net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #ifdef CONFIG_NF_NAT_NEEDED
  49 #include <net/netfilter/nf_nat_core.h>
  50 #include <net/netfilter/nf_nat_l4proto.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 MODULE_LICENSE("GPL");
  58
  59 static char __initdata version[] = "0.93";
  60
  61 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  62                                        const struct nf_conntrack_tuple *tuple,
  63                                        struct nf_conntrack_l4proto *l4proto)
  64 {
  65         int ret = 0;
  66         struct nlattr *nest_parms;
  67
  68         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  69         if (!nest_parms)
  70                 goto nla_put_failure;
  71         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  72                 goto nla_put_failure;
  73
  74         if (likely(l4proto->tuple_to_nlattr))
  75                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  76
  77         nla_nest_end(skb, nest_parms);
  78
  79         return ret;
  80
  81 nla_put_failure:
  82         return -1;
  83 }
  84
  85 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  86                                     const struct nf_conntrack_tuple *tuple,
  87                                     struct nf_conntrack_l3proto *l3proto)
  88 {
  89         int ret = 0;
  90         struct nlattr *nest_parms;
  91
  92         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  93         if (!nest_parms)
  94                 goto nla_put_failure;
  95
  96         if (likely(l3proto->tuple_to_nlattr))
  97                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  98
  99         nla_nest_end(skb, nest_parms);
 100
 101         return ret;
 102
 103 nla_put_failure:
 104         return -1;
 105 }
 106
 107 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 108                                  const struct nf_conntrack_tuple *tuple)
 109 {
 110         int ret;
 111         struct nf_conntrack_l3proto *l3proto;
 112         struct nf_conntrack_l4proto *l4proto;
 113
 114         rcu_read_lock();
 115         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 116         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 117
 118         if (ret >= 0) {
 119                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 120                                                tuple->dst.protonum);
 121                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 122         }
 123         rcu_read_unlock();
 124         return ret;
 125 }
 126
 127 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 128                                   const struct nf_conntrack_zone *zone, int dir)
 129 {
 130         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 131                 return 0;
 132         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 133                 goto nla_put_failure;
 134         return 0;
 135
 136 nla_put_failure:
 137         return -1;
 138 }
 139
 140 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 141 {
 142         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 143                 goto nla_put_failure;
 144         return 0;
 145
 146 nla_put_failure:
 147         return -1;
 148 }
 149
 150 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 151 {
 152         long timeout = nf_ct_expires(ct) / HZ;
 153
 154         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 155                 goto nla_put_failure;
 156         return 0;
 157
 158 nla_put_failure:
 159         return -1;
 160 }
 161
 162 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 163 {
 164         struct nf_conntrack_l4proto *l4proto;
 165         struct nlattr *nest_proto;
 166         int ret;
 167
 168         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 169         if (!l4proto->to_nlattr)
 170                 return 0;
 171
 172         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 173         if (!nest_proto)
 174                 goto nla_put_failure;
 175
 176         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 177
 178         nla_nest_end(skb, nest_proto);
 179
 180         return ret;
 181
 182 nla_put_failure:
 183         return -1;
 184 }
 185
 186 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 187                                    const struct nf_conn *ct)
 188 {
 189         struct nlattr *nest_helper;
 190         const struct nf_conn_help *help = nfct_help(ct);
 191         struct nf_conntrack_helper *helper;
 192
 193         if (!help)
 194                 return 0;
 195
 196         helper = rcu_dereference(help->helper);
 197         if (!helper)
 198                 goto out;
 199
 200         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 201         if (!nest_helper)
 202                 goto nla_put_failure;
 203         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 204                 goto nla_put_failure;
 205
 206         if (helper->to_nlattr)
 207                 helper->to_nlattr(skb, ct);
 208
 209         nla_nest_end(skb, nest_helper);
 210 out:
 211         return 0;
 212
 213 nla_put_failure:
 214         return -1;
 215 }
 216
 217 static int
 218 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 219               enum ip_conntrack_dir dir, int type)
 220 {
 221         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 222         struct nf_conn_counter *counter = acct->counter;
 223         struct nlattr *nest_count;
 224         u64 pkts, bytes;
 225
 226         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 227                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 228                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 229         } else {
 230                 pkts = atomic64_read(&counter[dir].packets);
 231                 bytes = atomic64_read(&counter[dir].bytes);
 232         }
 233
 234         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 235         if (!nest_count)
 236                 goto nla_put_failure;
 237
 238         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 239                          CTA_COUNTERS_PAD) ||
 240             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 241                          CTA_COUNTERS_PAD))
 242                 goto nla_put_failure;
 243
 244         nla_nest_end(skb, nest_count);
 245
 246         return 0;
 247
 248 nla_put_failure:
 249         return -1;
 250 }
 251
 252 static int
 253 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 254 {
 255         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 256
 257         if (!acct)
 258                 return 0;
 259
 260         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 261                 return -1;
 262         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 263                 return -1;
 264
 265         return 0;
 266 }
 267
 268 static int
 269 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 270 {
 271         struct nlattr *nest_count;
 272         const struct nf_conn_tstamp *tstamp;
 273
 274         tstamp = nf_conn_tstamp_find(ct);
 275         if (!tstamp)
 276                 return 0;
 277
 278         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 279         if (!nest_count)
 280                 goto nla_put_failure;
 281
 282         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 283                          CTA_TIMESTAMP_PAD) ||
 284             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 285                                                cpu_to_be64(tstamp->stop),
 286                                                CTA_TIMESTAMP_PAD)))
 287                 goto nla_put_failure;
 288         nla_nest_end(skb, nest_count);
 289
 290         return 0;
 291
 292 nla_put_failure:
 293         return -1;
 294 }
 295
 296 #ifdef CONFIG_NF_CONNTRACK_MARK
 297 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 298 {
 299         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 300                 goto nla_put_failure;
 301         return 0;
 302
 303 nla_put_failure:
 304         return -1;
 305 }
 306 #else
 307 #define ctnetlink_dump_mark(a, b) (0)
 308 #endif
 309
 310 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 311 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 312 {
 313         struct nlattr *nest_secctx;
 314         int len, ret;
 315         char *secctx;
 316
 317         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 318         if (ret)
 319                 return 0;
 320
 321         ret = -1;
 322         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 323         if (!nest_secctx)
 324                 goto nla_put_failure;
 325
 326         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 327                 goto nla_put_failure;
 328         nla_nest_end(skb, nest_secctx);
 329
 330         ret = 0;
 331 nla_put_failure:
 332         security_release_secctx(secctx, len);
 333         return ret;
 334 }
 335 #else
 336 #define ctnetlink_dump_secctx(a, b) (0)
 337 #endif
 338
 339 #ifdef CONFIG_NF_CONNTRACK_LABELS
 340 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 341 {
 342         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 343
 344         if (!labels)
 345                 return 0;
 346         return nla_total_size(sizeof(labels->bits));
 347 }
 348
 349 static int
 350 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 351 {
 352         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 353         unsigned int i;
 354
 355         if (!labels)
 356                 return 0;
 357
 358         i = 0;
 359         do {
 360                 if (labels->bits[i] != 0)
 361                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 362                                        labels->bits);
 363                 i++;
 364         } while (i < ARRAY_SIZE(labels->bits));
 365
 366         return 0;
 367 }
 368 #else
 369 #define ctnetlink_dump_labels(a, b) (0)
 370 #define ctnetlink_label_size(a) (0)
 371 #endif
 372
 373 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 374
 375 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 376 {
 377         struct nlattr *nest_parms;
 378
 379         if (!(ct->status & IPS_EXPECTED))
 380                 return 0;
 381
 382         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 383         if (!nest_parms)
 384                 goto nla_put_failure;
 385         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 386                 goto nla_put_failure;
 387         nla_nest_end(skb, nest_parms);
 388
 389         return 0;
 390
 391 nla_put_failure:
 392         return -1;
 393 }
 394
 395 static int
 396 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 397 {
 398         struct nlattr *nest_parms;
 399
 400         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 401         if (!nest_parms)
 402                 goto nla_put_failure;
 403
 404         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 405                          htonl(seq->correction_pos)) ||
 406             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 407                          htonl(seq->offset_before)) ||
 408             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 409                          htonl(seq->offset_after)))
 410                 goto nla_put_failure;
 411
 412         nla_nest_end(skb, nest_parms);
 413
 414         return 0;
 415
 416 nla_put_failure:
 417         return -1;
 418 }
 419
 420 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 421 {
 422         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 423         struct nf_ct_seqadj *seq;
 424
 425         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 426                 return 0;
 427
 428         spin_lock_bh(&ct->lock);
 429         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 430         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 431                 goto err;
 432
 433         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 434         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 435                 goto err;
 436
 437         spin_unlock_bh(&ct->lock);
 438         return 0;
 439 err:
 440         spin_unlock_bh(&ct->lock);
 441         return -1;
 442 }
 443
 444 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 445 {
 446         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 447                 goto nla_put_failure;
 448         return 0;
 449
 450 nla_put_failure:
 451         return -1;
 452 }
 453
 454 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 455 {
 456         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 457                 goto nla_put_failure;
 458         return 0;
 459
 460 nla_put_failure:
 461         return -1;
 462 }
 463
 464 static int
 465 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 466                     struct nf_conn *ct)
 467 {
 468         const struct nf_conntrack_zone *zone;
 469         struct nlmsghdr *nlh;
 470         struct nfgenmsg *nfmsg;
 471         struct nlattr *nest_parms;
 472         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 473
 474         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 475         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 476         if (nlh == NULL)
 477                 goto nlmsg_failure;
 478
 479         nfmsg = nlmsg_data(nlh);
 480         nfmsg->nfgen_family = nf_ct_l3num(ct);
