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 {
 172         long timeout = nf_ct_expires(ct) / HZ;
 173
 174         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 175                 goto nla_put_failure;
 176         return 0;
 177
 178 nla_put_failure:
 179         return -1;
 180 }
 181
 182 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 183 {
 184         const struct nf_conntrack_l4proto *l4proto;
 185         struct nlattr *nest_proto;
 186         int ret;
 187
 188         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 189         if (!l4proto->to_nlattr)
 190                 return 0;
 191
 192         nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 193         if (!nest_proto)
 194                 goto nla_put_failure;
 195
 196         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 197
 198         nla_nest_end(skb, nest_proto);
 199
 200         return ret;
 201
 202 nla_put_failure:
 203         return -1;
 204 }
 205
 206 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 207                                    const struct nf_conn *ct)
 208 {
 209         struct nlattr *nest_helper;
 210         const struct nf_conn_help *help = nfct_help(ct);
 211         struct nf_conntrack_helper *helper;
 212
 213         if (!help)
 214                 return 0;
 215
 216         helper = rcu_dereference(help->helper);
 217         if (!helper)
 218                 goto out;
 219
 220         nest_helper = nla_nest_start(skb, CTA_HELP);
 221         if (!nest_helper)
 222                 goto nla_put_failure;
 223         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 224                 goto nla_put_failure;
 225
 226         if (helper->to_nlattr)
 227                 helper->to_nlattr(skb, ct);
 228
 229         nla_nest_end(skb, nest_helper);
 230 out:
 231         return 0;
 232
 233 nla_put_failure:
 234         return -1;
 235 }
 236
 237 static int
 238 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 239               enum ip_conntrack_dir dir, int type)
 240 {
 241         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 242         struct nf_conn_counter *counter = acct->counter;
 243         struct nlattr *nest_count;
 244         u64 pkts, bytes;
 245
 246         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 247                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 248                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 249         } else {
 250                 pkts = atomic64_read(&counter[dir].packets);
 251                 bytes = atomic64_read(&counter[dir].bytes);
 252         }
 253
 254         nest_count = nla_nest_start(skb, attr);
 255         if (!nest_count)
 256                 goto nla_put_failure;
 257
 258         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 259                          CTA_COUNTERS_PAD) ||
 260             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 261                          CTA_COUNTERS_PAD))
 262                 goto nla_put_failure;
 263
 264         nla_nest_end(skb, nest_count);
 265
 266         return 0;
 267
 268 nla_put_failure:
 269         return -1;
 270 }
 271
 272 static int
 273 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 274 {
 275         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 276
 277         if (!acct)
 278                 return 0;
 279
 280         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 281                 return -1;
 282         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 283                 return -1;
 284
 285         return 0;
 286 }
 287
 288 static int
 289 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 290 {
 291         struct nlattr *nest_count;
 292         const struct nf_conn_tstamp *tstamp;
 293
 294         tstamp = nf_conn_tstamp_find(ct);
 295         if (!tstamp)
 296                 return 0;
 297
 298         nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 299         if (!nest_count)
 300                 goto nla_put_failure;
 301
 302         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 303                          CTA_TIMESTAMP_PAD) ||
 304             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 305                                                cpu_to_be64(tstamp->stop),
 306                                                CTA_TIMESTAMP_PAD)))
 307                 goto nla_put_failure;
 308         nla_nest_end(skb, nest_count);
 309
 310         return 0;
 311
 312 nla_put_failure:
 313         return -1;
 314 }
 315
 316 #ifdef CONFIG_NF_CONNTRACK_MARK
 317 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 318 {
 319         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 320                 goto nla_put_failure;
 321         return 0;
 322
 323 nla_put_failure:
 324         return -1;
 325 }
 326 #else
 327 #define ctnetlink_dump_mark(a, b) (0)
 328 #endif
 329
 330 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 331 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 332 {
 333         struct nlattr *nest_secctx;
 334         int len, ret;
 335         char *secctx;
 336
 337         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 338         if (ret)
 339                 return 0;
 340
 341         ret = -1;
 342         nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 343         if (!nest_secctx)
 344                 goto nla_put_failure;
 345
 346         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 347                 goto nla_put_failure;
 348         nla_nest_end(skb, nest_secctx);
 349
 350         ret = 0;
 351 nla_put_failure:
 352         security_release_secctx(secctx, len);
 353         return ret;
 354 }
 355 #else
 356 #define ctnetlink_dump_secctx(a, b) (0)
 357 #endif
 358
 359 #ifdef CONFIG_NF_CONNTRACK_LABELS
 360 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 361 {
 362         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 363
 364         if (!labels)
 365                 return 0;
 366         return nla_total_size(sizeof(labels->bits));
 367 }
 368
 369 static int
 370 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 371 {
 372         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 373         unsigned int i;
 374
 375         if (!labels)
 376                 return 0;
 377
 378         i = 0;
 379         do {
 380                 if (labels->bits[i] != 0)
 381                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 382                                        labels->bits);
 383                 i++;
 384         } while (i < ARRAY_SIZE(labels->bits));
 385
 386         return 0;
 387 }
 388 #else
 389 #define ctnetlink_dump_labels(a, b) (0)
 390 #define ctnetlink_label_size(a) (0)
 391 #endif
 392
 393 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 394
 395 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 396 {
 397         struct nlattr *nest_parms;
 398
 399         if (!(ct->status & IPS_EXPECTED))
 400                 return 0;
 401
 402         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 403         if (!nest_parms)
 404                 goto nla_put_failure;
 405         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 406                 goto nla_put_failure;
 407         nla_nest_end(skb, nest_parms);
 408
 409         return 0;
 410
 411 nla_put_failure:
 412         return -1;
 413 }
 414
 415 static int
 416 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 417 {
 418         struct nlattr *nest_parms;
 419
 420         nest_parms = nla_nest_start(skb, type);
 421         if (!nest_parms)
 422                 goto nla_put_failure;
 423
 424         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 425                          htonl(seq->correction_pos)) ||
 426             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 427                          htonl(seq->offset_before)) ||
 428             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 429                          htonl(seq->offset_after)))
 430                 goto nla_put_failure;
 431
 432         nla_nest_end(skb, nest_parms);
 433
 434         return 0;
 435
 436 nla_put_failure:
 437         return -1;
 438 }
 439
 440 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 441 {
 442         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 443         struct nf_ct_seqadj *seq;
 444
 445         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 446                 return 0;
 447
 448         spin_lock_bh(&ct->lock);
 449         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 450         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 451                 goto err;
 452
 453         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 454         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 455                 goto err;
 456
 457         spin_unlock_bh(&ct->lock);
 458         return 0;
 459 err:
 460         spin_unlock_bh(&ct->lock);
 461         return -1;
 462 }
 463
 464 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 465 {
 466         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 467         struct nlattr *nest_parms;
 468
 469         if (!synproxy)
 470                 return 0;
 471
 472         nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 473         if (!nest_parms)
 474                 goto nla_put_failure;
 475
 476         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 477             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 478             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 479                 goto nla_put_failure;
 480
 481         nla_nest_end(skb, nest_parms);
 482
 483         return 0;
 484
 485 nla_put_failure:
 486         return -1;
 487 }
 488
 489 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 490 {
 491         __be32 id = (__force __be32)nf_ct_get_id(ct);
 492
 493         if (nla_put_be32(skb, CTA_ID, id))
 494                 goto nla_put_failure;
 495         return 0;
 496
 497 nla_put_failure:
 498         return -1;
 499 }
 500
 501 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 502 {
 503         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 504                 goto nla_put_failure;
 505         return 0;
 506
 507 nla_put_failure:
 508         return -1;
 509 }
 510
 511 /* all these functions access ct->ext. Caller must either hold a reference
 512  * on ct or prevent its deletion by holding either the bucket spinlock or
 513  * pcpu dying list lock.
 514  */
 515 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 516                                   struct nf_conn *ct, u32 type)
 517 {
 518         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 519             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 520             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 521             ctnetlink_dump_labels(skb, ct) < 0 ||
 522             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 523             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 524                 return -1;
 525
 526         return 0;
 527 }
 528
 529 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 530 {
 531         if (ctnetlink_dump_status(skb, ct) < 0 ||
 532             ctnetlink_dump_mark(skb, ct) < 0 ||
 533             ctnetlink_dump_secctx(skb, ct) < 0 ||
 534             ctnetlink_dump_id(skb, ct) < 0 ||
 535             ctnetlink_dump_use(skb, ct) < 0 ||
 536             ctnetlink_dump_master(skb, ct) < 0)
 537                 return -1;
 538
 539         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 540             (ctnetlink_dump_timeout(skb, ct) < 0 ||
 541              ctnetlink_dump_protoinfo(skb, ct) < 0))
 542                 return -1;
 543
 544         return 0;
 545 }
 546
 547 static int
 548 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 549                     struct nf_conn *ct, bool extinfo, unsigned int flags)
 550 {
 551         const struct nf_conntrack_zone *zone;
 552         struct nlmsghdr *nlh;
 553         struct nfgenmsg *nfmsg;
 554         struct nlattr *nest_parms;
 555         unsigned int event;
 556
 557         if (portid)
 558                 flags |= NLM_F_MULTI;
 559         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 560         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 561         if (nlh == NULL)
 562                 goto nlmsg_failure;
 563
 564         nfmsg = nlmsg_data(nlh);
 565         nfmsg->nfgen_family = nf_ct_l3num(ct);
 566         nfmsg->version      = NFNETLINK_V0;
 567         nfmsg->res_id       = 0;
 568
 569         zone = nf_ct_zone(ct);
 570
 571         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 572         if (!nest_parms)
 573                 goto nla_put_failure;
 574         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 575                 goto nla_put_failure;
 576         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 577                                    NF_CT_ZONE_DIR_ORIG) < 0)
 578                 goto nla_put_failure;
 579         nla_nest_end(skb, nest_parms);
 580
 581         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 582         if (!nest_parms)
 583                 goto nla_put_failure;
 584         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 585                 goto nla_put_failure;
 586         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 587                                    NF_CT_ZONE_DIR_REPL) < 0)
 588                 goto nla_put_failure;
 589         nla_nest_end(skb, nest_parms);
 590
 591         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 592                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 593                 goto nla_put_failure;
 594
 595         if (ctnetlink_dump_info(skb, ct) < 0)
 596                 goto nla_put_failure;
 597         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 598                 goto nla_put_failure;
 599
 600         nlmsg_end(skb, nlh);
 601         return skb->len;
 602
 603 nlmsg_failure:
 604 nla_put_failure:
 605         nlmsg_cancel(skb, nlh);
 606         return -1;
 607 }
 608
 609 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 610         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 611         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 612         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 613         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 614 };
 615
 616 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 617 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 618 {
 619         const struct nf_conntrack_l4proto *l4proto;
 620         size_t len, len4 = 0;
 621
 622         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 623         len *= 3u; /* ORIG, REPLY, MASTER */
 624
 625         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 626         len += l4proto->nlattr_size;
 627         if (l4proto->nlattr_tuple_size) {
 628                 len4 = l4proto->nlattr_tuple_size();
 629                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 630         }
 631
 632         return len + len4;
 633 }
 634 #endif
 635
 636 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 637 {
 638         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 639                 return 0;
 640         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 641                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 642                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 643                ;
 644 }
 645
 646 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 647 {
 648 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 649         int len, ret;
 650
 651         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 652         if (ret)
 653                 return 0;
 654
 655         return nla_total_size(0) /* CTA_SECCTX */
 656                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 657 #else
 658         return 0;
 659 #endif
 660 }
 661
 662 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 663 {
 664 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 665         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 666                 return 0;
 667         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 668 #else
 669         return 0;
 670 #endif
 671 }
 672
 673 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 674 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 675 {
