net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32 #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_LABELS
 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
 395 static int
 396 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 397 {
 398         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 399         unsigned int i;
 400
 401         if (!labels)
 402                 return 0;
 403
 404         i = 0;
 405         do {
 406                 if (labels->bits[i] != 0)
 407                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 408                                        labels->bits);
 409                 i++;
 410         } while (i < ARRAY_SIZE(labels->bits));
 411
 412         return 0;
 413 }
 414 #else
 415 #define ctnetlink_dump_labels(a, b) (0)
 416 #define ctnetlink_label_size(a) (0)
 417 #endif
 418
 419 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 420
 421 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 422 {
 423         struct nlattr *nest_parms;
 424
 425         if (!(ct->status & IPS_EXPECTED))
 426                 return 0;
 427
 428         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 429         if (!nest_parms)
 430                 goto nla_put_failure;
 431         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 432                 goto nla_put_failure;
 433         nla_nest_end(skb, nest_parms);
 434
 435         return 0;
 436
 437 nla_put_failure:
 438         return -1;
 439 }
 440
 441 static int
 442 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 443 {
 444         struct nlattr *nest_parms;
 445
 446         nest_parms = nla_nest_start(skb, type);
 447         if (!nest_parms)
 448                 goto nla_put_failure;
 449
 450         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 451                          htonl(seq->correction_pos)) ||
 452             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 453                          htonl(seq->offset_before)) ||
 454             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 455                          htonl(seq->offset_after)))
 456                 goto nla_put_failure;
 457
 458         nla_nest_end(skb, nest_parms);
 459
 460         return 0;
 461
 462 nla_put_failure:
 463         return -1;
 464 }
 465
 466 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 467 {
 468         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 469         struct nf_ct_seqadj *seq;
 470
 471         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 472                 return 0;
 473
 474         spin_lock_bh(&ct->lock);
 475         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 476         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 477                 goto err;
 478
 479         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 480         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 481                 goto err;
 482
 483         spin_unlock_bh(&ct->lock);
 484         return 0;
 485 err:
 486         spin_unlock_bh(&ct->lock);
 487         return -1;
 488 }
 489
 490 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 491 {
 492         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 493         struct nlattr *nest_parms;
 494
 495         if (!synproxy)
 496                 return 0;
 497
 498         nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 499         if (!nest_parms)
 500                 goto nla_put_failure;
 501
 502         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 503             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 504             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 505                 goto nla_put_failure;
 506
 507         nla_nest_end(skb, nest_parms);
 508
 509         return 0;
 510
 511 nla_put_failure:
 512         return -1;
 513 }
 514
 515 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 516 {
 517         __be32 id = (__force __be32)nf_ct_get_id(ct);
 518
 519         if (nla_put_be32(skb, CTA_ID, id))
 520                 goto nla_put_failure;
 521         return 0;
 522
 523 nla_put_failure:
 524         return -1;
 525 }
 526
 527 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 528 {
 529         if (nla_put_be32(skb, CTA_USE, htonl(refcount_read(&ct->ct_general.use))))
 530                 goto nla_put_failure;
 531         return 0;
 532
 533 nla_put_failure:
 534         return -1;
 535 }
 536
 537 /* all these functions access ct->ext. Caller must either hold a reference
 538  * on ct or prevent its deletion by holding either the bucket spinlock or
 539  * pcpu dying list lock.
 540  */
 541 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 542                                   struct nf_conn *ct, u32 type)
 543 {
 544         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 545             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 546             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 547             ctnetlink_dump_labels(skb, ct) < 0 ||
 548             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 549             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 550                 return -1;
 551
 552         return 0;
 553 }
 554
 555 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 556 {
 557         if (ctnetlink_dump_status(skb, ct) < 0 ||
 558             ctnetlink_dump_mark(skb, ct, true) < 0 ||
 559             ctnetlink_dump_secctx(skb, ct) < 0 ||
 560             ctnetlink_dump_id(skb, ct) < 0 ||
 561             ctnetlink_dump_use(skb, ct) < 0 ||
 562             ctnetlink_dump_master(skb, ct) < 0)
 563                 return -1;
 564
 565         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 566             (ctnetlink_dump_timeout(skb, ct, false) < 0 ||
 567              ctnetlink_dump_protoinfo(skb, ct, false) < 0))
 568                 return -1;
 569
 570         return 0;
 571 }
 572
 573 static int
 574 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 575                     struct nf_conn *ct, bool extinfo, unsigned int flags)
 576 {
 577         const struct nf_conntrack_zone *zone;
 578         struct nlmsghdr *nlh;
 579         struct nlattr *nest_parms;
 580         unsigned int event;
 581
 582         if (portid)
 583                 flags |= NLM_F_MULTI;
 584         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 585         nlh = nfnl_msg_put(skb, portid, seq, event, flags, nf_ct_l3num(ct),
 586                            NFNETLINK_V0, 0);
 587         if (!nlh)
 588                 goto nlmsg_failure;
 589
 590         zone = nf_ct_zone(ct);
 591
 592         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 593         if (!nest_parms)
 594                 goto nla_put_failure;
 595         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 596                 goto nla_put_failure;
 597         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 598                                    NF_CT_ZONE_DIR_ORIG) < 0)
 599                 goto nla_put_failure;
 600         nla_nest_end(skb, nest_parms);
 601
 602         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 603         if (!nest_parms)
 604                 goto nla_put_failure;
 605         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 606                 goto nla_put_failure;
 607         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 608                                    NF_CT_ZONE_DIR_REPL) < 0)
 609                 goto nla_put_failure;
 610         nla_nest_end(skb, nest_parms);
 611
 612         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 613                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 614                 goto nla_put_failure;
 615
 616         if (ctnetlink_dump_info(skb, ct) < 0)
 617                 goto nla_put_failure;
 618         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 619                 goto nla_put_failure;
 620
 621         nlmsg_end(skb, nlh);
 622         return skb->len;
 623
 624 nlmsg_failure:
 625 nla_put_failure:
 626         nlmsg_cancel(skb, nlh);
 627         return -1;
 628 }
 629
 630 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 631         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 632         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 633         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 634         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 635 };
 636
 637 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 638 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 639 {
 640         const struct nf_conntrack_l4proto *l4proto;
 641         size_t len, len4 = 0;
 642
 643         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 644         len *= 3u; /* ORIG, REPLY, MASTER */
 645
 646         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 647         len += l4proto->nlattr_size;
 648         if (l4proto->nlattr_tuple_size) {
 649                 len4 = l4proto->nlattr_tuple_size();
 650                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 651         }
 652
 653         return len + len4;
 654 }
 655 #endif
 656
 657 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 658 {
 659         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 660                 return 0;
 661         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 662                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 663                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 664                ;
 665 }
 666
 667 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 668 {
 669 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 670         int len, ret;
 671
 672         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 673         if (ret)
 674                 return 0;
 675
 676         return nla_total_size(0) /* CTA_SECCTX */
 677                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 678 #else
 679         return 0;
 680 #endif
 681 }
 682
 683 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 684 {
 685 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 686         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 687                 return 0;
 688         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 689 #else
 690         return 0;
 691 #endif
 692 }
 693
 694 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 695 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 696 {
 697         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 698                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 699                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 700                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 701                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 702                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 703                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 704                + ctnetlink_acct_size(ct)
 705                + ctnetlink_timestamp_size(ct)
 706                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 707                + nla_total_size(0) /* CTA_PROTOINFO */
 708                + nla_total_size(0) /* CTA_HELP */
 709                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 710                + ctnetlink_secctx_size(ct)
 711 #if IS_ENABLED(CONFIG_NF_NAT)
 712                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 713                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 714 #endif
 715 #ifdef CONFIG_NF_CONNTRACK_MARK
 716                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 717 #endif
 718 #ifdef CONFIG_NF_CONNTRACK_ZONES
 719                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 720 #endif
 721                + ctnetlink_proto_size(ct)
 722                + ctnetlink_label_size(ct)
 723                ;
 724 }
 725
 726 static int
 727 ctnetlink_conntrack_event(unsigned int events, const struct nf_ct_event *item)
 728 {
 729         const struct nf_conntrack_zone *zone;
 730         struct net *net;
 731         struct nlmsghdr *nlh;
 732         struct nlattr *nest_parms;
 733         struct nf_conn *ct = item->ct;
 734         struct sk_buff *skb;
 735         unsigned int type;
 736         unsigned int flags = 0, group;
 737         int err;
 738
 739         if (events & (1 << IPCT_DESTROY)) {
 740                 type = IPCTNL_MSG_CT_DELETE;
 741                 group = NFNLGRP_CONNTRACK_DESTROY;
 742         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 743                 type = IPCTNL_MSG_CT_NEW;
 744                 flags = NLM_F_CREATE|NLM_F_EXCL;
 745                 group = NFNLGRP_CONNTRACK_NEW;
 746         } else if (events) {
 747                 type = IPCTNL_MSG_CT_NEW;
 748                 group = NFNLGRP_CONNTRACK_UPDATE;
 749         } else
 750                 return 0;
 751
 752         net = nf_ct_net(ct);
 753         if (!item->report && !nfnetlink_has_listeners(net, group))
 754                 return 0;
 755
 756         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 757         if (skb == NULL)
 758                 goto errout;
 759
 760         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 761         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags, nf_ct_l3num(ct),
 762                            NFNETLINK_V0, 0);
 763         if (!nlh)
 764                 goto nlmsg_failure;
 765
 766         zone = nf_ct_zone(ct);
 767
 768         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 769         if (!nest_parms)
 770                 goto nla_put_failure;
 771         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 772                 goto nla_put_failure;
 773         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 774                                    NF_CT_ZONE_DIR_ORIG) < 0)
 775                 goto nla_put_failure;
 776         nla_nest_end(skb, nest_parms);
 777
 778         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 779         if (!nest_parms)
 780                 goto nla_put_failure;
 781         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 782                 goto nla_put_failure;
 783         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 784                                    NF_CT_ZONE_DIR_REPL) < 0)
 785                 goto nla_put_failure;
 786         nla_nest_end(skb, nest_parms);
 787
 788         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 789                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 790                 goto nla_put_failure;
 791
 792         if (ctnetlink_dump_id(skb, ct) < 0)
 793                 goto nla_put_failure;
 794
 795         if (ctnetlink_dump_status(skb, ct) < 0)
 796                 goto nla_put_failure;
 797
 798         if (events & (1 << IPCT_DESTROY)) {
 799                 if (ctnetlink_dump_timeout(skb, ct, true) < 0)
 800                         goto nla_put_failure;
 801
 802                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 803                     ctnetlink_dump_timestamp(skb, ct) < 0 ||
 804                     ctnetlink_dump_protoinfo(skb, ct, true) < 0)
 805                         goto nla_put_failure;
 806         } else {
 807                 if (ctnetlink_dump_timeout(skb, ct, false) < 0)
 808                         goto nla_put_failure;
 809
 810                 if (events & (1 << IPCT_PROTOINFO) &&
 811                     ctnetlink_dump_protoinfo(skb, ct, false) < 0)
 812                         goto nla_put_failure;
 813
 814                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 815                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 816                         goto nla_put_failure;
 817
 818 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 819                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 820                     && ctnetlink_dump_secctx(skb, ct) < 0)
 821                         goto nla_put_failure;
 822 #endif
 823                 if (events & (1 << IPCT_LABEL) &&
 824                      ctnetlink_dump_labels(skb, ct) < 0)
 825                         goto nla_put_failure;
 826
 827                 if (events & (1 << IPCT_RELATED) &&
 828                     ctnetlink_dump_master(skb, ct) < 0)
 829                         goto nla_put_failure;
 830
 831                 if (events & (1 << IPCT_SEQADJ) &&
 832                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 833                         goto nla_put_failure;
 834
 835                 if (events & (1 << IPCT_SYNPROXY) &&
 836                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 837                         goto nla_put_failure;
 838         }
 839
 840 #ifdef CONFIG_NF_CONNTRACK_MARK
 841         if (ctnetlink_dump_mark(skb, ct, events & (1 << IPCT_MARK)))
 842                 goto nla_put_failure;
 843 #endif
 844         nlmsg_end(skb, nlh);
 845         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 846                              GFP_ATOMIC);
