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         if (cb->data)
 695                 kfree(cb->data);
 696         return 0;
 697 }
 698
 699 struct ctnetlink_dump_filter {
 700         struct {
 701                 u_int32_t val;
 702                 u_int32_t mask;
 703         } mark;
 704 };
 705
 706 static int
 707 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 708 {
 709         struct net *net = sock_net(skb->sk);
 710         struct nf_conn *ct, *last;
 711         struct nf_conntrack_tuple_hash *h;
 712         struct hlist_nulls_node *n;
 713         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 714         u_int8_t l3proto = nfmsg->nfgen_family;
 715 #ifdef CONFIG_NF_CONNTRACK_MARK
 716         const struct ctnetlink_dump_filter *filter = cb->data;
 717 #endif
 718         spin_lock_bh(&nf_conntrack_lock);
 719         last = (struct nf_conn *)cb->args[1];
 720         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 721 restart:
 722                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 723                                          hnnode) {
 724                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 725                                 continue;
 726                         ct = nf_ct_tuplehash_to_ctrack(h);
 727                         /* Dump entries of a given L3 protocol number.
 728                          * If it is not specified, ie. l3proto == 0,
 729                          * then dump everything. */
 730                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 731                                 continue;
 732                         if (cb->args[1]) {
 733                                 if (ct != last)
 734                                         continue;
 735                                 cb->args[1] = 0;
 736                         }
 737 #ifdef CONFIG_NF_CONNTRACK_MARK
 738                         if (filter && !((ct->mark & filter->mark.mask) ==
 739                                         filter->mark.val)) {
 740                                 continue;
 741                         }
 742 #endif
 743                         if (ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).pid,
 744                                                 cb->nlh->nlmsg_seq,
 745                                                 NFNL_MSG_TYPE(
 746                                                         cb->nlh->nlmsg_type),
 747                                                 ct) < 0) {
 748                                 nf_conntrack_get(&ct->ct_general);
 749                                 cb->args[1] = (unsigned long)ct;
 750                                 goto out;
 751                         }
 752                 }
 753                 if (cb->args[1]) {
 754                         cb->args[1] = 0;
 755                         goto restart;
 756                 }
 757         }
 758 out:
 759         spin_unlock_bh(&nf_conntrack_lock);
 760         if (last)
 761                 nf_ct_put(last);
 762
 763         return skb->len;
 764 }
 765
 766 static inline int
 767 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 768 {
 769         struct nlattr *tb[CTA_IP_MAX+1];
 770         struct nf_conntrack_l3proto *l3proto;
 771         int ret = 0;
 772
 773         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 774
 775         rcu_read_lock();
 776         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 777
 778         if (likely(l3proto->nlattr_to_tuple)) {
 779                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 780                                           l3proto->nla_policy);
 781                 if (ret == 0)
 782                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 783         }
 784
 785         rcu_read_unlock();
 786
 787         return ret;
 788 }
 789
 790 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 791         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 792 };
 793
 794 static inline int
 795 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 796                             struct nf_conntrack_tuple *tuple)
 797 {
 798         struct nlattr *tb[CTA_PROTO_MAX+1];
 799         struct nf_conntrack_l4proto *l4proto;
 800         int ret = 0;
 801
 802         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 803         if (ret < 0)
 804                 return ret;
 805
 806         if (!tb[CTA_PROTO_NUM])
 807                 return -EINVAL;
 808         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 809
 810         rcu_read_lock();
 811         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 812
 813         if (likely(l4proto->nlattr_to_tuple)) {
 814                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 815                                           l4proto->nla_policy);
 816                 if (ret == 0)
 817                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 818         }
 819
 820         rcu_read_unlock();
 821
 822         return ret;
 823 }
 824
 825 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 826         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 827         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 828 };
 829
 830 static int
 831 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 832                       struct nf_conntrack_tuple *tuple,
 833                       enum ctattr_type type, u_int8_t l3num)
 834 {
 835         struct nlattr *tb[CTA_TUPLE_MAX+1];
 836         int err;
 837
 838         memset(tuple, 0, sizeof(*tuple));
 839
 840         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 841
 842         if (!tb[CTA_TUPLE_IP])
 843                 return -EINVAL;
 844
 845         tuple->src.l3num = l3num;
 846
 847         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 848         if (err < 0)
 849                 return err;
 850
 851         if (!tb[CTA_TUPLE_PROTO])
 852                 return -EINVAL;
 853
 854         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 855         if (err < 0)
 856                 return err;
 857
 858         /* orig and expect tuples get DIR_ORIGINAL */
 859         if (type == CTA_TUPLE_REPLY)
 860                 tuple->dst.dir = IP_CT_DIR_REPLY;
 861         else
 862                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 863
 864         return 0;
 865 }
 866
 867 static int
 868 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 869 {
 870         if (attr)
 871 #ifdef CONFIG_NF_CONNTRACK_ZONES
 872                 *zone = ntohs(nla_get_be16(attr));
 873 #else
 874                 return -EOPNOTSUPP;
 875 #endif
 876         else
 877                 *zone = 0;
 878
 879         return 0;
 880 }
 881
 882 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 883         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING },
 884 };
 885
 886 static inline int
 887 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name)
 888 {
 889         struct nlattr *tb[CTA_HELP_MAX+1];
 890
 891         nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 892
 893         if (!tb[CTA_HELP_NAME])
 894                 return -EINVAL;
 895
 896         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 897
 898         return 0;
 899 }
 900
 901 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 902         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 903         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 904         [CTA_STATUS]            = { .type = NLA_U32 },
 905         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 906         [CTA_HELP]              = { .type = NLA_NESTED },
