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(refcount_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, const 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_u32 {
 856         u32 val;
 857         u32 mask;
 858 };
 859
 860 struct ctnetlink_filter {
 861         u8 family;
 862
 863         u_int32_t orig_flags;
 864         u_int32_t reply_flags;
 865
 866         struct nf_conntrack_tuple orig;
 867         struct nf_conntrack_tuple reply;
 868         struct nf_conntrack_zone zone;
 869
 870         struct ctnetlink_filter_u32 mark;
 871         struct ctnetlink_filter_u32 status;
 872 };
 873
 874 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 875         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 876         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 877 };
 878
 879 static int ctnetlink_parse_filter(const struct nlattr *attr,
 880                                   struct ctnetlink_filter *filter)
 881 {
 882         struct nlattr *tb[CTA_FILTER_MAX + 1];
 883         int ret = 0;
 884
 885         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 886                                NULL);
 887         if (ret)
 888                 return ret;
 889
 890         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 891                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 892                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 893                         return -EOPNOTSUPP;
 894         }
 895
 896         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 897                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 898                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 899                         return -EOPNOTSUPP;
 900         }
 901
 902         return 0;
 903 }
 904
 905 static int ctnetlink_parse_zone(const struct nlattr *attr,
 906                                 struct nf_conntrack_zone *zone);
 907 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 908                                          struct nf_conntrack_tuple *tuple,
 909                                          u32 type, u_int8_t l3num,
 910                                          struct nf_conntrack_zone *zone,
 911                                          u_int32_t flags);
 912
 913 static int ctnetlink_filter_parse_mark(struct ctnetlink_filter_u32 *mark,
 914                                        const struct nlattr * const cda[])
 915 {
 916 #ifdef CONFIG_NF_CONNTRACK_MARK
 917         if (cda[CTA_MARK]) {
 918                 mark->val = ntohl(nla_get_be32(cda[CTA_MARK]));
 919
 920                 if (cda[CTA_MARK_MASK])
 921                         mark->mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 922                 else
 923                         mark->mask = 0xffffffff;
 924         } else if (cda[CTA_MARK_MASK]) {
 925                 return -EINVAL;
 926         }
 927 #endif
 928         return 0;
 929 }
 930
 931 static int ctnetlink_filter_parse_status(struct ctnetlink_filter_u32 *status,
 932                                          const struct nlattr * const cda[])
 933 {
 934         if (cda[CTA_STATUS]) {
 935                 status->val = ntohl(nla_get_be32(cda[CTA_STATUS]));
 936                 if (cda[CTA_STATUS_MASK])
 937                         status->mask = ntohl(nla_get_be32(cda[CTA_STATUS_MASK]));
 938                 else
 939                         status->mask = status->val;
 940
 941                 /* status->val == 0? always true, else always false. */
 942                 if (status->mask == 0)
 943                         return -EINVAL;
 944         } else if (cda[CTA_STATUS_MASK]) {
 945                 return -EINVAL;
 946         }
 947
 948         /* CTA_STATUS is NLA_U32, if this fires UAPI needs to be extended */
 949         BUILD_BUG_ON(__IPS_MAX_BIT >= 32);
 950         return 0;
 951 }
 952
 953 static struct ctnetlink_filter *
 954 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 955 {
 956         struct ctnetlink_filter *filter;
 957         int err;
 958
 959 #ifndef CONFIG_NF_CONNTRACK_MARK
 960         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 961                 return ERR_PTR(-EOPNOTSUPP);
 962 #endif
 963
 964         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 965         if (filter == NULL)
 966                 return ERR_PTR(-ENOMEM);
 967
 968         filter->family = family;
 969
 970         err = ctnetlink_filter_parse_mark(&filter->mark, cda);
 971         if (err)
 972                 goto err_filter;
 973
 974         err = ctnetlink_filter_parse_status(&filter->status, cda);
 975         if (err)
 976                 goto err_filter;
 977
 978         if (!cda[CTA_FILTER])
 979                 return filter;
 980
 981         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 982         if (err < 0)
 983                 goto err_filter;
 984
 985         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
 986         if (err < 0)
 987                 goto err_filter;
 988
 989         if (filter->orig_flags) {
 990                 if (!cda[CTA_TUPLE_ORIG]) {
 991                         err = -EINVAL;
 992                         goto err_filter;
 993                 }
 994
 995                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
 996                                                    CTA_TUPLE_ORIG,
 997                                                    filter->family,
 998                                                    &filter->zone,
 999                                                    filter->orig_flags);
1000                 if (err < 0)
1001                         goto err_filter;
1002         }
1003
1004         if (filter->reply_flags) {
1005                 if (!cda[CTA_TUPLE_REPLY]) {
1006                         err = -EINVAL;
1007                         goto err_filter;
1008                 }
1009
1010                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
1011                                                    CTA_TUPLE_REPLY,
1012                                                    filter->family,
1013                                                    &filter->zone,
1014                                                    filter->reply_flags);
1015                 if (err < 0)
1016                         goto err_filter;
1017         }
1018
1019         return filter;
1020
1021 err_filter:
1022         kfree(filter);
1023
1024         return ERR_PTR(err);
1025 }
1026
1027 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1028 {
1029         return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS];
1030 }
1031
1032 static int ctnetlink_start(struct netlink_callback *cb)
1033 {
1034         const struct nlattr * const *cda = cb->data;
1035         struct ctnetlink_filter *filter = NULL;
1036         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1037         u8 family = nfmsg->nfgen_family;
1038
1039         if (ctnetlink_needs_filter(family, cda)) {
1040                 filter = ctnetlink_alloc_filter(cda, family);
1041                 if (IS_ERR(filter))
1042                         return PTR_ERR(filter);
1043         }
1044
1045         cb->data = filter;
1046         return 0;
1047 }
1048
1049 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1050                                         struct nf_conntrack_tuple *ct_tuple,
1051                                         u_int32_t flags, int family)
1052 {
1053         switch (family) {
1054         case NFPROTO_IPV4:
1055                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1056                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1057                         return  0;
1058
1059                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1060                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1061                         return  0;
1062                 break;
1063         case NFPROTO_IPV6:
1064                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1065                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1066                                    &ct_tuple->src.u3.in6))
1067                         return 0;
1068
1069                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1070                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1071                                    &ct_tuple->dst.u3.in6))
1072                         return 0;
1073                 break;
1074         }
1075
1076         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1077             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1078                 return 0;
1079
1080         switch (ct_tuple->dst.protonum) {
1081         case IPPROTO_TCP:
1082         case IPPROTO_UDP:
1083                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1084                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1085                         return 0;
1086
1087                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1088                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1089                         return 0;
1090                 break;
1091         case IPPROTO_ICMP:
1092                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1093                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1094                         return 0;
1095                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1096                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1097                         return 0;
1098                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1099                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1100                         return 0;
1101                 break;
1102         case IPPROTO_ICMPV6:
1103                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1104                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1105                         return 0;
1106                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1107                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1108                         return 0;
1109                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1110                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1111                         return 0;
1112                 break;
1113         }
1114
1115         return 1;
1116 }
1117
1118 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1119 {
1120         struct ctnetlink_filter *filter = data;
1121         struct nf_conntrack_tuple *tuple;
1122         u32 status;
1123
1124         if (filter == NULL)
1125                 goto out;
1126
1127         /* Match entries of a given L3 protocol number.
1128          * If it is not specified, ie. l3proto == 0,
1129          * then match everything.
1130          */
1131         if (filter->family && nf_ct_l3num(ct) != filter->family)
1132                 goto ignore_entry;
1133
1134         if (filter->orig_flags) {
1135                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1136                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1137                                                   filter->orig_flags,
1138                                                   filter->family))
1139                         goto ignore_entry;
1140         }
1141
1142         if (filter->reply_flags) {
1143                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1144                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1145                                                   filter->reply_flags,
1146                                                   filter->family))
1147                         goto ignore_entry;
1148         }
1149
1150 #ifdef CONFIG_NF_CONNTRACK_MARK
1151         if ((ct->mark & filter->mark.mask) != filter->mark.val)
1152                 goto ignore_entry;
1153 #endif
1154         status = (u32)READ_ONCE(ct->status);
1155         if ((status & filter->status.mask) != filter->status.val)
1156                 goto ignore_entry;
1157
1158 out:
1159         return 1;
1160
1161 ignore_entry:
1162         return 0;
1163 }
1164
1165 static int
1166 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1167 {
1168         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1169         struct net *net = sock_net(skb->sk);
1170         struct nf_conn *ct, *last;
1171         struct nf_conntrack_tuple_hash *h;
1172         struct hlist_nulls_node *n;
1173         struct nf_conn *nf_ct_evict[8];
1174         int res, i;
1175         spinlock_t *lockp;
1176
1177         last = (struct nf_conn *)cb->args[1];
1178         i = 0;
1179
1180         local_bh_disable();
1181         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1182 restart:
1183                 while (i) {
1184                         i--;
1185                         if (nf_ct_should_gc(nf_ct_evict[i]))
1186                                 nf_ct_kill(nf_ct_evict[i]);
1187                         nf_ct_put(nf_ct_evict[i]);
1188                 }
1189
1190                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1191                 nf_conntrack_lock(lockp);
1192                 if (cb->args[0] >= nf_conntrack_htable_size) {
1193                         spin_unlock(lockp);
1194                         goto out;
1195                 }
1196                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1197                                            hnnode) {
1198                         ct = nf_ct_tuplehash_to_ctrack(h);
1199                         if (nf_ct_is_expired(ct)) {
1200                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1201                                     refcount_inc_not_zero(&ct->ct_general.use))
1202                                         nf_ct_evict[i++] = ct;
1203                                 continue;
1204                         }
1205
1206                         if (!net_eq(net, nf_ct_net(ct)))
1207                                 continue;
1208
1209                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1210                                 continue;
1211
1212                         if (cb->args[1]) {
1213                                 if (ct != last)
1214                                         continue;
1215                                 cb->args[1] = 0;
1216                         }
1217                         if (!ctnetlink_filter_match(ct, cb->data))
1218                                 continue;
1219
1220                         res =
1221                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1222                                             cb->nlh->nlmsg_seq,
1223                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1224                                             ct, true, flags);
1225                         if (res < 0) {
1226                                 nf_conntrack_get(&ct->ct_general);
1227                                 cb->args[1] = (unsigned long)ct;
1228                                 spin_unlock(lockp);
1229                                 goto out;
1230                         }
1231                 }
1232                 spin_unlock(lockp);
1233                 if (cb->args[1]) {
1234                         cb->args[1] = 0;
1235                         goto restart;
1236                 }
1237         }
1238 out:
1239         local_bh_enable();
1240         if (last) {
1241                 /* nf ct hash resize happened, now clear the leftover. */
1242                 if ((struct nf_conn *)cb->args[1] == last)
1243                         cb->args[1] = 0;
1244
1245                 nf_ct_put(last);
1246         }
1247
1248         while (i) {
1249                 i--;
1250                 if (nf_ct_should_gc(nf_ct_evict[i]))
1251                         nf_ct_kill(nf_ct_evict[i]);
1252                 nf_ct_put(nf_ct_evict[i]);
1253         }
1254
1255         return skb->len;
1256 }
1257
1258 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1259                                 struct nf_conntrack_tuple *t,
1260                                 u_int32_t flags)
1261 {
1262         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1263                 if (!tb[CTA_IP_V4_SRC])
1264                         return -EINVAL;
1265
1266                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1267         }
1268
1269         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1270                 if (!tb[CTA_IP_V4_DST])
1271                         return -EINVAL;
1272
1273                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1274         }
1275
1276         return 0;
1277 }
1278
1279 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1280                                 struct nf_conntrack_tuple *t,
1281                                 u_int32_t flags)
1282 {
1283         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1284                 if (!tb[CTA_IP_V6_SRC])
1285                         return -EINVAL;
1286
1287                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1288         }