 481         nfmsg->version      = NFNETLINK_V0;
 482         nfmsg->res_id       = 0;
 483
 484         zone = nf_ct_zone(ct);
 485
 486         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 487         if (!nest_parms)
 488                 goto nla_put_failure;
 489         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 490                 goto nla_put_failure;
 491         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 492                                    NF_CT_ZONE_DIR_ORIG) < 0)
 493                 goto nla_put_failure;
 494         nla_nest_end(skb, nest_parms);
 495
 496         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 497         if (!nest_parms)
 498                 goto nla_put_failure;
 499         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 500                 goto nla_put_failure;
 501         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 502                                    NF_CT_ZONE_DIR_REPL) < 0)
 503                 goto nla_put_failure;
 504         nla_nest_end(skb, nest_parms);
 505
 506         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 507                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 508                 goto nla_put_failure;
 509
 510         if (ctnetlink_dump_status(skb, ct) < 0 ||
 511             ctnetlink_dump_timeout(skb, ct) < 0 ||
 512             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 513             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 514             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 515             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 516             ctnetlink_dump_mark(skb, ct) < 0 ||
 517             ctnetlink_dump_secctx(skb, ct) < 0 ||
 518             ctnetlink_dump_labels(skb, ct) < 0 ||
 519             ctnetlink_dump_id(skb, ct) < 0 ||
 520             ctnetlink_dump_use(skb, ct) < 0 ||
 521             ctnetlink_dump_master(skb, ct) < 0 ||
 522             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 523                 goto nla_put_failure;
 524
 525         nlmsg_end(skb, nlh);
 526         return skb->len;
 527
 528 nlmsg_failure:
 529 nla_put_failure:
 530         nlmsg_cancel(skb, nlh);
 531         return -1;
 532 }
 533
 534 static inline size_t ctnetlink_proto_size(const struct nf_conn *ct)
 535 {
 536         struct nf_conntrack_l3proto *l3proto;
 537         struct nf_conntrack_l4proto *l4proto;
 538         size_t len = 0;
 539
 540         rcu_read_lock();
 541         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 542         len += l3proto->nla_size;
 543
 544         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 545         len += l4proto->nla_size;
 546         rcu_read_unlock();
 547
 548         return len;
 549 }
 550
 551 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 552 {
 553         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 554                 return 0;
 555         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 556                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 557                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 558                ;
 559 }
 560
 561 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 562 {
 563 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 564         int len, ret;
 565
 566         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 567         if (ret)
 568                 return 0;
 569
 570         return nla_total_size(0) /* CTA_SECCTX */
 571                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 572 #else
 573         return 0;
 574 #endif
 575 }
 576
 577 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 578 {
 579 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 580         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 581                 return 0;
 582         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 583 #else
 584         return 0;
 585 #endif
 586 }
 587
 588 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 589 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 590 {
 591         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 592                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 593                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 594                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 595                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 596                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 597                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 598                + ctnetlink_acct_size(ct)
 599                + ctnetlink_timestamp_size(ct)
 600                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 601                + nla_total_size(0) /* CTA_PROTOINFO */
 602                + nla_total_size(0) /* CTA_HELP */
 603                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 604                + ctnetlink_secctx_size(ct)
 605 #ifdef CONFIG_NF_NAT_NEEDED
 606                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 607                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 608 #endif
 609 #ifdef CONFIG_NF_CONNTRACK_MARK
 610                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 611 #endif
 612 #ifdef CONFIG_NF_CONNTRACK_ZONES
 613                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 614 #endif
 615                + ctnetlink_proto_size(ct)
 616                + ctnetlink_label_size(ct)
 617                ;
 618 }
 619
 620 static int
 621 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 622 {
 623         const struct nf_conntrack_zone *zone;
 624         struct net *net;
 625         struct nlmsghdr *nlh;
 626         struct nfgenmsg *nfmsg;
 627         struct nlattr *nest_parms;
 628         struct nf_conn *ct = item->ct;
 629         struct sk_buff *skb;
 630         unsigned int type;
 631         unsigned int flags = 0, group;
 632         int err;
 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_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 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, 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, NULL);
 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                                NULL);
 943         if (ret < 0)
 944                 return ret;
 945
 946         if (!tb[CTA_PROTO_NUM])
 947                 return -EINVAL;
 948         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 949
 950         rcu_read_lock();
 951         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 952
 953         if (likely(l4proto->nlattr_to_tuple)) {
 954                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 955                                           l4proto->nla_policy, NULL);
 956                 if (ret == 0)
 957                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 958         }
 959
 960         rcu_read_unlock();
 961
 962         return ret;
 963 }
 964
 965 static int
 966 ctnetlink_parse_zone(const struct nlattr *attr,
 967                      struct nf_conntrack_zone *zone)
 968 {
 969         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
 970                         NF_CT_DEFAULT_ZONE_DIR, 0);
 971 #ifdef CONFIG_NF_CONNTRACK_ZONES
 972         if (attr)
 973                 zone->id = ntohs(nla_get_be16(attr));
 974 #else
 975         if (attr)
 976                 return -EOPNOTSUPP;
 977 #endif
 978         return 0;
 979 }
 980
 981 static int
 982 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
 983                            struct nf_conntrack_zone *zone)
 984 {
 985         int ret;
 986
 987         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
 988                 return -EINVAL;
 989
 990         ret = ctnetlink_parse_zone(attr, zone);
 991         if (ret < 0)
 992                 return ret;
 993
 994         if (type == CTA_TUPLE_REPLY)
 995                 zone->dir = NF_CT_ZONE_DIR_REPL;
 996         else
 997                 zone->dir = NF_CT_ZONE_DIR_ORIG;
 998
 999         return 0;
1000 }
1001
1002 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1003         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1004         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1005         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1006 };
1007
1008 static int
1009 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1010                       struct nf_conntrack_tuple *tuple,
1011                       enum ctattr_type type, u_int8_t l3num,
1012                       struct nf_conntrack_zone *zone)
1013 {
1014         struct nlattr *tb[CTA_TUPLE_MAX+1];
1015         int err;
1016
1017         memset(tuple, 0, sizeof(*tuple));
1018
1019         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1020                                NULL);
1021         if (err < 0)
1022                 return err;
1023
1024         if (!tb[CTA_TUPLE_IP])
1025                 return -EINVAL;
1026
1027         tuple->src.l3num = l3num;
1028
1029         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1030         if (err < 0)
1031                 return err;
1032
1033         if (!tb[CTA_TUPLE_PROTO])
1034                 return -EINVAL;
1035
1036         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1037         if (err < 0)
1038                 return err;
1039
1040         if (tb[CTA_TUPLE_ZONE]) {
1041                 if (!zone)
1042                         return -EINVAL;
1043
1044                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1045                                                  type, zone);
1046                 if (err < 0)
1047                         return err;
1048         }
1049
1050         /* orig and expect tuples get DIR_ORIGINAL */
1051         if (type == CTA_TUPLE_REPLY)
1052                 tuple->dst.dir = IP_CT_DIR_REPLY;
1053         else
1054                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1055
1056         return 0;
1057 }
1058
1059 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1060         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1061                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1062 };
1063
1064 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1065                                 struct nlattr **helpinfo)
1066 {
1067         int err;
1068         struct nlattr *tb[CTA_HELP_MAX+1];
1069
1070         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1071         if (err < 0)
1072                 return err;
1073
1074         if (!tb[CTA_HELP_NAME])
1075                 return -EINVAL;
1076
1077         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1078
1079         if (tb[CTA_HELP_INFO])
1080                 *helpinfo = tb[CTA_HELP_INFO];
1081
1082         return 0;
1083 }
1084
1085 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1086         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1087         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1088         [CTA_STATUS]            = { .type = NLA_U32 },
1089         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1090         [CTA_HELP]              = { .type = NLA_NESTED },
1091         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1092         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1093         [CTA_MARK]              = { .type = NLA_U32 },
1094         [CTA_ID]                = { .type = NLA_U32 },
1095         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1096         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1097         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1098         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1099         [CTA_ZONE]              = { .type = NLA_U16 },
1100         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1101         [CTA_LABELS]            = { .type = NLA_BINARY,
1102                                     .len = NF_CT_LABELS_MAX_SIZE },
1103         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1104                                     .len = NF_CT_LABELS_MAX_SIZE },
1105 };
1106
1107 static int ctnetlink_flush_conntrack(struct net *net,
1108                                      const struct nlattr * const cda[],
1109                                      u32 portid, int report)
1110 {
1111         struct ctnetlink_filter *filter = NULL;
1112
1113         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1114                 filter = ctnetlink_alloc_filter(cda);
1115                 if (IS_ERR(filter))
1116                         return PTR_ERR(filter);
1117         }
1118
1119         nf_ct_iterate_cleanup(net, ctnetlink_filter_match, filter,
1120                               portid, report);
1121         kfree(filter);
1122
1123         return 0;
1124 }
1125
1126 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1127                                    struct sk_buff *skb,
1128                                    const struct nlmsghdr *nlh,
1129                                    const struct nlattr * const cda[])
1130 {
1131         struct nf_conntrack_tuple_hash *h;
1132         struct nf_conntrack_tuple tuple;
1133         struct nf_conn *ct;
1134         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1135         u_int8_t u3 = nfmsg->nfgen_family;
1136         struct nf_conntrack_zone zone;
1137         int err;
1138
1139         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1140         if (err < 0)
