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 <[email protected]>
   5  * (C) 2002-2006 by Harald Welte <[email protected]>
   6  * (C) 2003 by Patrick Mchardy <[email protected]>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <[email protected]>
   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 #include <linux/siphash.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netlink.h>
  36 #include <net/sock.h>
  37 #include <net/netfilter/nf_conntrack.h>
  38 #include <net/netfilter/nf_conntrack_core.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_seqadj.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 #if IS_ENABLED(CONFIG_NF_NAT)
  50 #include <net/netfilter/nf_nat.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 #include "nf_internals.h"
  58
  59 MODULE_LICENSE("GPL");
  60 MODULE_DESCRIPTION("List and change connection tracking table");
  61
  62 struct ctnetlink_list_dump_ctx {
  63         struct nf_conn *last;
  64         unsigned int cpu;
  65         bool done;
  66 };
  67
  68 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  69                                 const struct nf_conntrack_tuple *tuple,
  70                                 const struct nf_conntrack_l4proto *l4proto)
  71 {
  72         int ret = 0;
  73         struct nlattr *nest_parms;
  74
  75         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO);
  76         if (!nest_parms)
  77                 goto nla_put_failure;
  78         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  79                 goto nla_put_failure;
  80
  81         if (likely(l4proto->tuple_to_nlattr))
  82                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  83
  84         nla_nest_end(skb, nest_parms);
  85
  86         return ret;
  87
  88 nla_put_failure:
  89         return -1;
  90 }
  91
  92 static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  93                                 const struct nf_conntrack_tuple *tuple)
  94 {
  95         if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  96             nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  97                 return -EMSGSIZE;
  98         return 0;
  99 }
 100
 101 static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
 102                                 const struct nf_conntrack_tuple *tuple)
 103 {
 104         if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
 105             nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
 106                 return -EMSGSIZE;
 107         return 0;
 108 }
 109
 110 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 111                                     const struct nf_conntrack_tuple *tuple)
 112 {
 113         int ret = 0;
 114         struct nlattr *nest_parms;
 115
 116         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP);
 117         if (!nest_parms)
 118                 goto nla_put_failure;
 119
 120         switch (tuple->src.l3num) {
 121         case NFPROTO_IPV4:
 122                 ret = ipv4_tuple_to_nlattr(skb, tuple);
 123                 break;
 124         case NFPROTO_IPV6:
 125                 ret = ipv6_tuple_to_nlattr(skb, tuple);
 126                 break;
 127         }
 128
 129         nla_nest_end(skb, nest_parms);
 130
 131         return ret;
 132
 133 nla_put_failure:
 134         return -1;
 135 }
 136
 137 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 138                                  const struct nf_conntrack_tuple *tuple)
 139 {
 140         const struct nf_conntrack_l4proto *l4proto;
 141         int ret;
 142
 143         rcu_read_lock();
 144         ret = ctnetlink_dump_tuples_ip(skb, tuple);
 145
 146         if (ret >= 0) {
 147                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 148                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 149         }
 150         rcu_read_unlock();
 151         return ret;
 152 }
 153
 154 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 155                                   const struct nf_conntrack_zone *zone, int dir)
 156 {
 157         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 158                 return 0;
 159         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 160                 goto nla_put_failure;
 161         return 0;
 162
 163 nla_put_failure:
 164         return -1;
 165 }
 166
 167 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 168 {
 169         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 170                 goto nla_put_failure;
 171         return 0;
 172
 173 nla_put_failure:
 174         return -1;
 175 }
 176
 177 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct,
 178                                   bool skip_zero)
 179 {
 180         long timeout;
 181
 182         if (nf_ct_is_confirmed(ct))
 183                 timeout = nf_ct_expires(ct) / HZ;
 184         else
 185                 timeout = ct->timeout / HZ;
 186
 187         if (skip_zero && timeout == 0)
 188                 return 0;
 189
 190         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 191                 goto nla_put_failure;
 192         return 0;
 193
 194 nla_put_failure:
 195         return -1;
 196 }
 197
 198 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct,
 199                                     bool destroy)
 200 {
 201         const struct nf_conntrack_l4proto *l4proto;
 202         struct nlattr *nest_proto;
 203         int ret;
 204
 205         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 206         if (!l4proto->to_nlattr)
 207                 return 0;
 208
 209         nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 210         if (!nest_proto)
 211                 goto nla_put_failure;
 212
 213         ret = l4proto->to_nlattr(skb, nest_proto, ct, destroy);
 214
 215         nla_nest_end(skb, nest_proto);
 216
 217         return ret;
 218
 219 nla_put_failure:
 220         return -1;
 221 }
 222
 223 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 224                                    const struct nf_conn *ct)
 225 {
 226         struct nlattr *nest_helper;
 227         const struct nf_conn_help *help = nfct_help(ct);
 228         struct nf_conntrack_helper *helper;
 229
 230         if (!help)
 231                 return 0;
 232
 233         rcu_read_lock();
 234         helper = rcu_dereference(help->helper);
 235         if (!helper)
 236                 goto out;
 237
 238         nest_helper = nla_nest_start(skb, CTA_HELP);
 239         if (!nest_helper)
 240                 goto nla_put_failure;
 241         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 242                 goto nla_put_failure;
 243
 244         if (helper->to_nlattr)
 245                 helper->to_nlattr(skb, ct);
 246
 247         nla_nest_end(skb, nest_helper);
 248 out:
 249         rcu_read_unlock();
 250         return 0;
 251
 252 nla_put_failure:
 253         rcu_read_unlock();
 254         return -1;
 255 }
 256
 257 static int
 258 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 259               enum ip_conntrack_dir dir, int type)
 260 {
 261         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 262         struct nf_conn_counter *counter = acct->counter;
 263         struct nlattr *nest_count;
 264         u64 pkts, bytes;
 265
 266         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 267                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 268                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 269         } else {
 270                 pkts = atomic64_read(&counter[dir].packets);
 271                 bytes = atomic64_read(&counter[dir].bytes);
 272         }
 273
 274         nest_count = nla_nest_start(skb, attr);
 275         if (!nest_count)
 276                 goto nla_put_failure;
 277
 278         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 279                          CTA_COUNTERS_PAD) ||
 280             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 281                          CTA_COUNTERS_PAD))
 282                 goto nla_put_failure;
 283
 284         nla_nest_end(skb, nest_count);
 285
 286         return 0;
 287
 288 nla_put_failure:
 289         return -1;
 290 }
 291
 292 static int
 293 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 294 {
 295         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 296
 297         if (!acct)
 298                 return 0;
 299
 300         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 301                 return -1;
 302         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 303                 return -1;
 304
 305         return 0;
 306 }
 307
 308 static int
 309 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 310 {
 311         struct nlattr *nest_count;
 312         const struct nf_conn_tstamp *tstamp;
 313
 314         tstamp = nf_conn_tstamp_find(ct);
 315         if (!tstamp)
 316                 return 0;
 317
 318         nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 319         if (!nest_count)
 320                 goto nla_put_failure;
 321
 322         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 323                          CTA_TIMESTAMP_PAD) ||
 324             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 325                                                cpu_to_be64(tstamp->stop),
 326                                                CTA_TIMESTAMP_PAD)))
 327                 goto nla_put_failure;
 328         nla_nest_end(skb, nest_count);
 329
 330         return 0;
 331
 332 nla_put_failure:
 333         return -1;
 334 }
 335
 336 #ifdef CONFIG_NF_CONNTRACK_MARK
 337 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct,
 338                                bool dump)
 339 {
 340         u32 mark = READ_ONCE(ct->mark);
 341
 342         if (!mark && !dump)
 343                 return 0;
 344
 345         if (nla_put_be32(skb, CTA_MARK, htonl(mark)))
 346                 goto nla_put_failure;
 347         return 0;
 348
 349 nla_put_failure:
 350         return -1;
 351 }
 352 #else
 353 #define ctnetlink_dump_mark(a, b, c) (0)
 354 #endif
 355
 356 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 357 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 358 {
 359         struct nlattr *nest_secctx;
 360         int len, ret;
 361         char *secctx;
 362
 363         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 364         if (ret)
 365                 return 0;
 366
 367         ret = -1;
 368         nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 369         if (!nest_secctx)
 370                 goto nla_put_failure;
 371
 372         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 373                 goto nla_put_failure;
 374         nla_nest_end(skb, nest_secctx);
 375
 376         ret = 0;
 377 nla_put_failure:
 378         security_release_secctx(secctx, len);
 379         return ret;
 380 }
 381 #else
 382 #define ctnetlink_dump_secctx(a, b) (0)
 383 #endif
 384
 385 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 386 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 387 {
 388         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 389
 390         if (!labels)
 391                 return 0;
 392         return nla_total_size(sizeof(labels->bits));
 393 }
 394 #endif
 395
 396 static int
 397 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 398 {
 399         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 400         unsigned int i;
 401
 402         if (!labels)
 403                 return 0;
 404
 405         i = 0;
 406         do {
 407                 if (labels->bits[i] != 0)
 408                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 409                                        labels->bits);
 410                 i++;
 411         } while (i < ARRAY_SIZE(labels->bits));
 412
 413         return 0;
 414 }
 415
 416 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 417
 418 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 419 {
 420         struct nlattr *nest_parms;
 421
 422         if (!(ct->status & IPS_EXPECTED))
 423                 return 0;
 424
 425         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 426         if (!nest_parms)
 427                 goto nla_put_failure;
 428         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 429                 goto nla_put_failure;
 430         nla_nest_end(skb, nest_parms);
 431
 432         return 0;
 433
 434 nla_put_failure:
 435         return -1;
 436 }
 437
 438 static int
 439 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 440 {
 441         struct nlattr *nest_parms;
 442
 443         nest_parms = nla_nest_start(skb, type);
 444         if (!nest_parms)
 445                 goto nla_put_failure;
 446
 447         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 448                          htonl(seq->correction_pos)) ||
 449             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 450                          htonl(seq->offset_before)) ||
 451             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 452                          htonl(seq->offset_after)))
 453                 goto nla_put_failure;
 454
 455         nla_nest_end(skb, nest_parms);
 456
 457         return 0;
 458
 459 nla_put_failure:
 460         return -1;
 461 }
 462
 463 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 464 {
 465         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 466         struct nf_ct_seqadj *seq;
 467
 468         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 469                 return 0;
 470
 471         spin_lock_bh(&ct->lock);
 472         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 473         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 474                 goto err;
 475
 476         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 477         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 478                 goto err;
 479
 480         spin_unlock_bh(&ct->lock);
 481         return 0;
 482 err:
 483         spin_unlock_bh(&ct->lock);
 484         return -1;
 485 }
 486
 487 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 488 {
 489         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 490         struct nlattr *nest_parms;
 491
 492         if (!synproxy)
 493                 return 0;
 494
 495         nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 496         if (!nest_parms)
 497                 goto nla_put_failure;
 498
 499         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 500             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 501             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 502                 goto nla_put_failure;
 503
 504         nla_nest_end(skb, nest_parms);
 505
 506         return 0;
 507
 508 nla_put_failure:
 509         return -1;
 510 }
 511
 512 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 513 {
 514         __be32 id = (__force __be32)nf_ct_get_id(ct);
 515
 516         if (nla_put_be32(skb, CTA_ID, id))
 517                 goto nla_put_failure;
 518         return 0;
 519
 520 nla_put_failure:
 521         return -1;
 522 }
 523
 524 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 525 {
 526         if (nla_put_be32(skb, CTA_USE, htonl(refcount_read(&ct->ct_general.use))))
 527                 goto nla_put_failure;
 528         return 0;
 529
 530 nla_put_failure:
 531         return -1;
 532 }
 533
 534 /* all these functions access ct->ext. Caller must either hold a reference
 535  * on ct or prevent its deletion by holding either the bucket spinlock or
 536  * pcpu dying list lock.
