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