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, const struct nf_conn *ct)
 332 {
 333         if (nla_put_be32(skb, CTA_MARK, htonl(ct->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, ct) < 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         int err;
 726
 727         if (events & (1 << IPCT_DESTROY)) {
 728                 type = IPCTNL_MSG_CT_DELETE;
 729                 group = NFNLGRP_CONNTRACK_DESTROY;
 730         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 731                 type = IPCTNL_MSG_CT_NEW;
 732                 flags = NLM_F_CREATE|NLM_F_EXCL;
 733                 group = NFNLGRP_CONNTRACK_NEW;
 734         } else if (events) {
 735                 type = IPCTNL_MSG_CT_NEW;
 736                 group = NFNLGRP_CONNTRACK_UPDATE;
 737         } else
 738                 return 0;
 739
 740         net = nf_ct_net(ct);
 741         if (!item->report && !nfnetlink_has_listeners(net, group))
 742                 return 0;
 743
 744         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 745         if (skb == NULL)
 746                 goto errout;
 747
 748         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 749         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags, nf_ct_l3num(ct),
 750                            NFNETLINK_V0, 0);
 751         if (!nlh)
 752                 goto nlmsg_failure;
 753
 754         zone = nf_ct_zone(ct);
 755
 756         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 757         if (!nest_parms)
 758                 goto nla_put_failure;
 759         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 760                 goto nla_put_failure;
 761         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 762                                    NF_CT_ZONE_DIR_ORIG) < 0)
 763                 goto nla_put_failure;
 764         nla_nest_end(skb, nest_parms);
 765
 766         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 767         if (!nest_parms)
 768                 goto nla_put_failure;
 769         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 770                 goto nla_put_failure;
 771         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 772                                    NF_CT_ZONE_DIR_REPL) < 0)
 773                 goto nla_put_failure;
 774         nla_nest_end(skb, nest_parms);
 775
 776         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 777                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 778                 goto nla_put_failure;
 779
 780         if (ctnetlink_dump_id(skb, ct) < 0)
 781                 goto nla_put_failure;
 782
 783         if (ctnetlink_dump_status(skb, ct) < 0)
 784                 goto nla_put_failure;
 785
 786         if (events & (1 << IPCT_DESTROY)) {
 787                 if (ctnetlink_dump_timeout(skb, ct, true) < 0)
 788                         goto nla_put_failure;
 789
 790                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 791                     ctnetlink_dump_timestamp(skb, ct) < 0 ||
 792                     ctnetlink_dump_protoinfo(skb, ct, true) < 0)
 793                         goto nla_put_failure;
 794         } else {
 795                 if (ctnetlink_dump_timeout(skb, ct, false) < 0)
 796                         goto nla_put_failure;
 797
 798                 if (events & (1 << IPCT_PROTOINFO) &&
 799                     ctnetlink_dump_protoinfo(skb, ct, false) < 0)
 800                         goto nla_put_failure;
 801
 802                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 803                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 804                         goto nla_put_failure;
 805
 806 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 807                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 808                     && ctnetlink_dump_secctx(skb, ct) < 0)
 809                         goto nla_put_failure;
 810 #endif
 811                 if (events & (1 << IPCT_LABEL) &&
 812                      ctnetlink_dump_labels(skb, ct) < 0)
 813                         goto nla_put_failure;
 814
 815                 if (events & (1 << IPCT_RELATED) &&
 816                     ctnetlink_dump_master(skb, ct) < 0)
 817                         goto nla_put_failure;
 818
 819                 if (events & (1 << IPCT_SEQADJ) &&
 820                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 821                         goto nla_put_failure;
 822
 823                 if (events & (1 << IPCT_SYNPROXY) &&
 824                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 825                         goto nla_put_failure;
 826         }
 827
 828 #ifdef CONFIG_NF_CONNTRACK_MARK
 829         if ((events & (1 << IPCT_MARK) || ct->mark)
 830             && ctnetlink_dump_mark(skb, ct) < 0)
 831                 goto nla_put_failure;
 832 #endif
 833         nlmsg_end(skb, nlh);
 834         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 835                              GFP_ATOMIC);
 836         if (err == -ENOBUFS || err == -EAGAIN)
 837                 return -ENOBUFS;
 838
 839         return 0;
 840
 841 nla_put_failure:
 842         nlmsg_cancel(skb, nlh);
 843 nlmsg_failure:
 844         kfree_skb(skb);
 845 errout:
 846         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 847                 return -ENOBUFS;
 848
 849         return 0;
 850 }
 851 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 852
 853 static int ctnetlink_done(struct netlink_callback *cb)
 854 {
 855         if (cb->args[1])
 856                 nf_ct_put((struct nf_conn *)cb->args[1]);
 857         kfree(cb->data);
 858         return 0;
 859 }
 860
 861 struct ctnetlink_filter_u32 {
 862         u32 val;
 863         u32 mask;
 864 };
 865
 866 struct ctnetlink_filter {
 867         u8 family;
 868
 869         u_int32_t orig_flags;
 870         u_int32_t reply_flags;
 871
 872         struct nf_conntrack_tuple orig;
 873         struct nf_conntrack_tuple reply;
 874         struct nf_conntrack_zone zone;
 875
 876         struct ctnetlink_filter_u32 mark;
 877         struct ctnetlink_filter_u32 status;
 878 };
 879
 880 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 881         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 882         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 883 };
 884
 885 static int ctnetlink_parse_filter(const struct nlattr *attr,
 886                                   struct ctnetlink_filter *filter)
 887 {
 888         struct nlattr *tb[CTA_FILTER_MAX + 1];
 889         int ret = 0;
 890
 891         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 892                                NULL);
 893         if (ret)
 894                 return ret;
 895
 896         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 897                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 898                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 899                         return -EOPNOTSUPP;
 900         }
 901
 902         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 903                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 904                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 905                         return -EOPNOTSUPP;
 906         }
 907
 908         return 0;
 909 }
 910
 911 static int ctnetlink_parse_zone(const struct nlattr *attr,
 912                                 struct nf_conntrack_zone *zone);
 913 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 914                                          struct nf_conntrack_tuple *tuple,
 915                                          u32 type, u_int8_t l3num,
 916                                          struct nf_conntrack_zone *zone,
 917                                          u_int32_t flags);
 918
 919 static int ctnetlink_filter_parse_mark(struct ctnetlink_filter_u32 *mark,
 920                                        const struct nlattr * const cda[])
 921 {
 922 #ifdef CONFIG_NF_CONNTRACK_MARK
 923         if (cda[CTA_MARK]) {
 924                 mark->val = ntohl(nla_get_be32(cda[CTA_MARK]));
 925
 926                 if (cda[CTA_MARK_MASK])
 927                         mark->mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 928                 else
 929                         mark->mask = 0xffffffff;
 930         } else if (cda[CTA_MARK_MASK]) {
 931                 return -EINVAL;
 932         }
 933 #endif
 934         return 0;
 935 }
 936
 937 static int ctnetlink_filter_parse_status(struct ctnetlink_filter_u32 *status,
 938                                          const struct nlattr * const cda[])
 939 {
 940         if (cda[CTA_STATUS]) {
 941                 status->val = ntohl(nla_get_be32(cda[CTA_STATUS]));
 942                 if (cda[CTA_STATUS_MASK])
 943                         status->mask = ntohl(nla_get_be32(cda[CTA_STATUS_MASK]));
 944                 else
 945                         status->mask = status->val;
 946
 947                 /* status->val == 0? always true, else always false. */
 948                 if (status->mask == 0)
 949                         return -EINVAL;
 950         } else if (cda[CTA_STATUS_MASK]) {
 951                 return -EINVAL;
 952         }
 953
 954         /* CTA_STATUS is NLA_U32, if this fires UAPI needs to be extended */
 955         BUILD_BUG_ON(__IPS_MAX_BIT >= 32);
 956         return 0;
 957 }
 958
 959 static struct ctnetlink_filter *
 960 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 961 {
 962         struct ctnetlink_filter *filter;
 963         int err;
 964
 965 #ifndef CONFIG_NF_CONNTRACK_MARK
 966         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 967                 return ERR_PTR(-EOPNOTSUPP);
 968 #endif
 969
 970         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 971         if (filter == NULL)
 972                 return ERR_PTR(-ENOMEM);
 973
 974         filter->family = family;
 975
 976         err = ctnetlink_filter_parse_mark(&filter->mark, cda);
 977         if (err)
 978                 goto err_filter;
 979
 980         err = ctnetlink_filter_parse_status(&filter->status, cda);
 981         if (err)
 982                 goto err_filter;
 983
 984         if (!cda[CTA_FILTER])
 985                 return filter;
 986
 987         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 988         if (err < 0)
 989                 goto err_filter;
 990
 991         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
 992         if (err < 0)
 993                 goto err_filter;
 994
 995         if (filter->orig_flags) {
 996                 if (!cda[CTA_TUPLE_ORIG]) {
 997                         err = -EINVAL;
 998                         goto err_filter;
 999                 }
1000
1001                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
1002                                                    CTA_TUPLE_ORIG,
1003                                                    filter->family,
1004                                                    &filter->zone,
1005                                                    filter->orig_flags);
1006                 if (err < 0)
1007                         goto err_filter;
1008         }
1009
1010         if (filter->reply_flags) {
1011                 if (!cda[CTA_TUPLE_REPLY]) {
1012                         err = -EINVAL;
1013                         goto err_filter;
1014                 }
1015
1016                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
1017                                                    CTA_TUPLE_REPLY,
1018                                                    filter->family,
1019                                                    &filter->zone,
1020                                                    filter->reply_flags);
1021                 if (err < 0)
1022                         goto err_filter;
1023         }
1024
1025         return filter;
1026
1027 err_filter:
1028         kfree(filter);
1029
1030         return ERR_PTR(err);
1031 }
1032
1033 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1034 {
1035         return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS];
1036 }
1037
1038 static int ctnetlink_start(struct netlink_callback *cb)
1039 {
1040         const struct nlattr * const *cda = cb->data;
1041         struct ctnetlink_filter *filter = NULL;
1042         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1043         u8 family = nfmsg->nfgen_family;
1044
1045         if (ctnetlink_needs_filter(family, cda)) {
1046                 filter = ctnetlink_alloc_filter(cda, family);
1047                 if (IS_ERR(filter))
1048                         return PTR_ERR(filter);
1049         }
1050
1051         cb->data = filter;
1052         return 0;
1053 }
1054
1055 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1056                                         struct nf_conntrack_tuple *ct_tuple,
1057                                         u_int32_t flags, int family)
1058 {
1059         switch (family) {
1060         case NFPROTO_IPV4:
1061                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1062                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1063                         return  0;
1064
1065                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1066                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1067                         return  0;
1068                 break;
1069         case NFPROTO_IPV6:
1070                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1071                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1072                                    &ct_tuple->src.u3.in6))
1073                         return 0;
1074
1075                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1076                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1077                                    &ct_tuple->dst.u3.in6))
1078                         return 0;
1079                 break;
1080         }
1081
1082         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1083             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1084                 return 0;
1085
1086         switch (ct_tuple->dst.protonum) {
1087         case IPPROTO_TCP:
1088         case IPPROTO_UDP:
1089                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1090                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1091                         return 0;
1092
1093                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1094                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1095                         return 0;
1096                 break;
1097         case IPPROTO_ICMP:
1098                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1099                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1100                         return 0;
1101                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1102                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1103                         return 0;
1104                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1105                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1106                         return 0;
1107                 break;
1108         case IPPROTO_ICMPV6:
1109                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1110                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1111                         return 0;
1112                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1113                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1114                         return 0;
1115                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1116                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1117                         return 0;
1118                 break;
1119         }
1120
1121         return 1;
1122 }
1123
1124 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1125 {
1126         struct ctnetlink_filter *filter = data;
1127         struct nf_conntrack_tuple *tuple;
1128         u32 status;
1129
1130         if (filter == NULL)
1131                 goto out;
1132
1133         /* Match entries of a given L3 protocol number.
