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         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1482         if (err < 0)
1483                 return err;
1484
1485         /* don't change helper of sibling connections */
1486         if (ct->master) {
1487                 /* If we try to change the helper to the same thing twice,
1488                  * treat the second attempt as a no-op instead of returning
1489                  * an error.
1490                  */
1491                 err = -EBUSY;
1492                 if (help) {
1493                         rcu_read_lock();
1494                         helper = rcu_dereference(help->helper);
1495                         if (helper && !strcmp(helper->name, helpname))
1496                                 err = 0;
1497                         rcu_read_unlock();
1498                 }
1499
1500                 return err;
1501         }
1502
1503         if (!strcmp(helpname, "")) {
1504                 if (help && help->helper) {
1505                         /* we had a helper before ... */
1506                         nf_ct_remove_expectations(ct);
1507                         RCU_INIT_POINTER(help->helper, NULL);
1508                 }
1509
1510                 return 0;
1511         }
1512
1513         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1514                                             nf_ct_protonum(ct));
1515         if (helper == NULL) {
1516 #ifdef CONFIG_MODULES
1517                 spin_unlock_bh(&nf_conntrack_expect_lock);
1518
1519                 if (request_module("nfct-helper-%s", helpname) < 0) {
1520                         spin_lock_bh(&nf_conntrack_expect_lock);
1521                         return -EOPNOTSUPP;
1522                 }
1523
1524                 spin_lock_bh(&nf_conntrack_expect_lock);
1525                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1526                                                     nf_ct_protonum(ct));
1527                 if (helper)
1528                         return -EAGAIN;
1529 #endif
1530                 return -EOPNOTSUPP;
1531         }
1532
1533         if (help) {
1534                 if (help->helper == helper) {
1535                         /* update private helper data if allowed. */
1536                         if (helper->from_nlattr)
1537                                 helper->from_nlattr(helpinfo, ct);
1538                         return 0;
1539                 } else
1540                         return -EBUSY;
1541         }
1542
1543         /* we cannot set a helper for an existing conntrack */
1544         return -EOPNOTSUPP;
1545 }
1546
1547 static int ctnetlink_change_timeout(struct nf_conn *ct,
1548                                     const struct nlattr * const cda[])
1549 {
1550         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1551
1552         ct->timeout = nfct_time_stamp + timeout * HZ;
1553
1554         if (test_bit(IPS_DYING_BIT, &ct->status))
1555                 return -ETIME;
1556
1557         return 0;
1558 }
1559
1560 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1561         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1562         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1563         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1564 };
1565
1566 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1567                                       const struct nlattr * const cda[])
1568 {
1569         const struct nlattr *attr = cda[CTA_PROTOINFO];
1570         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1571         struct nf_conntrack_l4proto *l4proto;
1572         int err = 0;
1573
1574         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1575         if (err < 0)
1576                 return err;
1577
1578         rcu_read_lock();
1579         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1580         if (l4proto->from_nlattr)
1581                 err = l4proto->from_nlattr(tb, ct);
1582         rcu_read_unlock();
1583
1584         return err;
1585 }
1586
1587 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1588         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1589         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1590         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1591 };
1592
1593 static int change_seq_adj(struct nf_ct_seqadj *seq,
1594                           const struct nlattr * const attr)
1595 {
1596         int err;
1597         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1598
1599         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1600         if (err < 0)
1601                 return err;
1602
1603         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1604                 return -EINVAL;
1605
1606         seq->correction_pos =
1607                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1608
1609         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1610                 return -EINVAL;
1611
1612         seq->offset_before =
1613                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1614
1615         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1616                 return -EINVAL;
1617
1618         seq->offset_after =
1619                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1620
1621         return 0;
1622 }
1623
1624 static int
1625 ctnetlink_change_seq_adj(struct nf_conn *ct,
1626                          const struct nlattr * const cda[])
1627 {
1628         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1629         int ret = 0;
1630
1631         if (!seqadj)
1632                 return 0;
1633
1634         if (cda[CTA_SEQ_ADJ_ORIG]) {
1635                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1636                                      cda[CTA_SEQ_ADJ_ORIG]);
1637                 if (ret < 0)
1638                         return ret;
1639
1640                 ct->status |= IPS_SEQ_ADJUST;
1641         }
1642
1643         if (cda[CTA_SEQ_ADJ_REPLY]) {
1644                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1645                                      cda[CTA_SEQ_ADJ_REPLY]);
1646                 if (ret < 0)
1647                         return ret;
1648
1649                 ct->status |= IPS_SEQ_ADJUST;
1650         }
1651
1652         return 0;
1653 }
1654
1655 static int
1656 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1657 {
1658 #ifdef CONFIG_NF_CONNTRACK_LABELS
1659         size_t len = nla_len(cda[CTA_LABELS]);
1660         const void *mask = cda[CTA_LABELS_MASK];
1661
1662         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1663                 return -EINVAL;
1664
1665         if (mask) {
1666                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1667                     nla_len(cda[CTA_LABELS_MASK]) != len)
1668                         return -EINVAL;
1669                 mask = nla_data(cda[CTA_LABELS_MASK]);
1670         }
1671
1672         len /= sizeof(u32);
1673
1674         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1675 #else
1676         return -EOPNOTSUPP;
1677 #endif
1678 }
1679
1680 static int
1681 ctnetlink_change_conntrack(struct nf_conn *ct,
1682                            const struct nlattr * const cda[])
1683 {
1684         int err;
1685
1686         /* only allow NAT changes and master assignation for new conntracks */
1687         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1688                 return -EOPNOTSUPP;
1689
1690         if (cda[CTA_HELP]) {
1691                 err = ctnetlink_change_helper(ct, cda);
1692                 if (err < 0)
1693                         return err;
1694         }
1695
1696         if (cda[CTA_TIMEOUT]) {
1697                 err = ctnetlink_change_timeout(ct, cda);
1698                 if (err < 0)
1699                         return err;
1700         }
1701
1702         if (cda[CTA_STATUS]) {
1703                 err = ctnetlink_change_status(ct, cda);
1704                 if (err < 0)
1705                         return err;
1706         }
1707
1708         if (cda[CTA_PROTOINFO]) {
1709                 err = ctnetlink_change_protoinfo(ct, cda);
1710                 if (err < 0)
1711                         return err;
1712         }
1713
1714 #if defined(CONFIG_NF_CONNTRACK_MARK)
1715         if (cda[CTA_MARK])
1716                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1717 #endif
1718
1719         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1720                 err = ctnetlink_change_seq_adj(ct, cda);
1721                 if (err < 0)
1722                         return err;
1723         }
1724
1725         if (cda[CTA_LABELS]) {
1726                 err = ctnetlink_attach_labels(ct, cda);
1727                 if (err < 0)
1728                         return err;
1729         }
1730
1731         return 0;
1732 }
1733
1734 static struct nf_conn *
1735 ctnetlink_create_conntrack(struct net *net,
1736                            const struct nf_conntrack_zone *zone,
1737                            const struct nlattr * const cda[],
1738                            struct nf_conntrack_tuple *otuple,
1739                            struct nf_conntrack_tuple *rtuple,
1740                            u8 u3)
1741 {
1742         struct nf_conn *ct;
1743         int err = -EINVAL;
1744         struct nf_conntrack_helper *helper;
1745         struct nf_conn_tstamp *tstamp;
1746
1747         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1748         if (IS_ERR(ct))
1749                 return ERR_PTR(-ENOMEM);
1750
1751         if (!cda[CTA_TIMEOUT])
1752                 goto err1;
1753
1754         ct->timeout = nfct_time_stamp + ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1755
1756         rcu_read_lock();
1757         if (cda[CTA_HELP]) {
1758                 char *helpname = NULL;
1759                 struct nlattr *helpinfo = NULL;
1760
1761                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1762                 if (err < 0)
1763                         goto err2;
1764
1765                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1766                                                     nf_ct_protonum(ct));
1767                 if (helper == NULL) {
1768                         rcu_read_unlock();
1769 #ifdef CONFIG_MODULES
1770                         if (request_module("nfct-helper-%s", helpname) < 0) {
1771                                 err = -EOPNOTSUPP;
1772                                 goto err1;
1773                         }
1774
1775                         rcu_read_lock();
