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 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  62                                 const struct nf_conntrack_tuple *tuple,
  63                                 const struct nf_conntrack_l4proto *l4proto)
  64 {
  65         int ret = 0;
  66         struct nlattr *nest_parms;
  67
  68         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO);
  69         if (!nest_parms)
  70                 goto nla_put_failure;
  71         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  72                 goto nla_put_failure;
  73
  74         if (likely(l4proto->tuple_to_nlattr))
  75                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  76
  77         nla_nest_end(skb, nest_parms);
  78
  79         return ret;
  80
  81 nla_put_failure:
  82         return -1;
  83 }
  84
  85 static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  86                                 const struct nf_conntrack_tuple *tuple)
  87 {
  88         if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  89             nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  90                 return -EMSGSIZE;
  91         return 0;
  92 }
  93
  94 static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
  95                                 const struct nf_conntrack_tuple *tuple)
  96 {
  97         if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
  98             nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
  99                 return -EMSGSIZE;
 100         return 0;
 101 }
 102
 103 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 104                                     const struct nf_conntrack_tuple *tuple)
 105 {
 106         int ret = 0;
 107         struct nlattr *nest_parms;
 108
 109         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP);
 110         if (!nest_parms)
 111                 goto nla_put_failure;
 112
 113         switch (tuple->src.l3num) {
 114         case NFPROTO_IPV4:
 115                 ret = ipv4_tuple_to_nlattr(skb, tuple);
 116                 break;
 117         case NFPROTO_IPV6:
 118                 ret = ipv6_tuple_to_nlattr(skb, tuple);
 119                 break;
 120         }
 121
 122         nla_nest_end(skb, nest_parms);
 123
 124         return ret;
 125
 126 nla_put_failure:
 127         return -1;
 128 }
 129
 130 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 131                                  const struct nf_conntrack_tuple *tuple)
 132 {
 133         const struct nf_conntrack_l4proto *l4proto;
 134         int ret;
 135
 136         rcu_read_lock();
 137         ret = ctnetlink_dump_tuples_ip(skb, tuple);
 138
 139         if (ret >= 0) {
 140                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 141                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 142         }
 143         rcu_read_unlock();
 144         return ret;
 145 }
 146
 147 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 148                                   const struct nf_conntrack_zone *zone, int dir)
 149 {
 150         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 151                 return 0;
 152         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 153                 goto nla_put_failure;
 154         return 0;
 155
 156 nla_put_failure:
 157         return -1;
 158 }
 159
 160 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 161 {
 162         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 163                 goto nla_put_failure;
 164         return 0;
 165
 166 nla_put_failure:
 167         return -1;
 168 }
 169
 170 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct,
 171                                   bool skip_zero)
 172 {
 173         long timeout = nf_ct_expires(ct) / HZ;
 174
 175         if (skip_zero && timeout == 0)
 176                 return 0;
 177
 178         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 179                 goto nla_put_failure;
 180         return 0;
 181
 182 nla_put_failure:
 183         return -1;
 184 }
 185
 186 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct,
 187                                     bool destroy)
 188 {
 189         const struct nf_conntrack_l4proto *l4proto;
 190         struct nlattr *nest_proto;
 191         int ret;
 192
 193         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 194         if (!l4proto->to_nlattr)
 195                 return 0;
 196
 197         nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 198         if (!nest_proto)
 199                 goto nla_put_failure;
 200
 201         ret = l4proto->to_nlattr(skb, nest_proto, ct, destroy);
 202
 203         nla_nest_end(skb, nest_proto);
 204
 205         return ret;
 206
 207 nla_put_failure:
 208         return -1;
 209 }
 210
 211 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 212                                    const struct nf_conn *ct)
 213 {
 214         struct nlattr *nest_helper;
 215         const struct nf_conn_help *help = nfct_help(ct);
 216         struct nf_conntrack_helper *helper;
 217
 218         if (!help)
 219                 return 0;
 220
 221         helper = rcu_dereference(help->helper);
 222         if (!helper)
 223                 goto out;
 224
 225         nest_helper = nla_nest_start(skb, CTA_HELP);
 226         if (!nest_helper)
 227                 goto nla_put_failure;
 228         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 229                 goto nla_put_failure;
 230
 231         if (helper->to_nlattr)
 232                 helper->to_nlattr(skb, ct);
 233
 234         nla_nest_end(skb, nest_helper);
 235 out:
 236         return 0;
 237
 238 nla_put_failure:
 239         return -1;
 240 }
 241
 242 static int
 243 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 244               enum ip_conntrack_dir dir, int type)
 245 {
 246         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 247         struct nf_conn_counter *counter = acct->counter;
 248         struct nlattr *nest_count;
 249         u64 pkts, bytes;
 250
 251         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 252                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 253                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 254         } else {
 255                 pkts = atomic64_read(&counter[dir].packets);
 256                 bytes = atomic64_read(&counter[dir].bytes);
 257         }
 258
 259         nest_count = nla_nest_start(skb, attr);
 260         if (!nest_count)
 261                 goto nla_put_failure;
 262
 263         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 264                          CTA_COUNTERS_PAD) ||
 265             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 266                          CTA_COUNTERS_PAD))
 267                 goto nla_put_failure;
 268
 269         nla_nest_end(skb, nest_count);
 270
 271         return 0;
 272
 273 nla_put_failure:
 274         return -1;
 275 }
 276
 277 static int
 278 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 279 {
 280         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 281
 282         if (!acct)
 283                 return 0;
 284
 285         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 286                 return -1;
 287         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 288                 return -1;
 289
 290         return 0;
 291 }
 292
 293 static int
 294 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 295 {
 296         struct nlattr *nest_count;
 297         const struct nf_conn_tstamp *tstamp;
 298
 299         tstamp = nf_conn_tstamp_find(ct);
 300         if (!tstamp)
 301                 return 0;
 302
 303         nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 304         if (!nest_count)
 305                 goto nla_put_failure;
 306
 307         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 308                          CTA_TIMESTAMP_PAD) ||
 309             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 310                                                cpu_to_be64(tstamp->stop),
 311                                                CTA_TIMESTAMP_PAD)))
 312                 goto nla_put_failure;
 313         nla_nest_end(skb, nest_count);
 314
 315         return 0;
 316
 317 nla_put_failure:
 318         return -1;
 319 }
 320
 321 #ifdef CONFIG_NF_CONNTRACK_MARK
 322 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 323 {
 324         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 325                 goto nla_put_failure;
 326         return 0;
 327
 328 nla_put_failure:
 329         return -1;
 330 }
 331 #else
 332 #define ctnetlink_dump_mark(a, b) (0)
 333 #endif
 334
 335 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 336 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 337 {
 338         struct nlattr *nest_secctx;
 339         int len, ret;
 340         char *secctx;
 341
 342         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 343         if (ret)
 344                 return 0;
 345
 346         ret = -1;
 347         nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 348         if (!nest_secctx)
 349                 goto nla_put_failure;
 350
 351         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 352                 goto nla_put_failure;
 353         nla_nest_end(skb, nest_secctx);
 354
 355         ret = 0;
 356 nla_put_failure:
 357         security_release_secctx(secctx, len);
 358         return ret;
 359 }
 360 #else
 361 #define ctnetlink_dump_secctx(a, b) (0)
 362 #endif
 363
 364 #ifdef CONFIG_NF_CONNTRACK_LABELS
 365 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 366 {
 367         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 368
 369         if (!labels)
 370                 return 0;
 371         return nla_total_size(sizeof(labels->bits));
 372 }
 373
 374 static int
 375 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 376 {
 377         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 378         unsigned int i;
 379
 380         if (!labels)
 381                 return 0;
 382
 383         i = 0;
 384         do {
 385                 if (labels->bits[i] != 0)
 386                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 387                                        labels->bits);
 388                 i++;
 389         } while (i < ARRAY_SIZE(labels->bits));
 390
 391         return 0;
 392 }
 393 #else
 394 #define ctnetlink_dump_labels(a, b) (0)
 395 #define ctnetlink_label_size(a) (0)
 396 #endif
 397
 398 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 399
 400 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 401 {
 402         struct nlattr *nest_parms;
 403
 404         if (!(ct->status & IPS_EXPECTED))
 405                 return 0;
 406
 407         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 408         if (!nest_parms)
 409                 goto nla_put_failure;
 410         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 411                 goto nla_put_failure;
 412         nla_nest_end(skb, nest_parms);
 413
 414         return 0;
 415
 416 nla_put_failure:
 417         return -1;
 418 }
 419
 420 static int
 421 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 422 {
 423         struct nlattr *nest_parms;
 424
 425         nest_parms = nla_nest_start(skb, type);
 426         if (!nest_parms)
 427                 goto nla_put_failure;
 428
 429         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 430                          htonl(seq->correction_pos)) ||
 431             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 432                          htonl(seq->offset_before)) ||
 433             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 434                          htonl(seq->offset_after)))
 435                 goto nla_put_failure;
 436
 437         nla_nest_end(skb, nest_parms);
 438
 439         return 0;
 440
 441 nla_put_failure:
 442         return -1;
 443 }
 444
 445 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 446 {
 447         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 448         struct nf_ct_seqadj *seq;
 449
 450         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 451                 return 0;
 452
 453         spin_lock_bh(&ct->lock);
 454         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 455         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 456                 goto err;
 457
 458         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 459         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 460                 goto err;
 461
 462         spin_unlock_bh(&ct->lock);
 463         return 0;
 464 err:
 465         spin_unlock_bh(&ct->lock);
 466         return -1;
 467 }
 468
 469 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 470 {
 471         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 472         struct nlattr *nest_parms;
 473
 474         if (!synproxy)
 475                 return 0;
 476
 477         nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 478         if (!nest_parms)
 479                 goto nla_put_failure;
 480
 481         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 482             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 483             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 484                 goto nla_put_failure;
 485
 486         nla_nest_end(skb, nest_parms);
 487
 488         return 0;
 489
 490 nla_put_failure:
 491         return -1;
 492 }
 493
 494 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 495 {
 496         __be32 id = (__force __be32)nf_ct_get_id(ct);
 497
 498         if (nla_put_be32(skb, CTA_ID, id))
 499                 goto nla_put_failure;
 500         return 0;
 501
 502 nla_put_failure:
 503         return -1;
 504 }
 505
 506 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 507 {
 508         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 509                 goto nla_put_failure;
 510         return 0;
 511
 512 nla_put_failure:
 513         return -1;
 514 }
 515
 516 /* all these functions access ct->ext. Caller must either hold a reference
 517  * on ct or prevent its deletion by holding either the bucket spinlock or
 518  * pcpu dying list lock.