 847         if (err == -ENOBUFS || err == -EAGAIN)
 848                 return -ENOBUFS;
 849
 850         return 0;
 851
 852 nla_put_failure:
 853         nlmsg_cancel(skb, nlh);
 854 nlmsg_failure:
 855         kfree_skb(skb);
 856 errout:
 857         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 858                 return -ENOBUFS;
 859
 860         return 0;
 861 }
 862 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 863
 864 static int ctnetlink_done(struct netlink_callback *cb)
 865 {
 866         if (cb->args[1])
 867                 nf_ct_put((struct nf_conn *)cb->args[1]);
 868         kfree(cb->data);
 869         return 0;
 870 }
 871
 872 struct ctnetlink_filter_u32 {
 873         u32 val;
 874         u32 mask;
 875 };
 876
 877 struct ctnetlink_filter {
 878         u8 family;
 879
 880         u_int32_t orig_flags;
 881         u_int32_t reply_flags;
 882
 883         struct nf_conntrack_tuple orig;
 884         struct nf_conntrack_tuple reply;
 885         struct nf_conntrack_zone zone;
 886
 887         struct ctnetlink_filter_u32 mark;
 888         struct ctnetlink_filter_u32 status;
 889 };
 890
 891 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 892         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 893         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 894 };
 895
 896 static int ctnetlink_parse_filter(const struct nlattr *attr,
 897                                   struct ctnetlink_filter *filter)
 898 {
 899         struct nlattr *tb[CTA_FILTER_MAX + 1];
 900         int ret = 0;
 901
 902         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 903                                NULL);
 904         if (ret)
 905                 return ret;
 906
 907         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 908                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 909                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 910                         return -EOPNOTSUPP;
 911         }
 912
 913         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 914                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 915                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 916                         return -EOPNOTSUPP;
 917         }
 918
 919         return 0;
 920 }
 921
 922 static int ctnetlink_parse_zone(const struct nlattr *attr,
 923                                 struct nf_conntrack_zone *zone);
 924 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 925                                          struct nf_conntrack_tuple *tuple,
 926                                          u32 type, u_int8_t l3num,
 927                                          struct nf_conntrack_zone *zone,
 928                                          u_int32_t flags);
 929
 930 static int ctnetlink_filter_parse_mark(struct ctnetlink_filter_u32 *mark,
 931                                        const struct nlattr * const cda[])
 932 {
 933 #ifdef CONFIG_NF_CONNTRACK_MARK
 934         if (cda[CTA_MARK]) {
 935                 mark->val = ntohl(nla_get_be32(cda[CTA_MARK]));
 936
 937                 if (cda[CTA_MARK_MASK])
 938                         mark->mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 939                 else
 940                         mark->mask = 0xffffffff;
 941         } else if (cda[CTA_MARK_MASK]) {
 942                 return -EINVAL;
 943         }
 944 #endif
 945         return 0;
 946 }
 947
 948 static int ctnetlink_filter_parse_status(struct ctnetlink_filter_u32 *status,
 949                                          const struct nlattr * const cda[])
 950 {
 951         if (cda[CTA_STATUS]) {
 952                 status->val = ntohl(nla_get_be32(cda[CTA_STATUS]));
 953                 if (cda[CTA_STATUS_MASK])
 954                         status->mask = ntohl(nla_get_be32(cda[CTA_STATUS_MASK]));
 955                 else
 956                         status->mask = status->val;
 957
 958                 /* status->val == 0? always true, else always false. */
 959                 if (status->mask == 0)
 960                         return -EINVAL;
 961         } else if (cda[CTA_STATUS_MASK]) {
 962                 return -EINVAL;
 963         }
 964
 965         /* CTA_STATUS is NLA_U32, if this fires UAPI needs to be extended */
 966         BUILD_BUG_ON(__IPS_MAX_BIT >= 32);
 967         return 0;
 968 }
 969
 970 static struct ctnetlink_filter *
 971 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 972 {
 973         struct ctnetlink_filter *filter;
 974         int err;
 975
 976 #ifndef CONFIG_NF_CONNTRACK_MARK
 977         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 978                 return ERR_PTR(-EOPNOTSUPP);
 979 #endif
 980
 981         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 982         if (filter == NULL)
 983                 return ERR_PTR(-ENOMEM);
 984
 985         filter->family = family;
 986
 987         err = ctnetlink_filter_parse_mark(&filter->mark, cda);
 988         if (err)
 989                 goto err_filter;
 990
 991         err = ctnetlink_filter_parse_status(&filter->status, cda);
 992         if (err)
 993                 goto err_filter;
 994
 995         if (!cda[CTA_FILTER])
 996                 return filter;
 997
 998         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 999         if (err < 0)
1000                 goto err_filter;
1001
1002         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
1003         if (err < 0)
1004                 goto err_filter;
1005
1006         if (filter->orig_flags) {
1007                 if (!cda[CTA_TUPLE_ORIG]) {
1008                         err = -EINVAL;
1009                         goto err_filter;
1010                 }
1011
1012                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
1013                                                    CTA_TUPLE_ORIG,
1014                                                    filter->family,
1015                                                    &filter->zone,
1016                                                    filter->orig_flags);
1017                 if (err < 0)
1018                         goto err_filter;
1019         }
1020
1021         if (filter->reply_flags) {
1022                 if (!cda[CTA_TUPLE_REPLY]) {
1023                         err = -EINVAL;
1024                         goto err_filter;
1025                 }
1026
1027                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
1028                                                    CTA_TUPLE_REPLY,
1029                                                    filter->family,
1030                                                    &filter->zone,
1031                                                    filter->reply_flags);
1032                 if (err < 0)
1033                         goto err_filter;
1034         }
1035
1036         return filter;
1037
1038 err_filter:
1039         kfree(filter);
1040
1041         return ERR_PTR(err);
1042 }
1043
1044 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1045 {
1046         return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS];
1047 }
1048
1049 static int ctnetlink_start(struct netlink_callback *cb)
1050 {
1051         const struct nlattr * const *cda = cb->data;
1052         struct ctnetlink_filter *filter = NULL;
1053         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1054         u8 family = nfmsg->nfgen_family;
1055
1056         if (ctnetlink_needs_filter(family, cda)) {
1057                 filter = ctnetlink_alloc_filter(cda, family);
1058                 if (IS_ERR(filter))
1059                         return PTR_ERR(filter);
1060         }
1061
1062         cb->data = filter;
1063         return 0;
1064 }
1065
1066 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1067                                         struct nf_conntrack_tuple *ct_tuple,
1068                                         u_int32_t flags, int family)
1069 {
1070         switch (family) {
1071         case NFPROTO_IPV4:
1072                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1073                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1074                         return  0;
1075
1076                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1077                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1078                         return  0;
1079                 break;
1080         case NFPROTO_IPV6:
1081                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1082                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1083                                    &ct_tuple->src.u3.in6))
1084                         return 0;
1085
1086                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1087                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1088                                    &ct_tuple->dst.u3.in6))
1089                         return 0;
1090                 break;
1091         }
1092
1093         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1094             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1095                 return 0;
1096
1097         switch (ct_tuple->dst.protonum) {
1098         case IPPROTO_TCP:
1099         case IPPROTO_UDP:
1100                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1101                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1102                         return 0;
1103
1104                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1105                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1106                         return 0;
1107                 break;
1108         case IPPROTO_ICMP:
1109                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1110                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1111                         return 0;
1112                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1113                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1114                         return 0;
1115                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1116                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1117                         return 0;
1118                 break;
1119         case IPPROTO_ICMPV6:
1120                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1121                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1122                         return 0;
1123                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1124                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1125                         return 0;
1126                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1127                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1128                         return 0;
1129                 break;
1130         }
1131
1132         return 1;
1133 }
1134
1135 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1136 {
1137         struct ctnetlink_filter *filter = data;
1138         struct nf_conntrack_tuple *tuple;
1139         u32 status;
1140
1141         if (filter == NULL)
1142                 goto out;
1143
1144         /* Match entries of a given L3 protocol number.
1145          * If it is not specified, ie. l3proto == 0,
1146          * then match everything.
1147          */
1148         if (filter->family && nf_ct_l3num(ct) != filter->family)
1149                 goto ignore_entry;
1150
1151         if (filter->orig_flags) {
1152                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1153                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1154                                                   filter->orig_flags,
1155                                                   filter->family))
1156                         goto ignore_entry;
1157         }
1158
1159         if (filter->reply_flags) {
1160                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1161                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1162                                                   filter->reply_flags,
1163                                                   filter->family))
1164                         goto ignore_entry;
1165         }
1166
1167 #ifdef CONFIG_NF_CONNTRACK_MARK
1168         if ((READ_ONCE(ct->mark) & filter->mark.mask) != filter->mark.val)
1169                 goto ignore_entry;
1170 #endif
1171         status = (u32)READ_ONCE(ct->status);
1172         if ((status & filter->status.mask) != filter->status.val)
1173                 goto ignore_entry;
1174
1175 out:
1176         return 1;
1177
1178 ignore_entry:
1179         return 0;
1180 }
1181
1182 static int
1183 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1184 {
1185         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1186         struct net *net = sock_net(skb->sk);
1187         struct nf_conn *ct, *last;
1188         struct nf_conntrack_tuple_hash *h;
1189         struct hlist_nulls_node *n;
1190         struct nf_conn *nf_ct_evict[8];
1191         int res, i;
1192         spinlock_t *lockp;
1193
1194         last = (struct nf_conn *)cb->args[1];
1195         i = 0;
1196
1197         local_bh_disable();
1198         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1199 restart:
1200                 while (i) {
1201                         i--;
1202                         if (nf_ct_should_gc(nf_ct_evict[i]))
1203                                 nf_ct_kill(nf_ct_evict[i]);
1204                         nf_ct_put(nf_ct_evict[i]);
1205                 }
1206
1207                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1208                 nf_conntrack_lock(lockp);
1209                 if (cb->args[0] >= nf_conntrack_htable_size) {
1210                         spin_unlock(lockp);
1211                         goto out;
1212                 }
1213                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1214                                            hnnode) {
1215                         ct = nf_ct_tuplehash_to_ctrack(h);
1216                         if (nf_ct_is_expired(ct)) {
1217                                 /* need to defer nf_ct_kill() until lock is released */
1218                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1219                                     refcount_inc_not_zero(&ct->ct_general.use))
1220                                         nf_ct_evict[i++] = ct;
1221                                 continue;
1222                         }
1223
1224                         if (!net_eq(net, nf_ct_net(ct)))
1225                                 continue;
1226
1227                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1228                                 continue;
1229
1230                         if (cb->args[1]) {
1231                                 if (ct != last)
1232                                         continue;
1233                                 cb->args[1] = 0;
1234                         }
1235                         if (!ctnetlink_filter_match(ct, cb->data))
1236                                 continue;
1237
1238                         res =
1239                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1240                                             cb->nlh->nlmsg_seq,
1241                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1242                                             ct, true, flags);
1243                         if (res < 0) {
1244                                 nf_conntrack_get(&ct->ct_general);
1245                                 cb->args[1] = (unsigned long)ct;
1246                                 spin_unlock(lockp);
1247                                 goto out;
1248                         }
1249                 }
1250                 spin_unlock(lockp);
1251                 if (cb->args[1]) {
1252                         cb->args[1] = 0;
1253                         goto restart;
1254                 }
1255         }
1256 out:
1257         local_bh_enable();
1258         if (last) {
1259                 /* nf ct hash resize happened, now clear the leftover. */
1260                 if ((struct nf_conn *)cb->args[1] == last)
1261                         cb->args[1] = 0;
1262
1263                 nf_ct_put(last);
1264         }
1265
1266         while (i) {
1267                 i--;
1268                 if (nf_ct_should_gc(nf_ct_evict[i]))
1269                         nf_ct_kill(nf_ct_evict[i]);
1270                 nf_ct_put(nf_ct_evict[i]);
1271         }
1272
1273         return skb->len;
1274 }
1275
1276 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1277                                 struct nf_conntrack_tuple *t,
1278                                 u_int32_t flags)
1279 {
1280         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1281                 if (!tb[CTA_IP_V4_SRC])
1282                         return -EINVAL;
1283
1284                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1285         }