1141                 return err;
1142
1143         if (cda[CTA_TUPLE_ORIG])
1144                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1145                                             u3, &zone);
1146         else if (cda[CTA_TUPLE_REPLY])
1147                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1148                                             u3, &zone);
1149         else {
1150                 return ctnetlink_flush_conntrack(net, cda,
1151                                                  NETLINK_CB(skb).portid,
1152                                                  nlmsg_report(nlh));
1153         }
1154
1155         if (err < 0)
1156                 return err;
1157
1158         h = nf_conntrack_find_get(net, &zone, &tuple);
1159         if (!h)
1160                 return -ENOENT;
1161
1162         ct = nf_ct_tuplehash_to_ctrack(h);
1163
1164         if (cda[CTA_ID]) {
1165                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1166                 if (id != (u32)(unsigned long)ct) {
1167                         nf_ct_put(ct);
1168                         return -ENOENT;
1169                 }
1170         }
1171
1172         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1173         nf_ct_put(ct);
1174
1175         return 0;
1176 }
1177
1178 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1179                                    struct sk_buff *skb,
1180                                    const struct nlmsghdr *nlh,
1181                                    const struct nlattr * const cda[])
1182 {
1183         struct nf_conntrack_tuple_hash *h;
1184         struct nf_conntrack_tuple tuple;
1185         struct nf_conn *ct;
1186         struct sk_buff *skb2 = NULL;
1187         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1188         u_int8_t u3 = nfmsg->nfgen_family;
1189         struct nf_conntrack_zone zone;
1190         int err;
1191
1192         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1193                 struct netlink_dump_control c = {
1194                         .dump = ctnetlink_dump_table,
1195                         .done = ctnetlink_done,
1196                 };
1197
1198                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1199                         struct ctnetlink_filter *filter;
1200
1201                         filter = ctnetlink_alloc_filter(cda);
1202                         if (IS_ERR(filter))
1203                                 return PTR_ERR(filter);
1204
1205                         c.data = filter;
1206                 }
1207                 return netlink_dump_start(ctnl, skb, nlh, &c);
1208         }
1209
1210         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1211         if (err < 0)
1212                 return err;
1213
1214         if (cda[CTA_TUPLE_ORIG])
1215                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1216                                             u3, &zone);
1217         else if (cda[CTA_TUPLE_REPLY])
1218                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1219                                             u3, &zone);
1220         else
1221                 return -EINVAL;
1222
1223         if (err < 0)
1224                 return err;
1225
1226         h = nf_conntrack_find_get(net, &zone, &tuple);
1227         if (!h)
1228                 return -ENOENT;
1229
1230         ct = nf_ct_tuplehash_to_ctrack(h);
1231
1232         err = -ENOMEM;
1233         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1234         if (skb2 == NULL) {
1235                 nf_ct_put(ct);
1236                 return -ENOMEM;
1237         }
1238
1239         rcu_read_lock();
1240         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1241                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1242         rcu_read_unlock();
1243         nf_ct_put(ct);
1244         if (err <= 0)
1245                 goto free;
1246
1247         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1248         if (err < 0)
1249                 goto out;
1250
1251         return 0;
1252
1253 free:
1254         kfree_skb(skb2);
1255 out:
1256         /* this avoids a loop in nfnetlink. */
1257         return err == -EAGAIN ? -ENOBUFS : err;
1258 }
1259
1260 static int ctnetlink_done_list(struct netlink_callback *cb)
1261 {
1262         if (cb->args[1])
1263                 nf_ct_put((struct nf_conn *)cb->args[1]);
1264         return 0;
1265 }
1266
1267 static int
1268 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1269 {
1270         struct nf_conn *ct, *last;
1271         struct nf_conntrack_tuple_hash *h;
1272         struct hlist_nulls_node *n;
1273         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1274         u_int8_t l3proto = nfmsg->nfgen_family;
1275         int res;
1276         int cpu;
1277         struct hlist_nulls_head *list;
1278         struct net *net = sock_net(skb->sk);
1279
1280         if (cb->args[2])
1281                 return 0;
1282
1283         last = (struct nf_conn *)cb->args[1];
1284
1285         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1286                 struct ct_pcpu *pcpu;
1287
1288                 if (!cpu_possible(cpu))
1289                         continue;
1290
1291                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1292                 spin_lock_bh(&pcpu->lock);
1293                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1294 restart:
1295                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1296                         ct = nf_ct_tuplehash_to_ctrack(h);
1297                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1298                                 continue;
1299                         if (cb->args[1]) {
1300                                 if (ct != last)
1301                                         continue;
1302                                 cb->args[1] = 0;
1303                         }
1304                         rcu_read_lock();
1305                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1306                                                   cb->nlh->nlmsg_seq,
1307                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1308                                                   ct);
1309                         rcu_read_unlock();
1310                         if (res < 0) {
1311                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1312                                         continue;
1313                                 cb->args[0] = cpu;
1314                                 cb->args[1] = (unsigned long)ct;
1315                                 spin_unlock_bh(&pcpu->lock);
1316                                 goto out;
1317                         }
1318                 }
1319                 if (cb->args[1]) {
1320                         cb->args[1] = 0;
1321                         goto restart;
1322                 }
1323                 spin_unlock_bh(&pcpu->lock);
1324         }
1325         cb->args[2] = 1;
1326 out:
1327         if (last)
1328                 nf_ct_put(last);
1329
1330         return skb->len;
1331 }
1332
1333 static int
1334 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1335 {
1336         return ctnetlink_dump_list(skb, cb, true);
1337 }
1338
1339 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1340                                   struct sk_buff *skb,
1341                                   const struct nlmsghdr *nlh,
1342                                   const struct nlattr * const cda[])
1343 {
1344         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1345                 struct netlink_dump_control c = {
1346                         .dump = ctnetlink_dump_dying,
1347                         .done = ctnetlink_done_list,
1348                 };
1349                 return netlink_dump_start(ctnl, skb, nlh, &c);
1350         }
1351
1352         return -EOPNOTSUPP;
1353 }
1354
1355 static int
1356 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1357 {
1358         return ctnetlink_dump_list(skb, cb, false);
1359 }
1360
1361 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1362                                         struct sk_buff *skb,
1363                                         const struct nlmsghdr *nlh,
1364                                         const struct nlattr * const cda[])
1365 {
1366         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1367                 struct netlink_dump_control c = {
1368                         .dump = ctnetlink_dump_unconfirmed,
1369                         .done = ctnetlink_done_list,
1370                 };
1371                 return netlink_dump_start(ctnl, skb, nlh, &c);
1372         }
1373
1374         return -EOPNOTSUPP;
1375 }
1376
1377 #ifdef CONFIG_NF_NAT_NEEDED
1378 static int
1379 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1380                           enum nf_nat_manip_type manip,
1381                           const struct nlattr *attr)
1382 {
1383         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1384         int err;
1385
1386         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1387         if (!parse_nat_setup) {
1388 #ifdef CONFIG_MODULES
1389                 rcu_read_unlock();
1390                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1391                 if (request_module("nf-nat") < 0) {
1392                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1393                         rcu_read_lock();
1394                         return -EOPNOTSUPP;
1395                 }
1396                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1397                 rcu_read_lock();
1398                 if (nfnetlink_parse_nat_setup_hook)
1399                         return -EAGAIN;
1400 #endif
1401                 return -EOPNOTSUPP;
1402         }
1403
1404         err = parse_nat_setup(ct, manip, attr);
1405         if (err == -EAGAIN) {
1406 #ifdef CONFIG_MODULES
1407                 rcu_read_unlock();
1408                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1409                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1410                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1411                         rcu_read_lock();
1412                         return -EOPNOTSUPP;
1413                 }
1414                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1415                 rcu_read_lock();
1416 #else
1417                 err = -EOPNOTSUPP;
1418 #endif
1419         }
1420         return err;
1421 }
1422 #endif
1423
1424 static void
1425 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1426                           unsigned long off)
1427 {
1428         unsigned int bit;
1429
1430         /* Ignore these unchangable bits */
1431         on &= ~IPS_UNCHANGEABLE_MASK;
1432         off &= ~IPS_UNCHANGEABLE_MASK;
1433
1434         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1435                 if (on & (1 << bit))
1436                         set_bit(bit, &ct->status);
1437                 else if (off & (1 << bit))
1438                         clear_bit(bit, &ct->status);
1439         }
1440 }
1441
1442 static int
1443 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1444 {
1445         unsigned long d;
1446         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1447         d = ct->status ^ status;
1448
1449         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1450                 /* unchangeable */
1451                 return -EBUSY;
1452
1453         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1454                 /* SEEN_REPLY bit can only be set */
1455                 return -EBUSY;
1456
1457         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1458                 /* ASSURED bit can only be set */
1459                 return -EBUSY;
1460
1461         __ctnetlink_change_status(ct, status, 0);
1462         return 0;
1463 }
1464
1465 static int
1466 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1467 {
1468 #ifdef CONFIG_NF_NAT_NEEDED
1469         int ret;
1470
1471         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1472                 return 0;
1473
1474         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1475                                         cda[CTA_NAT_DST]);
1476         if (ret < 0)
1477                 return ret;
1478
1479         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1480                                         cda[CTA_NAT_SRC]);
1481         return ret;
1482 #else
1483         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1484                 return 0;
1485         return -EOPNOTSUPP;
1486 #endif
1487 }
1488
1489 static int ctnetlink_change_helper(struct nf_conn *ct,
1490                                    const struct nlattr * const cda[])
1491 {
1492         struct nf_conntrack_helper *helper;
1493         struct nf_conn_help *help = nfct_help(ct);
1494         char *helpname = NULL;
1495         struct nlattr *helpinfo = NULL;
1496         int err;
1497
1498         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1499         if (err < 0)
1500                 return err;
1501
1502         /* don't change helper of sibling connections */
1503         if (ct->master) {
1504                 /* If we try to change the helper to the same thing twice,
1505                  * treat the second attempt as a no-op instead of returning
1506                  * an error.
