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                 spin_unlock_bh(&nf_conntrack_lock);
1045                 nfnl_unlock();
1046                 if (request_module("nf-nat-ipv4") < 0) {
1047                         nfnl_lock();
1048                         spin_lock_bh(&nf_conntrack_lock);
1049                         rcu_read_lock();
1050                         return -EOPNOTSUPP;
1051                 }
1052                 nfnl_lock();
1053                 spin_lock_bh(&nf_conntrack_lock);
1054                 rcu_read_lock();
1055                 if (nfnetlink_parse_nat_setup_hook)
1056                         return -EAGAIN;
1057 #endif
1058                 return -EOPNOTSUPP;
1059         }
1060
1061         return parse_nat_setup(ct, manip, attr);
1062 }
1063 #endif
1064
1065 static int
1066 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1067 {
1068         unsigned long d;
1069         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1070         d = ct->status ^ status;
1071
1072         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1073                 /* unchangeable */
1074                 return -EBUSY;
1075
1076         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1077                 /* SEEN_REPLY bit can only be set */
1078                 return -EBUSY;
1079
1080         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1081                 /* ASSURED bit can only be set */
1082                 return -EBUSY;
1083
1084         /* Be careful here, modifying NAT bits can screw up things,
1085          * so don't let users modify them directly if they don't pass
1086          * nf_nat_range. */
1087         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1088         return 0;
1089 }
1090
1091 static int
1092 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1093 {
1094 #ifdef CONFIG_NF_NAT_NEEDED
1095         int ret;
1096
1097         if (cda[CTA_NAT_DST]) {
1098                 ret = ctnetlink_parse_nat_setup(ct,
1099                                                 NF_NAT_MANIP_DST,
1100                                                 cda[CTA_NAT_DST]);
1101                 if (ret < 0)
1102                         return ret;
1103         }
1104         if (cda[CTA_NAT_SRC]) {
1105                 ret = ctnetlink_parse_nat_setup(ct,
1106                                                 NF_NAT_MANIP_SRC,
1107                                                 cda[CTA_NAT_SRC]);
1108                 if (ret < 0)
1109                         return ret;
1110         }
1111         return 0;
1112 #else
1113         return -EOPNOTSUPP;
1114 #endif
1115 }
1116
1117 static inline int
1118 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1119 {
1120         struct nf_conntrack_helper *helper;
1121         struct nf_conn_help *help = nfct_help(ct);
1122         char *helpname = NULL;
1123         int err;
1124
1125         /* don't change helper of sibling connections */
1126         if (ct->master)
1127                 return -EBUSY;
1128
1129         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1130         if (err < 0)
1131                 return err;
1132
1133         if (!strcmp(helpname, "")) {
1134                 if (help && help->helper) {
1135                         /* we had a helper before ... */
1136                         nf_ct_remove_expectations(ct);
1137                         RCU_INIT_POINTER(help->helper, NULL);
1138                 }
1139
1140                 return 0;
1141         }
1142
1143         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1144                                             nf_ct_protonum(ct));
1145         if (helper == NULL) {
1146 #ifdef CONFIG_MODULES
1147                 spin_unlock_bh(&nf_conntrack_lock);
1148
1149                 if (request_module("nfct-helper-%s", helpname) < 0) {
1150                         spin_lock_bh(&nf_conntrack_lock);
1151                         return -EOPNOTSUPP;
1152                 }
1153
1154                 spin_lock_bh(&nf_conntrack_lock);
1155                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1156                                                     nf_ct_protonum(ct));
1157                 if (helper)
1158                         return -EAGAIN;
1159 #endif
1160                 return -EOPNOTSUPP;
1161         }
1162
1163         if (help) {
1164                 if (help->helper == helper)
1165                         return 0;
1166                 if (help->helper)
1167                         return -EBUSY;
1168                 /* need to zero data of old helper */
1169                 memset(&help->help, 0, sizeof(help->help));
1170         } else {
1171                 /* we cannot set a helper for an existing conntrack */
1172                 return -EOPNOTSUPP;
1173         }
1174
1175         rcu_assign_pointer(help->helper, helper);
1176
1177         return 0;
1178 }
1179
1180 static inline int
1181 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1182 {
1183         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1184
1185         if (!del_timer(&ct->timeout))
1186                 return -ETIME;
1187
1188         ct->timeout.expires = jiffies + timeout * HZ;
1189         add_timer(&ct->timeout);
1190
1191         return 0;
1192 }
1193
1194 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1195         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1196         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1197         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1198 };
1199
1200 static inline int
1201 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1202 {
1203         const struct nlattr *attr = cda[CTA_PROTOINFO];
1204         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1205         struct nf_conntrack_l4proto *l4proto;