 676         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 677                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 678                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 679                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 680                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 681                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 682                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 683                + ctnetlink_acct_size(ct)
 684                + ctnetlink_timestamp_size(ct)
 685                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 686                + nla_total_size(0) /* CTA_PROTOINFO */
 687                + nla_total_size(0) /* CTA_HELP */
 688                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 689                + ctnetlink_secctx_size(ct)
 690 #if IS_ENABLED(CONFIG_NF_NAT)
 691                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 692                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 693 #endif
 694 #ifdef CONFIG_NF_CONNTRACK_MARK
 695                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 696 #endif
 697 #ifdef CONFIG_NF_CONNTRACK_ZONES
 698                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 699 #endif
 700                + ctnetlink_proto_size(ct)
 701                + ctnetlink_label_size(ct)
 702                ;
 703 }
 704
 705 static int
 706 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 707 {
 708         const struct nf_conntrack_zone *zone;
 709         struct net *net;
 710         struct nlmsghdr *nlh;
 711         struct nfgenmsg *nfmsg;
 712         struct nlattr *nest_parms;
 713         struct nf_conn *ct = item->ct;
 714         struct sk_buff *skb;
 715         unsigned int type;
 716         unsigned int flags = 0, group;
 717         int err;
 718
 719         if (events & (1 << IPCT_DESTROY)) {
 720                 type = IPCTNL_MSG_CT_DELETE;
 721                 group = NFNLGRP_CONNTRACK_DESTROY;
 722         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 723                 type = IPCTNL_MSG_CT_NEW;
 724                 flags = NLM_F_CREATE|NLM_F_EXCL;
 725                 group = NFNLGRP_CONNTRACK_NEW;
 726         } else if (events) {
 727                 type = IPCTNL_MSG_CT_NEW;
 728                 group = NFNLGRP_CONNTRACK_UPDATE;
 729         } else
 730                 return 0;
 731
 732         net = nf_ct_net(ct);
 733         if (!item->report && !nfnetlink_has_listeners(net, group))
 734                 return 0;
 735
 736         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 737         if (skb == NULL)
 738                 goto errout;
 739
 740         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 741         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 742         if (nlh == NULL)
 743                 goto nlmsg_failure;
 744
 745         nfmsg = nlmsg_data(nlh);
 746         nfmsg->nfgen_family = nf_ct_l3num(ct);
 747         nfmsg->version  = NFNETLINK_V0;
 748         nfmsg->res_id   = 0;
 749
 750         zone = nf_ct_zone(ct);
 751
 752         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 753         if (!nest_parms)
 754                 goto nla_put_failure;
 755         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 756                 goto nla_put_failure;
 757         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 758                                    NF_CT_ZONE_DIR_ORIG) < 0)
 759                 goto nla_put_failure;
 760         nla_nest_end(skb, nest_parms);
 761
 762         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 763         if (!nest_parms)
 764                 goto nla_put_failure;
 765         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 766                 goto nla_put_failure;
 767         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 768                                    NF_CT_ZONE_DIR_REPL) < 0)
 769                 goto nla_put_failure;
 770         nla_nest_end(skb, nest_parms);
 771
 772         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 773                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 774                 goto nla_put_failure;
 775
 776         if (ctnetlink_dump_id(skb, ct) < 0)
 777                 goto nla_put_failure;
 778
 779         if (ctnetlink_dump_status(skb, ct) < 0)
 780                 goto nla_put_failure;
 781
 782         if (events & (1 << IPCT_DESTROY)) {
 783                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 784                     ctnetlink_dump_timestamp(skb, ct) < 0)
 785                         goto nla_put_failure;
 786         } else {
 787                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 788                         goto nla_put_failure;
 789
 790                 if (events & (1 << IPCT_PROTOINFO)
 791                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 792                         goto nla_put_failure;
 793
 794                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 795                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 796                         goto nla_put_failure;
 797
 798 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 799                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 800                     && ctnetlink_dump_secctx(skb, ct) < 0)
 801                         goto nla_put_failure;
 802 #endif
 803                 if (events & (1 << IPCT_LABEL) &&
 804                      ctnetlink_dump_labels(skb, ct) < 0)
 805                         goto nla_put_failure;
 806
 807                 if (events & (1 << IPCT_RELATED) &&
 808                     ctnetlink_dump_master(skb, ct) < 0)
 809                         goto nla_put_failure;
 810
 811                 if (events & (1 << IPCT_SEQADJ) &&
 812                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 813                         goto nla_put_failure;
 814
 815                 if (events & (1 << IPCT_SYNPROXY) &&
 816                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 817                         goto nla_put_failure;
 818         }
 819
 820 #ifdef CONFIG_NF_CONNTRACK_MARK
 821         if ((events & (1 << IPCT_MARK) || ct->mark)
 822             && ctnetlink_dump_mark(skb, ct) < 0)
 823                 goto nla_put_failure;
 824 #endif
 825         nlmsg_end(skb, nlh);
 826         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 827                              GFP_ATOMIC);
 828         if (err == -ENOBUFS || err == -EAGAIN)
 829                 return -ENOBUFS;
 830
 831         return 0;
 832
 833 nla_put_failure:
 834         nlmsg_cancel(skb, nlh);
 835 nlmsg_failure:
 836         kfree_skb(skb);
 837 errout:
 838         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 839                 return -ENOBUFS;
 840
 841         return 0;
 842 }
 843 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 844
 845 static int ctnetlink_done(struct netlink_callback *cb)
 846 {
 847         if (cb->args[1])
 848                 nf_ct_put((struct nf_conn *)cb->args[1]);
 849         kfree(cb->data);
 850         return 0;
 851 }
 852
 853 struct ctnetlink_filter {
 854         u8 family;
 855
 856         u_int32_t orig_flags;
 857         u_int32_t reply_flags;
 858
 859         struct nf_conntrack_tuple orig;
 860         struct nf_conntrack_tuple reply;
 861         struct nf_conntrack_zone zone;
 862
 863         struct {
 864                 u_int32_t val;
 865                 u_int32_t mask;
 866         } mark;
 867 };
 868
 869 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 870         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 871         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 872 };
 873
 874 static int ctnetlink_parse_filter(const struct nlattr *attr,
 875                                   struct ctnetlink_filter *filter)
 876 {
 877         struct nlattr *tb[CTA_FILTER_MAX + 1];
 878         int ret = 0;
 879
 880         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 881                                NULL);
 882         if (ret)
 883                 return ret;
 884
 885         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 886                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 887                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 888                         return -EOPNOTSUPP;
 889         }
 890
 891         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 892                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 893                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 894                         return -EOPNOTSUPP;
 895         }
 896
 897         return 0;
 898 }
 899
 900 static int ctnetlink_parse_zone(const struct nlattr *attr,
 901                                 struct nf_conntrack_zone *zone);
 902 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 903                                          struct nf_conntrack_tuple *tuple,
 904                                          u32 type, u_int8_t l3num,
 905                                          struct nf_conntrack_zone *zone,
 906                                          u_int32_t flags);
 907
 908 static struct ctnetlink_filter *
 909 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 910 {
 911         struct ctnetlink_filter *filter;
 912         int err;
 913
 914 #ifndef CONFIG_NF_CONNTRACK_MARK
 915         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 916                 return ERR_PTR(-EOPNOTSUPP);
 917 #endif
 918
 919         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 920         if (filter == NULL)
 921                 return ERR_PTR(-ENOMEM);
 922
 923         filter->family = family;
 924
 925 #ifdef CONFIG_NF_CONNTRACK_MARK
 926         if (cda[CTA_MARK]) {
 927                 filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 928                 if (cda[CTA_MARK_MASK])
 929                         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 930                 else
 931                         filter->mark.mask = 0xffffffff;
 932         } else if (cda[CTA_MARK_MASK]) {
 933                 err = -EINVAL;
 934                 goto err_filter;
 935         }
 936 #endif
 937         if (!cda[CTA_FILTER])
 938                 return filter;
 939
 940         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 941         if (err < 0)
 942                 goto err_filter;
 943
 944         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
 945         if (err < 0)
 946                 goto err_filter;
 947
 948         if (filter->orig_flags) {
 949                 if (!cda[CTA_TUPLE_ORIG]) {
 950                         err = -EINVAL;
 951                         goto err_filter;
 952                 }
 953
 954                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
 955                                                    CTA_TUPLE_ORIG,
 956                                                    filter->family,
 957                                                    &filter->zone,
 958                                                    filter->orig_flags);
 959                 if (err < 0)
 960                         goto err_filter;
 961         }
 962
 963         if (filter->reply_flags) {
 964                 if (!cda[CTA_TUPLE_REPLY]) {
 965                         err = -EINVAL;
 966                         goto err_filter;
 967                 }
 968
 969                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
 970                                                    CTA_TUPLE_REPLY,
 971                                                    filter->family,
 972                                                    &filter->zone,
 973                                                    filter->orig_flags);
 974                 if (err < 0) {
 975                         err = -EINVAL;
 976                         goto err_filter;
 977                 }
 978         }
 979
 980         return filter;
 981
 982 err_filter:
 983         kfree(filter);
 984
 985         return ERR_PTR(err);
 986 }
 987
 988 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
 989 {
 990         return family || cda[CTA_MARK] || cda[CTA_FILTER];
 991 }
 992
 993 static int ctnetlink_start(struct netlink_callback *cb)
 994 {
 995         const struct nlattr * const *cda = cb->data;
 996         struct ctnetlink_filter *filter = NULL;
 997         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 998         u8 family = nfmsg->nfgen_family;
 999
1000         if (ctnetlink_needs_filter(family, cda)) {
1001                 filter = ctnetlink_alloc_filter(cda, family);
1002                 if (IS_ERR(filter))
1003                         return PTR_ERR(filter);
1004         }
1005
1006         cb->data = filter;
1007         return 0;
1008 }
1009
1010 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1011                                         struct nf_conntrack_tuple *ct_tuple,
1012                                         u_int32_t flags, int family)
1013 {
1014         switch (family) {
1015         case NFPROTO_IPV4:
1016                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1017                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1018                         return  0;
1019
1020                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1021                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1022                         return  0;
1023                 break;
1024         case NFPROTO_IPV6:
1025                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1026                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1027                                    &ct_tuple->src.u3.in6))
1028                         return 0;
1029
1030                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1031                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1032                                    &ct_tuple->dst.u3.in6))
1033                         return 0;
1034                 break;
1035         }
1036
1037         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1038             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1039                 return 0;
1040
1041         switch (ct_tuple->dst.protonum) {
1042         case IPPROTO_TCP:
1043         case IPPROTO_UDP:
1044                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1045                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1046                         return 0;
1047
1048                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1049                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1050                         return 0;
1051                 break;
1052         case IPPROTO_ICMP:
1053                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1054                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1055                         return 0;
1056                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1057                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1058                         return 0;
1059                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1060                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1061                         return 0;
1062                 break;
1063         case IPPROTO_ICMPV6:
1064                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1065                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1066                         return 0;
1067                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1068                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1069                         return 0;
1070                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1071                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1072                         return 0;
1073                 break;
1074         }
1075
1076         return 1;
1077 }
1078
1079 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1080 {
1081         struct ctnetlink_filter *filter = data;
1082         struct nf_conntrack_tuple *tuple;
1083
1084         if (filter == NULL)
1085                 goto out;
1086
1087         /* Match entries of a given L3 protocol number.
1088          * If it is not specified, ie. l3proto == 0,
1089          * then match everything.
