net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32 #include <linux/siphash.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netlink.h>
  36 #include <net/sock.h>
  37 #include <net/netfilter/nf_conntrack.h>
  38 #include <net/netfilter/nf_conntrack_core.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_seqadj.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #include <net/netfilter/nf_conntrack_synproxy.h>
  49 #if IS_ENABLED(CONFIG_NF_NAT)
  50 #include <net/netfilter/nf_nat.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 #include "nf_internals.h"
  58
  59 MODULE_LICENSE("GPL");
  60
  61 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  62                                 const struct nf_conntrack_tuple *tuple,
  63                                 const struct nf_conntrack_l4proto *l4proto)
  64 {
  65         int ret = 0;
  66         struct nlattr *nest_parms;
  67
  68         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO);
  69         if (!nest_parms)
  70                 goto nla_put_failure;
  71         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  72                 goto nla_put_failure;
  73
  74         if (likely(l4proto->tuple_to_nlattr))
  75                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  76
  77         nla_nest_end(skb, nest_parms);
  78
  79         return ret;
  80
  81 nla_put_failure:
  82         return -1;
  83 }
  84
  85 static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  86                                 const struct nf_conntrack_tuple *tuple)
  87 {
  88         if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  89             nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  90                 return -EMSGSIZE;
  91         return 0;
  92 }
  93
  94 static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
  95                                 const struct nf_conntrack_tuple *tuple)
  96 {
  97         if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
  98             nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
  99                 return -EMSGSIZE;
 100         return 0;
 101 }
 102
 103 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 104                                     const struct nf_conntrack_tuple *tuple)
 105 {
 106         int ret = 0;
 107         struct nlattr *nest_parms;
 108
 109         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP);
 110         if (!nest_parms)
 111                 goto nla_put_failure;
 112
 113         switch (tuple->src.l3num) {
 114         case NFPROTO_IPV4:
 115                 ret = ipv4_tuple_to_nlattr(skb, tuple);
 116                 break;
 117         case NFPROTO_IPV6:
 118                 ret = ipv6_tuple_to_nlattr(skb, tuple);
 119                 break;
 120         }
 121
 122         nla_nest_end(skb, nest_parms);
 123
 124         return ret;
 125
 126 nla_put_failure:
 127         return -1;
 128 }
 129
 130 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 131                                  const struct nf_conntrack_tuple *tuple)
 132 {
 133         const struct nf_conntrack_l4proto *l4proto;
 134         int ret;
 135
 136         rcu_read_lock();
 137         ret = ctnetlink_dump_tuples_ip(skb, tuple);
 138
 139         if (ret >= 0) {
 140                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 141                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 142         }
 143         rcu_read_unlock();
 144         return ret;
 145 }
 146
 147 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 148                                   const struct nf_conntrack_zone *zone, int dir)
 149 {
 150         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 151                 return 0;
 152         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 153                 goto nla_put_failure;
 154         return 0;
 155
 156 nla_put_failure:
 157         return -1;
 158 }
 159
 160 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 161 {
 162         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 163                 goto nla_put_failure;
 164         return 0;
 165
 166 nla_put_failure:
 167         return -1;
 168 }
 169
 170 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct,
 171                                   bool skip_zero)
 172 {
 173         long timeout = nf_ct_expires(ct) / HZ;
 174
 175         if (skip_zero && timeout == 0)
 176                 return 0;
 177
 178         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 179                 goto nla_put_failure;
 180         return 0;
 181
 182 nla_put_failure:
 183         return -1;
 184 }
 185
 186 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct,
 187                                     bool destroy)
 188 {
 189         const struct nf_conntrack_l4proto *l4proto;
 190         struct nlattr *nest_proto;
 191         int ret;
 192
 193         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 194         if (!l4proto->to_nlattr)
 195                 return 0;
 196
 197         nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 198         if (!nest_proto)
 199                 goto nla_put_failure;
 200
 201         ret = l4proto->to_nlattr(skb, nest_proto, ct, destroy);
 202
 203         nla_nest_end(skb, nest_proto);
 204
 205         return ret;
 206
 207 nla_put_failure:
 208         return -1;
 209 }
 210
 211 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 212                                    const struct nf_conn *ct)
 213 {
 214         struct nlattr *nest_helper;
 215         const struct nf_conn_help *help = nfct_help(ct);
 216         struct nf_conntrack_helper *helper;
 217
 218         if (!help)
 219                 return 0;
 220
 221         rcu_read_lock();
 222         helper = rcu_dereference(help->helper);
 223         if (!helper)
 224                 goto out;
 225
 226         nest_helper = nla_nest_start(skb, CTA_HELP);
 227         if (!nest_helper)
 228                 goto nla_put_failure;
 229         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 230                 goto nla_put_failure;
 231
 232         if (helper->to_nlattr)
 233                 helper->to_nlattr(skb, ct);
 234
 235         nla_nest_end(skb, nest_helper);
 236 out:
 237         rcu_read_unlock();
 238         return 0;
 239
 240 nla_put_failure:
 241         rcu_read_unlock();
 242         return -1;
 243 }
 244
 245 static int
 246 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 247               enum ip_conntrack_dir dir, int type)
 248 {
 249         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 250         struct nf_conn_counter *counter = acct->counter;
 251         struct nlattr *nest_count;
 252         u64 pkts, bytes;
 253
 254         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 255                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 256                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 257         } else {
 258                 pkts = atomic64_read(&counter[dir].packets);
 259                 bytes = atomic64_read(&counter[dir].bytes);
 260         }
 261
 262         nest_count = nla_nest_start(skb, attr);
 263         if (!nest_count)
 264                 goto nla_put_failure;
 265
 266         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 267                          CTA_COUNTERS_PAD) ||
 268             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 269                          CTA_COUNTERS_PAD))
 270                 goto nla_put_failure;
 271
 272         nla_nest_end(skb, nest_count);
 273
 274         return 0;
 275
 276 nla_put_failure:
 277         return -1;
 278 }
 279
 280 static int
 281 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 282 {
 283         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 284
 285         if (!acct)
 286                 return 0;
 287
 288         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 289                 return -1;
 290         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 291                 return -1;
 292
 293         return 0;
 294 }
 295
 296 static int
 297 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 298 {
 299         struct nlattr *nest_count;
 300         const struct nf_conn_tstamp *tstamp;
 301
 302         tstamp = nf_conn_tstamp_find(ct);
 303         if (!tstamp)
 304                 return 0;
 305
 306         nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 307         if (!nest_count)
 308                 goto nla_put_failure;
 309
 310         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 311                          CTA_TIMESTAMP_PAD) ||
 312             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 313                                                cpu_to_be64(tstamp->stop),
 314                                                CTA_TIMESTAMP_PAD)))
 315                 goto nla_put_failure;
 316         nla_nest_end(skb, nest_count);
 317
 318         return 0;
 319
 320 nla_put_failure:
 321         return -1;
 322 }
 323
 324 #ifdef CONFIG_NF_CONNTRACK_MARK
 325 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 326 {
 327         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 328                 goto nla_put_failure;
 329         return 0;
 330
 331 nla_put_failure:
 332         return -1;
 333 }
 334 #else
 335 #define ctnetlink_dump_mark(a, b) (0)
 336 #endif
 337
 338 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 339 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 340 {
 341         struct nlattr *nest_secctx;
 342         int len, ret;
 343         char *secctx;
 344
 345         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 346         if (ret)
 347                 return 0;
 348
 349         ret = -1;
 350         nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 351         if (!nest_secctx)
 352                 goto nla_put_failure;
 353
 354         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 355                 goto nla_put_failure;
 356         nla_nest_end(skb, nest_secctx);
 357
 358         ret = 0;
 359 nla_put_failure:
 360         security_release_secctx(secctx, len);
 361         return ret;
 362 }
 363 #else
 364 #define ctnetlink_dump_secctx(a, b) (0)
 365 #endif
 366
 367 #ifdef CONFIG_NF_CONNTRACK_LABELS
 368 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 369 {
 370         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 371
 372         if (!labels)
 373                 return 0;
 374         return nla_total_size(sizeof(labels->bits));
 375 }
 376
 377 static int
 378 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 379 {
 380         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 381         unsigned int i;
 382
 383         if (!labels)
 384                 return 0;
 385
 386         i = 0;
 387         do {
 388                 if (labels->bits[i] != 0)
 389                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 390                                        labels->bits);
 391                 i++;
 392         } while (i < ARRAY_SIZE(labels->bits));
 393
 394         return 0;
 395 }
 396 #else
 397 #define ctnetlink_dump_labels(a, b) (0)
 398 #define ctnetlink_label_size(a) (0)
 399 #endif
 400
 401 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 402
 403 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 404 {
 405         struct nlattr *nest_parms;
 406
 407         if (!(ct->status & IPS_EXPECTED))
 408                 return 0;
 409
 410         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 411         if (!nest_parms)
 412                 goto nla_put_failure;
 413         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 414                 goto nla_put_failure;
 415         nla_nest_end(skb, nest_parms);
 416
 417         return 0;
 418
 419 nla_put_failure:
 420         return -1;
 421 }
 422
 423 static int
 424 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 425 {
 426         struct nlattr *nest_parms;
 427
 428         nest_parms = nla_nest_start(skb, type);
 429         if (!nest_parms)
 430                 goto nla_put_failure;
 431
 432         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 433                          htonl(seq->correction_pos)) ||
 434             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 435                          htonl(seq->offset_before)) ||
 436             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 437                          htonl(seq->offset_after)))
 438                 goto nla_put_failure;
 439
 440         nla_nest_end(skb, nest_parms);
 441
 442         return 0;
 443
 444 nla_put_failure:
 445         return -1;
 446 }
 447
 448 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 449 {
 450         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 451         struct nf_ct_seqadj *seq;
 452
 453         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 454                 return 0;
 455
 456         spin_lock_bh(&ct->lock);
 457         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 458         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 459                 goto err;
 460
 461         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 462         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 463                 goto err;
 464
 465         spin_unlock_bh(&ct->lock);
 466         return 0;
 467 err:
 468         spin_unlock_bh(&ct->lock);
 469         return -1;
 470 }
 471
 472 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 473 {
 474         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 475         struct nlattr *nest_parms;
 476
 477         if (!synproxy)
 478                 return 0;
 479
 480         nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 481         if (!nest_parms)
 482                 goto nla_put_failure;
 483
 484         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 485             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 486             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 487                 goto nla_put_failure;
 488
 489         nla_nest_end(skb, nest_parms);
 490
 491         return 0;
 492
 493 nla_put_failure:
 494         return -1;
 495 }
 496
 497 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 498 {
 499         __be32 id = (__force __be32)nf_ct_get_id(ct);
 500
 501         if (nla_put_be32(skb, CTA_ID, id))
 502                 goto nla_put_failure;
 503         return 0;
 504
 505 nla_put_failure:
 506         return -1;
 507 }
 508
 509 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 510 {
 511         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 512                 goto nla_put_failure;
 513         return 0;
 514
 515 nla_put_failure:
 516         return -1;
 517 }
 518
 519 /* all these functions access ct->ext. Caller must either hold a reference
 520  * on ct or prevent its deletion by holding either the bucket spinlock or
 521  * pcpu dying list lock.
