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 #include <net/netfilter/nf_conntrack_seqadj.h>
  49 #include <net/netfilter/nf_conntrack_synproxy.h>
  50 #ifdef CONFIG_NF_NAT_NEEDED
  51 #include <net/netfilter/nf_nat_core.h>
  52 #include <net/netfilter/nf_nat_l4proto.h>
  53 #include <net/netfilter/nf_nat_helper.h>
  54 #endif
  55
  56 #include <linux/netfilter/nfnetlink.h>
  57 #include <linux/netfilter/nfnetlink_conntrack.h>
  58
  59 MODULE_LICENSE("GPL");
  60
  61 static char __initdata version[] = "0.93";
  62
  63 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  64                                        const struct nf_conntrack_tuple *tuple,
  65                                        struct nf_conntrack_l4proto *l4proto)
  66 {
  67         int ret = 0;
  68         struct nlattr *nest_parms;
  69
  70         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  71         if (!nest_parms)
  72                 goto nla_put_failure;
  73         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  74                 goto nla_put_failure;
  75
  76         if (likely(l4proto->tuple_to_nlattr))
  77                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  78
  79         nla_nest_end(skb, nest_parms);
  80
  81         return ret;
  82
  83 nla_put_failure:
  84         return -1;
  85 }
  86
  87 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  88                                     const struct nf_conntrack_tuple *tuple,
  89                                     struct nf_conntrack_l3proto *l3proto)
  90 {
  91         int ret = 0;
  92         struct nlattr *nest_parms;
  93
  94         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  95         if (!nest_parms)
  96                 goto nla_put_failure;
  97
  98         if (likely(l3proto->tuple_to_nlattr))
  99                 ret = l3proto->tuple_to_nlattr(skb, tuple);
 100
 101         nla_nest_end(skb, nest_parms);
 102
 103         return ret;
 104
 105 nla_put_failure:
 106         return -1;
 107 }
 108
 109 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 110                                  const struct nf_conntrack_tuple *tuple)
 111 {
 112         int ret;
 113         struct nf_conntrack_l3proto *l3proto;
 114         struct nf_conntrack_l4proto *l4proto;
 115
 116         rcu_read_lock();
 117         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 118         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 119
 120         if (ret >= 0) {
 121                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 122                                                tuple->dst.protonum);
 123                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 124         }
 125         rcu_read_unlock();
 126         return ret;
 127 }
 128
 129 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 130                                   const struct nf_conntrack_zone *zone, int dir)
 131 {
 132         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 133                 return 0;
 134         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 135                 goto nla_put_failure;
 136         return 0;
 137
 138 nla_put_failure:
 139         return -1;
 140 }
 141
 142 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 143 {
 144         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 145                 goto nla_put_failure;
 146         return 0;
 147
 148 nla_put_failure:
 149         return -1;
 150 }
 151
 152 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 153 {
 154         long timeout = nf_ct_expires(ct) / HZ;
 155
 156         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 157                 goto nla_put_failure;
 158         return 0;
 159
 160 nla_put_failure:
 161         return -1;
 162 }
 163
 164 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 165 {
 166         struct nf_conntrack_l4proto *l4proto;
 167         struct nlattr *nest_proto;
 168         int ret;
 169
 170         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 171         if (!l4proto->to_nlattr)
 172                 return 0;
 173
 174         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 175         if (!nest_proto)
 176                 goto nla_put_failure;
 177
 178         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 179
 180         nla_nest_end(skb, nest_proto);
 181
 182         return ret;
 183
 184 nla_put_failure:
 185         return -1;
 186 }
 187
 188 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 189                                    const struct nf_conn *ct)
 190 {
 191         struct nlattr *nest_helper;
 192         const struct nf_conn_help *help = nfct_help(ct);
 193         struct nf_conntrack_helper *helper;
 194
 195         if (!help)
 196                 return 0;
 197
 198         helper = rcu_dereference(help->helper);
 199         if (!helper)
 200                 goto out;
 201
 202         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 203         if (!nest_helper)
 204                 goto nla_put_failure;
 205         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 206                 goto nla_put_failure;
 207
 208         if (helper->to_nlattr)
 209                 helper->to_nlattr(skb, ct);
 210
 211         nla_nest_end(skb, nest_helper);
 212 out:
 213         return 0;
 214
 215 nla_put_failure:
 216         return -1;
 217 }
 218
 219 static int
 220 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 221               enum ip_conntrack_dir dir, int type)
 222 {
 223         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 224         struct nf_conn_counter *counter = acct->counter;
 225         struct nlattr *nest_count;
 226         u64 pkts, bytes;
 227
 228         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 229                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 230                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 231         } else {
 232                 pkts = atomic64_read(&counter[dir].packets);
 233                 bytes = atomic64_read(&counter[dir].bytes);
 234         }
 235
 236         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 237         if (!nest_count)
 238                 goto nla_put_failure;
 239
 240         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 241                          CTA_COUNTERS_PAD) ||
 242             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 243                          CTA_COUNTERS_PAD))
 244                 goto nla_put_failure;
 245
 246         nla_nest_end(skb, nest_count);
 247
 248         return 0;
 249
 250 nla_put_failure:
 251         return -1;
 252 }
 253
 254 static int
 255 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 256 {
 257         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 258
 259         if (!acct)
 260                 return 0;
 261
 262         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 263                 return -1;
 264         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 265                 return -1;
 266
 267         return 0;
 268 }
 269
 270 static int
 271 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 272 {
 273         struct nlattr *nest_count;
 274         const struct nf_conn_tstamp *tstamp;
 275
 276         tstamp = nf_conn_tstamp_find(ct);
 277         if (!tstamp)
 278                 return 0;
 279
 280         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 281         if (!nest_count)
 282                 goto nla_put_failure;
 283
 284         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 285                          CTA_TIMESTAMP_PAD) ||
 286             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 287                                                cpu_to_be64(tstamp->stop),
 288                                                CTA_TIMESTAMP_PAD)))
 289                 goto nla_put_failure;
 290         nla_nest_end(skb, nest_count);
 291
 292         return 0;
 293
 294 nla_put_failure:
 295         return -1;
 296 }
 297
 298 #ifdef CONFIG_NF_CONNTRACK_MARK
 299 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 300 {
 301         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 302                 goto nla_put_failure;
 303         return 0;
 304
 305 nla_put_failure:
 306         return -1;
 307 }
 308 #else
 309 #define ctnetlink_dump_mark(a, b) (0)
 310 #endif
 311
 312 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 313 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 314 {
 315         struct nlattr *nest_secctx;
 316         int len, ret;
 317         char *secctx;
 318
 319         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 320         if (ret)
 321                 return 0;
 322
 323         ret = -1;
 324         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 325         if (!nest_secctx)
 326                 goto nla_put_failure;
 327
 328         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 329                 goto nla_put_failure;
 330         nla_nest_end(skb, nest_secctx);
 331
 332         ret = 0;
 333 nla_put_failure:
 334         security_release_secctx(secctx, len);
 335         return ret;
 336 }
 337 #else
 338 #define ctnetlink_dump_secctx(a, b) (0)
 339 #endif
 340
 341 #ifdef CONFIG_NF_CONNTRACK_LABELS
 342 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 343 {
 344         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 345
 346         if (!labels)
 347                 return 0;
 348         return nla_total_size(sizeof(labels->bits));
 349 }
 350
 351 static int
 352 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 353 {
 354         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 355         unsigned int i;
 356
 357         if (!labels)
 358                 return 0;
 359
 360         i = 0;
 361         do {
 362                 if (labels->bits[i] != 0)
 363                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 364                                        labels->bits);
 365                 i++;
 366         } while (i < ARRAY_SIZE(labels->bits));
 367
 368         return 0;
 369 }
 370 #else
 371 #define ctnetlink_dump_labels(a, b) (0)
 372 #define ctnetlink_label_size(a) (0)
 373 #endif
 374
 375 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 376
 377 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 378 {
 379         struct nlattr *nest_parms;
 380
 381         if (!(ct->status & IPS_EXPECTED))
 382                 return 0;
 383
 384         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 385         if (!nest_parms)
 386                 goto nla_put_failure;
 387         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 388                 goto nla_put_failure;
 389         nla_nest_end(skb, nest_parms);
 390
 391         return 0;
 392
 393 nla_put_failure:
 394         return -1;
 395 }
 396
 397 static int
 398 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 399 {
 400         struct nlattr *nest_parms;
 401
 402         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 403         if (!nest_parms)
 404                 goto nla_put_failure;
 405
 406         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 407                          htonl(seq->correction_pos)) ||
 408             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 409                          htonl(seq->offset_before)) ||
 410             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 411                          htonl(seq->offset_after)))
 412                 goto nla_put_failure;
 413
 414         nla_nest_end(skb, nest_parms);
 415
 416         return 0;
 417
 418 nla_put_failure:
 419         return -1;
 420 }
 421
 422 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 423 {
 424         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 425         struct nf_ct_seqadj *seq;
 426
 427         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 428                 return 0;
 429
 430         spin_lock_bh(&ct->lock);
 431         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 432         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 433                 goto err;
 434
 435         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 436         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 437                 goto err;
 438
 439         spin_unlock_bh(&ct->lock);
 440         return 0;
 441 err:
 442         spin_unlock_bh(&ct->lock);
 443         return -1;
 444 }
 445
 446 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 447 {
 448         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 449                 goto nla_put_failure;
 450         return 0;
 451
 452 nla_put_failure:
 453         return -1;
 454 }
 455
 456 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 457 {
 458         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 459                 goto nla_put_failure;
 460         return 0;
 461
 462 nla_put_failure:
 463         return -1;
 464 }
 465
 466 static int
 467 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 468                     struct nf_conn *ct)
 469 {
 470         const struct nf_conntrack_zone *zone;
 471         struct nlmsghdr *nlh;
 472         struct nfgenmsg *nfmsg;
 473         struct nlattr *nest_parms;
 474         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 475
 476         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 477         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 478         if (nlh == NULL)
 479                 goto nlmsg_failure;
 480
 481         nfmsg = nlmsg_data(nlh);
 482         nfmsg->nfgen_family = nf_ct_l3num(ct);
 483         nfmsg->version      = NFNETLINK_V0;
 484         nfmsg->res_id       = 0;
 485
 486         zone = nf_ct_zone(ct);
 487
 488         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 489         if (!nest_parms)
 490                 goto nla_put_failure;
 491         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 492                 goto nla_put_failure;