1090          */
1091         if (filter->family && nf_ct_l3num(ct) != filter->family)
1092                 goto ignore_entry;
1093
1094         if (filter->orig_flags) {
1095                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1096                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1097                                                   filter->orig_flags,
1098                                                   filter->family))
1099                         goto ignore_entry;
1100         }
1101
1102         if (filter->reply_flags) {
1103                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1104                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1105                                                   filter->reply_flags,
1106                                                   filter->family))
1107                         goto ignore_entry;
1108         }
1109
1110 #ifdef CONFIG_NF_CONNTRACK_MARK
1111         if ((ct->mark & filter->mark.mask) != filter->mark.val)
1112                 goto ignore_entry;
1113 #endif
1114
1115 out:
1116         return 1;
1117
1118 ignore_entry:
1119         return 0;
1120 }
1121
1122 static int
1123 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1124 {
1125         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1126         struct net *net = sock_net(skb->sk);
1127         struct nf_conn *ct, *last;
1128         struct nf_conntrack_tuple_hash *h;
1129         struct hlist_nulls_node *n;
1130         struct nf_conn *nf_ct_evict[8];
1131         int res, i;
1132         spinlock_t *lockp;
1133
1134         last = (struct nf_conn *)cb->args[1];
1135         i = 0;
1136
1137         local_bh_disable();
1138         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1139 restart:
1140                 while (i) {
1141                         i--;
1142                         if (nf_ct_should_gc(nf_ct_evict[i]))
1143                                 nf_ct_kill(nf_ct_evict[i]);
1144                         nf_ct_put(nf_ct_evict[i]);
1145                 }
1146
1147                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1148                 nf_conntrack_lock(lockp);
1149                 if (cb->args[0] >= nf_conntrack_htable_size) {
1150                         spin_unlock(lockp);
1151                         goto out;
1152                 }
1153                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1154                                            hnnode) {
1155                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1156                                 continue;
1157                         ct = nf_ct_tuplehash_to_ctrack(h);
1158                         if (nf_ct_is_expired(ct)) {
1159                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1160                                     atomic_inc_not_zero(&ct->ct_general.use))
1161                                         nf_ct_evict[i++] = ct;
1162                                 continue;
1163                         }
1164
1165                         if (!net_eq(net, nf_ct_net(ct)))
1166                                 continue;
1167
1168                         if (cb->args[1]) {
1169                                 if (ct != last)
1170                                         continue;
1171                                 cb->args[1] = 0;
1172                         }
1173                         if (!ctnetlink_filter_match(ct, cb->data))
1174                                 continue;
1175
1176                         res =
1177                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1178                                             cb->nlh->nlmsg_seq,
1179                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1180                                             ct, true, flags);
1181                         if (res < 0) {
1182                                 nf_conntrack_get(&ct->ct_general);
1183                                 cb->args[1] = (unsigned long)ct;
1184                                 spin_unlock(lockp);
1185                                 goto out;
1186                         }
1187                 }
1188                 spin_unlock(lockp);
1189                 if (cb->args[1]) {
1190                         cb->args[1] = 0;
1191                         goto restart;
1192                 }
1193         }
1194 out:
1195         local_bh_enable();
1196         if (last) {
1197                 /* nf ct hash resize happened, now clear the leftover. */
1198                 if ((struct nf_conn *)cb->args[1] == last)
1199                         cb->args[1] = 0;
1200
1201                 nf_ct_put(last);
1202         }
1203
1204         while (i) {
1205                 i--;
1206                 if (nf_ct_should_gc(nf_ct_evict[i]))
1207                         nf_ct_kill(nf_ct_evict[i]);
1208                 nf_ct_put(nf_ct_evict[i]);
1209         }
1210
1211         return skb->len;
1212 }
1213
1214 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1215                                 struct nf_conntrack_tuple *t,
1216                                 u_int32_t flags)
1217 {
1218         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1219                 if (!tb[CTA_IP_V4_SRC])
1220                         return -EINVAL;
1221
1222                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1223         }
1224
1225         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1226                 if (!tb[CTA_IP_V4_DST])
1227                         return -EINVAL;
1228
1229                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1230         }
1231
1232         return 0;
1233 }
1234
1235 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1236                                 struct nf_conntrack_tuple *t,
1237                                 u_int32_t flags)
1238 {
1239         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1240                 if (!tb[CTA_IP_V6_SRC])
1241                         return -EINVAL;
1242
1243                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1244         }
1245
1246         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1247                 if (!tb[CTA_IP_V6_DST])
1248                         return -EINVAL;
1249
1250                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1251         }
1252
1253         return 0;
1254 }
1255
1256 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1257                                     struct nf_conntrack_tuple *tuple,
1258                                     u_int32_t flags)
1259 {
1260         struct nlattr *tb[CTA_IP_MAX+1];
1261         int ret = 0;
1262
1263         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1264         if (ret < 0)
1265                 return ret;
1266
1267         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1268                                              cta_ip_nla_policy, NULL);
1269         if (ret)
1270                 return ret;
1271
1272         switch (tuple->src.l3num) {
1273         case NFPROTO_IPV4:
1274                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1275                 break;
1276         case NFPROTO_IPV6:
1277                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1278                 break;
1279         }
1280
1281         return ret;
1282 }
1283
1284 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1285         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1286 };
1287
1288 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1289                                        struct nf_conntrack_tuple *tuple,
1290                                        u_int32_t flags)
1291 {
1292         const struct nf_conntrack_l4proto *l4proto;
1293         struct nlattr *tb[CTA_PROTO_MAX+1];
1294         int ret = 0;
1295
1296         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1297                                           proto_nla_policy, NULL);
1298         if (ret < 0)
1299                 return ret;
1300
1301         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1302                 return 0;
1303
1304         if (!tb[CTA_PROTO_NUM])
1305                 return -EINVAL;
1306
1307         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1308
1309         rcu_read_lock();
1310         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1311
1312         if (likely(l4proto->nlattr_to_tuple)) {
1313                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1314                                                      l4proto->nla_policy,
1315                                                      NULL);
1316                 if (ret == 0)
1317                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1318         }
1319
1320         rcu_read_unlock();
1321
1322         return ret;
1323 }
1324
1325 static int
1326 ctnetlink_parse_zone(const struct nlattr *attr,
1327                      struct nf_conntrack_zone *zone)
1328 {
1329         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1330                         NF_CT_DEFAULT_ZONE_DIR, 0);
1331 #ifdef CONFIG_NF_CONNTRACK_ZONES
1332         if (attr)
1333                 zone->id = ntohs(nla_get_be16(attr));
1334 #else
1335         if (attr)
1336                 return -EOPNOTSUPP;
1337 #endif
1338         return 0;
1339 }
1340
1341 static int
1342 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1343                            struct nf_conntrack_zone *zone)
1344 {
1345         int ret;
1346
1347         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1348                 return -EINVAL;
1349
1350         ret = ctnetlink_parse_zone(attr, zone);
1351         if (ret < 0)
1352                 return ret;
1353
1354         if (type == CTA_TUPLE_REPLY)
1355                 zone->dir = NF_CT_ZONE_DIR_REPL;
1356         else
1357                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1358
1359         return 0;
1360 }
1361
1362 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1363         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1364         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1365         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1366 };
1367
1368 #define CTA_FILTER_F_ALL_CTA_PROTO \
1369   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1370    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1371    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1372    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1373    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1374    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1375    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1376    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1377
1378 static int
1379 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1380                               struct nf_conntrack_tuple *tuple, u32 type,
1381                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1382                               u_int32_t flags)
1383 {
1384         struct nlattr *tb[CTA_TUPLE_MAX+1];
1385         int err;
1386
1387         memset(tuple, 0, sizeof(*tuple));
1388
1389         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1390                                           tuple_nla_policy, NULL);
1391         if (err < 0)
1392                 return err;
1393
1394         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1395                 return -EOPNOTSUPP;
1396         tuple->src.l3num = l3num;
1397
1398         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1399             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1400                 if (!tb[CTA_TUPLE_IP])
1401                         return -EINVAL;
1402
1403                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1404                 if (err < 0)
1405                         return err;
1406         }
1407
1408         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1409                 if (!tb[CTA_TUPLE_PROTO])
1410                         return -EINVAL;
1411
1412                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1413                 if (err < 0)
1414                         return err;
1415         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1416                 /* Can't manage proto flags without a protonum  */
1417                 return -EINVAL;
1418         }
1419
1420         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1421                 if (!zone)
1422                         return -EINVAL;
1423
1424                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1425                                                  type, zone);
1426                 if (err < 0)
1427                         return err;
1428         }
1429
1430         /* orig and expect tuples get DIR_ORIGINAL */
1431         if (type == CTA_TUPLE_REPLY)
1432                 tuple->dst.dir = IP_CT_DIR_REPLY;
1433         else
1434                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1435
1436         return 0;
1437 }
1438
1439 static int
1440 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1441                       struct nf_conntrack_tuple *tuple, u32 type,
1442                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1443 {
1444         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1445                                             CTA_FILTER_FLAG(ALL));
1446 }
1447
1448 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1449         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1450                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1451 };
1452
1453 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1454                                 struct nlattr **helpinfo)
1455 {
1456         int err;
1457         struct nlattr *tb[CTA_HELP_MAX+1];
1458
1459         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1460                                           help_nla_policy, NULL);
1461         if (err < 0)
1462                 return err;
1463
1464         if (!tb[CTA_HELP_NAME])
1465                 return -EINVAL;
1466
1467         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1468
1469         if (tb[CTA_HELP_INFO])
1470                 *helpinfo = tb[CTA_HELP_INFO];
1471
1472         return 0;
1473 }
1474
1475 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1476         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1477         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1478         [CTA_STATUS]            = { .type = NLA_U32 },
1479         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1480         [CTA_HELP]              = { .type = NLA_NESTED },
1481         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1482         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1483         [CTA_MARK]              = { .type = NLA_U32 },
1484         [CTA_ID]                = { .type = NLA_U32 },
1485         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1486         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1487         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1488         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1489         [CTA_ZONE]              = { .type = NLA_U16 },
1490         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1491         [CTA_LABELS]            = { .type = NLA_BINARY,
1492                                     .len = NF_CT_LABELS_MAX_SIZE },
1493         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1494                                     .len = NF_CT_LABELS_MAX_SIZE },
1495         [CTA_FILTER]            = { .type = NLA_NESTED },
1496 };
1497
1498 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1499 {
1500         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1501                 return 0;
1502
1503         return ctnetlink_filter_match(ct, data);
1504 }
1505
1506 static int ctnetlink_flush_conntrack(struct net *net,
1507                                      const struct nlattr * const cda[],
1508                                      u32 portid, int report, u8 family)
1509 {
1510         struct ctnetlink_filter *filter = NULL;
1511
1512         if (ctnetlink_needs_filter(family, cda)) {
1513                 if (cda[CTA_FILTER])
1514                         return -EOPNOTSUPP;
1515
1516                 filter = ctnetlink_alloc_filter(cda, family);
1517                 if (IS_ERR(filter))
1518                         return PTR_ERR(filter);
1519         }
1520
1521         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1522                                   portid, report);
1523         kfree(filter);
1524
1525         return 0;
1526 }
1527
1528 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1529                                    struct sk_buff *skb,
1530                                    const struct nlmsghdr *nlh,
1531                                    const struct nlattr * const cda[],
1532                                    struct netlink_ext_ack *extack)
1533 {
1534         struct nf_conntrack_tuple_hash *h;
1535         struct nf_conntrack_tuple tuple;
1536         struct nf_conn *ct;
1537         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1538         struct nf_conntrack_zone zone;
1539         int err;
1540
1541         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1542         if (err < 0)
1543                 return err;
1544
1545         if (cda[CTA_TUPLE_ORIG])
1546                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1547                                             nfmsg->nfgen_family, &zone);
1548         else if (cda[CTA_TUPLE_REPLY])
1549                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1550                                             nfmsg->nfgen_family, &zone);
1551         else {
1552                 u_int8_t u3 = nfmsg->version ? nfmsg->nfgen_family : AF_UNSPEC;
1553
1554                 return ctnetlink_flush_conntrack(net, cda,
1555                                                  NETLINK_CB(skb).portid,
1556                                                  nlmsg_report(nlh), u3);
1557         }
1558
1559         if (err < 0)
1560                 return err;
1561
1562         h = nf_conntrack_find_get(net, &zone, &tuple);
1563         if (!h)
1564                 return -ENOENT;
1565
1566         ct = nf_ct_tuplehash_to_ctrack(h);
1567
1568         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1569                 nf_ct_put(ct);
1570                 return -EBUSY;
1571         }
1572
1573         if (cda[CTA_ID]) {
1574                 __be32 id = nla_get_be32(cda[CTA_ID]);
1575
1576                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1577                         nf_ct_put(ct);
1578                         return -ENOENT;
1579                 }
1580         }
1581
1582         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1583         nf_ct_put(ct);
1584
1585         return 0;
1586 }
1587
1588 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1589                                    struct sk_buff *skb,
1590                                    const struct nlmsghdr *nlh,
1591                                    const struct nlattr * const cda[],
1592                                    struct netlink_ext_ack *extack)
1593 {
1594         struct nf_conntrack_tuple_hash *h;
1595         struct nf_conntrack_tuple tuple;
1596         struct nf_conn *ct;
1597         struct sk_buff *skb2 = NULL;
1598         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1599         u_int8_t u3 = nfmsg->nfgen_family;
1600         struct nf_conntrack_zone zone;
1601         int err;
1602
1603         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1604                 struct netlink_dump_control c = {
1605                         .start = ctnetlink_start,
1606                         .dump = ctnetlink_dump_table,
1607                         .done = ctnetlink_done,
1608                         .data = (void *)cda,
1609                 };
1610
1611                 return netlink_dump_start(ctnl, skb, nlh, &c);
1612         }
1613
1614         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1615         if (err < 0)
1616                 return err;
1617
1618         if (cda[CTA_TUPLE_ORIG])
1619                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1620                                             u3, &zone);
1621         else if (cda[CTA_TUPLE_REPLY])
1622                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1623                                             u3, &zone);
1624         else
1625                 return -EINVAL;
1626
1627         if (err < 0)
1628                 return err;
1629
1630         h = nf_conntrack_find_get(net, &zone, &tuple);
1631         if (!h)
1632                 return -ENOENT;
1633
1634         ct = nf_ct_tuplehash_to_ctrack(h);
1635
1636         err = -ENOMEM;
1637         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1638         if (skb2 == NULL) {
1639                 nf_ct_put(ct);
1640                 return -ENOMEM;
1641         }
1642
1643         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1644                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct, true, 0);
1645         nf_ct_put(ct);
1646         if (err <= 0)
1647                 goto free;
1648
1649         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1650         if (err < 0)
1651                 goto out;
1652
1653         return 0;
1654
1655 free:
1656         kfree_skb(skb2);
1657 out:
1658         /* this avoids a loop in nfnetlink. */
1659         return err == -EAGAIN ? -ENOBUFS : err;
1660 }
1661
1662 static int ctnetlink_done_list(struct netlink_callback *cb)
1663 {
1664         if (cb->args[1])
1665                 nf_ct_put((struct nf_conn *)cb->args[1]);
1666         return 0;
1667 }
1668
1669 static int
1670 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1671 {
1672         struct nf_conn *ct, *last;
1673         struct nf_conntrack_tuple_hash *h;
1674         struct hlist_nulls_node *n;
1675         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1676         u_int8_t l3proto = nfmsg->nfgen_family;
1677         int res;
1678         int cpu;
1679         struct hlist_nulls_head *list;
1680         struct net *net = sock_net(skb->sk);
1681
1682         if (cb->args[2])
1683                 return 0;
1684
1685         last = (struct nf_conn *)cb->args[1];
1686
1687         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1688                 struct ct_pcpu *pcpu;
1689
1690                 if (!cpu_possible(cpu))
1691                         continue;
1692
1693                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1694                 spin_lock_bh(&pcpu->lock);
1695                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1696 restart:
1697                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1698                         ct = nf_ct_tuplehash_to_ctrack(h);
1699                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1700                                 continue;
1701                         if (cb->args[1]) {
1702                                 if (ct != last)
1703                                         continue;
1704                                 cb->args[1] = 0;
1705                         }
1706
1707                         /* We can't dump extension info for the unconfirmed
1708                          * list because unconfirmed conntracks can have
1709                          * ct->ext reallocated (and thus freed).