1776                         helper = __nf_conntrack_helper_find(helpname,
1777                                                             nf_ct_l3num(ct),
1778                                                             nf_ct_protonum(ct));
1779                         if (helper) {
1780                                 err = -EAGAIN;
1781                                 goto err2;
1782                         }
1783                         rcu_read_unlock();
1784 #endif
1785                         err = -EOPNOTSUPP;
1786                         goto err1;
1787                 } else {
1788                         struct nf_conn_help *help;
1789
1790                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1791                         if (help == NULL) {
1792                                 err = -ENOMEM;
1793                                 goto err2;
1794                         }
1795                         /* set private helper data if allowed. */
1796                         if (helper->from_nlattr)
1797                                 helper->from_nlattr(helpinfo, ct);
1798
1799                         /* not in hash table yet so not strictly necessary */
1800                         RCU_INIT_POINTER(help->helper, helper);
1801                 }
1802         } else {
1803                 /* try an implicit helper assignation */
1804                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1805                 if (err < 0)
1806                         goto err2;
1807         }
1808
1809         err = ctnetlink_setup_nat(ct, cda);
1810         if (err < 0)
1811                 goto err2;
1812
1813         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1814         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1815         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1816         nf_ct_labels_ext_add(ct);
1817
1818         /* we must add conntrack extensions before confirmation. */
1819         ct->status |= IPS_CONFIRMED;
1820
1821         if (cda[CTA_STATUS]) {
1822                 err = ctnetlink_change_status(ct, cda);
1823                 if (err < 0)
1824                         goto err2;
1825         }
1826
1827         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1828                 err = ctnetlink_change_seq_adj(ct, cda);
1829                 if (err < 0)
1830                         goto err2;
1831         }
1832
1833         memset(&ct->proto, 0, sizeof(ct->proto));
1834         if (cda[CTA_PROTOINFO]) {
1835                 err = ctnetlink_change_protoinfo(ct, cda);
1836                 if (err < 0)
1837                         goto err2;
1838         }
1839
1840 #if defined(CONFIG_NF_CONNTRACK_MARK)
1841         if (cda[CTA_MARK])
1842                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1843 #endif
1844
1845         /* setup master conntrack: this is a confirmed expectation */
1846         if (cda[CTA_TUPLE_MASTER]) {
1847                 struct nf_conntrack_tuple master;
1848                 struct nf_conntrack_tuple_hash *master_h;
1849                 struct nf_conn *master_ct;
1850
1851                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1852                                             u3, NULL);
1853                 if (err < 0)
1854                         goto err2;
1855
1856                 master_h = nf_conntrack_find_get(net, zone, &master);
1857                 if (master_h == NULL) {
1858                         err = -ENOENT;
1859                         goto err2;
1860                 }
1861                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1862                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1863                 ct->master = master_ct;
1864         }
1865         tstamp = nf_conn_tstamp_find(ct);
1866         if (tstamp)
1867                 tstamp->start = ktime_get_real_ns();
1868
1869         err = nf_conntrack_hash_check_insert(ct);
1870         if (err < 0)
1871                 goto err2;
1872
1873         rcu_read_unlock();
1874
1875         return ct;
1876
1877 err2:
1878         rcu_read_unlock();
1879 err1:
1880         nf_conntrack_free(ct);
1881         return ERR_PTR(err);
1882 }
1883
1884 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
1885                                    struct sk_buff *skb,
1886                                    const struct nlmsghdr *nlh,
1887                                    const struct nlattr * const cda[])
1888 {
1889         struct nf_conntrack_tuple otuple, rtuple;
1890         struct nf_conntrack_tuple_hash *h = NULL;
1891         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1892         struct nf_conn *ct;
1893         u_int8_t u3 = nfmsg->nfgen_family;
1894         struct nf_conntrack_zone zone;
1895         int err;
1896
1897         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1898         if (err < 0)
1899                 return err;
1900
1901         if (cda[CTA_TUPLE_ORIG]) {
1902                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1903                                             u3, &zone);
1904                 if (err < 0)
1905                         return err;
1906         }
1907
1908         if (cda[CTA_TUPLE_REPLY]) {
1909                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1910                                             u3, &zone);
1911                 if (err < 0)
1912                         return err;
1913         }
1914
1915         if (cda[CTA_TUPLE_ORIG])
1916                 h = nf_conntrack_find_get(net, &zone, &otuple);
1917         else if (cda[CTA_TUPLE_REPLY])
1918                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1919
1920         if (h == NULL) {
1921                 err = -ENOENT;
1922                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1923                         enum ip_conntrack_events events;
1924
1925                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1926                                 return -EINVAL;
1927                         if (otuple.dst.protonum != rtuple.dst.protonum)
1928                                 return -EINVAL;
1929
1930                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1931                                                         &rtuple, u3);
1932                         if (IS_ERR(ct))
1933                                 return PTR_ERR(ct);
1934
1935                         err = 0;
1936                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1937                                 events = 1 << IPCT_RELATED;
1938                         else
1939                                 events = 1 << IPCT_NEW;
1940
1941                         if (cda[CTA_LABELS] &&
1942                             ctnetlink_attach_labels(ct, cda) == 0)
1943                                 events |= (1 << IPCT_LABEL);
1944
1945                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1946                                                       (1 << IPCT_ASSURED) |
1947                                                       (1 << IPCT_HELPER) |
1948                                                       (1 << IPCT_PROTOINFO) |
1949                                                       (1 << IPCT_SEQADJ) |
1950                                                       (1 << IPCT_MARK) | events,
1951                                                       ct, NETLINK_CB(skb).portid,
1952                                                       nlmsg_report(nlh));
1953                         nf_ct_put(ct);
1954                 }
1955
1956                 return err;
1957         }
1958         /* implicit 'else' */
1959
1960         err = -EEXIST;
1961         ct = nf_ct_tuplehash_to_ctrack(h);
1962         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1963                 spin_lock_bh(&nf_conntrack_expect_lock);
1964                 err = ctnetlink_change_conntrack(ct, cda);
1965                 spin_unlock_bh(&nf_conntrack_expect_lock);
1966                 if (err == 0) {
1967                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1968                                                       (1 << IPCT_ASSURED) |
1969                                                       (1 << IPCT_HELPER) |
1970                                                       (1 << IPCT_LABEL) |
1971                                                       (1 << IPCT_PROTOINFO) |
1972                                                       (1 << IPCT_SEQADJ) |
1973                                                       (1 << IPCT_MARK),
1974                                                       ct, NETLINK_CB(skb).portid,
1975                                                       nlmsg_report(nlh));
1976                 }
1977         }
1978
1979         nf_ct_put(ct);
1980         return err;
1981 }
1982
1983 static int
1984 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1985                                 __u16 cpu, const struct ip_conntrack_stat *st)
1986 {
1987         struct nlmsghdr *nlh;
1988         struct nfgenmsg *nfmsg;
1989         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1990
1991         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1992         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1993         if (nlh == NULL)
1994                 goto nlmsg_failure;
1995
1996         nfmsg = nlmsg_data(nlh);
1997         nfmsg->nfgen_family = AF_UNSPEC;
1998         nfmsg->version      = NFNETLINK_V0;
1999         nfmsg->res_id       = htons(cpu);
2000
2001         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2002             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2003             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2004             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2005             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2006                                 htonl(st->insert_failed)) ||
2007             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2008             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2009             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2010             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2011                                 htonl(st->search_restart)))
2012                 goto nla_put_failure;
2013
2014         nlmsg_end(skb, nlh);
2015         return skb->len;
2016
2017 nla_put_failure:
2018 nlmsg_failure:
2019         nlmsg_cancel(skb, nlh);
2020         return -1;
2021 }
2022
2023 static int
2024 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2025 {
2026         int cpu;
2027         struct net *net = sock_net(skb->sk);