 493         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 494                                    NF_CT_ZONE_DIR_ORIG) < 0)
 495                 goto nla_put_failure;
 496         nla_nest_end(skb, nest_parms);
 497
 498         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 499         if (!nest_parms)
 500                 goto nla_put_failure;
 501         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 502                 goto nla_put_failure;
 503         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 504                                    NF_CT_ZONE_DIR_REPL) < 0)
 505                 goto nla_put_failure;
 506         nla_nest_end(skb, nest_parms);
 507
 508         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 509                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 510                 goto nla_put_failure;
 511
 512         if (ctnetlink_dump_status(skb, ct) < 0 ||
 513             ctnetlink_dump_timeout(skb, ct) < 0 ||
 514             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 515             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 516             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 517             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 518             ctnetlink_dump_mark(skb, ct) < 0 ||
 519             ctnetlink_dump_secctx(skb, ct) < 0 ||
 520             ctnetlink_dump_labels(skb, ct) < 0 ||
 521             ctnetlink_dump_id(skb, ct) < 0 ||
 522             ctnetlink_dump_use(skb, ct) < 0 ||
 523             ctnetlink_dump_master(skb, ct) < 0 ||
 524             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 525                 goto nla_put_failure;
 526
 527         nlmsg_end(skb, nlh);
 528         return skb->len;
 529
 530 nlmsg_failure:
 531 nla_put_failure:
 532         nlmsg_cancel(skb, nlh);
 533         return -1;
 534 }
 535
 536 static inline size_t ctnetlink_proto_size(const struct nf_conn *ct)
 537 {
 538         struct nf_conntrack_l3proto *l3proto;
 539         struct nf_conntrack_l4proto *l4proto;
 540         size_t len = 0;
 541
 542         rcu_read_lock();
 543         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 544         len += l3proto->nla_size;
 545
 546         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 547         len += l4proto->nla_size;
 548         rcu_read_unlock();
 549
 550         return len;
 551 }
 552
 553 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 554 {
 555         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 556                 return 0;
 557         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 558                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 559                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 560                ;
 561 }
 562
 563 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 564 {
 565 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 566         int len, ret;
 567
 568         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 569         if (ret)
 570                 return 0;
 571
 572         return nla_total_size(0) /* CTA_SECCTX */
 573                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 574 #else
 575         return 0;
 576 #endif
 577 }
 578
 579 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 580 {
 581 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 582         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 583                 return 0;
 584         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 585 #else
 586         return 0;
 587 #endif
 588 }
 589
 590 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 591 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 592 {
 593         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 594                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 595                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 596                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 597                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 598                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 599                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 600                + ctnetlink_acct_size(ct)
 601                + ctnetlink_timestamp_size(ct)
 602                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 603                + nla_total_size(0) /* CTA_PROTOINFO */
 604                + nla_total_size(0) /* CTA_HELP */
 605                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 606                + ctnetlink_secctx_size(ct)
 607 #ifdef CONFIG_NF_NAT_NEEDED
 608                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 609                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 610 #endif
 611 #ifdef CONFIG_NF_CONNTRACK_MARK
 612                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 613 #endif
 614 #ifdef CONFIG_NF_CONNTRACK_ZONES
 615                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 616 #endif
 617                + ctnetlink_proto_size(ct)
 618                + ctnetlink_label_size(ct)
 619                ;
 620 }
 621
 622 static int
 623 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 624 {
 625         const struct nf_conntrack_zone *zone;
 626         struct net *net;
 627         struct nlmsghdr *nlh;
 628         struct nfgenmsg *nfmsg;
 629         struct nlattr *nest_parms;
 630         struct nf_conn *ct = item->ct;
 631         struct sk_buff *skb;
 632         unsigned int type;
 633         unsigned int flags = 0, group;
 634         int err;
 635
 636         if (events & (1 << IPCT_DESTROY)) {
 637                 type = IPCTNL_MSG_CT_DELETE;
 638                 group = NFNLGRP_CONNTRACK_DESTROY;
 639         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 640                 type = IPCTNL_MSG_CT_NEW;
 641                 flags = NLM_F_CREATE|NLM_F_EXCL;
 642                 group = NFNLGRP_CONNTRACK_NEW;
 643         } else if (events) {
 644                 type = IPCTNL_MSG_CT_NEW;
 645                 group = NFNLGRP_CONNTRACK_UPDATE;
 646         } else
 647                 return 0;
 648
 649         net = nf_ct_net(ct);
 650         if (!item->report && !nfnetlink_has_listeners(net, group))
 651                 return 0;
 652
 653         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 654         if (skb == NULL)
 655                 goto errout;
 656
 657         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 658         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 659         if (nlh == NULL)
 660                 goto nlmsg_failure;
 661
 662         nfmsg = nlmsg_data(nlh);
 663         nfmsg->nfgen_family = nf_ct_l3num(ct);
 664         nfmsg->version  = NFNETLINK_V0;
 665         nfmsg->res_id   = 0;
 666
 667         rcu_read_lock();
 668         zone = nf_ct_zone(ct);
 669
 670         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 671         if (!nest_parms)
 672                 goto nla_put_failure;
 673         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 674                 goto nla_put_failure;
 675         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 676                                    NF_CT_ZONE_DIR_ORIG) < 0)
 677                 goto nla_put_failure;
 678         nla_nest_end(skb, nest_parms);
 679
 680         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 681         if (!nest_parms)
 682                 goto nla_put_failure;
 683         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 684                 goto nla_put_failure;
 685         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 686                                    NF_CT_ZONE_DIR_REPL) < 0)
 687                 goto nla_put_failure;
 688         nla_nest_end(skb, nest_parms);
 689
 690         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 691                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 692                 goto nla_put_failure;
 693
 694         if (ctnetlink_dump_id(skb, ct) < 0)
 695                 goto nla_put_failure;
 696
 697         if (ctnetlink_dump_status(skb, ct) < 0)
 698                 goto nla_put_failure;
 699
 700         if (events & (1 << IPCT_DESTROY)) {
 701                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 702                     ctnetlink_dump_timestamp(skb, ct) < 0)
 703                         goto nla_put_failure;
 704         } else {
 705                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 706                         goto nla_put_failure;
 707
 708                 if (events & (1 << IPCT_PROTOINFO)
 709                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 710                         goto nla_put_failure;
 711
 712                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 713                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 714                         goto nla_put_failure;
 715
 716 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 717                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 718                     && ctnetlink_dump_secctx(skb, ct) < 0)
 719                         goto nla_put_failure;
 720 #endif
 721                 if (events & (1 << IPCT_LABEL) &&
 722                      ctnetlink_dump_labels(skb, ct) < 0)
 723                         goto nla_put_failure;
 724
 725                 if (events & (1 << IPCT_RELATED) &&
 726                     ctnetlink_dump_master(skb, ct) < 0)
 727                         goto nla_put_failure;
 728
 729                 if (events & (1 << IPCT_SEQADJ) &&
 730                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 731                         goto nla_put_failure;
 732         }
 733
 734 #ifdef CONFIG_NF_CONNTRACK_MARK
 735         if ((events & (1 << IPCT_MARK) || ct->mark)
 736             && ctnetlink_dump_mark(skb, ct) < 0)
 737                 goto nla_put_failure;
 738 #endif
 739         rcu_read_unlock();
 740
 741         nlmsg_end(skb, nlh);
 742         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 743                              GFP_ATOMIC);
 744         if (err == -ENOBUFS || err == -EAGAIN)
 745                 return -ENOBUFS;
 746
 747         return 0;
 748
 749 nla_put_failure:
 750         rcu_read_unlock();
 751         nlmsg_cancel(skb, nlh);
 752 nlmsg_failure:
 753         kfree_skb(skb);
 754 errout:
 755         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 756                 return -ENOBUFS;
 757
 758         return 0;
 759 }
 760 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 761
 762 static int ctnetlink_done(struct netlink_callback *cb)
 763 {
 764         if (cb->args[1])
 765                 nf_ct_put((struct nf_conn *)cb->args[1]);
 766         kfree(cb->data);
 767         return 0;
 768 }
 769
 770 struct ctnetlink_filter {
 771         struct {
 772                 u_int32_t val;
 773                 u_int32_t mask;
 774         } mark;
 775 };
 776
 777 static struct ctnetlink_filter *
 778 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 779 {
 780 #ifdef CONFIG_NF_CONNTRACK_MARK
 781         struct ctnetlink_filter *filter;
 782
 783         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 784         if (filter == NULL)
 785                 return ERR_PTR(-ENOMEM);
 786
 787         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 788         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 789
 790         return filter;
 791 #else
 792         return ERR_PTR(-EOPNOTSUPP);
 793 #endif
 794 }
 795
 796 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 797 {
 798         struct ctnetlink_filter *filter = data;
 799
 800         if (filter == NULL)
 801                 return 1;
 802
 803 #ifdef CONFIG_NF_CONNTRACK_MARK
 804         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 805                 return 1;
 806 #endif
 807
 808         return 0;
 809 }
 810
 811 static int
 812 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 813 {
 814         struct net *net = sock_net(skb->sk);
 815         struct nf_conn *ct, *last;
 816         struct nf_conntrack_tuple_hash *h;
 817         struct hlist_nulls_node *n;
 818         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 819         u_int8_t l3proto = nfmsg->nfgen_family;
 820         struct nf_conn *nf_ct_evict[8];
 821         int res, i;
 822         spinlock_t *lockp;
 823
 824         last = (struct nf_conn *)cb->args[1];
 825         i = 0;
 826
 827         local_bh_disable();
 828         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 829 restart:
 830                 while (i) {
 831                         i--;
 832                         if (nf_ct_should_gc(nf_ct_evict[i]))
 833                                 nf_ct_kill(nf_ct_evict[i]);
 834                         nf_ct_put(nf_ct_evict[i]);
 835                 }
 836
 837                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 838                 nf_conntrack_lock(lockp);
 839                 if (cb->args[0] >= nf_conntrack_htable_size) {
 840                         spin_unlock(lockp);
 841                         goto out;
 842                 }
 843                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 844                                            hnnode) {
 845                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 846                                 continue;
 847                         ct = nf_ct_tuplehash_to_ctrack(h);
 848                         if (nf_ct_is_expired(ct)) {
 849                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 850                                     atomic_inc_not_zero(&ct->ct_general.use))
 851                                         nf_ct_evict[i++] = ct;
 852                                 continue;
 853                         }
 854
 855                         if (!net_eq(net, nf_ct_net(ct)))
 856                                 continue;
 857
 858                         /* Dump entries of a given L3 protocol number.
