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