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