 522  */
 523 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 524                                   struct nf_conn *ct, u32 type)
 525 {
 526         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 527             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 528             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 529             ctnetlink_dump_labels(skb, ct) < 0 ||
 530             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 531             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 532                 return -1;
 533
 534         return 0;
 535 }
 536
 537 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 538 {
 539         if (ctnetlink_dump_status(skb, ct) < 0 ||
 540             ctnetlink_dump_mark(skb, ct) < 0 ||
 541             ctnetlink_dump_secctx(skb, ct) < 0 ||
 542             ctnetlink_dump_id(skb, ct) < 0 ||
 543             ctnetlink_dump_use(skb, ct) < 0 ||
 544             ctnetlink_dump_master(skb, ct) < 0)
 545                 return -1;
 546
 547         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 548             (ctnetlink_dump_timeout(skb, ct, false) < 0 ||
 549              ctnetlink_dump_protoinfo(skb, ct, false) < 0))
 550                 return -1;
 551
 552         return 0;
 553 }
 554
 555 static int
 556 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 557                     struct nf_conn *ct, bool extinfo, unsigned int flags)
 558 {
 559         const struct nf_conntrack_zone *zone;
 560         struct nlmsghdr *nlh;
 561         struct nlattr *nest_parms;
 562         unsigned int event;
 563
 564         if (portid)
 565                 flags |= NLM_F_MULTI;
 566         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 567         nlh = nfnl_msg_put(skb, portid, seq, event, flags, nf_ct_l3num(ct),
 568                            NFNETLINK_V0, 0);
 569         if (!nlh)
 570                 goto nlmsg_failure;
 571
 572         zone = nf_ct_zone(ct);
 573
 574         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 575         if (!nest_parms)
 576                 goto nla_put_failure;
 577         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 578                 goto nla_put_failure;
 579         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 580                                    NF_CT_ZONE_DIR_ORIG) < 0)
 581                 goto nla_put_failure;
 582         nla_nest_end(skb, nest_parms);
 583
 584         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 585         if (!nest_parms)
 586                 goto nla_put_failure;
 587         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 588                 goto nla_put_failure;
 589         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 590                                    NF_CT_ZONE_DIR_REPL) < 0)
 591                 goto nla_put_failure;
 592         nla_nest_end(skb, nest_parms);
 593
 594         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 595                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 596                 goto nla_put_failure;
 597
 598         if (ctnetlink_dump_info(skb, ct) < 0)
 599                 goto nla_put_failure;
 600         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 601                 goto nla_put_failure;
 602
 603         nlmsg_end(skb, nlh);
 604         return skb->len;
 605
 606 nlmsg_failure:
 607 nla_put_failure:
 608         nlmsg_cancel(skb, nlh);
 609         return -1;
 610 }
 611
 612 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 613         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 614         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 615         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 616         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 617 };
 618
 619 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 620 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 621 {
 622         const struct nf_conntrack_l4proto *l4proto;
 623         size_t len, len4 = 0;
 624
 625         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 626         len *= 3u; /* ORIG, REPLY, MASTER */
 627
 628         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 629         len += l4proto->nlattr_size;
 630         if (l4proto->nlattr_tuple_size) {
 631                 len4 = l4proto->nlattr_tuple_size();
 632                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 633         }
 634
 635         return len + len4;
 636 }
 637 #endif
 638
 639 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 640 {
 641         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 642                 return 0;
 643         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 644                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 645                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 646                ;
 647 }
 648
 649 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 650 {
 651 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 652         int len, ret;
 653
 654         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 655         if (ret)
 656                 return 0;
 657
 658         return nla_total_size(0) /* CTA_SECCTX */
 659                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 660 #else
 661         return 0;
 662 #endif
 663 }
 664
 665 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 666 {
 667 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 668         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 669                 return 0;
 670         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 671 #else
 672         return 0;
 673 #endif
 674 }
 675
 676 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 677 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 678 {
 679         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 680                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 681                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 682                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 683                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 684                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 685                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 686                + ctnetlink_acct_size(ct)
 687                + ctnetlink_timestamp_size(ct)
 688                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 689                + nla_total_size(0) /* CTA_PROTOINFO */
 690                + nla_total_size(0) /* CTA_HELP */
 691                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 692                + ctnetlink_secctx_size(ct)
 693 #if IS_ENABLED(CONFIG_NF_NAT)
 694                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 695                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 696 #endif
 697 #ifdef CONFIG_NF_CONNTRACK_MARK
 698                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 699 #endif
 700 #ifdef CONFIG_NF_CONNTRACK_ZONES
 701                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 702 #endif
 703                + ctnetlink_proto_size(ct)
 704                + ctnetlink_label_size(ct)
 705                ;
 706 }
 707
 708 static int
 709 ctnetlink_conntrack_event(unsigned int events, 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->orig_flags);
1015                 if (err < 0) {
1016                         err = -EINVAL;
1017                         goto err_filter;
1018                 }
1019         }
1020
1021         return filter;
1022
1023 err_filter:
1024         kfree(filter);
1025
1026         return ERR_PTR(err);
1027 }
1028
1029 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1030 {
1031         return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS];
1032 }
1033
1034 static int ctnetlink_start(struct netlink_callback *cb)
1035 {
1036         const struct nlattr * const *cda = cb->data;
1037         struct ctnetlink_filter *filter = NULL;
1038         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1039         u8 family = nfmsg->nfgen_family;
1040
1041         if (ctnetlink_needs_filter(family, cda)) {
1042                 filter = ctnetlink_alloc_filter(cda, family);
1043                 if (IS_ERR(filter))
1044                         return PTR_ERR(filter);
1045         }
1046
1047         cb->data = filter;
1048         return 0;
1049 }
1050
1051 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1052                                         struct nf_conntrack_tuple *ct_tuple,
1053                                         u_int32_t flags, int family)
1054 {
1055         switch (family) {
1056         case NFPROTO_IPV4:
1057                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1058                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1059                         return  0;
1060
1061                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1062                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1063                         return  0;
1064                 break;
1065         case NFPROTO_IPV6:
1066                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1067                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1068                                    &ct_tuple->src.u3.in6))
1069                         return 0;
1070
1071                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1072                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1073                                    &ct_tuple->dst.u3.in6))
1074                         return 0;
1075                 break;
1076         }
1077
1078         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1079             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1080                 return 0;
1081
1082         switch (ct_tuple->dst.protonum) {
1083         case IPPROTO_TCP:
1084         case IPPROTO_UDP:
1085                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1086                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1087                         return 0;
1088
1089                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1090                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1091                         return 0;
1092                 break;
1093         case IPPROTO_ICMP:
1094                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1095                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1096                         return 0;
1097                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1098                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1099                         return 0;
1100                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1101                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1102                         return 0;
1103                 break;
1104         case IPPROTO_ICMPV6:
1105                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1106                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1107                         return 0;
1108                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1109                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1110                         return 0;
1111                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1112                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1113                         return 0;
1114                 break;
1115         }
1116
1117         return 1;
1118 }
1119
1120 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1121 {
1122         struct ctnetlink_filter *filter = data;
1123         struct nf_conntrack_tuple *tuple;
1124         u32 status;
1125
1126         if (filter == NULL)
1127                 goto out;
1128
1129         /* Match entries of a given L3 protocol number.
1130          * If it is not specified, ie. l3proto == 0,
1131          * then match everything.
1132          */
1133         if (filter->family && nf_ct_l3num(ct) != filter->family)
1134                 goto ignore_entry;
1135
1136         if (filter->orig_flags) {
1137                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1138                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1139                                                   filter->orig_flags,
1140                                                   filter->family))
1141                         goto ignore_entry;
1142         }
1143
1144         if (filter->reply_flags) {
1145                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1146                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1147                                                   filter->reply_flags,
1148                                                   filter->family))
1149                         goto ignore_entry;
1150         }
1151
1152 #ifdef CONFIG_NF_CONNTRACK_MARK
1153         if ((ct->mark & filter->mark.mask) != filter->mark.val)
1154                 goto ignore_entry;
1155 #endif
1156         status = (u32)READ_ONCE(ct->status);
1157         if ((status & filter->status.mask) != filter->status.val)
1158                 goto ignore_entry;
1159
1160 out:
1161         return 1;
1162
1163 ignore_entry:
1164         return 0;
1165 }
1166
1167 static int
1168 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1169 {
1170         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1171         struct net *net = sock_net(skb->sk);
1172         struct nf_conn *ct, *last;
1173         struct nf_conntrack_tuple_hash *h;
1174         struct hlist_nulls_node *n;
1175         struct nf_conn *nf_ct_evict[8];
1176         int res, i;
1177         spinlock_t *lockp;
1178
1179         last = (struct nf_conn *)cb->args[1];
1180         i = 0;
1181
1182         local_bh_disable();
1183         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1184 restart:
1185                 while (i) {
1186                         i--;
1187                         if (nf_ct_should_gc(nf_ct_evict[i]))
1188                                 nf_ct_kill(nf_ct_evict[i]);
1189                         nf_ct_put(nf_ct_evict[i]);
1190                 }
1191
1192                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1193                 nf_conntrack_lock(lockp);
1194                 if (cb->args[0] >= nf_conntrack_htable_size) {
1195                         spin_unlock(lockp);
1196                         goto out;
1197                 }
1198                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1199                                            hnnode) {
1200                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1201                                 continue;
1202                         ct = nf_ct_tuplehash_to_ctrack(h);
1203                         if (nf_ct_is_expired(ct)) {
1204                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1205                                     atomic_inc_not_zero(&ct->ct_general.use))
1206                                         nf_ct_evict[i++] = ct;
1207                                 continue;
1208                         }
1209
1210                         if (!net_eq(net, nf_ct_net(ct)))
1211                                 continue;
1212
1213                         if (cb->args[1]) {
1214                                 if (ct != last)
1215                                         continue;
1216                                 cb->args[1] = 0;
1217                         }
1218                         if (!ctnetlink_filter_match(ct, cb->data))
1219                                 continue;
1220
1221                         res =
1222                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1223                                             cb->nlh->nlmsg_seq,
1224                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1225                                             ct, true, flags);
1226                         if (res < 0) {
1227                                 nf_conntrack_get(&ct->ct_general);
1228                                 cb->args[1] = (unsigned long)ct;
1229                                 spin_unlock(lockp);
1230                                 goto out;
1231                         }
1232                 }
1233                 spin_unlock(lockp);
1234                 if (cb->args[1]) {
1235                         cb->args[1] = 0;
1236                         goto restart;
1237                 }
1238         }
1239 out:
1240         local_bh_enable();
1241         if (last) {
1242                 /* nf ct hash resize happened, now clear the leftover. */
1243                 if ((struct nf_conn *)cb->args[1] == last)
1244                         cb->args[1] = 0;
1245
1246                 nf_ct_put(last);
1247         }
1248
1249         while (i) {
1250                 i--;
1251                 if (nf_ct_should_gc(nf_ct_evict[i]))
1252                         nf_ct_kill(nf_ct_evict[i]);
1253                 nf_ct_put(nf_ct_evict[i]);
1254         }
1255
1256         return skb->len;
1257 }
1258
1259 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1260                                 struct nf_conntrack_tuple *t,
1261                                 u_int32_t flags)
1262 {
1263         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1264                 if (!tb[CTA_IP_V4_SRC])
1265                         return -EINVAL;
1266
1267                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1268         }
1269
1270         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1271                 if (!tb[CTA_IP_V4_DST])
