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_synproxy.h>
  49 #ifdef CONFIG_NF_NAT_NEEDED
  50 #include <net/netfilter/nf_nat_core.h>
  51 #include <net/netfilter/nf_nat_l4proto.h>
  52 #include <net/netfilter/nf_nat_helper.h>
  53 #endif
  54
  55 #include <linux/netfilter/nfnetlink.h>
  56 #include <linux/netfilter/nfnetlink_conntrack.h>
  57
  58 MODULE_LICENSE("GPL");
  59
  60 static char __initdata version[] = "0.93";
  61
  62 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  63                                 const struct nf_conntrack_tuple *tuple,
  64                                 const struct nf_conntrack_l4proto *l4proto)
  65 {
  66         int ret = 0;
  67         struct nlattr *nest_parms;
  68
  69         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  70         if (!nest_parms)
  71                 goto nla_put_failure;
  72         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  73                 goto nla_put_failure;
  74
  75         if (likely(l4proto->tuple_to_nlattr))
  76                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  77
  78         nla_nest_end(skb, nest_parms);
  79
  80         return ret;
  81
  82 nla_put_failure:
  83         return -1;
  84 }
  85
  86 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  87                                     const struct nf_conntrack_tuple *tuple,
  88                                     const struct nf_conntrack_l3proto *l3proto)
  89 {
  90         int ret = 0;
  91         struct nlattr *nest_parms;
  92
  93         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  94         if (!nest_parms)
  95                 goto nla_put_failure;
  96
  97         if (likely(l3proto->tuple_to_nlattr))
  98                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  99
 100         nla_nest_end(skb, nest_parms);
 101
 102         return ret;
 103
 104 nla_put_failure:
 105         return -1;
 106 }
 107
 108 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 109                                  const struct nf_conntrack_tuple *tuple)
 110 {
 111         const struct nf_conntrack_l3proto *l3proto;
 112         const struct nf_conntrack_l4proto *l4proto;
 113         int ret;
 114
 115         rcu_read_lock();
 116         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 117         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 118
 119         if (ret >= 0) {
 120                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 121                                                tuple->dst.protonum);
 122                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 123         }
 124         rcu_read_unlock();
 125         return ret;
 126 }
 127
 128 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 129                                   const struct nf_conntrack_zone *zone, int dir)
 130 {
 131         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 132                 return 0;
 133         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 134                 goto nla_put_failure;
 135         return 0;
 136
 137 nla_put_failure:
 138         return -1;
 139 }
 140
 141 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 142 {
 143         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 144                 goto nla_put_failure;
 145         return 0;
 146
 147 nla_put_failure:
 148         return -1;
 149 }
 150
 151 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 152 {
 153         long timeout = nf_ct_expires(ct) / HZ;
 154
 155         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 156                 goto nla_put_failure;
 157         return 0;
 158
 159 nla_put_failure:
 160         return -1;
 161 }
 162
 163 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 164 {
 165         const struct nf_conntrack_l4proto *l4proto;
 166         struct nlattr *nest_proto;
 167         int ret;
 168
 169         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 170         if (!l4proto->to_nlattr)
 171                 return 0;
 172
 173         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 174         if (!nest_proto)
 175                 goto nla_put_failure;
 176
 177         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 178
 179         nla_nest_end(skb, nest_proto);
 180
 181         return ret;
 182
 183 nla_put_failure:
 184         return -1;
 185 }
 186
 187 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 188                                    const struct nf_conn *ct)
 189 {
 190         struct nlattr *nest_helper;
 191         const struct nf_conn_help *help = nfct_help(ct);
 192         struct nf_conntrack_helper *helper;
 193
 194         if (!help)
 195                 return 0;
 196
 197         helper = rcu_dereference(help->helper);
 198         if (!helper)
 199                 goto out;
 200
 201         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 202         if (!nest_helper)
 203                 goto nla_put_failure;
 204         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 205                 goto nla_put_failure;
 206
 207         if (helper->to_nlattr)
 208                 helper->to_nlattr(skb, ct);
 209
 210         nla_nest_end(skb, nest_helper);
 211 out:
 212         return 0;
 213
 214 nla_put_failure:
 215         return -1;
 216 }
 217
 218 static int
 219 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 220               enum ip_conntrack_dir dir, int type)
 221 {
 222         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 223         struct nf_conn_counter *counter = acct->counter;
 224         struct nlattr *nest_count;
 225         u64 pkts, bytes;
 226
 227         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 228                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 229                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 230         } else {
 231                 pkts = atomic64_read(&counter[dir].packets);
 232                 bytes = atomic64_read(&counter[dir].bytes);
 233         }
 234
 235         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 236         if (!nest_count)
 237                 goto nla_put_failure;
 238
 239         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 240                          CTA_COUNTERS_PAD) ||
 241             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 242                          CTA_COUNTERS_PAD))
 243                 goto nla_put_failure;
 244
 245         nla_nest_end(skb, nest_count);
 246
 247         return 0;
 248
 249 nla_put_failure:
 250         return -1;
 251 }
 252
 253 static int
 254 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 255 {
 256         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 257
 258         if (!acct)
 259                 return 0;
 260
 261         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 262                 return -1;
 263         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 264                 return -1;
 265
 266         return 0;
 267 }
 268
 269 static int
 270 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 271 {
 272         struct nlattr *nest_count;
 273         const struct nf_conn_tstamp *tstamp;
 274
 275         tstamp = nf_conn_tstamp_find(ct);
 276         if (!tstamp)
 277                 return 0;
 278
 279         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 280         if (!nest_count)
 281                 goto nla_put_failure;
 282
 283         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 284                          CTA_TIMESTAMP_PAD) ||
 285             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 286                                                cpu_to_be64(tstamp->stop),
 287                                                CTA_TIMESTAMP_PAD)))
 288                 goto nla_put_failure;
 289         nla_nest_end(skb, nest_count);
 290
 291         return 0;
 292
 293 nla_put_failure:
 294         return -1;
 295 }
 296
 297 #ifdef CONFIG_NF_CONNTRACK_MARK
 298 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 299 {
 300         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 301                 goto nla_put_failure;
 302         return 0;
 303
 304 nla_put_failure:
 305         return -1;
 306 }
 307 #else
 308 #define ctnetlink_dump_mark(a, b) (0)
 309 #endif
 310
 311 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 312 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 313 {
 314         struct nlattr *nest_secctx;
 315         int len, ret;
 316         char *secctx;
 317
 318         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 319         if (ret)
 320                 return 0;
 321
 322         ret = -1;
 323         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 324         if (!nest_secctx)
 325                 goto nla_put_failure;
 326
 327         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 328                 goto nla_put_failure;
 329         nla_nest_end(skb, nest_secctx);
 330
 331         ret = 0;
 332 nla_put_failure:
 333         security_release_secctx(secctx, len);
 334         return ret;
 335 }
 336 #else
 337 #define ctnetlink_dump_secctx(a, b) (0)
 338 #endif
 339
 340 #ifdef CONFIG_NF_CONNTRACK_LABELS
 341 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 342 {
 343         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 344
 345         if (!labels)
 346                 return 0;
 347         return nla_total_size(sizeof(labels->bits));
 348 }
 349
 350 static int
 351 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 352 {
 353         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 354         unsigned int i;
 355
 356         if (!labels)
 357                 return 0;
 358
 359         i = 0;
 360         do {
 361                 if (labels->bits[i] != 0)
 362                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 363                                        labels->bits);
 364                 i++;
 365         } while (i < ARRAY_SIZE(labels->bits));
 366
 367         return 0;
 368 }
 369 #else
 370 #define ctnetlink_dump_labels(a, b) (0)
 371 #define ctnetlink_label_size(a) (0)
 372 #endif
 373
 374 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 375
 376 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 377 {
 378         struct nlattr *nest_parms;
 379
 380         if (!(ct->status & IPS_EXPECTED))
 381                 return 0;
 382
 383         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 384         if (!nest_parms)
 385                 goto nla_put_failure;
 386         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 387                 goto nla_put_failure;
 388         nla_nest_end(skb, nest_parms);
 389
 390         return 0;
 391
 392 nla_put_failure:
 393         return -1;
 394 }
 395
 396 static int
 397 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 398 {
 399         struct nlattr *nest_parms;
 400
 401         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 402         if (!nest_parms)
 403                 goto nla_put_failure;
 404
 405         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 406                          htonl(seq->correction_pos)) ||
 407             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 408                          htonl(seq->offset_before)) ||
 409             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 410                          htonl(seq->offset_after)))
 411                 goto nla_put_failure;
 412
 413         nla_nest_end(skb, nest_parms);
 414
 415         return 0;
 416
 417 nla_put_failure:
 418         return -1;
 419 }
 420
 421 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 422 {
 423         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 424         struct nf_ct_seqadj *seq;
 425
 426         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 427                 return 0;
 428
 429         spin_lock_bh(&ct->lock);
 430         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 431         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 432                 goto err;
 433
 434         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 435         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 436                 goto err;
 437
 438         spin_unlock_bh(&ct->lock);
 439         return 0;
 440 err:
 441         spin_unlock_bh(&ct->lock);
 442         return -1;
 443 }
 444
 445 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 446 {
 447         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 448                 goto nla_put_failure;
 449         return 0;
 450
 451 nla_put_failure:
 452         return -1;
 453 }
 454
 455 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 456 {
 457         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 458                 goto nla_put_failure;
 459         return 0;
 460
 461 nla_put_failure:
 462         return -1;
 463 }
 464
 465 static int
 466 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 467                     struct nf_conn *ct)
 468 {
 469         const struct nf_conntrack_zone *zone;
 470         struct nlmsghdr *nlh;
 471         struct nfgenmsg *nfmsg;
 472         struct nlattr *nest_parms;
 473         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 474
 475         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 476         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 477         if (nlh == NULL)
 478                 goto nlmsg_failure;
 479
 480         nfmsg = nlmsg_data(nlh);
 481         nfmsg->nfgen_family = nf_ct_l3num(ct);
 482         nfmsg->version      = NFNETLINK_V0;
 483         nfmsg->res_id       = 0;
 484
 485         zone = nf_ct_zone(ct);
 486
 487         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 488         if (!nest_parms)
 489                 goto nla_put_failure;
 490         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 491                 goto nla_put_failure;
