure:
1771         kfree_skb(skb);
1772 errout:
1773         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
1774         return 0;
1775 }
1776 #endif
1777 static int ctnetlink_exp_done(struct netlink_callback *cb)
1778 {
1779         if (cb->args[1])
1780                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1781         return 0;
1782 }
1783
1784 static int
1785 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1786 {
1787         struct net *net = sock_net(skb->sk);
1788         struct nf_conntrack_expect *exp, *last;
1789         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1790         struct hlist_node *n;
1791         u_int8_t l3proto = nfmsg->nfgen_family;
1792
1793         rcu_read_lock();
1794         last = (struct nf_conntrack_expect *)cb->args[1];
1795         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1796 restart:
1797                 hlist_for_each_entry(exp, n, &net->ct.expect_hash[cb->args[0]],
1798                                      hnode) {
1799                         if (l3proto && exp->tuple.src.l3num != l3proto)
1800                                 continue;
1801                         if (cb->args[1]) {
1802                                 if (exp != last)
1803                                         continue;
1804                                 cb->args[1] = 0;
1805                         }
1806                         if (ctnetlink_exp_fill_info(skb,
1807                                                     NETLINK_CB(cb->skb).pid,
1808                                                     cb->nlh->nlmsg_seq,
1809                                                     IPCTNL_MSG_EXP_NEW,
1810                                                     exp) < 0) {
1811                                 if (!atomic_inc_not_zero(&exp->use))
1812                                         continue;
1813                                 cb->args[1] = (unsigned long)exp;
1814                                 goto out;
1815                         }
1816                 }
1817                 if (cb->args[1]) {
1818                         cb->args[1] = 0;
1819                         goto restart;
1820                 }
1821         }
1822 out:
1823         rcu_read_unlock();
1824         if (last)
1825                 nf_ct_expect_put(last);
1826
1827         return skb->len;
1828 }
1829
1830 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1831         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1832         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1833         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1834         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1835         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1836         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING },
1837         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1838         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1839 };
1840
1841 static int
1842 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1843                      const struct nlmsghdr *nlh,
1844                      const struct nlattr * const cda[])
1845 {
1846         struct net *net = sock_net(ctnl);
1847         struct nf_conntrack_tuple tuple;
1848         struct nf_conntrack_expect *exp;
1849         struct sk_buff *skb2;
1850         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1851         u_int8_t u3 = nfmsg->nfgen_family;
1852         u16 zone;
1853         int err;
1854
1855         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1856                 return netlink_dump_start(ctnl, skb, nlh,
1857                                           ctnetlink_exp_dump_table,
1858                                           ctnetlink_exp_done, 0);
1859         }
1860
1861         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1862         if (err < 0)
1863                 return err;
1864
1865         if (cda[CTA_EXPECT_TUPLE])
1866                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1867         else if (cda[CTA_EXPECT_MASTER])
1868                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1869         else
1870                 return -EINVAL;
1871
1872         if (err < 0)
1873                 return err;
1874
1875         exp = nf_ct_expect_find_get(net, zone, &tuple);
1876         if (!exp)
1877                 return -ENOENT;
1878
1879         if (cda[CTA_EXPECT_ID]) {
1880                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1881                 if (ntohl(id) != (u32)(unsigned long)exp) {
1882                         nf_ct_expect_put(exp);
1883                         return -ENOENT;
1884                 }
1885         }
1886
1887         err = -ENOMEM;
1888         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1889         if (skb2 == NULL)
1890                 goto out;
1891
1892         rcu_read_lock();
1893         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1894                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
1895         rcu_read_unlock();
1896         if (err <= 0)
1897                 goto free;
1898
1899         nf_ct_expect_put(exp);
1900
1901         return netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1902
1903 free:
1904         kfree_skb(skb2);
1905 out:
1906         nf_ct_expect_put(exp);
1907         return err;
1908 }
1909
1910 static int
1911 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1912                      const struct nlmsghdr *nlh,
1913                      const struct nlattr * const cda[])
1914 {
1915         struct net *net = sock_net(ctnl);
1916         struct nf_conntrack_expect *exp;
1917         struct nf_conntrack_tuple tuple;
1918         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1919         struct hlist_node *n, *next;
1920         u_int8_t u3 = nfmsg->nfgen_family;
1921         unsigned int i;
1922         u16 zone;
1923         int err;
1924
1925         if (cda[CTA_EXPECT_TUPLE]) {
1926                 /* delete a single expect by tuple */
1927                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1928                 if (err < 0)
1929                         return err;
1930
1931                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1932                 if (err < 0)
1933                         return err;
1934
1935                 /* bump usage count to 2 */
1936                 exp = nf_ct_expect_find_get(net, zone, &tuple);
1937                 if (!exp)
1938                         return -ENOENT;
1939
1940                 if (cda[CTA_EXPECT_ID]) {
1941                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1942                         if (ntohl(id) != (u32)(unsigned long)exp) {
1943                                 nf_ct_expect_put(exp);
1944                                 return -ENOENT;
1945                         }
1946                 }
1947
1948                 /* after list removal, usage count == 1 */
1949                 spin_lock_bh(&nf_conntrack_lock);
1950                 if (del_timer(&exp->timeout)) {
1951                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).pid,
1952                                                    nlmsg_report(nlh));
1953                         nf_ct_expect_put(exp);
1954                 }
1955                 spin_unlock_bh(&nf_conntrack_lock);
1956                 /* have to put what we 'get' above.
