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) 2005-2012 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_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #include <net/netfilter/nf_conntrack_seqadj.h>
  49 #include <net/netfilter/nf_conntrack_synproxy.h>
  50 #ifdef CONFIG_NF_NAT_NEEDED
  51 #include <net/netfilter/nf_nat_core.h>
  52 #include <net/netfilter/nf_nat_l4proto.h>
  53 #include <net/netfilter/nf_nat_helper.h>
  54 #endif
  55
  56 #include <linux/netfilter/nfnetlink.h>
  57 #include <linux/netfilter/nfnetlink_conntrack.h>
  58
  59 MODULE_LICENSE("GPL");
  60
  61 static char __initdata version[] = "0.93";
  62
  63 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  64                                        const struct nf_conntrack_tuple *tuple,
  65                                        struct nf_conntrack_l4proto *l4proto)
  66 {
  67         int ret = 0;
  68         struct nlattr *nest_parms;
  69
  70         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  71         if (!nest_parms)
  72                 goto nla_put_failure;
  73         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  74                 goto nla_put_failure;
  75
  76         if (likely(l4proto->tuple_to_nlattr))
  77                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  78
  79         nla_nest_end(skb, nest_parms);
  80
  81         return ret;
  82
  83 nla_put_failure:
  84         return -1;
  85 }
  86
  87 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  88                                     const struct nf_conntrack_tuple *tuple,
  89                                     struct nf_conntrack_l3proto *l3proto)
  90 {
  91         int ret = 0;
  92         struct nlattr *nest_parms;
  93
  94         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  95         if (!nest_parms)
  96                 goto nla_put_failure;
  97
  98         if (likely(l3proto->tuple_to_nlattr))
  99                 ret = l3proto->tuple_to_nlattr(skb, tuple);
 100
 101         nla_nest_end(skb, nest_parms);
 102
 103         return ret;
 104
 105 nla_put_failure:
 106         return -1;
 107 }
 108
 109 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 110                                  const struct nf_conntrack_tuple *tuple)
 111 {
 112         int ret;
 113         struct nf_conntrack_l3proto *l3proto;
 114         struct nf_conntrack_l4proto *l4proto;
 115
 116         rcu_read_lock();
 117         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 118         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 119
 120         if (ret >= 0) {
 121                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 122                                                tuple->dst.protonum);
 123                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 124         }
 125         rcu_read_unlock();
 126         return ret;
 127 }
 128
 129 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 130                                   const struct nf_conntrack_zone *zone, int dir)
 131 {
 132         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 133                 return 0;
 134         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 135                 goto nla_put_failure;
 136         return 0;
 137
 138 nla_put_failure:
 139         return -1;
 140 }
 141
 142 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 143 {
 144         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 145                 goto nla_put_failure;
 146         return 0;
 147
 148 nla_put_failure:
 149         return -1;
 150 }
 151
 152 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 153 {
 154         long timeout = nf_ct_expires(ct) / HZ;
 155
 156         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 157                 goto nla_put_failure;
 158         return 0;
 159
 160 nla_put_failure:
 161         return -1;
 162 }
 163
 164 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 165 {
 166         struct nf_conntrack_l4proto *l4proto;
 167         struct nlattr *nest_proto;
 168         int ret;
 169
 170         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 171         if (!l4proto->to_nlattr)
 172                 return 0;
 173
 174         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 175         if (!nest_proto)
 176                 goto nla_put_failure;
 177
 178         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 179
 180         nla_nest_end(skb, nest_proto);
 181
 182         return ret;
 183
 184 nla_put_failure:
 185         return -1;
 186 }
 187
 188 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 189                                    const struct nf_conn *ct)
 190 {
 191         struct nlattr *nest_helper;
 192         const struct nf_conn_help *help = nfct_help(ct);
 193         struct nf_conntrack_helper *helper;
 194
 195         if (!help)
 196                 return 0;
 197
 198         helper = rcu_dereference(help->helper);
 199         if (!helper)
 200                 goto out;
 201
 202         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 203         if (!nest_helper)
 204                 goto nla_put_failure;
 205         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 206                 goto nla_put_failure;
 207
 208         if (helper->to_nlattr)
 209                 helper->to_nlattr(skb, ct);
 210
 211         nla_nest_end(skb, nest_helper);
 212 out:
 213         return 0;
 214
 215 nla_put_failure:
 216         return -1;
 217 }
 218
 219 static int
 220 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 221               enum ip_conntrack_dir dir, int type)
 222 {
 223         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 224         struct nf_conn_counter *counter = acct->counter;
 225         struct nlattr *nest_count;
 226         u64 pkts, bytes;
 227
 228         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 229                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 230                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 231         } else {
 232                 pkts = atomic64_read(&counter[dir].packets);
 233                 bytes = atomic64_read(&counter[dir].bytes);
 234         }
 235
 236         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 237         if (!nest_count)
 238                 goto nla_put_failure;
 239
 240         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 241                          CTA_COUNTERS_PAD) ||
 242             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 243                          CTA_COUNTERS_PAD))
 244                 goto nla_put_failure;
 245
 246         nla_nest_end(skb, nest_count);
 247
 248         return 0;
 249
 250 nla_put_failure:
 251         return -1;
 252 }
 253
 254 static int
 255 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 256 {
 257         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 258
 259         if (!acct)
 260                 return 0;
 261
 262         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 263                 return -1;
 264         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 265                 return -1;
 266
 267         return 0;
 268 }
 269
 270 static int
 271 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 272 {
 273         struct nlattr *nest_count;
 274         const struct nf_conn_tstamp *tstamp;
 275
 276         tstamp = nf_conn_tstamp_find(ct);
 277         if (!tstamp)
 278                 return 0;
 279
 280         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 281         if (!nest_count)
 282                 goto nla_put_failure;
 283
 284         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 285                          CTA_TIMESTAMP_PAD) ||
 286             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 287                                                cpu_to_be64(tstamp->stop),
 288                                                CTA_TIMESTAMP_PAD)))
 289                 goto nla_put_failure;
 290         nla_nest_end(skb, nest_count);
 291
 292         return 0;
 293
 294 nla_put_failure:
 295         return -1;
 296 }
 297
 298 #ifdef CONFIG_NF_CONNTRACK_MARK
 299 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 300 {
 301         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 302                 goto nla_put_failure;
 303         return 0;
 304
 305 nla_put_failure:
 306         return -1;
 307 }
 308 #else
 309 #define ctnetlink_dump_mark(a, b) (0)
 310 #endif
 311
 312 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 313 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 314 {
 315         struct nlattr *nest_secctx;
 316         int len, ret;
 317         char *secctx;
 318
 319         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 320         if (ret)
 321                 return 0;
 322
 323         ret = -1;
 324         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 325         if (!nest_secctx)
 326                 goto nla_put_failure;
 327
 328         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 329                 goto nla_put_failure;
 330         nla_nest_end(skb, nest_secctx);
 331
 332         ret = 0;
 333 nla_put_failure:
 334         security_release_secctx(secctx, len);
 335         return ret;
 336 }
 337 #else
 338 #define ctnetlink_dump_secctx(a, b) (0)
 339 #endif
 340
 341 #ifdef CONFIG_NF_CONNTRACK_LABELS
 342 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 343 {
 344         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 345
 346         if (!labels)
 347                 return 0;
 348         return nla_total_size(sizeof(labels->bits));
 349 }
 350
 351 static int
 352 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 353 {
 354         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 355         unsigned int i;
 356
 357         if (!labels)
 358                 return 0;
 359
 360         i = 0;
 361         do {
 362                 if (labels->bits[i] != 0)
 363                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 364                                        labels->bits);
 365                 i++;
 366         } while (i < ARRAY_SIZE(labels->bits));
 367
 368         return 0;
 369 }
 370 #else
 371 #define ctnetlink_dump_labels(a, b) (0)
 372 #define ctnetlink_label_size(a) (0)
 373 #endif
 374
 375 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 376
 377 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 378 {
 379         struct nlattr *nest_parms;
 380
 381         if (!(ct->status & IPS_EXPECTED))
 382                 return 0;
 383
 384         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 385         if (!nest_parms)
 386                 goto nla_put_failure;
 387         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 388                 goto nla_put_failure;
 389         nla_nest_end(skb, nest_parms);
 390
 391         return 0;
 392
 393 nla_put_failure:
 394         return -1;
 395 }
 396
 397 static int
 398 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 399 {
 400         struct nlattr *nest_parms;
 401
 402         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 403         if (!nest_parms)
 404                 goto nla_put_failure;
 405
 406         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 407                          htonl(seq->correction_pos)) ||
 408             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 409                          htonl(seq->offset_before)) ||
 410             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 411                          htonl(seq->offset_after)))
 412                 goto nla_put_failure;
 413
 414         nla_nest_end(skb, nest_parms);
 415
 416         return 0;
 417
 418 nla_put_failure:
 419         return -1;
 420 }
 421
 422 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 423 {
 424         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 425         struct nf_ct_seqadj *seq;
 426
 427         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 428                 return 0;
 429
 430         spin_lock_bh(&ct->lock);
 431         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 432         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 433                 goto err;
 434
 435         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 436         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 437                 goto err;
 438
 439         spin_unlock_bh(&ct->lock);
 440         return 0;
 441 err:
 442         spin_unlock_bh(&ct->lock);
 443         return -1;
 444 }
 445
 446 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 447 {
 448         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 449                 goto nla_put_failure;
 450         return 0;
 451
 452 nla_put_failure:
 453         return -1;
 454 }
 455
 456 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 457 {
 458         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 459                 goto nla_put_failure;
 460         return 0;
 461
 462 nla_put_failure:
 463         return -1;
 464 }
 465
 466 static int
 467 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 468                     struct nf_conn *ct)
 469 {
 470         const struct nf_conntrack_zone *zone;
 471         struct nlmsghdr *nlh;
 472         struct nfgenmsg *nfmsg;
 473         struct nlattr *nest_parms;
 474         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 475
 476         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 477         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 478         if (nlh == NULL)
 479                 goto nlmsg_failure;
 480
 481         nfmsg = nlmsg_data(nlh);
 482         nfmsg->nfgen_family = nf_ct_l3num(ct);
 483         nfmsg->version      = NFNETLINK_V0;
 484         nfmsg->res_id       = 0;
 485
 486         zone = nf_ct_zone(ct);
 487
 488         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 489         if (!nest_parms)
 490                 goto nla_put_failure;
 491         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 492                 goto nla_put_failure;