 492         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 493                                    NF_CT_ZONE_DIR_ORIG) < 0)
 494                 goto nla_put_failure;
 495         nla_nest_end(skb, nest_parms);
 496
 497         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 498         if (!nest_parms)
 499                 goto nla_put_failure;
 500         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 501                 goto nla_put_failure;
 502         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 503                                    NF_CT_ZONE_DIR_REPL) < 0)
 504                 goto nla_put_failure;
 505         nla_nest_end(skb, nest_parms);
 506
 507         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 508                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 509                 goto nla_put_failure;
 510
 511         if (ctnetlink_dump_status(skb, ct) < 0 ||
 512             ctnetlink_dump_timeout(skb, ct) < 0 ||
 513             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 514             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 515             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 516             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 517             ctnetlink_dump_mark(skb, ct) < 0 ||
 518             ctnetlink_dump_secctx(skb, ct) < 0 ||
 519             ctnetlink_dump_labels(skb, ct) < 0 ||
 520             ctnetlink_dump_id(skb, ct) < 0 ||
 521             ctnetlink_dump_use(skb, ct) < 0 ||
 522             ctnetlink_dump_master(skb, ct) < 0 ||
 523             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 524                 goto nla_put_failure;
 525
 526         nlmsg_end(skb, nlh);
 527         return skb->len;
 528
 529 nlmsg_failure:
 530 nla_put_failure:
 531         nlmsg_cancel(skb, nlh);
 532         return -1;
 533 }
 534
 535 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 536 static 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, len4 = 0;
 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         if (l4proto->nlattr_tuple_size) {
 549                 len4 = l4proto->nlattr_tuple_size();
 550                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 551         }
 552
 553         return len + len4;
 554 }
 555 #endif
 556
 557 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 558 {
 559         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 560                 return 0;
 561         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 562                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 563                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 564                ;
 565 }
 566
 567 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 568 {
 569 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 570         int len, ret;
 571
 572         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 573         if (ret)
 574                 return 0;
 575
 576         return nla_total_size(0) /* CTA_SECCTX */
 577                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 578 #else
 579         return 0;
 580 #endif
 581 }
 582
 583 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 584 {
 585 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 586         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 587                 return 0;
 588         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 589 #else
 590         return 0;
 591 #endif
 592 }
 593
 594 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 595 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 596 {
 597         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 598                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 599                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 600                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 601                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 602                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 603                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 604                + ctnetlink_acct_size(ct)
 605                + ctnetlink_timestamp_size(ct)
 606                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 607                + nla_total_size(0) /* CTA_PROTOINFO */
 608                + nla_total_size(0) /* CTA_HELP */
 609                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 610                + ctnetlink_secctx_size(ct)
 611 #ifdef CONFIG_NF_NAT_NEEDED
 612                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 613                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 614 #endif
 615 #ifdef CONFIG_NF_CONNTRACK_MARK
 616                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 617 #endif
 618 #ifdef CONFIG_NF_CONNTRACK_ZONES
 619                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 620 #endif
 621                + ctnetlink_proto_size(ct)
 622                + ctnetlink_label_size(ct)
 623                ;
 624 }
 625
 626 static int
 627 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 628 {
 629         const struct nf_conntrack_zone *zone;
 630         struct net *net;
 631         struct nlmsghdr *nlh;
 632         struct nfgenmsg *nfmsg;
 633         struct nlattr *nest_parms;
 634         struct nf_conn *ct = item->ct;
 635         struct sk_buff *skb;
 636         unsigned int type;
 637         unsigned int flags = 0, group;
 638         int err;
 639
 640         if (events & (1 << IPCT_DESTROY)) {
 641                 type = IPCTNL_MSG_CT_DELETE;
 642                 group = NFNLGRP_CONNTRACK_DESTROY;
 643         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 644                 type = IPCTNL_MSG_CT_NEW;
 645                 flags = NLM_F_CREATE|NLM_F_EXCL;
 646                 group = NFNLGRP_CONNTRACK_NEW;
 647         } else if (events) {
 648                 type = IPCTNL_MSG_CT_NEW;
 649                 group = NFNLGRP_CONNTRACK_UPDATE;
 650         } else
 651                 return 0;
 652
 653         net = nf_ct_net(ct);
 654         if (!item->report && !nfnetlink_has_listeners(net, group))
 655                 return 0;
 656
 657         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 658         if (skb == NULL)
 659                 goto errout;
 660
 661         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 662         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 663         if (nlh == NULL)
 664                 goto nlmsg_failure;
 665
 666         nfmsg = nlmsg_data(nlh);
 667         nfmsg->nfgen_family = nf_ct_l3num(ct);
 668         nfmsg->version  = NFNETLINK_V0;
 669         nfmsg->res_id   = 0;
 670
 671         zone = nf_ct_zone(ct);
 672
 673         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 674         if (!nest_parms)
 675                 goto nla_put_failure;
 676         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 677                 goto nla_put_failure;
 678         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 679                                    NF_CT_ZONE_DIR_ORIG) < 0)
 680                 goto nla_put_failure;
 681         nla_nest_end(skb, nest_parms);
 682
 683         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 684         if (!nest_parms)
 685                 goto nla_put_failure;
 686         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 687                 goto nla_put_failure;
 688         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 689                                    NF_CT_ZONE_DIR_REPL) < 0)
 690                 goto nla_put_failure;
 691         nla_nest_end(skb, nest_parms);
 692
 693         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 694                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 695                 goto nla_put_failure;
 696
 697         if (ctnetlink_dump_id(skb, ct) < 0)
 698                 goto nla_put_failure;
 699
 700         if (ctnetlink_dump_status(skb, ct) < 0)
 701                 goto nla_put_failure;
 702
 703         if (events & (1 << IPCT_DESTROY)) {
 704                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 705                     ctnetlink_dump_timestamp(skb, ct) < 0)
 706                         goto nla_put_failure;
 707         } else {
 708                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 709                         goto nla_put_failure;
 710
 711                 if (events & (1 << IPCT_PROTOINFO)
 712                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 713                         goto nla_put_failure;
 714
 715                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 716                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 717                         goto nla_put_failure;
 718
 719 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 720                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 721                     && ctnetlink_dump_secctx(skb, ct) < 0)
 722                         goto nla_put_failure;
 723 #endif
 724                 if (events & (1 << IPCT_LABEL) &&
 725                      ctnetlink_dump_labels(skb, ct) < 0)
 726                         goto nla_put_failure;
 727
 728                 if (events & (1 << IPCT_RELATED) &&
 729                     ctnetlink_dump_master(skb, ct) < 0)
 730                         goto nla_put_failure;
 731
 732                 if (events & (1 << IPCT_SEQADJ) &&
 733                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 734                         goto nla_put_failure;
 735         }
 736
 737 #ifdef CONFIG_NF_CONNTRACK_MARK
 738         if ((events & (1 << IPCT_MARK) || ct->mark)
 739             && ctnetlink_dump_mark(skb, ct) < 0)
 740                 goto nla_put_failure;
 741 #endif
 742         nlmsg_end(skb, nlh);
 743         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 744                              GFP_ATOMIC);
 745         if (err == -ENOBUFS || err == -EAGAIN)
 746                 return -ENOBUFS;
 747
 748         return 0;
 749
 750 nla_put_failure:
 751         nlmsg_cancel(skb, nlh);
 752 nlmsg_failure:
 753         kfree_skb(skb);
 754 errout:
 755         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 756                 return -ENOBUFS;
 757
 758         return 0;
 759 }
 760 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 761
 762 static int ctnetlink_done(struct netlink_callback *cb)
 763 {
 764         if (cb->args[1])
 765                 nf_ct_put((struct nf_conn *)cb->args[1]);
 766         kfree(cb->data);
 767         return 0;
 768 }
 769
 770 struct ctnetlink_filter {
 771         struct {
 772                 u_int32_t val;
 773                 u_int32_t mask;
 774         } mark;
 775 };
 776
 777 static struct ctnetlink_filter *
 778 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 779 {
 780 #ifdef CONFIG_NF_CONNTRACK_MARK
 781         struct ctnetlink_filter *filter;
 782
 783         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 784         if (filter == NULL)
 785                 return ERR_PTR(-ENOMEM);
 786
 787         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 788         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 789
 790         return filter;
 791 #else
 792         return ERR_PTR(-EOPNOTSUPP);
 793 #endif
 794 }
 795
 796 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 797 {
 798         struct ctnetlink_filter *filter = data;
 799
 800         if (filter == NULL)
 801                 return 1;
 802
 803 #ifdef CONFIG_NF_CONNTRACK_MARK
 804         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 805                 return 1;
 806 #endif
 807
 808         return 0;
 809 }
 810
 811 static int
 812 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 813 {
 814         struct net *net = sock_net(skb->sk);
 815         struct nf_conn *ct, *last;
 816         struct nf_conntrack_tuple_hash *h;
 817         struct hlist_nulls_node *n;
 818         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 819         u_int8_t l3proto = nfmsg->nfgen_family;
 820         struct nf_conn *nf_ct_evict[8];
 821         int res, i;
 822         spinlock_t *lockp;
 823
 824         last = (struct nf_conn *)cb->args[1];
 825         i = 0;
 826
 827         local_bh_disable();
 828         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 829 restart:
 830                 while (i) {
 831                         i--;
 832                         if (nf_ct_should_gc(nf_ct_evict[i]))
 833                                 nf_ct_kill(nf_ct_evict[i]);
 834                         nf_ct_put(nf_ct_evict[i]);
 835                 }
 836
 837                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 838                 nf_conntrack_lock(lockp);
 839                 if (cb->args[0] >= nf_conntrack_htable_size) {
 840                         spin_unlock(lockp);
 841                         goto out;
 842                 }
 843                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 844                                            hnnode) {
 845                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 846                                 continue;
 847                         ct = nf_ct_tuplehash_to_ctrack(h);
 848                         if (nf_ct_is_expired(ct)) {
 849                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 850                                     atomic_inc_not_zero(&ct->ct_general.use))
 851                                         nf_ct_evict[i++] = ct;
 852                                 continue;
 853                         }
 854
 855                         if (!net_eq(net, nf_ct_net(ct)))
 856                                 continue;
 857
 858                         /* Dump entries of a given L3 protocol number.
