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 = ((long)ct->timeout.expires - (long)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 dump_counters(struct sk_buff *skb, u64 pkts, u64 bytes,
 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
 212         nest_count = nla_nest_start(skb, type | NLA_F_NESTED);
 213         if (!nest_count)
 214                 goto nla_put_failure;
 215
 216         NLA_PUT_BE64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts));
 217         NLA_PUT_BE64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes));
 218
 219         nla_nest_end(skb, nest_count);
 220
 221         return 0;
 222
 223 nla_put_failure:
 224         return -1;
 225 }
 226
 227 static int
 228 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 229                         enum ip_conntrack_dir dir, int type)
 230 {
 231         struct nf_conn_counter *acct;
 232         u64 pkts, bytes;
 233
 234         acct = nf_conn_acct_find(ct);
 235         if (!acct)
 236                 return 0;
 237
 238         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 239                 pkts = atomic64_xchg(&acct[dir].packets, 0);
 240                 bytes = atomic64_xchg(&acct[dir].bytes, 0);
 241         } else {
 242                 pkts = atomic64_read(&acct[dir].packets);
 243                 bytes = atomic64_read(&acct[dir].bytes);
 244         }
 245         return dump_counters(skb, pkts, bytes, dir);
 246 }
 247
 248 static int
 249 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 250 {
 251         struct nlattr *nest_count;
 252         const struct nf_conn_tstamp *tstamp;
 253
 254         tstamp = nf_conn_tstamp_find(ct);
 255         if (!tstamp)
 256                 return 0;
 257
 258         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 259         if (!nest_count)
 260                 goto nla_put_failure;
 261
 262         NLA_PUT_BE64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start));
 263         if (tstamp->stop != 0) {
 264                 NLA_PUT_BE64(skb, CTA_TIMESTAMP_STOP,
 265                              cpu_to_be64(tstamp->stop));
 266         }
 267         nla_nest_end(skb, nest_count);
 268
 269         return 0;
 270
 271 nla_put_failure:
 272         return -1;
 273 }
 274
 275 #ifdef CONFIG_NF_CONNTRACK_MARK
 276 static inline int
 277 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 278 {
 279         NLA_PUT_BE32(skb, CTA_MARK, htonl(ct->mark));
 280         return 0;
 281
 282 nla_put_failure:
 283         return -1;
 284 }
 285 #else
 286 #define ctnetlink_dump_mark(a, b) (0)
 287 #endif
 288
 289 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 290 static inline int
 291 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 292 {
 293         struct nlattr *nest_secctx;
 294         int len, ret;
 295         char *secctx;
 296
 297         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 298         if (ret)
 299                 return 0;
 300
 301         ret = -1;
 302         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 303         if (!nest_secctx)
 304                 goto nla_put_failure;
 305
 306         NLA_PUT_STRING(skb, CTA_SECCTX_NAME, secctx);
 307         nla_nest_end(skb, nest_secctx);
 308
 309         ret = 0;
 310 nla_put_failure:
 311         security_release_secctx(secctx, len);
 312         return ret;
 313 }
 314 #else
 315 #define ctnetlink_dump_secctx(a, b) (0)
 316 #endif
 317
 318 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 319
 320 static inline int
 321 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 322 {
 323         struct nlattr *nest_parms;
 324
 325         if (!(ct->status & IPS_EXPECTED))
 326                 return 0;
 327
 328         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 329         if (!nest_parms)
 330                 goto nla_put_failure;
 331         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 332                 goto nla_put_failure;
 333         nla_nest_end(skb, nest_parms);
 334
 335         return 0;
 336
 337 nla_put_failure:
 338         return -1;
 339 }
 340
 341 #ifdef CONFIG_NF_NAT_NEEDED
 342 static int
 343 dump_nat_seq_adj(struct sk_buff *skb, const struct nf_nat_seq *natseq, int type)
 344 {
 345         struct nlattr *nest_parms;
 346
 347         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 348         if (!nest_parms)
 349                 goto nla_put_failure;
 350
 351         NLA_PUT_BE32(skb, CTA_NAT_SEQ_CORRECTION_POS,
 352                      htonl(natseq->correction_pos));
 353         NLA_PUT_BE32(skb, CTA_NAT_SEQ_OFFSET_BEFORE,
 354                      htonl(natseq->offset_before));
 355         NLA_PUT_BE32(skb, CTA_NAT_SEQ_OFFSET_AFTER,
 356                      htonl(natseq->offset_after));
 357
 358         nla_nest_end(skb, nest_parms);
 359
 360         return 0;
 361
 362 nla_put_failure:
 363         return -1;
 364 }
 365
 366 static inline int
 367 ctnetlink_dump_nat_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 368 {
 369         struct nf_nat_seq *natseq;
 370         struct nf_conn_nat *nat = nfct_nat(ct);
 371
 372         if (!(ct->status & IPS_SEQ_ADJUST) || !nat)
 373                 return 0;
 374
 375         natseq = &nat->seq[IP_CT_DIR_ORIGINAL];
 376         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_ORIG) == -1)
 377                 return -1;
 378
 379         natseq = &nat->seq[IP_CT_DIR_REPLY];
 380         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_REPLY) == -1)
 381                 return -1;
 382
 383         return 0;
 384 }
 385 #else
 386 #define ctnetlink_dump_nat_seq_adj(a, b) (0)
 387 #endif
 388
 389 static inline int
 390 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 391 {
 392         NLA_PUT_BE32(skb, CTA_ID, htonl((unsigned long)ct));
 393         return 0;
 394
 395 nla_put_failure:
 396         return -1;
 397 }
 398
 399 static inline int
 400 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 401 {
 402         NLA_PUT_BE32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use)));
 403         return 0;
 404
 405 nla_put_failure:
 406         return -1;
 407 }
 408
 409 static int
 410 ctnetlink_fill_info(struct sk_buff *skb, u32 pid, u32 seq, u32 type,
 411                     struct nf_conn *ct)
 412 {
 413         struct nlmsghdr *nlh;
 414         struct nfgenmsg *nfmsg;
 415         struct nlattr *nest_parms;
 416         unsigned int flags = pid ? NLM_F_MULTI : 0, event;
 417
 418         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 419         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
 420         if (nlh == NULL)
 421                 goto nlmsg_failure;
 422
 423         nfmsg = nlmsg_data(nlh);
 424         nfmsg->nfgen_family = nf_ct_l3num(ct);
 425         nfmsg->version      = NFNETLINK_V0;
 426         nfmsg->res_id       = 0;
 427
 428         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 429         if (!nest_parms)
 430                 goto nla_put_failure;
 431         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 432                 goto nla_put_failure;
 433         nla_nest_end(skb, nest_parms);
 434
 435         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 436         if (!nest_parms)
 437                 goto nla_put_failure;
 438         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 439                 goto nla_put_failure;
 440         nla_nest_end(skb, nest_parms);
 441
 442         if (nf_ct_zone(ct))
 443                 NLA_PUT_BE16(skb, CTA_ZONE, htons(nf_ct_zone(ct)));
 444
 445         if (ctnetlink_dump_status(skb, ct) < 0 ||
 446             ctnetlink_dump_timeout(skb, ct) < 0 ||
 447             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL, type) < 0 ||
 448             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY, type) < 0 ||
 449             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 450             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 451             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 452             ctnetlink_dump_mark(skb, ct) < 0 ||
 453             ctnetlink_dump_secctx(skb, ct) < 0 ||
 454             ctnetlink_dump_id(skb, ct) < 0 ||
 455             ctnetlink_dump_use(skb, ct) < 0 ||
 456             ctnetlink_dump_master(skb, ct) < 0 ||
 457             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 458                 goto nla_put_failure;
 459
 460         nlmsg_end(skb, nlh);
 461         return skb->len;
 462
 463 nlmsg_failure:
 464 nla_put_failure:
 465         nlmsg_cancel(skb, nlh);
 466         return -1;
 467 }
 468
 469 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 470 static inline size_t
 471 ctnetlink_proto_size(const struct nf_conn *ct)
 472 {
 473         struct nf_conntrack_l3proto *l3proto;
 474         struct nf_conntrack_l4proto *l4proto;
 475         size_t len = 0;
 476
 477         rcu_read_lock();
 478         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 479         len += l3proto->nla_size;
 480