 493         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 494                                    NF_CT_ZONE_DIR_ORIG) < 0)
 495                 goto nla_put_failure;
 496         nla_nest_end(skb, nest_parms);
 497
 498         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 499         if (!nest_parms)
 500                 goto nla_put_failure;
 501         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 502                 goto nla_put_failure;
 503         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 504                                    NF_CT_ZONE_DIR_REPL) < 0)
 505                 goto nla_put_failure;
 506         nla_nest_end(skb, nest_parms);
 507
 508         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 509                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 510                 goto nla_put_failure;
 511
 512         if (ctnetlink_dump_status(skb, ct) < 0 ||
 513             ctnetlink_dump_timeout(skb, ct) < 0 ||
 514             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 515             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 516             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 517             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 518             ctnetlink_dump_mark(skb, ct) < 0 ||
 519             ctnetlink_dump_secctx(skb, ct) < 0 ||
 520             ctnetlink_dump_labels(skb, ct) < 0 ||
 521             ctnetlink_dump_id(skb, ct) < 0 ||
 522             ctnetlink_dump_use(skb, ct) < 0 ||
 523             ctnetlink_dump_master(skb, ct) < 0 ||
 524             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 525                 goto nla_put_failure;
 526
 527         nlmsg_end(skb, nlh);
 528         return skb->len;
 529
 530 nlmsg_failure:
 531 nla_put_failure:
 532         nlmsg_cancel(skb, nlh);
 533         return -1;
 534 }
 535
 536 static inline size_t ctnetlink_proto_size(const struct nf_conn *ct)
 537 {
 538         struct nf_conntrack_l3proto *l3proto;
 539         struct nf_conntrack_l4proto *l4proto;
 540         size_t len = 0;
 541
 542         rcu_read_lock();
 543         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 544         len += l3proto->nla_size;
 545
 546         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 547         len += l4proto->nla_size;
 548         rcu_read_unlock();
 549
 550         return len;
 551 }
 552
 553 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 554 {
 555         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 556                 return 0;
 557         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 558                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 559                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 560                ;
 561 }
 562
 563 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 564 {
 565 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 566         int len, ret;
 567
 568         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 569         if (ret)
 570                 return 0;
 571
 572         return nla_total_size(0) /* CTA_SECCTX */
 573                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 574 #else
 575         return 0;
 576 #endif
 577 }
 578
 579 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 580 {
 581 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 582         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 583                 return 0;
 584         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 585 #else
 586         return 0;
 587 #endif
 588 }
 589
 590 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 591 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 592 {
 593         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 594                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 595                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 596                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 597                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 598                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 599                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 600                + ctnetlink_acct_size(ct)
 601                + ctnetlink_timestamp_size(ct)
 602                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 603                + nla_total_size(0) /* CTA_PROTOINFO */
 604                + nla_total_size(0) /* CTA_HELP */
 605                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 606                + ctnetlink_secctx_size(ct)
 607 #ifdef CONFIG_NF_NAT_NEEDED
 608                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 609                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 610 #endif
 611 #ifdef CONFIG_NF_CONNTRACK_MARK
 612                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 613 #endif
 614 #ifdef CONFIG_NF_CONNTRACK_ZONES
 615                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 616 #endif
 617                + ctnetlink_proto_size(ct)
 618                + ctnetlink_label_size(ct)
 619                ;
 620 }
 621
 622 static int
 623 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 624 {
 625         const struct nf_conntrack_zone *zone;
 626         struct net *net;
 627         struct nlmsghdr *nlh;
 628         struct nfgenmsg *nfmsg;
 629         struct nlattr *nest_parms;
 630         struct nf_conn *ct = item->ct;
 631         struct sk_buff *skb;
 632         unsigned int type;
 633         unsigned int flags = 0, group;
 634         int err;
 635
 636         if (events & (1 << IPCT_DESTROY)) {
 637                 type = IPCTNL_MSG_CT_DELETE;
 638                 group = NFNLGRP_CONNTRACK_DESTROY;
 639         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 640                 type = IPCTNL_MSG_CT_NEW;
 641                 flags = NLM_F_CREATE|NLM_F_EXCL;
 642                 group = NFNLGRP_CONNTRACK_NEW;
 643         } else  if (events) {
 644                 type = IPCTNL_MSG_CT_NEW;
 645                 group = NFNLGRP_CONNTRACK_UPDATE;
 646         } else
 647                 return 0;
 648
 649         net = nf_ct_net(ct);
 650         if (!item->report && !nfnetlink_has_listeners(net, group))
 651                 return 0;
 652
 653         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 654         if (skb == NULL)
 655                 goto errout;
 656
 657         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 658         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 659         if (nlh == NULL)
 660                 goto nlmsg_failure;
 661
 662         nfmsg = nlmsg_data(nlh);
 663         nfmsg->nfgen_family = nf_ct_l3num(ct);
 664         nfmsg->version  = NFNETLINK_V0;
 665         nfmsg->res_id   = 0;
 666
 667         rcu_read_lock();
 668         zone = nf_ct_zone(ct);
 669
 670         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 671         if (!nest_parms)
 672                 goto nla_put_failure;
 673         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 674                 goto nla_put_failure;
 675         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 676                                    NF_CT_ZONE_DIR_ORIG) < 0)
 677                 goto nla_put_failure;
 678         nla_nest_end(skb, nest_parms);
 679
 680         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 681         if (!nest_parms)
 682                 goto nla_put_failure;
 683         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 684                 goto nla_put_failure;
 685         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 686                                    NF_CT_ZONE_DIR_REPL) < 0)
 687                 goto nla_put_failure;
 688         nla_nest_end(skb, nest_parms);
 689
 690         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 691                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 692                 goto nla_put_failure;
 693
 694         if (ctnetlink_dump_id(skb, ct) < 0)
 695                 goto nla_put_failure;
 696
 697         if (ctnetlink_dump_status(skb, ct) < 0)
 698                 goto nla_put_failure;
 699
 700         if (events & (1 << IPCT_DESTROY)) {
 701                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 702                     ctnetlink_dump_timestamp(skb, ct) < 0)
 703                         goto nla_put_failure;
 704         } else {
 705                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 706                         goto nla_put_failure;
 707
 708                 if (events & (1 << IPCT_PROTOINFO)
 709                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 710                         goto nla_put_failure;
 711
 712                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 713                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 714                         goto nla_put_failure;
 715
 716 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 717                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 718                     && ctnetlink_dump_secctx(skb, ct) < 0)
 719                         goto nla_put_failure;
 720 #endif
 721                 if (events & (1 << IPCT_LABEL) &&
 722                      ctnetlink_dump_labels(skb, ct) < 0)
 723                         goto nla_put_failure;
 724
 725                 if (events & (1 << IPCT_RELATED) &&
 726                     ctnetlink_dump_master(skb, ct) < 0)
 727                         goto nla_put_failure;
 728
 729                 if (events & (1 << IPCT_SEQADJ) &&
 730                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 731                         goto nla_put_failure;
 732         }
 733
 734 #ifdef CONFIG_NF_CONNTRACK_MARK
 735         if ((events & (1 << IPCT_MARK) || ct->mark)
 736             && ctnetlink_dump_mark(skb, ct) < 0)
 737                 goto nla_put_failure;
 738 #endif
 739         rcu_read_unlock();
 740
 741         nlmsg_end(skb, nlh);
 742         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 743                              GFP_ATOMIC);
 744         if (err == -ENOBUFS || err == -EAGAIN)
 745                 return -ENOBUFS;
 746
 747         return 0;
 748
 749 nla_put_failure:
 750         rcu_read_unlock();
 751         nlmsg_cancel(skb, nlh);
 752 nlmsg_failure:
 753         kfree_skb(skb);
 754 errout:
 755         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 756                 return -ENOBUFS;
 757
 758         return 0;
 759 }
 760 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 761
 762 static int ctnetlink_done(struct netlink_callback *cb)
 763 {
 764         if (cb->args[1])
 765                 nf_ct_put((struct nf_conn *)cb->args[1]);
 766         kfree(cb->data);
 767         return 0;
 768 }
 769
 770 struct ctnetlink_filter {
 771         struct {
 772                 u_int32_t val;
 773                 u_int32_t mask;
 774         } mark;
 775 };
 776
 777 static struct ctnetlink_filter *
 778 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 779 {
 780 #ifdef CONFIG_NF_CONNTRACK_MARK
 781         struct ctnetlink_filter *filter;
 782
 783         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 784         if (filter == NULL)
 785                 return ERR_PTR(-ENOMEM);
 786
 787         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 788         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 789
 790         return filter;
 791 #else
 792         return ERR_PTR(-EOPNOTSUPP);
 793 #endif
 794 }
 795
 796 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 797 {
 798         struct ctnetlink_filter *filter = data;
 799
 800         if (filter == NULL)
 801                 return 1;
 802
 803 #ifdef CONFIG_NF_CONNTRACK_MARK
 804         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 805                 return 1;
 806 #endif
 807
 808         return 0;
 809 }
 810
 811 static int
 812 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 813 {
 814         struct net *net = sock_net(skb->sk);
 815         struct nf_conn *ct, *last;
 816         struct nf_conntrack_tuple_hash *h;
 817         struct hlist_nulls_node *n;
 818         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 819         u_int8_t l3proto = nfmsg->nfgen_family;
 820         struct nf_conn *nf_ct_evict[8];
 821         int res, i;
 822         spinlock_t *lockp;
 823
 824         last = (struct nf_conn *)cb->args[1];
 825         i = 0;
 826
 827         local_bh_disable();
 828         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 829 restart:
 830                 while (i) {
 831                         i--;
 832                         if (nf_ct_should_gc(nf_ct_evict[i]))
 833                                 nf_ct_kill(nf_ct_evict[i]);
 834                         nf_ct_put(nf_ct_evict[i]);
 835                 }
 836
 837                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 838                 nf_conntrack_lock(lockp);
 839                 if (cb->args[0] >= nf_conntrack_htable_size) {
 840                         spin_unlock(lockp);
 841                         goto out;
 842                 }
 843                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 844                                            hnnode) {
 845                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 846                                 continue;
 847                         ct = nf_ct_tuplehash_to_ctrack(h);
 848                         if (nf_ct_is_expired(ct)) {
 849                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 850                                     atomic_inc_not_zero(&ct->ct_general.use))
 851                                         nf_ct_evict[i++] = ct;
 852                                 continue;
 853                         }
 854
 855                         if (!net_eq(net, nf_ct_net(ct)))
 856                                 continue;
 857
 858                         /* Dump entries of a given L3 protocol number.