1134          * If it is not specified, ie. l3proto == 0,
1135          * then match everything.
1136          */
1137         if (filter->family && nf_ct_l3num(ct) != filter->family)
1138                 goto ignore_entry;
1139
1140         if (filter->orig_flags) {
1141                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1142                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1143                                                   filter->orig_flags,
1144                                                   filter->family))
1145                         goto ignore_entry;
1146         }
1147
1148         if (filter->reply_flags) {
1149                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1150                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1151                                                   filter->reply_flags,
1152                                                   filter->family))
1153                         goto ignore_entry;
1154         }
1155
1156 #ifdef CONFIG_NF_CONNTRACK_MARK
1157         if ((ct->mark & filter->mark.mask) != filter->mark.val)
1158                 goto ignore_entry;
1159 #endif
1160         status = (u32)READ_ONCE(ct->status);
1161         if ((status & filter->status.mask) != filter->status.val)
1162                 goto ignore_entry;
1163
1164 out:
1165         return 1;
1166
1167 ignore_entry:
1168         return 0;
1169 }
1170
1171 static int
1172 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1173 {
1174         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1175         struct net *net = sock_net(skb->sk);
1176         struct nf_conn *ct, *last;
1177         struct nf_conntrack_tuple_hash *h;
1178         struct hlist_nulls_node *n;
1179         struct nf_conn *nf_ct_evict[8];
1180         int res, i;
1181         spinlock_t *lockp;
1182
1183         last = (struct nf_conn *)cb->args[1];
1184         i = 0;
1185
1186         local_bh_disable();
1187         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1188 restart:
1189                 while (i) {
1190                         i--;
1191                         if (nf_ct_should_gc(nf_ct_evict[i]))
1192                                 nf_ct_kill(nf_ct_evict[i]);
1193                         nf_ct_put(nf_ct_evict[i]);
1194                 }
1195
1196                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1197                 nf_conntrack_lock(lockp);
1198                 if (cb->args[0] >= nf_conntrack_htable_size) {
1199                         spin_unlock(lockp);
1200                         goto out;
1201                 }
1202                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1203                                            hnnode) {
1204                         ct = nf_ct_tuplehash_to_ctrack(h);
1205                         if (nf_ct_is_expired(ct)) {
1206                                 /* need to defer nf_ct_kill() until lock is released */
1207                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1208                                     refcount_inc_not_zero(&ct->ct_general.use))
1209                                         nf_ct_evict[i++] = ct;
1210                                 continue;
1211                         }
1212
1213                         if (!net_eq(net, nf_ct_net(ct)))
1214                                 continue;
1215
1216                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1217                                 continue;
1218
1219                         if (cb->args[1]) {
1220                                 if (ct != last)
1221                                         continue;
1222                                 cb->args[1] = 0;
1223                         }
1224                         if (!ctnetlink_filter_match(ct, cb->data))
1225                                 continue;
1226
1227                         res =
1228                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1229                                             cb->nlh->nlmsg_seq,
1230                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1231                                             ct, true, flags);
1232                         if (res < 0) {
1233                                 nf_conntrack_get(&ct->ct_general);
1234                                 cb->args[1] = (unsigned long)ct;
1235                                 spin_unlock(lockp);
1236                                 goto out;
1237                         }
1238                 }
1239                 spin_unlock(lockp);
1240                 if (cb->args[1]) {
1241                         cb->args[1] = 0;
1242                         goto restart;
1243                 }
1244         }
1245 out:
1246         local_bh_enable();
1247         if (last) {
1248                 /* nf ct hash resize happened, now clear the leftover. */
1249                 if ((struct nf_conn *)cb->args[1] == last)
1250                         cb->args[1] = 0;
1251
1252                 nf_ct_put(last);
1253         }
1254
1255         while (i) {
1256                 i--;
1257                 if (nf_ct_should_gc(nf_ct_evict[i]))
1258                         nf_ct_kill(nf_ct_evict[i]);
1259                 nf_ct_put(nf_ct_evict[i]);
1260         }
1261
1262         return skb->len;
1263 }
1264
1265 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1266                                 struct nf_conntrack_tuple *t,
1267                                 u_int32_t flags)
1268 {
1269         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1270                 if (!tb[CTA_IP_V4_SRC])
1271                         return -EINVAL;
1272
1273                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1274         }
1275
1276         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1277                 if (!tb[CTA_IP_V4_DST])
1278                         return -EINVAL;
1279
1280                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1281         }
1282
1283         return 0;
1284 }
1285
1286 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1287                                 struct nf_conntrack_tuple *t,
1288                                 u_int32_t flags)
1289 {
1290         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1291                 if (!tb[CTA_IP_V6_SRC])
1292                         return -EINVAL;
1293
1294                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1295         }
1296
1297         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1298                 if (!tb[CTA_IP_V6_DST])
1299                         return -EINVAL;
1300
1301                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1302         }
1303
1304         return 0;
1305 }
1306
1307 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1308                                     struct nf_conntrack_tuple *tuple,
1309                                     u_int32_t flags)
1310 {
1311         struct nlattr *tb[CTA_IP_MAX+1];
1312         int ret = 0;
1313
1314         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1315         if (ret < 0)
1316                 return ret;
1317
1318         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1319                                              cta_ip_nla_policy, NULL);
1320         if (ret)
1321                 return ret;
1322
1323         switch (tuple->src.l3num) {
1324         case NFPROTO_IPV4:
1325                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1326                 break;
1327         case NFPROTO_IPV6:
1328                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1329                 break;
1330         }
1331
1332         return ret;
1333 }
1334
1335 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1336         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1337 };
1338
1339 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1340                                        struct nf_conntrack_tuple *tuple,
1341                                        u_int32_t flags)
1342 {
1343         const struct nf_conntrack_l4proto *l4proto;
1344         struct nlattr *tb[CTA_PROTO_MAX+1];
1345         int ret = 0;
1346
1347         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1348                                           proto_nla_policy, NULL);
1349         if (ret < 0)
1350                 return ret;
1351
1352         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1353                 return 0;
1354
1355         if (!tb[CTA_PROTO_NUM])
1356                 return -EINVAL;
1357
1358         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1359
1360         rcu_read_lock();
1361         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1362
1363         if (likely(l4proto->nlattr_to_tuple)) {
1364                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1365                                                      l4proto->nla_policy,
1366                                                      NULL);
1367                 if (ret == 0)
1368                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1369         }
1370
1371         rcu_read_unlock();
1372
1373         return ret;
1374 }
1375
1376 static int
1377 ctnetlink_parse_zone(const struct nlattr *attr,
1378                      struct nf_conntrack_zone *zone)
1379 {
1380         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1381                         NF_CT_DEFAULT_ZONE_DIR, 0);
1382 #ifdef CONFIG_NF_CONNTRACK_ZONES
1383         if (attr)
1384                 zone->id = ntohs(nla_get_be16(attr));
1385 #else
1386         if (attr)
1387                 return -EOPNOTSUPP;
1388 #endif
1389         return 0;
1390 }
1391
1392 static int
1393 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1394                            struct nf_conntrack_zone *zone)
1395 {
1396         int ret;
1397
1398         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1399                 return -EINVAL;
1400
1401         ret = ctnetlink_parse_zone(attr, zone);
1402         if (ret < 0)
1403                 return ret;
1404
1405         if (type == CTA_TUPLE_REPLY)
1406                 zone->dir = NF_CT_ZONE_DIR_REPL;
1407         else
1408                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1409
1410         return 0;
1411 }
1412
1413 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1414         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1415         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1416         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1417 };
1418
1419 #define CTA_FILTER_F_ALL_CTA_PROTO \
1420   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1421    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1422    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1423    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1424    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1425    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1426    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1427    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1428
1429 static int
1430 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1431                               struct nf_conntrack_tuple *tuple, u32 type,
1432                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1433                               u_int32_t flags)
1434 {
1435         struct nlattr *tb[CTA_TUPLE_MAX+1];
1436         int err;
1437
1438         memset(tuple, 0, sizeof(*tuple));
1439
1440         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1441                                           tuple_nla_policy, NULL);
1442         if (err < 0)
1443                 return err;
1444
1445         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1446                 return -EOPNOTSUPP;
1447         tuple->src.l3num = l3num;
1448
1449         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1450             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1451                 if (!tb[CTA_TUPLE_IP])
1452                         return -EINVAL;
1453
1454                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1455                 if (err < 0)
1456                         return err;
1457         }
1458
1459         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1460                 if (!tb[CTA_TUPLE_PROTO])
1461                         return -EINVAL;
1462
1463                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1464                 if (err < 0)
1465                         return err;
1466         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1467                 /* Can't manage proto flags without a protonum  */
1468                 return -EINVAL;
1469         }
1470
1471         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1472                 if (!zone)
1473                         return -EINVAL;
1474
1475                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1476                                                  type, zone);
1477                 if (err < 0)
1478                         return err;
1479         }
1480
1481         /* orig and expect tuples get DIR_ORIGINAL */
1482         if (type == CTA_TUPLE_REPLY)
1483                 tuple->dst.dir = IP_CT_DIR_REPLY;
1484         else
1485                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1486
1487         return 0;
1488 }
1489
1490 static int
1491 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1492                       struct nf_conntrack_tuple *tuple, u32 type,
1493                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1494 {
1495         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1496                                             CTA_FILTER_FLAG(ALL));
1497 }
1498
1499 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1500         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1501                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1502 };
1503
1504 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1505                                 struct nlattr **helpinfo)
1506 {
1507         int err;
1508         struct nlattr *tb[CTA_HELP_MAX+1];
1509
1510         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1511                                           help_nla_policy, NULL);
1512         if (err < 0)
1513                 return err;
1514
1515         if (!tb[CTA_HELP_NAME])
1516                 return -EINVAL;
1517
1518         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1519
1520         if (tb[CTA_HELP_INFO])
1521                 *helpinfo = tb[CTA_HELP_INFO];
1522
1523         return 0;
1524 }
1525
1526 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1527         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1528         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1529         [CTA_STATUS]            = { .type = NLA_U32 },
1530         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1531         [CTA_HELP]              = { .type = NLA_NESTED },
1532         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1533         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1534         [CTA_MARK]              = { .type = NLA_U32 },
1535         [CTA_ID]                = { .type = NLA_U32 },
1536         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1537         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1538         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1539         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1540         [CTA_ZONE]              = { .type = NLA_U16 },