 907         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 908         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 909         [CTA_MARK]              = { .type = NLA_U32 },
 910         [CTA_ID]                = { .type = NLA_U32 },
 911         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 912         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 913         [CTA_ZONE]              = { .type = NLA_U16 },
 914         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 915 };
 916
 917 static int
 918 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 919                         const struct nlmsghdr *nlh,
 920                         const struct nlattr * const cda[])
 921 {
 922         struct net *net = sock_net(ctnl);
 923         struct nf_conntrack_tuple_hash *h;
 924         struct nf_conntrack_tuple tuple;
 925         struct nf_conn *ct;
 926         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 927         u_int8_t u3 = nfmsg->nfgen_family;
 928         u16 zone;
 929         int err;
 930
 931         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 932         if (err < 0)
 933                 return err;
 934
 935         if (cda[CTA_TUPLE_ORIG])
 936                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 937         else if (cda[CTA_TUPLE_REPLY])
 938                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 939         else {
 940                 /* Flush the whole table */
 941                 nf_conntrack_flush_report(net,
 942                                          NETLINK_CB(skb).pid,
 943                                          nlmsg_report(nlh));
 944                 return 0;
 945         }
 946
 947         if (err < 0)
 948                 return err;
 949
 950         h = nf_conntrack_find_get(net, zone, &tuple);
 951         if (!h)
 952                 return -ENOENT;
 953
 954         ct = nf_ct_tuplehash_to_ctrack(h);
 955
 956         if (cda[CTA_ID]) {
 957                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
 958                 if (id != (u32)(unsigned long)ct) {
 959                         nf_ct_put(ct);
 960                         return -ENOENT;
 961                 }
 962         }
 963
 964         if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 965                                       NETLINK_CB(skb).pid,
 966                                       nlmsg_report(nlh)) < 0) {
 967                 nf_ct_delete_from_lists(ct);
 968                 /* we failed to report the event, try later */
 969                 nf_ct_insert_dying_list(ct);
 970                 nf_ct_put(ct);
 971                 return 0;
 972         }
 973
 974         /* death_by_timeout would report the event again */
 975         set_bit(IPS_DYING_BIT, &ct->status);
 976
 977         nf_ct_kill(ct);
 978         nf_ct_put(ct);
 979
 980         return 0;
 981 }
 982
 983 static int
 984 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 985                         const struct nlmsghdr *nlh,
 986                         const struct nlattr * const cda[])
 987 {
 988         struct net *net = sock_net(ctnl);
 989         struct nf_conntrack_tuple_hash *h;
 990         struct nf_conntrack_tuple tuple;
 991         struct nf_conn *ct;
 992         struct sk_buff *skb2 = NULL;
 993         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 994         u_int8_t u3 = nfmsg->nfgen_family;
 995         u16 zone;
 996         int err;
 997
 998         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 999                 struct netlink_dump_control c = {
1000                         .dump = ctnetlink_dump_table,
1001                         .done = ctnetlink_done,
1002                 };
1003 #ifdef CONFIG_NF_CONNTRACK_MARK
1004                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1005                         struct ctnetlink_dump_filter *filter;
1006
1007                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1008                                          GFP_ATOMIC);
1009                         if (filter == NULL)
1010                                 return -ENOMEM;
1011
1012                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1013                         filter->mark.mask =
1014                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1015                         c.data = filter;
1016                 }
1017 #endif
1018                 return netlink_dump_start(ctnl, skb, nlh, &c);
1019         }
1020
1021         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1022         if (err < 0)
1023                 return err;
1024
1025         if (cda[CTA_TUPLE_ORIG])
1026                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1027         else if (cda[CTA_TUPLE_REPLY])
1028                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1029         else
1030                 return -EINVAL;
1031
1032         if (err < 0)
1033                 return err;
1034
1035         h = nf_conntrack_find_get(net, zone, &tuple);
1036         if (!h)
1037                 return -ENOENT;
1038
1039         ct = nf_ct_tuplehash_to_ctrack(h);
1040
1041         err = -ENOMEM;
1042         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1043         if (skb2 == NULL) {
1044                 nf_ct_put(ct);
1045                 return -ENOMEM;
1046         }
1047
1048         rcu_read_lock();
1049         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).pid, nlh->nlmsg_seq,
1050                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1051         rcu_read_unlock();
1052         nf_ct_put(ct);
1053         if (err <= 0)
1054                 goto free;
1055
1056         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1057         if (err < 0)
1058                 goto out;
1059
1060         return 0;
1061
1062 free:
1063         kfree_skb(skb2);
1064 out:
1065         /* this avoids a loop in nfnetlink. */
1066         return err == -EAGAIN ? -ENOBUFS : err;
1067 }
1068
1069 #ifdef CONFIG_NF_NAT_NEEDED
1070 static int
1071 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1072                           enum nf_nat_manip_type manip,
1073                           const struct nlattr *attr)
1074 {
1075         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1076
1077         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1078         if (!parse_nat_setup) {
1079 #ifdef CONFIG_MODULES
1080                 rcu_read_unlock();
1081                 spin_unlock_bh(&nf_conntrack_lock);
1082                 nfnl_unlock();
1083                 if (request_module("nf-nat-ipv4") < 0) {
1084                         nfnl_lock();
1085                         spin_lock_bh(&nf_conntrack_lock);
1086                         rcu_read_lock();
1087                         return -EOPNOTSUPP;
1088                 }
1089                 nfnl_lock();
1090                 spin_lock_bh(&nf_conntrack_lock);
1091                 rcu_read_lock();
1092                 if (nfnetlink_parse_nat_setup_hook)
1093                         return -EAGAIN;
1094 #endif
1095                 return -EOPNOTSUPP;
1096         }
1097
1098         return parse_nat_setup(ct, manip, attr);
1099 }
1100 #endif
1101
1102 static int
1103 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1104 {
1105         unsigned long d;
1106         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1107         d = ct->status ^ status;
1108
1109         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1110                 /* unchangeable */
