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         u_int32_t cta_flags;
 855         u8 family;
 856
 857         u_int32_t orig_flags;
 858         u_int32_t reply_flags;
 859
 860         struct nf_conntrack_tuple orig;
 861         struct nf_conntrack_tuple reply;
 862         struct nf_conntrack_zone zone;
 863
 864         struct {
 865                 u_int32_t val;
 866                 u_int32_t mask;
 867         } mark;
 868 };
 869
 870 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 871         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 872         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 873 };
 874
 875 static int ctnetlink_parse_filter(const struct nlattr *attr,
 876                                   struct ctnetlink_filter *filter)
 877 {
 878         struct nlattr *tb[CTA_FILTER_MAX + 1];
 879         int ret = 0;
 880
 881         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 882                                NULL);
 883         if (ret)
 884                 return ret;
 885
 886         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 887                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 888                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 889                         return -EOPNOTSUPP;
 890         }
 891
 892         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 893                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 894                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 895                         return -EOPNOTSUPP;
 896         }
 897
 898         return 0;
 899 }
 900
 901 static int ctnetlink_parse_zone(const struct nlattr *attr,
 902                                 struct nf_conntrack_zone *zone);
 903 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 904                                          struct nf_conntrack_tuple *tuple,
 905                                          u32 type, u_int8_t l3num,
 906                                          struct nf_conntrack_zone *zone,
 907                                          u_int32_t flags);
 908
 909 /* applied on filters */
 910 #define CTA_FILTER_F_CTA_MARK                   (1 << 0)
 911 #define CTA_FILTER_F_CTA_MARK_MASK              (1 << 1)
 912
 913 static struct ctnetlink_filter *
 914 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 915 {
 916         struct ctnetlink_filter *filter;
 917         int err;
 918
 919 #ifndef CONFIG_NF_CONNTRACK_MARK
 920         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 921                 return ERR_PTR(-EOPNOTSUPP);
 922 #endif
 923
 924         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 925         if (filter == NULL)
 926                 return ERR_PTR(-ENOMEM);
 927
 928         filter->family = family;
 929
 930 #ifdef CONFIG_NF_CONNTRACK_MARK
 931         if (cda[CTA_MARK]) {
 932                 filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 933                 filter->cta_flags |= CTA_FILTER_FLAG(CTA_MARK);
 934
 935                 if (cda[CTA_MARK_MASK]) {
 936                         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 937                         filter->cta_flags |= CTA_FILTER_FLAG(CTA_MARK_MASK);
 938                 } else {
 939                         filter->mark.mask = 0xffffffff;
 940                 }
 941         } else if (cda[CTA_MARK_MASK]) {
 942                 err = -EINVAL;
 943                 goto err_filter;
 944         }
 945 #endif
 946         if (!cda[CTA_FILTER])
 947                 return filter;
 948
 949         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 950         if (err < 0)
 951                 goto err_filter;
 952
 953         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
 954         if (err < 0)
 955                 goto err_filter;
 956
 957         if (filter->orig_flags) {
 958                 if (!cda[CTA_TUPLE_ORIG]) {
 959                         err = -EINVAL;
 960                         goto err_filter;
 961                 }
 962
 963                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
 964                                                    CTA_TUPLE_ORIG,
 965                                                    filter->family,
 966                                                    &filter->zone,
 967                                                    filter->orig_flags);
 968                 if (err < 0)
 969                         goto err_filter;
 970         }
 971
 972         if (filter->reply_flags) {
 973                 if (!cda[CTA_TUPLE_REPLY]) {
 974                         err = -EINVAL;
 975                         goto err_filter;
 976                 }
 977
 978                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
 979                                                    CTA_TUPLE_REPLY,
 980                                                    filter->family,
 981                                                    &filter->zone,
 982                                                    filter->orig_flags);
 983                 if (err < 0) {
 984                         err = -EINVAL;
 985                         goto err_filter;
 986                 }
 987         }
 988
 989         return filter;
 990
 991 err_filter:
 992         kfree(filter);
 993
 994         return ERR_PTR(err);
 995 }
 996
 997 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
 998 {
 999         return family || cda[CTA_MARK] || cda[CTA_FILTER];
1000 }
1001
1002 static int ctnetlink_start(struct netlink_callback *cb)
1003 {
1004         const struct nlattr * const *cda = cb->data;
1005         struct ctnetlink_filter *filter = NULL;
1006         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1007         u8 family = nfmsg->nfgen_family;
1008
1009         if (ctnetlink_needs_filter(family, cda)) {
1010                 filter = ctnetlink_alloc_filter(cda, family);
1011                 if (IS_ERR(filter))