 537  */
 538 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 539                                   struct nf_conn *ct, u32 type)
 540 {
 541         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 542             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 543             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 544             ctnetlink_dump_labels(skb, ct) < 0 ||
 545             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 546             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 547                 return -1;
 548
 549         return 0;
 550 }
 551
 552 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 553 {
 554         if (ctnetlink_dump_status(skb, ct) < 0 ||
 555             ctnetlink_dump_mark(skb, ct, true) < 0 ||
 556             ctnetlink_dump_secctx(skb, ct) < 0 ||
 557             ctnetlink_dump_id(skb, ct) < 0 ||
 558             ctnetlink_dump_use(skb, ct) < 0 ||
 559             ctnetlink_dump_master(skb, ct) < 0)
 560                 return -1;
 561
 562         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 563             (ctnetlink_dump_timeout(skb, ct, false) < 0 ||
 564              ctnetlink_dump_protoinfo(skb, ct, false) < 0))
 565                 return -1;
 566
 567         return 0;
 568 }
 569
 570 static int
 571 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 572                     struct nf_conn *ct, bool extinfo, unsigned int flags)
 573 {
 574         const struct nf_conntrack_zone *zone;
 575         struct nlmsghdr *nlh;
 576         struct nlattr *nest_parms;
 577         unsigned int event;
 578
 579         if (portid)
 580                 flags |= NLM_F_MULTI;
 581         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 582         nlh = nfnl_msg_put(skb, portid, seq, event, flags, nf_ct_l3num(ct),
 583                            NFNETLINK_V0, 0);
 584         if (!nlh)
 585                 goto nlmsg_failure;
 586
 587         zone = nf_ct_zone(ct);
 588
 589         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 590         if (!nest_parms)
 591                 goto nla_put_failure;
 592         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 593                 goto nla_put_failure;
 594         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 595                                    NF_CT_ZONE_DIR_ORIG) < 0)
 596                 goto nla_put_failure;
 597         nla_nest_end(skb, nest_parms);
 598
 599         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 600         if (!nest_parms)
 601                 goto nla_put_failure;
 602         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 603                 goto nla_put_failure;
 604         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 605                                    NF_CT_ZONE_DIR_REPL) < 0)
 606                 goto nla_put_failure;
 607         nla_nest_end(skb, nest_parms);
 608
 609         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 610                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 611                 goto nla_put_failure;
 612
 613         if (ctnetlink_dump_info(skb, ct) < 0)
 614                 goto nla_put_failure;
 615         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 616                 goto nla_put_failure;
 617
 618         nlmsg_end(skb, nlh);
 619         return skb->len;
 620
 621 nlmsg_failure:
 622 nla_put_failure:
 623         nlmsg_cancel(skb, nlh);
 624         return -1;
 625 }
 626
 627 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 628         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 629         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 630         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 631         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 632 };
 633
 634 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 635 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 636 {
 637         const struct nf_conntrack_l4proto *l4proto;
 638         size_t len, len4 = 0;
 639
 640         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 641         len *= 3u; /* ORIG, REPLY, MASTER */
 642
 643         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 644         len += l4proto->nlattr_size;
 645         if (l4proto->nlattr_tuple_size) {
 646                 len4 = l4proto->nlattr_tuple_size();
 647                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 648         }
 649
 650         return len + len4;
 651 }
 652
 653 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 654 {
 655         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 656                 return 0;
 657         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 658                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 659                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 660                ;
 661 }
 662
 663 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 664 {
 665 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 666         int len, ret;
 667
 668         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 669         if (ret)
 670                 return 0;
 671
 672         return nla_total_size(0) /* CTA_SECCTX */
 673                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 674 #else
 675         return 0;
 676 #endif
 677 }
 678
 679 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 680 {
 681 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 682         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 683                 return 0;
 684         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 685 #else
 686         return 0;
 687 #endif
 688 }
 689 #endif
 690
 691 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 692 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 693 {
 694         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 695                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 696                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 697                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 698                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 699                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 700                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 701                + ctnetlink_acct_size(ct)
 702                + ctnetlink_timestamp_size(ct)
 703                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 704                + nla_total_size(0) /* CTA_PROTOINFO */
 705                + nla_total_size(0) /* CTA_HELP */
 706                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 707                + ctnetlink_secctx_size(ct)
 708 #if IS_ENABLED(CONFIG_NF_NAT)
 709                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 710                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 711 #endif
 712 #ifdef CONFIG_NF_CONNTRACK_MARK
 713                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 714 #endif
 715 #ifdef CONFIG_NF_CONNTRACK_ZONES
 716                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 717 #endif
 718                + ctnetlink_proto_size(ct)
 719                + ctnetlink_label_size(ct)
 720                ;
 721 }
 722
 723 static int
 724 ctnetlink_conntrack_event(unsigned int events, const struct nf_ct_event *item)
 725 {
 726         const struct nf_conntrack_zone *zone;
 727         struct net *net;
 728         struct nlmsghdr *nlh;
 729         struct nlattr *nest_parms;
 730         struct nf_conn *ct = item->ct;
 731         struct sk_buff *skb;
 732         unsigned int type;
 733         unsigned int flags = 0, group;
 734         int err;
 735
 736         if (events & (1 << IPCT_DESTROY)) {
 737                 type = IPCTNL_MSG_CT_DELETE;
 738                 group = NFNLGRP_CONNTRACK_DESTROY;
 739         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 740                 type = IPCTNL_MSG_CT_NEW;
 741                 flags = NLM_F_CREATE|NLM_F_EXCL;
 742                 group = NFNLGRP_CONNTRACK_NEW;
 743         } else if (events) {
 744                 type = IPCTNL_MSG_CT_NEW;
 745                 group = NFNLGRP_CONNTRACK_UPDATE;
 746         } else
 747                 return 0;
 748
 749         net = nf_ct_net(ct);
 750         if (!item->report && !nfnetlink_has_listeners(net, group))
 751                 return 0;
 752
 753         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 754         if (skb == NULL)
 755                 goto errout;
 756
 757         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 758         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags, nf_ct_l3num(ct),
 759                            NFNETLINK_V0, 0);
 760         if (!nlh)
 761                 goto nlmsg_failure;
 762
 763         zone = nf_ct_zone(ct);
 764
 765         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 766         if (!nest_parms)
 767                 goto nla_put_failure;
 768         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 769                 goto nla_put_failure;
 770         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 771                                    NF_CT_ZONE_DIR_ORIG) < 0)
 772                 goto nla_put_failure;
 773         nla_nest_end(skb, nest_parms);
 774
 775         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 776         if (!nest_parms)
 777                 goto nla_put_failure;
 778         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 779                 goto nla_put_failure;
 780         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 781                                    NF_CT_ZONE_DIR_REPL) < 0)
 782                 goto nla_put_failure;
 783         nla_nest_end(skb, nest_parms);
 784
 785         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 786                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 787                 goto nla_put_failure;
 788
 789         if (ctnetlink_dump_id(skb, ct) < 0)
 790                 goto nla_put_failure;
 791
 792         if (ctnetlink_dump_status(skb, ct) < 0)
 793                 goto nla_put_failure;
 794
 795         if (events & (1 << IPCT_DESTROY)) {
 796                 if (ctnetlink_dump_timeout(skb, ct, true) < 0)
 797                         goto nla_put_failure;
 798
 799                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 800                     ctnetlink_dump_timestamp(skb, ct) < 0 ||
 801                     ctnetlink_dump_protoinfo(skb, ct, true) < 0)
 802                         goto nla_put_failure;
 803         } else {
 804                 if (ctnetlink_dump_timeout(skb, ct, false) < 0)
 805                         goto nla_put_failure;
 806
 807                 if (events & (1 << IPCT_PROTOINFO) &&
 808                     ctnetlink_dump_protoinfo(skb, ct, false) < 0)
 809                         goto nla_put_failure;
 810
 811                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 812                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 813                         goto nla_put_failure;
 814
 815 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 816                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 817                     && ctnetlink_dump_secctx(skb, ct) < 0)
 818                         goto nla_put_failure;
 819 #endif
 820                 if (events & (1 << IPCT_LABEL) &&
 821                      ctnetlink_dump_labels(skb, ct) < 0)
 822                         goto nla_put_failure;
 823
 824                 if (events & (1 << IPCT_RELATED) &&
 825                     ctnetlink_dump_master(skb, ct) < 0)
 826                         goto nla_put_failure;
 827
 828                 if (events & (1 << IPCT_SEQADJ) &&
 829                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 830                         goto nla_put_failure;
 831
 832                 if (events & (1 << IPCT_SYNPROXY) &&
 833                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 834                         goto nla_put_failure;
 835         }
 836
 837 #ifdef CONFIG_NF_CONNTRACK_MARK
 838         if (ctnetlink_dump_mark(skb, ct, events & (1 << IPCT_MARK)))
 839                 goto nla_put_failure;
 840 #endif
 841         nlmsg_end(skb, nlh);
 842         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 843                              GFP_ATOMIC);
 844         if (err == -ENOBUFS || err == -EAGAIN)
 845                 return -ENOBUFS;
 846
 847         return 0;
 848
 849 nla_put_failure:
 850         nlmsg_cancel(skb, nlh);
 851 nlmsg_failure:
 852         kfree_skb(skb);
 853 errout:
 854         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 855                 return -ENOBUFS;
 856
 857         return 0;
 858 }
 859 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 860
 861 static int ctnetlink_done(struct netlink_callback *cb)
 862 {
 863         if (cb->args[1])
 864                 nf_ct_put((struct nf_conn *)cb->args[1]);
 865         kfree(cb->data);
 866         return 0;
 867 }
 868
 869 struct ctnetlink_filter_u32 {
 870         u32 val;
 871         u32 mask;
 872 };
 873
 874 struct ctnetlink_filter {
 875         u8 family;
 876         bool zone_filter;
 877
 878         u_int32_t orig_flags;
 879         u_int32_t reply_flags;
 880
 881         struct nf_conntrack_tuple orig;
 882         struct nf_conntrack_tuple reply;
 883         struct nf_conntrack_zone zone;
 884
 885         struct ctnetlink_filter_u32 mark;
 886         struct ctnetlink_filter_u32 status;
 887 };
 888
 889 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 890         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 891         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 892 };
 893
 894 static int ctnetlink_parse_filter(const struct nlattr *attr,
 895                                   struct ctnetlink_filter *filter)
 896 {
 897         struct nlattr *tb[CTA_FILTER_MAX + 1];
 898         int ret = 0;
 899
 900         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 901                                NULL);
 902         if (ret)
 903                 return ret;
 904
 905         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 906                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 907                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 908                         return -EOPNOTSUPP;
 909         }
 910
 911         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 912                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 913                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 914                         return -EOPNOTSUPP;
 915         }
 916
 917         return 0;
 918 }
 919
 920 static int ctnetlink_parse_zone(const struct nlattr *attr,
 921                                 struct nf_conntrack_zone *zone);
 922 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 923                                          struct nf_conntrack_tuple *tuple,
 924                                          u32 type, u_int8_t l3num,
 925                                          struct nf_conntrack_zone *zone,
 926                                          u_int32_t flags);
 927
 928 static int ctnetlink_filter_parse_mark(struct ctnetlink_filter_u32 *mark,
 929                                        const struct nlattr * const cda[])
 930 {
 931 #ifdef CONFIG_NF_CONNTRACK_MARK
 932         if (cda[CTA_MARK]) {
 933                 mark->val = ntohl(nla_get_be32(cda[CTA_MARK]));
 934
 935                 if (cda[CTA_MARK_MASK])
 936                         mark->mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 937                 else
 938                         mark->mask = 0xffffffff;
 939         } else if (cda[CTA_MARK_MASK]) {
 940                 return -EINVAL;
 941         }
 942 #endif
 943         return 0;
 944 }
 945
 946 static int ctnetlink_filter_parse_status(struct ctnetlink_filter_u32 *status,
 947                                          const struct nlattr * const cda[])
 948 {
 949         if (cda[CTA_STATUS]) {
 950                 status->val = ntohl(nla_get_be32(cda[CTA_STATUS]));
 951                 if (cda[CTA_STATUS_MASK])
 952                         status->mask = ntohl(nla_get_be32(cda[CTA_STATUS_MASK]));
 953                 else
 954                         status->mask = status->val;
 955
 956                 /* status->val == 0? always true, else always false. */
 957                 if (status->mask == 0)
 958                         return -EINVAL;
 959         } else if (cda[CTA_STATUS_MASK]) {
 960                 return -EINVAL;
 961         }
 962
 963         /* CTA_STATUS is NLA_U32, if this fires UAPI needs to be extended */
 964         BUILD_BUG_ON(__IPS_MAX_BIT >= 32);
 965         return 0;
 966 }
 967
 968 static struct ctnetlink_filter *
 969 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 970 {
 971         struct ctnetlink_filter *filter;
 972         int err;
 973
 974 #ifndef CONFIG_NF_CONNTRACK_MARK
 975         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 976                 return ERR_PTR(-EOPNOTSUPP);
 977 #endif
 978
 979         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 980         if (filter == NULL)
 981                 return ERR_PTR(-ENOMEM);
 982
 983         filter->family = family;
 984
 985         err = ctnetlink_filter_parse_mark(&filter->mark, cda);
 986         if (err)
 987                 goto err_filter;
 988
 989         err = ctnetlink_filter_parse_status(&filter->status, cda);
 990         if (err)
 991                 goto err_filter;
 992
 993         if (cda[CTA_ZONE]) {
 994                 err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 995                 if (err < 0)
 996                         goto err_filter;
 997                 filter->zone_filter = true;
 998         }
 999
1000         if (!cda[CTA_FILTER])
1001                 return filter;
1002
1003         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
1004         if (err < 0)
1005                 goto err_filter;
1006
1007         if (filter->orig_flags) {
1008                 if (!cda[CTA_TUPLE_ORIG]) {
1009                         err = -EINVAL;
1010                         goto err_filter;
1011                 }
1012
1013                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
1014                                                    CTA_TUPLE_ORIG,
1015                                                    filter->family,
1016                                                    &filter->zone,
1017                                                    filter->orig_flags);
1018                 if (err < 0)
1019                         goto err_filter;
1020         }
1021
1022         if (filter->reply_flags) {
1023                 if (!cda[CTA_TUPLE_REPLY]) {
1024                         err = -EINVAL;
1025                         goto err_filter;
1026                 }
1027
1028                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
1029                                                    CTA_TUPLE_REPLY,
1030                                                    filter->family,
1031                                                    &filter->zone,
1032                                                    filter->reply_flags);
1033                 if (err < 0)
1034                         goto err_filter;
1035         }
1036
1037         return filter;
1038
1039 err_filter:
1040         kfree(filter);
1041
1042         return ERR_PTR(err);
1043 }
1044
1045 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1046 {
1047         return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS] || cda[CTA_ZONE];
1048 }
1049
1050 static int ctnetlink_start(struct netlink_callback *cb)
1051 {
1052         const struct nlattr * const *cda = cb->data;
1053         struct ctnetlink_filter *filter = NULL;
1054         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1055         u8 family = nfmsg->nfgen_family;
1056
1057         if (ctnetlink_needs_filter(family, cda)) {
1058                 filter = ctnetlink_alloc_filter(cda, family);
1059                 if (IS_ERR(filter))
1060                         return PTR_ERR(filter);
1061         }
1062
1063         cb->data = filter;
1064         return 0;
1065 }
1066
1067 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1068                                         struct nf_conntrack_tuple *ct_tuple,
1069                                         u_int32_t flags, int family)
1070 {
1071         switch (family) {
1072         case NFPROTO_IPV4:
1073                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1074                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1075                         return  0;
1076
1077                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1078                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1079                         return  0;
1080                 break;
1081         case NFPROTO_IPV6:
1082                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1083                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1084                                    &ct_tuple->src.u3.in6))
1085                         return 0;
1086
1087                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1088                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1089                                    &ct_tuple->dst.u3.in6))
1090                         return 0;
1091                 break;
1092         }
1093
1094         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1095             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1096                 return 0;
1097
1098         switch (ct_tuple->dst.protonum) {
1099         case IPPROTO_TCP:
1100         case IPPROTO_UDP:
1101                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1102                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1103                         return 0;
1104
1105                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1106                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1107                         return 0;
1108                 break;
1109         case IPPROTO_ICMP:
1110                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1111                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1112                         return 0;
1113                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1114                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1115                         return 0;
1116                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1117                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1118                         return 0;
1119                 break;
1120         case IPPROTO_ICMPV6:
1121                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1122                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1123                         return 0;
1124                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1125                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1126                         return 0;
1127                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1128                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1129                         return 0;
1130                 break;
1131         }
1132
1133         return 1;
1134 }
1135
1136 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1137 {
1138         struct ctnetlink_filter *filter = data;
1139         struct nf_conntrack_tuple *tuple;
1140         u32 status;
1141
1142         if (filter == NULL)
1143                 goto out;
1144
1145         /* Match entries of a given L3 protocol number.