1111                 return -EBUSY;
1112
1113         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1114                 /* SEEN_REPLY bit can only be set */
1115                 return -EBUSY;
1116
1117         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1118                 /* ASSURED bit can only be set */
1119                 return -EBUSY;
1120
1121         /* Be careful here, modifying NAT bits can screw up things,
1122          * so don't let users modify them directly if they don't pass
1123          * nf_nat_range. */
1124         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1125         return 0;
1126 }
1127
1128 static int
1129 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1130 {
1131 #ifdef CONFIG_NF_NAT_NEEDED
1132         int ret;
1133
1134         if (cda[CTA_NAT_DST]) {
1135                 ret = ctnetlink_parse_nat_setup(ct,
1136                                                 NF_NAT_MANIP_DST,
1137                                                 cda[CTA_NAT_DST]);
1138                 if (ret < 0)
1139                         return ret;
1140         }
1141         if (cda[CTA_NAT_SRC]) {
1142                 ret = ctnetlink_parse_nat_setup(ct,
1143                                                 NF_NAT_MANIP_SRC,
1144                                                 cda[CTA_NAT_SRC]);
1145                 if (ret < 0)
1146                         return ret;
1147         }
1148         return 0;
1149 #else
1150         return -EOPNOTSUPP;
1151 #endif
1152 }
1153
1154 static inline int
1155 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1156 {
1157         struct nf_conntrack_helper *helper;
1158         struct nf_conn_help *help = nfct_help(ct);
1159         char *helpname = NULL;
1160         int err;
1161
1162         /* don't change helper of sibling connections */
1163         if (ct->master)
1164                 return -EBUSY;
1165
1166         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1167         if (err < 0)
1168                 return err;
1169
1170         if (!strcmp(helpname, "")) {
1171                 if (help && help->helper) {
1172                         /* we had a helper before ... */
1173                         nf_ct_remove_expectations(ct);
1174                         RCU_INIT_POINTER(help->helper, NULL);
1175                 }
1176
1177                 return 0;
1178         }
1179
1180         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1181                                             nf_ct_protonum(ct));
1182         if (helper == NULL) {
1183 #ifdef CONFIG_MODULES
1184                 spin_unlock_bh(&nf_conntrack_lock);
1185
1186                 if (request_module("nfct-helper-%s", helpname) < 0) {
1187                         spin_lock_bh(&nf_conntrack_lock);
1188                         return -EOPNOTSUPP;
1189                 }
1190
1191                 spin_lock_bh(&nf_conntrack_lock);
1192                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1193                                                     nf_ct_protonum(ct));
1194                 if (helper)
1195                         return -EAGAIN;
1196 #endif
1197                 return -EOPNOTSUPP;
1198         }
1199
1200         if (help) {
1201                 if (help->helper == helper)
1202                         return 0;
1203                 if (help->helper)
1204                         return -EBUSY;
1205                 /* need to zero data of old helper */
1206                 memset(&help->help, 0, sizeof(help->help));
1207         } else {
1208                 /* we cannot set a helper for an existing conntrack */
1209                 return -EOPNOTSUPP;
1210         }
1211
1212         rcu_assign_pointer(help->helper, helper);
1213
1214         return 0;
1215 }
1216
1217 static inline int
1218 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1219 {
1220         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1221
1222         if (!del_timer(&ct->timeout))
1223                 return -ETIME;
1224
1225         ct->timeout.expires = jiffies + timeout * HZ;
1226         add_timer(&ct->timeout);
1227
1228         return 0;
1229 }
1230
1231 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1232         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1233         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1234         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1235 };
1236
1237 static inline int
1238 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1239 {
1240         const struct nlattr *attr = cda[CTA_PROTOINFO];
1241         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1242         struct nf_conntrack_l4proto *l4proto;
1243         int err = 0;
1244
1245         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1246
1247         rcu_read_lock();
1248         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1249         if (l4proto->from_nlattr)
1250                 err = l4proto->from_nlattr(tb, ct);
1251         rcu_read_unlock();
1252
1253         return err;
1254 }
1255
1256 #ifdef CONFIG_NF_NAT_NEEDED
1257 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1258         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1259         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1260         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1261 };
1262
1263 static inline int
1264 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1265 {
1266         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1267
1268         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1269
1270         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1271                 return -EINVAL;
1272
1273         natseq->correction_pos =
1274                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1275
1276         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1277                 return -EINVAL;
1278
1279         natseq->offset_before =
1280                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1281
1282         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1283                 return -EINVAL;
1284
1285         natseq->offset_after =
1286                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1287
1288         return 0;
1289 }
1290
1291 static int
1292 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1293                              const struct nlattr * const cda[])
1294 {
1295         int ret = 0;
1296         struct nf_conn_nat *nat = nfct_nat(ct);
1297
1298         if (!nat)
1299                 return 0;
1300
1301         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1302                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1303                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1304                 if (ret < 0)
1305                         return ret;
1306
1307                 ct->status |= IPS_SEQ_ADJUST;
1308         }
1309
1310         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1311                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1312                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1313                 if (ret < 0)
1314                         return ret;
1315
1316                 ct->status |= IPS_SEQ_ADJUST;
1317         }
1318
1319         return 0;
1320 }
1321 #endif
1322
1323 static int
1324 ctnetlink_change_conntrack(struct nf_conn *ct,
1325                            const struct nlattr * const cda[])
1326 {
1327         int err;
1328
1329         /* only allow NAT changes and master assignation for new conntracks */