 859                          * If it is not specified, ie. l3proto == 0,
 860                          * then dump everything. */
 861                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 862                                 continue;
 863                         if (cb->args[1]) {
 864                                 if (ct != last)
 865                                         continue;
 866                                 cb->args[1] = 0;
 867                         }
 868                         if (!ctnetlink_filter_match(ct, cb->data))
 869                                 continue;
 870
 871                         rcu_read_lock();
 872                         res =
 873                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 874                                             cb->nlh->nlmsg_seq,
 875                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 876                                             ct);
 877                         rcu_read_unlock();
 878                         if (res < 0) {
 879                                 nf_conntrack_get(&ct->ct_general);
 880                                 cb->args[1] = (unsigned long)ct;
 881                                 spin_unlock(lockp);
 882                                 goto out;
 883                         }
 884                 }
 885                 spin_unlock(lockp);
 886                 if (cb->args[1]) {
 887                         cb->args[1] = 0;
 888                         goto restart;
 889                 }
 890         }
 891 out:
 892         local_bh_enable();
 893         if (last) {
 894                 /* nf ct hash resize happened, now clear the leftover. */
 895                 if ((struct nf_conn *)cb->args[1] == last)
 896                         cb->args[1] = 0;
 897
 898                 nf_ct_put(last);
 899         }
 900
 901         while (i) {
 902                 i--;
 903                 if (nf_ct_should_gc(nf_ct_evict[i]))
 904                         nf_ct_kill(nf_ct_evict[i]);
 905                 nf_ct_put(nf_ct_evict[i]);
 906         }
 907
 908         return skb->len;
 909 }
 910
 911 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
 912                                     struct nf_conntrack_tuple *tuple)
 913 {
 914         struct nlattr *tb[CTA_IP_MAX+1];
 915         struct nf_conntrack_l3proto *l3proto;
 916         int ret = 0;
 917
 918         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL, NULL);
 919         if (ret < 0)
 920                 return ret;
 921
 922         rcu_read_lock();
 923         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 924
 925         if (likely(l3proto->nlattr_to_tuple)) {
 926                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 927                                           l3proto->nla_policy, NULL);
 928                 if (ret == 0)
 929                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 930         }
 931
 932         rcu_read_unlock();
 933
 934         return ret;
 935 }
 936
 937 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 938         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 939 };
 940
 941 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
 942                                        struct nf_conntrack_tuple *tuple)
 943 {
 944         const struct nf_conntrack_l4proto *l4proto;
 945         struct nlattr *tb[CTA_PROTO_MAX+1];
 946         int ret = 0;
 947
 948         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy,
 949                                NULL);
 950         if (ret < 0)
 951                 return ret;
 952
 953         if (!tb[CTA_PROTO_NUM])
 954                 return -EINVAL;
 955         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 956
 957         rcu_read_lock();
 958         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 959
 960         if (likely(l4proto->nlattr_to_tuple)) {
 961                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 962                                           l4proto->nla_policy, NULL);
 963                 if (ret == 0)
 964                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 965         }
 966
 967         rcu_read_unlock();
 968
 969         return ret;
 970 }
 971
 972 static int
 973 ctnetlink_parse_zone(const struct nlattr *attr,
 974                      struct nf_conntrack_zone *zone)
 975 {
 976         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
 977                         NF_CT_DEFAULT_ZONE_DIR, 0);
 978 #ifdef CONFIG_NF_CONNTRACK_ZONES
 979         if (attr)
 980                 zone->id = ntohs(nla_get_be16(attr));
 981 #else
 982         if (attr)
 983                 return -EOPNOTSUPP;
 984 #endif
 985         return 0;
 986 }
 987
 988 static int
 989 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
 990                            struct nf_conntrack_zone *zone)
 991 {
 992         int ret;
 993
 994         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
 995                 return -EINVAL;
 996
 997         ret = ctnetlink_parse_zone(attr, zone);
 998         if (ret < 0)
 999                 return ret;
1000
1001         if (type == CTA_TUPLE_REPLY)
1002                 zone->dir = NF_CT_ZONE_DIR_REPL;
1003         else
1004                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1005
1006         return 0;
1007 }
1008
1009 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1010         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1011         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1012         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1013 };
1014
1015 static int
1016 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1017                       struct nf_conntrack_tuple *tuple, u32 type,
1018                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1019 {
1020         struct nlattr *tb[CTA_TUPLE_MAX+1];
1021         int err;
1022
1023         memset(tuple, 0, sizeof(*tuple));
1024
1025         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1026                                NULL);
1027         if (err < 0)
1028                 return err;
1029
1030         if (!tb[CTA_TUPLE_IP])
1031                 return -EINVAL;
1032
1033         tuple->src.l3num = l3num;
1034
1035         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1036         if (err < 0)
1037                 return err;
1038
1039         if (!tb[CTA_TUPLE_PROTO])
1040                 return -EINVAL;
1041
1042         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1043         if (err < 0)
1044                 return err;
1045
1046         if (tb[CTA_TUPLE_ZONE]) {
1047                 if (!zone)
1048                         return -EINVAL;
1049
1050                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1051                                                  type, zone);
1052                 if (err < 0)
1053                         return err;
1054         }
1055
1056         /* orig and expect tuples get DIR_ORIGINAL */
1057         if (type == CTA_TUPLE_REPLY)
1058                 tuple->dst.dir = IP_CT_DIR_REPLY;
1059         else
1060                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1061
1062         return 0;
1063 }
1064
1065 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1066         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1067                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1068 };
1069
1070 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1071                                 struct nlattr **helpinfo)
1072 {
1073         int err;
1074         struct nlattr *tb[CTA_HELP_MAX+1];
1075
1076         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1077         if (err < 0)
1078                 return err;
1079
1080         if (!tb[CTA_HELP_NAME])
1081                 return -EINVAL;
1082
1083         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1084
1085         if (tb[CTA_HELP_INFO])
1086                 *helpinfo = tb[CTA_HELP_INFO];
1087
1088         return 0;
1089 }
1090
1091 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1092         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1093         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1094         [CTA_STATUS]            = { .type = NLA_U32 },
1095         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1096         [CTA_HELP]              = { .type = NLA_NESTED },
1097         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1098         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1099         [CTA_MARK]              = { .type = NLA_U32 },
1100         [CTA_ID]                = { .type = NLA_U32 },
1101         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1102         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1103         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1104         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1105         [CTA_ZONE]              = { .type = NLA_U16 },
1106         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1107         [CTA_LABELS]            = { .type = NLA_BINARY,
1108                                     .len = NF_CT_LABELS_MAX_SIZE },
1109         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1110                                     .len = NF_CT_LABELS_MAX_SIZE },
1111 };
1112
1113 static int ctnetlink_flush_conntrack(struct net *net,
1114                                      const struct nlattr * const cda[],
1115                                      u32 portid, int report)
1116 {
1117         struct ctnetlink_filter *filter = NULL;
1118
1119         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1120                 filter = ctnetlink_alloc_filter(cda);
1121                 if (IS_ERR(filter))
1122                         return PTR_ERR(filter);
1123         }
1124
1125         nf_ct_iterate_cleanup_net(net, ctnetlink_filter_match, filter,
1126                                   portid, report);
1127         kfree(filter);
1128
1129         return 0;
1130 }
1131
1132 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1133                                    struct sk_buff *skb,
1134                                    const struct nlmsghdr *nlh,
1135                                    const struct nlattr * const cda[],
1136                                    struct netlink_ext_ack *extack)
1137 {
1138         struct nf_conntrack_tuple_hash *h;
1139         struct nf_conntrack_tuple tuple;
1140         struct nf_conn *ct;
1141         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1142         u_int8_t u3 = nfmsg->nfgen_family;
1143         struct nf_conntrack_zone zone;
1144         int err;
1145
1146         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1147         if (err < 0)
1148                 return err;
1149
1150         if (cda[CTA_TUPLE_ORIG])
1151                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1152                                             u3, &zone);
1153         else if (cda[CTA_TUPLE_REPLY])
1154                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1155                                             u3, &zone);
1156         else {
1157                 return ctnetlink_flush_conntrack(net, cda,
1158                                                  NETLINK_CB(skb).portid,
1159                                                  nlmsg_report(nlh));
1160         }
1161
1162         if (err < 0)
1163                 return err;
1164
1165         h = nf_conntrack_find_get(net, &zone, &tuple);
1166         if (!h)
1167                 return -ENOENT;
1168
1169         ct = nf_ct_tuplehash_to_ctrack(h);
1170
1171         if (cda[CTA_ID]) {
1172                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1173                 if (id != (u32)(unsigned long)ct) {
1174                         nf_ct_put(ct);
1175                         return -ENOENT;
1176                 }
1177         }
1178
1179         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1180         nf_ct_put(ct);
1181
1182         return 0;
1183 }
1184
1185 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1186                                    struct sk_buff *skb,
1187                                    const struct nlmsghdr *nlh,
1188                                    const struct nlattr * const cda[],
1189                                    struct netlink_ext_ack *extack)
1190 {
1191         struct nf_conntrack_tuple_hash *h;
1192         struct nf_conntrack_tuple tuple;
1193         struct nf_conn *ct;
1194         struct sk_buff *skb2 = NULL;
1195         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1196         u_int8_t u3 = nfmsg->nfgen_family;
1197         struct nf_conntrack_zone zone;
1198         int err;
1199
1200         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1201                 struct netlink_dump_control c = {
1202                         .dump = ctnetlink_dump_table,
1203                         .done = ctnetlink_done,
1204                 };
1205
1206                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1207                         struct ctnetlink_filter *filter;
1208
1209                         filter = ctnetlink_alloc_filter(cda);
1210                         if (IS_ERR(filter))
1211                                 return PTR_ERR(filter);
1212
1213                         c.data = filter;
1214                 }
1215                 return netlink_dump_start(ctnl, skb, nlh, &c);
1216         }
1217
1218         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1219         if (err < 0)
1220                 return err;
1221
1222         if (cda[CTA_TUPLE_ORIG])
1223                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1224                                             u3, &zone);
1225         else if (cda[CTA_TUPLE_REPLY])
1226                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1227                                             u3, &zone);
1228         else
1229                 return -EINVAL;
1230
1231         if (err < 0)
1232                 return err;
1233
1234         h = nf_conntrack_find_get(net, &zone, &tuple);
1235         if (!h)
1236                 return -ENOENT;
1237
1238         ct = nf_ct_tuplehash_to_ctrack(h);
1239
1240         err = -ENOMEM;
1241         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1242         if (skb2 == NULL) {
1243                 nf_ct_put(ct);
1244                 return -ENOMEM;
1245         }
1246
1247         rcu_read_lock();
1248         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1249                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1250         rcu_read_unlock();
1251         nf_ct_put(ct);
1252         if (err <= 0)
1253                 goto free;
1254
1255         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1256         if (err < 0)
1257                 goto out;
1258
1259         return 0;
1260
1261 free:
1262         kfree_skb(skb2);
1263 out:
1264         /* this avoids a loop in nfnetlink. */
1265         return err == -EAGAIN ? -ENOBUFS : err;
1266 }
1267
1268 static int ctnetlink_done_list(struct netlink_callback *cb)
1269 {
1270         if (cb->args[1])
1271                 nf_ct_put((struct nf_conn *)cb->args[1]);
1272         return 0;
1273 }
1274
1275 static int
1276 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1277 {
1278         struct nf_conn *ct, *last;
1279         struct nf_conntrack_tuple_hash *h;
1280         struct hlist_nulls_node *n;
1281         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1282         u_int8_t l3proto = nfmsg->nfgen_family;
1283         int res;
1284         int cpu;
1285         struct hlist_nulls_head *list;
1286         struct net *net = sock_net(skb->sk);
1287
1288         if (cb->args[2])
1289                 return 0;
1290
1291         last = (struct nf_conn *)cb->args[1];
1292
1293         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1294                 struct ct_pcpu *pcpu;
1295
1296                 if (!cpu_possible(cpu))
1297                         continue;
1298
1299                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1300                 spin_lock_bh(&pcpu->lock);
1301                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1302 restart:
1303                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1304                         ct = nf_ct_tuplehash_to_ctrack(h);
1305                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1306                                 continue;
1307                         if (cb->args[1]) {
1308                                 if (ct != last)
1309                                         continue;
1310                                 cb->args[1] = 0;
1311                         }
1312                         rcu_read_lock();
1313                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1314                                                   cb->nlh->nlmsg_seq,
1315                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1316                                                   ct);
1317                         rcu_read_unlock();
1318                         if (res < 0) {
1319                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1320                                         continue;
1321                                 cb->args[0] = cpu;
1322                                 cb->args[1] = (unsigned long)ct;
1323                                 spin_unlock_bh(&pcpu->lock);
1324                                 goto out;
1325                         }
1326                 }
1327                 if (cb->args[1]) {
1328                         cb->args[1] = 0;
1329                         goto restart;
1330                 }
1331                 spin_unlock_bh(&pcpu->lock);
1332         }
1333         cb->args[2] = 1;
1334 out:
1335         if (last)
1336                 nf_ct_put(last);
1337
1338         return skb->len;
1339 }
1340
1341 static int
1342 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1343 {
1344         return ctnetlink_dump_list(skb, cb, true);
1345 }
1346
1347 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1348                                   struct sk_buff *skb,
1349                                   const struct nlmsghdr *nlh,
1350                                   const struct nlattr * const cda[],
1351                                   struct netlink_ext_ack *extack)
1352 {
1353         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1354                 struct netlink_dump_control c = {
1355                         .dump = ctnetlink_dump_dying,