 859                          * If it is not specified, ie. l3proto == 0,
 860                          * then dump everything. */
 861                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 862                                 continue;
 863                         if (cb->args[1]) {
 864                                 if (ct != last)
 865                                         continue;
 866                                 cb->args[1] = 0;
 867                         }
 868                         if (!ctnetlink_filter_match(ct, cb->data))
 869                                 continue;
 870
 871                         rcu_read_lock();
 872                         res =
 873                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 874                                             cb->nlh->nlmsg_seq,
 875                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 876                                             ct);
 877                         rcu_read_unlock();
 878                         if (res < 0) {
 879                                 nf_conntrack_get(&ct->ct_general);
 880                                 cb->args[1] = (unsigned long)ct;
 881                                 spin_unlock(lockp);
 882                                 goto out;
 883                         }
 884                 }
 885                 spin_unlock(lockp);
 886                 if (cb->args[1]) {
 887                         cb->args[1] = 0;
 888                         goto restart;
 889                 }
 890         }
 891 out:
 892         local_bh_enable();
 893         if (last) {
 894                 /* nf ct hash resize happened, now clear the leftover. */
 895                 if ((struct nf_conn *)cb->args[1] == last)
 896                         cb->args[1] = 0;
 897
 898                 nf_ct_put(last);
 899         }
 900
 901         while (i) {
 902                 i--;
 903                 if (nf_ct_should_gc(nf_ct_evict[i]))
 904                         nf_ct_kill(nf_ct_evict[i]);
 905                 nf_ct_put(nf_ct_evict[i]);
 906         }
 907
 908         return skb->len;
 909 }
 910
 911 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
 912                                     struct nf_conntrack_tuple *tuple)
 913 {
 914         struct nlattr *tb[CTA_IP_MAX+1];
 915         struct nf_conntrack_l3proto *l3proto;
 916         int ret = 0;
 917
 918         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL, NULL);
 919         if (ret < 0)
 920                 return ret;
 921
 922         rcu_read_lock();
 923         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 924
 925         if (likely(l3proto->nlattr_to_tuple)) {
 926                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 927                                           l3proto->nla_policy, NULL);
 928                 if (ret == 0)
 929                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 930         }
 931
 932         rcu_read_unlock();
 933
 934         return ret;
 935 }
 936
 937 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 938         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 939 };
 940
 941 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
 942                                        struct nf_conntrack_tuple *tuple)
 943 {
 944         struct nlattr *tb[CTA_PROTO_MAX+1];
 945         struct nf_conntrack_l4proto *l4proto;
 946         int ret = 0;
 947
 948         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy,
 949                                NULL);
 950         if (ret < 0)
 951                 return ret;
 952
 953         if (!tb[CTA_PROTO_NUM])
 954                 return -EINVAL;
 955         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 956
 957         rcu_read_lock();
 958         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 959
 960         if (likely(l4proto->nlattr_to_tuple)) {
 961                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 962                                           l4proto->nla_policy, NULL);
 963                 if (ret == 0)
 964                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 965         }
 966
 967         rcu_read_unlock();
 968
 969         return ret;
 970 }
 971
 972 static int
 973 ctnetlink_parse_zone(const struct nlattr *attr,
 974                      struct nf_conntrack_zone *zone)
 975 {
 976         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
 977                         NF_CT_DEFAULT_ZONE_DIR, 0);
 978 #ifdef CONFIG_NF_CONNTRACK_ZONES
 979         if (attr)
 980                 zone->id = ntohs(nla_get_be16(attr));
 981 #else
 982         if (attr)
 983                 return -EOPNOTSUPP;
 984 #endif
 985         return 0;
 986 }
 987
 988 static int
 989 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
 990                            struct nf_conntrack_zone *zone)
 991 {
 992         int ret;
 993
 994         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
 995                 return -EINVAL;
 996
 997         ret = ctnetlink_parse_zone(attr, zone);
 998         if (ret < 0)
 999                 return ret;
1000
1001         if (type == CTA_TUPLE_REPLY)
1002                 zone->dir = NF_CT_ZONE_DIR_REPL;
1003         else
1004                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1005
1006         return 0;
1007 }
1008
1009 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1010         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1011         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1012         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1013 };
1014
1015 static int
1016 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1017                       struct nf_conntrack_tuple *tuple, u32 type,
1018                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1019 {
1020         struct nlattr *tb[CTA_TUPLE_MAX+1];
1021         int err;
1022
1023         memset(tuple, 0, sizeof(*tuple));
1024
1025         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1026                                NULL);
1027         if (err < 0)
1028                 return err;
1029
1030         if (!tb[CTA_TUPLE_IP])
1031                 return -EINVAL;
1032
1033         tuple->src.l3num = l3num;
1034
1035         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1036         if (err < 0)
1037                 return err;
1038
1039         if (!tb[CTA_TUPLE_PROTO])
1040                 return -EINVAL;
1041
1042         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1043         if (err < 0)
1044                 return err;
1045
1046         if (tb[CTA_TUPLE_ZONE]) {
1047                 if (!zone)
1048                         return -EINVAL;
1049
1050                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1051                                                  type, zone);
1052                 if (err < 0)
1053                         return err;
1054         }
1055
1056         /* orig and expect tuples get DIR_ORIGINAL */
1057         if (type == CTA_TUPLE_REPLY)
1058                 tuple->dst.dir = IP_CT_DIR_REPLY;
1059         else
1060                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1061
1062         return 0;
1063 }
1064
1065 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1066         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1067                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1068 };
1069
1070 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1071                                 struct nlattr **helpinfo)
1072 {
1073         int err;
1074         struct nlattr *tb[CTA_HELP_MAX+1];
1075
1076         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1077         if (err < 0)
1078                 return err;
1079
1080         if (!tb[CTA_HELP_NAME])
1081                 return -EINVAL;
1082
1083         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1084
1085         if (tb[CTA_HELP_INFO])
1086                 *helpinfo = tb[CTA_HELP_INFO];
1087
1088         return 0;
1089 }
1090
1091 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1092         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1093         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1094         [CTA_STATUS]            = { .type = NLA_U32 },
1095         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1096         [CTA_HELP]              = { .type = NLA_NESTED },
1097         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1098         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1099         [CTA_MARK]              = { .type = NLA_U32 },
1100         [CTA_ID]                = { .type = NLA_U32 },
1101         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1102         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1103         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1104         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1105         [CTA_ZONE]              = { .type = NLA_U16 },
1106         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1107         [CTA_LABELS]            = { .type = NLA_BINARY,
1108                                     .len = NF_CT_LABELS_MAX_SIZE },
1109         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1110                                     .len = NF_CT_LABELS_MAX_SIZE },
1111 };
1112
1113 static int ctnetlink_flush_conntrack(struct net *net,
1114                                      const struct nlattr * const cda[],
1115                                      u32 portid, int report)
1116 {
1117         struct ctnetlink_filter *filter = NULL;
1118
1119         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1120                 filter = ctnetlink_alloc_filter(cda);
1121                 if (IS_ERR(filter))
1122                         return PTR_ERR(filter);
1123         }
1124
1125         nf_ct_iterate_cleanup_net(net, ctnetlink_filter_match, filter,
1126                                   portid, report);
1127         kfree(filter);
1128
1129         return 0;
1130 }
1131
1132 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1133                                    struct sk_buff *skb,
1134                                    const struct nlmsghdr *nlh,
1135                                    const struct nlattr * const cda[],
1136                                    struct netlink_ext_ack *extack)
1137 {
1138         struct nf_conntrack_tuple_hash *h;
1139         struct nf_conntrack_tuple tuple;
1140         struct nf_conn *ct;
1141         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1142         u_int8_t u3 = nfmsg->nfgen_family;
1143         struct nf_conntrack_zone zone;
1144         int err;
1145
1146         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1147         if (err < 0)
1148                 return err;
1149
1150         if (cda[CTA_TUPLE_ORIG])
1151                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1152                                             u3, &zone);
1153         else if (cda[CTA_TUPLE_REPLY])
1154                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1155                                             u3, &zone);
1156         else {
1157                 return ctnetlink_flush_conntrack(net, cda,
1158                                                  NETLINK_CB(skb).portid,
1159                                                  nlmsg_report(nlh));
1160         }
1161
1162         if (err < 0)
1163                 return err;
1164
1165         h = nf_conntrack_find_get(net, &zone, &tuple);
1166         if (!h)
1167                 return -ENOENT;
1168
1169         ct = nf_ct_tuplehash_to_ctrack(h);
1170
1171         if (cda[CTA_ID]) {
1172                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1173                 if (id != (u32)(unsigned long)ct) {
1174                         nf_ct_put(ct);
1175                         return -ENOENT;
1176                 }
1177         }
1178
1179         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1180         nf_ct_put(ct);
1181
1182         return 0;
1183 }
1184
1185 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1186                                    struct sk_buff *skb,
1187                                    const struct nlmsghdr *nlh,
1188                                    const struct nlattr * const cda[],
1189                                    struct netlink_ext_ack *extack)
1190 {
1191         struct nf_conntrack_tuple_hash *h;
1192         struct nf_conntrack_tuple tuple;
1193         struct nf_conn *ct;
1194         struct sk_buff *skb2 = NULL;
1195         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1196         u_int8_t u3 = nfmsg->nfgen_family;
1197         struct nf_conntrack_zone zone;
1198         int err;
1199
1200         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1201                 struct netlink_dump_control c = {
1202                         .dump = ctnetlink_dump_table,
1203                         .done = ctnetlink_done,
1204                 };
1205
1206                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1207                         struct ctnetlink_filter *filter;
1208
1209                         filter = ctnetlink_alloc_filter(cda);
1210                         if (IS_ERR(filter))
1211                                 return PTR_ERR(filter);
1212
1213                         c.data = filter;
1214                 }
1215                 return netlink_dump_start(ctnl, skb, nlh, &c);
1216         }
1217
1218         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1219         if (err < 0)
1220                 return err;
1221
1222         if (cda[CTA_TUPLE_ORIG])
1223                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1224                                             u3, &zone);
1225         else if (cda[CTA_TUPLE_REPLY])
1226                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1227                                             u3, &zone);
1228         else
1229                 return -EINVAL;
1230
1231         if (err < 0)
1232                 return err;
1233
1234         h = nf_conntrack_find_get(net, &zone, &tuple);
1235         if (!h)
1236                 return -ENOENT;
1237
1238         ct = nf_ct_tuplehash_to_ctrack(h);
1239
1240         err = -ENOMEM;
1241         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1242         if (skb2 == NULL) {
1243                 nf_ct_put(ct);
1244                 return -ENOMEM;
1245         }
1246
1247         rcu_read_lock();
1248         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1249                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1250         rcu_read_unlock();
1251         nf_ct_put(ct);
1252         if (err <= 0)
1253                 goto free;
1254
1255         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1256         if (err < 0)
1257                 goto out;
1258
1259         return 0;
1260
1261 free:
1262         kfree_skb(skb2);
1263 out:
1264         /* this avoids a loop in nfnetlink. */
1265         return err == -EAGAIN ? -ENOBUFS : err;
1266 }
1267
1268 static int ctnetlink_done_list(struct netlink_callback *cb)
1269 {
1270         if (cb->args[1])
1271                 nf_ct_put((struct nf_conn *)cb->args[1]);
1272         return 0;
1273 }
1274
1275 static int
1276 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1277 {
1278         struct nf_conn *ct, *last;
1279         struct nf_conntrack_tuple_hash *h;
1280         struct hlist_nulls_node *n;
1281         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1282         u_int8_t l3proto = nfmsg->nfgen_family;
1283         int res;
1284         int cpu;
1285         struct hlist_nulls_head *list;
1286         struct net *net = sock_net(skb->sk);
1287
1288         if (cb->args[2])
1289                 return 0;
1290
1291         last = (struct nf_conn *)cb->args[1];
1292
1293         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1294                 struct ct_pcpu *pcpu;
1295
1296                 if (!cpu_possible(cpu))
1297                         continue;
1298
1299                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1300                 spin_lock_bh(&pcpu->lock);
1301                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1302 restart:
1303                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1304                         ct = nf_ct_tuplehash_to_ctrack(h);
1305                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1306                                 continue;
1307                         if (cb->args[1]) {
1308                                 if (ct != last)
1309                                         continue;
1310                                 cb->args[1] = 0;
1311                         }
1312                         rcu_read_lock();
1313                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1314                                                   cb->nlh->nlmsg_seq,
1315                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1316                                                   ct);
1317                         rcu_read_unlock();
1318                         if (res < 0) {
1319                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1320                                         continue;
1321                                 cb->args[0] = cpu;
1322                                 cb->args[1] = (unsigned long)ct;
1323                                 spin_unlock_bh(&pcpu->lock);
1324                                 goto out;
1325                         }
1326                 }
1327                 if (cb->args[1]) {
1328                         cb->args[1] = 0;
1329                         goto restart;
1330                 }
1331                 spin_unlock_bh(&pcpu->lock);
1332         }
1333         cb->args[2] = 1;
1334 out:
1335         if (last)
1336                 nf_ct_put(last);
1337
1338         return skb->len;
1339 }
1340
1341 static int
1342 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1343 {
1344         return ctnetlink_dump_list(skb, cb, true);
1345 }
1346
1347 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1348                                   struct sk_buff *skb,
1349                                   const struct nlmsghdr *nlh,
1350                                   const struct nlattr * const cda[],
1351                                   struct netlink_ext_ack *extack)
1352 {
1353         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1354                 struct netlink_dump_control c = {
1355                         .dump = ctnetlink_dump_dying,
1356                         .done = ctnetlink_done_list,
1357                 };
1358                 return netlink_dump_start(ctnl, skb, nlh, &c);
1359         }
1360
1361         return -EOPNOTSUPP;
1362 }
1363
1364 static int
1365 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1366 {
1367         return ctnetlink_dump_list(skb, cb, false);
1368 }
1369
1370 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1371                                         struct sk_buff *skb,
1372                                         const struct nlmsghdr *nlh,
1373                                         const struct nlattr * const cda[],
1374                                         struct netlink_ext_ack *extack)
1375 {
1376         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1377                 struct netlink_dump_control c = {
1378                         .dump = ctnetlink_dump_unconfirmed,
1379                         .done = ctnetlink_done_list,
1380                 };
1381                 return netlink_dump_start(ctnl, skb, nlh, &c);
1382         }
1383
1384         return -EOPNOTSUPP;
1385 }
1386
1387 #ifdef CONFIG_NF_NAT_NEEDED
1388 static int
1389 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1390                           enum nf_nat_manip_type manip,
1391                           const struct nlattr *attr)
1392 {
1393         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1394         int err;
1395
1396         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1397         if (!parse_nat_setup) {
1398 #ifdef CONFIG_MODULES
1399                 rcu_read_unlock();
1400                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1401                 if (request_module("nf-nat") < 0) {
1402                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1403                         rcu_read_lock();
1404                         return -EOPNOTSUPP;
1405                 }
1406                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1407                 rcu_read_lock();
1408                 if (nfnetlink_parse_nat_setup_hook)
1409                         return -EAGAIN;
1410 #endif
1411                 return -EOPNOTSUPP;
1412         }
1413
1414         err = parse_nat_setup(ct, manip, attr);
1415         if (err == -EAGAIN) {
1416 #ifdef CONFIG_MODULES
1417                 rcu_read_unlock();
1418                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1419                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1420                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1421                         rcu_read_lock();
1422                         return -EOPNOTSUPP;
1423                 }
1424                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1425                 rcu_read_lock();
1426 #else
1427                 err = -EOPNOTSUPP;
1428 #endif
1429         }
1430         return err;
1431 }
1432 #endif
1433
1434 static void
1435 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1436                           unsigned long off)
1437 {
1438         unsigned int bit;
1439
1440         /* Ignore these unchangable bits */
1441         on &= ~IPS_UNCHANGEABLE_MASK;
1442         off &= ~IPS_UNCHANGEABLE_MASK;
1443
1444         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1445                 if (on & (1 << bit))
1446                         set_bit(bit, &ct->status);
1447                 else if (off & (1 << bit))
1448                         clear_bit(bit, &ct->status);
1449         }
1450 }
1451
1452 static int
1453 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1454 {
1455         unsigned long d;
1456         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1457         d = ct->status ^ status;
1458
1459         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1460                 /* unchangeable */
1461                 return -EBUSY;
1462
1463         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1464                 /* SEEN_REPLY bit can only be set */
1465                 return -EBUSY;
1466
1467         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1468                 /* ASSURED bit can only be set */
1469                 return -EBUSY;
1470
1471         __ctnetlink_change_status(ct, status, 0);
1472         return 0;
1473 }
1474
1475 static int
1476 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1477 {
1478 #ifdef CONFIG_NF_NAT_NEEDED
1479         int ret;
1480
1481         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1482                 return 0;
1483
1484         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1485                                         cda[CTA_NAT_DST]);
1486         if (ret < 0)
1487                 return ret;
1488
1489         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1490                                         cda[CTA_NAT_SRC]);
1491         return ret;
1492 #else
1493         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1494                 return 0;
1495         return -EOPNOTSUPP;
1496 #endif
1497 }
1498
1499 static int ctnetlink_change_helper(struct nf_conn *ct,
1500                                    const struct nlattr * const cda[])
1501 {
1502         struct nf_conntrack_helper *helper;
1503         struct nf_conn_help *help = nfct_help(ct);
1504         char *helpname = NULL;
1505         struct nlattr *helpinfo = NULL;
1506         int err;
1507
1508         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1509         if (err < 0)
1510                 return err;
1511
1512         /* don't change helper of sibling connections */
1513         if (ct->master) {
1514                 /* If we try to change the helper to the same thing twice,
1515                  * treat the second attempt as a no-op instead of returning
1516                  * an error.
