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         bool zone_filter;
 880
 881         u_int32_t orig_flags;
 882         u_int32_t reply_flags;
 883
 884         struct nf_conntrack_tuple orig;
 885         struct nf_conntrack_tuple reply;
 886         struct nf_conntrack_zone zone;
 887
 888         struct ctnetlink_filter_u32 mark;
 889         struct ctnetlink_filter_u32 status;
 890 };
 891
 892 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 893         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 894         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 895 };
 896
 897 static int ctnetlink_parse_filter(const struct nlattr *attr,
 898                                   struct ctnetlink_filter *filter)
 899 {
 900         struct nlattr *tb[CTA_FILTER_MAX + 1];
 901         int ret = 0;
 902
 903         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 904                                NULL);
 905         if (ret)
 906                 return ret;
 907
 908         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 909                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 910                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 911                         return -EOPNOTSUPP;
 912         }
 913
 914         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 915                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 916                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 917                         return -EOPNOTSUPP;
 918         }
 919
 920         return 0;
 921 }
 922
 923 static int ctnetlink_parse_zone(const struct nlattr *attr,
 924                                 struct nf_conntrack_zone *zone);
 925 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 926                                          struct nf_conntrack_tuple *tuple,
 927                                          u32 type, u_int8_t l3num,
 928                                          struct nf_conntrack_zone *zone,
 929                                          u_int32_t flags);
 930
 931 static int ctnetlink_filter_parse_mark(struct ctnetlink_filter_u32 *mark,
 932                                        const struct nlattr * const cda[])
 933 {
 934 #ifdef CONFIG_NF_CONNTRACK_MARK
 935         if (cda[CTA_MARK]) {
 936                 mark->val = ntohl(nla_get_be32(cda[CTA_MARK]));
 937
 938                 if (cda[CTA_MARK_MASK])
 939                         mark->mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 940                 else
 941                         mark->mask = 0xffffffff;
 942         } else if (cda[CTA_MARK_MASK]) {
 943                 return -EINVAL;
 944         }
 945 #endif
 946         return 0;
 947 }
 948
 949 static int ctnetlink_filter_parse_status(struct ctnetlink_filter_u32 *status,
 950                                          const struct nlattr * const cda[])
 951 {
 952         if (cda[CTA_STATUS]) {
 953                 status->val = ntohl(nla_get_be32(cda[CTA_STATUS]));
 954                 if (cda[CTA_STATUS_MASK])
 955                         status->mask = ntohl(nla_get_be32(cda[CTA_STATUS_MASK]));
 956                 else
 957                         status->mask = status->val;
 958
 959                 /* status->val == 0? always true, else always false. */
 960                 if (status->mask == 0)
 961                         return -EINVAL;
 962         } else if (cda[CTA_STATUS_MASK]) {
 963                 return -EINVAL;
 964         }
 965
 966         /* CTA_STATUS is NLA_U32, if this fires UAPI needs to be extended */
 967         BUILD_BUG_ON(__IPS_MAX_BIT >= 32);
 968         return 0;
 969 }
 970
 971 static struct ctnetlink_filter *
 972 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 973 {
 974         struct ctnetlink_filter *filter;
 975         int err;
 976
 977 #ifndef CONFIG_NF_CONNTRACK_MARK
 978         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 979                 return ERR_PTR(-EOPNOTSUPP);
 980 #endif
 981
 982         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 983         if (filter == NULL)
 984                 return ERR_PTR(-ENOMEM);
 985
 986         filter->family = family;
 987
 988         err = ctnetlink_filter_parse_mark(&filter->mark, cda);
 989         if (err)
 990                 goto err_filter;
 991
 992         err = ctnetlink_filter_parse_status(&filter->status, cda);
 993         if (err)
 994                 goto err_filter;
 995
 996         if (cda[CTA_ZONE]) {
 997                 err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 998                 if (err < 0)
 999                         goto err_filter;
1000                 filter->zone_filter = true;
1001         }
1002
1003         if (!cda[CTA_FILTER])
1004                 return filter;
1005
1006         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
1007         if (err < 0)
1008                 goto err_filter;
1009
1010         if (filter->orig_flags) {
1011                 if (!cda[CTA_TUPLE_ORIG]) {
1012                         err = -EINVAL;
1013                         goto err_filter;
1014                 }
1015
1016                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
1017                                                    CTA_TUPLE_ORIG,
1018                                                    filter->family,
1019                                                    &filter->zone,
1020                                                    filter->orig_flags);
1021                 if (err < 0)
1022                         goto err_filter;
1023         }
1024
1025         if (filter->reply_flags) {
1026                 if (!cda[CTA_TUPLE_REPLY]) {
1027                         err = -EINVAL;
1028                         goto err_filter;
1029                 }
1030
1031                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
1032                                                    CTA_TUPLE_REPLY,
1033                                                    filter->family,
1034                                                    &filter->zone,
1035                                                    filter->reply_flags);
1036                 if (err < 0)
1037                         goto err_filter;
1038         }
1039
1040         return filter;
1041
1042 err_filter:
1043         kfree(filter);
1044
1045         return ERR_PTR(err);
1046 }
1047
1048 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1049 {
1050         return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS] || cda[CTA_ZONE];
1051 }
1052
1053 static int ctnetlink_start(struct netlink_callback *cb)
1054 {
1055         const struct nlattr * const *cda = cb->data;
1056         struct ctnetlink_filter *filter = NULL;
1057         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1058         u8 family = nfmsg->nfgen_family;
1059
1060         if (ctnetlink_needs_filter(family, cda)) {
1061                 filter = ctnetlink_alloc_filter(cda, family);
1062                 if (IS_ERR(filter))
1063                         return PTR_ERR(filter);
1064         }
1065
1066         cb->data = filter;
1067         return 0;
1068 }
1069
1070 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1071                                         struct nf_conntrack_tuple *ct_tuple,
1072                                         u_int32_t flags, int family)
1073 {
1074         switch (family) {
1075         case NFPROTO_IPV4:
1076                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1077                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1078                         return  0;
1079
1080                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1081                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1082                         return  0;
1083                 break;
1084         case NFPROTO_IPV6:
1085                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1086                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1087                                    &ct_tuple->src.u3.in6))
1088                         return 0;
1089
1090                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1091                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1092                                    &ct_tuple->dst.u3.in6))
1093                         return 0;
1094                 break;
1095         }
1096
1097         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1098             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1099                 return 0;
1100
1101         switch (ct_tuple->dst.protonum) {
1102         case IPPROTO_TCP:
1103         case IPPROTO_UDP:
1104                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1105                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1106                         return 0;
1107
1108                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1109                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1110                         return 0;
1111                 break;
1112         case IPPROTO_ICMP:
1113                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1114                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1115                         return 0;
1116                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1117                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1118                         return 0;
1119                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1120                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1121                         return 0;
1122                 break;
1123         case IPPROTO_ICMPV6:
1124                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1125                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1126                         return 0;
1127                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1128                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1129                         return 0;
1130                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1131                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1132                         return 0;
1133                 break;
1134         }
1135
1136         return 1;
1137 }
1138
1139 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1140 {
1141         struct ctnetlink_filter *filter = data;
1142         struct nf_conntrack_tuple *tuple;
1143         u32 status;
1144
1145         if (filter == NULL)
1146                 goto out;
1147
1148         /* Match entries of a given L3 protocol number.
1149          * If it is not specified, ie. l3proto == 0,
1150          * then match everything.
1151          */
1152         if (filter->family && nf_ct_l3num(ct) != filter->family)
1153                 goto ignore_entry;
1154
1155         if (filter->zone_filter &&
1156             !nf_ct_zone_equal_any(ct, &filter->zone))
1157                 goto ignore_entry;
1158
1159         if (filter->orig_flags) {
1160                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1161                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1162                                                   filter->orig_flags,
1163                                                   filter->family))
1164                         goto ignore_entry;
1165         }
1166
1167         if (filter->reply_flags) {
1168                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1169                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1170                                                   filter->reply_flags,
1171                                                   filter->family))
1172                         goto ignore_entry;
1173         }
1174
1175 #ifdef CONFIG_NF_CONNTRACK_MARK
1176         if ((READ_ONCE(ct->mark) & filter->mark.mask) != filter->mark.val)
1177                 goto ignore_entry;
1178 #endif
1179         status = (u32)READ_ONCE(ct->status);
1180         if ((status & filter->status.mask) != filter->status.val)
1181                 goto ignore_entry;
1182
1183 out:
1184         return 1;
1185
1186 ignore_entry:
1187         return 0;
1188 }
1189
1190 static int
1191 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1192 {
1193         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1194         struct net *net = sock_net(skb->sk);
1195         struct nf_conn *ct, *last;
1196         struct nf_conntrack_tuple_hash *h;
1197         struct hlist_nulls_node *n;
1198         struct nf_conn *nf_ct_evict[8];
1199         int res, i;
1200         spinlock_t *lockp;
1201
1202         last = (struct nf_conn *)cb->args[1];
1203         i = 0;
1204
1205         local_bh_disable();
1206         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1207 restart:
1208                 while (i) {
1209                         i--;
1210                         if (nf_ct_should_gc(nf_ct_evict[i]))
1211                                 nf_ct_kill(nf_ct_evict[i]);
1212                         nf_ct_put(nf_ct_evict[i]);
1213                 }
1214
1215                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1216                 nf_conntrack_lock(lockp);
1217                 if (cb->args[0] >= nf_conntrack_htable_size) {
1218                         spin_unlock(lockp);
1219                         goto out;
1220                 }
1221                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1222                                            hnnode) {
1223                         ct = nf_ct_tuplehash_to_ctrack(h);
1224                         if (nf_ct_is_expired(ct)) {
1225                                 /* need to defer nf_ct_kill() until lock is released */
1226                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1227                                     refcount_inc_not_zero(&ct->ct_general.use))
1228                                         nf_ct_evict[i++] = ct;
1229                                 continue;
1230                         }
1231
1232                         if (!net_eq(net, nf_ct_net(ct)))
1233                                 continue;
1234
1235                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1236                                 continue;
1237
1238                         if (cb->args[1]) {
1239                                 if (ct != last)
1240                                         continue;
1241                                 cb->args[1] = 0;
1242                         }
1243                         if (!ctnetlink_filter_match(ct, cb->data))
1244                                 continue;
1245
1246                         res =
1247                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1248                                             cb->nlh->nlmsg_seq,
1249                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1250                                             ct, true, flags);
1251                         if (res < 0) {
1252                                 nf_conntrack_get(&ct->ct_general);
1253                                 cb->args[1] = (unsigned long)ct;
1254                                 spin_unlock(lockp);
1255                                 goto out;
1256                         }
1257                 }
1258                 spin_unlock(lockp);
1259                 if (cb->args[1]) {
1260                         cb->args[1] = 0;
1261                         goto restart;
1262                 }
1263         }
1264 out:
1265         local_bh_enable();
1266         if (last) {
1267                 /* nf ct hash resize happened, now clear the leftover. */
1268                 if ((struct nf_conn *)cb->args[1] == last)
1269                         cb->args[1] = 0;
1270
1271                 nf_ct_put(last);
1272         }
1273
1274         while (i) {
1275                 i--;
1276                 if (nf_ct_should_gc(nf_ct_evict[i]))
1277                         nf_ct_kill(nf_ct_evict[i]);
1278                 nf_ct_put(nf_ct_evict[i]);
1279         }
1280
1281         return skb->len;
1282 }
1283
1284 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1285                                 struct nf_conntrack_tuple *t,
1286                                 u_int32_t flags)
1287 {
1288         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1289                 if (!tb[CTA_IP_V4_SRC])
1290                         return -EINVAL;
1291
1292                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1293         }
1294
1295         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1296                 if (!tb[CTA_IP_V4_DST])
1297                         return -EINVAL;