 859                          * If it is not specified, ie. l3proto == 0,
 860                          * then dump everything. */
 861                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 862                                 continue;
 863                         if (cb->args[1]) {
 864                                 if (ct != last)
 865                                         continue;
 866                                 cb->args[1] = 0;
 867                         }
 868                         if (!ctnetlink_filter_match(ct, cb->data))
 869                                 continue;
 870
 871                         rcu_read_lock();
 872                         res =
 873                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 874                                             cb->nlh->nlmsg_seq,
 875                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 876                                             ct);
 877                         rcu_read_unlock();
 878                         if (res < 0) {
 879                                 nf_conntrack_get(&ct->ct_general);
 880                                 cb->args[1] = (unsigned long)ct;
 881                                 spin_unlock(lockp);
 882                                 goto out;
 883                         }
 884                 }
 885                 spin_unlock(lockp);
 886                 if (cb->args[1]) {
 887                         cb->args[1] = 0;
 888                         goto restart;
 889                 }
 890         }
 891 out:
 892         local_bh_enable();
 893         if (last)
 894                 nf_ct_put(last);
 895
 896         while (i) {
 897                 i--;
 898                 if (nf_ct_should_gc(nf_ct_evict[i]))
 899                         nf_ct_kill(nf_ct_evict[i]);
 900                 nf_ct_put(nf_ct_evict[i]);
 901         }
 902
 903         return skb->len;
 904 }
 905
 906 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
 907                                     struct nf_conntrack_tuple *tuple)
 908 {
 909         struct nlattr *tb[CTA_IP_MAX+1];
 910         struct nf_conntrack_l3proto *l3proto;
 911         int ret = 0;
 912
 913         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL, NULL);
 914         if (ret < 0)
 915                 return ret;
 916
 917         rcu_read_lock();
 918         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 919
 920         if (likely(l3proto->nlattr_to_tuple)) {
 921                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 922                                           l3proto->nla_policy, NULL);
 923                 if (ret == 0)
 924                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 925         }
 926
 927         rcu_read_unlock();
 928
 929         return ret;
 930 }
 931
 932 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 933         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 934 };
 935
 936 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
 937                                        struct nf_conntrack_tuple *tuple)
 938 {
 939         struct nlattr *tb[CTA_PROTO_MAX+1];
 940         struct nf_conntrack_l4proto *l4proto;
 941         int ret = 0;
 942
 943         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy,
 944                                NULL);
 945         if (ret < 0)
 946                 return ret;
 947
 948         if (!tb[CTA_PROTO_NUM])
 949                 return -EINVAL;
 950         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 951
 952         rcu_read_lock();
 953         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 954
 955         if (likely(l4proto->nlattr_to_tuple)) {
 956                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 957                                           l4proto->nla_policy, NULL);
 958                 if (ret == 0)
 959                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 960         }
 961
 962         rcu_read_unlock();
 963
 964         return ret;
 965 }
 966
 967 static int
 968 ctnetlink_parse_zone(const struct nlattr *attr,
 969                      struct nf_conntrack_zone *zone)
 970 {
 971         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
 972                         NF_CT_DEFAULT_ZONE_DIR, 0);
 973 #ifdef CONFIG_NF_CONNTRACK_ZONES
 974         if (attr)
 975                 zone->id = ntohs(nla_get_be16(attr));
 976 #else
 977         if (attr)
 978                 return -EOPNOTSUPP;
 979 #endif
 980         return 0;
 981 }
 982
 983 static int
 984 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
 985                            struct nf_conntrack_zone *zone)
 986 {
 987         int ret;
 988
 989         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
 990                 return -EINVAL;
 991
 992         ret = ctnetlink_parse_zone(attr, zone);
 993         if (ret < 0)
 994                 return ret;
 995
 996         if (type == CTA_TUPLE_REPLY)
 997                 zone->dir = NF_CT_ZONE_DIR_REPL;
 998         else
 999                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1000
1001         return 0;
1002 }
1003
1004 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1005         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1006         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1007         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1008 };
1009
1010 static int
1011 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1012                       struct nf_conntrack_tuple *tuple, u32 type,
1013                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1014 {
1015         struct nlattr *tb[CTA_TUPLE_MAX+1];
1016         int err;
1017
1018         memset(tuple, 0, sizeof(*tuple));
1019
1020         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1021                                NULL);
1022         if (err < 0)
1023                 return err;
1024
1025         if (!tb[CTA_TUPLE_IP])
1026                 return -EINVAL;
1027
1028         tuple->src.l3num = l3num;
1029
1030         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1031         if (err < 0)
1032                 return err;
1033
1034         if (!tb[CTA_TUPLE_PROTO])
1035                 return -EINVAL;
1036
1037         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1038         if (err < 0)
1039                 return err;
1040
1041         if (tb[CTA_TUPLE_ZONE]) {
1042                 if (!zone)
1043                         return -EINVAL;
1044
1045                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1046                                                  type, zone);
1047                 if (err < 0)
1048                         return err;
1049         }
1050
1051         /* orig and expect tuples get DIR_ORIGINAL */
1052         if (type == CTA_TUPLE_REPLY)
1053                 tuple->dst.dir = IP_CT_DIR_REPLY;
1054         else
1055                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1056
1057         return 0;
1058 }
1059
1060 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1061         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1062                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1063 };
1064
1065 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1066                                 struct nlattr **helpinfo)
1067 {
1068         int err;
1069         struct nlattr *tb[CTA_HELP_MAX+1];
1070
1071         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1072         if (err < 0)
1073                 return err;
1074
1075         if (!tb[CTA_HELP_NAME])
1076                 return -EINVAL;
1077
1078         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1079
1080         if (tb[CTA_HELP_INFO])
1081                 *helpinfo = tb[CTA_HELP_INFO];
1082
1083         return 0;
1084 }
1085
1086 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1087         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1088         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1089         [CTA_STATUS]            = { .type = NLA_U32 },
1090         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1091         [CTA_HELP]              = { .type = NLA_NESTED },
1092         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1093         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1094         [CTA_MARK]              = { .type = NLA_U32 },
1095         [CTA_ID]                = { .type = NLA_U32 },
1096         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1097         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1098         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1099         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1100         [CTA_ZONE]              = { .type = NLA_U16 },
1101         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1102         [CTA_LABELS]            = { .type = NLA_BINARY,
1103                                     .len = NF_CT_LABELS_MAX_SIZE },
1104         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1105                                     .len = NF_CT_LABELS_MAX_SIZE },
1106 };
1107
1108 static int ctnetlink_flush_conntrack(struct net *net,
1109                                      const struct nlattr * const cda[],
1110                                      u32 portid, int report)
1111 {
1112         struct ctnetlink_filter *filter = NULL;
1113
1114         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1115                 filter = ctnetlink_alloc_filter(cda);
1116                 if (IS_ERR(filter))
1117                         return PTR_ERR(filter);
1118         }
1119
1120         nf_ct_iterate_cleanup(net, ctnetlink_filter_match, filter,
1121                               portid, report);
1122         kfree(filter);
1123
1124         return 0;
1125 }
1126
1127 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1128                                    struct sk_buff *skb,
1129                                    const struct nlmsghdr *nlh,
1130                                    const struct nlattr * const cda[])
1131 {
1132         struct nf_conntrack_tuple_hash *h;
1133         struct nf_conntrack_tuple tuple;
1134         struct nf_conn *ct;
1135         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1136         u_int8_t u3 = nfmsg->nfgen_family;
1137         struct nf_conntrack_zone zone;
1138         int err;
1139
1140         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1141         if (err < 0)
1142                 return err;
1143
1144         if (cda[CTA_TUPLE_ORIG])
1145                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1146                                             u3, &zone);
1147         else if (cda[CTA_TUPLE_REPLY])
1148                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1149                                             u3, &zone);
1150         else {
1151                 return ctnetlink_flush_conntrack(net, cda,
1152                                                  NETLINK_CB(skb).portid,
1153                                                  nlmsg_report(nlh));
1154         }
1155
1156         if (err < 0)
1157                 return err;
1158
1159         h = nf_conntrack_find_get(net, &zone, &tuple);
1160         if (!h)
1161                 return -ENOENT;
1162
1163         ct = nf_ct_tuplehash_to_ctrack(h);
1164
1165         if (cda[CTA_ID]) {
1166                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1167                 if (id != (u32)(unsigned long)ct) {
1168                         nf_ct_put(ct);
1169                         return -ENOENT;
1170                 }
1171         }
1172
1173         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1174         nf_ct_put(ct);
1175
1176         return 0;
1177 }
1178
1179 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1180                                    struct sk_buff *skb,
1181                                    const struct nlmsghdr *nlh,
1182                                    const struct nlattr * const cda[])
1183 {
1184         struct nf_conntrack_tuple_hash *h;
1185         struct nf_conntrack_tuple tuple;
1186         struct nf_conn *ct;
1187         struct sk_buff *skb2 = NULL;
1188         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1189         u_int8_t u3 = nfmsg->nfgen_family;
1190         struct nf_conntrack_zone zone;
1191         int err;
1192
1193         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1194                 struct netlink_dump_control c = {
1195                         .dump = ctnetlink_dump_table,
1196                         .done = ctnetlink_done,
1197                 };
1198
1199                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1200                         struct ctnetlink_filter *filter;
1201
1202                         filter = ctnetlink_alloc_filter(cda);
1203                         if (IS_ERR(filter))
1204                                 return PTR_ERR(filter);
1205
1206                         c.data = filter;
1207                 }
1208                 return netlink_dump_start(ctnl, skb, nlh, &c);
1209         }
1210
1211         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1212         if (err < 0)
1213                 return err;
1214
1215         if (cda[CTA_TUPLE_ORIG])
1216                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1217                                             u3, &zone);
1218         else if (cda[CTA_TUPLE_REPLY])
1219                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1220                                             u3, &zone);
1221         else
1222                 return -EINVAL;
1223
1224         if (err < 0)
1225                 return err;
1226
1227         h = nf_conntrack_find_get(net, &zone, &tuple);
1228         if (!h)
1229                 return -ENOENT;
1230
1231         ct = nf_ct_tuplehash_to_ctrack(h);
1232
1233         err = -ENOMEM;
1234         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1235         if (skb2 == NULL) {
1236                 nf_ct_put(ct);
1237                 return -ENOMEM;
1238         }
1239
1240         rcu_read_lock();
1241         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1242                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1243         rcu_read_unlock();
1244         nf_ct_put(ct);
1245         if (err <= 0)
1246                 goto free;
1247
1248         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1249         if (err < 0)
1250                 goto out;
1251
1252         return 0;
1253
1254 free:
1255         kfree_skb(skb2);
1256 out:
1257         /* this avoids a loop in nfnetlink. */
1258         return err == -EAGAIN ? -ENOBUFS : err;
1259 }
1260
1261 static int ctnetlink_done_list(struct netlink_callback *cb)
1262 {
1263         if (cb->args[1])
1264                 nf_ct_put((struct nf_conn *)cb->args[1]);
1265         return 0;
1266 }
1267
1268 static int
1269 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1270 {
1271         struct nf_conn *ct, *last;
1272         struct nf_conntrack_tuple_hash *h;
1273         struct hlist_nulls_node *n;
1274         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1275         u_int8_t l3proto = nfmsg->nfgen_family;
1276         int res;
1277         int cpu;
1278         struct hlist_nulls_head *list;
1279         struct net *net = sock_net(skb->sk);
1280
1281         if (cb->args[2])
1282                 return 0;
1283
1284         last = (struct nf_conn *)cb->args[1];
1285
1286         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1287                 struct ct_pcpu *pcpu;
1288
1289                 if (!cpu_possible(cpu))
1290                         continue;
1291
1292                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1293                 spin_lock_bh(&pcpu->lock);
1294                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1295 restart:
1296                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1297                         ct = nf_ct_tuplehash_to_ctrack(h);
1298                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1299                                 continue;
1300                         if (cb->args[1]) {
1301                                 if (ct != last)
1302                                         continue;
1303                                 cb->args[1] = 0;
1304                         }
1305                         rcu_read_lock();
1306                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1307                                                   cb->nlh->nlmsg_seq,
1308                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1309                                                   ct);
1310                         rcu_read_unlock();
1311                         if (res < 0) {
1312                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1313                                         continue;
1314                                 cb->args[0] = cpu;
1315                                 cb->args[1] = (unsigned long)ct;
1316                                 spin_unlock_bh(&pcpu->lock);
1317                                 goto out;
1318                         }
1319                 }
1320                 if (cb->args[1]) {
1321                         cb->args[1] = 0;
1322                         goto restart;
1323                 }
1324                 spin_unlock_bh(&pcpu->lock);
1325         }
1326         cb->args[2] = 1;
1327 out:
1328         if (last)
1329                 nf_ct_put(last);
1330
1331         return skb->len;
1332 }
1333
1334 static int
1335 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1336 {
1337         return ctnetlink_dump_list(skb, cb, true);
1338 }
1339
1340 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1341                                   struct sk_buff *skb,
1342                                   const struct nlmsghdr *nlh,
1343                                   const struct nlattr * const cda[])
1344 {
1345         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1346                 struct netlink_dump_control c = {
1347                         .dump = ctnetlink_dump_dying,
1348                         .done = ctnetlink_done_list,
1349                 };
1350                 return netlink_dump_start(ctnl, skb, nlh, &c);
1351         }
1352
1353         return -EOPNOTSUPP;
1354 }
1355
1356 static int
1357 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1358 {
1359         return ctnetlink_dump_list(skb, cb, false);
1360 }
1361
1362 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1363                                         struct sk_buff *skb,
1364                                         const struct nlmsghdr *nlh,
1365                                         const struct nlattr * const cda[])
1366 {
1367         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1368                 struct netlink_dump_control c = {
1369                         .dump = ctnetlink_dump_unconfirmed,
1370                         .done = ctnetlink_done_list,
1371                 };
1372                 return netlink_dump_start(ctnl, skb, nlh, &c);
1373         }
1374
1375         return -EOPNOTSUPP;
1376 }
1377
1378 #ifdef CONFIG_NF_NAT_NEEDED
1379 static int
1380 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1381                           enum nf_nat_manip_type manip,
1382                           const struct nlattr *attr)
1383 {
1384         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1385         int err;
1386
1387         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1388         if (!parse_nat_setup) {
1389 #ifdef CONFIG_MODULES
1390                 rcu_read_unlock();
1391                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1392                 if (request_module("nf-nat") < 0) {
1393                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1394                         rcu_read_lock();
1395                         return -EOPNOTSUPP;
1396                 }
1397                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1398                 rcu_read_lock();
1399                 if (nfnetlink_parse_nat_setup_hook)
1400                         return -EAGAIN;
1401 #endif
1402                 return -EOPNOTSUPP;
1403         }
1404
1405         err = parse_nat_setup(ct, manip, attr);
1406         if (err == -EAGAIN) {
1407 #ifdef CONFIG_MODULES
1408                 rcu_read_unlock();
1409                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1410                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1411                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1412                         rcu_read_lock();
1413                         return -EOPNOTSUPP;
1414                 }
1415                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1416                 rcu_read_lock();
1417 #else
1418                 err = -EOPNOTSUPP;
1419 #endif
1420         }
1421         return err;
1422 }
1423 #endif
1424
1425 static void
1426 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1427                           unsigned long off)
1428 {
1429         unsigned int bit;
1430
1431         /* Ignore these unchangable bits */
1432         on &= ~IPS_UNCHANGEABLE_MASK;
1433         off &= ~IPS_UNCHANGEABLE_MASK;
1434
1435         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1436                 if (on & (1 << bit))
1437                         set_bit(bit, &ct->status);
1438                 else if (off & (1 << bit))
1439                         clear_bit(bit, &ct->status);
1440         }
1441 }
1442
1443 static int
1444 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1445 {
1446         unsigned long d;
1447         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1448         d = ct->status ^ status;
1449
1450         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1451                 /* unchangeable */
1452                 return -EBUSY;
1453
1454         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1455                 /* SEEN_REPLY bit can only be set */
1456                 return -EBUSY;
1457
1458         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1459                 /* ASSURED bit can only be set */
1460                 return -EBUSY;
1461
1462         __ctnetlink_change_status(ct, status, 0);
1463         return 0;
1464 }
1465
1466 static int
1467 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1468 {
1469 #ifdef CONFIG_NF_NAT_NEEDED
1470         int ret;
1471
1472         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1473                 return 0;
1474
1475         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1476                                         cda[CTA_NAT_DST]);
1477         if (ret < 0)
1478                 return ret;
1479
1480         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1481                                         cda[CTA_NAT_SRC]);
1482         return ret;
1483 #else
1484         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1485                 return 0;
1486         return -EOPNOTSUPP;
1487 #endif
1488 }
1489
1490 static int ctnetlink_change_helper(struct nf_conn *ct,
1491                                    const struct nlattr * const cda[])
1492 {
1493         struct nf_conntrack_helper *helper;
1494         struct nf_conn_help *help = nfct_help(ct);
1495         char *helpname = NULL;
1496         struct nlattr *helpinfo = NULL;
1497         int err;
1498
1499         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1500         if (err < 0)
1501                 return err;
1502
1503         /* don't change helper of sibling connections */
1504         if (ct->master) {
1505                 /* If we try to change the helper to the same thing twice,
1506                  * treat the second attempt as a no-op instead of returning
1507                  * an error.