1330         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1331                 return -EOPNOTSUPP;
1332
1333         if (cda[CTA_HELP]) {
1334                 err = ctnetlink_change_helper(ct, cda);
1335                 if (err < 0)
1336                         return err;
1337         }
1338
1339         if (cda[CTA_TIMEOUT]) {
1340                 err = ctnetlink_change_timeout(ct, cda);
1341                 if (err < 0)
1342                         return err;
1343         }
1344
1345         if (cda[CTA_STATUS]) {
1346                 err = ctnetlink_change_status(ct, cda);
1347                 if (err < 0)
1348                         return err;
1349         }
1350
1351         if (cda[CTA_PROTOINFO]) {
1352                 err = ctnetlink_change_protoinfo(ct, cda);
1353                 if (err < 0)
1354                         return err;
1355         }
1356
1357 #if defined(CONFIG_NF_CONNTRACK_MARK)
1358         if (cda[CTA_MARK])
1359                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1360 #endif
1361
1362 #ifdef CONFIG_NF_NAT_NEEDED
1363         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1364                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1365                 if (err < 0)
1366                         return err;
1367         }
1368 #endif
1369
1370         return 0;
1371 }
1372
1373 static struct nf_conn *
1374 ctnetlink_create_conntrack(struct net *net, u16 zone,
1375                            const struct nlattr * const cda[],
1376                            struct nf_conntrack_tuple *otuple,
1377                            struct nf_conntrack_tuple *rtuple,
1378                            u8 u3)
1379 {
1380         struct nf_conn *ct;
1381         int err = -EINVAL;
1382         struct nf_conntrack_helper *helper;
1383         struct nf_conn_tstamp *tstamp;
1384
1385         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1386         if (IS_ERR(ct))
1387                 return ERR_PTR(-ENOMEM);
1388
1389         if (!cda[CTA_TIMEOUT])
1390                 goto err1;
1391         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1392
1393         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1394
1395         rcu_read_lock();
1396         if (cda[CTA_HELP]) {
1397                 char *helpname = NULL;
1398
1399                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1400                 if (err < 0)
1401                         goto err2;
1402
1403                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1404                                                     nf_ct_protonum(ct));
1405                 if (helper == NULL) {
1406                         rcu_read_unlock();
1407 #ifdef CONFIG_MODULES
1408                         if (request_module("nfct-helper-%s", helpname) < 0) {
1409                                 err = -EOPNOTSUPP;
1410                                 goto err1;
1411                         }
1412
1413                         rcu_read_lock();
1414                         helper = __nf_conntrack_helper_find(helpname,
1415                                                             nf_ct_l3num(ct),
1416                                                             nf_ct_protonum(ct));
1417                         if (helper) {
1418                                 err = -EAGAIN;
1419                                 goto err2;
1420                         }
1421                         rcu_read_unlock();
1422 #endif
1423                         err = -EOPNOTSUPP;
1424                         goto err1;
1425                 } else {
1426                         struct nf_conn_help *help;
1427
1428                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1429                         if (help == NULL) {
1430                                 err = -ENOMEM;
1431                                 goto err2;
1432                         }
1433
1434                         /* not in hash table yet so not strictly necessary */
1435                         RCU_INIT_POINTER(help->helper, helper);
1436                 }
1437         } else {
1438                 /* try an implicit helper assignation */
1439                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1440                 if (err < 0)
1441                         goto err2;
1442         }
1443
1444         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1445                 err = ctnetlink_change_nat(ct, cda);
1446                 if (err < 0)
1447                         goto err2;
1448         }
1449
1450         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1451         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1452         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1453         /* we must add conntrack extensions before confirmation. */
1454         ct->status |= IPS_CONFIRMED;
1455
1456         if (cda[CTA_STATUS]) {
1457                 err = ctnetlink_change_status(ct, cda);
1458                 if (err < 0)
1459                         goto err2;
1460         }
1461
1462 #ifdef CONFIG_NF_NAT_NEEDED
1463         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1464                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1465                 if (err < 0)
1466                         goto err2;
1467         }
1468 #endif
1469
1470         memset(&ct->proto, 0, sizeof(ct->proto));
1471         if (cda[CTA_PROTOINFO]) {
1472                 err = ctnetlink_change_protoinfo(ct, cda);
1473                 if (err < 0)
1474                         goto err2;
1475         }
1476
1477 #if defined(CONFIG_NF_CONNTRACK_MARK)
1478         if (cda[CTA_MARK])
1479                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1480 #endif
1481
1482         /* setup master conntrack: this is a confirmed expectation */
1483         if (cda[CTA_TUPLE_MASTER]) {
1484                 struct nf_conntrack_tuple master;
1485                 struct nf_conntrack_tuple_hash *master_h;
1486                 struct nf_conn *master_ct;
1487
1488                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1489                 if (err < 0)
1490                         goto err2;
1491
1492                 master_h = nf_conntrack_find_get(net, zone, &master);
1493                 if (master_h == NULL) {
1494                         err = -ENOENT;
1495                         goto err2;
1496                 }
1497                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1498                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1499                 ct->master = master_ct;
1500         }
1501         tstamp = nf_conn_tstamp_find(ct);
1502         if (tstamp)
1503                 tstamp->start = ktime_to_ns(ktime_get_real());
1504
1505         err = nf_conntrack_hash_check_insert(ct);
1506         if (err < 0)
1507                 goto err2;
1508
1509         rcu_read_unlock();
1510
1511         return ct;
1512
1513 err2:
1514         rcu_read_unlock();
1515 err1:
1516         nf_conntrack_free(ct);
1517         return ERR_PTR(err);
1518 }
1519
1520 static int
1521 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1522                         const struct nlmsghdr *nlh,
1523                         const struct nlattr * const cda[])
1524 {
1525         struct net *net = sock_net(ctnl);
1526         struct nf_conntrack_tuple otuple, rtuple;
1527         struct nf_conntrack_tuple_hash *h = NULL;
1528         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1529         struct nf_conn *ct;
1530         u_int8_t u3 = nfmsg->nfgen_family;
1531         u16 zone;
1532         int err;
1533