1957                  * after this line usage count == 0 */
1958                 nf_ct_expect_put(exp);
1959         } else if (cda[CTA_EXPECT_HELP_NAME]) {
1960                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
1961                 struct nf_conn_help *m_help;
1962
1963                 /* delete all expectations for this helper */
1964                 spin_lock_bh(&nf_conntrack_lock);
1965                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1966                         hlist_for_each_entry_safe(exp, n, next,
1967                                                   &net->ct.expect_hash[i],
1968                                                   hnode) {
1969                                 m_help = nfct_help(exp->master);
1970                                 if (!strcmp(m_help->helper->name, name) &&
1971                                     del_timer(&exp->timeout)) {
1972                                         nf_ct_unlink_expect_report(exp,
1973                                                         NETLINK_CB(skb).pid,
1974                                                         nlmsg_report(nlh));
1975                                         nf_ct_expect_put(exp);
1976                                 }
1977                         }
1978                 }
1979                 spin_unlock_bh(&nf_conntrack_lock);
1980         } else {
1981                 /* This basically means we have to flush everything*/
1982                 spin_lock_bh(&nf_conntrack_lock);
1983                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1984                         hlist_for_each_entry_safe(exp, n, next,
1985                                                   &net->ct.expect_hash[i],
1986                                                   hnode) {
1987                                 if (del_timer(&exp->timeout)) {
1988                                         nf_ct_unlink_expect_report(exp,
1989                                                         NETLINK_CB(skb).pid,
1990                                                         nlmsg_report(nlh));
1991                                         nf_ct_expect_put(exp);
1992                                 }
1993                         }
1994                 }
1995                 spin_unlock_bh(&nf_conntrack_lock);
1996         }
1997
1998         return 0;
1999 }
2000 static int
2001 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2002                         const struct nlattr * const cda[])
2003 {
2004         return -EOPNOTSUPP;
2005 }
2006
2007 static int
2008 ctnetlink_create_expect(struct net *net, u16 zone,
2009                         const struct nlattr * const cda[],
2010                         u_int8_t u3,
2011                         u32 pid, int report)
2012 {
2013         struct nf_conntrack_tuple tuple, mask, master_tuple;
2014         struct nf_conntrack_tuple_hash *h = NULL;
2015         struct nf_conntrack_expect *exp;
2016         struct nf_conn *ct;
2017         struct nf_conn_help *help;
2018         int err = 0;
2019
2020         /* caller guarantees that those three CTA_EXPECT_* exist */
2021         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2022         if (err < 0)
2023                 return err;
2024         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2025         if (err < 0)
2026                 return err;
2027         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2028         if (err < 0)
2029                 return err;
2030
2031         /* Look for master conntrack of this expectation */
2032         h = nf_conntrack_find_get(net, zone, &master_tuple);
2033         if (!h)
2034                 return -ENOENT;
2035         ct = nf_ct_tuplehash_to_ctrack(h);
2036         exp = nf_ct_expect_alloc(ct);
2037         if (!exp) {
2038                 err = -ENOMEM;
2039                 goto out;
2040         }
2041         help = nfct_help(ct);
2042         if (!help) {
2043                 err = -EOPNOTSUPP;
2044                 goto out;
2045         }
2046         if (test_bit(IPS_USERSPACE_HELPER_BIT, &ct->status)) {
2047                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2048                         err = -EINVAL;
2049                         goto out;
2050                 }
2051                 exp->timeout.expires =
2052                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2053
2054                 exp->flags = NF_CT_EXPECT_USERSPACE;
2055                 if (cda[CTA_EXPECT_FLAGS]) {
2056                         exp->flags |=
2057                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2058                 }
2059         } else {
2060                 if (cda[CTA_EXPECT_FLAGS]) {
2061                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2062                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2063                 } else
2064                         exp->flags = 0;
2065         }
2066
2067         exp->class = 0;
2068         exp->expectfn = NULL;
2069         exp->master = ct;
2070         exp->helper = NULL;
2071         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2072         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2073         exp->mask.src.u.all = mask.src.u.all;
2074
2075         err = nf_ct_expect_related_report(exp, pid, report);
2076         nf_ct_expect_put(exp);
2077
2078 out:
2079         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2080         return err;
2081 }
2082
2083 static int
2084 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2085                      const struct nlmsghdr *nlh,
2086                      const struct nlattr * const cda[])
2087 {
2088         struct net *net = sock_net(ctnl);
2089         struct nf_conntrack_tuple tuple;
2090         struct nf_conntrack_expect *exp;
2091         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2092         u_int8_t u3 = nfmsg->nfgen_family;
2093         u16 zone;
2094         int err;
2095
2096         if (!cda[CTA_EXPECT_TUPLE]
2097             || !cda[CTA_EXPECT_MASK]
2098             || !cda[CTA_EXPECT_MASTER])
2099                 return -EINVAL;
2100
2101         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2102         if (err < 0)
2103                 return err;
2104
2105         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2106         if (err < 0)
2107                 return err;
2108
2109         spin_lock_bh(&nf_conntrack_lock);
2110         exp = __nf_ct_expect_find(net, zone, &tuple);
2111