1710                          *
1711                          * In the dying list case ct->ext can't be free'd
1712                          * until after we drop pcpu->lock.
1713                          */
1714                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1715                                                   cb->nlh->nlmsg_seq,
1716                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1717                                                   ct, dying ? true : false, 0);
1718                         if (res < 0) {
1719                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1720                                         continue;
1721                                 cb->args[0] = cpu;
1722                                 cb->args[1] = (unsigned long)ct;
1723                                 spin_unlock_bh(&pcpu->lock);
1724                                 goto out;
1725                         }
1726                 }
1727                 if (cb->args[1]) {
1728                         cb->args[1] = 0;
1729                         goto restart;
1730                 }
1731                 spin_unlock_bh(&pcpu->lock);
1732         }
1733         cb->args[2] = 1;
1734 out:
1735         if (last)
1736                 nf_ct_put(last);
1737
1738         return skb->len;
1739 }
1740
1741 static int
1742 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1743 {
1744         return ctnetlink_dump_list(skb, cb, true);
1745 }
1746
1747 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1748                                   struct sk_buff *skb,
1749                                   const struct nlmsghdr *nlh,
1750                                   const struct nlattr * const cda[],
1751                                   struct netlink_ext_ack *extack)
1752 {
1753         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1754                 struct netlink_dump_control c = {
1755                         .dump = ctnetlink_dump_dying,
1756                         .done = ctnetlink_done_list,
1757                 };
1758                 return netlink_dump_start(ctnl, skb, nlh, &c);
1759         }
1760
1761         return -EOPNOTSUPP;
1762 }
1763
1764 static int
1765 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1766 {
1767         return ctnetlink_dump_list(skb, cb, false);
1768 }
1769
1770 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1771                                         struct sk_buff *skb,
1772                                         const struct nlmsghdr *nlh,
1773                                         const struct nlattr * const cda[],
1774                                         struct netlink_ext_ack *extack)
1775 {
1776         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1777                 struct netlink_dump_control c = {
1778                         .dump = ctnetlink_dump_unconfirmed,
1779                         .done = ctnetlink_done_list,
1780                 };
1781                 return netlink_dump_start(ctnl, skb, nlh, &c);
1782         }
1783
1784         return -EOPNOTSUPP;
1785 }
1786
1787 #if IS_ENABLED(CONFIG_NF_NAT)
1788 static int
1789 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1790                           enum nf_nat_manip_type manip,
1791                           const struct nlattr *attr)
1792         __must_hold(RCU)
1793 {
1794         struct nf_nat_hook *nat_hook;
1795         int err;
1796
1797         nat_hook = rcu_dereference(nf_nat_hook);
1798         if (!nat_hook) {
1799 #ifdef CONFIG_MODULES
1800                 rcu_read_unlock();
1801                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1802                 if (request_module("nf-nat") < 0) {
1803                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1804                         rcu_read_lock();
1805                         return -EOPNOTSUPP;
1806                 }
1807                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1808                 rcu_read_lock();
1809                 nat_hook = rcu_dereference(nf_nat_hook);
1810                 if (nat_hook)
1811                         return -EAGAIN;
1812 #endif
1813                 return -EOPNOTSUPP;
1814         }
1815
1816         err = nat_hook->parse_nat_setup(ct, manip, attr);
1817         if (err == -EAGAIN) {
1818 #ifdef CONFIG_MODULES
1819                 rcu_read_unlock();
1820                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1821                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1822                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1823                         rcu_read_lock();
1824                         return -EOPNOTSUPP;
1825                 }
1826                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1827                 rcu_read_lock();
1828 #else
1829                 err = -EOPNOTSUPP;
1830 #endif
1831         }
1832         return err;
1833 }
1834 #endif
1835
1836 static void
1837 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1838                           unsigned long off)
1839 {
1840         unsigned int bit;
1841
1842         /* Ignore these unchangable bits */
1843         on &= ~IPS_UNCHANGEABLE_MASK;
1844         off &= ~IPS_UNCHANGEABLE_MASK;
1845
1846         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1847                 if (on & (1 << bit))
1848                         set_bit(bit, &ct->status);
1849                 else if (off & (1 << bit))
1850                         clear_bit(bit, &ct->status);
1851         }
1852 }
1853
1854 static int
1855 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1856 {
1857         unsigned long d;
1858         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1859         d = ct->status ^ status;
1860
1861         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1862                 /* unchangeable */
1863                 return -EBUSY;
1864
1865         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1866                 /* SEEN_REPLY bit can only be set */
1867                 return -EBUSY;
1868
1869         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1870                 /* ASSURED bit can only be set */
1871                 return -EBUSY;
1872
1873         __ctnetlink_change_status(ct, status, 0);
1874         return 0;
1875 }
1876
1877 static int
1878 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1879 {
1880 #if IS_ENABLED(CONFIG_NF_NAT)
1881         int ret;
1882
1883         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1884                 return 0;
1885
1886         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1887                                         cda[CTA_NAT_DST]);
1888         if (ret < 0)
1889                 return ret;
1890
1891         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1892                                          cda[CTA_NAT_SRC]);
1893 #else
1894         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1895                 return 0;
1896         return -EOPNOTSUPP;
1897 #endif
1898 }
1899
1900 static int ctnetlink_change_helper(struct nf_conn *ct,
1901                                    const struct nlattr * const cda[])
1902 {
1903         struct nf_conntrack_helper *helper;
1904         struct nf_conn_help *help = nfct_help(ct);
1905         char *helpname = NULL;
1906         struct nlattr *helpinfo = NULL;
1907         int err;
1908
1909         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1910         if (err < 0)
1911                 return err;
1912
1913         /* don't change helper of sibling connections */
1914         if (ct->master) {
1915                 /* If we try to change the helper to the same thing twice,
1916                  * treat the second attempt as a no-op instead of returning
1917                  * an error.
1918                  */
1919                 err = -EBUSY;
1920                 if (help) {
1921                         rcu_read_lock();
1922                         helper = rcu_dereference(help->helper);
1923                         if (helper && !strcmp(helper->name, helpname))
1924                                 err = 0;
1925                         rcu_read_unlock();
1926                 }
1927
1928                 return err;
1929         }
1930
1931         if (!strcmp(helpname, "")) {
1932                 if (help && help->helper) {
1933                         /* we had a helper before ... */
1934                         nf_ct_remove_expectations(ct);
1935                         RCU_INIT_POINTER(help->helper, NULL);
1936                 }
1937
1938                 return 0;
1939         }
1940
1941         rcu_read_lock();
1942         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1943                                             nf_ct_protonum(ct));
1944         if (helper == NULL) {
1945                 rcu_read_unlock();
1946                 return -EOPNOTSUPP;
1947         }
1948
1949         if (help) {
1950                 if (help->helper == helper) {
1951                         /* update private helper data if allowed. */
1952                         if (helper->from_nlattr)
1953                                 helper->from_nlattr(helpinfo, ct);
1954                         err = 0;
1955                 } else
1956                         err = -EBUSY;
1957         } else {
1958                 /* we cannot set a helper for an existing conntrack */
1959                 err = -EOPNOTSUPP;
1960         }
1961
1962         rcu_read_unlock();
1963         return err;
1964 }
1965
1966 static int ctnetlink_change_timeout(struct nf_conn *ct,
1967                                     const struct nlattr * const cda[])
1968 {
1969         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1970
1971         if (timeout > INT_MAX)
1972                 timeout = INT_MAX;
1973         ct->timeout = nfct_time_stamp + (u32)timeout;
1974
1975         if (test_bit(IPS_DYING_BIT, &ct->status))
1976                 return -ETIME;
1977
1978         return 0;
1979 }
1980
1981 #if defined(CONFIG_NF_CONNTRACK_MARK)
1982 static void ctnetlink_change_mark(struct nf_conn *ct,
1983                                     const struct nlattr * const cda[])
1984 {
1985         u32 mark, newmark, mask = 0;
1986
1987         if (cda[CTA_MARK_MASK])
1988                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1989
1990         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1991         newmark = (ct->mark & mask) ^ mark;
1992         if (newmark != ct->mark)
1993                 ct->mark = newmark;
1994 }
1995 #endif
1996
1997 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1998         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1999         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2000         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2001 };
2002
2003 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2004                                       const struct nlattr * const cda[])
2005 {
2006         const struct nlattr *attr = cda[CTA_PROTOINFO];
2007         const struct nf_conntrack_l4proto *l4proto;
2008         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2009         int err = 0;
2010
2011         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2012                                           protoinfo_policy, NULL);
2013         if (err < 0)
2014                 return err;
2015
2016         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2017         if (l4proto->from_nlattr)
2018                 err = l4proto->from_nlattr(tb, ct);
2019
2020         return err;
2021 }
2022
2023 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2024         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2025         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2026         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2027 };
2028
2029 static int change_seq_adj(struct nf_ct_seqadj *seq,
2030                           const struct nlattr * const attr)
2031 {
2032         int err;
2033         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2034
2035         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2036                                           seqadj_policy, NULL);
2037         if (err < 0)
2038                 return err;
2039
2040         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2041                 return -EINVAL;
2042
2043         seq->correction_pos =
2044                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2045
2046         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2047                 return -EINVAL;
2048
2049         seq->offset_before =
2050                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2051
2052         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2053                 return -EINVAL;
2054
2055         seq->offset_after =
2056                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2057
2058         return 0;
2059 }
2060
2061 static int
2062 ctnetlink_change_seq_adj(struct nf_conn *ct,
2063                          const struct nlattr * const cda[])
2064 {
2065         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2066         int ret = 0;
2067
2068         if (!seqadj)
2069                 return 0;
2070
2071         spin_lock_bh(&ct->lock);
2072         if (cda[CTA_SEQ_ADJ_ORIG]) {
2073                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2074                                      cda[CTA_SEQ_ADJ_ORIG]);
2075                 if (ret < 0)
2076                         goto err;
2077
2078                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2079         }
2080
2081         if (cda[CTA_SEQ_ADJ_REPLY]) {
2082                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2083                                      cda[CTA_SEQ_ADJ_REPLY]);
2084                 if (ret < 0)
2085                         goto err;
2086
2087                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2088         }
2089
2090         spin_unlock_bh(&ct->lock);
2091         return 0;
2092 err:
2093         spin_unlock_bh(&ct->lock);
2094         return ret;
2095 }
2096
2097 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2098         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2099         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2100         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2101 };
2102
2103 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2104                                      const struct nlattr * const cda[])
2105 {
2106         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2107         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2108         int err;
2109
2110         if (!synproxy)
2111                 return 0;
2112
2113         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2114                                           cda[CTA_SYNPROXY], synproxy_policy,
2115                                           NULL);
2116         if (err < 0)
2117                 return err;
2118
2119         if (!tb[CTA_SYNPROXY_ISN] ||
2120             !tb[CTA_SYNPROXY_ITS] ||
2121             !tb[CTA_SYNPROXY_TSOFF])
2122                 return -EINVAL;
2123
2124         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2125         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2126         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2127
2128         return 0;
2129 }
2130
2131 static int
2132 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2133 {
2134 #ifdef CONFIG_NF_CONNTRACK_LABELS
2135         size_t len = nla_len(cda[CTA_LABELS]);
2136         const void *mask = cda[CTA_LABELS_MASK];
2137
2138         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2139                 return -EINVAL;
2140
2141         if (mask) {
2142                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2143                     nla_len(cda[CTA_LABELS_MASK]) != len)
2144                         return -EINVAL;
2145                 mask = nla_data(cda[CTA_LABELS_MASK]);
2146         }
2147
2148         len /= sizeof(u32);
2149
2150         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2151 #else
2152         return -EOPNOTSUPP;
2153 #endif
2154 }
2155
2156 static int
2157 ctnetlink_change_conntrack(struct nf_conn *ct,
2158                            const struct nlattr * const cda[])
2159 {
2160         int err;
2161
2162         /* only allow NAT changes and master assignation for new conntracks */
2163         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2164                 return -EOPNOTSUPP;