1146          * If it is not specified, ie. l3proto == 0,
1147          * then match everything.
1148          */
1149         if (filter->family && nf_ct_l3num(ct) != filter->family)
1150                 goto ignore_entry;
1151
1152         if (filter->zone_filter &&
1153             !nf_ct_zone_equal_any(ct, &filter->zone))
1154                 goto ignore_entry;
1155
1156         if (filter->orig_flags) {
1157                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1158                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1159                                                   filter->orig_flags,
1160                                                   filter->family))
1161                         goto ignore_entry;
1162         }
1163
1164         if (filter->reply_flags) {
1165                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1166                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1167                                                   filter->reply_flags,
1168                                                   filter->family))
1169                         goto ignore_entry;
1170         }
1171
1172 #ifdef CONFIG_NF_CONNTRACK_MARK
1173         if ((READ_ONCE(ct->mark) & filter->mark.mask) != filter->mark.val)
1174                 goto ignore_entry;
1175 #endif
1176         status = (u32)READ_ONCE(ct->status);
1177         if ((status & filter->status.mask) != filter->status.val)
1178                 goto ignore_entry;
1179
1180 out:
1181         return 1;
1182
1183 ignore_entry:
1184         return 0;
1185 }
1186
1187 static int
1188 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1189 {
1190         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1191         struct net *net = sock_net(skb->sk);
1192         struct nf_conn *ct, *last;
1193         struct nf_conntrack_tuple_hash *h;
1194         struct hlist_nulls_node *n;
1195         struct nf_conn *nf_ct_evict[8];
1196         int res, i;
1197         spinlock_t *lockp;
1198
1199         last = (struct nf_conn *)cb->args[1];
1200         i = 0;
1201
1202         local_bh_disable();
1203         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1204 restart:
1205                 while (i) {
1206                         i--;
1207                         if (nf_ct_should_gc(nf_ct_evict[i]))
1208                                 nf_ct_kill(nf_ct_evict[i]);
1209                         nf_ct_put(nf_ct_evict[i]);
1210                 }
1211
1212                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1213                 nf_conntrack_lock(lockp);
1214                 if (cb->args[0] >= nf_conntrack_htable_size) {
1215                         spin_unlock(lockp);
1216                         goto out;
1217                 }
1218                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1219                                            hnnode) {
1220                         ct = nf_ct_tuplehash_to_ctrack(h);
1221                         if (nf_ct_is_expired(ct)) {
1222                                 /* need to defer nf_ct_kill() until lock is released */
1223                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1224                                     refcount_inc_not_zero(&ct->ct_general.use))
1225                                         nf_ct_evict[i++] = ct;
1226                                 continue;
1227                         }
1228
1229                         if (!net_eq(net, nf_ct_net(ct)))
1230                                 continue;
1231
1232                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1233                                 continue;
1234
1235                         if (cb->args[1]) {
1236                                 if (ct != last)
1237                                         continue;
1238                                 cb->args[1] = 0;
1239                         }
1240                         if (!ctnetlink_filter_match(ct, cb->data))
1241                                 continue;
1242
1243                         res =
1244                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1245                                             cb->nlh->nlmsg_seq,
1246                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1247                                             ct, true, flags);
1248                         if (res < 0) {
1249                                 nf_conntrack_get(&ct->ct_general);
1250                                 cb->args[1] = (unsigned long)ct;
1251                                 spin_unlock(lockp);
1252                                 goto out;
1253                         }
1254                 }
1255                 spin_unlock(lockp);
1256                 if (cb->args[1]) {
1257                         cb->args[1] = 0;
1258                         goto restart;
1259                 }
1260         }
1261 out:
1262         local_bh_enable();
1263         if (last) {
1264                 /* nf ct hash resize happened, now clear the leftover. */
1265                 if ((struct nf_conn *)cb->args[1] == last)
1266                         cb->args[1] = 0;
1267
1268                 nf_ct_put(last);
1269         }
1270
1271         while (i) {
1272                 i--;
1273                 if (nf_ct_should_gc(nf_ct_evict[i]))
1274                         nf_ct_kill(nf_ct_evict[i]);
1275                 nf_ct_put(nf_ct_evict[i]);
1276         }
1277
1278         return skb->len;
1279 }
1280
1281 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1282                                 struct nf_conntrack_tuple *t,
1283                                 u_int32_t flags)
1284 {
1285         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1286                 if (!tb[CTA_IP_V4_SRC])
1287                         return -EINVAL;
1288
1289                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1290         }
1291
1292         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1293                 if (!tb[CTA_IP_V4_DST])
1294                         return -EINVAL;
1295
1296                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1297         }
1298
1299         return 0;
1300 }
1301
1302 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1303                                 struct nf_conntrack_tuple *t,
1304                                 u_int32_t flags)
1305 {
1306         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1307                 if (!tb[CTA_IP_V6_SRC])
1308                         return -EINVAL;
1309
1310                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1311         }
1312
1313         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1314                 if (!tb[CTA_IP_V6_DST])
1315                         return -EINVAL;
1316
1317                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1318         }
1319
1320         return 0;
1321 }
1322
1323 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1324                                     struct nf_conntrack_tuple *tuple,
1325                                     u_int32_t flags)
1326 {
1327         struct nlattr *tb[CTA_IP_MAX+1];
1328         int ret = 0;
1329
1330         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr,
1331                                           cta_ip_nla_policy, NULL);
1332         if (ret < 0)
1333                 return ret;
1334
1335         switch (tuple->src.l3num) {
1336         case NFPROTO_IPV4:
1337                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1338                 break;
1339         case NFPROTO_IPV6:
1340                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1341                 break;
1342         }
1343
1344         return ret;
1345 }
1346
1347 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1348         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1349 };
1350
1351 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1352                                        struct nf_conntrack_tuple *tuple,
1353                                        u_int32_t flags)
1354 {
1355         const struct nf_conntrack_l4proto *l4proto;
1356         struct nlattr *tb[CTA_PROTO_MAX+1];
1357         int ret = 0;
1358
1359         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1360                                           proto_nla_policy, NULL);
1361         if (ret < 0)
1362                 return ret;
1363
1364         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1365                 return 0;
1366
1367         if (!tb[CTA_PROTO_NUM])
1368                 return -EINVAL;
1369
1370         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1371
1372         rcu_read_lock();
1373         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1374
1375         if (likely(l4proto->nlattr_to_tuple)) {
1376                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1377                                                      l4proto->nla_policy,
1378                                                      NULL);
1379                 if (ret == 0)
1380                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1381         }
1382
1383         rcu_read_unlock();
1384
1385         return ret;
1386 }
1387
1388 static int
1389 ctnetlink_parse_zone(const struct nlattr *attr,
1390                      struct nf_conntrack_zone *zone)
1391 {
1392         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1393                         NF_CT_DEFAULT_ZONE_DIR, 0);
1394 #ifdef CONFIG_NF_CONNTRACK_ZONES
1395         if (attr)
1396                 zone->id = ntohs(nla_get_be16(attr));
1397 #else
1398         if (attr)
1399                 return -EOPNOTSUPP;
1400 #endif
1401         return 0;
1402 }
1403
1404 static int
1405 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1406                            struct nf_conntrack_zone *zone)
1407 {
1408         int ret;
1409
1410         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1411                 return -EINVAL;
1412
1413         ret = ctnetlink_parse_zone(attr, zone);
1414         if (ret < 0)
1415                 return ret;
1416
1417         if (type == CTA_TUPLE_REPLY)
1418                 zone->dir = NF_CT_ZONE_DIR_REPL;
1419         else
1420                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1421
1422         return 0;
1423 }
1424
1425 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1426         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1427         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1428         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1429 };
1430
1431 #define CTA_FILTER_F_ALL_CTA_PROTO \
1432   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1433    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1434    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1435    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1436    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1437    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1438    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1439    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1440
1441 static int
1442 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1443                               struct nf_conntrack_tuple *tuple, u32 type,
1444                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1445                               u_int32_t flags)
1446 {
1447         struct nlattr *tb[CTA_TUPLE_MAX+1];
1448         int err;
1449
1450         memset(tuple, 0, sizeof(*tuple));
1451
1452         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1453                                           tuple_nla_policy, NULL);
1454         if (err < 0)
1455                 return err;
1456
1457         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1458                 return -EOPNOTSUPP;
1459         tuple->src.l3num = l3num;
1460
1461         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1462             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1463                 if (!tb[CTA_TUPLE_IP])
1464                         return -EINVAL;
1465
1466                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1467                 if (err < 0)
1468                         return err;
1469         }
1470
1471         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1472                 if (!tb[CTA_TUPLE_PROTO])
1473                         return -EINVAL;
1474
1475                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1476                 if (err < 0)
1477                         return err;
1478         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1479                 /* Can't manage proto flags without a protonum  */
1480                 return -EINVAL;
1481         }
1482
1483         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1484                 if (!zone)
1485                         return -EINVAL;
1486
1487                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1488                                                  type, zone);
1489                 if (err < 0)
1490                         return err;
1491         }
1492
1493         /* orig and expect tuples get DIR_ORIGINAL */
1494         if (type == CTA_TUPLE_REPLY)
1495                 tuple->dst.dir = IP_CT_DIR_REPLY;
1496         else
1497                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1498
1499         return 0;
1500 }
1501
1502 static int
1503 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1504                       struct nf_conntrack_tuple *tuple, u32 type,
1505                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1506 {
1507         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1508                                             CTA_FILTER_FLAG(ALL));
1509 }
1510
1511 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1512         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1513                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1514 };
1515
1516 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1517                                 struct nlattr **helpinfo)
1518 {
1519         int err;
1520         struct nlattr *tb[CTA_HELP_MAX+1];
1521
1522         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1523                                           help_nla_policy, NULL);
1524         if (err < 0)
1525                 return err;
1526
1527         if (!tb[CTA_HELP_NAME])
1528                 return -EINVAL;
1529
1530         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1531
1532         if (tb[CTA_HELP_INFO])
1533                 *helpinfo = tb[CTA_HELP_INFO];
1534
1535         return 0;
1536 }
1537
1538 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1539         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1540         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1541         [CTA_STATUS]            = { .type = NLA_U32 },