1289
1290         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1291                 if (!tb[CTA_IP_V6_DST])
1292                         return -EINVAL;
1293
1294                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1295         }
1296
1297         return 0;
1298 }
1299
1300 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1301                                     struct nf_conntrack_tuple *tuple,
1302                                     u_int32_t flags)
1303 {
1304         struct nlattr *tb[CTA_IP_MAX+1];
1305         int ret = 0;
1306
1307         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1308         if (ret < 0)
1309                 return ret;
1310
1311         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1312                                              cta_ip_nla_policy, NULL);
1313         if (ret)
1314                 return ret;
1315
1316         switch (tuple->src.l3num) {
1317         case NFPROTO_IPV4:
1318                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1319                 break;
1320         case NFPROTO_IPV6:
1321                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1322                 break;
1323         }
1324
1325         return ret;
1326 }
1327
1328 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1329         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1330 };
1331
1332 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1333                                        struct nf_conntrack_tuple *tuple,
1334                                        u_int32_t flags)
1335 {
1336         const struct nf_conntrack_l4proto *l4proto;
1337         struct nlattr *tb[CTA_PROTO_MAX+1];
1338         int ret = 0;
1339
1340         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1341                                           proto_nla_policy, NULL);
1342         if (ret < 0)
1343                 return ret;
1344
1345         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1346                 return 0;
1347
1348         if (!tb[CTA_PROTO_NUM])
1349                 return -EINVAL;
1350
1351         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1352
1353         rcu_read_lock();
1354         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1355
1356         if (likely(l4proto->nlattr_to_tuple)) {
1357                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1358                                                      l4proto->nla_policy,
1359                                                      NULL);
1360                 if (ret == 0)
1361                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1362         }
1363
1364         rcu_read_unlock();
1365
1366         return ret;
1367 }
1368
1369 static int
1370 ctnetlink_parse_zone(const struct nlattr *attr,
1371                      struct nf_conntrack_zone *zone)
1372 {
1373         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1374                         NF_CT_DEFAULT_ZONE_DIR, 0);
1375 #ifdef CONFIG_NF_CONNTRACK_ZONES
1376         if (attr)
1377                 zone->id = ntohs(nla_get_be16(attr));
1378 #else
1379         if (attr)
1380                 return -EOPNOTSUPP;
1381 #endif
1382         return 0;
1383 }
1384
1385 static int
1386 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1387                            struct nf_conntrack_zone *zone)
1388 {
1389         int ret;
1390
1391         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1392                 return -EINVAL;
1393
1394         ret = ctnetlink_parse_zone(attr, zone);
1395         if (ret < 0)
1396                 return ret;
1397
1398         if (type == CTA_TUPLE_REPLY)
1399                 zone->dir = NF_CT_ZONE_DIR_REPL;
1400         else
1401                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1402
1403         return 0;
1404 }
1405
1406 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1407         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1408         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1409         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1410 };
1411
1412 #define CTA_FILTER_F_ALL_CTA_PROTO \
1413   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1414    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1415    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1416    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1417    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1418    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1419    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1420    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1421
1422 static int
1423 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1424                               struct nf_conntrack_tuple *tuple, u32 type,
1425                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1426                               u_int32_t flags)
1427 {
1428         struct nlattr *tb[CTA_TUPLE_MAX+1];
1429         int err;
1430
1431         memset(tuple, 0, sizeof(*tuple));
1432
1433         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1434                                           tuple_nla_policy, NULL);
1435         if (err < 0)
1436                 return err;
1437
1438         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1439                 return -EOPNOTSUPP;
1440         tuple->src.l3num = l3num;
1441
1442         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1443             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1444                 if (!tb[CTA_TUPLE_IP])
1445                         return -EINVAL;
1446
1447                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1448                 if (err < 0)
1449                         return err;
1450         }
1451
1452         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1453                 if (!tb[CTA_TUPLE_PROTO])
1454                         return -EINVAL;
1455
1456                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1457                 if (err < 0)
1458                         return err;
1459         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1460                 /* Can't manage proto flags without a protonum  */
1461                 return -EINVAL;
1462         }
1463
1464         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1465                 if (!zone)
1466                         return -EINVAL;
1467
1468                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1469                                                  type, zone);
1470                 if (err < 0)
1471                         return err;
1472         }
1473
1474         /* orig and expect tuples get DIR_ORIGINAL */
1475         if (type == CTA_TUPLE_REPLY)
1476                 tuple->dst.dir = IP_CT_DIR_REPLY;
1477         else
1478                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1479
1480         return 0;
1481 }
1482
1483 static int
1484 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1485                       struct nf_conntrack_tuple *tuple, u32 type,
1486                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1487 {
1488         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1489                                             CTA_FILTER_FLAG(ALL));
1490 }
1491
1492 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1493         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1494                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1495 };
1496
1497 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1498                                 struct nlattr **helpinfo)
1499 {
1500         int err;
1501         struct nlattr *tb[CTA_HELP_MAX+1];
1502
1503         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1504                                           help_nla_policy, NULL);
1505         if (err < 0)
1506                 return err;
1507
1508         if (!tb[CTA_HELP_NAME])
1509                 return -EINVAL;
1510
1511         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1512
1513         if (tb[CTA_HELP_INFO])
1514                 *helpinfo = tb[CTA_HELP_INFO];
1515
1516         return 0;
1517 }
1518
1519 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1520         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1521         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1522         [CTA_STATUS]            = { .type = NLA_U32 },
1523         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1524         [CTA_HELP]              = { .type = NLA_NESTED },
1525         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1526         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1527         [CTA_MARK]              = { .type = NLA_U32 },
1528         [CTA_ID]                = { .type = NLA_U32 },
1529         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1530         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1531         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1532         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1533         [CTA_ZONE]              = { .type = NLA_U16 },
1534         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1535         [CTA_LABELS]            = { .type = NLA_BINARY,
1536                                     .len = NF_CT_LABELS_MAX_SIZE },
1537         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1538                                     .len = NF_CT_LABELS_MAX_SIZE },
1539         [CTA_FILTER]            = { .type = NLA_NESTED },
1540         [CTA_STATUS_MASK]       = { .type = NLA_U32 },
1541 };
1542
1543 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1544 {
1545         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1546                 return 0;
1547
1548         return ctnetlink_filter_match(ct, data);
1549 }
1550
1551 static int ctnetlink_flush_conntrack(struct net *net,
1552                                      const struct nlattr * const cda[],
1553                                      u32 portid, int report, u8 family)
1554 {
1555         struct ctnetlink_filter *filter = NULL;
1556
1557         if (ctnetlink_needs_filter(family, cda)) {
1558                 if (cda[CTA_FILTER])
1559                         return -EOPNOTSUPP;
1560
1561                 filter = ctnetlink_alloc_filter(cda, family);
1562                 if (IS_ERR(filter))
1563                         return PTR_ERR(filter);
1564         }
1565
1566         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1567                                   portid, report);
1568         kfree(filter);
1569
1570         return 0;
1571 }
1572
1573 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1574                                    const struct nfnl_info *info,
1575                                    const struct nlattr * const cda[])
1576 {
1577         u8 family = info->nfmsg->nfgen_family;
1578         struct nf_conntrack_tuple_hash *h;
1579         struct nf_conntrack_tuple tuple;
1580         struct nf_conntrack_zone zone;
1581         struct nf_conn *ct;
1582         int err;
1583
1584         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1585         if (err < 0)
1586                 return err;
1587
1588         if (cda[CTA_TUPLE_ORIG])
1589                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1590                                             family, &zone);
1591         else if (cda[CTA_TUPLE_REPLY])
1592                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1593                                             family, &zone);
1594         else {
1595                 u_int8_t u3 = info->nfmsg->version ? family : AF_UNSPEC;
1596
1597                 return ctnetlink_flush_conntrack(info->net, cda,
1598                                                  NETLINK_CB(skb).portid,
1599                                                  nlmsg_report(info->nlh), u3);
1600         }
1601
1602         if (err < 0)
1603                 return err;
1604
1605         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1606         if (!h)
1607                 return -ENOENT;
1608
1609         ct = nf_ct_tuplehash_to_ctrack(h);
1610
1611         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1612                 nf_ct_put(ct);
1613                 return -EBUSY;
1614         }
1615
1616         if (cda[CTA_ID]) {
1617                 __be32 id = nla_get_be32(cda[CTA_ID]);
1618
1619                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1620                         nf_ct_put(ct);
1621                         return -ENOENT;
1622                 }
1623         }
1624
1625         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1626         nf_ct_put(ct);
1627
1628         return 0;
1629 }
1630
1631 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1632                                    const struct nfnl_info *info,
1633                                    const struct nlattr * const cda[])
1634 {
1635         u_int8_t u3 = info->nfmsg->nfgen_family;
1636         struct nf_conntrack_tuple_hash *h;
1637         struct nf_conntrack_tuple tuple;
1638         struct nf_conntrack_zone zone;
1639         struct sk_buff *skb2;
1640         struct nf_conn *ct;
1641         int err;
1642
1643         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1644                 struct netlink_dump_control c = {
1645                         .start = ctnetlink_start,
1646                         .dump = ctnetlink_dump_table,
1647                         .done = ctnetlink_done,
1648                         .data = (void *)cda,
1649                 };
1650
1651                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1652         }
1653
1654         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1655         if (err < 0)
1656                 return err;
1657
1658         if (cda[CTA_TUPLE_ORIG])
1659                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1660                                             u3, &zone);
1661         else if (cda[CTA_TUPLE_REPLY])
1662                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1663                                             u3, &zone);
1664         else
1665                 return -EINVAL;
1666
1667         if (err < 0)
1668                 return err;
1669
1670         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1671         if (!h)
1672                 return -ENOENT;
1673
1674         ct = nf_ct_tuplehash_to_ctrack(h);
1675
1676         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1677         if (!skb2) {
1678                 nf_ct_put(ct);
1679                 return -ENOMEM;
1680         }
1681
1682         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1683                                   info->nlh->nlmsg_seq,
1684                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1685                                   true, 0);
1686         nf_ct_put(ct);
1687         if (err <= 0) {
1688                 kfree_skb(skb2);
1689                 return -ENOMEM;
1690         }
1691
1692         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1693 }
1694
1695 static int ctnetlink_done_list(struct netlink_callback *cb)
1696 {
1697         if (cb->args[1])
1698                 nf_ct_put((struct nf_conn *)cb->args[1]);
1699         return 0;
1700 }
1701
1702 static int
1703 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1704 {
1705         struct nf_conn *ct, *last;
1706         struct nf_conntrack_tuple_hash *h;
1707         struct hlist_nulls_node *n;
1708         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1709         u_int8_t l3proto = nfmsg->nfgen_family;
1710         int res;
1711         int cpu;
1712         struct hlist_nulls_head *list;
1713         struct net *net = sock_net(skb->sk);
1714
1715         if (cb->args[2])
1716                 return 0;
1717
1718         last = (struct nf_conn *)cb->args[1];
1719
1720         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1721                 struct ct_pcpu *pcpu;
1722
1723                 if (!cpu_possible(cpu))
1724                         continue;
1725
1726                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1727                 spin_lock_bh(&pcpu->lock);
1728                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1729 restart:
1730                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1731                         ct = nf_ct_tuplehash_to_ctrack(h);
1732                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1733                                 continue;
1734                         if (cb->args[1]) {
1735                                 if (ct != last)
1736                                         continue;
1737                                 cb->args[1] = 0;
1738                         }
1739
1740                         /* We can't dump extension info for the unconfirmed
1741                          * list because unconfirmed conntracks can have
1742                          * ct->ext reallocated (and thus freed).