1272                         return -EINVAL;
1273
1274                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1275         }
1276
1277         return 0;
1278 }
1279
1280 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1281                                 struct nf_conntrack_tuple *t,
1282                                 u_int32_t flags)
1283 {
1284         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1285                 if (!tb[CTA_IP_V6_SRC])
1286                         return -EINVAL;
1287
1288                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1289         }
1290
1291         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1292                 if (!tb[CTA_IP_V6_DST])
1293                         return -EINVAL;
1294
1295                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1296         }
1297
1298         return 0;
1299 }
1300
1301 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1302                                     struct nf_conntrack_tuple *tuple,
1303                                     u_int32_t flags)
1304 {
1305         struct nlattr *tb[CTA_IP_MAX+1];
1306         int ret = 0;
1307
1308         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1309         if (ret < 0)
1310                 return ret;
1311
1312         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1313                                              cta_ip_nla_policy, NULL);
1314         if (ret)
1315                 return ret;
1316
1317         switch (tuple->src.l3num) {
1318         case NFPROTO_IPV4:
1319                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1320                 break;
1321         case NFPROTO_IPV6:
1322                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1323                 break;
1324         }
1325
1326         return ret;
1327 }
1328
1329 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1330         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1331 };
1332
1333 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1334                                        struct nf_conntrack_tuple *tuple,
1335                                        u_int32_t flags)
1336 {
1337         const struct nf_conntrack_l4proto *l4proto;
1338         struct nlattr *tb[CTA_PROTO_MAX+1];
1339         int ret = 0;
1340
1341         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1342                                           proto_nla_policy, NULL);
1343         if (ret < 0)
1344                 return ret;
1345
1346         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1347                 return 0;
1348
1349         if (!tb[CTA_PROTO_NUM])
1350                 return -EINVAL;
1351
1352         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1353
1354         rcu_read_lock();
1355         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1356
1357         if (likely(l4proto->nlattr_to_tuple)) {
1358                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1359                                                      l4proto->nla_policy,
1360                                                      NULL);
1361                 if (ret == 0)
1362                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1363         }
1364
1365         rcu_read_unlock();
1366
1367         return ret;
1368 }
1369
1370 static int
1371 ctnetlink_parse_zone(const struct nlattr *attr,
1372                      struct nf_conntrack_zone *zone)
1373 {
1374         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1375                         NF_CT_DEFAULT_ZONE_DIR, 0);
1376 #ifdef CONFIG_NF_CONNTRACK_ZONES
1377         if (attr)
1378                 zone->id = ntohs(nla_get_be16(attr));
1379 #else
1380         if (attr)
1381                 return -EOPNOTSUPP;
1382 #endif
1383         return 0;
1384 }
1385
1386 static int
1387 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1388                            struct nf_conntrack_zone *zone)
1389 {
1390         int ret;
1391
1392         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1393                 return -EINVAL;
1394
1395         ret = ctnetlink_parse_zone(attr, zone);
1396         if (ret < 0)
1397                 return ret;
1398
1399         if (type == CTA_TUPLE_REPLY)
1400                 zone->dir = NF_CT_ZONE_DIR_REPL;
1401         else
1402                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1403
1404         return 0;
1405 }
1406
1407 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1408         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1409         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1410         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1411 };
1412
1413 #define CTA_FILTER_F_ALL_CTA_PROTO \
1414   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1415    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1416    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1417    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1418    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1419    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1420    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1421    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1422
1423 static int
1424 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1425                               struct nf_conntrack_tuple *tuple, u32 type,
1426                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1427                               u_int32_t flags)
1428 {
1429         struct nlattr *tb[CTA_TUPLE_MAX+1];
1430         int err;
1431
1432         memset(tuple, 0, sizeof(*tuple));
1433
1434         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1435                                           tuple_nla_policy, NULL);
1436         if (err < 0)
1437                 return err;
1438
1439         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1440                 return -EOPNOTSUPP;
1441         tuple->src.l3num = l3num;
1442
1443         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1444             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1445                 if (!tb[CTA_TUPLE_IP])
1446                         return -EINVAL;
1447
1448                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1449                 if (err < 0)
1450                         return err;
1451         }
1452
1453         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1454                 if (!tb[CTA_TUPLE_PROTO])
1455                         return -EINVAL;
1456
1457                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1458                 if (err < 0)
1459                         return err;
1460         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1461                 /* Can't manage proto flags without a protonum  */
1462                 return -EINVAL;
1463         }
1464
1465         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1466                 if (!zone)
1467                         return -EINVAL;
1468
1469                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1470                                                  type, zone);
1471                 if (err < 0)
1472                         return err;
1473         }
1474
1475         /* orig and expect tuples get DIR_ORIGINAL */
1476         if (type == CTA_TUPLE_REPLY)
1477                 tuple->dst.dir = IP_CT_DIR_REPLY;
1478         else
1479                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1480
1481         return 0;
1482 }
1483
1484 static int
1485 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1486                       struct nf_conntrack_tuple *tuple, u32 type,
1487                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1488 {
1489         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1490                                             CTA_FILTER_FLAG(ALL));
1491 }
1492
1493 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1494         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1495                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1496 };
1497
1498 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1499                                 struct nlattr **helpinfo)
1500 {
1501         int err;
1502         struct nlattr *tb[CTA_HELP_MAX+1];
1503
1504         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1505                                           help_nla_policy, NULL);
1506         if (err < 0)
1507                 return err;
1508
1509         if (!tb[CTA_HELP_NAME])
1510                 return -EINVAL;
1511
1512         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1513
1514         if (tb[CTA_HELP_INFO])
1515                 *helpinfo = tb[CTA_HELP_INFO];
1516
1517         return 0;
1518 }
1519
1520 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1521         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1522         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1523         [CTA_STATUS]            = { .type = NLA_U32 },
1524         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1525         [CTA_HELP]              = { .type = NLA_NESTED },
1526         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1527         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1528         [CTA_MARK]              = { .type = NLA_U32 },
1529         [CTA_ID]                = { .type = NLA_U32 },
1530         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1531         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1532         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1533         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1534         [CTA_ZONE]              = { .type = NLA_U16 },
1535         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1536         [CTA_LABELS]            = { .type = NLA_BINARY,
1537                                     .len = NF_CT_LABELS_MAX_SIZE },
1538         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1539                                     .len = NF_CT_LABELS_MAX_SIZE },
1540         [CTA_FILTER]            = { .type = NLA_NESTED },
1541         [CTA_STATUS_MASK]       = { .type = NLA_U32 },
1542 };
1543
1544 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1545 {
1546         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1547                 return 0;
1548
1549         return ctnetlink_filter_match(ct, data);
1550 }
1551
1552 static int ctnetlink_flush_conntrack(struct net *net,
1553                                      const struct nlattr * const cda[],
1554                                      u32 portid, int report, u8 family)
1555 {
1556         struct ctnetlink_filter *filter = NULL;
1557
1558         if (ctnetlink_needs_filter(family, cda)) {
1559                 if (cda[CTA_FILTER])
1560                         return -EOPNOTSUPP;
1561
1562                 filter = ctnetlink_alloc_filter(cda, family);
1563                 if (IS_ERR(filter))
1564                         return PTR_ERR(filter);
1565         }
1566
1567         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1568                                   portid, report);
1569         kfree(filter);
1570
1571         return 0;
1572 }
1573
1574 static int ctnetlink_del_conntrack(struct sk_buff *skb,
1575                                    const struct nfnl_info *info,
1576                                    const struct nlattr * const cda[])
1577 {
1578         u8 family = info->nfmsg->nfgen_family;
1579         struct nf_conntrack_tuple_hash *h;
1580         struct nf_conntrack_tuple tuple;
1581         struct nf_conntrack_zone zone;
1582         struct nf_conn *ct;
1583         int err;
1584
1585         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1586         if (err < 0)
1587                 return err;
1588
1589         if (cda[CTA_TUPLE_ORIG])
1590                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1591                                             family, &zone);
1592         else if (cda[CTA_TUPLE_REPLY])
1593                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1594                                             family, &zone);
1595         else {
1596                 u_int8_t u3 = info->nfmsg->version ? family : AF_UNSPEC;
1597
1598                 return ctnetlink_flush_conntrack(info->net, cda,
1599                                                  NETLINK_CB(skb).portid,
1600                                                  nlmsg_report(info->nlh), u3);
1601         }
1602
1603         if (err < 0)
1604                 return err;
1605
1606         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1607         if (!h)
1608                 return -ENOENT;
1609
1610         ct = nf_ct_tuplehash_to_ctrack(h);
1611
1612         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1613                 nf_ct_put(ct);
1614                 return -EBUSY;
1615         }
1616
1617         if (cda[CTA_ID]) {
1618                 __be32 id = nla_get_be32(cda[CTA_ID]);
1619
1620                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1621                         nf_ct_put(ct);
1622                         return -ENOENT;
1623                 }
1624         }
1625
1626         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1627         nf_ct_put(ct);
1628
1629         return 0;
1630 }
1631
1632 static int ctnetlink_get_conntrack(struct sk_buff *skb,
1633                                    const struct nfnl_info *info,
1634                                    const struct nlattr * const cda[])
1635 {
1636         u_int8_t u3 = info->nfmsg->nfgen_family;
1637         struct nf_conntrack_tuple_hash *h;
1638         struct nf_conntrack_tuple tuple;
1639         struct nf_conntrack_zone zone;
1640         struct sk_buff *skb2;
1641         struct nf_conn *ct;
1642         int err;
1643
1644         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1645                 struct netlink_dump_control c = {
1646                         .start = ctnetlink_start,
1647                         .dump = ctnetlink_dump_table,
1648                         .done = ctnetlink_done,
1649                         .data = (void *)cda,
1650                 };
1651
1652                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1653         }
1654
1655         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1656         if (err < 0)
1657                 return err;
1658
1659         if (cda[CTA_TUPLE_ORIG])
1660                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1661                                             u3, &zone);
1662         else if (cda[CTA_TUPLE_REPLY])
1663                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1664                                             u3, &zone);
1665         else
1666                 return -EINVAL;
1667
1668         if (err < 0)
1669                 return err;
1670
1671         h = nf_conntrack_find_get(info->net, &zone, &tuple);
1672         if (!h)
1673                 return -ENOENT;
1674
1675         ct = nf_ct_tuplehash_to_ctrack(h);
1676
1677         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1678         if (!skb2) {
1679                 nf_ct_put(ct);
1680                 return -ENOMEM;
1681         }
1682
1683         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1684                                   info->nlh->nlmsg_seq,
1685                                   NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1686                                   true, 0);
1687         nf_ct_put(ct);
1688         if (err <= 0) {
1689                 kfree_skb(skb2);
1690                 return -ENOMEM;
1691         }
1692
1693         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1694 }
1695
1696 static int ctnetlink_done_list(struct netlink_callback *cb)
1697 {
1698         if (cb->args[1])
1699                 nf_ct_put((struct nf_conn *)cb->args[1]);
1700         return 0;
1701 }
1702
1703 static int
1704 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1705 {
1706         struct nf_conn *ct, *last;
1707         struct nf_conntrack_tuple_hash *h;
1708         struct hlist_nulls_node *n;
1709         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1710         u_int8_t l3proto = nfmsg->nfgen_family;
1711         int res;
1712         int cpu;
1713         struct hlist_nulls_head *list;
1714         struct net *net = sock_net(skb->sk);
1715
1716         if (cb->args[2])
1717                 return 0;
1718
1719         last = (struct nf_conn *)cb->args[1];
1720
1721         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1722                 struct ct_pcpu *pcpu;
1723
1724                 if (!cpu_possible(cpu))
1725                         continue;
1726
1727                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1728                 spin_lock_bh(&pcpu->lock);
1729                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1730 restart:
1731                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1732                         ct = nf_ct_tuplehash_to_ctrack(h);
1733                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1734                                 continue;
1735                         if (cb->args[1]) {
1736                                 if (ct != last)
1737                                         continue;
1738                                 cb->args[1] = 0;
1739                         }
1740
1741                         /* We can't dump extension info for the unconfirmed
1742                          * list because unconfirmed conntracks can have
1743                          * ct->ext reallocated (and thus freed).