2165
2166         if (cda[CTA_HELP]) {
2167                 err = ctnetlink_change_helper(ct, cda);
2168                 if (err < 0)
2169                         return err;
2170         }
2171
2172         if (cda[CTA_TIMEOUT]) {
2173                 err = ctnetlink_change_timeout(ct, cda);
2174                 if (err < 0)
2175                         return err;
2176         }
2177
2178         if (cda[CTA_STATUS]) {
2179                 err = ctnetlink_change_status(ct, cda);
2180                 if (err < 0)
2181                         return err;
2182         }
2183
2184         if (cda[CTA_PROTOINFO]) {
2185                 err = ctnetlink_change_protoinfo(ct, cda);
2186                 if (err < 0)
2187                         return err;
2188         }
2189
2190 #if defined(CONFIG_NF_CONNTRACK_MARK)
2191         if (cda[CTA_MARK])
2192                 ctnetlink_change_mark(ct, cda);
2193 #endif
2194
2195         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2196                 err = ctnetlink_change_seq_adj(ct, cda);
2197                 if (err < 0)
2198                         return err;
2199         }
2200
2201         if (cda[CTA_SYNPROXY]) {
2202                 err = ctnetlink_change_synproxy(ct, cda);
2203                 if (err < 0)
2204                         return err;
2205         }
2206
2207         if (cda[CTA_LABELS]) {
2208                 err = ctnetlink_attach_labels(ct, cda);
2209                 if (err < 0)
2210                         return err;
2211         }
2212
2213         return 0;
2214 }
2215
2216 static struct nf_conn *
2217 ctnetlink_create_conntrack(struct net *net,
2218                            const struct nf_conntrack_zone *zone,
2219                            const struct nlattr * const cda[],
2220                            struct nf_conntrack_tuple *otuple,
2221                            struct nf_conntrack_tuple *rtuple,
2222                            u8 u3)
2223 {
2224         struct nf_conn *ct;
2225         int err = -EINVAL;
2226         struct nf_conntrack_helper *helper;
2227         struct nf_conn_tstamp *tstamp;
2228         u64 timeout;
2229
2230         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2231         if (IS_ERR(ct))
2232                 return ERR_PTR(-ENOMEM);
2233
2234         if (!cda[CTA_TIMEOUT])
2235                 goto err1;
2236
2237         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2238         if (timeout > INT_MAX)
2239                 timeout = INT_MAX;
2240         ct->timeout = (u32)timeout + nfct_time_stamp;
2241
2242         rcu_read_lock();
2243         if (cda[CTA_HELP]) {
2244                 char *helpname = NULL;
2245                 struct nlattr *helpinfo = NULL;
2246
2247                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2248                 if (err < 0)
2249                         goto err2;
2250
2251                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2252                                                     nf_ct_protonum(ct));
2253                 if (helper == NULL) {
2254                         rcu_read_unlock();
2255 #ifdef CONFIG_MODULES
2256                         if (request_module("nfct-helper-%s", helpname) < 0) {
2257                                 err = -EOPNOTSUPP;
2258                                 goto err1;
2259                         }
2260
2261                         rcu_read_lock();
2262                         helper = __nf_conntrack_helper_find(helpname,
2263                                                             nf_ct_l3num(ct),
2264                                                             nf_ct_protonum(ct));
2265                         if (helper) {
2266                                 err = -EAGAIN;
2267                                 goto err2;
2268                         }
2269                         rcu_read_unlock();
2270 #endif
2271                         err = -EOPNOTSUPP;
2272                         goto err1;
2273                 } else {
2274                         struct nf_conn_help *help;
2275
2276                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2277                         if (help == NULL) {
2278                                 err = -ENOMEM;
2279                                 goto err2;
2280                         }
2281                         /* set private helper data if allowed. */
2282                         if (helper->from_nlattr)
2283                                 helper->from_nlattr(helpinfo, ct);
2284
2285                         /* not in hash table yet so not strictly necessary */
2286                         RCU_INIT_POINTER(help->helper, helper);
2287                 }
2288         } else {
2289                 /* try an implicit helper assignation */
2290                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2291                 if (err < 0)
2292                         goto err2;
2293         }
2294
2295         err = ctnetlink_setup_nat(ct, cda);
2296         if (err < 0)
2297                 goto err2;
2298
2299         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2300         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2301         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2302         nf_ct_labels_ext_add(ct);
2303         nfct_seqadj_ext_add(ct);
2304         nfct_synproxy_ext_add(ct);
2305
2306         /* we must add conntrack extensions before confirmation. */
2307         ct->status |= IPS_CONFIRMED;
2308
2309         if (cda[CTA_STATUS]) {
2310                 err = ctnetlink_change_status(ct, cda);
2311                 if (err < 0)
2312                         goto err2;
2313         }
2314
2315         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2316                 err = ctnetlink_change_seq_adj(ct, cda);
2317                 if (err < 0)
2318                         goto err2;
2319         }
2320
2321         memset(&ct->proto, 0, sizeof(ct->proto));
2322         if (cda[CTA_PROTOINFO]) {
2323                 err = ctnetlink_change_protoinfo(ct, cda);
2324                 if (err < 0)
2325                         goto err2;
2326         }
2327
2328         if (cda[CTA_SYNPROXY]) {
2329                 err = ctnetlink_change_synproxy(ct, cda);
2330                 if (err < 0)
2331                         goto err2;
2332         }
2333
2334 #if defined(CONFIG_NF_CONNTRACK_MARK)
2335         if (cda[CTA_MARK])
2336                 ctnetlink_change_mark(ct, cda);
2337 #endif
2338
2339         /* setup master conntrack: this is a confirmed expectation */
2340         if (cda[CTA_TUPLE_MASTER]) {
2341                 struct nf_conntrack_tuple master;
2342                 struct nf_conntrack_tuple_hash *master_h;
2343                 struct nf_conn *master_ct;
2344
2345                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2346                                             u3, NULL);
2347                 if (err < 0)
2348                         goto err2;
2349
2350                 master_h = nf_conntrack_find_get(net, zone, &master);
2351                 if (master_h == NULL) {
2352                         err = -ENOENT;
2353                         goto err2;
2354                 }
2355                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2356                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2357                 ct->master = master_ct;
2358         }
2359         tstamp = nf_conn_tstamp_find(ct);
2360         if (tstamp)
2361                 tstamp->start = ktime_get_real_ns();
2362
2363         err = nf_conntrack_hash_check_insert(ct);
2364         if (err < 0)
2365                 goto err2;
2366
2367         rcu_read_unlock();
2368
2369         return ct;
2370
2371 err2:
2372         rcu_read_unlock();
2373 err1:
2374         nf_conntrack_free(ct);
2375         return ERR_PTR(err);
2376 }
2377
2378 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2379                                    struct sk_buff *skb,
2380                                    const struct nlmsghdr *nlh,
2381                                    const struct nlattr * const cda[],
2382                                    struct netlink_ext_ack *extack)
2383 {
2384         struct nf_conntrack_tuple otuple, rtuple;
2385         struct nf_conntrack_tuple_hash *h = NULL;
2386         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2387         struct nf_conn *ct;
2388         u_int8_t u3 = nfmsg->nfgen_family;
2389         struct nf_conntrack_zone zone;
2390         int err;
2391
2392         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2393         if (err < 0)
2394                 return err;
2395
2396         if (cda[CTA_TUPLE_ORIG]) {
2397                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2398                                             u3, &zone);
2399                 if (err < 0)
2400                         return err;
2401         }
2402
2403         if (cda[CTA_TUPLE_REPLY]) {
2404                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2405                                             u3, &zone);
2406                 if (err < 0)
2407                         return err;
2408         }
2409
2410         if (cda[CTA_TUPLE_ORIG])
2411                 h = nf_conntrack_find_get(net, &zone, &otuple);
2412         else if (cda[CTA_TUPLE_REPLY])
2413                 h = nf_conntrack_find_get(net, &zone, &rtuple);
2414
2415         if (h == NULL) {
2416                 err = -ENOENT;
2417                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2418                         enum ip_conntrack_events events;
2419
2420                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2421                                 return -EINVAL;
2422                         if (otuple.dst.protonum != rtuple.dst.protonum)
2423                                 return -EINVAL;
2424
2425                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2426                                                         &rtuple, u3);
2427                         if (IS_ERR(ct))
2428                                 return PTR_ERR(ct);
2429
2430                         err = 0;
2431                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2432                                 events = 1 << IPCT_RELATED;
2433                         else
2434                                 events = 1 << IPCT_NEW;
2435
2436                         if (cda[CTA_LABELS] &&
2437                             ctnetlink_attach_labels(ct, cda) == 0)
2438                                 events |= (1 << IPCT_LABEL);
2439
2440                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2441                                                       (1 << IPCT_ASSURED) |
2442                                                       (1 << IPCT_HELPER) |
2443                                                       (1 << IPCT_PROTOINFO) |
2444                                                       (1 << IPCT_SEQADJ) |
2445                                                       (1 << IPCT_MARK) |
2446                                                       (1 << IPCT_SYNPROXY) |
2447                                                       events,
2448                                                       ct, NETLINK_CB(skb).portid,
2449                                                       nlmsg_report(nlh));
2450                         nf_ct_put(ct);
2451                 }
2452
2453                 return err;
2454         }
2455         /* implicit 'else' */
2456
2457         err = -EEXIST;
2458         ct = nf_ct_tuplehash_to_ctrack(h);
2459         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2460                 err = ctnetlink_change_conntrack(ct, cda);
2461                 if (err == 0) {
2462                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2463                                                       (1 << IPCT_ASSURED) |
2464                                                       (1 << IPCT_HELPER) |
2465                                                       (1 << IPCT_LABEL) |
2466                                                       (1 << IPCT_PROTOINFO) |
2467                                                       (1 << IPCT_SEQADJ) |
2468                                                       (1 << IPCT_MARK) |
2469                                                       (1 << IPCT_SYNPROXY),
2470                                                       ct, NETLINK_CB(skb).portid,
2471                                                       nlmsg_report(nlh));
2472                 }
2473         }
2474
2475         nf_ct_put(ct);
2476         return err;
2477 }
2478
2479 static int
2480 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2481                                 __u16 cpu, const struct ip_conntrack_stat *st)
2482 {
2483         struct nlmsghdr *nlh;
2484         struct nfgenmsg *nfmsg;
2485         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2486
2487         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2488                               IPCTNL_MSG_CT_GET_STATS_CPU);
2489         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2490         if (nlh == NULL)
2491                 goto nlmsg_failure;
2492
2493         nfmsg = nlmsg_data(nlh);
2494         nfmsg->nfgen_family = AF_UNSPEC;
2495         nfmsg->version      = NFNETLINK_V0;
2496         nfmsg->res_id       = htons(cpu);
2497
2498         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2499             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2500             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2501             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2502             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2503                                 htonl(st->insert_failed)) ||
2504             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2505             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2506             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2507             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2508                                 htonl(st->search_restart)))
2509                 goto nla_put_failure;
2510
2511         nlmsg_end(skb, nlh);
2512         return skb->len;
2513
2514 nla_put_failure:
2515 nlmsg_failure:
2516         nlmsg_cancel(skb, nlh);
2517         return -1;
2518 }
2519
2520 static int
2521 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2522 {
2523         int cpu;
2524         struct net *net = sock_net(skb->sk);
2525
2526         if (cb->args[0] == nr_cpu_ids)
2527                 return 0;
2528
2529         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2530                 const struct ip_conntrack_stat *st;
2531
2532                 if (!cpu_possible(cpu))
2533                         continue;
2534
2535                 st = per_cpu_ptr(net->ct.stat, cpu);
2536                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2537                                                     NETLINK_CB(cb->skb).portid,
2538                                                     cb->nlh->nlmsg_seq,
2539                                                     cpu, st) < 0)
2540                                 break;
2541         }
2542         cb->args[0] = cpu;
2543
2544         return skb->len;
2545 }
2546
2547 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2548                                  struct sk_buff *skb,
2549                                  const struct nlmsghdr *nlh,
2550                                  const struct nlattr * const cda[],
2551                                  struct netlink_ext_ack *extack)
2552 {
2553         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2554                 struct netlink_dump_control c = {
2555                         .dump = ctnetlink_ct_stat_cpu_dump,
2556                 };
2557                 return netlink_dump_start(ctnl, skb, nlh, &c);
2558         }
2559
2560         return 0;
2561 }
2562
2563 static int
2564 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2565                             struct net *net)
2566 {
2567         struct nlmsghdr *nlh;
2568         struct nfgenmsg *nfmsg;
2569         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2570         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2571
2572         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2573         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2574         if (nlh == NULL)
2575                 goto nlmsg_failure;