1517                  */
1518                 err = -EBUSY;
1519                 if (help) {
1520                         rcu_read_lock();
1521                         helper = rcu_dereference(help->helper);
1522                         if (helper && !strcmp(helper->name, helpname))
1523                                 err = 0;
1524                         rcu_read_unlock();
1525                 }
1526
1527                 return err;
1528         }
1529
1530         if (!strcmp(helpname, "")) {
1531                 if (help && help->helper) {
1532                         /* we had a helper before ... */
1533                         nf_ct_remove_expectations(ct);
1534                         RCU_INIT_POINTER(help->helper, NULL);
1535                 }
1536
1537                 return 0;
1538         }
1539
1540         rcu_read_lock();
1541         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1542                                             nf_ct_protonum(ct));
1543         if (helper == NULL) {
1544                 rcu_read_unlock();
1545                 return -EOPNOTSUPP;
1546         }
1547
1548         if (help) {
1549                 if (help->helper == helper) {
1550                         /* update private helper data if allowed. */
1551                         if (helper->from_nlattr)
1552                                 helper->from_nlattr(helpinfo, ct);
1553                         err = 0;
1554                 } else
1555                         err = -EBUSY;
1556         } else {
1557                 /* we cannot set a helper for an existing conntrack */
1558                 err = -EOPNOTSUPP;
1559         }
1560
1561         rcu_read_unlock();
1562         return err;
1563 }
1564
1565 static int ctnetlink_change_timeout(struct nf_conn *ct,
1566                                     const struct nlattr * const cda[])
1567 {
1568         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1569
1570         ct->timeout = nfct_time_stamp + timeout * HZ;
1571
1572         if (test_bit(IPS_DYING_BIT, &ct->status))
1573                 return -ETIME;
1574
1575         return 0;
1576 }
1577
1578 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1579         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1580         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1581         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1582 };
1583
1584 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1585                                       const struct nlattr * const cda[])
1586 {
1587         const struct nlattr *attr = cda[CTA_PROTOINFO];
1588         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1589         struct nf_conntrack_l4proto *l4proto;
1590         int err = 0;
1591
1592         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1593                                NULL);
1594         if (err < 0)
1595                 return err;
1596
1597         rcu_read_lock();
1598         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1599         if (l4proto->from_nlattr)
1600                 err = l4proto->from_nlattr(tb, ct);
1601         rcu_read_unlock();
1602
1603         return err;
1604 }
1605
1606 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1607         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1608         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1609         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1610 };
1611
1612 static int change_seq_adj(struct nf_ct_seqadj *seq,
1613                           const struct nlattr * const attr)
1614 {
1615         int err;
1616         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1617
1618         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1619         if (err < 0)
1620                 return err;
1621
1622         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1623                 return -EINVAL;
1624
1625         seq->correction_pos =
1626                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1627
1628         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1629                 return -EINVAL;
1630
1631         seq->offset_before =
1632                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1633
1634         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1635                 return -EINVAL;
1636
1637         seq->offset_after =
1638                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1639
1640         return 0;
1641 }
1642
1643 static int
1644 ctnetlink_change_seq_adj(struct nf_conn *ct,
1645                          const struct nlattr * const cda[])
1646 {
1647         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1648         int ret = 0;
1649
1650         if (!seqadj)
1651                 return 0;
1652
1653         spin_lock_bh(&ct->lock);
1654         if (cda[CTA_SEQ_ADJ_ORIG]) {
1655                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1656                                      cda[CTA_SEQ_ADJ_ORIG]);
1657                 if (ret < 0)
1658                         goto err;
1659
1660                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1661         }
1662
1663         if (cda[CTA_SEQ_ADJ_REPLY]) {
1664                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1665                                      cda[CTA_SEQ_ADJ_REPLY]);
1666                 if (ret < 0)
1667                         goto err;
1668
1669                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1670         }
1671
1672         spin_unlock_bh(&ct->lock);
1673         return 0;
1674 err:
1675         spin_unlock_bh(&ct->lock);
1676         return ret;
1677 }
1678
1679 static int
1680 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1681 {
1682 #ifdef CONFIG_NF_CONNTRACK_LABELS
1683         size_t len = nla_len(cda[CTA_LABELS]);
1684         const void *mask = cda[CTA_LABELS_MASK];
1685
1686         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1687                 return -EINVAL;
1688
1689         if (mask) {
1690                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1691                     nla_len(cda[CTA_LABELS_MASK]) != len)
1692                         return -EINVAL;
1693                 mask = nla_data(cda[CTA_LABELS_MASK]);
1694         }
1695
1696         len /= sizeof(u32);
1697
1698         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1699 #else
1700         return -EOPNOTSUPP;
1701 #endif
1702 }
1703
1704 static int
1705 ctnetlink_change_conntrack(struct nf_conn *ct,
1706                            const struct nlattr * const cda[])
1707 {
1708         int err;
1709
1710         /* only allow NAT changes and master assignation for new conntracks */
1711         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1712                 return -EOPNOTSUPP;
1713
1714         if (cda[CTA_HELP]) {
1715                 err = ctnetlink_change_helper(ct, cda);
1716                 if (err < 0)
1717                         return err;
1718         }
1719
1720         if (cda[CTA_TIMEOUT]) {
1721                 err = ctnetlink_change_timeout(ct, cda);
1722                 if (err < 0)
1723                         return err;
1724         }
1725
1726         if (cda[CTA_STATUS]) {
1727                 err = ctnetlink_change_status(ct, cda);
1728                 if (err < 0)
1729                         return err;
1730         }
1731
1732         if (cda[CTA_PROTOINFO]) {
1733                 err = ctnetlink_change_protoinfo(ct, cda);
1734                 if (err < 0)
1735                         return err;
1736         }
1737
1738 #if defined(CONFIG_NF_CONNTRACK_MARK)
1739         if (cda[CTA_MARK])
1740                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1741 #endif
1742
1743         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1744                 err = ctnetlink_change_seq_adj(ct, cda);
1745                 if (err < 0)
1746                         return err;
1747         }
1748
1749         if (cda[CTA_LABELS]) {
1750                 err = ctnetlink_attach_labels(ct, cda);
1751                 if (err < 0)
1752                         return err;
1753         }
1754
1755         return 0;
1756 }
1757
1758 static struct nf_conn *
1759 ctnetlink_create_conntrack(struct net *net,
1760                            const struct nf_conntrack_zone *zone,
1761                            const struct nlattr * const cda[],
1762                            struct nf_conntrack_tuple *otuple,
1763                            struct nf_conntrack_tuple *rtuple,
1764                            u8 u3)
1765 {
1766         struct nf_conn *ct;
1767         int err = -EINVAL;
1768         struct nf_conntrack_helper *helper;
1769         struct nf_conn_tstamp *tstamp;
1770
1771         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1772         if (IS_ERR(ct))
1773                 return ERR_PTR(-ENOMEM);
1774
1775         if (!cda[CTA_TIMEOUT])
1776                 goto err1;
1777
1778         ct->timeout = nfct_time_stamp + ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1779
1780         rcu_read_lock();
1781         if (cda[CTA_HELP]) {
1782                 char *helpname = NULL;
1783                 struct nlattr *helpinfo = NULL;
1784
1785                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1786                 if (err < 0)
1787                         goto err2;
1788
1789                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1790                                                     nf_ct_protonum(ct));
1791                 if (helper == NULL) {
1792                         rcu_read_unlock();
1793 #ifdef CONFIG_MODULES
1794                         if (request_module("nfct-helper-%s", helpname) < 0) {
1795                                 err = -EOPNOTSUPP;
1796                                 goto err1;
1797                         }
1798
1799                         rcu_read_lock();
1800                         helper = __nf_conntrack_helper_find(helpname,
1801                                                             nf_ct_l3num(ct),
1802                                                             nf_ct_protonum(ct));
1803                         if (helper) {
1804                                 err = -EAGAIN;
1805                                 goto err2;
1806                         }
1807                         rcu_read_unlock();
1808 #endif
1809                         err = -EOPNOTSUPP;
1810                         goto err1;
1811                 } else {
1812                         struct nf_conn_help *help;
1813
1814                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1815                         if (help == NULL) {
1816                                 err = -ENOMEM;
1817                                 goto err2;
1818                         }
1819                         /* set private helper data if allowed. */
1820                         if (helper->from_nlattr)
1821                                 helper->from_nlattr(helpinfo, ct);
1822
1823                         /* not in hash table yet so not strictly necessary */
1824                         RCU_INIT_POINTER(help->helper, helper);
1825                 }
1826         } else {
1827                 /* try an implicit helper assignation */
1828                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1829                 if (err < 0)
1830                         goto err2;
1831         }
1832
1833         err = ctnetlink_setup_nat(ct, cda);
1834         if (err < 0)
1835                 goto err2;
1836
1837         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1838         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1839         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1840         nf_ct_labels_ext_add(ct);
1841         nfct_seqadj_ext_add(ct);
1842         nfct_synproxy_ext_add(ct);
1843
1844         /* we must add conntrack extensions before confirmation. */
1845         ct->status |= IPS_CONFIRMED;
1846
1847         if (cda[CTA_STATUS]) {
1848                 err = ctnetlink_change_status(ct, cda);
1849                 if (err < 0)
1850                         goto err2;
1851         }
1852
1853         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1854                 err = ctnetlink_change_seq_adj(ct, cda);
1855                 if (err < 0)
1856                         goto err2;
1857         }
1858
1859         memset(&ct->proto, 0, sizeof(ct->proto));
1860         if (cda[CTA_PROTOINFO]) {
1861                 err = ctnetlink_change_protoinfo(ct, cda);
1862                 if (err < 0)
1863                         goto err2;
1864         }
1865
1866 #if defined(CONFIG_NF_CONNTRACK_MARK)
1867         if (cda[CTA_MARK])
1868                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1869 #endif
1870
1871         /* setup master conntrack: this is a confirmed expectation */
1872         if (cda[CTA_TUPLE_MASTER]) {
1873                 struct nf_conntrack_tuple master;
1874                 struct nf_conntrack_tuple_hash *master_h;
1875                 struct nf_conn *master_ct;
1876
1877                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1878                                             u3, NULL);
1879                 if (err < 0)
1880                         goto err2;
1881
1882                 master_h = nf_conntrack_find_get(net, zone, &master);
1883                 if (master_h == NULL) {
1884                         err = -ENOENT;
1885                         goto err2;
1886                 }
1887                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1888                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1889                 ct->master = master_ct;
1890         }
1891         tstamp = nf_conn_tstamp_find(ct);
1892         if (tstamp)
1893                 tstamp->start = ktime_get_real_ns();
1894
1895         err = nf_conntrack_hash_check_insert(ct);
1896         if (err < 0)
1897                 goto err2;
1898
1899         rcu_read_unlock();
1900
1901         return ct;
1902
1903 err2:
1904         rcu_read_unlock();
1905 err1:
1906         nf_conntrack_free(ct);
1907         return ERR_PTR(err);
1908 }
1909
1910 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
1911                                    struct sk_buff *skb,
1912                                    const struct nlmsghdr *nlh,
1913                                    const struct nlattr * const cda[],
1914                                    struct netlink_ext_ack *extack)
1915 {
1916         struct nf_conntrack_tuple otuple, rtuple;
1917         struct nf_conntrack_tuple_hash *h = NULL;
1918         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1919         struct nf_conn *ct;
1920         u_int8_t u3 = nfmsg->nfgen_family;
1921         struct nf_conntrack_zone zone;
1922         int err;
1923
1924         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1925         if (err < 0)
1926                 return err;
1927
1928         if (cda[CTA_TUPLE_ORIG]) {
1929                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1930                                             u3, &zone);
1931                 if (err < 0)
1932                         return err;
1933         }
1934
1935         if (cda[CTA_TUPLE_REPLY]) {
1936                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1937                                             u3, &zone);
1938                 if (err < 0)
1939                         return err;
1940         }
1941
1942         if (cda[CTA_TUPLE_ORIG])
1943                 h = nf_conntrack_find_get(net, &zone, &otuple);
1944         else if (cda[CTA_TUPLE_REPLY])