1012                         return PTR_ERR(filter);
1013         }
1014
1015         cb->data = filter;
1016         return 0;
1017 }
1018
1019 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1020                                         struct nf_conntrack_tuple *ct_tuple,
1021                                         u_int32_t flags, int family)
1022 {
1023         switch (family) {
1024         case NFPROTO_IPV4:
1025                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1026                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1027                         return  0;
1028
1029                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1030                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1031                         return  0;
1032                 break;
1033         case NFPROTO_IPV6:
1034                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1035                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1036                                    &ct_tuple->src.u3.in6))
1037                         return 0;
1038
1039                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1040                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1041                                    &ct_tuple->dst.u3.in6))
1042                         return 0;
1043                 break;
1044         }
1045
1046         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1047             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1048                 return 0;
1049
1050         switch (ct_tuple->dst.protonum) {
1051         case IPPROTO_TCP:
1052         case IPPROTO_UDP:
1053                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1054                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1055                         return 0;
1056
1057                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1058                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1059                         return 0;
1060                 break;
1061         case IPPROTO_ICMP:
1062                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1063                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1064                         return 0;
1065                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1066                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1067                         return 0;
1068                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1069                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1070                         return 0;
1071                 break;
1072         case IPPROTO_ICMPV6:
1073                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1074                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1075                         return 0;
1076                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1077                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1078                         return 0;
1079                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1080                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1081                         return 0;
1082                 break;
1083         }
1084
1085         return 1;
1086 }
1087
1088 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1089 {
1090         struct ctnetlink_filter *filter = data;
1091         struct nf_conntrack_tuple *tuple;
1092
1093         if (filter == NULL)
1094                 goto out;
1095
1096         /* Match entries of a given L3 protocol number.
1097          * If it is not specified, ie. l3proto == 0,
1098          * then match everything.
1099          */
1100         if (filter->family && nf_ct_l3num(ct) != filter->family)
1101                 goto ignore_entry;
1102
1103         if (filter->orig_flags) {
1104                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1105                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1106                                                   filter->orig_flags,
1107                                                   filter->family))
1108                         goto ignore_entry;
1109         }
1110
1111         if (filter->reply_flags) {
1112                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1113                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1114                                                   filter->reply_flags,
1115                                                   filter->family))
1116                         goto ignore_entry;
1117         }
1118
1119 #ifdef CONFIG_NF_CONNTRACK_MARK
1120         if ((filter->cta_flags & CTA_FILTER_FLAG(CTA_MARK_MASK)) &&
1121             (ct->mark & filter->mark.mask) != filter->mark.val)
1122                 goto ignore_entry;
1123         else if ((filter->cta_flags & CTA_FILTER_FLAG(CTA_MARK)) &&
1124                  ct->mark != filter->mark.val)
1125                 goto ignore_entry;
1126 #endif
1127
1128 out:
1129         return 1;
1130
1131 ignore_entry:
1132         return 0;
1133 }
1134
1135 static int
1136 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1137 {
1138         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1139         struct net *net = sock_net(skb->sk);
1140         struct nf_conn *ct, *last;
1141         struct nf_conntrack_tuple_hash *h;
1142         struct hlist_nulls_node *n;
1143         struct nf_conn *nf_ct_evict[8];
1144         int res, i;
1145         spinlock_t *lockp;
1146
1147         last = (struct nf_conn *)cb->args[1];
1148         i = 0;
1149
1150         local_bh_disable();
1151         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1152 restart:
1153                 while (i) {
1154                         i--;
1155                         if (nf_ct_should_gc(nf_ct_evict[i]))
1156                                 nf_ct_kill(nf_ct_evict[i]);