1542         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1543         [CTA_HELP]              = { .type = NLA_NESTED },
1544         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1545         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1546         [CTA_MARK]              = { .type = NLA_U32 },
1547         [CTA_ID]                = { .type = NLA_U32 },
1548         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1549         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1550         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1551         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1552         [CTA_ZONE]              = { .type = NLA_U16 },
1553         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1554         [CTA_LABELS]            = { .type = NLA_BINARY,
1555                                     .len = NF_CT_LABELS_MAX_SIZE },
1556         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1557                                     .len = NF_CT_LABELS_MAX_SIZE },
1558         [CTA_FILTER]            = { .type = NLA_NESTED },
1559         [CTA_STATUS_MASK]       = { .type = NLA_U32 },
1560 };
1561
1562 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1563 {
1564         return ctnetlink_filter_match(ct, data);
1565 }
1566
1567 static int ctnetlink_flush_conntrack(struct net *net,
1568                                      const struct nlattr * const cda[],
1569                                      u32 portid, int report, u8 family)
1570 {
1571         struct ctnetlink_filter *filter = NULL;
1572         struct nf_ct_iter_data iter = {
1573                 .net            = net,
1574                 .portid         = portid,
1575                 .report         = report,
1576         };
1577
1578         if (ctnetlink_needs_filter(family, cda)) {
1579                 filter = ctnetlink_alloc_filter(cda, family);
1580                 if (IS_ERR(filter))
1581                         return PTR_ERR(filter);
1582
1583                 iter.data = filter;
1584         }
1585
1586         nf_ct_iterate_cleanup_net(ctnetlink_flush_iterate, &iter);
1587         kfree(filter);
1588
1589         return 0;
1590 }
1591
1592 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1593                                    const struct nfnl_info *info,
1594                                    const struct nlattr * const cda[])
1595 {
1596         u8 family = info->nfmsg->nfgen_family;
1597         struct nf_conntrack_tuple_hash *h;
1598         struct nf_conntrack_tuple tuple;
1599         struct nf_conntrack_zone zone;
1600         struct nf_conn *ct;
1601         int err;
1602
1603         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1604         if (err < 0)
1605                 return err;
1606
1607         if (cda[CTA_TUPLE_ORIG] && !cda[CTA_FILTER])
1608                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1609                                             family, &zone);
1610         else if (cda[CTA_TUPLE_REPLY] && !cda[CTA_FILTER])
1611                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1612                                             family, &zone);
1613         else {
1614                 u8 u3 = info->nfmsg->version || cda[CTA_FILTER] ? family : AF_UNSPEC;
1615
1616                 return ctnetlink_flush_conntrack(info->net, cda,
1617                                                  NETLINK_CB(skb).portid,
1618                                                  nlmsg_report(info->nlh), u3);
1619         }
1620
1621         if (err < 0)
1622                 return err;
1623
1624         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1625         if (!h)
1626                 return -ENOENT;
1627
1628         ct = nf_ct_tuplehash_to_ctrack(h);
1629
1630         if (cda[CTA_ID]) {
1631                 __be32 id = nla_get_be32(cda[CTA_ID]);
1632
1633                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1634                         nf_ct_put(ct);
1635                         return -ENOENT;
1636                 }
1637         }
1638
1639         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1640         nf_ct_put(ct);
1641
1642         return 0;
1643 }
1644
1645 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1646                                    const struct nfnl_info *info,
1647                                    const struct nlattr * const cda[])
1648 {
1649         u_int8_t u3 = info->nfmsg->nfgen_family;
1650         struct nf_conntrack_tuple_hash *h;
1651         struct nf_conntrack_tuple tuple;
1652         struct nf_conntrack_zone zone;
1653         struct sk_buff *skb2;
1654         struct nf_conn *ct;
1655         int err;
1656
1657         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1658                 struct netlink_dump_control c = {
1659                         .start = ctnetlink_start,
1660                         .dump = ctnetlink_dump_table,
1661                         .done = ctnetlink_done,
1662                         .data = (void *)cda,
1663                 };
1664
1665                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1666         }
1667
1668         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1669         if (err < 0)
1670                 return err;
1671
1672         if (cda[CTA_TUPLE_ORIG])
1673                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1674                                             u3, &zone);
1675         else if (cda[CTA_TUPLE_REPLY])
1676                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1677                                             u3, &zone);
1678         else
1679                 return -EINVAL;
1680
1681         if (err < 0)
1682                 return err;
1683
1684         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1685         if (!h)
1686                 return -ENOENT;
1687
1688         ct = nf_ct_tuplehash_to_ctrack(h);
1689
1690         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1691         if (!skb2) {
1692                 nf_ct_put(ct);
1693                 return -ENOMEM;
1694         }
1695
1696         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1697                                   info->nlh->nlmsg_seq,
1698                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1699                                   true, 0);
1700         nf_ct_put(ct);
1701         if (err <= 0) {
1702                 kfree_skb(skb2);
1703                 return -ENOMEM;
1704         }
1705
1706         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1707 }
1708
1709 static int ctnetlink_done_list(struct netlink_callback *cb)
1710 {
1711         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1712
1713         if (ctx->last)
1714                 nf_ct_put(ctx->last);
1715
1716         return 0;
1717 }
1718
1719 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1720 static int ctnetlink_dump_one_entry(struct sk_buff *skb,
1721                                     struct netlink_callback *cb,
1722                                     struct nf_conn *ct,
1723                                     bool dying)
1724 {
1725         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1726         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1727         u8 l3proto = nfmsg->nfgen_family;
1728         int res;
1729
1730         if (l3proto && nf_ct_l3num(ct) != l3proto)
1731                 return 0;
1732
1733         if (ctx->last) {
1734                 if (ct != ctx->last)
1735                         return 0;
1736
1737                 ctx->last = NULL;
1738         }
1739
1740         /* We can't dump extension info for the unconfirmed
1741          * list because unconfirmed conntracks can have
1742          * ct->ext reallocated (and thus freed).
1743          *
1744          * In the dying list case ct->ext can't be free'd
1745          * until after we drop pcpu->lock.
1746          */
1747         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1748                                   cb->nlh->nlmsg_seq,
1749                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1750                                   ct, dying, 0);
1751         if (res < 0) {
1752                 if (!refcount_inc_not_zero(&ct->ct_general.use))
1753                         return 0;
1754
1755                 ctx->last = ct;
1756         }
1757
1758         return res;
1759 }
1760 #endif
1761
1762 static int
1763 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1764 {
1765         return 0;
1766 }
1767
1768 static int
1769 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1770 {
1771         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1772         struct nf_conn *last = ctx->last;
1773 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1774         const struct net *net = sock_net(skb->sk);
1775         struct nf_conntrack_net_ecache *ecache_net;
1776         struct nf_conntrack_tuple_hash *h;
1777         struct hlist_nulls_node *n;
1778 #endif
1779
1780         if (ctx->done)
1781                 return 0;
1782
1783         ctx->last = NULL;
1784
1785 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1786         ecache_net = nf_conn_pernet_ecache(net);
1787         spin_lock_bh(&ecache_net->dying_lock);
1788
1789         hlist_nulls_for_each_entry(h, n, &ecache_net->dying_list, hnnode) {
1790                 struct nf_conn *ct;
1791                 int res;
1792
1793                 ct = nf_ct_tuplehash_to_ctrack(h);
1794                 if (last && last != ct)
1795                         continue;
1796
1797                 res = ctnetlink_dump_one_entry(skb, cb, ct, true);
1798                 if (res < 0) {
1799                         spin_unlock_bh(&ecache_net->dying_lock);
1800                         nf_ct_put(last);
1801                         return skb->len;
1802                 }
1803
1804                 nf_ct_put(last);
1805                 last = NULL;
1806         }
1807
1808         spin_unlock_bh(&ecache_net->dying_lock);
1809 #endif
1810         ctx->done = true;
1811         nf_ct_put(last);
1812
1813         return skb->len;
1814 }
1815
1816 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1817                                   const struct nfnl_info *info,
1818                                   const struct nlattr * const cda[])
1819 {
1820         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1821                 struct netlink_dump_control c = {
1822                         .dump = ctnetlink_dump_dying,
1823                         .done = ctnetlink_done_list,
1824                 };
1825                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1826         }
1827
1828         return -EOPNOTSUPP;
1829 }
1830
1831 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1832                                         const struct nfnl_info *info,
1833                                         const struct nlattr * const cda[])
1834 {
1835         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1836                 struct netlink_dump_control c = {
1837                         .dump = ctnetlink_dump_unconfirmed,
1838                         .done = ctnetlink_done_list,
1839                 };
1840                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1841         }
1842
1843         return -EOPNOTSUPP;
1844 }
1845
1846 #if IS_ENABLED(CONFIG_NF_NAT)
1847 static int
1848 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1849                           enum nf_nat_manip_type manip,
1850                           const struct nlattr *attr)
1851         __must_hold(RCU)
1852 {
1853         const struct nf_nat_hook *nat_hook;
1854         int err;
1855
1856         nat_hook = rcu_dereference(nf_nat_hook);
1857         if (!nat_hook) {
1858 #ifdef CONFIG_MODULES
1859                 rcu_read_unlock();
1860                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1861                 if (request_module("nf-nat") < 0) {
1862                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1863                         rcu_read_lock();
1864                         return -EOPNOTSUPP;
1865                 }
1866                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1867                 rcu_read_lock();
1868                 nat_hook = rcu_dereference(nf_nat_hook);
1869                 if (nat_hook)
1870                         return -EAGAIN;
1871 #endif
1872                 return -EOPNOTSUPP;
1873         }
1874
1875         err = nat_hook->parse_nat_setup(ct, manip, attr);
1876         if (err == -EAGAIN) {
1877 #ifdef CONFIG_MODULES
1878                 rcu_read_unlock();
1879                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1880                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1881                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1882                         rcu_read_lock();
1883                         return -EOPNOTSUPP;
1884                 }
1885                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1886                 rcu_read_lock();
1887 #else
1888                 err = -EOPNOTSUPP;
1889 #endif
1890         }
1891         return err;
1892 }
1893 #endif
1894
1895 static int
1896 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1897 {
1898         return nf_ct_change_status_common(ct, ntohl(nla_get_be32(cda[CTA_STATUS])));
1899 }
1900
1901 static int
1902 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1903 {
1904 #if IS_ENABLED(CONFIG_NF_NAT)
1905         int ret;
1906
1907         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1908                 return 0;
1909
1910         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1911                                         cda[CTA_NAT_DST]);
1912         if (ret < 0)
1913                 return ret;
1914
1915         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1916                                          cda[CTA_NAT_SRC]);
1917 #else
1918         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1919                 return 0;
1920         return -EOPNOTSUPP;
1921 #endif
1922 }
1923
1924 static int ctnetlink_change_helper(struct nf_conn *ct,
1925                                    const struct nlattr * const cda[])
1926 {
1927         struct nf_conntrack_helper *helper;
1928         struct nf_conn_help *help = nfct_help(ct);
1929         char *helpname = NULL;
1930         struct nlattr *helpinfo = NULL;
1931         int err;
1932
1933         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1934         if (err < 0)
1935                 return err;
1936
1937         /* don't change helper of sibling connections */
1938         if (ct->master) {
1939                 /* If we try to change the helper to the same thing twice,
1940                  * treat the second attempt as a no-op instead of returning
1941                  * an error.