 519  */
 520 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 521                                   struct nf_conn *ct, u32 type)
 522 {
 523         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 524             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 525             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 526             ctnetlink_dump_labels(skb, ct) < 0 ||
 527             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 528             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 529                 return -1;
 530
 531         return 0;
 532 }
 533
 534 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 535 {
 536         if (ctnetlink_dump_status(skb, ct) < 0 ||
 537             ctnetlink_dump_mark(skb, ct) < 0 ||
 538             ctnetlink_dump_secctx(skb, ct) < 0 ||
 539             ctnetlink_dump_id(skb, ct) < 0 ||
 540             ctnetlink_dump_use(skb, ct) < 0 ||
 541             ctnetlink_dump_master(skb, ct) < 0)
 542                 return -1;
 543
 544         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 545             (ctnetlink_dump_timeout(skb, ct, false) < 0 ||
 546              ctnetlink_dump_protoinfo(skb, ct, false) < 0))
 547                 return -1;
 548
 549         return 0;
 550 }
 551
 552 static int
 553 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 554                     struct nf_conn *ct, bool extinfo, unsigned int flags)
 555 {
 556         const struct nf_conntrack_zone *zone;
 557         struct nlmsghdr *nlh;
 558         struct nlattr *nest_parms;
 559         unsigned int event;
 560
 561         if (portid)
 562                 flags |= NLM_F_MULTI;
 563         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 564         nlh = nfnl_msg_put(skb, portid, seq, event, flags, nf_ct_l3num(ct),
 565                            NFNETLINK_V0, 0);
 566         if (!nlh)
 567                 goto nlmsg_failure;
 568
 569         zone = nf_ct_zone(ct);
 570
 571         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 572         if (!nest_parms)
 573                 goto nla_put_failure;
 574         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 575                 goto nla_put_failure;
 576         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 577                                    NF_CT_ZONE_DIR_ORIG) < 0)
 578                 goto nla_put_failure;
 579         nla_nest_end(skb, nest_parms);
 580
 581         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 582         if (!nest_parms)
 583                 goto nla_put_failure;
 584         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 585                 goto nla_put_failure;
 586         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 587                                    NF_CT_ZONE_DIR_REPL) < 0)
 588                 goto nla_put_failure;
 589         nla_nest_end(skb, nest_parms);
 590
 591         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 592                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 593                 goto nla_put_failure;
 594
 595         if (ctnetlink_dump_info(skb, ct) < 0)
 596                 goto nla_put_failure;
 597         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 598                 goto nla_put_failure;
 599
 600         nlmsg_end(skb, nlh);
 601         return skb->len;
 602
 603 nlmsg_failure:
 604 nla_put_failure:
 605         nlmsg_cancel(skb, nlh);
 606         return -1;
 607 }
 608
 609 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 610         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 611         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 612         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 613         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 614 };
 615
 616 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 617 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 618 {
 619         const struct nf_conntrack_l4proto *l4proto;
 620         size_t len, len4 = 0;
 621
 622         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 623         len *= 3u; /* ORIG, REPLY, MASTER */
 624
 625         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 626         len += l4proto->nlattr_size;
 627         if (l4proto->nlattr_tuple_size) {
 628                 len4 = l4proto->nlattr_tuple_size();
 629                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 630         }
 631
 632         return len + len4;
 633 }
 634 #endif
 635
 636 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 637 {
 638         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 639                 return 0;
 640         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 641                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 642                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 643                ;
 644 }
 645
 646 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 647 {
 648 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 649         int len, ret;
 650
 651         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 652         if (ret)
 653                 return 0;
 654
 655         return nla_total_size(0) /* CTA_SECCTX */
 656                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 657 #else
 658         return 0;
 659 #endif
 660 }
 661
 662 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 663 {
 664 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 665         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 666                 return 0;
 667         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 668 #else
 669         return 0;
 670 #endif
 671 }
 672
 673 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 674 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 675 {
 676         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 677                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 678                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 679                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 680                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 681                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 682                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 683                + ctnetlink_acct_size(ct)
 684                + ctnetlink_timestamp_size(ct)
 685                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 686                + nla_total_size(0) /* CTA_PROTOINFO */
 687                + nla_total_size(0) /* CTA_HELP */
 688                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 689                + ctnetlink_secctx_size(ct)
 690 #if IS_ENABLED(CONFIG_NF_NAT)
 691                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 692                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 693 #endif
 694 #ifdef CONFIG_NF_CONNTRACK_MARK
 695                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 696 #endif
 697 #ifdef CONFIG_NF_CONNTRACK_ZONES
 698                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 699 #endif
 700                + ctnetlink_proto_size(ct)
 701                + ctnetlink_label_size(ct)
 702                ;
 703 }
 704
 705 static int
 706 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 707 {
 708         const struct nf_conntrack_zone *zone;
 709         struct net *net;
 710         struct nlmsghdr *nlh;
 711         struct nlattr *nest_parms;
 712         struct nf_conn *ct = item->ct;
 713         struct sk_buff *skb;
 714         unsigned int type;
 715         unsigned int flags = 0, group;
 716         int err;
 717
 718         if (events & (1 << IPCT_DESTROY)) {
 719                 type = IPCTNL_MSG_CT_DELETE;
 720                 group = NFNLGRP_CONNTRACK_DESTROY;
 721         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 722                 type = IPCTNL_MSG_CT_NEW;
 723                 flags = NLM_F_CREATE|NLM_F_EXCL;
 724                 group = NFNLGRP_CONNTRACK_NEW;
 725         } else if (events) {
 726                 type = IPCTNL_MSG_CT_NEW;
 727                 group = NFNLGRP_CONNTRACK_UPDATE;
 728         } else
 729                 return 0;
 730
 731         net = nf_ct_net(ct);
 732         if (!item->report && !nfnetlink_has_listeners(net, group))
 733                 return 0;
 734
 735         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 736         if (skb == NULL)
 737                 goto errout;
 738
 739         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 740         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags, nf_ct_l3num(ct),
 741                            NFNETLINK_V0, 0);
 742         if (!nlh)
 743                 goto nlmsg_failure;
 744
 745         zone = nf_ct_zone(ct);
 746
 747         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 748         if (!nest_parms)
 749                 goto nla_put_failure;
 750         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 751                 goto nla_put_failure;
 752         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 753                                    NF_CT_ZONE_DIR_ORIG) < 0)
 754                 goto nla_put_failure;
 755         nla_nest_end(skb, nest_parms);
 756
 757         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 758         if (!nest_parms)
 759                 goto nla_put_failure;
 760         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 761                 goto nla_put_failure;
 762         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 763                                    NF_CT_ZONE_DIR_REPL) < 0)
 764                 goto nla_put_failure;
 765         nla_nest_end(skb, nest_parms);
 766
 767         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 768                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 769                 goto nla_put_failure;
 770
 771         if (ctnetlink_dump_id(skb, ct) < 0)
 772                 goto nla_put_failure;
 773
 774         if (ctnetlink_dump_status(skb, ct) < 0)
 775                 goto nla_put_failure;
 776
 777         if (events & (1 << IPCT_DESTROY)) {
 778                 if (ctnetlink_dump_timeout(skb, ct, true) < 0)
 779                         goto nla_put_failure;
 780
 781                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 782                     ctnetlink_dump_timestamp(skb, ct) < 0 ||
 783                     ctnetlink_dump_protoinfo(skb, ct, true) < 0)
 784                         goto nla_put_failure;
 785         } else {
 786                 if (ctnetlink_dump_timeout(skb, ct, false) < 0)
 787                         goto nla_put_failure;
 788
 789                 if (events & (1 << IPCT_PROTOINFO) &&
 790                     ctnetlink_dump_protoinfo(skb, ct, false) < 0)
 791                         goto nla_put_failure;
 792
 793                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 794                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 795                         goto nla_put_failure;
 796
 797 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 798                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 799                     && ctnetlink_dump_secctx(skb, ct) < 0)
 800                         goto nla_put_failure;
 801 #endif
 802                 if (events & (1 << IPCT_LABEL) &&
 803                      ctnetlink_dump_labels(skb, ct) < 0)
 804                         goto nla_put_failure;
 805
 806                 if (events & (1 << IPCT_RELATED) &&
 807                     ctnetlink_dump_master(skb, ct) < 0)
 808                         goto nla_put_failure;
 809
 810                 if (events & (1 << IPCT_SEQADJ) &&
 811                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 812                         goto nla_put_failure;
 813
 814                 if (events & (1 << IPCT_SYNPROXY) &&
 815                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 816                         goto nla_put_failure;
 817         }
 818
 819 #ifdef CONFIG_NF_CONNTRACK_MARK
 820         if ((events & (1 << IPCT_MARK) || ct->mark)
 821             && ctnetlink_dump_mark(skb, ct) < 0)
 822                 goto nla_put_failure;
 823 #endif
 824         nlmsg_end(skb, nlh);
 825         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 826                              GFP_ATOMIC);
 827         if (err == -ENOBUFS || err == -EAGAIN)
 828                 return -ENOBUFS;
 829
 830         return 0;
 831
 832 nla_put_failure:
 833         nlmsg_cancel(skb, nlh);
 834 nlmsg_failure:
 835         kfree_skb(skb);
 836 errout:
 837         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 838                 return -ENOBUFS;
 839
 840         return 0;
 841 }
 842 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 843
 844 static int ctnetlink_done(struct netlink_callback *cb)
 845 {
 846         if (cb->args[1])
 847                 nf_ct_put((struct nf_conn *)cb->args[1]);
 848         kfree(cb->data);
 849         return 0;
 850 }
 851
 852 struct ctnetlink_filter {
 853         u8 family;
 854
 855         u_int32_t orig_flags;
 856         u_int32_t reply_flags;
 857
 858         struct nf_conntrack_tuple orig;
 859         struct nf_conntrack_tuple reply;
 860         struct nf_conntrack_zone zone;
 861
 862         struct {
 863                 u_int32_t val;
 864                 u_int32_t mask;
 865         } mark;
 866 };
 867
 868 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 869         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 870         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 871 };
 872
 873 static int ctnetlink_parse_filter(const struct nlattr *attr,
 874                                   struct ctnetlink_filter *filter)
 875 {
 876         struct nlattr *tb[CTA_FILTER_MAX + 1];
 877         int ret = 0;
 878
 879         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 880                                NULL);
 881         if (ret)
 882                 return ret;
 883
 884         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 885                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 886                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 887                         return -EOPNOTSUPP;
 888         }
 889
 890         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 891                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 892                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 893                         return -EOPNOTSUPP;
 894         }
 895
 896         return 0;
 897 }
 898
 899 static int ctnetlink_parse_zone(const struct nlattr *attr,
 900                                 struct nf_conntrack_zone *zone);
 901 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 902                                          struct nf_conntrack_tuple *tuple,
 903                                          u32 type, u_int8_t l3num,
 904                                          struct nf_conntrack_zone *zone,
 905                                          u_int32_t flags);
 906
 907 static struct ctnetlink_filter *
 908 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 909 {
 910         struct ctnetlink_filter *filter;
 911         int err;
 912
 913 #ifndef CONFIG_NF_CONNTRACK_MARK
 914         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 915                 return ERR_PTR(-EOPNOTSUPP);
 916 #endif
 917
 918         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 919         if (filter == NULL)
 920                 return ERR_PTR(-ENOMEM);
 921
 922         filter->family = family;
 923
 924 #ifdef CONFIG_NF_CONNTRACK_MARK
 925         if (cda[CTA_MARK]) {
 926                 filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 927                 if (cda[CTA_MARK_MASK])
 928                         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 929                 else
 930                         filter->mark.mask = 0xffffffff;
 931         } else if (cda[CTA_MARK_MASK]) {
 932                 err = -EINVAL;
 933                 goto err_filter;
 934         }
 935 #endif
 936         if (!cda[CTA_FILTER])
 937                 return filter;
 938
 939         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 940         if (err < 0)
 941                 goto err_filter;
 942
 943         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
 944         if (err < 0)
 945                 goto err_filter;
 946
 947         if (filter->orig_flags) {
 948                 if (!cda[CTA_TUPLE_ORIG]) {
 949                         err = -EINVAL;
 950                         goto err_filter;
 951                 }
 952
 953                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
 954                                                    CTA_TUPLE_ORIG,
 955                                                    filter->family,
 956                                                    &filter->zone,
 957                                                    filter->orig_flags);
 958                 if (err < 0)
 959                         goto err_filter;
 960         }
 961
 962         if (filter->reply_flags) {
 963                 if (!cda[CTA_TUPLE_REPLY]) {
 964                         err = -EINVAL;
 965                         goto err_filter;
 966                 }
 967
 968                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
 969                                                    CTA_TUPLE_REPLY,
 970                                                    filter->family,
 971                                                    &filter->zone,
 972                                                    filter->orig_flags);
 973                 if (err < 0) {
 974                         err = -EINVAL;
 975                         goto err_filter;
 976                 }
 977         }
 978
 979         return filter;
 980
 981 err_filter:
 982         kfree(filter);
 983
 984         return ERR_PTR(err);
 985 }
 986
 987 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
 988 {
 989         return family || cda[CTA_MARK] || cda[CTA_FILTER];
 990 }
 991
 992 static int ctnetlink_start(struct netlink_callback *cb)
 993 {
 994         const struct nlattr * const *cda = cb->data;
 995         struct ctnetlink_filter *filter = NULL;
 996         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 997         u8 family = nfmsg->nfgen_family;
 998
 999         if (ctnetlink_needs_filter(family, cda)) {
1000                 filter = ctnetlink_alloc_filter(cda, family);
1001                 if (IS_ERR(filter))
1002                         return PTR_ERR(filter);
1003         }
1004
1005         cb->data = filter;
1006         return 0;
1007 }
1008
1009 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1010                                         struct nf_conntrack_tuple *ct_tuple,
1011                                         u_int32_t flags, int family)
1012 {
1013         switch (family) {
1014         case NFPROTO_IPV4:
1015                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1016                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1017                         return  0;
1018
1019                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1020                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1021                         return  0;
1022                 break;
1023         case NFPROTO_IPV6:
1024                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1025                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1026                                    &ct_tuple->src.u3.in6))
1027                         return 0;
1028
1029                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1030                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1031                                    &ct_tuple->dst.u3.in6))
1032                         return 0;
1033                 break;
1034         }
1035
1036         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1037             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1038                 return 0;
1039
1040         switch (ct_tuple->dst.protonum) {
1041         case IPPROTO_TCP:
1042         case IPPROTO_UDP:
1043                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1044                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1045                         return 0;
1046
1047                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1048                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1049                         return 0;
1050                 break;
1051         case IPPROTO_ICMP:
1052                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1053                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1054                         return 0;
1055                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1056                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1057                         return 0;
1058                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1059                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1060                         return 0;
1061                 break;
1062         case IPPROTO_ICMPV6:
1063                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1064                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1065                         return 0;
1066                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1067                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1068                         return 0;
1069                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1070                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1071                         return 0;
1072                 break;
1073         }
1074
1075         return 1;
1076 }
1077
1078 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1079 {
1080         struct ctnetlink_filter *filter = data;
1081         struct nf_conntrack_tuple *tuple;
1082
1083         if (filter == NULL)
1084                 goto out;
1085
1086         /* Match entries of a given L3 protocol number.
