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