2028
2029         if (cb->args[0] == nr_cpu_ids)
2030                 return 0;
2031
2032         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2033                 const struct ip_conntrack_stat *st;
2034
2035                 if (!cpu_possible(cpu))
2036                         continue;
2037
2038                 st = per_cpu_ptr(net->ct.stat, cpu);
2039                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2040                                                     NETLINK_CB(cb->skb).portid,
2041                                                     cb->nlh->nlmsg_seq,
2042                                                     cpu, st) < 0)
2043                                 break;
2044         }
2045         cb->args[0] = cpu;
2046
2047         return skb->len;
2048 }
2049
2050 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2051                                  struct sk_buff *skb,
2052                                  const struct nlmsghdr *nlh,
2053                                  const struct nlattr * const cda[])
2054 {
2055         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2056                 struct netlink_dump_control c = {
2057                         .dump = ctnetlink_ct_stat_cpu_dump,
2058                 };
2059                 return netlink_dump_start(ctnl, skb, nlh, &c);
2060         }
2061
2062         return 0;
2063 }
2064
2065 static int
2066 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2067                             struct net *net)
2068 {
2069         struct nlmsghdr *nlh;
2070         struct nfgenmsg *nfmsg;
2071         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2072         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2073
2074         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
2075         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2076         if (nlh == NULL)
2077                 goto nlmsg_failure;
2078
2079         nfmsg = nlmsg_data(nlh);
2080         nfmsg->nfgen_family = AF_UNSPEC;
2081         nfmsg->version      = NFNETLINK_V0;
2082         nfmsg->res_id       = 0;
2083
2084         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2085                 goto nla_put_failure;
2086
2087         nlmsg_end(skb, nlh);
2088         return skb->len;
2089
2090 nla_put_failure:
2091 nlmsg_failure:
2092         nlmsg_cancel(skb, nlh);
2093         return -1;
2094 }
2095
2096 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2097                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2098                              const struct nlattr * const cda[])
2099 {
2100         struct sk_buff *skb2;
2101         int err;
2102
2103         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2104         if (skb2 == NULL)
2105                 return -ENOMEM;
2106
2107         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2108                                           nlh->nlmsg_seq,
2109                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2110                                           sock_net(skb->sk));
2111         if (err <= 0)
2112                 goto free;
2113
2114         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2115         if (err < 0)
2116                 goto out;
2117
2118         return 0;
2119
2120 free:
2121         kfree_skb(skb2);
2122 out:
2123         /* this avoids a loop in nfnetlink. */
2124         return err == -EAGAIN ? -ENOBUFS : err;
2125 }
2126
2127 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2128         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2129         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2130         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2131         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2132         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2133         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2134                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2135         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2136         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2137         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2138         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2139         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2140 };
2141
2142 static struct nf_conntrack_expect *
2143 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2144                        struct nf_conntrack_helper *helper,
2145                        struct nf_conntrack_tuple *tuple,
2146                        struct nf_conntrack_tuple *mask);
2147
2148 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2149 static size_t
2150 ctnetlink_glue_build_size(const struct nf_conn *ct)
2151 {
2152         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2153                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2154                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2155                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2156                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2157                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2158                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2159                + nla_total_size(0) /* CTA_PROTOINFO */
2160                + nla_total_size(0) /* CTA_HELP */
2161                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2162                + ctnetlink_secctx_size(ct)
2163 #ifdef CONFIG_NF_NAT_NEEDED
2164                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2165                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2166 #endif
2167 #ifdef CONFIG_NF_CONNTRACK_MARK
2168                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2169 #endif
2170 #ifdef CONFIG_NF_CONNTRACK_ZONES
2171                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2172 #endif
2173                + ctnetlink_proto_size(ct)
2174                ;
2175 }
2176
2177 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2178                                              enum ip_conntrack_info *ctinfo)
2179 {
2180         struct nf_conn *ct;
2181
2182         ct = nf_ct_get(skb, ctinfo);
2183         if (ct && nf_ct_is_untracked(ct))
2184                 ct = NULL;
2185
2186         return ct;
2187 }
2188
2189 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2190 {
2191         const struct nf_conntrack_zone *zone;
2192         struct nlattr *nest_parms;
2193
2194         rcu_read_lock();
2195         zone = nf_ct_zone(ct);
2196
2197         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2198         if (!nest_parms)
2199                 goto nla_put_failure;
2200         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2201                 goto nla_put_failure;
2202         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2203                                    NF_CT_ZONE_DIR_ORIG) < 0)
2204                 goto nla_put_failure;
2205         nla_nest_end(skb, nest_parms);
2206
2207         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2208         if (!nest_parms)
2209                 goto nla_put_failure;
2210         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2211                 goto nla_put_failure;
2212         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2213                                    NF_CT_ZONE_DIR_REPL) < 0)
2214                 goto nla_put_failure;
2215         nla_nest_end(skb, nest_parms);
2216
2217         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2218                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2219                 goto nla_put_failure;
2220
2221         if (ctnetlink_dump_id(skb, ct) < 0)
2222                 goto nla_put_failure;
2223
2224         if (ctnetlink_dump_status(skb, ct) < 0)
2225                 goto nla_put_failure;
2226
2227         if (ctnetlink_dump_timeout(skb, ct) < 0)
2228                 goto nla_put_failure;
2229
2230         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2231                 goto nla_put_failure;
2232
2233         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2234                 goto nla_put_failure;
2235
2236 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2237         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2238                 goto nla_put_failure;
2239 #endif
2240         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2241                 goto nla_put_failure;
2242
2243         if ((ct->status & IPS_SEQ_ADJUST) &&
2244             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2245                 goto nla_put_failure;
2246
2247 #ifdef CONFIG_NF_CONNTRACK_MARK
2248         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2249                 goto nla_put_failure;
2250 #endif
2251         if (ctnetlink_dump_labels(skb, ct) < 0)
2252                 goto nla_put_failure;
2253         rcu_read_unlock();
2254         return 0;
2255
2256 nla_put_failure:
2257         rcu_read_unlock();
2258         return -ENOSPC;
2259 }
2260
2261 static int
2262 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2263                      enum ip_conntrack_info ctinfo,
2264                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2265 {
2266         struct nlattr *nest_parms;
2267
2268         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2269         if (!nest_parms)
2270                 goto nla_put_failure;
2271
2272         if (__ctnetlink_glue_build(skb, ct) < 0)
2273                 goto nla_put_failure;
2274
2275         nla_nest_end(skb, nest_parms);
2276
2277         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2278                 goto nla_put_failure;
2279
2280         return 0;
2281
2282 nla_put_failure:
2283         return -ENOSPC;
2284 }
2285
2286 static int
2287 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2288 {
2289         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2290         unsigned long d = ct->status ^ status;
2291
2292         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2293                 /* SEEN_REPLY bit can only be set */
2294                 return -EBUSY;
2295
2296         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2297                 /* ASSURED bit can only be set */
2298                 return -EBUSY;
2299
2300         /* This check is less strict than ctnetlink_change_status()
2301          * because callers often flip IPS_EXPECTED bits when sending
2302          * an NFQA_CT attribute to the kernel.  So ignore the
2303          * unchangeable bits but do not error out.