1298
1299                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1300         }
1301
1302         return 0;
1303 }
1304
1305 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1306                                 struct nf_conntrack_tuple *t,
1307                                 u_int32_t flags)
1308 {
1309         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1310                 if (!tb[CTA_IP_V6_SRC])
1311                         return -EINVAL;
1312
1313                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1314         }
1315
1316         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1317                 if (!tb[CTA_IP_V6_DST])
1318                         return -EINVAL;
1319
1320                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1321         }
1322
1323         return 0;
1324 }
1325
1326 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1327                                     struct nf_conntrack_tuple *tuple,
1328                                     u_int32_t flags)
1329 {
1330         struct nlattr *tb[CTA_IP_MAX+1];
1331         int ret = 0;
1332
1333         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr,
1334                                           cta_ip_nla_policy, NULL);
1335         if (ret < 0)
1336                 return ret;
1337
1338         switch (tuple->src.l3num) {
1339         case NFPROTO_IPV4:
1340                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1341                 break;
1342         case NFPROTO_IPV6:
1343                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1344                 break;
1345         }
1346
1347         return ret;
1348 }
1349
1350 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1351         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1352 };
1353
1354 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1355                                        struct nf_conntrack_tuple *tuple,
1356                                        u_int32_t flags)
1357 {
1358         const struct nf_conntrack_l4proto *l4proto;
1359         struct nlattr *tb[CTA_PROTO_MAX+1];
1360         int ret = 0;
1361
1362         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1363                                           proto_nla_policy, NULL);
1364         if (ret < 0)
1365                 return ret;
1366
1367         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1368                 return 0;
1369
1370         if (!tb[CTA_PROTO_NUM])
1371                 return -EINVAL;
1372
1373         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1374
1375         rcu_read_lock();
1376         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1377
1378         if (likely(l4proto->nlattr_to_tuple)) {
1379                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1380                                                      l4proto->nla_policy,
1381                                                      NULL);
1382                 if (ret == 0)
1383                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1384         }
1385
1386         rcu_read_unlock();
1387
1388         return ret;
1389 }
1390
1391 static int
1392 ctnetlink_parse_zone(const struct nlattr *attr,
1393                      struct nf_conntrack_zone *zone)
1394 {
1395         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1396                         NF_CT_DEFAULT_ZONE_DIR, 0);
1397 #ifdef CONFIG_NF_CONNTRACK_ZONES
1398         if (attr)
1399                 zone->id = ntohs(nla_get_be16(attr));
1400 #else
1401         if (attr)
1402                 return -EOPNOTSUPP;
1403 #endif
1404         return 0;
1405 }
1406
1407 static int
1408 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1409                            struct nf_conntrack_zone *zone)
1410 {
1411         int ret;
1412
1413         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1414                 return -EINVAL;
1415
1416         ret = ctnetlink_parse_zone(attr, zone);
1417         if (ret < 0)
1418                 return ret;
1419
1420         if (type == CTA_TUPLE_REPLY)
1421                 zone->dir = NF_CT_ZONE_DIR_REPL;
1422         else
1423                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1424
1425         return 0;
1426 }
1427
1428 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1429         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1430         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1431         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1432 };
1433
1434 #define CTA_FILTER_F_ALL_CTA_PROTO \
1435   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1436    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1437    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1438    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1439    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1440    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1441    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1442    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1443
1444 static int
1445 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1446                               struct nf_conntrack_tuple *tuple, u32 type,
1447                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1448                               u_int32_t flags)
1449 {
1450         struct nlattr *tb[CTA_TUPLE_MAX+1];
1451         int err;
1452
1453         memset(tuple, 0, sizeof(*tuple));
1454
1455         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1456                                           tuple_nla_policy, NULL);
1457         if (err < 0)
1458                 return err;
1459
1460         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1461                 return -EOPNOTSUPP;
1462         tuple->src.l3num = l3num;
1463
1464         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1465             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1466                 if (!tb[CTA_TUPLE_IP])
1467                         return -EINVAL;
1468
1469                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1470                 if (err < 0)
1471                         return err;
1472         }
1473
1474         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1475                 if (!tb[CTA_TUPLE_PROTO])
1476                         return -EINVAL;
1477
1478                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1479                 if (err < 0)
1480                         return err;
1481         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1482                 /* Can't manage proto flags without a protonum  */
1483                 return -EINVAL;
1484         }
1485
1486         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1487                 if (!zone)
1488                         return -EINVAL;
1489
1490                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1491                                                  type, zone);
1492                 if (err < 0)
1493                         return err;
1494         }
1495
1496         /* orig and expect tuples get DIR_ORIGINAL */
1497         if (type == CTA_TUPLE_REPLY)
1498                 tuple->dst.dir = IP_CT_DIR_REPLY;
1499         else
1500                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1501
1502         return 0;
1503 }
1504
1505 static int
1506 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1507                       struct nf_conntrack_tuple *tuple, u32 type,
1508                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1509 {
1510         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1511                                             CTA_FILTER_FLAG(ALL));
1512 }
1513
1514 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1515         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1516                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1517 };
1518
1519 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1520                                 struct nlattr **helpinfo)
1521 {
1522         int err;
1523         struct nlattr *tb[CTA_HELP_MAX+1];
1524
1525         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1526                                           help_nla_policy, NULL);
1527         if (err < 0)
1528                 return err;
1529
1530         if (!tb[CTA_HELP_NAME])
1531                 return -EINVAL;
1532
1533         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1534
1535         if (tb[CTA_HELP_INFO])
1536                 *helpinfo = tb[CTA_HELP_INFO];
1537
1538         return 0;
1539 }
1540
1541 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1542         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1543         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1544         [CTA_STATUS]            = { .type = NLA_U32 },
1545         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1546         [CTA_HELP]              = { .type = NLA_NESTED },
1547         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1548         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1549         [CTA_MARK]              = { .type = NLA_U32 },
1550         [CTA_ID]                = { .type = NLA_U32 },
1551         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1552         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1553         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1554         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1555         [CTA_ZONE]              = { .type = NLA_U16 },
1556         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1557         [CTA_LABELS]            = { .type = NLA_BINARY,
1558                                     .len = NF_CT_LABELS_MAX_SIZE },
1559         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1560                                     .len = NF_CT_LABELS_MAX_SIZE },
1561         [CTA_FILTER]            = { .type = NLA_NESTED },
1562         [CTA_STATUS_MASK]       = { .type = NLA_U32 },
1563 };
1564
1565 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1566 {
1567         return ctnetlink_filter_match(ct, data);
1568 }
1569
1570 static int ctnetlink_flush_conntrack(struct net *net,
1571                                      const struct nlattr * const cda[],
1572                                      u32 portid, int report, u8 family)
1573 {
1574         struct ctnetlink_filter *filter = NULL;
1575         struct nf_ct_iter_data iter = {
1576                 .net            = net,
1577                 .portid         = portid,
1578                 .report         = report,
1579         };
1580
1581         if (ctnetlink_needs_filter(family, cda)) {
1582                 if (cda[CTA_FILTER])
1583                         return -EOPNOTSUPP;
1584
1585                 filter = ctnetlink_alloc_filter(cda, family);
1586                 if (IS_ERR(filter))
1587                         return PTR_ERR(filter);
1588
1589                 iter.data = filter;
1590         }
1591
1592         nf_ct_iterate_cleanup_net(ctnetlink_flush_iterate, &iter);
1593         kfree(filter);
1594
1595         return 0;
1596 }
1597
1598 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1599                                    const struct nfnl_info *info,
1600                                    const struct nlattr * const cda[])
1601 {
1602         u8 family = info->nfmsg->nfgen_family;
1603         struct nf_conntrack_tuple_hash *h;
1604         struct nf_conntrack_tuple tuple;
1605         struct nf_conntrack_zone zone;
1606         struct nf_conn *ct;
1607         int err;
1608
1609         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1610         if (err < 0)
1611                 return err;
1612
1613         if (cda[CTA_TUPLE_ORIG])
1614                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1615                                             family, &zone);
1616         else if (cda[CTA_TUPLE_REPLY])
1617                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1618                                             family, &zone);
1619         else {
1620                 u_int8_t u3 = info->nfmsg->version ? family : AF_UNSPEC;
1621
1622                 return ctnetlink_flush_conntrack(info->net, cda,
1623                                                  NETLINK_CB(skb).portid,
1624                                                  nlmsg_report(info->nlh), u3);
1625         }
1626
1627         if (err < 0)
1628                 return err;
1629
1630         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1631         if (!h)
1632                 return -ENOENT;
1633
1634         ct = nf_ct_tuplehash_to_ctrack(h);
1635
1636         if (cda[CTA_ID]) {
1637                 __be32 id = nla_get_be32(cda[CTA_ID]);
1638
1639                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1640                         nf_ct_put(ct);
1641                         return -ENOENT;
1642                 }
1643         }
1644
1645         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1646         nf_ct_put(ct);
1647
1648         return 0;
1649 }
1650
1651 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1652                                    const struct nfnl_info *info,
1653                                    const struct nlattr * const cda[])
1654 {
1655         u_int8_t u3 = info->nfmsg->nfgen_family;
1656         struct nf_conntrack_tuple_hash *h;
1657         struct nf_conntrack_tuple tuple;
1658         struct nf_conntrack_zone zone;
1659         struct sk_buff *skb2;
1660         struct nf_conn *ct;
1661         int err;
1662
1663         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1664                 struct netlink_dump_control c = {
1665                         .start = ctnetlink_start,
1666                         .dump = ctnetlink_dump_table,
1667                         .done = ctnetlink_done,
1668                         .data = (void *)cda,
1669                 };
1670
1671                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1672         }
1673
1674         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1675         if (err < 0)
1676                 return err;
1677
1678         if (cda[CTA_TUPLE_ORIG])
1679                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1680                                             u3, &zone);
1681         else if (cda[CTA_TUPLE_REPLY])
1682                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1683                                             u3, &zone);
1684         else
1685                 return -EINVAL;
1686
1687         if (err < 0)
1688                 return err;
1689
1690         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1691         if (!h)
1692                 return -ENOENT;
1693
1694         ct = nf_ct_tuplehash_to_ctrack(h);
1695
1696         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1697         if (!skb2) {
1698                 nf_ct_put(ct);
1699                 return -ENOMEM;
1700         }
1701
1702         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1703                                   info->nlh->nlmsg_seq,
1704                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1705                                   true, 0);
1706         nf_ct_put(ct);
1707         if (err <= 0) {
1708                 kfree_skb(skb2);
1709                 return -ENOMEM;
1710         }
1711
1712         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1713 }
1714
1715 static int ctnetlink_done_list(struct netlink_callback *cb)
1716 {
1717         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1718
1719         if (ctx->last)
1720                 nf_ct_put(ctx->last);
1721
1722         return 0;
1723 }
1724
1725 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1726 static int ctnetlink_dump_one_entry(struct sk_buff *skb,
1727                                     struct netlink_callback *cb,
1728                                     struct nf_conn *ct,
1729                                     bool dying)
1730 {
1731         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1732         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1733         u8 l3proto = nfmsg->nfgen_family;
1734         int res;
1735
1736         if (l3proto && nf_ct_l3num(ct) != l3proto)
1737                 return 0;
1738
1739         if (ctx->last) {
1740                 if (ct != ctx->last)
1741                         return 0;
1742
1743                 ctx->last = NULL;
1744         }
1745
1746         /* We can't dump extension info for the unconfirmed
1747          * list because unconfirmed conntracks can have
1748          * ct->ext reallocated (and thus freed).