1945                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1946
1947         if (h == NULL) {
1948                 err = -ENOENT;
1949                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1950                         enum ip_conntrack_events events;
1951
1952                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1953                                 return -EINVAL;
1954                         if (otuple.dst.protonum != rtuple.dst.protonum)
1955                                 return -EINVAL;
1956
1957                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1958                                                         &rtuple, u3);
1959                         if (IS_ERR(ct))
1960                                 return PTR_ERR(ct);
1961
1962                         err = 0;
1963                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1964                                 events = 1 << IPCT_RELATED;
1965                         else
1966                                 events = 1 << IPCT_NEW;
1967
1968                         if (cda[CTA_LABELS] &&
1969                             ctnetlink_attach_labels(ct, cda) == 0)
1970                                 events |= (1 << IPCT_LABEL);
1971
1972                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1973                                                       (1 << IPCT_ASSURED) |
1974                                                       (1 << IPCT_HELPER) |
1975                                                       (1 << IPCT_PROTOINFO) |
1976                                                       (1 << IPCT_SEQADJ) |
1977                                                       (1 << IPCT_MARK) | events,
1978                                                       ct, NETLINK_CB(skb).portid,
1979                                                       nlmsg_report(nlh));
1980                         nf_ct_put(ct);
1981                 }
1982
1983                 return err;
1984         }
1985         /* implicit 'else' */
1986
1987         err = -EEXIST;
1988         ct = nf_ct_tuplehash_to_ctrack(h);
1989         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1990                 err = ctnetlink_change_conntrack(ct, cda);
1991                 if (err == 0) {
1992                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1993                                                       (1 << IPCT_ASSURED) |
1994                                                       (1 << IPCT_HELPER) |
1995                                                       (1 << IPCT_LABEL) |
1996                                                       (1 << IPCT_PROTOINFO) |
1997                                                       (1 << IPCT_SEQADJ) |
1998                                                       (1 << IPCT_MARK),
1999                                                       ct, NETLINK_CB(skb).portid,
2000                                                       nlmsg_report(nlh));
2001                 }
2002         }
2003
2004         nf_ct_put(ct);
2005         return err;
2006 }
2007
2008 static int
2009 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2010                                 __u16 cpu, const struct ip_conntrack_stat *st)
2011 {
2012         struct nlmsghdr *nlh;
2013         struct nfgenmsg *nfmsg;
2014         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2015
2016         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2017                               IPCTNL_MSG_CT_GET_STATS_CPU);
2018         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2019         if (nlh == NULL)
2020                 goto nlmsg_failure;
2021
2022         nfmsg = nlmsg_data(nlh);
2023         nfmsg->nfgen_family = AF_UNSPEC;
2024         nfmsg->version      = NFNETLINK_V0;
2025         nfmsg->res_id       = htons(cpu);
2026
2027         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2028             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2029             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2030             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2031             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2032                                 htonl(st->insert_failed)) ||
2033             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2034             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2035             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2036             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2037                                 htonl(st->search_restart)))
2038                 goto nla_put_failure;
2039
2040         nlmsg_end(skb, nlh);
2041         return skb->len;
2042
2043 nla_put_failure:
2044 nlmsg_failure:
2045         nlmsg_cancel(skb, nlh);
2046         return -1;
2047 }
2048
2049 static int
2050 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2051 {
2052         int cpu;
2053         struct net *net = sock_net(skb->sk);
2054
2055         if (cb->args[0] == nr_cpu_ids)
2056                 return 0;
2057
2058         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2059                 const struct ip_conntrack_stat *st;
2060
2061                 if (!cpu_possible(cpu))
2062                         continue;
2063
2064                 st = per_cpu_ptr(net->ct.stat, cpu);
2065                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2066                                                     NETLINK_CB(cb->skb).portid,
2067                                                     cb->nlh->nlmsg_seq,
2068                                                     cpu, st) < 0)
2069                                 break;
2070         }
2071         cb->args[0] = cpu;
2072
2073         return skb->len;
2074 }
2075
2076 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2077                                  struct sk_buff *skb,
2078                                  const struct nlmsghdr *nlh,
2079                                  const struct nlattr * const cda[],
2080                                  struct netlink_ext_ack *extack)
2081 {
2082         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2083                 struct netlink_dump_control c = {
2084                         .dump = ctnetlink_ct_stat_cpu_dump,
2085                 };
2086                 return netlink_dump_start(ctnl, skb, nlh, &c);
2087         }
2088
2089         return 0;
2090 }
2091
2092 static int
2093 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2094                             struct net *net)
2095 {
2096         struct nlmsghdr *nlh;
2097         struct nfgenmsg *nfmsg;
2098         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2099         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2100
2101         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2102         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2103         if (nlh == NULL)
2104                 goto nlmsg_failure;
2105
2106         nfmsg = nlmsg_data(nlh);
2107         nfmsg->nfgen_family = AF_UNSPEC;
2108         nfmsg->version      = NFNETLINK_V0;
2109         nfmsg->res_id       = 0;
2110
2111         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2112                 goto nla_put_failure;
2113
2114         nlmsg_end(skb, nlh);
2115         return skb->len;
2116
2117 nla_put_failure:
2118 nlmsg_failure:
2119         nlmsg_cancel(skb, nlh);
2120         return -1;
2121 }
2122
2123 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2124                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2125                              const struct nlattr * const cda[],
2126                              struct netlink_ext_ack *extack)
2127 {
2128         struct sk_buff *skb2;
2129         int err;
2130
2131         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2132         if (skb2 == NULL)
2133                 return -ENOMEM;
2134
2135         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2136                                           nlh->nlmsg_seq,
2137                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2138                                           sock_net(skb->sk));
2139         if (err <= 0)
2140                 goto free;
2141
2142         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2143         if (err < 0)
2144                 goto out;
2145
2146         return 0;
2147
2148 free:
2149         kfree_skb(skb2);
2150 out:
2151         /* this avoids a loop in nfnetlink. */
2152         return err == -EAGAIN ? -ENOBUFS : err;
2153 }
2154
2155 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2156         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2157         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2158         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2159         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2160         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2161         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2162                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2163         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2164         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2165         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2166         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2167         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2168 };
2169
2170 static struct nf_conntrack_expect *
2171 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2172                        struct nf_conntrack_helper *helper,
2173                        struct nf_conntrack_tuple *tuple,
2174                        struct nf_conntrack_tuple *mask);
2175
2176 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2177 static size_t
2178 ctnetlink_glue_build_size(const struct nf_conn *ct)
2179 {
2180         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2181                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2182                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2183                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2184                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2185                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2186                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2187                + nla_total_size(0) /* CTA_PROTOINFO */
2188                + nla_total_size(0) /* CTA_HELP */
2189                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2190                + ctnetlink_secctx_size(ct)
2191 #ifdef CONFIG_NF_NAT_NEEDED
2192                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2193                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2194 #endif
2195 #ifdef CONFIG_NF_CONNTRACK_MARK
2196                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2197 #endif
2198 #ifdef CONFIG_NF_CONNTRACK_ZONES
2199                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2200 #endif
2201                + ctnetlink_proto_size(ct)
2202                ;
2203 }
2204
2205 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2206                                              enum ip_conntrack_info *ctinfo)
2207 {
2208         return nf_ct_get(skb, ctinfo);
2209 }
2210
2211 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2212 {
2213         const struct nf_conntrack_zone *zone;
2214         struct nlattr *nest_parms;
2215
2216         rcu_read_lock();
2217         zone = nf_ct_zone(ct);
2218
2219         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2220         if (!nest_parms)
2221                 goto nla_put_failure;
2222         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2223                 goto nla_put_failure;
2224         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2225                                    NF_CT_ZONE_DIR_ORIG) < 0)
2226                 goto nla_put_failure;
2227         nla_nest_end(skb, nest_parms);
2228
2229         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2230         if (!nest_parms)
2231                 goto nla_put_failure;
2232         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2233                 goto nla_put_failure;
2234         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2235                                    NF_CT_ZONE_DIR_REPL) < 0)
2236                 goto nla_put_failure;
2237         nla_nest_end(skb, nest_parms);
2238
2239         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2240                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2241                 goto nla_put_failure;
2242
2243         if (ctnetlink_dump_id(skb, ct) < 0)
2244                 goto nla_put_failure;
2245
2246         if (ctnetlink_dump_status(skb, ct) < 0)
2247                 goto nla_put_failure;
2248
2249         if (ctnetlink_dump_timeout(skb, ct) < 0)
2250                 goto nla_put_failure;
2251
2252         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2253                 goto nla_put_failure;
2254
2255         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2256                 goto nla_put_failure;
2257
2258 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2259         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2260                 goto nla_put_failure;
2261 #endif
2262         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2263                 goto nla_put_failure;
2264
2265         if ((ct->status & IPS_SEQ_ADJUST) &&
2266             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2267                 goto nla_put_failure;
2268
2269 #ifdef CONFIG_NF_CONNTRACK_MARK
2270         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2271                 goto nla_put_failure;
2272 #endif
2273         if (ctnetlink_dump_labels(skb, ct) < 0)
2274                 goto nla_put_failure;
2275         rcu_read_unlock();
2276         return 0;
2277
2278 nla_put_failure:
2279         rcu_read_unlock();
2280         return -ENOSPC;
2281 }
2282
2283 static int
2284 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2285                      enum ip_conntrack_info ctinfo,
2286                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2287 {
2288         struct nlattr *nest_parms;
2289
2290         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2291         if (!nest_parms)
2292                 goto nla_put_failure;
2293
2294         if (__ctnetlink_glue_build(skb, ct) < 0)
2295                 goto nla_put_failure;
2296
2297         nla_nest_end(skb, nest_parms);
2298
2299         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2300                 goto nla_put_failure;
2301
2302         return 0;
2303
2304 nla_put_failure:
2305         return -ENOSPC;
2306 }
2307
2308 static int
2309 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2310 {
2311         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2312         unsigned long d = ct->status ^ status;
2313
2314         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2315                 /* SEEN_REPLY bit can only be set */
2316                 return -EBUSY;
2317
2318         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2319                 /* ASSURED bit can only be set */
2320                 return -EBUSY;
2321
2322         /* This check is less strict than ctnetlink_change_status()
2323          * because callers often flip IPS_EXPECTED bits when sending
2324          * an NFQA_CT attribute to the kernel.  So ignore the
2325          * unchangeable bits but do not error out. Also user programs
2326          * are allowed to clear the bits that they are allowed to change.