1508                  */
1509                 err = -EBUSY;
1510                 if (help) {
1511                         rcu_read_lock();
1512                         helper = rcu_dereference(help->helper);
1513                         if (helper && !strcmp(helper->name, helpname))
1514                                 err = 0;
1515                         rcu_read_unlock();
1516                 }
1517
1518                 return err;
1519         }
1520
1521         if (!strcmp(helpname, "")) {
1522                 if (help && help->helper) {
1523                         /* we had a helper before ... */
1524                         nf_ct_remove_expectations(ct);
1525                         RCU_INIT_POINTER(help->helper, NULL);
1526                 }
1527
1528                 return 0;
1529         }
1530
1531         rcu_read_lock();
1532         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1533                                             nf_ct_protonum(ct));
1534         if (helper == NULL) {
1535                 rcu_read_unlock();
1536                 return -EOPNOTSUPP;
1537         }
1538
1539         if (help) {
1540                 if (help->helper == helper) {
1541                         /* update private helper data if allowed. */
1542                         if (helper->from_nlattr)
1543                                 helper->from_nlattr(helpinfo, ct);
1544                         err = 0;
1545                 } else
1546                         err = -EBUSY;
1547         } else {
1548                 /* we cannot set a helper for an existing conntrack */
1549                 err = -EOPNOTSUPP;
1550         }
1551
1552         rcu_read_unlock();
1553         return err;
1554 }
1555
1556 static int ctnetlink_change_timeout(struct nf_conn *ct,
1557                                     const struct nlattr * const cda[])
1558 {
1559         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1560
1561         ct->timeout = nfct_time_stamp + timeout * HZ;
1562
1563         if (test_bit(IPS_DYING_BIT, &ct->status))
1564                 return -ETIME;
1565
1566         return 0;
1567 }
1568
1569 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1570         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1571         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1572         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1573 };
1574
1575 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1576                                       const struct nlattr * const cda[])
1577 {
1578         const struct nlattr *attr = cda[CTA_PROTOINFO];
1579         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1580         struct nf_conntrack_l4proto *l4proto;
1581         int err = 0;
1582
1583         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1584                                NULL);
1585         if (err < 0)
1586                 return err;
1587
1588         rcu_read_lock();
1589         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1590         if (l4proto->from_nlattr)
1591                 err = l4proto->from_nlattr(tb, ct);
1592         rcu_read_unlock();
1593
1594         return err;
1595 }
1596
1597 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1598         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1599         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1600         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1601 };
1602
1603 static int change_seq_adj(struct nf_ct_seqadj *seq,
1604                           const struct nlattr * const attr)
1605 {
1606         int err;
1607         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1608
1609         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1610         if (err < 0)
1611                 return err;
1612
1613         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1614                 return -EINVAL;
1615
1616         seq->correction_pos =
1617                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1618
1619         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1620                 return -EINVAL;
1621
1622         seq->offset_before =
1623                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1624
1625         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1626                 return -EINVAL;
1627
1628         seq->offset_after =
1629                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1630
1631         return 0;
1632 }
1633
1634 static int
1635 ctnetlink_change_seq_adj(struct nf_conn *ct,
1636                          const struct nlattr * const cda[])
1637 {
1638         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1639         int ret = 0;
1640
1641         if (!seqadj)
1642                 return 0;
1643
1644         spin_lock_bh(&ct->lock);
1645         if (cda[CTA_SEQ_ADJ_ORIG]) {
1646                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1647                                      cda[CTA_SEQ_ADJ_ORIG]);
1648                 if (ret < 0)
1649                         goto err;
1650
1651                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1652         }
1653
1654         if (cda[CTA_SEQ_ADJ_REPLY]) {
1655                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1656                                      cda[CTA_SEQ_ADJ_REPLY]);
1657                 if (ret < 0)
1658                         goto err;
1659
1660                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1661         }
1662
1663         spin_unlock_bh(&ct->lock);
1664         return 0;
1665 err:
1666         spin_unlock_bh(&ct->lock);
1667         return ret;
1668 }
1669
1670 static int
1671 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1672 {
1673 #ifdef CONFIG_NF_CONNTRACK_LABELS
1674         size_t len = nla_len(cda[CTA_LABELS]);
1675         const void *mask = cda[CTA_LABELS_MASK];
1676
1677         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1678                 return -EINVAL;
1679
1680         if (mask) {
1681                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1682                     nla_len(cda[CTA_LABELS_MASK]) != len)
1683                         return -EINVAL;
1684                 mask = nla_data(cda[CTA_LABELS_MASK]);
1685         }
1686
1687         len /= sizeof(u32);
1688
1689         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1690 #else
1691         return -EOPNOTSUPP;
1692 #endif
1693 }
1694
1695 static int
1696 ctnetlink_change_conntrack(struct nf_conn *ct,
1697                            const struct nlattr * const cda[])
1698 {
1699         int err;
1700
1701         /* only allow NAT changes and master assignation for new conntracks */
1702         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1703                 return -EOPNOTSUPP;
1704
1705         if (cda[CTA_HELP]) {
1706                 err = ctnetlink_change_helper(ct, cda);
1707                 if (err < 0)
1708                         return err;
1709         }
1710
1711         if (cda[CTA_TIMEOUT]) {
1712                 err = ctnetlink_change_timeout(ct, cda);
1713                 if (err < 0)
1714                         return err;
1715         }
1716
1717         if (cda[CTA_STATUS]) {
1718                 err = ctnetlink_change_status(ct, cda);
1719                 if (err < 0)
1720                         return err;
1721         }
1722
1723         if (cda[CTA_PROTOINFO]) {
1724                 err = ctnetlink_change_protoinfo(ct, cda);
1725                 if (err < 0)
1726                         return err;
1727         }
1728
1729 #if defined(CONFIG_NF_CONNTRACK_MARK)
1730         if (cda[CTA_MARK])
1731                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1732 #endif
1733
1734         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1735                 err = ctnetlink_change_seq_adj(ct, cda);
1736                 if (err < 0)
1737                         return err;
1738         }
1739
1740         if (cda[CTA_LABELS]) {
1741                 err = ctnetlink_attach_labels(ct, cda);
1742                 if (err < 0)
1743                         return err;
1744         }
1745
1746         return 0;
1747 }
1748
1749 static struct nf_conn *
1750 ctnetlink_create_conntrack(struct net *net,
1751                            const struct nf_conntrack_zone *zone,
1752                            const struct nlattr * const cda[],
1753                            struct nf_conntrack_tuple *otuple,
1754                            struct nf_conntrack_tuple *rtuple,
1755                            u8 u3)
1756 {
1757         struct nf_conn *ct;
1758         int err = -EINVAL;
1759         struct nf_conntrack_helper *helper;
1760         struct nf_conn_tstamp *tstamp;
1761
1762         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1763         if (IS_ERR(ct))
1764                 return ERR_PTR(-ENOMEM);
1765
1766         if (!cda[CTA_TIMEOUT])
1767                 goto err1;
1768
1769         ct->timeout = nfct_time_stamp + ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1770
1771         rcu_read_lock();
1772         if (cda[CTA_HELP]) {
1773                 char *helpname = NULL;
1774                 struct nlattr *helpinfo = NULL;
1775
1776                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1777                 if (err < 0)
1778                         goto err2;
1779
1780                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1781                                                     nf_ct_protonum(ct));
1782                 if (helper == NULL) {
1783                         rcu_read_unlock();
1784 #ifdef CONFIG_MODULES
1785                         if (request_module("nfct-helper-%s", helpname) < 0) {
1786                                 err = -EOPNOTSUPP;
1787                                 goto err1;
1788                         }
1789
1790                         rcu_read_lock();
1791                         helper = __nf_conntrack_helper_find(helpname,
1792                                                             nf_ct_l3num(ct),
1793                                                             nf_ct_protonum(ct));
1794                         if (helper) {
1795                                 err = -EAGAIN;
1796                                 goto err2;
1797                         }
1798                         rcu_read_unlock();
1799 #endif
1800                         err = -EOPNOTSUPP;
1801                         goto err1;
1802                 } else {
1803                         struct nf_conn_help *help;
1804
1805                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1806                         if (help == NULL) {
1807                                 err = -ENOMEM;
1808                                 goto err2;
1809                         }
1810                         /* set private helper data if allowed. */
1811                         if (helper->from_nlattr)
1812                                 helper->from_nlattr(helpinfo, ct);
1813
1814                         /* not in hash table yet so not strictly necessary */
1815                         RCU_INIT_POINTER(help->helper, helper);
1816                 }
1817         } else {
1818                 /* try an implicit helper assignation */
1819                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1820                 if (err < 0)
1821                         goto err2;
1822         }
1823
1824         err = ctnetlink_setup_nat(ct, cda);
1825         if (err < 0)
1826                 goto err2;
1827
1828         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1829         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1830         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1831         nf_ct_labels_ext_add(ct);
1832         nfct_seqadj_ext_add(ct);
1833         nfct_synproxy_ext_add(ct);
1834
1835         /* we must add conntrack extensions before confirmation. */
1836         ct->status |= IPS_CONFIRMED;
1837
1838         if (cda[CTA_STATUS]) {
1839                 err = ctnetlink_change_status(ct, cda);
1840                 if (err < 0)
1841                         goto err2;
1842         }
1843
1844         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1845                 err = ctnetlink_change_seq_adj(ct, cda);
1846                 if (err < 0)
1847                         goto err2;
1848         }
1849
1850         memset(&ct->proto, 0, sizeof(ct->proto));
1851         if (cda[CTA_PROTOINFO]) {
1852                 err = ctnetlink_change_protoinfo(ct, cda);
1853                 if (err < 0)
1854                         goto err2;
1855         }
1856
1857 #if defined(CONFIG_NF_CONNTRACK_MARK)
1858         if (cda[CTA_MARK])
1859                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1860 #endif
1861
1862         /* setup master conntrack: this is a confirmed expectation */
1863         if (cda[CTA_TUPLE_MASTER]) {
1864                 struct nf_conntrack_tuple master;
1865                 struct nf_conntrack_tuple_hash *master_h;
1866                 struct nf_conn *master_ct;
1867
1868                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1869                                             u3, NULL);
1870                 if (err < 0)
1871                         goto err2;
1872
1873                 master_h = nf_conntrack_find_get(net, zone, &master);
1874                 if (master_h == NULL) {
1875                         err = -ENOENT;
1876                         goto err2;
1877                 }
1878                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1879                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1880                 ct->master = master_ct;
1881         }
1882         tstamp = nf_conn_tstamp_find(ct);
1883         if (tstamp)
1884                 tstamp->start = ktime_get_real_ns();
1885
1886         err = nf_conntrack_hash_check_insert(ct);
1887         if (err < 0)
1888                 goto err2;
1889
1890         rcu_read_unlock();
1891
1892         return ct;
1893
1894 err2:
1895         rcu_read_unlock();
1896 err1:
1897         nf_conntrack_free(ct);
1898         return ERR_PTR(err);
1899 }
1900
1901 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
1902                                    struct sk_buff *skb,
1903                                    const struct nlmsghdr *nlh,
1904                                    const struct nlattr * const cda[])
1905 {
1906         struct nf_conntrack_tuple otuple, rtuple;
1907         struct nf_conntrack_tuple_hash *h = NULL;
1908         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1909         struct nf_conn *ct;
1910         u_int8_t u3 = nfmsg->nfgen_family;
1911         struct nf_conntrack_zone zone;
1912         int err;
1913
1914         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1915         if (err < 0)
1916                 return err;
1917
1918         if (cda[CTA_TUPLE_ORIG]) {
1919                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1920                                             u3, &zone);
1921                 if (err < 0)
1922                         return err;
1923         }
1924
1925         if (cda[CTA_TUPLE_REPLY]) {
1926                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1927                                             u3, &zone);
1928                 if (err < 0)
1929                         return err;
1930         }
1931
1932         if (cda[CTA_TUPLE_ORIG])
1933                 h = nf_conntrack_find_get(net, &zone, &otuple);
1934         else if (cda[CTA_TUPLE_REPLY])
