net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32 #include <linux/siphash.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netlink.h>
  36 #include <net/sock.h>
  37 #include <net/netfilter/nf_conntrack.h>
  38 #include <net/netfilter/nf_conntrack_core.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_seqadj.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #include <net/netfilter/nf_conntrack_synproxy.h>
  49 #if IS_ENABLED(CONFIG_NF_NAT)
  50 #include <net/netfilter/nf_nat.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 #include "nf_internals.h"
  58
  59 MODULE_LICENSE("GPL");
  60
  61 struct ctnetlink_list_dump_ctx {
  62         struct nf_conn *last;
  63         unsigned int cpu;
  64         bool done;
  65 };
  66
  67 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  68                                 const struct nf_conntrack_tuple *tuple,
  69                                 const struct nf_conntrack_l4proto *l4proto)
  70 {
  71         int ret = 0;
  72         struct nlattr *nest_parms;
  73
  74         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO);
  75         if (!nest_parms)
  76                 goto nla_put_failure;
  77         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  78                 goto nla_put_failure;
  79
  80         if (likely(l4proto->tuple_to_nlattr))
  81                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  82
  83         nla_nest_end(skb, nest_parms);
  84
  85         return ret;
  86
  87 nla_put_failure:
  88         return -1;
  89 }
  90
  91 static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  92                                 const struct nf_conntrack_tuple *tuple)
  93 {
  94         if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  95             nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  96                 return -EMSGSIZE;
  97         return 0;
  98 }
  99
 100 static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
 101                                 const struct nf_conntrack_tuple *tuple)
 102 {
 103         if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
 104             nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
 105                 return -EMSGSIZE;
 106         return 0;
 107 }
 108
 109 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 110                                     const struct nf_conntrack_tuple *tuple)
 111 {
 112         int ret = 0;
 113         struct nlattr *nest_parms;
 114
 115         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP);
 116         if (!nest_parms)
 117                 goto nla_put_failure;
 118
 119         switch (tuple->src.l3num) {
 120         case NFPROTO_IPV4:
 121                 ret = ipv4_tuple_to_nlattr(skb, tuple);
 122                 break;
 123         case NFPROTO_IPV6:
 124                 ret = ipv6_tuple_to_nlattr(skb, tuple);
 125                 break;
 126         }
 127
 128         nla_nest_end(skb, nest_parms);
 129
 130         return ret;
 131
 132 nla_put_failure:
 133         return -1;
 134 }
 135
 136 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 137                                  const struct nf_conntrack_tuple *tuple)
 138 {
 139         const struct nf_conntrack_l4proto *l4proto;
 140         int ret;
 141
 142         rcu_read_lock();
 143         ret = ctnetlink_dump_tuples_ip(skb, tuple);
 144
 145         if (ret >= 0) {
 146                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 147                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 148         }
 149         rcu_read_unlock();
 150         return ret;
 151 }
 152
 153 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 154                                   const struct nf_conntrack_zone *zone, int dir)
 155 {
 156         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 157                 return 0;
 158         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 159                 goto nla_put_failure;
 160         return 0;
 161
 162 nla_put_failure:
 163         return -1;
 164 }
 165
 166 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 167 {
 168         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 169                 goto nla_put_failure;
 170         return 0;
 171
 172 nla_put_failure:
 173         return -1;
 174 }
 175
 176 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct,
 177                                   bool skip_zero)
 178 {
 179         long timeout = nf_ct_expires(ct) / HZ;
 180
 181         if (skip_zero && timeout == 0)
 182                 return 0;
 183
 184         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 185                 goto nla_put_failure;
 186         return 0;
 187
 188 nla_put_failure:
 189         return -1;
 190 }
 191
 192 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct,
 193                                     bool destroy)
 194 {
 195         const struct nf_conntrack_l4proto *l4proto;
 196         struct nlattr *nest_proto;
 197         int ret;
 198
 199         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 200         if (!l4proto->to_nlattr)
 201                 return 0;
 202
 203         nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 204         if (!nest_proto)
 205                 goto nla_put_failure;
 206
 207         ret = l4proto->to_nlattr(skb, nest_proto, ct, destroy);
 208
 209         nla_nest_end(skb, nest_proto);
 210
 211         return ret;
 212
 213 nla_put_failure:
 214         return -1;
 215 }
 216
 217 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 218                                    const struct nf_conn *ct)
 219 {
 220         struct nlattr *nest_helper;
 221         const struct nf_conn_help *help = nfct_help(ct);
 222         struct nf_conntrack_helper *helper;
 223
 224         if (!help)
 225                 return 0;
 226
 227         rcu_read_lock();
 228         helper = rcu_dereference(help->helper);
 229         if (!helper)
 230                 goto out;
 231
 232         nest_helper = nla_nest_start(skb, CTA_HELP);
 233         if (!nest_helper)
 234                 goto nla_put_failure;
 235         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 236                 goto nla_put_failure;
 237
 238         if (helper->to_nlattr)
 239                 helper->to_nlattr(skb, ct);
 240
 241         nla_nest_end(skb, nest_helper);
 242 out:
 243         rcu_read_unlock();
 244         return 0;
 245
 246 nla_put_failure:
 247         rcu_read_unlock();
 248         return -1;
 249 }
 250
 251 static int
 252 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 253               enum ip_conntrack_dir dir, int type)
 254 {
 255         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 256         struct nf_conn_counter *counter = acct->counter;
 257         struct nlattr *nest_count;
 258         u64 pkts, bytes;
 259
 260         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 261                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 262                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 263         } else {
 264                 pkts = atomic64_read(&counter[dir].packets);
 265                 bytes = atomic64_read(&counter[dir].bytes);
 266         }
 267
 268         nest_count = nla_nest_start(skb, attr);
 269         if (!nest_count)
 270                 goto nla_put_failure;
 271
 272         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 273                          CTA_COUNTERS_PAD) ||
 274             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 275                          CTA_COUNTERS_PAD))
 276                 goto nla_put_failure;
 277
 278         nla_nest_end(skb, nest_count);
 279
 280         return 0;
 281
 282 nla_put_failure:
 283         return -1;
 284 }
 285
 286 static int
 287 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 288 {
 289         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 290
 291         if (!acct)
 292                 return 0;
 293
 294         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 295                 return -1;
 296         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 297                 return -1;
 298
 299         return 0;
 300 }
 301
 302 static int
 303 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 304 {
 305         struct nlattr *nest_count;
 306         const struct nf_conn_tstamp *tstamp;
 307
 308         tstamp = nf_conn_tstamp_find(ct);
 309         if (!tstamp)
 310                 return 0;
 311
 312         nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 313         if (!nest_count)
 314                 goto nla_put_failure;
 315
 316         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 317                          CTA_TIMESTAMP_PAD) ||
 318             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 319                                                cpu_to_be64(tstamp->stop),
 320                                                CTA_TIMESTAMP_PAD)))
 321                 goto nla_put_failure;
 322         nla_nest_end(skb, nest_count);
 323
 324         return 0;
 325
 326 nla_put_failure:
 327         return -1;
 328 }
 329
 330 #ifdef CONFIG_NF_CONNTRACK_MARK
 331 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 332 {
 333         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 334                 goto nla_put_failure;
 335         return 0;
 336
 337 nla_put_failure:
 338         return -1;
 339 }
 340 #else
 341 #define ctnetlink_dump_mark(a, b) (0)
 342 #endif
 343
 344 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 345 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 346 {
 347         struct nlattr *nest_secctx;
 348         int len, ret;
 349         char *secctx;
 350
 351         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 352         if (ret)
 353                 return 0;
 354
 355         ret = -1;
 356         nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 357         if (!nest_secctx)
 358                 goto nla_put_failure;
 359
 360         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 361                 goto nla_put_failure;
 362         nla_nest_end(skb, nest_secctx);
 363
 364         ret = 0;
 365 nla_put_failure:
 366         security_release_secctx(secctx, len);
 367         return ret;
 368 }
 369 #else
 370 #define ctnetlink_dump_secctx(a, b) (0)
 371 #endif
 372
 373 #ifdef CONFIG_NF_CONNTRACK_LABELS
 374 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 375 {
 376         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 377
 378         if (!labels)
 379                 return 0;
 380         return nla_total_size(sizeof(labels->bits));
 381 }
 382
 383 static int
 384 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 385 {
 386         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 387         unsigned int i;
 388
 389         if (!labels)
 390                 return 0;
 391
 392         i = 0;
 393         do {
 394                 if (labels->bits[i] != 0)
 395                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 396                                        labels->bits);
 397                 i++;
 398         } while (i < ARRAY_SIZE(labels->bits));
 399
 400         return 0;
 401 }
 402 #else
 403 #define ctnetlink_dump_labels(a, b) (0)
 404 #define ctnetlink_label_size(a) (0)
 405 #endif
 406
 407 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 408
 409 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 410 {
 411         struct nlattr *nest_parms;
 412
 413         if (!(ct->status & IPS_EXPECTED))
 414                 return 0;
 415
 416         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 417         if (!nest_parms)
 418                 goto nla_put_failure;
 419         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 420                 goto nla_put_failure;
 421         nla_nest_end(skb, nest_parms);
 422
 423         return 0;
 424
 425 nla_put_failure:
 426         return -1;
 427 }
 428
 429 static int
 430 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 431 {
 432         struct nlattr *nest_parms;
 433
 434         nest_parms = nla_nest_start(skb, type);
 435         if (!nest_parms)
 436                 goto nla_put_failure;
 437
 438         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 439                          htonl(seq->correction_pos)) ||
 440             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 441                          htonl(seq->offset_before)) ||
 442             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 443                          htonl(seq->offset_after)))
 444                 goto nla_put_failure;
 445
 446         nla_nest_end(skb, nest_parms);
 447
 448         return 0;
 449
 450 nla_put_failure:
 451         return -1;
 452 }
 453
 454 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 455 {
 456         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 457         struct nf_ct_seqadj *seq;
 458
 459         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 460                 return 0;
 461
 462         spin_lock_bh(&ct->lock);
 463         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 464         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 465                 goto err;
 466
 467         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 468         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 469                 goto err;
 470
 471         spin_unlock_bh(&ct->lock);
 472         return 0;
 473 err:
 474         spin_unlock_bh(&ct->lock);
 475         return -1;
 476 }
 477
 478 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 479 {
 480         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 481         struct nlattr *nest_parms;
 482
 483         if (!synproxy)
 484                 return 0;
 485
 486         nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 487         if (!nest_parms)
 488                 goto nla_put_failure;
 489
 490         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 491             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 492             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 493                 goto nla_put_failure;
 494
 495         nla_nest_end(skb, nest_parms);
 496
 497         return 0;
 498
 499 nla_put_failure:
 500         return -1;
 501 }
 502
 503 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 504 {
 505         __be32 id = (__force __be32)nf_ct_get_id(ct);
 506
 507         if (nla_put_be32(skb, CTA_ID, id))
 508                 goto nla_put_failure;
 509         return 0;
 510
 511 nla_put_failure:
 512         return -1;
 513 }
 514
 515 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 516 {
 517         if (nla_put_be32(skb, CTA_USE, htonl(refcount_read(&ct->ct_general.use))))
 518                 goto nla_put_failure;
 519         return 0;
 520
 521 nla_put_failure:
 522         return -1;
 523 }
 524
 525 /* all these functions access ct->ext. Caller must either hold a reference
 526  * on ct or prevent its deletion by holding either the bucket spinlock or
 527  * pcpu dying list lock.