2327          */
2328         __ctnetlink_change_status(ct, status, ~status);
2329         return 0;
2330 }
2331
2332 static int
2333 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2334 {
2335         int err;
2336
2337         if (cda[CTA_TIMEOUT]) {
2338                 err = ctnetlink_change_timeout(ct, cda);
2339                 if (err < 0)
2340                         return err;
2341         }
2342         if (cda[CTA_STATUS]) {
2343                 err = ctnetlink_update_status(ct, cda);
2344                 if (err < 0)
2345                         return err;
2346         }
2347         if (cda[CTA_HELP]) {
2348                 err = ctnetlink_change_helper(ct, cda);
2349                 if (err < 0)
2350                         return err;
2351         }
2352         if (cda[CTA_LABELS]) {
2353                 err = ctnetlink_attach_labels(ct, cda);
2354                 if (err < 0)
2355                         return err;
2356         }
2357 #if defined(CONFIG_NF_CONNTRACK_MARK)
2358         if (cda[CTA_MARK]) {
2359                 u32 mask = 0, mark, newmark;
2360                 if (cda[CTA_MARK_MASK])
2361                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2362
2363                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2364                 newmark = (ct->mark & mask) ^ mark;
2365                 if (newmark != ct->mark)
2366                         ct->mark = newmark;
2367         }
2368 #endif
2369         return 0;
2370 }
2371
2372 static int
2373 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2374 {
2375         struct nlattr *cda[CTA_MAX+1];
2376         int ret;
2377
2378         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2379         if (ret < 0)
2380                 return ret;
2381
2382         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2383 }
2384
2385 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2386                                     const struct nf_conn *ct,
2387                                     struct nf_conntrack_tuple *tuple,
2388                                     struct nf_conntrack_tuple *mask)
2389 {
2390         int err;
2391
2392         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2393                                     nf_ct_l3num(ct), NULL);
2394         if (err < 0)
2395                 return err;
2396
2397         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2398                                      nf_ct_l3num(ct), NULL);
2399 }
2400
2401 static int
2402 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2403                              u32 portid, u32 report)
2404 {
2405         struct nlattr *cda[CTA_EXPECT_MAX+1];
2406         struct nf_conntrack_tuple tuple, mask;
2407         struct nf_conntrack_helper *helper = NULL;
2408         struct nf_conntrack_expect *exp;
2409         int err;
2410
2411         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2412                                NULL);
2413         if (err < 0)
2414                 return err;
2415
2416         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2417                                        ct, &tuple, &mask);
2418         if (err < 0)
2419                 return err;
2420
2421         if (cda[CTA_EXPECT_HELP_NAME]) {
2422                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2423
2424                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2425                                                     nf_ct_protonum(ct));
2426                 if (helper == NULL)
2427                         return -EOPNOTSUPP;
2428         }
2429
2430         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2431                                      helper, &tuple, &mask);
2432         if (IS_ERR(exp))
2433                 return PTR_ERR(exp);
2434
2435         err = nf_ct_expect_related_report(exp, portid, report);
2436         nf_ct_expect_put(exp);
2437         return err;
2438 }
2439
2440 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2441                                   enum ip_conntrack_info ctinfo, int diff)
2442 {
2443         if (!(ct->status & IPS_NAT_MASK))
2444                 return;
2445
2446         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2447 }
2448
2449 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2450         .get_ct         = ctnetlink_glue_get_ct,
2451         .build_size     = ctnetlink_glue_build_size,
2452         .build          = ctnetlink_glue_build,
2453         .parse          = ctnetlink_glue_parse,
2454         .attach_expect  = ctnetlink_glue_attach_expect,
2455         .seq_adjust     = ctnetlink_glue_seqadj,
2456 };
2457 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2458
2459 /***********************************************************************
2460  * EXPECT
2461  ***********************************************************************/
2462
2463 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2464                                     const struct nf_conntrack_tuple *tuple,
2465                                     u32 type)
2466 {
2467         struct nlattr *nest_parms;
2468
2469         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2470         if (!nest_parms)
2471                 goto nla_put_failure;
2472         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2473                 goto nla_put_failure;
2474         nla_nest_end(skb, nest_parms);
2475
2476         return 0;
2477
2478 nla_put_failure:
2479         return -1;
2480 }
2481
2482 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2483                                    const struct nf_conntrack_tuple *tuple,
2484                                    const struct nf_conntrack_tuple_mask *mask)
2485 {
2486         int ret;
2487         struct nf_conntrack_l3proto *l3proto;
2488         struct nf_conntrack_l4proto *l4proto;
2489         struct nf_conntrack_tuple m;
2490         struct nlattr *nest_parms;
2491
2492         memset(&m, 0xFF, sizeof(m));
2493         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2494         m.src.u.all = mask->src.u.all;
2495         m.dst.protonum = tuple->dst.protonum;
2496
2497         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2498         if (!nest_parms)
2499                 goto nla_put_failure;
2500
2501         rcu_read_lock();
2502         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2503         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2504         if (ret >= 0) {
2505                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2506                                                tuple->dst.protonum);
2507         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2508         }
2509         rcu_read_unlock();
2510
2511         if (unlikely(ret < 0))
2512                 goto nla_put_failure;
2513
2514         nla_nest_end(skb, nest_parms);
2515
2516         return 0;
2517
2518 nla_put_failure:
2519         return -1;
2520 }
2521
2522 static const union nf_inet_addr any_addr;
2523
2524 static int
2525 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2526                           const struct nf_conntrack_expect *exp)
2527 {
2528         struct nf_conn *master = exp->master;
2529         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2530         struct nf_conn_help *help;
2531 #ifdef CONFIG_NF_NAT_NEEDED
2532         struct nlattr *nest_parms;
2533         struct nf_conntrack_tuple nat_tuple = {};
2534 #endif
2535         struct nf_ct_helper_expectfn *expfn;
2536
2537         if (timeout < 0)
2538                 timeout = 0;
2539
2540         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2541                 goto nla_put_failure;
2542         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2543                 goto nla_put_failure;
2544         if (ctnetlink_exp_dump_tuple(skb,
2545                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2546                                  CTA_EXPECT_MASTER) < 0)
2547                 goto nla_put_failure;
2548
2549 #ifdef CONFIG_NF_NAT_NEEDED
2550         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2551             exp->saved_proto.all) {
2552                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2553                 if (!nest_parms)
2554                         goto nla_put_failure;
2555
2556                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2557                         goto nla_put_failure;
2558
2559                 nat_tuple.src.l3num = nf_ct_l3num(master);
2560                 nat_tuple.src.u3 = exp->saved_addr;
2561                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2562                 nat_tuple.src.u = exp->saved_proto;
2563
2564                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2565                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2566                         goto nla_put_failure;
2567                 nla_nest_end(skb, nest_parms);
2568         }
2569 #endif
2570         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2571             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2572             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2573             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2574                 goto nla_put_failure;
2575         help = nfct_help(master);
2576         if (help) {
2577                 struct nf_conntrack_helper *helper;
2578
2579                 helper = rcu_dereference(help->helper);
2580                 if (helper &&
2581                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2582                         goto nla_put_failure;
2583         }
2584         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2585         if (expfn != NULL &&
2586             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2587                 goto nla_put_failure;
2588
2589         return 0;
2590
2591 nla_put_failure:
2592         return -1;
2593 }
2594
2595 static int
2596 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2597                         int event, const struct nf_conntrack_expect *exp)
2598 {
2599         struct nlmsghdr *nlh;
2600         struct nfgenmsg *nfmsg;
2601         unsigned int flags = portid ? NLM_F_MULTI : 0;
2602
2603         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2604         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2605         if (nlh == NULL)
2606                 goto nlmsg_failure;
2607
2608         nfmsg = nlmsg_data(nlh);
2609         nfmsg->nfgen_family = exp->tuple.src.l3num;
2610         nfmsg->version      = NFNETLINK_V0;
2611         nfmsg->res_id       = 0;
2612
2613         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2614                 goto nla_put_failure;
2615
2616         nlmsg_end(skb, nlh);
2617         return skb->len;
2618
2619 nlmsg_failure:
2620 nla_put_failure:
2621         nlmsg_cancel(skb, nlh);
2622         return -1;
2623 }
2624
2625 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2626 static int
2627 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2628 {
2629         struct nf_conntrack_expect *exp = item->exp;
2630         struct net *net = nf_ct_exp_net(exp);
2631         struct nlmsghdr *nlh;
2632         struct nfgenmsg *nfmsg;
2633         struct sk_buff *skb;
2634         unsigned int type, group;
2635         int flags = 0;
2636
2637         if (events & (1 << IPEXP_DESTROY)) {
2638                 type = IPCTNL_MSG_EXP_DELETE;
2639                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2640         } else if (events & (1 << IPEXP_NEW)) {
2641                 type = IPCTNL_MSG_EXP_NEW;
2642                 flags = NLM_F_CREATE|NLM_F_EXCL;
2643                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2644         } else
2645                 return 0;
2646
2647         if (!item->report && !nfnetlink_has_listeners(net, group))
2648                 return 0;
2649