2576
2577         nfmsg = nlmsg_data(nlh);
2578         nfmsg->nfgen_family = AF_UNSPEC;
2579         nfmsg->version      = NFNETLINK_V0;
2580         nfmsg->res_id       = 0;
2581
2582         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2583                 goto nla_put_failure;
2584
2585         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2586                 goto nla_put_failure;
2587
2588         nlmsg_end(skb, nlh);
2589         return skb->len;
2590
2591 nla_put_failure:
2592 nlmsg_failure:
2593         nlmsg_cancel(skb, nlh);
2594         return -1;
2595 }
2596
2597 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2598                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2599                              const struct nlattr * const cda[],
2600                              struct netlink_ext_ack *extack)
2601 {
2602         struct sk_buff *skb2;
2603         int err;
2604
2605         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2606         if (skb2 == NULL)
2607                 return -ENOMEM;
2608
2609         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2610                                           nlh->nlmsg_seq,
2611                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2612                                           sock_net(skb->sk));
2613         if (err <= 0)
2614                 goto free;
2615
2616         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2617         if (err < 0)
2618                 goto out;
2619
2620         return 0;
2621
2622 free:
2623         kfree_skb(skb2);
2624 out:
2625         /* this avoids a loop in nfnetlink. */
2626         return err == -EAGAIN ? -ENOBUFS : err;
2627 }
2628
2629 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2630         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2631         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2632         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2633         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2634         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2635         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2636                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2637         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2638         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2639         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2640         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2641         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2642 };
2643
2644 static struct nf_conntrack_expect *
2645 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2646                        struct nf_conntrack_helper *helper,
2647                        struct nf_conntrack_tuple *tuple,
2648                        struct nf_conntrack_tuple *mask);
2649
2650 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2651 static size_t
2652 ctnetlink_glue_build_size(const struct nf_conn *ct)
2653 {
2654         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2655                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2656                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2657                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2658                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2659                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2660                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2661                + nla_total_size(0) /* CTA_PROTOINFO */
2662                + nla_total_size(0) /* CTA_HELP */
2663                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2664                + ctnetlink_secctx_size(ct)
2665 #if IS_ENABLED(CONFIG_NF_NAT)
2666                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2667                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2668 #endif
2669 #ifdef CONFIG_NF_CONNTRACK_MARK
2670                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2671 #endif
2672 #ifdef CONFIG_NF_CONNTRACK_ZONES
2673                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2674 #endif
2675                + ctnetlink_proto_size(ct)
2676                ;
2677 }
2678
2679 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2680                                              enum ip_conntrack_info *ctinfo)
2681 {
2682         return nf_ct_get(skb, ctinfo);
2683 }
2684
2685 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2686 {
2687         const struct nf_conntrack_zone *zone;
2688         struct nlattr *nest_parms;
2689
2690         zone = nf_ct_zone(ct);
2691
2692         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2693         if (!nest_parms)
2694                 goto nla_put_failure;
2695         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2696                 goto nla_put_failure;
2697         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2698                                    NF_CT_ZONE_DIR_ORIG) < 0)
2699                 goto nla_put_failure;
2700         nla_nest_end(skb, nest_parms);
2701
2702         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2703         if (!nest_parms)
2704                 goto nla_put_failure;
2705         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2706                 goto nla_put_failure;
2707         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2708                                    NF_CT_ZONE_DIR_REPL) < 0)
2709                 goto nla_put_failure;
2710         nla_nest_end(skb, nest_parms);
2711
2712         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2713                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2714                 goto nla_put_failure;
2715
2716         if (ctnetlink_dump_id(skb, ct) < 0)
2717                 goto nla_put_failure;
2718
2719         if (ctnetlink_dump_status(skb, ct) < 0)
2720                 goto nla_put_failure;
2721
2722         if (ctnetlink_dump_timeout(skb, ct) < 0)
2723                 goto nla_put_failure;
2724
2725         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2726                 goto nla_put_failure;
2727
2728         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2729                 goto nla_put_failure;
2730
2731 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2732         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2733                 goto nla_put_failure;
2734 #endif
2735         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2736                 goto nla_put_failure;
2737
2738         if ((ct->status & IPS_SEQ_ADJUST) &&
2739             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2740                 goto nla_put_failure;
2741
2742         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2743                 goto nla_put_failure;
2744
2745 #ifdef CONFIG_NF_CONNTRACK_MARK
2746         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2747                 goto nla_put_failure;
2748 #endif
2749         if (ctnetlink_dump_labels(skb, ct) < 0)
2750                 goto nla_put_failure;
2751         return 0;
2752
2753 nla_put_failure:
2754         return -ENOSPC;
2755 }
2756
2757 static int
2758 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2759                      enum ip_conntrack_info ctinfo,
2760                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2761 {
2762         struct nlattr *nest_parms;
2763
2764         nest_parms = nla_nest_start(skb, ct_attr);
2765         if (!nest_parms)
2766                 goto nla_put_failure;
2767
2768         if (__ctnetlink_glue_build(skb, ct) < 0)
2769                 goto nla_put_failure;
2770
2771         nla_nest_end(skb, nest_parms);
2772
2773         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2774                 goto nla_put_failure;
2775
2776         return 0;
2777
2778 nla_put_failure:
2779         return -ENOSPC;
2780 }
2781
2782 static int
2783 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2784 {
2785         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2786         unsigned long d = ct->status ^ status;
2787
2788         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2789                 /* SEEN_REPLY bit can only be set */
2790                 return -EBUSY;
2791
2792         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2793                 /* ASSURED bit can only be set */
2794                 return -EBUSY;
2795
2796         /* This check is less strict than ctnetlink_change_status()
2797          * because callers often flip IPS_EXPECTED bits when sending
2798          * an NFQA_CT attribute to the kernel.  So ignore the
2799          * unchangeable bits but do not error out. Also user programs
2800          * are allowed to clear the bits that they are allowed to change.
2801          */
2802         __ctnetlink_change_status(ct, status, ~status);
2803         return 0;
2804 }
2805
2806 static int
2807 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2808 {
2809         int err;
2810
2811         if (cda[CTA_TIMEOUT]) {
2812                 err = ctnetlink_change_timeout(ct, cda);
2813                 if (err < 0)
2814                         return err;
2815         }
2816         if (cda[CTA_STATUS]) {
2817                 err = ctnetlink_update_status(ct, cda);
2818                 if (err < 0)
2819                         return err;
2820         }
2821         if (cda[CTA_HELP]) {
2822                 err = ctnetlink_change_helper(ct, cda);
2823                 if (err < 0)
2824                         return err;
2825         }
2826         if (cda[CTA_LABELS]) {
2827                 err = ctnetlink_attach_labels(ct, cda);
2828                 if (err < 0)
2829                         return err;
2830         }
2831 #if defined(CONFIG_NF_CONNTRACK_MARK)
2832         if (cda[CTA_MARK]) {
2833                 ctnetlink_change_mark(ct, cda);
2834         }
2835 #endif
2836         return 0;
2837 }
2838
2839 static int
2840 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2841 {
2842         struct nlattr *cda[CTA_MAX+1];
2843         int ret;
2844
2845         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2846                                           NULL);
2847         if (ret < 0)
2848                 return ret;
2849
2850         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2851 }
2852
2853 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2854                                     const struct nf_conn *ct,
2855                                     struct nf_conntrack_tuple *tuple,
2856                                     struct nf_conntrack_tuple *mask)
2857 {
2858         int err;
2859
2860         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2861                                     nf_ct_l3num(ct), NULL);
2862         if (err < 0)
2863                 return err;
2864
2865         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2866                                      nf_ct_l3num(ct), NULL);
2867 }
2868
2869 static int
2870 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2871                              u32 portid, u32 report)
2872 {
2873         struct nlattr *cda[CTA_EXPECT_MAX+1];
2874         struct nf_conntrack_tuple tuple, mask;
2875         struct nf_conntrack_helper *helper = NULL;
2876         struct nf_conntrack_expect *exp;
2877         int err;
2878
2879         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2880                                           exp_nla_policy, NULL);
2881         if (err < 0)
2882                 return err;
2883
2884         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2885                                        ct, &tuple, &mask);
2886         if (err < 0)
2887                 return err;
2888
2889         if (cda[CTA_EXPECT_HELP_NAME]) {
2890                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2891
2892                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2893                                                     nf_ct_protonum(ct));
2894                 if (helper == NULL)
2895                         return -EOPNOTSUPP;
2896         }
2897
2898         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2899                                      helper, &tuple, &mask);
2900         if (IS_ERR(exp))
2901                 return PTR_ERR(exp);
2902
2903         err = nf_ct_expect_related_report(exp, portid, report, 0);
2904         nf_ct_expect_put(exp);
2905         return err;
2906 }
2907
2908 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2909                                   enum ip_conntrack_info ctinfo, int diff)
2910 {
2911         if (!(ct->status & IPS_NAT_MASK))
2912                 return;
2913
2914         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2915 }
2916
2917 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2918         .get_ct         = ctnetlink_glue_get_ct,
2919         .build_size     = ctnetlink_glue_build_size,
2920         .build          = ctnetlink_glue_build,
2921         .parse          = ctnetlink_glue_parse,
2922         .attach_expect  = ctnetlink_glue_attach_expect,
2923         .seq_adjust     = ctnetlink_glue_seqadj,
2924 };
2925 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2926
2927 /***********************************************************************
2928  * EXPECT
2929  ***********************************************************************/
2930
2931 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2932                                     const struct nf_conntrack_tuple *tuple,
2933                                     u32 type)
2934 {
2935         struct nlattr *nest_parms;
2936
2937         nest_parms = nla_nest_start(skb, type);
2938         if (!nest_parms)
2939                 goto nla_put_failure;
2940         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2941                 goto nla_put_failure;
2942         nla_nest_end(skb, nest_parms);
2943
2944         return 0;
2945
2946 nla_put_failure:
2947         return -1;
2948 }
2949
2950 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2951                                    const struct nf_conntrack_tuple *tuple,
2952                                    const struct nf_conntrack_tuple_mask *mask)
2953 {
2954         const struct nf_conntrack_l4proto *l4proto;
2955         struct nf_conntrack_tuple m;
2956         struct nlattr *nest_parms;
2957         int ret;
2958
2959         memset(&m, 0xFF, sizeof(m));
2960         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2961         m.src.u.all = mask->src.u.all;
2962         m.dst.protonum = tuple->dst.protonum;
2963
2964         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2965         if (!nest_parms)
2966                 goto nla_put_failure;
2967
2968         rcu_read_lock();
2969         ret = ctnetlink_dump_tuples_ip(skb, &m);
2970         if (ret >= 0) {
2971                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2972                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2973         }
2974         rcu_read_unlock();
2975
2976         if (unlikely(ret < 0))
2977                 goto nla_put_failure;
2978
2979         nla_nest_end(skb, nest_parms);
2980
2981         return 0;
2982
2983 nla_put_failure:
2984         return -1;
2985 }
2986
2987 static const union nf_inet_addr any_addr;
2988
2989 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2990 {
2991         static __read_mostly siphash_key_t exp_id_seed;
2992         unsigned long a, b, c, d;
2993
2994         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2995
2996         a = (unsigned long)exp;
2997         b = (unsigned long)exp->helper;
2998         c = (unsigned long)exp->master;
2999         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
3000
3001 #ifdef CONFIG_64BIT
3002         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
3003 #else
3004         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
3005 #endif
3006 }
3007
3008 static int
3009 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3010                           const struct nf_conntrack_expect *exp)
3011 {
3012         struct nf_conn *master = exp->master;
3013         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3014         struct nf_conn_help *help;
3015 #if IS_ENABLED(CONFIG_NF_NAT)
3016         struct nlattr *nest_parms;
3017         struct nf_conntrack_tuple nat_tuple = {};
3018 #endif
3019         struct nf_ct_helper_expectfn *expfn;
3020
3021         if (timeout < 0)
3022                 timeout = 0;