1356                         .done = ctnetlink_done_list,
1357                 };
1358                 return netlink_dump_start(ctnl, skb, nlh, &c);
1359         }
1360
1361         return -EOPNOTSUPP;
1362 }
1363
1364 static int
1365 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1366 {
1367         return ctnetlink_dump_list(skb, cb, false);
1368 }
1369
1370 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1371                                         struct sk_buff *skb,
1372                                         const struct nlmsghdr *nlh,
1373                                         const struct nlattr * const cda[],
1374                                         struct netlink_ext_ack *extack)
1375 {
1376         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1377                 struct netlink_dump_control c = {
1378                         .dump = ctnetlink_dump_unconfirmed,
1379                         .done = ctnetlink_done_list,
1380                 };
1381                 return netlink_dump_start(ctnl, skb, nlh, &c);
1382         }
1383
1384         return -EOPNOTSUPP;
1385 }
1386
1387 #ifdef CONFIG_NF_NAT_NEEDED
1388 static int
1389 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1390                           enum nf_nat_manip_type manip,
1391                           const struct nlattr *attr)
1392 {
1393         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1394         int err;
1395
1396         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1397         if (!parse_nat_setup) {
1398 #ifdef CONFIG_MODULES
1399                 rcu_read_unlock();
1400                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1401                 if (request_module("nf-nat") < 0) {
1402                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1403                         rcu_read_lock();
1404                         return -EOPNOTSUPP;
1405                 }
1406                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1407                 rcu_read_lock();
1408                 if (nfnetlink_parse_nat_setup_hook)
1409                         return -EAGAIN;
1410 #endif
1411                 return -EOPNOTSUPP;
1412         }
1413
1414         err = parse_nat_setup(ct, manip, attr);
1415         if (err == -EAGAIN) {
1416 #ifdef CONFIG_MODULES
1417                 rcu_read_unlock();
1418                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1419                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1420                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1421                         rcu_read_lock();
1422                         return -EOPNOTSUPP;
1423                 }
1424                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1425                 rcu_read_lock();
1426 #else
1427                 err = -EOPNOTSUPP;
1428 #endif
1429         }
1430         return err;
1431 }
1432 #endif
1433
1434 static void
1435 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1436                           unsigned long off)
1437 {
1438         unsigned int bit;
1439
1440         /* Ignore these unchangable bits */
1441         on &= ~IPS_UNCHANGEABLE_MASK;
1442         off &= ~IPS_UNCHANGEABLE_MASK;
1443
1444         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1445                 if (on & (1 << bit))
1446                         set_bit(bit, &ct->status);
1447                 else if (off & (1 << bit))
1448                         clear_bit(bit, &ct->status);
1449         }
1450 }
1451
1452 static int
1453 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1454 {
1455         unsigned long d;
1456         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1457         d = ct->status ^ status;
1458
1459         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1460                 /* unchangeable */
1461                 return -EBUSY;
1462
1463         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1464                 /* SEEN_REPLY bit can only be set */
1465                 return -EBUSY;
1466
1467         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1468                 /* ASSURED bit can only be set */
1469                 return -EBUSY;
1470
1471         __ctnetlink_change_status(ct, status, 0);
1472         return 0;
1473 }
1474
1475 static int
1476 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1477 {
1478 #ifdef CONFIG_NF_NAT_NEEDED
1479         int ret;
1480
1481         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1482                 return 0;
1483
1484         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1485                                         cda[CTA_NAT_DST]);
1486         if (ret < 0)
1487                 return ret;
1488
1489         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1490                                         cda[CTA_NAT_SRC]);
1491         return ret;
1492 #else
1493         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1494                 return 0;
1495         return -EOPNOTSUPP;
1496 #endif
1497 }
1498
1499 static int ctnetlink_change_helper(struct nf_conn *ct,
1500                                    const struct nlattr * const cda[])
1501 {
1502         struct nf_conntrack_helper *helper;
1503         struct nf_conn_help *help = nfct_help(ct);
1504         char *helpname = NULL;
1505         struct nlattr *helpinfo = NULL;
1506         int err;
1507
1508         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1509         if (err < 0)
1510                 return err;
1511
1512         /* don't change helper of sibling connections */
1513         if (ct->master) {
1514                 /* If we try to change the helper to the same thing twice,
1515                  * treat the second attempt as a no-op instead of returning
1516                  * an error.
1517                  */
1518                 err = -EBUSY;
1519                 if (help) {
1520                         rcu_read_lock();
1521                         helper = rcu_dereference(help->helper);
1522                         if (helper && !strcmp(helper->name, helpname))
1523                                 err = 0;
1524                         rcu_read_unlock();
1525                 }
1526
1527                 return err;
1528         }
1529
1530         if (!strcmp(helpname, "")) {
1531                 if (help && help->helper) {
1532                         /* we had a helper before ... */
1533                         nf_ct_remove_expectations(ct);
1534                         RCU_INIT_POINTER(help->helper, NULL);
1535                 }
1536
1537                 return 0;
1538         }
1539
1540         rcu_read_lock();
1541         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1542                                             nf_ct_protonum(ct));
1543         if (helper == NULL) {
1544                 rcu_read_unlock();
1545                 return -EOPNOTSUPP;
1546         }
1547
1548         if (help) {
1549                 if (help->helper == helper) {
1550                         /* update private helper data if allowed. */
1551                         if (helper->from_nlattr)
1552                                 helper->from_nlattr(helpinfo, ct);
1553                         err = 0;
1554                 } else
1555                         err = -EBUSY;
1556         } else {
1557                 /* we cannot set a helper for an existing conntrack */
1558                 err = -EOPNOTSUPP;
1559         }
1560
1561         rcu_read_unlock();
1562         return err;
1563 }
1564
1565 static int ctnetlink_change_timeout(struct nf_conn *ct,
1566                                     const struct nlattr * const cda[])
1567 {
1568         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1569
1570         if (timeout > INT_MAX)
1571                 timeout = INT_MAX;
1572         ct->timeout = nfct_time_stamp + (u32)timeout;
1573
1574         if (test_bit(IPS_DYING_BIT, &ct->status))
1575                 return -ETIME;
1576
1577         return 0;
1578 }
1579
1580 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1581         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1582         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1583         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1584 };
1585
1586 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1587                                       const struct nlattr * const cda[])
1588 {
1589         const struct nlattr *attr = cda[CTA_PROTOINFO];
1590         const struct nf_conntrack_l4proto *l4proto;
1591         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1592         int err = 0;
1593
1594         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1595                                NULL);
1596         if (err < 0)
1597                 return err;
1598
1599         rcu_read_lock();
1600         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1601         if (l4proto->from_nlattr)
1602                 err = l4proto->from_nlattr(tb, ct);
1603         rcu_read_unlock();
1604
1605         return err;
1606 }
1607
1608 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1609         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1610         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1611         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1612 };
1613
1614 static int change_seq_adj(struct nf_ct_seqadj *seq,
1615                           const struct nlattr * const attr)
1616 {
1617         int err;
1618         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1619
1620         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1621         if (err < 0)
1622                 return err;
1623
1624         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1625                 return -EINVAL;
1626
1627         seq->correction_pos =
1628                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1629
1630         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1631                 return -EINVAL;
1632
1633         seq->offset_before =
1634                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1635
1636         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1637                 return -EINVAL;
1638
1639         seq->offset_after =
1640                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1641
1642         return 0;
1643 }
1644
1645 static int
1646 ctnetlink_change_seq_adj(struct nf_conn *ct,
1647                          const struct nlattr * const cda[])
1648 {
1649         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1650         int ret = 0;
1651
1652         if (!seqadj)
1653                 return 0;
1654
1655         spin_lock_bh(&ct->lock);
1656         if (cda[CTA_SEQ_ADJ_ORIG]) {
1657                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1658                                      cda[CTA_SEQ_ADJ_ORIG]);
1659                 if (ret < 0)
1660                         goto err;
1661
1662                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1663         }
1664
1665         if (cda[CTA_SEQ_ADJ_REPLY]) {
1666                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1667                                      cda[CTA_SEQ_ADJ_REPLY]);
1668                 if (ret < 0)
1669                         goto err;
1670
1671                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1672         }
1673
1674         spin_unlock_bh(&ct->lock);
1675         return 0;
1676 err:
1677         spin_unlock_bh(&ct->lock);
1678         return ret;
1679 }
1680
1681 static int
1682 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1683 {
1684 #ifdef CONFIG_NF_CONNTRACK_LABELS
1685         size_t len = nla_len(cda[CTA_LABELS]);
1686         const void *mask = cda[CTA_LABELS_MASK];
1687
1688         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1689                 return -EINVAL;
1690
1691         if (mask) {
1692                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1693                     nla_len(cda[CTA_LABELS_MASK]) != len)
1694                         return -EINVAL;
1695                 mask = nla_data(cda[CTA_LABELS_MASK]);
1696         }
1697
1698         len /= sizeof(u32);
1699
1700         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1701 #else
1702         return -EOPNOTSUPP;
1703 #endif
1704 }
1705
1706 static int
1707 ctnetlink_change_conntrack(struct nf_conn *ct,
1708                            const struct nlattr * const cda[])
1709 {
1710         int err;
1711
1712         /* only allow NAT changes and master assignation for new conntracks */
1713         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1714                 return -EOPNOTSUPP;
1715
1716         if (cda[CTA_HELP]) {
1717                 err = ctnetlink_change_helper(ct, cda);
1718                 if (err < 0)
1719                         return err;
1720         }
1721
1722         if (cda[CTA_TIMEOUT]) {
1723                 err = ctnetlink_change_timeout(ct, cda);
1724                 if (err < 0)
1725                         return err;
1726         }
1727
1728         if (cda[CTA_STATUS]) {
1729                 err = ctnetlink_change_status(ct, cda);
1730                 if (err < 0)
1731                         return err;
1732         }
1733
1734         if (cda[CTA_PROTOINFO]) {
1735                 err = ctnetlink_change_protoinfo(ct, cda);
1736                 if (err < 0)
1737                         return err;
1738         }
1739
1740 #if defined(CONFIG_NF_CONNTRACK_MARK)
1741         if (cda[CTA_MARK])
1742                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1743 #endif
1744
1745         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1746                 err = ctnetlink_change_seq_adj(ct, cda);
1747                 if (err < 0)
1748                         return err;
1749         }
1750
1751         if (cda[CTA_LABELS]) {
1752                 err = ctnetlink_attach_labels(ct, cda);
1753                 if (err < 0)
1754                         return err;
1755         }
1756
1757         return 0;
1758 }
1759
1760 static struct nf_conn *
1761 ctnetlink_create_conntrack(struct net *net,
1762                            const struct nf_conntrack_zone *zone,
1763                            const struct nlattr * const cda[],
1764                            struct nf_conntrack_tuple *otuple,
1765                            struct nf_conntrack_tuple *rtuple,
1766                            u8 u3)
1767 {
1768         struct nf_conn *ct;
1769         int err = -EINVAL;
1770         struct nf_conntrack_helper *helper;
1771         struct nf_conn_tstamp *tstamp;
1772         u64 timeout;
1773
1774         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1775         if (IS_ERR(ct))
1776                 return ERR_PTR(-ENOMEM);
1777
1778         if (!cda[CTA_TIMEOUT])
1779                 goto err1;
1780
1781         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1782         if (timeout > INT_MAX)
1783                 timeout = INT_MAX;
1784         ct->timeout = (u32)timeout + nfct_time_stamp;
1785
1786         rcu_read_lock();
1787         if (cda[CTA_HELP]) {
1788                 char *helpname = NULL;
1789                 struct nlattr *helpinfo = NULL;
1790
1791                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1792                 if (err < 0)
1793                         goto err2;
1794
1795                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1796                                                     nf_ct_protonum(ct));
1797                 if (helper == NULL) {
1798                         rcu_read_unlock();
1799 #ifdef CONFIG_MODULES
1800                         if (request_module("nfct-helper-%s", helpname) < 0) {
1801                                 err = -EOPNOTSUPP;
1802                                 goto err1;
1803                         }
1804
1805                         rcu_read_lock();
1806                         helper = __nf_conntrack_helper_find(helpname,
1807                                                             nf_ct_l3num(ct),
1808                                                             nf_ct_protonum(ct));
1809                         if (helper) {
1810                                 err = -EAGAIN;
1811                                 goto err2;
1812                         }
1813                         rcu_read_unlock();
1814 #endif
1815                         err = -EOPNOTSUPP;
1816                         goto err1;
1817                 } else {
1818                         struct nf_conn_help *help;
1819
1820                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1821                         if (help == NULL) {
1822                                 err = -ENOMEM;
1823                                 goto err2;
1824                         }
1825                         /* set private helper data if allowed. */
1826                         if (helper->from_nlattr)
1827                                 helper->from_nlattr(helpinfo, ct);
1828
1829                         /* not in hash table yet so not strictly necessary */
1830                         RCU_INIT_POINTER(help->helper, helper);
1831                 }
1832         } else {
1833                 /* try an implicit helper assignation */
1834                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1835                 if (err < 0)
1836                         goto err2;
1837         }
1838
1839         err = ctnetlink_setup_nat(ct, cda);
1840         if (err < 0)
1841                 goto err2;
1842
1843         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1844         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1845         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1846         nf_ct_labels_ext_add(ct);
1847         nfct_seqadj_ext_add(ct);
1848         nfct_synproxy_ext_add(ct);
1849
1850         /* we must add conntrack extensions before confirmation. */
1851         ct->status |= IPS_CONFIRMED;
1852
1853         if (cda[CTA_STATUS]) {
1854                 err = ctnetlink_change_status(ct, cda);
1855                 if (err < 0)
1856                         goto err2;
1857         }
1858
1859         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1860                 err = ctnetlink_change_seq_adj(ct, cda);
1861                 if (err < 0)
1862                         goto err2;
1863         }
1864
1865         memset(&ct->proto, 0, sizeof(ct->proto));
1866         if (cda[CTA_PROTOINFO]) {
1867                 err = ctnetlink_change_protoinfo(ct, cda);
1868                 if (err < 0)
1869                         goto err2;