1157                         nf_ct_put(nf_ct_evict[i]);
1158                 }
1159
1160                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1161                 nf_conntrack_lock(lockp);
1162                 if (cb->args[0] >= nf_conntrack_htable_size) {
1163                         spin_unlock(lockp);
1164                         goto out;
1165                 }
1166                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1167                                            hnnode) {
1168                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1169                                 continue;
1170                         ct = nf_ct_tuplehash_to_ctrack(h);
1171                         if (nf_ct_is_expired(ct)) {
1172                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1173                                     atomic_inc_not_zero(&ct->ct_general.use))
1174                                         nf_ct_evict[i++] = ct;
1175                                 continue;
1176                         }
1177
1178                         if (!net_eq(net, nf_ct_net(ct)))
1179                                 continue;
1180
1181                         if (cb->args[1]) {
1182                                 if (ct != last)
1183                                         continue;
1184                                 cb->args[1] = 0;
1185                         }
1186                         if (!ctnetlink_filter_match(ct, cb->data))
1187                                 continue;
1188
1189                         res =
1190                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1191                                             cb->nlh->nlmsg_seq,
1192                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1193                                             ct, true, flags);
1194                         if (res < 0) {
1195                                 nf_conntrack_get(&ct->ct_general);
1196                                 cb->args[1] = (unsigned long)ct;
1197                                 spin_unlock(lockp);
1198                                 goto out;
1199                         }
1200                 }
1201                 spin_unlock(lockp);
1202                 if (cb->args[1]) {
1203                         cb->args[1] = 0;
1204                         goto restart;
1205                 }
1206         }
1207 out:
1208         local_bh_enable();
1209         if (last) {
1210                 /* nf ct hash resize happened, now clear the leftover. */
1211                 if ((struct nf_conn *)cb->args[1] == last)
1212                         cb->args[1] = 0;
1213
1214                 nf_ct_put(last);
1215         }
1216
1217         while (i) {
1218                 i--;
1219                 if (nf_ct_should_gc(nf_ct_evict[i]))
1220                         nf_ct_kill(nf_ct_evict[i]);
1221                 nf_ct_put(nf_ct_evict[i]);
1222         }
1223
1224         return skb->len;
1225 }
1226
1227 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1228                                 struct nf_conntrack_tuple *t,
1229                                 u_int32_t flags)
1230 {
1231         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1232                 if (!tb[CTA_IP_V4_SRC])
1233                         return -EINVAL;
1234
1235                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1236         }
1237
1238         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1239                 if (!tb[CTA_IP_V4_DST])
1240                         return -EINVAL;
1241
1242                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1243         }
1244
1245         return 0;
1246 }
1247
1248 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1249                                 struct nf_conntrack_tuple *t,
1250                                 u_int32_t flags)
1251 {
1252         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1253                 if (!tb[CTA_IP_V6_SRC])
1254                         return -EINVAL;
1255
1256                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1257         }
1258
1259         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1260                 if (!tb[CTA_IP_V6_DST])
1261                         return -EINVAL;
1262
1263                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1264         }
1265
1266         return 0;
1267 }
1268
1269 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1270                                     struct nf_conntrack_tuple *tuple,
1271                                     u_int32_t flags)
1272 {
1273         struct nlattr *tb[CTA_IP_MAX+1];
1274         int ret = 0;
1275
1276         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1277         if (ret < 0)
1278                 return ret;
1279
1280         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1281                                              cta_ip_nla_policy, NULL);
1282         if (ret)
1283                 return ret;
1284
1285         switch (tuple->src.l3num) {
1286         case NFPROTO_IPV4:
1287                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1288                 break;
1289         case NFPROTO_IPV6:
1290                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1291                 break;
1292         }
1293
1294         return ret;
1295 }
1296
1297 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1298         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1299 };
1300
1301 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1302                                        struct nf_conntrack_tuple *tuple,
1303                                        u_int32_t flags)
1304 {
1305         const struct nf_conntrack_l4proto *l4proto;
1306         struct nlattr *tb[CTA_PROTO_MAX+1];
1307         int ret = 0;
1308
1309         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1310                                           proto_nla_policy, NULL);