1743                          *
1744                          * In the dying list case ct->ext can't be free'd
1745                          * until after we drop pcpu->lock.
1746                          */
1747                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1748                                                   cb->nlh->nlmsg_seq,
1749                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1750                                                   ct, dying, 0);
1751                         if (res < 0) {
1752                                 if (!refcount_inc_not_zero(&ct->ct_general.use))
1753                                         continue;
1754                                 cb->args[0] = cpu;
1755                                 cb->args[1] = (unsigned long)ct;
1756                                 spin_unlock_bh(&pcpu->lock);
1757                                 goto out;
1758                         }
1759                 }
1760                 if (cb->args[1]) {
1761                         cb->args[1] = 0;
1762                         goto restart;
1763                 }
1764                 spin_unlock_bh(&pcpu->lock);
1765         }
1766         cb->args[2] = 1;
1767 out:
1768         if (last)
1769                 nf_ct_put(last);
1770
1771         return skb->len;
1772 }
1773
1774 static int
1775 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1776 {
1777         return ctnetlink_dump_list(skb, cb, true);
1778 }
1779
1780 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1781                                   const struct nfnl_info *info,
1782                                   const struct nlattr * const cda[])
1783 {
1784         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1785                 struct netlink_dump_control c = {
1786                         .dump = ctnetlink_dump_dying,
1787                         .done = ctnetlink_done_list,
1788                 };
1789                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1790         }
1791
1792         return -EOPNOTSUPP;
1793 }
1794
1795 static int
1796 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1797 {
1798         return ctnetlink_dump_list(skb, cb, false);
1799 }
1800
1801 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1802                                         const struct nfnl_info *info,
1803                                         const struct nlattr * const cda[])
1804 {
1805         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1806                 struct netlink_dump_control c = {
1807                         .dump = ctnetlink_dump_unconfirmed,
1808                         .done = ctnetlink_done_list,
1809                 };
1810                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1811         }
1812
1813         return -EOPNOTSUPP;
1814 }
1815
1816 #if IS_ENABLED(CONFIG_NF_NAT)
1817 static int
1818 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1819                           enum nf_nat_manip_type manip,
1820                           const struct nlattr *attr)
1821         __must_hold(RCU)
1822 {
1823         const struct nf_nat_hook *nat_hook;
1824         int err;
1825
1826         nat_hook = rcu_dereference(nf_nat_hook);
1827         if (!nat_hook) {
1828 #ifdef CONFIG_MODULES
1829                 rcu_read_unlock();
1830                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1831                 if (request_module("nf-nat") < 0) {
1832                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1833                         rcu_read_lock();
1834                         return -EOPNOTSUPP;
1835                 }
1836                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1837                 rcu_read_lock();
1838                 nat_hook = rcu_dereference(nf_nat_hook);
1839                 if (nat_hook)
1840                         return -EAGAIN;
1841 #endif
1842                 return -EOPNOTSUPP;
1843         }
1844
1845         err = nat_hook->parse_nat_setup(ct, manip, attr);
1846         if (err == -EAGAIN) {
1847 #ifdef CONFIG_MODULES
1848                 rcu_read_unlock();
1849                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1850                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1851                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1852                         rcu_read_lock();
1853                         return -EOPNOTSUPP;
1854                 }
1855                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1856                 rcu_read_lock();
1857 #else
1858                 err = -EOPNOTSUPP;
1859 #endif
1860         }
1861         return err;
1862 }
1863 #endif
1864
1865 static void
1866 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1867                           unsigned long off)
1868 {
1869         unsigned int bit;
1870
1871         /* Ignore these unchangable bits */
1872         on &= ~IPS_UNCHANGEABLE_MASK;
1873         off &= ~IPS_UNCHANGEABLE_MASK;
1874
1875         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1876                 if (on & (1 << bit))
1877                         set_bit(bit, &ct->status);
1878                 else if (off & (1 << bit))
1879                         clear_bit(bit, &ct->status);
1880         }
1881 }
1882
1883 static int
1884 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1885 {
1886         unsigned long d;
1887         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1888         d = ct->status ^ status;
1889
1890         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1891                 /* unchangeable */
1892                 return -EBUSY;
1893
1894         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1895                 /* SEEN_REPLY bit can only be set */
1896                 return -EBUSY;
1897
1898         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1899                 /* ASSURED bit can only be set */
1900                 return -EBUSY;
1901
1902         __ctnetlink_change_status(ct, status, 0);
1903         return 0;
1904 }
1905
1906 static int
1907 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1908 {
1909 #if IS_ENABLED(CONFIG_NF_NAT)
1910         int ret;
1911
1912         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1913                 return 0;
1914
1915         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1916                                         cda[CTA_NAT_DST]);
1917         if (ret < 0)
1918                 return ret;
1919
1920         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1921                                          cda[CTA_NAT_SRC]);
1922 #else
1923         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1924                 return 0;
1925         return -EOPNOTSUPP;
1926 #endif
1927 }
1928
1929 static int ctnetlink_change_helper(struct nf_conn *ct,
1930                                    const struct nlattr * const cda[])
1931 {
1932         struct nf_conntrack_helper *helper;
1933         struct nf_conn_help *help = nfct_help(ct);
1934         char *helpname = NULL;
1935         struct nlattr *helpinfo = NULL;
1936         int err;
1937
1938         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1939         if (err < 0)
1940                 return err;
1941
1942         /* don't change helper of sibling connections */
1943         if (ct->master) {
1944                 /* If we try to change the helper to the same thing twice,
1945                  * treat the second attempt as a no-op instead of returning
1946                  * an error.