1749          *
1750          * In the dying list case ct->ext can't be free'd
1751          * until after we drop pcpu->lock.
1752          */
1753         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1754                                   cb->nlh->nlmsg_seq,
1755                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1756                                   ct, dying, 0);
1757         if (res < 0) {
1758                 if (!refcount_inc_not_zero(&ct->ct_general.use))
1759                         return 0;
1760
1761                 ctx->last = ct;
1762         }
1763
1764         return res;
1765 }
1766 #endif
1767
1768 static int
1769 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1770 {
1771         return 0;
1772 }
1773
1774 static int
1775 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1776 {
1777         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1778         struct nf_conn *last = ctx->last;
1779 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1780         const struct net *net = sock_net(skb->sk);
1781         struct nf_conntrack_net_ecache *ecache_net;
1782         struct nf_conntrack_tuple_hash *h;
1783         struct hlist_nulls_node *n;
1784 #endif
1785
1786         if (ctx->done)
1787                 return 0;
1788
1789         ctx->last = NULL;
1790
1791 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1792         ecache_net = nf_conn_pernet_ecache(net);
1793         spin_lock_bh(&ecache_net->dying_lock);
1794
1795         hlist_nulls_for_each_entry(h, n, &ecache_net->dying_list, hnnode) {
1796                 struct nf_conn *ct;
1797                 int res;
1798
1799                 ct = nf_ct_tuplehash_to_ctrack(h);
1800                 if (last && last != ct)
1801                         continue;
1802
1803                 res = ctnetlink_dump_one_entry(skb, cb, ct, true);
1804                 if (res < 0) {
1805                         spin_unlock_bh(&ecache_net->dying_lock);
1806                         nf_ct_put(last);
1807                         return skb->len;
1808                 }
1809
1810                 nf_ct_put(last);
1811                 last = NULL;
1812         }
1813
1814         spin_unlock_bh(&ecache_net->dying_lock);
1815 #endif
1816         ctx->done = true;
1817         nf_ct_put(last);
1818
1819         return skb->len;
1820 }
1821
1822 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1823                                   const struct nfnl_info *info,
1824                                   const struct nlattr * const cda[])
1825 {
1826         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1827                 struct netlink_dump_control c = {
1828                         .dump = ctnetlink_dump_dying,
1829                         .done = ctnetlink_done_list,
1830                 };
1831                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1832         }
1833
1834         return -EOPNOTSUPP;
1835 }
1836
1837 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1838                                         const struct nfnl_info *info,
1839                                         const struct nlattr * const cda[])
1840 {
1841         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1842                 struct netlink_dump_control c = {
1843                         .dump = ctnetlink_dump_unconfirmed,
1844                         .done = ctnetlink_done_list,
1845                 };
1846                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1847         }
1848
1849         return -EOPNOTSUPP;
1850 }
1851
1852 #if IS_ENABLED(CONFIG_NF_NAT)
1853 static int
1854 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1855                           enum nf_nat_manip_type manip,
1856                           const struct nlattr *attr)
1857         __must_hold(RCU)
1858 {
1859         const struct nf_nat_hook *nat_hook;
1860         int err;
1861
1862         nat_hook = rcu_dereference(nf_nat_hook);
1863         if (!nat_hook) {
1864 #ifdef CONFIG_MODULES
1865                 rcu_read_unlock();
1866                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1867                 if (request_module("nf-nat") < 0) {
1868                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1869                         rcu_read_lock();
1870                         return -EOPNOTSUPP;
1871                 }
1872                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1873                 rcu_read_lock();
1874                 nat_hook = rcu_dereference(nf_nat_hook);
1875                 if (nat_hook)
1876                         return -EAGAIN;
1877 #endif
1878                 return -EOPNOTSUPP;
1879         }
1880
1881         err = nat_hook->parse_nat_setup(ct, manip, attr);
1882         if (err == -EAGAIN) {
1883 #ifdef CONFIG_MODULES
1884                 rcu_read_unlock();
1885                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1886                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1887                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1888                         rcu_read_lock();
1889                         return -EOPNOTSUPP;
1890                 }
1891                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1892                 rcu_read_lock();
1893 #else
1894                 err = -EOPNOTSUPP;
1895 #endif
1896         }
1897         return err;
1898 }
1899 #endif
1900
1901 static int
1902 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1903 {
1904         return nf_ct_change_status_common(ct, ntohl(nla_get_be32(cda[CTA_STATUS])));
1905 }
1906
1907 static int
1908 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1909 {
1910 #if IS_ENABLED(CONFIG_NF_NAT)
1911         int ret;
1912
1913         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1914                 return 0;
1915
1916         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1917                                         cda[CTA_NAT_DST]);
1918         if (ret < 0)
1919                 return ret;
1920
1921         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1922                                          cda[CTA_NAT_SRC]);
1923 #else
1924         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1925                 return 0;
1926         return -EOPNOTSUPP;
1927 #endif
1928 }
1929
1930 static int ctnetlink_change_helper(struct nf_conn *ct,
1931                                    const struct nlattr * const cda[])
1932 {
1933         struct nf_conntrack_helper *helper;
1934         struct nf_conn_help *help = nfct_help(ct);
1935         char *helpname = NULL;
1936         struct nlattr *helpinfo = NULL;
1937         int err;
1938
1939         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1940         if (err < 0)
1941                 return err;
1942
1943         /* don't change helper of sibling connections */
1944         if (ct->master) {
1945                 /* If we try to change the helper to the same thing twice,
1946                  * treat the second attempt as a no-op instead of returning
1947                  * an error.
1948                  */
1949                 err = -EBUSY;
1950                 if (help) {
1951                         rcu_read_lock();
1952                         helper = rcu_dereference(help->helper);
1953                         if (helper && !strcmp(helper->name, helpname))
1954                                 err = 0;
1955                         rcu_read_unlock();
1956                 }
1957
1958                 return err;
1959         }
1960
1961         if (!strcmp(helpname, "")) {
1962                 if (help && help->helper) {
1963                         /* we had a helper before ... */
1964                         nf_ct_remove_expectations(ct);
1965                         RCU_INIT_POINTER(help->helper, NULL);
1966                 }
1967
1968                 return 0;
1969         }
1970
1971         rcu_read_lock();
1972         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1973                                             nf_ct_protonum(ct));
1974         if (helper == NULL) {
1975                 rcu_read_unlock();
1976                 return -EOPNOTSUPP;
1977         }
1978
1979         if (help) {
1980                 if (rcu_access_pointer(help->helper) == helper) {
1981                         /* update private helper data if allowed. */
1982                         if (helper->from_nlattr)
1983                                 helper->from_nlattr(helpinfo, ct);
1984                         err = 0;
1985                 } else
1986                         err = -EBUSY;
1987         } else {
1988                 /* we cannot set a helper for an existing conntrack */
1989                 err = -EOPNOTSUPP;
1990         }
1991
1992         rcu_read_unlock();
1993         return err;
1994 }
1995
1996 static int ctnetlink_change_timeout(struct nf_conn *ct,
1997                                     const struct nlattr * const cda[])
1998 {
1999         return __nf_ct_change_timeout(ct, (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ);
2000 }
2001
2002 #if defined(CONFIG_NF_CONNTRACK_MARK)
2003 static void ctnetlink_change_mark(struct nf_conn *ct,
2004                                     const struct nlattr * const cda[])
2005 {
2006         u32 mark, newmark, mask = 0;
2007
2008         if (cda[CTA_MARK_MASK])
2009                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2010
2011         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2012         newmark = (READ_ONCE(ct->mark) & mask) ^ mark;
2013         if (newmark != READ_ONCE(ct->mark))
2014                 WRITE_ONCE(ct->mark, newmark);
2015 }
2016 #endif
2017
2018 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2019         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2020         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2021         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2022 };
2023
2024 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2025                                       const struct nlattr * const cda[])
2026 {
2027         const struct nlattr *attr = cda[CTA_PROTOINFO];
2028         const struct nf_conntrack_l4proto *l4proto;
2029         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2030         int err = 0;
2031
2032         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2033                                           protoinfo_policy, NULL);
2034         if (err < 0)
2035                 return err;
2036
2037         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2038         if (l4proto->from_nlattr)
2039                 err = l4proto->from_nlattr(tb, ct);
2040
2041         return err;
2042 }
2043
2044 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2045         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2046         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2047         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2048 };
2049
2050 static int change_seq_adj(struct nf_ct_seqadj *seq,
2051                           const struct nlattr * const attr)
2052 {
2053         int err;
2054         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2055
2056         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2057                                           seqadj_policy, NULL);
2058         if (err < 0)
2059                 return err;
2060
2061         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2062                 return -EINVAL;
2063
2064         seq->correction_pos =
2065                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2066
2067         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2068                 return -EINVAL;
2069
2070         seq->offset_before =
2071                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2072
2073         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2074                 return -EINVAL;
2075
2076         seq->offset_after =
2077                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2078
2079         return 0;
2080 }
2081
2082 static int
2083 ctnetlink_change_seq_adj(struct nf_conn *ct,
2084                          const struct nlattr * const cda[])
2085 {
2086         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2087         int ret = 0;
2088
2089         if (!seqadj)
2090                 return 0;
2091
2092         spin_lock_bh(&ct->lock);
2093         if (cda[CTA_SEQ_ADJ_ORIG]) {
2094                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2095                                      cda[CTA_SEQ_ADJ_ORIG]);
2096                 if (ret < 0)
2097                         goto err;
2098
2099                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2100         }
2101
2102         if (cda[CTA_SEQ_ADJ_REPLY]) {
2103                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2104                                      cda[CTA_SEQ_ADJ_REPLY]);
2105                 if (ret < 0)
2106                         goto err;
2107
2108                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2109         }
2110
2111         spin_unlock_bh(&ct->lock);
2112         return 0;
2113 err:
2114         spin_unlock_bh(&ct->lock);
2115         return ret;
2116 }
2117
2118 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2119         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2120         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2121         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2122 };