1507                  */
1508                 err = -EBUSY;
1509                 if (help) {
1510                         rcu_read_lock();
1511                         helper = rcu_dereference(help->helper);
1512                         if (helper && !strcmp(helper->name, helpname))
1513                                 err = 0;
1514                         rcu_read_unlock();
1515                 }
1516
1517                 return err;
1518         }
1519
1520         if (!strcmp(helpname, "")) {
1521                 if (help && help->helper) {
1522                         /* we had a helper before ... */
1523                         nf_ct_remove_expectations(ct);
1524                         RCU_INIT_POINTER(help->helper, NULL);
1525                 }
1526
1527                 return 0;
1528         }
1529
1530         rcu_read_lock();
1531         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1532                                             nf_ct_protonum(ct));
1533         if (helper == NULL) {
1534                 rcu_read_unlock();
1535                 return -EOPNOTSUPP;
1536         }
1537
1538         if (help) {
1539                 if (help->helper == helper) {
1540                         /* update private helper data if allowed. */
1541                         if (helper->from_nlattr)
1542                                 helper->from_nlattr(helpinfo, ct);
1543                         err = 0;
1544                 } else
1545                         err = -EBUSY;
1546         } else {
1547                 /* we cannot set a helper for an existing conntrack */
1548                 err = -EOPNOTSUPP;
1549         }
1550
1551         rcu_read_unlock();
1552         return err;
1553 }
1554
1555 static int ctnetlink_change_timeout(struct nf_conn *ct,
1556                                     const struct nlattr * const cda[])
1557 {
1558         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1559
1560         ct->timeout = nfct_time_stamp + timeout * HZ;
1561
1562         if (test_bit(IPS_DYING_BIT, &ct->status))
1563                 return -ETIME;
1564
1565         return 0;
1566 }
1567
1568 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1569         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1570         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1571         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1572 };
1573
1574 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1575                                       const struct nlattr * const cda[])
1576 {
1577         const struct nlattr *attr = cda[CTA_PROTOINFO];
1578         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1579         struct nf_conntrack_l4proto *l4proto;
1580         int err = 0;
1581
1582         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1583                                NULL);
1584         if (err < 0)
1585                 return err;
1586
1587         rcu_read_lock();
1588         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1589         if (l4proto->from_nlattr)
1590                 err = l4proto->from_nlattr(tb, ct);
1591         rcu_read_unlock();
1592
1593         return err;
1594 }
1595
1596 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1597         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1598         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1599         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1600 };
1601
1602 static int change_seq_adj(struct nf_ct_seqadj *seq,
1603                           const struct nlattr * const attr)
1604 {
1605         int err;
1606         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1607
1608         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1609         if (err < 0)
1610                 return err;
1611
1612         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1613                 return -EINVAL;
1614
1615         seq->correction_pos =
1616                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1617
1618         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1619                 return -EINVAL;
1620
1621         seq->offset_before =
1622                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1623
1624         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1625                 return -EINVAL;
1626
1627         seq->offset_after =
1628                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1629
1630         return 0;
1631 }
1632
1633 static int
1634 ctnetlink_change_seq_adj(struct nf_conn *ct,
1635                          const struct nlattr * const cda[])
1636 {
1637         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1638         int ret = 0;
1639
1640         if (!seqadj)
1641                 return 0;
1642
1643         spin_lock_bh(&ct->lock);
1644         if (cda[CTA_SEQ_ADJ_ORIG]) {
1645                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1646                                      cda[CTA_SEQ_ADJ_ORIG]);
1647                 if (ret < 0)
1648                         goto err;
1649
1650                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1651         }
1652
1653         if (cda[CTA_SEQ_ADJ_REPLY]) {
1654                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1655                                      cda[CTA_SEQ_ADJ_REPLY]);
1656                 if (ret < 0)
1657                         goto err;
1658
1659                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1660         }
1661
1662         spin_unlock_bh(&ct->lock);
1663         return 0;
1664 err:
1665         spin_unlock_bh(&ct->lock);
1666         return ret;
1667 }
1668
1669 static int
1670 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1671 {
1672 #ifdef CONFIG_NF_CONNTRACK_LABELS
1673         size_t len = nla_len(cda[CTA_LABELS]);
1674         const void *mask = cda[CTA_LABELS_MASK];
1675
1676         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1677                 return -EINVAL;
1678
1679         if (mask) {
1680                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1681                     nla_len(cda[CTA_LABELS_MASK]) != len)
1682                         return -EINVAL;
1683                 mask = nla_data(cda[CTA_LABELS_MASK]);
1684         }
1685
1686         len /= sizeof(u32);
1687
1688         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1689 #else
1690         return -EOPNOTSUPP;
1691 #endif
1692 }
1693
1694 static int
1695 ctnetlink_change_conntrack(struct nf_conn *ct,
1696                            const struct nlattr * const cda[])
1697 {
1698         int err;
1699
1700         /* only allow NAT changes and master assignation for new conntracks */
1701         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1702                 return -EOPNOTSUPP;
1703
1704         if (cda[CTA_HELP]) {
1705                 err = ctnetlink_change_helper(ct, cda);
1706                 if (err < 0)
1707                         return err;
1708         }
1709
1710         if (cda[CTA_TIMEOUT]) {
1711                 err = ctnetlink_change_timeout(ct, cda);
1712                 if (err < 0)
1713                         return err;
1714         }
1715
1716         if (cda[CTA_STATUS]) {
1717                 err = ctnetlink_change_status(ct, cda);
1718                 if (err < 0)
1719                         return err;
1720         }
1721
1722         if (cda[CTA_PROTOINFO]) {
1723                 err = ctnetlink_change_protoinfo(ct, cda);
1724                 if (err < 0)
1725                         return err;
1726         }
1727
1728 #if defined(CONFIG_NF_CONNTRACK_MARK)
1729         if (cda[CTA_MARK])
1730                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1731 #endif
1732
1733         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1734                 err = ctnetlink_change_seq_adj(ct, cda);
1735                 if (err < 0)
1736                         return err;
1737         }
1738
1739         if (cda[CTA_LABELS]) {
1740                 err = ctnetlink_attach_labels(ct, cda);
1741                 if (err < 0)
1742                         return err;
1743         }
1744
1745         return 0;
1746 }
1747
1748 static struct nf_conn *
1749 ctnetlink_create_conntrack(struct net *net,
1750                            const struct nf_conntrack_zone *zone,
1751                            const struct nlattr * const cda[],
1752                            struct nf_conntrack_tuple *otuple,
1753                            struct nf_conntrack_tuple *rtuple,
1754                            u8 u3)
1755 {
1756         struct nf_conn *ct;
1757         int err = -EINVAL;
1758         struct nf_conntrack_helper *helper;
1759         struct nf_conn_tstamp *tstamp;
1760
1761         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1762         if (IS_ERR(ct))
1763                 return ERR_PTR(-ENOMEM);
1764
1765         if (!cda[CTA_TIMEOUT])
1766                 goto err1;
1767
1768         ct->timeout = nfct_time_stamp + ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1769
1770         rcu_read_lock();
1771         if (cda[CTA_HELP]) {
1772                 char *helpname = NULL;
1773                 struct nlattr *helpinfo = NULL;
1774
1775                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1776                 if (err < 0)
1777                         goto err2;
1778
1779                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1780                                                     nf_ct_protonum(ct));
1781                 if (helper == NULL) {
1782                         rcu_read_unlock();
1783 #ifdef CONFIG_MODULES
1784                         if (request_module("nfct-helper-%s", helpname) < 0) {
1785                                 err = -EOPNOTSUPP;
1786                                 goto err1;
1787                         }
1788
1789                         rcu_read_lock();
1790                         helper = __nf_conntrack_helper_find(helpname,
1791                                                             nf_ct_l3num(ct),
1792                                                             nf_ct_protonum(ct));
1793                         if (helper) {
1794                                 err = -EAGAIN;
1795                                 goto err2;
1796                         }
1797                         rcu_read_unlock();
1798 #endif
1799                         err = -EOPNOTSUPP;
1800                         goto err1;
1801                 } else {
1802                         struct nf_conn_help *help;
1803
1804                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1805                         if (help == NULL) {
1806                                 err = -ENOMEM;
1807                                 goto err2;
1808                         }
1809                         /* set private helper data if allowed. */
1810                         if (helper->from_nlattr)
1811                                 helper->from_nlattr(helpinfo, ct);
1812
1813                         /* not in hash table yet so not strictly necessary */
1814                         RCU_INIT_POINTER(help->helper, helper);
1815                 }
1816         } else {
1817                 /* try an implicit helper assignation */
1818                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1819                 if (err < 0)
1820                         goto err2;
1821         }
1822
1823         err = ctnetlink_setup_nat(ct, cda);
1824         if (err < 0)
1825                 goto err2;
1826
1827         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1828         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1829         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1830         nf_ct_labels_ext_add(ct);
1831
1832         /* we must add conntrack extensions before confirmation. */
1833         ct->status |= IPS_CONFIRMED;
1834
1835         if (cda[CTA_STATUS]) {
1836                 err = ctnetlink_change_status(ct, cda);
1837                 if (err < 0)
1838                         goto err2;
1839         }
1840
1841         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1842                 err = ctnetlink_change_seq_adj(ct, cda);
1843                 if (err < 0)
1844                         goto err2;
1845         }
1846
1847         memset(&ct->proto, 0, sizeof(ct->proto));
1848         if (cda[CTA_PROTOINFO]) {
1849                 err = ctnetlink_change_protoinfo(ct, cda);
1850                 if (err < 0)
1851                         goto err2;
1852         }
1853
1854 #if defined(CONFIG_NF_CONNTRACK_MARK)
1855         if (cda[CTA_MARK])
1856                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1857 #endif
1858
1859         /* setup master conntrack: this is a confirmed expectation */
1860         if (cda[CTA_TUPLE_MASTER]) {
1861                 struct nf_conntrack_tuple master;
1862                 struct nf_conntrack_tuple_hash *master_h;
1863                 struct nf_conn *master_ct;
1864
1865                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1866                                             u3, NULL);
1867                 if (err < 0)
1868                         goto err2;
1869
1870                 master_h = nf_conntrack_find_get(net, zone, &master);
1871                 if (master_h == NULL) {
1872                         err = -ENOENT;
1873                         goto err2;
1874                 }
1875                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1876                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1877                 ct->master = master_ct;
1878         }
1879         tstamp = nf_conn_tstamp_find(ct);
1880         if (tstamp)
1881                 tstamp->start = ktime_get_real_ns();
1882
1883         err = nf_conntrack_hash_check_insert(ct);
1884         if (err < 0)
1885                 goto err2;
1886
1887         rcu_read_unlock();
1888
1889         return ct;
1890
1891 err2:
1892         rcu_read_unlock();
1893 err1:
1894         nf_conntrack_free(ct);
1895         return ERR_PTR(err);
1896 }
1897
1898 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
1899                                    struct sk_buff *skb,
1900                                    const struct nlmsghdr *nlh,
1901                                    const struct nlattr * const cda[])
1902 {
1903         struct nf_conntrack_tuple otuple, rtuple;
1904         struct nf_conntrack_tuple_hash *h = NULL;
1905         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1906         struct nf_conn *ct;
1907         u_int8_t u3 = nfmsg->nfgen_family;
1908         struct nf_conntrack_zone zone;
1909         int err;
1910
1911         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1912         if (err < 0)
1913                 return err;
1914
1915         if (cda[CTA_TUPLE_ORIG]) {
1916                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1917                                             u3, &zone);
1918                 if (err < 0)
1919                         return err;
1920         }
1921
1922         if (cda[CTA_TUPLE_REPLY]) {
1923                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1924                                             u3, &zone);
1925                 if (err < 0)
1926                         return err;
1927         }
1928
1929         if (cda[CTA_TUPLE_ORIG])
1930                 h = nf_conntrack_find_get(net, &zone, &otuple);
1931         else if (cda[CTA_TUPLE_REPLY])
1932                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1933