1087          * If it is not specified, ie. l3proto == 0,
1088          * then match everything.
1089          */
1090         if (filter->family && nf_ct_l3num(ct) != filter->family)
1091                 goto ignore_entry;
1092
1093         if (filter->orig_flags) {
1094                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1095                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1096                                                   filter->orig_flags,
1097                                                   filter->family))
1098                         goto ignore_entry;
1099         }
1100
1101         if (filter->reply_flags) {
1102                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1103                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1104                                                   filter->reply_flags,
1105                                                   filter->family))
1106                         goto ignore_entry;
1107         }
1108
1109 #ifdef CONFIG_NF_CONNTRACK_MARK
1110         if ((ct->mark & filter->mark.mask) != filter->mark.val)
1111                 goto ignore_entry;
1112 #endif
1113
1114 out:
1115         return 1;
1116
1117 ignore_entry:
1118         return 0;
1119 }
1120
1121 static int
1122 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1123 {
1124         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1125         struct net *net = sock_net(skb->sk);
1126         struct nf_conn *ct, *last;
1127         struct nf_conntrack_tuple_hash *h;
1128         struct hlist_nulls_node *n;
1129         struct nf_conn *nf_ct_evict[8];
1130         int res, i;
1131         spinlock_t *lockp;
1132
1133         last = (struct nf_conn *)cb->args[1];
1134         i = 0;
1135
1136         local_bh_disable();
1137         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1138 restart:
1139                 while (i) {
1140                         i--;
1141                         if (nf_ct_should_gc(nf_ct_evict[i]))
1142                                 nf_ct_kill(nf_ct_evict[i]);
1143                         nf_ct_put(nf_ct_evict[i]);
1144                 }
1145
1146                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1147                 nf_conntrack_lock(lockp);
1148                 if (cb->args[0] >= nf_conntrack_htable_size) {
1149                         spin_unlock(lockp);
1150                         goto out;
1151                 }
1152                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1153                                            hnnode) {
1154                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1155                                 continue;
1156                         ct = nf_ct_tuplehash_to_ctrack(h);
1157                         if (nf_ct_is_expired(ct)) {
1158                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1159                                     atomic_inc_not_zero(&ct->ct_general.use))
1160                                         nf_ct_evict[i++] = ct;
1161                                 continue;
1162                         }
1163
1164                         if (!net_eq(net, nf_ct_net(ct)))
1165                                 continue;
1166
1167                         if (cb->args[1]) {
1168                                 if (ct != last)
1169                                         continue;
1170                                 cb->args[1] = 0;
1171                         }
1172                         if (!ctnetlink_filter_match(ct, cb->data))
1173                                 continue;
1174
1175                         res =
1176                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1177                                             cb->nlh->nlmsg_seq,
1178                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1179                                             ct, true, flags);
1180                         if (res < 0) {
1181                                 nf_conntrack_get(&ct->ct_general);
1182                                 cb->args[1] = (unsigned long)ct;
1183                                 spin_unlock(lockp);
1184                                 goto out;
1185                         }
1186                 }
1187                 spin_unlock(lockp);
1188                 if (cb->args[1]) {
1189                         cb->args[1] = 0;
1190                         goto restart;
1191                 }
1192         }
1193 out:
1194         local_bh_enable();
1195         if (last) {
1196                 /* nf ct hash resize happened, now clear the leftover. */
1197                 if ((struct nf_conn *)cb->args[1] == last)
1198                         cb->args[1] = 0;
1199
1200                 nf_ct_put(last);
1201         }
1202
1203         while (i) {
1204                 i--;
1205                 if (nf_ct_should_gc(nf_ct_evict[i]))
1206                         nf_ct_kill(nf_ct_evict[i]);
1207                 nf_ct_put(nf_ct_evict[i]);
1208         }
1209
1210         return skb->len;
1211 }
1212
1213 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1214                                 struct nf_conntrack_tuple *t,
1215                                 u_int32_t flags)
1216 {
1217         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1218                 if (!tb[CTA_IP_V4_SRC])
1219                         return -EINVAL;
1220
1221                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1222         }
1223
1224         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1225                 if (!tb[CTA_IP_V4_DST])
1226                         return -EINVAL;
1227
1228                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1229         }
1230
1231         return 0;
1232 }
1233
1234 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1235                                 struct nf_conntrack_tuple *t,
1236                                 u_int32_t flags)
1237 {
1238         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1239                 if (!tb[CTA_IP_V6_SRC])
1240                         return -EINVAL;
1241
1242                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1243         }
1244
1245         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1246                 if (!tb[CTA_IP_V6_DST])
1247                         return -EINVAL;
1248
1249                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1250         }
1251
1252         return 0;
1253 }
1254
1255 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1256                                     struct nf_conntrack_tuple *tuple,
1257                                     u_int32_t flags)
1258 {
1259         struct nlattr *tb[CTA_IP_MAX+1];
1260         int ret = 0;
1261
1262         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1263         if (ret < 0)
1264                 return ret;
1265
1266         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1267                                              cta_ip_nla_policy, NULL);
1268         if (ret)
1269                 return ret;
1270
1271         switch (tuple->src.l3num) {
1272         case NFPROTO_IPV4:
1273                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1274                 break;
1275         case NFPROTO_IPV6:
1276                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1277                 break;
1278         }
1279
1280         return ret;
1281 }
1282
1283 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1284         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1285 };
1286
1287 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1288                                        struct nf_conntrack_tuple *tuple,
1289                                        u_int32_t flags)
1290 {
1291         const struct nf_conntrack_l4proto *l4proto;
1292         struct nlattr *tb[CTA_PROTO_MAX+1];
1293         int ret = 0;
1294
1295         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1296                                           proto_nla_policy, NULL);
1297         if (ret < 0)
1298                 return ret;
1299
1300         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1301                 return 0;
1302
1303         if (!tb[CTA_PROTO_NUM])
1304                 return -EINVAL;
1305
1306         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1307
1308         rcu_read_lock();
1309         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1310
1311         if (likely(l4proto->nlattr_to_tuple)) {
1312                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1313                                                      l4proto->nla_policy,
1314                                                      NULL);
1315                 if (ret == 0)
1316                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1317         }
1318
1319         rcu_read_unlock();
1320
1321         return ret;
1322 }
1323
1324 static int
1325 ctnetlink_parse_zone(const struct nlattr *attr,
1326                      struct nf_conntrack_zone *zone)
1327 {
1328         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1329                         NF_CT_DEFAULT_ZONE_DIR, 0);
1330 #ifdef CONFIG_NF_CONNTRACK_ZONES
1331         if (attr)
1332                 zone->id = ntohs(nla_get_be16(attr));
1333 #else
1334         if (attr)
1335                 return -EOPNOTSUPP;
1336 #endif
1337         return 0;
1338 }
1339
1340 static int
1341 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1342                            struct nf_conntrack_zone *zone)
1343 {
1344         int ret;
1345
1346         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1347                 return -EINVAL;
1348
1349         ret = ctnetlink_parse_zone(attr, zone);
1350         if (ret < 0)
1351                 return ret;
1352
1353         if (type == CTA_TUPLE_REPLY)
1354                 zone->dir = NF_CT_ZONE_DIR_REPL;
1355         else
1356                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1357
1358         return 0;
1359 }
1360
1361 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1362         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1363         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1364         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1365 };
1366
1367 #define CTA_FILTER_F_ALL_CTA_PROTO \
1368   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1369    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1370    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1371    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1372    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1373    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1374    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1375    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1376
1377 static int
1378 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1379                               struct nf_conntrack_tuple *tuple, u32 type,
1380                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1381                               u_int32_t flags)
1382 {
1383         struct nlattr *tb[CTA_TUPLE_MAX+1];
1384         int err;
1385
1386         memset(tuple, 0, sizeof(*tuple));
1387
1388         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1389                                           tuple_nla_policy, NULL);
1390         if (err < 0)
1391                 return err;
1392
1393         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1394                 return -EOPNOTSUPP;
1395         tuple->src.l3num = l3num;
1396
1397         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1398             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1399                 if (!tb[CTA_TUPLE_IP])
1400                         return -EINVAL;
1401
1402                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1403                 if (err < 0)
1404                         return err;
1405         }
1406
1407         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1408                 if (!tb[CTA_TUPLE_PROTO])
1409                         return -EINVAL;
1410
1411                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1412                 if (err < 0)
1413                         return err;
1414         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1415                 /* Can't manage proto flags without a protonum  */
1416                 return -EINVAL;
1417         }
1418
1419         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1420                 if (!zone)
1421                         return -EINVAL;
1422
1423                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1424                                                  type, zone);
1425                 if (err < 0)
1426                         return err;
1427         }
1428
1429         /* orig and expect tuples get DIR_ORIGINAL */
1430         if (type == CTA_TUPLE_REPLY)
1431                 tuple->dst.dir = IP_CT_DIR_REPLY;
1432         else
1433                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1434
1435         return 0;
1436 }
1437
1438 static int
1439 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1440                       struct nf_conntrack_tuple *tuple, u32 type,
1441                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1442 {
1443         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1444                                             CTA_FILTER_FLAG(ALL));
1445 }
1446
1447 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1448         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1449                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1450 };
1451
1452 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1453                                 struct nlattr **helpinfo)
1454 {
1455         int err;
1456         struct nlattr *tb[CTA_HELP_MAX+1];
1457
1458         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1459                                           help_nla_policy, NULL);
1460         if (err < 0)
1461                 return err;
1462
1463         if (!tb[CTA_HELP_NAME])
1464                 return -EINVAL;
1465
1466         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1467
1468         if (tb[CTA_HELP_INFO])
1469                 *helpinfo = tb[CTA_HELP_INFO];
1470
1471         return 0;
1472 }
1473
1474 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1475         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1476         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1477         [CTA_STATUS]            = { .type = NLA_U32 },
1478         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1479         [CTA_HELP]              = { .type = NLA_NESTED },
1480         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1481         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1482         [CTA_MARK]              = { .type = NLA_U32 },
1483         [CTA_ID]                = { .type = NLA_U32 },
1484         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1485         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1486         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1487         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1488         [CTA_ZONE]              = { .type = NLA_U16 },
1489         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1490         [CTA_LABELS]            = { .type = NLA_BINARY,
1491                                     .len = NF_CT_LABELS_MAX_SIZE },