1947                  */
1948                 err = -EBUSY;
1949                 if (help) {
1950                         rcu_read_lock();
1951                         helper = rcu_dereference(help->helper);
1952                         if (helper && !strcmp(helper->name, helpname))
1953                                 err = 0;
1954                         rcu_read_unlock();
1955                 }
1956
1957                 return err;
1958         }
1959
1960         if (!strcmp(helpname, "")) {
1961                 if (help && help->helper) {
1962                         /* we had a helper before ... */
1963                         nf_ct_remove_expectations(ct);
1964                         RCU_INIT_POINTER(help->helper, NULL);
1965                 }
1966
1967                 return 0;
1968         }
1969
1970         rcu_read_lock();
1971         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1972                                             nf_ct_protonum(ct));
1973         if (helper == NULL) {
1974                 rcu_read_unlock();
1975                 return -EOPNOTSUPP;
1976         }
1977
1978         if (help) {
1979                 if (help->helper == helper) {
1980                         /* update private helper data if allowed. */
1981                         if (helper->from_nlattr)
1982                                 helper->from_nlattr(helpinfo, ct);
1983                         err = 0;
1984                 } else
1985                         err = -EBUSY;
1986         } else {
1987                 /* we cannot set a helper for an existing conntrack */
1988                 err = -EOPNOTSUPP;
1989         }
1990
1991         rcu_read_unlock();
1992         return err;
1993 }
1994
1995 static int ctnetlink_change_timeout(struct nf_conn *ct,
1996                                     const struct nlattr * const cda[])
1997 {
1998         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1999
2000         if (timeout > INT_MAX)
2001                 timeout = INT_MAX;
2002         WRITE_ONCE(ct->timeout, nfct_time_stamp + (u32)timeout);
2003
2004         if (test_bit(IPS_DYING_BIT, &ct->status))
2005                 return -ETIME;
2006
2007         return 0;
2008 }
2009
2010 #if defined(CONFIG_NF_CONNTRACK_MARK)
2011 static void ctnetlink_change_mark(struct nf_conn *ct,
2012                                     const struct nlattr * const cda[])
2013 {
2014         u32 mark, newmark, mask = 0;
2015
2016         if (cda[CTA_MARK_MASK])
2017                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2018
2019         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2020         newmark = (ct->mark & mask) ^ mark;
2021         if (newmark != ct->mark)
2022                 ct->mark = newmark;
2023 }
2024 #endif
2025
2026 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2027         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2028         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2029         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2030 };
2031
2032 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2033                                       const struct nlattr * const cda[])
2034 {
2035         const struct nlattr *attr = cda[CTA_PROTOINFO];
2036         const struct nf_conntrack_l4proto *l4proto;
2037         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2038         int err = 0;
2039
2040         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2041                                           protoinfo_policy, NULL);
2042         if (err < 0)
2043                 return err;
2044
2045         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2046         if (l4proto->from_nlattr)
2047                 err = l4proto->from_nlattr(tb, ct);
2048
2049         return err;
2050 }
2051
2052 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2053         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2054         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2055         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2056 };
2057
2058 static int change_seq_adj(struct nf_ct_seqadj *seq,
2059                           const struct nlattr * const attr)
2060 {
2061         int err;
2062         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2063
2064         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2065                                           seqadj_policy, NULL);
2066         if (err < 0)
2067                 return err;
2068
2069         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2070                 return -EINVAL;
2071
2072         seq->correction_pos =
2073                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2074
2075         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2076                 return -EINVAL;
2077
2078         seq->offset_before =
2079                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2080
2081         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2082                 return -EINVAL;
2083
2084         seq->offset_after =
2085                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2086
2087         return 0;
2088 }
2089
2090 static int
2091 ctnetlink_change_seq_adj(struct nf_conn *ct,
2092                          const struct nlattr * const cda[])
2093 {
2094         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2095         int ret = 0;
2096
2097         if (!seqadj)
2098                 return 0;
2099
2100         spin_lock_bh(&ct->lock);
2101         if (cda[CTA_SEQ_ADJ_ORIG]) {
2102                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2103                                      cda[CTA_SEQ_ADJ_ORIG]);
2104                 if (ret < 0)
2105                         goto err;
2106
2107                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2108         }
2109
2110         if (cda[CTA_SEQ_ADJ_REPLY]) {
2111                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2112                                      cda[CTA_SEQ_ADJ_REPLY]);
2113                 if (ret < 0)
2114                         goto err;
2115
2116                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2117         }
2118
2119         spin_unlock_bh(&ct->lock);
2120         return 0;
2121 err:
2122         spin_unlock_bh(&ct->lock);
2123         return ret;
2124 }
2125
2126 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2127         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2128         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2129         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2130 };
2131
2132 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2133                                      const struct nlattr * const cda[])
2134 {
2135         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2136         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2137         int err;
2138
2139         if (!synproxy)
2140                 return 0;
2141
2142         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2143                                           cda[CTA_SYNPROXY], synproxy_policy,
2144                                           NULL);
2145         if (err < 0)
2146                 return err;
2147
2148         if (!tb[CTA_SYNPROXY_ISN] ||
2149             !tb[CTA_SYNPROXY_ITS] ||
2150             !tb[CTA_SYNPROXY_TSOFF])
2151                 return -EINVAL;
2152
2153         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2154         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2155         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2156
2157         return 0;
2158 }
2159
2160 static int
2161 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2162 {
2163 #ifdef CONFIG_NF_CONNTRACK_LABELS
2164         size_t len = nla_len(cda[CTA_LABELS]);
2165         const void *mask = cda[CTA_LABELS_MASK];
2166
2167         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2168                 return -EINVAL;
2169
2170         if (mask) {
2171                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2172                     nla_len(cda[CTA_LABELS_MASK]) != len)
2173                         return -EINVAL;
2174                 mask = nla_data(cda[CTA_LABELS_MASK]);
2175         }
2176
2177         len /= sizeof(u32);
2178
2179         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2180 #else
2181         return -EOPNOTSUPP;
2182 #endif
2183 }
2184
2185 static int
2186 ctnetlink_change_conntrack(struct nf_conn *ct,
2187                            const struct nlattr * const cda[])
2188 {
2189         int err;
2190
2191         /* only allow NAT changes and master assignation for new conntracks */
2192         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2193                 return -EOPNOTSUPP;
2194
2195         if (cda[CTA_HELP]) {
2196                 err = ctnetlink_change_helper(ct, cda);
2197                 if (err < 0)
2198                         return err;
2199         }
2200
2201         if (cda[CTA_TIMEOUT]) {
2202                 err = ctnetlink_change_timeout(ct, cda);
2203                 if (err < 0)
2204                         return err;
2205         }
2206
2207         if (cda[CTA_STATUS]) {
2208                 err = ctnetlink_change_status(ct, cda);
2209                 if (err < 0)
2210                         return err;
2211         }
2212
2213         if (cda[CTA_PROTOINFO]) {
2214                 err = ctnetlink_change_protoinfo(ct, cda);
2215                 if (err < 0)
2216                         return err;
2217         }
2218
2219 #if defined(CONFIG_NF_CONNTRACK_MARK)
2220         if (cda[CTA_MARK])
2221                 ctnetlink_change_mark(ct, cda);
2222 #endif
2223
2224         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2225                 err = ctnetlink_change_seq_adj(ct, cda);
2226                 if (err < 0)
2227                         return err;
2228         }
2229
2230         if (cda[CTA_SYNPROXY]) {
2231                 err = ctnetlink_change_synproxy(ct, cda);
2232                 if (err < 0)
2233                         return err;
2234         }
2235
2236         if (cda[CTA_LABELS]) {
2237                 err = ctnetlink_attach_labels(ct, cda);
2238                 if (err < 0)
2239                         return err;
2240         }
2241
2242         return 0;
2243 }
2244
2245 static struct nf_conn *
2246 ctnetlink_create_conntrack(struct net *net,
2247                            const struct nf_conntrack_zone *zone,
2248                            const struct nlattr * const cda[],
2249                            struct nf_conntrack_tuple *otuple,
2250                            struct nf_conntrack_tuple *rtuple,
2251                            u8 u3)
2252 {
2253         struct nf_conn *ct;
2254         int err = -EINVAL;
2255         struct nf_conntrack_helper *helper;
2256         struct nf_conn_tstamp *tstamp;
2257         u64 timeout;
2258
2259         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2260         if (IS_ERR(ct))
2261                 return ERR_PTR(-ENOMEM);
2262
2263         if (!cda[CTA_TIMEOUT])
2264                 goto err1;
2265
2266         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2267         if (timeout > INT_MAX)
2268                 timeout = INT_MAX;
2269         ct->timeout = (u32)timeout + nfct_time_stamp;
2270
2271         rcu_read_lock();
2272         if (cda[CTA_HELP]) {
2273                 char *helpname = NULL;
2274                 struct nlattr *helpinfo = NULL;
2275
2276                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2277                 if (err < 0)
2278                         goto err2;
2279
2280                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2281                                                     nf_ct_protonum(ct));
2282                 if (helper == NULL) {
2283                         rcu_read_unlock();
2284 #ifdef CONFIG_MODULES
2285                         if (request_module("nfct-helper-%s", helpname) < 0) {
2286                                 err = -EOPNOTSUPP;
2287                                 goto err1;
2288                         }
2289
2290                         rcu_read_lock();
2291                         helper = __nf_conntrack_helper_find(helpname,
2292                                                             nf_ct_l3num(ct),
2293                                                             nf_ct_protonum(ct));
2294                         if (helper) {
2295                                 err = -EAGAIN;
2296                                 goto err2;
2297                         }
2298                         rcu_read_unlock();
2299 #endif
2300                         err = -EOPNOTSUPP;
2301                         goto err1;
2302                 } else {
2303                         struct nf_conn_help *help;
2304
2305                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2306                         if (help == NULL) {
2307                                 err = -ENOMEM;
2308                                 goto err2;
2309                         }
2310                         /* set private helper data if allowed. */
2311                         if (helper->from_nlattr)
2312                                 helper->from_nlattr(helpinfo, ct);
2313
2314                         /* disable helper auto-assignment for this entry */
2315                         ct->status |= IPS_HELPER;
2316                         RCU_INIT_POINTER(help->helper, helper);
2317                 }
2318         } else {
2319                 /* try an implicit helper assignation */
2320                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2321                 if (err < 0)
2322                         goto err2;
2323         }
2324
2325         err = ctnetlink_setup_nat(ct, cda);
2326         if (err < 0)
2327                 goto err2;
2328
2329         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2330         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2331         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2332         nf_ct_labels_ext_add(ct);
2333         nfct_seqadj_ext_add(ct);
2334         nfct_synproxy_ext_add(ct);
2335
2336         /* we must add conntrack extensions before confirmation. */
2337         ct->status |= IPS_CONFIRMED;
2338
2339         if (cda[CTA_STATUS]) {
2340                 err = ctnetlink_change_status(ct, cda);
2341                 if (err < 0)
2342                         goto err2;
2343         }
2344
2345         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2346                 err = ctnetlink_change_seq_adj(ct, cda);
2347                 if (err < 0)
2348                         goto err2;
2349         }
2350
2351         memset(&ct->proto, 0, sizeof(ct->proto));
2352         if (cda[CTA_PROTOINFO]) {
2353                 err = ctnetlink_change_protoinfo(ct, cda);
2354                 if (err < 0)
2355                         goto err2;
2356         }
2357
2358         if (cda[CTA_SYNPROXY]) {
2359                 err = ctnetlink_change_synproxy(ct, cda);
2360                 if (err < 0)
2361                         goto err2;
2362         }
2363
2364 #if defined(CONFIG_NF_CONNTRACK_MARK)
2365         if (cda[CTA_MARK])
2366                 ctnetlink_change_mark(ct, cda);
2367 #endif
2368
2369         /* setup master conntrack: this is a confirmed expectation */
2370         if (cda[CTA_TUPLE_MASTER]) {
2371                 struct nf_conntrack_tuple master;
2372                 struct nf_conntrack_tuple_hash *master_h;
2373                 struct nf_conn *master_ct;
2374
2375                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2376                                             u3, NULL);
2377                 if (err < 0)
2378                         goto err2;
2379
2380                 master_h = nf_conntrack_find_get(net, zone, &master);
2381                 if (master_h == NULL) {
2382                         err = -ENOENT;
2383                         goto err2;
2384                 }
2385                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2386                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2387                 ct->master = master_ct;
2388         }
2389         tstamp = nf_conn_tstamp_find(ct);
2390         if (tstamp)
2391                 tstamp->start = ktime_get_real_ns();
2392
2393         err = nf_conntrack_hash_check_insert(ct);
2394         if (err < 0)
2395                 goto err2;
2396
2397         rcu_read_unlock();
2398
2399         return ct;
2400
2401 err2:
2402         rcu_read_unlock();
2403 err1:
2404         nf_conntrack_free(ct);
2405         return ERR_PTR(err);
2406 }
2407
2408 static int ctnetlink_new_conntrack(struct sk_buff *skb,
2409                                    const struct nfnl_info *info,