 528  */
 529 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 530                                   struct nf_conn *ct, u32 type)
 531 {
 532         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 533             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 534             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 535             ctnetlink_dump_labels(skb, ct) < 0 ||
 536             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 537             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 538                 return -1;
 539
 540         return 0;
 541 }
 542
 543 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 544 {
 545         if (ctnetlink_dump_status(skb, ct) < 0 ||
 546             ctnetlink_dump_mark(skb, ct) < 0 ||
 547             ctnetlink_dump_secctx(skb, ct) < 0 ||
 548             ctnetlink_dump_id(skb, ct) < 0 ||
 549             ctnetlink_dump_use(skb, ct) < 0 ||
 550             ctnetlink_dump_master(skb, ct) < 0)
 551                 return -1;
 552
 553         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 554             (ctnetlink_dump_timeout(skb, ct, false) < 0 ||
 555              ctnetlink_dump_protoinfo(skb, ct, false) < 0))
 556                 return -1;
 557
 558         return 0;
 559 }
 560
 561 static int
 562 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 563                     struct nf_conn *ct, bool extinfo, unsigned int flags)
 564 {
 565         const struct nf_conntrack_zone *zone;
 566         struct nlmsghdr *nlh;
 567         struct nlattr *nest_parms;
 568         unsigned int event;
 569
 570         if (portid)
 571                 flags |= NLM_F_MULTI;
 572         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 573         nlh = nfnl_msg_put(skb, portid, seq, event, flags, nf_ct_l3num(ct),
 574                            NFNETLINK_V0, 0);
 575         if (!nlh)
 576                 goto nlmsg_failure;
 577
 578         zone = nf_ct_zone(ct);
 579
 580         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 581         if (!nest_parms)
 582                 goto nla_put_failure;
 583         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 584                 goto nla_put_failure;
 585         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 586                                    NF_CT_ZONE_DIR_ORIG) < 0)
 587                 goto nla_put_failure;
 588         nla_nest_end(skb, nest_parms);
 589
 590         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 591         if (!nest_parms)
 592                 goto nla_put_failure;
 593         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 594                 goto nla_put_failure;
 595         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 596                                    NF_CT_ZONE_DIR_REPL) < 0)
 597                 goto nla_put_failure;
 598         nla_nest_end(skb, nest_parms);
 599
 600         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 601                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 602                 goto nla_put_failure;
 603
 604         if (ctnetlink_dump_info(skb, ct) < 0)
 605                 goto nla_put_failure;
 606         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 607                 goto nla_put_failure;
 608
 609         nlmsg_end(skb, nlh);
 610         return skb->len;
 611
 612 nlmsg_failure:
 613 nla_put_failure:
 614         nlmsg_cancel(skb, nlh);
 615         return -1;
 616 }
 617
 618 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 619         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 620         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 621         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 622         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 623 };
 624
 625 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 626 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 627 {
 628         const struct nf_conntrack_l4proto *l4proto;
 629         size_t len, len4 = 0;
 630
 631         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 632         len *= 3u; /* ORIG, REPLY, MASTER */
 633
 634         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 635         len += l4proto->nlattr_size;
 636         if (l4proto->nlattr_tuple_size) {
 637                 len4 = l4proto->nlattr_tuple_size();
 638                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 639         }
 640
 641         return len + len4;
 642 }
 643 #endif
 644
 645 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 646 {
 647         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 648                 return 0;
 649         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 650                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 651                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 652                ;
 653 }
 654
 655 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 656 {
 657 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 658         int len, ret;
 659
 660         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 661         if (ret)
 662                 return 0;
 663
 664         return nla_total_size(0) /* CTA_SECCTX */
 665                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 666 #else
 667         return 0;
 668 #endif
 669 }
 670
 671 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 672 {
 673 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 674         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 675                 return 0;
 676         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 677 #else
 678         return 0;
 679 #endif
 680 }
 681
 682 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 683 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 684 {
 685         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 686                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 687                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 688                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 689                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 690                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 691                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 692                + ctnetlink_acct_size(ct)
 693                + ctnetlink_timestamp_size(ct)
 694                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 695                + nla_total_size(0) /* CTA_PROTOINFO */
 696                + nla_total_size(0) /* CTA_HELP */
 697                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 698                + ctnetlink_secctx_size(ct)
 699 #if IS_ENABLED(CONFIG_NF_NAT)
 700                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 701                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 702 #endif
 703 #ifdef CONFIG_NF_CONNTRACK_MARK
 704                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 705 #endif
 706 #ifdef CONFIG_NF_CONNTRACK_ZONES
 707                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 708 #endif
 709                + ctnetlink_proto_size(ct)
 710                + ctnetlink_label_size(ct)
 711                ;
 712 }
 713
 714 static int
 715 ctnetlink_conntrack_event(unsigned int events, const struct nf_ct_event *item)
 716 {
 717         const struct nf_conntrack_zone *zone;
 718         struct net *net;
 719         struct nlmsghdr *nlh;
 720         struct nlattr *nest_parms;
 721         struct nf_conn *ct = item->ct;
 722         struct sk_buff *skb;
 723         unsigned int type;
 724         unsigned int flags = 0, group;
 725         int err;
 726
 727         if (events & (1 << IPCT_DESTROY)) {
 728                 type = IPCTNL_MSG_CT_DELETE;
 729                 group = NFNLGRP_CONNTRACK_DESTROY;
 730         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 731                 type = IPCTNL_MSG_CT_NEW;
 732                 flags = NLM_F_CREATE|NLM_F_EXCL;
 733                 group = NFNLGRP_CONNTRACK_NEW;
 734         } else if (events) {
 735                 type = IPCTNL_MSG_CT_NEW;
 736                 group = NFNLGRP_CONNTRACK_UPDATE;
 737         } else
 738                 return 0;
 739
 740         net = nf_ct_net(ct);
 741         if (!item->report && !nfnetlink_has_listeners(net, group))
 742                 return 0;
 743
 744         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 745         if (skb == NULL)
 746                 goto errout;
 747
 748         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 749         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags, nf_ct_l3num(ct),
 750                            NFNETLINK_V0, 0);
 751         if (!nlh)
 752                 goto nlmsg_failure;
 753
 754         zone = nf_ct_zone(ct);
 755
 756         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 757         if (!nest_parms)
 758                 goto nla_put_failure;
 759         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 760                 goto nla_put_failure;
 761         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 762                                    NF_CT_ZONE_DIR_ORIG) < 0)
 763                 goto nla_put_failure;
 764         nla_nest_end(skb, nest_parms);
 765
 766         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 767         if (!nest_parms)
 768                 goto nla_put_failure;
 769         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 770                 goto nla_put_failure;
 771         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 772                                    NF_CT_ZONE_DIR_REPL) < 0)
 773                 goto nla_put_failure;
 774         nla_nest_end(skb, nest_parms);
 775
 776         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 777                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 778                 goto nla_put_failure;
 779
 780         if (ctnetlink_dump_id(skb, ct) < 0)
 781                 goto nla_put_failure;
 782
 783         if (ctnetlink_dump_status(skb, ct) < 0)
 784                 goto nla_put_failure;
 785
 786         if (events & (1 << IPCT_DESTROY)) {
 787                 if (ctnetlink_dump_timeout(skb, ct, true) < 0)
 788                         goto nla_put_failure;
 789
 790                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 791                     ctnetlink_dump_timestamp(skb, ct) < 0 ||
 792                     ctnetlink_dump_protoinfo(skb, ct, true) < 0)
 793                         goto nla_put_failure;
 794         } else {
 795                 if (ctnetlink_dump_timeout(skb, ct, false) < 0)
 796                         goto nla_put_failure;
 797
 798                 if (events & (1 << IPCT_PROTOINFO) &&
 799                     ctnetlink_dump_protoinfo(skb, ct, false) < 0)
 800                         goto nla_put_failure;
 801
 802                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 803                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 804                         goto nla_put_failure;
 805
 806 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 807                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 808                     && ctnetlink_dump_secctx(skb, ct) < 0)
 809                         goto nla_put_failure;
 810 #endif
 811                 if (events & (1 << IPCT_LABEL) &&
 812                      ctnetlink_dump_labels(skb, ct) < 0)
 813                         goto nla_put_failure;
 814
 815                 if (events & (1 << IPCT_RELATED) &&
 816                     ctnetlink_dump_master(skb, ct) < 0)
 817                         goto nla_put_failure;
 818
 819                 if (events & (1 << IPCT_SEQADJ) &&
 820                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 821                         goto nla_put_failure;
 822
 823                 if (events & (1 << IPCT_SYNPROXY) &&
 824                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 825                         goto nla_put_failure;
 826         }
 827
 828 #ifdef CONFIG_NF_CONNTRACK_MARK
 829         if ((events & (1 << IPCT_MARK) || ct->mark)
 830             && ctnetlink_dump_mark(skb, ct) < 0)
 831                 goto nla_put_failure;
 832 #endif
 833         nlmsg_end(skb, nlh);
 834         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 835                              GFP_ATOMIC);
 836         if (err == -ENOBUFS || err == -EAGAIN)
 837                 return -ENOBUFS;
 838
 839         return 0;
 840
 841 nla_put_failure:
 842         nlmsg_cancel(skb, nlh);
 843 nlmsg_failure:
 844         kfree_skb(skb);
 845 errout:
 846         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 847                 return -ENOBUFS;
 848
 849         return 0;
 850 }
 851 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 852
 853 static int ctnetlink_done(struct netlink_callback *cb)
 854 {
 855         if (cb->args[1])
 856                 nf_ct_put((struct nf_conn *)cb->args[1]);
 857         kfree(cb->data);
 858         return 0;
 859 }
 860
 861 struct ctnetlink_filter_u32 {
 862         u32 val;
 863         u32 mask;
 864 };
 865
 866 struct ctnetlink_filter {
 867         u8 family;
 868
 869         u_int32_t orig_flags;
 870         u_int32_t reply_flags;
 871
 872         struct nf_conntrack_tuple orig;
 873         struct nf_conntrack_tuple reply;
 874         struct nf_conntrack_zone zone;
 875
 876         struct ctnetlink_filter_u32 mark;
 877         struct ctnetlink_filter_u32 status;
 878 };
 879
 880 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 881         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 882         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 883 };
 884
 885 static int ctnetlink_parse_filter(const struct nlattr *attr,
 886                                   struct ctnetlink_filter *filter)
 887 {
 888         struct nlattr *tb[CTA_FILTER_MAX + 1];
 889         int ret = 0;
 890
 891         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 892                                NULL);
 893         if (ret)
 894                 return ret;
 895
 896         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 897                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 898                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 899                         return -EOPNOTSUPP;
 900         }
 901
 902         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 903                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 904                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 905                         return -EOPNOTSUPP;
 906         }
 907
 908         return 0;
 909 }
 910
 911 static int ctnetlink_parse_zone(const struct nlattr *attr,
 912                                 struct nf_conntrack_zone *zone);
 913 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 914                                          struct nf_conntrack_tuple *tuple,
 915                                          u32 type, u_int8_t l3num,
 916                                          struct nf_conntrack_zone *zone,
 917                                          u_int32_t flags);
 918
 919 static int ctnetlink_filter_parse_mark(struct ctnetlink_filter_u32 *mark,
 920                                        const struct nlattr * const cda[])
 921 {
 922 #ifdef CONFIG_NF_CONNTRACK_MARK
 923         if (cda[CTA_MARK]) {
 924                 mark->val = ntohl(nla_get_be32(cda[CTA_MARK]));
 925
 926                 if (cda[CTA_MARK_MASK])
 927                         mark->mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 928                 else
 929                         mark->mask = 0xffffffff;
 930         } else if (cda[CTA_MARK_MASK]) {
 931                 return -EINVAL;
 932         }
 933 #endif
 934         return 0;
 935 }
 936
 937 static int ctnetlink_filter_parse_status(struct ctnetlink_filter_u32 *status,
 938                                          const struct nlattr * const cda[])
 939 {
 940         if (cda[CTA_STATUS]) {
 941                 status->val = ntohl(nla_get_be32(cda[CTA_STATUS]));
 942                 if (cda[CTA_STATUS_MASK])
 943                         status->mask = ntohl(nla_get_be32(cda[CTA_STATUS_MASK]));
 944                 else
 945                         status->mask = status->val;
 946
 947                 /* status->val == 0? always true, else always false. */
 948                 if (status->mask == 0)
 949                         return -EINVAL;
 950         } else if (cda[CTA_STATUS_MASK]) {
 951                 return -EINVAL;
 952         }
 953
 954         /* CTA_STATUS is NLA_U32, if this fires UAPI needs to be extended */
 955         BUILD_BUG_ON(__IPS_MAX_BIT >= 32);
 956         return 0;
 957 }
 958
 959 static struct ctnetlink_filter *
 960 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 961 {
 962         struct ctnetlink_filter *filter;
 963         int err;
 964
 965 #ifndef CONFIG_NF_CONNTRACK_MARK
 966         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 967                 return ERR_PTR(-EOPNOTSUPP);
 968 #endif
 969
 970         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 971         if (filter == NULL)
 972                 return ERR_PTR(-ENOMEM);
 973
 974         filter->family = family;
 975
 976         err = ctnetlink_filter_parse_mark(&filter->mark, cda);
 977         if (err)
 978                 goto err_filter;
 979
 980         err = ctnetlink_filter_parse_status(&filter->status, cda);
 981         if (err)
 982                 goto err_filter;
 983
 984         if (!cda[CTA_FILTER])
 985                 return filter;
 986
 987         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 988         if (err < 0)
 989                 goto err_filter;
 990
 991         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
 992         if (err < 0)
 993                 goto err_filter;
 994
 995         if (filter->orig_flags) {
 996                 if (!cda[CTA_TUPLE_ORIG]) {
 997                         err = -EINVAL;
 998                         goto err_filter;
 999                 }
1000
1001                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
1002                                                    CTA_TUPLE_ORIG,
1003                                                    filter->family,
1004                                                    &filter->zone,
1005                                                    filter->orig_flags);
1006                 if (err < 0)
1007                         goto err_filter;
1008         }
1009
1010         if (filter->reply_flags) {
1011                 if (!cda[CTA_TUPLE_REPLY]) {
1012                         err = -EINVAL;
1013                         goto err_filter;
1014                 }
1015
1016                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
1017                                                    CTA_TUPLE_REPLY,
1018                                                    filter->family,
1019                                                    &filter->zone,
1020                                                    filter->reply_flags);
1021                 if (err < 0)
1022                         goto err_filter;
1023         }
1024
1025         return filter;
1026
1027 err_filter:
1028         kfree(filter);
1029
1030         return ERR_PTR(err);
1031 }
1032
1033 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1034 {
1035         return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS];
1036 }
1037
1038 static int ctnetlink_start(struct netlink_callback *cb)
1039 {
1040         const struct nlattr * const *cda = cb->data;
1041         struct ctnetlink_filter *filter = NULL;
1042         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1043         u8 family = nfmsg->nfgen_family;
1044
1045         if (ctnetlink_needs_filter(family, cda)) {
1046                 filter = ctnetlink_alloc_filter(cda, family);
1047                 if (IS_ERR(filter))
1048                         return PTR_ERR(filter);
1049         }
1050
1051         cb->data = filter;
1052         return 0;
1053 }
1054
1055 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1056                                         struct nf_conntrack_tuple *ct_tuple,
1057                                         u_int32_t flags, int family)
1058 {
1059         switch (family) {
1060         case NFPROTO_IPV4:
1061                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1062                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1063                         return  0;
1064
1065                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1066                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1067                         return  0;
1068                 break;
1069         case NFPROTO_IPV6:
1070                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1071                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1072                                    &ct_tuple->src.u3.in6))
1073                         return 0;
1074
1075                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1076                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1077                                    &ct_tuple->dst.u3.in6))
1078                         return 0;
1079                 break;
1080         }
1081
1082         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1083             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1084                 return 0;
1085
1086         switch (ct_tuple->dst.protonum) {
1087         case IPPROTO_TCP:
1088         case IPPROTO_UDP:
1089                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1090                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1091                         return 0;
1092
1093                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1094                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1095                         return 0;
1096                 break;
1097         case IPPROTO_ICMP:
1098                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1099                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1100                         return 0;
1101                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1102                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1103                         return 0;
1104                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1105                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1106                         return 0;
1107                 break;
1108         case IPPROTO_ICMPV6:
1109                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1110                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1111                         return 0;
1112                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1113                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1114                         return 0;
1115                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1116                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1117                         return 0;
1118                 break;
1119         }
1120
1121         return 1;
1122 }
1123
1124 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1125 {
1126         struct ctnetlink_filter *filter = data;
1127         struct nf_conntrack_tuple *tuple;
1128         u32 status;
1129
1130         if (filter == NULL)
1131                 goto out;
1132
1133         /* Match entries of a given L3 protocol number.