1492         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1493                                     .len = NF_CT_LABELS_MAX_SIZE },
1494         [CTA_FILTER]            = { .type = NLA_NESTED },
1495 };
1496
1497 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1498 {
1499         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1500                 return 0;
1501
1502         return ctnetlink_filter_match(ct, data);
1503 }
1504
1505 static int ctnetlink_flush_conntrack(struct net *net,
1506                                      const struct nlattr * const cda[],
1507                                      u32 portid, int report, u8 family)
1508 {
1509         struct ctnetlink_filter *filter = NULL;
1510
1511         if (ctnetlink_needs_filter(family, cda)) {
1512                 if (cda[CTA_FILTER])
1513                         return -EOPNOTSUPP;
1514
1515                 filter = ctnetlink_alloc_filter(cda, family);
1516                 if (IS_ERR(filter))
1517                         return PTR_ERR(filter);
1518         }
1519
1520         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1521                                   portid, report);
1522         kfree(filter);
1523
1524         return 0;
1525 }
1526
1527 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1528                                    const struct nfnl_info *info,
1529                                    const struct nlattr * const cda[])
1530 {
1531         struct nfgenmsg *nfmsg = nlmsg_data(info->nlh);
1532         struct nf_conntrack_tuple_hash *h;
1533         struct nf_conntrack_tuple tuple;
1534         struct nf_conntrack_zone zone;
1535         struct nf_conn *ct;
1536         int err;
1537
1538         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1539         if (err < 0)
1540                 return err;
1541
1542         if (cda[CTA_TUPLE_ORIG])
1543                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1544                                             nfmsg->nfgen_family, &zone);
1545         else if (cda[CTA_TUPLE_REPLY])
1546                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1547                                             nfmsg->nfgen_family, &zone);
1548         else {
1549                 u_int8_t u3 = nfmsg->version ? nfmsg->nfgen_family : AF_UNSPEC;
1550
1551                 return ctnetlink_flush_conntrack(info->net, cda,
1552                                                  NETLINK_CB(skb).portid,
1553                                                  nlmsg_report(info->nlh), u3);
1554         }
1555
1556         if (err < 0)
1557                 return err;
1558
1559         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1560         if (!h)
1561                 return -ENOENT;
1562
1563         ct = nf_ct_tuplehash_to_ctrack(h);
1564
1565         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1566                 nf_ct_put(ct);
1567                 return -EBUSY;
1568         }
1569
1570         if (cda[CTA_ID]) {
1571                 __be32 id = nla_get_be32(cda[CTA_ID]);
1572
1573                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1574                         nf_ct_put(ct);
1575                         return -ENOENT;
1576                 }
1577         }
1578
1579         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1580         nf_ct_put(ct);
1581
1582         return 0;
1583 }
1584
1585 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1586                                    const struct nfnl_info *info,
1587                                    const struct nlattr * const cda[])
1588 {
1589         struct nfgenmsg *nfmsg = nlmsg_data(info->nlh);
1590         u_int8_t u3 = nfmsg->nfgen_family;
1591         struct nf_conntrack_tuple_hash *h;
1592         struct nf_conntrack_tuple tuple;
1593         struct nf_conntrack_zone zone;
1594         struct sk_buff *skb2;
1595         struct nf_conn *ct;
1596         int err;
1597
1598         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1599                 struct netlink_dump_control c = {
1600                         .start = ctnetlink_start,
1601                         .dump = ctnetlink_dump_table,
1602                         .done = ctnetlink_done,
1603                         .data = (void *)cda,
1604                 };
1605
1606                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1607         }
1608
1609         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1610         if (err < 0)
1611                 return err;
1612
1613         if (cda[CTA_TUPLE_ORIG])
1614                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1615                                             u3, &zone);
1616         else if (cda[CTA_TUPLE_REPLY])
1617                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1618                                             u3, &zone);
1619         else
1620                 return -EINVAL;
1621
1622         if (err < 0)
1623                 return err;
1624
1625         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1626         if (!h)
1627                 return -ENOENT;
1628
1629         ct = nf_ct_tuplehash_to_ctrack(h);
1630
1631         err = -ENOMEM;
1632         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1633         if (skb2 == NULL) {
1634                 nf_ct_put(ct);
1635                 return -ENOMEM;
1636         }
1637
1638         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1639                                   info->nlh->nlmsg_seq,
1640                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1641                                   true, 0);
1642         nf_ct_put(ct);
1643         if (err <= 0)
1644                 goto free;
1645
1646         err = netlink_unicast(info->sk, skb2, NETLINK_CB(skb).portid,
1647                               MSG_DONTWAIT);
1648         if (err < 0)
1649                 goto out;
1650
1651         return 0;
1652
1653 free:
1654         kfree_skb(skb2);
1655 out:
1656         /* this avoids a loop in nfnetlink. */
1657         return err == -EAGAIN ? -ENOBUFS : err;
1658 }
1659
1660 static int ctnetlink_done_list(struct netlink_callback *cb)
1661 {
1662         if (cb->args[1])
1663                 nf_ct_put((struct nf_conn *)cb->args[1]);
1664         return 0;
1665 }
1666
1667 static int
1668 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1669 {
1670         struct nf_conn *ct, *last;
1671         struct nf_conntrack_tuple_hash *h;
1672         struct hlist_nulls_node *n;
1673         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1674         u_int8_t l3proto = nfmsg->nfgen_family;
1675         int res;
1676         int cpu;
1677         struct hlist_nulls_head *list;
1678         struct net *net = sock_net(skb->sk);
1679
1680         if (cb->args[2])
1681                 return 0;
1682
1683         last = (struct nf_conn *)cb->args[1];
1684
1685         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1686                 struct ct_pcpu *pcpu;
1687
1688                 if (!cpu_possible(cpu))
1689                         continue;
1690
1691                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1692                 spin_lock_bh(&pcpu->lock);
1693                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1694 restart:
1695                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1696                         ct = nf_ct_tuplehash_to_ctrack(h);
1697                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1698                                 continue;
1699                         if (cb->args[1]) {
1700                                 if (ct != last)
1701                                         continue;
1702                                 cb->args[1] = 0;
1703                         }
1704
1705                         /* We can't dump extension info for the unconfirmed
1706                          * list because unconfirmed conntracks can have
1707                          * ct->ext reallocated (and thus freed).
1708                          *
1709                          * In the dying list case ct->ext can't be free'd
1710                          * until after we drop pcpu->lock.
1711                          */
1712                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1713                                                   cb->nlh->nlmsg_seq,
1714                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1715                                                   ct, dying ? true : false, 0);
1716                         if (res < 0) {
1717                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1718                                         continue;
1719                                 cb->args[0] = cpu;
1720                                 cb->args[1] = (unsigned long)ct;
1721                                 spin_unlock_bh(&pcpu->lock);
1722                                 goto out;
1723                         }
1724                 }
1725                 if (cb->args[1]) {
1726                         cb->args[1] = 0;
1727                         goto restart;
1728                 }
1729                 spin_unlock_bh(&pcpu->lock);
1730         }
1731         cb->args[2] = 1;
1732 out:
1733         if (last)
1734                 nf_ct_put(last);
1735
1736         return skb->len;
1737 }
1738
1739 static int
1740 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1741 {
1742         return ctnetlink_dump_list(skb, cb, true);
1743 }
1744
1745 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1746                                   const struct nfnl_info *info,
1747                                   const struct nlattr * const cda[])
1748 {
1749         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1750                 struct netlink_dump_control c = {
1751                         .dump = ctnetlink_dump_dying,
1752                         .done = ctnetlink_done_list,
1753                 };
1754                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1755         }
1756
1757         return -EOPNOTSUPP;
1758 }
1759
1760 static int
1761 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1762 {
1763         return ctnetlink_dump_list(skb, cb, false);
1764 }
1765
1766 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1767                                         const struct nfnl_info *info,
1768                                         const struct nlattr * const cda[])
1769 {
1770         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1771                 struct netlink_dump_control c = {
1772                         .dump = ctnetlink_dump_unconfirmed,
1773                         .done = ctnetlink_done_list,
1774                 };
1775                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1776         }
1777
1778         return -EOPNOTSUPP;
1779 }
1780
1781 #if IS_ENABLED(CONFIG_NF_NAT)
1782 static int
1783 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1784                           enum nf_nat_manip_type manip,
1785                           const struct nlattr *attr)
1786         __must_hold(RCU)
1787 {
1788         struct nf_nat_hook *nat_hook;
1789         int err;
1790
1791         nat_hook = rcu_dereference(nf_nat_hook);
1792         if (!nat_hook) {
1793 #ifdef CONFIG_MODULES
1794                 rcu_read_unlock();
1795                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1796                 if (request_module("nf-nat") < 0) {
1797                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1798                         rcu_read_lock();
1799                         return -EOPNOTSUPP;
1800                 }
1801                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1802                 rcu_read_lock();
1803                 nat_hook = rcu_dereference(nf_nat_hook);
1804                 if (nat_hook)
1805                         return -EAGAIN;
1806 #endif
1807                 return -EOPNOTSUPP;
1808         }
1809
1810         err = nat_hook->parse_nat_setup(ct, manip, attr);
1811         if (err == -EAGAIN) {
1812 #ifdef CONFIG_MODULES
1813                 rcu_read_unlock();
1814                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1815                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1816                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1817                         rcu_read_lock();
1818                         return -EOPNOTSUPP;
1819                 }
1820                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1821                 rcu_read_lock();
1822 #else
1823                 err = -EOPNOTSUPP;
1824 #endif
1825         }
1826         return err;
1827 }
1828 #endif
1829
1830 static void
1831 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1832                           unsigned long off)
1833 {
1834         unsigned int bit;
1835
1836         /* Ignore these unchangable bits */
1837         on &= ~IPS_UNCHANGEABLE_MASK;
1838         off &= ~IPS_UNCHANGEABLE_MASK;
1839
1840         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1841                 if (on & (1 << bit))
1842                         set_bit(bit, &ct->status);
1843                 else if (off & (1 << bit))
1844                         clear_bit(bit, &ct->status);
1845         }
1846 }
1847
1848 static int
1849 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1850 {
1851         unsigned long d;
1852         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1853         d = ct->status ^ status;
1854
1855         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1856                 /* unchangeable */
1857                 return -EBUSY;
1858
1859         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1860                 /* SEEN_REPLY bit can only be set */
1861                 return -EBUSY;
1862
1863         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1864                 /* ASSURED bit can only be set */
1865                 return -EBUSY;
1866
1867         __ctnetlink_change_status(ct, status, 0);
1868         return 0;
1869 }
1870
1871 static int
1872 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1873 {
1874 #if IS_ENABLED(CONFIG_NF_NAT)
1875         int ret;
1876
1877         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1878                 return 0;
1879
1880         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1881                                         cda[CTA_NAT_DST]);
1882         if (ret < 0)
1883                 return ret;
1884
1885         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1886                                          cda[CTA_NAT_SRC]);
1887 #else
1888         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1889                 return 0;
1890         return -EOPNOTSUPP;
1891 #endif
1892 }
1893
1894 static int ctnetlink_change_helper(struct nf_conn *ct,
1895                                    const struct nlattr * const cda[])
1896 {
1897         struct nf_conntrack_helper *helper;
1898         struct nf_conn_help *help = nfct_help(ct);
1899         char *helpname = NULL;
1900         struct nlattr *helpinfo = NULL;
1901         int err;
1902
1903         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1904         if (err < 0)
1905                 return err;
1906
1907         /* don't change helper of sibling connections */
1908         if (ct->master) {
1909                 /* If we try to change the helper to the same thing twice,
1910                  * treat the second attempt as a no-op instead of returning
1911                  * an error.