3023
3024         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3025                 goto nla_put_failure;
3026         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3027                 goto nla_put_failure;
3028         if (ctnetlink_exp_dump_tuple(skb,
3029                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3030                                  CTA_EXPECT_MASTER) < 0)
3031                 goto nla_put_failure;
3032
3033 #if IS_ENABLED(CONFIG_NF_NAT)
3034         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3035             exp->saved_proto.all) {
3036                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3037                 if (!nest_parms)
3038                         goto nla_put_failure;
3039
3040                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3041                         goto nla_put_failure;
3042
3043                 nat_tuple.src.l3num = nf_ct_l3num(master);
3044                 nat_tuple.src.u3 = exp->saved_addr;
3045                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3046                 nat_tuple.src.u = exp->saved_proto;
3047
3048                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3049                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3050                         goto nla_put_failure;
3051                 nla_nest_end(skb, nest_parms);
3052         }
3053 #endif
3054         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3055             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3056             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3057             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3058                 goto nla_put_failure;
3059         help = nfct_help(master);
3060         if (help) {
3061                 struct nf_conntrack_helper *helper;
3062
3063                 helper = rcu_dereference(help->helper);
3064                 if (helper &&
3065                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3066                         goto nla_put_failure;
3067         }
3068         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3069         if (expfn != NULL &&
3070             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3071                 goto nla_put_failure;
3072
3073         return 0;
3074
3075 nla_put_failure:
3076         return -1;
3077 }
3078
3079 static int
3080 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3081                         int event, const struct nf_conntrack_expect *exp)
3082 {
3083         struct nlmsghdr *nlh;
3084         struct nfgenmsg *nfmsg;
3085         unsigned int flags = portid ? NLM_F_MULTI : 0;
3086
3087         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3088         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3089         if (nlh == NULL)
3090                 goto nlmsg_failure;
3091
3092         nfmsg = nlmsg_data(nlh);
3093         nfmsg->nfgen_family = exp->tuple.src.l3num;
3094         nfmsg->version      = NFNETLINK_V0;
3095         nfmsg->res_id       = 0;
3096
3097         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3098                 goto nla_put_failure;
3099
3100         nlmsg_end(skb, nlh);
3101         return skb->len;
3102
3103 nlmsg_failure:
3104 nla_put_failure:
3105         nlmsg_cancel(skb, nlh);
3106         return -1;
3107 }
3108
3109 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3110 static int
3111 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
3112 {
3113         struct nf_conntrack_expect *exp = item->exp;
3114         struct net *net = nf_ct_exp_net(exp);
3115         struct nlmsghdr *nlh;
3116         struct nfgenmsg *nfmsg;
3117         struct sk_buff *skb;
3118         unsigned int type, group;
3119         int flags = 0;
3120
3121         if (events & (1 << IPEXP_DESTROY)) {
3122                 type = IPCTNL_MSG_EXP_DELETE;
3123                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3124         } else if (events & (1 << IPEXP_NEW)) {
3125                 type = IPCTNL_MSG_EXP_NEW;
3126                 flags = NLM_F_CREATE|NLM_F_EXCL;
3127                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3128         } else
3129                 return 0;
3130
3131         if (!item->report && !nfnetlink_has_listeners(net, group))
3132                 return 0;
3133
3134         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3135         if (skb == NULL)
3136                 goto errout;
3137
3138         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3139         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
3140         if (nlh == NULL)
3141                 goto nlmsg_failure;
3142
3143         nfmsg = nlmsg_data(nlh);
3144         nfmsg->nfgen_family = exp->tuple.src.l3num;
3145         nfmsg->version      = NFNETLINK_V0;
3146         nfmsg->res_id       = 0;
3147
3148         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3149                 goto nla_put_failure;
3150
3151         nlmsg_end(skb, nlh);
3152         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3153         return 0;
3154
3155 nla_put_failure:
3156         nlmsg_cancel(skb, nlh);
3157 nlmsg_failure:
3158         kfree_skb(skb);
3159 errout:
3160         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3161         return 0;
3162 }
3163 #endif
3164 static int ctnetlink_exp_done(struct netlink_callback *cb)
3165 {
3166         if (cb->args[1])
3167                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3168         return 0;
3169 }
3170
3171 static int
3172 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3173 {
3174         struct net *net = sock_net(skb->sk);
3175         struct nf_conntrack_expect *exp, *last;
3176         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3177         u_int8_t l3proto = nfmsg->nfgen_family;
3178
3179         rcu_read_lock();
3180         last = (struct nf_conntrack_expect *)cb->args[1];
3181         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3182 restart:
3183                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3184                                          hnode) {
3185                         if (l3proto && exp->tuple.src.l3num != l3proto)
3186                                 continue;
3187
3188                         if (!net_eq(nf_ct_net(exp->master), net))
3189                                 continue;
3190
3191                         if (cb->args[1]) {
3192                                 if (exp != last)
3193                                         continue;
3194                                 cb->args[1] = 0;
3195                         }
3196                         if (ctnetlink_exp_fill_info(skb,
3197                                                     NETLINK_CB(cb->skb).portid,
3198                                                     cb->nlh->nlmsg_seq,
3199                                                     IPCTNL_MSG_EXP_NEW,
3200                                                     exp) < 0) {
3201                                 if (!refcount_inc_not_zero(&exp->use))
3202                                         continue;
3203                                 cb->args[1] = (unsigned long)exp;
3204                                 goto out;
3205                         }
3206                 }
3207                 if (cb->args[1]) {
3208                         cb->args[1] = 0;
3209                         goto restart;
3210                 }
3211         }
3212 out:
3213         rcu_read_unlock();
3214         if (last)
3215                 nf_ct_expect_put(last);
3216
3217         return skb->len;
3218 }
3219
3220 static int
3221 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3222 {
3223         struct nf_conntrack_expect *exp, *last;
3224         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3225         struct nf_conn *ct = cb->data;
3226         struct nf_conn_help *help = nfct_help(ct);
3227         u_int8_t l3proto = nfmsg->nfgen_family;
3228
3229         if (cb->args[0])
3230                 return 0;
3231
3232         rcu_read_lock();
3233         last = (struct nf_conntrack_expect *)cb->args[1];
3234 restart:
3235         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3236                 if (l3proto && exp->tuple.src.l3num != l3proto)
3237                         continue;
3238                 if (cb->args[1]) {
3239                         if (exp != last)
3240                                 continue;
3241                         cb->args[1] = 0;
3242                 }
3243                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3244                                             cb->nlh->nlmsg_seq,
3245                                             IPCTNL_MSG_EXP_NEW,
3246                                             exp) < 0) {
3247                         if (!refcount_inc_not_zero(&exp->use))
3248                                 continue;
3249                         cb->args[1] = (unsigned long)exp;
3250                         goto out;
3251                 }
3252         }
3253         if (cb->args[1]) {
3254                 cb->args[1] = 0;
3255                 goto restart;
3256         }
3257         cb->args[0] = 1;
3258 out:
3259         rcu_read_unlock();
3260         if (last)
3261                 nf_ct_expect_put(last);
3262
3263         return skb->len;
3264 }
3265
3266 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3267                                  struct sk_buff *skb,
3268                                  const struct nlmsghdr *nlh,
3269                                  const struct nlattr * const cda[],
3270                                  struct netlink_ext_ack *extack)
3271 {
3272         int err;
3273         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3274         u_int8_t u3 = nfmsg->nfgen_family;
3275         struct nf_conntrack_tuple tuple;
3276         struct nf_conntrack_tuple_hash *h;
3277         struct nf_conn *ct;
3278         struct nf_conntrack_zone zone;
3279         struct netlink_dump_control c = {
3280                 .dump = ctnetlink_exp_ct_dump_table,
3281                 .done = ctnetlink_exp_done,
3282         };
3283
3284         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3285                                     u3, NULL);
3286         if (err < 0)
3287                 return err;
3288
3289         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3290         if (err < 0)
3291                 return err;
3292
3293         h = nf_conntrack_find_get(net, &zone, &tuple);
3294         if (!h)
3295                 return -ENOENT;
3296
3297         ct = nf_ct_tuplehash_to_ctrack(h);
3298         /* No expectation linked to this connection tracking. */
3299         if (!nfct_help(ct)) {
3300                 nf_ct_put(ct);
3301                 return 0;
3302         }
3303
3304         c.data = ct;
3305
3306         err = netlink_dump_start(ctnl, skb, nlh, &c);
3307         nf_ct_put(ct);
3308
3309         return err;
3310 }
3311
3312 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
3313                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3314                                 const struct nlattr * const cda[],
3315                                 struct netlink_ext_ack *extack)
3316 {
3317         struct nf_conntrack_tuple tuple;
3318         struct nf_conntrack_expect *exp;
3319         struct sk_buff *skb2;
3320         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3321         u_int8_t u3 = nfmsg->nfgen_family;
3322         struct nf_conntrack_zone zone;
3323         int err;
3324
3325         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3326                 if (cda[CTA_EXPECT_MASTER])
3327                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
3328                                                      extack);
3329                 else {
3330                         struct netlink_dump_control c = {
3331                                 .dump = ctnetlink_exp_dump_table,
3332                                 .done = ctnetlink_exp_done,
3333                         };
3334                         return netlink_dump_start(ctnl, skb, nlh, &c);
3335                 }
3336         }
3337
3338         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3339         if (err < 0)
3340                 return err;
3341
3342         if (cda[CTA_EXPECT_TUPLE])
3343                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3344                                             u3, NULL);
3345         else if (cda[CTA_EXPECT_MASTER])
3346                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3347                                             u3, NULL);
3348         else
3349                 return -EINVAL;
3350
3351         if (err < 0)
3352                 return err;
3353
3354         exp = nf_ct_expect_find_get(net, &zone, &tuple);
3355         if (!exp)
3356                 return -ENOENT;
3357
3358         if (cda[CTA_EXPECT_ID]) {
3359                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3360
3361                 if (id != nf_expect_get_id(exp)) {
3362                         nf_ct_expect_put(exp);
3363                         return -ENOENT;
3364                 }
3365         }
3366
3367         err = -ENOMEM;
3368         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3369         if (skb2 == NULL) {
3370                 nf_ct_expect_put(exp);
3371                 goto out;
3372         }
3373
3374         rcu_read_lock();
3375         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3376                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
3377         rcu_read_unlock();
3378         nf_ct_expect_put(exp);
3379         if (err <= 0)
3380                 goto free;
3381
3382         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
3383         if (err < 0)
3384                 goto out;
3385
3386         return 0;
3387
3388 free:
3389         kfree_skb(skb2);
3390 out:
3391         /* this avoids a loop in nfnetlink. */
3392         return err == -EAGAIN ? -ENOBUFS : err;
3393 }
3394
3395 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3396 {
3397         const struct nf_conn_help *m_help;
3398         const char *name = data;
3399
3400         m_help = nfct_help(exp->master);
3401
3402         return strcmp(m_help->helper->name, name) == 0;
3403 }
3404
3405 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3406 {
3407         return true;
3408 }
3409
3410 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3411                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3412                                 const struct nlattr * const cda[],
3413                                 struct netlink_ext_ack *extack)
3414 {
3415         struct nf_conntrack_expect *exp;
3416         struct nf_conntrack_tuple tuple;
3417         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3418         u_int8_t u3 = nfmsg->nfgen_family;
3419         struct nf_conntrack_zone zone;
3420         int err;
3421
3422         if (cda[CTA_EXPECT_TUPLE]) {
3423                 /* delete a single expect by tuple */
3424                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3425                 if (err < 0)
3426                         return err;
3427
3428                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3429                                             u3, NULL);
3430                 if (err < 0)
3431                         return err;
3432
3433                 /* bump usage count to 2 */
3434                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
3435                 if (!exp)
3436                         return -ENOENT;
3437
3438                 if (cda[CTA_EXPECT_ID]) {
3439                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3440                         if (ntohl(id) != (u32)(unsigned long)exp) {
3441                                 nf_ct_expect_put(exp);
3442                                 return -ENOENT;
3443                         }
3444                 }
3445
3446                 /* after list removal, usage count == 1 */
3447                 spin_lock_bh(&nf_conntrack_expect_lock);
3448                 if (del_timer(&exp->timeout)) {
3449                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3450                                                    nlmsg_report(nlh));
3451                         nf_ct_expect_put(exp);
3452                 }
3453                 spin_unlock_bh(&nf_conntrack_expect_lock);
3454                 /* have to put what we 'get' above.