1206         int err = 0;
1207
1208         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1209
1210         rcu_read_lock();
1211         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1212         if (l4proto->from_nlattr)
1213                 err = l4proto->from_nlattr(tb, ct);
1214         rcu_read_unlock();
1215
1216         return err;
1217 }
1218
1219 #ifdef CONFIG_NF_NAT_NEEDED
1220 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1221         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1222         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1223         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1224 };
1225
1226 static inline int
1227 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1228 {
1229         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1230
1231         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1232
1233         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1234                 return -EINVAL;
1235
1236         natseq->correction_pos =
1237                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1238
1239         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1240                 return -EINVAL;
1241
1242         natseq->offset_before =
1243                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1244
1245         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1246                 return -EINVAL;
1247
1248         natseq->offset_after =
1249                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1250
1251         return 0;
1252 }
1253
1254 static int
1255 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1256                              const struct nlattr * const cda[])
1257 {
1258         int ret = 0;
1259         struct nf_conn_nat *nat = nfct_nat(ct);
1260
1261         if (!nat)
1262                 return 0;
1263
1264         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1265                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1266                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1267                 if (ret < 0)
1268                         return ret;
1269
1270                 ct->status |= IPS_SEQ_ADJUST;
1271         }
1272
1273         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1274                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1275                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1276                 if (ret < 0)
1277                         return ret;
1278
1279                 ct->status |= IPS_SEQ_ADJUST;
1280         }
1281
1282         return 0;
1283 }
1284 #endif
1285
1286 static int
1287 ctnetlink_change_conntrack(struct nf_conn *ct,
1288                            const struct nlattr * const cda[])
1289 {
1290         int err;
1291
1292         /* only allow NAT changes and master assignation for new conntracks */
1293         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1294                 return -EOPNOTSUPP;
1295
1296         if (cda[CTA_HELP]) {
1297                 err = ctnetlink_change_helper(ct, cda);
1298                 if (err < 0)
1299                         return err;
1300         }
1301
1302         if (cda[CTA_TIMEOUT]) {
1303                 err = ctnetlink_change_timeout(ct, cda);
1304                 if (err < 0)
1305                         return err;
1306         }
1307
1308         if (cda[CTA_STATUS]) {
1309                 err = ctnetlink_change_status(ct, cda);
1310                 if (err < 0)
1311                         return err;
1312         }
1313
1314         if (cda[CTA_PROTOINFO]) {
1315                 err = ctnetlink_change_protoinfo(ct, cda);
1316                 if (err < 0)
1317                         return err;
1318         }
1319
1320 #if defined(CONFIG_NF_CONNTRACK_MARK)
1321         if (cda[CTA_MARK])
1322                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1323 #endif
1324
1325 #ifdef CONFIG_NF_NAT_NEEDED
1326         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1327                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1328                 if (err < 0)
1329                         return err;
1330         }
1331 #endif
1332
1333         return 0;
1334 }
1335
1336 static struct nf_conn *
1337 ctnetlink_create_conntrack(struct net *net, u16 zone,
1338                            const struct nlattr * const cda[],
1339                            struct nf_conntrack_tuple *otuple,
1340                            struct nf_conntrack_tuple *rtuple,
1341                            u8 u3)
1342 {
1343         struct nf_conn *ct;
1344         int err = -EINVAL;
1345         struct nf_conntrack_helper *helper;
1346         struct nf_conn_tstamp *tstamp;
1347
1348         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1349         if (IS_ERR(ct))
1350                 return ERR_PTR(-ENOMEM);
1351
1352         if (!cda[CTA_TIMEOUT])
1353                 goto err1;
1354         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1355
1356         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1357
1358         rcu_read_lock();
1359         if (cda[CTA_HELP]) {
1360                 char *helpname = NULL;
1361
1362                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname);
1363                 if (err < 0)
1364                         goto err2;
1365
1366                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1367                                                     nf_ct_protonum(ct));
1368                 if (helper == NULL) {
1369                         rcu_read_unlock();
1370                         spin_unlock_bh(&nf_conntrack_lock);
1371 #ifdef CONFIG_MODULES
1372                         if (request_module("nfct-helper-%s", helpname) < 0) {