1541         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1542         [CTA_LABELS]            = { .type = NLA_BINARY,
1543                                     .len = NF_CT_LABELS_MAX_SIZE },
1544         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1545                                     .len = NF_CT_LABELS_MAX_SIZE },
1546         [CTA_FILTER]            = { .type = NLA_NESTED },
1547         [CTA_STATUS_MASK]       = { .type = NLA_U32 },
1548 };
1549
1550 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1551 {
1552         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1553                 return 0;
1554
1555         return ctnetlink_filter_match(ct, data);
1556 }
1557
1558 static int ctnetlink_flush_conntrack(struct net *net,
1559                                      const struct nlattr * const cda[],
1560                                      u32 portid, int report, u8 family)
1561 {
1562         struct ctnetlink_filter *filter = NULL;
1563         struct nf_ct_iter_data iter = {
1564                 .net            = net,
1565                 .portid         = portid,
1566                 .report         = report,
1567         };
1568
1569         if (ctnetlink_needs_filter(family, cda)) {
1570                 if (cda[CTA_FILTER])
1571                         return -EOPNOTSUPP;
1572
1573                 filter = ctnetlink_alloc_filter(cda, family);
1574                 if (IS_ERR(filter))
1575                         return PTR_ERR(filter);
1576
1577                 iter.data = filter;
1578         }
1579
1580         nf_ct_iterate_cleanup_net(ctnetlink_flush_iterate, &iter);
1581         kfree(filter);
1582
1583         return 0;
1584 }
1585
1586 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1587                                    const struct nfnl_info *info,
1588                                    const struct nlattr * const cda[])
1589 {
1590         u8 family = info->nfmsg->nfgen_family;
1591         struct nf_conntrack_tuple_hash *h;
1592         struct nf_conntrack_tuple tuple;
1593         struct nf_conntrack_zone zone;
1594         struct nf_conn *ct;
1595         int err;
1596
1597         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1598         if (err < 0)
1599                 return err;
1600
1601         if (cda[CTA_TUPLE_ORIG])
1602                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1603                                             family, &zone);
1604         else if (cda[CTA_TUPLE_REPLY])
1605                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1606                                             family, &zone);
1607         else {
1608                 u_int8_t u3 = info->nfmsg->version ? family : AF_UNSPEC;
1609
1610                 return ctnetlink_flush_conntrack(info->net, cda,
1611                                                  NETLINK_CB(skb).portid,
1612                                                  nlmsg_report(info->nlh), u3);
1613         }
1614
1615         if (err < 0)
1616                 return err;
1617
1618         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1619         if (!h)
1620                 return -ENOENT;
1621
1622         ct = nf_ct_tuplehash_to_ctrack(h);
1623
1624         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1625                 nf_ct_put(ct);
1626                 return -EBUSY;
1627         }
1628
1629         if (cda[CTA_ID]) {
1630                 __be32 id = nla_get_be32(cda[CTA_ID]);
1631
1632                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1633                         nf_ct_put(ct);
1634                         return -ENOENT;
1635                 }
1636         }
1637
1638         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1639         nf_ct_put(ct);
1640
1641         return 0;
1642 }
1643
1644 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1645                                    const struct nfnl_info *info,
1646                                    const struct nlattr * const cda[])
1647 {
1648         u_int8_t u3 = info->nfmsg->nfgen_family;
1649         struct nf_conntrack_tuple_hash *h;
1650         struct nf_conntrack_tuple tuple;
1651         struct nf_conntrack_zone zone;
1652         struct sk_buff *skb2;
1653         struct nf_conn *ct;
1654         int err;
1655
1656         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1657                 struct netlink_dump_control c = {
1658                         .start = ctnetlink_start,
1659                         .dump = ctnetlink_dump_table,
1660                         .done = ctnetlink_done,
1661                         .data = (void *)cda,
1662                 };
1663
1664                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1665         }
1666
1667         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1668         if (err < 0)
1669                 return err;
1670
1671         if (cda[CTA_TUPLE_ORIG])
1672                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1673                                             u3, &zone);
1674         else if (cda[CTA_TUPLE_REPLY])
1675                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1676                                             u3, &zone);
1677         else
1678                 return -EINVAL;
1679
1680         if (err < 0)
1681                 return err;
1682
1683         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1684         if (!h)
1685                 return -ENOENT;
1686
1687         ct = nf_ct_tuplehash_to_ctrack(h);
1688
1689         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1690         if (!skb2) {
1691                 nf_ct_put(ct);
1692                 return -ENOMEM;
1693         }
1694
1695         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1696                                   info->nlh->nlmsg_seq,
1697                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1698                                   true, 0);
1699         nf_ct_put(ct);
1700         if (err <= 0) {
1701                 kfree_skb(skb2);
1702                 return -ENOMEM;
1703         }
1704
1705         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1706 }
1707
1708 static int ctnetlink_done_list(struct netlink_callback *cb)
1709 {
1710         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1711
1712         if (ctx->last)
1713                 nf_ct_put(ctx->last);
1714
1715         return 0;
1716 }
1717
1718 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1719 static int ctnetlink_dump_one_entry(struct sk_buff *skb,
1720                                     struct netlink_callback *cb,
1721                                     struct nf_conn *ct,
1722                                     bool dying)
1723 {
1724         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1725         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1726         u8 l3proto = nfmsg->nfgen_family;
1727         int res;
1728
1729         if (l3proto && nf_ct_l3num(ct) != l3proto)
1730                 return 0;
1731
1732         if (ctx->last) {
1733                 if (ct != ctx->last)
1734                         return 0;
1735
1736                 ctx->last = NULL;
1737         }
1738
1739         /* We can't dump extension info for the unconfirmed
1740          * list because unconfirmed conntracks can have
1741          * ct->ext reallocated (and thus freed).
1742          *
1743          * In the dying list case ct->ext can't be free'd
1744          * until after we drop pcpu->lock.
1745          */
1746         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1747                                   cb->nlh->nlmsg_seq,
1748                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1749                                   ct, dying, 0);
1750         if (res < 0) {
1751                 if (!refcount_inc_not_zero(&ct->ct_general.use))
1752                         return 0;
1753
1754                 ctx->last = ct;
1755         }
1756
1757         return res;
1758 }
1759 #endif
1760
1761 static int
1762 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1763 {
1764         return 0;
1765 }
1766
1767 static int
1768 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1769 {
1770         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1771         struct nf_conn *last = ctx->last;
1772 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1773         const struct net *net = sock_net(skb->sk);
1774         struct nf_conntrack_net_ecache *ecache_net;
1775         struct nf_conntrack_tuple_hash *h;
1776         struct hlist_nulls_node *n;
1777 #endif
1778
1779         if (ctx->done)
1780                 return 0;
1781
1782         ctx->last = NULL;
1783
1784 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1785         ecache_net = nf_conn_pernet_ecache(net);
1786         spin_lock_bh(&ecache_net->dying_lock);
1787
1788         hlist_nulls_for_each_entry(h, n, &ecache_net->dying_list, hnnode) {
1789                 struct nf_conn *ct;
1790                 int res;
1791
1792                 ct = nf_ct_tuplehash_to_ctrack(h);
1793                 if (last && last != ct)
1794                         continue;
1795
1796                 res = ctnetlink_dump_one_entry(skb, cb, ct, true);
1797                 if (res < 0) {
1798                         spin_unlock_bh(&ecache_net->dying_lock);
1799                         nf_ct_put(last);
1800                         return skb->len;
1801                 }
1802
1803                 nf_ct_put(last);
1804                 last = NULL;
1805         }
1806
1807         spin_unlock_bh(&ecache_net->dying_lock);
1808 #endif
1809         ctx->done = true;
1810         nf_ct_put(last);
1811
1812         return skb->len;
1813 }
1814
1815 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1816                                   const struct nfnl_info *info,
1817                                   const struct nlattr * const cda[])
1818 {
1819         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1820                 struct netlink_dump_control c = {
1821                         .dump = ctnetlink_dump_dying,
1822                         .done = ctnetlink_done_list,
1823                 };
1824                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1825         }
1826
1827         return -EOPNOTSUPP;
1828 }
1829
1830 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1831                                         const struct nfnl_info *info,
1832                                         const struct nlattr * const cda[])
1833 {
1834         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1835                 struct netlink_dump_control c = {
1836                         .dump = ctnetlink_dump_unconfirmed,
1837                         .done = ctnetlink_done_list,
1838                 };
1839                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1840         }
1841
1842         return -EOPNOTSUPP;
1843 }
1844
1845 #if IS_ENABLED(CONFIG_NF_NAT)
1846 static int
1847 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1848                           enum nf_nat_manip_type manip,
1849                           const struct nlattr *attr)
1850         __must_hold(RCU)
1851 {
1852         const struct nf_nat_hook *nat_hook;
1853         int err;
1854
1855         nat_hook = rcu_dereference(nf_nat_hook);
1856         if (!nat_hook) {
1857 #ifdef CONFIG_MODULES
1858                 rcu_read_unlock();
1859                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1860                 if (request_module("nf-nat") < 0) {
1861                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1862                         rcu_read_lock();
1863                         return -EOPNOTSUPP;
1864                 }
1865                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1866                 rcu_read_lock();
1867                 nat_hook = rcu_dereference(nf_nat_hook);
1868                 if (nat_hook)
1869                         return -EAGAIN;
1870 #endif
1871                 return -EOPNOTSUPP;
1872         }
1873
1874         err = nat_hook->parse_nat_setup(ct, manip, attr);
1875         if (err == -EAGAIN) {
1876 #ifdef CONFIG_MODULES
1877                 rcu_read_unlock();
1878                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1879                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1880                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1881                         rcu_read_lock();
1882                         return -EOPNOTSUPP;
1883                 }
1884                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1885                 rcu_read_lock();
1886 #else
1887                 err = -EOPNOTSUPP;
1888 #endif
1889         }
1890         return err;
1891 }
1892 #endif
1893
1894 static void
1895 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1896                           unsigned long off)
1897 {
1898         unsigned int bit;
1899
1900         /* Ignore these unchangable bits */
1901         on &= ~IPS_UNCHANGEABLE_MASK;
1902         off &= ~IPS_UNCHANGEABLE_MASK;
1903
1904         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1905                 if (on & (1 << bit))
1906                         set_bit(bit, &ct->status);
1907                 else if (off & (1 << bit))
1908                         clear_bit(bit, &ct->status);
1909         }
1910 }
1911
1912 static int
1913 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1914 {
1915         unsigned long d;
1916         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1917         d = ct->status ^ status;
1918
1919         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1920                 /* unchangeable */
1921                 return -EBUSY;
1922
1923         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1924                 /* SEEN_REPLY bit can only be set */
1925                 return -EBUSY;
1926
1927         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1928                 /* ASSURED bit can only be set */
1929                 return -EBUSY;
1930
1931         __ctnetlink_change_status(ct, status, 0);
1932         return 0;
1933 }
1934
1935 static int
1936 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1937 {
1938 #if IS_ENABLED(CONFIG_NF_NAT)
1939         int ret;
1940
1941         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1942                 return 0;
1943
1944         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1945                                         cda[CTA_NAT_DST]);
1946         if (ret < 0)
1947                 return ret;
1948
1949         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1950                                          cda[CTA_NAT_SRC]);
1951 #else
1952         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1953                 return 0;
1954         return -EOPNOTSUPP;
1955 #endif
1956 }
1957
1958 static int ctnetlink_change_helper(struct nf_conn *ct,
1959                                    const struct nlattr * const cda[])
1960 {
1961         struct nf_conntrack_helper *helper;
1962         struct nf_conn_help *help = nfct_help(ct);
1963         char *helpname = NULL;
1964         struct nlattr *helpinfo = NULL;
1965         int err;
1966
1967         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1968         if (err < 0)
1969                 return err;
1970
1971         /* don't change helper of sibling connections */
1972         if (ct->master) {
1973                 /* If we try to change the helper to the same thing twice,
1974                  * treat the second attempt as a no-op instead of returning
1975                  * an error.