1134          * If it is not specified, ie. l3proto == 0,
1135          * then match everything.
1136          */
1137         if (filter->family && nf_ct_l3num(ct) != filter->family)
1138                 goto ignore_entry;
1139
1140         if (filter->orig_flags) {
1141                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1142                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1143                                                   filter->orig_flags,
1144                                                   filter->family))
1145                         goto ignore_entry;
1146         }
1147
1148         if (filter->reply_flags) {
1149                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1150                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1151                                                   filter->reply_flags,
1152                                                   filter->family))
1153                         goto ignore_entry;
1154         }
1155
1156 #ifdef CONFIG_NF_CONNTRACK_MARK
1157         if ((ct->mark & filter->mark.mask) != filter->mark.val)
1158                 goto ignore_entry;
1159 #endif
1160         status = (u32)READ_ONCE(ct->status);
1161         if ((status & filter->status.mask) != filter->status.val)
1162                 goto ignore_entry;
1163
1164 out:
1165         return 1;
1166
1167 ignore_entry:
1168         return 0;
1169 }
1170
1171 static int
1172 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1173 {
1174         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1175         struct net *net = sock_net(skb->sk);
1176         struct nf_conn *ct, *last;
1177         struct nf_conntrack_tuple_hash *h;
1178         struct hlist_nulls_node *n;
1179         struct nf_conn *nf_ct_evict[8];
1180         int res, i;
1181         spinlock_t *lockp;
1182
1183         last = (struct nf_conn *)cb->args[1];
1184         i = 0;
1185
1186         local_bh_disable();
1187         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1188 restart:
1189                 while (i) {
1190                         i--;
1191                         if (nf_ct_should_gc(nf_ct_evict[i]))
1192                                 nf_ct_kill(nf_ct_evict[i]);
1193                         nf_ct_put(nf_ct_evict[i]);
1194                 }
1195
1196                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1197                 nf_conntrack_lock(lockp);
1198                 if (cb->args[0] >= nf_conntrack_htable_size) {
1199                         spin_unlock(lockp);
1200                         goto out;
1201                 }
1202                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1203                                            hnnode) {
1204                         ct = nf_ct_tuplehash_to_ctrack(h);
1205                         if (nf_ct_is_expired(ct)) {
1206                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1207                                     refcount_inc_not_zero(&ct->ct_general.use))
1208                                         nf_ct_evict[i++] = ct;
1209                                 continue;
1210                         }
1211
1212                         if (!net_eq(net, nf_ct_net(ct)))
1213                                 continue;
1214
1215                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1216                                 continue;
1217
1218                         if (cb->args[1]) {
1219                                 if (ct != last)
1220                                         continue;
1221                                 cb->args[1] = 0;
1222                         }
1223                         if (!ctnetlink_filter_match(ct, cb->data))
1224                                 continue;
1225
1226                         res =
1227                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1228                                             cb->nlh->nlmsg_seq,
1229                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1230                                             ct, true, flags);
1231                         if (res < 0) {
1232                                 nf_conntrack_get(&ct->ct_general);
1233                                 cb->args[1] = (unsigned long)ct;
1234                                 spin_unlock(lockp);
1235                                 goto out;
1236                         }
1237                 }
1238                 spin_unlock(lockp);
1239                 if (cb->args[1]) {
1240                         cb->args[1] = 0;
1241                         goto restart;
1242                 }
1243         }
1244 out:
1245         local_bh_enable();
1246         if (last) {
1247                 /* nf ct hash resize happened, now clear the leftover. */
1248                 if ((struct nf_conn *)cb->args[1] == last)
1249                         cb->args[1] = 0;
1250
1251                 nf_ct_put(last);
1252         }
1253
1254         while (i) {
1255                 i--;
1256                 if (nf_ct_should_gc(nf_ct_evict[i]))
1257                         nf_ct_kill(nf_ct_evict[i]);
1258                 nf_ct_put(nf_ct_evict[i]);
1259         }
1260
1261         return skb->len;
1262 }
1263
1264 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1265                                 struct nf_conntrack_tuple *t,
1266                                 u_int32_t flags)
1267 {
1268         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1269                 if (!tb[CTA_IP_V4_SRC])
1270                         return -EINVAL;
1271
1272                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1273         }
1274
1275         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1276                 if (!tb[CTA_IP_V4_DST])
1277                         return -EINVAL;
1278
1279                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1280         }
1281
1282         return 0;
1283 }
1284
1285 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1286                                 struct nf_conntrack_tuple *t,
1287                                 u_int32_t flags)
1288 {
1289         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1290                 if (!tb[CTA_IP_V6_SRC])
1291                         return -EINVAL;
1292
1293                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1294         }
1295
1296         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1297                 if (!tb[CTA_IP_V6_DST])
1298                         return -EINVAL;
1299
1300                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1301         }
1302
1303         return 0;
1304 }
1305
1306 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1307                                     struct nf_conntrack_tuple *tuple,
1308                                     u_int32_t flags)
1309 {
1310         struct nlattr *tb[CTA_IP_MAX+1];
1311         int ret = 0;
1312
1313         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1314         if (ret < 0)
1315                 return ret;
1316
1317         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1318                                              cta_ip_nla_policy, NULL);
1319         if (ret)
1320                 return ret;
1321
1322         switch (tuple->src.l3num) {
1323         case NFPROTO_IPV4:
1324                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1325                 break;
1326         case NFPROTO_IPV6:
1327                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1328                 break;
1329         }
1330
1331         return ret;
1332 }
1333
1334 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1335         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1336 };
1337
1338 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1339                                        struct nf_conntrack_tuple *tuple,
1340                                        u_int32_t flags)
1341 {
1342         const struct nf_conntrack_l4proto *l4proto;
1343         struct nlattr *tb[CTA_PROTO_MAX+1];
1344         int ret = 0;
1345
1346         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1347                                           proto_nla_policy, NULL);
1348         if (ret < 0)
1349                 return ret;
1350
1351         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1352                 return 0;
1353
1354         if (!tb[CTA_PROTO_NUM])
1355                 return -EINVAL;
1356
1357         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1358
1359         rcu_read_lock();
1360         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1361
1362         if (likely(l4proto->nlattr_to_tuple)) {
1363                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1364                                                      l4proto->nla_policy,
1365                                                      NULL);
1366                 if (ret == 0)
1367                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1368         }
1369
1370         rcu_read_unlock();
1371
1372         return ret;
1373 }
1374
1375 static int
1376 ctnetlink_parse_zone(const struct nlattr *attr,
1377                      struct nf_conntrack_zone *zone)
1378 {
1379         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1380                         NF_CT_DEFAULT_ZONE_DIR, 0);
1381 #ifdef CONFIG_NF_CONNTRACK_ZONES
1382         if (attr)
1383                 zone->id = ntohs(nla_get_be16(attr));
1384 #else
1385         if (attr)
1386                 return -EOPNOTSUPP;
1387 #endif
1388         return 0;
1389 }
1390
1391 static int
1392 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1393                            struct nf_conntrack_zone *zone)
1394 {
1395         int ret;
1396
1397         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1398                 return -EINVAL;
1399
1400         ret = ctnetlink_parse_zone(attr, zone);
1401         if (ret < 0)
1402                 return ret;
1403
1404         if (type == CTA_TUPLE_REPLY)
1405                 zone->dir = NF_CT_ZONE_DIR_REPL;
1406         else
1407                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1408
1409         return 0;
1410 }
1411
1412 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1413         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1414         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1415         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1416 };
1417
1418 #define CTA_FILTER_F_ALL_CTA_PROTO \
1419   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1420    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1421    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1422    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1423    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1424    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1425    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1426    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1427
1428 static int
1429 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1430                               struct nf_conntrack_tuple *tuple, u32 type,
1431                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1432                               u_int32_t flags)
1433 {
1434         struct nlattr *tb[CTA_TUPLE_MAX+1];
1435         int err;
1436
1437         memset(tuple, 0, sizeof(*tuple));
1438
1439         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1440                                           tuple_nla_policy, NULL);
1441         if (err < 0)
1442                 return err;
1443
1444         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1445                 return -EOPNOTSUPP;
1446         tuple->src.l3num = l3num;
1447
1448         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1449             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1450                 if (!tb[CTA_TUPLE_IP])
1451                         return -EINVAL;
1452
1453                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1454                 if (err < 0)
1455                         return err;
1456         }
1457
1458         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1459                 if (!tb[CTA_TUPLE_PROTO])
1460                         return -EINVAL;
1461
1462                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1463                 if (err < 0)
1464                         return err;
1465         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1466                 /* Can't manage proto flags without a protonum  */
1467                 return -EINVAL;
1468         }
1469
1470         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1471                 if (!zone)
1472                         return -EINVAL;
1473
1474                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1475                                                  type, zone);
1476                 if (err < 0)
1477                         return err;
1478         }
1479
1480         /* orig and expect tuples get DIR_ORIGINAL */
1481         if (type == CTA_TUPLE_REPLY)
1482                 tuple->dst.dir = IP_CT_DIR_REPLY;
1483         else
1484                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1485
1486         return 0;
1487 }
1488
1489 static int
1490 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1491                       struct nf_conntrack_tuple *tuple, u32 type,
1492                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1493 {
1494         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1495                                             CTA_FILTER_FLAG(ALL));
1496 }
1497
1498 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1499         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1500                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1501 };
1502
1503 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1504                                 struct nlattr **helpinfo)
1505 {
1506         int err;
1507         struct nlattr *tb[CTA_HELP_MAX+1];
1508
1509         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1510                                           help_nla_policy, NULL);
1511         if (err < 0)
1512                 return err;
1513
1514         if (!tb[CTA_HELP_NAME])
1515                 return -EINVAL;
1516
1517         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1518
1519         if (tb[CTA_HELP_INFO])
1520                 *helpinfo = tb[CTA_HELP_INFO];
1521
1522         return 0;
1523 }
1524
1525 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1526         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1527         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1528         [CTA_STATUS]            = { .type = NLA_U32 },
1529         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1530         [CTA_HELP]              = { .type = NLA_NESTED },