2123
2124 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2125                                      const struct nlattr * const cda[])
2126 {
2127         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2128         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2129         int err;
2130
2131         if (!synproxy)
2132                 return 0;
2133
2134         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2135                                           cda[CTA_SYNPROXY], synproxy_policy,
2136                                           NULL);
2137         if (err < 0)
2138                 return err;
2139
2140         if (!tb[CTA_SYNPROXY_ISN] ||
2141             !tb[CTA_SYNPROXY_ITS] ||
2142             !tb[CTA_SYNPROXY_TSOFF])
2143                 return -EINVAL;
2144
2145         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2146         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2147         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2148
2149         return 0;
2150 }
2151
2152 static int
2153 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2154 {
2155 #ifdef CONFIG_NF_CONNTRACK_LABELS
2156         size_t len = nla_len(cda[CTA_LABELS]);
2157         const void *mask = cda[CTA_LABELS_MASK];
2158
2159         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2160                 return -EINVAL;
2161
2162         if (mask) {
2163                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2164                     nla_len(cda[CTA_LABELS_MASK]) != len)
2165                         return -EINVAL;
2166                 mask = nla_data(cda[CTA_LABELS_MASK]);
2167         }
2168
2169         len /= sizeof(u32);
2170
2171         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2172 #else
2173         return -EOPNOTSUPP;
2174 #endif
2175 }
2176
2177 static int
2178 ctnetlink_change_conntrack(struct nf_conn *ct,
2179                            const struct nlattr * const cda[])
2180 {
2181         int err;
2182
2183         /* only allow NAT changes and master assignation for new conntracks */
2184         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2185                 return -EOPNOTSUPP;
2186
2187         if (cda[CTA_HELP]) {
2188                 err = ctnetlink_change_helper(ct, cda);
2189                 if (err < 0)
2190                         return err;
2191         }
2192
2193         if (cda[CTA_TIMEOUT]) {
2194                 err = ctnetlink_change_timeout(ct, cda);
2195                 if (err < 0)
2196                         return err;
2197         }
2198
2199         if (cda[CTA_STATUS]) {
2200                 err = ctnetlink_change_status(ct, cda);
2201                 if (err < 0)
2202                         return err;
2203         }
2204
2205         if (cda[CTA_PROTOINFO]) {
2206                 err = ctnetlink_change_protoinfo(ct, cda);
2207                 if (err < 0)
2208                         return err;
2209         }
2210
2211 #if defined(CONFIG_NF_CONNTRACK_MARK)
2212         if (cda[CTA_MARK])
2213                 ctnetlink_change_mark(ct, cda);
2214 #endif
2215
2216         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2217                 err = ctnetlink_change_seq_adj(ct, cda);
2218                 if (err < 0)
2219                         return err;
2220         }
2221
2222         if (cda[CTA_SYNPROXY]) {
2223                 err = ctnetlink_change_synproxy(ct, cda);
2224                 if (err < 0)
2225                         return err;
2226         }
2227
2228         if (cda[CTA_LABELS]) {
2229                 err = ctnetlink_attach_labels(ct, cda);
2230                 if (err < 0)
2231                         return err;
2232         }
2233
2234         return 0;
2235 }
2236
2237 static struct nf_conn *
2238 ctnetlink_create_conntrack(struct net *net,
2239                            const struct nf_conntrack_zone *zone,
2240                            const struct nlattr * const cda[],
2241                            struct nf_conntrack_tuple *otuple,
2242                            struct nf_conntrack_tuple *rtuple,
2243                            u8 u3)
2244 {
2245         struct nf_conn *ct;
2246         int err = -EINVAL;
2247         struct nf_conntrack_helper *helper;
2248         struct nf_conn_tstamp *tstamp;
2249         u64 timeout;
2250
2251         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2252         if (IS_ERR(ct))
2253                 return ERR_PTR(-ENOMEM);
2254
2255         if (!cda[CTA_TIMEOUT])
2256                 goto err1;
2257
2258         rcu_read_lock();
2259         if (cda[CTA_HELP]) {
2260                 char *helpname = NULL;
2261                 struct nlattr *helpinfo = NULL;
2262
2263                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2264                 if (err < 0)
2265                         goto err2;
2266
2267                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2268                                                     nf_ct_protonum(ct));
2269                 if (helper == NULL) {
2270                         rcu_read_unlock();
2271 #ifdef CONFIG_MODULES
2272                         if (request_module("nfct-helper-%s", helpname) < 0) {
2273                                 err = -EOPNOTSUPP;
2274                                 goto err1;
2275                         }
2276
2277                         rcu_read_lock();
2278                         helper = __nf_conntrack_helper_find(helpname,
2279                                                             nf_ct_l3num(ct),
2280                                                             nf_ct_protonum(ct));
2281                         if (helper) {
2282                                 err = -EAGAIN;
2283                                 goto err2;
2284                         }
2285                         rcu_read_unlock();
2286 #endif
2287                         err = -EOPNOTSUPP;
2288                         goto err1;
2289                 } else {
2290                         struct nf_conn_help *help;
2291
2292                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2293                         if (help == NULL) {
2294                                 err = -ENOMEM;
2295                                 goto err2;
2296                         }
2297                         /* set private helper data if allowed. */
2298                         if (helper->from_nlattr)
2299                                 helper->from_nlattr(helpinfo, ct);
2300
2301                         /* disable helper auto-assignment for this entry */
2302                         ct->status |= IPS_HELPER;
2303                         RCU_INIT_POINTER(help->helper, helper);
2304                 }
2305         }
2306
2307         err = ctnetlink_setup_nat(ct, cda);
2308         if (err < 0)
2309                 goto err2;
2310
2311         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2312         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2313         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2314         nf_ct_labels_ext_add(ct);
2315         nfct_seqadj_ext_add(ct);
2316         nfct_synproxy_ext_add(ct);
2317
2318         /* we must add conntrack extensions before confirmation. */
2319         ct->status |= IPS_CONFIRMED;
2320
2321         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2322         __nf_ct_set_timeout(ct, timeout);
2323
2324         if (cda[CTA_STATUS]) {
2325                 err = ctnetlink_change_status(ct, cda);
2326                 if (err < 0)
2327                         goto err2;
2328         }
2329
2330         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2331                 err = ctnetlink_change_seq_adj(ct, cda);
2332                 if (err < 0)
2333                         goto err2;
2334         }
2335
2336         memset(&ct->proto, 0, sizeof(ct->proto));
2337         if (cda[CTA_PROTOINFO]) {
2338                 err = ctnetlink_change_protoinfo(ct, cda);
2339                 if (err < 0)
2340                         goto err2;
2341         }
2342
2343         if (cda[CTA_SYNPROXY]) {
2344                 err = ctnetlink_change_synproxy(ct, cda);
2345                 if (err < 0)
2346                         goto err2;
2347         }
2348
2349 #if defined(CONFIG_NF_CONNTRACK_MARK)
2350         if (cda[CTA_MARK])
2351                 ctnetlink_change_mark(ct, cda);
2352 #endif
2353
2354         /* setup master conntrack: this is a confirmed expectation */
2355         if (cda[CTA_TUPLE_MASTER]) {
2356                 struct nf_conntrack_tuple master;
2357                 struct nf_conntrack_tuple_hash *master_h;
2358                 struct nf_conn *master_ct;
2359
2360                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2361                                             u3, NULL);
2362                 if (err < 0)
2363                         goto err2;
2364
2365                 master_h = nf_conntrack_find_get(net, zone, &master);
2366                 if (master_h == NULL) {
2367                         err = -ENOENT;
2368                         goto err2;
2369                 }
2370                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2371                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2372                 ct->master = master_ct;
2373         }
2374         tstamp = nf_conn_tstamp_find(ct);
2375         if (tstamp)
2376                 tstamp->start = ktime_get_real_ns();
2377
2378         err = nf_conntrack_hash_check_insert(ct);
2379         if (err < 0)
2380                 goto err3;
2381
2382         rcu_read_unlock();
2383
2384         return ct;
2385
2386 err3:
2387         if (ct->master)
2388                 nf_ct_put(ct->master);
2389 err2:
2390         rcu_read_unlock();
2391 err1:
2392         nf_conntrack_free(ct);
2393         return ERR_PTR(err);
2394 }
2395
2396 static int ctnetlink_new_conntrack(struct sk_buff *skb,
2397                                    const struct nfnl_info *info,
2398                                    const struct nlattr * const cda[])
2399 {
2400         struct nf_conntrack_tuple otuple, rtuple;
2401         struct nf_conntrack_tuple_hash *h = NULL;
2402         u_int8_t u3 = info->nfmsg->nfgen_family;
2403         struct nf_conntrack_zone zone;
2404         struct nf_conn *ct;
2405         int err;
2406
2407         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2408         if (err < 0)
2409                 return err;
2410
2411         if (cda[CTA_TUPLE_ORIG]) {
2412                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2413                                             u3, &zone);
2414                 if (err < 0)
2415                         return err;
2416         }
2417
2418         if (cda[CTA_TUPLE_REPLY]) {
2419                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2420                                             u3, &zone);
2421                 if (err < 0)
2422                         return err;
2423         }
2424
2425         if (cda[CTA_TUPLE_ORIG])
2426                 h = nf_conntrack_find_get(info->net, &zone, &otuple);
2427         else if (cda[CTA_TUPLE_REPLY])
2428                 h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2429
2430         if (h == NULL) {
2431                 err = -ENOENT;
2432                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2433                         enum ip_conntrack_events events;
2434
2435                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2436                                 return -EINVAL;
2437                         if (otuple.dst.protonum != rtuple.dst.protonum)
2438                                 return -EINVAL;
2439
2440                         ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2441                                                         &otuple, &rtuple, u3);
2442                         if (IS_ERR(ct))
2443                                 return PTR_ERR(ct);
2444
2445                         err = 0;
2446                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2447                                 events = 1 << IPCT_RELATED;
2448                         else
2449                                 events = 1 << IPCT_NEW;
2450
2451                         if (cda[CTA_LABELS] &&
2452                             ctnetlink_attach_labels(ct, cda) == 0)
2453                                 events |= (1 << IPCT_LABEL);
2454
2455                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2456                                                       (1 << IPCT_ASSURED) |
2457                                                       (1 << IPCT_HELPER) |
2458                                                       (1 << IPCT_PROTOINFO) |
2459                                                       (1 << IPCT_SEQADJ) |
2460                                                       (1 << IPCT_MARK) |
2461                                                       (1 << IPCT_SYNPROXY) |
2462                                                       events,
2463                                                       ct, NETLINK_CB(skb).portid,
2464                                                       nlmsg_report(info->nlh));
2465                         nf_ct_put(ct);
2466                 }
2467
2468                 return err;