1286
1287         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1288                 if (!tb[CTA_IP_V4_DST])
1289                         return -EINVAL;
1290
1291                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1292         }
1293
1294         return 0;
1295 }
1296
1297 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1298                                 struct nf_conntrack_tuple *t,
1299                                 u_int32_t flags)
1300 {
1301         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1302                 if (!tb[CTA_IP_V6_SRC])
1303                         return -EINVAL;
1304
1305                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1306         }
1307
1308         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1309                 if (!tb[CTA_IP_V6_DST])
1310                         return -EINVAL;
1311
1312                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1313         }
1314
1315         return 0;
1316 }
1317
1318 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1319                                     struct nf_conntrack_tuple *tuple,
1320                                     u_int32_t flags)
1321 {
1322         struct nlattr *tb[CTA_IP_MAX+1];
1323         int ret = 0;
1324
1325         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr,
1326                                           cta_ip_nla_policy, NULL);
1327         if (ret < 0)
1328                 return ret;
1329
1330         switch (tuple->src.l3num) {
1331         case NFPROTO_IPV4:
1332                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1333                 break;
1334         case NFPROTO_IPV6:
1335                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1336                 break;
1337         }
1338
1339         return ret;
1340 }
1341
1342 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1343         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1344 };
1345
1346 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1347                                        struct nf_conntrack_tuple *tuple,
1348                                        u_int32_t flags)
1349 {
1350         const struct nf_conntrack_l4proto *l4proto;
1351         struct nlattr *tb[CTA_PROTO_MAX+1];
1352         int ret = 0;
1353
1354         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1355                                           proto_nla_policy, NULL);
1356         if (ret < 0)
1357                 return ret;
1358
1359         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1360                 return 0;
1361
1362         if (!tb[CTA_PROTO_NUM])
1363                 return -EINVAL;
1364
1365         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1366
1367         rcu_read_lock();
1368         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1369
1370         if (likely(l4proto->nlattr_to_tuple)) {
1371                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1372                                                      l4proto->nla_policy,
1373                                                      NULL);
1374                 if (ret == 0)
1375                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1376         }
1377
1378         rcu_read_unlock();
1379
1380         return ret;
1381 }
1382
1383 static int
1384 ctnetlink_parse_zone(const struct nlattr *attr,
1385                      struct nf_conntrack_zone *zone)
1386 {
1387         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1388                         NF_CT_DEFAULT_ZONE_DIR, 0);
1389 #ifdef CONFIG_NF_CONNTRACK_ZONES
1390         if (attr)
1391                 zone->id = ntohs(nla_get_be16(attr));
1392 #else
1393         if (attr)
1394                 return -EOPNOTSUPP;
1395 #endif
1396         return 0;
1397 }
1398
1399 static int
1400 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1401                            struct nf_conntrack_zone *zone)
1402 {
1403         int ret;
1404
1405         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1406                 return -EINVAL;
1407
1408         ret = ctnetlink_parse_zone(attr, zone);
1409         if (ret < 0)
1410                 return ret;
1411
1412         if (type == CTA_TUPLE_REPLY)
1413                 zone->dir = NF_CT_ZONE_DIR_REPL;
1414         else
1415                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1416
1417         return 0;
1418 }
1419
1420 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1421         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1422         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1423         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1424 };
1425
1426 #define CTA_FILTER_F_ALL_CTA_PROTO \
1427   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1428    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1429    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1430    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1431    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1432    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1433    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1434    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1435
1436 static int
1437 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1438                               struct nf_conntrack_tuple *tuple, u32 type,
1439                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1440                               u_int32_t flags)
1441 {
1442         struct nlattr *tb[CTA_TUPLE_MAX+1];
1443         int err;
1444
1445         memset(tuple, 0, sizeof(*tuple));
1446
1447         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1448                                           tuple_nla_policy, NULL);
1449         if (err < 0)
1450                 return err;
1451
1452         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1453                 return -EOPNOTSUPP;
1454         tuple->src.l3num = l3num;
1455
1456         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1457             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1458                 if (!tb[CTA_TUPLE_IP])
1459                         return -EINVAL;
1460
1461                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1462                 if (err < 0)
1463                         return err;
1464         }
1465
1466         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1467                 if (!tb[CTA_TUPLE_PROTO])
1468                         return -EINVAL;
1469
1470                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1471                 if (err < 0)
1472                         return err;
1473         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1474                 /* Can't manage proto flags without a protonum  */
1475                 return -EINVAL;
1476         }
1477
1478         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1479                 if (!zone)
1480                         return -EINVAL;
1481
1482                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1483                                                  type, zone);
1484                 if (err < 0)
1485                         return err;
1486         }
1487
1488         /* orig and expect tuples get DIR_ORIGINAL */
1489         if (type == CTA_TUPLE_REPLY)
1490                 tuple->dst.dir = IP_CT_DIR_REPLY;
1491         else
1492                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1493
1494         return 0;
1495 }
1496
1497 static int
1498 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1499                       struct nf_conntrack_tuple *tuple, u32 type,
1500                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1501 {
1502         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1503                                             CTA_FILTER_FLAG(ALL));
1504 }
1505
1506 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1507         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1508                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1509 };
1510
1511 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1512                                 struct nlattr **helpinfo)
1513 {
1514         int err;
1515         struct nlattr *tb[CTA_HELP_MAX+1];
1516
1517         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1518                                           help_nla_policy, NULL);
1519         if (err < 0)
1520                 return err;
1521
1522         if (!tb[CTA_HELP_NAME])
1523                 return -EINVAL;
1524
1525         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1526
1527         if (tb[CTA_HELP_INFO])
1528                 *helpinfo = tb[CTA_HELP_INFO];
1529
1530         return 0;
1531 }
1532
1533 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1534         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1535         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1536         [CTA_STATUS]            = { .type = NLA_U32 },
1537         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1538         [CTA_HELP]              = { .type = NLA_NESTED },
1539         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1540         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1541         [CTA_MARK]              = { .type = NLA_U32 },
1542         [CTA_ID]                = { .type = NLA_U32 },
1543         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1544         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1545         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1546         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1547         [CTA_ZONE]              = { .type = NLA_U16 },
1548         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1549         [CTA_LABELS]            = { .type = NLA_BINARY,
1550                                     .len = NF_CT_LABELS_MAX_SIZE },
1551         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1552                                     .len = NF_CT_LABELS_MAX_SIZE },
1553         [CTA_FILTER]            = { .type = NLA_NESTED },
1554         [CTA_STATUS_MASK]       = { .type = NLA_U32 },
1555 };
1556
1557 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1558 {
1559         return ctnetlink_filter_match(ct, data);
1560 }
1561
1562 static int ctnetlink_flush_conntrack(struct net *net,
1563                                      const struct nlattr * const cda[],
1564                                      u32 portid, int report, u8 family)
1565 {
1566         struct ctnetlink_filter *filter = NULL;
1567         struct nf_ct_iter_data iter = {
1568                 .net            = net,
1569                 .portid         = portid,
1570                 .report         = report,
1571         };
1572
1573         if (ctnetlink_needs_filter(family, cda)) {
1574                 if (cda[CTA_FILTER])
1575                         return -EOPNOTSUPP;
1576
1577                 filter = ctnetlink_alloc_filter(cda, family);
1578                 if (IS_ERR(filter))
1579                         return PTR_ERR(filter);
1580
1581                 iter.data = filter;
1582         }
1583
1584         nf_ct_iterate_cleanup_net(ctnetlink_flush_iterate, &iter);
1585         kfree(filter);
1586
1587         return 0;
1588 }
1589
1590 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1591                                    const struct nfnl_info *info,
1592                                    const struct nlattr * const cda[])
1593 {
1594         u8 family = info->nfmsg->nfgen_family;
1595         struct nf_conntrack_tuple_hash *h;
1596         struct nf_conntrack_tuple tuple;
1597         struct nf_conntrack_zone zone;
1598         struct nf_conn *ct;
1599         int err;
1600
1601         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1602         if (err < 0)
1603                 return err;
1604
1605         if (cda[CTA_TUPLE_ORIG])
1606                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1607                                             family, &zone);
1608         else if (cda[CTA_TUPLE_REPLY])
1609                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1610                                             family, &zone);
1611         else {
1612                 u_int8_t u3 = info->nfmsg->version ? family : AF_UNSPEC;
1613
1614                 return ctnetlink_flush_conntrack(info->net, cda,
1615                                                  NETLINK_CB(skb).portid,
1616                                                  nlmsg_report(info->nlh), u3);
1617         }
1618
1619         if (err < 0)
1620                 return err;
1621
1622         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1623         if (!h)
1624                 return -ENOENT;
1625
1626         ct = nf_ct_tuplehash_to_ctrack(h);
1627
1628         if (cda[CTA_ID]) {
1629                 __be32 id = nla_get_be32(cda[CTA_ID]);
1630
1631                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1632                         nf_ct_put(ct);
1633                         return -ENOENT;
1634                 }
1635         }
1636
1637         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1638         nf_ct_put(ct);
1639
1640         return 0;
1641 }
1642
1643 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1644                                    const struct nfnl_info *info,
1645                                    const struct nlattr * const cda[])
1646 {
1647         u_int8_t u3 = info->nfmsg->nfgen_family;
1648         struct nf_conntrack_tuple_hash *h;
1649         struct nf_conntrack_tuple tuple;
1650         struct nf_conntrack_zone zone;
1651         struct sk_buff *skb2;
1652         struct nf_conn *ct;
1653         int err;
1654
1655         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1656                 struct netlink_dump_control c = {
1657                         .start = ctnetlink_start,
1658                         .dump = ctnetlink_dump_table,
1659                         .done = ctnetlink_done,
1660                         .data = (void *)cda,
1661                 };
1662
1663                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1664         }
1665
1666         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1667         if (err < 0)
1668                 return err;
1669
1670         if (cda[CTA_TUPLE_ORIG])
1671                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1672                                             u3, &zone);
1673         else if (cda[CTA_TUPLE_REPLY])
1674                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1675                                             u3, &zone);
1676         else
1677                 return -EINVAL;
1678
1679         if (err < 0)
1680                 return err;
1681
1682         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1683         if (!h)
1684                 return -ENOENT;
1685
1686         ct = nf_ct_tuplehash_to_ctrack(h);
1687
1688         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1689         if (!skb2) {
1690                 nf_ct_put(ct);
1691                 return -ENOMEM;
1692         }
1693
1694         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1695                                   info->nlh->nlmsg_seq,
1696                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1697                                   true, 0);
1698         nf_ct_put(ct);
1699         if (err <= 0) {
1700                 kfree_skb(skb2);
1701                 return -ENOMEM;
1702         }
1703
1704         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1705 }
1706
1707 static int ctnetlink_done_list(struct netlink_callback *cb)
1708 {
1709         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1710
1711         if (ctx->last)
1712                 nf_ct_put(ctx->last);
1713
1714         return 0;
1715 }
1716
1717 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1718 static int ctnetlink_dump_one_entry(struct sk_buff *skb,
1719                                     struct netlink_callback *cb,
1720                                     struct nf_conn *ct,
1721                                     bool dying)
1722 {
1723         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1724         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1725         u8 l3proto = nfmsg->nfgen_family;
1726         int res;
1727
1728         if (l3proto && nf_ct_l3num(ct) != l3proto)
1729                 return 0;
1730
1731         if (ctx->last) {
1732                 if (ct != ctx->last)
1733                         return 0;
1734
1735                 ctx->last = NULL;
1736         }
1737
1738         /* We can't dump extension info for the unconfirmed
1739          * list because unconfirmed conntracks can have
1740          * ct->ext reallocated (and thus freed).
1741          *
1742          * In the dying list case ct->ext can't be free'd
1743          * until after we drop pcpu->lock.