1531         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1532         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1533         [CTA_MARK]              = { .type = NLA_U32 },
1534         [CTA_ID]                = { .type = NLA_U32 },
1535         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1536         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1537         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1538         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1539         [CTA_ZONE]              = { .type = NLA_U16 },
1540         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1541         [CTA_LABELS]            = { .type = NLA_BINARY,
1542                                     .len = NF_CT_LABELS_MAX_SIZE },
1543         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1544                                     .len = NF_CT_LABELS_MAX_SIZE },
1545         [CTA_FILTER]            = { .type = NLA_NESTED },
1546         [CTA_STATUS_MASK]       = { .type = NLA_U32 },
1547 };
1548
1549 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1550 {
1551         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1552                 return 0;
1553
1554         return ctnetlink_filter_match(ct, data);
1555 }
1556
1557 static int ctnetlink_flush_conntrack(struct net *net,
1558                                      const struct nlattr * const cda[],
1559                                      u32 portid, int report, u8 family)
1560 {
1561         struct ctnetlink_filter *filter = NULL;
1562         struct nf_ct_iter_data iter = {
1563                 .net            = net,
1564                 .portid         = portid,
1565                 .report         = report,
1566         };
1567
1568         if (ctnetlink_needs_filter(family, cda)) {
1569                 if (cda[CTA_FILTER])
1570                         return -EOPNOTSUPP;
1571
1572                 filter = ctnetlink_alloc_filter(cda, family);
1573                 if (IS_ERR(filter))
1574                         return PTR_ERR(filter);
1575
1576                 iter.data = filter;
1577         }
1578
1579         nf_ct_iterate_cleanup_net(ctnetlink_flush_iterate, &iter);
1580         kfree(filter);
1581
1582         return 0;
1583 }
1584
1585 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1586                                    const struct nfnl_info *info,
1587                                    const struct nlattr * const cda[])
1588 {
1589         u8 family = info->nfmsg->nfgen_family;
1590         struct nf_conntrack_tuple_hash *h;
1591         struct nf_conntrack_tuple tuple;
1592         struct nf_conntrack_zone zone;
1593         struct nf_conn *ct;
1594         int err;
1595
1596         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1597         if (err < 0)
1598                 return err;
1599
1600         if (cda[CTA_TUPLE_ORIG])
1601                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1602                                             family, &zone);
1603         else if (cda[CTA_TUPLE_REPLY])
1604                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1605                                             family, &zone);
1606         else {
1607                 u_int8_t u3 = info->nfmsg->version ? family : AF_UNSPEC;
1608
1609                 return ctnetlink_flush_conntrack(info->net, cda,
1610                                                  NETLINK_CB(skb).portid,
1611                                                  nlmsg_report(info->nlh), u3);
1612         }
1613
1614         if (err < 0)
1615                 return err;
1616
1617         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1618         if (!h)
1619                 return -ENOENT;
1620
1621         ct = nf_ct_tuplehash_to_ctrack(h);
1622
1623         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1624                 nf_ct_put(ct);
1625                 return -EBUSY;
1626         }
1627
1628         if (cda[CTA_ID]) {
1629                 __be32 id = nla_get_be32(cda[CTA_ID]);
1630
1631                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1632                         nf_ct_put(ct);
1633                         return -ENOENT;
1634                 }
1635         }
1636
1637         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1638         nf_ct_put(ct);
1639
1640         return 0;
1641 }
1642
1643 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1644                                    const struct nfnl_info *info,
1645                                    const struct nlattr * const cda[])
1646 {
1647         u_int8_t u3 = info->nfmsg->nfgen_family;
1648         struct nf_conntrack_tuple_hash *h;
1649         struct nf_conntrack_tuple tuple;
1650         struct nf_conntrack_zone zone;
1651         struct sk_buff *skb2;
1652         struct nf_conn *ct;
1653         int err;
1654
1655         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1656                 struct netlink_dump_control c = {
1657                         .start = ctnetlink_start,
1658                         .dump = ctnetlink_dump_table,
1659                         .done = ctnetlink_done,
1660                         .data = (void *)cda,
1661                 };
1662
1663                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1664         }
1665
1666         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1667         if (err < 0)
1668                 return err;
1669
1670         if (cda[CTA_TUPLE_ORIG])
1671                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1672                                             u3, &zone);
1673         else if (cda[CTA_TUPLE_REPLY])
1674                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1675                                             u3, &zone);
1676         else
1677                 return -EINVAL;
1678
1679         if (err < 0)
1680                 return err;
1681
1682         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1683         if (!h)
1684                 return -ENOENT;
1685
1686         ct = nf_ct_tuplehash_to_ctrack(h);
1687
1688         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1689         if (!skb2) {
1690                 nf_ct_put(ct);
1691                 return -ENOMEM;
1692         }
1693
1694         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1695                                   info->nlh->nlmsg_seq,
1696                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1697                                   true, 0);
1698         nf_ct_put(ct);
1699         if (err <= 0) {
1700                 kfree_skb(skb2);
1701                 return -ENOMEM;
1702         }
1703
1704         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1705 }
1706
1707 static int ctnetlink_done_list(struct netlink_callback *cb)
1708 {
1709         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1710
1711         if (ctx->last)
1712                 nf_ct_put(ctx->last);
1713
1714         return 0;
1715 }
1716
1717 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1718 static int ctnetlink_dump_one_entry(struct sk_buff *skb,
1719                                     struct netlink_callback *cb,
1720                                     struct nf_conn *ct,
1721                                     bool dying)
1722 {
1723         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1724         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1725         u8 l3proto = nfmsg->nfgen_family;
1726         int res;
1727
1728         if (l3proto && nf_ct_l3num(ct) != l3proto)
1729                 return 0;
1730
1731         if (ctx->last) {
1732                 if (ct != ctx->last)
1733                         return 0;
1734
1735                 ctx->last = NULL;
1736         }
1737
1738         /* We can't dump extension info for the unconfirmed
1739          * list because unconfirmed conntracks can have
1740          * ct->ext reallocated (and thus freed).
1741          *
1742          * In the dying list case ct->ext can't be free'd
1743          * until after we drop pcpu->lock.
1744          */
1745         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1746                                   cb->nlh->nlmsg_seq,
1747                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1748                                   ct, dying, 0);
1749         if (res < 0) {
1750                 if (!refcount_inc_not_zero(&ct->ct_general.use))
1751                         return 0;
1752
1753                 ctx->last = ct;
1754         }
1755
1756         return res;
1757 }
1758 #endif
1759
1760 static int
1761 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1762 {
1763         return 0;
1764 }
1765
1766 static int
1767 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1768 {
1769         struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1770         struct nf_conn *last = ctx->last;
1771 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1772         const struct net *net = sock_net(skb->sk);
1773         struct nf_conntrack_net_ecache *ecache_net;
1774         struct nf_conntrack_tuple_hash *h;
1775         struct hlist_nulls_node *n;
1776 #endif
1777
1778         if (ctx->done)
1779                 return 0;
1780
1781         ctx->last = NULL;
1782
1783 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1784         ecache_net = nf_conn_pernet_ecache(net);
1785         spin_lock_bh(&ecache_net->dying_lock);
1786
1787         hlist_nulls_for_each_entry(h, n, &ecache_net->dying_list, hnnode) {
1788                 struct nf_conn *ct;
1789                 int res;
1790
1791                 ct = nf_ct_tuplehash_to_ctrack(h);
1792                 if (last && last != ct)
1793                         continue;
1794
1795                 res = ctnetlink_dump_one_entry(skb, cb, ct, true);
1796                 if (res < 0) {
1797                         spin_unlock_bh(&ecache_net->dying_lock);
1798                         nf_ct_put(last);
1799                         return skb->len;
1800                 }
1801
1802                 nf_ct_put(last);
1803                 last = NULL;
1804         }
1805
1806         spin_unlock_bh(&ecache_net->dying_lock);
1807 #endif
1808         ctx->done = true;
1809         nf_ct_put(last);
1810
1811         return skb->len;
1812 }
1813
1814 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1815                                   const struct nfnl_info *info,
1816                                   const struct nlattr * const cda[])
1817 {
1818         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1819                 struct netlink_dump_control c = {
1820                         .dump = ctnetlink_dump_dying,
1821                         .done = ctnetlink_done_list,
1822                 };
1823                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1824         }
1825
1826         return -EOPNOTSUPP;
1827 }
1828
1829 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1830                                         const struct nfnl_info *info,
1831                                         const struct nlattr * const cda[])
1832 {
1833         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1834                 struct netlink_dump_control c = {
1835                         .dump = ctnetlink_dump_unconfirmed,
1836                         .done = ctnetlink_done_list,
1837                 };
1838                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1839         }
1840
1841         return -EOPNOTSUPP;
1842 }
1843
1844 #if IS_ENABLED(CONFIG_NF_NAT)
1845 static int
1846 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1847                           enum nf_nat_manip_type manip,
1848                           const struct nlattr *attr)
1849         __must_hold(RCU)
1850 {
1851         const struct nf_nat_hook *nat_hook;
1852         int err;
1853
1854         nat_hook = rcu_dereference(nf_nat_hook);
1855         if (!nat_hook) {
1856 #ifdef CONFIG_MODULES
1857                 rcu_read_unlock();
1858                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1859                 if (request_module("nf-nat") < 0) {
1860                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1861                         rcu_read_lock();
1862                         return -EOPNOTSUPP;
1863                 }
1864                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1865                 rcu_read_lock();
1866                 nat_hook = rcu_dereference(nf_nat_hook);
1867                 if (nat_hook)
1868                         return -EAGAIN;
1869 #endif
1870                 return -EOPNOTSUPP;
1871         }
1872
1873         err = nat_hook->parse_nat_setup(ct, manip, attr);
1874         if (err == -EAGAIN) {
1875 #ifdef CONFIG_MODULES
1876                 rcu_read_unlock();
1877                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1878                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1879                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1880                         rcu_read_lock();
1881                         return -EOPNOTSUPP;
1882                 }
1883                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1884                 rcu_read_lock();
1885 #else
1886                 err = -EOPNOTSUPP;
1887 #endif
1888         }
1889         return err;
1890 }
1891 #endif
1892
1893 static int
1894 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1895 {
1896         return nf_ct_change_status_common(ct, ntohl(nla_get_be32(cda[CTA_STATUS])));
1897 }
1898
1899 static int
1900 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1901 {
1902 #if IS_ENABLED(CONFIG_NF_NAT)
1903         int ret;
1904
1905         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1906                 return 0;
1907
1908         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1909                                         cda[CTA_NAT_DST]);
1910         if (ret < 0)
1911                 return ret;
1912
1913         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1914                                          cda[CTA_NAT_SRC]);
1915 #else
1916         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1917                 return 0;
1918         return -EOPNOTSUPP;
1919 #endif
1920 }
1921
1922 static int ctnetlink_change_helper(struct nf_conn *ct,
1923                                    const struct nlattr * const cda[])
1924 {
1925         struct nf_conntrack_helper *helper;
1926         struct nf_conn_help *help = nfct_help(ct);
1927         char *helpname = NULL;
1928         struct nlattr *helpinfo = NULL;
1929         int err;
1930
1931         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1932         if (err < 0)
1933                 return err;
1934
1935         /* don't change helper of sibling connections */
1936         if (ct->master) {
1937                 /* If we try to change the helper to the same thing twice,
1938                  * treat the second attempt as a no-op instead of returning
1939                  * an error.