1744                          *
1745                          * In the dying list case ct->ext can't be free'd
1746                          * until after we drop pcpu->lock.
1747                          */
1748                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1749                                                   cb->nlh->nlmsg_seq,
1750                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1751                                                   ct, dying ? true : false, 0);
1752                         if (res < 0) {
1753                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1754                                         continue;
1755                                 cb->args[0] = cpu;
1756                                 cb->args[1] = (unsigned long)ct;
1757                                 spin_unlock_bh(&pcpu->lock);
1758                                 goto out;
1759                         }
1760                 }
1761                 if (cb->args[1]) {
1762                         cb->args[1] = 0;
1763                         goto restart;
1764                 }
1765                 spin_unlock_bh(&pcpu->lock);
1766         }
1767         cb->args[2] = 1;
1768 out:
1769         if (last)
1770                 nf_ct_put(last);
1771
1772         return skb->len;
1773 }
1774
1775 static int
1776 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1777 {
1778         return ctnetlink_dump_list(skb, cb, true);
1779 }
1780
1781 static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1782                                   const struct nfnl_info *info,
1783                                   const struct nlattr * const cda[])
1784 {
1785         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1786                 struct netlink_dump_control c = {
1787                         .dump = ctnetlink_dump_dying,
1788                         .done = ctnetlink_done_list,
1789                 };
1790                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1791         }
1792
1793         return -EOPNOTSUPP;
1794 }
1795
1796 static int
1797 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1798 {
1799         return ctnetlink_dump_list(skb, cb, false);
1800 }
1801
1802 static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1803                                         const struct nfnl_info *info,
1804                                         const struct nlattr * const cda[])
1805 {
1806         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1807                 struct netlink_dump_control c = {
1808                         .dump = ctnetlink_dump_unconfirmed,
1809                         .done = ctnetlink_done_list,
1810                 };
1811                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
1812         }
1813
1814         return -EOPNOTSUPP;
1815 }
1816
1817 #if IS_ENABLED(CONFIG_NF_NAT)
1818 static int
1819 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1820                           enum nf_nat_manip_type manip,
1821                           const struct nlattr *attr)
1822         __must_hold(RCU)
1823 {
1824         struct nf_nat_hook *nat_hook;
1825         int err;
1826
1827         nat_hook = rcu_dereference(nf_nat_hook);
1828         if (!nat_hook) {
1829 #ifdef CONFIG_MODULES
1830                 rcu_read_unlock();
1831                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1832                 if (request_module("nf-nat") < 0) {
1833                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1834                         rcu_read_lock();
1835                         return -EOPNOTSUPP;
1836                 }
1837                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1838                 rcu_read_lock();
1839                 nat_hook = rcu_dereference(nf_nat_hook);
1840                 if (nat_hook)
1841                         return -EAGAIN;
1842 #endif
1843                 return -EOPNOTSUPP;
1844         }
1845
1846         err = nat_hook->parse_nat_setup(ct, manip, attr);
1847         if (err == -EAGAIN) {
1848 #ifdef CONFIG_MODULES
1849                 rcu_read_unlock();
1850                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1851                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1852                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1853                         rcu_read_lock();
1854                         return -EOPNOTSUPP;
1855                 }
1856                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1857                 rcu_read_lock();
1858 #else
1859                 err = -EOPNOTSUPP;
1860 #endif
1861         }
1862         return err;
1863 }
1864 #endif
1865
1866 static void
1867 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1868                           unsigned long off)
1869 {
1870         unsigned int bit;
1871
1872         /* Ignore these unchangable bits */
1873         on &= ~IPS_UNCHANGEABLE_MASK;
1874         off &= ~IPS_UNCHANGEABLE_MASK;
1875
1876         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1877                 if (on & (1 << bit))
1878                         set_bit(bit, &ct->status);
1879                 else if (off & (1 << bit))
1880                         clear_bit(bit, &ct->status);
1881         }
1882 }
1883
1884 static int
1885 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1886 {
1887         unsigned long d;
1888         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1889         d = ct->status ^ status;
1890
1891         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1892                 /* unchangeable */
1893                 return -EBUSY;
1894
1895         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1896                 /* SEEN_REPLY bit can only be set */
1897                 return -EBUSY;
1898
1899         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1900                 /* ASSURED bit can only be set */
1901                 return -EBUSY;
1902
1903         __ctnetlink_change_status(ct, status, 0);
1904         return 0;
1905 }
1906
1907 static int
1908 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1909 {
1910 #if IS_ENABLED(CONFIG_NF_NAT)
1911         int ret;
1912
1913         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1914                 return 0;
1915
1916         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1917                                         cda[CTA_NAT_DST]);
1918         if (ret < 0)
1919                 return ret;
1920
1921         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1922                                          cda[CTA_NAT_SRC]);
1923 #else
1924         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1925                 return 0;
1926         return -EOPNOTSUPP;
1927 #endif
1928 }
1929
1930 static int ctnetlink_change_helper(struct nf_conn *ct,
1931                                    const struct nlattr * const cda[])
1932 {
1933         struct nf_conntrack_helper *helper;
1934         struct nf_conn_help *help = nfct_help(ct);
1935         char *helpname = NULL;
1936         struct nlattr *helpinfo = NULL;
1937         int err;
1938
1939         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1940         if (err < 0)
1941                 return err;
1942
1943         /* don't change helper of sibling connections */
1944         if (ct->master) {
1945                 /* If we try to change the helper to the same thing twice,
1946                  * treat the second attempt as a no-op instead of returning
1947                  * an error.
1948                  */
1949                 err = -EBUSY;
1950                 if (help) {
1951                         rcu_read_lock();
1952                         helper = rcu_dereference(help->helper);
1953                         if (helper && !strcmp(helper->name, helpname))
1954                                 err = 0;
1955                         rcu_read_unlock();
1956                 }
1957
1958                 return err;
1959         }
1960
1961         if (!strcmp(helpname, "")) {
1962                 if (help && help->helper) {
1963                         /* we had a helper before ... */
1964                         nf_ct_remove_expectations(ct);
1965                         RCU_INIT_POINTER(help->helper, NULL);
1966                 }
1967
1968                 return 0;
1969         }
1970
1971         rcu_read_lock();
1972         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1973                                             nf_ct_protonum(ct));
1974         if (helper == NULL) {
1975                 rcu_read_unlock();
1976                 return -EOPNOTSUPP;
1977         }
1978
1979         if (help) {
1980                 if (help->helper == helper) {
1981                         /* update private helper data if allowed. */
1982                         if (helper->from_nlattr)
1983                                 helper->from_nlattr(helpinfo, ct);
1984                         err = 0;
1985                 } else
1986                         err = -EBUSY;
1987         } else {
1988                 /* we cannot set a helper for an existing conntrack */
1989                 err = -EOPNOTSUPP;
1990         }
1991
1992         rcu_read_unlock();
1993         return err;
1994 }
1995
1996 static int ctnetlink_change_timeout(struct nf_conn *ct,
1997                                     const struct nlattr * const cda[])
1998 {
1999         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2000
2001         if (timeout > INT_MAX)
2002                 timeout = INT_MAX;
2003         ct->timeout = nfct_time_stamp + (u32)timeout;
2004
2005         if (test_bit(IPS_DYING_BIT, &ct->status))
2006                 return -ETIME;
2007
2008         return 0;
2009 }
2010
2011 #if defined(CONFIG_NF_CONNTRACK_MARK)
2012 static void ctnetlink_change_mark(struct nf_conn *ct,
2013                                     const struct nlattr * const cda[])
2014 {
2015         u32 mark, newmark, mask = 0;
2016
2017         if (cda[CTA_MARK_MASK])
2018                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2019
2020         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2021         newmark = (ct->mark & mask) ^ mark;
2022         if (newmark != ct->mark)
2023                 ct->mark = newmark;
2024 }
2025 #endif
2026
2027 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2028         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2029         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2030         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2031 };
2032
2033 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2034                                       const struct nlattr * const cda[])
2035 {
2036         const struct nlattr *attr = cda[CTA_PROTOINFO];
2037         const struct nf_conntrack_l4proto *l4proto;
2038         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2039         int err = 0;
2040
2041         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2042                                           protoinfo_policy, NULL);
2043         if (err < 0)
2044                 return err;
2045
2046         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2047         if (l4proto->from_nlattr)
2048                 err = l4proto->from_nlattr(tb, ct);
2049
2050         return err;
2051 }
2052
2053 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2054         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2055         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2056         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2057 };
2058
2059 static int change_seq_adj(struct nf_ct_seqadj *seq,
2060                           const struct nlattr * const attr)
2061 {
2062         int err;
2063         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2064
2065         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2066                                           seqadj_policy, NULL);
2067         if (err < 0)
2068                 return err;
2069
2070         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2071                 return -EINVAL;
2072
2073         seq->correction_pos =
2074                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2075
2076         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2077                 return -EINVAL;
2078
2079         seq->offset_before =
2080                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2081
2082         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2083                 return -EINVAL;
2084
2085         seq->offset_after =
2086                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2087
2088         return 0;
2089 }
2090
2091 static int
2092 ctnetlink_change_seq_adj(struct nf_conn *ct,
2093                          const struct nlattr * const cda[])
2094 {
2095         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2096         int ret = 0;
2097
2098         if (!seqadj)
2099                 return 0;
2100
2101         spin_lock_bh(&ct->lock);
2102         if (cda[CTA_SEQ_ADJ_ORIG]) {
2103                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2104                                      cda[CTA_SEQ_ADJ_ORIG]);
2105                 if (ret < 0)
2106                         goto err;
2107
2108                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2109         }
2110
2111         if (cda[CTA_SEQ_ADJ_REPLY]) {
2112                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2113                                      cda[CTA_SEQ_ADJ_REPLY]);
2114                 if (ret < 0)
2115                         goto err;
2116
2117                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2118         }
2119
2120         spin_unlock_bh(&ct->lock);
2121         return 0;
2122 err:
2123         spin_unlock_bh(&ct->lock);
2124         return ret;
2125 }
2126
2127 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2128         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2129         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2130         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2131 };
2132
2133 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2134                                      const struct nlattr * const cda[])
2135 {
2136         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2137         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2138         int err;
2139
2140         if (!synproxy)
2141                 return 0;
2142
2143         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2144                                           cda[CTA_SYNPROXY], synproxy_policy,
2145                                           NULL);
2146         if (err < 0)
2147                 return err;
2148
2149         if (!tb[CTA_SYNPROXY_ISN] ||
2150             !tb[CTA_SYNPROXY_ITS] ||
2151             !tb[CTA_SYNPROXY_TSOFF])
2152                 return -EINVAL;
2153
2154         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2155         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2156         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2157
2158         return 0;
2159 }
2160
2161 static int
2162 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2163 {
2164 #ifdef CONFIG_NF_CONNTRACK_LABELS
2165         size_t len = nla_len(cda[CTA_LABELS]);
2166         const void *mask = cda[CTA_LABELS_MASK];
2167
2168         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2169                 return -EINVAL;
2170
2171         if (mask) {
2172                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2173                     nla_len(cda[CTA_LABELS_MASK]) != len)
2174                         return -EINVAL;
2175                 mask = nla_data(cda[CTA_LABELS_MASK]);
2176         }
2177
2178         len /= sizeof(u32);
2179