 481         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 482         len += l4proto->nla_size;
 483         rcu_read_unlock();
 484
 485         return len;
 486 }
 487
 488 static inline size_t
 489 ctnetlink_counters_size(const struct nf_conn *ct)
 490 {
 491         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 492                 return 0;
 493         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 494                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 495                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 496                ;
 497 }
 498
 499 static inline int
 500 ctnetlink_secctx_size(const struct nf_conn *ct)
 501 {
 502 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 503         int len, ret;
 504
 505         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 506         if (ret)
 507                 return 0;
 508
 509         return nla_total_size(0) /* CTA_SECCTX */
 510                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 511 #else
 512         return 0;
 513 #endif
 514 }
 515
 516 static inline size_t
 517 ctnetlink_timestamp_size(const struct nf_conn *ct)
 518 {
 519 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 520         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 521                 return 0;
 522         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 523 #else
 524         return 0;
 525 #endif
 526 }
 527
 528 static inline size_t
 529 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 530 {
 531         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 532                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 533                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 534                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 535                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 536                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 537                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 538                + ctnetlink_counters_size(ct)
 539                + ctnetlink_timestamp_size(ct)
 540                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 541                + nla_total_size(0) /* CTA_PROTOINFO */
 542                + nla_total_size(0) /* CTA_HELP */
 543                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 544                + ctnetlink_secctx_size(ct)
 545 #ifdef CONFIG_NF_NAT_NEEDED
 546                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 547                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 548 #endif
 549 #ifdef CONFIG_NF_CONNTRACK_MARK
 550                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 551 #endif
 552                + ctnetlink_proto_size(ct)
 553                ;
 554 }
 555
 556 static int
 557 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 558 {
 559         struct net *net;
 560         struct nlmsghdr *nlh;
 561         struct nfgenmsg *nfmsg;
 562         struct nlattr *nest_parms;
 563         struct nf_conn *ct = item->ct;
 564         struct sk_buff *skb;
 565         unsigned int type;
 566         unsigned int flags = 0, group;
 567         int err;
 568
 569         /* ignore our fake conntrack entry */
 570         if (nf_ct_is_untracked(ct))
 571                 return 0;
 572
 573         if (events & (1 << IPCT_DESTROY)) {
 574                 type = IPCTNL_MSG_CT_DELETE;
 575                 group = NFNLGRP_CONNTRACK_DESTROY;
 576         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 577                 type = IPCTNL_MSG_CT_NEW;
 578                 flags = NLM_F_CREATE|NLM_F_EXCL;
 579                 group = NFNLGRP_CONNTRACK_NEW;
 580         } else  if (events) {
 581                 type = IPCTNL_MSG_CT_NEW;
 582                 group = NFNLGRP_CONNTRACK_UPDATE;
 583         } else
 584                 return 0;
 585
 586         net = nf_ct_net(ct);
 587         if (!item->report && !nfnetlink_has_listeners(net, group))
 588                 return 0;
 589
 590         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 591         if (skb == NULL)
 592                 goto errout;
 593
 594         type |= NFNL_SUBSYS_CTNETLINK << 8;
 595         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
 596         if (nlh == NULL)
 597                 goto nlmsg_failure;
 598
 599         nfmsg = nlmsg_data(nlh);
 600         nfmsg->nfgen_family = nf_ct_l3num(ct);
 601         nfmsg->version  = NFNETLINK_V0;
 602         nfmsg->res_id   = 0;
 603
 604         rcu_read_lock();
 605         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 606         if (!nest_parms)
 607                 goto nla_put_failure;
 608         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 609                 goto nla_put_failure;
 610         nla_nest_end(skb, nest_parms);
 611
 612         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 613         if (!nest_parms)
 614                 goto nla_put_failure;
 615         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 616                 goto nla_put_failure;
 617         nla_nest_end(skb, nest_parms);
 618
 619         if (nf_ct_zone(ct))
 620                 NLA_PUT_BE16(skb, CTA_ZONE, htons(nf_ct_zone(ct)));
 621
 622         if (ctnetlink_dump_id(skb, ct) < 0)
 623                 goto nla_put_failure;
 624
 625         if (ctnetlink_dump_status(skb, ct) < 0)
 626                 goto nla_put_failure;
 627
 628         if (events & (1 << IPCT_DESTROY)) {
 629                 if (ctnetlink_dump_counters(skb, ct,
 630                                             IP_CT_DIR_ORIGINAL, type) < 0 ||
 631                     ctnetlink_dump_counters(skb, ct,
 632                                             IP_CT_DIR_REPLY, type) < 0 ||
 633                     ctnetlink_dump_timestamp(skb, ct) < 0)
 634                         goto nla_put_failure;
 635         } else {
 636                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 637                         goto nla_put_failure;
 638
 639                 if (events & (1 << IPCT_PROTOINFO)
 640                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 641                         goto nla_put_failure;
 642
 643                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 644                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 645                         goto nla_put_failure;
 646
 647 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 648                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 649                     && ctnetlink_dump_secctx(skb, ct) < 0)
 650                         goto nla_put_failure;
 651 #endif
 652
 653                 if (events & (1 << IPCT_RELATED) &&
 654                     ctnetlink_dump_master(skb, ct) < 0)
 655                         goto nla_put_failure;
 656
 657                 if (events & (1 << IPCT_NATSEQADJ) &&
 658                     ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 659                         goto nla_put_failure;
 660         }
 661
 662 #ifdef CONFIG_NF_CONNTRACK_MARK
 663         if ((events & (1 << IPCT_MARK) || ct->mark)
 664             && ctnetlink_dump_mark(skb, ct) < 0)
 665                 goto nla_put_failure;
 666 #endif
 667         rcu_read_unlock();
 668
 669         nlmsg_end(skb, nlh);
 670         err = nfnetlink_send(skb, net, item->pid, group, item->report,
 671                              GFP_ATOMIC);
 672         if (err == -ENOBUFS || err == -EAGAIN)
 673                 return -ENOBUFS;
 674
 675         return 0;
 676
 677 nla_put_failure:
 678         rcu_read_unlock();
 679         nlmsg_cancel(skb, nlh);
 680 nlmsg_failure:
 681         kfree_skb(skb);
 682 errout:
 683         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 684                 return -ENOBUFS;
 685
 686         return 0;
 687 }
 688 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 689
 690 static int ctnetlink_done(struct netlink_callback *cb)
 691 {
 692         if (cb->args[1])
 693                 nf_ct_put((struct nf_conn *)cb->args[1]);
 694         return 0;
 695 }
 696
 697 static int
 698 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 699 {
 700         struct net *net = sock_net(skb->sk);
 701         struct nf_conn *ct, *last;
 702         struct nf_conntrack_tuple_hash *h;
 703         struct hlist_nulls_node *n;
 704         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 705         u_int8_t l3proto = nfmsg->nfgen_family;
 706
 707         spin_lock_bh(&nf_conntrack_lock);
 708         last = (struct nf_conn *)cb->args[1];
 709         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 710 restart:
 711                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 712                                          hnnode) {
 713                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 714                                 continue;
 715                         ct = nf_ct_tuplehash_to_ctrack(h);
 716                         /* Dump entries of a given L3 protocol number.