1934         if (h == NULL) {
1935                 err = -ENOENT;
1936                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1937                         enum ip_conntrack_events events;
1938
1939                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1940                                 return -EINVAL;
1941                         if (otuple.dst.protonum != rtuple.dst.protonum)
1942                                 return -EINVAL;
1943
1944                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1945                                                         &rtuple, u3);
1946                         if (IS_ERR(ct))
1947                                 return PTR_ERR(ct);
1948
1949                         err = 0;
1950                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1951                                 events = 1 << IPCT_RELATED;
1952                         else
1953                                 events = 1 << IPCT_NEW;
1954
1955                         if (cda[CTA_LABELS] &&
1956                             ctnetlink_attach_labels(ct, cda) == 0)
1957                                 events |= (1 << IPCT_LABEL);
1958
1959                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1960                                                       (1 << IPCT_ASSURED) |
1961                                                       (1 << IPCT_HELPER) |
1962                                                       (1 << IPCT_PROTOINFO) |
1963                                                       (1 << IPCT_SEQADJ) |
1964                                                       (1 << IPCT_MARK) | events,
1965                                                       ct, NETLINK_CB(skb).portid,
1966                                                       nlmsg_report(nlh));
1967                         nf_ct_put(ct);
1968                 }
1969
1970                 return err;
1971         }
1972         /* implicit 'else' */
1973
1974         err = -EEXIST;
1975         ct = nf_ct_tuplehash_to_ctrack(h);
1976         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1977                 err = ctnetlink_change_conntrack(ct, cda);
1978                 if (err == 0) {
1979                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1980                                                       (1 << IPCT_ASSURED) |
1981                                                       (1 << IPCT_HELPER) |
1982                                                       (1 << IPCT_LABEL) |
1983                                                       (1 << IPCT_PROTOINFO) |
1984                                                       (1 << IPCT_SEQADJ) |
1985                                                       (1 << IPCT_MARK),
1986                                                       ct, NETLINK_CB(skb).portid,
1987                                                       nlmsg_report(nlh));
1988                 }
1989         }
1990
1991         nf_ct_put(ct);
1992         return err;
1993 }
1994
1995 static int
1996 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1997                                 __u16 cpu, const struct ip_conntrack_stat *st)
1998 {
1999         struct nlmsghdr *nlh;
2000         struct nfgenmsg *nfmsg;
2001         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2002
2003         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2004                               IPCTNL_MSG_CT_GET_STATS_CPU);
2005         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2006         if (nlh == NULL)
2007                 goto nlmsg_failure;
2008
2009         nfmsg = nlmsg_data(nlh);
2010         nfmsg->nfgen_family = AF_UNSPEC;
2011         nfmsg->version      = NFNETLINK_V0;
2012         nfmsg->res_id       = htons(cpu);
2013
2014         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2015             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2016             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2017             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2018             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2019                                 htonl(st->insert_failed)) ||
2020             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2021             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2022             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2023             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2024                                 htonl(st->search_restart)))
2025                 goto nla_put_failure;
2026
2027         nlmsg_end(skb, nlh);
2028         return skb->len;
2029
2030 nla_put_failure:
2031 nlmsg_failure:
2032         nlmsg_cancel(skb, nlh);
2033         return -1;
2034 }
2035
2036 static int
2037 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2038 {
2039         int cpu;
2040         struct net *net = sock_net(skb->sk);
2041
2042         if (cb->args[0] == nr_cpu_ids)
2043                 return 0;
2044
2045         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2046                 const struct ip_conntrack_stat *st;
2047
2048                 if (!cpu_possible(cpu))
2049                         continue;
2050
2051                 st = per_cpu_ptr(net->ct.stat, cpu);
2052                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2053                                                     NETLINK_CB(cb->skb).portid,
2054                                                     cb->nlh->nlmsg_seq,
2055                                                     cpu, st) < 0)
2056                                 break;
2057         }
2058         cb->args[0] = cpu;
2059
2060         return skb->len;
2061 }
2062
2063 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2064                                  struct sk_buff *skb,
2065                                  const struct nlmsghdr *nlh,
2066                                  const struct nlattr * const cda[])
2067 {
2068         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2069                 struct netlink_dump_control c = {
2070                         .dump = ctnetlink_ct_stat_cpu_dump,
2071                 };
2072                 return netlink_dump_start(ctnl, skb, nlh, &c);
2073         }
2074
2075         return 0;
2076 }
2077
2078 static int
2079 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2080                             struct net *net)
2081 {
2082         struct nlmsghdr *nlh;
2083         struct nfgenmsg *nfmsg;
2084         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2085         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2086
2087         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2088         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2089         if (nlh == NULL)
2090                 goto nlmsg_failure;
2091
2092         nfmsg = nlmsg_data(nlh);
2093         nfmsg->nfgen_family = AF_UNSPEC;
2094         nfmsg->version      = NFNETLINK_V0;
2095         nfmsg->res_id       = 0;
2096
2097         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2098                 goto nla_put_failure;
2099
2100         nlmsg_end(skb, nlh);
2101         return skb->len;
2102
2103 nla_put_failure:
2104 nlmsg_failure:
2105         nlmsg_cancel(skb, nlh);
2106         return -1;
2107 }
2108
2109 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2110                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2111                              const struct nlattr * const cda[])
2112 {
2113         struct sk_buff *skb2;
2114         int err;
2115
2116         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2117         if (skb2 == NULL)
2118                 return -ENOMEM;
2119
2120         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2121                                           nlh->nlmsg_seq,
2122                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2123                                           sock_net(skb->sk));
2124         if (err <= 0)
2125                 goto free;
2126
2127         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2128         if (err < 0)
2129                 goto out;
2130
2131         return 0;
2132
2133 free:
2134         kfree_skb(skb2);
2135 out:
2136         /* this avoids a loop in nfnetlink. */
2137         return err == -EAGAIN ? -ENOBUFS : err;
2138 }
2139
2140 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2141         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2142         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2143         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2144         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2145         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2146         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2147                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2148         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2149         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2150         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2151         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2152         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2153 };
2154
2155 static struct nf_conntrack_expect *
2156 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2157                        struct nf_conntrack_helper *helper,
2158                        struct nf_conntrack_tuple *tuple,
2159                        struct nf_conntrack_tuple *mask);
2160
2161 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2162 static size_t
2163 ctnetlink_glue_build_size(const struct nf_conn *ct)
2164 {
2165         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2166                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2167                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2168                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2169                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2170                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2171                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2172                + nla_total_size(0) /* CTA_PROTOINFO */
2173                + nla_total_size(0) /* CTA_HELP */
2174                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2175                + ctnetlink_secctx_size(ct)
2176 #ifdef CONFIG_NF_NAT_NEEDED
2177                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2178                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2179 #endif
2180 #ifdef CONFIG_NF_CONNTRACK_MARK
2181                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2182 #endif
2183 #ifdef CONFIG_NF_CONNTRACK_ZONES
2184                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2185 #endif
2186                + ctnetlink_proto_size(ct)
2187                ;
2188 }
2189
2190 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2191                                              enum ip_conntrack_info *ctinfo)
2192 {
2193         return nf_ct_get(skb, ctinfo);
2194 }
2195
2196 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2197 {
2198         const struct nf_conntrack_zone *zone;
2199         struct nlattr *nest_parms;
2200
2201         rcu_read_lock();
2202         zone = nf_ct_zone(ct);
2203
2204         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2205         if (!nest_parms)
2206                 goto nla_put_failure;
2207         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2208                 goto nla_put_failure;
2209         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2210                                    NF_CT_ZONE_DIR_ORIG) < 0)
2211                 goto nla_put_failure;
2212         nla_nest_end(skb, nest_parms);
2213
2214         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2215         if (!nest_parms)
2216                 goto nla_put_failure;
2217         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2218                 goto nla_put_failure;
2219         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2220                                    NF_CT_ZONE_DIR_REPL) < 0)
2221                 goto nla_put_failure;
2222         nla_nest_end(skb, nest_parms);
2223
2224         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2225                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2226                 goto nla_put_failure;
2227
2228         if (ctnetlink_dump_id(skb, ct) < 0)
2229                 goto nla_put_failure;
2230
2231         if (ctnetlink_dump_status(skb, ct) < 0)
2232                 goto nla_put_failure;
2233
2234         if (ctnetlink_dump_timeout(skb, ct) < 0)
2235                 goto nla_put_failure;
2236
2237         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2238                 goto nla_put_failure;
2239
2240         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2241                 goto nla_put_failure;
2242
2243 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2244         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2245                 goto nla_put_failure;
2246 #endif
2247         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2248                 goto nla_put_failure;
2249
2250         if ((ct->status & IPS_SEQ_ADJUST) &&
2251             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2252                 goto nla_put_failure;
2253
2254 #ifdef CONFIG_NF_CONNTRACK_MARK
2255         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2256                 goto nla_put_failure;
2257 #endif
2258         if (ctnetlink_dump_labels(skb, ct) < 0)
2259                 goto nla_put_failure;
2260         rcu_read_unlock();
2261         return 0;
2262
2263 nla_put_failure:
2264         rcu_read_unlock();
2265         return -ENOSPC;
2266 }
2267
2268 static int
2269 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2270                      enum ip_conntrack_info ctinfo,
2271                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2272 {
2273         struct nlattr *nest_parms;
2274
2275         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2276         if (!nest_parms)
2277                 goto nla_put_failure;
2278
2279         if (__ctnetlink_glue_build(skb, ct) < 0)
2280                 goto nla_put_failure;
2281
2282         nla_nest_end(skb, nest_parms);
2283
2284         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2285                 goto nla_put_failure;
2286
2287         return 0;
2288
2289 nla_put_failure:
2290         return -ENOSPC;
2291 }
2292
2293 static int
2294 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2295 {
2296         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2297         unsigned long d = ct->status ^ status;
2298
2299         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2300                 /* SEEN_REPLY bit can only be set */
2301                 return -EBUSY;
2302
2303         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2304                 /* ASSURED bit can only be set */
2305                 return -EBUSY;
2306
2307         /* This check is less strict than ctnetlink_change_status()
2308          * because callers often flip IPS_EXPECTED bits when sending
2309          * an NFQA_CT attribute to the kernel.  So ignore the
2310          * unchangeable bits but do not error out. Also user programs
2311          * are allowed to clear the bits that they are allowed to change.