2469         }
2470         /* implicit 'else' */
2471
2472         err = -EEXIST;
2473         ct = nf_ct_tuplehash_to_ctrack(h);
2474         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2475                 err = ctnetlink_change_conntrack(ct, cda);
2476                 if (err == 0) {
2477                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2478                                                       (1 << IPCT_ASSURED) |
2479                                                       (1 << IPCT_HELPER) |
2480                                                       (1 << IPCT_LABEL) |
2481                                                       (1 << IPCT_PROTOINFO) |
2482                                                       (1 << IPCT_SEQADJ) |
2483                                                       (1 << IPCT_MARK) |
2484                                                       (1 << IPCT_SYNPROXY),
2485                                                       ct, NETLINK_CB(skb).portid,
2486                                                       nlmsg_report(info->nlh));
2487                 }
2488         }
2489
2490         nf_ct_put(ct);
2491         return err;
2492 }
2493
2494 static int
2495 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2496                                 __u16 cpu, const struct ip_conntrack_stat *st)
2497 {
2498         struct nlmsghdr *nlh;
2499         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2500
2501         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2502                               IPCTNL_MSG_CT_GET_STATS_CPU);
2503         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2504                            NFNETLINK_V0, htons(cpu));
2505         if (!nlh)
2506                 goto nlmsg_failure;
2507
2508         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2509             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2510             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2511             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2512                                 htonl(st->insert_failed)) ||
2513             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2514             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2515             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2516             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2517                                 htonl(st->search_restart)) ||
2518             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2519                                 htonl(st->clash_resolve)) ||
2520             nla_put_be32(skb, CTA_STATS_CHAIN_TOOLONG,
2521                          htonl(st->chaintoolong)))
2522                 goto nla_put_failure;
2523
2524         nlmsg_end(skb, nlh);
2525         return skb->len;
2526
2527 nla_put_failure:
2528 nlmsg_failure:
2529         nlmsg_cancel(skb, nlh);
2530         return -1;
2531 }
2532
2533 static int
2534 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2535 {
2536         int cpu;
2537         struct net *net = sock_net(skb->sk);
2538
2539         if (cb->args[0] == nr_cpu_ids)
2540                 return 0;
2541
2542         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2543                 const struct ip_conntrack_stat *st;
2544
2545                 if (!cpu_possible(cpu))
2546                         continue;
2547
2548                 st = per_cpu_ptr(net->ct.stat, cpu);
2549                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2550                                                     NETLINK_CB(cb->skb).portid,
2551                                                     cb->nlh->nlmsg_seq,
2552                                                     cpu, st) < 0)
2553                                 break;
2554         }
2555         cb->args[0] = cpu;
2556
2557         return skb->len;
2558 }
2559
2560 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2561                                  const struct nfnl_info *info,
2562                                  const struct nlattr * const cda[])
2563 {
2564         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2565                 struct netlink_dump_control c = {
2566                         .dump = ctnetlink_ct_stat_cpu_dump,
2567                 };
2568                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2569         }
2570
2571         return 0;
2572 }
2573
2574 static int
2575 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2576                             struct net *net)
2577 {
2578         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2579         unsigned int nr_conntracks;
2580         struct nlmsghdr *nlh;
2581
2582         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2583         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2584                            NFNETLINK_V0, 0);
2585         if (!nlh)
2586                 goto nlmsg_failure;
2587
2588         nr_conntracks = nf_conntrack_count(net);
2589         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2590                 goto nla_put_failure;
2591
2592         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2593                 goto nla_put_failure;
2594
2595         nlmsg_end(skb, nlh);
2596         return skb->len;
2597
2598 nla_put_failure:
2599 nlmsg_failure:
2600         nlmsg_cancel(skb, nlh);
2601         return -1;
2602 }
2603
2604 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2605                              const struct nlattr * const cda[])
2606 {
2607         struct sk_buff *skb2;
2608         int err;
2609
2610         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2611         if (skb2 == NULL)
2612                 return -ENOMEM;
2613
2614         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2615                                           info->nlh->nlmsg_seq,
2616                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2617                                           sock_net(skb->sk));
2618         if (err <= 0) {
2619                 kfree_skb(skb2);
2620                 return -ENOMEM;
2621         }
2622
2623         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2624 }
2625
2626 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2627         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2628         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2629         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2630         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2631         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2632         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2633                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2634         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2635         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2636         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2637         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2638         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2639 };
2640
2641 static struct nf_conntrack_expect *
2642 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2643                        struct nf_conntrack_helper *helper,
2644                        struct nf_conntrack_tuple *tuple,
2645                        struct nf_conntrack_tuple *mask);
2646
2647 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2648 static size_t
2649 ctnetlink_glue_build_size(const struct nf_conn *ct)
2650 {
2651         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2652                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2653                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2654                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2655                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2656                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2657                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2658                + nla_total_size(0) /* CTA_PROTOINFO */
2659                + nla_total_size(0) /* CTA_HELP */
2660                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2661                + ctnetlink_secctx_size(ct)
2662                + ctnetlink_acct_size(ct)
2663                + ctnetlink_timestamp_size(ct)
2664 #if IS_ENABLED(CONFIG_NF_NAT)
2665                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2666                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2667 #endif
2668 #ifdef CONFIG_NF_CONNTRACK_MARK
2669                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2670 #endif
2671 #ifdef CONFIG_NF_CONNTRACK_ZONES
2672                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2673 #endif
2674                + ctnetlink_proto_size(ct)
2675                ;
2676 }
2677
2678 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2679 {
2680         const struct nf_conntrack_zone *zone;
2681         struct nlattr *nest_parms;
2682
2683         zone = nf_ct_zone(ct);
2684
2685         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2686         if (!nest_parms)
2687                 goto nla_put_failure;
2688         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2689                 goto nla_put_failure;
2690         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2691                                    NF_CT_ZONE_DIR_ORIG) < 0)
2692                 goto nla_put_failure;
2693         nla_nest_end(skb, nest_parms);
2694
2695         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2696         if (!nest_parms)
2697                 goto nla_put_failure;
2698         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2699                 goto nla_put_failure;
2700         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2701                                    NF_CT_ZONE_DIR_REPL) < 0)
2702                 goto nla_put_failure;
2703         nla_nest_end(skb, nest_parms);
2704
2705         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2706                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2707                 goto nla_put_failure;
2708
2709         if (ctnetlink_dump_id(skb, ct) < 0)
2710                 goto nla_put_failure;
2711
2712         if (ctnetlink_dump_status(skb, ct) < 0)
2713                 goto nla_put_failure;
2714
2715         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2716                 goto nla_put_failure;
2717
2718         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2719                 goto nla_put_failure;
2720
2721         if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2722             ctnetlink_dump_timestamp(skb, ct) < 0)
2723                 goto nla_put_failure;
2724
2725         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2726                 goto nla_put_failure;
2727
2728 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2729         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2730                 goto nla_put_failure;
2731 #endif
2732         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2733                 goto nla_put_failure;
2734
2735         if ((ct->status & IPS_SEQ_ADJUST) &&
2736             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2737                 goto nla_put_failure;
2738
2739         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2740                 goto nla_put_failure;
2741
2742 #ifdef CONFIG_NF_CONNTRACK_MARK
2743         if (ctnetlink_dump_mark(skb, ct, true) < 0)
2744                 goto nla_put_failure;
2745 #endif
2746         if (ctnetlink_dump_labels(skb, ct) < 0)
2747                 goto nla_put_failure;
2748         return 0;
2749
2750 nla_put_failure:
2751         return -ENOSPC;
2752 }
2753
2754 static int
2755 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2756                      enum ip_conntrack_info ctinfo,
2757                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2758 {
2759         struct nlattr *nest_parms;
2760
2761         nest_parms = nla_nest_start(skb, ct_attr);
2762         if (!nest_parms)
2763                 goto nla_put_failure;
2764
2765         if (__ctnetlink_glue_build(skb, ct) < 0)
2766                 goto nla_put_failure;
2767
2768         nla_nest_end(skb, nest_parms);
2769
2770         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2771                 goto nla_put_failure;
2772
2773         return 0;
2774
2775 nla_put_failure:
2776         return -ENOSPC;
2777 }
2778
2779 static int
2780 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2781 {
2782         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2783         unsigned long d = ct->status ^ status;
2784
2785         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2786                 /* SEEN_REPLY bit can only be set */
2787                 return -EBUSY;
2788
2789         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2790                 /* ASSURED bit can only be set */
2791                 return -EBUSY;
2792
2793         /* This check is less strict than ctnetlink_change_status()
2794          * because callers often flip IPS_EXPECTED bits when sending
2795          * an NFQA_CT attribute to the kernel.  So ignore the
2796          * unchangeable bits but do not error out. Also user programs
2797          * are allowed to clear the bits that they are allowed to change.