2180         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2181 #else
2182         return -EOPNOTSUPP;
2183 #endif
2184 }
2185
2186 static int
2187 ctnetlink_change_conntrack(struct nf_conn *ct,
2188                            const struct nlattr * const cda[])
2189 {
2190         int err;
2191
2192         /* only allow NAT changes and master assignation for new conntracks */
2193         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2194                 return -EOPNOTSUPP;
2195
2196         if (cda[CTA_HELP]) {
2197                 err = ctnetlink_change_helper(ct, cda);
2198                 if (err < 0)
2199                         return err;
2200         }
2201
2202         if (cda[CTA_TIMEOUT]) {
2203                 err = ctnetlink_change_timeout(ct, cda);
2204                 if (err < 0)
2205                         return err;
2206         }
2207
2208         if (cda[CTA_STATUS]) {
2209                 err = ctnetlink_change_status(ct, cda);
2210                 if (err < 0)
2211                         return err;
2212         }
2213
2214         if (cda[CTA_PROTOINFO]) {
2215                 err = ctnetlink_change_protoinfo(ct, cda);
2216                 if (err < 0)
2217                         return err;
2218         }
2219
2220 #if defined(CONFIG_NF_CONNTRACK_MARK)
2221         if (cda[CTA_MARK])
2222                 ctnetlink_change_mark(ct, cda);
2223 #endif
2224
2225         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2226                 err = ctnetlink_change_seq_adj(ct, cda);
2227                 if (err < 0)
2228                         return err;
2229         }
2230
2231         if (cda[CTA_SYNPROXY]) {
2232                 err = ctnetlink_change_synproxy(ct, cda);
2233                 if (err < 0)
2234                         return err;
2235         }
2236
2237         if (cda[CTA_LABELS]) {
2238                 err = ctnetlink_attach_labels(ct, cda);
2239                 if (err < 0)
2240                         return err;
2241         }
2242
2243         return 0;
2244 }
2245
2246 static struct nf_conn *
2247 ctnetlink_create_conntrack(struct net *net,
2248                            const struct nf_conntrack_zone *zone,
2249                            const struct nlattr * const cda[],
2250                            struct nf_conntrack_tuple *otuple,
2251                            struct nf_conntrack_tuple *rtuple,
2252                            u8 u3)
2253 {
2254         struct nf_conn *ct;
2255         int err = -EINVAL;
2256         struct nf_conntrack_helper *helper;
2257         struct nf_conn_tstamp *tstamp;
2258         u64 timeout;
2259
2260         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2261         if (IS_ERR(ct))
2262                 return ERR_PTR(-ENOMEM);
2263
2264         if (!cda[CTA_TIMEOUT])
2265                 goto err1;
2266
2267         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2268         if (timeout > INT_MAX)
2269                 timeout = INT_MAX;
2270         ct->timeout = (u32)timeout + nfct_time_stamp;
2271
2272         rcu_read_lock();
2273         if (cda[CTA_HELP]) {
2274                 char *helpname = NULL;
2275                 struct nlattr *helpinfo = NULL;
2276
2277                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2278                 if (err < 0)
2279                         goto err2;
2280
2281                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2282                                                     nf_ct_protonum(ct));
2283                 if (helper == NULL) {
2284                         rcu_read_unlock();
2285 #ifdef CONFIG_MODULES
2286                         if (request_module("nfct-helper-%s", helpname) < 0) {
2287                                 err = -EOPNOTSUPP;
2288                                 goto err1;
2289                         }
2290
2291                         rcu_read_lock();
2292                         helper = __nf_conntrack_helper_find(helpname,
2293                                                             nf_ct_l3num(ct),
2294                                                             nf_ct_protonum(ct));
2295                         if (helper) {
2296                                 err = -EAGAIN;
2297                                 goto err2;
2298                         }
2299                         rcu_read_unlock();
2300 #endif
2301                         err = -EOPNOTSUPP;
2302                         goto err1;
2303                 } else {
2304                         struct nf_conn_help *help;
2305
2306                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2307                         if (help == NULL) {
2308                                 err = -ENOMEM;
2309                                 goto err2;
2310                         }
2311                         /* set private helper data if allowed. */
2312                         if (helper->from_nlattr)
2313                                 helper->from_nlattr(helpinfo, ct);
2314
2315                         /* not in hash table yet so not strictly necessary */
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                 goto nla_put_failure;
2533
2534         nlmsg_end(skb, nlh);
2535         return skb->len;
2536
2537 nla_put_failure:
2538 nlmsg_failure:
2539         nlmsg_cancel(skb, nlh);
2540         return -1;
2541 }
2542
2543 static int
2544 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2545 {
2546         int cpu;
2547         struct net *net = sock_net(skb->sk);
2548
2549         if (cb->args[0] == nr_cpu_ids)
2550                 return 0;
2551
2552         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2553                 const struct ip_conntrack_stat *st;
2554
2555                 if (!cpu_possible(cpu))
2556                         continue;
2557
2558                 st = per_cpu_ptr(net->ct.stat, cpu);
2559                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2560                                                     NETLINK_CB(cb->skb).portid,
2561                                                     cb->nlh->nlmsg_seq,
2562                                                     cpu, st) < 0)
2563                                 break;
2564         }
2565         cb->args[0] = cpu;
2566
2567         return skb->len;
2568 }
2569
2570 static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2571                                  const struct nfnl_info *info,
2572                                  const struct nlattr * const cda[])
2573 {
2574         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2575                 struct netlink_dump_control c = {
2576                         .dump = ctnetlink_ct_stat_cpu_dump,
2577                 };
2578                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
2579         }
2580
2581         return 0;
2582 }
2583
2584 static int
2585 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2586                             struct net *net)
2587 {
2588         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2589         unsigned int nr_conntracks;
2590         struct nlmsghdr *nlh;
2591
2592         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2593         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2594                            NFNETLINK_V0, 0);
2595         if (!nlh)
2596                 goto nlmsg_failure;
2597
2598         nr_conntracks = nf_conntrack_count(net);
2599         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2600                 goto nla_put_failure;
2601
2602         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2603                 goto nla_put_failure;
2604
2605         nlmsg_end(skb, nlh);
2606         return skb->len;
2607
2608 nla_put_failure:
2609 nlmsg_failure:
2610         nlmsg_cancel(skb, nlh);
2611         return -1;
2612 }
2613
2614 static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2615                              const struct nlattr * const cda[])
2616 {
2617         struct sk_buff *skb2;
2618         int err;
2619
2620         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2621         if (skb2 == NULL)
2622                 return -ENOMEM;
2623
2624         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2625                                           info->nlh->nlmsg_seq,
2626                                           NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2627                                           sock_net(skb->sk));
2628         if (err <= 0) {
2629                 kfree_skb(skb2);
2630                 return -ENOMEM;
2631         }
2632
2633         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2634 }
2635
2636 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2637         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2638         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2639         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2640         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2641         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2642         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2643                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2644         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2645         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2646         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2647         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2648         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2649 };
2650
2651 static struct nf_conntrack_expect *
2652 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2653                        struct nf_conntrack_helper *helper,
2654                        struct nf_conntrack_tuple *tuple,
2655                        struct nf_conntrack_tuple *mask);
2656
2657 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2658 static size_t
2659 ctnetlink_glue_build_size(const struct nf_conn *ct)
2660 {
2661         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2662                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2663                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2664                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2665                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2666                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2667                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2668                + nla_total_size(0) /* CTA_PROTOINFO */
2669                + nla_total_size(0) /* CTA_HELP */
2670                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2671                + ctnetlink_secctx_size(ct)
2672                + ctnetlink_acct_size(ct)
2673                + ctnetlink_timestamp_size(ct)
2674 #if IS_ENABLED(CONFIG_NF_NAT)
2675                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2676                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2677 #endif
2678 #ifdef CONFIG_NF_CONNTRACK_MARK
2679                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2680 #endif
2681 #ifdef CONFIG_NF_CONNTRACK_ZONES
2682                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2683 #endif
2684                + ctnetlink_proto_size(ct)
2685                ;
2686 }
2687
2688 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2689 {
2690         const struct nf_conntrack_zone *zone;
2691         struct nlattr *nest_parms;
2692
2693         zone = nf_ct_zone(ct);
2694
2695         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2696         if (!nest_parms)
2697                 goto nla_put_failure;
2698         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2699                 goto nla_put_failure;
2700         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2701                                    NF_CT_ZONE_DIR_ORIG) < 0)
2702                 goto nla_put_failure;
2703         nla_nest_end(skb, nest_parms);
2704
2705         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2706         if (!nest_parms)
2707                 goto nla_put_failure;
2708         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2709                 goto nla_put_failure;
2710         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2711                                    NF_CT_ZONE_DIR_REPL) < 0)
2712                 goto nla_put_failure;
2713         nla_nest_end(skb, nest_parms);
2714
2715         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2716                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2717                 goto nla_put_failure;
2718
2719         if (ctnetlink_dump_id(skb, ct) < 0)
2720                 goto nla_put_failure;
2721
2722         if (ctnetlink_dump_status(skb, ct) < 0)
2723                 goto nla_put_failure;
2724
2725         if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2726                 goto nla_put_failure;
2727
2728         if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2729                 goto nla_put_failure;
2730
2731         if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2732             ctnetlink_dump_timestamp(skb, ct) < 0)
2733                 goto nla_put_failure;
2734
2735         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2736                 goto nla_put_failure;
2737
2738 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2739         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2740                 goto nla_put_failure;
2741 #endif
2742         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2743                 goto nla_put_failure;
2744
2745         if ((ct->status & IPS_SEQ_ADJUST) &&
2746             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2747                 goto nla_put_failure;
2748
2749         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2750                 goto nla_put_failure;
2751
2752 #ifdef CONFIG_NF_CONNTRACK_MARK
2753         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2754                 goto nla_put_failure;
2755 #endif
2756         if (ctnetlink_dump_labels(skb, ct) < 0)
2757                 goto nla_put_failure;
2758         return 0;
2759
2760 nla_put_failure:
2761         return -ENOSPC;
2762 }
2763
2764 static int
2765 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2766                      enum ip_conntrack_info ctinfo,
2767                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2768 {
2769         struct nlattr *nest_parms;
2770
2771         nest_parms = nla_nest_start(skb, ct_attr);
2772         if (!nest_parms)
2773                 goto nla_put_failure;
2774
2775         if (__ctnetlink_glue_build(skb, ct) < 0)
2776                 goto nla_put_failure;
2777
2778         nla_nest_end(skb, nest_parms);
2779
2780         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2781                 goto nla_put_failure;
2782
2783         return 0;
2784
2785 nla_put_failure:
2786         return -ENOSPC;
2787 }
2788
2789 static int
2790 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2791 {
2792         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2793         unsigned long d = ct->status ^ status;
2794
2795         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2796                 /* SEEN_REPLY bit can only be set */
2797                 return -EBUSY;
2798
2799         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2800                 /* ASSURED bit can only be set */
2801                 return -EBUSY;
2802
2803         /* This check is less strict than ctnetlink_change_status()
2804          * because callers often flip IPS_EXPECTED bits when sending
2805          * an NFQA_CT attribute to the kernel.  So ignore the
2806          * unchangeable bits but do not error out. Also user programs
2807          * are allowed to clear the bits that they are allowed to change.