1942                  */
1943                 err = -EBUSY;
1944                 if (help) {
1945                         rcu_read_lock();
1946                         helper = rcu_dereference(help->helper);
1947                         if (helper && !strcmp(helper->name, helpname))
1948                                 err = 0;
1949                         rcu_read_unlock();
1950                 }
1951
1952                 return err;
1953         }
1954
1955         if (!strcmp(helpname, "")) {
1956                 if (help && help->helper) {
1957                         /* we had a helper before ... */
1958                         nf_ct_remove_expectations(ct);
1959                         RCU_INIT_POINTER(help->helper, NULL);
1960                 }
1961
1962                 return 0;
1963         }
1964
1965         rcu_read_lock();
1966         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1967                                             nf_ct_protonum(ct));
1968         if (helper == NULL) {
1969                 rcu_read_unlock();
1970                 return -EOPNOTSUPP;
1971         }
1972
1973         if (help) {
1974                 if (rcu_access_pointer(help->helper) == helper) {
1975                         /* update private helper data if allowed. */
1976                         if (helper->from_nlattr)
1977                                 helper->from_nlattr(helpinfo, ct);
1978                         err = 0;
1979                 } else
1980                         err = -EBUSY;
1981         } else {
1982                 /* we cannot set a helper for an existing conntrack */
1983                 err = -EOPNOTSUPP;
1984         }
1985
1986         rcu_read_unlock();
1987         return err;
1988 }
1989
1990 static int ctnetlink_change_timeout(struct nf_conn *ct,
1991                                     const struct nlattr * const cda[])
1992 {
1993         return __nf_ct_change_timeout(ct, (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ);
1994 }
1995
1996 #if defined(CONFIG_NF_CONNTRACK_MARK)
1997 static void ctnetlink_change_mark(struct nf_conn *ct,
1998                                     const struct nlattr * const cda[])
1999 {
2000         u32 mark, newmark, mask = 0;
2001
2002         if (cda[CTA_MARK_MASK])
2003                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2004
2005         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2006         newmark = (READ_ONCE(ct->mark) & mask) ^ mark;
2007         if (newmark != READ_ONCE(ct->mark))
2008                 WRITE_ONCE(ct->mark, newmark);
2009 }
2010 #endif
2011
2012 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2013         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2014         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2015         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2016 };
2017
2018 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2019                                       const struct nlattr * const cda[])
2020 {
2021         const struct nlattr *attr = cda[CTA_PROTOINFO];
2022         const struct nf_conntrack_l4proto *l4proto;
2023         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2024         int err = 0;
2025
2026         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2027                                           protoinfo_policy, NULL);
2028         if (err < 0)
2029                 return err;
2030
2031         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2032         if (l4proto->from_nlattr)
2033                 err = l4proto->from_nlattr(tb, ct);
2034
2035         return err;
2036 }
2037
2038 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2039         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2040         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2041         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2042 };
2043
2044 static int change_seq_adj(struct nf_ct_seqadj *seq,
2045                           const struct nlattr * const attr)
2046 {
2047         int err;
2048         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2049
2050         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2051                                           seqadj_policy, NULL);
2052         if (err < 0)
2053                 return err;
2054
2055         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2056                 return -EINVAL;
2057
2058         seq->correction_pos =
2059                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2060
2061         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2062                 return -EINVAL;
2063
2064         seq->offset_before =
2065                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2066
2067         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2068                 return -EINVAL;
2069
2070         seq->offset_after =
2071                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2072
2073         return 0;
2074 }
2075
2076 static int
2077 ctnetlink_change_seq_adj(struct nf_conn *ct,
2078                          const struct nlattr * const cda[])
2079 {
2080         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2081         int ret = 0;
2082
2083         if (!seqadj)
2084                 return 0;
2085
2086         spin_lock_bh(&ct->lock);
2087         if (cda[CTA_SEQ_ADJ_ORIG]) {
2088                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2089                                      cda[CTA_SEQ_ADJ_ORIG]);
2090                 if (ret < 0)
2091                         goto err;
2092
2093                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2094         }
2095
2096         if (cda[CTA_SEQ_ADJ_REPLY]) {
2097                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2098                                      cda[CTA_SEQ_ADJ_REPLY]);
2099                 if (ret < 0)
2100                         goto err;
2101
2102                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2103         }
2104
2105         spin_unlock_bh(&ct->lock);
2106         return 0;
2107 err:
2108         spin_unlock_bh(&ct->lock);
2109         return ret;
2110 }
2111
2112 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2113         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2114         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2115         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2116 };
2117
2118 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2119                                      const struct nlattr * const cda[])
2120 {
2121         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2122         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2123         int err;
2124
2125         if (!synproxy)
2126                 return 0;
2127
2128         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2129                                           cda[CTA_SYNPROXY], synproxy_policy,
2130                                           NULL);
2131         if (err < 0)
2132                 return err;
2133
2134         if (!tb[CTA_SYNPROXY_ISN] ||
2135             !tb[CTA_SYNPROXY_ITS] ||
2136             !tb[CTA_SYNPROXY_TSOFF])
2137                 return -EINVAL;
2138
2139         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2140         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2141         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2142
2143         return 0;
2144 }
2145
2146 static int
2147 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2148 {
2149 #ifdef CONFIG_NF_CONNTRACK_LABELS
2150         size_t len = nla_len(cda[CTA_LABELS]);
2151         const void *mask = cda[CTA_LABELS_MASK];
2152
2153         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2154                 return -EINVAL;
2155
2156         if (mask) {
2157                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2158                     nla_len(cda[CTA_LABELS_MASK]) != len)
2159                         return -EINVAL;
2160                 mask = nla_data(cda[CTA_LABELS_MASK]);
2161         }
2162
2163         len /= sizeof(u32);
2164
2165         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2166 #else
2167         return -EOPNOTSUPP;
2168 #endif
2169 }
2170
2171 static int
2172 ctnetlink_change_conntrack(struct nf_conn *ct,
2173                            const struct nlattr * const cda[])
2174 {
2175         int err;
2176
2177         /* only allow NAT changes and master assignation for new conntracks */
2178         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2179                 return -EOPNOTSUPP;
2180
2181         if (cda[CTA_HELP]) {
2182                 err = ctnetlink_change_helper(ct, cda);
2183                 if (err < 0)
2184                         return err;
2185         }
2186
2187         if (cda[CTA_TIMEOUT]) {
2188                 err = ctnetlink_change_timeout(ct, cda);
2189                 if (err < 0)
2190                         return err;
2191         }
2192
2193         if (cda[CTA_STATUS]) {
2194                 err = ctnetlink_change_status(ct, cda);
2195                 if (err < 0)
2196                         return err;
2197         }
2198
2199         if (cda[CTA_PROTOINFO]) {
2200                 err = ctnetlink_change_protoinfo(ct, cda);
2201                 if (err < 0)
2202                         return err;
2203         }
2204
2205 #if defined(CONFIG_NF_CONNTRACK_MARK)
2206         if (cda[CTA_MARK])
2207                 ctnetlink_change_mark(ct, cda);
2208 #endif
2209
2210         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2211                 err = ctnetlink_change_seq_adj(ct, cda);
2212                 if (err < 0)
2213                         return err;
2214         }
2215
2216         if (cda[CTA_SYNPROXY]) {
2217                 err = ctnetlink_change_synproxy(ct, cda);
2218                 if (err < 0)
2219                         return err;
2220         }
2221
2222         if (cda[CTA_LABELS]) {
2223                 err = ctnetlink_attach_labels(ct, cda);
2224                 if (err < 0)
2225                         return err;
2226         }
2227
2228         return 0;
2229 }
2230
2231 static struct nf_conn *
2232 ctnetlink_create_conntrack(struct net *net,
2233                            const struct nf_conntrack_zone *zone,
2234                            const struct nlattr * const cda[],
2235                            struct nf_conntrack_tuple *otuple,
2236                            struct nf_conntrack_tuple *rtuple,
2237                            u8 u3)
2238 {
2239         struct nf_conn *ct;
2240         int err = -EINVAL;
2241         struct nf_conntrack_helper *helper;
2242         struct nf_conn_tstamp *tstamp;
2243         u64 timeout;
2244
2245         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2246         if (IS_ERR(ct))
2247                 return ERR_PTR(-ENOMEM);
2248
2249         if (!cda[CTA_TIMEOUT])
2250                 goto err1;
2251
2252         rcu_read_lock();
2253         if (cda[CTA_HELP]) {
2254                 char *helpname = NULL;
2255                 struct nlattr *helpinfo = NULL;
2256
2257                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2258                 if (err < 0)
2259                         goto err2;
2260
2261                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2262                                                     nf_ct_protonum(ct));
2263                 if (helper == NULL) {
2264                         rcu_read_unlock();
2265 #ifdef CONFIG_MODULES
2266                         if (request_module("nfct-helper-%s", helpname) < 0) {
2267                                 err = -EOPNOTSUPP;
2268                                 goto err1;
2269                         }
2270
2271                         rcu_read_lock();
2272                         helper = __nf_conntrack_helper_find(helpname,
2273                                                             nf_ct_l3num(ct),
2274                                                             nf_ct_protonum(ct));
2275                         if (helper) {
2276                                 err = -EAGAIN;
2277                                 goto err2;
2278                         }
2279                         rcu_read_unlock();
2280 #endif
2281                         err = -EOPNOTSUPP;
2282                         goto err1;
2283                 } else {
2284                         struct nf_conn_help *help;
2285
2286                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2287                         if (help == NULL) {
2288                                 err = -ENOMEM;
2289                                 goto err2;
2290                         }
2291                         /* set private helper data if allowed. */
2292                         if (helper->from_nlattr)
2293                                 helper->from_nlattr(helpinfo, ct);
2294
2295                         /* disable helper auto-assignment for this entry */
2296                         ct->status |= IPS_HELPER;
2297                         RCU_INIT_POINTER(help->helper, helper);
2298                 }
2299         }
2300
2301         err = ctnetlink_setup_nat(ct, cda);
2302         if (err < 0)
2303                 goto err2;
2304
2305         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2306         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2307         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2308         nf_ct_labels_ext_add(ct);
2309         nfct_seqadj_ext_add(ct);
2310         nfct_synproxy_ext_add(ct);
2311
2312         /* we must add conntrack extensions before confirmation. */
2313         ct->status |= IPS_CONFIRMED;
2314
2315         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2316         __nf_ct_set_timeout(ct, timeout);
2317
2318         if (cda[CTA_STATUS]) {
2319                 err = ctnetlink_change_status(ct, cda);
2320                 if (err < 0)
2321                         goto err2;
2322         }
2323
2324         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2325                 err = ctnetlink_change_seq_adj(ct, cda);
2326                 if (err < 0)
2327                         goto err2;
2328         }
2329
2330         memset(&ct->proto, 0, sizeof(ct->proto));
2331         if (cda[CTA_PROTOINFO]) {
2332                 err = ctnetlink_change_protoinfo(ct, cda);
2333                 if (err < 0)
2334                         goto err2;
2335         }
2336
2337         if (cda[CTA_SYNPROXY]) {
2338                 err = ctnetlink_change_synproxy(ct, cda);
2339                 if (err < 0)
2340                         goto err2;
2341         }
2342
2343 #if defined(CONFIG_NF_CONNTRACK_MARK)
2344         if (cda[CTA_MARK])
2345                 ctnetlink_change_mark(ct, cda);
2346 #endif
2347
2348         /* setup master conntrack: this is a confirmed expectation */
2349         if (cda[CTA_TUPLE_MASTER]) {
2350                 struct nf_conntrack_tuple master;
2351                 struct nf_conntrack_tuple_hash *master_h;
2352                 struct nf_conn *master_ct;
2353
2354                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2355                                             u3, NULL);
2356                 if (err < 0)
2357                         goto err2;
2358
2359                 master_h = nf_conntrack_find_get(net, zone, &master);
2360                 if (master_h == NULL) {
2361                         err = -ENOENT;
2362                         goto err2;
2363                 }
2364                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2365                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2366                 ct->master = master_ct;
2367         }
2368         tstamp = nf_conn_tstamp_find(ct);
2369         if (tstamp)
2370                 tstamp->start = ktime_get_real_ns();
2371
2372         err = nf_conntrack_hash_check_insert(ct);
2373         if (err < 0)
2374                 goto err3;
2375
2376         rcu_read_unlock();
2377
2378         return ct;
2379
2380 err3:
2381         if (ct->master)
2382                 nf_ct_put(ct->master);
2383 err2:
2384         rcu_read_unlock();
2385 err1:
2386         nf_conntrack_free(ct);
2387         return ERR_PTR(err);
2388 }
2389
2390 static int ctnetlink_new_conntrack(struct sk_buff *skb,
2391                                    const struct nfnl_info *info,
2392                                    const struct nlattr * const cda[])
2393 {
2394         struct nf_conntrack_tuple otuple, rtuple;
2395         struct nf_conntrack_tuple_hash *h = NULL;
2396         u_int8_t u3 = info->nfmsg->nfgen_family;
2397         struct nf_conntrack_zone zone;
2398         struct nf_conn *ct;
2399         int err;
2400
2401         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2402         if (err < 0)
2403                 return err;
2404
2405         if (cda[CTA_TUPLE_ORIG]) {
2406                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2407                                             u3, &zone);
2408                 if (err < 0)
2409                         return err;
2410         }
2411
2412         if (cda[CTA_TUPLE_REPLY]) {
2413                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2414                                             u3, &zone);
2415                 if (err < 0)
2416                         return err;
2417         }
2418
2419         if (cda[CTA_TUPLE_ORIG])
2420                 h = nf_conntrack_find_get(info->net, &zone, &otuple);
2421         else if (cda[CTA_TUPLE_REPLY])
2422                 h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2423
2424         if (h == NULL) {
2425                 err = -ENOENT;
2426                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2427                         enum ip_conntrack_events events;
2428
2429                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2430                                 return -EINVAL;
2431                         if (otuple.dst.protonum != rtuple.dst.protonum)
2432                                 return -EINVAL;
2433
2434                         ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2435                                                         &otuple, &rtuple, u3);
2436                         if (IS_ERR(ct))
2437                                 return PTR_ERR(ct);
2438
2439                         err = 0;
2440                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2441                                 events = 1 << IPCT_RELATED;
2442                         else
2443                                 events = 1 << IPCT_NEW;
2444
2445                         if (cda[CTA_LABELS] &&
2446                             ctnetlink_attach_labels(ct, cda) == 0)
2447                                 events |= (1 << IPCT_LABEL);
2448
2449                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2450                                                       (1 << IPCT_ASSURED) |
