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