1935                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1936
1937         if (h == NULL) {
1938                 err = -ENOENT;
1939                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1940                         enum ip_conntrack_events events;
1941
1942                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1943                                 return -EINVAL;
1944                         if (otuple.dst.protonum != rtuple.dst.protonum)
1945                                 return -EINVAL;
1946
1947                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1948                                                         &rtuple, u3);
1949                         if (IS_ERR(ct))
1950                                 return PTR_ERR(ct);
1951
1952                         err = 0;
1953                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1954                                 events = 1 << IPCT_RELATED;
1955                         else
1956                                 events = 1 << IPCT_NEW;
1957
1958                         if (cda[CTA_LABELS] &&
1959                             ctnetlink_attach_labels(ct, cda) == 0)
1960                                 events |= (1 << IPCT_LABEL);
1961
1962                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1963                                                       (1 << IPCT_ASSURED) |
1964                                                       (1 << IPCT_HELPER) |
1965                                                       (1 << IPCT_PROTOINFO) |
1966                                                       (1 << IPCT_SEQADJ) |
1967                                                       (1 << IPCT_MARK) | events,
1968                                                       ct, NETLINK_CB(skb).portid,
1969                                                       nlmsg_report(nlh));
1970                         nf_ct_put(ct);
1971                 }
1972
1973                 return err;
1974         }
1975         /* implicit 'else' */
1976
1977         err = -EEXIST;
1978         ct = nf_ct_tuplehash_to_ctrack(h);
1979         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1980                 err = ctnetlink_change_conntrack(ct, cda);
1981                 if (err == 0) {
1982                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1983                                                       (1 << IPCT_ASSURED) |
1984                                                       (1 << IPCT_HELPER) |
1985                                                       (1 << IPCT_LABEL) |
1986                                                       (1 << IPCT_PROTOINFO) |
1987                                                       (1 << IPCT_SEQADJ) |
1988                                                       (1 << IPCT_MARK),
1989                                                       ct, NETLINK_CB(skb).portid,
1990                                                       nlmsg_report(nlh));
1991                 }
1992         }
1993
1994         nf_ct_put(ct);
1995         return err;
1996 }
1997
1998 static int
1999 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2000                                 __u16 cpu, const struct ip_conntrack_stat *st)
2001 {
2002         struct nlmsghdr *nlh;
2003         struct nfgenmsg *nfmsg;
2004         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2005
2006         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2007                               IPCTNL_MSG_CT_GET_STATS_CPU);
2008         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2009         if (nlh == NULL)
2010                 goto nlmsg_failure;
2011
2012         nfmsg = nlmsg_data(nlh);
2013         nfmsg->nfgen_family = AF_UNSPEC;
2014         nfmsg->version      = NFNETLINK_V0;
2015         nfmsg->res_id       = htons(cpu);
2016
2017         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2018             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2019             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2020             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2021             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2022                                 htonl(st->insert_failed)) ||
2023             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2024             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2025             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2026             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2027                                 htonl(st->search_restart)))
2028                 goto nla_put_failure;
2029
2030         nlmsg_end(skb, nlh);
2031         return skb->len;
2032
2033 nla_put_failure:
2034 nlmsg_failure:
2035         nlmsg_cancel(skb, nlh);
2036         return -1;
2037 }
2038
2039 static int
2040 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2041 {
2042         int cpu;
2043         struct net *net = sock_net(skb->sk);
2044
2045         if (cb->args[0] == nr_cpu_ids)
2046                 return 0;
2047
2048         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2049                 const struct ip_conntrack_stat *st;
2050
2051                 if (!cpu_possible(cpu))
2052                         continue;
2053
2054                 st = per_cpu_ptr(net->ct.stat, cpu);
2055                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2056                                                     NETLINK_CB(cb->skb).portid,
2057                                                     cb->nlh->nlmsg_seq,
2058                                                     cpu, st) < 0)
2059                                 break;
2060         }
2061         cb->args[0] = cpu;
2062
2063         return skb->len;
2064 }
2065
2066 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2067                                  struct sk_buff *skb,
2068                                  const struct nlmsghdr *nlh,
2069                                  const struct nlattr * const cda[])
2070 {
2071         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2072                 struct netlink_dump_control c = {
2073                         .dump = ctnetlink_ct_stat_cpu_dump,
2074                 };
2075                 return netlink_dump_start(ctnl, skb, nlh, &c);
2076         }
2077
2078         return 0;
2079 }
2080
2081 static int
2082 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2083                             struct net *net)
2084 {
2085         struct nlmsghdr *nlh;
2086         struct nfgenmsg *nfmsg;
2087         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2088         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2089
2090         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2091         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2092         if (nlh == NULL)
2093                 goto nlmsg_failure;
2094
2095         nfmsg = nlmsg_data(nlh);
2096         nfmsg->nfgen_family = AF_UNSPEC;
2097         nfmsg->version      = NFNETLINK_V0;
2098         nfmsg->res_id       = 0;
2099
2100         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2101                 goto nla_put_failure;
2102
2103         nlmsg_end(skb, nlh);
2104         return skb->len;
2105
2106 nla_put_failure:
2107 nlmsg_failure:
2108         nlmsg_cancel(skb, nlh);
2109         return -1;
2110 }
2111
2112 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2113                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2114                              const struct nlattr * const cda[])
2115 {
2116         struct sk_buff *skb2;
2117         int err;
2118
2119         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2120         if (skb2 == NULL)
2121                 return -ENOMEM;
2122
2123         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2124                                           nlh->nlmsg_seq,
2125                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2126                                           sock_net(skb->sk));
2127         if (err <= 0)
2128                 goto free;
2129
2130         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2131         if (err < 0)
2132                 goto out;
2133
2134         return 0;
2135
2136 free:
2137         kfree_skb(skb2);
2138 out:
2139         /* this avoids a loop in nfnetlink. */
2140         return err == -EAGAIN ? -ENOBUFS : err;
2141 }
2142
2143 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2144         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2145         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2146         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2147         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2148         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2149         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2150                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2151         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2152         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2153         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2154         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2155         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2156 };
2157
2158 static struct nf_conntrack_expect *
2159 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2160                        struct nf_conntrack_helper *helper,
2161                        struct nf_conntrack_tuple *tuple,
2162                        struct nf_conntrack_tuple *mask);
2163
2164 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2165 static size_t
2166 ctnetlink_glue_build_size(const struct nf_conn *ct)
2167 {
2168         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2169                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2170                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2171                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2172                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2173                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2174                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2175                + nla_total_size(0) /* CTA_PROTOINFO */
2176                + nla_total_size(0) /* CTA_HELP */
2177                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2178                + ctnetlink_secctx_size(ct)
2179 #ifdef CONFIG_NF_NAT_NEEDED
2180                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2181                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2182 #endif
2183 #ifdef CONFIG_NF_CONNTRACK_MARK
2184                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2185 #endif
2186 #ifdef CONFIG_NF_CONNTRACK_ZONES
2187                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2188 #endif
2189                + ctnetlink_proto_size(ct)
2190                ;
2191 }
2192
2193 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2194                                              enum ip_conntrack_info *ctinfo)
2195 {
2196         return nf_ct_get(skb, ctinfo);
2197 }
2198
2199 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2200 {
2201         const struct nf_conntrack_zone *zone;
2202         struct nlattr *nest_parms;
2203
2204         rcu_read_lock();
2205         zone = nf_ct_zone(ct);
2206
2207         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2208         if (!nest_parms)
2209                 goto nla_put_failure;
2210         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2211                 goto nla_put_failure;
2212         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2213                                    NF_CT_ZONE_DIR_ORIG) < 0)
2214                 goto nla_put_failure;
2215         nla_nest_end(skb, nest_parms);
2216
2217         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2218         if (!nest_parms)
2219                 goto nla_put_failure;
2220         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2221                 goto nla_put_failure;
2222         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2223                                    NF_CT_ZONE_DIR_REPL) < 0)
2224                 goto nla_put_failure;
2225         nla_nest_end(skb, nest_parms);
2226
2227         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2228                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2229                 goto nla_put_failure;
2230
2231         if (ctnetlink_dump_id(skb, ct) < 0)
2232                 goto nla_put_failure;
2233
2234         if (ctnetlink_dump_status(skb, ct) < 0)
2235                 goto nla_put_failure;
2236
2237         if (ctnetlink_dump_timeout(skb, ct) < 0)
2238                 goto nla_put_failure;
2239
2240         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2241                 goto nla_put_failure;
2242
2243         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2244                 goto nla_put_failure;
2245
2246 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2247         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2248                 goto nla_put_failure;
2249 #endif
2250         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2251                 goto nla_put_failure;
2252
2253         if ((ct->status & IPS_SEQ_ADJUST) &&
2254             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2255                 goto nla_put_failure;
2256
2257 #ifdef CONFIG_NF_CONNTRACK_MARK
2258         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2259                 goto nla_put_failure;
2260 #endif
2261         if (ctnetlink_dump_labels(skb, ct) < 0)
2262                 goto nla_put_failure;
2263         rcu_read_unlock();
2264         return 0;
2265
2266 nla_put_failure:
2267         rcu_read_unlock();
2268         return -ENOSPC;
2269 }
2270
2271 static int
2272 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2273                      enum ip_conntrack_info ctinfo,
2274                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2275 {
2276         struct nlattr *nest_parms;
2277
2278         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2279         if (!nest_parms)
2280                 goto nla_put_failure;
2281
2282         if (__ctnetlink_glue_build(skb, ct) < 0)
2283                 goto nla_put_failure;
2284
2285         nla_nest_end(skb, nest_parms);
2286
2287         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2288                 goto nla_put_failure;
2289
2290         return 0;
2291
2292 nla_put_failure:
2293         return -ENOSPC;
2294 }
2295
2296 static int
2297 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2298 {
2299         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2300         unsigned long d = ct->status ^ status;
2301
2302         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2303                 /* SEEN_REPLY bit can only be set */
2304                 return -EBUSY;
2305
2306         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2307                 /* ASSURED bit can only be set */
2308                 return -EBUSY;
2309
2310         /* This check is less strict than ctnetlink_change_status()
2311          * because callers often flip IPS_EXPECTED bits when sending
2312          * an NFQA_CT attribute to the kernel.  So ignore the
2313          * unchangeable bits but do not error out. Also user programs
2314          * are allowed to clear the bits that they are allowed to change.