2451                                                       (1 << IPCT_HELPER) |
2452                                                       (1 << IPCT_PROTOINFO) |
2453                                                       (1 << IPCT_SEQADJ) |
2454                                                       (1 << IPCT_MARK) |
2455                                                       (1 << IPCT_SYNPROXY) |
2456                                                       events,
2457                                                       ct, NETLINK_CB(skb).portid,
2458                                                       nlmsg_report(info->nlh));
2459                         nf_ct_put(ct);
2460                 }
2461
2462                 return err;
2463         }
2464         /* implicit 'else' */
2465
2466         err = -EEXIST;
2467         ct = nf_ct_tuplehash_to_ctrack(h);
2468         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2469                 err = ctnetlink_change_conntrack(ct, cda);
2470                 if (err == 0) {
2471                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2472                                                       (1 << IPCT_ASSURED) |
2473                                                       (1 << IPCT_HELPER) |
2474                                                       (1 << IPCT_LABEL) |
2475                                                       (1 << IPCT_PROTOINFO) |
2476                                                       (1 << IPCT_SEQADJ) |
2477                                                       (1 << IPCT_MARK) |
2478                                                       (1 << IPCT_SYNPROXY),
2479                                                       ct, NETLINK_CB(skb).portid,
2480                                                       nlmsg_report(info->nlh));
2481                 }
2482         }
2483
2484         nf_ct_put(ct);
2485         return err;
2486 }
2487
2488 static int
2489 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2490                                 __u16 cpu, const struct ip_conntrack_stat *st)
2491 {
2492         struct nlmsghdr *nlh;
2493         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2494
2495         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2496                               IPCTNL_MSG_CT_GET_STATS_CPU);
2497         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2498                            NFNETLINK_V0, htons(cpu));
2499         if (!nlh)
2500                 goto nlmsg_failure;
2501
2502         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2503             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2504             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2505             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2506                                 htonl(st->insert_failed)) ||
2507             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2508             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2509             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2510             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2511                                 htonl(st->search_restart)) ||
2512             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2513                                 htonl(st->clash_resolve)) ||
2514             nla_put_be32(skb, CTA_STATS_CHAIN_TOOLONG,
2515                          htonl(st->chaintoolong)))
2516                 goto nla_put_failure;
2517
2518         nlmsg_end(skb, nlh);
2519         return skb->len;
2520
2521 nla_put_failure:
2522 nlmsg_failure:
2523         nlmsg_cancel(skb, nlh);
2524         return -1;
2525 }
2526
2527 static int
2528 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2529 {
2530         int cpu;
2531         struct net *net = sock_net(skb->sk);
2532
2533         if (cb->args[0] == nr_cpu_ids)
2534                 return 0;
2535
2536         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2537                 const struct ip_conntrack_stat *st;
2538
2539                 if (!cpu_possible(cpu))
2540                         continue;
2541
2542                 st = per_cpu_ptr(net->ct.stat, cpu);
2543                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2544                                                     NETLINK_CB(cb->skb).portid,
2545                                                     cb->nlh->nlmsg_seq,
2546                                                     cpu, st) < 0)
2547                                 break;
2548         }
2549         cb->args[0] = cpu;
2550
2551         return skb->len;
2552 }
2553
2554 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2555                                  const struct nfnl_info *info,
2556                                  const struct nlattr * const cda[])
2557 {
2558         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2559                 struct netlink_dump_control c = {
2560                         .dump = ctnetlink_ct_stat_cpu_dump,
2561                 };
2562                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2563         }
2564
2565         return 0;
2566 }
2567
2568 static int
2569 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2570                             struct net *net)
2571 {
2572         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2573         unsigned int nr_conntracks;
2574         struct nlmsghdr *nlh;
2575
2576         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2577         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2578                            NFNETLINK_V0, 0);
2579         if (!nlh)
2580                 goto nlmsg_failure;
2581
2582         nr_conntracks = nf_conntrack_count(net);
2583         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2584                 goto nla_put_failure;
2585
2586         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2587                 goto nla_put_failure;
2588
2589         nlmsg_end(skb, nlh);
2590         return skb->len;
2591
2592 nla_put_failure:
2593 nlmsg_failure:
2594         nlmsg_cancel(skb, nlh);
2595         return -1;
2596 }
2597
2598 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2599                              const struct nlattr * const cda[])
2600 {
2601         struct sk_buff *skb2;
2602         int err;
2603
2604         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2605         if (skb2 == NULL)
2606                 return -ENOMEM;
2607
2608         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2609                                           info->nlh->nlmsg_seq,
2610                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2611                                           sock_net(skb->sk));
2612         if (err <= 0) {
2613                 kfree_skb(skb2);
2614                 return -ENOMEM;
2615         }
2616
2617         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2618 }
2619
2620 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2621         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2622         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2623         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2624         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2625         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2626         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2627                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2628         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2629         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2630         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2631         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2632         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2633 };
2634
2635 static struct nf_conntrack_expect *
2636 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2637                        struct nf_conntrack_helper *helper,
2638                        struct nf_conntrack_tuple *tuple,
2639                        struct nf_conntrack_tuple *mask);
2640
2641 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2642 static size_t
2643 ctnetlink_glue_build_size(const struct nf_conn *ct)
2644 {
2645         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2646                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2647                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2648                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2649                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2650                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2651                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2652                + nla_total_size(0) /* CTA_PROTOINFO */
2653                + nla_total_size(0) /* CTA_HELP */
2654                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2655                + ctnetlink_secctx_size(ct)
2656                + ctnetlink_acct_size(ct)
2657                + ctnetlink_timestamp_size(ct)
2658 #if IS_ENABLED(CONFIG_NF_NAT)
2659                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2660                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2661 #endif
2662 #ifdef CONFIG_NF_CONNTRACK_MARK
2663                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2664 #endif
2665 #ifdef CONFIG_NF_CONNTRACK_ZONES
2666                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2667 #endif
2668                + ctnetlink_proto_size(ct)
2669                ;
2670 }
2671
2672 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2673 {
2674         const struct nf_conntrack_zone *zone;
2675         struct nlattr *nest_parms;
2676
2677         zone = nf_ct_zone(ct);
2678
2679         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2680         if (!nest_parms)
2681                 goto nla_put_failure;
2682         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2683                 goto nla_put_failure;
2684         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2685                                    NF_CT_ZONE_DIR_ORIG) < 0)
2686                 goto nla_put_failure;
2687         nla_nest_end(skb, nest_parms);
2688
2689         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2690         if (!nest_parms)
2691                 goto nla_put_failure;
2692         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2693                 goto nla_put_failure;
2694         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2695                                    NF_CT_ZONE_DIR_REPL) < 0)
2696                 goto nla_put_failure;
2697         nla_nest_end(skb, nest_parms);
2698
2699         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2700                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2701                 goto nla_put_failure;
2702
2703         if (ctnetlink_dump_id(skb, ct) < 0)
2704                 goto nla_put_failure;
2705
2706         if (ctnetlink_dump_status(skb, ct) < 0)
2707                 goto nla_put_failure;
2708
2709         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2710                 goto nla_put_failure;
2711
2712         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2713                 goto nla_put_failure;
2714
2715         if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2716             ctnetlink_dump_timestamp(skb, ct) < 0)
2717                 goto nla_put_failure;
2718
2719         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2720                 goto nla_put_failure;
2721
2722 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2723         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2724                 goto nla_put_failure;
2725 #endif
2726         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2727                 goto nla_put_failure;
2728
2729         if ((ct->status & IPS_SEQ_ADJUST) &&
2730             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2731                 goto nla_put_failure;
2732
2733         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2734                 goto nla_put_failure;
2735
2736 #ifdef CONFIG_NF_CONNTRACK_MARK
2737         if (ctnetlink_dump_mark(skb, ct, true) < 0)
2738                 goto nla_put_failure;
2739 #endif
2740         if (ctnetlink_dump_labels(skb, ct) < 0)
2741                 goto nla_put_failure;
2742         return 0;
2743
2744 nla_put_failure:
2745         return -ENOSPC;
2746 }
2747
2748 static int
2749 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2750                      enum ip_conntrack_info ctinfo,
2751                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2752 {
2753         struct nlattr *nest_parms;
2754
2755         nest_parms = nla_nest_start(skb, ct_attr);
2756         if (!nest_parms)
2757                 goto nla_put_failure;
2758
2759         if (__ctnetlink_glue_build(skb, ct) < 0)
2760                 goto nla_put_failure;
2761
2762         nla_nest_end(skb, nest_parms);
2763
2764         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2765                 goto nla_put_failure;
2766
2767         return 0;
2768
2769 nla_put_failure:
2770         return -ENOSPC;
2771 }
2772
2773 static int
2774 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2775 {
2776         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2777         unsigned long d = ct->status ^ status;
2778
2779         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2780                 /* SEEN_REPLY bit can only be set */
2781                 return -EBUSY;
2782
2783         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2784                 /* ASSURED bit can only be set */
2785                 return -EBUSY;
2786
2787         /* This check is less strict than ctnetlink_change_status()
2788          * because callers often flip IPS_EXPECTED bits when sending
2789          * an NFQA_CT attribute to the kernel.  So ignore the
2790          * unchangeable bits but do not error out. Also user programs
2791          * are allowed to clear the bits that they are allowed to change.
2792          */
2793         __nf_ct_change_status(ct, status, ~status);
2794         return 0;
2795 }
2796
2797 static int
2798 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2799 {
2800         int err;
2801
2802         if (cda[CTA_TIMEOUT]) {
2803                 err = ctnetlink_change_timeout(ct, cda);
2804                 if (err < 0)
2805                         return err;
2806         }
2807         if (cda[CTA_STATUS]) {
2808                 err = ctnetlink_update_status(ct, cda);
2809                 if (err < 0)
2810                         return err;
2811         }
2812         if (cda[CTA_HELP]) {
2813                 err = ctnetlink_change_helper(ct, cda);
2814                 if (err < 0)
2815                         return err;
2816         }
2817         if (cda[CTA_LABELS]) {
2818                 err = ctnetlink_attach_labels(ct, cda);
2819                 if (err < 0)
2820                         return err;
2821         }
2822 #if defined(CONFIG_NF_CONNTRACK_MARK)
2823         if (cda[CTA_MARK]) {
2824                 ctnetlink_change_mark(ct, cda);
2825         }
2826 #endif
2827         return 0;
2828 }
2829
2830 static int
2831 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2832 {
2833         struct nlattr *cda[CTA_MAX+1];
2834         int ret;
2835
2836         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2837                                           NULL);
2838         if (ret < 0)
2839                 return ret;
2840
2841         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2842 }
2843
2844 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2845                                     const struct nf_conn *ct,
2846                                     struct nf_conntrack_tuple *tuple,
2847                                     struct nf_conntrack_tuple *mask)
2848 {
2849         int err;
2850
2851         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2852                                     nf_ct_l3num(ct), NULL);
2853         if (err < 0)
2854                 return err;
2855
2856         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2857                                      nf_ct_l3num(ct), NULL);
2858 }
2859
2860 static int
2861 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2862                              u32 portid, u32 report)
2863 {
2864         struct nlattr *cda[CTA_EXPECT_MAX+1];
2865         struct nf_conntrack_tuple tuple, mask;
2866         struct nf_conntrack_helper *helper = NULL;
2867         struct nf_conntrack_expect *exp;
2868         int err;
2869
2870         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2871                                           exp_nla_policy, NULL);
2872         if (err < 0)
2873                 return err;
2874
2875         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2876                                        ct, &tuple, &mask);
2877         if (err < 0)
2878                 return err;
2879
2880         if (cda[CTA_EXPECT_HELP_NAME]) {
2881                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2882
2883                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2884                                                     nf_ct_protonum(ct));
2885                 if (helper == NULL)
2886                         return -EOPNOTSUPP;
2887         }
2888
2889         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2890                                      helper, &tuple, &mask);
2891         if (IS_ERR(exp))
2892                 return PTR_ERR(exp);
2893
2894         err = nf_ct_expect_related_report(exp, portid, report, 0);
2895         nf_ct_expect_put(exp);
2896         return err;
2897 }
2898
2899 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2900                                   enum ip_conntrack_info ctinfo, int diff)
2901 {
2902         if (!(ct->status & IPS_NAT_MASK))
2903                 return;
2904
2905         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2906 }
2907
2908 static const struct nfnl_ct_hook ctnetlink_glue_hook = {
2909         .build_size     = ctnetlink_glue_build_size,
2910         .build          = ctnetlink_glue_build,
2911         .parse          = ctnetlink_glue_parse,
2912         .attach_expect  = ctnetlink_glue_attach_expect,
2913         .seq_adjust     = ctnetlink_glue_seqadj,
2914 };
2915 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2916
2917 /***********************************************************************
2918  * EXPECT
2919  ***********************************************************************/
2920
2921 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2922                                     const struct nf_conntrack_tuple *tuple,
2923                                     u32 type)
2924 {
2925         struct nlattr *nest_parms;
2926