1534         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1535         if (err < 0)
1536                 return err;
1537
1538         if (cda[CTA_TUPLE_ORIG]) {
1539                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1540                 if (err < 0)
1541                         return err;
1542         }
1543
1544         if (cda[CTA_TUPLE_REPLY]) {
1545                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1546                 if (err < 0)
1547                         return err;
1548         }
1549
1550         if (cda[CTA_TUPLE_ORIG])
1551                 h = nf_conntrack_find_get(net, zone, &otuple);
1552         else if (cda[CTA_TUPLE_REPLY])
1553                 h = nf_conntrack_find_get(net, zone, &rtuple);
1554
1555         if (h == NULL) {
1556                 err = -ENOENT;
1557                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1558                         enum ip_conntrack_events events;
1559
1560                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1561                                                         &rtuple, u3);
1562                         if (IS_ERR(ct))
1563                                 return PTR_ERR(ct);
1564
1565                         err = 0;
1566                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1567                                 events = IPCT_RELATED;
1568                         else
1569                                 events = IPCT_NEW;
1570
1571                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1572                                                       (1 << IPCT_ASSURED) |
1573                                                       (1 << IPCT_HELPER) |
1574                                                       (1 << IPCT_PROTOINFO) |
1575                                                       (1 << IPCT_NATSEQADJ) |
1576                                                       (1 << IPCT_MARK) | events,
1577                                                       ct, NETLINK_CB(skb).pid,
1578                                                       nlmsg_report(nlh));
1579                         nf_ct_put(ct);
1580                 }
1581
1582                 return err;
1583         }
1584         /* implicit 'else' */
1585
1586         err = -EEXIST;
1587         ct = nf_ct_tuplehash_to_ctrack(h);
1588         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1589                 spin_lock_bh(&nf_conntrack_lock);
1590                 err = ctnetlink_change_conntrack(ct, cda);
1591                 spin_unlock_bh(&nf_conntrack_lock);
1592                 if (err == 0) {
1593                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1594                                                       (1 << IPCT_ASSURED) |
1595                                                       (1 << IPCT_HELPER) |
1596                                                       (1 << IPCT_PROTOINFO) |
1597                                                       (1 << IPCT_NATSEQADJ) |
1598                                                       (1 << IPCT_MARK),
1599                                                       ct, NETLINK_CB(skb).pid,
1600                                                       nlmsg_report(nlh));
1601                 }
1602         }
1603
1604         nf_ct_put(ct);
1605         return err;
1606 }
1607
1608 /***********************************************************************
1609  * EXPECT
1610  ***********************************************************************/
1611
1612 static inline int
1613 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1614                          const struct nf_conntrack_tuple *tuple,
1615                          enum ctattr_expect type)
1616 {
1617         struct nlattr *nest_parms;
1618
1619         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
1620         if (!nest_parms)
1621                 goto nla_put_failure;
1622         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1623                 goto nla_put_failure;
1624         nla_nest_end(skb, nest_parms);
1625
1626         return 0;
1627
1628 nla_put_failure:
1629         return -1;
1630 }
1631
1632 static inline int
1633 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1634                         const struct nf_conntrack_tuple *tuple,
1635                         const struct nf_conntrack_tuple_mask *mask)
1636 {
1637         int ret;
1638         struct nf_conntrack_l3proto *l3proto;
1639         struct nf_conntrack_l4proto *l4proto;
1640         struct nf_conntrack_tuple m;
1641         struct nlattr *nest_parms;
1642
1643         memset(&m, 0xFF, sizeof(m));
1644         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1645         m.src.u.all = mask->src.u.all;
1646         m.dst.protonum = tuple->dst.protonum;
1647
1648         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
1649         if (!nest_parms)
1650                 goto nla_put_failure;
1651
1652         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
1653         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1654
1655         if (unlikely(ret < 0))
1656                 goto nla_put_failure;
1657
1658         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
1659         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1660         if (unlikely(ret < 0))
1661                 goto nla_put_failure;
1662
1663         nla_nest_end(skb, nest_parms);
1664
1665         return 0;
1666
1667 nla_put_failure:
1668         return -1;
1669 }
1670
1671 static int
1672 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1673                           const struct nf_conntrack_expect *exp)
1674 {
1675         struct nf_conn *master = exp->master;
1676         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
1677         struct nf_conn_help *help;
1678
1679         if (timeout < 0)
1680                 timeout = 0;
1681
1682         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1683                 goto nla_put_failure;
1684         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1685                 goto nla_put_failure;
1686         if (ctnetlink_exp_dump_tuple(skb,
1687                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1688                                  CTA_EXPECT_MASTER) < 0)
1689                 goto nla_put_failure;
1690
1691         NLA_PUT_BE32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout));
1692         NLA_PUT_BE32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp));
1693         NLA_PUT_BE32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags));
1694         help = nfct_help(master);
1695         if (help) {
1696                 struct nf_conntrack_helper *helper;
1697
1698                 helper = rcu_dereference(help->helper);
1699                 if (helper)
1700                         NLA_PUT_STRING(skb, CTA_EXPECT_HELP_NAME, helper->name);
1701         }
1702
1703         return 0;
1704
1705 nla_put_failure:
1706         return -1;
1707 }
1708
1709 static int
1710 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1711                         int event, const struct nf_conntrack_expect *exp)
1712 {
1713         struct nlmsghdr *nlh;
1714         struct nfgenmsg *nfmsg;
1715         unsigned int flags = pid ? NLM_F_MULTI : 0;
1716
1717         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1718         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
1719         if (nlh == NULL)
1720                 goto nlmsg_failure;
1721
1722         nfmsg = nlmsg_data(nlh);
1723         nfmsg->nfgen_family = exp->tuple.src.l3num;