2798          */
2799         __nf_ct_change_status(ct, status, ~status);
2800         return 0;
2801 }
2802
2803 static int
2804 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2805 {
2806         int err;
2807
2808         if (cda[CTA_TIMEOUT]) {
2809                 err = ctnetlink_change_timeout(ct, cda);
2810                 if (err < 0)
2811                         return err;
2812         }
2813         if (cda[CTA_STATUS]) {
2814                 err = ctnetlink_update_status(ct, cda);
2815                 if (err < 0)
2816                         return err;
2817         }
2818         if (cda[CTA_HELP]) {
2819                 err = ctnetlink_change_helper(ct, cda);
2820                 if (err < 0)
2821                         return err;
2822         }
2823         if (cda[CTA_LABELS]) {
2824                 err = ctnetlink_attach_labels(ct, cda);
2825                 if (err < 0)
2826                         return err;
2827         }
2828 #if defined(CONFIG_NF_CONNTRACK_MARK)
2829         if (cda[CTA_MARK]) {
2830                 ctnetlink_change_mark(ct, cda);
2831         }
2832 #endif
2833         return 0;
2834 }
2835
2836 static int
2837 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2838 {
2839         struct nlattr *cda[CTA_MAX+1];
2840         int ret;
2841
2842         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2843                                           NULL);
2844         if (ret < 0)
2845                 return ret;
2846
2847         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2848 }
2849
2850 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2851                                     const struct nf_conn *ct,
2852                                     struct nf_conntrack_tuple *tuple,
2853                                     struct nf_conntrack_tuple *mask)
2854 {
2855         int err;
2856
2857         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2858                                     nf_ct_l3num(ct), NULL);
2859         if (err < 0)
2860                 return err;
2861
2862         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2863                                      nf_ct_l3num(ct), NULL);
2864 }
2865
2866 static int
2867 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2868                              u32 portid, u32 report)
2869 {
2870         struct nlattr *cda[CTA_EXPECT_MAX+1];
2871         struct nf_conntrack_tuple tuple, mask;
2872         struct nf_conntrack_helper *helper = NULL;
2873         struct nf_conntrack_expect *exp;
2874         int err;
2875
2876         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2877                                           exp_nla_policy, NULL);
2878         if (err < 0)
2879                 return err;
2880
2881         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2882                                        ct, &tuple, &mask);
2883         if (err < 0)
2884                 return err;
2885
2886         if (cda[CTA_EXPECT_HELP_NAME]) {
2887                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2888
2889                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2890                                                     nf_ct_protonum(ct));
2891                 if (helper == NULL)
2892                         return -EOPNOTSUPP;
2893         }
2894
2895         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2896                                      helper, &tuple, &mask);
2897         if (IS_ERR(exp))
2898                 return PTR_ERR(exp);
2899
2900         err = nf_ct_expect_related_report(exp, portid, report, 0);
2901         nf_ct_expect_put(exp);
2902         return err;
2903 }
2904
2905 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2906                                   enum ip_conntrack_info ctinfo, int diff)
2907 {
2908         if (!(ct->status & IPS_NAT_MASK))
2909                 return;
2910
2911         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2912 }
2913
2914 static const struct nfnl_ct_hook ctnetlink_glue_hook = {
2915         .build_size     = ctnetlink_glue_build_size,
2916         .build          = ctnetlink_glue_build,
2917         .parse          = ctnetlink_glue_parse,
2918         .attach_expect  = ctnetlink_glue_attach_expect,
2919         .seq_adjust     = ctnetlink_glue_seqadj,
2920 };
2921 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2922
2923 /***********************************************************************
2924  * EXPECT
2925  ***********************************************************************/
2926
2927 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2928                                     const struct nf_conntrack_tuple *tuple,
2929                                     u32 type)
2930 {
2931         struct nlattr *nest_parms;
2932
2933         nest_parms = nla_nest_start(skb, type);
2934         if (!nest_parms)
2935                 goto nla_put_failure;
2936         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2937                 goto nla_put_failure;
2938         nla_nest_end(skb, nest_parms);
2939
2940         return 0;
2941
2942 nla_put_failure:
2943         return -1;
2944 }
2945
2946 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2947                                    const struct nf_conntrack_tuple *tuple,
2948                                    const struct nf_conntrack_tuple_mask *mask)
2949 {
2950         const struct nf_conntrack_l4proto *l4proto;
2951         struct nf_conntrack_tuple m;
2952         struct nlattr *nest_parms;
2953         int ret;
2954
2955         memset(&m, 0xFF, sizeof(m));
2956         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2957         m.src.u.all = mask->src.u.all;
2958         m.src.l3num = tuple->src.l3num;
2959         m.dst.protonum = tuple->dst.protonum;
2960
2961         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2962         if (!nest_parms)
2963                 goto nla_put_failure;
2964
2965         rcu_read_lock();
2966         ret = ctnetlink_dump_tuples_ip(skb, &m);
2967         if (ret >= 0) {
2968                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2969                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2970         }
2971         rcu_read_unlock();
2972
2973         if (unlikely(ret < 0))
2974                 goto nla_put_failure;
2975
2976         nla_nest_end(skb, nest_parms);
2977
2978         return 0;
2979
2980 nla_put_failure:
2981         return -1;
2982 }
2983
2984 #if IS_ENABLED(CONFIG_NF_NAT)
2985 static const union nf_inet_addr any_addr;
2986 #endif
2987
2988 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2989 {
2990         static siphash_aligned_key_t exp_id_seed;
2991         unsigned long a, b, c, d;
2992
2993         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2994
2995         a = (unsigned long)exp;
2996         b = (unsigned long)exp->helper;
2997         c = (unsigned long)exp->master;
2998         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2999
3000 #ifdef CONFIG_64BIT
3001         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
3002 #else
3003         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
3004 #endif
3005 }
3006
3007 static int
3008 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3009                           const struct nf_conntrack_expect *exp)
3010 {
3011         struct nf_conn *master = exp->master;
3012         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3013         struct nf_conn_help *help;
3014 #if IS_ENABLED(CONFIG_NF_NAT)
3015         struct nlattr *nest_parms;
3016         struct nf_conntrack_tuple nat_tuple = {};
3017 #endif
3018         struct nf_ct_helper_expectfn *expfn;
3019
3020         if (timeout < 0)
3021                 timeout = 0;
3022
3023         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3024                 goto nla_put_failure;
3025         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3026                 goto nla_put_failure;
3027         if (ctnetlink_exp_dump_tuple(skb,
3028                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3029                                  CTA_EXPECT_MASTER) < 0)
3030                 goto nla_put_failure;
3031
3032 #if IS_ENABLED(CONFIG_NF_NAT)
3033         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3034             exp->saved_proto.all) {
3035                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3036                 if (!nest_parms)
3037                         goto nla_put_failure;
3038
3039                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3040                         goto nla_put_failure;
3041
3042                 nat_tuple.src.l3num = nf_ct_l3num(master);
3043                 nat_tuple.src.u3 = exp->saved_addr;
3044                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3045                 nat_tuple.src.u = exp->saved_proto;
3046
3047                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3048                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3049                         goto nla_put_failure;
3050                 nla_nest_end(skb, nest_parms);
3051         }
3052 #endif
3053         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3054             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3055             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3056             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3057                 goto nla_put_failure;
3058         help = nfct_help(master);
3059         if (help) {
3060                 struct nf_conntrack_helper *helper;
3061
3062                 helper = rcu_dereference(help->helper);
3063                 if (helper &&
3064                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3065                         goto nla_put_failure;
3066         }
3067         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3068         if (expfn != NULL &&
3069             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3070                 goto nla_put_failure;
3071
3072         return 0;
3073
3074 nla_put_failure:
3075         return -1;
3076 }
3077
3078 static int
3079 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3080                         int event, const struct nf_conntrack_expect *exp)
3081 {
3082         struct nlmsghdr *nlh;
3083         unsigned int flags = portid ? NLM_F_MULTI : 0;
3084
3085         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3086         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3087                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3088         if (!nlh)
3089                 goto nlmsg_failure;
3090
3091         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3092                 goto nla_put_failure;
3093
3094         nlmsg_end(skb, nlh);
3095         return skb->len;
3096
3097 nlmsg_failure:
3098 nla_put_failure:
3099         nlmsg_cancel(skb, nlh);
3100         return -1;
3101 }
3102
3103 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3104 static int
3105 ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3106 {
3107         struct nf_conntrack_expect *exp = item->exp;
3108         struct net *net = nf_ct_exp_net(exp);
3109         struct nlmsghdr *nlh;
3110         struct sk_buff *skb;
3111         unsigned int type, group;
3112         int flags = 0;
3113
3114         if (events & (1 << IPEXP_DESTROY)) {
3115                 type = IPCTNL_MSG_EXP_DELETE;
3116                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3117         } else if (events & (1 << IPEXP_NEW)) {
3118                 type = IPCTNL_MSG_EXP_NEW;
3119                 flags = NLM_F_CREATE|NLM_F_EXCL;
3120                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3121         } else
3122                 return 0;
3123
3124         if (!item->report && !nfnetlink_has_listeners(net, group))
3125                 return 0;
3126
3127         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3128         if (skb == NULL)
3129                 goto errout;
3130
3131         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3132         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3133                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3134         if (!nlh)
3135                 goto nlmsg_failure;
3136
3137         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3138                 goto nla_put_failure;
3139
3140         nlmsg_end(skb, nlh);
3141         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3142         return 0;
3143
3144 nla_put_failure:
3145         nlmsg_cancel(skb, nlh);
3146 nlmsg_failure:
3147         kfree_skb(skb);
3148 errout:
3149         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3150         return 0;
3151 }
3152 #endif
3153 static int ctnetlink_exp_done(struct netlink_callback *cb)
3154 {
3155         if (cb->args[1])
3156                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3157         return 0;
3158 }
3159
3160 static int
3161 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3162 {
3163         struct net *net = sock_net(skb->sk);
3164         struct nf_conntrack_expect *exp, *last;
3165         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3166         u_int8_t l3proto = nfmsg->nfgen_family;
3167
3168         rcu_read_lock();
3169         last = (struct nf_conntrack_expect *)cb->args[1];
3170         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3171 restart:
3172                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3173                                          hnode) {
3174                         if (l3proto && exp->tuple.src.l3num != l3proto)
3175                                 continue;
3176
3177                         if (!net_eq(nf_ct_net(exp->master), net))
3178                                 continue;
3179
3180                         if (cb->args[1]) {
3181                                 if (exp != last)
3182                                         continue;
3183                                 cb->args[1] = 0;
3184                         }
3185                         if (ctnetlink_exp_fill_info(skb,