2312          */
2313         __ctnetlink_change_status(ct, status, ~status);
2314         return 0;
2315 }
2316
2317 static int
2318 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2319 {
2320         int err;
2321
2322         if (cda[CTA_TIMEOUT]) {
2323                 err = ctnetlink_change_timeout(ct, cda);
2324                 if (err < 0)
2325                         return err;
2326         }
2327         if (cda[CTA_STATUS]) {
2328                 err = ctnetlink_update_status(ct, cda);
2329                 if (err < 0)
2330                         return err;
2331         }
2332         if (cda[CTA_HELP]) {
2333                 err = ctnetlink_change_helper(ct, cda);
2334                 if (err < 0)
2335                         return err;
2336         }
2337         if (cda[CTA_LABELS]) {
2338                 err = ctnetlink_attach_labels(ct, cda);
2339                 if (err < 0)
2340                         return err;
2341         }
2342 #if defined(CONFIG_NF_CONNTRACK_MARK)
2343         if (cda[CTA_MARK]) {
2344                 u32 mask = 0, mark, newmark;
2345                 if (cda[CTA_MARK_MASK])
2346                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2347
2348                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2349                 newmark = (ct->mark & mask) ^ mark;
2350                 if (newmark != ct->mark)
2351                         ct->mark = newmark;
2352         }
2353 #endif
2354         return 0;
2355 }
2356
2357 static int
2358 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2359 {
2360         struct nlattr *cda[CTA_MAX+1];
2361         int ret;
2362
2363         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2364         if (ret < 0)
2365                 return ret;
2366
2367         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2368 }
2369
2370 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2371                                     const struct nf_conn *ct,
2372                                     struct nf_conntrack_tuple *tuple,
2373                                     struct nf_conntrack_tuple *mask)
2374 {
2375         int err;
2376
2377         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2378                                     nf_ct_l3num(ct), NULL);
2379         if (err < 0)
2380                 return err;
2381
2382         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2383                                      nf_ct_l3num(ct), NULL);
2384 }
2385
2386 static int
2387 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2388                              u32 portid, u32 report)
2389 {
2390         struct nlattr *cda[CTA_EXPECT_MAX+1];
2391         struct nf_conntrack_tuple tuple, mask;
2392         struct nf_conntrack_helper *helper = NULL;
2393         struct nf_conntrack_expect *exp;
2394         int err;
2395
2396         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2397                                NULL);
2398         if (err < 0)
2399                 return err;
2400
2401         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2402                                        ct, &tuple, &mask);
2403         if (err < 0)
2404                 return err;
2405
2406         if (cda[CTA_EXPECT_HELP_NAME]) {
2407                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2408
2409                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2410                                                     nf_ct_protonum(ct));
2411                 if (helper == NULL)
2412                         return -EOPNOTSUPP;
2413         }
2414
2415         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2416                                      helper, &tuple, &mask);
2417         if (IS_ERR(exp))
2418                 return PTR_ERR(exp);
2419
2420         err = nf_ct_expect_related_report(exp, portid, report);
2421         nf_ct_expect_put(exp);
2422         return err;
2423 }
2424
2425 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2426                                   enum ip_conntrack_info ctinfo, int diff)
2427 {
2428         if (!(ct->status & IPS_NAT_MASK))
2429                 return;
2430
2431         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2432 }
2433
2434 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2435         .get_ct         = ctnetlink_glue_get_ct,
2436         .build_size     = ctnetlink_glue_build_size,
2437         .build          = ctnetlink_glue_build,
2438         .parse          = ctnetlink_glue_parse,
2439         .attach_expect  = ctnetlink_glue_attach_expect,
2440         .seq_adjust     = ctnetlink_glue_seqadj,
2441 };
2442 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2443
2444 /***********************************************************************
2445  * EXPECT
2446  ***********************************************************************/
2447
2448 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2449                                     const struct nf_conntrack_tuple *tuple,
2450                                     enum ctattr_expect type)
2451 {
2452         struct nlattr *nest_parms;
2453
2454         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2455         if (!nest_parms)
2456                 goto nla_put_failure;
2457         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2458                 goto nla_put_failure;
2459         nla_nest_end(skb, nest_parms);
2460
2461         return 0;
2462
2463 nla_put_failure:
2464         return -1;
2465 }
2466
2467 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2468                                    const struct nf_conntrack_tuple *tuple,
2469                                    const struct nf_conntrack_tuple_mask *mask)
2470 {
2471         int ret;
2472         struct nf_conntrack_l3proto *l3proto;
2473         struct nf_conntrack_l4proto *l4proto;
2474         struct nf_conntrack_tuple m;
2475         struct nlattr *nest_parms;
2476
2477         memset(&m, 0xFF, sizeof(m));
2478         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2479         m.src.u.all = mask->src.u.all;
2480         m.dst.protonum = tuple->dst.protonum;
2481
2482         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2483         if (!nest_parms)
2484                 goto nla_put_failure;
2485
2486         rcu_read_lock();
2487         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2488         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2489         if (ret >= 0) {
2490                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2491                                                tuple->dst.protonum);
2492         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2493         }
2494         rcu_read_unlock();
2495
2496         if (unlikely(ret < 0))
2497                 goto nla_put_failure;
2498
2499         nla_nest_end(skb, nest_parms);
2500
2501         return 0;
2502
2503 nla_put_failure:
2504         return -1;
2505 }
2506
2507 static const union nf_inet_addr any_addr;
2508
2509 static int
2510 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2511                           const struct nf_conntrack_expect *exp)
2512 {
2513         struct nf_conn *master = exp->master;
2514         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2515         struct nf_conn_help *help;
2516 #ifdef CONFIG_NF_NAT_NEEDED
2517         struct nlattr *nest_parms;
2518         struct nf_conntrack_tuple nat_tuple = {};
2519 #endif
2520         struct nf_ct_helper_expectfn *expfn;
2521
2522         if (timeout < 0)
2523                 timeout = 0;
2524
2525         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2526                 goto nla_put_failure;
2527         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2528                 goto nla_put_failure;
2529         if (ctnetlink_exp_dump_tuple(skb,
2530                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2531                                  CTA_EXPECT_MASTER) < 0)
2532                 goto nla_put_failure;
2533
2534 #ifdef CONFIG_NF_NAT_NEEDED
2535         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2536             exp->saved_proto.all) {
2537                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2538                 if (!nest_parms)
2539                         goto nla_put_failure;
2540
2541                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2542                         goto nla_put_failure;
2543
2544                 nat_tuple.src.l3num = nf_ct_l3num(master);
2545                 nat_tuple.src.u3 = exp->saved_addr;
2546                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2547                 nat_tuple.src.u = exp->saved_proto;
2548
2549                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2550                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2551                         goto nla_put_failure;
2552                 nla_nest_end(skb, nest_parms);
2553         }
2554 #endif
2555         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2556             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2557             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2558             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2559                 goto nla_put_failure;
2560         help = nfct_help(master);
2561         if (help) {
2562                 struct nf_conntrack_helper *helper;
2563
2564                 helper = rcu_dereference(help->helper);
2565                 if (helper &&
2566                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2567                         goto nla_put_failure;
2568         }
2569         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2570         if (expfn != NULL &&
2571             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2572                 goto nla_put_failure;
2573
2574         return 0;
2575
2576 nla_put_failure:
2577         return -1;
2578 }
2579
2580 static int
2581 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2582                         int event, const struct nf_conntrack_expect *exp)
2583 {
2584         struct nlmsghdr *nlh;
2585         struct nfgenmsg *nfmsg;
2586         unsigned int flags = portid ? NLM_F_MULTI : 0;
2587
2588         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2589         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2590         if (nlh == NULL)
2591                 goto nlmsg_failure;
2592
2593         nfmsg = nlmsg_data(nlh);
2594         nfmsg->nfgen_family = exp->tuple.src.l3num;
2595         nfmsg->version      = NFNETLINK_V0;
2596         nfmsg->res_id       = 0;
2597
2598         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2599                 goto nla_put_failure;
2600
2601         nlmsg_end(skb, nlh);
2602         return skb->len;
2603
2604 nlmsg_failure:
2605 nla_put_failure:
2606         nlmsg_cancel(skb, nlh);
2607         return -1;
2608 }
2609
2610 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2611 static int
2612 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2613 {
2614         struct nf_conntrack_expect *exp = item->exp;
2615         struct net *net = nf_ct_exp_net(exp);
2616         struct nlmsghdr *nlh;
2617         struct nfgenmsg *nfmsg;
2618         struct sk_buff *skb;
2619         unsigned int type, group;
2620         int flags = 0;
2621
2622         if (events & (1 << IPEXP_DESTROY)) {
2623                 type = IPCTNL_MSG_EXP_DELETE;
2624                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2625         } else if (events & (1 << IPEXP_NEW)) {
2626                 type = IPCTNL_MSG_EXP_NEW;
2627                 flags = NLM_F_CREATE|NLM_F_EXCL;
2628                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2629         } else
2630                 return 0;
2631
2632         if (!item->report && !nfnetlink_has_listeners(net, group))
2633                 return 0;
2634
2635         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2636         if (skb == NULL)
2637                 goto errout;
2638
2639         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2640         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2641         if (nlh == NULL)
2642                 goto nlmsg_failure;
2643
2644         nfmsg = nlmsg_data(nlh);
2645         nfmsg->nfgen_family = exp->tuple.src.l3num;
2646         nfmsg->version      = NFNETLINK_V0;
2647         nfmsg->res_id       = 0;
2648
2649         rcu_read_lock();
2650         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2651                 goto nla_put_failure;
2652         rcu_read_unlock();
2653
2654         nlmsg_end(skb, nlh);
2655         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2656         return 0;
2657
2658 nla_put_failure:
2659         rcu_read_unlock();
2660         nlmsg_cancel(skb, nlh);
2661 nlmsg_failure:
2662         kfree_skb(skb);
2663 errout:
2664         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2665         return 0;
2666 }
2667 #endif
2668 static int ctnetlink_exp_done(struct netlink_callback *cb)
2669 {
2670         if (cb->args[1])
2671                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2672         return 0;
2673 }
2674
2675 static int
2676 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2677 {
2678         struct net *net = sock_net(skb->sk);
2679         struct nf_conntrack_expect *exp, *last;
2680         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2681         u_int8_t l3proto = nfmsg->nfgen_family;
2682
2683         rcu_read_lock();
2684         last = (struct nf_conntrack_expect *)cb->args[1];
2685         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2686 restart:
2687                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2688                                          hnode) {
2689                         if (l3proto && exp->tuple.src.l3num != l3proto)
2690                                 continue;
2691
2692                         if (!net_eq(nf_ct_net(exp->master), net))
2693                                 continue;
2694
2695                         if (cb->args[1]) {
2696                                 if (exp != last)
2697                                         continue;
2698                                 cb->args[1] = 0;
2699                         }
2700                         if (ctnetlink_exp_fill_info(skb,
2701                                                     NETLINK_CB(cb->skb).portid,
2702                                                     cb->nlh->nlmsg_seq,
2703                                                     IPCTNL_MSG_EXP_NEW,
2704                                                     exp) < 0) {
2705                                 if (!refcount_inc_not_zero(&exp->use))
2706                                         continue;