2808          */
2809         __ctnetlink_change_status(ct, status, ~status);
2810         return 0;
2811 }
2812
2813 static int
2814 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2815 {
2816         int err;
2817
2818         if (cda[CTA_TIMEOUT]) {
2819                 err = ctnetlink_change_timeout(ct, cda);
2820                 if (err < 0)
2821                         return err;
2822         }
2823         if (cda[CTA_STATUS]) {
2824                 err = ctnetlink_update_status(ct, cda);
2825                 if (err < 0)
2826                         return err;
2827         }
2828         if (cda[CTA_HELP]) {
2829                 err = ctnetlink_change_helper(ct, cda);
2830                 if (err < 0)
2831                         return err;
2832         }
2833         if (cda[CTA_LABELS]) {
2834                 err = ctnetlink_attach_labels(ct, cda);
2835                 if (err < 0)
2836                         return err;
2837         }
2838 #if defined(CONFIG_NF_CONNTRACK_MARK)
2839         if (cda[CTA_MARK]) {
2840                 ctnetlink_change_mark(ct, cda);
2841         }
2842 #endif
2843         return 0;
2844 }
2845
2846 static int
2847 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2848 {
2849         struct nlattr *cda[CTA_MAX+1];
2850         int ret;
2851
2852         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2853                                           NULL);
2854         if (ret < 0)
2855                 return ret;
2856
2857         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2858 }
2859
2860 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2861                                     const struct nf_conn *ct,
2862                                     struct nf_conntrack_tuple *tuple,
2863                                     struct nf_conntrack_tuple *mask)
2864 {
2865         int err;
2866
2867         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2868                                     nf_ct_l3num(ct), NULL);
2869         if (err < 0)
2870                 return err;
2871
2872         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2873                                      nf_ct_l3num(ct), NULL);
2874 }
2875
2876 static int
2877 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2878                              u32 portid, u32 report)
2879 {
2880         struct nlattr *cda[CTA_EXPECT_MAX+1];
2881         struct nf_conntrack_tuple tuple, mask;
2882         struct nf_conntrack_helper *helper = NULL;
2883         struct nf_conntrack_expect *exp;
2884         int err;
2885
2886         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2887                                           exp_nla_policy, NULL);
2888         if (err < 0)
2889                 return err;
2890
2891         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2892                                        ct, &tuple, &mask);
2893         if (err < 0)
2894                 return err;
2895
2896         if (cda[CTA_EXPECT_HELP_NAME]) {
2897                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2898
2899                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2900                                                     nf_ct_protonum(ct));
2901                 if (helper == NULL)
2902                         return -EOPNOTSUPP;
2903         }
2904
2905         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2906                                      helper, &tuple, &mask);
2907         if (IS_ERR(exp))
2908                 return PTR_ERR(exp);
2909
2910         err = nf_ct_expect_related_report(exp, portid, report, 0);
2911         nf_ct_expect_put(exp);
2912         return err;
2913 }
2914
2915 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2916                                   enum ip_conntrack_info ctinfo, int diff)
2917 {
2918         if (!(ct->status & IPS_NAT_MASK))
2919                 return;
2920
2921         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2922 }
2923
2924 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2925         .build_size     = ctnetlink_glue_build_size,
2926         .build          = ctnetlink_glue_build,
2927         .parse          = ctnetlink_glue_parse,
2928         .attach_expect  = ctnetlink_glue_attach_expect,
2929         .seq_adjust     = ctnetlink_glue_seqadj,
2930 };
2931 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2932
2933 /***********************************************************************
2934  * EXPECT
2935  ***********************************************************************/
2936
2937 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2938                                     const struct nf_conntrack_tuple *tuple,
2939                                     u32 type)
2940 {
2941         struct nlattr *nest_parms;
2942
2943         nest_parms = nla_nest_start(skb, type);
2944         if (!nest_parms)
2945                 goto nla_put_failure;
2946         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2947                 goto nla_put_failure;
2948         nla_nest_end(skb, nest_parms);
2949
2950         return 0;
2951
2952 nla_put_failure:
2953         return -1;
2954 }
2955
2956 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2957                                    const struct nf_conntrack_tuple *tuple,
2958                                    const struct nf_conntrack_tuple_mask *mask)
2959 {
2960         const struct nf_conntrack_l4proto *l4proto;
2961         struct nf_conntrack_tuple m;
2962         struct nlattr *nest_parms;
2963         int ret;
2964
2965         memset(&m, 0xFF, sizeof(m));
2966         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2967         m.src.u.all = mask->src.u.all;
2968         m.src.l3num = tuple->src.l3num;
2969         m.dst.protonum = tuple->dst.protonum;
2970
2971         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2972         if (!nest_parms)
2973                 goto nla_put_failure;
2974
2975         rcu_read_lock();
2976         ret = ctnetlink_dump_tuples_ip(skb, &m);
2977         if (ret >= 0) {
2978                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2979                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2980         }
2981         rcu_read_unlock();
2982
2983         if (unlikely(ret < 0))
2984                 goto nla_put_failure;
2985
2986         nla_nest_end(skb, nest_parms);
2987
2988         return 0;
2989
2990 nla_put_failure:
2991         return -1;
2992 }
2993
2994 static const union nf_inet_addr any_addr;
2995
2996 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2997 {
2998         static __read_mostly siphash_key_t exp_id_seed;
2999         unsigned long a, b, c, d;
3000
3001         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
3002
3003         a = (unsigned long)exp;
3004         b = (unsigned long)exp->helper;
3005         c = (unsigned long)exp->master;
3006         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
3007
3008 #ifdef CONFIG_64BIT
3009         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
3010 #else
3011         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
3012 #endif
3013 }
3014
3015 static int
3016 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3017                           const struct nf_conntrack_expect *exp)
3018 {
3019         struct nf_conn *master = exp->master;
3020         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3021         struct nf_conn_help *help;
3022 #if IS_ENABLED(CONFIG_NF_NAT)
3023         struct nlattr *nest_parms;
3024         struct nf_conntrack_tuple nat_tuple = {};
3025 #endif
3026         struct nf_ct_helper_expectfn *expfn;
3027
3028         if (timeout < 0)
3029                 timeout = 0;
3030
3031         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3032                 goto nla_put_failure;
3033         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3034                 goto nla_put_failure;
3035         if (ctnetlink_exp_dump_tuple(skb,
3036                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3037                                  CTA_EXPECT_MASTER) < 0)
3038                 goto nla_put_failure;
3039
3040 #if IS_ENABLED(CONFIG_NF_NAT)
3041         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3042             exp->saved_proto.all) {
3043                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3044                 if (!nest_parms)
3045                         goto nla_put_failure;
3046
3047                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3048                         goto nla_put_failure;
3049
3050                 nat_tuple.src.l3num = nf_ct_l3num(master);
3051                 nat_tuple.src.u3 = exp->saved_addr;
3052                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3053                 nat_tuple.src.u = exp->saved_proto;
3054
3055                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3056                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3057                         goto nla_put_failure;
3058                 nla_nest_end(skb, nest_parms);
3059         }
3060 #endif
3061         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3062             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3063             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3064             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3065                 goto nla_put_failure;
3066         help = nfct_help(master);
3067         if (help) {
3068                 struct nf_conntrack_helper *helper;
3069
3070                 helper = rcu_dereference(help->helper);
3071                 if (helper &&
3072                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3073                         goto nla_put_failure;
3074         }
3075         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3076         if (expfn != NULL &&
3077             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3078                 goto nla_put_failure;
3079
3080         return 0;
3081
3082 nla_put_failure:
3083         return -1;
3084 }
3085
3086 static int
3087 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3088                         int event, const struct nf_conntrack_expect *exp)
3089 {
3090         struct nlmsghdr *nlh;
3091         unsigned int flags = portid ? NLM_F_MULTI : 0;
3092
3093         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3094         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3095                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3096         if (!nlh)
3097                 goto nlmsg_failure;
3098
3099         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3100                 goto nla_put_failure;
3101
3102         nlmsg_end(skb, nlh);
3103         return skb->len;
3104
3105 nlmsg_failure:
3106 nla_put_failure:
3107         nlmsg_cancel(skb, nlh);
3108         return -1;
3109 }
3110
3111 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3112 static int
3113 ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3114 {
3115         struct nf_conntrack_expect *exp = item->exp;
3116         struct net *net = nf_ct_exp_net(exp);
3117         struct nlmsghdr *nlh;
3118         struct sk_buff *skb;
3119         unsigned int type, group;
3120         int flags = 0;
3121
3122         if (events & (1 << IPEXP_DESTROY)) {
3123                 type = IPCTNL_MSG_EXP_DELETE;
3124                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3125         } else if (events & (1 << IPEXP_NEW)) {
3126                 type = IPCTNL_MSG_EXP_NEW;
3127                 flags = NLM_F_CREATE|NLM_F_EXCL;
3128                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3129         } else
3130                 return 0;
3131
3132         if (!item->report && !nfnetlink_has_listeners(net, group))
3133                 return 0;
3134
3135         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3136         if (skb == NULL)
3137                 goto errout;
3138
3139         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3140         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3141                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3142         if (!nlh)
3143                 goto nlmsg_failure;
3144
3145         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3146                 goto nla_put_failure;
3147
3148         nlmsg_end(skb, nlh);
3149         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3150         return 0;
3151
3152 nla_put_failure:
3153         nlmsg_cancel(skb, nlh);
3154 nlmsg_failure:
3155         kfree_skb(skb);
3156 errout:
3157         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3158         return 0;
3159 }
3160 #endif
3161 static int ctnetlink_exp_done(struct netlink_callback *cb)
3162 {
3163         if (cb->args[1])
3164                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3165         return 0;
3166 }
3167
3168 static int
3169 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3170 {
3171         struct net *net = sock_net(skb->sk);
3172         struct nf_conntrack_expect *exp, *last;
3173         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3174         u_int8_t l3proto = nfmsg->nfgen_family;
3175
3176         rcu_read_lock();
3177         last = (struct nf_conntrack_expect *)cb->args[1];
3178         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3179 restart:
3180                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3181                                          hnode) {
3182                         if (l3proto && exp->tuple.src.l3num != l3proto)
3183                                 continue;
3184
3185                         if (!net_eq(nf_ct_net(exp->master), net))
3186                                 continue;
3187
3188                         if (cb->args[1]) {
3189                                 if (exp != last)
3190                                         continue;
3191                                 cb->args[1] = 0;
3192                         }
3193                         if (ctnetlink_exp_fill_info(skb,