1744          */
1745         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1746                                   cb->nlh->nlmsg_seq,
1747                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1748                                   ct, dying, 0);
1749         if (res < 0) {
1750                 if (!refcount_inc_not_zero(&ct->ct_general.use))
1751                         return 0;
1752
1753                 ctx->last = ct;
1754         }
1755
1756         return res;
1757 }
1758 #endif
1759
1760 static int
1761 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1762 {
1763         return 0;
1764 }
1765
1766 static int
1767 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1768 {
1769         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1770         struct nf_conn *last = ctx->last;
1771 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1772         const struct net *net = sock_net(skb->sk);
1773         struct nf_conntrack_net_ecache *ecache_net;
1774         struct nf_conntrack_tuple_hash *h;
1775         struct hlist_nulls_node *n;
1776 #endif
1777
1778         if (ctx->done)
1779                 return 0;
1780
1781         ctx->last = NULL;
1782
1783 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1784         ecache_net = nf_conn_pernet_ecache(net);
1785         spin_lock_bh(&ecache_net->dying_lock);
1786
1787         hlist_nulls_for_each_entry(h, n, &ecache_net->dying_list, hnnode) {
1788                 struct nf_conn *ct;
1789                 int res;
1790
1791                 ct = nf_ct_tuplehash_to_ctrack(h);
1792                 if (last && last != ct)
1793                         continue;
1794
1795                 res = ctnetlink_dump_one_entry(skb, cb, ct, true);
1796                 if (res < 0) {
1797                         spin_unlock_bh(&ecache_net->dying_lock);
1798                         nf_ct_put(last);
1799                         return skb->len;
1800                 }
1801
1802                 nf_ct_put(last);
1803                 last = NULL;
1804         }
1805
1806         spin_unlock_bh(&ecache_net->dying_lock);
1807 #endif
1808         ctx->done = true;
1809         nf_ct_put(last);
1810
1811         return skb->len;
1812 }
1813
1814 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1815                                   const struct nfnl_info *info,
1816                                   const struct nlattr * const cda[])
1817 {
1818         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1819                 struct netlink_dump_control c = {
1820                         .dump = ctnetlink_dump_dying,
1821                         .done = ctnetlink_done_list,
1822                 };
1823                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1824         }
1825
1826         return -EOPNOTSUPP;
1827 }
1828
1829 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1830                                         const struct nfnl_info *info,
1831                                         const struct nlattr * const cda[])
1832 {
1833         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1834                 struct netlink_dump_control c = {
1835                         .dump = ctnetlink_dump_unconfirmed,
1836                         .done = ctnetlink_done_list,
1837                 };
1838                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1839         }
1840
1841         return -EOPNOTSUPP;
1842 }
1843
1844 #if IS_ENABLED(CONFIG_NF_NAT)
1845 static int
1846 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1847                           enum nf_nat_manip_type manip,
1848                           const struct nlattr *attr)
1849         __must_hold(RCU)
1850 {
1851         const struct nf_nat_hook *nat_hook;
1852         int err;
1853
1854         nat_hook = rcu_dereference(nf_nat_hook);
1855         if (!nat_hook) {
1856 #ifdef CONFIG_MODULES
1857                 rcu_read_unlock();
1858                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1859                 if (request_module("nf-nat") < 0) {
1860                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1861                         rcu_read_lock();
1862                         return -EOPNOTSUPP;
1863                 }
1864                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1865                 rcu_read_lock();
1866                 nat_hook = rcu_dereference(nf_nat_hook);
1867                 if (nat_hook)
1868                         return -EAGAIN;
1869 #endif
1870                 return -EOPNOTSUPP;
1871         }
1872
1873         err = nat_hook->parse_nat_setup(ct, manip, attr);
1874         if (err == -EAGAIN) {
1875 #ifdef CONFIG_MODULES
1876                 rcu_read_unlock();
1877                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1878                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1879                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1880                         rcu_read_lock();
1881                         return -EOPNOTSUPP;
1882                 }
1883                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1884                 rcu_read_lock();
1885 #else
1886                 err = -EOPNOTSUPP;
1887 #endif
1888         }
1889         return err;
1890 }
1891 #endif
1892
1893 static int
1894 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1895 {
1896         return nf_ct_change_status_common(ct, ntohl(nla_get_be32(cda[CTA_STATUS])));
1897 }
1898
1899 static int
1900 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1901 {
1902 #if IS_ENABLED(CONFIG_NF_NAT)
1903         int ret;
1904
1905         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1906                 return 0;
1907
1908         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1909                                         cda[CTA_NAT_DST]);
1910         if (ret < 0)
1911                 return ret;
1912
1913         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1914                                          cda[CTA_NAT_SRC]);
1915 #else
1916         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1917                 return 0;
1918         return -EOPNOTSUPP;
1919 #endif
1920 }
1921
1922 static int ctnetlink_change_helper(struct nf_conn *ct,
1923                                    const struct nlattr * const cda[])
1924 {
1925         struct nf_conntrack_helper *helper;
1926         struct nf_conn_help *help = nfct_help(ct);
1927         char *helpname = NULL;
1928         struct nlattr *helpinfo = NULL;
1929         int err;
1930
1931         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1932         if (err < 0)
1933                 return err;
1934
1935         /* don't change helper of sibling connections */
1936         if (ct->master) {
1937                 /* If we try to change the helper to the same thing twice,
1938                  * treat the second attempt as a no-op instead of returning
1939                  * an error.
1940                  */
1941                 err = -EBUSY;
1942                 if (help) {
1943                         rcu_read_lock();
1944                         helper = rcu_dereference(help->helper);
1945                         if (helper && !strcmp(helper->name, helpname))
1946                                 err = 0;
1947                         rcu_read_unlock();
1948                 }
1949
1950                 return err;
1951         }
1952
1953         if (!strcmp(helpname, "")) {
1954                 if (help && help->helper) {
1955                         /* we had a helper before ... */
1956                         nf_ct_remove_expectations(ct);
1957                         RCU_INIT_POINTER(help->helper, NULL);
1958                 }
1959
1960                 return 0;
1961         }
1962
1963         rcu_read_lock();
1964         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1965                                             nf_ct_protonum(ct));
1966         if (helper == NULL) {
1967                 rcu_read_unlock();
1968                 return -EOPNOTSUPP;
1969         }
1970
1971         if (help) {
1972                 if (rcu_access_pointer(help->helper) == helper) {
1973                         /* update private helper data if allowed. */
1974                         if (helper->from_nlattr)
1975                                 helper->from_nlattr(helpinfo, ct);
1976                         err = 0;
1977                 } else
1978                         err = -EBUSY;
1979         } else {
1980                 /* we cannot set a helper for an existing conntrack */
1981                 err = -EOPNOTSUPP;
1982         }
1983
1984         rcu_read_unlock();
1985         return err;
1986 }
1987
1988 static int ctnetlink_change_timeout(struct nf_conn *ct,
1989                                     const struct nlattr * const cda[])
1990 {
1991         return __nf_ct_change_timeout(ct, (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ);
1992 }
1993
1994 #if defined(CONFIG_NF_CONNTRACK_MARK)
1995 static void ctnetlink_change_mark(struct nf_conn *ct,
1996                                     const struct nlattr * const cda[])
1997 {
1998         u32 mark, newmark, mask = 0;
1999
2000         if (cda[CTA_MARK_MASK])
2001                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2002
2003         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2004         newmark = (READ_ONCE(ct->mark) & mask) ^ mark;
2005         if (newmark != READ_ONCE(ct->mark))
2006                 WRITE_ONCE(ct->mark, newmark);
2007 }
2008 #endif
2009
2010 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2011         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2012         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2013         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2014 };
2015
2016 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2017                                       const struct nlattr * const cda[])
2018 {
2019         const struct nlattr *attr = cda[CTA_PROTOINFO];
2020         const struct nf_conntrack_l4proto *l4proto;
2021         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2022         int err = 0;
2023
2024         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2025                                           protoinfo_policy, NULL);
2026         if (err < 0)
2027                 return err;
2028
2029         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2030         if (l4proto->from_nlattr)
2031                 err = l4proto->from_nlattr(tb, ct);
2032
2033         return err;
2034 }
2035
2036 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2037         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2038         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2039         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2040 };
2041
2042 static int change_seq_adj(struct nf_ct_seqadj *seq,
2043                           const struct nlattr * const attr)
2044 {
2045         int err;
2046         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2047
2048         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2049                                           seqadj_policy, NULL);
2050         if (err < 0)
2051                 return err;
2052
2053         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2054                 return -EINVAL;
2055
2056         seq->correction_pos =
2057                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2058
2059         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2060                 return -EINVAL;
2061
2062         seq->offset_before =
2063                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2064
2065         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2066                 return -EINVAL;
2067
2068         seq->offset_after =
2069                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2070
2071         return 0;
2072 }
2073
2074 static int
2075 ctnetlink_change_seq_adj(struct nf_conn *ct,
2076                          const struct nlattr * const cda[])
2077 {
2078         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2079         int ret = 0;
2080
2081         if (!seqadj)
2082                 return 0;
2083
2084         spin_lock_bh(&ct->lock);
2085         if (cda[CTA_SEQ_ADJ_ORIG]) {
2086                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2087                                      cda[CTA_SEQ_ADJ_ORIG]);
2088                 if (ret < 0)
2089                         goto err;
2090
2091                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2092         }
2093
2094         if (cda[CTA_SEQ_ADJ_REPLY]) {
2095                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2096                                      cda[CTA_SEQ_ADJ_REPLY]);
2097                 if (ret < 0)
2098                         goto err;
2099
2100                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2101         }
2102
2103         spin_unlock_bh(&ct->lock);
2104         return 0;
2105 err:
2106         spin_unlock_bh(&ct->lock);
2107         return ret;
2108 }
2109
2110 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2111         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2112         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2113         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2114 };
2115
2116 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2117                                      const struct nlattr * const cda[])
2118 {
2119         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2120         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2121         int err;
2122
2123         if (!synproxy)
2124                 return 0;
2125
2126         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2127                                           cda[CTA_SYNPROXY], synproxy_policy,
2128                                           NULL);
2129         if (err < 0)
2130                 return err;
2131
2132         if (!tb[CTA_SYNPROXY_ISN] ||
2133             !tb[CTA_SYNPROXY_ITS] ||
2134             !tb[CTA_SYNPROXY_TSOFF])
2135                 return -EINVAL;
2136
2137         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2138         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2139         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2140
2141         return 0;
2142 }
2143
2144 static int
2145 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2146 {
2147 #ifdef CONFIG_NF_CONNTRACK_LABELS
2148         size_t len = nla_len(cda[CTA_LABELS]);
2149         const void *mask = cda[CTA_LABELS_MASK];
2150
2151         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2152                 return -EINVAL;
2153
2154         if (mask) {
2155                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2156                     nla_len(cda[CTA_LABELS_MASK]) != len)
2157                         return -EINVAL;
2158                 mask = nla_data(cda[CTA_LABELS_MASK]);
2159         }
2160
2161         len /= sizeof(u32);
2162
2163         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2164 #else
2165         return -EOPNOTSUPP;
2166 #endif
2167 }
2168
2169 static int
2170 ctnetlink_change_conntrack(struct nf_conn *ct,
2171                            const struct nlattr * const cda[])
2172 {
2173         int err;
2174
2175         /* only allow NAT changes and master assignation for new conntracks */
2176         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2177                 return -EOPNOTSUPP;
2178
2179         if (cda[CTA_HELP]) {
2180                 err = ctnetlink_change_helper(ct, cda);
2181                 if (err < 0)
2182                         return err;
2183         }
2184
2185         if (cda[CTA_TIMEOUT]) {
2186                 err = ctnetlink_change_timeout(ct, cda);
2187                 if (err < 0)
2188                         return err;
2189         }
2190
2191         if (cda[CTA_STATUS]) {