3186                                                     NETLINK_CB(cb->skb).portid,
3187                                                     cb->nlh->nlmsg_seq,
3188                                                     IPCTNL_MSG_EXP_NEW,
3189                                                     exp) < 0) {
3190                                 if (!refcount_inc_not_zero(&exp->use))
3191                                         continue;
3192                                 cb->args[1] = (unsigned long)exp;
3193                                 goto out;
3194                         }
3195                 }
3196                 if (cb->args[1]) {
3197                         cb->args[1] = 0;
3198                         goto restart;
3199                 }
3200         }
3201 out:
3202         rcu_read_unlock();
3203         if (last)
3204                 nf_ct_expect_put(last);
3205
3206         return skb->len;
3207 }
3208
3209 static int
3210 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3211 {
3212         struct nf_conntrack_expect *exp, *last;
3213         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3214         struct nf_conn *ct = cb->data;
3215         struct nf_conn_help *help = nfct_help(ct);
3216         u_int8_t l3proto = nfmsg->nfgen_family;
3217
3218         if (cb->args[0])
3219                 return 0;
3220
3221         rcu_read_lock();
3222         last = (struct nf_conntrack_expect *)cb->args[1];
3223 restart:
3224         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3225                 if (l3proto && exp->tuple.src.l3num != l3proto)
3226                         continue;
3227                 if (cb->args[1]) {
3228                         if (exp != last)
3229                                 continue;
3230                         cb->args[1] = 0;
3231                 }
3232                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3233                                             cb->nlh->nlmsg_seq,
3234                                             IPCTNL_MSG_EXP_NEW,
3235                                             exp) < 0) {
3236                         if (!refcount_inc_not_zero(&exp->use))
3237                                 continue;
3238                         cb->args[1] = (unsigned long)exp;
3239                         goto out;
3240                 }
3241         }
3242         if (cb->args[1]) {
3243                 cb->args[1] = 0;
3244                 goto restart;
3245         }
3246         cb->args[0] = 1;
3247 out:
3248         rcu_read_unlock();
3249         if (last)
3250                 nf_ct_expect_put(last);
3251
3252         return skb->len;
3253 }
3254
3255 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3256                                  struct sk_buff *skb,
3257                                  const struct nlmsghdr *nlh,
3258                                  const struct nlattr * const cda[],
3259                                  struct netlink_ext_ack *extack)
3260 {
3261         int err;
3262         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3263         u_int8_t u3 = nfmsg->nfgen_family;
3264         struct nf_conntrack_tuple tuple;
3265         struct nf_conntrack_tuple_hash *h;
3266         struct nf_conn *ct;
3267         struct nf_conntrack_zone zone;
3268         struct netlink_dump_control c = {
3269                 .dump = ctnetlink_exp_ct_dump_table,
3270                 .done = ctnetlink_exp_done,
3271         };
3272
3273         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3274                                     u3, NULL);
3275         if (err < 0)
3276                 return err;
3277
3278         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3279         if (err < 0)
3280                 return err;
3281
3282         h = nf_conntrack_find_get(net, &zone, &tuple);
3283         if (!h)
3284                 return -ENOENT;
3285
3286         ct = nf_ct_tuplehash_to_ctrack(h);
3287         /* No expectation linked to this connection tracking. */
3288         if (!nfct_help(ct)) {
3289                 nf_ct_put(ct);
3290                 return 0;
3291         }
3292
3293         c.data = ct;
3294
3295         err = netlink_dump_start(ctnl, skb, nlh, &c);
3296         nf_ct_put(ct);
3297
3298         return err;
3299 }
3300
3301 static int ctnetlink_get_expect(struct sk_buff *skb,
3302                                 const struct nfnl_info *info,
3303                                 const struct nlattr * const cda[])
3304 {
3305         u_int8_t u3 = info->nfmsg->nfgen_family;
3306         struct nf_conntrack_tuple tuple;
3307         struct nf_conntrack_expect *exp;
3308         struct nf_conntrack_zone zone;
3309         struct sk_buff *skb2;
3310         int err;
3311
3312         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3313                 if (cda[CTA_EXPECT_MASTER])
3314                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3315                                                      info->nlh, cda,
3316                                                      info->extack);
3317                 else {
3318                         struct netlink_dump_control c = {
3319                                 .dump = ctnetlink_exp_dump_table,
3320                                 .done = ctnetlink_exp_done,
3321                         };
3322                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3323                 }
3324         }
3325
3326         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3327         if (err < 0)
3328                 return err;
3329
3330         if (cda[CTA_EXPECT_TUPLE])
3331                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3332                                             u3, NULL);
3333         else if (cda[CTA_EXPECT_MASTER])
3334                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3335                                             u3, NULL);
3336         else
3337                 return -EINVAL;
3338
3339         if (err < 0)
3340                 return err;
3341
3342         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3343         if (!exp)
3344                 return -ENOENT;
3345
3346         if (cda[CTA_EXPECT_ID]) {
3347                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3348
3349                 if (id != nf_expect_get_id(exp)) {
3350                         nf_ct_expect_put(exp);
3351                         return -ENOENT;
3352                 }
3353         }
3354
3355         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3356         if (!skb2) {
3357                 nf_ct_expect_put(exp);
3358                 return -ENOMEM;
3359         }
3360
3361         rcu_read_lock();
3362         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3363                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3364                                       exp);
3365         rcu_read_unlock();
3366         nf_ct_expect_put(exp);
3367         if (err <= 0) {
3368                 kfree_skb(skb2);
3369                 return -ENOMEM;
3370         }
3371
3372         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3373 }
3374
3375 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3376 {
3377         struct nf_conntrack_helper *helper;
3378         const struct nf_conn_help *m_help;
3379         const char *name = data;
3380
3381         m_help = nfct_help(exp->master);
3382
3383         helper = rcu_dereference(m_help->helper);
3384         if (!helper)
3385                 return false;
3386
3387         return strcmp(helper->name, name) == 0;
3388 }
3389
3390 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3391 {
3392         return true;
3393 }
3394
3395 static int ctnetlink_del_expect(struct sk_buff *skb,
3396                                 const struct nfnl_info *info,
3397                                 const struct nlattr * const cda[])
3398 {
3399         u_int8_t u3 = info->nfmsg->nfgen_family;
3400         struct nf_conntrack_expect *exp;
3401         struct nf_conntrack_tuple tuple;
3402         struct nf_conntrack_zone zone;
3403         int err;
3404
3405         if (cda[CTA_EXPECT_TUPLE]) {
3406                 /* delete a single expect by tuple */
3407                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3408                 if (err < 0)
3409                         return err;
3410
3411                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3412                                             u3, NULL);
3413                 if (err < 0)
3414                         return err;
3415
3416                 /* bump usage count to 2 */
3417                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3418                 if (!exp)
3419                         return -ENOENT;
3420
3421                 if (cda[CTA_EXPECT_ID]) {
3422                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3423                         if (ntohl(id) != (u32)(unsigned long)exp) {
3424                                 nf_ct_expect_put(exp);
3425                                 return -ENOENT;
3426                         }
3427                 }
3428
3429                 /* after list removal, usage count == 1 */
3430                 spin_lock_bh(&nf_conntrack_expect_lock);
3431                 if (del_timer(&exp->timeout)) {
3432                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3433                                                    nlmsg_report(info->nlh));
3434                         nf_ct_expect_put(exp);
3435                 }
3436                 spin_unlock_bh(&nf_conntrack_expect_lock);
3437                 /* have to put what we 'get' above.
3438                  * after this line usage count == 0 */
3439                 nf_ct_expect_put(exp);
3440         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3441                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3442
3443                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3444                                          NETLINK_CB(skb).portid,
3445                                          nlmsg_report(info->nlh));
3446         } else {
3447                 /* This basically means we have to flush everything*/
3448                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3449                                          NETLINK_CB(skb).portid,
3450                                          nlmsg_report(info->nlh));
3451         }
3452
3453         return 0;
3454 }
3455 static int
3456 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3457                         const struct nlattr * const cda[])
3458 {
3459         if (cda[CTA_EXPECT_TIMEOUT]) {
3460                 if (!del_timer(&x->timeout))
3461                         return -ETIME;
3462
3463                 x->timeout.expires = jiffies +
3464                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3465                 add_timer(&x->timeout);
3466         }
3467         return 0;
3468 }
3469
3470 #if IS_ENABLED(CONFIG_NF_NAT)
3471 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3472         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3473         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3474 };
3475 #endif
3476
3477 static int
3478 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3479                            struct nf_conntrack_expect *exp,
3480                            u_int8_t u3)
3481 {
3482 #if IS_ENABLED(CONFIG_NF_NAT)
3483         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3484         struct nf_conntrack_tuple nat_tuple = {};
3485         int err;
3486
3487         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3488                                           exp_nat_nla_policy, NULL);
3489         if (err < 0)
3490                 return err;
3491
3492         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3493                 return -EINVAL;
3494
3495         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3496                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3497                                     u3, NULL);
3498         if (err < 0)
3499                 return err;
3500
3501         exp->saved_addr = nat_tuple.src.u3;
3502         exp->saved_proto = nat_tuple.src.u;
3503         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3504
3505         return 0;
3506 #else
3507         return -EOPNOTSUPP;
3508 #endif
3509 }
3510
3511 static struct nf_conntrack_expect *
3512 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3513                        struct nf_conntrack_helper *helper,
3514                        struct nf_conntrack_tuple *tuple,
3515                        struct nf_conntrack_tuple *mask)
3516 {
3517         u_int32_t class = 0;
3518         struct nf_conntrack_expect *exp;
3519         struct nf_conn_help *help;
3520         int err;
3521
3522         help = nfct_help(ct);
3523         if (!help)
3524                 return ERR_PTR(-EOPNOTSUPP);
3525
3526         if (cda[CTA_EXPECT_CLASS] && helper) {
3527                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3528                 if (class > helper->expect_class_max)
3529                         return ERR_PTR(-EINVAL);
3530         }
3531         exp = nf_ct_expect_alloc(ct);
3532         if (!exp)
3533                 return ERR_PTR(-ENOMEM);
3534
3535         if (cda[CTA_EXPECT_FLAGS]) {
3536                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3537                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3538         } else {
3539                 exp->flags = 0;
3540         }
3541         if (cda[CTA_EXPECT_FN]) {
3542                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3543                 struct nf_ct_helper_expectfn *expfn;
3544
3545                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3546                 if (expfn == NULL) {
3547                         err = -EINVAL;
3548                         goto err_out;
3549                 }
3550                 exp->expectfn = expfn->expectfn;
3551         } else
3552                 exp->expectfn = NULL;
3553
3554         exp->class = class;
3555         exp->master = ct;