2304          */
2305         ct->status = (status & ~IPS_UNCHANGEABLE_MASK) |
2306                      (ct->status & IPS_UNCHANGEABLE_MASK);
2307         return 0;
2308 }
2309
2310 static int
2311 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2312 {
2313         int err;
2314
2315         if (cda[CTA_TIMEOUT]) {
2316                 err = ctnetlink_change_timeout(ct, cda);
2317                 if (err < 0)
2318                         return err;
2319         }
2320         if (cda[CTA_STATUS]) {
2321                 err = ctnetlink_update_status(ct, cda);
2322                 if (err < 0)
2323                         return err;
2324         }
2325         if (cda[CTA_HELP]) {
2326                 err = ctnetlink_change_helper(ct, cda);
2327                 if (err < 0)
2328                         return err;
2329         }
2330         if (cda[CTA_LABELS]) {
2331                 err = ctnetlink_attach_labels(ct, cda);
2332                 if (err < 0)
2333                         return err;
2334         }
2335 #if defined(CONFIG_NF_CONNTRACK_MARK)
2336         if (cda[CTA_MARK]) {
2337                 u32 mask = 0, mark, newmark;
2338                 if (cda[CTA_MARK_MASK])
2339                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2340
2341                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2342                 newmark = (ct->mark & mask) ^ mark;
2343                 if (newmark != ct->mark)
2344                         ct->mark = newmark;
2345         }
2346 #endif
2347         return 0;
2348 }
2349
2350 static int
2351 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2352 {
2353         struct nlattr *cda[CTA_MAX+1];
2354         int ret;
2355
2356         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2357         if (ret < 0)
2358                 return ret;
2359
2360         spin_lock_bh(&nf_conntrack_expect_lock);
2361         ret = ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2362         spin_unlock_bh(&nf_conntrack_expect_lock);
2363
2364         return ret;
2365 }
2366
2367 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2368                                     const struct nf_conn *ct,
2369                                     struct nf_conntrack_tuple *tuple,
2370                                     struct nf_conntrack_tuple *mask)
2371 {
2372         int err;
2373
2374         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2375                                     nf_ct_l3num(ct), NULL);
2376         if (err < 0)
2377                 return err;
2378
2379         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2380                                      nf_ct_l3num(ct), NULL);
2381 }
2382
2383 static int
2384 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2385                              u32 portid, u32 report)
2386 {
2387         struct nlattr *cda[CTA_EXPECT_MAX+1];
2388         struct nf_conntrack_tuple tuple, mask;
2389         struct nf_conntrack_helper *helper = NULL;
2390         struct nf_conntrack_expect *exp;
2391         int err;
2392
2393         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2394         if (err < 0)
2395                 return err;
2396
2397         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2398                                        ct, &tuple, &mask);
2399         if (err < 0)
2400                 return err;
2401
2402         if (cda[CTA_EXPECT_HELP_NAME]) {
2403                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2404
2405                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2406                                                     nf_ct_protonum(ct));
2407                 if (helper == NULL)
2408                         return -EOPNOTSUPP;
2409         }
2410
2411         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2412                                      helper, &tuple, &mask);
2413         if (IS_ERR(exp))
2414                 return PTR_ERR(exp);
2415
2416         err = nf_ct_expect_related_report(exp, portid, report);
2417         nf_ct_expect_put(exp);
2418         return err;
2419 }
2420
2421 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2422                                   enum ip_conntrack_info ctinfo, int diff)
2423 {
2424         if (!(ct->status & IPS_NAT_MASK))
2425                 return;
2426
2427         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2428 }
2429
2430 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2431         .get_ct         = ctnetlink_glue_get_ct,
2432         .build_size     = ctnetlink_glue_build_size,
2433         .build          = ctnetlink_glue_build,
2434         .parse          = ctnetlink_glue_parse,
2435         .attach_expect  = ctnetlink_glue_attach_expect,
2436         .seq_adjust     = ctnetlink_glue_seqadj,
2437 };
2438 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2439
2440 /***********************************************************************
2441  * EXPECT
2442  ***********************************************************************/
2443
2444 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2445                                     const struct nf_conntrack_tuple *tuple,
2446                                     enum ctattr_expect type)
2447 {
2448         struct nlattr *nest_parms;
2449
2450         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2451         if (!nest_parms)
2452                 goto nla_put_failure;
2453         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2454                 goto nla_put_failure;
2455         nla_nest_end(skb, nest_parms);
2456
2457         return 0;
2458
2459 nla_put_failure:
2460         return -1;
2461 }
2462
2463 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2464                                    const struct nf_conntrack_tuple *tuple,
2465                                    const struct nf_conntrack_tuple_mask *mask)
2466 {
2467         int ret;
2468         struct nf_conntrack_l3proto *l3proto;
2469         struct nf_conntrack_l4proto *l4proto;
2470         struct nf_conntrack_tuple m;
2471         struct nlattr *nest_parms;
2472
2473         memset(&m, 0xFF, sizeof(m));
2474         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2475         m.src.u.all = mask->src.u.all;
2476         m.dst.protonum = tuple->dst.protonum;
2477
2478         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2479         if (!nest_parms)
2480                 goto nla_put_failure;
2481
2482         rcu_read_lock();
2483         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2484         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2485         if (ret >= 0) {
2486                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2487                                                tuple->dst.protonum);
2488         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2489         }
2490         rcu_read_unlock();
2491
2492         if (unlikely(ret < 0))
2493                 goto nla_put_failure;
2494
2495         nla_nest_end(skb, nest_parms);
2496
2497         return 0;
2498
2499 nla_put_failure:
2500         return -1;
2501 }
2502
2503 static const union nf_inet_addr any_addr;
2504
2505 static int
2506 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2507                           const struct nf_conntrack_expect *exp)
2508 {
2509         struct nf_conn *master = exp->master;
2510         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2511         struct nf_conn_help *help;
2512 #ifdef CONFIG_NF_NAT_NEEDED
2513         struct nlattr *nest_parms;
2514         struct nf_conntrack_tuple nat_tuple = {};
2515 #endif
2516         struct nf_ct_helper_expectfn *expfn;
2517
2518         if (timeout < 0)
2519                 timeout = 0;
2520
2521         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2522                 goto nla_put_failure;
2523         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2524                 goto nla_put_failure;