2315          */
2316         __ctnetlink_change_status(ct, status, ~status);
2317         return 0;
2318 }
2319
2320 static int
2321 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2322 {
2323         int err;
2324
2325         if (cda[CTA_TIMEOUT]) {
2326                 err = ctnetlink_change_timeout(ct, cda);
2327                 if (err < 0)
2328                         return err;
2329         }
2330         if (cda[CTA_STATUS]) {
2331                 err = ctnetlink_update_status(ct, cda);
2332                 if (err < 0)
2333                         return err;
2334         }
2335         if (cda[CTA_HELP]) {
2336                 err = ctnetlink_change_helper(ct, cda);
2337                 if (err < 0)
2338                         return err;
2339         }
2340         if (cda[CTA_LABELS]) {
2341                 err = ctnetlink_attach_labels(ct, cda);
2342                 if (err < 0)
2343                         return err;
2344         }
2345 #if defined(CONFIG_NF_CONNTRACK_MARK)
2346         if (cda[CTA_MARK]) {
2347                 u32 mask = 0, mark, newmark;
2348                 if (cda[CTA_MARK_MASK])
2349                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2350
2351                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2352                 newmark = (ct->mark & mask) ^ mark;
2353                 if (newmark != ct->mark)
2354                         ct->mark = newmark;
2355         }
2356 #endif
2357         return 0;
2358 }
2359
2360 static int
2361 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2362 {
2363         struct nlattr *cda[CTA_MAX+1];
2364         int ret;
2365
2366         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2367         if (ret < 0)
2368                 return ret;
2369
2370         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2371 }
2372
2373 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2374                                     const struct nf_conn *ct,
2375                                     struct nf_conntrack_tuple *tuple,
2376                                     struct nf_conntrack_tuple *mask)
2377 {
2378         int err;
2379
2380         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2381                                     nf_ct_l3num(ct), NULL);
2382         if (err < 0)
2383                 return err;
2384
2385         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2386                                      nf_ct_l3num(ct), NULL);
2387 }
2388
2389 static int
2390 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2391                              u32 portid, u32 report)
2392 {
2393         struct nlattr *cda[CTA_EXPECT_MAX+1];
2394         struct nf_conntrack_tuple tuple, mask;
2395         struct nf_conntrack_helper *helper = NULL;
2396         struct nf_conntrack_expect *exp;
2397         int err;
2398
2399         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2400                                NULL);
2401         if (err < 0)
2402                 return err;
2403
2404         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2405                                        ct, &tuple, &mask);
2406         if (err < 0)
2407                 return err;
2408
2409         if (cda[CTA_EXPECT_HELP_NAME]) {
2410                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2411
2412                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2413                                                     nf_ct_protonum(ct));
2414                 if (helper == NULL)
2415                         return -EOPNOTSUPP;
2416         }
2417
2418         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2419                                      helper, &tuple, &mask);
2420         if (IS_ERR(exp))
2421                 return PTR_ERR(exp);
2422
2423         err = nf_ct_expect_related_report(exp, portid, report);
2424         nf_ct_expect_put(exp);
2425         return err;
2426 }
2427
2428 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2429                                   enum ip_conntrack_info ctinfo, int diff)
2430 {
2431         if (!(ct->status & IPS_NAT_MASK))
2432                 return;
2433
2434         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2435 }
2436
2437 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2438         .get_ct         = ctnetlink_glue_get_ct,
2439         .build_size     = ctnetlink_glue_build_size,
2440         .build          = ctnetlink_glue_build,
2441         .parse          = ctnetlink_glue_parse,
2442         .attach_expect  = ctnetlink_glue_attach_expect,
2443         .seq_adjust     = ctnetlink_glue_seqadj,
2444 };
2445 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2446
2447 /***********************************************************************
2448  * EXPECT
2449  ***********************************************************************/
2450
2451 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2452                                     const struct nf_conntrack_tuple *tuple,
2453                                     u32 type)
2454 {
2455         struct nlattr *nest_parms;
2456
2457         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2458         if (!nest_parms)
2459                 goto nla_put_failure;
2460         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2461                 goto nla_put_failure;
2462         nla_nest_end(skb, nest_parms);
2463
2464         return 0;
2465
2466 nla_put_failure:
2467         return -1;
2468 }
2469
2470 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2471                                    const struct nf_conntrack_tuple *tuple,
2472                                    const struct nf_conntrack_tuple_mask *mask)
2473 {
2474         int ret;
2475         struct nf_conntrack_l3proto *l3proto;
2476         struct nf_conntrack_l4proto *l4proto;
2477         struct nf_conntrack_tuple m;
2478         struct nlattr *nest_parms;
2479
2480         memset(&m, 0xFF, sizeof(m));
2481         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2482         m.src.u.all = mask->src.u.all;
2483         m.dst.protonum = tuple->dst.protonum;
2484
2485         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2486         if (!nest_parms)
2487                 goto nla_put_failure;
2488
2489         rcu_read_lock();
2490         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2491         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2492         if (ret >= 0) {
2493                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2494                                                tuple->dst.protonum);
2495         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2496         }
2497         rcu_read_unlock();
2498
2499         if (unlikely(ret < 0))
2500                 goto nla_put_failure;
2501
2502         nla_nest_end(skb, nest_parms);
2503
2504         return 0;
2505
2506 nla_put_failure:
2507         return -1;
2508 }
2509
2510 static const union nf_inet_addr any_addr;
2511
2512 static int
2513 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2514                           const struct nf_conntrack_expect *exp)
2515 {
2516         struct nf_conn *master = exp->master;
2517         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2518         struct nf_conn_help *help;
2519 #ifdef CONFIG_NF_NAT_NEEDED
2520         struct nlattr *nest_parms;
2521         struct nf_conntrack_tuple nat_tuple = {};
2522 #endif
2523         struct nf_ct_helper_expectfn *expfn;
2524
2525         if (timeout < 0)
2526                 timeout = 0;
2527
2528         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2529                 goto nla_put_failure;
2530         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2531                 goto nla_put_failure;
2532         if (ctnetlink_exp_dump_tuple(skb,
2533                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2534                                  CTA_EXPECT_MASTER) < 0)
2535                 goto nla_put_failure;
2536
2537 #ifdef CONFIG_NF_NAT_NEEDED
2538         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2539             exp->saved_proto.all) {
2540                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2541                 if (!nest_parms)
2542                         goto nla_put_failure;
2543
2544                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2545                         goto nla_put_failure;
2546
2547                 nat_tuple.src.l3num = nf_ct_l3num(master);
2548                 nat_tuple.src.u3 = exp->saved_addr;
2549                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2550                 nat_tuple.src.u = exp->saved_proto;
2551
2552                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2553                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2554                         goto nla_put_failure;
2555                 nla_nest_end(skb, nest_parms);
2556         }
2557 #endif
2558         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2559             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2560             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2561             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2562                 goto nla_put_failure;
2563         help = nfct_help(master);
2564         if (help) {
2565                 struct nf_conntrack_helper *helper;
2566
2567                 helper = rcu_dereference(help->helper);
2568                 if (helper &&
2569                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2570                         goto nla_put_failure;
2571         }
2572         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2573         if (expfn != NULL &&
2574             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2575                 goto nla_put_failure;
2576
2577         return 0;
2578
2579 nla_put_failure:
2580         return -1;
2581 }
2582
2583 static int
2584 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2585                         int event, const struct nf_conntrack_expect *exp)
2586 {
2587         struct nlmsghdr *nlh;
2588         struct nfgenmsg *nfmsg;
2589         unsigned int flags = portid ? NLM_F_MULTI : 0;
2590
2591         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2592         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2593         if (nlh == NULL)
2594                 goto nlmsg_failure;
2595
2596         nfmsg = nlmsg_data(nlh);
2597         nfmsg->nfgen_family = exp->tuple.src.l3num;
2598         nfmsg->version      = NFNETLINK_V0;
2599         nfmsg->res_id       = 0;
2600
2601         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2602                 goto nla_put_failure;
2603
2604         nlmsg_end(skb, nlh);
2605         return skb->len;
2606
2607 nlmsg_failure:
2608 nla_put_failure:
2609         nlmsg_cancel(skb, nlh);
2610         return -1;
2611 }
2612
2613 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2614 static int
2615 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2616 {
2617         struct nf_conntrack_expect *exp = item->exp;
2618         struct net *net = nf_ct_exp_net(exp);
2619         struct nlmsghdr *nlh;
2620         struct nfgenmsg *nfmsg;
2621         struct sk_buff *skb;
2622         unsigned int type, group;
2623         int flags = 0;
2624
2625         if (events & (1 << IPEXP_DESTROY)) {
2626                 type = IPCTNL_MSG_EXP_DELETE;
2627                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2628         } else if (events & (1 << IPEXP_NEW)) {
2629                 type = IPCTNL_MSG_EXP_NEW;
2630                 flags = NLM_F_CREATE|NLM_F_EXCL;
2631                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2632         } else
2633                 return 0;
2634