1373                                 spin_lock_bh(&nf_conntrack_lock);
1374                                 err = -EOPNOTSUPP;
1375                                 goto err1;
1376                         }
1377
1378                         spin_lock_bh(&nf_conntrack_lock);
1379                         rcu_read_lock();
1380                         helper = __nf_conntrack_helper_find(helpname,
1381                                                             nf_ct_l3num(ct),
1382                                                             nf_ct_protonum(ct));
1383                         if (helper) {
1384                                 err = -EAGAIN;
1385                                 goto err2;
1386                         }
1387                         rcu_read_unlock();
1388 #endif
1389                         err = -EOPNOTSUPP;
1390                         goto err1;
1391                 } else {
1392                         struct nf_conn_help *help;
1393
1394                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1395                         if (help == NULL) {
1396                                 err = -ENOMEM;
1397                                 goto err2;
1398                         }
1399
1400                         /* not in hash table yet so not strictly necessary */
1401                         RCU_INIT_POINTER(help->helper, helper);
1402                 }
1403         } else {
1404                 /* try an implicit helper assignation */
1405                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1406                 if (err < 0)
1407                         goto err2;
1408         }
1409
1410         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1411                 err = ctnetlink_change_nat(ct, cda);
1412                 if (err < 0)
1413                         goto err2;
1414         }
1415
1416         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1417         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1418         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1419         /* we must add conntrack extensions before confirmation. */
1420         ct->status |= IPS_CONFIRMED;
1421
1422         if (cda[CTA_STATUS]) {
1423                 err = ctnetlink_change_status(ct, cda);
1424                 if (err < 0)
1425                         goto err2;
1426         }
1427
1428 #ifdef CONFIG_NF_NAT_NEEDED
1429         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1430                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1431                 if (err < 0)
1432                         goto err2;
1433         }
1434 #endif
1435
1436         memset(&ct->proto, 0, sizeof(ct->proto));
1437         if (cda[CTA_PROTOINFO]) {
1438                 err = ctnetlink_change_protoinfo(ct, cda);
1439                 if (err < 0)
1440                         goto err2;
1441         }
1442
1443 #if defined(CONFIG_NF_CONNTRACK_MARK)
1444         if (cda[CTA_MARK])
1445                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1446 #endif
1447
1448         /* setup master conntrack: this is a confirmed expectation */
1449         if (cda[CTA_TUPLE_MASTER]) {
1450                 struct nf_conntrack_tuple master;
1451                 struct nf_conntrack_tuple_hash *master_h;
1452                 struct nf_conn *master_ct;
1453
1454                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1455                 if (err < 0)
1456                         goto err2;
1457
1458                 master_h = nf_conntrack_find_get(net, zone, &master);
1459                 if (master_h == NULL) {
1460                         err = -ENOENT;
1461                         goto err2;
1462                 }
1463                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1464                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1465                 ct->master = master_ct;
1466         }
1467         tstamp = nf_conn_tstamp_find(ct);
1468         if (tstamp)
1469                 tstamp->start = ktime_to_ns(ktime_get_real());
1470
1471         add_timer(&ct->timeout);
1472         nf_conntrack_hash_insert(ct);
1473         rcu_read_unlock();
1474
1475         return ct;
1476
1477 err2:
1478         rcu_read_unlock();
1479 err1:
1480         nf_conntrack_free(ct);
1481         return ERR_PTR(err);
1482 }
1483
1484 static int
1485 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1486                         const struct nlmsghdr *nlh,
1487                         const struct nlattr * const cda[])
1488 {
1489         struct net *net = sock_net(ctnl);
1490         struct nf_conntrack_tuple otuple, rtuple;
1491         struct nf_conntrack_tuple_hash *h = NULL;
1492         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1493         u_int8_t u3 = nfmsg->nfgen_family;
1494         u16 zone;
1495         int err;
1496
1497         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1498         if (err < 0)
1499                 return err;
1500
1501         if (cda[CTA_TUPLE_ORIG]) {
1502                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1503                 if (err < 0)
1504                         return err;
1505         }
1506
1507         if (cda[CTA_TUPLE_REPLY]) {
1508                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1509                 if (err < 0)
1510                         return err;
1511         }
1512
1513         spin_lock_bh(&nf_conntrack_lock);
1514         if (cda[CTA_TUPLE_ORIG])
1515                 h = __nf_conntrack_find(net, zone, &otuple);
1516         else if (cda[CTA_TUPLE_REPLY])
1517                 h = __nf_conntrack_find(net, zone, &rtuple);
1518
1519         if (h == NULL) {
1520                 err = -ENOENT;