2410                                    const struct nlattr * const cda[])
2411 {
2412         struct nf_conntrack_tuple otuple, rtuple;
2413         struct nf_conntrack_tuple_hash *h = NULL;
2414         u_int8_t u3 = info->nfmsg->nfgen_family;
2415         struct nf_conntrack_zone zone;
2416         struct nf_conn *ct;
2417         int err;
2418
2419         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2420         if (err < 0)
2421                 return err;
2422
2423         if (cda[CTA_TUPLE_ORIG]) {
2424                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2425                                             u3, &zone);
2426                 if (err < 0)
2427                         return err;
2428         }
2429
2430         if (cda[CTA_TUPLE_REPLY]) {
2431                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2432                                             u3, &zone);
2433                 if (err < 0)
2434                         return err;
2435         }
2436
2437         if (cda[CTA_TUPLE_ORIG])
2438                 h = nf_conntrack_find_get(info->net, &zone, &otuple);
2439         else if (cda[CTA_TUPLE_REPLY])
2440                 h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2441
2442         if (h == NULL) {
2443                 err = -ENOENT;
2444                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2445                         enum ip_conntrack_events events;
2446
2447                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2448                                 return -EINVAL;
2449                         if (otuple.dst.protonum != rtuple.dst.protonum)
2450                                 return -EINVAL;
2451
2452                         ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2453                                                         &otuple, &rtuple, u3);
2454                         if (IS_ERR(ct))
2455                                 return PTR_ERR(ct);
2456
2457                         err = 0;
2458                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2459                                 events = 1 << IPCT_RELATED;
2460                         else
2461                                 events = 1 << IPCT_NEW;
2462
2463                         if (cda[CTA_LABELS] &&
2464                             ctnetlink_attach_labels(ct, cda) == 0)
2465                                 events |= (1 << IPCT_LABEL);
2466
2467                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2468                                                       (1 << IPCT_ASSURED) |
2469                                                       (1 << IPCT_HELPER) |
2470                                                       (1 << IPCT_PROTOINFO) |
2471                                                       (1 << IPCT_SEQADJ) |
2472                                                       (1 << IPCT_MARK) |
2473                                                       (1 << IPCT_SYNPROXY) |
2474                                                       events,
2475                                                       ct, NETLINK_CB(skb).portid,
2476                                                       nlmsg_report(info->nlh));
2477                         nf_ct_put(ct);
2478                 }
2479
2480                 return err;
2481         }
2482         /* implicit 'else' */
2483
2484         err = -EEXIST;
2485         ct = nf_ct_tuplehash_to_ctrack(h);
2486         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2487                 err = ctnetlink_change_conntrack(ct, cda);
2488                 if (err == 0) {
2489                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2490                                                       (1 << IPCT_ASSURED) |
2491                                                       (1 << IPCT_HELPER) |
2492                                                       (1 << IPCT_LABEL) |
2493                                                       (1 << IPCT_PROTOINFO) |
2494                                                       (1 << IPCT_SEQADJ) |
2495                                                       (1 << IPCT_MARK) |
2496                                                       (1 << IPCT_SYNPROXY),
2497                                                       ct, NETLINK_CB(skb).portid,
2498                                                       nlmsg_report(info->nlh));
2499                 }
2500         }
2501
2502         nf_ct_put(ct);
2503         return err;
2504 }
2505
2506 static int
2507 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2508                                 __u16 cpu, const struct ip_conntrack_stat *st)
2509 {
2510         struct nlmsghdr *nlh;
2511         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2512
2513         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2514                               IPCTNL_MSG_CT_GET_STATS_CPU);
2515         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2516                            NFNETLINK_V0, htons(cpu));
2517         if (!nlh)
2518                 goto nlmsg_failure;
2519
2520         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2521             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2522             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2523             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2524                                 htonl(st->insert_failed)) ||
2525             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2526             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2527             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2528             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2529                                 htonl(st->search_restart)) ||
2530             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2531                                 htonl(st->clash_resolve)) ||
2532             nla_put_be32(skb, CTA_STATS_CHAIN_TOOLONG,
2533                          htonl(st->chaintoolong)))
2534                 goto nla_put_failure;
2535
2536         nlmsg_end(skb, nlh);
2537         return skb->len;
2538
2539 nla_put_failure:
2540 nlmsg_failure:
2541         nlmsg_cancel(skb, nlh);
2542         return -1;
2543 }
2544
2545 static int
2546 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2547 {
2548         int cpu;
2549         struct net *net = sock_net(skb->sk);
2550
2551         if (cb->args[0] == nr_cpu_ids)
2552                 return 0;
2553
2554         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2555                 const struct ip_conntrack_stat *st;
2556
2557                 if (!cpu_possible(cpu))
2558                         continue;
2559
2560                 st = per_cpu_ptr(net->ct.stat, cpu);
2561                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2562                                                     NETLINK_CB(cb->skb).portid,
2563                                                     cb->nlh->nlmsg_seq,
2564                                                     cpu, st) < 0)
2565                                 break;
2566         }
2567         cb->args[0] = cpu;
2568
2569         return skb->len;
2570 }
2571
2572 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2573                                  const struct nfnl_info *info,
2574                                  const struct nlattr * const cda[])
2575 {
2576         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2577                 struct netlink_dump_control c = {
2578                         .dump = ctnetlink_ct_stat_cpu_dump,
2579                 };
2580                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2581         }
2582
2583         return 0;
2584 }
2585
2586 static int
2587 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2588                             struct net *net)
2589 {
2590         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2591         unsigned int nr_conntracks;
2592         struct nlmsghdr *nlh;
2593
2594         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2595         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2596                            NFNETLINK_V0, 0);
2597         if (!nlh)
2598                 goto nlmsg_failure;
2599
2600         nr_conntracks = nf_conntrack_count(net);
2601         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2602                 goto nla_put_failure;
2603
2604         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2605                 goto nla_put_failure;
2606
2607         nlmsg_end(skb, nlh);
2608         return skb->len;
2609
2610 nla_put_failure:
2611 nlmsg_failure:
2612         nlmsg_cancel(skb, nlh);
2613         return -1;
2614 }
2615
2616 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2617                              const struct nlattr * const cda[])
2618 {
2619         struct sk_buff *skb2;
2620         int err;
2621
2622         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2623         if (skb2 == NULL)
2624                 return -ENOMEM;
2625
2626         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2627                                           info->nlh->nlmsg_seq,
2628                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2629                                           sock_net(skb->sk));
2630         if (err <= 0) {
2631                 kfree_skb(skb2);
2632                 return -ENOMEM;
2633         }
2634
2635         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2636 }
2637
2638 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2639         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2640         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2641         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2642         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2643         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2644         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2645                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2646         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2647         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2648         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2649         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2650         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2651 };
2652
2653 static struct nf_conntrack_expect *
2654 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2655                        struct nf_conntrack_helper *helper,
2656                        struct nf_conntrack_tuple *tuple,
2657                        struct nf_conntrack_tuple *mask);
2658
2659 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2660 static size_t
2661 ctnetlink_glue_build_size(const struct nf_conn *ct)
2662 {
2663         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2664                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2665                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2666                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2667                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2668                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2669                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2670                + nla_total_size(0) /* CTA_PROTOINFO */
2671                + nla_total_size(0) /* CTA_HELP */
2672                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2673                + ctnetlink_secctx_size(ct)
2674                + ctnetlink_acct_size(ct)
2675                + ctnetlink_timestamp_size(ct)
2676 #if IS_ENABLED(CONFIG_NF_NAT)
2677                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2678                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2679 #endif
2680 #ifdef CONFIG_NF_CONNTRACK_MARK
2681                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2682 #endif
2683 #ifdef CONFIG_NF_CONNTRACK_ZONES
2684                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2685 #endif
2686                + ctnetlink_proto_size(ct)
2687                ;
2688 }
2689
2690 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2691 {
2692         const struct nf_conntrack_zone *zone;
2693         struct nlattr *nest_parms;
2694
2695         zone = nf_ct_zone(ct);
2696
2697         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2698         if (!nest_parms)
2699                 goto nla_put_failure;
2700         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2701                 goto nla_put_failure;
2702         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2703                                    NF_CT_ZONE_DIR_ORIG) < 0)
2704                 goto nla_put_failure;
2705         nla_nest_end(skb, nest_parms);
2706
2707         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2708         if (!nest_parms)
2709                 goto nla_put_failure;
2710         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2711                 goto nla_put_failure;
2712         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2713                                    NF_CT_ZONE_DIR_REPL) < 0)
2714                 goto nla_put_failure;
2715         nla_nest_end(skb, nest_parms);
2716
2717         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2718                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2719                 goto nla_put_failure;
2720
2721         if (ctnetlink_dump_id(skb, ct) < 0)
2722                 goto nla_put_failure;
2723
2724         if (ctnetlink_dump_status(skb, ct) < 0)
2725                 goto nla_put_failure;
2726
2727         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2728                 goto nla_put_failure;
2729
2730         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2731                 goto nla_put_failure;
2732
2733         if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2734             ctnetlink_dump_timestamp(skb, ct) < 0)
2735                 goto nla_put_failure;
2736
2737         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2738                 goto nla_put_failure;
2739
2740 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2741         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2742                 goto nla_put_failure;
2743 #endif
2744         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2745                 goto nla_put_failure;
2746
2747         if ((ct->status & IPS_SEQ_ADJUST) &&
2748             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2749                 goto nla_put_failure;
2750
2751         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2752                 goto nla_put_failure;
2753
2754 #ifdef CONFIG_NF_CONNTRACK_MARK
2755         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2756                 goto nla_put_failure;
2757 #endif
2758         if (ctnetlink_dump_labels(skb, ct) < 0)
2759                 goto nla_put_failure;
2760         return 0;
2761
2762 nla_put_failure:
2763         return -ENOSPC;
2764 }
2765
2766 static int
2767 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2768                      enum ip_conntrack_info ctinfo,
2769                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2770 {
2771         struct nlattr *nest_parms;
2772
2773         nest_parms = nla_nest_start(skb, ct_attr);
2774         if (!nest_parms)
2775                 goto nla_put_failure;
2776
2777         if (__ctnetlink_glue_build(skb, ct) < 0)
2778                 goto nla_put_failure;
2779
2780         nla_nest_end(skb, nest_parms);
2781
2782         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2783                 goto nla_put_failure;
2784
2785         return 0;
2786
2787 nla_put_failure:
2788         return -ENOSPC;
2789 }
2790
2791 static int
2792 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2793 {
2794         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2795         unsigned long d = ct->status ^ status;
2796
2797         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2798                 /* SEEN_REPLY bit can only be set */
2799                 return -EBUSY;
2800
2801         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2802                 /* ASSURED bit can only be set */
2803                 return -EBUSY;
2804
2805         /* This check is less strict than ctnetlink_change_status()
2806          * because callers often flip IPS_EXPECTED bits when sending
2807          * an NFQA_CT attribute to the kernel.  So ignore the
2808          * unchangeable bits but do not error out. Also user programs
2809          * are allowed to clear the bits that they are allowed to change.