2650         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2651         if (skb == NULL)
2652                 goto errout;
2653
2654         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2655         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2656         if (nlh == NULL)
2657                 goto nlmsg_failure;
2658
2659         nfmsg = nlmsg_data(nlh);
2660         nfmsg->nfgen_family = exp->tuple.src.l3num;
2661         nfmsg->version      = NFNETLINK_V0;
2662         nfmsg->res_id       = 0;
2663
2664         rcu_read_lock();
2665         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2666                 goto nla_put_failure;
2667         rcu_read_unlock();
2668
2669         nlmsg_end(skb, nlh);
2670         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2671         return 0;
2672
2673 nla_put_failure:
2674         rcu_read_unlock();
2675         nlmsg_cancel(skb, nlh);
2676 nlmsg_failure:
2677         kfree_skb(skb);
2678 errout:
2679         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2680         return 0;
2681 }
2682 #endif
2683 static int ctnetlink_exp_done(struct netlink_callback *cb)
2684 {
2685         if (cb->args[1])
2686                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2687         return 0;
2688 }
2689
2690 static int
2691 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2692 {
2693         struct net *net = sock_net(skb->sk);
2694         struct nf_conntrack_expect *exp, *last;
2695         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2696         u_int8_t l3proto = nfmsg->nfgen_family;
2697
2698         rcu_read_lock();
2699         last = (struct nf_conntrack_expect *)cb->args[1];
2700         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2701 restart:
2702                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2703                                          hnode) {
2704                         if (l3proto && exp->tuple.src.l3num != l3proto)
2705                                 continue;
2706
2707                         if (!net_eq(nf_ct_net(exp->master), net))
2708                                 continue;
2709
2710                         if (cb->args[1]) {
2711                                 if (exp != last)
2712                                         continue;
2713                                 cb->args[1] = 0;
2714                         }
2715                         if (ctnetlink_exp_fill_info(skb,
2716                                                     NETLINK_CB(cb->skb).portid,
2717                                                     cb->nlh->nlmsg_seq,
2718                                                     IPCTNL_MSG_EXP_NEW,
2719                                                     exp) < 0) {
2720                                 if (!refcount_inc_not_zero(&exp->use))
2721                                         continue;
2722                                 cb->args[1] = (unsigned long)exp;
2723                                 goto out;
2724                         }
2725                 }
2726                 if (cb->args[1]) {
2727                         cb->args[1] = 0;
2728                         goto restart;
2729                 }
2730         }
2731 out:
2732         rcu_read_unlock();
2733         if (last)
2734                 nf_ct_expect_put(last);
2735
2736         return skb->len;
2737 }
2738
2739 static int
2740 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2741 {
2742         struct nf_conntrack_expect *exp, *last;
2743         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2744         struct nf_conn *ct = cb->data;
2745         struct nf_conn_help *help = nfct_help(ct);
2746         u_int8_t l3proto = nfmsg->nfgen_family;
2747
2748         if (cb->args[0])
2749                 return 0;
2750
2751         rcu_read_lock();
2752         last = (struct nf_conntrack_expect *)cb->args[1];
2753 restart:
2754         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2755                 if (l3proto && exp->tuple.src.l3num != l3proto)
2756                         continue;
2757                 if (cb->args[1]) {
2758                         if (exp != last)
2759                                 continue;
2760                         cb->args[1] = 0;
2761                 }
2762                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2763                                             cb->nlh->nlmsg_seq,
2764                                             IPCTNL_MSG_EXP_NEW,
2765                                             exp) < 0) {
2766                         if (!refcount_inc_not_zero(&exp->use))
2767                                 continue;
2768                         cb->args[1] = (unsigned long)exp;
2769                         goto out;
2770                 }
2771         }
2772         if (cb->args[1]) {
2773                 cb->args[1] = 0;
2774                 goto restart;
2775         }
2776         cb->args[0] = 1;
2777 out:
2778         rcu_read_unlock();
2779         if (last)
2780                 nf_ct_expect_put(last);
2781
2782         return skb->len;
2783 }
2784
2785 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2786                                  struct sk_buff *skb,
2787                                  const struct nlmsghdr *nlh,
2788                                  const struct nlattr * const cda[],
2789                                  struct netlink_ext_ack *extack)
2790 {
2791         int err;
2792         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2793         u_int8_t u3 = nfmsg->nfgen_family;
2794         struct nf_conntrack_tuple tuple;
2795         struct nf_conntrack_tuple_hash *h;
2796         struct nf_conn *ct;
2797         struct nf_conntrack_zone zone;
2798         struct netlink_dump_control c = {
2799                 .dump = ctnetlink_exp_ct_dump_table,
2800                 .done = ctnetlink_exp_done,
2801         };
2802
2803         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2804                                     u3, NULL);
2805         if (err < 0)
2806                 return err;
2807
2808         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2809         if (err < 0)
2810                 return err;
2811
2812         h = nf_conntrack_find_get(net, &zone, &tuple);
2813         if (!h)
2814                 return -ENOENT;
2815
2816         ct = nf_ct_tuplehash_to_ctrack(h);
2817         /* No expectation linked to this connection tracking. */
2818         if (!nfct_help(ct)) {
2819                 nf_ct_put(ct);
2820                 return 0;
2821         }
2822
2823         c.data = ct;
2824
2825         err = netlink_dump_start(ctnl, skb, nlh, &c);
2826         nf_ct_put(ct);
2827
2828         return err;
2829 }
2830
2831 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
2832                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2833                                 const struct nlattr * const cda[],
2834                                 struct netlink_ext_ack *extack)
2835 {
2836         struct nf_conntrack_tuple tuple;
2837         struct nf_conntrack_expect *exp;
2838         struct sk_buff *skb2;
2839         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2840         u_int8_t u3 = nfmsg->nfgen_family;
2841         struct nf_conntrack_zone zone;
2842         int err;
2843
2844         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2845                 if (cda[CTA_EXPECT_MASTER])
2846                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
2847                                                      extack);
2848                 else {
2849                         struct netlink_dump_control c = {
2850                                 .dump = ctnetlink_exp_dump_table,
2851                                 .done = ctnetlink_exp_done,
2852                         };
2853                         return netlink_dump_start(ctnl, skb, nlh, &c);
2854                 }
2855         }
2856
2857         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2858         if (err < 0)
2859                 return err;
2860
2861         if (cda[CTA_EXPECT_TUPLE])
2862                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2863                                             u3, NULL);
2864         else if (cda[CTA_EXPECT_MASTER])
2865                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2866                                             u3, NULL);
2867         else
2868                 return -EINVAL;
2869
2870         if (err < 0)
2871                 return err;
2872
2873         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2874         if (!exp)
2875                 return -ENOENT;
2876
2877         if (cda[CTA_EXPECT_ID]) {
2878                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2879                 if (ntohl(id) != (u32)(unsigned long)exp) {
2880                         nf_ct_expect_put(exp);
2881                         return -ENOENT;
2882                 }
2883         }
2884
2885         err = -ENOMEM;
2886         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2887         if (skb2 == NULL) {
2888                 nf_ct_expect_put(exp);
2889                 goto out;
2890         }
2891
2892         rcu_read_lock();
2893         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2894                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2895         rcu_read_unlock();
2896         nf_ct_expect_put(exp);
2897         if (err <= 0)
2898                 goto free;
2899
2900         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2901         if (err < 0)
2902                 goto out;
2903
2904         return 0;
2905
2906 free:
2907         kfree_skb(skb2);
2908 out:
2909         /* this avoids a loop in nfnetlink. */
2910         return err == -EAGAIN ? -ENOBUFS : err;
2911 }
2912
2913 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
2914                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2915                                 const struct nlattr * const cda[],
2916                                 struct netlink_ext_ack *extack)
2917 {
2918         struct nf_conntrack_expect *exp;
2919         struct nf_conntrack_tuple tuple;
2920         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2921         struct hlist_node *next;
2922         u_int8_t u3 = nfmsg->nfgen_family;
2923         struct nf_conntrack_zone zone;
2924         unsigned int i;
2925         int err;
2926
2927         if (cda[CTA_EXPECT_TUPLE]) {
2928                 /* delete a single expect by tuple */
2929                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2930                 if (err < 0)
2931                         return err;
2932
2933                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2934                                             u3, NULL);
2935                 if (err < 0)
2936                         return err;
2937
2938                 /* bump usage count to 2 */
2939                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2940                 if (!exp)
2941                         return -ENOENT;
2942
2943                 if (cda[CTA_EXPECT_ID]) {
2944                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2945                         if (ntohl(id) != (u32)(unsigned long)exp) {
2946                                 nf_ct_expect_put(exp);
2947                                 return -ENOENT;
2948                         }
2949                 }
2950
2951                 /* after list removal, usage count == 1 */
2952                 spin_lock_bh(&nf_conntrack_expect_lock);
2953                 if (del_timer(&exp->timeout)) {
2954                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2955                                                    nlmsg_report(nlh));
2956                         nf_ct_expect_put(exp);
2957                 }
2958                 spin_unlock_bh(&nf_conntrack_expect_lock);
2959                 /* have to put what we 'get' above.