1912                  */
1913                 err = -EBUSY;
1914                 if (help) {
1915                         rcu_read_lock();
1916                         helper = rcu_dereference(help->helper);
1917                         if (helper && !strcmp(helper->name, helpname))
1918                                 err = 0;
1919                         rcu_read_unlock();
1920                 }
1921
1922                 return err;
1923         }
1924
1925         if (!strcmp(helpname, "")) {
1926                 if (help && help->helper) {
1927                         /* we had a helper before ... */
1928                         nf_ct_remove_expectations(ct);
1929                         RCU_INIT_POINTER(help->helper, NULL);
1930                 }
1931
1932                 return 0;
1933         }
1934
1935         rcu_read_lock();
1936         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1937                                             nf_ct_protonum(ct));
1938         if (helper == NULL) {
1939                 rcu_read_unlock();
1940                 return -EOPNOTSUPP;
1941         }
1942
1943         if (help) {
1944                 if (help->helper == helper) {
1945                         /* update private helper data if allowed. */
1946                         if (helper->from_nlattr)
1947                                 helper->from_nlattr(helpinfo, ct);
1948                         err = 0;
1949                 } else
1950                         err = -EBUSY;
1951         } else {
1952                 /* we cannot set a helper for an existing conntrack */
1953                 err = -EOPNOTSUPP;
1954         }
1955
1956         rcu_read_unlock();
1957         return err;
1958 }
1959
1960 static int ctnetlink_change_timeout(struct nf_conn *ct,
1961                                     const struct nlattr * const cda[])
1962 {
1963         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1964
1965         if (timeout > INT_MAX)
1966                 timeout = INT_MAX;
1967         ct->timeout = nfct_time_stamp + (u32)timeout;
1968
1969         if (test_bit(IPS_DYING_BIT, &ct->status))
1970                 return -ETIME;
1971
1972         return 0;
1973 }
1974
1975 #if defined(CONFIG_NF_CONNTRACK_MARK)
1976 static void ctnetlink_change_mark(struct nf_conn *ct,
1977                                     const struct nlattr * const cda[])
1978 {
1979         u32 mark, newmark, mask = 0;
1980
1981         if (cda[CTA_MARK_MASK])
1982                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1983
1984         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1985         newmark = (ct->mark & mask) ^ mark;
1986         if (newmark != ct->mark)
1987                 ct->mark = newmark;
1988 }
1989 #endif
1990
1991 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1992         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1993         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1994         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1995 };
1996
1997 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1998                                       const struct nlattr * const cda[])
1999 {
2000         const struct nlattr *attr = cda[CTA_PROTOINFO];
2001         const struct nf_conntrack_l4proto *l4proto;
2002         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2003         int err = 0;
2004
2005         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2006                                           protoinfo_policy, NULL);
2007         if (err < 0)
2008                 return err;
2009
2010         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2011         if (l4proto->from_nlattr)
2012                 err = l4proto->from_nlattr(tb, ct);
2013
2014         return err;
2015 }
2016
2017 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2018         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2019         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2020         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2021 };
2022
2023 static int change_seq_adj(struct nf_ct_seqadj *seq,
2024                           const struct nlattr * const attr)
2025 {
2026         int err;
2027         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2028
2029         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2030                                           seqadj_policy, NULL);
2031         if (err < 0)
2032                 return err;
2033
2034         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2035                 return -EINVAL;
2036
2037         seq->correction_pos =
2038                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2039
2040         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2041                 return -EINVAL;
2042
2043         seq->offset_before =
2044                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2045
2046         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2047                 return -EINVAL;
2048
2049         seq->offset_after =
2050                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2051
2052         return 0;
2053 }
2054
2055 static int
2056 ctnetlink_change_seq_adj(struct nf_conn *ct,
2057                          const struct nlattr * const cda[])
2058 {
2059         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2060         int ret = 0;
2061
2062         if (!seqadj)
2063                 return 0;
2064
2065         spin_lock_bh(&ct->lock);
2066         if (cda[CTA_SEQ_ADJ_ORIG]) {
2067                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2068                                      cda[CTA_SEQ_ADJ_ORIG]);
2069                 if (ret < 0)
2070                         goto err;
2071
2072                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2073         }
2074
2075         if (cda[CTA_SEQ_ADJ_REPLY]) {
2076                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2077                                      cda[CTA_SEQ_ADJ_REPLY]);
2078                 if (ret < 0)
2079                         goto err;
2080
2081                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2082         }
2083
2084         spin_unlock_bh(&ct->lock);
2085         return 0;
2086 err:
2087         spin_unlock_bh(&ct->lock);
2088         return ret;
2089 }
2090
2091 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2092         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2093         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2094         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2095 };
2096
2097 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2098                                      const struct nlattr * const cda[])
2099 {
2100         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2101         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2102         int err;
2103
2104         if (!synproxy)
2105                 return 0;
2106
2107         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2108                                           cda[CTA_SYNPROXY], synproxy_policy,
2109                                           NULL);
2110         if (err < 0)
2111                 return err;
2112
2113         if (!tb[CTA_SYNPROXY_ISN] ||
2114             !tb[CTA_SYNPROXY_ITS] ||
2115             !tb[CTA_SYNPROXY_TSOFF])
2116                 return -EINVAL;
2117
2118         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2119         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2120         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2121
2122         return 0;
2123 }
2124
2125 static int
2126 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2127 {
2128 #ifdef CONFIG_NF_CONNTRACK_LABELS
2129         size_t len = nla_len(cda[CTA_LABELS]);
2130         const void *mask = cda[CTA_LABELS_MASK];
2131
2132         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2133                 return -EINVAL;
2134
2135         if (mask) {
2136                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2137                     nla_len(cda[CTA_LABELS_MASK]) != len)
2138                         return -EINVAL;
2139                 mask = nla_data(cda[CTA_LABELS_MASK]);
2140         }
2141
2142         len /= sizeof(u32);
2143
2144         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2145 #else
2146         return -EOPNOTSUPP;
2147 #endif
2148 }
2149
2150 static int
2151 ctnetlink_change_conntrack(struct nf_conn *ct,
2152                            const struct nlattr * const cda[])
2153 {
2154         int err;
2155
2156         /* only allow NAT changes and master assignation for new conntracks */
2157         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2158                 return -EOPNOTSUPP;
2159
2160         if (cda[CTA_HELP]) {
2161                 err = ctnetlink_change_helper(ct, cda);
2162                 if (err < 0)
2163                         return err;
2164         }
2165
2166         if (cda[CTA_TIMEOUT]) {
2167                 err = ctnetlink_change_timeout(ct, cda);
2168                 if (err < 0)
2169                         return err;
2170         }
2171
2172         if (cda[CTA_STATUS]) {
2173                 err = ctnetlink_change_status(ct, cda);
2174                 if (err < 0)
2175                         return err;
2176         }
2177
2178         if (cda[CTA_PROTOINFO]) {
2179                 err = ctnetlink_change_protoinfo(ct, cda);
2180                 if (err < 0)
2181                         return err;
2182         }
2183
2184 #if defined(CONFIG_NF_CONNTRACK_MARK)
2185         if (cda[CTA_MARK])
2186                 ctnetlink_change_mark(ct, cda);
2187 #endif
2188
2189         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2190                 err = ctnetlink_change_seq_adj(ct, cda);
2191                 if (err < 0)
2192                         return err;
2193         }
2194
2195         if (cda[CTA_SYNPROXY]) {
2196                 err = ctnetlink_change_synproxy(ct, cda);
2197                 if (err < 0)
2198                         return err;
2199         }
2200
2201         if (cda[CTA_LABELS]) {
2202                 err = ctnetlink_attach_labels(ct, cda);
2203                 if (err < 0)
2204                         return err;
2205         }
2206
2207         return 0;
2208 }
2209
2210 static struct nf_conn *
2211 ctnetlink_create_conntrack(struct net *net,
2212                            const struct nf_conntrack_zone *zone,
2213                            const struct nlattr * const cda[],
2214                            struct nf_conntrack_tuple *otuple,
2215                            struct nf_conntrack_tuple *rtuple,
2216                            u8 u3)
2217 {
2218         struct nf_conn *ct;
2219         int err = -EINVAL;
2220         struct nf_conntrack_helper *helper;
2221         struct nf_conn_tstamp *tstamp;
2222         u64 timeout;
2223
2224         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2225         if (IS_ERR(ct))
2226                 return ERR_PTR(-ENOMEM);
2227
2228         if (!cda[CTA_TIMEOUT])
2229                 goto err1;
2230
2231         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2232         if (timeout > INT_MAX)
2233                 timeout = INT_MAX;
2234         ct->timeout = (u32)timeout + nfct_time_stamp;
2235
2236         rcu_read_lock();
2237         if (cda[CTA_HELP]) {
2238                 char *helpname = NULL;
2239                 struct nlattr *helpinfo = NULL;
2240
2241                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2242                 if (err < 0)
2243                         goto err2;
2244
2245                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2246                                                     nf_ct_protonum(ct));
2247                 if (helper == NULL) {
2248                         rcu_read_unlock();
2249 #ifdef CONFIG_MODULES
2250                         if (request_module("nfct-helper-%s", helpname) < 0) {
2251                                 err = -EOPNOTSUPP;
2252                                 goto err1;
2253                         }
2254
2255                         rcu_read_lock();
2256                         helper = __nf_conntrack_helper_find(helpname,
2257                                                             nf_ct_l3num(ct),
2258                                                             nf_ct_protonum(ct));
2259                         if (helper) {
2260                                 err = -EAGAIN;
2261                                 goto err2;
2262                         }
2263                         rcu_read_unlock();
2264 #endif
2265                         err = -EOPNOTSUPP;
2266                         goto err1;
2267                 } else {
2268                         struct nf_conn_help *help;
2269
2270                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2271                         if (help == NULL) {
2272                                 err = -ENOMEM;
2273                                 goto err2;
2274                         }
2275                         /* set private helper data if allowed. */
2276                         if (helper->from_nlattr)
2277                                 helper->from_nlattr(helpinfo, ct);
2278
2279                         /* not in hash table yet so not strictly necessary */
2280                         RCU_INIT_POINTER(help->helper, helper);
2281                 }
2282         } else {
2283                 /* try an implicit helper assignation */
2284                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2285                 if (err < 0)
2286                         goto err2;
2287         }
2288
2289         err = ctnetlink_setup_nat(ct, cda);
2290         if (err < 0)
2291                 goto err2;
2292
2293         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2294         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2295         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2296         nf_ct_labels_ext_add(ct);
2297         nfct_seqadj_ext_add(ct);
2298         nfct_synproxy_ext_add(ct);
2299
2300         /* we must add conntrack extensions before confirmation. */
2301         ct->status |= IPS_CONFIRMED;
2302
2303         if (cda[CTA_STATUS]) {
2304                 err = ctnetlink_change_status(ct, cda);
2305                 if (err < 0)
2306                         goto err2;
2307         }
2308
2309         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2310                 err = ctnetlink_change_seq_adj(ct, cda);
2311                 if (err < 0)
2312                         goto err2;
2313         }
2314
2315         memset(&ct->proto, 0, sizeof(ct->proto));
2316         if (cda[CTA_PROTOINFO]) {
2317                 err = ctnetlink_change_protoinfo(ct, cda);
2318                 if (err < 0)
2319                         goto err2;
2320         }
2321
2322         if (cda[CTA_SYNPROXY]) {
2323                 err = ctnetlink_change_synproxy(ct, cda);
2324                 if (err < 0)
2325                         goto err2;
2326         }
2327
2328 #if defined(CONFIG_NF_CONNTRACK_MARK)
2329         if (cda[CTA_MARK])
2330                 ctnetlink_change_mark(ct, cda);
2331 #endif
2332
2333         /* setup master conntrack: this is a confirmed expectation */
2334         if (cda[CTA_TUPLE_MASTER]) {
2335                 struct nf_conntrack_tuple master;
2336                 struct nf_conntrack_tuple_hash *master_h;
2337                 struct nf_conn *master_ct;
2338
2339                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2340                                             u3, NULL);
2341                 if (err < 0)
2342                         goto err2;
2343
2344                 master_h = nf_conntrack_find_get(net, zone, &master);
2345                 if (master_h == NULL) {
2346                         err = -ENOENT;
2347                         goto err2;
2348                 }
2349                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2350                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2351                 ct->master = master_ct;
2352         }
2353         tstamp = nf_conn_tstamp_find(ct);
2354         if (tstamp)
2355                 tstamp->start = ktime_get_real_ns();
2356
2357         err = nf_conntrack_hash_check_insert(ct);
2358         if (err < 0)
2359                 goto err2;
2360
2361         rcu_read_unlock();
2362
2363         return ct;
2364
2365 err2:
2366         rcu_read_unlock();
2367 err1:
2368         nf_conntrack_free(ct);
2369         return ERR_PTR(err);
2370 }
2371
2372 static int ctnetlink_new_conntrack(struct sk_buff *skb,
2373                                    const struct nfnl_info *info,
2374                                    const struct nlattr * const cda[])
2375 {
2376         struct nfgenmsg *nfmsg = nlmsg_data(info->nlh);
2377         struct nf_conntrack_tuple otuple, rtuple;
2378         struct nf_conntrack_tuple_hash *h = NULL;
2379         u_int8_t u3 = nfmsg->nfgen_family;
2380         struct nf_conntrack_zone zone;
2381         struct nf_conn *ct;
2382         int err;
2383
2384         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2385         if (err < 0)
2386                 return err;
2387
2388         if (cda[CTA_TUPLE_ORIG]) {
2389                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2390                                             u3, &zone);
2391                 if (err < 0)
2392                         return err;
2393         }
2394
2395         if (cda[CTA_TUPLE_REPLY]) {
2396                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2397                                             u3, &zone);
2398                 if (err < 0)
2399                         return err;
2400         }
2401
2402         if (cda[CTA_TUPLE_ORIG])
2403                 h = nf_conntrack_find_get(info->net, &zone, &otuple);
2404         else if (cda[CTA_TUPLE_REPLY])
2405                 h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2406
2407         if (h == NULL) {
2408                 err = -ENOENT;
2409                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2410                         enum ip_conntrack_events events;
2411
2412                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2413                                 return -EINVAL;
2414                         if (otuple.dst.protonum != rtuple.dst.protonum)
2415                                 return -EINVAL;
2416
2417                         ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2418                                                         &otuple, &rtuple, u3);
2419                         if (IS_ERR(ct))
2420                                 return PTR_ERR(ct);
2421
2422                         err = 0;
2423                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2424                                 events = 1 << IPCT_RELATED;
2425                         else
2426                                 events = 1 << IPCT_NEW;
2427
2428                         if (cda[CTA_LABELS] &&
2429                             ctnetlink_attach_labels(ct, cda) == 0)