 717                          * If it is not specified, ie. l3proto == 0,
 718                          * then dump everything. */
 719                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 720                                 continue;
 721                         if (cb->args[1]) {
 722                                 if (ct != last)
 723                                         continue;
 724                                 cb->args[1] = 0;
 725                         }
 726                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 727                                                 cb->nlh->nlmsg_seq,
 728                                                 NFNL_MSG_TYPE(
 729                                                         cb->nlh->nlmsg_type),
 730                                                 ct) < 0) {
 731                                 nf_conntrack_get(&ct->ct_general);
 732                                 cb->args[1] = (unsigned long)ct;
 733                                 goto out;
 734                         }
 735                 }
 736                 if (cb->args[1]) {
 737                         cb->args[1] = 0;
 738                         goto restart;
 739                 }
 740         }
 741 out:
 742         spin_unlock_bh(&nf_conntrack_lock);
 743         if (last)
 744                 nf_ct_put(last);
 745
 746         return skb->len;
 747 }
 748
 749 static inline int
 750 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 751 {
 752         struct nlattr *tb[CTA_IP_MAX+1];
 753         struct nf_conntrack_l3proto *l3proto;
 754         int ret = 0;
 755
 756         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 757
 758         rcu_read_lock();
 759         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 760
 761         if (likely(l3proto->nlattr_to_tuple)) {
 762                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 763                                           l3proto->nla_policy);
 764                 if (ret == 0)
 765                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 766         }
 767
 768         rcu_read_unlock();
 769
 770         return ret;
 771 }
 772
 773 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 774         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 775 };
 776
 777 static inline int
 778 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 779                             struct nf_conntrack_tuple *tuple)
 780 {
 781         struct nlattr *tb[CTA_PROTO_MAX+1];
 782         struct nf_conntrack_l4proto *l4proto;
 783         int ret = 0;
 784
 785         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 786         if (ret < 0)
 787                 return ret;
 788
 789         if (!tb[CTA_PROTO_NUM])
 790                 return -EINVAL;
 791         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 792
 793         rcu_read_lock();
 794         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 795
 796         if (likely(l4proto->nlattr_to_tuple)) {
 797                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 798                                           l4proto->nla_policy);
 799                 if (ret == 0)
 800                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 801         }
 802
 803         rcu_read_unlock();
 804
 805         return ret;
 806 }
 807
 808 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 809         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 810         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 811 };
 812
 813 static int
 814 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 815                       struct nf_conntrack_tuple *tuple,
 816                       enum ctattr_type type, u_int8_t l3num)
 817 {
 818         struct nlattr *tb[CTA_TUPLE_MAX+1];
 819         int err;
 820
 821         memset(tuple, 0, sizeof(*tuple));
 822
 823         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 824
 825         if (!tb[CTA_TUPLE_IP])
 826                 return -EINVAL;
 827
 828         tuple->src.l3num = l3num;
 829
 830         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 831         if (err < 0)
 832                 return err;
 833
 834         if (!tb[CTA_TUPLE_PROTO])
 835                 return -EINVAL;
 836
 837         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 838         if (err < 0)
 839                 return err;
 840
 841         /* orig and expect tuples get DIR_ORIGINAL */
 842         if (type == CTA_TUPLE_REPLY)
 843                 tuple->dst.dir = IP_CT_DIR_REPLY;
 844         else
 845                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 846
 847         return 0;
 848 }
 849
 850 static int
 851 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 852 {
 853         if (attr)
 854 #ifdef CONFIG_NF_CONNTRACK_ZONES
 855                 *zone = ntohs(nla_get_be16(attr));
 856 #else
 857                 return -EOPNOTSUPP;
 858 #endif
 859         else
 860                 *zone = 0;
 861
 862         return 0;
 863 }
 864
 865 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 866         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING },
 867 };
 868
 869 static inline int
 870 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name)
 871 {
 872         struct nlattr *tb[CTA_HELP_MAX+1];
 873
 874         nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 875
 876         if (!tb[CTA_HELP_NAME])
 877                 return -EINVAL;
 878
 879         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 880
 881         return 0;
 882 }
 883
 884 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 885         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 886         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 887         [CTA_STATUS]            = { .type = NLA_U32 },
 888         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 889         [CTA_HELP]              = { .type = NLA_NESTED },
 890         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 891         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 892         [CTA_MARK]              = { .type = NLA_U32 },
 893         [CTA_ID]                = { .type = NLA_U32 },
 894         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 895         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 896         [CTA_ZONE]              = { .type = NLA_U16 },
 897 };
 898
 899 static int
 900 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 901                         const struct nlmsghdr *nlh,
 902                         const struct nlattr * const cda[])
 903 {
 904         struct net *net = sock_net(ctnl);
 905         struct nf_conntrack_tuple_hash *h;
 906         struct nf_conntrack_tuple tuple;
 907         struct nf_conn *ct;
 908         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 909         u_int8_t u3 = nfmsg->nfgen_family;
 910         u16 zone;
 911         int err;
 912
 913         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 914         if (err < 0)
 915                 return err;
 916
 917         if (cda[CTA_TUPLE_ORIG])
 918                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 919         else if (cda[CTA_TUPLE_REPLY])
 920                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 921         else {
 922                 /* Flush the whole table */
 923                 nf_conntrack_flush_report(net,
 924                                          NETLINK_CB(skb).pid,
 925                                          nlmsg_report(nlh));
 926                 return 0;
 927         }
 928
 929         if (err < 0)
 930                 return err;
 931
 932         h = nf_conntrack_find_get(net, zone, &tuple);
 933         if (!h)
 934                 return -ENOENT;
 935
 936         ct = nf_ct_tuplehash_to_ctrack(h);
 937
 938         if (cda[CTA_ID]) {
 939                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
 940                 if (id != (u32)(unsigned long)ct) {
 941                         nf_ct_put(ct);
 942                         return -ENOENT;
 943                 }
 944         }
 945
 946         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 947                                       NETLINK_CB(skb).pid,
 948                                       nlmsg_report(nlh)) < 0) {
 949                 nf_ct_delete_from_lists(ct);
 950                 /* we failed to report the event, try later */
 951                 nf_ct_insert_dying_list(ct);
 952                 nf_ct_put(ct);
 953                 return 0;
 954         }
 955
 956         /* death_by_timeout would report the event again */
 957         set_bit(IPS_DYING_BIT, &ct->status);
 958
 959         nf_ct_kill(ct);
 960         nf_ct_put(ct);
 961
 962         return 0;
 963 }
 964
 965 static int
 966 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 967                         const struct nlmsghdr *nlh,
 968                         const struct nlattr * const cda[])
 969 {