1976                  */
1977                 err = -EBUSY;
1978                 if (help) {
1979                         rcu_read_lock();
1980                         helper = rcu_dereference(help->helper);
1981                         if (helper && !strcmp(helper->name, helpname))
1982                                 err = 0;
1983                         rcu_read_unlock();
1984                 }
1985
1986                 return err;
1987         }
1988
1989         if (!strcmp(helpname, "")) {
1990                 if (help && help->helper) {
1991                         /* we had a helper before ... */
1992                         nf_ct_remove_expectations(ct);
1993                         RCU_INIT_POINTER(help->helper, NULL);
1994                 }
1995
1996                 return 0;
1997         }
1998
1999         rcu_read_lock();
2000         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2001                                             nf_ct_protonum(ct));
2002         if (helper == NULL) {
2003                 rcu_read_unlock();
2004                 return -EOPNOTSUPP;
2005         }
2006
2007         if (help) {
2008                 if (rcu_access_pointer(help->helper) == helper) {
2009                         /* update private helper data if allowed. */
2010                         if (helper->from_nlattr)
2011                                 helper->from_nlattr(helpinfo, ct);
2012                         err = 0;
2013                 } else
2014                         err = -EBUSY;
2015         } else {
2016                 /* we cannot set a helper for an existing conntrack */
2017                 err = -EOPNOTSUPP;
2018         }
2019
2020         rcu_read_unlock();
2021         return err;
2022 }
2023
2024 static int ctnetlink_change_timeout(struct nf_conn *ct,
2025                                     const struct nlattr * const cda[])
2026 {
2027         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2028
2029         if (timeout > INT_MAX)
2030                 timeout = INT_MAX;
2031         WRITE_ONCE(ct->timeout, nfct_time_stamp + (u32)timeout);
2032
2033         if (test_bit(IPS_DYING_BIT, &ct->status))
2034                 return -ETIME;
2035
2036         return 0;
2037 }
2038
2039 #if defined(CONFIG_NF_CONNTRACK_MARK)
2040 static void ctnetlink_change_mark(struct nf_conn *ct,
2041                                     const struct nlattr * const cda[])
2042 {
2043         u32 mark, newmark, mask = 0;
2044
2045         if (cda[CTA_MARK_MASK])
2046                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2047
2048         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2049         newmark = (ct->mark & mask) ^ mark;
2050         if (newmark != ct->mark)
2051                 ct->mark = newmark;
2052 }
2053 #endif
2054
2055 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2056         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2057         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2058         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2059 };
2060
2061 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2062                                       const struct nlattr * const cda[])
2063 {
2064         const struct nlattr *attr = cda[CTA_PROTOINFO];
2065         const struct nf_conntrack_l4proto *l4proto;
2066         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2067         int err = 0;
2068
2069         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2070                                           protoinfo_policy, NULL);
2071         if (err < 0)
2072                 return err;
2073
2074         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2075         if (l4proto->from_nlattr)
2076                 err = l4proto->from_nlattr(tb, ct);
2077
2078         return err;
2079 }
2080
2081 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2082         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2083         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2084         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2085 };
2086
2087 static int change_seq_adj(struct nf_ct_seqadj *seq,
2088                           const struct nlattr * const attr)
2089 {
2090         int err;
2091         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2092
2093         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2094                                           seqadj_policy, NULL);
2095         if (err < 0)
2096                 return err;
2097
2098         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2099                 return -EINVAL;
2100
2101         seq->correction_pos =
2102                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2103
2104         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2105                 return -EINVAL;
2106
2107         seq->offset_before =
2108                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2109
2110         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2111                 return -EINVAL;
2112
2113         seq->offset_after =
2114                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2115
2116         return 0;
2117 }
2118
2119 static int
2120 ctnetlink_change_seq_adj(struct nf_conn *ct,
2121                          const struct nlattr * const cda[])
2122 {
2123         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2124         int ret = 0;
2125
2126         if (!seqadj)
2127                 return 0;
2128
2129         spin_lock_bh(&ct->lock);
2130         if (cda[CTA_SEQ_ADJ_ORIG]) {
2131                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2132                                      cda[CTA_SEQ_ADJ_ORIG]);
2133                 if (ret < 0)
2134                         goto err;
2135
2136                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2137         }
2138
2139         if (cda[CTA_SEQ_ADJ_REPLY]) {
2140                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2141                                      cda[CTA_SEQ_ADJ_REPLY]);
2142                 if (ret < 0)
2143                         goto err;
2144
2145                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2146         }
2147
2148         spin_unlock_bh(&ct->lock);
2149         return 0;
2150 err:
2151         spin_unlock_bh(&ct->lock);
2152         return ret;
2153 }
2154
2155 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2156         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2157         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2158         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2159 };
2160
2161 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2162                                      const struct nlattr * const cda[])
2163 {
2164         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2165         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2166         int err;
2167
2168         if (!synproxy)
2169                 return 0;
2170
2171         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2172                                           cda[CTA_SYNPROXY], synproxy_policy,
2173                                           NULL);
2174         if (err < 0)
2175                 return err;
2176
2177         if (!tb[CTA_SYNPROXY_ISN] ||
2178             !tb[CTA_SYNPROXY_ITS] ||
2179             !tb[CTA_SYNPROXY_TSOFF])
2180                 return -EINVAL;
2181
2182         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2183         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2184         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2185
2186         return 0;
2187 }
2188
2189 static int
2190 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2191 {
2192 #ifdef CONFIG_NF_CONNTRACK_LABELS
2193         size_t len = nla_len(cda[CTA_LABELS]);
2194         const void *mask = cda[CTA_LABELS_MASK];
2195
2196         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2197                 return -EINVAL;
2198
2199         if (mask) {
2200                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2201                     nla_len(cda[CTA_LABELS_MASK]) != len)
2202                         return -EINVAL;
2203                 mask = nla_data(cda[CTA_LABELS_MASK]);
2204         }
2205
2206         len /= sizeof(u32);
2207
2208         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2209 #else
2210         return -EOPNOTSUPP;
2211 #endif
2212 }
2213
2214 static int
2215 ctnetlink_change_conntrack(struct nf_conn *ct,
2216                            const struct nlattr * const cda[])
2217 {
2218         int err;
2219
2220         /* only allow NAT changes and master assignation for new conntracks */
2221         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2222                 return -EOPNOTSUPP;
2223
2224         if (cda[CTA_HELP]) {
2225                 err = ctnetlink_change_helper(ct, cda);
2226                 if (err < 0)
2227                         return err;
2228         }
2229
2230         if (cda[CTA_TIMEOUT]) {
2231                 err = ctnetlink_change_timeout(ct, cda);
2232                 if (err < 0)
2233                         return err;
2234         }
2235
2236         if (cda[CTA_STATUS]) {
2237                 err = ctnetlink_change_status(ct, cda);
2238                 if (err < 0)
2239                         return err;
2240         }
2241
2242         if (cda[CTA_PROTOINFO]) {
2243                 err = ctnetlink_change_protoinfo(ct, cda);
2244                 if (err < 0)
2245                         return err;
2246         }
2247
2248 #if defined(CONFIG_NF_CONNTRACK_MARK)
2249         if (cda[CTA_MARK])
2250                 ctnetlink_change_mark(ct, cda);
2251 #endif
2252
2253         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2254                 err = ctnetlink_change_seq_adj(ct, cda);
2255                 if (err < 0)
2256                         return err;
2257         }
2258
2259         if (cda[CTA_SYNPROXY]) {
2260                 err = ctnetlink_change_synproxy(ct, cda);
2261                 if (err < 0)
2262                         return err;
2263         }
2264
2265         if (cda[CTA_LABELS]) {
2266                 err = ctnetlink_attach_labels(ct, cda);
2267                 if (err < 0)
2268                         return err;
2269         }
2270
2271         return 0;
2272 }
2273
2274 static struct nf_conn *
2275 ctnetlink_create_conntrack(struct net *net,
2276                            const struct nf_conntrack_zone *zone,
2277                            const struct nlattr * const cda[],
2278                            struct nf_conntrack_tuple *otuple,
2279                            struct nf_conntrack_tuple *rtuple,
2280                            u8 u3)
2281 {
2282         struct nf_conn *ct;
2283         int err = -EINVAL;
2284         struct nf_conntrack_helper *helper;
2285         struct nf_conn_tstamp *tstamp;
2286         u64 timeout;
2287
2288         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2289         if (IS_ERR(ct))
2290                 return ERR_PTR(-ENOMEM);
2291
2292         if (!cda[CTA_TIMEOUT])
2293                 goto err1;
2294
2295         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2296         if (timeout > INT_MAX)
2297                 timeout = INT_MAX;
2298         ct->timeout = (u32)timeout + nfct_time_stamp;
2299
2300         rcu_read_lock();
2301         if (cda[CTA_HELP]) {
2302                 char *helpname = NULL;
2303                 struct nlattr *helpinfo = NULL;
2304
2305                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2306                 if (err < 0)
2307                         goto err2;
2308
2309                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2310                                                     nf_ct_protonum(ct));
2311                 if (helper == NULL) {
2312                         rcu_read_unlock();
2313 #ifdef CONFIG_MODULES
2314                         if (request_module("nfct-helper-%s", helpname) < 0) {
2315                                 err = -EOPNOTSUPP;
2316                                 goto err1;
2317                         }
2318
2319                         rcu_read_lock();
2320                         helper = __nf_conntrack_helper_find(helpname,
2321                                                             nf_ct_l3num(ct),
2322                                                             nf_ct_protonum(ct));
2323                         if (helper) {
2324                                 err = -EAGAIN;
2325                                 goto err2;
2326                         }
2327                         rcu_read_unlock();
2328 #endif
2329                         err = -EOPNOTSUPP;
2330                         goto err1;
2331                 } else {
2332                         struct nf_conn_help *help;
2333
2334                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2335                         if (help == NULL) {
2336                                 err = -ENOMEM;
2337                                 goto err2;
2338                         }
2339                         /* set private helper data if allowed. */
2340                         if (helper->from_nlattr)
2341                                 helper->from_nlattr(helpinfo, ct);
2342
2343                         /* disable helper auto-assignment for this entry */
2344                         ct->status |= IPS_HELPER;
2345                         RCU_INIT_POINTER(help->helper, helper);
2346                 }
2347         } else {
2348                 /* try an implicit helper assignation */
2349                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2350                 if (err < 0)
2351                         goto err2;
2352         }
2353
2354         err = ctnetlink_setup_nat(ct, cda);
2355         if (err < 0)
2356                 goto err2;
2357
2358         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2359         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2360         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2361         nf_ct_labels_ext_add(ct);
2362         nfct_seqadj_ext_add(ct);
2363         nfct_synproxy_ext_add(ct);
2364
2365         /* we must add conntrack extensions before confirmation. */
2366         ct->status |= IPS_CONFIRMED;
2367
2368         if (cda[CTA_STATUS]) {
2369                 err = ctnetlink_change_status(ct, cda);
2370                 if (err < 0)
2371                         goto err2;
2372         }
2373
2374         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2375                 err = ctnetlink_change_seq_adj(ct, cda);
2376                 if (err < 0)
2377                         goto err2;
2378         }
2379
2380         memset(&ct->proto, 0, sizeof(ct->proto));
2381         if (cda[CTA_PROTOINFO]) {
2382                 err = ctnetlink_change_protoinfo(ct, cda);
2383                 if (err < 0)
2384                         goto err2;
2385         }
2386
2387         if (cda[CTA_SYNPROXY]) {
2388                 err = ctnetlink_change_synproxy(ct, cda);
2389                 if (err < 0)
2390                         goto err2;
2391         }
2392
2393 #if defined(CONFIG_NF_CONNTRACK_MARK)
2394         if (cda[CTA_MARK])
2395                 ctnetlink_change_mark(ct, cda);
2396 #endif
2397
2398         /* setup master conntrack: this is a confirmed expectation */
2399         if (cda[CTA_TUPLE_MASTER]) {
2400                 struct nf_conntrack_tuple master;
2401                 struct nf_conntrack_tuple_hash *master_h;
2402                 struct nf_conn *master_ct;
2403
2404                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2405                                             u3, NULL);
2406                 if (err < 0)
2407                         goto err2;
2408
2409                 master_h = nf_conntrack_find_get(net, zone, &master);
2410                 if (master_h == NULL) {
2411                         err = -ENOENT;
2412                         goto err2;
2413                 }
2414                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2415                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2416                 ct->master = master_ct;
2417         }
2418         tstamp = nf_conn_tstamp_find(ct);
2419         if (tstamp)
2420                 tstamp->start = ktime_get_real_ns();
2421
2422         err = nf_conntrack_hash_check_insert(ct);
2423         if (err < 0)
2424                 goto err2;
2425
2426         rcu_read_unlock();
2427
2428         return ct;
2429
2430 err2:
2431         rcu_read_unlock();
2432 err1:
2433         nf_conntrack_free(ct);
2434         return ERR_PTR(err);
2435 }
2436
2437 static int ctnetlink_new_conntrack(struct sk_buff *skb,
2438                                    const struct nfnl_info *info,