2635         if (!item->report && !nfnetlink_has_listeners(net, group))
2636                 return 0;
2637
2638         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2639         if (skb == NULL)
2640                 goto errout;
2641
2642         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2643         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2644         if (nlh == NULL)
2645                 goto nlmsg_failure;
2646
2647         nfmsg = nlmsg_data(nlh);
2648         nfmsg->nfgen_family = exp->tuple.src.l3num;
2649         nfmsg->version      = NFNETLINK_V0;
2650         nfmsg->res_id       = 0;
2651
2652         rcu_read_lock();
2653         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2654                 goto nla_put_failure;
2655         rcu_read_unlock();
2656
2657         nlmsg_end(skb, nlh);
2658         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2659         return 0;
2660
2661 nla_put_failure:
2662         rcu_read_unlock();
2663         nlmsg_cancel(skb, nlh);
2664 nlmsg_failure:
2665         kfree_skb(skb);
2666 errout:
2667         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2668         return 0;
2669 }
2670 #endif
2671 static int ctnetlink_exp_done(struct netlink_callback *cb)
2672 {
2673         if (cb->args[1])
2674                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2675         return 0;
2676 }
2677
2678 static int
2679 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2680 {
2681         struct net *net = sock_net(skb->sk);
2682         struct nf_conntrack_expect *exp, *last;
2683         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2684         u_int8_t l3proto = nfmsg->nfgen_family;
2685
2686         rcu_read_lock();
2687         last = (struct nf_conntrack_expect *)cb->args[1];
2688         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2689 restart:
2690                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2691                                          hnode) {
2692                         if (l3proto && exp->tuple.src.l3num != l3proto)
2693                                 continue;
2694
2695                         if (!net_eq(nf_ct_net(exp->master), net))
2696                                 continue;
2697
2698                         if (cb->args[1]) {
2699                                 if (exp != last)
2700                                         continue;
2701                                 cb->args[1] = 0;
2702                         }
2703                         if (ctnetlink_exp_fill_info(skb,
2704                                                     NETLINK_CB(cb->skb).portid,
2705                                                     cb->nlh->nlmsg_seq,
2706                                                     IPCTNL_MSG_EXP_NEW,
2707                                                     exp) < 0) {
2708                                 if (!refcount_inc_not_zero(&exp->use))
2709                                         continue;
2710                                 cb->args[1] = (unsigned long)exp;
2711                                 goto out;
2712                         }
2713                 }
2714                 if (cb->args[1]) {
2715                         cb->args[1] = 0;
2716                         goto restart;
2717                 }
2718         }
2719 out:
2720         rcu_read_unlock();
2721         if (last)
2722                 nf_ct_expect_put(last);
2723
2724         return skb->len;
2725 }
2726
2727 static int
2728 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2729 {
2730         struct nf_conntrack_expect *exp, *last;
2731         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2732         struct nf_conn *ct = cb->data;
2733         struct nf_conn_help *help = nfct_help(ct);
2734         u_int8_t l3proto = nfmsg->nfgen_family;
2735
2736         if (cb->args[0])
2737                 return 0;
2738
2739         rcu_read_lock();
2740         last = (struct nf_conntrack_expect *)cb->args[1];
2741 restart:
2742         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2743                 if (l3proto && exp->tuple.src.l3num != l3proto)
2744                         continue;
2745                 if (cb->args[1]) {
2746                         if (exp != last)
2747                                 continue;
2748                         cb->args[1] = 0;
2749                 }
2750                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2751                                             cb->nlh->nlmsg_seq,
2752                                             IPCTNL_MSG_EXP_NEW,
2753                                             exp) < 0) {
2754                         if (!refcount_inc_not_zero(&exp->use))
2755                                 continue;
2756                         cb->args[1] = (unsigned long)exp;
2757                         goto out;
2758                 }
2759         }
2760         if (cb->args[1]) {
2761                 cb->args[1] = 0;
2762                 goto restart;
2763         }
2764         cb->args[0] = 1;
2765 out:
2766         rcu_read_unlock();
2767         if (last)
2768                 nf_ct_expect_put(last);
2769
2770         return skb->len;
2771 }
2772
2773 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2774                                  struct sk_buff *skb,
2775                                  const struct nlmsghdr *nlh,
2776                                  const struct nlattr * const cda[])
2777 {
2778         int err;
2779         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2780         u_int8_t u3 = nfmsg->nfgen_family;
2781         struct nf_conntrack_tuple tuple;
2782         struct nf_conntrack_tuple_hash *h;
2783         struct nf_conn *ct;
2784         struct nf_conntrack_zone zone;
2785         struct netlink_dump_control c = {
2786                 .dump = ctnetlink_exp_ct_dump_table,
2787                 .done = ctnetlink_exp_done,
2788         };
2789
2790         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2791                                     u3, NULL);
2792         if (err < 0)
2793                 return err;
2794
2795         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2796         if (err < 0)
2797                 return err;
2798
2799         h = nf_conntrack_find_get(net, &zone, &tuple);
2800         if (!h)
2801                 return -ENOENT;
2802
2803         ct = nf_ct_tuplehash_to_ctrack(h);
2804         /* No expectation linked to this connection tracking. */
2805         if (!nfct_help(ct)) {
2806                 nf_ct_put(ct);
2807                 return 0;
2808         }
2809
2810         c.data = ct;
2811
2812         err = netlink_dump_start(ctnl, skb, nlh, &c);
2813         nf_ct_put(ct);
2814
2815         return err;
2816 }
2817
2818 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
2819                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2820                                 const struct nlattr * const cda[])
2821 {
2822         struct nf_conntrack_tuple tuple;
2823         struct nf_conntrack_expect *exp;
2824         struct sk_buff *skb2;
2825         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2826         u_int8_t u3 = nfmsg->nfgen_family;
2827         struct nf_conntrack_zone zone;
2828         int err;
2829
2830         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2831                 if (cda[CTA_EXPECT_MASTER])
2832                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda);
2833                 else {
2834                         struct netlink_dump_control c = {
2835                                 .dump = ctnetlink_exp_dump_table,
2836                                 .done = ctnetlink_exp_done,
2837                         };
2838                         return netlink_dump_start(ctnl, skb, nlh, &c);
2839                 }
2840         }
2841
2842         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2843         if (err < 0)
2844                 return err;
2845
2846         if (cda[CTA_EXPECT_TUPLE])
2847                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2848                                             u3, NULL);
2849         else if (cda[CTA_EXPECT_MASTER])
2850                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2851                                             u3, NULL);
2852         else
2853                 return -EINVAL;
2854
2855         if (err < 0)
2856                 return err;
2857
2858         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2859         if (!exp)
2860                 return -ENOENT;
2861
2862         if (cda[CTA_EXPECT_ID]) {
2863                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2864                 if (ntohl(id) != (u32)(unsigned long)exp) {
2865                         nf_ct_expect_put(exp);
2866                         return -ENOENT;
2867                 }
2868         }
2869
2870         err = -ENOMEM;
2871         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2872         if (skb2 == NULL) {
2873                 nf_ct_expect_put(exp);
2874                 goto out;
2875         }
2876
2877         rcu_read_lock();
2878         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2879                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2880         rcu_read_unlock();
2881         nf_ct_expect_put(exp);
2882         if (err <= 0)
2883                 goto free;
2884
2885         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2886         if (err < 0)
2887                 goto out;
2888
2889         return 0;
2890
2891 free:
2892         kfree_skb(skb2);
2893 out:
2894         /* this avoids a loop in nfnetlink. */
2895         return err == -EAGAIN ? -ENOBUFS : err;
2896 }
2897
2898 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
2899                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2900                                 const struct nlattr * const cda[])
2901 {
2902         struct nf_conntrack_expect *exp;
2903         struct nf_conntrack_tuple tuple;
2904         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2905         struct hlist_node *next;
2906         u_int8_t u3 = nfmsg->nfgen_family;
2907         struct nf_conntrack_zone zone;
2908         unsigned int i;
2909         int err;
2910
2911         if (cda[CTA_EXPECT_TUPLE]) {
2912                 /* delete a single expect by tuple */
2913                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2914                 if (err < 0)
2915                         return err;
2916
2917                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2918                                             u3, NULL);
2919                 if (err < 0)
2920                         return err;
2921
2922                 /* bump usage count to 2 */
2923                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2924                 if (!exp)
2925                         return -ENOENT;
2926
2927                 if (cda[CTA_EXPECT_ID]) {
2928                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2929                         if (ntohl(id) != (u32)(unsigned long)exp) {
2930                                 nf_ct_expect_put(exp);
2931                                 return -ENOENT;
2932                         }
2933                 }
2934
2935                 /* after list removal, usage count == 1 */
2936                 spin_lock_bh(&nf_conntrack_expect_lock);
2937                 if (del_timer(&exp->timeout)) {
2938                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2939                                                    nlmsg_report(nlh));
2940                         nf_ct_expect_put(exp);
2941                 }
2942                 spin_unlock_bh(&nf_conntrack_expect_lock);
2943                 /* have to put what we 'get' above.