1521                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1522                         struct nf_conn *ct;
1523                         enum ip_conntrack_events events;
1524
1525                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1526                                                         &rtuple, u3);
1527                         if (IS_ERR(ct)) {
1528                                 err = PTR_ERR(ct);
1529                                 goto out_unlock;
1530                         }
1531                         err = 0;
1532                         nf_conntrack_get(&ct->ct_general);
1533                         spin_unlock_bh(&nf_conntrack_lock);
1534                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1535                                 events = IPCT_RELATED;
1536                         else
1537                                 events = IPCT_NEW;
1538
1539                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1540                                                       (1 << IPCT_ASSURED) |
1541                                                       (1 << IPCT_HELPER) |
1542                                                       (1 << IPCT_PROTOINFO) |
1543                                                       (1 << IPCT_NATSEQADJ) |
1544                                                       (1 << IPCT_MARK) | events,
1545                                                       ct, NETLINK_CB(skb).pid,
1546                                                       nlmsg_report(nlh));
1547                         nf_ct_put(ct);
1548                 } else
1549                         spin_unlock_bh(&nf_conntrack_lock);
1550
1551                 return err;
1552         }
1553         /* implicit 'else' */
1554
1555         /* We manipulate the conntrack inside the global conntrack table lock,
1556          * so there's no need to increase the refcount */
1557         err = -EEXIST;
1558         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1559                 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1560
1561                 err = ctnetlink_change_conntrack(ct, cda);
1562                 if (err == 0) {
1563                         nf_conntrack_get(&ct->ct_general);
1564                         spin_unlock_bh(&nf_conntrack_lock);
1565                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1566                                                       (1 << IPCT_ASSURED) |
1567                                                       (1 << IPCT_HELPER) |
1568                                                       (1 << IPCT_PROTOINFO) |
1569                                                       (1 << IPCT_NATSEQADJ) |
1570                                                       (1 << IPCT_MARK),
1571                                                       ct, NETLINK_CB(skb).pid,
1572                                                       nlmsg_report(nlh));
1573                         nf_ct_put(ct);
1574                 } else
1575                         spin_unlock_bh(&nf_conntrack_lock);
1576
1577                 return err;
1578         }
1579
1580 out_unlock:
1581         spin_unlock_bh(&nf_conntrack_lock);
1582         return err;
1583 }
1584
1585 /***********************************************************************
1586  * EXPECT
1587  ***********************************************************************/
1588
1589 static inline int
1590 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
1591                          const struct nf_conntrack_tuple *tuple,
1592                          enum ctattr_expect type)
1593 {
1594         struct nlattr *nest_parms;
1595
1596         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
1597         if (!nest_parms)
1598                 goto nla_put_failure;
1599         if (ctnetlink_dump_tuples(skb, tuple) < 0)
1600                 goto nla_put_failure;
1601         nla_nest_end(skb, nest_parms);
1602
1603         return 0;
1604
1605 nla_put_failure:
1606         return -1;
1607 }
1608
1609 static inline int
1610 ctnetlink_exp_dump_mask(struct sk_buff *skb,
1611                         const struct nf_conntrack_tuple *tuple,
1612                         const struct nf_conntrack_tuple_mask *mask)
1613 {
1614         int ret;
1615         struct nf_conntrack_l3proto *l3proto;
1616         struct nf_conntrack_l4proto *l4proto;
1617         struct nf_conntrack_tuple m;
1618         struct nlattr *nest_parms;
1619
1620         memset(&m, 0xFF, sizeof(m));
1621         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
1622         m.src.u.all = mask->src.u.all;
1623         m.dst.protonum = tuple->dst.protonum;
1624
1625         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
1626         if (!nest_parms)
1627                 goto nla_put_failure;
1628
1629         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
1630         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
1631
1632         if (unlikely(ret < 0))
1633                 goto nla_put_failure;
1634
1635         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
1636         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
1637         if (unlikely(ret < 0))
1638                 goto nla_put_failure;
1639
1640         nla_nest_end(skb, nest_parms);
1641
1642         return 0;
1643
1644 nla_put_failure:
1645         return -1;
1646 }
1647
1648 static int
1649 ctnetlink_exp_dump_expect(struct sk_buff *skb,
1650                           const struct nf_conntrack_expect *exp)
1651 {
1652         struct nf_conn *master = exp->master;
1653         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
1654         struct nf_conn_help *help;
1655
1656         if (timeout < 0)
1657                 timeout = 0;
1658
1659         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