2439                                    const struct nlattr * const cda[])
2440 {
2441         struct nf_conntrack_tuple otuple, rtuple;
2442         struct nf_conntrack_tuple_hash *h = NULL;
2443         u_int8_t u3 = info->nfmsg->nfgen_family;
2444         struct nf_conntrack_zone zone;
2445         struct nf_conn *ct;
2446         int err;
2447
2448         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2449         if (err < 0)
2450                 return err;
2451
2452         if (cda[CTA_TUPLE_ORIG]) {
2453                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2454                                             u3, &zone);
2455                 if (err < 0)
2456                         return err;
2457         }
2458
2459         if (cda[CTA_TUPLE_REPLY]) {
2460                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2461                                             u3, &zone);
2462                 if (err < 0)
2463                         return err;
2464         }
2465
2466         if (cda[CTA_TUPLE_ORIG])
2467                 h = nf_conntrack_find_get(info->net, &zone, &otuple);
2468         else if (cda[CTA_TUPLE_REPLY])
2469                 h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2470
2471         if (h == NULL) {
2472                 err = -ENOENT;
2473                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2474                         enum ip_conntrack_events events;
2475
2476                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2477                                 return -EINVAL;
2478                         if (otuple.dst.protonum != rtuple.dst.protonum)
2479                                 return -EINVAL;
2480
2481                         ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2482                                                         &otuple, &rtuple, u3);
2483                         if (IS_ERR(ct))
2484                                 return PTR_ERR(ct);
2485
2486                         err = 0;
2487                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2488                                 events = 1 << IPCT_RELATED;
2489                         else
2490                                 events = 1 << IPCT_NEW;
2491
2492                         if (cda[CTA_LABELS] &&
2493                             ctnetlink_attach_labels(ct, cda) == 0)
2494                                 events |= (1 << IPCT_LABEL);
2495
2496                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2497                                                       (1 << IPCT_ASSURED) |
2498                                                       (1 << IPCT_HELPER) |
2499                                                       (1 << IPCT_PROTOINFO) |
2500                                                       (1 << IPCT_SEQADJ) |
2501                                                       (1 << IPCT_MARK) |
2502                                                       (1 << IPCT_SYNPROXY) |
2503                                                       events,
2504                                                       ct, NETLINK_CB(skb).portid,
2505                                                       nlmsg_report(info->nlh));
2506                         nf_ct_put(ct);
2507                 }
2508
2509                 return err;
2510         }
2511         /* implicit 'else' */
2512
2513         err = -EEXIST;
2514         ct = nf_ct_tuplehash_to_ctrack(h);
2515         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2516                 err = ctnetlink_change_conntrack(ct, cda);
2517                 if (err == 0) {
2518                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2519                                                       (1 << IPCT_ASSURED) |
2520                                                       (1 << IPCT_HELPER) |
2521                                                       (1 << IPCT_LABEL) |
2522                                                       (1 << IPCT_PROTOINFO) |
2523                                                       (1 << IPCT_SEQADJ) |
2524                                                       (1 << IPCT_MARK) |
2525                                                       (1 << IPCT_SYNPROXY),
2526                                                       ct, NETLINK_CB(skb).portid,
2527                                                       nlmsg_report(info->nlh));
2528                 }
2529         }
2530
2531         nf_ct_put(ct);
2532         return err;
2533 }
2534
2535 static int
2536 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2537                                 __u16 cpu, const struct ip_conntrack_stat *st)
2538 {
2539         struct nlmsghdr *nlh;
2540         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2541
2542         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2543                               IPCTNL_MSG_CT_GET_STATS_CPU);
2544         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2545                            NFNETLINK_V0, htons(cpu));
2546         if (!nlh)
2547                 goto nlmsg_failure;
2548
2549         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2550             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2551             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2552             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2553                                 htonl(st->insert_failed)) ||
2554             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2555             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2556             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2557             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2558                                 htonl(st->search_restart)) ||
2559             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2560                                 htonl(st->clash_resolve)) ||
2561             nla_put_be32(skb, CTA_STATS_CHAIN_TOOLONG,
2562                          htonl(st->chaintoolong)))
2563                 goto nla_put_failure;
2564
2565         nlmsg_end(skb, nlh);
2566         return skb->len;
2567
2568 nla_put_failure:
2569 nlmsg_failure:
2570         nlmsg_cancel(skb, nlh);
2571         return -1;
2572 }
2573
2574 static int
2575 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2576 {
2577         int cpu;
2578         struct net *net = sock_net(skb->sk);
2579
2580         if (cb->args[0] == nr_cpu_ids)
2581                 return 0;
2582
2583         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2584                 const struct ip_conntrack_stat *st;
2585
2586                 if (!cpu_possible(cpu))
2587                         continue;
2588
2589                 st = per_cpu_ptr(net->ct.stat, cpu);
2590                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2591                                                     NETLINK_CB(cb->skb).portid,
2592                                                     cb->nlh->nlmsg_seq,
2593                                                     cpu, st) < 0)
2594                                 break;
2595         }
2596         cb->args[0] = cpu;
2597
2598         return skb->len;
2599 }
2600
2601 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2602                                  const struct nfnl_info *info,
2603                                  const struct nlattr * const cda[])
2604 {
2605         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2606                 struct netlink_dump_control c = {
2607                         .dump = ctnetlink_ct_stat_cpu_dump,
2608                 };
2609                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2610         }
2611
2612         return 0;
2613 }
2614
2615 static int
2616 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2617                             struct net *net)
2618 {
2619         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2620         unsigned int nr_conntracks;
2621         struct nlmsghdr *nlh;
2622
2623         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2624         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2625                            NFNETLINK_V0, 0);
2626         if (!nlh)
2627                 goto nlmsg_failure;
2628
2629         nr_conntracks = nf_conntrack_count(net);
2630         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2631                 goto nla_put_failure;
2632
2633         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2634                 goto nla_put_failure;
2635
2636         nlmsg_end(skb, nlh);
2637         return skb->len;
2638
2639 nla_put_failure:
2640 nlmsg_failure:
2641         nlmsg_cancel(skb, nlh);
2642         return -1;
2643 }
2644
2645 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2646                              const struct nlattr * const cda[])
2647 {
2648         struct sk_buff *skb2;
2649         int err;
2650
2651         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2652         if (skb2 == NULL)
2653                 return -ENOMEM;
2654
2655         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2656                                           info->nlh->nlmsg_seq,
2657                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2658                                           sock_net(skb->sk));
2659         if (err <= 0) {
2660                 kfree_skb(skb2);
2661                 return -ENOMEM;
2662         }
2663
2664         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2665 }
2666
2667 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2668         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2669         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2670         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2671         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2672         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2673         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2674                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2675         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2676         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2677         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2678         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2679         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2680 };
2681
2682 static struct nf_conntrack_expect *
2683 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2684                        struct nf_conntrack_helper *helper,
2685                        struct nf_conntrack_tuple *tuple,
2686                        struct nf_conntrack_tuple *mask);
2687
2688 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2689 static size_t
2690 ctnetlink_glue_build_size(const struct nf_conn *ct)
2691 {
2692         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2693                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2694                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2695                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2696                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2697                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2698                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2699                + nla_total_size(0) /* CTA_PROTOINFO */
2700                + nla_total_size(0) /* CTA_HELP */
2701                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2702                + ctnetlink_secctx_size(ct)
2703                + ctnetlink_acct_size(ct)
2704                + ctnetlink_timestamp_size(ct)
2705 #if IS_ENABLED(CONFIG_NF_NAT)
2706                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2707                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2708 #endif
2709 #ifdef CONFIG_NF_CONNTRACK_MARK
2710                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2711 #endif
2712 #ifdef CONFIG_NF_CONNTRACK_ZONES
2713                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2714 #endif
2715                + ctnetlink_proto_size(ct)
2716                ;
2717 }
2718
2719 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2720 {
2721         const struct nf_conntrack_zone *zone;
2722         struct nlattr *nest_parms;
2723
2724         zone = nf_ct_zone(ct);
2725
2726         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2727         if (!nest_parms)
2728                 goto nla_put_failure;
2729         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2730                 goto nla_put_failure;
2731         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2732                                    NF_CT_ZONE_DIR_ORIG) < 0)
2733                 goto nla_put_failure;
2734         nla_nest_end(skb, nest_parms);
2735
2736         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2737         if (!nest_parms)
2738                 goto nla_put_failure;
2739         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2740                 goto nla_put_failure;
2741         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2742                                    NF_CT_ZONE_DIR_REPL) < 0)
2743                 goto nla_put_failure;
2744         nla_nest_end(skb, nest_parms);
2745
2746         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2747                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2748                 goto nla_put_failure;
2749
2750         if (ctnetlink_dump_id(skb, ct) < 0)
2751                 goto nla_put_failure;
2752
2753         if (ctnetlink_dump_status(skb, ct) < 0)
2754                 goto nla_put_failure;
2755
2756         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2757                 goto nla_put_failure;
2758
2759         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2760                 goto nla_put_failure;
2761
2762         if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2763             ctnetlink_dump_timestamp(skb, ct) < 0)
2764                 goto nla_put_failure;
2765
2766         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2767                 goto nla_put_failure;
2768
2769 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2770         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2771                 goto nla_put_failure;
2772 #endif
2773         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2774                 goto nla_put_failure;
2775
2776         if ((ct->status & IPS_SEQ_ADJUST) &&
2777             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2778                 goto nla_put_failure;
2779
2780         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2781                 goto nla_put_failure;
2782
2783 #ifdef CONFIG_NF_CONNTRACK_MARK
2784         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2785                 goto nla_put_failure;
2786 #endif
2787         if (ctnetlink_dump_labels(skb, ct) < 0)
2788                 goto nla_put_failure;
2789         return 0;
2790
2791 nla_put_failure:
2792         return -ENOSPC;
2793 }
2794
2795 static int
2796 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2797                      enum ip_conntrack_info ctinfo,
2798                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2799 {
2800         struct nlattr *nest_parms;
2801
2802         nest_parms = nla_nest_start(skb, ct_attr);
2803         if (!nest_parms)
2804                 goto nla_put_failure;
2805
2806         if (__ctnetlink_glue_build(skb, ct) < 0)
2807                 goto nla_put_failure;
2808
2809         nla_nest_end(skb, nest_parms);
2810
2811         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2812                 goto nla_put_failure;
2813
2814         return 0;
2815
2816 nla_put_failure:
2817         return -ENOSPC;
2818 }
2819
2820 static int
2821 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2822 {
2823         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2824         unsigned long d = ct->status ^ status;
2825
2826         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2827                 /* SEEN_REPLY bit can only be set */
2828                 return -EBUSY;
2829
2830         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2831                 /* ASSURED bit can only be set */
2832                 return -EBUSY;
2833
2834         /* This check is less strict than ctnetlink_change_status()
2835          * because callers often flip IPS_EXPECTED bits when sending
2836          * an NFQA_CT attribute to the kernel.  So ignore the
2837          * unchangeable bits but do not error out. Also user programs
2838          * are allowed to clear the bits that they are allowed to change.
