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