1660                 goto nla_put_failure;
1661         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
1662                 goto nla_put_failure;
1663         if (ctnetlink_exp_dump_tuple(skb,
1664                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1665                                  CTA_EXPECT_MASTER) < 0)
1666                 goto nla_put_failure;
1667
1668         NLA_PUT_BE32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout));
1669         NLA_PUT_BE32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp));
1670         NLA_PUT_BE32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags));
1671         help = nfct_help(master);
1672         if (help) {
1673                 struct nf_conntrack_helper *helper;
1674
1675                 helper = rcu_dereference(help->helper);
1676                 if (helper)
1677                         NLA_PUT_STRING(skb, CTA_EXPECT_HELP_NAME, helper->name);
1678         }
1679
1680         return 0;
1681
1682 nla_put_failure:
1683         return -1;
1684 }
1685
1686 static int
1687 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 pid, u32 seq,
1688                         int event, const struct nf_conntrack_expect *exp)
1689 {
1690         struct nlmsghdr *nlh;
1691         struct nfgenmsg *nfmsg;
1692         unsigned int flags = pid ? NLM_F_MULTI : 0;
1693
1694         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1695         nlh = nlmsg_put(skb, pid, seq, event, sizeof(*nfmsg), flags);
1696         if (nlh == NULL)
1697                 goto nlmsg_failure;
1698
1699         nfmsg = nlmsg_data(nlh);
1700         nfmsg->nfgen_family = exp->tuple.src.l3num;
1701         nfmsg->version      = NFNETLINK_V0;
1702         nfmsg->res_id       = 0;
1703
1704         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1705                 goto nla_put_failure;
1706
1707         nlmsg_end(skb, nlh);
1708         return skb->len;
1709
1710 nlmsg_failure:
1711 nla_put_failure:
1712         nlmsg_cancel(skb, nlh);
1713         return -1;
1714 }
1715
1716 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1717 static int
1718 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
1719 {
1720         struct nf_conntrack_expect *exp = item->exp;
1721         struct net *net = nf_ct_exp_net(exp);
1722         struct nlmsghdr *nlh;
1723         struct nfgenmsg *nfmsg;
1724         struct sk_buff *skb;
1725         unsigned int type, group;
1726         int flags = 0;
1727
1728         if (events & (1 << IPEXP_DESTROY)) {
1729                 type = IPCTNL_MSG_EXP_DELETE;
1730                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
1731         } else if (events & (1 << IPEXP_NEW)) {
1732                 type = IPCTNL_MSG_EXP_NEW;
1733                 flags = NLM_F_CREATE|NLM_F_EXCL;
1734                 group = NFNLGRP_CONNTRACK_EXP_NEW;
1735         } else
1736                 return 0;
1737
1738         if (!item->report && !nfnetlink_has_listeners(net, group))
1739                 return 0;
1740
1741         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
1742         if (skb == NULL)
1743                 goto errout;
1744
1745         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
1746         nlh = nlmsg_put(skb, item->pid, 0, type, sizeof(*nfmsg), flags);
1747         if (nlh == NULL)
1748                 goto nlmsg_failure;
1749
1750         nfmsg = nlmsg_data(nlh);
1751         nfmsg->nfgen_family = exp->tuple.src.l3num;
1752         nfmsg->version      = NFNETLINK_V0;
1753         nfmsg->res_id       = 0;
1754
1755         rcu_read_lock();
1756         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
1757                 goto nla_put_failure;
1758         rcu_read_unlock();
1759
1760         nlmsg_end(skb, nlh);
1761         nfnetlink_send(skb, net, item->pid, group, item->report, GFP_ATOMIC);
1762         return 0;
1763
1764 nla_put_failure:
1765         rcu_read_unlock();
1766         nlmsg_cancel(skb, nlh);
1767 nlmsg_failure:
1768         kfree_skb(skb);
1769 errout:
1770         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
1771         return 0;
1772 }
1773 #endif
1774 static int ctnetlink_exp_done(struct netlink_callback *cb)
1775 {
1776         if (cb->args[1])
1777                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
1778         return 0;
1779 }
1780
1781 static int
1782 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1783 {
1784         struct net *net = sock_net(skb->sk);
1785         struct nf_conntrack_expect *exp, *last;
1786         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1787         struct hlist_node *n;
1788         u_int8_t l3proto = nfmsg->nfgen_family;
1789
1790         rcu_read_lock();
1791         last = (struct nf_conntrack_expect *)cb->args[1];
1792         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
1793 restart:
1794                 hlist_for_each_entry(exp, n, &net->ct.expect_hash[cb->args[0]],
1795                                      hnode) {
1796                         if (l3proto && exp->tuple.src.l3num != l3proto)
1797                                 continue;
1798                         if (cb->args[1]) {
1799                                 if (exp != last)
1800                                         continue;
1801                                 cb->args[1] = 0;
1802                         }
1803                         if (ctnetlink_exp_fill_info(skb,
1804                                                     NETLINK_CB(cb->skb).pid,
1805                                                     cb->nlh->nlmsg_seq,
1806                                                     IPCTNL_MSG_EXP_NEW,
1807                                                     exp) < 0) {
1808                                 if (!atomic_inc_not_zero(&exp->use))