2839          */
2840         __ctnetlink_change_status(ct, status, ~status);
2841         return 0;
2842 }
2843
2844 static int
2845 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2846 {
2847         int err;
2848
2849         if (cda[CTA_TIMEOUT]) {
2850                 err = ctnetlink_change_timeout(ct, cda);
2851                 if (err < 0)
2852                         return err;
2853         }
2854         if (cda[CTA_STATUS]) {
2855                 err = ctnetlink_update_status(ct, cda);
2856                 if (err < 0)
2857                         return err;
2858         }
2859         if (cda[CTA_HELP]) {
2860                 err = ctnetlink_change_helper(ct, cda);
2861                 if (err < 0)
2862                         return err;
2863         }
2864         if (cda[CTA_LABELS]) {
2865                 err = ctnetlink_attach_labels(ct, cda);
2866                 if (err < 0)
2867                         return err;
2868         }
2869 #if defined(CONFIG_NF_CONNTRACK_MARK)
2870         if (cda[CTA_MARK]) {
2871                 ctnetlink_change_mark(ct, cda);
2872         }
2873 #endif
2874         return 0;
2875 }
2876
2877 static int
2878 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2879 {
2880         struct nlattr *cda[CTA_MAX+1];
2881         int ret;
2882
2883         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2884                                           NULL);
2885         if (ret < 0)
2886                 return ret;
2887
2888         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2889 }
2890
2891 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2892                                     const struct nf_conn *ct,
2893                                     struct nf_conntrack_tuple *tuple,
2894                                     struct nf_conntrack_tuple *mask)
2895 {
2896         int err;
2897
2898         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2899                                     nf_ct_l3num(ct), NULL);
2900         if (err < 0)
2901                 return err;
2902
2903         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2904                                      nf_ct_l3num(ct), NULL);
2905 }
2906
2907 static int
2908 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2909                              u32 portid, u32 report)
2910 {
2911         struct nlattr *cda[CTA_EXPECT_MAX+1];
2912         struct nf_conntrack_tuple tuple, mask;
2913         struct nf_conntrack_helper *helper = NULL;
2914         struct nf_conntrack_expect *exp;
2915         int err;
2916
2917         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2918                                           exp_nla_policy, NULL);
2919         if (err < 0)
2920                 return err;
2921
2922         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2923                                        ct, &tuple, &mask);
2924         if (err < 0)
2925                 return err;
2926
2927         if (cda[CTA_EXPECT_HELP_NAME]) {
2928                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2929
2930                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2931                                                     nf_ct_protonum(ct));
2932                 if (helper == NULL)
2933                         return -EOPNOTSUPP;
2934         }
2935
2936         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2937                                      helper, &tuple, &mask);
2938         if (IS_ERR(exp))
2939                 return PTR_ERR(exp);
2940
2941         err = nf_ct_expect_related_report(exp, portid, report, 0);
2942         nf_ct_expect_put(exp);
2943         return err;
2944 }
2945
2946 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2947                                   enum ip_conntrack_info ctinfo, int diff)
2948 {
2949         if (!(ct->status & IPS_NAT_MASK))
2950                 return;
2951
2952         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2953 }
2954
2955 static const struct nfnl_ct_hook ctnetlink_glue_hook = {
2956         .build_size     = ctnetlink_glue_build_size,
2957         .build          = ctnetlink_glue_build,
2958         .parse          = ctnetlink_glue_parse,
2959         .attach_expect  = ctnetlink_glue_attach_expect,
2960         .seq_adjust     = ctnetlink_glue_seqadj,
2961 };
2962 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2963
2964 /***********************************************************************
2965  * EXPECT
2966  ***********************************************************************/
2967
2968 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2969                                     const struct nf_conntrack_tuple *tuple,
2970                                     u32 type)
2971 {
2972         struct nlattr *nest_parms;
2973
2974         nest_parms = nla_nest_start(skb, type);
2975         if (!nest_parms)
2976                 goto nla_put_failure;
2977         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2978                 goto nla_put_failure;
2979         nla_nest_end(skb, nest_parms);
2980
2981         return 0;
2982
2983 nla_put_failure:
2984         return -1;
2985 }
2986
2987 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2988                                    const struct nf_conntrack_tuple *tuple,
2989                                    const struct nf_conntrack_tuple_mask *mask)
2990 {
2991         const struct nf_conntrack_l4proto *l4proto;
2992         struct nf_conntrack_tuple m;
2993         struct nlattr *nest_parms;
2994         int ret;
2995
2996         memset(&m, 0xFF, sizeof(m));
2997         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2998         m.src.u.all = mask->src.u.all;
2999         m.src.l3num = tuple->src.l3num;
3000         m.dst.protonum = tuple->dst.protonum;
3001
3002         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
3003         if (!nest_parms)
3004                 goto nla_put_failure;
3005
3006         rcu_read_lock();
3007         ret = ctnetlink_dump_tuples_ip(skb, &m);
3008         if (ret >= 0) {
3009                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
3010                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
3011         }
3012         rcu_read_unlock();
3013
3014         if (unlikely(ret < 0))
3015                 goto nla_put_failure;
3016
3017         nla_nest_end(skb, nest_parms);
3018
3019         return 0;
3020
3021 nla_put_failure:
3022         return -1;
3023 }
3024
3025 static const union nf_inet_addr any_addr;
3026
3027 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
3028 {
3029         static siphash_aligned_key_t exp_id_seed;
3030         unsigned long a, b, c, d;
3031
3032         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
3033
3034         a = (unsigned long)exp;
3035         b = (unsigned long)exp->helper;
3036         c = (unsigned long)exp->master;
3037         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
3038
3039 #ifdef CONFIG_64BIT
3040         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
3041 #else
3042         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
3043 #endif
3044 }
3045
3046 static int
3047 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3048                           const struct nf_conntrack_expect *exp)
3049 {
3050         struct nf_conn *master = exp->master;
3051         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3052         struct nf_conn_help *help;
3053 #if IS_ENABLED(CONFIG_NF_NAT)
3054         struct nlattr *nest_parms;
3055         struct nf_conntrack_tuple nat_tuple = {};
3056 #endif
3057         struct nf_ct_helper_expectfn *expfn;
3058
3059         if (timeout < 0)
3060                 timeout = 0;
3061
3062         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3063                 goto nla_put_failure;
3064         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3065                 goto nla_put_failure;
3066         if (ctnetlink_exp_dump_tuple(skb,
3067                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3068                                  CTA_EXPECT_MASTER) < 0)
3069                 goto nla_put_failure;
3070
3071 #if IS_ENABLED(CONFIG_NF_NAT)
3072         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3073             exp->saved_proto.all) {
3074                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3075                 if (!nest_parms)
3076                         goto nla_put_failure;
3077
3078                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3079                         goto nla_put_failure;
3080
3081                 nat_tuple.src.l3num = nf_ct_l3num(master);
3082                 nat_tuple.src.u3 = exp->saved_addr;
3083                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3084                 nat_tuple.src.u = exp->saved_proto;
3085
3086                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3087                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3088                         goto nla_put_failure;
3089                 nla_nest_end(skb, nest_parms);
3090         }
3091 #endif
3092         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3093             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3094             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3095             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3096                 goto nla_put_failure;
3097         help = nfct_help(master);
3098         if (help) {
3099                 struct nf_conntrack_helper *helper;
3100
3101                 helper = rcu_dereference(help->helper);
3102                 if (helper &&
3103                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3104                         goto nla_put_failure;
3105         }
3106         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3107         if (expfn != NULL &&
3108             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3109                 goto nla_put_failure;
3110
3111         return 0;
3112
3113 nla_put_failure:
3114         return -1;
3115 }
3116
3117 static int
3118 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3119                         int event, const struct nf_conntrack_expect *exp)
3120 {
3121         struct nlmsghdr *nlh;
3122         unsigned int flags = portid ? NLM_F_MULTI : 0;
3123
3124         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3125         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3126                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3127         if (!nlh)
3128                 goto nlmsg_failure;
3129
3130         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3131                 goto nla_put_failure;
3132
3133         nlmsg_end(skb, nlh);
3134         return skb->len;
3135
3136 nlmsg_failure:
3137 nla_put_failure:
3138         nlmsg_cancel(skb, nlh);
3139         return -1;
3140 }
3141
3142 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3143 static int
3144 ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3145 {
3146         struct nf_conntrack_expect *exp = item->exp;
3147         struct net *net = nf_ct_exp_net(exp);
3148         struct nlmsghdr *nlh;
3149         struct sk_buff *skb;
3150         unsigned int type, group;
3151         int flags = 0;
3152
3153         if (events & (1 << IPEXP_DESTROY)) {
3154                 type = IPCTNL_MSG_EXP_DELETE;
3155                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3156         } else if (events & (1 << IPEXP_NEW)) {
3157                 type = IPCTNL_MSG_EXP_NEW;
3158                 flags = NLM_F_CREATE|NLM_F_EXCL;
3159                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3160         } else
3161                 return 0;
3162
3163         if (!item->report && !nfnetlink_has_listeners(net, group))
3164                 return 0;
3165
3166         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3167         if (skb == NULL)
3168                 goto errout;
3169
3170         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3171         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3172                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3173         if (!nlh)
3174                 goto nlmsg_failure;
3175
3176         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3177                 goto nla_put_failure;
3178
3179         nlmsg_end(skb, nlh);
3180         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3181         return 0;
3182
3183 nla_put_failure:
3184         nlmsg_cancel(skb, nlh);
3185 nlmsg_failure:
3186         kfree_skb(skb);
3187 errout:
3188         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3189         return 0;
3190 }
3191 #endif
3192 static int ctnetlink_exp_done(struct netlink_callback *cb)
3193 {
3194         if (cb->args[1])
3195                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3196         return 0;
3197 }
3198
3199 static int
3200 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3201 {
3202         struct net *net = sock_net(skb->sk);
3203         struct nf_conntrack_expect *exp, *last;
3204         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3205         u_int8_t l3proto = nfmsg->nfgen_family;
3206
3207         rcu_read_lock();
3208         last = (struct nf_conntrack_expect *)cb->args[1];
3209         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3210 restart:
3211                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3212                                          hnode) {
3213                         if (l3proto && exp->tuple.src.l3num != l3proto)
3214                                 continue;
3215
3216                         if (!net_eq(nf_ct_net(exp->master), net))
3217                                 continue;
3218
3219                         if (cb->args[1]) {
3220                                 if (exp != last)
3221                                         continue;
3222                                 cb->args[1] = 0;
3223                         }
3224                         if (ctnetlink_exp_fill_info(skb,