2525         if (ctnetlink_exp_dump_tuple(skb,
2526                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2527                                  CTA_EXPECT_MASTER) < 0)
2528                 goto nla_put_failure;
2529
2530 #ifdef CONFIG_NF_NAT_NEEDED
2531         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2532             exp->saved_proto.all) {
2533                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2534                 if (!nest_parms)
2535                         goto nla_put_failure;
2536
2537                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2538                         goto nla_put_failure;
2539
2540                 nat_tuple.src.l3num = nf_ct_l3num(master);
2541                 nat_tuple.src.u3 = exp->saved_addr;
2542                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2543                 nat_tuple.src.u = exp->saved_proto;
2544
2545                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2546                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2547                         goto nla_put_failure;
2548                 nla_nest_end(skb, nest_parms);
2549         }
2550 #endif
2551         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2552             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2553             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2554             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2555                 goto nla_put_failure;
2556         help = nfct_help(master);
2557         if (help) {
2558                 struct nf_conntrack_helper *helper;
2559
2560                 helper = rcu_dereference(help->helper);
2561                 if (helper &&
2562                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2563                         goto nla_put_failure;
2564         }
2565         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2566         if (expfn != NULL &&
2567             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2568                 goto nla_put_failure;
2569
2570         return 0;
2571
2572 nla_put_failure:
2573         return -1;
2574 }
2575
2576 static int
2577 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2578                         int event, const struct nf_conntrack_expect *exp)
2579 {
2580         struct nlmsghdr *nlh;
2581         struct nfgenmsg *nfmsg;
2582         unsigned int flags = portid ? NLM_F_MULTI : 0;
2583
2584         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2585         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2586         if (nlh == NULL)
2587                 goto nlmsg_failure;
2588
2589         nfmsg = nlmsg_data(nlh);
2590         nfmsg->nfgen_family = exp->tuple.src.l3num;
2591         nfmsg->version      = NFNETLINK_V0;
2592         nfmsg->res_id       = 0;
2593
2594         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2595                 goto nla_put_failure;
2596
2597         nlmsg_end(skb, nlh);
2598         return skb->len;
2599
2600 nlmsg_failure:
2601 nla_put_failure:
2602         nlmsg_cancel(skb, nlh);
2603         return -1;
2604 }
2605
2606 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2607 static int
2608 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2609 {
2610         struct nf_conntrack_expect *exp = item->exp;
2611         struct net *net = nf_ct_exp_net(exp);
2612         struct nlmsghdr *nlh;
2613         struct nfgenmsg *nfmsg;
2614         struct sk_buff *skb;
2615         unsigned int type, group;
2616         int flags = 0;
2617
2618         if (events & (1 << IPEXP_DESTROY)) {
2619                 type = IPCTNL_MSG_EXP_DELETE;
2620                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2621         } else if (events & (1 << IPEXP_NEW)) {
2622                 type = IPCTNL_MSG_EXP_NEW;
2623                 flags = NLM_F_CREATE|NLM_F_EXCL;
2624                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2625         } else
2626                 return 0;
2627
2628         if (!item->report && !nfnetlink_has_listeners(net, group))
2629                 return 0;
2630
2631         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2632         if (skb == NULL)
2633                 goto errout;
2634
2635         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2636         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2637         if (nlh == NULL)
2638                 goto nlmsg_failure;
2639
2640         nfmsg = nlmsg_data(nlh);
2641         nfmsg->nfgen_family = exp->tuple.src.l3num;
2642         nfmsg->version      = NFNETLINK_V0;
2643         nfmsg->res_id       = 0;
2644
2645         rcu_read_lock();
2646         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2647                 goto nla_put_failure;
2648         rcu_read_unlock();
2649
2650         nlmsg_end(skb, nlh);
2651         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2652         return 0;
2653
2654 nla_put_failure:
2655         rcu_read_unlock();
2656         nlmsg_cancel(skb, nlh);
2657 nlmsg_failure:
2658         kfree_skb(skb);
2659 errout:
2660         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2661         return 0;
2662 }
2663 #endif
2664 static int ctnetlink_exp_done(struct netlink_callback *cb)
2665 {
2666         if (cb->args[1])
2667                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2668         return 0;
2669 }
2670
2671 static int
2672 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2673 {
2674         struct net *net = sock_net(skb->sk);
2675         struct nf_conntrack_expect *exp, *last;
2676         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2677         u_int8_t l3proto = nfmsg->nfgen_family;
2678
2679         rcu_read_lock();
2680         last = (struct nf_conntrack_expect *)cb->args[1];
2681         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2682 restart:
2683                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2684                                          hnode) {
2685                         if (l3proto && exp->tuple.src.l3num != l3proto)
2686                                 continue;
2687
2688                         if (!net_eq(nf_ct_net(exp->master), net))
2689                                 continue;
2690
2691                         if (cb->args[1]) {
2692                                 if (exp != last)
2693                                         continue;
2694                                 cb->args[1] = 0;
2695                         }
2696                         if (ctnetlink_exp_fill_info(skb,
2697                                                     NETLINK_CB(cb->skb).portid,
2698                                                     cb->nlh->nlmsg_seq,
2699                                                     IPCTNL_MSG_EXP_NEW,
2700                                                     exp) < 0) {
2701                                 if (!atomic_inc_not_zero(&exp->use))
2702                                         continue;
2703                                 cb->args[1] = (unsigned long)exp;
2704                                 goto out;
2705                         }
2706                 }
2707                 if (cb->args[1]) {
2708                         cb->args[1] = 0;
2709                         goto restart;
2710                 }
2711         }
2712 out:
2713         rcu_read_unlock();
2714         if (last)
2715                 nf_ct_expect_put(last);
2716
2717         return skb->len;
2718 }
2719
2720 static int
2721 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2722 {
2723         struct nf_conntrack_expect *exp, *last;
2724         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2725         struct nf_conn *ct = cb->data;
2726         struct nf_conn_help *help = nfct_help(ct);
2727         u_int8_t l3proto = nfmsg->nfgen_family;
2728
2729         if (cb->args[0])
2730                 return 0;