1809                                         continue;
1810                                 cb->args[1] = (unsigned long)exp;
1811                                 goto out;
1812                         }
1813                 }
1814                 if (cb->args[1]) {
1815                         cb->args[1] = 0;
1816                         goto restart;
1817                 }
1818         }
1819 out:
1820         rcu_read_unlock();
1821         if (last)
1822                 nf_ct_expect_put(last);
1823
1824         return skb->len;
1825 }
1826
1827 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1828         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1829         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
1830         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
1831         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
1832         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
1833         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING },
1834         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
1835         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
1836 };
1837
1838 static int
1839 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
1840                      const struct nlmsghdr *nlh,
1841                      const struct nlattr * const cda[])
1842 {
1843         struct net *net = sock_net(ctnl);
1844         struct nf_conntrack_tuple tuple;
1845         struct nf_conntrack_expect *exp;
1846         struct sk_buff *skb2;
1847         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1848         u_int8_t u3 = nfmsg->nfgen_family;
1849         u16 zone;
1850         int err;
1851
1852         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1853                 return netlink_dump_start(ctnl, skb, nlh,
1854                                           ctnetlink_exp_dump_table,
1855                                           ctnetlink_exp_done, 0);
1856         }
1857
1858         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1859         if (err < 0)
1860                 return err;
1861
1862         if (cda[CTA_EXPECT_TUPLE])
1863                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1864         else if (cda[CTA_EXPECT_MASTER])
1865                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
1866         else
1867                 return -EINVAL;
1868
1869         if (err < 0)
1870                 return err;
1871
1872         exp = nf_ct_expect_find_get(net, zone, &tuple);
1873         if (!exp)
1874                 return -ENOENT;
1875
1876         if (cda[CTA_EXPECT_ID]) {
1877                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1878                 if (ntohl(id) != (u32)(unsigned long)exp) {
1879                         nf_ct_expect_put(exp);
1880                         return -ENOENT;
1881                 }
1882         }
1883
1884         err = -ENOMEM;
1885         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1886         if (skb2 == NULL) {
1887                 nf_ct_expect_put(exp);
1888                 goto out;
1889         }
1890
1891         rcu_read_lock();
1892         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).pid,
1893                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
1894         rcu_read_unlock();
1895         nf_ct_expect_put(exp);
1896         if (err <= 0)
1897                 goto free;
1898
1899         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).pid, MSG_DONTWAIT);
1900         if (err < 0)
1901                 goto out;
1902
1903         return 0;
1904
1905 free:
1906         kfree_skb(skb2);
1907 out:
1908         /* this avoids a loop in nfnetlink. */
1909         return err == -EAGAIN ? -ENOBUFS : err;
1910 }
1911
1912 static int
1913 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
1914                      const struct nlmsghdr *nlh,
1915                      const struct nlattr * const cda[])
1916 {
1917         struct net *net = sock_net(ctnl);
1918         struct nf_conntrack_expect *exp;
1919         struct nf_conntrack_tuple tuple;
1920         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1921         struct hlist_node *n, *next;
1922         u_int8_t u3 = nfmsg->nfgen_family;
1923         unsigned int i;
1924         u16 zone;
1925         int err;
1926
1927         if (cda[CTA_EXPECT_TUPLE]) {
1928                 /* delete a single expect by tuple */
1929                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
1930                 if (err < 0)
1931                         return err;
1932
1933                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
1934                 if (err < 0)
1935                         return err;
1936
1937                 /* bump usage count to 2 */
1938                 exp = nf_ct_expect_find_get(net, zone, &tuple);
1939                 if (!exp)
1940                         return -ENOENT;
1941
1942                 if (cda[CTA_EXPECT_ID]) {
1943                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
1944                         if (ntohl(id) != (u32)(unsigned long)exp) {
1945                                 nf_ct_expect_put(exp);
1946                                 return -ENOENT;
1947                         }
1948                 }
1949
1950                 /* after list removal, usage count == 1 */
1951                 spin_lock_bh(&nf_conntrack_lock);
1952                 if (del_timer(&exp->timeout)) {
1953                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).pid,
1954                                                    nlmsg_report(nlh));
1955                         nf_ct_expect_put(exp);
1956                 }
1957                 spin_unlock_bh(&nf_conntrack_lock);
1958                 /* have to put what we 'get' above.