 970         struct net *net = sock_net(ctnl);
 971         struct nf_conntrack_tuple_hash *h;
 972         struct nf_conntrack_tuple tuple;
 973         struct nf_conn *ct;
 974         struct sk_buff *skb2 = NULL;
 975         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 976         u_int8_t u3 = nfmsg->nfgen_family;
 977         u16 zone;
 978         int err;
 979
 980         if (nlh->nlmsg_flags & NLM_F_DUMP)
 981                 return netlink_dump_start(ctnl, skb, nlh, ctnetlink_dump_table,
 982                                           ctnetlink_done, 0);
 983
 984         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 985         if (err < 0)
 986                 return err;
 987
 988         if (cda[CTA_TUPLE_ORIG])
 989                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 990         else if (cda[CTA_TUPLE_REPLY])
 991                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 992         else
 993                 return -EINVAL;
 994
 995         if (err < 0)
 996                 return err;
 997
 998         h = nf_conntrack_find_get(net, zone, &tuple);
 999         if (!h)
1000                 return -ENOENT;
1001
1002         ct = nf_ct_tuplehash_to_ctrack(h);
1003
1004         err = -ENOMEM;
1005         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1006         if (skb2 == NULL) {
1007                 nf_ct_put(ct);
1008                 return -ENOMEM;
1009         }
1010
1011         rcu_read_lock();
1012         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
1013                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1014         rcu_read_unlock();
1015         nf_ct_put(ct);
1016         if (err <= 0)
1017                 goto free;
1018
1019         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1020         if (err < 0)
1021                 goto out;
1022
1023         return 0;
1024
1025 free:
1026         kfree_skb(skb2);
1027 out:
1028         /* this avoids a loop in nfnetlink. */
1029         return err == -EAGAIN ? -ENOBUFS : err;
1030 }
1031
1032 #ifdef CONFIG_NF_NAT_NEEDED
1033 static int
1034 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1035                           enum nf_nat_manip_type manip,
1036                           const struct nlattr *attr)
1037 {
1038         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1039
1040         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1041         if (!parse_nat_setup) {
1042 #ifdef CONFIG_MODULES
1043                 rcu_read_unlock();
1044                 nfnl_unlock();
1045                 if (request_module("nf-nat-ipv4") < 0) {
1046                         nfnl_lock();
1047                         rcu_read_lock();
1048                         return -EOPNOTSUPP;
1049                 }
1050                 nfnl_lock();
1051                 rcu_read_lock();
1052                 if (nfnetlink_parse_nat_setup_hook)
1053                         return -EAGAIN;
1054 #endif
1055                 return -EOPNOTSUPP;
1056         }
1057
1058         return parse_nat_setup(ct, manip, attr);
1059 }
1060 #endif
1061
1062 static int
1063 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1064 {
1065         unsigned long d;
1066         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1067         d = ct->status ^ status;
1068
1069         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1070                 /* unchangeable */
1071                 return -EBUSY;
1072
1073         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1074                 /* SEEN_REPLY bit can only be set */
1075                 return -EBUSY;
1076
1077         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1078                 /* ASSURED bit can only be set */
1079                 return -EBUSY;
1080
1081         /* Be careful here, modifying NAT bits can screw up things,
1082          * so don't let users modify them directly if they don't pass
1083          * nf_nat_range. */
1084         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1085         return 0;
1086 }
1087
1088 static int
1089 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1090 {
1091 #ifdef CONFIG_NF_NAT_NEEDED
1092         int ret;
1093
1094         if (cda[CTA_NAT_DST]) {
1095                 ret = ctnetlink_parse_nat_setup(ct,
1096                                                 NF_NAT_MANIP_DST,
1097                                                 cda[CTA_NAT_DST]);
1098                 if (ret < 0)
1099                         return ret;
1100         }
1101         if (cda[CTA_NAT_SRC]) {
1102                 ret = ctnetlink_parse_nat_setup(ct,
1103                                                 NF_NAT_MANIP_SRC,
1104                                                 cda[CTA_NAT_SRC]);
1105                 if (ret < 0)
1106                         return ret;
1107         }
1108         return 0;
1109 #else
1110         return -EOPNOTSUPP;
1111 #endif
1112 }
1113
1114 static inline int
1115 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1116 {
1117         struct nf_conntrack_helper *helper;
1118         struct nf_conn_help *help = nfct_help(ct);
1119         char *helpname = NULL;
1120         int err;
1121
1122         /* don't change helper of sibling connections */
1123         if (ct->master)
1124                 return -EBUSY;
1125
1126         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1127         if (err < 0)
1128                 return err;
1129
1130         if (!strcmp(helpname, "")) {
1131                 if (help && help->helper) {
1132                         /* we had a helper before ... */
1133                         nf_ct_remove_expectations(ct);
1134                         RCU_INIT_POINTER(help->helper, NULL);
1135                 }
1136
1137                 return 0;
1138         }
1139
1140         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1141                                             nf_ct_protonum(ct));
1142         if (helper == NULL) {
1143 #ifdef CONFIG_MODULES
1144                 spin_unlock_bh(&nf_conntrack_lock);
1145
1146                 if (request_module("nfct-helper-%s", helpname) < 0) {
1147                         spin_lock_bh(&nf_conntrack_lock);
1148                         return -EOPNOTSUPP;
1149                 }
1150
1151                 spin_lock_bh(&nf_conntrack_lock);
1152                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1153                                                     nf_ct_protonum(ct));
1154                 if (helper)
1155                         return -EAGAIN;
1156 #endif
1157                 return -EOPNOTSUPP;
1158         }
1159
1160         if (help) {
1161                 if (help->helper == helper)
1162                         return 0;
1163                 if (help->helper)
1164                         return -EBUSY;
1165                 /* need to zero data of old helper */
1166                 memset(&help->help, 0, sizeof(help->help));
1167         } else {
1168                 /* we cannot set a helper for an existing conntrack */
1169                 return -EOPNOTSUPP;
1170         }
1171
1172         rcu_assign_pointer(help->helper, helper);
1173
1174         return 0;
1175 }
1176
1177 static inline int
1178 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1179 {
1180         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1181
1182         if (!del_timer(&ct->timeout))
1183                 return -ETIME;
1184
1185         ct->timeout.expires = jiffies + timeout * HZ;
1186         add_timer(&ct->timeout);
1187
1188         return 0;
1189 }
1190
1191 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1192         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1193         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1194         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1195 };
1196
1197 static inline int
1198 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1199 {
1200         const struct nlattr *attr = cda[CTA_PROTOINFO];
1201         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1202         struct nf_conntrack_l4proto *l4proto;
1203         int err = 0;
1204
1205         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1206
1207         rcu_read_lock();
1208         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1209         if (l4proto->from_nlattr)
1210                 err = l4proto->from_nlattr(tb, ct);
1211         rcu_read_unlock();
1212
1213         return err;
1214 }
1215
1216 #ifdef CONFIG_NF_NAT_NEEDED
1217 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1218         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1219         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1220         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1221 };
1222
1223 static inline int
1224 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1225 {
1226         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1227
1228         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1229