2731
2732         rcu_read_lock();
2733         last = (struct nf_conntrack_expect *)cb->args[1];
2734 restart:
2735         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2736                 if (l3proto && exp->tuple.src.l3num != l3proto)
2737                         continue;
2738                 if (cb->args[1]) {
2739                         if (exp != last)
2740                                 continue;
2741                         cb->args[1] = 0;
2742                 }
2743                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2744                                             cb->nlh->nlmsg_seq,
2745                                             IPCTNL_MSG_EXP_NEW,
2746                                             exp) < 0) {
2747                         if (!atomic_inc_not_zero(&exp->use))
2748                                 continue;
2749                         cb->args[1] = (unsigned long)exp;
2750                         goto out;
2751                 }
2752         }
2753         if (cb->args[1]) {
2754                 cb->args[1] = 0;
2755                 goto restart;
2756         }
2757         cb->args[0] = 1;
2758 out:
2759         rcu_read_unlock();
2760         if (last)
2761                 nf_ct_expect_put(last);
2762
2763         return skb->len;
2764 }
2765
2766 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2767                                  struct sk_buff *skb,
2768                                  const struct nlmsghdr *nlh,
2769                                  const struct nlattr * const cda[])
2770 {
2771         int err;
2772         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2773         u_int8_t u3 = nfmsg->nfgen_family;
2774         struct nf_conntrack_tuple tuple;
2775         struct nf_conntrack_tuple_hash *h;
2776         struct nf_conn *ct;
2777         struct nf_conntrack_zone zone;
2778         struct netlink_dump_control c = {
2779                 .dump = ctnetlink_exp_ct_dump_table,
2780                 .done = ctnetlink_exp_done,
2781         };
2782
2783         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2784                                     u3, NULL);
2785         if (err < 0)
2786                 return err;
2787
2788         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2789         if (err < 0)
2790                 return err;
2791
2792         h = nf_conntrack_find_get(net, &zone, &tuple);
2793         if (!h)
2794                 return -ENOENT;
2795
2796         ct = nf_ct_tuplehash_to_ctrack(h);
2797         /* No expectation linked to this connection tracking. */
2798         if (!nfct_help(ct)) {
2799                 nf_ct_put(ct);
2800                 return 0;
2801         }
2802
2803         c.data = ct;
2804
2805         err = netlink_dump_start(ctnl, skb, nlh, &c);
2806         nf_ct_put(ct);
2807
2808         return err;
2809 }
2810
2811 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
2812                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2813                                 const struct nlattr * const cda[])
2814 {
2815         struct nf_conntrack_tuple tuple;
2816         struct nf_conntrack_expect *exp;
2817         struct sk_buff *skb2;
2818         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2819         u_int8_t u3 = nfmsg->nfgen_family;
2820         struct nf_conntrack_zone zone;
2821         int err;
2822
2823         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2824                 if (cda[CTA_EXPECT_MASTER])
2825                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda);
2826                 else {
2827                         struct netlink_dump_control c = {
2828                                 .dump = ctnetlink_exp_dump_table,
2829                                 .done = ctnetlink_exp_done,
2830                         };
2831                         return netlink_dump_start(ctnl, skb, nlh, &c);
2832                 }
2833         }
2834
2835         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2836         if (err < 0)
2837                 return err;
2838
2839         if (cda[CTA_EXPECT_TUPLE])
2840                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2841                                             u3, NULL);
2842         else if (cda[CTA_EXPECT_MASTER])
2843                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2844                                             u3, NULL);
2845         else
2846                 return -EINVAL;
2847
2848         if (err < 0)
2849                 return err;
2850
2851         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2852         if (!exp)
2853                 return -ENOENT;
2854
2855         if (cda[CTA_EXPECT_ID]) {
2856                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2857                 if (ntohl(id) != (u32)(unsigned long)exp) {
2858                         nf_ct_expect_put(exp);
2859                         return -ENOENT;
2860                 }
2861         }
2862
2863         err = -ENOMEM;
2864         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2865         if (skb2 == NULL) {
2866                 nf_ct_expect_put(exp);
2867                 goto out;
2868         }
2869
2870         rcu_read_lock();
2871         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2872                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2873         rcu_read_unlock();
2874         nf_ct_expect_put(exp);
2875         if (err <= 0)
2876                 goto free;
2877
2878         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2879         if (err < 0)
2880                 goto out;
2881
2882         return 0;
2883
2884 free:
2885         kfree_skb(skb2);
2886 out:
2887         /* this avoids a loop in nfnetlink. */
2888         return err == -EAGAIN ? -ENOBUFS : err;
2889 }
2890
2891 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
2892                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2893                                 const struct nlattr * const cda[])
2894 {
2895         struct nf_conntrack_expect *exp;
2896         struct nf_conntrack_tuple tuple;
2897         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2898         struct hlist_node *next;
2899         u_int8_t u3 = nfmsg->nfgen_family;
2900         struct nf_conntrack_zone zone;
2901         unsigned int i;
2902         int err;
2903
2904         if (cda[CTA_EXPECT_TUPLE]) {
2905                 /* delete a single expect by tuple */
2906                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2907                 if (err < 0)
2908                         return err;
2909
2910                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2911                                             u3, NULL);
2912                 if (err < 0)
2913                         return err;
2914
2915                 /* bump usage count to 2 */
2916                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2917                 if (!exp)
2918                         return -ENOENT;
2919
2920                 if (cda[CTA_EXPECT_ID]) {
2921                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2922                         if (ntohl(id) != (u32)(unsigned long)exp) {
2923                                 nf_ct_expect_put(exp);
2924                                 return -ENOENT;
2925                         }
2926                 }
2927
2928                 /* after list removal, usage count == 1 */
2929                 spin_lock_bh(&nf_conntrack_expect_lock);
2930                 if (del_timer(&exp->timeout)) {
2931                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2932                                                    nlmsg_report(nlh));
2933                         nf_ct_expect_put(exp);
2934                 }
2935                 spin_unlock_bh(&nf_conntrack_expect_lock);
2936                 /* have to put what we 'get' above.