2944                  * after this line usage count == 0 */
2945                 nf_ct_expect_put(exp);
2946         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2947                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2948                 struct nf_conn_help *m_help;
2949
2950                 /* delete all expectations for this helper */
2951                 spin_lock_bh(&nf_conntrack_expect_lock);
2952                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2953                         hlist_for_each_entry_safe(exp, next,
2954                                                   &nf_ct_expect_hash[i],
2955                                                   hnode) {
2956
2957                                 if (!net_eq(nf_ct_exp_net(exp), net))
2958                                         continue;
2959
2960                                 m_help = nfct_help(exp->master);
2961                                 if (!strcmp(m_help->helper->name, name) &&
2962                                     del_timer(&exp->timeout)) {
2963                                         nf_ct_unlink_expect_report(exp,
2964                                                         NETLINK_CB(skb).portid,
2965                                                         nlmsg_report(nlh));
2966                                         nf_ct_expect_put(exp);
2967                                 }
2968                         }
2969                 }
2970                 spin_unlock_bh(&nf_conntrack_expect_lock);
2971         } else {
2972                 /* This basically means we have to flush everything*/
2973                 spin_lock_bh(&nf_conntrack_expect_lock);
2974                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2975                         hlist_for_each_entry_safe(exp, next,
2976                                                   &nf_ct_expect_hash[i],
2977                                                   hnode) {
2978
2979                                 if (!net_eq(nf_ct_exp_net(exp), net))
2980                                         continue;
2981
2982                                 if (del_timer(&exp->timeout)) {
2983                                         nf_ct_unlink_expect_report(exp,
2984                                                         NETLINK_CB(skb).portid,
2985                                                         nlmsg_report(nlh));
2986                                         nf_ct_expect_put(exp);
2987                                 }
2988                         }
2989                 }
2990                 spin_unlock_bh(&nf_conntrack_expect_lock);
2991         }
2992
2993         return 0;
2994 }
2995 static int
2996 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2997                         const struct nlattr * const cda[])
2998 {
2999         if (cda[CTA_EXPECT_TIMEOUT]) {
3000                 if (!del_timer(&x->timeout))
3001                         return -ETIME;
3002
3003                 x->timeout.expires = jiffies +
3004                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3005                 add_timer(&x->timeout);
3006         }
3007         return 0;
3008 }
3009
3010 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3011         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3012         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3013 };
3014
3015 static int
3016 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3017                            struct nf_conntrack_expect *exp,
3018                            u_int8_t u3)
3019 {
3020 #ifdef CONFIG_NF_NAT_NEEDED
3021         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3022         struct nf_conntrack_tuple nat_tuple = {};
3023         int err;
3024
3025         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3026                                exp_nat_nla_policy, NULL);
3027         if (err < 0)
3028                 return err;
3029
3030         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3031                 return -EINVAL;
3032
3033         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3034                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3035                                     u3, NULL);
3036         if (err < 0)
3037                 return err;
3038
3039         exp->saved_addr = nat_tuple.src.u3;
3040         exp->saved_proto = nat_tuple.src.u;
3041         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3042
3043         return 0;
3044 #else
3045         return -EOPNOTSUPP;
3046 #endif
3047 }
3048
3049 static struct nf_conntrack_expect *
3050 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3051                        struct nf_conntrack_helper *helper,
3052                        struct nf_conntrack_tuple *tuple,
3053                        struct nf_conntrack_tuple *mask)
3054 {
3055         u_int32_t class = 0;
3056         struct nf_conntrack_expect *exp;
3057         struct nf_conn_help *help;
3058         int err;
3059
3060         help = nfct_help(ct);
3061         if (!help)
3062                 return ERR_PTR(-EOPNOTSUPP);
3063
3064         if (cda[CTA_EXPECT_CLASS] && helper) {
3065                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3066                 if (class > helper->expect_class_max)
3067                         return ERR_PTR(-EINVAL);
3068         }
3069         exp = nf_ct_expect_alloc(ct);
3070         if (!exp)
3071                 return ERR_PTR(-ENOMEM);
3072
3073         if (cda[CTA_EXPECT_FLAGS]) {
3074                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3075                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3076         } else {
3077                 exp->flags = 0;
3078         }
3079         if (cda[CTA_EXPECT_FN]) {
3080                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3081                 struct nf_ct_helper_expectfn *expfn;
3082
3083                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3084                 if (expfn == NULL) {
3085                         err = -EINVAL;
3086                         goto err_out;
3087                 }
3088                 exp->expectfn = expfn->expectfn;
3089         } else
3090                 exp->expectfn = NULL;
3091
3092         exp->class = class;
3093         exp->master = ct;
3094         exp->helper = helper;
3095         exp->tuple = *tuple;
3096         exp->mask.src.u3 = mask->src.u3;
3097         exp->mask.src.u.all = mask->src.u.all;
3098
3099         if (cda[CTA_EXPECT_NAT]) {
3100                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3101                                                  exp, nf_ct_l3num(ct));
3102                 if (err < 0)
3103                         goto err_out;
3104         }
3105         return exp;
3106 err_out:
3107         nf_ct_expect_put(exp);
3108         return ERR_PTR(err);
3109 }
3110
3111 static int
3112 ctnetlink_create_expect(struct net *net,
3113                         const struct nf_conntrack_zone *zone,
3114                         const struct nlattr * const cda[],
3115                         u_int8_t u3, u32 portid, int report)
3116 {
3117         struct nf_conntrack_tuple tuple, mask, master_tuple;
3118         struct nf_conntrack_tuple_hash *h = NULL;
3119         struct nf_conntrack_helper *helper = NULL;
3120         struct nf_conntrack_expect *exp;
3121         struct nf_conn *ct;
3122         int err;
3123
3124         /* caller guarantees that those three CTA_EXPECT_* exist */
3125         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3126                                     u3, NULL);
3127         if (err < 0)
3128                 return err;
3129         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3130                                     u3, NULL);
3131         if (err < 0)
3132                 return err;
3133         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3134                                     u3, NULL);
3135         if (err < 0)
3136                 return err;
3137
3138         /* Look for master conntrack of this expectation */
3139         h = nf_conntrack_find_get(net, zone, &master_tuple);
3140         if (!h)
3141                 return -ENOENT;
3142         ct = nf_ct_tuplehash_to_ctrack(h);
3143
3144         rcu_read_lock();
3145         if (cda[CTA_EXPECT_HELP_NAME]) {
3146                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3147
3148                 helper = __nf_conntrack_helper_find(helpname, u3,
3149                                                     nf_ct_protonum(ct));
3150                 if (helper == NULL) {
3151                         rcu_read_unlock();
3152 #ifdef CONFIG_MODULES
3153                         if (request_module("nfct-helper-%s", helpname) < 0) {
3154                                 err = -EOPNOTSUPP;
3155                                 goto err_ct;
3156                         }
3157                         rcu_read_lock();
3158                         helper = __nf_conntrack_helper_find(helpname, u3,
3159                                                             nf_ct_protonum(ct));
3160                         if (helper) {
3161                                 err = -EAGAIN;
3162                                 goto err_rcu;
3163                         }
3164                         rcu_read_unlock();
3165 #endif
3166                         err = -EOPNOTSUPP;
3167                         goto err_ct;
3168                 }
3169         }
3170
3171         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3172         if (IS_ERR(exp)) {
3173                 err = PTR_ERR(exp);
3174                 goto err_rcu;
3175         }
3176
3177         err = nf_ct_expect_related_report(exp, portid, report);
3178         nf_ct_expect_put(exp);
3179 err_rcu:
3180         rcu_read_unlock();
3181 err_ct:
3182         nf_ct_put(ct);
3183         return err;
3184 }
3185
3186 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3187                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3188                                 const struct nlattr * const cda[])
3189 {
3190         struct nf_conntrack_tuple tuple;
3191         struct nf_conntrack_expect *exp;
3192         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3193         u_int8_t u3 = nfmsg->nfgen_family;
3194         struct nf_conntrack_zone zone;
3195         int err;
3196
3197         if (!cda[CTA_EXPECT_TUPLE]
3198             || !cda[CTA_EXPECT_MASK]
3199             || !cda[CTA_EXPECT_MASTER])
3200                 return -EINVAL;
3201
3202         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3203         if (err < 0)
3204                 return err;
3205
3206         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3207                                     u3, NULL);
3208         if (err < 0)
3209                 return err;
3210
3211         spin_lock_bh(&nf_conntrack_expect_lock);
3212         exp = __nf_ct_expect_find(net, &zone, &tuple);
3213         if (!exp) {
3214                 spin_unlock_bh(&nf_conntrack_expect_lock);
3215                 err = -ENOENT;
3216                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3217                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3218                                                       NETLINK_CB(skb).portid,
3219                                                       nlmsg_report(nlh));
3220                 }
3221                 return err;
3222         }
3223
3224         err = -EEXIST;
3225         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3226                 err = ctnetlink_change_expect(exp, cda);
3227         spin_unlock_bh(&nf_conntrack_expect_lock);
3228
3229         return err;
3230 }
3231
3232 static int
3233 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3234                              const struct ip_conntrack_stat *st)
3235 {
3236         struct nlmsghdr *nlh;
3237         struct nfgenmsg *nfmsg;
3238         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3239
3240         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3241                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3242         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3243         if (nlh == NULL)
3244                 goto nlmsg_failure;
3245
3246         nfmsg = nlmsg_data(nlh);
3247         nfmsg->nfgen_family = AF_UNSPEC;
3248         nfmsg->version      = NFNETLINK_V0;
3249         nfmsg->res_id       = htons(cpu);
3250
3251         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3252             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3253             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3254                 goto nla_put_failure;
3255
3256         nlmsg_end(skb, nlh);
3257         return skb->len;
3258
3259 nla_put_failure:
3260 nlmsg_failure:
3261         nlmsg_cancel(skb, nlh);
3262         return -1;
3263 }
3264
3265 static int
3266 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3267 {
3268         int cpu;
3269         struct net *net = sock_net(skb->sk);
3270
3271         if (cb->args[0] == nr_cpu_ids)
3272                 return 0;
3273
3274         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3275                 const struct ip_conntrack_stat *st;
3276
3277                 if (!cpu_possible(cpu))
3278                         continue;
3279
3280                 st = per_cpu_ptr(net->ct.stat, cpu);
3281                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3282                                                  cb->nlh->nlmsg_seq,
3283                                                  cpu, st) < 0)
3284                         break;
3285         }
3286         cb->args[0] = cpu;
3287
3288         return skb->len;
3289 }
3290
3291 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3292                                   struct sk_buff *skb,
3293                                   const struct nlmsghdr *nlh,
3294                                   const struct nlattr * const cda[])
3295 {
3296         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3297                 struct netlink_dump_control c = {
3298                         .dump = ctnetlink_exp_stat_cpu_dump,
3299                 };
3300                 return netlink_dump_start(ctnl, skb, nlh, &c);
3301         }
3302
3303         return 0;
3304 }
3305
3306 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3307 static struct nf_ct_event_notifier ctnl_notifier = {
3308         .fcn = ctnetlink_conntrack_event,
3309 };
3310
3311 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3312         .fcn = ctnetlink_expect_event,
3313 };
3314 #endif
3315
3316 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3317         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3318                                             .attr_count = CTA_MAX,
3319                                             .policy = ct_nla_policy },
3320         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3321                                             .attr_count = CTA_MAX,
3322                                             .policy = ct_nla_policy },
3323         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3324                                             .attr_count = CTA_MAX,
3325                                             .policy = ct_nla_policy },
3326         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3327                                             .attr_count = CTA_MAX,
3328                                             .policy = ct_nla_policy },
3329         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3330         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3331         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3332         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3333 };
3334
3335 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3336         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3337                                             .attr_count = CTA_EXPECT_MAX,
3338                                             .policy = exp_nla_policy },
3339         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3340                                             .attr_count = CTA_EXPECT_MAX,
3341                                             .policy = exp_nla_policy },
3342         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3343                                             .attr_count = CTA_EXPECT_MAX,
3344                                             .policy = exp_nla_policy },
3345         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3346 };
3347
3348 static const struct nfnetlink_subsystem ctnl_subsys = {
3349         .name                           = "conntrack",
3350         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3351         .cb_count                       = IPCTNL_MSG_MAX,
3352         .cb                             = ctnl_cb,
3353 };
3354
3355 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3356         .name                           = "conntrack_expect",
3357         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3358         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3359         .cb                             = ctnl_exp_cb,
3360 };
3361
3362 MODULE_ALIAS("ip_conntrack_netlink");
3363 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3364 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3365
3366 static int __net_init ctnetlink_net_init(struct net *net)
3367 {
3368 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3369         int ret;
3370
3371         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3372         if (ret < 0) {
3373                 pr_err("ctnetlink_init: cannot register notifier.\n");
3374                 goto err_out;
3375         }
3376
3377         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3378         if (ret < 0) {
3379                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3380                 goto err_unreg_notifier;
3381         }
3382 #endif
3383         return 0;
3384
3385 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3386 err_unreg_notifier:
3387         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3388 err_out:
3389         return ret;
3390 #endif
3391 }
3392
3393 static void ctnetlink_net_exit(struct net *net)
3394 {
3395 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3396         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3397         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3398 #endif
3399 }
3400
3401 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3402 {
3403         struct net *net;
3404
3405         list_for_each_entry(net, net_exit_list, exit_list)
3406                 ctnetlink_net_exit(net);
3407 }
3408
3409 static struct pernet_operations ctnetlink_net_ops = {
3410         .init           = ctnetlink_net_init,
3411         .exit_batch     = ctnetlink_net_exit_batch,
3412 };
3413
3414 static int __init ctnetlink_init(void)
3415 {
3416         int ret;
3417
3418         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3419         ret = nfnetlink_subsys_register(&ctnl_subsys);
3420         if (ret < 0) {
3421                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3422                 goto err_out;
3423         }
3424
3425         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3426         if (ret < 0) {
3427                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3428                 goto err_unreg_subsys;
3429         }
3430
3431         ret = register_pernet_subsys(&ctnetlink_net_ops);
3432         if (ret < 0) {
3433                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3434                 goto err_unreg_exp_subsys;
3435         }
3436 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3437         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3438         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3439 #endif
3440         return 0;
3441
3442 err_unreg_exp_subsys:
3443         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3444 err_unreg_subsys:
3445         nfnetlink_subsys_unregister(&ctnl_subsys);
3446 err_out:
3447         return ret;
3448 }
3449
3450 static void __exit ctnetlink_exit(void)
3451 {
3452         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3453
3454         unregister_pernet_subsys(&ctnetlink_net_ops);
3455         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3456         nfnetlink_subsys_unregister(&ctnl_subsys);
3457 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3458         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3459 #endif
3460         synchronize_rcu();
3461 }
3462
3463 module_init(ctnetlink_init);
3464 module_exit(ctnetlink_exit);