1940                  */
1941                 err = -EBUSY;
1942                 if (help) {
1943                         rcu_read_lock();
1944                         helper = rcu_dereference(help->helper);
1945                         if (helper && !strcmp(helper->name, helpname))
1946                                 err = 0;
1947                         rcu_read_unlock();
1948                 }
1949
1950                 return err;
1951         }
1952
1953         if (!strcmp(helpname, "")) {
1954                 if (help && help->helper) {
1955                         /* we had a helper before ... */
1956                         nf_ct_remove_expectations(ct);
1957                         RCU_INIT_POINTER(help->helper, NULL);
1958                 }
1959
1960                 return 0;
1961         }
1962
1963         rcu_read_lock();
1964         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1965                                             nf_ct_protonum(ct));
1966         if (helper == NULL) {
1967                 rcu_read_unlock();
1968                 return -EOPNOTSUPP;
1969         }
1970
1971         if (help) {
1972                 if (help->helper == helper) {
1973                         /* update private helper data if allowed. */
1974                         if (helper->from_nlattr)
1975                                 helper->from_nlattr(helpinfo, ct);
1976                         err = 0;
1977                 } else
1978                         err = -EBUSY;
1979         } else {
1980                 /* we cannot set a helper for an existing conntrack */
1981                 err = -EOPNOTSUPP;
1982         }
1983
1984         rcu_read_unlock();
1985         return err;
1986 }
1987
1988 static int ctnetlink_change_timeout(struct nf_conn *ct,
1989                                     const struct nlattr * const cda[])
1990 {
1991         return __nf_ct_change_timeout(ct, (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ);
1992 }
1993
1994 #if defined(CONFIG_NF_CONNTRACK_MARK)
1995 static void ctnetlink_change_mark(struct nf_conn *ct,
1996                                     const struct nlattr * const cda[])
1997 {
1998         u32 mark, newmark, mask = 0;
1999
2000         if (cda[CTA_MARK_MASK])
2001                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2002
2003         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2004         newmark = (ct->mark & mask) ^ mark;
2005         if (newmark != ct->mark)
2006                 ct->mark = newmark;
2007 }
2008 #endif
2009
2010 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2011         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2012         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2013         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2014 };
2015
2016 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2017                                       const struct nlattr * const cda[])
2018 {
2019         const struct nlattr *attr = cda[CTA_PROTOINFO];
2020         const struct nf_conntrack_l4proto *l4proto;
2021         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2022         int err = 0;
2023
2024         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2025                                           protoinfo_policy, NULL);
2026         if (err < 0)
2027                 return err;
2028
2029         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2030         if (l4proto->from_nlattr)
2031                 err = l4proto->from_nlattr(tb, ct);
2032
2033         return err;
2034 }
2035
2036 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2037         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2038         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2039         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2040 };
2041
2042 static int change_seq_adj(struct nf_ct_seqadj *seq,
2043                           const struct nlattr * const attr)
2044 {
2045         int err;
2046         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2047
2048         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2049                                           seqadj_policy, NULL);
2050         if (err < 0)
2051                 return err;
2052
2053         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2054                 return -EINVAL;
2055
2056         seq->correction_pos =
2057                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2058
2059         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2060                 return -EINVAL;
2061
2062         seq->offset_before =
2063                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2064
2065         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2066                 return -EINVAL;
2067
2068         seq->offset_after =
2069                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2070
2071         return 0;
2072 }
2073
2074 static int
2075 ctnetlink_change_seq_adj(struct nf_conn *ct,
2076                          const struct nlattr * const cda[])
2077 {
2078         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2079         int ret = 0;
2080
2081         if (!seqadj)
2082                 return 0;
2083
2084         spin_lock_bh(&ct->lock);
2085         if (cda[CTA_SEQ_ADJ_ORIG]) {
2086                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2087                                      cda[CTA_SEQ_ADJ_ORIG]);
2088                 if (ret < 0)
2089                         goto err;
2090
2091                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2092         }
2093
2094         if (cda[CTA_SEQ_ADJ_REPLY]) {
2095                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2096                                      cda[CTA_SEQ_ADJ_REPLY]);
2097                 if (ret < 0)
2098                         goto err;
2099
2100                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2101         }
2102
2103         spin_unlock_bh(&ct->lock);
2104         return 0;
2105 err:
2106         spin_unlock_bh(&ct->lock);
2107         return ret;
2108 }
2109
2110 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2111         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2112         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2113         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2114 };
2115
2116 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2117                                      const struct nlattr * const cda[])
2118 {
2119         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2120         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2121         int err;
2122
2123         if (!synproxy)
2124                 return 0;
2125
2126         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2127                                           cda[CTA_SYNPROXY], synproxy_policy,
2128                                           NULL);
2129         if (err < 0)
2130                 return err;
2131
2132         if (!tb[CTA_SYNPROXY_ISN] ||
2133             !tb[CTA_SYNPROXY_ITS] ||
2134             !tb[CTA_SYNPROXY_TSOFF])
2135                 return -EINVAL;
2136
2137         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2138         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2139         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2140
2141         return 0;
2142 }
2143
2144 static int
2145 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2146 {
2147 #ifdef CONFIG_NF_CONNTRACK_LABELS
2148         size_t len = nla_len(cda[CTA_LABELS]);
2149         const void *mask = cda[CTA_LABELS_MASK];
2150
2151         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2152                 return -EINVAL;
2153
2154         if (mask) {
2155                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2156                     nla_len(cda[CTA_LABELS_MASK]) != len)
2157                         return -EINVAL;
2158                 mask = nla_data(cda[CTA_LABELS_MASK]);
2159         }
2160
2161         len /= sizeof(u32);
2162
2163         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2164 #else
2165         return -EOPNOTSUPP;
2166 #endif
2167 }
2168
2169 static int
2170 ctnetlink_change_conntrack(struct nf_conn *ct,
2171                            const struct nlattr * const cda[])
2172 {
2173         int err;
2174
2175         /* only allow NAT changes and master assignation for new conntracks */
2176         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2177                 return -EOPNOTSUPP;
2178
2179         if (cda[CTA_HELP]) {
2180                 err = ctnetlink_change_helper(ct, cda);
2181                 if (err < 0)
2182                         return err;
2183         }
2184
2185         if (cda[CTA_TIMEOUT]) {
2186                 err = ctnetlink_change_timeout(ct, cda);
2187                 if (err < 0)
2188                         return err;
2189         }
2190
2191         if (cda[CTA_STATUS]) {
2192                 err = ctnetlink_change_status(ct, cda);
2193                 if (err < 0)
2194                         return err;
2195         }
2196
2197         if (cda[CTA_PROTOINFO]) {
2198                 err = ctnetlink_change_protoinfo(ct, cda);
2199                 if (err < 0)
2200                         return err;
2201         }
2202
2203 #if defined(CONFIG_NF_CONNTRACK_MARK)
2204         if (cda[CTA_MARK])
2205                 ctnetlink_change_mark(ct, cda);
2206 #endif
2207
2208         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2209                 err = ctnetlink_change_seq_adj(ct, cda);
2210                 if (err < 0)
2211                         return err;
2212         }
2213
2214         if (cda[CTA_SYNPROXY]) {
2215                 err = ctnetlink_change_synproxy(ct, cda);
2216                 if (err < 0)
2217                         return err;
2218         }
2219
2220         if (cda[CTA_LABELS]) {
2221                 err = ctnetlink_attach_labels(ct, cda);
2222                 if (err < 0)
2223                         return err;
2224         }
2225
2226         return 0;
2227 }
2228
2229 static struct nf_conn *
2230 ctnetlink_create_conntrack(struct net *net,
2231                            const struct nf_conntrack_zone *zone,
2232                            const struct nlattr * const cda[],
2233                            struct nf_conntrack_tuple *otuple,
2234                            struct nf_conntrack_tuple *rtuple,
2235                            u8 u3)
2236 {
2237         struct nf_conn *ct;
2238         int err = -EINVAL;
2239         struct nf_conntrack_helper *helper;
2240         struct nf_conn_tstamp *tstamp;
2241         u64 timeout;
2242
2243         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2244         if (IS_ERR(ct))
2245                 return ERR_PTR(-ENOMEM);
2246
2247         if (!cda[CTA_TIMEOUT])
2248                 goto err1;
2249
2250         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2251         __nf_ct_set_timeout(ct, timeout);
2252
2253         rcu_read_lock();
2254         if (cda[CTA_HELP]) {
2255                 char *helpname = NULL;
2256                 struct nlattr *helpinfo = NULL;
2257
2258                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2259                 if (err < 0)
2260                         goto err2;
2261
2262                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2263                                                     nf_ct_protonum(ct));
2264                 if (helper == NULL) {
2265                         rcu_read_unlock();
2266 #ifdef CONFIG_MODULES
2267                         if (request_module("nfct-helper-%s", helpname) < 0) {
2268                                 err = -EOPNOTSUPP;
2269                                 goto err1;
2270                         }
2271
2272                         rcu_read_lock();
2273                         helper = __nf_conntrack_helper_find(helpname,
2274                                                             nf_ct_l3num(ct),
2275                                                             nf_ct_protonum(ct));
2276                         if (helper) {
2277                                 err = -EAGAIN;
2278                                 goto err2;
2279                         }
2280                         rcu_read_unlock();
2281 #endif
2282                         err = -EOPNOTSUPP;
2283                         goto err1;
2284                 } else {
2285                         struct nf_conn_help *help;
2286
2287                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2288                         if (help == NULL) {
2289                                 err = -ENOMEM;
2290                                 goto err2;
2291                         }
2292                         /* set private helper data if allowed. */
2293                         if (helper->from_nlattr)
2294                                 helper->from_nlattr(helpinfo, ct);
2295
2296                         /* disable helper auto-assignment for this entry */
2297                         ct->status |= IPS_HELPER;
2298                         RCU_INIT_POINTER(help->helper, helper);
2299                 }
2300         } else {
2301                 /* try an implicit helper assignation */
2302                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2303                 if (err < 0)
2304                         goto err2;
2305         }
2306
2307         err = ctnetlink_setup_nat(ct, cda);
2308         if (err < 0)
2309                 goto err2;
2310
2311         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2312         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2313         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2314         nf_ct_labels_ext_add(ct);
2315         nfct_seqadj_ext_add(ct);
2316         nfct_synproxy_ext_add(ct);
2317
2318         /* we must add conntrack extensions before confirmation. */
2319         ct->status |= IPS_CONFIRMED;
2320
2321         if (cda[CTA_STATUS]) {
2322                 err = ctnetlink_change_status(ct, cda);
2323                 if (err < 0)
2324                         goto err2;
2325         }
2326
2327         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2328                 err = ctnetlink_change_seq_adj(ct, cda);
2329                 if (err < 0)
2330                         goto err2;
2331         }
2332
2333         memset(&ct->proto, 0, sizeof(ct->proto));
2334         if (cda[CTA_PROTOINFO]) {
2335                 err = ctnetlink_change_protoinfo(ct, cda);
2336                 if (err < 0)
2337                         goto err2;
2338         }
2339
2340         if (cda[CTA_SYNPROXY]) {
2341                 err = ctnetlink_change_synproxy(ct, cda);
2342                 if (err < 0)
2343                         goto err2;
2344         }
2345
2346 #if defined(CONFIG_NF_CONNTRACK_MARK)
2347         if (cda[CTA_MARK])
2348                 ctnetlink_change_mark(ct, cda);
2349 #endif
2350
2351         /* setup master conntrack: this is a confirmed expectation */
2352         if (cda[CTA_TUPLE_MASTER]) {
2353                 struct nf_conntrack_tuple master;
2354                 struct nf_conntrack_tuple_hash *master_h;
2355                 struct nf_conn *master_ct;
2356
2357                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2358                                             u3, NULL);
2359                 if (err < 0)
2360                         goto err2;
2361
2362                 master_h = nf_conntrack_find_get(net, zone, &master);
2363                 if (master_h == NULL) {
2364                         err = -ENOENT;
2365                         goto err2;
2366                 }
2367                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2368                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2369                 ct->master = master_ct;
2370         }
2371         tstamp = nf_conn_tstamp_find(ct);
2372         if (tstamp)
2373                 tstamp->start = ktime_get_real_ns();
2374
2375         err = nf_conntrack_hash_check_insert(ct);
2376         if (err < 0)
2377                 goto err2;
2378
2379         rcu_read_unlock();
2380
2381         return ct;
2382
2383 err2:
2384         rcu_read_unlock();
2385 err1:
2386         nf_conntrack_free(ct);
2387         return ERR_PTR(err);
2388 }
2389
2390 static int ctnetlink_new_conntrack(struct sk_buff *skb,
2391                                    const struct nfnl_info *info,
2392                                    const struct nlattr * const cda[])
2393 {
2394         struct nf_conntrack_tuple otuple, rtuple;
2395         struct nf_conntrack_tuple_hash *h = NULL;
2396         u_int8_t u3 = info->nfmsg->nfgen_family;
2397         struct nf_conntrack_zone zone;
2398         struct nf_conn *ct;
2399         int err;
2400
2401         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2402         if (err < 0)
2403                 return err;
2404
2405         if (cda[CTA_TUPLE_ORIG]) {
2406                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2407                                             u3, &zone);
2408                 if (err < 0)
2409                         return err;
2410         }
2411
2412         if (cda[CTA_TUPLE_REPLY]) {
2413                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2414                                             u3, &zone);
2415                 if (err < 0)
2416                         return err;
2417         }
2418
2419         if (cda[CTA_TUPLE_ORIG])
2420                 h = nf_conntrack_find_get(info->net, &zone, &otuple);
2421         else if (cda[CTA_TUPLE_REPLY])
2422                 h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2423
2424         if (h == NULL) {
2425                 err = -ENOENT;
2426                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2427                         enum ip_conntrack_events events;
2428
2429                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2430                                 return -EINVAL;
2431                         if (otuple.dst.protonum != rtuple.dst.protonum)
2432                                 return -EINVAL;
2433
2434                         ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2435                                                         &otuple, &rtuple, u3);
2436                         if (IS_ERR(ct))
2437                                 return PTR_ERR(ct);
2438
2439                         err = 0;
2440                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2441                                 events = 1 << IPCT_RELATED;
2442                         else
2443                                 events = 1 << IPCT_NEW;
2444
2445                         if (cda[CTA_LABELS] &&
2446                             ctnetlink_attach_labels(ct, cda) == 0)
2447                                 events |= (1 << IPCT_LABEL);
2448