2937                  * after this line usage count == 0 */
2938                 nf_ct_expect_put(exp);
2939         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2940                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2941                 struct nf_conn_help *m_help;
2942
2943                 /* delete all expectations for this helper */
2944                 spin_lock_bh(&nf_conntrack_expect_lock);
2945                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2946                         hlist_for_each_entry_safe(exp, next,
2947                                                   &nf_ct_expect_hash[i],
2948                                                   hnode) {
2949
2950                                 if (!net_eq(nf_ct_exp_net(exp), net))
2951                                         continue;
2952
2953                                 m_help = nfct_help(exp->master);
2954                                 if (!strcmp(m_help->helper->name, name) &&
2955                                     del_timer(&exp->timeout)) {
2956                                         nf_ct_unlink_expect_report(exp,
2957                                                         NETLINK_CB(skb).portid,
2958                                                         nlmsg_report(nlh));
2959                                         nf_ct_expect_put(exp);
2960                                 }
2961                         }
2962                 }
2963                 spin_unlock_bh(&nf_conntrack_expect_lock);
2964         } else {
2965                 /* This basically means we have to flush everything*/
2966                 spin_lock_bh(&nf_conntrack_expect_lock);
2967                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2968                         hlist_for_each_entry_safe(exp, next,
2969                                                   &nf_ct_expect_hash[i],
2970                                                   hnode) {
2971
2972                                 if (!net_eq(nf_ct_exp_net(exp), net))
2973                                         continue;
2974
2975                                 if (del_timer(&exp->timeout)) {
2976                                         nf_ct_unlink_expect_report(exp,
2977                                                         NETLINK_CB(skb).portid,
2978                                                         nlmsg_report(nlh));
2979                                         nf_ct_expect_put(exp);
2980                                 }
2981                         }
2982                 }
2983                 spin_unlock_bh(&nf_conntrack_expect_lock);
2984         }
2985
2986         return 0;
2987 }
2988 static int
2989 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2990                         const struct nlattr * const cda[])
2991 {
2992         if (cda[CTA_EXPECT_TIMEOUT]) {
2993                 if (!del_timer(&x->timeout))
2994                         return -ETIME;
2995
2996                 x->timeout.expires = jiffies +
2997                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2998                 add_timer(&x->timeout);
2999         }
3000         return 0;
3001 }
3002
3003 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3004         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3005         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3006 };
3007
3008 static int
3009 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3010                            struct nf_conntrack_expect *exp,
3011                            u_int8_t u3)
3012 {
3013 #ifdef CONFIG_NF_NAT_NEEDED
3014         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3015         struct nf_conntrack_tuple nat_tuple = {};
3016         int err;
3017
3018         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
3019         if (err < 0)
3020                 return err;
3021
3022         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3023                 return -EINVAL;
3024
3025         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3026                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3027                                     u3, NULL);
3028         if (err < 0)
3029                 return err;
3030
3031         exp->saved_addr = nat_tuple.src.u3;
3032         exp->saved_proto = nat_tuple.src.u;
3033         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3034
3035         return 0;
3036 #else
3037         return -EOPNOTSUPP;
3038 #endif
3039 }
3040
3041 static struct nf_conntrack_expect *
3042 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3043                        struct nf_conntrack_helper *helper,
3044                        struct nf_conntrack_tuple *tuple,
3045                        struct nf_conntrack_tuple *mask)
3046 {
3047         u_int32_t class = 0;
3048         struct nf_conntrack_expect *exp;
3049         struct nf_conn_help *help;
3050         int err;
3051
3052         if (cda[CTA_EXPECT_CLASS] && helper) {
3053                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3054                 if (class > helper->expect_class_max)
3055                         return ERR_PTR(-EINVAL);
3056         }
3057         exp = nf_ct_expect_alloc(ct);
3058         if (!exp)
3059                 return ERR_PTR(-ENOMEM);
3060
3061         help = nfct_help(ct);
3062         if (!help) {
3063                 if (!cda[CTA_EXPECT_TIMEOUT]) {
3064                         err = -EINVAL;
3065                         goto err_out;
3066                 }
3067                 exp->timeout.expires =
3068                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3069
3070                 exp->flags = NF_CT_EXPECT_USERSPACE;
3071                 if (cda[CTA_EXPECT_FLAGS]) {
3072                         exp->flags |=
3073                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3074                 }
3075         } else {
3076                 if (cda[CTA_EXPECT_FLAGS]) {
3077                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3078                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3079                 } else
3080                         exp->flags = 0;
3081         }
3082         if (cda[CTA_EXPECT_FN]) {
3083                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3084                 struct nf_ct_helper_expectfn *expfn;
3085
3086                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3087                 if (expfn == NULL) {
3088                         err = -EINVAL;
3089                         goto err_out;
3090                 }
3091                 exp->expectfn = expfn->expectfn;
3092         } else
3093                 exp->expectfn = NULL;
3094
3095         exp->class = class;
3096         exp->master = ct;
3097         exp->helper = helper;
3098         exp->tuple = *tuple;
3099         exp->mask.src.u3 = mask->src.u3;
3100         exp->mask.src.u.all = mask->src.u.all;
3101
3102         if (cda[CTA_EXPECT_NAT]) {
3103                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3104                                                  exp, nf_ct_l3num(ct));
3105                 if (err < 0)
3106                         goto err_out;
3107         }
3108         return exp;
3109 err_out:
3110         nf_ct_expect_put(exp);
3111         return ERR_PTR(err);
3112 }
3113
3114 static int
3115 ctnetlink_create_expect(struct net *net,
3116                         const struct nf_conntrack_zone *zone,
3117                         const struct nlattr * const cda[],
3118                         u_int8_t u3, u32 portid, int report)
3119 {
3120         struct nf_conntrack_tuple tuple, mask, master_tuple;
3121         struct nf_conntrack_tuple_hash *h = NULL;
3122         struct nf_conntrack_helper *helper = NULL;
3123         struct nf_conntrack_expect *exp;
3124         struct nf_conn *ct;
3125         int err;
3126
3127         /* caller guarantees that those three CTA_EXPECT_* exist */
3128         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3129                                     u3, NULL);
3130         if (err < 0)
3131                 return err;
3132         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3133                                     u3, NULL);
3134         if (err < 0)
3135                 return err;
3136         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3137                                     u3, NULL);
3138         if (err < 0)
3139                 return err;
3140
3141         /* Look for master conntrack of this expectation */
3142         h = nf_conntrack_find_get(net, zone, &master_tuple);
3143         if (!h)
3144                 return -ENOENT;
3145         ct = nf_ct_tuplehash_to_ctrack(h);
3146
3147         rcu_read_lock();
3148         if (cda[CTA_EXPECT_HELP_NAME]) {
3149                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3150
3151                 helper = __nf_conntrack_helper_find(helpname, u3,
3152                                                     nf_ct_protonum(ct));
3153                 if (helper == NULL) {
3154                         rcu_read_unlock();
3155 #ifdef CONFIG_MODULES
3156                         if (request_module("nfct-helper-%s", helpname) < 0) {
3157                                 err = -EOPNOTSUPP;
3158                                 goto err_ct;
3159                         }
3160                         rcu_read_lock();
3161                         helper = __nf_conntrack_helper_find(helpname, u3,
3162                                                             nf_ct_protonum(ct));
3163                         if (helper) {
3164                                 err = -EAGAIN;
3165                                 goto err_rcu;
3166                         }
3167                         rcu_read_unlock();
3168 #endif
3169                         err = -EOPNOTSUPP;
3170                         goto err_ct;
3171                 }
3172         }
3173
3174         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3175         if (IS_ERR(exp)) {
3176                 err = PTR_ERR(exp);
3177                 goto err_rcu;
3178         }
3179
3180         err = nf_ct_expect_related_report(exp, portid, report);
3181         nf_ct_expect_put(exp);
3182 err_rcu:
3183         rcu_read_unlock();
3184 err_ct:
3185         nf_ct_put(ct);
3186         return err;
3187 }
3188
3189 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3190                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3191                                 const struct nlattr * const cda[])
3192 {