1724         nfmsg->version      = NFNETLINK_V0;
1725         nfmsg->res_id       = 0;
1726
1727         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1728                 goto nla_put_failure;
1729
1730         nlmsg_end(skb, nlh);
1731         return skb->len;
1732
1733 nlmsg_failure:
1734 nla_put_failure:
1735         nlmsg_cancel(skb, nlh);
1736         return -1;
1737 }
1738
1739 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1740 static int
1741 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
1742 {
1743         struct nf_conntrack_expect *exp = item->exp;
1744         struct net *net = nf_ct_exp_net(exp);
1745         struct nlmsghdr *nlh;
1746         struct nfgenmsg *nfmsg;
1747         struct sk_buff *skb;
1748         unsigned int type, group;
1749         int flags = 0;
1750
1751         if (events & (1 << IPEXP_DESTROY)) {
1752                 type = IPCTNL_MSG_EXP_DELETE;
1753                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
1754         } else if (events & (1 << IPEXP_NEW)) {
1755                 type = IPCTNL_MSG_EXP_NEW;
1756                 flags = NLM_F_CREATE|NLM_F_EXCL;
1757                 group = NFNLGRP_CONNTRACK_EXP_NEW;
1758         } else
1759                 return 0;
1760
1761         if (!item->report && !nfnetlink_has_listeners(net, group))
1762                 return 0;
1763
1764         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
1765         if (skb == NULL)
1766                 goto errout;
1767
1768         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1769         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
1770         if (nlh == NULL)
1771                 goto nlmsg_failure;
1772
1773         nfmsg = nlmsg_data(nlh);
1774         nfmsg->nfgen_family = exp->tuple.src.l3num;
1775         nfmsg->version      = NFNETLINK_V0;
1776         nfmsg->res_id       = 0;
1777
1778         rcu_read_lock();
1779         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1780                 goto nla_put_failure;
1781         rcu_read_unlock();
1782
1783         nlmsg_end(skb, nlh);
1784         nfnetlink_send(skb, net, item->pid, group, item->report, GFP_ATOMIC);
1785         return 0;
1786
1787 nla_put_failure:
1788         rcu_read_unlock();
1789         nlmsg_cancel(skb, nlh);
1790 nlmsg_failure:
1791         kfree_skb(skb);
1792 errout:
1793         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
1794         return 0;
1795 }
1796 #endif
1797 static int ctnetlink_exp_done(struct netlink_callback *cb)
1798 {
1799         if (cb->args[1])
1800                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1801         return 0;
1802 }
1803
1804 static int
1805 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1806 {
1807         struct net *net = sock_net(skb->sk);
1808         struct nf_conntrack_expect *exp, *last;
1809         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1810         struct hlist_node *n;
1811         u_int8_t l3proto = nfmsg->nfgen_family;
1812
1813         rcu_read_lock();
1814         last = (struct nf_conntrack_expect *)cb->args[1];
1815         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1816 restart:
1817                 hlist_for_each_entry(exp, n, &net->ct.expect_hash[cb->args[0]],
1818                                      hnode) {
1819                         if (l3proto && exp->tuple.src.l3num != l3proto)
1820                                 continue;
1821                         if (cb->args[1]) {
1822                                 if (exp != last)
1823                                         continue;
1824                                 cb->args[1] = 0;
1825                         }
1826                         if (ctnetlink_exp_fill_info(skb,
1827                                                     NETLINK_CB(cb->skb).pid,
1828                                                     cb->nlh->nlmsg_seq,
1829                                                     IPCTNL_MSG_EXP_NEW,
1830                                                     exp) < 0) {
1831                                 if (!atomic_inc_not_zero(&exp->use))
1832                                         continue;
1833                                 cb->args[1] = (unsigned long)exp;
1834                                 goto out;
1835                         }
1836                 }
1837                 if (cb->args[1]) {
1838                         cb->args[1] = 0;
1839                         goto restart;
1840                 }
1841         }
1842 out:
1843         rcu_read_unlock();
1844         if (last)
1845                 nf_ct_expect_put(last);
1846
1847         return skb->len;
1848 }
1849
1850 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1851         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1852         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1853         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1854         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1855         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1856         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING },
1857         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1858         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1859 };
1860
1861 static int
1862 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1863                      const struct nlmsghdr *nlh,
1864                      const struct nlattr * const cda[])
1865 {
1866         struct net *net = sock_net(ctnl);
1867         struct nf_conntrack_tuple tuple;
1868         struct nf_conntrack_expect *exp;
1869         struct sk_buff *skb2;
1870         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1871         u_int8_t u3 = nfmsg->nfgen_family;
1872         u16 zone;
1873         int err;
1874
1875         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1876                 struct netlink_dump_control c = {
1877                         .dump = ctnetlink_exp_dump_table,
1878                         .done = ctnetlink_exp_done,
1879                 };
1880                 return netlink_dump_start(ctnl, skb, nlh, &c);
1881         }
1882
1883         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1884         if (err < 0)
1885                 return err;
1886
1887         if (cda[CTA_EXPECT_TUPLE])
1888                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1889         else if (cda[CTA_EXPECT_MASTER])
1890                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1891         else
1892                 return -EINVAL;
1893
1894         if (err < 0)
1895                 return err;
1896
1897         exp = nf_ct_expect_find_get(net, zone, &tuple);
1898         if (!exp)
1899                 return -ENOENT;
1900
1901         if (cda[CTA_EXPECT_ID]) {
1902                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1903                 if (ntohl(id) != (u32)(unsigned long)exp) {
1904                         nf_ct_expect_put(exp);
1905                         return -ENOENT;
1906                 }
1907         }
1908
1909         err = -ENOMEM;
1910         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1911         if (skb2 == NULL) {
1912                 nf_ct_expect_put(exp);
1913                 goto out;
1914         }
1915
1916         rcu_read_lock();
1917         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1918                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
1919         rcu_read_unlock();
1920         nf_ct_expect_put(exp);
1921         if (err <= 0)
1922                 goto free;
1923