1870         }
1871
1872 #if defined(CONFIG_NF_CONNTRACK_MARK)
1873         if (cda[CTA_MARK])
1874                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1875 #endif
1876
1877         /* setup master conntrack: this is a confirmed expectation */
1878         if (cda[CTA_TUPLE_MASTER]) {
1879                 struct nf_conntrack_tuple master;
1880                 struct nf_conntrack_tuple_hash *master_h;
1881                 struct nf_conn *master_ct;
1882
1883                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1884                                             u3, NULL);
1885                 if (err < 0)
1886                         goto err2;
1887
1888                 master_h = nf_conntrack_find_get(net, zone, &master);
1889                 if (master_h == NULL) {
1890                         err = -ENOENT;
1891                         goto err2;
1892                 }
1893                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1894                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1895                 ct->master = master_ct;
1896         }
1897         tstamp = nf_conn_tstamp_find(ct);
1898         if (tstamp)
1899                 tstamp->start = ktime_get_real_ns();
1900
1901         err = nf_conntrack_hash_check_insert(ct);
1902         if (err < 0)
1903                 goto err2;
1904
1905         rcu_read_unlock();
1906
1907         return ct;
1908
1909 err2:
1910         rcu_read_unlock();
1911 err1:
1912         nf_conntrack_free(ct);
1913         return ERR_PTR(err);
1914 }
1915
1916 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
1917                                    struct sk_buff *skb,
1918                                    const struct nlmsghdr *nlh,
1919                                    const struct nlattr * const cda[],
1920                                    struct netlink_ext_ack *extack)
1921 {
1922         struct nf_conntrack_tuple otuple, rtuple;
1923         struct nf_conntrack_tuple_hash *h = NULL;
1924         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1925         struct nf_conn *ct;
1926         u_int8_t u3 = nfmsg->nfgen_family;
1927         struct nf_conntrack_zone zone;
1928         int err;
1929
1930         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1931         if (err < 0)
1932                 return err;
1933
1934         if (cda[CTA_TUPLE_ORIG]) {
1935                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1936                                             u3, &zone);
1937                 if (err < 0)
1938                         return err;
1939         }
1940
1941         if (cda[CTA_TUPLE_REPLY]) {
1942                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1943                                             u3, &zone);
1944                 if (err < 0)
1945                         return err;
1946         }
1947
1948         if (cda[CTA_TUPLE_ORIG])
1949                 h = nf_conntrack_find_get(net, &zone, &otuple);
1950         else if (cda[CTA_TUPLE_REPLY])
1951                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1952
1953         if (h == NULL) {
1954                 err = -ENOENT;
1955                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1956                         enum ip_conntrack_events events;
1957
1958                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1959                                 return -EINVAL;
1960                         if (otuple.dst.protonum != rtuple.dst.protonum)
1961                                 return -EINVAL;
1962
1963                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1964                                                         &rtuple, u3);
1965                         if (IS_ERR(ct))
1966                                 return PTR_ERR(ct);
1967
1968                         err = 0;
1969                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1970                                 events = 1 << IPCT_RELATED;
1971                         else
1972                                 events = 1 << IPCT_NEW;
1973
1974                         if (cda[CTA_LABELS] &&
1975                             ctnetlink_attach_labels(ct, cda) == 0)
1976                                 events |= (1 << IPCT_LABEL);
1977
1978                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1979                                                       (1 << IPCT_ASSURED) |
1980                                                       (1 << IPCT_HELPER) |
1981                                                       (1 << IPCT_PROTOINFO) |
1982                                                       (1 << IPCT_SEQADJ) |
1983                                                       (1 << IPCT_MARK) | events,
1984                                                       ct, NETLINK_CB(skb).portid,
1985                                                       nlmsg_report(nlh));
1986                         nf_ct_put(ct);
1987                 }
1988
1989                 return err;
1990         }
1991         /* implicit 'else' */
1992
1993         err = -EEXIST;
1994         ct = nf_ct_tuplehash_to_ctrack(h);
1995         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1996                 err = ctnetlink_change_conntrack(ct, cda);
1997                 if (err == 0) {
1998                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1999                                                       (1 << IPCT_ASSURED) |
2000                                                       (1 << IPCT_HELPER) |
2001                                                       (1 << IPCT_LABEL) |
2002                                                       (1 << IPCT_PROTOINFO) |
2003                                                       (1 << IPCT_SEQADJ) |
2004                                                       (1 << IPCT_MARK),
2005                                                       ct, NETLINK_CB(skb).portid,
2006                                                       nlmsg_report(nlh));
2007                 }
2008         }
2009
2010         nf_ct_put(ct);
2011         return err;
2012 }
2013
2014 static int
2015 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2016                                 __u16 cpu, const struct ip_conntrack_stat *st)
2017 {
2018         struct nlmsghdr *nlh;
2019         struct nfgenmsg *nfmsg;
2020         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2021
2022         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2023                               IPCTNL_MSG_CT_GET_STATS_CPU);
2024         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2025         if (nlh == NULL)
2026                 goto nlmsg_failure;
2027
2028         nfmsg = nlmsg_data(nlh);
2029         nfmsg->nfgen_family = AF_UNSPEC;
2030         nfmsg->version      = NFNETLINK_V0;
2031         nfmsg->res_id       = htons(cpu);
2032
2033         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2034             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2035             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2036             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2037             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2038                                 htonl(st->insert_failed)) ||
2039             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2040             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2041             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2042             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2043                                 htonl(st->search_restart)))
2044                 goto nla_put_failure;
2045
2046         nlmsg_end(skb, nlh);
2047         return skb->len;
2048
2049 nla_put_failure:
2050 nlmsg_failure:
2051         nlmsg_cancel(skb, nlh);
2052         return -1;
2053 }
2054
2055 static int
2056 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2057 {
2058         int cpu;
2059         struct net *net = sock_net(skb->sk);
2060
2061         if (cb->args[0] == nr_cpu_ids)
2062                 return 0;
2063
2064         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2065                 const struct ip_conntrack_stat *st;
2066
2067                 if (!cpu_possible(cpu))
2068                         continue;
2069
2070                 st = per_cpu_ptr(net->ct.stat, cpu);
2071                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2072                                                     NETLINK_CB(cb->skb).portid,
2073                                                     cb->nlh->nlmsg_seq,
2074                                                     cpu, st) < 0)
2075                                 break;
2076         }
2077         cb->args[0] = cpu;
2078
2079         return skb->len;
2080 }
2081
2082 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2083                                  struct sk_buff *skb,
2084                                  const struct nlmsghdr *nlh,
2085                                  const struct nlattr * const cda[],
2086                                  struct netlink_ext_ack *extack)
2087 {
2088         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2089                 struct netlink_dump_control c = {
2090                         .dump = ctnetlink_ct_stat_cpu_dump,
2091                 };
2092                 return netlink_dump_start(ctnl, skb, nlh, &c);
2093         }
2094
2095         return 0;
2096 }
2097
2098 static int
2099 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2100                             struct net *net)
2101 {
2102         struct nlmsghdr *nlh;
2103         struct nfgenmsg *nfmsg;
2104         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2105         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2106
2107         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2108         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2109         if (nlh == NULL)
2110                 goto nlmsg_failure;
2111
2112         nfmsg = nlmsg_data(nlh);
2113         nfmsg->nfgen_family = AF_UNSPEC;
2114         nfmsg->version      = NFNETLINK_V0;
2115         nfmsg->res_id       = 0;
2116
2117         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2118                 goto nla_put_failure;
2119
2120         nlmsg_end(skb, nlh);
2121         return skb->len;
2122
2123 nla_put_failure:
2124 nlmsg_failure:
2125         nlmsg_cancel(skb, nlh);
2126         return -1;
2127 }
2128
2129 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2130                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2131                              const struct nlattr * const cda[],
2132                              struct netlink_ext_ack *extack)
2133 {
2134         struct sk_buff *skb2;
2135         int err;
2136
2137         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2138         if (skb2 == NULL)
2139                 return -ENOMEM;
2140
2141         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2142                                           nlh->nlmsg_seq,
2143                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2144                                           sock_net(skb->sk));
2145         if (err <= 0)
2146                 goto free;
2147
2148         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2149         if (err < 0)
2150                 goto out;
2151
2152         return 0;
2153
2154 free:
2155         kfree_skb(skb2);
2156 out:
2157         /* this avoids a loop in nfnetlink. */
2158         return err == -EAGAIN ? -ENOBUFS : err;
2159 }
2160
2161 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2162         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2163         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2164         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2165         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2166         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2167         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2168                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2169         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2170         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2171         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2172         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2173         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2174 };
2175
2176 static struct nf_conntrack_expect *
2177 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2178                        struct nf_conntrack_helper *helper,
2179                        struct nf_conntrack_tuple *tuple,
2180                        struct nf_conntrack_tuple *mask);
2181
2182 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2183 static size_t
2184 ctnetlink_glue_build_size(const struct nf_conn *ct)
2185 {
2186         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2187                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2188                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2189                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2190                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2191                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2192                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2193                + nla_total_size(0) /* CTA_PROTOINFO */
2194                + nla_total_size(0) /* CTA_HELP */
2195                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2196                + ctnetlink_secctx_size(ct)
2197 #ifdef CONFIG_NF_NAT_NEEDED
2198                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2199                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2200 #endif
2201 #ifdef CONFIG_NF_CONNTRACK_MARK
2202                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2203 #endif
2204 #ifdef CONFIG_NF_CONNTRACK_ZONES
2205                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2206 #endif
2207                + ctnetlink_proto_size(ct)
2208                ;
2209 }
2210
2211 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2212                                              enum ip_conntrack_info *ctinfo)
2213 {
2214         return nf_ct_get(skb, ctinfo);
2215 }
2216
2217 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2218 {
2219         const struct nf_conntrack_zone *zone;
2220         struct nlattr *nest_parms;
2221
2222         zone = nf_ct_zone(ct);
2223
2224         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2225         if (!nest_parms)
2226                 goto nla_put_failure;
2227         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2228                 goto nla_put_failure;
2229         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2230                                    NF_CT_ZONE_DIR_ORIG) < 0)
2231                 goto nla_put_failure;
2232         nla_nest_end(skb, nest_parms);
2233
2234         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2235         if (!nest_parms)
2236                 goto nla_put_failure;
2237         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2238                 goto nla_put_failure;
2239         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2240                                    NF_CT_ZONE_DIR_REPL) < 0)
2241                 goto nla_put_failure;
2242         nla_nest_end(skb, nest_parms);
2243
2244         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2245                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2246                 goto nla_put_failure;
2247
2248         if (ctnetlink_dump_id(skb, ct) < 0)
2249                 goto nla_put_failure;
2250
2251         if (ctnetlink_dump_status(skb, ct) < 0)
2252                 goto nla_put_failure;
2253
2254         if (ctnetlink_dump_timeout(skb, ct) < 0)
2255                 goto nla_put_failure;
2256
2257         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2258                 goto nla_put_failure;
2259
2260         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2261                 goto nla_put_failure;
2262
2263 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2264         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2265                 goto nla_put_failure;
2266 #endif
2267         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2268                 goto nla_put_failure;
2269
2270         if ((ct->status & IPS_SEQ_ADJUST) &&
2271             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2272                 goto nla_put_failure;
2273
2274 #ifdef CONFIG_NF_CONNTRACK_MARK
2275         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2276                 goto nla_put_failure;
2277 #endif
2278         if (ctnetlink_dump_labels(skb, ct) < 0)
2279                 goto nla_put_failure;
2280         return 0;
2281
2282 nla_put_failure:
2283         return -ENOSPC;
2284 }
2285
2286 static int
2287 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2288                      enum ip_conntrack_info ctinfo,
2289                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2290 {
2291         struct nlattr *nest_parms;
2292
2293         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2294         if (!nest_parms)
2295                 goto nla_put_failure;
2296
2297         if (__ctnetlink_glue_build(skb, ct) < 0)
2298                 goto nla_put_failure;
2299
2300         nla_nest_end(skb, nest_parms);
2301
2302         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2303                 goto nla_put_failure;
2304
2305         return 0;
2306
2307 nla_put_failure:
2308         return -ENOSPC;
2309 }
2310
2311 static int
2312 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2313 {
2314         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2315         unsigned long d = ct->status ^ status;
2316
2317         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2318                 /* SEEN_REPLY bit can only be set */
2319                 return -EBUSY;
2320
2321         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2322                 /* ASSURED bit can only be set */
2323                 return -EBUSY;
2324
2325         /* This check is less strict than ctnetlink_change_status()
2326          * because callers often flip IPS_EXPECTED bits when sending
2327          * an NFQA_CT attribute to the kernel.  So ignore the
2328          * unchangeable bits but do not error out. Also user programs
2329          * are allowed to clear the bits that they are allowed to change.