3194                                                     NETLINK_CB(cb->skb).portid,
3195                                                     cb->nlh->nlmsg_seq,
3196                                                     IPCTNL_MSG_EXP_NEW,
3197                                                     exp) < 0) {
3198                                 if (!refcount_inc_not_zero(&exp->use))
3199                                         continue;
3200                                 cb->args[1] = (unsigned long)exp;
3201                                 goto out;
3202                         }
3203                 }
3204                 if (cb->args[1]) {
3205                         cb->args[1] = 0;
3206                         goto restart;
3207                 }
3208         }
3209 out:
3210         rcu_read_unlock();
3211         if (last)
3212                 nf_ct_expect_put(last);
3213
3214         return skb->len;
3215 }
3216
3217 static int
3218 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3219 {
3220         struct nf_conntrack_expect *exp, *last;
3221         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3222         struct nf_conn *ct = cb->data;
3223         struct nf_conn_help *help = nfct_help(ct);
3224         u_int8_t l3proto = nfmsg->nfgen_family;
3225
3226         if (cb->args[0])
3227                 return 0;
3228
3229         rcu_read_lock();
3230         last = (struct nf_conntrack_expect *)cb->args[1];
3231 restart:
3232         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3233                 if (l3proto && exp->tuple.src.l3num != l3proto)
3234                         continue;
3235                 if (cb->args[1]) {
3236                         if (exp != last)
3237                                 continue;
3238                         cb->args[1] = 0;
3239                 }
3240                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3241                                             cb->nlh->nlmsg_seq,
3242                                             IPCTNL_MSG_EXP_NEW,
3243                                             exp) < 0) {
3244                         if (!refcount_inc_not_zero(&exp->use))
3245                                 continue;
3246                         cb->args[1] = (unsigned long)exp;
3247                         goto out;
3248                 }
3249         }
3250         if (cb->args[1]) {
3251                 cb->args[1] = 0;
3252                 goto restart;
3253         }
3254         cb->args[0] = 1;
3255 out:
3256         rcu_read_unlock();
3257         if (last)
3258                 nf_ct_expect_put(last);
3259
3260         return skb->len;
3261 }
3262
3263 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3264                                  struct sk_buff *skb,
3265                                  const struct nlmsghdr *nlh,
3266                                  const struct nlattr * const cda[],
3267                                  struct netlink_ext_ack *extack)
3268 {
3269         int err;
3270         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3271         u_int8_t u3 = nfmsg->nfgen_family;
3272         struct nf_conntrack_tuple tuple;
3273         struct nf_conntrack_tuple_hash *h;
3274         struct nf_conn *ct;
3275         struct nf_conntrack_zone zone;
3276         struct netlink_dump_control c = {
3277                 .dump = ctnetlink_exp_ct_dump_table,
3278                 .done = ctnetlink_exp_done,
3279         };
3280
3281         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3282                                     u3, NULL);
3283         if (err < 0)
3284                 return err;
3285
3286         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3287         if (err < 0)
3288                 return err;
3289
3290         h = nf_conntrack_find_get(net, &zone, &tuple);
3291         if (!h)
3292                 return -ENOENT;
3293
3294         ct = nf_ct_tuplehash_to_ctrack(h);
3295         /* No expectation linked to this connection tracking. */
3296         if (!nfct_help(ct)) {
3297                 nf_ct_put(ct);
3298                 return 0;
3299         }
3300
3301         c.data = ct;
3302
3303         err = netlink_dump_start(ctnl, skb, nlh, &c);
3304         nf_ct_put(ct);
3305
3306         return err;
3307 }
3308
3309 static int ctnetlink_get_expect(struct sk_buff *skb,
3310                                 const struct nfnl_info *info,
3311                                 const struct nlattr * const cda[])
3312 {
3313         u_int8_t u3 = info->nfmsg->nfgen_family;
3314         struct nf_conntrack_tuple tuple;
3315         struct nf_conntrack_expect *exp;
3316         struct nf_conntrack_zone zone;
3317         struct sk_buff *skb2;
3318         int err;
3319
3320         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3321                 if (cda[CTA_EXPECT_MASTER])
3322                         return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3323                                                      info->nlh, cda,
3324                                                      info->extack);
3325                 else {
3326                         struct netlink_dump_control c = {
3327                                 .dump = ctnetlink_exp_dump_table,
3328                                 .done = ctnetlink_exp_done,
3329                         };
3330                         return netlink_dump_start(info->sk, skb, info->nlh, &c);
3331                 }
3332         }
3333
3334         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3335         if (err < 0)
3336                 return err;
3337
3338         if (cda[CTA_EXPECT_TUPLE])
3339                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3340                                             u3, NULL);
3341         else if (cda[CTA_EXPECT_MASTER])
3342                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3343                                             u3, NULL);
3344         else
3345                 return -EINVAL;
3346
3347         if (err < 0)
3348                 return err;
3349
3350         exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3351         if (!exp)
3352                 return -ENOENT;
3353
3354         if (cda[CTA_EXPECT_ID]) {
3355                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3356
3357                 if (id != nf_expect_get_id(exp)) {
3358                         nf_ct_expect_put(exp);
3359                         return -ENOENT;
3360                 }
3361         }
3362
3363         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3364         if (!skb2) {
3365                 nf_ct_expect_put(exp);
3366                 return -ENOMEM;
3367         }
3368
3369         rcu_read_lock();
3370         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3371                                       info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3372                                       exp);
3373         rcu_read_unlock();
3374         nf_ct_expect_put(exp);
3375         if (err <= 0) {
3376                 kfree_skb(skb2);
3377                 return -ENOMEM;
3378         }
3379
3380         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3381 }
3382
3383 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3384 {
3385         const struct nf_conn_help *m_help;
3386         const char *name = data;
3387
3388         m_help = nfct_help(exp->master);
3389
3390         return strcmp(m_help->helper->name, name) == 0;
3391 }
3392
3393 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3394 {
3395         return true;
3396 }
3397
3398 static int ctnetlink_del_expect(struct sk_buff *skb,
3399                                 const struct nfnl_info *info,
3400                                 const struct nlattr * const cda[])
3401 {
3402         u_int8_t u3 = info->nfmsg->nfgen_family;
3403         struct nf_conntrack_expect *exp;
3404         struct nf_conntrack_tuple tuple;
3405         struct nf_conntrack_zone zone;
3406         int err;
3407
3408         if (cda[CTA_EXPECT_TUPLE]) {
3409                 /* delete a single expect by tuple */
3410                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3411                 if (err < 0)
3412                         return err;
3413
3414                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3415                                             u3, NULL);
3416                 if (err < 0)
3417                         return err;
3418
3419                 /* bump usage count to 2 */
3420                 exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3421                 if (!exp)
3422                         return -ENOENT;
3423
3424                 if (cda[CTA_EXPECT_ID]) {
3425                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3426                         if (ntohl(id) != (u32)(unsigned long)exp) {
3427                                 nf_ct_expect_put(exp);
3428                                 return -ENOENT;
3429                         }
3430                 }
3431
3432                 /* after list removal, usage count == 1 */
3433                 spin_lock_bh(&nf_conntrack_expect_lock);
3434                 if (del_timer(&exp->timeout)) {
3435                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3436                                                    nlmsg_report(info->nlh));
3437                         nf_ct_expect_put(exp);
3438                 }
3439                 spin_unlock_bh(&nf_conntrack_expect_lock);
3440                 /* have to put what we 'get' above.
3441                  * after this line usage count == 0 */
3442                 nf_ct_expect_put(exp);
3443         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3444                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3445
3446                 nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3447                                          NETLINK_CB(skb).portid,
3448                                          nlmsg_report(info->nlh));
3449         } else {
3450                 /* This basically means we have to flush everything*/
3451                 nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3452                                          NETLINK_CB(skb).portid,
3453                                          nlmsg_report(info->nlh));
3454         }
3455
3456         return 0;
3457 }
3458 static int
3459 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3460                         const struct nlattr * const cda[])
3461 {
3462         if (cda[CTA_EXPECT_TIMEOUT]) {
3463                 if (!del_timer(&x->timeout))
3464                         return -ETIME;
3465
3466                 x->timeout.expires = jiffies +
3467                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3468                 add_timer(&x->timeout);
3469         }
3470         return 0;
3471 }
3472
3473 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3474         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3475         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3476 };
3477
3478 static int
3479 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3480                            struct nf_conntrack_expect *exp,
3481                            u_int8_t u3)
3482 {
3483 #if IS_ENABLED(CONFIG_NF_NAT)
3484         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3485         struct nf_conntrack_tuple nat_tuple = {};
3486         int err;
3487
3488         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3489                                           exp_nat_nla_policy, NULL);
3490         if (err < 0)
3491                 return err;
3492
3493         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3494                 return -EINVAL;
3495
3496         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3497                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3498                                     u3, NULL);
3499         if (err < 0)
3500                 return err;
3501
3502         exp->saved_addr = nat_tuple.src.u3;
3503         exp->saved_proto = nat_tuple.src.u;
3504         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3505
3506         return 0;
3507 #else
3508         return -EOPNOTSUPP;
3509 #endif
3510 }
3511
3512 static struct nf_conntrack_expect *
3513 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3514                        struct nf_conntrack_helper *helper,
3515                        struct nf_conntrack_tuple *tuple,
3516                        struct nf_conntrack_tuple *mask)
3517 {
3518         u_int32_t class = 0;
3519         struct nf_conntrack_expect *exp;
3520         struct nf_conn_help *help;
3521         int err;
3522
3523         help = nfct_help(ct);
3524         if (!help)
3525                 return ERR_PTR(-EOPNOTSUPP);
3526
3527         if (cda[CTA_EXPECT_CLASS] && helper) {
3528                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3529                 if (class > helper->expect_class_max)
3530                         return ERR_PTR(-EINVAL);
3531         }
3532         exp = nf_ct_expect_alloc(ct);
3533         if (!exp)
3534                 return ERR_PTR(-ENOMEM);
3535
3536         if (cda[CTA_EXPECT_FLAGS]) {
3537                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3538                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3539         } else {
3540                 exp->flags = 0;
3541         }
3542         if (cda[CTA_EXPECT_FN]) {
3543                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3544                 struct nf_ct_helper_expectfn *expfn;
3545
3546                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3547                 if (expfn == NULL) {
3548                         err = -EINVAL;
3549                         goto err_out;
3550                 }
3551                 exp->expectfn = expfn->expectfn;
3552         } else
3553                 exp->expectfn = NULL;
3554
3555         exp->class = class;
3556         exp->master = ct;
3557         exp->helper = helper;