1924         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1925         if (err < 0)
1926                 goto out;
1927
1928         return 0;
1929
1930 free:
1931         kfree_skb(skb2);
1932 out:
1933         /* this avoids a loop in nfnetlink. */
1934         return err == -EAGAIN ? -ENOBUFS : err;
1935 }
1936
1937 static int
1938 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1939                      const struct nlmsghdr *nlh,
1940                      const struct nlattr * const cda[])
1941 {
1942         struct net *net = sock_net(ctnl);
1943         struct nf_conntrack_expect *exp;
1944         struct nf_conntrack_tuple tuple;
1945         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1946         struct hlist_node *n, *next;
1947         u_int8_t u3 = nfmsg->nfgen_family;
1948         unsigned int i;
1949         u16 zone;
1950         int err;
1951
1952         if (cda[CTA_EXPECT_TUPLE]) {
1953                 /* delete a single expect by tuple */
1954                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1955                 if (err < 0)
1956                         return err;
1957
1958                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1959                 if (err < 0)
1960                         return err;
1961
1962                 /* bump usage count to 2 */
1963                 exp = nf_ct_expect_find_get(net, zone, &tuple);
1964                 if (!exp)
1965                         return -ENOENT;
1966
1967                 if (cda[CTA_EXPECT_ID]) {
1968                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1969                         if (ntohl(id) != (u32)(unsigned long)exp) {
1970                                 nf_ct_expect_put(exp);
1971                                 return -ENOENT;
1972                         }
1973                 }
1974
1975                 /* after list removal, usage count == 1 */
1976                 spin_lock_bh(&nf_conntrack_lock);
1977                 if (del_timer(&exp->timeout)) {
1978                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).pid,
1979                                                    nlmsg_report(nlh));
1980                         nf_ct_expect_put(exp);
1981                 }
1982                 spin_unlock_bh(&nf_conntrack_lock);
1983                 /* have to put what we 'get' above.
1984                  * after this line usage count == 0 */
1985                 nf_ct_expect_put(exp);
1986         } else if (cda[CTA_EXPECT_HELP_NAME]) {
1987                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
1988                 struct nf_conn_help *m_help;
1989
1990                 /* delete all expectations for this helper */
1991                 spin_lock_bh(&nf_conntrack_lock);
1992                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1993                         hlist_for_each_entry_safe(exp, n, next,
1994                                                   &net->ct.expect_hash[i],
1995                                                   hnode) {
1996                                 m_help = nfct_help(exp->master);
1997                                 if (!strcmp(m_help->helper->name, name) &&
1998                                     del_timer(&exp->timeout)) {
1999                                         nf_ct_unlink_expect_report(exp,
2000                                                         NETLINK_CB(skb).pid,
2001                                                         nlmsg_report(nlh));
2002                                         nf_ct_expect_put(exp);
2003                                 }
2004                         }
2005                 }
2006                 spin_unlock_bh(&nf_conntrack_lock);
2007         } else {
2008                 /* This basically means we have to flush everything*/
2009                 spin_lock_bh(&nf_conntrack_lock);
2010                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2011                         hlist_for_each_entry_safe(exp, n, next,
2012                                                   &net->ct.expect_hash[i],
2013                                                   hnode) {
2014                                 if (del_timer(&exp->timeout)) {
2015                                         nf_ct_unlink_expect_report(exp,
2016                                                         NETLINK_CB(skb).pid,
2017                                                         nlmsg_report(nlh));
2018                                         nf_ct_expect_put(exp);
2019                                 }
2020                         }
2021                 }
2022                 spin_unlock_bh(&nf_conntrack_lock);
2023         }
2024
2025         return 0;
2026 }
2027 static int
2028 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2029                         const struct nlattr * const cda[])
2030 {
2031         return -EOPNOTSUPP;
2032 }
2033
2034 static int
2035 ctnetlink_create_expect(struct net *net, u16 zone,
2036                         const struct nlattr * const cda[],
2037                         u_int8_t u3,
2038                         u32 pid, int report)
2039 {
2040         struct nf_conntrack_tuple tuple, mask, master_tuple;
2041         struct nf_conntrack_tuple_hash *h = NULL;
2042         struct nf_conntrack_expect *exp;
2043         struct nf_conn *ct;
2044         struct nf_conn_help *help;
2045         int err = 0;
2046
2047         /* caller guarantees that those three CTA_EXPECT_* exist */
2048         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2049         if (err < 0)
2050                 return err;
2051         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2052         if (err < 0)
2053                 return err;
2054         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2055         if (err < 0)
2056                 return err;
2057
2058         /* Look for master conntrack of this expectation */
2059         h = nf_conntrack_find_get(net, zone, &master_tuple);
2060         if (!h)
2061                 return -ENOENT;
2062         ct = nf_ct_tuplehash_to_ctrack(h);
2063         exp = nf_ct_expect_alloc(ct);
2064         if (!exp) {
2065                 err = -ENOMEM;
2066                 goto out;
2067         }
2068         help = nfct_help(ct);
2069         if (!help) {
2070                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2071                         err = -EINVAL;
2072                         goto out;
2073                 }
2074                 exp->timeout.expires =
2075                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2076
2077                 exp->flags = NF_CT_EXPECT_USERSPACE;
2078                 if (cda[CTA_EXPECT_FLAGS]) {
2079                         exp->flags |=
2080                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2081                 }
2082         } else {
2083                 if (cda[CTA_EXPECT_FLAGS]) {
2084                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2085                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2086                 } else
2087                         exp->flags = 0;
2088         }
2089
2090         exp->class = 0;
2091         exp->expectfn = NULL;
2092         exp->master = ct;
2093         exp->helper = NULL;
2094         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2095         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2096         exp->mask.src.u.all = mask.src.u.all;
2097
2098         err = nf_ct_expect_related_report(exp, pid, report);
2099         nf_ct_expect_put(exp);
2100
2101 out:
2102         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2103         return err;