2112         if (!exp) {
2113                 spin_unlock_bh(&nf_conntrack_lock);
2114                 err = -ENOENT;
2115                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2116                         err = ctnetlink_create_expect(net, zone, cda,
2117                                                       u3,
2118                                                       NETLINK_CB(skb).pid,
2119                                                       nlmsg_report(nlh));
2120                 }
2121                 return err;
2122         }
2123
2124         err = -EEXIST;
2125         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2126                 err = ctnetlink_change_expect(exp, cda);
2127         spin_unlock_bh(&nf_conntrack_lock);
2128
2129         return err;
2130 }
2131
2132 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2133 static struct nf_ct_event_notifier ctnl_notifier = {
2134         .fcn = ctnetlink_conntrack_event,
2135 };
2136
2137 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2138         .fcn = ctnetlink_expect_event,
2139 };
2140 #endif
2141
2142 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2143         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2144                                             .attr_count = CTA_MAX,
2145                                             .policy = ct_nla_policy },
2146         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2147                                             .attr_count = CTA_MAX,
2148                                             .policy = ct_nla_policy },
2149         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2150                                             .attr_count = CTA_MAX,
2151                                             .policy = ct_nla_policy },
2152         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
2153                                             .attr_count = CTA_MAX,
2154                                             .policy = ct_nla_policy },
2155 };
2156
2157 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
2158         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
2159                                             .attr_count = CTA_EXPECT_MAX,
2160                                             .policy = exp_nla_policy },
2161         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
2162                                             .attr_count = CTA_EXPECT_MAX,
2163                                             .policy = exp_nla_policy },
2164         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
2165                                             .attr_count = CTA_EXPECT_MAX,
2166                                             .policy = exp_nla_policy },
2167 };
2168
2169 static const struct nfnetlink_subsystem ctnl_subsys = {
2170         .name                           = "conntrack",
2171         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
2172         .cb_count                       = IPCTNL_MSG_MAX,
2173         .cb                             = ctnl_cb,
2174 };
2175
2176 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
2177         .name                           = "conntrack_expect",
2178         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
2179         .cb_count                       = IPCTNL_MSG_EXP_MAX,
2180         .cb                             = ctnl_exp_cb,
2181 };
2182
2183 MODULE_ALIAS("ip_conntrack_netlink");
2184 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
2185 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
2186
2187 static int __net_init ctnetlink_net_init(struct net *net)
2188 {
2189 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2190         int ret;
2191
2192         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
2193         if (ret < 0) {
2194                 pr_err("ctnetlink_init: cannot register notifier.\n");
2195                 goto err_out;
2196         }
2197
2198         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
2199         if (ret < 0) {
2200                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
2201                 goto err_unreg_notifier;
2202         }
2203 #endif
2204         return 0;
2205
2206 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2207 err_unreg_notifier:
2208         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2209 err_out:
2210         return ret;
2211 #endif
2212 }
2213
2214 static void ctnetlink_net_exit(struct net *net)
2215 {
2216 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2217         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
2218         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2219 #endif
2220 }
2221
2222 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
2223 {
2224         struct net *net;
2225
2226         list_for_each_entry(net, net_exit_list, exit_list)
2227                 ctnetlink_net_exit(net);
2228 }
2229
2230 static struct pernet_operations ctnetlink_net_ops = {
2231         .init           = ctnetlink_net_init,
2232         .exit_batch     = ctnetlink_net_exit_batch,
2233 };
2234
2235 static int __init ctnetlink_init(void)
2236 {
2237         int ret;
2238
2239         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
2240         ret = nfnetlink_subsys_register(&ctnl_subsys);
2241         if (ret < 0) {
2242                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
2243                 goto err_out;
2244         }
2245
2246         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
2247         if (ret < 0) {
2248                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
2249                 goto err_unreg_subsys;
2250         }
2251
2252         if (register_pernet_subsys(&ctnetlink_net_ops)) {
2253                 pr_err("ctnetlink_init: cannot register pernet operations\n");
2254                 goto err_unreg_exp_subsys;
2255         }
2256
2257         return 0;
2258
2259 err_unreg_exp_subsys:
2260         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2261 err_unreg_subsys:
2262         nfnetlink_subsys_unregister(&ctnl_subsys);
2263 err_out:
2264         return ret;
2265 }
2266
2267 static void __exit ctnetlink_exit(void)
2268 {
2269         pr_info("ctnetlink: unregistering from nfnetlink.\n");
2270
2271         unregister_pernet_subsys(&ctnetlink_net_ops);
2272         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2273         nfnetlink_subsys_unregister(&ctnl_subsys);
2274 }
2275
2276 module_init(ctnetlink_init);
2277 module_exit(ctnetlink_exit);