1959                  * after this line usage count == 0 */
1960                 nf_ct_expect_put(exp);
1961         } else if (cda[CTA_EXPECT_HELP_NAME]) {
1962                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
1963                 struct nf_conn_help *m_help;
1964
1965                 /* delete all expectations for this helper */
1966                 spin_lock_bh(&nf_conntrack_lock);
1967                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1968                         hlist_for_each_entry_safe(exp, n, next,
1969                                                   &net->ct.expect_hash[i],
1970                                                   hnode) {
1971                                 m_help = nfct_help(exp->master);
1972                                 if (!strcmp(m_help->helper->name, name) &&
1973                                     del_timer(&exp->timeout)) {
1974                                         nf_ct_unlink_expect_report(exp,
1975                                                         NETLINK_CB(skb).pid,
1976                                                         nlmsg_report(nlh));
1977                                         nf_ct_expect_put(exp);
1978                                 }
1979                         }
1980                 }
1981                 spin_unlock_bh(&nf_conntrack_lock);
1982         } else {
1983                 /* This basically means we have to flush everything*/
1984                 spin_lock_bh(&nf_conntrack_lock);
1985                 for (i = 0; i < nf_ct_expect_hsize; i++) {
1986                         hlist_for_each_entry_safe(exp, n, next,
1987                                                   &net->ct.expect_hash[i],
1988                                                   hnode) {
1989                                 if (del_timer(&exp->timeout)) {
1990                                         nf_ct_unlink_expect_report(exp,
1991                                                         NETLINK_CB(skb).pid,
1992                                                         nlmsg_report(nlh));
1993                                         nf_ct_expect_put(exp);
1994                                 }
1995                         }
1996                 }
1997                 spin_unlock_bh(&nf_conntrack_lock);
1998         }
1999
2000         return 0;
2001 }
2002 static int
2003 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2004                         const struct nlattr * const cda[])
2005 {
2006         return -EOPNOTSUPP;
2007 }
2008
2009 static int
2010 ctnetlink_create_expect(struct net *net, u16 zone,
2011                         const struct nlattr * const cda[],
2012                         u_int8_t u3,
2013                         u32 pid, int report)
2014 {
2015         struct nf_conntrack_tuple tuple, mask, master_tuple;
2016         struct nf_conntrack_tuple_hash *h = NULL;
2017         struct nf_conntrack_expect *exp;
2018         struct nf_conn *ct;
2019         struct nf_conn_help *help;
2020         int err = 0;
2021
2022         /* caller guarantees that those three CTA_EXPECT_* exist */
2023         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2024         if (err < 0)
2025                 return err;
2026         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2027         if (err < 0)
2028                 return err;
2029         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2030         if (err < 0)
2031                 return err;
2032
2033         /* Look for master conntrack of this expectation */
2034         h = nf_conntrack_find_get(net, zone, &master_tuple);
2035         if (!h)
2036                 return -ENOENT;
2037         ct = nf_ct_tuplehash_to_ctrack(h);
2038         exp = nf_ct_expect_alloc(ct);
2039         if (!exp) {
2040                 err = -ENOMEM;
2041                 goto out;
2042         }
2043         help = nfct_help(ct);
2044         if (!help) {
2045                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2046                         err = -EINVAL;
2047                         goto out;
2048                 }
2049                 exp->timeout.expires =
2050                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2051
2052                 exp->flags = NF_CT_EXPECT_USERSPACE;
2053                 if (cda[CTA_EXPECT_FLAGS]) {
2054                         exp->flags |=
2055                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2056                 }
2057         } else {
2058                 if (cda[CTA_EXPECT_FLAGS]) {
2059                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2060                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2061                 } else
2062                         exp->flags = 0;
2063         }
2064
2065         exp->class = 0;
2066         exp->expectfn = NULL;
2067         exp->master = ct;
2068         exp->helper = NULL;
2069         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2070         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2071         exp->mask.src.u.all = mask.src.u.all;
2072
2073         err = nf_ct_expect_related_report(exp, pid, report);
2074         nf_ct_expect_put(exp);
2075
2076 out:
2077         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2078         return err;
2079 }
2080
2081 static int
2082 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2083                      const struct nlmsghdr *nlh,
2084                      const struct nlattr * const cda[])
2085 {
2086         struct net *net = sock_net(ctnl);
2087         struct nf_conntrack_tuple tuple;
2088         struct nf_conntrack_expect *exp;
2089         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2090         u_int8_t u3 = nfmsg->nfgen_family;
2091         u16 zone;
2092         int err;
2093
2094         if (!cda[CTA_EXPECT_TUPLE]
2095             || !cda[CTA_EXPECT_MASK]
2096             || !cda[CTA_EXPECT_MASTER])
2097                 return -EINVAL;
2098
2099         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2100         if (err < 0)
2101                 return err;
2102
2103         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2104         if (err < 0)
2105                 return err;
2106
2107         spin_lock_bh(&nf_conntrack_lock);
2108         exp = __nf_ct_expect_find(net, zone, &tuple);
2109
2110         if (!exp) {
2111                 spin_unlock_bh(&nf_conntrack_lock);