2192                 err = ctnetlink_change_status(ct, cda);
2193                 if (err < 0)
2194                         return err;
2195         }
2196
2197         if (cda[CTA_PROTOINFO]) {
2198                 err = ctnetlink_change_protoinfo(ct, cda);
2199                 if (err < 0)
2200                         return err;
2201         }
2202
2203 #if defined(CONFIG_NF_CONNTRACK_MARK)
2204         if (cda[CTA_MARK])
2205                 ctnetlink_change_mark(ct, cda);
2206 #endif
2207
2208         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2209                 err = ctnetlink_change_seq_adj(ct, cda);
2210                 if (err < 0)
2211                         return err;
2212         }
2213
2214         if (cda[CTA_SYNPROXY]) {
2215                 err = ctnetlink_change_synproxy(ct, cda);
2216                 if (err < 0)
2217                         return err;
2218         }
2219
2220         if (cda[CTA_LABELS]) {
2221                 err = ctnetlink_attach_labels(ct, cda);
2222                 if (err < 0)
2223                         return err;
2224         }
2225
2226         return 0;
2227 }
2228
2229 static struct nf_conn *
2230 ctnetlink_create_conntrack(struct net *net,
2231                            const struct nf_conntrack_zone *zone,
2232                            const struct nlattr * const cda[],
2233                            struct nf_conntrack_tuple *otuple,
2234                            struct nf_conntrack_tuple *rtuple,
2235                            u8 u3)
2236 {
2237         struct nf_conn *ct;
2238         int err = -EINVAL;
2239         struct nf_conntrack_helper *helper;
2240         struct nf_conn_tstamp *tstamp;
2241         u64 timeout;
2242
2243         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2244         if (IS_ERR(ct))
2245                 return ERR_PTR(-ENOMEM);
2246
2247         if (!cda[CTA_TIMEOUT])
2248                 goto err1;
2249
2250         rcu_read_lock();
2251         if (cda[CTA_HELP]) {
2252                 char *helpname = NULL;
2253                 struct nlattr *helpinfo = NULL;
2254
2255                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2256                 if (err < 0)
2257                         goto err2;
2258
2259                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2260                                                     nf_ct_protonum(ct));
2261                 if (helper == NULL) {
2262                         rcu_read_unlock();
2263 #ifdef CONFIG_MODULES
2264                         if (request_module("nfct-helper-%s", helpname) < 0) {
2265                                 err = -EOPNOTSUPP;
2266                                 goto err1;
2267                         }
2268
2269                         rcu_read_lock();
2270                         helper = __nf_conntrack_helper_find(helpname,
2271                                                             nf_ct_l3num(ct),
2272                                                             nf_ct_protonum(ct));
2273                         if (helper) {
2274                                 err = -EAGAIN;
2275                                 goto err2;
2276                         }
2277                         rcu_read_unlock();
2278 #endif
2279                         err = -EOPNOTSUPP;
2280                         goto err1;
2281                 } else {
2282                         struct nf_conn_help *help;
2283
2284                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2285                         if (help == NULL) {
2286                                 err = -ENOMEM;
2287                                 goto err2;
2288                         }
2289                         /* set private helper data if allowed. */
2290                         if (helper->from_nlattr)
2291                                 helper->from_nlattr(helpinfo, ct);
2292
2293                         /* disable helper auto-assignment for this entry */
2294                         ct->status |= IPS_HELPER;
2295                         RCU_INIT_POINTER(help->helper, helper);
2296                 }
2297         }
2298
2299         err = ctnetlink_setup_nat(ct, cda);
2300         if (err < 0)
2301                 goto err2;
2302
2303         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2304         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2305         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2306         nf_ct_labels_ext_add(ct);
2307         nfct_seqadj_ext_add(ct);
2308         nfct_synproxy_ext_add(ct);
2309
2310         /* we must add conntrack extensions before confirmation. */
2311         ct->status |= IPS_CONFIRMED;
2312
2313         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2314         __nf_ct_set_timeout(ct, timeout);
2315
2316         if (cda[CTA_STATUS]) {
2317                 err = ctnetlink_change_status(ct, cda);
2318                 if (err < 0)
2319                         goto err2;
2320         }
2321
2322         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2323                 err = ctnetlink_change_seq_adj(ct, cda);
2324                 if (err < 0)
2325                         goto err2;
2326         }
2327
2328         memset(&ct->proto, 0, sizeof(ct->proto));
2329         if (cda[CTA_PROTOINFO]) {
2330                 err = ctnetlink_change_protoinfo(ct, cda);
2331                 if (err < 0)
2332                         goto err2;
2333         }
2334
2335         if (cda[CTA_SYNPROXY]) {
2336                 err = ctnetlink_change_synproxy(ct, cda);
2337                 if (err < 0)
2338                         goto err2;
2339         }
2340
2341 #if defined(CONFIG_NF_CONNTRACK_MARK)
2342         if (cda[CTA_MARK])
2343                 ctnetlink_change_mark(ct, cda);
2344 #endif
2345
2346         /* setup master conntrack: this is a confirmed expectation */
2347         if (cda[CTA_TUPLE_MASTER]) {
2348                 struct nf_conntrack_tuple master;
2349                 struct nf_conntrack_tuple_hash *master_h;
2350                 struct nf_conn *master_ct;
2351
2352                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2353                                             u3, NULL);
2354                 if (err < 0)
2355                         goto err2;
2356
2357                 master_h = nf_conntrack_find_get(net, zone, &master);
2358                 if (master_h == NULL) {
2359                         err = -ENOENT;
2360                         goto err2;
2361                 }
2362                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2363                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2364                 ct->master = master_ct;
2365         }
2366         tstamp = nf_conn_tstamp_find(ct);
2367         if (tstamp)
2368                 tstamp->start = ktime_get_real_ns();
2369
2370         err = nf_conntrack_hash_check_insert(ct);
2371         if (err < 0)
2372                 goto err3;
2373
2374         rcu_read_unlock();
2375
2376         return ct;
2377
2378 err3:
2379         if (ct->master)
2380                 nf_ct_put(ct->master);
2381 err2:
2382         rcu_read_unlock();
2383 err1:
2384         nf_conntrack_free(ct);
2385         return ERR_PTR(err);
2386 }
2387
2388 static int ctnetlink_new_conntrack(struct sk_buff *skb,
2389                                    const struct nfnl_info *info,
2390                                    const struct nlattr * const cda[])
2391 {
2392         struct nf_conntrack_tuple otuple, rtuple;
2393         struct nf_conntrack_tuple_hash *h = NULL;
2394         u_int8_t u3 = info->nfmsg->nfgen_family;
2395         struct nf_conntrack_zone zone;
2396         struct nf_conn *ct;
2397         int err;
2398
2399         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2400         if (err < 0)
2401                 return err;
2402
2403         if (cda[CTA_TUPLE_ORIG]) {
2404                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2405                                             u3, &zone);
2406                 if (err < 0)
2407                         return err;
2408         }
2409
2410         if (cda[CTA_TUPLE_REPLY]) {
2411                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2412                                             u3, &zone);
2413                 if (err < 0)
2414                         return err;
2415         }
2416
2417         if (cda[CTA_TUPLE_ORIG])
2418                 h = nf_conntrack_find_get(info->net, &zone, &otuple);
2419         else if (cda[CTA_TUPLE_REPLY])
2420                 h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2421
2422         if (h == NULL) {
2423                 err = -ENOENT;
2424                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2425                         enum ip_conntrack_events events;
2426
2427                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2428                                 return -EINVAL;
2429                         if (otuple.dst.protonum != rtuple.dst.protonum)
2430                                 return -EINVAL;
2431
2432                         ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2433                                                         &otuple, &rtuple, u3);
2434                         if (IS_ERR(ct))
2435                                 return PTR_ERR(ct);
2436
2437                         err = 0;
2438                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2439                                 events = 1 << IPCT_RELATED;
2440                         else
2441                                 events = 1 << IPCT_NEW;
2442
2443                         if (cda[CTA_LABELS] &&
2444                             ctnetlink_attach_labels(ct, cda) == 0)
2445                                 events |= (1 << IPCT_LABEL);
2446
2447                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2448                                                       (1 << IPCT_ASSURED) |
2449                                                       (1 << IPCT_HELPER) |
2450                                                       (1 << IPCT_PROTOINFO) |
2451                                                       (1 << IPCT_SEQADJ) |
2452                                                       (1 << IPCT_MARK) |
2453                                                       (1 << IPCT_SYNPROXY) |
2454                                                       events,
2455                                                       ct, NETLINK_CB(skb).portid,
2456                                                       nlmsg_report(info->nlh));
2457                         nf_ct_put(ct);
2458                 }
2459
2460                 return err;
2461         }
2462         /* implicit 'else' */
2463
2464         err = -EEXIST;
2465         ct = nf_ct_tuplehash_to_ctrack(h);
2466         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2467                 err = ctnetlink_change_conntrack(ct, cda);
2468                 if (err == 0) {
2469                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2470                                                       (1 << IPCT_ASSURED) |
2471                                                       (1 << IPCT_HELPER) |
2472                                                       (1 << IPCT_LABEL) |
2473                                                       (1 << IPCT_PROTOINFO) |
2474                                                       (1 << IPCT_SEQADJ) |
2475                                                       (1 << IPCT_MARK) |
2476                                                       (1 << IPCT_SYNPROXY),
2477                                                       ct, NETLINK_CB(skb).portid,
2478                                                       nlmsg_report(info->nlh));
2479                 }
2480         }
2481
2482         nf_ct_put(ct);
2483         return err;
2484 }
2485
2486 static int
2487 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2488                                 __u16 cpu, const struct ip_conntrack_stat *st)
2489 {
2490         struct nlmsghdr *nlh;
2491         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2492
2493         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2494                               IPCTNL_MSG_CT_GET_STATS_CPU);
2495         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2496                            NFNETLINK_V0, htons(cpu));
2497         if (!nlh)
2498                 goto nlmsg_failure;
2499
2500         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2501             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2502             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2503             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2504                                 htonl(st->insert_failed)) ||
2505             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2506             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2507             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2508             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2509                                 htonl(st->search_restart)) ||
2510             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2511                                 htonl(st->clash_resolve)) ||
2512             nla_put_be32(skb, CTA_STATS_CHAIN_TOOLONG,
2513                          htonl(st->chaintoolong)))
2514                 goto nla_put_failure;
2515
2516         nlmsg_end(skb, nlh);
2517         return skb->len;
2518
2519 nla_put_failure:
2520 nlmsg_failure:
2521         nlmsg_cancel(skb, nlh);
2522         return -1;
2523 }
2524
2525 static int
2526 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2527 {
2528         int cpu;
2529         struct net *net = sock_net(skb->sk);
2530
2531         if (cb->args[0] == nr_cpu_ids)
2532                 return 0;
2533
2534         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2535                 const struct ip_conntrack_stat *st;
2536
2537                 if (!cpu_possible(cpu))
2538                         continue;
2539
2540                 st = per_cpu_ptr(net->ct.stat, cpu);
2541                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2542                                                     NETLINK_CB(cb->skb).portid,
2543                                                     cb->nlh->nlmsg_seq,
2544                                                     cpu, st) < 0)
2545                                 break;
2546         }
2547         cb->args[0] = cpu;
2548
2549         return skb->len;
2550 }
2551
2552 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2553                                  const struct nfnl_info *info,
2554                                  const struct nlattr * const cda[])
2555 {
2556         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2557                 struct netlink_dump_control c = {
2558                         .dump = ctnetlink_ct_stat_cpu_dump,
2559                 };
2560                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2561         }
2562
2563         return 0;
2564 }
2565
2566 static int
2567 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2568                             struct net *net)
2569 {
2570         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2571         unsigned int nr_conntracks;
2572         struct nlmsghdr *nlh;
2573
2574         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2575         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2576                            NFNETLINK_V0, 0);
2577         if (!nlh)
2578                 goto nlmsg_failure;
2579
2580         nr_conntracks = nf_conntrack_count(net);
2581         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2582                 goto nla_put_failure;
2583
2584         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2585                 goto nla_put_failure;
2586
2587         nlmsg_end(skb, nlh);