2449                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2450                                                       (1 << IPCT_ASSURED) |
2451                                                       (1 << IPCT_HELPER) |
2452                                                       (1 << IPCT_PROTOINFO) |
2453                                                       (1 << IPCT_SEQADJ) |
2454                                                       (1 << IPCT_MARK) |
2455                                                       (1 << IPCT_SYNPROXY) |
2456                                                       events,
2457                                                       ct, NETLINK_CB(skb).portid,
2458                                                       nlmsg_report(info->nlh));
2459                         nf_ct_put(ct);
2460                 }
2461
2462                 return err;
2463         }
2464         /* implicit 'else' */
2465
2466         err = -EEXIST;
2467         ct = nf_ct_tuplehash_to_ctrack(h);
2468         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2469                 err = ctnetlink_change_conntrack(ct, cda);
2470                 if (err == 0) {
2471                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2472                                                       (1 << IPCT_ASSURED) |
2473                                                       (1 << IPCT_HELPER) |
2474                                                       (1 << IPCT_LABEL) |
2475                                                       (1 << IPCT_PROTOINFO) |
2476                                                       (1 << IPCT_SEQADJ) |
2477                                                       (1 << IPCT_MARK) |
2478                                                       (1 << IPCT_SYNPROXY),
2479                                                       ct, NETLINK_CB(skb).portid,
2480                                                       nlmsg_report(info->nlh));
2481                 }
2482         }
2483
2484         nf_ct_put(ct);
2485         return err;
2486 }
2487
2488 static int
2489 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2490                                 __u16 cpu, const struct ip_conntrack_stat *st)
2491 {
2492         struct nlmsghdr *nlh;
2493         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2494
2495         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2496                               IPCTNL_MSG_CT_GET_STATS_CPU);
2497         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2498                            NFNETLINK_V0, htons(cpu));
2499         if (!nlh)
2500                 goto nlmsg_failure;
2501
2502         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2503             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2504             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2505             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2506                                 htonl(st->insert_failed)) ||
2507             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2508             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2509             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2510             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2511                                 htonl(st->search_restart)) ||
2512             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2513                                 htonl(st->clash_resolve)) ||
2514             nla_put_be32(skb, CTA_STATS_CHAIN_TOOLONG,
2515                          htonl(st->chaintoolong)))
2516                 goto nla_put_failure;
2517
2518         nlmsg_end(skb, nlh);
2519         return skb->len;
2520
2521 nla_put_failure:
2522 nlmsg_failure:
2523         nlmsg_cancel(skb, nlh);
2524         return -1;
2525 }
2526
2527 static int
2528 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2529 {
2530         int cpu;
2531         struct net *net = sock_net(skb->sk);
2532
2533         if (cb->args[0] == nr_cpu_ids)
2534                 return 0;
2535
2536         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2537                 const struct ip_conntrack_stat *st;
2538
2539                 if (!cpu_possible(cpu))
2540                         continue;
2541
2542                 st = per_cpu_ptr(net->ct.stat, cpu);
2543                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2544                                                     NETLINK_CB(cb->skb).portid,
2545                                                     cb->nlh->nlmsg_seq,
2546                                                     cpu, st) < 0)
2547                                 break;
2548         }
2549         cb->args[0] = cpu;
2550
2551         return skb->len;
2552 }
2553
2554 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2555                                  const struct nfnl_info *info,
2556                                  const struct nlattr * const cda[])
2557 {
2558         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2559                 struct netlink_dump_control c = {
2560                         .dump = ctnetlink_ct_stat_cpu_dump,
2561                 };
2562                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2563         }
2564
2565         return 0;
2566 }
2567
2568 static int
2569 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2570                             struct net *net)
2571 {
2572         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2573         unsigned int nr_conntracks;
2574         struct nlmsghdr *nlh;
2575
2576         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2577         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2578                            NFNETLINK_V0, 0);
2579         if (!nlh)
2580                 goto nlmsg_failure;
2581
2582         nr_conntracks = nf_conntrack_count(net);
2583         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2584                 goto nla_put_failure;
2585
2586         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2587                 goto nla_put_failure;
2588
2589         nlmsg_end(skb, nlh);
2590         return skb->len;
2591
2592 nla_put_failure:
2593 nlmsg_failure:
2594         nlmsg_cancel(skb, nlh);
2595         return -1;
2596 }
2597
2598 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2599                              const struct nlattr * const cda[])
2600 {
2601         struct sk_buff *skb2;
2602         int err;
2603
2604         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2605         if (skb2 == NULL)
2606                 return -ENOMEM;
2607
2608         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2609                                           info->nlh->nlmsg_seq,
2610                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2611                                           sock_net(skb->sk));
2612         if (err <= 0) {
2613                 kfree_skb(skb2);
2614                 return -ENOMEM;
2615         }
2616
2617         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2618 }
2619
2620 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2621         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2622         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2623         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2624         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2625         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2626         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2627                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2628         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2629         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2630         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2631         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2632         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2633 };
2634
2635 static struct nf_conntrack_expect *
2636 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2637                        struct nf_conntrack_helper *helper,
2638                        struct nf_conntrack_tuple *tuple,
2639                        struct nf_conntrack_tuple *mask);
2640
2641 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2642 static size_t
2643 ctnetlink_glue_build_size(const struct nf_conn *ct)
2644 {
2645         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2646                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2647                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2648                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2649                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2650                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2651                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2652                + nla_total_size(0) /* CTA_PROTOINFO */
2653                + nla_total_size(0) /* CTA_HELP */
2654                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2655                + ctnetlink_secctx_size(ct)
2656                + ctnetlink_acct_size(ct)
2657                + ctnetlink_timestamp_size(ct)
2658 #if IS_ENABLED(CONFIG_NF_NAT)
2659                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2660                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2661 #endif
2662 #ifdef CONFIG_NF_CONNTRACK_MARK
2663                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2664 #endif
2665 #ifdef CONFIG_NF_CONNTRACK_ZONES
2666                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2667 #endif
2668                + ctnetlink_proto_size(ct)
2669                ;
2670 }
2671
2672 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2673 {
2674         const struct nf_conntrack_zone *zone;
2675         struct nlattr *nest_parms;
2676
2677         zone = nf_ct_zone(ct);
2678
2679         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2680         if (!nest_parms)
2681                 goto nla_put_failure;
2682         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2683                 goto nla_put_failure;
2684         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2685                                    NF_CT_ZONE_DIR_ORIG) < 0)
2686                 goto nla_put_failure;
2687         nla_nest_end(skb, nest_parms);
2688
2689         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2690         if (!nest_parms)
2691                 goto nla_put_failure;
2692         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2693                 goto nla_put_failure;
2694         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2695                                    NF_CT_ZONE_DIR_REPL) < 0)
2696                 goto nla_put_failure;
2697         nla_nest_end(skb, nest_parms);
2698
2699         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2700                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2701                 goto nla_put_failure;
2702
2703         if (ctnetlink_dump_id(skb, ct) < 0)
2704                 goto nla_put_failure;
2705
2706         if (ctnetlink_dump_status(skb, ct) < 0)
2707                 goto nla_put_failure;
2708
2709         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2710                 goto nla_put_failure;
2711
2712         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2713                 goto nla_put_failure;
2714
2715         if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2716             ctnetlink_dump_timestamp(skb, ct) < 0)
2717                 goto nla_put_failure;
2718
2719         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2720                 goto nla_put_failure;
2721
2722 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2723         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2724                 goto nla_put_failure;
2725 #endif
2726         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2727                 goto nla_put_failure;
2728
2729         if ((ct->status & IPS_SEQ_ADJUST) &&
2730             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2731                 goto nla_put_failure;
2732
2733         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2734                 goto nla_put_failure;
2735
2736 #ifdef CONFIG_NF_CONNTRACK_MARK
2737         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2738                 goto nla_put_failure;
2739 #endif
2740         if (ctnetlink_dump_labels(skb, ct) < 0)
2741                 goto nla_put_failure;
2742         return 0;
2743
2744 nla_put_failure:
2745         return -ENOSPC;
2746 }
2747
2748 static int
2749 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2750                      enum ip_conntrack_info ctinfo,
2751                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2752 {
2753         struct nlattr *nest_parms;
2754
2755         nest_parms = nla_nest_start(skb, ct_attr);
2756         if (!nest_parms)
2757                 goto nla_put_failure;
2758
2759         if (__ctnetlink_glue_build(skb, ct) < 0)
2760                 goto nla_put_failure;
2761
2762         nla_nest_end(skb, nest_parms);
2763
2764         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2765                 goto nla_put_failure;
2766
2767         return 0;
2768
2769 nla_put_failure:
2770         return -ENOSPC;
2771 }
2772
2773 static int
2774 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2775 {
2776         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2777         unsigned long d = ct->status ^ status;
2778
2779         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2780                 /* SEEN_REPLY bit can only be set */
2781                 return -EBUSY;
2782
2783         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2784                 /* ASSURED bit can only be set */
2785                 return -EBUSY;
2786
2787         /* This check is less strict than ctnetlink_change_status()
2788          * because callers often flip IPS_EXPECTED bits when sending
2789          * an NFQA_CT attribute to the kernel.  So ignore the
2790          * unchangeable bits but do not error out. Also user programs
2791          * are allowed to clear the bits that they are allowed to change.
2792          */
2793         __nf_ct_change_status(ct, status, ~status);
2794         return 0;
2795 }
2796
2797 static int
2798 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2799 {
2800         int err;
2801
2802         if (cda[CTA_TIMEOUT]) {
2803                 err = ctnetlink_change_timeout(ct, cda);
2804                 if (err < 0)
2805                         return err;
2806         }
2807         if (cda[CTA_STATUS]) {
2808                 err = ctnetlink_update_status(ct, cda);
2809                 if (err < 0)
2810                         return err;
2811         }
2812         if (cda[CTA_HELP]) {
2813                 err = ctnetlink_change_helper(ct, cda);
2814                 if (err < 0)
2815                         return err;
2816         }
2817         if (cda[CTA_LABELS]) {
2818                 err = ctnetlink_attach_labels(ct, cda);
2819                 if (err < 0)
2820                         return err;
2821         }
2822 #if defined(CONFIG_NF_CONNTRACK_MARK)
2823         if (cda[CTA_MARK]) {
2824                 ctnetlink_change_mark(ct, cda);
2825         }
2826 #endif
2827         return 0;
2828 }
2829
2830 static int
2831 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2832 {
2833         struct nlattr *cda[CTA_MAX+1];
2834         int ret;
2835
2836         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2837                                           NULL);
2838         if (ret < 0)
2839                 return ret;
2840
2841         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2842 }
2843
2844 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2845                                     const struct nf_conn *ct,
2846                                     struct nf_conntrack_tuple *tuple,
2847                                     struct nf_conntrack_tuple *mask)
2848 {
2849         int err;
2850
2851         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2852                                     nf_ct_l3num(ct), NULL);
2853         if (err < 0)
2854                 return err;
2855
2856         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2857                                      nf_ct_l3num(ct), NULL);
2858 }
2859
2860 static int
2861 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2862                              u32 portid, u32 report)
2863 {
2864         struct nlattr *cda[CTA_EXPECT_MAX+1];
2865         struct nf_conntrack_tuple tuple, mask;
2866         struct nf_conntrack_helper *helper = NULL;
2867         struct nf_conntrack_expect *exp;
2868         int err;
2869
2870         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2871                                           exp_nla_policy, NULL);
2872         if (err < 0)
2873                 return err;
2874
2875         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2876                                        ct, &tuple, &mask);
2877         if (err < 0)
2878                 return err;
2879
2880         if (cda[CTA_EXPECT_HELP_NAME]) {
2881                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2882
2883                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2884                                                     nf_ct_protonum(ct));
2885                 if (helper == NULL)
2886                         return -EOPNOTSUPP;
2887         }
2888
2889         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2890                                      helper, &tuple, &mask);
2891         if (IS_ERR(exp))
2892                 return PTR_ERR(exp);
2893
2894         err = nf_ct_expect_related_report(exp, portid, report, 0);
2895         nf_ct_expect_put(exp);
2896         return err;
2897 }
2898
2899 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2900                                   enum ip_conntrack_info ctinfo, int diff)
2901 {
2902         if (!(ct->status & IPS_NAT_MASK))
2903                 return;
2904
2905         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2906 }
2907
2908 static const struct nfnl_ct_hook ctnetlink_glue_hook = {
2909         .build_size     = ctnetlink_glue_build_size,
2910         .build          = ctnetlink_glue_build,
2911         .parse          = ctnetlink_glue_parse,
2912         .attach_expect  = ctnetlink_glue_attach_expect,
2913         .seq_adjust     = ctnetlink_glue_seqadj,
2914 };
2915 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2916
2917 /***********************************************************************
2918  * EXPECT
2919  ***********************************************************************/
2920
2921 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2922                                     const struct nf_conntrack_tuple *tuple,
