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