2588         return skb->len;
2589
2590 nla_put_failure:
2591 nlmsg_failure:
2592         nlmsg_cancel(skb, nlh);
2593         return -1;
2594 }
2595
2596 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2597                              const struct nlattr * const cda[])
2598 {
2599         struct sk_buff *skb2;
2600         int err;
2601
2602         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2603         if (skb2 == NULL)
2604                 return -ENOMEM;
2605
2606         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2607                                           info->nlh->nlmsg_seq,
2608                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2609                                           sock_net(skb->sk));
2610         if (err <= 0) {
2611                 kfree_skb(skb2);
2612                 return -ENOMEM;
2613         }
2614
2615         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2616 }
2617
2618 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2619         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2620         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2621         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2622         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2623         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2624         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2625                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2626         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2627         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2628         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2629         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2630         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2631 };
2632
2633 static struct nf_conntrack_expect *
2634 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2635                        struct nf_conntrack_helper *helper,
2636                        struct nf_conntrack_tuple *tuple,
2637                        struct nf_conntrack_tuple *mask);
2638
2639 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2640 static size_t
2641 ctnetlink_glue_build_size(const struct nf_conn *ct)
2642 {
2643         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2644                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2645                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2646                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2647                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2648                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2649                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2650                + nla_total_size(0) /* CTA_PROTOINFO */
2651                + nla_total_size(0) /* CTA_HELP */
2652                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2653                + ctnetlink_secctx_size(ct)
2654                + ctnetlink_acct_size(ct)
2655                + ctnetlink_timestamp_size(ct)
2656 #if IS_ENABLED(CONFIG_NF_NAT)
2657                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2658                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2659 #endif
2660 #ifdef CONFIG_NF_CONNTRACK_MARK
2661                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2662 #endif
2663 #ifdef CONFIG_NF_CONNTRACK_ZONES
2664                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2665 #endif
2666                + ctnetlink_proto_size(ct)
2667                ;
2668 }
2669
2670 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2671 {
2672         const struct nf_conntrack_zone *zone;
2673         struct nlattr *nest_parms;
2674
2675         zone = nf_ct_zone(ct);
2676
2677         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2678         if (!nest_parms)
2679                 goto nla_put_failure;
2680         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2681                 goto nla_put_failure;
2682         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2683                                    NF_CT_ZONE_DIR_ORIG) < 0)
2684                 goto nla_put_failure;
2685         nla_nest_end(skb, nest_parms);
2686
2687         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2688         if (!nest_parms)
2689                 goto nla_put_failure;
2690         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2691                 goto nla_put_failure;
2692         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2693                                    NF_CT_ZONE_DIR_REPL) < 0)
2694                 goto nla_put_failure;
2695         nla_nest_end(skb, nest_parms);
2696
2697         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2698                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2699                 goto nla_put_failure;
2700
2701         if (ctnetlink_dump_id(skb, ct) < 0)
2702                 goto nla_put_failure;
2703
2704         if (ctnetlink_dump_status(skb, ct) < 0)
2705                 goto nla_put_failure;
2706
2707         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2708                 goto nla_put_failure;
2709
2710         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2711                 goto nla_put_failure;
2712
2713         if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2714             ctnetlink_dump_timestamp(skb, ct) < 0)
2715                 goto nla_put_failure;
2716
2717         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2718                 goto nla_put_failure;
2719
2720 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2721         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2722                 goto nla_put_failure;
2723 #endif
2724         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2725                 goto nla_put_failure;
2726
2727         if ((ct->status & IPS_SEQ_ADJUST) &&
2728             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2729                 goto nla_put_failure;
2730
2731         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2732                 goto nla_put_failure;
2733
2734 #ifdef CONFIG_NF_CONNTRACK_MARK
2735         if (ctnetlink_dump_mark(skb, ct, true) < 0)
2736                 goto nla_put_failure;
2737 #endif
2738         if (ctnetlink_dump_labels(skb, ct) < 0)
2739                 goto nla_put_failure;
2740         return 0;
2741
2742 nla_put_failure:
2743         return -ENOSPC;
2744 }
2745
2746 static int
2747 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2748                      enum ip_conntrack_info ctinfo,
2749                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2750 {
2751         struct nlattr *nest_parms;
2752
2753         nest_parms = nla_nest_start(skb, ct_attr);
2754         if (!nest_parms)
2755                 goto nla_put_failure;
2756
2757         if (__ctnetlink_glue_build(skb, ct) < 0)
2758                 goto nla_put_failure;
2759
2760         nla_nest_end(skb, nest_parms);
2761
2762         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2763                 goto nla_put_failure;
2764
2765         return 0;
2766
2767 nla_put_failure:
2768         return -ENOSPC;
2769 }
2770
2771 static int
2772 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2773 {
2774         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2775         unsigned long d = ct->status ^ status;
2776
2777         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2778                 /* SEEN_REPLY bit can only be set */
2779                 return -EBUSY;
2780
2781         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2782                 /* ASSURED bit can only be set */
2783                 return -EBUSY;
2784
2785         /* This check is less strict than ctnetlink_change_status()
2786          * because callers often flip IPS_EXPECTED bits when sending
2787          * an NFQA_CT attribute to the kernel.  So ignore the
2788          * unchangeable bits but do not error out. Also user programs
2789          * are allowed to clear the bits that they are allowed to change.
2790          */
2791         __nf_ct_change_status(ct, status, ~status);
2792         return 0;
2793 }
2794
2795 static int
2796 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2797 {
2798         int err;
2799
2800         if (cda[CTA_TIMEOUT]) {
2801                 err = ctnetlink_change_timeout(ct, cda);
2802                 if (err < 0)
2803                         return err;
2804         }
2805         if (cda[CTA_STATUS]) {
2806                 err = ctnetlink_update_status(ct, cda);
2807                 if (err < 0)
2808                         return err;
2809         }
2810         if (cda[CTA_HELP]) {
2811                 err = ctnetlink_change_helper(ct, cda);
2812                 if (err < 0)
2813                         return err;
2814         }
2815         if (cda[CTA_LABELS]) {
2816                 err = ctnetlink_attach_labels(ct, cda);
2817                 if (err < 0)
2818                         return err;
2819         }
2820 #if defined(CONFIG_NF_CONNTRACK_MARK)
2821         if (cda[CTA_MARK]) {
2822                 ctnetlink_change_mark(ct, cda);
2823         }
2824 #endif
2825         return 0;
2826 }
2827
2828 static int
2829 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2830 {
2831         struct nlattr *cda[CTA_MAX+1];
2832         int ret;
2833
2834         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2835                                           NULL);
2836         if (ret < 0)
2837                 return ret;
2838
2839         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2840 }
2841
2842 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2843                                     const struct nf_conn *ct,
2844                                     struct nf_conntrack_tuple *tuple,
2845                                     struct nf_conntrack_tuple *mask)
2846 {
2847         int err;
2848
2849         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2850                                     nf_ct_l3num(ct), NULL);
2851         if (err < 0)
2852                 return err;
2853
2854         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2855                                      nf_ct_l3num(ct), NULL);
2856 }
2857
2858 static int
2859 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2860                              u32 portid, u32 report)
2861 {
2862         struct nlattr *cda[CTA_EXPECT_MAX+1];
2863         struct nf_conntrack_tuple tuple, mask;
2864         struct nf_conntrack_helper *helper = NULL;
2865         struct nf_conntrack_expect *exp;
2866         int err;
2867
2868         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2869                                           exp_nla_policy, NULL);
2870         if (err < 0)
2871                 return err;
2872
2873         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2874                                        ct, &tuple, &mask);
2875         if (err < 0)
2876                 return err;
2877
2878         if (cda[CTA_EXPECT_HELP_NAME]) {
2879                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2880
2881                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2882                                                     nf_ct_protonum(ct));
2883                 if (helper == NULL)
2884                         return -EOPNOTSUPP;
2885         }
2886
2887         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2888                                      helper, &tuple, &mask);
2889         if (IS_ERR(exp))
2890                 return PTR_ERR(exp);
2891
2892         err = nf_ct_expect_related_report(exp, portid, report, 0);
2893         nf_ct_expect_put(exp);
2894         return err;
2895 }
2896
2897 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2898                                   enum ip_conntrack_info ctinfo, int diff)
2899 {
2900         if (!(ct->status & IPS_NAT_MASK))
2901                 return;
2902
2903         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2904 }
2905
2906 static const struct nfnl_ct_hook ctnetlink_glue_hook = {
2907         .build_size     = ctnetlink_glue_build_size,
2908         .build          = ctnetlink_glue_build,
2909         .parse          = ctnetlink_glue_parse,
2910         .attach_expect  = ctnetlink_glue_attach_expect,
2911         .seq_adjust     = ctnetlink_glue_seqadj,
2912 };
2913 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2914
2915 /***********************************************************************
2916  * EXPECT
2917  ***********************************************************************/
2918
2919 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2920                                     const struct nf_conntrack_tuple *tuple,
2921                                     u32 type)
2922 {
2923         struct nlattr *nest_parms;
2924
2925         nest_parms = nla_nest_start(skb, type);
2926         if (!nest_parms)
2927                 goto nla_put_failure;
2928         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2929                 goto nla_put_failure;
2930         nla_nest_end(skb, nest_parms);
2931
2932         return 0;
2933
2934 nla_put_failure:
2935         return -1;
2936 }
2937
2938 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2939                                    const struct nf_conntrack_tuple *tuple,
2940                                    const struct nf_conntrack_tuple_mask *mask)
2941 {
2942         const struct nf_conntrack_l4proto *l4proto;
2943         struct nf_conntrack_tuple m;
2944         struct nlattr *nest_parms;
2945         int ret;
2946
2947         memset(&m, 0xFF, sizeof(m));
2948         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2949         m.src.u.all = mask->src.u.all;
2950         m.src.l3num = tuple->src.l3num;
2951         m.dst.protonum = tuple->dst.protonum;
2952
2953         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2954         if (!nest_parms)
2955                 goto nla_put_failure;
2956
2957         rcu_read_lock();
2958         ret = ctnetlink_dump_tuples_ip(skb, &m);
2959         if (ret >= 0) {
2960                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2961                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2962         }
2963         rcu_read_unlock();
2964
2965         if (unlikely(ret < 0))
2966                 goto nla_put_failure;
2967
2968         nla_nest_end(skb, nest_parms);
2969
2970         return 0;
2971
2972 nla_put_failure:
2973         return -1;
2974 }
2975
2976 #if IS_ENABLED(CONFIG_NF_NAT)
2977 static const union nf_inet_addr any_addr;
2978 #endif
2979
2980 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2981 {
2982         static siphash_aligned_key_t exp_id_seed;
2983         unsigned long a, b, c, d;
2984
2985         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2986
2987         a = (unsigned long)exp;
2988         b = (unsigned long)exp->helper;
2989         c = (unsigned long)exp->master;
2990         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2991
2992 #ifdef CONFIG_64BIT
2993         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2994 #else
2995         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2996 #endif
2997 }
2998
2999 static int
3000 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3001                           const struct nf_conntrack_expect *exp)
3002 {
3003         struct nf_conn *master = exp->master;
3004         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3005         struct nf_conn_help *help;
3006 #if IS_ENABLED(CONFIG_NF_NAT)
3007         struct nlattr *nest_parms;
3008         struct nf_conntrack_tuple nat_tuple = {};
3009 #endif
3010         struct nf_ct_helper_expectfn *expfn;
3011
3012         if (timeout < 0)
3013                 timeout = 0;
3014
3015         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3016                 goto nla_put_failure;