2927         nest_parms = nla_nest_start(skb, type);
2928         if (!nest_parms)
2929                 goto nla_put_failure;
2930         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2931                 goto nla_put_failure;
2932         nla_nest_end(skb, nest_parms);
2933
2934         return 0;
2935
2936 nla_put_failure:
2937         return -1;
2938 }
2939
2940 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2941                                    const struct nf_conntrack_tuple *tuple,
2942                                    const struct nf_conntrack_tuple_mask *mask)
2943 {
2944         const struct nf_conntrack_l4proto *l4proto;
2945         struct nf_conntrack_tuple m;
2946         struct nlattr *nest_parms;
2947         int ret;
2948
2949         memset(&m, 0xFF, sizeof(m));
2950         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2951         m.src.u.all = mask->src.u.all;
2952         m.src.l3num = tuple->src.l3num;
2953         m.dst.protonum = tuple->dst.protonum;
2954
2955         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2956         if (!nest_parms)
2957                 goto nla_put_failure;
2958
2959         rcu_read_lock();
2960         ret = ctnetlink_dump_tuples_ip(skb, &m);
2961         if (ret >= 0) {
2962                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2963                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2964         }
2965         rcu_read_unlock();
2966
2967         if (unlikely(ret < 0))
2968                 goto nla_put_failure;
2969
2970         nla_nest_end(skb, nest_parms);
2971
2972         return 0;
2973
2974 nla_put_failure:
2975         return -1;
2976 }
2977
2978 #if IS_ENABLED(CONFIG_NF_NAT)
2979 static const union nf_inet_addr any_addr;
2980 #endif
2981
2982 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2983 {
2984         static siphash_aligned_key_t exp_id_seed;
2985         unsigned long a, b, c, d;
2986
2987         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2988
2989         a = (unsigned long)exp;
2990         b = (unsigned long)exp->helper;
2991         c = (unsigned long)exp->master;
2992         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2993
2994 #ifdef CONFIG_64BIT
2995         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2996 #else
2997         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2998 #endif
2999 }
3000
3001 static int
3002 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3003                           const struct nf_conntrack_expect *exp)
3004 {
3005         struct nf_conn *master = exp->master;
3006         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3007         struct nf_conn_help *help;
3008 #if IS_ENABLED(CONFIG_NF_NAT)
3009         struct nlattr *nest_parms;
3010         struct nf_conntrack_tuple nat_tuple = {};
3011 #endif
3012         struct nf_ct_helper_expectfn *expfn;
3013
3014         if (timeout < 0)
3015                 timeout = 0;
3016
3017         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3018                 goto nla_put_failure;
3019         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3020                 goto nla_put_failure;
3021         if (ctnetlink_exp_dump_tuple(skb,
3022                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3023                                  CTA_EXPECT_MASTER) < 0)
3024                 goto nla_put_failure;
3025
3026 #if IS_ENABLED(CONFIG_NF_NAT)
3027         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3028             exp->saved_proto.all) {
3029                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3030                 if (!nest_parms)
3031                         goto nla_put_failure;
3032
3033                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3034                         goto nla_put_failure;
3035
3036                 nat_tuple.src.l3num = nf_ct_l3num(master);
3037                 nat_tuple.src.u3 = exp->saved_addr;
3038                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3039                 nat_tuple.src.u = exp->saved_proto;
3040
3041                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3042                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3043                         goto nla_put_failure;
3044                 nla_nest_end(skb, nest_parms);
3045         }
3046 #endif
3047         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3048             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3049             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3050             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3051                 goto nla_put_failure;
3052         help = nfct_help(master);
3053         if (help) {
3054                 struct nf_conntrack_helper *helper;
3055
3056                 helper = rcu_dereference(help->helper);
3057                 if (helper &&
3058                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3059                         goto nla_put_failure;
3060         }
3061         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3062         if (expfn != NULL &&
3063             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3064                 goto nla_put_failure;
3065
3066         return 0;
3067
3068 nla_put_failure:
3069         return -1;
3070 }
3071
3072 static int
3073 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3074                         int event, const struct nf_conntrack_expect *exp)
3075 {
3076         struct nlmsghdr *nlh;
3077         unsigned int flags = portid ? NLM_F_MULTI : 0;
3078
3079         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3080         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3081                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3082         if (!nlh)
3083                 goto nlmsg_failure;
3084
3085         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3086                 goto nla_put_failure;
3087
3088         nlmsg_end(skb, nlh);
3089         return skb->len;
3090
3091 nlmsg_failure:
3092 nla_put_failure:
3093         nlmsg_cancel(skb, nlh);
3094         return -1;
3095 }
3096
3097 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3098 static int
3099 ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3100 {
3101         struct nf_conntrack_expect *exp = item->exp;
3102         struct net *net = nf_ct_exp_net(exp);
3103         struct nlmsghdr *nlh;
3104         struct sk_buff *skb;
3105         unsigned int type, group;
3106         int flags = 0;
3107
3108         if (events & (1 << IPEXP_DESTROY)) {
3109                 type = IPCTNL_MSG_EXP_DELETE;
3110                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3111         } else if (events & (1 << IPEXP_NEW)) {
3112                 type = IPCTNL_MSG_EXP_NEW;
3113                 flags = NLM_F_CREATE|NLM_F_EXCL;
3114                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3115         } else
3116                 return 0;
3117
3118         if (!item->report && !nfnetlink_has_listeners(net, group))
3119                 return 0;
3120
3121         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3122         if (skb == NULL)
3123                 goto errout;
3124
3125         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3126         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3127                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3128         if (!nlh)
3129                 goto nlmsg_failure;
3130
3131         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3132                 goto nla_put_failure;
3133
3134         nlmsg_end(skb, nlh);
3135         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3136         return 0;
3137
3138 nla_put_failure:
3139         nlmsg_cancel(skb, nlh);
3140 nlmsg_failure:
3141         kfree_skb(skb);
3142 errout:
3143         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3144         return 0;
3145 }
3146 #endif
3147 static int ctnetlink_exp_done(struct netlink_callback *cb)
3148 {
3149         if (cb->args[1])
3150                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3151         return 0;
3152 }
3153
3154 static int
3155 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3156 {
3157         struct net *net = sock_net(skb->sk);
3158         struct nf_conntrack_expect *exp, *last;
3159         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3160         u_int8_t l3proto = nfmsg->nfgen_family;
3161
3162         rcu_read_lock();
3163         last = (struct nf_conntrack_expect *)cb->args[1];
3164         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3165 restart:
3166                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3167                                          hnode) {
3168                         if (l3proto && exp->tuple.src.l3num != l3proto)
3169                                 continue;
3170
3171                         if (!net_eq(nf_ct_net(exp->master), net))
3172                                 continue;
3173
3174                         if (cb->args[1]) {
3175                                 if (exp != last)
3176                                         continue;
3177                                 cb->args[1] = 0;
3178                         }
3179                         if (ctnetlink_exp_fill_info(skb,
3180                                                     NETLINK_CB(cb->skb).portid,
3181                                                     cb->nlh->nlmsg_seq,
3182                                                     IPCTNL_MSG_EXP_NEW,
3183                                                     exp) < 0) {
3184                                 if (!refcount_inc_not_zero(&exp->use))
3185                                         continue;
3186                                 cb->args[1] = (unsigned long)exp;
3187                                 goto out;
3188                         }
3189                 }
3190                 if (cb->args[1]) {
3191                         cb->args[1] = 0;
3192                         goto restart;
3193                 }
3194         }
3195 out:
3196         rcu_read_unlock();
3197         if (last)
3198                 nf_ct_expect_put(last);
3199
3200         return skb->len;
3201 }
3202
3203 static int
3204 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3205 {
3206         struct nf_conntrack_expect *exp, *last;
3207         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3208         struct nf_conn *ct = cb->data;
3209         struct nf_conn_help *help = nfct_help(ct);
3210         u_int8_t l3proto = nfmsg->nfgen_family;
3211
3212         if (cb->args[0])
3213                 return 0;
3214
3215         rcu_read_lock();
3216         last = (struct nf_conntrack_expect *)cb->args[1];
3217 restart:
3218         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3219                 if (l3proto && exp->tuple.src.l3num != l3proto)
3220                         continue;
3221                 if (cb->args[1]) {
3222                         if (exp != last)
3223                                 continue;
3224                         cb->args[1] = 0;
3225                 }
3226                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3227                                             cb->nlh->nlmsg_seq,
3228                                             IPCTNL_MSG_EXP_NEW,
3229                                             exp) < 0) {
3230                         if (!refcount_inc_not_zero(&exp->use))
3231                                 continue;
3232                         cb->args[1] = (unsigned long)exp;
3233                         goto out;
3234                 }
3235         }
3236         if (cb->args[1]) {
3237                 cb->args[1] = 0;
3238                 goto restart;
3239         }
3240         cb->args[0] = 1;
3241 out:
3242         rcu_read_unlock();
3243         if (last)
3244                 nf_ct_expect_put(last);
3245
3246         return skb->len;
3247 }
3248
3249 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3250                                  struct sk_buff *skb,
3251                                  const struct nlmsghdr *nlh,
3252                                  const struct nlattr * const cda[],
3253                                  struct netlink_ext_ack *extack)
3254 {
3255         int err;
3256         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3257         u_int8_t u3 = nfmsg->nfgen_family;
3258         struct nf_conntrack_tuple tuple;
3259         struct nf_conntrack_tuple_hash *h;
3260         struct nf_conn *ct;
3261         struct nf_conntrack_zone zone;
3262         struct netlink_dump_control c = {
3263                 .dump = ctnetlink_exp_ct_dump_table,
3264                 .done = ctnetlink_exp_done,
3265         };
3266
3267         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3268                                     u3, NULL);
3269         if (err < 0)
3270                 return err;
3271
3272         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3273         if (err < 0)
3274                 return err;
3275
3276         h = nf_conntrack_find_get(net, &zone, &tuple);
3277         if (!h)
3278                 return -ENOENT;
3279
3280         ct = nf_ct_tuplehash_to_ctrack(h);
3281         /* No expectation linked to this connection tracking. */
3282         if (!nfct_help(ct)) {
3283                 nf_ct_put(ct);
3284                 return 0;
3285         }
3286
3287         c.data = ct;
3288
3289         err = netlink_dump_start(ctnl, skb, nlh, &c);
3290         nf_ct_put(ct);
3291
3292         return err;
3293 }
3294
3295 static int ctnetlink_get_expect(struct sk_buff *skb,
3296                                 const struct nfnl_info *info,
3297                                 const struct nlattr * const cda[])
3298 {
3299         u_int8_t u3 = info->nfmsg->nfgen_family;
3300         struct nf_conntrack_tuple tuple;
3301         struct nf_conntrack_expect *exp;
3302         struct nf_conntrack_zone zone;
3303         struct sk_buff *skb2;
3304         int err;
3305
3306         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3307                 if (cda[CTA_EXPECT_MASTER])
3308                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3309                                                      info->nlh, cda,
3310                                                      info->extack);
3311                 else {
3312                         struct netlink_dump_control c = {
3313                                 .dump = ctnetlink_exp_dump_table,
3314                                 .done = ctnetlink_exp_done,
3315                         };
3316                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3317                 }
3318         }
3319
3320         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3321         if (err < 0)
3322                 return err;
3323
3324         if (cda[CTA_EXPECT_TUPLE])
3325                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3326                                             u3, NULL);
3327         else if (cda[CTA_EXPECT_MASTER])
3328                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3329                                             u3, NULL);
3330         else
3331                 return -EINVAL;
3332
3333         if (err < 0)
3334                 return err;
3335
3336         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3337         if (!exp)
3338                 return -ENOENT;
3339
3340         if (cda[CTA_EXPECT_ID]) {
3341                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3342
3343                 if (id != nf_expect_get_id(exp)) {
3344                         nf_ct_expect_put(exp);
3345                         return -ENOENT;
3346                 }
3347         }
3348
3349         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3350         if (!skb2) {
3351                 nf_ct_expect_put(exp);
3352                 return -ENOMEM;
3353         }
3354
3355         rcu_read_lock();
3356         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3357                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3358                                       exp);
3359         rcu_read_unlock();
3360         nf_ct_expect_put(exp);
3361         if (err <= 0) {
3362                 kfree_skb(skb2);
3363                 return -ENOMEM;
3364         }
3365
3366         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3367 }
3368
3369 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3370 {
3371         struct nf_conntrack_helper *helper;
3372         const struct nf_conn_help *m_help;
3373         const char *name = data;
3374
3375         m_help = nfct_help(exp->master);
3376
3377         helper = rcu_dereference(m_help->helper);
3378         if (!helper)
3379                 return false;
3380
3381         return strcmp(helper->name, name) == 0;
3382 }
3383
3384 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3385 {
3386         return true;
3387 }
3388
3389 static int ctnetlink_del_expect(struct sk_buff *skb,
3390                                 const struct nfnl_info *info,
3391                                 const struct nlattr * const cda[])
3392 {
3393         u_int8_t u3 = info->nfmsg->nfgen_family;
3394         struct nf_conntrack_expect *exp;
3395         struct nf_conntrack_tuple tuple;
3396         struct nf_conntrack_zone zone;
3397         int err;
3398
3399         if (cda[CTA_EXPECT_TUPLE]) {
3400                 /* delete a single expect by tuple */
3401                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3402                 if (err < 0)
3403                         return err;
3404
3405                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3406                                             u3, NULL);
3407                 if (err < 0)
3408                         return err;
3409
3410                 /* bump usage count to 2 */
3411                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3412                 if (!exp)
3413                         return -ENOENT;
3414
3415                 if (cda[CTA_EXPECT_ID]) {
3416                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3417
3418                         if (id != nf_expect_get_id(exp)) {
3419                                 nf_ct_expect_put(exp);
3420                                 return -ENOENT;
3421                         }
3422                 }
3423
3424                 /* after list removal, usage count == 1 */
3425                 spin_lock_bh(&nf_conntrack_expect_lock);
3426                 if (del_timer(&exp->timeout)) {
3427                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3428                                                    nlmsg_report(info->nlh));
3429                         nf_ct_expect_put(exp);
3430                 }
3431                 spin_unlock_bh(&nf_conntrack_expect_lock);
3432                 /* have to put what we 'get' above.