1311         if (ret < 0)
1312                 return ret;
1313
1314         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1315                 return 0;
1316
1317         if (!tb[CTA_PROTO_NUM])
1318                 return -EINVAL;
1319
1320         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1321
1322         rcu_read_lock();
1323         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1324
1325         if (likely(l4proto->nlattr_to_tuple)) {
1326                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1327                                                      l4proto->nla_policy,
1328                                                      NULL);
1329                 if (ret == 0)
1330                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1331         }
1332
1333         rcu_read_unlock();
1334
1335         return ret;
1336 }
1337
1338 static int
1339 ctnetlink_parse_zone(const struct nlattr *attr,
1340                      struct nf_conntrack_zone *zone)
1341 {
1342         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1343                         NF_CT_DEFAULT_ZONE_DIR, 0);
1344 #ifdef CONFIG_NF_CONNTRACK_ZONES
1345         if (attr)
1346                 zone->id = ntohs(nla_get_be16(attr));
1347 #else
1348         if (attr)
1349                 return -EOPNOTSUPP;
1350 #endif
1351         return 0;
1352 }
1353
1354 static int
1355 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1356                            struct nf_conntrack_zone *zone)
1357 {
1358         int ret;
1359
1360         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1361                 return -EINVAL;
1362
1363         ret = ctnetlink_parse_zone(attr, zone);
1364         if (ret < 0)
1365                 return ret;
1366
1367         if (type == CTA_TUPLE_REPLY)
1368                 zone->dir = NF_CT_ZONE_DIR_REPL;
1369         else
1370                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1371
1372         return 0;
1373 }
1374
1375 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1376         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1377         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1378         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1379 };
1380
1381 #define CTA_FILTER_F_ALL_CTA_PROTO \
1382   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1383    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1384    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1385    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1386    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1387    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1388    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1389    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1390
1391 static int
1392 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1393                               struct nf_conntrack_tuple *tuple, u32 type,
1394                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1395                               u_int32_t flags)
1396 {
1397         struct nlattr *tb[CTA_TUPLE_MAX+1];
1398         int err;
1399
1400         memset(tuple, 0, sizeof(*tuple));
1401
1402         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1403                                           tuple_nla_policy, NULL);
1404         if (err < 0)
1405                 return err;
1406
1407
1408         tuple->src.l3num = l3num;
1409
1410         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1411             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1412                 if (!tb[CTA_TUPLE_IP])
1413                         return -EINVAL;
1414
1415                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1416                 if (err < 0)
1417                         return err;
1418         }
1419
1420         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1421                 if (!tb[CTA_TUPLE_PROTO])
1422                         return -EINVAL;
1423
1424                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1425                 if (err < 0)
1426                         return err;
1427         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1428                 /* Can't manage proto flags without a protonum  */
1429                 return -EINVAL;
1430         }
1431
1432         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1433                 if (!zone)
1434                         return -EINVAL;
1435
1436                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1437                                                  type, zone);
1438                 if (err < 0)
1439                         return err;
1440         }
1441
1442         /* orig and expect tuples get DIR_ORIGINAL */
1443         if (type == CTA_TUPLE_REPLY)
1444                 tuple->dst.dir = IP_CT_DIR_REPLY;
1445         else
1446                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1447
1448         return 0;
1449 }
1450
1451 static int
1452 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1453                       struct nf_conntrack_tuple *tuple, u32 type,
1454                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1455 {
1456         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1457                                             CTA_FILTER_FLAG(ALL));
1458 }
1459
1460 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1461         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1462                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1463 };
1464
1465 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1466                                 struct nlattr **helpinfo)
1467 {
1468         int err;
1469         struct nlattr *tb[CTA_HELP_MAX+1];
1470
1471         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1472                                           help_nla_policy, NULL);
1473         if (err < 0)