2810          */
2811         __ctnetlink_change_status(ct, status, ~status);
2812         return 0;
2813 }
2814
2815 static int
2816 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2817 {
2818         int err;
2819
2820         if (cda[CTA_TIMEOUT]) {
2821                 err = ctnetlink_change_timeout(ct, cda);
2822                 if (err < 0)
2823                         return err;
2824         }
2825         if (cda[CTA_STATUS]) {
2826                 err = ctnetlink_update_status(ct, cda);
2827                 if (err < 0)
2828                         return err;
2829         }
2830         if (cda[CTA_HELP]) {
2831                 err = ctnetlink_change_helper(ct, cda);
2832                 if (err < 0)
2833                         return err;
2834         }
2835         if (cda[CTA_LABELS]) {
2836                 err = ctnetlink_attach_labels(ct, cda);
2837                 if (err < 0)
2838                         return err;
2839         }
2840 #if defined(CONFIG_NF_CONNTRACK_MARK)
2841         if (cda[CTA_MARK]) {
2842                 ctnetlink_change_mark(ct, cda);
2843         }
2844 #endif
2845         return 0;
2846 }
2847
2848 static int
2849 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2850 {
2851         struct nlattr *cda[CTA_MAX+1];
2852         int ret;
2853
2854         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2855                                           NULL);
2856         if (ret < 0)
2857                 return ret;
2858
2859         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2860 }
2861
2862 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2863                                     const struct nf_conn *ct,
2864                                     struct nf_conntrack_tuple *tuple,
2865                                     struct nf_conntrack_tuple *mask)
2866 {
2867         int err;
2868
2869         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2870                                     nf_ct_l3num(ct), NULL);
2871         if (err < 0)
2872                 return err;
2873
2874         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2875                                      nf_ct_l3num(ct), NULL);
2876 }
2877
2878 static int
2879 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2880                              u32 portid, u32 report)
2881 {
2882         struct nlattr *cda[CTA_EXPECT_MAX+1];
2883         struct nf_conntrack_tuple tuple, mask;
2884         struct nf_conntrack_helper *helper = NULL;
2885         struct nf_conntrack_expect *exp;
2886         int err;
2887
2888         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2889                                           exp_nla_policy, NULL);
2890         if (err < 0)
2891                 return err;
2892
2893         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2894                                        ct, &tuple, &mask);
2895         if (err < 0)
2896                 return err;
2897
2898         if (cda[CTA_EXPECT_HELP_NAME]) {
2899                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2900
2901                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2902                                                     nf_ct_protonum(ct));
2903                 if (helper == NULL)
2904                         return -EOPNOTSUPP;
2905         }
2906
2907         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2908                                      helper, &tuple, &mask);
2909         if (IS_ERR(exp))
2910                 return PTR_ERR(exp);
2911
2912         err = nf_ct_expect_related_report(exp, portid, report, 0);
2913         nf_ct_expect_put(exp);
2914         return err;
2915 }
2916
2917 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2918                                   enum ip_conntrack_info ctinfo, int diff)
2919 {
2920         if (!(ct->status & IPS_NAT_MASK))
2921                 return;
2922
2923         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2924 }
2925
2926 static const struct nfnl_ct_hook ctnetlink_glue_hook = {
2927         .build_size     = ctnetlink_glue_build_size,
2928         .build          = ctnetlink_glue_build,
2929         .parse          = ctnetlink_glue_parse,
2930         .attach_expect  = ctnetlink_glue_attach_expect,
2931         .seq_adjust     = ctnetlink_glue_seqadj,
2932 };
2933 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2934
2935 /***********************************************************************
2936  * EXPECT
2937  ***********************************************************************/
2938
2939 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2940                                     const struct nf_conntrack_tuple *tuple,
2941                                     u32 type)
2942 {
2943         struct nlattr *nest_parms;
2944
2945         nest_parms = nla_nest_start(skb, type);
2946         if (!nest_parms)
2947                 goto nla_put_failure;
2948         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2949                 goto nla_put_failure;
2950         nla_nest_end(skb, nest_parms);
2951
2952         return 0;
2953
2954 nla_put_failure:
2955         return -1;
2956 }
2957
2958 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2959                                    const struct nf_conntrack_tuple *tuple,
2960                                    const struct nf_conntrack_tuple_mask *mask)
2961 {
2962         const struct nf_conntrack_l4proto *l4proto;
2963         struct nf_conntrack_tuple m;
2964         struct nlattr *nest_parms;
2965         int ret;
2966
2967         memset(&m, 0xFF, sizeof(m));
2968         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2969         m.src.u.all = mask->src.u.all;
2970         m.src.l3num = tuple->src.l3num;
2971         m.dst.protonum = tuple->dst.protonum;
2972
2973         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2974         if (!nest_parms)
2975                 goto nla_put_failure;
2976
2977         rcu_read_lock();
2978         ret = ctnetlink_dump_tuples_ip(skb, &m);
2979         if (ret >= 0) {
2980                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2981                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2982         }
2983         rcu_read_unlock();
2984
2985         if (unlikely(ret < 0))
2986                 goto nla_put_failure;
2987
2988         nla_nest_end(skb, nest_parms);
2989
2990         return 0;
2991
2992 nla_put_failure:
2993         return -1;
2994 }
2995
2996 static const union nf_inet_addr any_addr;
2997
2998 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2999 {
3000         static siphash_aligned_key_t exp_id_seed;
3001         unsigned long a, b, c, d;
3002
3003         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
3004
3005         a = (unsigned long)exp;
3006         b = (unsigned long)exp->helper;
3007         c = (unsigned long)exp->master;
3008         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
3009
3010 #ifdef CONFIG_64BIT
3011         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
3012 #else
3013         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
3014 #endif
3015 }
3016
3017 static int
3018 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3019                           const struct nf_conntrack_expect *exp)
3020 {
3021         struct nf_conn *master = exp->master;
3022         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3023         struct nf_conn_help *help;
3024 #if IS_ENABLED(CONFIG_NF_NAT)
3025         struct nlattr *nest_parms;
3026         struct nf_conntrack_tuple nat_tuple = {};
3027 #endif
3028         struct nf_ct_helper_expectfn *expfn;
3029
3030         if (timeout < 0)
3031                 timeout = 0;
3032
3033         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3034                 goto nla_put_failure;
3035         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3036                 goto nla_put_failure;
3037         if (ctnetlink_exp_dump_tuple(skb,
3038                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3039                                  CTA_EXPECT_MASTER) < 0)
3040                 goto nla_put_failure;
3041
3042 #if IS_ENABLED(CONFIG_NF_NAT)
3043         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3044             exp->saved_proto.all) {
3045                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3046                 if (!nest_parms)
3047                         goto nla_put_failure;
3048
3049                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3050                         goto nla_put_failure;
3051
3052                 nat_tuple.src.l3num = nf_ct_l3num(master);
3053                 nat_tuple.src.u3 = exp->saved_addr;
3054                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3055                 nat_tuple.src.u = exp->saved_proto;
3056
3057                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3058                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3059                         goto nla_put_failure;
3060                 nla_nest_end(skb, nest_parms);
3061         }
3062 #endif
3063         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3064             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3065             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3066             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3067                 goto nla_put_failure;
3068         help = nfct_help(master);
3069         if (help) {
3070                 struct nf_conntrack_helper *helper;
3071
3072                 helper = rcu_dereference(help->helper);
3073                 if (helper &&
3074                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3075                         goto nla_put_failure;
3076         }
3077         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3078         if (expfn != NULL &&
3079             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3080                 goto nla_put_failure;
3081
3082         return 0;
3083
3084 nla_put_failure:
3085         return -1;
3086 }
3087
3088 static int
3089 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3090                         int event, const struct nf_conntrack_expect *exp)
3091 {
3092         struct nlmsghdr *nlh;
3093         unsigned int flags = portid ? NLM_F_MULTI : 0;
3094
3095         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3096         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3097                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3098         if (!nlh)
3099                 goto nlmsg_failure;
3100
3101         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3102                 goto nla_put_failure;
3103
3104         nlmsg_end(skb, nlh);
3105         return skb->len;
3106
3107 nlmsg_failure:
3108 nla_put_failure:
3109         nlmsg_cancel(skb, nlh);
3110         return -1;
3111 }
3112
3113 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3114 static int
3115 ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3116 {
3117         struct nf_conntrack_expect *exp = item->exp;
3118         struct net *net = nf_ct_exp_net(exp);
3119         struct nlmsghdr *nlh;
3120         struct sk_buff *skb;
3121         unsigned int type, group;
3122         int flags = 0;
3123
3124         if (events & (1 << IPEXP_DESTROY)) {
3125                 type = IPCTNL_MSG_EXP_DELETE;
3126                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3127         } else if (events & (1 << IPEXP_NEW)) {
3128                 type = IPCTNL_MSG_EXP_NEW;
3129                 flags = NLM_F_CREATE|NLM_F_EXCL;
3130                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3131         } else
3132                 return 0;
3133
3134         if (!item->report && !nfnetlink_has_listeners(net, group))
3135                 return 0;
3136
3137         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3138         if (skb == NULL)
3139                 goto errout;
3140
3141         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3142         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3143                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3144         if (!nlh)
3145                 goto nlmsg_failure;
3146
3147         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3148                 goto nla_put_failure;
3149
3150         nlmsg_end(skb, nlh);
3151         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3152         return 0;
3153
3154 nla_put_failure:
3155         nlmsg_cancel(skb, nlh);
3156 nlmsg_failure:
3157         kfree_skb(skb);
3158 errout:
3159         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3160         return 0;
3161 }
3162 #endif
3163 static int ctnetlink_exp_done(struct netlink_callback *cb)
3164 {
3165         if (cb->args[1])
3166                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3167         return 0;
3168 }
3169
3170 static int
3171 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3172 {
3173         struct net *net = sock_net(skb->sk);
3174         struct nf_conntrack_expect *exp, *last;
3175         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3176         u_int8_t l3proto = nfmsg->nfgen_family;
3177
3178         rcu_read_lock();
3179         last = (struct nf_conntrack_expect *)cb->args[1];
3180         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3181 restart:
3182                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3183                                          hnode) {
3184                         if (l3proto && exp->tuple.src.l3num != l3proto)
3185                                 continue;
3186
3187                         if (!net_eq(nf_ct_net(exp->master), net))
3188                                 continue;
3189
3190                         if (cb->args[1]) {
3191                                 if (exp != last)
3192                                         continue;
3193                                 cb->args[1] = 0;
3194                         }