3225                                                     NETLINK_CB(cb->skb).portid,
3226                                                     cb->nlh->nlmsg_seq,
3227                                                     IPCTNL_MSG_EXP_NEW,
3228                                                     exp) < 0) {
3229                                 if (!refcount_inc_not_zero(&exp->use))
3230                                         continue;
3231                                 cb->args[1] = (unsigned long)exp;
3232                                 goto out;
3233                         }
3234                 }
3235                 if (cb->args[1]) {
3236                         cb->args[1] = 0;
3237                         goto restart;
3238                 }
3239         }
3240 out:
3241         rcu_read_unlock();
3242         if (last)
3243                 nf_ct_expect_put(last);
3244
3245         return skb->len;
3246 }
3247
3248 static int
3249 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3250 {
3251         struct nf_conntrack_expect *exp, *last;
3252         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3253         struct nf_conn *ct = cb->data;
3254         struct nf_conn_help *help = nfct_help(ct);
3255         u_int8_t l3proto = nfmsg->nfgen_family;
3256
3257         if (cb->args[0])
3258                 return 0;
3259
3260         rcu_read_lock();
3261         last = (struct nf_conntrack_expect *)cb->args[1];
3262 restart:
3263         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3264                 if (l3proto && exp->tuple.src.l3num != l3proto)
3265                         continue;
3266                 if (cb->args[1]) {
3267                         if (exp != last)
3268                                 continue;
3269                         cb->args[1] = 0;
3270                 }
3271                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3272                                             cb->nlh->nlmsg_seq,
3273                                             IPCTNL_MSG_EXP_NEW,
3274                                             exp) < 0) {
3275                         if (!refcount_inc_not_zero(&exp->use))
3276                                 continue;
3277                         cb->args[1] = (unsigned long)exp;
3278                         goto out;
3279                 }
3280         }
3281         if (cb->args[1]) {
3282                 cb->args[1] = 0;
3283                 goto restart;
3284         }
3285         cb->args[0] = 1;
3286 out:
3287         rcu_read_unlock();
3288         if (last)
3289                 nf_ct_expect_put(last);
3290
3291         return skb->len;
3292 }
3293
3294 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3295                                  struct sk_buff *skb,
3296                                  const struct nlmsghdr *nlh,
3297                                  const struct nlattr * const cda[],
3298                                  struct netlink_ext_ack *extack)
3299 {
3300         int err;
3301         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3302         u_int8_t u3 = nfmsg->nfgen_family;
3303         struct nf_conntrack_tuple tuple;
3304         struct nf_conntrack_tuple_hash *h;
3305         struct nf_conn *ct;
3306         struct nf_conntrack_zone zone;
3307         struct netlink_dump_control c = {
3308                 .dump = ctnetlink_exp_ct_dump_table,
3309                 .done = ctnetlink_exp_done,
3310         };
3311
3312         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3313                                     u3, NULL);
3314         if (err < 0)
3315                 return err;
3316
3317         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3318         if (err < 0)
3319                 return err;
3320
3321         h = nf_conntrack_find_get(net, &zone, &tuple);
3322         if (!h)
3323                 return -ENOENT;
3324
3325         ct = nf_ct_tuplehash_to_ctrack(h);
3326         /* No expectation linked to this connection tracking. */
3327         if (!nfct_help(ct)) {
3328                 nf_ct_put(ct);
3329                 return 0;
3330         }
3331
3332         c.data = ct;
3333
3334         err = netlink_dump_start(ctnl, skb, nlh, &c);
3335         nf_ct_put(ct);
3336
3337         return err;
3338 }
3339
3340 static int ctnetlink_get_expect(struct sk_buff *skb,
3341                                 const struct nfnl_info *info,
3342                                 const struct nlattr * const cda[])
3343 {
3344         u_int8_t u3 = info->nfmsg->nfgen_family;
3345         struct nf_conntrack_tuple tuple;
3346         struct nf_conntrack_expect *exp;
3347         struct nf_conntrack_zone zone;
3348         struct sk_buff *skb2;
3349         int err;
3350
3351         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3352                 if (cda[CTA_EXPECT_MASTER])
3353                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3354                                                      info->nlh, cda,
3355                                                      info->extack);
3356                 else {
3357                         struct netlink_dump_control c = {
3358                                 .dump = ctnetlink_exp_dump_table,
3359                                 .done = ctnetlink_exp_done,
3360                         };
3361                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3362                 }
3363         }
3364
3365         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3366         if (err < 0)
3367                 return err;
3368
3369         if (cda[CTA_EXPECT_TUPLE])
3370                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3371                                             u3, NULL);
3372         else if (cda[CTA_EXPECT_MASTER])
3373                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3374                                             u3, NULL);
3375         else
3376                 return -EINVAL;
3377
3378         if (err < 0)
3379                 return err;
3380
3381         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3382         if (!exp)
3383                 return -ENOENT;
3384
3385         if (cda[CTA_EXPECT_ID]) {
3386                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3387
3388                 if (id != nf_expect_get_id(exp)) {
3389                         nf_ct_expect_put(exp);
3390                         return -ENOENT;
3391                 }
3392         }
3393
3394         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3395         if (!skb2) {
3396                 nf_ct_expect_put(exp);
3397                 return -ENOMEM;
3398         }
3399
3400         rcu_read_lock();
3401         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3402                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3403                                       exp);
3404         rcu_read_unlock();
3405         nf_ct_expect_put(exp);
3406         if (err <= 0) {
3407                 kfree_skb(skb2);
3408                 return -ENOMEM;
3409         }
3410
3411         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3412 }
3413
3414 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3415 {
3416         struct nf_conntrack_helper *helper;
3417         const struct nf_conn_help *m_help;
3418         const char *name = data;
3419
3420         m_help = nfct_help(exp->master);
3421
3422         helper = rcu_dereference(m_help->helper);
3423         if (!helper)
3424                 return false;
3425
3426         return strcmp(helper->name, name) == 0;
3427 }
3428
3429 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3430 {
3431         return true;
3432 }
3433
3434 static int ctnetlink_del_expect(struct sk_buff *skb,
3435                                 const struct nfnl_info *info,
3436                                 const struct nlattr * const cda[])
3437 {
3438         u_int8_t u3 = info->nfmsg->nfgen_family;
3439         struct nf_conntrack_expect *exp;
3440         struct nf_conntrack_tuple tuple;
3441         struct nf_conntrack_zone zone;
3442         int err;
3443
3444         if (cda[CTA_EXPECT_TUPLE]) {
3445                 /* delete a single expect by tuple */
3446                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3447                 if (err < 0)
3448                         return err;
3449
3450                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3451                                             u3, NULL);
3452                 if (err < 0)
3453                         return err;
3454
3455                 /* bump usage count to 2 */
3456                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3457                 if (!exp)
3458                         return -ENOENT;
3459
3460                 if (cda[CTA_EXPECT_ID]) {
3461                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3462                         if (ntohl(id) != (u32)(unsigned long)exp) {
3463                                 nf_ct_expect_put(exp);
3464                                 return -ENOENT;
3465                         }
3466                 }
3467
3468                 /* after list removal, usage count == 1 */
3469                 spin_lock_bh(&nf_conntrack_expect_lock);
3470                 if (del_timer(&exp->timeout)) {
3471                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3472                                                    nlmsg_report(info->nlh));
3473                         nf_ct_expect_put(exp);
3474                 }
3475                 spin_unlock_bh(&nf_conntrack_expect_lock);
3476                 /* have to put what we 'get' above.
3477                  * after this line usage count == 0 */
3478                 nf_ct_expect_put(exp);
3479         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3480                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3481
3482                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3483                                          NETLINK_CB(skb).portid,
3484                                          nlmsg_report(info->nlh));
3485         } else {
3486                 /* This basically means we have to flush everything*/
3487                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3488                                          NETLINK_CB(skb).portid,
3489                                          nlmsg_report(info->nlh));
3490         }
3491
3492         return 0;
3493 }
3494 static int
3495 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3496                         const struct nlattr * const cda[])
3497 {
3498         if (cda[CTA_EXPECT_TIMEOUT]) {
3499                 if (!del_timer(&x->timeout))
3500                         return -ETIME;
3501
3502                 x->timeout.expires = jiffies +
3503                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3504                 add_timer(&x->timeout);
3505         }
3506         return 0;
3507 }
3508
3509 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3510         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3511         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3512 };
3513
3514 static int
3515 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3516                            struct nf_conntrack_expect *exp,
3517                            u_int8_t u3)
3518 {
3519 #if IS_ENABLED(CONFIG_NF_NAT)
3520         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3521         struct nf_conntrack_tuple nat_tuple = {};
3522         int err;
3523
3524         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3525                                           exp_nat_nla_policy, NULL);
3526         if (err < 0)
3527                 return err;
3528
3529         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3530                 return -EINVAL;
3531
3532         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3533                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3534                                     u3, NULL);
3535         if (err < 0)
3536                 return err;
3537
3538         exp->saved_addr = nat_tuple.src.u3;
3539         exp->saved_proto = nat_tuple.src.u;
3540         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3541
3542         return 0;
3543 #else
3544         return -EOPNOTSUPP;
3545 #endif
3546 }
3547
3548 static struct nf_conntrack_expect *
3549 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3550                        struct nf_conntrack_helper *helper,
3551                        struct nf_conntrack_tuple *tuple,
3552                        struct nf_conntrack_tuple *mask)
3553 {
3554         u_int32_t class = 0;
3555         struct nf_conntrack_expect *exp;
3556         struct nf_conn_help *help;
3557         int err;
3558
3559         help = nfct_help(ct);
3560         if (!help)
3561                 return ERR_PTR(-EOPNOTSUPP);
3562
3563         if (cda[CTA_EXPECT_CLASS] && helper) {
3564                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3565                 if (class > helper->expect_class_max)
3566                         return ERR_PTR(-EINVAL);
3567         }
3568         exp = nf_ct_expect_alloc(ct);
3569         if (!exp)
3570                 return ERR_PTR(-ENOMEM);
3571
3572         if (cda[CTA_EXPECT_FLAGS]) {
3573                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3574                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3575         } else {
3576                 exp->flags = 0;
3577         }
3578         if (cda[CTA_EXPECT_FN]) {
3579                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3580                 struct nf_ct_helper_expectfn *expfn;
3581
3582                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3583                 if (expfn == NULL) {
3584                         err = -EINVAL;
3585                         goto err_out;
3586                 }
3587                 exp->expectfn = expfn->expectfn;
3588         } else
3589                 exp->expectfn = NULL;
3590
3591         exp->class = class;
3592         exp->master = ct;
3593         exp->helper = helper;
3594         exp->tuple = *tuple;