3455                  * after this line usage count == 0 */
3456                 nf_ct_expect_put(exp);
3457         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3458                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3459
3460                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
3461                                          NETLINK_CB(skb).portid,
3462                                          nlmsg_report(nlh));
3463         } else {
3464                 /* This basically means we have to flush everything*/
3465                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3466                                          NETLINK_CB(skb).portid,
3467                                          nlmsg_report(nlh));
3468         }
3469
3470         return 0;
3471 }
3472 static int
3473 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3474                         const struct nlattr * const cda[])
3475 {
3476         if (cda[CTA_EXPECT_TIMEOUT]) {
3477                 if (!del_timer(&x->timeout))
3478                         return -ETIME;
3479
3480                 x->timeout.expires = jiffies +
3481                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3482                 add_timer(&x->timeout);
3483         }
3484         return 0;
3485 }
3486
3487 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3488         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3489         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3490 };
3491
3492 static int
3493 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3494                            struct nf_conntrack_expect *exp,
3495                            u_int8_t u3)
3496 {
3497 #if IS_ENABLED(CONFIG_NF_NAT)
3498         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3499         struct nf_conntrack_tuple nat_tuple = {};
3500         int err;
3501
3502         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3503                                           exp_nat_nla_policy, NULL);
3504         if (err < 0)
3505                 return err;
3506
3507         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3508                 return -EINVAL;
3509
3510         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3511                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3512                                     u3, NULL);
3513         if (err < 0)
3514                 return err;
3515
3516         exp->saved_addr = nat_tuple.src.u3;
3517         exp->saved_proto = nat_tuple.src.u;
3518         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3519
3520         return 0;
3521 #else
3522         return -EOPNOTSUPP;
3523 #endif
3524 }
3525
3526 static struct nf_conntrack_expect *
3527 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3528                        struct nf_conntrack_helper *helper,
3529                        struct nf_conntrack_tuple *tuple,
3530                        struct nf_conntrack_tuple *mask)
3531 {
3532         u_int32_t class = 0;
3533         struct nf_conntrack_expect *exp;
3534         struct nf_conn_help *help;
3535         int err;
3536
3537         help = nfct_help(ct);
3538         if (!help)
3539                 return ERR_PTR(-EOPNOTSUPP);
3540
3541         if (cda[CTA_EXPECT_CLASS] && helper) {
3542                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3543                 if (class > helper->expect_class_max)
3544                         return ERR_PTR(-EINVAL);
3545         }
3546         exp = nf_ct_expect_alloc(ct);
3547         if (!exp)
3548                 return ERR_PTR(-ENOMEM);
3549
3550         if (cda[CTA_EXPECT_FLAGS]) {
3551                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3552                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3553         } else {
3554                 exp->flags = 0;
3555         }
3556         if (cda[CTA_EXPECT_FN]) {
3557                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3558                 struct nf_ct_helper_expectfn *expfn;
3559
3560                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3561                 if (expfn == NULL) {
3562                         err = -EINVAL;
3563                         goto err_out;
3564                 }
3565                 exp->expectfn = expfn->expectfn;
3566         } else
3567                 exp->expectfn = NULL;
3568
3569         exp->class = class;
3570         exp->master = ct;
3571         exp->helper = helper;
3572         exp->tuple = *tuple;
3573         exp->mask.src.u3 = mask->src.u3;
3574         exp->mask.src.u.all = mask->src.u.all;
3575
3576         if (cda[CTA_EXPECT_NAT]) {
3577                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3578                                                  exp, nf_ct_l3num(ct));
3579                 if (err < 0)
3580                         goto err_out;
3581         }
3582         return exp;
3583 err_out:
3584         nf_ct_expect_put(exp);
3585         return ERR_PTR(err);
3586 }
3587
3588 static int
3589 ctnetlink_create_expect(struct net *net,
3590                         const struct nf_conntrack_zone *zone,
3591                         const struct nlattr * const cda[],
3592                         u_int8_t u3, u32 portid, int report)
3593 {
3594         struct nf_conntrack_tuple tuple, mask, master_tuple;
3595         struct nf_conntrack_tuple_hash *h = NULL;
3596         struct nf_conntrack_helper *helper = NULL;
3597         struct nf_conntrack_expect *exp;
3598         struct nf_conn *ct;
3599         int err;
3600
3601         /* caller guarantees that those three CTA_EXPECT_* exist */
3602         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3603                                     u3, NULL);
3604         if (err < 0)
3605                 return err;
3606         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3607                                     u3, NULL);
3608         if (err < 0)
3609                 return err;
3610         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3611                                     u3, NULL);
3612         if (err < 0)
3613                 return err;
3614
3615         /* Look for master conntrack of this expectation */
3616         h = nf_conntrack_find_get(net, zone, &master_tuple);
3617         if (!h)
3618                 return -ENOENT;
3619         ct = nf_ct_tuplehash_to_ctrack(h);
3620
3621         rcu_read_lock();
3622         if (cda[CTA_EXPECT_HELP_NAME]) {
3623                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3624
3625                 helper = __nf_conntrack_helper_find(helpname, u3,
3626                                                     nf_ct_protonum(ct));
3627                 if (helper == NULL) {
3628                         rcu_read_unlock();
3629 #ifdef CONFIG_MODULES
3630                         if (request_module("nfct-helper-%s", helpname) < 0) {
3631                                 err = -EOPNOTSUPP;
3632                                 goto err_ct;
3633                         }
3634                         rcu_read_lock();
3635                         helper = __nf_conntrack_helper_find(helpname, u3,
3636                                                             nf_ct_protonum(ct));
3637                         if (helper) {
3638                                 err = -EAGAIN;
3639                                 goto err_rcu;
3640                         }
3641                         rcu_read_unlock();
3642 #endif
3643                         err = -EOPNOTSUPP;
3644                         goto err_ct;
3645                 }
3646         }
3647
3648         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3649         if (IS_ERR(exp)) {
3650                 err = PTR_ERR(exp);
3651                 goto err_rcu;
3652         }
3653
3654         err = nf_ct_expect_related_report(exp, portid, report, 0);
3655         nf_ct_expect_put(exp);
3656 err_rcu:
3657         rcu_read_unlock();
3658 err_ct:
3659         nf_ct_put(ct);
3660         return err;
3661 }
3662
3663 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3664                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3665                                 const struct nlattr * const cda[],
3666                                 struct netlink_ext_ack *extack)
3667 {
3668         struct nf_conntrack_tuple tuple;
3669         struct nf_conntrack_expect *exp;
3670         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3671         u_int8_t u3 = nfmsg->nfgen_family;
3672         struct nf_conntrack_zone zone;
3673         int err;
3674
3675         if (!cda[CTA_EXPECT_TUPLE]
3676             || !cda[CTA_EXPECT_MASK]
3677             || !cda[CTA_EXPECT_MASTER])
3678                 return -EINVAL;
3679
3680         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3681         if (err < 0)
3682                 return err;
3683
3684         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3685                                     u3, NULL);
3686         if (err < 0)
3687                 return err;
3688
3689         spin_lock_bh(&nf_conntrack_expect_lock);
3690         exp = __nf_ct_expect_find(net, &zone, &tuple);
3691         if (!exp) {
3692                 spin_unlock_bh(&nf_conntrack_expect_lock);
3693                 err = -ENOENT;
3694                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3695                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3696                                                       NETLINK_CB(skb).portid,
3697                                                       nlmsg_report(nlh));
3698                 }
3699                 return err;
3700         }
3701
3702         err = -EEXIST;
3703         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3704                 err = ctnetlink_change_expect(exp, cda);
3705         spin_unlock_bh(&nf_conntrack_expect_lock);
3706
3707         return err;
3708 }
3709
3710 static int
3711 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3712                              const struct ip_conntrack_stat *st)
3713 {
3714         struct nlmsghdr *nlh;
3715         struct nfgenmsg *nfmsg;
3716         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3717
3718         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3719                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3720         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3721         if (nlh == NULL)
3722                 goto nlmsg_failure;
3723
3724         nfmsg = nlmsg_data(nlh);
3725         nfmsg->nfgen_family = AF_UNSPEC;
3726         nfmsg->version      = NFNETLINK_V0;
3727         nfmsg->res_id       = htons(cpu);
3728
3729         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3730             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3731             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3732                 goto nla_put_failure;
3733
3734         nlmsg_end(skb, nlh);
3735         return skb->len;
3736
3737 nla_put_failure:
3738 nlmsg_failure:
3739         nlmsg_cancel(skb, nlh);
3740         return -1;
3741 }
3742
3743 static int
3744 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3745 {
3746         int cpu;
3747         struct net *net = sock_net(skb->sk);
3748
3749         if (cb->args[0] == nr_cpu_ids)
3750                 return 0;
3751
3752         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3753                 const struct ip_conntrack_stat *st;
3754
3755                 if (!cpu_possible(cpu))
3756                         continue;
3757
3758                 st = per_cpu_ptr(net->ct.stat, cpu);
3759                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3760                                                  cb->nlh->nlmsg_seq,
3761                                                  cpu, st) < 0)
3762                         break;
3763         }
3764         cb->args[0] = cpu;
3765
3766         return skb->len;
3767 }
3768
3769 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3770                                   struct sk_buff *skb,
3771                                   const struct nlmsghdr *nlh,
3772                                   const struct nlattr * const cda[],
3773                                   struct netlink_ext_ack *extack)
3774 {
3775         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3776                 struct netlink_dump_control c = {
3777                         .dump = ctnetlink_exp_stat_cpu_dump,
3778                 };
3779                 return netlink_dump_start(ctnl, skb, nlh, &c);
3780         }
3781
3782         return 0;
3783 }
3784
3785 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3786 static struct nf_ct_event_notifier ctnl_notifier = {
3787         .fcn = ctnetlink_conntrack_event,
3788 };
3789
3790 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3791         .fcn = ctnetlink_expect_event,
3792 };
3793 #endif
3794
3795 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3796         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3797                                             .attr_count = CTA_MAX,
3798                                             .policy = ct_nla_policy },
3799         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3800                                             .attr_count = CTA_MAX,
3801                                             .policy = ct_nla_policy },
3802         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3803                                             .attr_count = CTA_MAX,
3804                                             .policy = ct_nla_policy },
3805         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3806                                             .attr_count = CTA_MAX,
3807                                             .policy = ct_nla_policy },
3808         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3809         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3810         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3811         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3812 };
3813
3814 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3815         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3816                                             .attr_count = CTA_EXPECT_MAX,
3817                                             .policy = exp_nla_policy },
3818         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3819                                             .attr_count = CTA_EXPECT_MAX,
3820                                             .policy = exp_nla_policy },
3821         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3822                                             .attr_count = CTA_EXPECT_MAX,
3823                                             .policy = exp_nla_policy },
3824         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3825 };
3826
3827 static const struct nfnetlink_subsystem ctnl_subsys = {
3828         .name                           = "conntrack",
3829         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3830         .cb_count                       = IPCTNL_MSG_MAX,
3831         .cb                             = ctnl_cb,
3832 };
3833
3834 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3835         .name                           = "conntrack_expect",
3836         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3837         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3838         .cb                             = ctnl_exp_cb,
3839 };
3840
3841 MODULE_ALIAS("ip_conntrack_netlink");
3842 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3843 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3844
3845 static int __net_init ctnetlink_net_init(struct net *net)
3846 {
3847 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3848         int ret;
3849
3850         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3851         if (ret < 0) {
3852                 pr_err("ctnetlink_init: cannot register notifier.\n");
3853                 goto err_out;
3854         }
3855
3856         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3857         if (ret < 0) {
3858                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3859                 goto err_unreg_notifier;
3860         }
3861 #endif
3862         return 0;
3863
3864 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3865 err_unreg_notifier:
3866         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3867 err_out:
3868         return ret;
3869 #endif
3870 }
3871
3872 static void ctnetlink_net_exit(struct net *net)
3873 {
3874 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3875         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3876         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3877 #endif
3878 }
3879
3880 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3881 {
3882         struct net *net;
3883
3884         list_for_each_entry(net, net_exit_list, exit_list)
3885                 ctnetlink_net_exit(net);
3886
3887         /* wait for other cpus until they are done with ctnl_notifiers */
3888         synchronize_rcu();
3889 }
3890
3891 static struct pernet_operations ctnetlink_net_ops = {
3892         .init           = ctnetlink_net_init,
3893         .exit_batch     = ctnetlink_net_exit_batch,
3894 };
3895
3896 static int __init ctnetlink_init(void)
3897 {
3898         int ret;
3899
3900         ret = nfnetlink_subsys_register(&ctnl_subsys);
3901         if (ret < 0) {
3902                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3903                 goto err_out;
3904         }
3905
3906         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3907         if (ret < 0) {
3908                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3909                 goto err_unreg_subsys;
3910         }
3911
3912         ret = register_pernet_subsys(&ctnetlink_net_ops);
3913         if (ret < 0) {
3914                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3915                 goto err_unreg_exp_subsys;
3916         }
3917 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3918         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3919         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3920 #endif
3921         return 0;
3922
3923 err_unreg_exp_subsys:
3924         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3925 err_unreg_subsys:
3926         nfnetlink_subsys_unregister(&ctnl_subsys);
3927 err_out:
3928         return ret;
3929 }
3930
3931 static void __exit ctnetlink_exit(void)
3932 {
3933         unregister_pernet_subsys(&ctnetlink_net_ops);
3934         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3935         nfnetlink_subsys_unregister(&ctnl_subsys);
3936 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3937         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3938 #endif
3939         synchronize_rcu();
3940 }
3941
3942 module_init(ctnetlink_init);
3943 module_exit(ctnetlink_exit);