3017         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3018                 goto nla_put_failure;
3019         if (ctnetlink_exp_dump_tuple(skb,
3020                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3021                                  CTA_EXPECT_MASTER) < 0)
3022                 goto nla_put_failure;
3023
3024 #if IS_ENABLED(CONFIG_NF_NAT)
3025         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3026             exp->saved_proto.all) {
3027                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3028                 if (!nest_parms)
3029                         goto nla_put_failure;
3030
3031                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3032                         goto nla_put_failure;
3033
3034                 nat_tuple.src.l3num = nf_ct_l3num(master);
3035                 nat_tuple.src.u3 = exp->saved_addr;
3036                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3037                 nat_tuple.src.u = exp->saved_proto;
3038
3039                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3040                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3041                         goto nla_put_failure;
3042                 nla_nest_end(skb, nest_parms);
3043         }
3044 #endif
3045         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3046             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3047             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3048             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3049                 goto nla_put_failure;
3050         help = nfct_help(master);
3051         if (help) {
3052                 struct nf_conntrack_helper *helper;
3053
3054                 helper = rcu_dereference(help->helper);
3055                 if (helper &&
3056                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3057                         goto nla_put_failure;
3058         }
3059         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3060         if (expfn != NULL &&
3061             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3062                 goto nla_put_failure;
3063
3064         return 0;
3065
3066 nla_put_failure:
3067         return -1;
3068 }
3069
3070 static int
3071 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3072                         int event, const struct nf_conntrack_expect *exp)
3073 {
3074         struct nlmsghdr *nlh;
3075         unsigned int flags = portid ? NLM_F_MULTI : 0;
3076
3077         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3078         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3079                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3080         if (!nlh)
3081                 goto nlmsg_failure;
3082
3083         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3084                 goto nla_put_failure;
3085
3086         nlmsg_end(skb, nlh);
3087         return skb->len;
3088
3089 nlmsg_failure:
3090 nla_put_failure:
3091         nlmsg_cancel(skb, nlh);
3092         return -1;
3093 }
3094
3095 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3096 static int
3097 ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3098 {
3099         struct nf_conntrack_expect *exp = item->exp;
3100         struct net *net = nf_ct_exp_net(exp);
3101         struct nlmsghdr *nlh;
3102         struct sk_buff *skb;
3103         unsigned int type, group;
3104         int flags = 0;
3105
3106         if (events & (1 << IPEXP_DESTROY)) {
3107                 type = IPCTNL_MSG_EXP_DELETE;
3108                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3109         } else if (events & (1 << IPEXP_NEW)) {
3110                 type = IPCTNL_MSG_EXP_NEW;
3111                 flags = NLM_F_CREATE|NLM_F_EXCL;
3112                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3113         } else
3114                 return 0;
3115
3116         if (!item->report && !nfnetlink_has_listeners(net, group))
3117                 return 0;
3118
3119         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3120         if (skb == NULL)
3121                 goto errout;
3122
3123         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3124         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3125                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3126         if (!nlh)
3127                 goto nlmsg_failure;
3128
3129         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3130                 goto nla_put_failure;
3131
3132         nlmsg_end(skb, nlh);
3133         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3134         return 0;
3135
3136 nla_put_failure:
3137         nlmsg_cancel(skb, nlh);
3138 nlmsg_failure:
3139         kfree_skb(skb);
3140 errout:
3141         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3142         return 0;
3143 }
3144 #endif
3145 static int ctnetlink_exp_done(struct netlink_callback *cb)
3146 {
3147         if (cb->args[1])
3148                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3149         return 0;
3150 }
3151
3152 static int
3153 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3154 {
3155         struct net *net = sock_net(skb->sk);
3156         struct nf_conntrack_expect *exp, *last;
3157         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3158         u_int8_t l3proto = nfmsg->nfgen_family;
3159
3160         rcu_read_lock();
3161         last = (struct nf_conntrack_expect *)cb->args[1];
3162         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3163 restart:
3164                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3165                                          hnode) {
3166                         if (l3proto && exp->tuple.src.l3num != l3proto)
3167                                 continue;
3168
3169                         if (!net_eq(nf_ct_net(exp->master), net))
3170                                 continue;
3171
3172                         if (cb->args[1]) {
3173                                 if (exp != last)
3174                                         continue;
3175                                 cb->args[1] = 0;
3176                         }
3177                         if (ctnetlink_exp_fill_info(skb,
3178                                                     NETLINK_CB(cb->skb).portid,
3179                                                     cb->nlh->nlmsg_seq,
3180                                                     IPCTNL_MSG_EXP_NEW,
3181                                                     exp) < 0) {
3182                                 if (!refcount_inc_not_zero(&exp->use))
3183                                         continue;
3184                                 cb->args[1] = (unsigned long)exp;
3185                                 goto out;
3186                         }
3187                 }
3188                 if (cb->args[1]) {
3189                         cb->args[1] = 0;
3190                         goto restart;
3191                 }
3192         }
3193 out:
3194         rcu_read_unlock();
3195         if (last)
3196                 nf_ct_expect_put(last);
3197
3198         return skb->len;
3199 }
3200
3201 static int
3202 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3203 {
3204         struct nf_conntrack_expect *exp, *last;
3205         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3206         struct nf_conn *ct = cb->data;
3207         struct nf_conn_help *help = nfct_help(ct);
3208         u_int8_t l3proto = nfmsg->nfgen_family;
3209
3210         if (cb->args[0])
3211                 return 0;
3212
3213         rcu_read_lock();
3214         last = (struct nf_conntrack_expect *)cb->args[1];
3215 restart:
3216         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3217                 if (l3proto && exp->tuple.src.l3num != l3proto)
3218                         continue;
3219                 if (cb->args[1]) {
3220                         if (exp != last)
3221                                 continue;
3222                         cb->args[1] = 0;
3223                 }
3224                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3225                                             cb->nlh->nlmsg_seq,
3226                                             IPCTNL_MSG_EXP_NEW,
3227                                             exp) < 0) {
3228                         if (!refcount_inc_not_zero(&exp->use))
3229                                 continue;
3230                         cb->args[1] = (unsigned long)exp;
3231                         goto out;
3232                 }
3233         }
3234         if (cb->args[1]) {
3235                 cb->args[1] = 0;
3236                 goto restart;
3237         }
3238         cb->args[0] = 1;
3239 out:
3240         rcu_read_unlock();
3241         if (last)
3242                 nf_ct_expect_put(last);
3243
3244         return skb->len;
3245 }
3246
3247 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3248                                  struct sk_buff *skb,
3249                                  const struct nlmsghdr *nlh,
3250                                  const struct nlattr * const cda[],
3251                                  struct netlink_ext_ack *extack)
3252 {
3253         int err;
3254         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3255         u_int8_t u3 = nfmsg->nfgen_family;
3256         struct nf_conntrack_tuple tuple;
3257         struct nf_conntrack_tuple_hash *h;
3258         struct nf_conn *ct;
3259         struct nf_conntrack_zone zone;
3260         struct netlink_dump_control c = {
3261                 .dump = ctnetlink_exp_ct_dump_table,
3262                 .done = ctnetlink_exp_done,
3263         };
3264
3265         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3266                                     u3, NULL);
3267         if (err < 0)
3268                 return err;
3269
3270         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3271         if (err < 0)
3272                 return err;
3273
3274         h = nf_conntrack_find_get(net, &zone, &tuple);
3275         if (!h)
3276                 return -ENOENT;
3277
3278         ct = nf_ct_tuplehash_to_ctrack(h);
3279         /* No expectation linked to this connection tracking. */
3280         if (!nfct_help(ct)) {
3281                 nf_ct_put(ct);
3282                 return 0;
3283         }
3284
3285         c.data = ct;
3286
3287         err = netlink_dump_start(ctnl, skb, nlh, &c);
3288         nf_ct_put(ct);
3289
3290         return err;
3291 }
3292
3293 static int ctnetlink_get_expect(struct sk_buff *skb,
3294                                 const struct nfnl_info *info,
3295                                 const struct nlattr * const cda[])
3296 {
3297         u_int8_t u3 = info->nfmsg->nfgen_family;
3298         struct nf_conntrack_tuple tuple;
3299         struct nf_conntrack_expect *exp;
3300         struct nf_conntrack_zone zone;
3301         struct sk_buff *skb2;
3302         int err;
3303
3304         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3305                 if (cda[CTA_EXPECT_MASTER])
3306                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3307                                                      info->nlh, cda,
3308                                                      info->extack);
3309                 else {
3310                         struct netlink_dump_control c = {
3311                                 .dump = ctnetlink_exp_dump_table,
3312                                 .done = ctnetlink_exp_done,
3313                         };
3314                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3315                 }
3316         }
3317
3318         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3319         if (err < 0)
3320                 return err;
3321
3322         if (cda[CTA_EXPECT_TUPLE])
3323                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3324                                             u3, NULL);
3325         else if (cda[CTA_EXPECT_MASTER])
3326                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3327                                             u3, NULL);
3328         else
3329                 return -EINVAL;
3330
3331         if (err < 0)
3332                 return err;
3333
3334         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3335         if (!exp)
3336                 return -ENOENT;
3337
3338         if (cda[CTA_EXPECT_ID]) {
3339                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3340
3341                 if (id != nf_expect_get_id(exp)) {
3342                         nf_ct_expect_put(exp);
3343                         return -ENOENT;
3344                 }
3345         }
3346
3347         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3348         if (!skb2) {
3349                 nf_ct_expect_put(exp);
3350                 return -ENOMEM;
3351         }
3352
3353         rcu_read_lock();
3354         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3355                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3356                                       exp);
3357         rcu_read_unlock();
3358         nf_ct_expect_put(exp);
3359         if (err <= 0) {
3360                 kfree_skb(skb2);
3361                 return -ENOMEM;
3362         }
3363
3364         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3365 }
3366
3367 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3368 {
3369         struct nf_conntrack_helper *helper;
3370         const struct nf_conn_help *m_help;
3371         const char *name = data;
3372
3373         m_help = nfct_help(exp->master);
3374
3375         helper = rcu_dereference(m_help->helper);
3376         if (!helper)
3377                 return false;
3378
3379         return strcmp(helper->name, name) == 0;
3380 }
3381
3382 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3383 {
3384         return true;
3385 }
3386
3387 static int ctnetlink_del_expect(struct sk_buff *skb,
3388                                 const struct nfnl_info *info,
3389                                 const struct nlattr * const cda[])
3390 {
3391         u_int8_t u3 = info->nfmsg->nfgen_family;
3392         struct nf_conntrack_expect *exp;
3393         struct nf_conntrack_tuple tuple;
3394         struct nf_conntrack_zone zone;
3395         int err;
3396
3397         if (cda[CTA_EXPECT_TUPLE]) {
3398                 /* delete a single expect by tuple */
3399                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3400                 if (err < 0)
3401                         return err;
3402
3403                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3404                                             u3, NULL);
3405                 if (err < 0)
3406                         return err;
3407
3408                 /* bump usage count to 2 */
3409                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3410                 if (!exp)
3411                         return -ENOENT;
3412
3413                 if (cda[CTA_EXPECT_ID]) {
3414                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3415                         if (ntohl(id) != (u32)(unsigned long)exp) {
3416                                 nf_ct_expect_put(exp);
3417                                 return -ENOENT;
3418                         }
3419                 }
3420
3421                 /* after list removal, usage count == 1 */
3422                 spin_lock_bh(&nf_conntrack_expect_lock);
3423                 if (del_timer(&exp->timeout)) {
3424                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3425                                                    nlmsg_report(info->nlh));
3426                         nf_ct_expect_put(exp);
3427                 }
3428                 spin_unlock_bh(&nf_conntrack_expect_lock);
3429                 /* have to put what we 'get' above.
