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