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