2923                                     u32 type)
2924 {
2925         struct nlattr *nest_parms;
2926
2927         nest_parms = nla_nest_start(skb, type);
2928         if (!nest_parms)
2929                 goto nla_put_failure;
2930         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2931                 goto nla_put_failure;
2932         nla_nest_end(skb, nest_parms);
2933
2934         return 0;
2935
2936 nla_put_failure:
2937         return -1;
2938 }
2939
2940 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2941                                    const struct nf_conntrack_tuple *tuple,
2942                                    const struct nf_conntrack_tuple_mask *mask)
2943 {
2944         const struct nf_conntrack_l4proto *l4proto;
2945         struct nf_conntrack_tuple m;
2946         struct nlattr *nest_parms;
2947         int ret;
2948
2949         memset(&m, 0xFF, sizeof(m));
2950         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2951         m.src.u.all = mask->src.u.all;
2952         m.src.l3num = tuple->src.l3num;
2953         m.dst.protonum = tuple->dst.protonum;
2954
2955         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2956         if (!nest_parms)
2957                 goto nla_put_failure;
2958
2959         rcu_read_lock();
2960         ret = ctnetlink_dump_tuples_ip(skb, &m);
2961         if (ret >= 0) {
2962                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2963                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2964         }
2965         rcu_read_unlock();
2966
2967         if (unlikely(ret < 0))
2968                 goto nla_put_failure;
2969
2970         nla_nest_end(skb, nest_parms);
2971
2972         return 0;
2973
2974 nla_put_failure:
2975         return -1;
2976 }
2977
2978 static const union nf_inet_addr any_addr;
2979
2980 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2981 {
2982         static siphash_aligned_key_t exp_id_seed;
2983         unsigned long a, b, c, d;
2984
2985         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2986
2987         a = (unsigned long)exp;
2988         b = (unsigned long)exp->helper;
2989         c = (unsigned long)exp->master;
2990         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2991
2992 #ifdef CONFIG_64BIT
2993         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2994 #else
2995         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2996 #endif
2997 }
2998
2999 static int
3000 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3001                           const struct nf_conntrack_expect *exp)
3002 {
3003         struct nf_conn *master = exp->master;
3004         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3005         struct nf_conn_help *help;
3006 #if IS_ENABLED(CONFIG_NF_NAT)
3007         struct nlattr *nest_parms;
3008         struct nf_conntrack_tuple nat_tuple = {};
3009 #endif
3010         struct nf_ct_helper_expectfn *expfn;
3011
3012         if (timeout < 0)
3013                 timeout = 0;
3014
3015         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3016                 goto nla_put_failure;
3017         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3018                 goto nla_put_failure;
3019         if (ctnetlink_exp_dump_tuple(skb,
3020                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3021                                  CTA_EXPECT_MASTER) < 0)
3022                 goto nla_put_failure;
3023
3024 #if IS_ENABLED(CONFIG_NF_NAT)
3025         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3026             exp->saved_proto.all) {
3027                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3028                 if (!nest_parms)
3029                         goto nla_put_failure;
3030
3031                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3032                         goto nla_put_failure;
3033
3034                 nat_tuple.src.l3num = nf_ct_l3num(master);
3035                 nat_tuple.src.u3 = exp->saved_addr;
3036                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3037                 nat_tuple.src.u = exp->saved_proto;
3038
3039                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3040                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3041                         goto nla_put_failure;
3042                 nla_nest_end(skb, nest_parms);
3043         }
3044 #endif
3045         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3046             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3047             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3048             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3049                 goto nla_put_failure;
3050         help = nfct_help(master);
3051         if (help) {
3052                 struct nf_conntrack_helper *helper;
3053
3054                 helper = rcu_dereference(help->helper);
3055                 if (helper &&
3056                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3057                         goto nla_put_failure;
3058         }
3059         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3060         if (expfn != NULL &&
3061             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3062                 goto nla_put_failure;
3063
3064         return 0;
3065
3066 nla_put_failure:
3067         return -1;
3068 }
3069
3070 static int
3071 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3072                         int event, const struct nf_conntrack_expect *exp)
3073 {
3074         struct nlmsghdr *nlh;
3075         unsigned int flags = portid ? NLM_F_MULTI : 0;
3076
3077         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3078         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3079                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3080         if (!nlh)
3081                 goto nlmsg_failure;
3082
3083         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3084                 goto nla_put_failure;
3085
3086         nlmsg_end(skb, nlh);
3087         return skb->len;
3088
3089 nlmsg_failure:
3090 nla_put_failure:
3091         nlmsg_cancel(skb, nlh);
3092         return -1;
3093 }
3094
3095 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3096 static int
3097 ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3098 {
3099         struct nf_conntrack_expect *exp = item->exp;
3100         struct net *net = nf_ct_exp_net(exp);
3101         struct nlmsghdr *nlh;
3102         struct sk_buff *skb;
3103         unsigned int type, group;
3104         int flags = 0;
3105
3106         if (events & (1 << IPEXP_DESTROY)) {
3107                 type = IPCTNL_MSG_EXP_DELETE;
3108                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3109         } else if (events & (1 << IPEXP_NEW)) {
3110                 type = IPCTNL_MSG_EXP_NEW;
3111                 flags = NLM_F_CREATE|NLM_F_EXCL;
3112                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3113         } else
3114                 return 0;
3115
3116         if (!item->report && !nfnetlink_has_listeners(net, group))
3117                 return 0;
3118
3119         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3120         if (skb == NULL)
3121                 goto errout;
3122
3123         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3124         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3125                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3126         if (!nlh)
3127                 goto nlmsg_failure;
3128
3129         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3130                 goto nla_put_failure;
3131
3132         nlmsg_end(skb, nlh);
3133         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3134         return 0;
3135
3136 nla_put_failure:
3137         nlmsg_cancel(skb, nlh);
3138 nlmsg_failure:
3139         kfree_skb(skb);
3140 errout:
3141         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3142         return 0;
3143 }
3144 #endif
3145 static int ctnetlink_exp_done(struct netlink_callback *cb)
3146 {
3147         if (cb->args[1])
3148                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3149         return 0;
3150 }
3151
3152 static int
3153 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3154 {
3155         struct net *net = sock_net(skb->sk);
3156         struct nf_conntrack_expect *exp, *last;
3157         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3158         u_int8_t l3proto = nfmsg->nfgen_family;
3159
3160         rcu_read_lock();
3161         last = (struct nf_conntrack_expect *)cb->args[1];
3162         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3163 restart:
3164                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3165                                          hnode) {
3166                         if (l3proto && exp->tuple.src.l3num != l3proto)
3167                                 continue;
3168
3169                         if (!net_eq(nf_ct_net(exp->master), net))
3170                                 continue;
3171
3172                         if (cb->args[1]) {
3173                                 if (exp != last)
3174                                         continue;
3175                                 cb->args[1] = 0;
3176                         }
3177                         if (ctnetlink_exp_fill_info(skb,
3178                                                     NETLINK_CB(cb->skb).portid,
3179                                                     cb->nlh->nlmsg_seq,
3180                                                     IPCTNL_MSG_EXP_NEW,
3181                                                     exp) < 0) {
3182                                 if (!refcount_inc_not_zero(&exp->use))
3183                                         continue;
3184                                 cb->args[1] = (unsigned long)exp;
3185                                 goto out;
3186                         }
3187                 }
3188                 if (cb->args[1]) {
3189                         cb->args[1] = 0;
3190                         goto restart;
3191                 }
3192         }
3193 out:
3194         rcu_read_unlock();
3195         if (last)
3196                 nf_ct_expect_put(last);
3197
3198         return skb->len;
3199 }
3200
3201 static int
3202 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3203 {
3204         struct nf_conntrack_expect *exp, *last;
3205         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3206         struct nf_conn *ct = cb->data;
3207         struct nf_conn_help *help = nfct_help(ct);
3208         u_int8_t l3proto = nfmsg->nfgen_family;
3209
3210         if (cb->args[0])
3211                 return 0;
3212
3213         rcu_read_lock();
3214         last = (struct nf_conntrack_expect *)cb->args[1];
3215 restart:
3216         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3217                 if (l3proto && exp->tuple.src.l3num != l3proto)
3218                         continue;
3219                 if (cb->args[1]) {
3220                         if (exp != last)
3221                                 continue;
3222                         cb->args[1] = 0;
3223                 }
3224                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3225                                             cb->nlh->nlmsg_seq,
3226                                             IPCTNL_MSG_EXP_NEW,
3227                                             exp) < 0) {
3228                         if (!refcount_inc_not_zero(&exp->use))
3229                                 continue;
3230                         cb->args[1] = (unsigned long)exp;
3231                         goto out;
3232                 }
3233         }
3234         if (cb->args[1]) {
3235                 cb->args[1] = 0;
3236                 goto restart;
3237         }
3238         cb->args[0] = 1;
3239 out:
3240         rcu_read_unlock();
3241         if (last)
3242                 nf_ct_expect_put(last);
3243
3244         return skb->len;
3245 }
3246
3247 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3248                                  struct sk_buff *skb,
3249                                  const struct nlmsghdr *nlh,
3250                                  const struct nlattr * const cda[],
3251                                  struct netlink_ext_ack *extack)
3252 {
3253         int err;
3254         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3255         u_int8_t u3 = nfmsg->nfgen_family;
3256         struct nf_conntrack_tuple tuple;
3257         struct nf_conntrack_tuple_hash *h;
3258         struct nf_conn *ct;
3259         struct nf_conntrack_zone zone;
3260         struct netlink_dump_control c = {
3261                 .dump = ctnetlink_exp_ct_dump_table,
3262                 .done = ctnetlink_exp_done,
3263         };
3264
3265         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3266                                     u3, NULL);
3267         if (err < 0)
3268                 return err;
3269
3270         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3271         if (err < 0)
3272                 return err;
3273
3274         h = nf_conntrack_find_get(net, &zone, &tuple);
3275         if (!h)
3276                 return -ENOENT;
3277
3278         ct = nf_ct_tuplehash_to_ctrack(h);
3279         /* No expectation linked to this connection tracking. */
3280         if (!nfct_help(ct)) {
3281                 nf_ct_put(ct);
3282                 return 0;
3283         }
3284
3285         c.data = ct;
3286
3287         err = netlink_dump_start(ctnl, skb, nlh, &c);
3288         nf_ct_put(ct);
3289
3290         return err;
3291 }
3292
3293 static int ctnetlink_get_expect(struct sk_buff *skb,
3294                                 const struct nfnl_info *info,
3295                                 const struct nlattr * const cda[])
3296 {
3297         u_int8_t u3 = info->nfmsg->nfgen_family;
3298         struct nf_conntrack_tuple tuple;
3299         struct nf_conntrack_expect *exp;
3300         struct nf_conntrack_zone zone;
3301         struct sk_buff *skb2;
3302         int err;
3303
3304         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3305                 if (cda[CTA_EXPECT_MASTER])
3306                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3307                                                      info->nlh, cda,
3308                                                      info->extack);
3309                 else {
3310                         struct netlink_dump_control c = {
3311                                 .dump = ctnetlink_exp_dump_table,
3312                                 .done = ctnetlink_exp_done,
3313                         };
3314                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3315                 }
3316         }
3317
3318         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3319         if (err < 0)
3320                 return err;
3321
3322         if (cda[CTA_EXPECT_TUPLE])
3323                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3324                                             u3, NULL);
3325         else if (cda[CTA_EXPECT_MASTER])
3326                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3327                                             u3, NULL);
3328         else
3329                 return -EINVAL;
3330
3331         if (err < 0)
3332                 return err;
3333
3334         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3335         if (!exp)
3336                 return -ENOENT;
3337
3338         if (cda[CTA_EXPECT_ID]) {
3339                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3340
3341                 if (id != nf_expect_get_id(exp)) {
3342                         nf_ct_expect_put(exp);
3343                         return -ENOENT;
3344                 }
3345         }
3346
3347         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3348         if (!skb2) {
3349                 nf_ct_expect_put(exp);
3350                 return -ENOMEM;
3351         }
3352
3353         rcu_read_lock();
3354         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3355                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3356                                       exp);
3357         rcu_read_unlock();
3358         nf_ct_expect_put(exp);
3359         if (err <= 0) {
3360                 kfree_skb(skb2);
3361                 return -ENOMEM;
3362         }
3363
3364         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3365 }
3366
3367 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3368 {
3369         const struct nf_conn_help *m_help;
3370         const char *name = data;
3371
3372         m_help = nfct_help(exp->master);
3373
3374         return strcmp(m_help->helper->name, name) == 0;
3375 }
3376
3377 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3378 {
3379         return true;
3380 }
3381
3382 static int ctnetlink_del_expect(struct sk_buff *skb,
3383                                 const struct nfnl_info *info,
3384                                 const struct nlattr * const cda[])
3385 {
3386         u_int8_t u3 = info->nfmsg->nfgen_family;
3387         struct nf_conntrack_expect *exp;
3388         struct nf_conntrack_tuple tuple;
3389         struct nf_conntrack_zone zone;
3390         int err;
3391
3392         if (cda[CTA_EXPECT_TUPLE]) {
3393                 /* delete a single expect by tuple */
3394                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3395                 if (err < 0)
3396                         return err;
3397
3398                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3399                                             u3, NULL);
3400                 if (err < 0)
3401                         return err;
3402
3403                 /* bump usage count to 2 */
3404                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3405                 if (!exp)
3406                         return -ENOENT;
3407
3408                 if (cda[CTA_EXPECT_ID]) {
3409                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3410                         if (ntohl(id) != (u32)(unsigned long)exp) {
3411                                 nf_ct_expect_put(exp);
3412                                 return -ENOENT;
3413                         }
3414                 }
3415
3416                 /* after list removal, usage count == 1 */
3417                 spin_lock_bh(&nf_conntrack_expect_lock);
3418                 if (del_timer(&exp->timeout)) {
3419                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3420                                                    nlmsg_report(info->nlh));
3421                         nf_ct_expect_put(exp);
3422                 }
3423                 spin_unlock_bh(&nf_conntrack_expect_lock);
3424                 /* have to put what we 'get' above.