2330          */
2331         __ctnetlink_change_status(ct, status, ~status);
2332         return 0;
2333 }
2334
2335 static int
2336 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2337 {
2338         int err;
2339
2340         if (cda[CTA_TIMEOUT]) {
2341                 err = ctnetlink_change_timeout(ct, cda);
2342                 if (err < 0)
2343                         return err;
2344         }
2345         if (cda[CTA_STATUS]) {
2346                 err = ctnetlink_update_status(ct, cda);
2347                 if (err < 0)
2348                         return err;
2349         }
2350         if (cda[CTA_HELP]) {
2351                 err = ctnetlink_change_helper(ct, cda);
2352                 if (err < 0)
2353                         return err;
2354         }
2355         if (cda[CTA_LABELS]) {
2356                 err = ctnetlink_attach_labels(ct, cda);
2357                 if (err < 0)
2358                         return err;
2359         }
2360 #if defined(CONFIG_NF_CONNTRACK_MARK)
2361         if (cda[CTA_MARK]) {
2362                 u32 mask = 0, mark, newmark;
2363                 if (cda[CTA_MARK_MASK])
2364                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2365
2366                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2367                 newmark = (ct->mark & mask) ^ mark;
2368                 if (newmark != ct->mark)
2369                         ct->mark = newmark;
2370         }
2371 #endif
2372         return 0;
2373 }
2374
2375 static int
2376 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2377 {
2378         struct nlattr *cda[CTA_MAX+1];
2379         int ret;
2380
2381         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2382         if (ret < 0)
2383                 return ret;
2384
2385         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2386 }
2387
2388 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2389                                     const struct nf_conn *ct,
2390                                     struct nf_conntrack_tuple *tuple,
2391                                     struct nf_conntrack_tuple *mask)
2392 {
2393         int err;
2394
2395         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2396                                     nf_ct_l3num(ct), NULL);
2397         if (err < 0)
2398                 return err;
2399
2400         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2401                                      nf_ct_l3num(ct), NULL);
2402 }
2403
2404 static int
2405 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2406                              u32 portid, u32 report)
2407 {
2408         struct nlattr *cda[CTA_EXPECT_MAX+1];
2409         struct nf_conntrack_tuple tuple, mask;
2410         struct nf_conntrack_helper *helper = NULL;
2411         struct nf_conntrack_expect *exp;
2412         int err;
2413
2414         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2415                                NULL);
2416         if (err < 0)
2417                 return err;
2418
2419         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2420                                        ct, &tuple, &mask);
2421         if (err < 0)
2422                 return err;
2423
2424         if (cda[CTA_EXPECT_HELP_NAME]) {
2425                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2426
2427                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2428                                                     nf_ct_protonum(ct));
2429                 if (helper == NULL)
2430                         return -EOPNOTSUPP;
2431         }
2432
2433         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2434                                      helper, &tuple, &mask);
2435         if (IS_ERR(exp))
2436                 return PTR_ERR(exp);
2437
2438         err = nf_ct_expect_related_report(exp, portid, report);
2439         nf_ct_expect_put(exp);
2440         return err;
2441 }
2442
2443 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2444                                   enum ip_conntrack_info ctinfo, int diff)
2445 {
2446         if (!(ct->status & IPS_NAT_MASK))
2447                 return;
2448
2449         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2450 }
2451
2452 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2453         .get_ct         = ctnetlink_glue_get_ct,
2454         .build_size     = ctnetlink_glue_build_size,
2455         .build          = ctnetlink_glue_build,
2456         .parse          = ctnetlink_glue_parse,
2457         .attach_expect  = ctnetlink_glue_attach_expect,
2458         .seq_adjust     = ctnetlink_glue_seqadj,
2459 };
2460 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2461
2462 /***********************************************************************
2463  * EXPECT
2464  ***********************************************************************/
2465
2466 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2467                                     const struct nf_conntrack_tuple *tuple,
2468                                     u32 type)
2469 {
2470         struct nlattr *nest_parms;
2471
2472         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2473         if (!nest_parms)
2474                 goto nla_put_failure;
2475         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2476                 goto nla_put_failure;
2477         nla_nest_end(skb, nest_parms);
2478
2479         return 0;
2480
2481 nla_put_failure:
2482         return -1;
2483 }
2484
2485 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2486                                    const struct nf_conntrack_tuple *tuple,
2487                                    const struct nf_conntrack_tuple_mask *mask)
2488 {
2489         const struct nf_conntrack_l3proto *l3proto;
2490         const struct nf_conntrack_l4proto *l4proto;
2491         struct nf_conntrack_tuple m;
2492         struct nlattr *nest_parms;
2493         int ret;
2494
2495         memset(&m, 0xFF, sizeof(m));
2496         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2497         m.src.u.all = mask->src.u.all;
2498         m.dst.protonum = tuple->dst.protonum;
2499
2500         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2501         if (!nest_parms)
2502                 goto nla_put_failure;
2503
2504         rcu_read_lock();
2505         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2506         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2507         if (ret >= 0) {
2508                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2509                                                tuple->dst.protonum);
2510         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2511         }
2512         rcu_read_unlock();
2513
2514         if (unlikely(ret < 0))
2515                 goto nla_put_failure;
2516
2517         nla_nest_end(skb, nest_parms);
2518
2519         return 0;
2520
2521 nla_put_failure:
2522         return -1;
2523 }
2524
2525 static const union nf_inet_addr any_addr;
2526
2527 static int
2528 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2529                           const struct nf_conntrack_expect *exp)
2530 {
2531         struct nf_conn *master = exp->master;
2532         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2533         struct nf_conn_help *help;
2534 #ifdef CONFIG_NF_NAT_NEEDED
2535         struct nlattr *nest_parms;
2536         struct nf_conntrack_tuple nat_tuple = {};
2537 #endif
2538         struct nf_ct_helper_expectfn *expfn;
2539
2540         if (timeout < 0)
2541                 timeout = 0;
2542
2543         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2544                 goto nla_put_failure;
2545         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2546                 goto nla_put_failure;
2547         if (ctnetlink_exp_dump_tuple(skb,
2548                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2549                                  CTA_EXPECT_MASTER) < 0)
2550                 goto nla_put_failure;
2551
2552 #ifdef CONFIG_NF_NAT_NEEDED
2553         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2554             exp->saved_proto.all) {
2555                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2556                 if (!nest_parms)
2557                         goto nla_put_failure;
2558
2559                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2560                         goto nla_put_failure;
2561
2562                 nat_tuple.src.l3num = nf_ct_l3num(master);
2563                 nat_tuple.src.u3 = exp->saved_addr;
2564                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2565                 nat_tuple.src.u = exp->saved_proto;
2566
2567                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2568                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2569                         goto nla_put_failure;
2570                 nla_nest_end(skb, nest_parms);
2571         }
2572 #endif
2573         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2574             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2575             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2576             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2577                 goto nla_put_failure;
2578         help = nfct_help(master);
2579         if (help) {
2580                 struct nf_conntrack_helper *helper;
2581
2582                 helper = rcu_dereference(help->helper);
2583                 if (helper &&
2584                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2585                         goto nla_put_failure;
2586         }
2587         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2588         if (expfn != NULL &&
2589             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2590                 goto nla_put_failure;
2591
2592         return 0;
2593
2594 nla_put_failure:
2595         return -1;
2596 }
2597
2598 static int
2599 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2600                         int event, const struct nf_conntrack_expect *exp)
2601 {
2602         struct nlmsghdr *nlh;
2603         struct nfgenmsg *nfmsg;
2604         unsigned int flags = portid ? NLM_F_MULTI : 0;
2605
2606         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2607         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2608         if (nlh == NULL)
2609                 goto nlmsg_failure;
2610
2611         nfmsg = nlmsg_data(nlh);
2612         nfmsg->nfgen_family = exp->tuple.src.l3num;
2613         nfmsg->version      = NFNETLINK_V0;
2614         nfmsg->res_id       = 0;
2615
2616         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2617                 goto nla_put_failure;
2618
2619         nlmsg_end(skb, nlh);
2620         return skb->len;
2621
2622 nlmsg_failure:
2623 nla_put_failure:
2624         nlmsg_cancel(skb, nlh);
2625         return -1;
2626 }
2627
2628 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2629 static int
2630 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2631 {
2632         struct nf_conntrack_expect *exp = item->exp;
2633         struct net *net = nf_ct_exp_net(exp);
2634         struct nlmsghdr *nlh;
2635         struct nfgenmsg *nfmsg;
2636         struct sk_buff *skb;
2637         unsigned int type, group;
2638         int flags = 0;
2639
2640         if (events & (1 << IPEXP_DESTROY)) {
2641                 type = IPCTNL_MSG_EXP_DELETE;
2642                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2643         } else if (events & (1 << IPEXP_NEW)) {
2644                 type = IPCTNL_MSG_EXP_NEW;
2645                 flags = NLM_F_CREATE|NLM_F_EXCL;
2646                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2647         } else
2648                 return 0;
2649
2650         if (!item->report && !nfnetlink_has_listeners(net, group))
2651                 return 0;
2652
2653         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2654         if (skb == NULL)
2655                 goto errout;
2656