1230         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1231                 return -EINVAL;
1232
1233         natseq->correction_pos =
1234                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1235
1236         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1237                 return -EINVAL;
1238
1239         natseq->offset_before =
1240                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1241
1242         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1243                 return -EINVAL;
1244
1245         natseq->offset_after =
1246                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1247
1248         return 0;
1249 }
1250
1251 static int
1252 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1253                              const struct nlattr * const cda[])
1254 {
1255         int ret = 0;
1256         struct nf_conn_nat *nat = nfct_nat(ct);
1257
1258         if (!nat)
1259                 return 0;
1260
1261         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1262                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1263                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1264                 if (ret < 0)
1265                         return ret;
1266
1267                 ct->status |= IPS_SEQ_ADJUST;
1268         }
1269
1270         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1271                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1272                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1273                 if (ret < 0)
1274                         return ret;
1275
1276                 ct->status |= IPS_SEQ_ADJUST;
1277         }
1278
1279         return 0;
1280 }
1281 #endif
1282
1283 static int
1284 ctnetlink_change_conntrack(struct nf_conn *ct,
1285                            const struct nlattr * const cda[])
1286 {
1287         int err;
1288
1289         /* only allow NAT changes and master assignation for new conntracks */
1290         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1291                 return -EOPNOTSUPP;
1292
1293         if (cda[CTA_HELP]) {
1294                 err = ctnetlink_change_helper(ct, cda);
1295                 if (err < 0)
1296                         return err;
1297         }
1298
1299         if (cda[CTA_TIMEOUT]) {
1300                 err = ctnetlink_change_timeout(ct, cda);
1301                 if (err < 0)
1302                         return err;
1303         }
1304
1305         if (cda[CTA_STATUS]) {
1306                 err = ctnetlink_change_status(ct, cda);
1307                 if (err < 0)
1308                         return err;
1309         }
1310
1311         if (cda[CTA_PROTOINFO]) {
1312                 err = ctnetlink_change_protoinfo(ct, cda);
1313                 if (err < 0)
1314                         return err;
1315         }
1316
1317 #if defined(CONFIG_NF_CONNTRACK_MARK)
1318         if (cda[CTA_MARK])
1319                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1320 #endif
1321
1322 #ifdef CONFIG_NF_NAT_NEEDED
1323         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1324                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1325                 if (err < 0)
1326                         return err;
1327         }
1328 #endif
1329
1330         return 0;
1331 }
1332
1333 static struct nf_conn *
1334 ctnetlink_create_conntrack(struct net *net, u16 zone,
1335                            const struct nlattr * const cda[],
1336                            struct nf_conntrack_tuple *otuple,
1337                            struct nf_conntrack_tuple *rtuple,
1338                            u8 u3)
1339 {
1340         struct nf_conn *ct;
1341         int err = -EINVAL;
1342         struct nf_conntrack_helper *helper;
1343         struct nf_conn_tstamp *tstamp;
1344
1345         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1346         if (IS_ERR(ct))
1347                 return ERR_PTR(-ENOMEM);
1348
1349         if (!cda[CTA_TIMEOUT])
1350                 goto err1;
1351         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1352
1353         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1354
1355         rcu_read_lock();
1356         if (cda[CTA_HELP]) {
1357                 char *helpname = NULL;
1358
1359                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1360                 if (err < 0)
1361                         goto err2;
1362
1363                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1364                                                     nf_ct_protonum(ct));
1365                 if (helper == NULL) {
1366                         rcu_read_unlock();
1367 #ifdef CONFIG_MODULES
1368                         if (request_module("nfct-helper-%s", helpname) < 0) {
1369                                 err = -EOPNOTSUPP;
1370                                 goto err1;
1371                         }
1372
1373                         rcu_read_lock();
1374                         helper = __nf_conntrack_helper_find(helpname,
1375                                                             nf_ct_l3num(ct),
1376                                                             nf_ct_protonum(ct));
1377                         if (helper) {
1378                                 err = -EAGAIN;
1379                                 goto err2;
1380                         }
1381                         rcu_read_unlock();
1382 #endif
1383                         err = -EOPNOTSUPP;
1384                         goto err1;
1385                 } else {
1386                         struct nf_conn_help *help;
1387
1388                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1389                         if (help == NULL) {
1390                                 err = -ENOMEM;
1391                                 goto err2;
1392                         }
1393
1394                         /* not in hash table yet so not strictly necessary */
1395                         RCU_INIT_POINTER(help->helper, helper);
1396                 }
1397         } else {
1398                 /* try an implicit helper assignation */
1399                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1400                 if (err < 0)
1401                         goto err2;
1402         }
1403
1404         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1405                 err = ctnetlink_change_nat(ct, cda);
1406                 if (err < 0)
1407                         goto err2;
1408         }
1409
1410         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1411         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1412         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1413         /* we must add conntrack extensions before confirmation. */
1414         ct->status |= IPS_CONFIRMED;
1415
1416         if (cda[CTA_STATUS]) {
1417                 err = ctnetlink_change_status(ct, cda);
1418                 if (err < 0)
1419                         goto err2;
1420         }
1421
1422 #ifdef CONFIG_NF_NAT_NEEDED
1423         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1424                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1425                 if (err < 0)
1426                         goto err2;
1427         }
1428 #endif
1429
1430         memset(&ct->proto, 0, sizeof(ct->proto));
1431         if (cda[CTA_PROTOINFO]) {
1432                 err = ctnetlink_change_protoinfo(ct, cda);
1433                 if (err < 0)
1434                         goto err2;
1435         }
1436
1437 #if defined(CONFIG_NF_CONNTRACK_MARK)
1438         if (cda[CTA_MARK])
1439                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1440 #endif
1441
1442         /* setup master conntrack: this is a confirmed expectation */
1443         if (cda[CTA_TUPLE_MASTER]) {
1444                 struct nf_conntrack_tuple master;
1445                 struct nf_conntrack_tuple_hash *master_h;
1446                 struct nf_conn *master_ct;
1447
1448                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1449                 if (err < 0)
1450                         goto err2;
1451
1452                 master_h = nf_conntrack_find_get(net, zone, &master);
1453                 if (master_h == NULL) {
1454                         err = -ENOENT;
1455                         goto err2;
1456                 }
1457                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1458                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1459                 ct->master = master_ct;
1460         }
1461         tstamp = nf_conn_tstamp_find(ct);
1462         if (tstamp)
1463                 tstamp->start = ktime_to_ns(ktime_get_real());
1464
1465         err = nf_conntrack_hash_check_insert(ct);
1466         if (err < 0)
1467                 goto err2;
1468
1469         rcu_read_unlock();
1470
1471         return ct;
1472
1473 err2:
1474         rcu_read_unlock();
1475 err1:
1476         nf_conntrack_free(ct);
1477         return ERR_PTR(err);