2104 }
2105
2106 static int
2107 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2108                      const struct nlmsghdr *nlh,
2109                      const struct nlattr * const cda[])
2110 {
2111         struct net *net = sock_net(ctnl);
2112         struct nf_conntrack_tuple tuple;
2113         struct nf_conntrack_expect *exp;
2114         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2115         u_int8_t u3 = nfmsg->nfgen_family;
2116         u16 zone;
2117         int err;
2118
2119         if (!cda[CTA_EXPECT_TUPLE]
2120             || !cda[CTA_EXPECT_MASK]
2121             || !cda[CTA_EXPECT_MASTER])
2122                 return -EINVAL;
2123
2124         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2125         if (err < 0)
2126                 return err;
2127
2128         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2129         if (err < 0)
2130                 return err;
2131
2132         spin_lock_bh(&nf_conntrack_lock);
2133         exp = __nf_ct_expect_find(net, zone, &tuple);
2134
2135         if (!exp) {
2136                 spin_unlock_bh(&nf_conntrack_lock);
2137                 err = -ENOENT;
2138                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2139                         err = ctnetlink_create_expect(net, zone, cda,
2140                                                       u3,
2141                                                       NETLINK_CB(skb).pid,
2142                                                       nlmsg_report(nlh));
2143                 }
2144                 return err;
2145         }
2146
2147         err = -EEXIST;
2148         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2149                 err = ctnetlink_change_expect(exp, cda);
2150         spin_unlock_bh(&nf_conntrack_lock);
2151
2152         return err;
2153 }
2154
2155 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2156 static struct nf_ct_event_notifier ctnl_notifier = {
2157         .fcn = ctnetlink_conntrack_event,
2158 };
2159
2160 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2161         .fcn = ctnetlink_expect_event,
2162 };
2163 #endif
2164
2165 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2166         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2167                                             .attr_count = CTA_MAX,
2168                                             .policy = ct_nla_policy },
2169         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2170                                             .attr_count = CTA_MAX,
2171                                             .policy = ct_nla_policy },
2172         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2173                                             .attr_count = CTA_MAX,
2174                                             .policy = ct_nla_policy },
2175         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
2176                                             .attr_count = CTA_MAX,
2177                                             .policy = ct_nla_policy },
2178 };
2179
2180 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
2181         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
2182                                             .attr_count = CTA_EXPECT_MAX,
2183                                             .policy = exp_nla_policy },
2184         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
2185                                             .attr_count = CTA_EXPECT_MAX,
2186                                             .policy = exp_nla_policy },
2187         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
2188                                             .attr_count = CTA_EXPECT_MAX,
2189                                             .policy = exp_nla_policy },
2190 };
2191
2192 static const struct nfnetlink_subsystem ctnl_subsys = {
2193         .name                           = "conntrack",
2194         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
2195         .cb_count                       = IPCTNL_MSG_MAX,
2196         .cb                             = ctnl_cb,
2197 };
2198
2199 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
2200         .name                           = "conntrack_expect",
2201         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
2202         .cb_count                       = IPCTNL_MSG_EXP_MAX,
2203         .cb                             = ctnl_exp_cb,
2204 };
2205
2206 MODULE_ALIAS("ip_conntrack_netlink");
2207 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
2208 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
2209
2210 static int __net_init ctnetlink_net_init(struct net *net)
2211 {
2212 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2213         int ret;
2214
2215         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
2216         if (ret < 0) {
2217                 pr_err("ctnetlink_init: cannot register notifier.\n");
2218                 goto err_out;
2219         }
2220
2221         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
2222         if (ret < 0) {
2223                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
2224                 goto err_unreg_notifier;
2225         }
2226 #endif
2227         return 0;
2228
2229 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2230 err_unreg_notifier:
2231         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2232 err_out:
2233         return ret;
2234 #endif
2235 }
2236
2237 static void ctnetlink_net_exit(struct net *net)
2238 {
2239 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2240         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
2241         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2242 #endif
2243 }
2244
2245 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
2246 {
2247         struct net *net;
2248
2249         list_for_each_entry(net, net_exit_list, exit_list)
2250                 ctnetlink_net_exit(net);
2251 }
2252
2253 static struct pernet_operations ctnetlink_net_ops = {
2254         .init           = ctnetlink_net_init,
2255         .exit_batch     = ctnetlink_net_exit_batch,
2256 };
2257
2258 static int __init ctnetlink_init(void)
2259 {
2260         int ret;
2261
2262         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
2263         ret = nfnetlink_subsys_register(&ctnl_subsys);
2264         if (ret < 0) {
2265                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
2266                 goto err_out;
2267         }
2268
2269         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
2270         if (ret < 0) {
2271                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
2272                 goto err_unreg_subsys;
2273         }
2274
2275         if (register_pernet_subsys(&ctnetlink_net_ops)) {
2276                 pr_err("ctnetlink_init: cannot register pernet operations\n");
2277                 goto err_unreg_exp_subsys;
2278         }
2279
2280         return 0;
2281
2282 err_unreg_exp_subsys:
2283         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2284 err_unreg_subsys:
2285         nfnetlink_subsys_unregister(&ctnl_subsys);
2286 err_out:
2287         return ret;
2288 }
2289
2290 static void __exit ctnetlink_exit(void)
2291 {
2292         pr_info("ctnetlink: unregistering from nfnetlink.\n");
2293
2294         unregister_pernet_subsys(&ctnetlink_net_ops);
2295         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2296         nfnetlink_subsys_unregister(&ctnl_subsys);
2297 }
2298
2299 module_init(ctnetlink_init);
2300 module_exit(ctnetlink_exit);