2430                                 events |= (1 << IPCT_LABEL);
2431
2432                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2433                                                       (1 << IPCT_ASSURED) |
2434                                                       (1 << IPCT_HELPER) |
2435                                                       (1 << IPCT_PROTOINFO) |
2436                                                       (1 << IPCT_SEQADJ) |
2437                                                       (1 << IPCT_MARK) |
2438                                                       (1 << IPCT_SYNPROXY) |
2439                                                       events,
2440                                                       ct, NETLINK_CB(skb).portid,
2441                                                       nlmsg_report(info->nlh));
2442                         nf_ct_put(ct);
2443                 }
2444
2445                 return err;
2446         }
2447         /* implicit 'else' */
2448
2449         err = -EEXIST;
2450         ct = nf_ct_tuplehash_to_ctrack(h);
2451         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2452                 err = ctnetlink_change_conntrack(ct, cda);
2453                 if (err == 0) {
2454                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2455                                                       (1 << IPCT_ASSURED) |
2456                                                       (1 << IPCT_HELPER) |
2457                                                       (1 << IPCT_LABEL) |
2458                                                       (1 << IPCT_PROTOINFO) |
2459                                                       (1 << IPCT_SEQADJ) |
2460                                                       (1 << IPCT_MARK) |
2461                                                       (1 << IPCT_SYNPROXY),
2462                                                       ct, NETLINK_CB(skb).portid,
2463                                                       nlmsg_report(info->nlh));
2464                 }
2465         }
2466
2467         nf_ct_put(ct);
2468         return err;
2469 }
2470
2471 static int
2472 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2473                                 __u16 cpu, const struct ip_conntrack_stat *st)
2474 {
2475         struct nlmsghdr *nlh;
2476         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2477
2478         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2479                               IPCTNL_MSG_CT_GET_STATS_CPU);
2480         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2481                            NFNETLINK_V0, htons(cpu));
2482         if (!nlh)
2483                 goto nlmsg_failure;
2484
2485         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2486             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2487             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2488             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2489                                 htonl(st->insert_failed)) ||
2490             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2491             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2492             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2493             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2494                                 htonl(st->search_restart)) ||
2495             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2496                                 htonl(st->clash_resolve)))
2497                 goto nla_put_failure;
2498
2499         nlmsg_end(skb, nlh);
2500         return skb->len;
2501
2502 nla_put_failure:
2503 nlmsg_failure:
2504         nlmsg_cancel(skb, nlh);
2505         return -1;
2506 }
2507
2508 static int
2509 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2510 {
2511         int cpu;
2512         struct net *net = sock_net(skb->sk);
2513
2514         if (cb->args[0] == nr_cpu_ids)
2515                 return 0;
2516
2517         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2518                 const struct ip_conntrack_stat *st;
2519
2520                 if (!cpu_possible(cpu))
2521                         continue;
2522
2523                 st = per_cpu_ptr(net->ct.stat, cpu);
2524                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2525                                                     NETLINK_CB(cb->skb).portid,
2526                                                     cb->nlh->nlmsg_seq,
2527                                                     cpu, st) < 0)
2528                                 break;
2529         }
2530         cb->args[0] = cpu;
2531
2532         return skb->len;
2533 }
2534
2535 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2536                                  const struct nfnl_info *info,
2537                                  const struct nlattr * const cda[])
2538 {
2539         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2540                 struct netlink_dump_control c = {
2541                         .dump = ctnetlink_ct_stat_cpu_dump,
2542                 };
2543                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2544         }
2545
2546         return 0;
2547 }
2548
2549 static int
2550 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2551                             struct net *net)
2552 {
2553         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2554         unsigned int nr_conntracks;
2555         struct nlmsghdr *nlh;
2556
2557         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2558         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2559                            NFNETLINK_V0, 0);
2560         if (!nlh)
2561                 goto nlmsg_failure;
2562
2563         nr_conntracks = nf_conntrack_count(net);
2564         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2565                 goto nla_put_failure;
2566
2567         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2568                 goto nla_put_failure;
2569
2570         nlmsg_end(skb, nlh);
2571         return skb->len;
2572
2573 nla_put_failure:
2574 nlmsg_failure:
2575         nlmsg_cancel(skb, nlh);
2576         return -1;
2577 }
2578
2579 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2580                              const struct nlattr * const cda[])
2581 {
2582         struct sk_buff *skb2;
2583         int err;
2584
2585         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2586         if (skb2 == NULL)
2587                 return -ENOMEM;
2588
2589         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2590                                           info->nlh->nlmsg_seq,
2591                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2592                                           sock_net(skb->sk));
2593         if (err <= 0)
2594                 goto free;
2595
2596         err = netlink_unicast(info->sk, skb2, NETLINK_CB(skb).portid,
2597                               MSG_DONTWAIT);
2598         if (err < 0)
2599                 goto out;
2600
2601         return 0;
2602
2603 free:
2604         kfree_skb(skb2);
2605 out:
2606         /* this avoids a loop in nfnetlink. */
2607         return err == -EAGAIN ? -ENOBUFS : err;
2608 }
2609
2610 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2611         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2612         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2613         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2614         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2615         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2616         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2617                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2618         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2619         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2620         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2621         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2622         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2623 };
2624
2625 static struct nf_conntrack_expect *
2626 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2627                        struct nf_conntrack_helper *helper,
2628                        struct nf_conntrack_tuple *tuple,
2629                        struct nf_conntrack_tuple *mask);
2630
2631 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2632 static size_t
2633 ctnetlink_glue_build_size(const struct nf_conn *ct)
2634 {
2635         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2636                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2637                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2638                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2639                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2640                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2641                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2642                + nla_total_size(0) /* CTA_PROTOINFO */
2643                + nla_total_size(0) /* CTA_HELP */
2644                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2645                + ctnetlink_secctx_size(ct)
2646 #if IS_ENABLED(CONFIG_NF_NAT)
2647                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2648                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2649 #endif
2650 #ifdef CONFIG_NF_CONNTRACK_MARK
2651                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2652 #endif
2653 #ifdef CONFIG_NF_CONNTRACK_ZONES
2654                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2655 #endif
2656                + ctnetlink_proto_size(ct)
2657                ;
2658 }
2659
2660 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2661 {
2662         const struct nf_conntrack_zone *zone;
2663         struct nlattr *nest_parms;
2664
2665         zone = nf_ct_zone(ct);
2666
2667         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2668         if (!nest_parms)
2669                 goto nla_put_failure;
2670         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2671                 goto nla_put_failure;
2672         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2673                                    NF_CT_ZONE_DIR_ORIG) < 0)
2674                 goto nla_put_failure;
2675         nla_nest_end(skb, nest_parms);
2676
2677         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2678         if (!nest_parms)
2679                 goto nla_put_failure;
2680         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2681                 goto nla_put_failure;
2682         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2683                                    NF_CT_ZONE_DIR_REPL) < 0)
2684                 goto nla_put_failure;
2685         nla_nest_end(skb, nest_parms);
2686
2687         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2688                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2689                 goto nla_put_failure;
2690
2691         if (ctnetlink_dump_id(skb, ct) < 0)
2692                 goto nla_put_failure;
2693
2694         if (ctnetlink_dump_status(skb, ct) < 0)
2695                 goto nla_put_failure;
2696
2697         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2698                 goto nla_put_failure;
2699
2700         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2701                 goto nla_put_failure;
2702
2703         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2704                 goto nla_put_failure;
2705
2706 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2707         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2708                 goto nla_put_failure;
2709 #endif
2710         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2711                 goto nla_put_failure;
2712
2713         if ((ct->status & IPS_SEQ_ADJUST) &&
2714             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2715                 goto nla_put_failure;
2716
2717         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2718                 goto nla_put_failure;
2719
2720 #ifdef CONFIG_NF_CONNTRACK_MARK
2721         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2722                 goto nla_put_failure;
2723 #endif
2724         if (ctnetlink_dump_labels(skb, ct) < 0)
2725                 goto nla_put_failure;
2726         return 0;
2727
2728 nla_put_failure:
2729         return -ENOSPC;
2730 }
2731
2732 static int
2733 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2734                      enum ip_conntrack_info ctinfo,
2735                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2736 {
2737         struct nlattr *nest_parms;
2738
2739         nest_parms = nla_nest_start(skb, ct_attr);
2740         if (!nest_parms)
2741                 goto nla_put_failure;
2742
2743         if (__ctnetlink_glue_build(skb, ct) < 0)
2744                 goto nla_put_failure;
2745
2746         nla_nest_end(skb, nest_parms);
2747
2748         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2749                 goto nla_put_failure;
2750
2751         return 0;
2752
2753 nla_put_failure:
2754         return -ENOSPC;
2755 }
2756
2757 static int
2758 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2759 {
2760         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2761         unsigned long d = ct->status ^ status;
2762
2763         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2764                 /* SEEN_REPLY bit can only be set */
2765                 return -EBUSY;
2766
2767         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2768                 /* ASSURED bit can only be set */
2769                 return -EBUSY;
2770
2771         /* This check is less strict than ctnetlink_change_status()
2772          * because callers often flip IPS_EXPECTED bits when sending
2773          * an NFQA_CT attribute to the kernel.  So ignore the
2774          * unchangeable bits but do not error out. Also user programs
2775          * are allowed to clear the bits that they are allowed to change.
2776          */
2777         __ctnetlink_change_status(ct, status, ~status);
2778         return 0;
2779 }
2780
2781 static int
2782 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2783 {
2784         int err;
2785
2786         if (cda[CTA_TIMEOUT]) {
2787                 err = ctnetlink_change_timeout(ct, cda);
2788                 if (err < 0)
2789                         return err;
2790         }
2791         if (cda[CTA_STATUS]) {
2792                 err = ctnetlink_update_status(ct, cda);
2793                 if (err < 0)
2794                         return err;
2795         }
2796         if (cda[CTA_HELP]) {
2797                 err = ctnetlink_change_helper(ct, cda);
2798                 if (err < 0)
2799                         return err;
2800         }
2801         if (cda[CTA_LABELS]) {
2802                 err = ctnetlink_attach_labels(ct, cda);
2803                 if (err < 0)
2804                         return err;
2805         }
2806 #if defined(CONFIG_NF_CONNTRACK_MARK)
2807         if (cda[CTA_MARK]) {
2808                 ctnetlink_change_mark(ct, cda);
2809         }
2810 #endif
2811         return 0;
2812 }
2813
2814 static int
2815 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2816 {
2817         struct nlattr *cda[CTA_MAX+1];
2818         int ret;
2819
2820         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2821                                           NULL);
2822         if (ret < 0)
2823                 return ret;
2824
2825         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2826 }
2827
2828 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2829                                     const struct nf_conn *ct,
2830                                     struct nf_conntrack_tuple *tuple,
2831                                     struct nf_conntrack_tuple *mask)
2832 {
2833         int err;
2834
2835         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2836                                     nf_ct_l3num(ct), NULL);
2837         if (err < 0)
2838                 return err;
2839
2840         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2841                                      nf_ct_l3num(ct), NULL);
2842 }
2843
2844 static int
2845 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2846                              u32 portid, u32 report)
2847 {
2848         struct nlattr *cda[CTA_EXPECT_MAX+1];
2849         struct nf_conntrack_tuple tuple, mask;
2850         struct nf_conntrack_helper *helper = NULL;
2851         struct nf_conntrack_expect *exp;
2852         int err;
2853
2854         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2855                                           exp_nla_policy, NULL);
2856         if (err < 0)
2857                 return err;
2858
2859         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2860                                        ct, &tuple, &mask);
2861         if (err < 0)
2862                 return err;
2863
2864         if (cda[CTA_EXPECT_HELP_NAME]) {
2865                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2866
2867                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2868                                                     nf_ct_protonum(ct));
2869                 if (helper == NULL)
2870                         return -EOPNOTSUPP;
2871         }
2872
2873         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2874                                      helper, &tuple, &mask);
2875         if (IS_ERR(exp))
2876                 return PTR_ERR(exp);
2877
2878         err = nf_ct_expect_related_report(exp, portid, report, 0);
2879         nf_ct_expect_put(exp);
2880         return err;
2881 }
2882
2883 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2884                                   enum ip_conntrack_info ctinfo, int diff)
2885 {
2886         if (!(ct->status & IPS_NAT_MASK))
2887                 return;
2888
2889         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2890 }
2891
2892 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2893         .build_size     = ctnetlink_glue_build_size,
2894         .build          = ctnetlink_glue_build,
2895         .parse          = ctnetlink_glue_parse,
2896         .attach_expect  = ctnetlink_glue_attach_expect,
2897         .seq_adjust     = ctnetlink_glue_seqadj,
2898 };
2899 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2900
2901 /***********************************************************************
2902  * EXPECT
2903  ***********************************************************************/
2904
2905 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2906                                     const struct nf_conntrack_tuple *tuple,
2907                                     u32 type)
2908 {
2909         struct nlattr *nest_parms;
2910
2911         nest_parms = nla_nest_start(skb, type);