3195                         if (ctnetlink_exp_fill_info(skb,
3196                                                     NETLINK_CB(cb->skb).portid,
3197                                                     cb->nlh->nlmsg_seq,
3198                                                     IPCTNL_MSG_EXP_NEW,
3199                                                     exp) < 0) {
3200                                 if (!refcount_inc_not_zero(&exp->use))
3201                                         continue;
3202                                 cb->args[1] = (unsigned long)exp;
3203                                 goto out;
3204                         }
3205                 }
3206                 if (cb->args[1]) {
3207                         cb->args[1] = 0;
3208                         goto restart;
3209                 }
3210         }
3211 out:
3212         rcu_read_unlock();
3213         if (last)
3214                 nf_ct_expect_put(last);
3215
3216         return skb->len;
3217 }
3218
3219 static int
3220 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3221 {
3222         struct nf_conntrack_expect *exp, *last;
3223         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3224         struct nf_conn *ct = cb->data;
3225         struct nf_conn_help *help = nfct_help(ct);
3226         u_int8_t l3proto = nfmsg->nfgen_family;
3227
3228         if (cb->args[0])
3229                 return 0;
3230
3231         rcu_read_lock();
3232         last = (struct nf_conntrack_expect *)cb->args[1];
3233 restart:
3234         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3235                 if (l3proto && exp->tuple.src.l3num != l3proto)
3236                         continue;
3237                 if (cb->args[1]) {
3238                         if (exp != last)
3239                                 continue;
3240                         cb->args[1] = 0;
3241                 }
3242                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3243                                             cb->nlh->nlmsg_seq,
3244                                             IPCTNL_MSG_EXP_NEW,
3245                                             exp) < 0) {
3246                         if (!refcount_inc_not_zero(&exp->use))
3247                                 continue;
3248                         cb->args[1] = (unsigned long)exp;
3249                         goto out;
3250                 }
3251         }
3252         if (cb->args[1]) {
3253                 cb->args[1] = 0;
3254                 goto restart;
3255         }
3256         cb->args[0] = 1;
3257 out:
3258         rcu_read_unlock();
3259         if (last)
3260                 nf_ct_expect_put(last);
3261
3262         return skb->len;
3263 }
3264
3265 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3266                                  struct sk_buff *skb,
3267                                  const struct nlmsghdr *nlh,
3268                                  const struct nlattr * const cda[],
3269                                  struct netlink_ext_ack *extack)
3270 {
3271         int err;
3272         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3273         u_int8_t u3 = nfmsg->nfgen_family;
3274         struct nf_conntrack_tuple tuple;
3275         struct nf_conntrack_tuple_hash *h;
3276         struct nf_conn *ct;
3277         struct nf_conntrack_zone zone;
3278         struct netlink_dump_control c = {
3279                 .dump = ctnetlink_exp_ct_dump_table,
3280                 .done = ctnetlink_exp_done,
3281         };
3282
3283         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3284                                     u3, NULL);
3285         if (err < 0)
3286                 return err;
3287
3288         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3289         if (err < 0)
3290                 return err;
3291
3292         h = nf_conntrack_find_get(net, &zone, &tuple);
3293         if (!h)
3294                 return -ENOENT;
3295
3296         ct = nf_ct_tuplehash_to_ctrack(h);
3297         /* No expectation linked to this connection tracking. */
3298         if (!nfct_help(ct)) {
3299                 nf_ct_put(ct);
3300                 return 0;
3301         }
3302
3303         c.data = ct;
3304
3305         err = netlink_dump_start(ctnl, skb, nlh, &c);
3306         nf_ct_put(ct);
3307
3308         return err;
3309 }
3310
3311 static int ctnetlink_get_expect(struct sk_buff *skb,
3312                                 const struct nfnl_info *info,
3313                                 const struct nlattr * const cda[])
3314 {
3315         u_int8_t u3 = info->nfmsg->nfgen_family;
3316         struct nf_conntrack_tuple tuple;
3317         struct nf_conntrack_expect *exp;
3318         struct nf_conntrack_zone zone;
3319         struct sk_buff *skb2;
3320         int err;
3321
3322         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3323                 if (cda[CTA_EXPECT_MASTER])
3324                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3325                                                      info->nlh, cda,
3326                                                      info->extack);
3327                 else {
3328                         struct netlink_dump_control c = {
3329                                 .dump = ctnetlink_exp_dump_table,
3330                                 .done = ctnetlink_exp_done,
3331                         };
3332                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3333                 }
3334         }
3335
3336         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3337         if (err < 0)
3338                 return err;
3339
3340         if (cda[CTA_EXPECT_TUPLE])
3341                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3342                                             u3, NULL);
3343         else if (cda[CTA_EXPECT_MASTER])
3344                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3345                                             u3, NULL);
3346         else
3347                 return -EINVAL;
3348
3349         if (err < 0)
3350                 return err;
3351
3352         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3353         if (!exp)
3354                 return -ENOENT;
3355
3356         if (cda[CTA_EXPECT_ID]) {
3357                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3358
3359                 if (id != nf_expect_get_id(exp)) {
3360                         nf_ct_expect_put(exp);
3361                         return -ENOENT;
3362                 }
3363         }
3364
3365         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3366         if (!skb2) {
3367                 nf_ct_expect_put(exp);
3368                 return -ENOMEM;
3369         }
3370
3371         rcu_read_lock();
3372         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3373                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3374                                       exp);
3375         rcu_read_unlock();
3376         nf_ct_expect_put(exp);
3377         if (err <= 0) {
3378                 kfree_skb(skb2);
3379                 return -ENOMEM;
3380         }
3381
3382         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3383 }
3384
3385 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3386 {
3387         const struct nf_conn_help *m_help;
3388         const char *name = data;
3389
3390         m_help = nfct_help(exp->master);
3391
3392         return strcmp(m_help->helper->name, name) == 0;
3393 }
3394
3395 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3396 {
3397         return true;
3398 }
3399
3400 static int ctnetlink_del_expect(struct sk_buff *skb,
3401                                 const struct nfnl_info *info,
3402                                 const struct nlattr * const cda[])
3403 {
3404         u_int8_t u3 = info->nfmsg->nfgen_family;
3405         struct nf_conntrack_expect *exp;
3406         struct nf_conntrack_tuple tuple;
3407         struct nf_conntrack_zone zone;
3408         int err;
3409
3410         if (cda[CTA_EXPECT_TUPLE]) {
3411                 /* delete a single expect by tuple */
3412                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3413                 if (err < 0)
3414                         return err;
3415
3416                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3417                                             u3, NULL);
3418                 if (err < 0)
3419                         return err;
3420
3421                 /* bump usage count to 2 */
3422                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3423                 if (!exp)
3424                         return -ENOENT;
3425
3426                 if (cda[CTA_EXPECT_ID]) {
3427                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3428                         if (ntohl(id) != (u32)(unsigned long)exp) {
3429                                 nf_ct_expect_put(exp);
3430                                 return -ENOENT;
3431                         }
3432                 }
3433
3434                 /* after list removal, usage count == 1 */
3435                 spin_lock_bh(&nf_conntrack_expect_lock);
3436                 if (del_timer(&exp->timeout)) {
3437                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3438                                                    nlmsg_report(info->nlh));
3439                         nf_ct_expect_put(exp);
3440                 }
3441                 spin_unlock_bh(&nf_conntrack_expect_lock);
3442                 /* have to put what we 'get' above.
3443                  * after this line usage count == 0 */
3444                 nf_ct_expect_put(exp);
3445         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3446                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3447
3448                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3449                                          NETLINK_CB(skb).portid,
3450                                          nlmsg_report(info->nlh));
3451         } else {
3452                 /* This basically means we have to flush everything*/
3453                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3454                                          NETLINK_CB(skb).portid,
3455                                          nlmsg_report(info->nlh));
3456         }
3457
3458         return 0;
3459 }
3460 static int
3461 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3462                         const struct nlattr * const cda[])
3463 {
3464         if (cda[CTA_EXPECT_TIMEOUT]) {
3465                 if (!del_timer(&x->timeout))
3466                         return -ETIME;
3467
3468                 x->timeout.expires = jiffies +
3469                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3470                 add_timer(&x->timeout);
3471         }
3472         return 0;
3473 }
3474
3475 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3476         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3477         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3478 };
3479
3480 static int
3481 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3482                            struct nf_conntrack_expect *exp,
3483                            u_int8_t u3)
3484 {
3485 #if IS_ENABLED(CONFIG_NF_NAT)
3486         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3487         struct nf_conntrack_tuple nat_tuple = {};
3488         int err;
3489
3490         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3491                                           exp_nat_nla_policy, NULL);
3492         if (err < 0)
3493                 return err;
3494
3495         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3496                 return -EINVAL;
3497
3498         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3499                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3500                                     u3, NULL);
3501         if (err < 0)
3502                 return err;
3503
3504         exp->saved_addr = nat_tuple.src.u3;
3505         exp->saved_proto = nat_tuple.src.u;
3506         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3507
3508         return 0;
3509 #else
3510         return -EOPNOTSUPP;
3511 #endif
3512 }
3513
3514 static struct nf_conntrack_expect *
3515 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3516                        struct nf_conntrack_helper *helper,
3517                        struct nf_conntrack_tuple *tuple,
3518                        struct nf_conntrack_tuple *mask)
3519 {
3520         u_int32_t class = 0;
3521         struct nf_conntrack_expect *exp;
3522         struct nf_conn_help *help;
3523         int err;
3524
3525         help = nfct_help(ct);
3526         if (!help)
3527                 return ERR_PTR(-EOPNOTSUPP);
3528
3529         if (cda[CTA_EXPECT_CLASS] && helper) {
3530                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3531                 if (class > helper->expect_class_max)
3532                         return ERR_PTR(-EINVAL);
3533         }
3534         exp = nf_ct_expect_alloc(ct);
3535         if (!exp)
3536                 return ERR_PTR(-ENOMEM);
3537
3538         if (cda[CTA_EXPECT_FLAGS]) {
3539                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3540                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3541         } else {
3542                 exp->flags = 0;
3543         }
3544         if (cda[CTA_EXPECT_FN]) {
3545                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3546                 struct nf_ct_helper_expectfn *expfn;
3547
3548                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3549                 if (expfn == NULL) {
3550                         err = -EINVAL;
3551                         goto err_out;
3552                 }
3553                 exp->expectfn = expfn->expectfn;
3554         } else
3555                 exp->expectfn = NULL;
3556
3557         exp->class = class;
3558         exp->master = ct;
3559         exp->helper = helper;
