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