2912         if (!nest_parms)
2913                 goto nla_put_failure;
2914         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2915                 goto nla_put_failure;
2916         nla_nest_end(skb, nest_parms);
2917
2918         return 0;
2919
2920 nla_put_failure:
2921         return -1;
2922 }
2923
2924 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2925                                    const struct nf_conntrack_tuple *tuple,
2926                                    const struct nf_conntrack_tuple_mask *mask)
2927 {
2928         const struct nf_conntrack_l4proto *l4proto;
2929         struct nf_conntrack_tuple m;
2930         struct nlattr *nest_parms;
2931         int ret;
2932
2933         memset(&m, 0xFF, sizeof(m));
2934         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2935         m.src.u.all = mask->src.u.all;
2936         m.src.l3num = tuple->src.l3num;
2937         m.dst.protonum = tuple->dst.protonum;
2938
2939         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2940         if (!nest_parms)
2941                 goto nla_put_failure;
2942
2943         rcu_read_lock();
2944         ret = ctnetlink_dump_tuples_ip(skb, &m);
2945         if (ret >= 0) {
2946                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2947                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2948         }
2949         rcu_read_unlock();
2950
2951         if (unlikely(ret < 0))
2952                 goto nla_put_failure;
2953
2954         nla_nest_end(skb, nest_parms);
2955
2956         return 0;
2957
2958 nla_put_failure:
2959         return -1;
2960 }
2961
2962 static const union nf_inet_addr any_addr;
2963
2964 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2965 {
2966         static __read_mostly siphash_key_t exp_id_seed;
2967         unsigned long a, b, c, d;
2968
2969         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2970
2971         a = (unsigned long)exp;
2972         b = (unsigned long)exp->helper;
2973         c = (unsigned long)exp->master;
2974         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2975
2976 #ifdef CONFIG_64BIT
2977         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2978 #else
2979         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2980 #endif
2981 }
2982
2983 static int
2984 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2985                           const struct nf_conntrack_expect *exp)
2986 {
2987         struct nf_conn *master = exp->master;
2988         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2989         struct nf_conn_help *help;
2990 #if IS_ENABLED(CONFIG_NF_NAT)
2991         struct nlattr *nest_parms;
2992         struct nf_conntrack_tuple nat_tuple = {};
2993 #endif
2994         struct nf_ct_helper_expectfn *expfn;
2995
2996         if (timeout < 0)
2997                 timeout = 0;
2998
2999         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3000                 goto nla_put_failure;
3001         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3002                 goto nla_put_failure;
3003         if (ctnetlink_exp_dump_tuple(skb,
3004                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3005                                  CTA_EXPECT_MASTER) < 0)
3006                 goto nla_put_failure;
3007
3008 #if IS_ENABLED(CONFIG_NF_NAT)
3009         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3010             exp->saved_proto.all) {
3011                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3012                 if (!nest_parms)
3013                         goto nla_put_failure;
3014
3015                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3016                         goto nla_put_failure;
3017
3018                 nat_tuple.src.l3num = nf_ct_l3num(master);
3019                 nat_tuple.src.u3 = exp->saved_addr;
3020                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3021                 nat_tuple.src.u = exp->saved_proto;
3022
3023                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3024                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3025                         goto nla_put_failure;
3026                 nla_nest_end(skb, nest_parms);
3027         }
3028 #endif
3029         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3030             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3031             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3032             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3033                 goto nla_put_failure;
3034         help = nfct_help(master);
3035         if (help) {
3036                 struct nf_conntrack_helper *helper;
3037
3038                 helper = rcu_dereference(help->helper);
3039                 if (helper &&
3040                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3041                         goto nla_put_failure;
3042         }
3043         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3044         if (expfn != NULL &&
3045             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3046                 goto nla_put_failure;
3047
3048         return 0;
3049
3050 nla_put_failure:
3051         return -1;
3052 }
3053
3054 static int
3055 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3056                         int event, const struct nf_conntrack_expect *exp)
3057 {
3058         struct nlmsghdr *nlh;
3059         unsigned int flags = portid ? NLM_F_MULTI : 0;
3060
3061         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3062         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3063                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3064         if (!nlh)
3065                 goto nlmsg_failure;
3066
3067         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3068                 goto nla_put_failure;
3069
3070         nlmsg_end(skb, nlh);
3071         return skb->len;
3072
3073 nlmsg_failure:
3074 nla_put_failure:
3075         nlmsg_cancel(skb, nlh);
3076         return -1;
3077 }
3078
3079 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3080 static int
3081 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
3082 {
3083         struct nf_conntrack_expect *exp = item->exp;
3084         struct net *net = nf_ct_exp_net(exp);
3085         struct nlmsghdr *nlh;
3086         struct sk_buff *skb;
3087         unsigned int type, group;
3088         int flags = 0;
3089
3090         if (events & (1 << IPEXP_DESTROY)) {
3091                 type = IPCTNL_MSG_EXP_DELETE;
3092                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3093         } else if (events & (1 << IPEXP_NEW)) {
3094                 type = IPCTNL_MSG_EXP_NEW;
3095                 flags = NLM_F_CREATE|NLM_F_EXCL;
3096                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3097         } else
3098                 return 0;
3099
3100         if (!item->report && !nfnetlink_has_listeners(net, group))
3101                 return 0;
3102
3103         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3104         if (skb == NULL)
3105                 goto errout;
3106
3107         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3108         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3109                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3110         if (!nlh)
3111                 goto nlmsg_failure;
3112
3113         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3114                 goto nla_put_failure;
3115
3116         nlmsg_end(skb, nlh);
3117         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3118         return 0;
3119
3120 nla_put_failure:
3121         nlmsg_cancel(skb, nlh);
3122 nlmsg_failure:
3123         kfree_skb(skb);
3124 errout:
3125         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3126         return 0;
3127 }
3128 #endif
3129 static int ctnetlink_exp_done(struct netlink_callback *cb)
3130 {
3131         if (cb->args[1])
3132                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3133         return 0;
3134 }
3135
3136 static int
3137 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3138 {
3139         struct net *net = sock_net(skb->sk);
3140         struct nf_conntrack_expect *exp, *last;
3141         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3142         u_int8_t l3proto = nfmsg->nfgen_family;
3143
3144         rcu_read_lock();
3145         last = (struct nf_conntrack_expect *)cb->args[1];
3146         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3147 restart:
3148                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3149                                          hnode) {
3150                         if (l3proto && exp->tuple.src.l3num != l3proto)
3151                                 continue;
3152
3153                         if (!net_eq(nf_ct_net(exp->master), net))
3154                                 continue;
3155
3156                         if (cb->args[1]) {
3157                                 if (exp != last)
3158                                         continue;
3159                                 cb->args[1] = 0;
3160                         }
3161                         if (ctnetlink_exp_fill_info(skb,
3162                                                     NETLINK_CB(cb->skb).portid,
3163                                                     cb->nlh->nlmsg_seq,
3164                                                     IPCTNL_MSG_EXP_NEW,
3165                                                     exp) < 0) {
3166                                 if (!refcount_inc_not_zero(&exp->use))
3167                                         continue;
3168                                 cb->args[1] = (unsigned long)exp;
3169                                 goto out;
3170                         }
3171                 }
3172                 if (cb->args[1]) {
3173                         cb->args[1] = 0;
3174                         goto restart;
3175                 }
3176         }
3177 out:
3178         rcu_read_unlock();
3179         if (last)
3180                 nf_ct_expect_put(last);
3181
3182         return skb->len;
3183 }
3184
3185 static int
3186 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3187 {
3188         struct nf_conntrack_expect *exp, *last;
3189         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3190         struct nf_conn *ct = cb->data;
3191         struct nf_conn_help *help = nfct_help(ct);
3192         u_int8_t l3proto = nfmsg->nfgen_family;
3193
3194         if (cb->args[0])
3195                 return 0;
3196
3197         rcu_read_lock();
3198         last = (struct nf_conntrack_expect *)cb->args[1];
3199 restart:
3200         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3201                 if (l3proto && exp->tuple.src.l3num != l3proto)
3202                         continue;
3203                 if (cb->args[1]) {
3204                         if (exp != last)
3205                                 continue;
3206                         cb->args[1] = 0;
3207                 }
3208                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3209                                             cb->nlh->nlmsg_seq,
3210                                             IPCTNL_MSG_EXP_NEW,
3211                                             exp) < 0) {
3212                         if (!refcount_inc_not_zero(&exp->use))
3213                                 continue;
3214                         cb->args[1] = (unsigned long)exp;
3215                         goto out;
3216                 }
3217         }
3218         if (cb->args[1]) {
3219                 cb->args[1] = 0;
3220                 goto restart;
3221         }
3222         cb->args[0] = 1;
3223 out:
3224         rcu_read_unlock();
3225         if (last)
3226                 nf_ct_expect_put(last);
3227
3228         return skb->len;
3229 }
3230
3231 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3232                                  struct sk_buff *skb,
3233                                  const struct nlmsghdr *nlh,
3234                                  const struct nlattr * const cda[],
3235                                  struct netlink_ext_ack *extack)
3236 {
3237         int err;
3238         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3239         u_int8_t u3 = nfmsg->nfgen_family;
3240         struct nf_conntrack_tuple tuple;
3241         struct nf_conntrack_tuple_hash *h;
3242         struct nf_conn *ct;
3243         struct nf_conntrack_zone zone;
3244         struct netlink_dump_control c = {
3245                 .dump = ctnetlink_exp_ct_dump_table,
3246                 .done = ctnetlink_exp_done,
3247         };
3248
3249         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3250                                     u3, NULL);
3251         if (err < 0)
3252                 return err;
3253
3254         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3255         if (err < 0)
3256                 return err;
3257
3258         h = nf_conntrack_find_get(net, &zone, &tuple);
3259         if (!h)
3260                 return -ENOENT;
3261
3262         ct = nf_ct_tuplehash_to_ctrack(h);
3263         /* No expectation linked to this connection tracking. */
3264         if (!nfct_help(ct)) {
3265                 nf_ct_put(ct);
3266                 return 0;
3267         }
3268
3269         c.data = ct;
3270
3271         err = netlink_dump_start(ctnl, skb, nlh, &c);
3272         nf_ct_put(ct);
3273
3274         return err;
3275 }
3276
3277 static int ctnetlink_get_expect(struct sk_buff *skb,
3278                                 const struct nfnl_info *info,
3279                                 const struct nlattr * const cda[])
3280 {
3281         struct nfgenmsg *nfmsg = nlmsg_data(info->nlh);
3282         u_int8_t u3 = nfmsg->nfgen_family;
3283         struct nf_conntrack_tuple tuple;
3284         struct nf_conntrack_expect *exp;
3285         struct nf_conntrack_zone zone;
3286         struct sk_buff *skb2;
3287         int err;
3288
3289         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3290                 if (cda[CTA_EXPECT_MASTER])
3291                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3292                                                      info->nlh, cda,
3293                                                      info->extack);
3294                 else {
3295                         struct netlink_dump_control c = {
3296                                 .dump = ctnetlink_exp_dump_table,
3297                                 .done = ctnetlink_exp_done,
3298                         };
3299                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3300                 }
3301         }
3302
3303         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3304         if (err < 0)
3305                 return err;
3306
3307         if (cda[CTA_EXPECT_TUPLE])
3308                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3309                                             u3, NULL);
3310         else if (cda[CTA_EXPECT_MASTER])
3311                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3312                                             u3, NULL);
3313         else
3314                 return -EINVAL;
3315
3316         if (err < 0)
3317                 return err;
3318
3319         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3320         if (!exp)
3321                 return -ENOENT;
3322
3323         if (cda[CTA_EXPECT_ID]) {
3324                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3325
3326                 if (id != nf_expect_get_id(exp)) {
3327                         nf_ct_expect_put(exp);
3328                         return -ENOENT;
3329                 }
3330         }
3331
3332         err = -ENOMEM;
3333         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3334         if (skb2 == NULL) {
3335                 nf_ct_expect_put(exp);
3336                 goto out;
3337         }
3338
3339         rcu_read_lock();
3340         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3341                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3342                                       exp);
3343         rcu_read_unlock();
3344         nf_ct_expect_put(exp);
3345         if (err <= 0)
3346                 goto free;
3347
3348         err = netlink_unicast(info->sk, skb2, NETLINK_CB(skb).portid,
3349                               MSG_DONTWAIT);
3350         if (err < 0)
3351                 goto out;
3352
3353         return 0;
3354
3355 free:
3356         kfree_skb(skb2);
3357 out:
3358         /* this avoids a loop in nfnetlink. */
3359         return err == -EAGAIN ? -ENOBUFS : err;
3360 }
3361
3362 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3363 {
3364         const struct nf_conn_help *m_help;
3365         const char *name = data;
3366
3367         m_help = nfct_help(exp->master);
3368
3369         return strcmp(m_help->helper->name, name) == 0;
3370 }
3371
3372 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3373 {
3374         return true;
3375 }
3376
3377 static int ctnetlink_del_expect(struct sk_buff *skb,
3378                                 const struct nfnl_info *info,
3379                                 const struct nlattr * const cda[])
3380 {
3381         struct nfgenmsg *nfmsg = nlmsg_data(info->nlh);
3382         u_int8_t u3 = nfmsg->nfgen_family;
3383         struct nf_conntrack_expect *exp;
3384         struct nf_conntrack_tuple tuple;
3385         struct nf_conntrack_zone zone;
3386         int err;
3387
3388         if (cda[CTA_EXPECT_TUPLE]) {
3389                 /* delete a single expect by tuple */
3390                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3391                 if (err < 0)
3392                         return err;
3393
3394                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3395                                             u3, NULL);
3396                 if (err < 0)
3397                         return err;
3398
3399                 /* bump usage count to 2 */
3400                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3401                 if (!exp)
3402                         return -ENOENT;
3403
3404                 if (cda[CTA_EXPECT_ID]) {
3405                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3406                         if (ntohl(id) != (u32)(unsigned long)exp) {
3407                                 nf_ct_expect_put(exp);
3408                                 return -ENOENT;
3409                         }
3410                 }
3411
3412                 /* after list removal, usage count == 1 */
3413                 spin_lock_bh(&nf_conntrack_expect_lock);
3414                 if (del_timer(&exp->timeout)) {
3415                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3416                                                    nlmsg_report(info->nlh));
3417                         nf_ct_expect_put(exp);
3418                 }
3419                 spin_unlock_bh(&nf_conntrack_expect_lock);
3420                 /* have to put what we 'get' above.