3560         exp->tuple = *tuple;
3561         exp->mask.src.u3 = mask->src.u3;
3562         exp->mask.src.u.all = mask->src.u.all;
3563
3564         if (cda[CTA_EXPECT_NAT]) {
3565                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3566                                                  exp, nf_ct_l3num(ct));
3567                 if (err < 0)
3568                         goto err_out;
3569         }
3570         return exp;
3571 err_out:
3572         nf_ct_expect_put(exp);
3573         return ERR_PTR(err);
3574 }
3575
3576 static int
3577 ctnetlink_create_expect(struct net *net,
3578                         const struct nf_conntrack_zone *zone,
3579                         const struct nlattr * const cda[],
3580                         u_int8_t u3, u32 portid, int report)
3581 {
3582         struct nf_conntrack_tuple tuple, mask, master_tuple;
3583         struct nf_conntrack_tuple_hash *h = NULL;
3584         struct nf_conntrack_helper *helper = NULL;
3585         struct nf_conntrack_expect *exp;
3586         struct nf_conn *ct;
3587         int err;
3588
3589         /* caller guarantees that those three CTA_EXPECT_* exist */
3590         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3591                                     u3, NULL);
3592         if (err < 0)
3593                 return err;
3594         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3595                                     u3, NULL);
3596         if (err < 0)
3597                 return err;
3598         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3599                                     u3, NULL);
3600         if (err < 0)
3601                 return err;
3602
3603         /* Look for master conntrack of this expectation */
3604         h = nf_conntrack_find_get(net, zone, &master_tuple);
3605         if (!h)
3606                 return -ENOENT;
3607         ct = nf_ct_tuplehash_to_ctrack(h);
3608
3609         rcu_read_lock();
3610         if (cda[CTA_EXPECT_HELP_NAME]) {
3611                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3612
3613                 helper = __nf_conntrack_helper_find(helpname, u3,
3614                                                     nf_ct_protonum(ct));
3615                 if (helper == NULL) {
3616                         rcu_read_unlock();
3617 #ifdef CONFIG_MODULES
3618                         if (request_module("nfct-helper-%s", helpname) < 0) {
3619                                 err = -EOPNOTSUPP;
3620                                 goto err_ct;
3621                         }
3622                         rcu_read_lock();
3623                         helper = __nf_conntrack_helper_find(helpname, u3,
3624                                                             nf_ct_protonum(ct));
3625                         if (helper) {
3626                                 err = -EAGAIN;
3627                                 goto err_rcu;
3628                         }
3629                         rcu_read_unlock();
3630 #endif
3631                         err = -EOPNOTSUPP;
3632                         goto err_ct;
3633                 }
3634         }
3635
3636         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3637         if (IS_ERR(exp)) {
3638                 err = PTR_ERR(exp);
3639                 goto err_rcu;
3640         }
3641
3642         err = nf_ct_expect_related_report(exp, portid, report, 0);
3643         nf_ct_expect_put(exp);
3644 err_rcu:
3645         rcu_read_unlock();
3646 err_ct:
3647         nf_ct_put(ct);
3648         return err;
3649 }
3650
3651 static int ctnetlink_new_expect(struct sk_buff *skb,
3652                                 const struct nfnl_info *info,
3653                                 const struct nlattr * const cda[])
3654 {
3655         u_int8_t u3 = info->nfmsg->nfgen_family;
3656         struct nf_conntrack_tuple tuple;
3657         struct nf_conntrack_expect *exp;
3658         struct nf_conntrack_zone zone;
3659         int err;
3660
3661         if (!cda[CTA_EXPECT_TUPLE]
3662             || !cda[CTA_EXPECT_MASK]
3663             || !cda[CTA_EXPECT_MASTER])
3664                 return -EINVAL;
3665
3666         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3667         if (err < 0)
3668                 return err;
3669
3670         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3671                                     u3, NULL);
3672         if (err < 0)
3673                 return err;
3674
3675         spin_lock_bh(&nf_conntrack_expect_lock);
3676         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3677         if (!exp) {
3678                 spin_unlock_bh(&nf_conntrack_expect_lock);
3679                 err = -ENOENT;
3680                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3681                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3682                                                       NETLINK_CB(skb).portid,
3683                                                       nlmsg_report(info->nlh));
3684                 }
3685                 return err;
3686         }
3687
3688         err = -EEXIST;
3689         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3690                 err = ctnetlink_change_expect(exp, cda);
3691         spin_unlock_bh(&nf_conntrack_expect_lock);
3692
3693         return err;
3694 }
3695
3696 static int
3697 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3698                              const struct ip_conntrack_stat *st)
3699 {
3700         struct nlmsghdr *nlh;
3701         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3702
3703         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3704                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3705         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3706                            NFNETLINK_V0, htons(cpu));
3707         if (!nlh)
3708                 goto nlmsg_failure;
3709
3710         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3711             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3712             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3713                 goto nla_put_failure;
3714
3715         nlmsg_end(skb, nlh);
3716         return skb->len;
3717
3718 nla_put_failure:
3719 nlmsg_failure:
3720         nlmsg_cancel(skb, nlh);
3721         return -1;
3722 }
3723
3724 static int
3725 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3726 {
3727         int cpu;
3728         struct net *net = sock_net(skb->sk);
3729
3730         if (cb->args[0] == nr_cpu_ids)
3731                 return 0;
3732
3733         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3734                 const struct ip_conntrack_stat *st;
3735
3736                 if (!cpu_possible(cpu))
3737                         continue;
3738
3739                 st = per_cpu_ptr(net->ct.stat, cpu);
3740                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3741                                                  cb->nlh->nlmsg_seq,
3742                                                  cpu, st) < 0)
3743                         break;
3744         }
3745         cb->args[0] = cpu;
3746
3747         return skb->len;
3748 }
3749
3750 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3751                                   const struct nfnl_info *info,
3752                                   const struct nlattr * const cda[])
3753 {
3754         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3755                 struct netlink_dump_control c = {
3756                         .dump = ctnetlink_exp_stat_cpu_dump,
3757                 };
3758                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3759         }
3760
3761         return 0;
3762 }
3763
3764 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3765 static struct nf_ct_event_notifier ctnl_notifier = {
3766         .ct_event = ctnetlink_conntrack_event,
3767         .exp_event = ctnetlink_expect_event,
3768 };
3769 #endif
3770
3771 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3772         [IPCTNL_MSG_CT_NEW]     = {
3773                 .call           = ctnetlink_new_conntrack,
3774                 .type           = NFNL_CB_MUTEX,
3775                 .attr_count     = CTA_MAX,
3776                 .policy         = ct_nla_policy
3777         },
3778         [IPCTNL_MSG_CT_GET]     = {
3779                 .call           = ctnetlink_get_conntrack,
3780                 .type           = NFNL_CB_MUTEX,
3781                 .attr_count     = CTA_MAX,
3782                 .policy         = ct_nla_policy
3783         },
3784         [IPCTNL_MSG_CT_DELETE]  = {
3785                 .call           = ctnetlink_del_conntrack,
3786                 .type           = NFNL_CB_MUTEX,
3787                 .attr_count     = CTA_MAX,
3788                 .policy         = ct_nla_policy
3789         },
3790         [IPCTNL_MSG_CT_GET_CTRZERO] = {
3791                 .call           = ctnetlink_get_conntrack,
3792                 .type           = NFNL_CB_MUTEX,
3793                 .attr_count     = CTA_MAX,
3794                 .policy         = ct_nla_policy
3795         },
3796         [IPCTNL_MSG_CT_GET_STATS_CPU] = {
3797                 .call           = ctnetlink_stat_ct_cpu,
3798                 .type           = NFNL_CB_MUTEX,
3799         },
3800         [IPCTNL_MSG_CT_GET_STATS] = {
3801                 .call           = ctnetlink_stat_ct,
3802                 .type           = NFNL_CB_MUTEX,
3803         },
3804         [IPCTNL_MSG_CT_GET_DYING] = {
3805                 .call           = ctnetlink_get_ct_dying,
3806                 .type           = NFNL_CB_MUTEX,
3807         },
3808         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = {
3809                 .call           = ctnetlink_get_ct_unconfirmed,
3810                 .type           = NFNL_CB_MUTEX,
3811         },
3812 };
3813
3814 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3815         [IPCTNL_MSG_EXP_GET] = {
3816                 .call           = ctnetlink_get_expect,
3817                 .type           = NFNL_CB_MUTEX,
3818                 .attr_count     = CTA_EXPECT_MAX,
3819                 .policy         = exp_nla_policy
3820         },
3821         [IPCTNL_MSG_EXP_NEW] = {
3822                 .call           = ctnetlink_new_expect,
3823                 .type           = NFNL_CB_MUTEX,
3824                 .attr_count     = CTA_EXPECT_MAX,
3825                 .policy         = exp_nla_policy
3826         },
3827         [IPCTNL_MSG_EXP_DELETE] = {
3828                 .call           = ctnetlink_del_expect,
3829                 .type           = NFNL_CB_MUTEX,
3830                 .attr_count     = CTA_EXPECT_MAX,
3831                 .policy         = exp_nla_policy
3832         },
3833         [IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3834                 .call           = ctnetlink_stat_exp_cpu,
3835                 .type           = NFNL_CB_MUTEX,
3836         },
3837 };
3838
3839 static const struct nfnetlink_subsystem ctnl_subsys = {
3840         .name                           = "conntrack",
3841         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3842         .cb_count                       = IPCTNL_MSG_MAX,
3843         .cb                             = ctnl_cb,
3844 };
3845
3846 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3847         .name                           = "conntrack_expect",
3848         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3849         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3850         .cb                             = ctnl_exp_cb,
3851 };
3852
3853 MODULE_ALIAS("ip_conntrack_netlink");
3854 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3855 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3856
3857 static int __net_init ctnetlink_net_init(struct net *net)
3858 {
3859 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3860         nf_conntrack_register_notifier(net, &ctnl_notifier);
3861 #endif
3862         return 0;
3863 }
3864
3865 static void ctnetlink_net_pre_exit(struct net *net)
3866 {
3867 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3868         nf_conntrack_unregister_notifier(net);
3869 #endif
3870 }
3871
3872 static struct pernet_operations ctnetlink_net_ops = {
3873         .init           = ctnetlink_net_init,
3874         .pre_exit       = ctnetlink_net_pre_exit,
3875 };
3876
3877 static int __init ctnetlink_init(void)
3878 {
3879         int ret;
3880
3881         ret = nfnetlink_subsys_register(&ctnl_subsys);
3882         if (ret < 0) {
3883                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3884                 goto err_out;
3885         }
3886
3887         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3888         if (ret < 0) {
3889                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3890                 goto err_unreg_subsys;
3891         }
3892
3893         ret = register_pernet_subsys(&ctnetlink_net_ops);
3894         if (ret < 0) {
3895                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3896                 goto err_unreg_exp_subsys;
3897         }
3898 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3899         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3900         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3901 #endif
3902         return 0;
3903
3904 err_unreg_exp_subsys:
3905         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3906 err_unreg_subsys:
3907         nfnetlink_subsys_unregister(&ctnl_subsys);
3908 err_out:
3909         return ret;
3910 }
3911
3912 static void __exit ctnetlink_exit(void)
3913 {
3914         unregister_pernet_subsys(&ctnetlink_net_ops);
3915         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3916         nfnetlink_subsys_unregister(&ctnl_subsys);
3917 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3918         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3919 #endif
3920         synchronize_rcu();
3921 }
3922
3923 module_init(ctnetlink_init);
3924 module_exit(ctnetlink_exit);