2960                  * after this line usage count == 0 */
2961                 nf_ct_expect_put(exp);
2962         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2963                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2964                 struct nf_conn_help *m_help;
2965
2966                 /* delete all expectations for this helper */
2967                 spin_lock_bh(&nf_conntrack_expect_lock);
2968                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2969                         hlist_for_each_entry_safe(exp, next,
2970                                                   &nf_ct_expect_hash[i],
2971                                                   hnode) {
2972
2973                                 if (!net_eq(nf_ct_exp_net(exp), net))
2974                                         continue;
2975
2976                                 m_help = nfct_help(exp->master);
2977                                 if (!strcmp(m_help->helper->name, name) &&
2978                                     del_timer(&exp->timeout)) {
2979                                         nf_ct_unlink_expect_report(exp,
2980                                                         NETLINK_CB(skb).portid,
2981                                                         nlmsg_report(nlh));
2982                                         nf_ct_expect_put(exp);
2983                                 }
2984                         }
2985                 }
2986                 spin_unlock_bh(&nf_conntrack_expect_lock);
2987         } else {
2988                 /* This basically means we have to flush everything*/
2989                 spin_lock_bh(&nf_conntrack_expect_lock);
2990                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2991                         hlist_for_each_entry_safe(exp, next,
2992                                                   &nf_ct_expect_hash[i],
2993                                                   hnode) {
2994
2995                                 if (!net_eq(nf_ct_exp_net(exp), net))
2996                                         continue;
2997
2998                                 if (del_timer(&exp->timeout)) {
2999                                         nf_ct_unlink_expect_report(exp,
3000                                                         NETLINK_CB(skb).portid,
3001                                                         nlmsg_report(nlh));
3002                                         nf_ct_expect_put(exp);
3003                                 }
3004                         }
3005                 }
3006                 spin_unlock_bh(&nf_conntrack_expect_lock);
3007         }
3008
3009         return 0;
3010 }
3011 static int
3012 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3013                         const struct nlattr * const cda[])
3014 {
3015         if (cda[CTA_EXPECT_TIMEOUT]) {
3016                 if (!del_timer(&x->timeout))
3017                         return -ETIME;
3018
3019                 x->timeout.expires = jiffies +
3020                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3021                 add_timer(&x->timeout);
3022         }
3023         return 0;
3024 }
3025
3026 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3027         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3028         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3029 };
3030
3031 static int
3032 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3033                            struct nf_conntrack_expect *exp,
3034                            u_int8_t u3)
3035 {
3036 #ifdef CONFIG_NF_NAT_NEEDED
3037         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3038         struct nf_conntrack_tuple nat_tuple = {};
3039         int err;
3040
3041         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3042                                exp_nat_nla_policy, NULL);
3043         if (err < 0)
3044                 return err;
3045
3046         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3047                 return -EINVAL;
3048
3049         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3050                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3051                                     u3, NULL);
3052         if (err < 0)
3053                 return err;
3054
3055         exp->saved_addr = nat_tuple.src.u3;
3056         exp->saved_proto = nat_tuple.src.u;
3057         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3058
3059         return 0;
3060 #else
3061         return -EOPNOTSUPP;
3062 #endif
3063 }
3064
3065 static struct nf_conntrack_expect *
3066 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3067                        struct nf_conntrack_helper *helper,
3068                        struct nf_conntrack_tuple *tuple,
3069                        struct nf_conntrack_tuple *mask)
3070 {
3071         u_int32_t class = 0;
3072         struct nf_conntrack_expect *exp;
3073         struct nf_conn_help *help;
3074         int err;
3075
3076         help = nfct_help(ct);
3077         if (!help)
3078                 return ERR_PTR(-EOPNOTSUPP);
3079
3080         if (cda[CTA_EXPECT_CLASS] && helper) {
3081                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3082                 if (class > helper->expect_class_max)
3083                         return ERR_PTR(-EINVAL);
3084         }
3085         exp = nf_ct_expect_alloc(ct);
3086         if (!exp)
3087                 return ERR_PTR(-ENOMEM);
3088
3089         if (cda[CTA_EXPECT_FLAGS]) {
3090                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3091                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3092         } else {
3093                 exp->flags = 0;
3094         }
3095         if (cda[CTA_EXPECT_FN]) {
3096                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3097                 struct nf_ct_helper_expectfn *expfn;
3098
3099                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3100                 if (expfn == NULL) {
3101                         err = -EINVAL;
3102                         goto err_out;
3103                 }
3104                 exp->expectfn = expfn->expectfn;
3105         } else
3106                 exp->expectfn = NULL;
3107
3108         exp->class = class;
3109         exp->master = ct;
3110         exp->helper = helper;
3111         exp->tuple = *tuple;
3112         exp->mask.src.u3 = mask->src.u3;
3113         exp->mask.src.u.all = mask->src.u.all;
3114
3115         if (cda[CTA_EXPECT_NAT]) {
3116                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3117                                                  exp, nf_ct_l3num(ct));
3118                 if (err < 0)
3119                         goto err_out;
3120         }
3121         return exp;
3122 err_out:
3123         nf_ct_expect_put(exp);
3124         return ERR_PTR(err);
3125 }
3126
3127 static int
3128 ctnetlink_create_expect(struct net *net,
3129                         const struct nf_conntrack_zone *zone,
3130                         const struct nlattr * const cda[],
3131                         u_int8_t u3, u32 portid, int report)
3132 {
3133         struct nf_conntrack_tuple tuple, mask, master_tuple;
3134         struct nf_conntrack_tuple_hash *h = NULL;
3135         struct nf_conntrack_helper *helper = NULL;
3136         struct nf_conntrack_expect *exp;
3137         struct nf_conn *ct;
3138         int err;
3139
3140         /* caller guarantees that those three CTA_EXPECT_* exist */
3141         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3142                                     u3, NULL);
3143         if (err < 0)
3144                 return err;
3145         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3146                                     u3, NULL);
3147         if (err < 0)
3148                 return err;
3149         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3150                                     u3, NULL);
3151         if (err < 0)
3152                 return err;
3153
3154         /* Look for master conntrack of this expectation */
3155         h = nf_conntrack_find_get(net, zone, &master_tuple);
3156         if (!h)
3157                 return -ENOENT;
3158         ct = nf_ct_tuplehash_to_ctrack(h);
3159
3160         rcu_read_lock();
3161         if (cda[CTA_EXPECT_HELP_NAME]) {
3162                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3163
3164                 helper = __nf_conntrack_helper_find(helpname, u3,
3165                                                     nf_ct_protonum(ct));
3166                 if (helper == NULL) {
3167                         rcu_read_unlock();
3168 #ifdef CONFIG_MODULES
3169                         if (request_module("nfct-helper-%s", helpname) < 0) {
3170                                 err = -EOPNOTSUPP;
3171                                 goto err_ct;
3172                         }
3173                         rcu_read_lock();
3174                         helper = __nf_conntrack_helper_find(helpname, u3,
3175                                                             nf_ct_protonum(ct));
3176                         if (helper) {
3177                                 err = -EAGAIN;
3178                                 goto err_rcu;
3179                         }
3180                         rcu_read_unlock();
3181 #endif
3182                         err = -EOPNOTSUPP;
3183                         goto err_ct;
3184                 }
3185         }
3186
3187         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3188         if (IS_ERR(exp)) {
3189                 err = PTR_ERR(exp);
3190                 goto err_rcu;
3191         }
3192
3193         err = nf_ct_expect_related_report(exp, portid, report);
3194         nf_ct_expect_put(exp);
3195 err_rcu:
3196         rcu_read_unlock();
3197 err_ct:
3198         nf_ct_put(ct);
3199         return err;
3200 }
3201
3202 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3203                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3204                                 const struct nlattr * const cda[],
3205                                 struct netlink_ext_ack *extack)
3206 {
3207         struct nf_conntrack_tuple tuple;
3208         struct nf_conntrack_expect *exp;
3209         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3210         u_int8_t u3 = nfmsg->nfgen_family;
3211         struct nf_conntrack_zone zone;
3212         int err;
3213
3214         if (!cda[CTA_EXPECT_TUPLE]
3215             || !cda[CTA_EXPECT_MASK]
3216             || !cda[CTA_EXPECT_MASTER])
3217                 return -EINVAL;
3218
3219         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3220         if (err < 0)
3221                 return err;
3222
3223         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3224                                     u3, NULL);
3225         if (err < 0)
3226                 return err;
3227
3228         spin_lock_bh(&nf_conntrack_expect_lock);
3229         exp = __nf_ct_expect_find(net, &zone, &tuple);
3230         if (!exp) {
3231                 spin_unlock_bh(&nf_conntrack_expect_lock);
3232                 err = -ENOENT;
3233                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3234                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3235                                                       NETLINK_CB(skb).portid,
3236                                                       nlmsg_report(nlh));
3237                 }
3238                 return err;
3239         }
3240
3241         err = -EEXIST;
3242         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3243                 err = ctnetlink_change_expect(exp, cda);
3244         spin_unlock_bh(&nf_conntrack_expect_lock);
3245
3246         return err;
3247 }
3248
3249 static int
3250 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3251                              const struct ip_conntrack_stat *st)
3252 {
3253         struct nlmsghdr *nlh;
3254         struct nfgenmsg *nfmsg;
3255         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3256
3257         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3258                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3259         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3260         if (nlh == NULL)
3261                 goto nlmsg_failure;
3262
3263         nfmsg = nlmsg_data(nlh);
3264         nfmsg->nfgen_family = AF_UNSPEC;
3265         nfmsg->version      = NFNETLINK_V0;
3266         nfmsg->res_id       = htons(cpu);
3267
3268         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3269             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3270             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3271                 goto nla_put_failure;
3272
3273         nlmsg_end(skb, nlh);
3274         return skb->len;
3275
3276 nla_put_failure:
3277 nlmsg_failure:
3278         nlmsg_cancel(skb, nlh);
3279         return -1;
3280 }
3281
3282 static int
3283 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3284 {
3285         int cpu;
3286         struct net *net = sock_net(skb->sk);
3287
3288         if (cb->args[0] == nr_cpu_ids)
3289                 return 0;
3290
3291         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3292                 const struct ip_conntrack_stat *st;
3293
3294                 if (!cpu_possible(cpu))
3295                         continue;
3296
3297                 st = per_cpu_ptr(net->ct.stat, cpu);
3298                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3299                                                  cb->nlh->nlmsg_seq,
3300                                                  cpu, st) < 0)
3301                         break;
3302         }
3303         cb->args[0] = cpu;
3304
3305         return skb->len;
3306 }
3307
3308 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3309                                   struct sk_buff *skb,
3310                                   const struct nlmsghdr *nlh,
3311                                   const struct nlattr * const cda[],
3312                                   struct netlink_ext_ack *extack)
3313 {
3314         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3315                 struct netlink_dump_control c = {
3316                         .dump = ctnetlink_exp_stat_cpu_dump,
3317                 };
3318                 return netlink_dump_start(ctnl, skb, nlh, &c);
3319         }
3320
3321         return 0;
3322 }
3323
3324 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3325 static struct nf_ct_event_notifier ctnl_notifier = {
3326         .fcn = ctnetlink_conntrack_event,
3327 };
3328
3329 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3330         .fcn = ctnetlink_expect_event,
3331 };
3332 #endif
3333
3334 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3335         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3336                                             .attr_count = CTA_MAX,
3337                                             .policy = ct_nla_policy },
3338         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3339                                             .attr_count = CTA_MAX,
3340                                             .policy = ct_nla_policy },
3341         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3342                                             .attr_count = CTA_MAX,
3343                                             .policy = ct_nla_policy },
3344         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3345                                             .attr_count = CTA_MAX,
3346                                             .policy = ct_nla_policy },
3347         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3348         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3349         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3350         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3351 };
3352
3353 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3354         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3355                                             .attr_count = CTA_EXPECT_MAX,
3356                                             .policy = exp_nla_policy },
3357         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3358                                             .attr_count = CTA_EXPECT_MAX,
3359                                             .policy = exp_nla_policy },
3360         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3361                                             .attr_count = CTA_EXPECT_MAX,
3362                                             .policy = exp_nla_policy },
3363         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3364 };
3365
3366 static const struct nfnetlink_subsystem ctnl_subsys = {
3367         .name                           = "conntrack",
3368         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3369         .cb_count                       = IPCTNL_MSG_MAX,
3370         .cb                             = ctnl_cb,
3371 };
3372
3373 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3374         .name                           = "conntrack_expect",
3375         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3376         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3377         .cb                             = ctnl_exp_cb,
3378 };
3379
3380 MODULE_ALIAS("ip_conntrack_netlink");
3381 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3382 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3383
3384 static int __net_init ctnetlink_net_init(struct net *net)
3385 {
3386 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3387         int ret;
3388
3389         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3390         if (ret < 0) {
3391                 pr_err("ctnetlink_init: cannot register notifier.\n");
3392                 goto err_out;
3393         }
3394
3395         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3396         if (ret < 0) {
3397                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3398                 goto err_unreg_notifier;
3399         }
3400 #endif
3401         return 0;
3402
3403 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3404 err_unreg_notifier:
3405         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3406 err_out:
3407         return ret;
3408 #endif
3409 }
3410
3411 static void ctnetlink_net_exit(struct net *net)
3412 {
3413 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3414         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3415         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3416 #endif
3417 }
3418
3419 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3420 {
3421         struct net *net;
3422
3423         list_for_each_entry(net, net_exit_list, exit_list)
3424                 ctnetlink_net_exit(net);
3425 }
3426
3427 static struct pernet_operations ctnetlink_net_ops = {
3428         .init           = ctnetlink_net_init,
3429         .exit_batch     = ctnetlink_net_exit_batch,
3430 };
3431
3432 static int __init ctnetlink_init(void)
3433 {
3434         int ret;
3435
3436         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3437         ret = nfnetlink_subsys_register(&ctnl_subsys);
3438         if (ret < 0) {
3439                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3440                 goto err_out;
3441         }
3442
3443         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3444         if (ret < 0) {
3445                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3446                 goto err_unreg_subsys;
3447         }
3448
3449         ret = register_pernet_subsys(&ctnetlink_net_ops);
3450         if (ret < 0) {
3451                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3452                 goto err_unreg_exp_subsys;
3453         }
3454 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3455         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3456         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3457 #endif
3458         return 0;
3459
3460 err_unreg_exp_subsys:
3461         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3462 err_unreg_subsys:
3463         nfnetlink_subsys_unregister(&ctnl_subsys);
3464 err_out:
3465         return ret;
3466 }
3467
3468 static void __exit ctnetlink_exit(void)
3469 {
3470         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3471
3472         unregister_pernet_subsys(&ctnetlink_net_ops);
3473         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3474         nfnetlink_subsys_unregister(&ctnl_subsys);
3475 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3476         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3477 #endif
3478         synchronize_rcu();
3479 }
3480
3481 module_init(ctnetlink_init);
3482 module_exit(ctnetlink_exit);