3556         exp->helper = helper;
3557         exp->tuple = *tuple;
3558         exp->mask.src.u3 = mask->src.u3;
3559         exp->mask.src.u.all = mask->src.u.all;
3560
3561         if (cda[CTA_EXPECT_NAT]) {
3562                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3563                                                  exp, nf_ct_l3num(ct));
3564                 if (err < 0)
3565                         goto err_out;
3566         }
3567         return exp;
3568 err_out:
3569         nf_ct_expect_put(exp);
3570         return ERR_PTR(err);
3571 }
3572
3573 static int
3574 ctnetlink_create_expect(struct net *net,
3575                         const struct nf_conntrack_zone *zone,
3576                         const struct nlattr * const cda[],
3577                         u_int8_t u3, u32 portid, int report)
3578 {
3579         struct nf_conntrack_tuple tuple, mask, master_tuple;
3580         struct nf_conntrack_tuple_hash *h = NULL;
3581         struct nf_conntrack_helper *helper = NULL;
3582         struct nf_conntrack_expect *exp;
3583         struct nf_conn *ct;
3584         int err;
3585
3586         /* caller guarantees that those three CTA_EXPECT_* exist */
3587         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3588                                     u3, NULL);
3589         if (err < 0)
3590                 return err;
3591         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3592                                     u3, NULL);
3593         if (err < 0)
3594                 return err;
3595         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3596                                     u3, NULL);
3597         if (err < 0)
3598                 return err;
3599
3600         /* Look for master conntrack of this expectation */
3601         h = nf_conntrack_find_get(net, zone, &master_tuple);
3602         if (!h)
3603                 return -ENOENT;
3604         ct = nf_ct_tuplehash_to_ctrack(h);
3605
3606         rcu_read_lock();
3607         if (cda[CTA_EXPECT_HELP_NAME]) {
3608                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3609
3610                 helper = __nf_conntrack_helper_find(helpname, u3,
3611                                                     nf_ct_protonum(ct));
3612                 if (helper == NULL) {
3613                         rcu_read_unlock();
3614 #ifdef CONFIG_MODULES
3615                         if (request_module("nfct-helper-%s", helpname) < 0) {
3616                                 err = -EOPNOTSUPP;
3617                                 goto err_ct;
3618                         }
3619                         rcu_read_lock();
3620                         helper = __nf_conntrack_helper_find(helpname, u3,
3621                                                             nf_ct_protonum(ct));
3622                         if (helper) {
3623                                 err = -EAGAIN;
3624                                 goto err_rcu;
3625                         }
3626                         rcu_read_unlock();
3627 #endif
3628                         err = -EOPNOTSUPP;
3629                         goto err_ct;
3630                 }
3631         }
3632
3633         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3634         if (IS_ERR(exp)) {
3635                 err = PTR_ERR(exp);
3636                 goto err_rcu;
3637         }
3638
3639         err = nf_ct_expect_related_report(exp, portid, report, 0);
3640         nf_ct_expect_put(exp);
3641 err_rcu:
3642         rcu_read_unlock();
3643 err_ct:
3644         nf_ct_put(ct);
3645         return err;
3646 }
3647
3648 static int ctnetlink_new_expect(struct sk_buff *skb,
3649                                 const struct nfnl_info *info,
3650                                 const struct nlattr * const cda[])
3651 {
3652         u_int8_t u3 = info->nfmsg->nfgen_family;
3653         struct nf_conntrack_tuple tuple;
3654         struct nf_conntrack_expect *exp;
3655         struct nf_conntrack_zone zone;
3656         int err;
3657
3658         if (!cda[CTA_EXPECT_TUPLE]
3659             || !cda[CTA_EXPECT_MASK]
3660             || !cda[CTA_EXPECT_MASTER])
3661                 return -EINVAL;
3662
3663         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3664         if (err < 0)
3665                 return err;
3666
3667         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3668                                     u3, NULL);
3669         if (err < 0)
3670                 return err;
3671
3672         spin_lock_bh(&nf_conntrack_expect_lock);
3673         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3674         if (!exp) {
3675                 spin_unlock_bh(&nf_conntrack_expect_lock);
3676                 err = -ENOENT;
3677                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3678                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3679                                                       NETLINK_CB(skb).portid,
3680                                                       nlmsg_report(info->nlh));
3681                 }
3682                 return err;
3683         }
3684
3685         err = -EEXIST;
3686         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3687                 err = ctnetlink_change_expect(exp, cda);
3688         spin_unlock_bh(&nf_conntrack_expect_lock);
3689
3690         return err;
3691 }
3692
3693 static int
3694 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3695                              const struct ip_conntrack_stat *st)
3696 {
3697         struct nlmsghdr *nlh;
3698         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3699
3700         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3701                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3702         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3703                            NFNETLINK_V0, htons(cpu));
3704         if (!nlh)
3705                 goto nlmsg_failure;
3706
3707         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3708             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3709             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3710                 goto nla_put_failure;
3711
3712         nlmsg_end(skb, nlh);
3713         return skb->len;
3714
3715 nla_put_failure:
3716 nlmsg_failure:
3717         nlmsg_cancel(skb, nlh);
3718         return -1;
3719 }
3720
3721 static int
3722 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3723 {
3724         int cpu;
3725         struct net *net = sock_net(skb->sk);
3726
3727         if (cb->args[0] == nr_cpu_ids)
3728                 return 0;
3729
3730         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3731                 const struct ip_conntrack_stat *st;
3732
3733                 if (!cpu_possible(cpu))
3734                         continue;
3735
3736                 st = per_cpu_ptr(net->ct.stat, cpu);
3737                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3738                                                  cb->nlh->nlmsg_seq,
3739                                                  cpu, st) < 0)
3740                         break;
3741         }
3742         cb->args[0] = cpu;
3743
3744         return skb->len;
3745 }
3746
3747 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3748                                   const struct nfnl_info *info,
3749                                   const struct nlattr * const cda[])
3750 {
3751         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3752                 struct netlink_dump_control c = {
3753                         .dump = ctnetlink_exp_stat_cpu_dump,
3754                 };
3755                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3756         }
3757
3758         return 0;
3759 }
3760
3761 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3762 static struct nf_ct_event_notifier ctnl_notifier = {
3763         .ct_event = ctnetlink_conntrack_event,
3764         .exp_event = ctnetlink_expect_event,
3765 };
3766 #endif
3767
3768 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3769         [IPCTNL_MSG_CT_NEW]     = {
3770                 .call           = ctnetlink_new_conntrack,
3771                 .type           = NFNL_CB_MUTEX,
3772                 .attr_count     = CTA_MAX,
3773                 .policy         = ct_nla_policy
3774         },
3775         [IPCTNL_MSG_CT_GET]     = {
3776                 .call           = ctnetlink_get_conntrack,
3777                 .type           = NFNL_CB_MUTEX,
3778                 .attr_count     = CTA_MAX,
3779                 .policy         = ct_nla_policy
3780         },
3781         [IPCTNL_MSG_CT_DELETE]  = {
3782                 .call           = ctnetlink_del_conntrack,
3783                 .type           = NFNL_CB_MUTEX,
3784                 .attr_count     = CTA_MAX,
3785                 .policy         = ct_nla_policy
3786         },
3787         [IPCTNL_MSG_CT_GET_CTRZERO] = {
3788                 .call           = ctnetlink_get_conntrack,
3789                 .type           = NFNL_CB_MUTEX,
3790                 .attr_count     = CTA_MAX,
3791                 .policy         = ct_nla_policy
3792         },
3793         [IPCTNL_MSG_CT_GET_STATS_CPU] = {
3794                 .call           = ctnetlink_stat_ct_cpu,
3795                 .type           = NFNL_CB_MUTEX,
3796         },
3797         [IPCTNL_MSG_CT_GET_STATS] = {
3798                 .call           = ctnetlink_stat_ct,
3799                 .type           = NFNL_CB_MUTEX,
3800         },
3801         [IPCTNL_MSG_CT_GET_DYING] = {
3802                 .call           = ctnetlink_get_ct_dying,
3803                 .type           = NFNL_CB_MUTEX,
3804         },
3805         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = {
3806                 .call           = ctnetlink_get_ct_unconfirmed,
3807                 .type           = NFNL_CB_MUTEX,
3808         },
3809 };
3810
3811 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3812         [IPCTNL_MSG_EXP_GET] = {
3813                 .call           = ctnetlink_get_expect,
3814                 .type           = NFNL_CB_MUTEX,
3815                 .attr_count     = CTA_EXPECT_MAX,
3816                 .policy         = exp_nla_policy
3817         },
3818         [IPCTNL_MSG_EXP_NEW] = {
3819                 .call           = ctnetlink_new_expect,
3820                 .type           = NFNL_CB_MUTEX,
3821                 .attr_count     = CTA_EXPECT_MAX,
3822                 .policy         = exp_nla_policy
3823         },
3824         [IPCTNL_MSG_EXP_DELETE] = {
3825                 .call           = ctnetlink_del_expect,
3826                 .type           = NFNL_CB_MUTEX,
3827                 .attr_count     = CTA_EXPECT_MAX,
3828                 .policy         = exp_nla_policy
3829         },
3830         [IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3831                 .call           = ctnetlink_stat_exp_cpu,
3832                 .type           = NFNL_CB_MUTEX,
3833         },
3834 };
3835
3836 static const struct nfnetlink_subsystem ctnl_subsys = {
3837         .name                           = "conntrack",
3838         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3839         .cb_count                       = IPCTNL_MSG_MAX,
3840         .cb                             = ctnl_cb,
3841 };
3842
3843 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3844         .name                           = "conntrack_expect",
3845         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3846         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3847         .cb                             = ctnl_exp_cb,
3848 };
3849
3850 MODULE_ALIAS("ip_conntrack_netlink");
3851 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3852 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3853
3854 static int __net_init ctnetlink_net_init(struct net *net)
3855 {
3856 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3857         nf_conntrack_register_notifier(net, &ctnl_notifier);
3858 #endif
3859         return 0;
3860 }
3861
3862 static void ctnetlink_net_pre_exit(struct net *net)
3863 {
3864 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3865         nf_conntrack_unregister_notifier(net);
3866 #endif
3867 }
3868
3869 static struct pernet_operations ctnetlink_net_ops = {
3870         .init           = ctnetlink_net_init,
3871         .pre_exit       = ctnetlink_net_pre_exit,
3872 };
3873
3874 static int __init ctnetlink_init(void)
3875 {
3876         int ret;
3877
3878         NL_ASSERT_DUMP_CTX_FITS(struct ctnetlink_list_dump_ctx);
3879
3880         ret = nfnetlink_subsys_register(&ctnl_subsys);
3881         if (ret < 0) {
3882                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3883                 goto err_out;
3884         }
3885
3886         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3887         if (ret < 0) {
3888                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3889                 goto err_unreg_subsys;
3890         }
3891
3892         ret = register_pernet_subsys(&ctnetlink_net_ops);
3893         if (ret < 0) {
3894                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3895                 goto err_unreg_exp_subsys;
3896         }
3897 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3898         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3899         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3900 #endif
3901         return 0;
3902
3903 err_unreg_exp_subsys:
3904         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3905 err_unreg_subsys:
3906         nfnetlink_subsys_unregister(&ctnl_subsys);
3907 err_out:
3908         return ret;
3909 }
3910
3911 static void __exit ctnetlink_exit(void)
3912 {
3913         unregister_pernet_subsys(&ctnetlink_net_ops);
3914         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3915         nfnetlink_subsys_unregister(&ctnl_subsys);
3916 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3917         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3918 #endif
3919         synchronize_rcu();
3920 }
3921
3922 module_init(ctnetlink_init);
3923 module_exit(ctnetlink_exit);