3425                  * after this line usage count == 0 */
3426                 nf_ct_expect_put(exp);
3427         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3428                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3429
3430                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3431                                          NETLINK_CB(skb).portid,
3432                                          nlmsg_report(info->nlh));
3433         } else {
3434                 /* This basically means we have to flush everything*/
3435                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3436                                          NETLINK_CB(skb).portid,
3437                                          nlmsg_report(info->nlh));
3438         }
3439
3440         return 0;
3441 }
3442 static int
3443 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3444                         const struct nlattr * const cda[])
3445 {
3446         if (cda[CTA_EXPECT_TIMEOUT]) {
3447                 if (!del_timer(&x->timeout))
3448                         return -ETIME;
3449
3450                 x->timeout.expires = jiffies +
3451                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3452                 add_timer(&x->timeout);
3453         }
3454         return 0;
3455 }
3456
3457 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3458         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3459         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3460 };
3461
3462 static int
3463 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3464                            struct nf_conntrack_expect *exp,
3465                            u_int8_t u3)
3466 {
3467 #if IS_ENABLED(CONFIG_NF_NAT)
3468         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3469         struct nf_conntrack_tuple nat_tuple = {};
3470         int err;
3471
3472         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3473                                           exp_nat_nla_policy, NULL);
3474         if (err < 0)
3475                 return err;
3476
3477         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3478                 return -EINVAL;
3479
3480         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3481                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3482                                     u3, NULL);
3483         if (err < 0)
3484                 return err;
3485
3486         exp->saved_addr = nat_tuple.src.u3;
3487         exp->saved_proto = nat_tuple.src.u;
3488         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3489
3490         return 0;
3491 #else
3492         return -EOPNOTSUPP;
3493 #endif
3494 }
3495
3496 static struct nf_conntrack_expect *
3497 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3498                        struct nf_conntrack_helper *helper,
3499                        struct nf_conntrack_tuple *tuple,
3500                        struct nf_conntrack_tuple *mask)
3501 {
3502         u_int32_t class = 0;
3503         struct nf_conntrack_expect *exp;
3504         struct nf_conn_help *help;
3505         int err;
3506
3507         help = nfct_help(ct);
3508         if (!help)
3509                 return ERR_PTR(-EOPNOTSUPP);
3510
3511         if (cda[CTA_EXPECT_CLASS] && helper) {
3512                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3513                 if (class > helper->expect_class_max)
3514                         return ERR_PTR(-EINVAL);
3515         }
3516         exp = nf_ct_expect_alloc(ct);
3517         if (!exp)
3518                 return ERR_PTR(-ENOMEM);
3519
3520         if (cda[CTA_EXPECT_FLAGS]) {
3521                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3522                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3523         } else {
3524                 exp->flags = 0;
3525         }
3526         if (cda[CTA_EXPECT_FN]) {
3527                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3528                 struct nf_ct_helper_expectfn *expfn;
3529
3530                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3531                 if (expfn == NULL) {
3532                         err = -EINVAL;
3533                         goto err_out;
3534                 }
3535                 exp->expectfn = expfn->expectfn;
3536         } else
3537                 exp->expectfn = NULL;
3538
3539         exp->class = class;
3540         exp->master = ct;
3541         exp->helper = helper;
3542         exp->tuple = *tuple;
3543         exp->mask.src.u3 = mask->src.u3;
3544         exp->mask.src.u.all = mask->src.u.all;
3545
3546         if (cda[CTA_EXPECT_NAT]) {
3547                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3548                                                  exp, nf_ct_l3num(ct));
3549                 if (err < 0)
3550                         goto err_out;
3551         }
3552         return exp;
3553 err_out:
3554         nf_ct_expect_put(exp);
3555         return ERR_PTR(err);
3556 }
3557
3558 static int
3559 ctnetlink_create_expect(struct net *net,
3560                         const struct nf_conntrack_zone *zone,
3561                         const struct nlattr * const cda[],
3562                         u_int8_t u3, u32 portid, int report)
3563 {
3564         struct nf_conntrack_tuple tuple, mask, master_tuple;
3565         struct nf_conntrack_tuple_hash *h = NULL;
3566         struct nf_conntrack_helper *helper = NULL;
3567         struct nf_conntrack_expect *exp;
3568         struct nf_conn *ct;
3569         int err;
3570
3571         /* caller guarantees that those three CTA_EXPECT_* exist */
3572         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3573                                     u3, NULL);
3574         if (err < 0)
3575                 return err;
3576         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3577                                     u3, NULL);
3578         if (err < 0)
3579                 return err;
3580         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3581                                     u3, NULL);
3582         if (err < 0)
3583                 return err;
3584
3585         /* Look for master conntrack of this expectation */
3586         h = nf_conntrack_find_get(net, zone, &master_tuple);
3587         if (!h)
3588                 return -ENOENT;
3589         ct = nf_ct_tuplehash_to_ctrack(h);
3590
3591         rcu_read_lock();
3592         if (cda[CTA_EXPECT_HELP_NAME]) {
3593                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3594
3595                 helper = __nf_conntrack_helper_find(helpname, u3,
3596                                                     nf_ct_protonum(ct));
3597                 if (helper == NULL) {
3598                         rcu_read_unlock();
3599 #ifdef CONFIG_MODULES
3600                         if (request_module("nfct-helper-%s", helpname) < 0) {
3601                                 err = -EOPNOTSUPP;
3602                                 goto err_ct;
3603                         }
3604                         rcu_read_lock();
3605                         helper = __nf_conntrack_helper_find(helpname, u3,
3606                                                             nf_ct_protonum(ct));
3607                         if (helper) {
3608                                 err = -EAGAIN;
3609                                 goto err_rcu;
3610                         }
3611                         rcu_read_unlock();
3612 #endif
3613                         err = -EOPNOTSUPP;
3614                         goto err_ct;
3615                 }
3616         }
3617
3618         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3619         if (IS_ERR(exp)) {
3620                 err = PTR_ERR(exp);
3621                 goto err_rcu;
3622         }
3623
3624         err = nf_ct_expect_related_report(exp, portid, report, 0);
3625         nf_ct_expect_put(exp);
3626 err_rcu:
3627         rcu_read_unlock();
3628 err_ct:
3629         nf_ct_put(ct);
3630         return err;
3631 }
3632
3633 static int ctnetlink_new_expect(struct sk_buff *skb,
3634                                 const struct nfnl_info *info,
3635                                 const struct nlattr * const cda[])
3636 {
3637         u_int8_t u3 = info->nfmsg->nfgen_family;
3638         struct nf_conntrack_tuple tuple;
3639         struct nf_conntrack_expect *exp;
3640         struct nf_conntrack_zone zone;
3641         int err;
3642
3643         if (!cda[CTA_EXPECT_TUPLE]
3644             || !cda[CTA_EXPECT_MASK]
3645             || !cda[CTA_EXPECT_MASTER])
3646                 return -EINVAL;
3647
3648         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3649         if (err < 0)
3650                 return err;
3651
3652         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3653                                     u3, NULL);
3654         if (err < 0)
3655                 return err;
3656
3657         spin_lock_bh(&nf_conntrack_expect_lock);
3658         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3659         if (!exp) {
3660                 spin_unlock_bh(&nf_conntrack_expect_lock);
3661                 err = -ENOENT;
3662                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3663                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3664                                                       NETLINK_CB(skb).portid,
3665                                                       nlmsg_report(info->nlh));
3666                 }
3667                 return err;
3668         }
3669
3670         err = -EEXIST;
3671         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3672                 err = ctnetlink_change_expect(exp, cda);
3673         spin_unlock_bh(&nf_conntrack_expect_lock);
3674
3675         return err;
3676 }
3677
3678 static int
3679 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3680                              const struct ip_conntrack_stat *st)
3681 {
3682         struct nlmsghdr *nlh;
3683         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3684
3685         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3686                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3687         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3688                            NFNETLINK_V0, htons(cpu));
3689         if (!nlh)
3690                 goto nlmsg_failure;
3691
3692         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3693             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3694             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3695                 goto nla_put_failure;
3696
3697         nlmsg_end(skb, nlh);
3698         return skb->len;
3699
3700 nla_put_failure:
3701 nlmsg_failure:
3702         nlmsg_cancel(skb, nlh);
3703         return -1;
3704 }
3705
3706 static int
3707 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3708 {
3709         int cpu;
3710         struct net *net = sock_net(skb->sk);
3711
3712         if (cb->args[0] == nr_cpu_ids)
3713                 return 0;
3714
3715         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3716                 const struct ip_conntrack_stat *st;
3717
3718                 if (!cpu_possible(cpu))
3719                         continue;
3720
3721                 st = per_cpu_ptr(net->ct.stat, cpu);
3722                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3723                                                  cb->nlh->nlmsg_seq,
3724                                                  cpu, st) < 0)
3725                         break;
3726         }
3727         cb->args[0] = cpu;
3728
3729         return skb->len;
3730 }
3731
3732 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3733                                   const struct nfnl_info *info,
3734                                   const struct nlattr * const cda[])
3735 {
3736         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3737                 struct netlink_dump_control c = {
3738                         .dump = ctnetlink_exp_stat_cpu_dump,
3739                 };
3740                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3741         }
3742
3743         return 0;
3744 }
3745
3746 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3747 static struct nf_ct_event_notifier ctnl_notifier = {
3748         .ct_event = ctnetlink_conntrack_event,
3749         .exp_event = ctnetlink_expect_event,
3750 };
3751 #endif
3752
3753 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3754         [IPCTNL_MSG_CT_NEW]     = {
3755                 .call           = ctnetlink_new_conntrack,
3756                 .type           = NFNL_CB_MUTEX,
3757                 .attr_count     = CTA_MAX,
3758                 .policy         = ct_nla_policy
3759         },
3760         [IPCTNL_MSG_CT_GET]     = {
3761                 .call           = ctnetlink_get_conntrack,
3762                 .type           = NFNL_CB_MUTEX,
3763                 .attr_count     = CTA_MAX,
3764                 .policy         = ct_nla_policy
3765         },
3766         [IPCTNL_MSG_CT_DELETE]  = {
3767                 .call           = ctnetlink_del_conntrack,
3768                 .type           = NFNL_CB_MUTEX,
3769                 .attr_count     = CTA_MAX,
3770                 .policy         = ct_nla_policy
3771         },
3772         [IPCTNL_MSG_CT_GET_CTRZERO] = {
3773                 .call           = ctnetlink_get_conntrack,
3774                 .type           = NFNL_CB_MUTEX,
3775                 .attr_count     = CTA_MAX,
3776                 .policy         = ct_nla_policy
3777         },
3778         [IPCTNL_MSG_CT_GET_STATS_CPU] = {
3779                 .call           = ctnetlink_stat_ct_cpu,
3780                 .type           = NFNL_CB_MUTEX,
3781         },
3782         [IPCTNL_MSG_CT_GET_STATS] = {
3783                 .call           = ctnetlink_stat_ct,
3784                 .type           = NFNL_CB_MUTEX,
3785         },
3786         [IPCTNL_MSG_CT_GET_DYING] = {
3787                 .call           = ctnetlink_get_ct_dying,
3788                 .type           = NFNL_CB_MUTEX,
3789         },
3790         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = {
3791                 .call           = ctnetlink_get_ct_unconfirmed,
3792                 .type           = NFNL_CB_MUTEX,
3793         },
3794 };
3795
3796 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3797         [IPCTNL_MSG_EXP_GET] = {
3798                 .call           = ctnetlink_get_expect,
3799                 .type           = NFNL_CB_MUTEX,
3800                 .attr_count     = CTA_EXPECT_MAX,
3801                 .policy         = exp_nla_policy
3802         },
3803         [IPCTNL_MSG_EXP_NEW] = {
3804                 .call           = ctnetlink_new_expect,
3805                 .type           = NFNL_CB_MUTEX,
3806                 .attr_count     = CTA_EXPECT_MAX,
3807                 .policy         = exp_nla_policy
3808         },
3809         [IPCTNL_MSG_EXP_DELETE] = {
3810                 .call           = ctnetlink_del_expect,
3811                 .type           = NFNL_CB_MUTEX,
3812                 .attr_count     = CTA_EXPECT_MAX,
3813                 .policy         = exp_nla_policy
3814         },
3815         [IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3816                 .call           = ctnetlink_stat_exp_cpu,
3817                 .type           = NFNL_CB_MUTEX,
3818         },
3819 };
3820
3821 static const struct nfnetlink_subsystem ctnl_subsys = {
3822         .name                           = "conntrack",
3823         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3824         .cb_count                       = IPCTNL_MSG_MAX,
3825         .cb                             = ctnl_cb,
3826 };
3827
3828 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3829         .name                           = "conntrack_expect",
3830         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3831         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3832         .cb                             = ctnl_exp_cb,
3833 };
3834
3835 MODULE_ALIAS("ip_conntrack_netlink");
3836 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3837 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3838
3839 static int __net_init ctnetlink_net_init(struct net *net)
3840 {
3841 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3842         nf_conntrack_register_notifier(net, &ctnl_notifier);
3843 #endif
3844         return 0;
3845 }
3846
3847 static void ctnetlink_net_pre_exit(struct net *net)
3848 {
3849 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3850         nf_conntrack_unregister_notifier(net);
3851 #endif
3852 }
3853
3854 static struct pernet_operations ctnetlink_net_ops = {
3855         .init           = ctnetlink_net_init,
3856         .pre_exit       = ctnetlink_net_pre_exit,
3857 };
3858
3859 static int __init ctnetlink_init(void)
3860 {
3861         int ret;
3862
3863         BUILD_BUG_ON(sizeof(struct ctnetlink_list_dump_ctx) > sizeof_field(struct netlink_callback, ctx));
3864
3865         ret = nfnetlink_subsys_register(&ctnl_subsys);
3866         if (ret < 0) {
3867                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3868                 goto err_out;
3869         }
3870
3871         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3872         if (ret < 0) {
3873                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3874                 goto err_unreg_subsys;
3875         }
3876
3877         ret = register_pernet_subsys(&ctnetlink_net_ops);
3878         if (ret < 0) {
3879                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3880                 goto err_unreg_exp_subsys;
3881         }
3882 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3883         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3884         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3885 #endif
3886         return 0;
3887
3888 err_unreg_exp_subsys:
3889         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3890 err_unreg_subsys:
3891         nfnetlink_subsys_unregister(&ctnl_subsys);
3892 err_out:
3893         return ret;
3894 }
3895
3896 static void __exit ctnetlink_exit(void)
3897 {
3898         unregister_pernet_subsys(&ctnetlink_net_ops);
3899         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3900         nfnetlink_subsys_unregister(&ctnl_subsys);
3901 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3902         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3903 #endif
3904         synchronize_rcu();
3905 }
3906
3907 module_init(ctnetlink_init);
3908 module_exit(ctnetlink_exit);