2707                                 cb->args[1] = (unsigned long)exp;
2708                                 goto out;
2709                         }
2710                 }
2711                 if (cb->args[1]) {
2712                         cb->args[1] = 0;
2713                         goto restart;
2714                 }
2715         }
2716 out:
2717         rcu_read_unlock();
2718         if (last)
2719                 nf_ct_expect_put(last);
2720
2721         return skb->len;
2722 }
2723
2724 static int
2725 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2726 {
2727         struct nf_conntrack_expect *exp, *last;
2728         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2729         struct nf_conn *ct = cb->data;
2730         struct nf_conn_help *help = nfct_help(ct);
2731         u_int8_t l3proto = nfmsg->nfgen_family;
2732
2733         if (cb->args[0])
2734                 return 0;
2735
2736         rcu_read_lock();
2737         last = (struct nf_conntrack_expect *)cb->args[1];
2738 restart:
2739         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2740                 if (l3proto && exp->tuple.src.l3num != l3proto)
2741                         continue;
2742                 if (cb->args[1]) {
2743                         if (exp != last)
2744                                 continue;
2745                         cb->args[1] = 0;
2746                 }
2747                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2748                                             cb->nlh->nlmsg_seq,
2749                                             IPCTNL_MSG_EXP_NEW,
2750                                             exp) < 0) {
2751                         if (!refcount_inc_not_zero(&exp->use))
2752                                 continue;
2753                         cb->args[1] = (unsigned long)exp;
2754                         goto out;
2755                 }
2756         }
2757         if (cb->args[1]) {
2758                 cb->args[1] = 0;
2759                 goto restart;
2760         }
2761         cb->args[0] = 1;
2762 out:
2763         rcu_read_unlock();
2764         if (last)
2765                 nf_ct_expect_put(last);
2766
2767         return skb->len;
2768 }
2769
2770 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2771                                  struct sk_buff *skb,
2772                                  const struct nlmsghdr *nlh,
2773                                  const struct nlattr * const cda[])
2774 {
2775         int err;
2776         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2777         u_int8_t u3 = nfmsg->nfgen_family;
2778         struct nf_conntrack_tuple tuple;
2779         struct nf_conntrack_tuple_hash *h;
2780         struct nf_conn *ct;
2781         struct nf_conntrack_zone zone;
2782         struct netlink_dump_control c = {
2783                 .dump = ctnetlink_exp_ct_dump_table,
2784                 .done = ctnetlink_exp_done,
2785         };
2786
2787         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2788                                     u3, NULL);
2789         if (err < 0)
2790                 return err;
2791
2792         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2793         if (err < 0)
2794                 return err;
2795
2796         h = nf_conntrack_find_get(net, &zone, &tuple);
2797         if (!h)
2798                 return -ENOENT;
2799
2800         ct = nf_ct_tuplehash_to_ctrack(h);
2801         /* No expectation linked to this connection tracking. */
2802         if (!nfct_help(ct)) {
2803                 nf_ct_put(ct);
2804                 return 0;
2805         }
2806
2807         c.data = ct;
2808
2809         err = netlink_dump_start(ctnl, skb, nlh, &c);
2810         nf_ct_put(ct);
2811
2812         return err;
2813 }
2814
2815 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
2816                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2817                                 const struct nlattr * const cda[])
2818 {
2819         struct nf_conntrack_tuple tuple;
2820         struct nf_conntrack_expect *exp;
2821         struct sk_buff *skb2;
2822         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2823         u_int8_t u3 = nfmsg->nfgen_family;
2824         struct nf_conntrack_zone zone;
2825         int err;
2826
2827         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2828                 if (cda[CTA_EXPECT_MASTER])
2829                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda);
2830                 else {
2831                         struct netlink_dump_control c = {
2832                                 .dump = ctnetlink_exp_dump_table,
2833                                 .done = ctnetlink_exp_done,
2834                         };
2835                         return netlink_dump_start(ctnl, skb, nlh, &c);
2836                 }
2837         }
2838
2839         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2840         if (err < 0)
2841                 return err;
2842
2843         if (cda[CTA_EXPECT_TUPLE])
2844                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2845                                             u3, NULL);
2846         else if (cda[CTA_EXPECT_MASTER])
2847                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2848                                             u3, NULL);
2849         else
2850                 return -EINVAL;
2851
2852         if (err < 0)
2853                 return err;
2854
2855         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2856         if (!exp)
2857                 return -ENOENT;
2858
2859         if (cda[CTA_EXPECT_ID]) {
2860                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2861                 if (ntohl(id) != (u32)(unsigned long)exp) {
2862                         nf_ct_expect_put(exp);
2863                         return -ENOENT;
2864                 }
2865         }
2866
2867         err = -ENOMEM;
2868         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2869         if (skb2 == NULL) {
2870                 nf_ct_expect_put(exp);
2871                 goto out;
2872         }
2873
2874         rcu_read_lock();
2875         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2876                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2877         rcu_read_unlock();
2878         nf_ct_expect_put(exp);
2879         if (err <= 0)
2880                 goto free;
2881
2882         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2883         if (err < 0)
2884                 goto out;
2885
2886         return 0;
2887
2888 free:
2889         kfree_skb(skb2);
2890 out:
2891         /* this avoids a loop in nfnetlink. */
2892         return err == -EAGAIN ? -ENOBUFS : err;
2893 }
2894
2895 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
2896                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2897                                 const struct nlattr * const cda[])
2898 {
2899         struct nf_conntrack_expect *exp;
2900         struct nf_conntrack_tuple tuple;
2901         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2902         struct hlist_node *next;
2903         u_int8_t u3 = nfmsg->nfgen_family;
2904         struct nf_conntrack_zone zone;
2905         unsigned int i;
2906         int err;
2907
2908         if (cda[CTA_EXPECT_TUPLE]) {
2909                 /* delete a single expect by tuple */
2910                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2911                 if (err < 0)
2912                         return err;
2913
2914                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2915                                             u3, NULL);
2916                 if (err < 0)
2917                         return err;
2918
2919                 /* bump usage count to 2 */
2920                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2921                 if (!exp)
2922                         return -ENOENT;
2923
2924                 if (cda[CTA_EXPECT_ID]) {
2925                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2926                         if (ntohl(id) != (u32)(unsigned long)exp) {
2927                                 nf_ct_expect_put(exp);
2928                                 return -ENOENT;
2929                         }
2930                 }
2931
2932                 /* after list removal, usage count == 1 */
2933                 spin_lock_bh(&nf_conntrack_expect_lock);
2934                 if (del_timer(&exp->timeout)) {
2935                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2936                                                    nlmsg_report(nlh));
2937                         nf_ct_expect_put(exp);
2938                 }
2939                 spin_unlock_bh(&nf_conntrack_expect_lock);
2940                 /* have to put what we 'get' above.
2941                  * after this line usage count == 0 */
2942                 nf_ct_expect_put(exp);
2943         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2944                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2945                 struct nf_conn_help *m_help;
2946
2947                 /* delete all expectations for this helper */
2948                 spin_lock_bh(&nf_conntrack_expect_lock);
2949                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2950                         hlist_for_each_entry_safe(exp, next,
2951                                                   &nf_ct_expect_hash[i],
2952                                                   hnode) {
2953
2954                                 if (!net_eq(nf_ct_exp_net(exp), net))
2955                                         continue;
2956
2957                                 m_help = nfct_help(exp->master);
2958                                 if (!strcmp(m_help->helper->name, name) &&
2959                                     del_timer(&exp->timeout)) {
2960                                         nf_ct_unlink_expect_report(exp,
2961                                                         NETLINK_CB(skb).portid,
2962                                                         nlmsg_report(nlh));
2963                                         nf_ct_expect_put(exp);
2964                                 }
2965                         }
2966                 }
2967                 spin_unlock_bh(&nf_conntrack_expect_lock);
2968         } else {
2969                 /* This basically means we have to flush everything*/
2970                 spin_lock_bh(&nf_conntrack_expect_lock);
2971                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2972                         hlist_for_each_entry_safe(exp, next,
2973                                                   &nf_ct_expect_hash[i],
2974                                                   hnode) {
2975
2976                                 if (!net_eq(nf_ct_exp_net(exp), net))
2977                                         continue;
2978
2979                                 if (del_timer(&exp->timeout)) {
2980                                         nf_ct_unlink_expect_report(exp,
2981                                                         NETLINK_CB(skb).portid,
2982                                                         nlmsg_report(nlh));
2983                                         nf_ct_expect_put(exp);
2984                                 }
2985                         }
2986                 }
2987                 spin_unlock_bh(&nf_conntrack_expect_lock);
2988         }
2989
2990         return 0;
2991 }
2992 static int
2993 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2994                         const struct nlattr * const cda[])
2995 {
2996         if (cda[CTA_EXPECT_TIMEOUT]) {
2997                 if (!del_timer(&x->timeout))
2998                         return -ETIME;
2999
3000                 x->timeout.expires = jiffies +
3001                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3002                 add_timer(&x->timeout);
3003         }
3004         return 0;
3005 }
3006
3007 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3008         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3009         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3010 };
3011
3012 static int
3013 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3014                            struct nf_conntrack_expect *exp,
3015                            u_int8_t u3)
3016 {
3017 #ifdef CONFIG_NF_NAT_NEEDED
3018         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3019         struct nf_conntrack_tuple nat_tuple = {};
3020         int err;
3021
3022         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3023                                exp_nat_nla_policy, NULL);
3024         if (err < 0)
3025                 return err;
3026
3027         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3028                 return -EINVAL;
3029
3030         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3031                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3032                                     u3, NULL);
3033         if (err < 0)
3034                 return err;
3035
3036         exp->saved_addr = nat_tuple.src.u3;
3037         exp->saved_proto = nat_tuple.src.u;
3038         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3039
3040         return 0;
3041 #else
3042         return -EOPNOTSUPP;
3043 #endif
3044 }
3045
3046 static struct nf_conntrack_expect *
3047 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3048                        struct nf_conntrack_helper *helper,
3049                        struct nf_conntrack_tuple *tuple,
3050                        struct nf_conntrack_tuple *mask)
3051 {
3052         u_int32_t class = 0;
3053         struct nf_conntrack_expect *exp;
3054         struct nf_conn_help *help;
3055         int err;
3056
3057         help = nfct_help(ct);
3058         if (!help)
3059                 return ERR_PTR(-EOPNOTSUPP);
3060
3061         if (cda[CTA_EXPECT_CLASS] && helper) {
3062                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3063                 if (class > helper->expect_class_max)
3064                         return ERR_PTR(-EINVAL);
3065         }
3066         exp = nf_ct_expect_alloc(ct);
3067         if (!exp)
3068                 return ERR_PTR(-ENOMEM);
3069
3070         if (cda[CTA_EXPECT_FLAGS]) {
3071                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3072                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3073         } else {
3074                 exp->flags = 0;
3075         }
3076         if (cda[CTA_EXPECT_FN]) {
3077                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3078                 struct nf_ct_helper_expectfn *expfn;
3079