1478 }
1479
1480 static int
1481 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1482                         const struct nlmsghdr *nlh,
1483                         const struct nlattr * const cda[])
1484 {
1485         struct net *net = sock_net(ctnl);
1486         struct nf_conntrack_tuple otuple, rtuple;
1487         struct nf_conntrack_tuple_hash *h = NULL;
1488         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1489         struct nf_conn *ct;
1490         u_int8_t u3 = nfmsg->nfgen_family;
1491         u16 zone;
1492         int err;
1493
1494         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1495         if (err < 0)
1496                 return err;
1497
1498         if (cda[CTA_TUPLE_ORIG]) {
1499                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1500                 if (err < 0)
1501                         return err;
1502         }
1503
1504         if (cda[CTA_TUPLE_REPLY]) {
1505                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1506                 if (err < 0)
1507                         return err;
1508         }
1509
1510         if (cda[CTA_TUPLE_ORIG])
1511                 h = nf_conntrack_find_get(net, zone, &otuple);
1512         else if (cda[CTA_TUPLE_REPLY])
1513                 h = nf_conntrack_find_get(net, zone, &rtuple);
1514
1515         if (h == NULL) {
1516                 err = -ENOENT;
1517                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1518                         enum ip_conntrack_events events;
1519
1520                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1521                                                         &rtuple, u3);
1522                         if (IS_ERR(ct))
1523                                 return PTR_ERR(ct);
1524
1525                         err = 0;
1526                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1527                                 events = IPCT_RELATED;
1528                         else
1529                                 events = IPCT_NEW;
1530
1531                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1532                                                       (1 << IPCT_ASSURED) |
1533                                                       (1 << IPCT_HELPER) |
1534                                                       (1 << IPCT_PROTOINFO) |
1535                                                       (1 << IPCT_NATSEQADJ) |
1536                                                       (1 << IPCT_MARK) | events,
1537                                                       ct, NETLINK_CB(skb).pid,
1538                                                       nlmsg_report(nlh));
1539                         nf_ct_put(ct);
1540                 }
1541
1542                 return err;
1543         }
1544         /* implicit 'else' */
1545
1546         err = -EEXIST;
1547         ct = nf_ct_tuplehash_to_ctrack(h);
1548         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1549                 spin_lock_bh(&nf_conntrack_lock);
1550                 err = ctnetlink_change_conntrack(ct, cda);
1551                 spin_unlock_bh(&nf_conntrack_lock);
1552                 if (err == 0) {
1553                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1554                                                       (1 << IPCT_ASSURED) |
1555                                                       (1 << IPCT_HELPER) |
1556                                                       (1 << IPCT_PROTOINFO) |
1557                                                       (1 << IPCT_NATSEQADJ) |
1558                                                       (1 << IPCT_MARK),
1559                                                       ct, NETLINK_CB(skb).pid,
1560                                                       nlmsg_report(nlh));
1561                 }
1562         }
1563
1564         nf_ct_put(ct);
1565         return err;
1566 }
1567
1568 /***********************************************************************
1569  * EXPECT
1570  ***********************************************************************/
1571
1572 static inline int
1573 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1574                          const struct nf_conntrack_tuple *tuple,
1575                          enum ctattr_expect type)
1576 {
1577         struct nlattr *nest_parms;
1578
1579         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
1580         if (!nest_parms)
1581                 goto nla_put_failure;
1582         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1583                 goto nla_put_failure;
1584         nla_nest_end(skb, nest_parms);
1585
1586         return 0;
1587
1588 nla_put_failure:
1589         return -1;
1590 }
1591
1592 static inline int
1593 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1594                         const struct nf_conntrack_tuple *tuple,
1595                         const struct nf_conntrack_tuple_mask *mask)
1596 {
1597         int ret;
1598         struct nf_conntrack_l3proto *l3proto;
1599         struct nf_conntrack_l4proto *l4proto;
1600         struct nf_conntrack_tuple m;
1601         struct nlattr *nest_parms;
1602
1603         memset(&m, 0xFF, sizeof(m));
1604         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1605         m.src.u.all = mask->src.u.all;
1606         m.dst.protonum = tuple->dst.protonum;
1607
1608         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
1609         if (!nest_parms)
1610                 goto nla_put_failure;
1611
1612         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
1613         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1614
1615         if (unlikely(ret < 0))
1616                 goto nla_put_failure;
1617
1618         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
1619         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1620         if (unlikely(ret < 0))
1621                 goto nla_put_failure;
1622
1623         nla_nest_end(skb, nest_parms);
1624
1625         return 0;
1626
1627 nla_put_failure:
1628         return -1;
1629 }
1630
1631 static int
1632 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1633                           const struct nf_conntrack_expect *exp)
1634 {
1635         struct nf_conn *master = exp->master;
1636         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
1637         struct nf_conn_help *help;
1638
1639         if (timeout < 0)
1640                 timeout = 0;
1641
1642         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1643                 goto nla_put_failure;
1644         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1645                 goto nla_put_failure;
1646         if (ctnetlink_exp_dump_tuple(skb,
1647                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1648                                  CTA_EXPECT_MASTER) < 0)
1649                 goto nla_put_failure;
1650
1651         NLA_PUT_BE32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout));
1652         NLA_PUT_BE32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp));
1653         NLA_PUT_BE32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags));
1654         help = nfct_help(master);
1655         if (help) {
1656                 struct nf_conntrack_helper *helper;
1657
1658                 helper = rcu_dereference(help->helper);
1659                 if (helper)
1660                         NLA_PUT_STRING(skb, CTA_EXPECT_HELP_NAME, helper->name);
1661         }
1662
1663         return 0;
1664
1665 nla_put_failure:
1666         return -1;
1667 }
1668
1669 static int
1670 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1671                         int event, const struct nf_conntrack_expect *exp)
1672 {
1673         struct nlmsghdr *nlh;
1674         struct nfgenmsg *nfmsg;
1675         unsigned int flags = pid ? NLM_F_MULTI : 0;
1676
1677         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1678         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
1679         if (nlh == NULL)
1680                 goto nlmsg_failure;
1681
1682         nfmsg = nlmsg_data(nlh);
1683         nfmsg->nfgen_family = exp->tuple.src.l3num;
1684         nfmsg->version      = NFNETLINK_V0;
1685         nfmsg->res_id       = 0;
1686
1687         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1688                 goto nla_put_failure;
1689
1690         nlmsg_end(skb, nlh);
1691         return skb->len;
1692
1693 nlmsg_failure:
1694 nla_put_failure:
1695         nlmsg_cancel(skb, nlh);
1696         return -1;
1697 }
1698
1699 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1700 static int
1701 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
1702 {
1703         struct nf_conntrack_expect *exp = item->exp;
1704         struct net *net = nf_ct_exp_net(exp);
1705         struct nlmsghdr *nlh;
1706         struct nfgenmsg *nfmsg;
1707         struct sk_buff *skb;
1708         unsigned int type, group;
1709         int flags = 0;
1710
1711         if (events & (1 << IPEXP_DESTROY)) {