3595         exp->mask.src.u3 = mask->src.u3;
3596         exp->mask.src.u.all = mask->src.u.all;
3597
3598         if (cda[CTA_EXPECT_NAT]) {
3599                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3600                                                  exp, nf_ct_l3num(ct));
3601                 if (err < 0)
3602                         goto err_out;
3603         }
3604         return exp;
3605 err_out:
3606         nf_ct_expect_put(exp);
3607         return ERR_PTR(err);
3608 }
3609
3610 static int
3611 ctnetlink_create_expect(struct net *net,
3612                         const struct nf_conntrack_zone *zone,
3613                         const struct nlattr * const cda[],
3614                         u_int8_t u3, u32 portid, int report)
3615 {
3616         struct nf_conntrack_tuple tuple, mask, master_tuple;
3617         struct nf_conntrack_tuple_hash *h = NULL;
3618         struct nf_conntrack_helper *helper = NULL;
3619         struct nf_conntrack_expect *exp;
3620         struct nf_conn *ct;
3621         int err;
3622
3623         /* caller guarantees that those three CTA_EXPECT_* exist */
3624         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3625                                     u3, NULL);
3626         if (err < 0)
3627                 return err;
3628         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3629                                     u3, NULL);
3630         if (err < 0)
3631                 return err;
3632         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3633                                     u3, NULL);
3634         if (err < 0)
3635                 return err;
3636
3637         /* Look for master conntrack of this expectation */
3638         h = nf_conntrack_find_get(net, zone, &master_tuple);
3639         if (!h)
3640                 return -ENOENT;
3641         ct = nf_ct_tuplehash_to_ctrack(h);
3642
3643         rcu_read_lock();
3644         if (cda[CTA_EXPECT_HELP_NAME]) {
3645                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3646
3647                 helper = __nf_conntrack_helper_find(helpname, u3,
3648                                                     nf_ct_protonum(ct));
3649                 if (helper == NULL) {
3650                         rcu_read_unlock();
3651 #ifdef CONFIG_MODULES
3652                         if (request_module("nfct-helper-%s", helpname) < 0) {
3653                                 err = -EOPNOTSUPP;
3654                                 goto err_ct;
3655                         }
3656                         rcu_read_lock();
3657                         helper = __nf_conntrack_helper_find(helpname, u3,
3658                                                             nf_ct_protonum(ct));
3659                         if (helper) {
3660                                 err = -EAGAIN;
3661                                 goto err_rcu;
3662                         }
3663                         rcu_read_unlock();
3664 #endif
3665                         err = -EOPNOTSUPP;
3666                         goto err_ct;
3667                 }
3668         }
3669
3670         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3671         if (IS_ERR(exp)) {
3672                 err = PTR_ERR(exp);
3673                 goto err_rcu;
3674         }
3675
3676         err = nf_ct_expect_related_report(exp, portid, report, 0);
3677         nf_ct_expect_put(exp);
3678 err_rcu:
3679         rcu_read_unlock();
3680 err_ct:
3681         nf_ct_put(ct);
3682         return err;
3683 }
3684
3685 static int ctnetlink_new_expect(struct sk_buff *skb,
3686                                 const struct nfnl_info *info,
3687                                 const struct nlattr * const cda[])
3688 {
3689         u_int8_t u3 = info->nfmsg->nfgen_family;
3690         struct nf_conntrack_tuple tuple;
3691         struct nf_conntrack_expect *exp;
3692         struct nf_conntrack_zone zone;
3693         int err;
3694
3695         if (!cda[CTA_EXPECT_TUPLE]
3696             || !cda[CTA_EXPECT_MASK]
3697             || !cda[CTA_EXPECT_MASTER])
3698                 return -EINVAL;
3699
3700         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3701         if (err < 0)
3702                 return err;
3703
3704         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3705                                     u3, NULL);
3706         if (err < 0)
3707                 return err;
3708
3709         spin_lock_bh(&nf_conntrack_expect_lock);
3710         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3711         if (!exp) {
3712                 spin_unlock_bh(&nf_conntrack_expect_lock);
3713                 err = -ENOENT;
3714                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3715                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3716                                                       NETLINK_CB(skb).portid,
3717                                                       nlmsg_report(info->nlh));
3718                 }
3719                 return err;
3720         }
3721
3722         err = -EEXIST;
3723         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3724                 err = ctnetlink_change_expect(exp, cda);
3725         spin_unlock_bh(&nf_conntrack_expect_lock);
3726
3727         return err;
3728 }
3729
3730 static int
3731 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3732                              const struct ip_conntrack_stat *st)
3733 {
3734         struct nlmsghdr *nlh;
3735         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3736
3737         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3738                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3739         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3740                            NFNETLINK_V0, htons(cpu));
3741         if (!nlh)
3742                 goto nlmsg_failure;
3743
3744         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3745             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3746             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3747                 goto nla_put_failure;
3748
3749         nlmsg_end(skb, nlh);
3750         return skb->len;
3751
3752 nla_put_failure:
3753 nlmsg_failure:
3754         nlmsg_cancel(skb, nlh);
3755         return -1;
3756 }
3757
3758 static int
3759 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3760 {
3761         int cpu;
3762         struct net *net = sock_net(skb->sk);
3763
3764         if (cb->args[0] == nr_cpu_ids)
3765                 return 0;
3766
3767         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3768                 const struct ip_conntrack_stat *st;
3769
3770                 if (!cpu_possible(cpu))
3771                         continue;
3772
3773                 st = per_cpu_ptr(net->ct.stat, cpu);
3774                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3775                                                  cb->nlh->nlmsg_seq,
3776                                                  cpu, st) < 0)
3777                         break;
3778         }
3779         cb->args[0] = cpu;
3780
3781         return skb->len;
3782 }
3783
3784 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3785                                   const struct nfnl_info *info,
3786                                   const struct nlattr * const cda[])
3787 {
3788         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3789                 struct netlink_dump_control c = {
3790                         .dump = ctnetlink_exp_stat_cpu_dump,
3791                 };
3792                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3793         }
3794
3795         return 0;
3796 }
3797
3798 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3799 static struct nf_ct_event_notifier ctnl_notifier = {
3800         .ct_event = ctnetlink_conntrack_event,
3801         .exp_event = ctnetlink_expect_event,
3802 };
3803 #endif
3804
3805 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3806         [IPCTNL_MSG_CT_NEW]     = {
3807                 .call           = ctnetlink_new_conntrack,
3808                 .type           = NFNL_CB_MUTEX,
3809                 .attr_count     = CTA_MAX,
3810                 .policy         = ct_nla_policy
3811         },
3812         [IPCTNL_MSG_CT_GET]     = {
3813                 .call           = ctnetlink_get_conntrack,
3814                 .type           = NFNL_CB_MUTEX,
3815                 .attr_count     = CTA_MAX,
3816                 .policy         = ct_nla_policy
3817         },
3818         [IPCTNL_MSG_CT_DELETE]  = {
3819                 .call           = ctnetlink_del_conntrack,
3820                 .type           = NFNL_CB_MUTEX,
3821                 .attr_count     = CTA_MAX,
3822                 .policy         = ct_nla_policy
3823         },
3824         [IPCTNL_MSG_CT_GET_CTRZERO] = {
3825                 .call           = ctnetlink_get_conntrack,
3826                 .type           = NFNL_CB_MUTEX,
3827                 .attr_count     = CTA_MAX,
3828                 .policy         = ct_nla_policy
3829         },
3830         [IPCTNL_MSG_CT_GET_STATS_CPU] = {
3831                 .call           = ctnetlink_stat_ct_cpu,
3832                 .type           = NFNL_CB_MUTEX,
3833         },
3834         [IPCTNL_MSG_CT_GET_STATS] = {
3835                 .call           = ctnetlink_stat_ct,
3836                 .type           = NFNL_CB_MUTEX,
3837         },
3838         [IPCTNL_MSG_CT_GET_DYING] = {
3839                 .call           = ctnetlink_get_ct_dying,
3840                 .type           = NFNL_CB_MUTEX,
3841         },
3842         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = {
3843                 .call           = ctnetlink_get_ct_unconfirmed,
3844                 .type           = NFNL_CB_MUTEX,
3845         },
3846 };
3847
3848 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3849         [IPCTNL_MSG_EXP_GET] = {
3850                 .call           = ctnetlink_get_expect,
3851                 .type           = NFNL_CB_MUTEX,
3852                 .attr_count     = CTA_EXPECT_MAX,
3853                 .policy         = exp_nla_policy
3854         },
3855         [IPCTNL_MSG_EXP_NEW] = {
3856                 .call           = ctnetlink_new_expect,
3857                 .type           = NFNL_CB_MUTEX,
3858                 .attr_count     = CTA_EXPECT_MAX,
3859                 .policy         = exp_nla_policy
3860         },
3861         [IPCTNL_MSG_EXP_DELETE] = {
3862                 .call           = ctnetlink_del_expect,
3863                 .type           = NFNL_CB_MUTEX,
3864                 .attr_count     = CTA_EXPECT_MAX,
3865                 .policy         = exp_nla_policy
3866         },
3867         [IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3868                 .call           = ctnetlink_stat_exp_cpu,
3869                 .type           = NFNL_CB_MUTEX,
3870         },
3871 };
3872
3873 static const struct nfnetlink_subsystem ctnl_subsys = {
3874         .name                           = "conntrack",
3875         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3876         .cb_count                       = IPCTNL_MSG_MAX,
3877         .cb                             = ctnl_cb,
3878 };
3879
3880 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3881         .name                           = "conntrack_expect",
3882         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3883         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3884         .cb                             = ctnl_exp_cb,
3885 };
3886
3887 MODULE_ALIAS("ip_conntrack_netlink");
3888 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3889 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3890
3891 static int __net_init ctnetlink_net_init(struct net *net)
3892 {
3893 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3894         nf_conntrack_register_notifier(net, &ctnl_notifier);
3895 #endif
3896         return 0;
3897 }
3898
3899 static void ctnetlink_net_pre_exit(struct net *net)
3900 {
3901 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3902         nf_conntrack_unregister_notifier(net);
3903 #endif
3904 }
3905
3906 static struct pernet_operations ctnetlink_net_ops = {
3907         .init           = ctnetlink_net_init,
3908         .pre_exit       = ctnetlink_net_pre_exit,
3909 };
3910
3911 static int __init ctnetlink_init(void)
3912 {
3913         int ret;
3914
3915         BUILD_BUG_ON(sizeof(struct ctnetlink_list_dump_ctx) > sizeof_field(struct netlink_callback, ctx));
3916
3917         ret = nfnetlink_subsys_register(&ctnl_subsys);
3918         if (ret < 0) {
3919                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3920                 goto err_out;
3921         }
3922
3923         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3924         if (ret < 0) {
3925                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3926                 goto err_unreg_subsys;
3927         }
3928
3929         ret = register_pernet_subsys(&ctnetlink_net_ops);
3930         if (ret < 0) {
3931                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3932                 goto err_unreg_exp_subsys;
3933         }
3934 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3935         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3936         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3937 #endif
3938         return 0;
3939
3940 err_unreg_exp_subsys:
3941         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3942 err_unreg_subsys:
3943         nfnetlink_subsys_unregister(&ctnl_subsys);
3944 err_out:
3945         return ret;
3946 }
3947
3948 static void __exit ctnetlink_exit(void)
3949 {
3950         unregister_pernet_subsys(&ctnetlink_net_ops);
3951         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3952         nfnetlink_subsys_unregister(&ctnl_subsys);
3953 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3954         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3955 #endif
3956         synchronize_rcu();
3957 }
3958
3959 module_init(ctnetlink_init);
3960 module_exit(ctnetlink_exit);