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