3421                  * after this line usage count == 0 */
3422                 nf_ct_expect_put(exp);
3423         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3424                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3425
3426                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3427                                          NETLINK_CB(skb).portid,
3428                                          nlmsg_report(info->nlh));
3429         } else {
3430                 /* This basically means we have to flush everything*/
3431                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3432                                          NETLINK_CB(skb).portid,
3433                                          nlmsg_report(info->nlh));
3434         }
3435
3436         return 0;
3437 }
3438 static int
3439 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3440                         const struct nlattr * const cda[])
3441 {
3442         if (cda[CTA_EXPECT_TIMEOUT]) {
3443                 if (!del_timer(&x->timeout))
3444                         return -ETIME;
3445
3446                 x->timeout.expires = jiffies +
3447                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3448                 add_timer(&x->timeout);
3449         }
3450         return 0;
3451 }
3452
3453 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3454         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3455         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3456 };
3457
3458 static int
3459 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3460                            struct nf_conntrack_expect *exp,
3461                            u_int8_t u3)
3462 {
3463 #if IS_ENABLED(CONFIG_NF_NAT)
3464         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3465         struct nf_conntrack_tuple nat_tuple = {};
3466         int err;
3467
3468         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3469                                           exp_nat_nla_policy, NULL);
3470         if (err < 0)
3471                 return err;
3472
3473         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3474                 return -EINVAL;
3475
3476         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3477                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3478                                     u3, NULL);
3479         if (err < 0)
3480                 return err;
3481
3482         exp->saved_addr = nat_tuple.src.u3;
3483         exp->saved_proto = nat_tuple.src.u;
3484         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3485
3486         return 0;
3487 #else
3488         return -EOPNOTSUPP;
3489 #endif
3490 }
3491
3492 static struct nf_conntrack_expect *
3493 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3494                        struct nf_conntrack_helper *helper,
3495                        struct nf_conntrack_tuple *tuple,
3496                        struct nf_conntrack_tuple *mask)
3497 {
3498         u_int32_t class = 0;
3499         struct nf_conntrack_expect *exp;
3500         struct nf_conn_help *help;
3501         int err;
3502
3503         help = nfct_help(ct);
3504         if (!help)
3505                 return ERR_PTR(-EOPNOTSUPP);
3506
3507         if (cda[CTA_EXPECT_CLASS] && helper) {
3508                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3509                 if (class > helper->expect_class_max)
3510                         return ERR_PTR(-EINVAL);
3511         }
3512         exp = nf_ct_expect_alloc(ct);
3513         if (!exp)
3514                 return ERR_PTR(-ENOMEM);
3515
3516         if (cda[CTA_EXPECT_FLAGS]) {
3517                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3518                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3519         } else {
3520                 exp->flags = 0;
3521         }
3522         if (cda[CTA_EXPECT_FN]) {
3523                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3524                 struct nf_ct_helper_expectfn *expfn;
3525
3526                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3527                 if (expfn == NULL) {
3528                         err = -EINVAL;
3529                         goto err_out;
3530                 }
3531                 exp->expectfn = expfn->expectfn;
3532         } else
3533                 exp->expectfn = NULL;
3534
3535         exp->class = class;
3536         exp->master = ct;
3537         exp->helper = helper;
3538         exp->tuple = *tuple;
3539         exp->mask.src.u3 = mask->src.u3;
3540         exp->mask.src.u.all = mask->src.u.all;
3541
3542         if (cda[CTA_EXPECT_NAT]) {
3543                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3544                                                  exp, nf_ct_l3num(ct));
3545                 if (err < 0)
3546                         goto err_out;
3547         }
3548         return exp;
3549 err_out:
3550         nf_ct_expect_put(exp);
3551         return ERR_PTR(err);
3552 }
3553
3554 static int
3555 ctnetlink_create_expect(struct net *net,
3556                         const struct nf_conntrack_zone *zone,
3557                         const struct nlattr * const cda[],
3558                         u_int8_t u3, u32 portid, int report)
3559 {
3560         struct nf_conntrack_tuple tuple, mask, master_tuple;
3561         struct nf_conntrack_tuple_hash *h = NULL;
3562         struct nf_conntrack_helper *helper = NULL;
3563         struct nf_conntrack_expect *exp;
3564         struct nf_conn *ct;
3565         int err;
3566
3567         /* caller guarantees that those three CTA_EXPECT_* exist */
3568         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3569                                     u3, NULL);
3570         if (err < 0)
3571                 return err;
3572         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3573                                     u3, NULL);
3574         if (err < 0)
3575                 return err;
3576         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3577                                     u3, NULL);
3578         if (err < 0)
3579                 return err;
3580
3581         /* Look for master conntrack of this expectation */
3582         h = nf_conntrack_find_get(net, zone, &master_tuple);
3583         if (!h)
3584                 return -ENOENT;
3585         ct = nf_ct_tuplehash_to_ctrack(h);
3586
3587         rcu_read_lock();
3588         if (cda[CTA_EXPECT_HELP_NAME]) {
3589                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3590
3591                 helper = __nf_conntrack_helper_find(helpname, u3,
3592                                                     nf_ct_protonum(ct));
3593                 if (helper == NULL) {
3594                         rcu_read_unlock();
3595 #ifdef CONFIG_MODULES
3596                         if (request_module("nfct-helper-%s", helpname) < 0) {
3597                                 err = -EOPNOTSUPP;
3598                                 goto err_ct;
3599                         }
3600                         rcu_read_lock();
3601                         helper = __nf_conntrack_helper_find(helpname, u3,
3602                                                             nf_ct_protonum(ct));
3603                         if (helper) {
3604                                 err = -EAGAIN;
3605                                 goto err_rcu;
3606                         }
3607                         rcu_read_unlock();
3608 #endif
3609                         err = -EOPNOTSUPP;
3610                         goto err_ct;
3611                 }
3612         }
3613
3614         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3615         if (IS_ERR(exp)) {
3616                 err = PTR_ERR(exp);
3617                 goto err_rcu;
3618         }
3619
3620         err = nf_ct_expect_related_report(exp, portid, report, 0);
3621         nf_ct_expect_put(exp);
3622 err_rcu:
3623         rcu_read_unlock();
3624 err_ct:
3625         nf_ct_put(ct);
3626         return err;
3627 }
3628
3629 static int ctnetlink_new_expect(struct sk_buff *skb,
3630                                 const struct nfnl_info *info,
3631                                 const struct nlattr * const cda[])
3632 {
3633         struct nfgenmsg *nfmsg = nlmsg_data(info->nlh);
3634         u_int8_t u3 = nfmsg->nfgen_family;
3635         struct nf_conntrack_tuple tuple;
3636         struct nf_conntrack_expect *exp;
3637         struct nf_conntrack_zone zone;
3638         int err;
3639
3640         if (!cda[CTA_EXPECT_TUPLE]
3641             || !cda[CTA_EXPECT_MASK]
3642             || !cda[CTA_EXPECT_MASTER])
3643                 return -EINVAL;
3644
3645         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3646         if (err < 0)
3647                 return err;
3648
3649         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3650                                     u3, NULL);
3651         if (err < 0)
3652                 return err;
3653
3654         spin_lock_bh(&nf_conntrack_expect_lock);
3655         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3656         if (!exp) {
3657                 spin_unlock_bh(&nf_conntrack_expect_lock);
3658                 err = -ENOENT;
3659                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3660                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3661                                                       NETLINK_CB(skb).portid,
3662                                                       nlmsg_report(info->nlh));
3663                 }
3664                 return err;
3665         }
3666
3667         err = -EEXIST;
3668         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3669                 err = ctnetlink_change_expect(exp, cda);
3670         spin_unlock_bh(&nf_conntrack_expect_lock);
3671
3672         return err;
3673 }
3674
3675 static int
3676 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3677                              const struct ip_conntrack_stat *st)
3678 {
3679         struct nlmsghdr *nlh;
3680         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3681
3682         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3683                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3684         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3685                            NFNETLINK_V0, htons(cpu));
3686         if (!nlh)
3687                 goto nlmsg_failure;
3688
3689         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3690             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3691             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3692                 goto nla_put_failure;
3693
3694         nlmsg_end(skb, nlh);
3695         return skb->len;
3696
3697 nla_put_failure:
3698 nlmsg_failure:
3699         nlmsg_cancel(skb, nlh);
3700         return -1;
3701 }
3702
3703 static int
3704 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3705 {
3706         int cpu;
3707         struct net *net = sock_net(skb->sk);
3708
3709         if (cb->args[0] == nr_cpu_ids)
3710                 return 0;
3711
3712         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3713                 const struct ip_conntrack_stat *st;
3714
3715                 if (!cpu_possible(cpu))
3716                         continue;
3717
3718                 st = per_cpu_ptr(net->ct.stat, cpu);
3719                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3720                                                  cb->nlh->nlmsg_seq,
3721                                                  cpu, st) < 0)
3722                         break;
3723         }
3724         cb->args[0] = cpu;
3725
3726         return skb->len;
3727 }
3728
3729 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3730                                   const struct nfnl_info *info,
3731                                   const struct nlattr * const cda[])
3732 {
3733         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3734                 struct netlink_dump_control c = {
3735                         .dump = ctnetlink_exp_stat_cpu_dump,
3736                 };
3737                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3738         }
3739
3740         return 0;
3741 }
3742
3743 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3744 static struct nf_ct_event_notifier ctnl_notifier = {
3745         .fcn = ctnetlink_conntrack_event,
3746 };
3747
3748 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3749         .fcn = 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         int ret;
3843
3844         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3845         if (ret < 0) {
3846                 pr_err("ctnetlink_init: cannot register notifier.\n");
3847                 goto err_out;
3848         }
3849
3850         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3851         if (ret < 0) {
3852                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3853                 goto err_unreg_notifier;
3854         }
3855 #endif
3856         return 0;
3857
3858 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3859 err_unreg_notifier:
3860         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3861 err_out:
3862         return ret;
3863 #endif
3864 }
3865
3866 static void ctnetlink_net_exit(struct net *net)
3867 {
3868 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3869         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3870         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3871 #endif
3872 }
3873
3874 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3875 {
3876         struct net *net;
3877
3878         list_for_each_entry(net, net_exit_list, exit_list)
3879                 ctnetlink_net_exit(net);
3880
3881         /* wait for other cpus until they are done with ctnl_notifiers */
3882         synchronize_rcu();
3883 }
3884
3885 static struct pernet_operations ctnetlink_net_ops = {
3886         .init           = ctnetlink_net_init,
3887         .exit_batch     = ctnetlink_net_exit_batch,
3888 };
3889
3890 static int __init ctnetlink_init(void)
3891 {
3892         int ret;
3893
3894         ret = nfnetlink_subsys_register(&ctnl_subsys);
3895         if (ret < 0) {
3896                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3897                 goto err_out;
3898         }
3899
3900         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3901         if (ret < 0) {
3902                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3903                 goto err_unreg_subsys;
3904         }
3905
3906         ret = register_pernet_subsys(&ctnetlink_net_ops);
3907         if (ret < 0) {
3908                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3909                 goto err_unreg_exp_subsys;
3910         }
3911 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3912         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3913         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3914 #endif
3915         return 0;
3916
3917 err_unreg_exp_subsys:
3918         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3919 err_unreg_subsys:
3920         nfnetlink_subsys_unregister(&ctnl_subsys);
3921 err_out:
3922         return ret;
3923 }
3924
3925 static void __exit ctnetlink_exit(void)
3926 {
3927         unregister_pernet_subsys(&ctnetlink_net_ops);
3928         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3929         nfnetlink_subsys_unregister(&ctnl_subsys);
3930 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3931         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3932 #endif
3933         synchronize_rcu();
3934 }
3935
3936 module_init(ctnetlink_init);
3937 module_exit(ctnetlink_exit);