3558         exp->tuple = *tuple;
3559         exp->mask.src.u3 = mask->src.u3;
3560         exp->mask.src.u.all = mask->src.u.all;
3561
3562         if (cda[CTA_EXPECT_NAT]) {
3563                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3564                                                  exp, nf_ct_l3num(ct));
3565                 if (err < 0)
3566                         goto err_out;
3567         }
3568         return exp;
3569 err_out:
3570         nf_ct_expect_put(exp);
3571         return ERR_PTR(err);
3572 }
3573
3574 static int
3575 ctnetlink_create_expect(struct net *net,
3576                         const struct nf_conntrack_zone *zone,
3577                         const struct nlattr * const cda[],
3578                         u_int8_t u3, u32 portid, int report)
3579 {
3580         struct nf_conntrack_tuple tuple, mask, master_tuple;
3581         struct nf_conntrack_tuple_hash *h = NULL;
3582         struct nf_conntrack_helper *helper = NULL;
3583         struct nf_conntrack_expect *exp;
3584         struct nf_conn *ct;
3585         int err;
3586
3587         /* caller guarantees that those three CTA_EXPECT_* exist */
3588         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3589                                     u3, NULL);
3590         if (err < 0)
3591                 return err;
3592         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3593                                     u3, NULL);
3594         if (err < 0)
3595                 return err;
3596         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3597                                     u3, NULL);
3598         if (err < 0)
3599                 return err;
3600
3601         /* Look for master conntrack of this expectation */
3602         h = nf_conntrack_find_get(net, zone, &master_tuple);
3603         if (!h)
3604                 return -ENOENT;
3605         ct = nf_ct_tuplehash_to_ctrack(h);
3606
3607         rcu_read_lock();
3608         if (cda[CTA_EXPECT_HELP_NAME]) {
3609                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3610
3611                 helper = __nf_conntrack_helper_find(helpname, u3,
3612                                                     nf_ct_protonum(ct));
3613                 if (helper == NULL) {
3614                         rcu_read_unlock();
3615 #ifdef CONFIG_MODULES
3616                         if (request_module("nfct-helper-%s", helpname) < 0) {
3617                                 err = -EOPNOTSUPP;
3618                                 goto err_ct;
3619                         }
3620                         rcu_read_lock();
3621                         helper = __nf_conntrack_helper_find(helpname, u3,
3622                                                             nf_ct_protonum(ct));
3623                         if (helper) {
3624                                 err = -EAGAIN;
3625                                 goto err_rcu;
3626                         }
3627                         rcu_read_unlock();
3628 #endif
3629                         err = -EOPNOTSUPP;
3630                         goto err_ct;
3631                 }
3632         }
3633
3634         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3635         if (IS_ERR(exp)) {
3636                 err = PTR_ERR(exp);
3637                 goto err_rcu;
3638         }
3639
3640         err = nf_ct_expect_related_report(exp, portid, report, 0);
3641         nf_ct_expect_put(exp);
3642 err_rcu:
3643         rcu_read_unlock();
3644 err_ct:
3645         nf_ct_put(ct);
3646         return err;
3647 }
3648
3649 static int ctnetlink_new_expect(struct sk_buff *skb,
3650                                 const struct nfnl_info *info,
3651                                 const struct nlattr * const cda[])
3652 {
3653         u_int8_t u3 = info->nfmsg->nfgen_family;
3654         struct nf_conntrack_tuple tuple;
3655         struct nf_conntrack_expect *exp;
3656         struct nf_conntrack_zone zone;
3657         int err;
3658
3659         if (!cda[CTA_EXPECT_TUPLE]
3660             || !cda[CTA_EXPECT_MASK]
3661             || !cda[CTA_EXPECT_MASTER])
3662                 return -EINVAL;
3663
3664         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3665         if (err < 0)
3666                 return err;
3667
3668         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3669                                     u3, NULL);
3670         if (err < 0)
3671                 return err;
3672
3673         spin_lock_bh(&nf_conntrack_expect_lock);
3674         exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3675         if (!exp) {
3676                 spin_unlock_bh(&nf_conntrack_expect_lock);
3677                 err = -ENOENT;
3678                 if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3679                         err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3680                                                       NETLINK_CB(skb).portid,
3681                                                       nlmsg_report(info->nlh));
3682                 }
3683                 return err;
3684         }
3685
3686         err = -EEXIST;
3687         if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3688                 err = ctnetlink_change_expect(exp, cda);
3689         spin_unlock_bh(&nf_conntrack_expect_lock);
3690
3691         return err;
3692 }
3693
3694 static int
3695 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3696                              const struct ip_conntrack_stat *st)
3697 {
3698         struct nlmsghdr *nlh;
3699         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3700
3701         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3702                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3703         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3704                            NFNETLINK_V0, htons(cpu));
3705         if (!nlh)
3706                 goto nlmsg_failure;
3707
3708         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3709             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3710             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3711                 goto nla_put_failure;
3712
3713         nlmsg_end(skb, nlh);
3714         return skb->len;
3715
3716 nla_put_failure:
3717 nlmsg_failure:
3718         nlmsg_cancel(skb, nlh);
3719         return -1;
3720 }
3721
3722 static int
3723 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3724 {
3725         int cpu;
3726         struct net *net = sock_net(skb->sk);
3727
3728         if (cb->args[0] == nr_cpu_ids)
3729                 return 0;
3730
3731         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3732                 const struct ip_conntrack_stat *st;
3733
3734                 if (!cpu_possible(cpu))
3735                         continue;
3736
3737                 st = per_cpu_ptr(net->ct.stat, cpu);
3738                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3739                                                  cb->nlh->nlmsg_seq,
3740                                                  cpu, st) < 0)
3741                         break;
3742         }
3743         cb->args[0] = cpu;
3744
3745         return skb->len;
3746 }
3747
3748 static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3749                                   const struct nfnl_info *info,
3750                                   const struct nlattr * const cda[])
3751 {
3752         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3753                 struct netlink_dump_control c = {
3754                         .dump = ctnetlink_exp_stat_cpu_dump,
3755                 };
3756                 return netlink_dump_start(info->sk, skb, info->nlh, &c);
3757         }
3758
3759         return 0;
3760 }
3761
3762 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3763 static struct nf_ct_event_notifier ctnl_notifier = {
3764         .ct_event = ctnetlink_conntrack_event,
3765         .exp_event = ctnetlink_expect_event,
3766 };
3767 #endif
3768
3769 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3770         [IPCTNL_MSG_CT_NEW]     = {
3771                 .call           = ctnetlink_new_conntrack,
3772                 .type           = NFNL_CB_MUTEX,
3773                 .attr_count     = CTA_MAX,
3774                 .policy         = ct_nla_policy
3775         },
3776         [IPCTNL_MSG_CT_GET]     = {
3777                 .call           = ctnetlink_get_conntrack,
3778                 .type           = NFNL_CB_MUTEX,
3779                 .attr_count     = CTA_MAX,
3780                 .policy         = ct_nla_policy
3781         },
3782         [IPCTNL_MSG_CT_DELETE]  = {
3783                 .call           = ctnetlink_del_conntrack,
3784                 .type           = NFNL_CB_MUTEX,
3785                 .attr_count     = CTA_MAX,
3786                 .policy         = ct_nla_policy
3787         },
3788         [IPCTNL_MSG_CT_GET_CTRZERO] = {
3789                 .call           = ctnetlink_get_conntrack,
3790                 .type           = NFNL_CB_MUTEX,
3791                 .attr_count     = CTA_MAX,
3792                 .policy         = ct_nla_policy
3793         },
3794         [IPCTNL_MSG_CT_GET_STATS_CPU] = {
3795                 .call           = ctnetlink_stat_ct_cpu,
3796                 .type           = NFNL_CB_MUTEX,
3797         },
3798         [IPCTNL_MSG_CT_GET_STATS] = {
3799                 .call           = ctnetlink_stat_ct,
3800                 .type           = NFNL_CB_MUTEX,
3801         },
3802         [IPCTNL_MSG_CT_GET_DYING] = {
3803                 .call           = ctnetlink_get_ct_dying,
3804                 .type           = NFNL_CB_MUTEX,
3805         },
3806         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = {
3807                 .call           = ctnetlink_get_ct_unconfirmed,
3808                 .type           = NFNL_CB_MUTEX,
3809         },
3810 };
3811
3812 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3813         [IPCTNL_MSG_EXP_GET] = {
3814                 .call           = ctnetlink_get_expect,
3815                 .type           = NFNL_CB_MUTEX,
3816                 .attr_count     = CTA_EXPECT_MAX,
3817                 .policy         = exp_nla_policy
3818         },
3819         [IPCTNL_MSG_EXP_NEW] = {
3820                 .call           = ctnetlink_new_expect,
3821                 .type           = NFNL_CB_MUTEX,
3822                 .attr_count     = CTA_EXPECT_MAX,
3823                 .policy         = exp_nla_policy
3824         },
3825         [IPCTNL_MSG_EXP_DELETE] = {
3826                 .call           = ctnetlink_del_expect,
3827                 .type           = NFNL_CB_MUTEX,
3828                 .attr_count     = CTA_EXPECT_MAX,
3829                 .policy         = exp_nla_policy
3830         },
3831         [IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3832                 .call           = ctnetlink_stat_exp_cpu,
3833                 .type           = NFNL_CB_MUTEX,
3834         },
3835 };
3836
3837 static const struct nfnetlink_subsystem ctnl_subsys = {
3838         .name                           = "conntrack",
3839         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3840         .cb_count                       = IPCTNL_MSG_MAX,
3841         .cb                             = ctnl_cb,
3842 };
3843
3844 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3845         .name                           = "conntrack_expect",
3846         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3847         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3848         .cb                             = ctnl_exp_cb,
3849 };
3850
3851 MODULE_ALIAS("ip_conntrack_netlink");
3852 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3853 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3854
3855 static int __net_init ctnetlink_net_init(struct net *net)
3856 {
3857 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3858         nf_conntrack_register_notifier(net, &ctnl_notifier);
3859 #endif
3860         return 0;
3861 }
3862
3863 static void ctnetlink_net_pre_exit(struct net *net)
3864 {
3865 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3866         nf_conntrack_unregister_notifier(net);
3867 #endif
3868 }
3869
3870 static struct pernet_operations ctnetlink_net_ops = {
3871         .init           = ctnetlink_net_init,
3872         .pre_exit       = ctnetlink_net_pre_exit,
3873 };
3874
3875 static int __init ctnetlink_init(void)
3876 {
3877         int ret;
3878
3879         ret = nfnetlink_subsys_register(&ctnl_subsys);
3880         if (ret < 0) {
3881                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3882                 goto err_out;
3883         }
3884
3885         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3886         if (ret < 0) {
3887                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3888                 goto err_unreg_subsys;
3889         }
3890
3891         ret = register_pernet_subsys(&ctnetlink_net_ops);
3892         if (ret < 0) {
3893                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3894                 goto err_unreg_exp_subsys;
3895         }
3896 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3897         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3898         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3899 #endif
3900         return 0;
3901
3902 err_unreg_exp_subsys:
3903         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3904 err_unreg_subsys:
3905         nfnetlink_subsys_unregister(&ctnl_subsys);
3906 err_out:
3907         return ret;
3908 }
3909
3910 static void __exit ctnetlink_exit(void)
3911 {
3912         unregister_pernet_subsys(&ctnetlink_net_ops);
3913         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3914         nfnetlink_subsys_unregister(&ctnl_subsys);
3915 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3916         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3917 #endif
3918         synchronize_rcu();
3919 }
3920
3921 module_init(ctnetlink_init);
3922 module_exit(ctnetlink_exit);