2112                 err = -ENOENT;
2113                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2114                         err = ctnetlink_create_expect(net, zone, cda,
2115                                                       u3,
2116                                                       NETLINK_CB(skb).pid,
2117                                                       nlmsg_report(nlh));
2118                 }
2119                 return err;
2120         }
2121
2122         err = -EEXIST;
2123         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2124                 err = ctnetlink_change_expect(exp, cda);
2125         spin_unlock_bh(&nf_conntrack_lock);
2126
2127         return err;
2128 }
2129
2130 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2131 static struct nf_ct_event_notifier ctnl_notifier = {
2132         .fcn = ctnetlink_conntrack_event,
2133 };
2134
2135 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2136         .fcn = ctnetlink_expect_event,
2137 };
2138 #endif
2139
2140 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2141         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2142                                             .attr_count = CTA_MAX,
2143                                             .policy = ct_nla_policy },
2144         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2145                                             .attr_count = CTA_MAX,
2146                                             .policy = ct_nla_policy },
2147         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2148                                             .attr_count = CTA_MAX,
2149                                             .policy = ct_nla_policy },
2150         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
2151                                             .attr_count = CTA_MAX,
2152                                             .policy = ct_nla_policy },
2153 };
2154
2155 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
2156         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
2157                                             .attr_count = CTA_EXPECT_MAX,
2158                                             .policy = exp_nla_policy },
2159         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
2160                                             .attr_count = CTA_EXPECT_MAX,
2161                                             .policy = exp_nla_policy },
2162         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
2163                                             .attr_count = CTA_EXPECT_MAX,
2164                                             .policy = exp_nla_policy },
2165 };
2166
2167 static const struct nfnetlink_subsystem ctnl_subsys = {
2168         .name                           = "conntrack",
2169         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
2170         .cb_count                       = IPCTNL_MSG_MAX,
2171         .cb                             = ctnl_cb,
2172 };
2173
2174 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
2175         .name                           = "conntrack_expect",
2176         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
2177         .cb_count                       = IPCTNL_MSG_EXP_MAX,
2178         .cb                             = ctnl_exp_cb,
2179 };
2180
2181 MODULE_ALIAS("ip_conntrack_netlink");
2182 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
2183 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
2184
2185 static int __net_init ctnetlink_net_init(struct net *net)
2186 {
2187 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2188         int ret;
2189
2190         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
2191         if (ret < 0) {
2192                 pr_err("ctnetlink_init: cannot register notifier.\n");
2193                 goto err_out;
2194         }
2195
2196         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
2197         if (ret < 0) {
2198                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
2199                 goto err_unreg_notifier;
2200         }
2201 #endif
2202         return 0;
2203
2204 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2205 err_unreg_notifier:
2206         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2207 err_out:
2208         return ret;
2209 #endif
2210 }
2211
2212 static void ctnetlink_net_exit(struct net *net)
2213 {
2214 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2215         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
2216         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2217 #endif
2218 }
2219
2220 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
2221 {
2222         struct net *net;
2223
2224         list_for_each_entry(net, net_exit_list, exit_list)
2225                 ctnetlink_net_exit(net);
2226 }
2227
2228 static struct pernet_operations ctnetlink_net_ops = {
2229         .init           = ctnetlink_net_init,
2230         .exit_batch     = ctnetlink_net_exit_batch,
2231 };
2232
2233 static int __init ctnetlink_init(void)
2234 {
2235         int ret;
2236
2237         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
2238         ret = nfnetlink_subsys_register(&ctnl_subsys);
2239         if (ret < 0) {
2240                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
2241                 goto err_out;
2242         }
2243
2244         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
2245         if (ret < 0) {
2246                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
2247                 goto err_unreg_subsys;
2248         }
2249
2250         if (register_pernet_subsys(&ctnetlink_net_ops)) {
2251                 pr_err("ctnetlink_init: cannot register pernet operations\n");
2252                 goto err_unreg_exp_subsys;
2253         }
2254
2255         return 0;
2256
2257 err_unreg_exp_subsys:
2258         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2259 err_unreg_subsys:
2260         nfnetlink_subsys_unregister(&ctnl_subsys);
2261 err_out:
2262         return ret;
2263 }
2264
2265 static void __exit ctnetlink_exit(void)
2266 {
2267         pr_info("ctnetlink: unregistering from nfnetlink.\n");
2268
2269         unregister_pernet_subsys(&ctnetlink_net_ops);
2270         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2271         nfnetlink_subsys_unregister(&ctnl_subsys);
2272 }
2273
2274 module_init(ctnetlink_init);
2275 module_exit(ctnetlink_exit);