3193         struct nf_conntrack_tuple tuple;
3194         struct nf_conntrack_expect *exp;
3195         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3196         u_int8_t u3 = nfmsg->nfgen_family;
3197         struct nf_conntrack_zone zone;
3198         int err;
3199
3200         if (!cda[CTA_EXPECT_TUPLE]
3201             || !cda[CTA_EXPECT_MASK]
3202             || !cda[CTA_EXPECT_MASTER])
3203                 return -EINVAL;
3204
3205         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3206         if (err < 0)
3207                 return err;
3208
3209         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3210                                     u3, NULL);
3211         if (err < 0)
3212                 return err;
3213
3214         spin_lock_bh(&nf_conntrack_expect_lock);
3215         exp = __nf_ct_expect_find(net, &zone, &tuple);
3216         if (!exp) {
3217                 spin_unlock_bh(&nf_conntrack_expect_lock);
3218                 err = -ENOENT;
3219                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3220                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3221                                                       NETLINK_CB(skb).portid,
3222                                                       nlmsg_report(nlh));
3223                 }
3224                 return err;
3225         }
3226
3227         err = -EEXIST;
3228         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3229                 err = ctnetlink_change_expect(exp, cda);
3230         spin_unlock_bh(&nf_conntrack_expect_lock);
3231
3232         return err;
3233 }
3234
3235 static int
3236 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3237                              const struct ip_conntrack_stat *st)
3238 {
3239         struct nlmsghdr *nlh;
3240         struct nfgenmsg *nfmsg;
3241         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3242
3243         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3244         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3245         if (nlh == NULL)
3246                 goto nlmsg_failure;
3247
3248         nfmsg = nlmsg_data(nlh);
3249         nfmsg->nfgen_family = AF_UNSPEC;
3250         nfmsg->version      = NFNETLINK_V0;
3251         nfmsg->res_id       = htons(cpu);
3252
3253         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3254             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3255             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3256                 goto nla_put_failure;
3257
3258         nlmsg_end(skb, nlh);
3259         return skb->len;
3260
3261 nla_put_failure:
3262 nlmsg_failure:
3263         nlmsg_cancel(skb, nlh);
3264         return -1;
3265 }
3266
3267 static int
3268 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3269 {
3270         int cpu;
3271         struct net *net = sock_net(skb->sk);
3272
3273         if (cb->args[0] == nr_cpu_ids)
3274                 return 0;
3275
3276         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3277                 const struct ip_conntrack_stat *st;
3278
3279                 if (!cpu_possible(cpu))
3280                         continue;
3281
3282                 st = per_cpu_ptr(net->ct.stat, cpu);
3283                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3284                                                  cb->nlh->nlmsg_seq,
3285                                                  cpu, st) < 0)
3286                         break;
3287         }
3288         cb->args[0] = cpu;
3289
3290         return skb->len;
3291 }
3292
3293 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3294                                   struct sk_buff *skb,
3295                                   const struct nlmsghdr *nlh,
3296                                   const struct nlattr * const cda[])
3297 {
3298         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3299                 struct netlink_dump_control c = {
3300                         .dump = ctnetlink_exp_stat_cpu_dump,
3301                 };
3302                 return netlink_dump_start(ctnl, skb, nlh, &c);
3303         }
3304
3305         return 0;
3306 }
3307
3308 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3309 static struct nf_ct_event_notifier ctnl_notifier = {
3310         .fcn = ctnetlink_conntrack_event,
3311 };
3312
3313 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3314         .fcn = ctnetlink_expect_event,
3315 };
3316 #endif
3317
3318 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3319         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3320                                             .attr_count = CTA_MAX,
3321                                             .policy = ct_nla_policy },
3322         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3323                                             .attr_count = CTA_MAX,
3324                                             .policy = ct_nla_policy },
3325         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3326                                             .attr_count = CTA_MAX,
3327                                             .policy = ct_nla_policy },
3328         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3329                                             .attr_count = CTA_MAX,
3330                                             .policy = ct_nla_policy },
3331         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3332         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3333         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3334         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3335 };
3336
3337 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3338         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3339                                             .attr_count = CTA_EXPECT_MAX,
3340                                             .policy = exp_nla_policy },
3341         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3342                                             .attr_count = CTA_EXPECT_MAX,
3343                                             .policy = exp_nla_policy },
3344         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3345                                             .attr_count = CTA_EXPECT_MAX,
3346                                             .policy = exp_nla_policy },
3347         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3348 };
3349
3350 static const struct nfnetlink_subsystem ctnl_subsys = {
3351         .name                           = "conntrack",
3352         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3353         .cb_count                       = IPCTNL_MSG_MAX,
3354         .cb                             = ctnl_cb,
3355 };
3356
3357 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3358         .name                           = "conntrack_expect",
3359         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3360         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3361         .cb                             = ctnl_exp_cb,
3362 };
3363
3364 MODULE_ALIAS("ip_conntrack_netlink");
3365 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3366 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3367
3368 static int __net_init ctnetlink_net_init(struct net *net)
3369 {
3370 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3371         int ret;
3372
3373         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3374         if (ret < 0) {
3375                 pr_err("ctnetlink_init: cannot register notifier.\n");
3376                 goto err_out;
3377         }
3378
3379         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3380         if (ret < 0) {
3381                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3382                 goto err_unreg_notifier;
3383         }
3384 #endif
3385         return 0;
3386
3387 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3388 err_unreg_notifier:
3389         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3390 err_out:
3391         return ret;
3392 #endif
3393 }
3394
3395 static void ctnetlink_net_exit(struct net *net)
3396 {
3397 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3398         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3399         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3400 #endif
3401 }
3402
3403 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3404 {
3405         struct net *net;
3406
3407         list_for_each_entry(net, net_exit_list, exit_list)
3408                 ctnetlink_net_exit(net);
3409 }
3410
3411 static struct pernet_operations ctnetlink_net_ops = {
3412         .init           = ctnetlink_net_init,
3413         .exit_batch     = ctnetlink_net_exit_batch,
3414 };
3415
3416 static int __init ctnetlink_init(void)
3417 {
3418         int ret;
3419
3420         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3421         ret = nfnetlink_subsys_register(&ctnl_subsys);
3422         if (ret < 0) {
3423                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3424                 goto err_out;
3425         }
3426
3427         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3428         if (ret < 0) {
3429                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3430                 goto err_unreg_subsys;
3431         }
3432
3433         ret = register_pernet_subsys(&ctnetlink_net_ops);
3434         if (ret < 0) {
3435                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3436                 goto err_unreg_exp_subsys;
3437         }
3438 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3439         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3440         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3441 #endif
3442         return 0;
3443
3444 err_unreg_exp_subsys:
3445         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3446 err_unreg_subsys:
3447         nfnetlink_subsys_unregister(&ctnl_subsys);
3448 err_out:
3449         return ret;
3450 }
3451
3452 static void __exit ctnetlink_exit(void)
3453 {
3454         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3455
3456         unregister_pernet_subsys(&ctnetlink_net_ops);
3457         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3458         nfnetlink_subsys_unregister(&ctnl_subsys);
3459 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3460         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3461 #endif
3462         synchronize_rcu();
3463 }
3464
3465 module_init(ctnetlink_init);
3466 module_exit(ctnetlink_exit);