3433                  * after this line usage count == 0 */
3434                 nf_ct_expect_put(exp);
3435         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3436                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3437
3438                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3439                                          NETLINK_CB(skb).portid,
3440                                          nlmsg_report(info->nlh));
3441         } else {
3442                 /* This basically means we have to flush everything*/
3443                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3444                                          NETLINK_CB(skb).portid,
3445                                          nlmsg_report(info->nlh));
3446         }
3447
3448         return 0;
3449 }
3450 static int
3451 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3452                         const struct nlattr * const cda[])
3453 {
3454         if (cda[CTA_EXPECT_TIMEOUT]) {
3455                 if (!del_timer(&x->timeout))
3456                         return -ETIME;
3457
3458                 x->timeout.expires = jiffies +
3459                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3460                 add_timer(&x->timeout);
3461         }
3462         return 0;
3463 }
3464
3465 #if IS_ENABLED(CONFIG_NF_NAT)
3466 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3467         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3468         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3469 };
3470 #endif
3471
3472 static int
3473 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3474                            struct nf_conntrack_expect *exp,
3475                            u_int8_t u3)
3476 {
3477 #if IS_ENABLED(CONFIG_NF_NAT)
3478         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3479         struct nf_conntrack_tuple nat_tuple = {};
3480         int err;
3481
3482         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3483                                           exp_nat_nla_policy, NULL);
3484         if (err < 0)
3485                 return err;
3486
3487         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3488                 return -EINVAL;
3489
3490         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3491                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3492                                     u3, NULL);
3493         if (err < 0)
3494                 return err;
3495
3496         exp->saved_addr = nat_tuple.src.u3;
3497         exp->saved_proto = nat_tuple.src.u;
3498         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3499
3500         return 0;
3501 #else
3502         return -EOPNOTSUPP;
3503 #endif
3504 }
3505
3506 static struct nf_conntrack_expect *
3507 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3508                        struct nf_conntrack_helper *helper,
3509                        struct nf_conntrack_tuple *tuple,
3510                        struct nf_conntrack_tuple *mask)
3511 {
3512         u_int32_t class = 0;
3513         struct nf_conntrack_expect *exp;
3514         struct nf_conn_help *help;
3515         int err;
3516
3517         help = nfct_help(ct);
3518         if (!help)
3519                 return ERR_PTR(-EOPNOTSUPP);
3520
3521         if (cda[CTA_EXPECT_CLASS] && helper) {
3522                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3523                 if (class > helper->expect_class_max)
3524                         return ERR_PTR(-EINVAL);
3525         }
3526         exp = nf_ct_expect_alloc(ct);
3527         if (!exp)
3528                 return ERR_PTR(-ENOMEM);
3529
3530         if (cda[CTA_EXPECT_FLAGS]) {
3531                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3532                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3533         } else {
3534                 exp->flags = 0;
3535         }
3536         if (cda[CTA_EXPECT_FN]) {
3537                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3538                 struct nf_ct_helper_expectfn *expfn;
3539
3540                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3541                 if (expfn == NULL) {
3542                         err = -EINVAL;
3543                         goto err_out;
3544                 }
3545                 exp->expectfn = expfn->expectfn;
3546         } else
3547                 exp->expectfn = NULL;
3548
3549         exp->class = class;
3550         exp->master = ct;
3551         exp->helper = helper;
3552         exp->tuple = *tuple;
3553         exp->mask.src.u3 = mask->src.u3;
3554         exp->mask.src.u.all = mask->src.u.all;
3555
3556         if (cda[CTA_EXPECT_NAT]) {
3557                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3558                                                  exp, nf_ct_l3num(ct));
3559                 if (err < 0)
3560                         goto err_out;
3561         }
3562         return exp;
3563 err_out:
3564         nf_ct_expect_put(exp);
3565         return ERR_PTR(err);
3566 }
3567
3568 static int
3569 ctnetlink_create_expect(struct net *net,
3570                         const struct nf_conntrack_zone *zone,
3571                         const struct nlattr * const cda[],
3572                         u_int8_t u3, u32 portid, int report)
3573 {
3574         struct nf_conntrack_tuple tuple, mask, master_tuple;
3575         struct nf_conntrack_tuple_hash *h = NULL;
3576         struct nf_conntrack_helper *helper = NULL;
3577         struct nf_conntrack_expect *exp;
3578         struct nf_conn *ct;
3579         int err;
3580
3581         /* caller guarantees that those three CTA_EXPECT_* exist */
3582         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3583                                     u3, NULL);
3584         if (err < 0)
3585                 return err;
3586         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3587                                     u3, NULL);
3588         if (err < 0)
3589                 return err;
3590         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3591                                     u3, NULL);
3592         if (err < 0)
3593                 return err;
3594
3595         /* Look for master conntrack of this expectation */
3596         h = nf_conntrack_find_get(net, zone, &master_tuple);
3597         if (!h)
3598                 return -ENOENT;
3599         ct = nf_ct_tuplehash_to_ctrack(h);
3600
3601         rcu_read_lock();
3602         if (cda[CTA_EXPECT_HELP_NAME]) {
3603                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3604
3605                 helper = __nf_conntrack_helper_find(helpname, u3,
3606                                                     nf_ct_protonum(ct));
3607                 if (helper == NULL) {
3608                         rcu_read_unlock();
3609 #ifdef CONFIG_MODULES
3610                         if (request_module("nfct-helper-%s", helpname) < 0) {
3611                                 err = -EOPNOTSUPP;
3612                                 goto err_ct;
3613                         }
3614                         rcu_read_lock();
3615                         helper = __nf_conntrack_helper_find(helpname, u3,
3616                                                             nf_ct_protonum(ct));
3617                         if (helper) {
3618                                 err = -EAGAIN;
3619                                 goto err_rcu;
3620                         }
3621                         rcu_read_unlock();
3622 #endif
3623                         err = -EOPNOTSUPP;
3624                         goto err_ct;
3625                 }
3626         }
3627
3628         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3629         if (IS_ERR(exp)) {
3630                 err = PTR_ERR(exp);
3631                 goto err_rcu;
3632         }
3633
3634         err = nf_ct_expect_related_report(exp, portid, report, 0);
3635         nf_ct_expect_put(exp);
3636 err_rcu:
3637         rcu_read_unlock();
3638 err_ct:
3639         nf_ct_put(ct);
3640         return err;
3641 }
3642
3643 static int ctnetlink_new_expect(struct sk_buff *skb,
3644                                 const struct nfnl_info *info,
3645                                 const struct nlattr * const cda[])
3646 {
3647         u_int8_t u3 = info->nfmsg->nfgen_family;
3648         struct nf_conntrack_tuple tuple;
3649         struct nf_conntrack_expect *exp;
3650         struct nf_conntrack_zone zone;
3651         int err;
3652
3653         if (!cda[CTA_EXPECT_TUPLE]
3654             || !cda[CTA_EXPECT_MASK]
3655             || !cda[CTA_EXPECT_MASTER])
3656                 return -EINVAL;
3657
3658         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3659         if (err < 0)
3660                 return err;
3661
3662         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3663                                     u3, NULL);
3664         if (err < 0)
3665                 return err;
3666
3667         spin_lock_bh(&nf_conntrack_expect_lock);
3668         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3669         if (!exp) {
3670                 spin_unlock_bh(&nf_conntrack_expect_lock);
3671                 err = -ENOENT;
3672                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3673                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3674                                                       NETLINK_CB(skb).portid,
3675                                                       nlmsg_report(info->nlh));
3676                 }
3677                 return err;
3678         }
3679
3680         err = -EEXIST;
3681         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3682                 err = ctnetlink_change_expect(exp, cda);
3683         spin_unlock_bh(&nf_conntrack_expect_lock);
3684
3685         return err;
3686 }
3687
3688 static int
3689 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3690                              const struct ip_conntrack_stat *st)
3691 {
3692         struct nlmsghdr *nlh;
3693         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3694
3695         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3696                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3697         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3698                            NFNETLINK_V0, htons(cpu));
3699         if (!nlh)
3700                 goto nlmsg_failure;
3701
3702         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3703             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3704             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3705                 goto nla_put_failure;
3706
3707         nlmsg_end(skb, nlh);
3708         return skb->len;
3709
3710 nla_put_failure:
3711 nlmsg_failure:
3712         nlmsg_cancel(skb, nlh);
3713         return -1;
3714 }
3715
3716 static int
3717 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3718 {
3719         int cpu;
3720         struct net *net = sock_net(skb->sk);
3721
3722         if (cb->args[0] == nr_cpu_ids)
3723                 return 0;
3724
3725         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3726                 const struct ip_conntrack_stat *st;
3727
3728                 if (!cpu_possible(cpu))
3729                         continue;
3730
3731                 st = per_cpu_ptr(net->ct.stat, cpu);
3732                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3733                                                  cb->nlh->nlmsg_seq,
3734                                                  cpu, st) < 0)
3735                         break;
3736         }
3737         cb->args[0] = cpu;
3738
3739         return skb->len;
3740 }
3741
3742 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3743                                   const struct nfnl_info *info,
3744                                   const struct nlattr * const cda[])
3745 {
3746         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3747                 struct netlink_dump_control c = {
3748                         .dump = ctnetlink_exp_stat_cpu_dump,
3749                 };
3750                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3751         }
3752
3753         return 0;
3754 }
3755
3756 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3757 static struct nf_ct_event_notifier ctnl_notifier = {
3758         .ct_event = ctnetlink_conntrack_event,
3759         .exp_event = ctnetlink_expect_event,
3760 };
3761 #endif
3762
3763 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3764         [IPCTNL_MSG_CT_NEW]     = {
3765                 .call           = ctnetlink_new_conntrack,
3766                 .type           = NFNL_CB_MUTEX,
3767                 .attr_count     = CTA_MAX,
3768                 .policy         = ct_nla_policy
3769         },
3770         [IPCTNL_MSG_CT_GET]     = {
3771                 .call           = ctnetlink_get_conntrack,
3772                 .type           = NFNL_CB_MUTEX,
3773                 .attr_count     = CTA_MAX,
3774                 .policy         = ct_nla_policy
3775         },
3776         [IPCTNL_MSG_CT_DELETE]  = {
3777                 .call           = ctnetlink_del_conntrack,
3778                 .type           = NFNL_CB_MUTEX,
3779                 .attr_count     = CTA_MAX,
3780                 .policy         = ct_nla_policy
3781         },
3782         [IPCTNL_MSG_CT_GET_CTRZERO] = {
3783                 .call           = ctnetlink_get_conntrack,
3784                 .type           = NFNL_CB_MUTEX,
3785                 .attr_count     = CTA_MAX,
3786                 .policy         = ct_nla_policy
3787         },
3788         [IPCTNL_MSG_CT_GET_STATS_CPU] = {
3789                 .call           = ctnetlink_stat_ct_cpu,
3790                 .type           = NFNL_CB_MUTEX,
3791         },
3792         [IPCTNL_MSG_CT_GET_STATS] = {
3793                 .call           = ctnetlink_stat_ct,
3794                 .type           = NFNL_CB_MUTEX,
3795         },
3796         [IPCTNL_MSG_CT_GET_DYING] = {
3797                 .call           = ctnetlink_get_ct_dying,
3798                 .type           = NFNL_CB_MUTEX,
3799         },
3800         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = {
3801                 .call           = ctnetlink_get_ct_unconfirmed,
3802                 .type           = NFNL_CB_MUTEX,
3803         },
3804 };
3805
3806 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3807         [IPCTNL_MSG_EXP_GET] = {
3808                 .call           = ctnetlink_get_expect,
3809                 .type           = NFNL_CB_MUTEX,
3810                 .attr_count     = CTA_EXPECT_MAX,
3811                 .policy         = exp_nla_policy
3812         },
3813         [IPCTNL_MSG_EXP_NEW] = {
3814                 .call           = ctnetlink_new_expect,
3815                 .type           = NFNL_CB_MUTEX,
3816                 .attr_count     = CTA_EXPECT_MAX,
3817                 .policy         = exp_nla_policy
3818         },
3819         [IPCTNL_MSG_EXP_DELETE] = {
3820                 .call           = ctnetlink_del_expect,
3821                 .type           = NFNL_CB_MUTEX,
3822                 .attr_count     = CTA_EXPECT_MAX,
3823                 .policy         = exp_nla_policy
3824         },
3825         [IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3826                 .call           = ctnetlink_stat_exp_cpu,
3827                 .type           = NFNL_CB_MUTEX,
3828         },
3829 };
3830
3831 static const struct nfnetlink_subsystem ctnl_subsys = {
3832         .name                           = "conntrack",
3833         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3834         .cb_count                       = IPCTNL_MSG_MAX,
3835         .cb                             = ctnl_cb,
3836 };
3837
3838 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3839         .name                           = "conntrack_expect",
3840         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3841         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3842         .cb                             = ctnl_exp_cb,
3843 };
3844
3845 MODULE_ALIAS("ip_conntrack_netlink");
3846 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3847 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3848
3849 static int __net_init ctnetlink_net_init(struct net *net)
3850 {
3851 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3852         nf_conntrack_register_notifier(net, &ctnl_notifier);
3853 #endif
3854         return 0;
3855 }
3856
3857 static void ctnetlink_net_pre_exit(struct net *net)
3858 {
3859 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3860         nf_conntrack_unregister_notifier(net);
3861 #endif
3862 }
3863
3864 static struct pernet_operations ctnetlink_net_ops = {
3865         .init           = ctnetlink_net_init,
3866         .pre_exit       = ctnetlink_net_pre_exit,
3867 };
3868
3869 static int __init ctnetlink_init(void)
3870 {
3871         int ret;
3872
3873         NL_ASSERT_CTX_FITS(struct ctnetlink_list_dump_ctx);
3874
3875         ret = nfnetlink_subsys_register(&ctnl_subsys);
3876         if (ret < 0) {
3877                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3878                 goto err_out;
3879         }
3880
3881         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3882         if (ret < 0) {
3883                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3884                 goto err_unreg_subsys;
3885         }
3886
3887         ret = register_pernet_subsys(&ctnetlink_net_ops);
3888         if (ret < 0) {
3889                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3890                 goto err_unreg_exp_subsys;
3891         }
3892 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3893         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3894         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3895 #endif
3896         return 0;
3897
3898 err_unreg_exp_subsys:
3899         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3900 err_unreg_subsys:
3901         nfnetlink_subsys_unregister(&ctnl_subsys);
3902 err_out:
3903         return ret;
3904 }
3905
3906 static void __exit ctnetlink_exit(void)
3907 {
3908         unregister_pernet_subsys(&ctnetlink_net_ops);
3909         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3910         nfnetlink_subsys_unregister(&ctnl_subsys);
3911 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3912         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3913 #endif
3914         synchronize_rcu();
3915 }
3916
3917 module_init(ctnetlink_init);
3918 module_exit(ctnetlink_exit);