2657         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2658         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2659         if (nlh == NULL)
2660                 goto nlmsg_failure;
2661
2662         nfmsg = nlmsg_data(nlh);
2663         nfmsg->nfgen_family = exp->tuple.src.l3num;
2664         nfmsg->version      = NFNETLINK_V0;
2665         nfmsg->res_id       = 0;
2666
2667         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2668                 goto nla_put_failure;
2669
2670         nlmsg_end(skb, nlh);
2671         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2672         return 0;
2673
2674 nla_put_failure:
2675         nlmsg_cancel(skb, nlh);
2676 nlmsg_failure:
2677         kfree_skb(skb);
2678 errout:
2679         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2680         return 0;
2681 }
2682 #endif
2683 static int ctnetlink_exp_done(struct netlink_callback *cb)
2684 {
2685         if (cb->args[1])
2686                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2687         return 0;
2688 }
2689
2690 static int
2691 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2692 {
2693         struct net *net = sock_net(skb->sk);
2694         struct nf_conntrack_expect *exp, *last;
2695         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2696         u_int8_t l3proto = nfmsg->nfgen_family;
2697
2698         rcu_read_lock();
2699         last = (struct nf_conntrack_expect *)cb->args[1];
2700         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2701 restart:
2702                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2703                                          hnode) {
2704                         if (l3proto && exp->tuple.src.l3num != l3proto)
2705                                 continue;
2706
2707                         if (!net_eq(nf_ct_net(exp->master), net))
2708                                 continue;
2709
2710                         if (cb->args[1]) {
2711                                 if (exp != last)
2712                                         continue;
2713                                 cb->args[1] = 0;
2714                         }
2715                         if (ctnetlink_exp_fill_info(skb,
2716                                                     NETLINK_CB(cb->skb).portid,
2717                                                     cb->nlh->nlmsg_seq,
2718                                                     IPCTNL_MSG_EXP_NEW,
2719                                                     exp) < 0) {
2720                                 if (!refcount_inc_not_zero(&exp->use))
2721                                         continue;
2722                                 cb->args[1] = (unsigned long)exp;
2723                                 goto out;
2724                         }
2725                 }
2726                 if (cb->args[1]) {
2727                         cb->args[1] = 0;
2728                         goto restart;
2729                 }
2730         }
2731 out:
2732         rcu_read_unlock();
2733         if (last)
2734                 nf_ct_expect_put(last);
2735
2736         return skb->len;
2737 }
2738
2739 static int
2740 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2741 {
2742         struct nf_conntrack_expect *exp, *last;
2743         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2744         struct nf_conn *ct = cb->data;
2745         struct nf_conn_help *help = nfct_help(ct);
2746         u_int8_t l3proto = nfmsg->nfgen_family;
2747
2748         if (cb->args[0])
2749                 return 0;
2750
2751         rcu_read_lock();
2752         last = (struct nf_conntrack_expect *)cb->args[1];
2753 restart:
2754         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2755                 if (l3proto && exp->tuple.src.l3num != l3proto)
2756                         continue;
2757                 if (cb->args[1]) {
2758                         if (exp != last)
2759                                 continue;
2760                         cb->args[1] = 0;
2761                 }
2762                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2763                                             cb->nlh->nlmsg_seq,
2764                                             IPCTNL_MSG_EXP_NEW,
2765                                             exp) < 0) {
2766                         if (!refcount_inc_not_zero(&exp->use))
2767                                 continue;
2768                         cb->args[1] = (unsigned long)exp;
2769                         goto out;
2770                 }
2771         }
2772         if (cb->args[1]) {
2773                 cb->args[1] = 0;
2774                 goto restart;
2775         }
2776         cb->args[0] = 1;
2777 out:
2778         rcu_read_unlock();
2779         if (last)
2780                 nf_ct_expect_put(last);
2781
2782         return skb->len;
2783 }
2784
2785 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2786                                  struct sk_buff *skb,
2787                                  const struct nlmsghdr *nlh,
2788                                  const struct nlattr * const cda[],
2789                                  struct netlink_ext_ack *extack)
2790 {
2791         int err;
2792         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2793         u_int8_t u3 = nfmsg->nfgen_family;
2794         struct nf_conntrack_tuple tuple;
2795         struct nf_conntrack_tuple_hash *h;
2796         struct nf_conn *ct;
2797         struct nf_conntrack_zone zone;
2798         struct netlink_dump_control c = {
2799                 .dump = ctnetlink_exp_ct_dump_table,
2800                 .done = ctnetlink_exp_done,
2801         };
2802
2803         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2804                                     u3, NULL);
2805         if (err < 0)
2806                 return err;
2807
2808         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2809         if (err < 0)
2810                 return err;
2811
2812         h = nf_conntrack_find_get(net, &zone, &tuple);
2813         if (!h)
2814                 return -ENOENT;
2815
2816         ct = nf_ct_tuplehash_to_ctrack(h);
2817         /* No expectation linked to this connection tracking. */
2818         if (!nfct_help(ct)) {
2819                 nf_ct_put(ct);
2820                 return 0;
2821         }
2822
2823         c.data = ct;
2824
2825         err = netlink_dump_start(ctnl, skb, nlh, &c);
2826         nf_ct_put(ct);
2827
2828         return err;
2829 }
2830
2831 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
2832                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2833                                 const struct nlattr * const cda[],
2834                                 struct netlink_ext_ack *extack)
2835 {
2836         struct nf_conntrack_tuple tuple;
2837         struct nf_conntrack_expect *exp;
2838         struct sk_buff *skb2;
2839         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2840         u_int8_t u3 = nfmsg->nfgen_family;
2841         struct nf_conntrack_zone zone;
2842         int err;
2843
2844         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2845                 if (cda[CTA_EXPECT_MASTER])
2846                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
2847                                                      extack);
2848                 else {
2849                         struct netlink_dump_control c = {
2850                                 .dump = ctnetlink_exp_dump_table,
2851                                 .done = ctnetlink_exp_done,
2852                         };
2853                         return netlink_dump_start(ctnl, skb, nlh, &c);
2854                 }
2855         }
2856
2857         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2858         if (err < 0)
2859                 return err;
2860
2861         if (cda[CTA_EXPECT_TUPLE])
2862                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2863                                             u3, NULL);
2864         else if (cda[CTA_EXPECT_MASTER])
2865                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2866                                             u3, NULL);
2867         else
2868                 return -EINVAL;
2869
2870         if (err < 0)
2871                 return err;
2872
2873         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2874         if (!exp)
2875                 return -ENOENT;
2876
2877         if (cda[CTA_EXPECT_ID]) {
2878                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2879                 if (ntohl(id) != (u32)(unsigned long)exp) {
2880                         nf_ct_expect_put(exp);
2881                         return -ENOENT;
2882                 }
2883         }
2884
2885         err = -ENOMEM;
2886         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2887         if (skb2 == NULL) {
2888                 nf_ct_expect_put(exp);
2889                 goto out;
2890         }
2891
2892         rcu_read_lock();
2893         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2894                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2895         rcu_read_unlock();
2896         nf_ct_expect_put(exp);
2897         if (err <= 0)
2898                 goto free;
2899
2900         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2901         if (err < 0)
2902                 goto out;
2903
2904         return 0;
2905
2906 free:
2907         kfree_skb(skb2);
2908 out:
2909         /* this avoids a loop in nfnetlink. */
2910         return err == -EAGAIN ? -ENOBUFS : err;
2911 }
2912
2913 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
2914 {
2915         const struct nf_conn_help *m_help;
2916         const char *name = data;
2917
2918         m_help = nfct_help(exp->master);
2919
2920         return strcmp(m_help->helper->name, name) == 0;
2921 }
2922
2923 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
2924 {
2925         return true;
2926 }
2927
2928 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
2929                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2930                                 const struct nlattr * const cda[],
2931                                 struct netlink_ext_ack *extack)
2932 {
2933         struct nf_conntrack_expect *exp;
2934         struct nf_conntrack_tuple tuple;
2935         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2936         u_int8_t u3 = nfmsg->nfgen_family;
2937         struct nf_conntrack_zone zone;
2938         int err;
2939
2940         if (cda[CTA_EXPECT_TUPLE]) {
2941                 /* delete a single expect by tuple */
2942                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2943                 if (err < 0)
2944                         return err;
2945
2946                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2947                                             u3, NULL);
2948                 if (err < 0)
2949                         return err;
2950
2951                 /* bump usage count to 2 */
2952                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2953                 if (!exp)
2954                         return -ENOENT;
2955
2956                 if (cda[CTA_EXPECT_ID]) {
2957                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2958                         if (ntohl(id) != (u32)(unsigned long)exp) {
2959                                 nf_ct_expect_put(exp);
2960                                 return -ENOENT;
2961                         }
2962                 }
2963
2964                 /* after list removal, usage count == 1 */
2965                 spin_lock_bh(&nf_conntrack_expect_lock);
2966                 if (del_timer(&exp->timeout)) {
2967                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2968                                                    nlmsg_report(nlh));
2969                         nf_ct_expect_put(exp);
2970                 }
2971                 spin_unlock_bh(&nf_conntrack_expect_lock);
2972                 /* have to put what we 'get' above.