3080                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3081                 if (expfn == NULL) {
3082                         err = -EINVAL;
3083                         goto err_out;
3084                 }
3085                 exp->expectfn = expfn->expectfn;
3086         } else
3087                 exp->expectfn = NULL;
3088
3089         exp->class = class;
3090         exp->master = ct;
3091         exp->helper = helper;
3092         exp->tuple = *tuple;
3093         exp->mask.src.u3 = mask->src.u3;
3094         exp->mask.src.u.all = mask->src.u.all;
3095
3096         if (cda[CTA_EXPECT_NAT]) {
3097                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3098                                                  exp, nf_ct_l3num(ct));
3099                 if (err < 0)
3100                         goto err_out;
3101         }
3102         return exp;
3103 err_out:
3104         nf_ct_expect_put(exp);
3105         return ERR_PTR(err);
3106 }
3107
3108 static int
3109 ctnetlink_create_expect(struct net *net,
3110                         const struct nf_conntrack_zone *zone,
3111                         const struct nlattr * const cda[],
3112                         u_int8_t u3, u32 portid, int report)
3113 {
3114         struct nf_conntrack_tuple tuple, mask, master_tuple;
3115         struct nf_conntrack_tuple_hash *h = NULL;
3116         struct nf_conntrack_helper *helper = NULL;
3117         struct nf_conntrack_expect *exp;
3118         struct nf_conn *ct;
3119         int err;
3120
3121         /* caller guarantees that those three CTA_EXPECT_* exist */
3122         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3123                                     u3, NULL);
3124         if (err < 0)
3125                 return err;
3126         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3127                                     u3, NULL);
3128         if (err < 0)
3129                 return err;
3130         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3131                                     u3, NULL);
3132         if (err < 0)
3133                 return err;
3134
3135         /* Look for master conntrack of this expectation */
3136         h = nf_conntrack_find_get(net, zone, &master_tuple);
3137         if (!h)
3138                 return -ENOENT;
3139         ct = nf_ct_tuplehash_to_ctrack(h);
3140
3141         rcu_read_lock();
3142         if (cda[CTA_EXPECT_HELP_NAME]) {
3143                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3144
3145                 helper = __nf_conntrack_helper_find(helpname, u3,
3146                                                     nf_ct_protonum(ct));
3147                 if (helper == NULL) {
3148                         rcu_read_unlock();
3149 #ifdef CONFIG_MODULES
3150                         if (request_module("nfct-helper-%s", helpname) < 0) {
3151                                 err = -EOPNOTSUPP;
3152                                 goto err_ct;
3153                         }
3154                         rcu_read_lock();
3155                         helper = __nf_conntrack_helper_find(helpname, u3,
3156                                                             nf_ct_protonum(ct));
3157                         if (helper) {
3158                                 err = -EAGAIN;
3159                                 goto err_rcu;
3160                         }
3161                         rcu_read_unlock();
3162 #endif
3163                         err = -EOPNOTSUPP;
3164                         goto err_ct;
3165                 }
3166         }
3167
3168         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3169         if (IS_ERR(exp)) {
3170                 err = PTR_ERR(exp);
3171                 goto err_rcu;
3172         }
3173
3174         err = nf_ct_expect_related_report(exp, portid, report);
3175         nf_ct_expect_put(exp);
3176 err_rcu:
3177         rcu_read_unlock();
3178 err_ct:
3179         nf_ct_put(ct);
3180         return err;
3181 }
3182
3183 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3184                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3185                                 const struct nlattr * const cda[])
3186 {
3187         struct nf_conntrack_tuple tuple;
3188         struct nf_conntrack_expect *exp;
3189         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3190         u_int8_t u3 = nfmsg->nfgen_family;
3191         struct nf_conntrack_zone zone;
3192         int err;
3193
3194         if (!cda[CTA_EXPECT_TUPLE]
3195             || !cda[CTA_EXPECT_MASK]
3196             || !cda[CTA_EXPECT_MASTER])
3197                 return -EINVAL;
3198
3199         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3200         if (err < 0)
3201                 return err;
3202
3203         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3204                                     u3, NULL);
3205         if (err < 0)
3206                 return err;
3207
3208         spin_lock_bh(&nf_conntrack_expect_lock);
3209         exp = __nf_ct_expect_find(net, &zone, &tuple);
3210         if (!exp) {
3211                 spin_unlock_bh(&nf_conntrack_expect_lock);
3212                 err = -ENOENT;
3213                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3214                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3215                                                       NETLINK_CB(skb).portid,
3216                                                       nlmsg_report(nlh));
3217                 }
3218                 return err;
3219         }
3220
3221         err = -EEXIST;
3222         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3223                 err = ctnetlink_change_expect(exp, cda);
3224         spin_unlock_bh(&nf_conntrack_expect_lock);
3225
3226         return err;
3227 }
3228
3229 static int
3230 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3231                              const struct ip_conntrack_stat *st)
3232 {
3233         struct nlmsghdr *nlh;
3234         struct nfgenmsg *nfmsg;
3235         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3236
3237         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3238                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3239         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3240         if (nlh == NULL)
3241                 goto nlmsg_failure;
3242
3243         nfmsg = nlmsg_data(nlh);
3244         nfmsg->nfgen_family = AF_UNSPEC;
3245         nfmsg->version      = NFNETLINK_V0;
3246         nfmsg->res_id       = htons(cpu);
3247
3248         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3249             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3250             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3251                 goto nla_put_failure;
3252
3253         nlmsg_end(skb, nlh);
3254         return skb->len;
3255
3256 nla_put_failure:
3257 nlmsg_failure:
3258         nlmsg_cancel(skb, nlh);
3259         return -1;
3260 }
3261
3262 static int
3263 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3264 {
3265         int cpu;
3266         struct net *net = sock_net(skb->sk);
3267
3268         if (cb->args[0] == nr_cpu_ids)
3269                 return 0;
3270
3271         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3272                 const struct ip_conntrack_stat *st;
3273
3274                 if (!cpu_possible(cpu))
3275                         continue;
3276
3277                 st = per_cpu_ptr(net->ct.stat, cpu);
3278                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3279                                                  cb->nlh->nlmsg_seq,
3280                                                  cpu, st) < 0)
3281                         break;
3282         }
3283         cb->args[0] = cpu;
3284
3285         return skb->len;
3286 }
3287
3288 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3289                                   struct sk_buff *skb,
3290                                   const struct nlmsghdr *nlh,
3291                                   const struct nlattr * const cda[])
3292 {
3293         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3294                 struct netlink_dump_control c = {
3295                         .dump = ctnetlink_exp_stat_cpu_dump,
3296                 };
3297                 return netlink_dump_start(ctnl, skb, nlh, &c);
3298         }
3299
3300         return 0;
3301 }
3302
3303 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3304 static struct nf_ct_event_notifier ctnl_notifier = {
3305         .fcn = ctnetlink_conntrack_event,
3306 };
3307
3308 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3309         .fcn = ctnetlink_expect_event,
3310 };
3311 #endif
3312
3313 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3314         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3315                                             .attr_count = CTA_MAX,
3316                                             .policy = ct_nla_policy },
3317         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3318                                             .attr_count = CTA_MAX,
3319                                             .policy = ct_nla_policy },
3320         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3321                                             .attr_count = CTA_MAX,
3322                                             .policy = ct_nla_policy },
3323         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3324                                             .attr_count = CTA_MAX,
3325                                             .policy = ct_nla_policy },
3326         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3327         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3328         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3329         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3330 };
3331
3332 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3333         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3334                                             .attr_count = CTA_EXPECT_MAX,
3335                                             .policy = exp_nla_policy },
3336         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3337                                             .attr_count = CTA_EXPECT_MAX,
3338                                             .policy = exp_nla_policy },
3339         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3340                                             .attr_count = CTA_EXPECT_MAX,
3341                                             .policy = exp_nla_policy },
3342         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3343 };
3344
3345 static const struct nfnetlink_subsystem ctnl_subsys = {
3346         .name                           = "conntrack",
3347         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3348         .cb_count                       = IPCTNL_MSG_MAX,
3349         .cb                             = ctnl_cb,
3350 };
3351
3352 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3353         .name                           = "conntrack_expect",
3354         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3355         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3356         .cb                             = ctnl_exp_cb,
3357 };
3358
3359 MODULE_ALIAS("ip_conntrack_netlink");
3360 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3361 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3362
3363 static int __net_init ctnetlink_net_init(struct net *net)
3364 {
3365 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3366         int ret;
3367
3368         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3369         if (ret < 0) {
3370                 pr_err("ctnetlink_init: cannot register notifier.\n");
3371                 goto err_out;
3372         }
3373
3374         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3375         if (ret < 0) {
3376                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3377                 goto err_unreg_notifier;
3378         }
3379 #endif
3380         return 0;
3381
3382 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3383 err_unreg_notifier:
3384         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3385 err_out:
3386         return ret;
3387 #endif
3388 }
3389
3390 static void ctnetlink_net_exit(struct net *net)
3391 {
3392 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3393         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3394         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3395 #endif
3396 }
3397
3398 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3399 {
3400         struct net *net;
3401
3402         list_for_each_entry(net, net_exit_list, exit_list)
3403                 ctnetlink_net_exit(net);
3404 }
3405
3406 static struct pernet_operations ctnetlink_net_ops = {
3407         .init           = ctnetlink_net_init,
3408         .exit_batch     = ctnetlink_net_exit_batch,
3409 };
3410
3411 static int __init ctnetlink_init(void)
3412 {
3413         int ret;
3414
3415         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3416         ret = nfnetlink_subsys_register(&ctnl_subsys);
3417         if (ret < 0) {
3418                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3419                 goto err_out;
3420         }
3421
3422         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3423         if (ret < 0) {
3424                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3425                 goto err_unreg_subsys;
3426         }
3427
3428         ret = register_pernet_subsys(&ctnetlink_net_ops);
3429         if (ret < 0) {
3430                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3431                 goto err_unreg_exp_subsys;
3432         }
3433 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3434         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3435         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3436 #endif
3437         return 0;
3438
3439 err_unreg_exp_subsys:
3440         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3441 err_unreg_subsys:
3442         nfnetlink_subsys_unregister(&ctnl_subsys);
3443 err_out:
3444         return ret;
3445 }
3446
3447 static void __exit ctnetlink_exit(void)
3448 {
3449         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3450
3451         unregister_pernet_subsys(&ctnetlink_net_ops);
3452         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3453         nfnetlink_subsys_unregister(&ctnl_subsys);
3454 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3455         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3456 #endif
3457         synchronize_rcu();
3458 }
3459
3460 module_init(ctnetlink_init);
3461 module_exit(ctnetlink_exit);