3430                  * after this line usage count == 0 */
3431                 nf_ct_expect_put(exp);
3432         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3433                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3434
3435                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3436                                          NETLINK_CB(skb).portid,
3437                                          nlmsg_report(info->nlh));
3438         } else {
3439                 /* This basically means we have to flush everything*/
3440                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3441                                          NETLINK_CB(skb).portid,
3442                                          nlmsg_report(info->nlh));
3443         }
3444
3445         return 0;
3446 }
3447 static int
3448 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3449                         const struct nlattr * const cda[])
3450 {
3451         if (cda[CTA_EXPECT_TIMEOUT]) {
3452                 if (!del_timer(&x->timeout))
3453                         return -ETIME;
3454
3455                 x->timeout.expires = jiffies +
3456                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3457                 add_timer(&x->timeout);
3458         }
3459         return 0;
3460 }
3461
3462 #if IS_ENABLED(CONFIG_NF_NAT)
3463 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3464         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3465         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3466 };
3467 #endif
3468
3469 static int
3470 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3471                            struct nf_conntrack_expect *exp,
3472                            u_int8_t u3)
3473 {
3474 #if IS_ENABLED(CONFIG_NF_NAT)
3475         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3476         struct nf_conntrack_tuple nat_tuple = {};
3477         int err;
3478
3479         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3480                                           exp_nat_nla_policy, NULL);
3481         if (err < 0)
3482                 return err;
3483
3484         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3485                 return -EINVAL;
3486
3487         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3488                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3489                                     u3, NULL);
3490         if (err < 0)
3491                 return err;
3492
3493         exp->saved_addr = nat_tuple.src.u3;
3494         exp->saved_proto = nat_tuple.src.u;
3495         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3496
3497         return 0;
3498 #else
3499         return -EOPNOTSUPP;
3500 #endif
3501 }
3502
3503 static struct nf_conntrack_expect *
3504 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3505                        struct nf_conntrack_helper *helper,
3506                        struct nf_conntrack_tuple *tuple,
3507                        struct nf_conntrack_tuple *mask)
3508 {
3509         u_int32_t class = 0;
3510         struct nf_conntrack_expect *exp;
3511         struct nf_conn_help *help;
3512         int err;
3513
3514         help = nfct_help(ct);
3515         if (!help)
3516                 return ERR_PTR(-EOPNOTSUPP);
3517
3518         if (cda[CTA_EXPECT_CLASS] && helper) {
3519                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3520                 if (class > helper->expect_class_max)
3521                         return ERR_PTR(-EINVAL);
3522         }
3523         exp = nf_ct_expect_alloc(ct);
3524         if (!exp)
3525                 return ERR_PTR(-ENOMEM);
3526
3527         if (cda[CTA_EXPECT_FLAGS]) {
3528                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3529                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3530         } else {
3531                 exp->flags = 0;
3532         }
3533         if (cda[CTA_EXPECT_FN]) {
3534                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3535                 struct nf_ct_helper_expectfn *expfn;
3536
3537                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3538                 if (expfn == NULL) {
3539                         err = -EINVAL;
3540                         goto err_out;
3541                 }
3542                 exp->expectfn = expfn->expectfn;
3543         } else
3544                 exp->expectfn = NULL;
3545
3546         exp->class = class;
3547         exp->master = ct;
3548         exp->helper = helper;
3549         exp->tuple = *tuple;
3550         exp->mask.src.u3 = mask->src.u3;
3551         exp->mask.src.u.all = mask->src.u.all;
3552
3553         if (cda[CTA_EXPECT_NAT]) {
3554                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3555                                                  exp, nf_ct_l3num(ct));
3556                 if (err < 0)
3557                         goto err_out;
3558         }
3559         return exp;
3560 err_out:
3561         nf_ct_expect_put(exp);
3562         return ERR_PTR(err);
3563 }
3564
3565 static int
3566 ctnetlink_create_expect(struct net *net,
3567                         const struct nf_conntrack_zone *zone,
3568                         const struct nlattr * const cda[],
3569                         u_int8_t u3, u32 portid, int report)
3570 {
3571         struct nf_conntrack_tuple tuple, mask, master_tuple;
3572         struct nf_conntrack_tuple_hash *h = NULL;
3573         struct nf_conntrack_helper *helper = NULL;
3574         struct nf_conntrack_expect *exp;
3575         struct nf_conn *ct;
3576         int err;
3577
3578         /* caller guarantees that those three CTA_EXPECT_* exist */
3579         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3580                                     u3, NULL);
3581         if (err < 0)
3582                 return err;
3583         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3584                                     u3, NULL);
3585         if (err < 0)
3586                 return err;
3587         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3588                                     u3, NULL);
3589         if (err < 0)
3590                 return err;
3591
3592         /* Look for master conntrack of this expectation */
3593         h = nf_conntrack_find_get(net, zone, &master_tuple);
3594         if (!h)
3595                 return -ENOENT;
3596         ct = nf_ct_tuplehash_to_ctrack(h);
3597
3598         rcu_read_lock();
3599         if (cda[CTA_EXPECT_HELP_NAME]) {
3600                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3601
3602                 helper = __nf_conntrack_helper_find(helpname, u3,
3603                                                     nf_ct_protonum(ct));
3604                 if (helper == NULL) {
3605                         rcu_read_unlock();
3606 #ifdef CONFIG_MODULES
3607                         if (request_module("nfct-helper-%s", helpname) < 0) {
3608                                 err = -EOPNOTSUPP;
3609                                 goto err_ct;
3610                         }
3611                         rcu_read_lock();
3612                         helper = __nf_conntrack_helper_find(helpname, u3,
3613                                                             nf_ct_protonum(ct));
3614                         if (helper) {
3615                                 err = -EAGAIN;
3616                                 goto err_rcu;
3617                         }
3618                         rcu_read_unlock();
3619 #endif
3620                         err = -EOPNOTSUPP;
3621                         goto err_ct;
3622                 }
3623         }
3624
3625         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3626         if (IS_ERR(exp)) {
3627                 err = PTR_ERR(exp);
3628                 goto err_rcu;
3629         }
3630
3631         err = nf_ct_expect_related_report(exp, portid, report, 0);
3632         nf_ct_expect_put(exp);
3633 err_rcu:
3634         rcu_read_unlock();
3635 err_ct:
3636         nf_ct_put(ct);
3637         return err;
3638 }
3639
3640 static int ctnetlink_new_expect(struct sk_buff *skb,
3641                                 const struct nfnl_info *info,
3642                                 const struct nlattr * const cda[])
3643 {
3644         u_int8_t u3 = info->nfmsg->nfgen_family;
3645         struct nf_conntrack_tuple tuple;
3646         struct nf_conntrack_expect *exp;
3647         struct nf_conntrack_zone zone;
3648         int err;
3649
3650         if (!cda[CTA_EXPECT_TUPLE]
3651             || !cda[CTA_EXPECT_MASK]
3652             || !cda[CTA_EXPECT_MASTER])
3653                 return -EINVAL;
3654
3655         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3656         if (err < 0)
3657                 return err;
3658
3659         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3660                                     u3, NULL);
3661         if (err < 0)
3662                 return err;
3663
3664         spin_lock_bh(&nf_conntrack_expect_lock);
3665         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3666         if (!exp) {
3667                 spin_unlock_bh(&nf_conntrack_expect_lock);
3668                 err = -ENOENT;
3669                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3670                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3671                                                       NETLINK_CB(skb).portid,
3672                                                       nlmsg_report(info->nlh));
3673                 }
3674                 return err;
3675         }
3676
3677         err = -EEXIST;
3678         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3679                 err = ctnetlink_change_expect(exp, cda);
3680         spin_unlock_bh(&nf_conntrack_expect_lock);
3681
3682         return err;
3683 }
3684
3685 static int
3686 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3687                              const struct ip_conntrack_stat *st)
3688 {
3689         struct nlmsghdr *nlh;
3690         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3691
3692         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3693                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3694         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3695                            NFNETLINK_V0, htons(cpu));
3696         if (!nlh)
3697                 goto nlmsg_failure;
3698
3699         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3700             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3701             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3702                 goto nla_put_failure;
3703
3704         nlmsg_end(skb, nlh);
3705         return skb->len;
3706
3707 nla_put_failure:
3708 nlmsg_failure:
3709         nlmsg_cancel(skb, nlh);
3710         return -1;
3711 }
3712
3713 static int
3714 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3715 {
3716         int cpu;
3717         struct net *net = sock_net(skb->sk);
3718
3719         if (cb->args[0] == nr_cpu_ids)
3720                 return 0;
3721
3722         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3723                 const struct ip_conntrack_stat *st;
3724
3725                 if (!cpu_possible(cpu))
3726                         continue;
3727
3728                 st = per_cpu_ptr(net->ct.stat, cpu);
3729                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3730                                                  cb->nlh->nlmsg_seq,
3731                                                  cpu, st) < 0)
3732                         break;
3733         }
3734         cb->args[0] = cpu;
3735
3736         return skb->len;
3737 }
3738
3739 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3740                                   const struct nfnl_info *info,
3741                                   const struct nlattr * const cda[])
3742 {
3743         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3744                 struct netlink_dump_control c = {
3745                         .dump = ctnetlink_exp_stat_cpu_dump,
3746                 };
3747                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3748         }
3749
3750         return 0;
3751 }
3752
3753 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3754 static struct nf_ct_event_notifier ctnl_notifier = {
3755         .ct_event = ctnetlink_conntrack_event,
3756         .exp_event = ctnetlink_expect_event,
3757 };
3758 #endif
3759
3760 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3761         [IPCTNL_MSG_CT_NEW]     = {
3762                 .call           = ctnetlink_new_conntrack,
3763                 .type           = NFNL_CB_MUTEX,
3764                 .attr_count     = CTA_MAX,
3765                 .policy         = ct_nla_policy
3766         },
3767         [IPCTNL_MSG_CT_GET]     = {
3768                 .call           = ctnetlink_get_conntrack,
3769                 .type           = NFNL_CB_MUTEX,
3770                 .attr_count     = CTA_MAX,
3771                 .policy         = ct_nla_policy
3772         },
3773         [IPCTNL_MSG_CT_DELETE]  = {
3774                 .call           = ctnetlink_del_conntrack,
3775                 .type           = NFNL_CB_MUTEX,
3776                 .attr_count     = CTA_MAX,
3777                 .policy         = ct_nla_policy
3778         },
3779         [IPCTNL_MSG_CT_GET_CTRZERO] = {
3780                 .call           = ctnetlink_get_conntrack,
3781                 .type           = NFNL_CB_MUTEX,
3782                 .attr_count     = CTA_MAX,
3783                 .policy         = ct_nla_policy
3784         },
3785         [IPCTNL_MSG_CT_GET_STATS_CPU] = {
3786                 .call           = ctnetlink_stat_ct_cpu,
3787                 .type           = NFNL_CB_MUTEX,
3788         },
3789         [IPCTNL_MSG_CT_GET_STATS] = {
3790                 .call           = ctnetlink_stat_ct,
3791                 .type           = NFNL_CB_MUTEX,
3792         },
3793         [IPCTNL_MSG_CT_GET_DYING] = {
3794                 .call           = ctnetlink_get_ct_dying,
3795                 .type           = NFNL_CB_MUTEX,
3796         },
3797         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = {
3798                 .call           = ctnetlink_get_ct_unconfirmed,
3799                 .type           = NFNL_CB_MUTEX,
3800         },
3801 };
3802
3803 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3804         [IPCTNL_MSG_EXP_GET] = {
3805                 .call           = ctnetlink_get_expect,
3806                 .type           = NFNL_CB_MUTEX,
3807                 .attr_count     = CTA_EXPECT_MAX,
3808                 .policy         = exp_nla_policy
3809         },
3810         [IPCTNL_MSG_EXP_NEW] = {
3811                 .call           = ctnetlink_new_expect,
3812                 .type           = NFNL_CB_MUTEX,
3813                 .attr_count     = CTA_EXPECT_MAX,
3814                 .policy         = exp_nla_policy
3815         },
3816         [IPCTNL_MSG_EXP_DELETE] = {
3817                 .call           = ctnetlink_del_expect,
3818                 .type           = NFNL_CB_MUTEX,
3819                 .attr_count     = CTA_EXPECT_MAX,
3820                 .policy         = exp_nla_policy
3821         },
3822         [IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3823                 .call           = ctnetlink_stat_exp_cpu,
3824                 .type           = NFNL_CB_MUTEX,
3825         },
3826 };
3827
3828 static const struct nfnetlink_subsystem ctnl_subsys = {
3829         .name                           = "conntrack",
3830         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3831         .cb_count                       = IPCTNL_MSG_MAX,
3832         .cb                             = ctnl_cb,
3833 };
3834
3835 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3836         .name                           = "conntrack_expect",
3837         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3838         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3839         .cb                             = ctnl_exp_cb,
3840 };
3841
3842 MODULE_ALIAS("ip_conntrack_netlink");
3843 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3844 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3845
3846 static int __net_init ctnetlink_net_init(struct net *net)
3847 {
3848 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3849         nf_conntrack_register_notifier(net, &ctnl_notifier);
3850 #endif
3851         return 0;
3852 }
3853
3854 static void ctnetlink_net_pre_exit(struct net *net)
3855 {
3856 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3857         nf_conntrack_unregister_notifier(net);
3858 #endif
3859 }
3860
3861 static struct pernet_operations ctnetlink_net_ops = {
3862         .init           = ctnetlink_net_init,
3863         .pre_exit       = ctnetlink_net_pre_exit,
3864 };
3865
3866 static int __init ctnetlink_init(void)
3867 {
3868         int ret;
3869
3870         NL_ASSERT_DUMP_CTX_FITS(struct ctnetlink_list_dump_ctx);
3871
3872         ret = nfnetlink_subsys_register(&ctnl_subsys);
3873         if (ret < 0) {
3874                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3875                 goto err_out;
3876         }
3877
3878         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3879         if (ret < 0) {
3880                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3881                 goto err_unreg_subsys;
3882         }
3883
3884         ret = register_pernet_subsys(&ctnetlink_net_ops);
3885         if (ret < 0) {
3886                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3887                 goto err_unreg_exp_subsys;
3888         }
3889 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3890         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3891         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3892 #endif
3893         return 0;
3894
3895 err_unreg_exp_subsys:
3896         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3897 err_unreg_subsys:
3898         nfnetlink_subsys_unregister(&ctnl_subsys);
3899 err_out:
3900         return ret;
3901 }
3902
3903 static void __exit ctnetlink_exit(void)
3904 {
3905         unregister_pernet_subsys(&ctnetlink_net_ops);
3906         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3907         nfnetlink_subsys_unregister(&ctnl_subsys);
3908 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3909         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3910 #endif
3911         synchronize_rcu();
3912 }
3913
3914 module_init(ctnetlink_init);
3915 module_exit(ctnetlink_exit);