2973                  * after this line usage count == 0 */
2974                 nf_ct_expect_put(exp);
2975         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2976                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2977
2978                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
2979                                          NETLINK_CB(skb).portid,
2980                                          nlmsg_report(nlh));
2981         } else {
2982                 /* This basically means we have to flush everything*/
2983                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
2984                                          NETLINK_CB(skb).portid,
2985                                          nlmsg_report(nlh));
2986         }
2987
2988         return 0;
2989 }
2990 static int
2991 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2992                         const struct nlattr * const cda[])
2993 {
2994         if (cda[CTA_EXPECT_TIMEOUT]) {
2995                 if (!del_timer(&x->timeout))
2996                         return -ETIME;
2997
2998                 x->timeout.expires = jiffies +
2999                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3000                 add_timer(&x->timeout);
3001         }
3002         return 0;
3003 }
3004
3005 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3006         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3007         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3008 };
3009
3010 static int
3011 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3012                            struct nf_conntrack_expect *exp,
3013                            u_int8_t u3)
3014 {
3015 #ifdef CONFIG_NF_NAT_NEEDED
3016         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3017         struct nf_conntrack_tuple nat_tuple = {};
3018         int err;
3019
3020         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3021                                exp_nat_nla_policy, NULL);
3022         if (err < 0)
3023                 return err;
3024
3025         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3026                 return -EINVAL;
3027
3028         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3029                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3030                                     u3, NULL);
3031         if (err < 0)
3032                 return err;
3033
3034         exp->saved_addr = nat_tuple.src.u3;
3035         exp->saved_proto = nat_tuple.src.u;
3036         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3037
3038         return 0;
3039 #else
3040         return -EOPNOTSUPP;
3041 #endif
3042 }
3043
3044 static struct nf_conntrack_expect *
3045 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3046                        struct nf_conntrack_helper *helper,
3047                        struct nf_conntrack_tuple *tuple,
3048                        struct nf_conntrack_tuple *mask)
3049 {
3050         u_int32_t class = 0;
3051         struct nf_conntrack_expect *exp;
3052         struct nf_conn_help *help;
3053         int err;
3054
3055         help = nfct_help(ct);
3056         if (!help)
3057                 return ERR_PTR(-EOPNOTSUPP);
3058
3059         if (cda[CTA_EXPECT_CLASS] && helper) {
3060                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3061                 if (class > helper->expect_class_max)
3062                         return ERR_PTR(-EINVAL);
3063         }
3064         exp = nf_ct_expect_alloc(ct);
3065         if (!exp)
3066                 return ERR_PTR(-ENOMEM);
3067
3068         if (cda[CTA_EXPECT_FLAGS]) {
3069                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3070                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3071         } else {
3072                 exp->flags = 0;
3073         }
3074         if (cda[CTA_EXPECT_FN]) {
3075                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3076                 struct nf_ct_helper_expectfn *expfn;
3077
3078                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3079                 if (expfn == NULL) {
3080                         err = -EINVAL;
3081                         goto err_out;
3082                 }
3083                 exp->expectfn = expfn->expectfn;
3084         } else
3085                 exp->expectfn = NULL;
3086
3087         exp->class = class;
3088         exp->master = ct;
3089         exp->helper = helper;
3090         exp->tuple = *tuple;
3091         exp->mask.src.u3 = mask->src.u3;
3092         exp->mask.src.u.all = mask->src.u.all;
3093
3094         if (cda[CTA_EXPECT_NAT]) {
3095                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3096                                                  exp, nf_ct_l3num(ct));
3097                 if (err < 0)
3098                         goto err_out;
3099         }
3100         return exp;
3101 err_out:
3102         nf_ct_expect_put(exp);
3103         return ERR_PTR(err);
3104 }
3105
3106 static int
3107 ctnetlink_create_expect(struct net *net,
3108                         const struct nf_conntrack_zone *zone,
3109                         const struct nlattr * const cda[],
3110                         u_int8_t u3, u32 portid, int report)
3111 {
3112         struct nf_conntrack_tuple tuple, mask, master_tuple;
3113         struct nf_conntrack_tuple_hash *h = NULL;
3114         struct nf_conntrack_helper *helper = NULL;
3115         struct nf_conntrack_expect *exp;
3116         struct nf_conn *ct;
3117         int err;
3118
3119         /* caller guarantees that those three CTA_EXPECT_* exist */
3120         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3121                                     u3, NULL);
3122         if (err < 0)
3123                 return err;
3124         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3125                                     u3, NULL);
3126         if (err < 0)
3127                 return err;
3128         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3129                                     u3, NULL);
3130         if (err < 0)
3131                 return err;
3132
3133         /* Look for master conntrack of this expectation */
3134         h = nf_conntrack_find_get(net, zone, &master_tuple);
3135         if (!h)
3136                 return -ENOENT;
3137         ct = nf_ct_tuplehash_to_ctrack(h);
3138
3139         rcu_read_lock();
3140         if (cda[CTA_EXPECT_HELP_NAME]) {
3141                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3142
3143                 helper = __nf_conntrack_helper_find(helpname, u3,
3144                                                     nf_ct_protonum(ct));
3145                 if (helper == NULL) {
3146                         rcu_read_unlock();
3147 #ifdef CONFIG_MODULES
3148                         if (request_module("nfct-helper-%s", helpname) < 0) {
3149                                 err = -EOPNOTSUPP;
3150                                 goto err_ct;
3151                         }
3152                         rcu_read_lock();
3153                         helper = __nf_conntrack_helper_find(helpname, u3,
3154                                                             nf_ct_protonum(ct));
3155                         if (helper) {
3156                                 err = -EAGAIN;
3157                                 goto err_rcu;
3158                         }
3159                         rcu_read_unlock();
3160 #endif
3161                         err = -EOPNOTSUPP;
3162                         goto err_ct;
3163                 }
3164         }
3165
3166         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3167         if (IS_ERR(exp)) {
3168                 err = PTR_ERR(exp);
3169                 goto err_rcu;
3170         }
3171
3172         err = nf_ct_expect_related_report(exp, portid, report);
3173         nf_ct_expect_put(exp);
3174 err_rcu:
3175         rcu_read_unlock();
3176 err_ct:
3177         nf_ct_put(ct);
3178         return err;
3179 }
3180
3181 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3182                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3183                                 const struct nlattr * const cda[],
3184                                 struct netlink_ext_ack *extack)
3185 {
3186         struct nf_conntrack_tuple tuple;
3187         struct nf_conntrack_expect *exp;
3188         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3189         u_int8_t u3 = nfmsg->nfgen_family;
3190         struct nf_conntrack_zone zone;
3191         int err;
3192
3193         if (!cda[CTA_EXPECT_TUPLE]
3194             || !cda[CTA_EXPECT_MASK]
3195             || !cda[CTA_EXPECT_MASTER])
3196                 return -EINVAL;
3197
3198         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3199         if (err < 0)
3200                 return err;
3201
3202         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3203                                     u3, NULL);
3204         if (err < 0)
3205                 return err;
3206
3207         spin_lock_bh(&nf_conntrack_expect_lock);
3208         exp = __nf_ct_expect_find(net, &zone, &tuple);
3209         if (!exp) {
3210                 spin_unlock_bh(&nf_conntrack_expect_lock);
3211                 err = -ENOENT;
3212                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3213                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3214                                                       NETLINK_CB(skb).portid,
3215                                                       nlmsg_report(nlh));
3216                 }
3217                 return err;
3218         }
3219
3220         err = -EEXIST;
3221         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3222                 err = ctnetlink_change_expect(exp, cda);
3223         spin_unlock_bh(&nf_conntrack_expect_lock);
3224
3225         return err;
3226 }
3227
3228 static int
3229 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3230                              const struct ip_conntrack_stat *st)
3231 {
3232         struct nlmsghdr *nlh;
3233         struct nfgenmsg *nfmsg;
3234         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3235
3236         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3237                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3238         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3239         if (nlh == NULL)
3240                 goto nlmsg_failure;
3241
3242         nfmsg = nlmsg_data(nlh);
3243         nfmsg->nfgen_family = AF_UNSPEC;
3244         nfmsg->version      = NFNETLINK_V0;
3245         nfmsg->res_id       = htons(cpu);
3246
3247         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3248             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3249             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3250                 goto nla_put_failure;
3251
3252         nlmsg_end(skb, nlh);
3253         return skb->len;
3254
3255 nla_put_failure:
3256 nlmsg_failure:
3257         nlmsg_cancel(skb, nlh);
3258         return -1;
3259 }
3260
3261 static int
3262 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3263 {
3264         int cpu;
3265         struct net *net = sock_net(skb->sk);
3266
3267         if (cb->args[0] == nr_cpu_ids)
3268                 return 0;
3269
3270         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3271                 const struct ip_conntrack_stat *st;
3272
3273                 if (!cpu_possible(cpu))
3274                         continue;
3275
3276                 st = per_cpu_ptr(net->ct.stat, cpu);
3277                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3278                                                  cb->nlh->nlmsg_seq,
3279                                                  cpu, st) < 0)
3280                         break;
3281         }
3282         cb->args[0] = cpu;
3283
3284         return skb->len;
3285 }
3286
3287 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3288                                   struct sk_buff *skb,
3289                                   const struct nlmsghdr *nlh,
3290                                   const struct nlattr * const cda[],
3291                                   struct netlink_ext_ack *extack)
3292 {
3293         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3294                 struct netlink_dump_control c = {
3295                         .dump = ctnetlink_exp_stat_cpu_dump,
3296                 };
3297                 return netlink_dump_start(ctnl, skb, nlh, &c);
3298         }
3299
3300         return 0;
3301 }
3302
3303 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3304 static struct nf_ct_event_notifier ctnl_notifier = {
3305         .fcn = ctnetlink_conntrack_event,
3306 };
3307
3308 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3309         .fcn = ctnetlink_expect_event,
3310 };
3311 #endif
3312
3313 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3314         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3315                                             .attr_count = CTA_MAX,
3316                                             .policy = ct_nla_policy },
3317         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3318                                             .attr_count = CTA_MAX,
3319                                             .policy = ct_nla_policy },
3320         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3321                                             .attr_count = CTA_MAX,
3322                                             .policy = ct_nla_policy },
3323         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3324                                             .attr_count = CTA_MAX,
3325                                             .policy = ct_nla_policy },
3326         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3327         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3328         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3329         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3330 };
3331
3332 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3333         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3334                                             .attr_count = CTA_EXPECT_MAX,
3335                                             .policy = exp_nla_policy },
3336         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3337                                             .attr_count = CTA_EXPECT_MAX,
3338                                             .policy = exp_nla_policy },
3339         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3340                                             .attr_count = CTA_EXPECT_MAX,
3341                                             .policy = exp_nla_policy },
3342         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3343 };
3344
3345 static const struct nfnetlink_subsystem ctnl_subsys = {
3346         .name                           = "conntrack",
3347         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3348         .cb_count                       = IPCTNL_MSG_MAX,
3349         .cb                             = ctnl_cb,
3350 };
3351
3352 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3353         .name                           = "conntrack_expect",
3354         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3355         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3356         .cb                             = ctnl_exp_cb,
3357 };
3358
3359 MODULE_ALIAS("ip_conntrack_netlink");
3360 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3361 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3362
3363 static int __net_init ctnetlink_net_init(struct net *net)
3364 {
3365 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3366         int ret;
3367
3368         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3369         if (ret < 0) {
3370                 pr_err("ctnetlink_init: cannot register notifier.\n");
3371                 goto err_out;
3372         }
3373
3374         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3375         if (ret < 0) {
3376                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3377                 goto err_unreg_notifier;
3378         }
3379 #endif
3380         return 0;
3381
3382 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3383 err_unreg_notifier:
3384         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3385 err_out:
3386         return ret;
3387 #endif
3388 }
3389
3390 static void ctnetlink_net_exit(struct net *net)
3391 {
3392 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3393         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3394         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3395 #endif
3396 }
3397
3398 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3399 {
3400         struct net *net;
3401
3402         list_for_each_entry(net, net_exit_list, exit_list)
3403                 ctnetlink_net_exit(net);
3404 }
3405
3406 static struct pernet_operations ctnetlink_net_ops = {
3407         .init           = ctnetlink_net_init,
3408         .exit_batch     = ctnetlink_net_exit_batch,
3409 };
3410
3411 static int __init ctnetlink_init(void)
3412 {
3413         int ret;
3414
3415         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3416         ret = nfnetlink_subsys_register(&ctnl_subsys);
3417         if (ret < 0) {
3418                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3419                 goto err_out;
3420         }
3421
3422         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3423         if (ret < 0) {
3424                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3425                 goto err_unreg_subsys;
3426         }
3427
3428         ret = register_pernet_subsys(&ctnetlink_net_ops);
3429         if (ret < 0) {
3430                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3431                 goto err_unreg_exp_subsys;
3432         }
3433 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3434         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3435         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3436 #endif
3437         return 0;
3438
3439 err_unreg_exp_subsys:
3440         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3441 err_unreg_subsys:
3442         nfnetlink_subsys_unregister(&ctnl_subsys);
3443 err_out:
3444         return ret;
3445 }
3446
3447 static void __exit ctnetlink_exit(void)
3448 {
3449         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3450
3451         unregister_pernet_subsys(&ctnetlink_net_ops);
3452         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3453         nfnetlink_subsys_unregister(&ctnl_subsys);
3454 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3455         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3456 #endif
3457         synchronize_rcu();
3458 }
3459
3460 module_init(ctnetlink_init);
3461 module_exit(ctnetlink_exit);