1712                 type = IPCTNL_MSG_EXP_DELETE;
1713                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
1714         } else if (events & (1 << IPEXP_NEW)) {
1715                 type = IPCTNL_MSG_EXP_NEW;
1716                 flags = NLM_F_CREATE|NLM_F_EXCL;
1717                 group = NFNLGRP_CONNTRACK_EXP_NEW;
1718         } else
1719                 return 0;
1720
1721         if (!item->report && !nfnetlink_has_listeners(net, group))
1722                 return 0;
1723
1724         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
1725         if (skb == NULL)
1726                 goto errout;
1727
1728         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1729         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
1730         if (nlh == NULL)
1731                 goto nlmsg_failure;
1732
1733         nfmsg = nlmsg_data(nlh);
1734         nfmsg->nfgen_family = exp->tuple.src.l3num;
1735         nfmsg->version      = NFNETLINK_V0;
1736         nfmsg->res_id       = 0;
1737
1738         rcu_read_lock();
1739         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1740                 goto nla_put_failure;
1741         rcu_read_unlock();
1742
1743         nlmsg_end(skb, nlh);
1744         nfnetlink_send(skb, net, item->pid, group, item->report, GFP_ATOMIC);
1745         return 0;
1746
1747 nla_put_failure:
1748         rcu_read_unlock();
1749         nlmsg_cancel(skb, nlh);
1750 nlmsg_failure:
1751         kfree_skb(skb);
1752 errout:
1753         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
1754         return 0;
1755 }
1756 #endif
1757 static int ctnetlink_exp_done(struct netlink_callback *cb)
1758 {
1759         if (cb->args[1])
1760                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1761         return 0;
1762 }
1763
1764 static int
1765 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1766 {
1767         struct net *net = sock_net(skb->sk);
1768         struct nf_conntrack_expect *exp, *last;
1769         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1770         struct hlist_node *n;
1771         u_int8_t l3proto = nfmsg->nfgen_family;
1772
1773         rcu_read_lock();
1774         last = (struct nf_conntrack_expect *)cb->args[1];
1775         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1776 restart:
1777                 hlist_for_each_entry(exp, n, &net->ct.expect_hash[cb->args[0]],
1778                                      hnode) {
1779                         if (l3proto && exp->tuple.src.l3num != l3proto)
1780                                 continue;
1781                         if (cb->args[1]) {
1782                                 if (exp != last)
1783                                         continue;
1784                                 cb->args[1] = 0;
1785                         }
1786                         if (ctnetlink_exp_fill_info(skb,
1787                                                     NETLINK_CB(cb->skb).pid,
1788                                                     cb->nlh->nlmsg_seq,
1789                                                     IPCTNL_MSG_EXP_NEW,
1790                                                     exp) < 0) {
1791                                 if (!atomic_inc_not_zero(&exp->use))
1792                                         continue;
1793                                 cb->args[1] = (unsigned long)exp;
1794                                 goto out;
1795                         }
1796                 }
1797                 if (cb->args[1]) {
1798                         cb->args[1] = 0;
1799                         goto restart;
1800                 }
1801         }
1802 out:
1803         rcu_read_unlock();
1804         if (last)
1805                 nf_ct_expect_put(last);
1806
1807         return skb->len;
1808 }
1809
1810 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1811         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1812         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1813         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1814         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1815         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1816         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING },
1817         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1818         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1819 };
1820
1821 static int
1822 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1823                      const struct nlmsghdr *nlh,
1824                      const struct nlattr * const cda[])
1825 {
1826         struct net *net = sock_net(ctnl);
1827         struct nf_conntrack_tuple tuple;
1828         struct nf_conntrack_expect *exp;
1829         struct sk_buff *skb2;
1830         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1831         u_int8_t u3 = nfmsg->nfgen_family;
1832         u16 zone;
1833         int err;
1834
1835         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1836                 return netlink_dump_start(ctnl, skb, nlh,
1837                                           ctnetlink_exp_dump_table,
1838                                           ctnetlink_exp_done, 0);
1839         }
1840
1841         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1842         if (err < 0)
1843                 return err;
1844
1845         if (cda[CTA_EXPECT_TUPLE])
1846                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1847         else if (cda[CTA_EXPECT_MASTER])
1848                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1849         else
1850                 return -EINVAL;
1851
1852         if (err < 0)
1853                 return err;
1854
1855         exp = nf_ct_expect_find_get(net, zone, &tuple);
1856         if (!exp)
1857                 return -ENOENT;
1858
1859         if (cda[CTA_EXPECT_ID]) {
1860                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1861                 if (ntohl(id) != (u32)(unsigned long)exp) {
1862                         nf_ct_expect_put(exp);
1863                         return -ENOENT;
1864                 }
1865         }
1866
1867         err = -ENOMEM;
1868         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1869         if (skb2 == NULL) {
1870                 nf_ct_expect_put(exp);
1871                 goto out;
1872         }
1873
1874         rcu_read_lock();
1875         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1876                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
1877         rcu_read_unlock();
1878         nf_ct_expect_put(exp);
1879         if (err <= 0)
1880                 goto free;
1881
1882         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1883         if (err < 0)
1884                 goto out;
1885
1886         return 0;
1887
1888 free:
1889         kfree_skb(skb2);
1890 out:
1891         /* this avoids a loop in nfnetlink. */
1892         return err == -EAGAIN ? -ENOBUFS : err;
1893 }
1894
1895 static int
1896 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1897                      const struct nlmsghdr *nlh,
1898                      const struct nlattr * const cda[])
1899 {
1900         struct net *net = sock_net(ctnl);
1901         struct nf_conntrack_expect *exp;
1902         struct nf_conntrack_tuple tuple;
1903         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1904         struct hlist_node *n, *next;
1905         u_int8_t u3 = nfmsg->nfgen_family;
1906         unsigned int i;
1907         u16 zone;
1908         int err;
1909
1910         if (cda[CTA_EXPECT_TUPLE]) {
1911                 /* delete a single expect by tuple */
1912                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1913                 if (err < 0)
1914                         return err;
1915
1916                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1917                 if (err < 0)
1918                         return err;
1919
1920                 /* bump usage count to 2 */
1921                 exp = nf_ct_expect_find_get(net, zone, &tuple);
1922                 if (!exp)
1923                         return -ENOENT;
1924
1925                 if (cda[CTA_EXPECT_ID]) {
1926                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1927                         if (ntohl(id) != (u32)(unsigned long)exp) {
1928                                 nf_ct_expect_put(exp);
1929                                 return -ENOENT;
1930                         }
1931                 }
1932
1933                 /* after list removal, usage count == 1 */
1934                 spin_lock_bh(&nf_conntrack_lock);
1935                 if (del_timer(&exp->timeout)) {
1936                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).pid,
1937                                                    nlmsg_report(nlh));
1938                         nf_ct_expect_put(exp);
1939                 }
1940                 spin_unlock_bh(&nf_conntrack_lock);
1941                 /* have to put what we 'get' above.
1942                  * after this line usage count == 0 */
1943                 nf_ct_expect_put(exp);
1944         } else if (cda[CTA_EXPECT_HELP_NAME]) {
1945                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
1946                 struct nf_conn_help *m_help;
1947
1948                 /* delete all expectations for this helper */
1949                 spin_lock_bh(&nf_conntrack_lock);
1950                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1951                         hlist_for_each_entry_safe(exp, n, next,
1952                                                   &net->ct.expect_hash[i],
1953                                                   hnode) {
1954                                 m_help = nfct_help(exp->master);
1955                                 if (!strcmp(m_help->helper->name, name) &&
1956                                     del_timer(&exp->timeout)) {
1957                                         nf_ct_unlink_expect_report(exp,
1958                                                         NETLINK_CB(skb).pid,
1959                                                         nlmsg_report(nlh));
1960                                         nf_ct_expect_put(exp);
1961                                 }
1962                         }
1963                 }
1964                 spin_unlock_bh(&nf_conntrack_lock);
1965         } else {
1966                 /* This basically means we have to flush everything*/
1967                 spin_lock_bh(&nf_conntrack_lock);
1968                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1969                         hlist_for_each_entry_safe(exp, n, next,
1970                                                   &net->ct.expect_hash[i],
1971                                                   hnode) {
1972                                 if (del_timer(&exp->timeout)) {
1973                                         nf_ct_unlink_expect_report(exp,
1974                                                         NETLINK_CB(skb).pid,
1975                                                         nlmsg_report(nlh));
1976                                         nf_ct_expect_put(exp);
1977                                 }
1978                         }
1979                 }
1980                 spin_unlock_bh(&nf_conntrack_lock);
1981         }
1982
1983         return 0;
1984 }
1985 static int
1986 ctnetlink_change_expect(struct nf_conntrack_expect *x,
1987                         const struct nlattr * const cda[])
1988 {
1989         return -EOPNOTSUPP;
1990 }
1991
1992 static int
1993 ctnetlink_create_expect(struct net *net, u16 zone,
1994                         const struct nlattr * const cda[],
1995                         u_int8_t u3,
1996                         u32 pid, int report)
1997 {
1998         struct nf_conntrack_tuple tuple, mask, master_tuple;
1999         struct nf_conntrack_tuple_hash *h = NULL;
2000         struct nf_conntrack_expect *exp;
2001         struct nf_conn *ct;
2002         struct nf_conn_help *help;
2003         int err = 0;
2004
2005         /* caller guarantees that those three CTA_EXPECT_* exist */
2006         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2007         if (err < 0)
2008                 return err;
2009         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2010         if (err < 0)
2011                 return err;
2012         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2013         if (err < 0)
2014                 return err;
2015
2016         /* Look for master conntrack of this expectation */
2017         h = nf_conntrack_find_get(net, zone, &master_tuple);
2018         if (!h)
2019                 return -ENOENT;
2020         ct = nf_ct_tuplehash_to_ctrack(h);
2021         exp = nf_ct_expect_alloc(ct);
2022         if (!exp) {
2023                 err = -ENOMEM;
2024                 goto out;
2025         }
2026         help = nfct_help(ct);
2027         if (!help) {
2028                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2029                         err = -EINVAL;
2030                         goto out;
2031                 }
2032                 exp->timeout.expires =
2033                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2034
2035                 exp->flags = NF_CT_EXPECT_USERSPACE;
2036                 if (cda[CTA_EXPECT_FLAGS]) {
2037                         exp->flags |=
2038                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2039                 }
2040         } else {
2041                 if (cda[CTA_EXPECT_FLAGS]) {
2042                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2043                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2044                 } else
2045                         exp->flags = 0;
2046         }
2047
2048         exp->class = 0;
2049         exp->expectfn = NULL;
2050         exp->master = ct;
2051         exp->helper = NULL;
2052         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2053         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2054         exp->mask.src.u.all = mask.src.u.all;
2055
2056         err = nf_ct_expect_related_report(exp, pid, report);
2057         nf_ct_expect_put(exp);
2058
2059 out:
2060         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2061         return err;
2062 }
2063
2064 static int
2065 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2066                      const struct nlmsghdr *nlh,
2067                      const struct nlattr * const cda[])
2068 {
2069         struct net *net = sock_net(ctnl);
2070         struct nf_conntrack_tuple tuple;
2071         struct nf_conntrack_expect *exp;
2072         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2073         u_int8_t u3 = nfmsg->nfgen_family;
2074         u16 zone;
2075         int err;
2076
2077         if (!cda[CTA_EXPECT_TUPLE]
2078             || !cda[CTA_EXPECT_MASK]
2079             || !cda[CTA_EXPECT_MASTER])
2080                 return -EINVAL;
2081
2082         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2083         if (err < 0)
2084                 return err;
2085
2086         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2087         if (err < 0)
2088                 return err;
2089
2090         spin_lock_bh(&nf_conntrack_lock);
2091         exp = __nf_ct_expect_find(net, zone, &tuple);
2092
2093         if (!exp) {
2094                 spin_unlock_bh(&nf_conntrack_lock);
2095                 err = -ENOENT;
2096                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2097                         err = ctnetlink_create_expect(net, zone, cda,
2098                                                       u3,
2099                                                       NETLINK_CB(skb).pid,
2100                                                       nlmsg_report(nlh));
2101                 }
2102                 return err;
2103         }
2104
2105         err = -EEXIST;
2106         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2107                 err = ctnetlink_change_expect(exp, cda);
2108         spin_unlock_bh(&nf_conntrack_lock);
2109
2110         return err;
2111 }
2112
2113 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2114 static struct nf_ct_event_notifier ctnl_notifier = {
2115         .fcn = ctnetlink_conntrack_event,
2116 };
2117
2118 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2119         .fcn = ctnetlink_expect_event,
2120 };
2121 #endif
2122
2123 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2124         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2125                                             .attr_count = CTA_MAX,
2126                                             .policy = ct_nla_policy },
2127         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2128                                             .attr_count = CTA_MAX,
2129                                             .policy = ct_nla_policy },
2130         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2131                                             .attr_count = CTA_MAX,
2132                                             .policy = ct_nla_policy },
2133         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
2134                                             .attr_count = CTA_MAX,
2135                                             .policy = ct_nla_policy },
2136 };
2137
2138 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
2139         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
2140                                             .attr_count = CTA_EXPECT_MAX,
2141                                             .policy = exp_nla_policy },
2142         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
2143                                             .attr_count = CTA_EXPECT_MAX,
2144                                             .policy = exp_nla_policy },
2145         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
2146                                             .attr_count = CTA_EXPECT_MAX,
2147                                             .policy = exp_nla_policy },
2148 };
2149
2150 static const struct nfnetlink_subsystem ctnl_subsys = {
2151         .name                           = "conntrack",
2152         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
2153         .cb_count                       = IPCTNL_MSG_MAX,
2154         .cb                             = ctnl_cb,
2155 };
2156
2157 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
2158         .name                           = "conntrack_expect",
2159         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
2160         .cb_count                       = IPCTNL_MSG_EXP_MAX,
2161         .cb                             = ctnl_exp_cb,
2162 };
2163
2164 MODULE_ALIAS("ip_conntrack_netlink");
2165 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
2166 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
2167
2168 static int __net_init ctnetlink_net_init(struct net *net)
2169 {
2170 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2171         int ret;
2172
2173         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
2174         if (ret < 0) {
2175                 pr_err("ctnetlink_init: cannot register notifier.\n");
2176                 goto err_out;
2177         }
2178
2179         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
2180         if (ret < 0) {
2181                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
2182                 goto err_unreg_notifier;
2183         }
2184 #endif
2185         return 0;
2186
2187 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2188 err_unreg_notifier:
2189         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2190 err_out:
2191         return ret;
2192 #endif
2193 }
2194
2195 static void ctnetlink_net_exit(struct net *net)
2196 {
2197 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2198         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
2199         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2200 #endif
2201 }
2202
2203 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
2204 {
2205         struct net *net;
2206
2207         list_for_each_entry(net, net_exit_list, exit_list)
2208                 ctnetlink_net_exit(net);
2209 }
2210
2211 static struct pernet_operations ctnetlink_net_ops = {
2212         .init           = ctnetlink_net_init,
2213         .exit_batch     = ctnetlink_net_exit_batch,
2214 };
2215
2216 static int __init ctnetlink_init(void)
2217 {
2218         int ret;
2219
2220         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
2221         ret = nfnetlink_subsys_register(&ctnl_subsys);
2222         if (ret < 0) {
2223                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
2224                 goto err_out;
2225         }
2226
2227         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
2228         if (ret < 0) {
2229                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
2230                 goto err_unreg_subsys;
2231         }
2232
2233         if (register_pernet_subsys(&ctnetlink_net_ops)) {
2234                 pr_err("ctnetlink_init: cannot register pernet operations\n");
2235                 goto err_unreg_exp_subsys;
2236         }
2237
2238         return 0;
2239
2240 err_unreg_exp_subsys:
2241         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2242 err_unreg_subsys:
2243         nfnetlink_subsys_unregister(&ctnl_subsys);
2244 err_out:
2245         return ret;
2246 }
2247
2248 static void __exit ctnetlink_exit(void)
2249 {
2250         pr_info("ctnetlink: unregistering from nfnetlink.\n");
2251
2252         unregister_pernet_subsys(&ctnetlink_net_ops);
2253         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2254         nfnetlink_subsys_unregister(&ctnl_subsys);
2255 }
2256
2257 module_init(ctnetlink_init);
2258 module_exit(ctnetlink_exit);