1474                 return err;
1475
1476         if (!tb[CTA_HELP_NAME])
1477                 return -EINVAL;
1478
1479         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1480
1481         if (tb[CTA_HELP_INFO])
1482                 *helpinfo = tb[CTA_HELP_INFO];
1483
1484         return 0;
1485 }
1486
1487 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1488         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1489         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1490         [CTA_STATUS]            = { .type = NLA_U32 },
1491         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1492         [CTA_HELP]              = { .type = NLA_NESTED },
1493         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1494         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1495         [CTA_MARK]              = { .type = NLA_U32 },
1496         [CTA_ID]                = { .type = NLA_U32 },
1497         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1498         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1499         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1500         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1501         [CTA_ZONE]              = { .type = NLA_U16 },
1502         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1503         [CTA_LABELS]            = { .type = NLA_BINARY,
1504                                     .len = NF_CT_LABELS_MAX_SIZE },
1505         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1506                                     .len = NF_CT_LABELS_MAX_SIZE },
1507         [CTA_FILTER]            = { .type = NLA_NESTED },
1508 };
1509
1510 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1511 {
1512         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1513                 return 0;
1514
1515         return ctnetlink_filter_match(ct, data);
1516 }
1517
1518 static int ctnetlink_flush_conntrack(struct net *net,
1519                                      const struct nlattr * const cda[],
1520                                      u32 portid, int report, u8 family)
1521 {
1522         struct ctnetlink_filter *filter = NULL;
1523
1524         if (ctnetlink_needs_filter(family, cda)) {
1525                 if (cda[CTA_FILTER])
1526                         return -EOPNOTSUPP;
1527
1528                 filter = ctnetlink_alloc_filter(cda, family);
1529                 if (IS_ERR(filter))
1530                         return PTR_ERR(filter);
1531         }
1532
1533         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1534                                   portid, report);
1535         kfree(filter);
1536
1537         return 0;
1538 }
1539
1540 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1541                                    struct sk_buff *skb,
1542                                    const struct nlmsghdr *nlh,
1543                                    const struct nlattr * const cda[],
1544                                    struct netlink_ext_ack *extack)
1545 {
1546         struct nf_conntrack_tuple_hash *h;
1547         struct nf_conntrack_tuple tuple;
1548         struct nf_conn *ct;
1549         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1550         struct nf_conntrack_zone zone;
1551         int err;
1552
1553         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1554         if (err < 0)
1555                 return err;
1556
1557         if (cda[CTA_TUPLE_ORIG])
1558                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1559                                             nfmsg->nfgen_family, &zone);
1560         else if (cda[CTA_TUPLE_REPLY])
1561                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1562                                             nfmsg->nfgen_family, &zone);
1563         else {
1564                 u_int8_t u3 = nfmsg->version ? nfmsg->nfgen_family : AF_UNSPEC;
1565
1566                 return ctnetlink_flush_conntrack(net, cda,
1567                                                  NETLINK_CB(skb).portid,
1568                                                  nlmsg_report(nlh), u3);
1569         }
1570
1571         if (err < 0)
1572                 return err;
1573
1574         h = nf_conntrack_find_get(net, &zone, &tuple);
1575         if (!h)
1576                 return -ENOENT;
1577
1578         ct = nf_ct_tuplehash_to_ctrack(h);
1579
1580         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1581                 nf_ct_put(ct);
1582                 return -EBUSY;
1583         }
1584
1585         if (cda[CTA_ID]) {
1586                 __be32 id = nla_get_be32(cda[CTA_ID]);
1587
1588                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1589                         nf_ct_put(ct);
1590                         return -ENOENT;
1591                 }
1592         }
1593
1594         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1595         nf_ct_put(ct);
1596
1597         return 0;
1598 }
1599
1600 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1601                                    struct sk_buff *skb,
1602                                    const struct nlmsghdr *nlh,
1603                                    const struct nlattr * const cda[],
1604                                    struct netlink_ext_ack *extack)
1605 {
1606         struct nf_conntrack_tuple_hash *h;
1607         struct nf_conntrack_tuple tuple;
1608         struct nf_conn *ct;
1609         struct sk_buff *skb2 = NULL;
1610         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1611         u_int8_t u3 = nfmsg->nfgen_family;
1612         struct nf_conntrack_zone zone;
1613         int err;
1614
1615         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1616                 struct netlink_dump_control c = {
1617                         .start = ctnetlink_start,
1618                         .dump = ctnetlink_dump_table,
1619                         .done = ctnetlink_done,
1620                         .data = (void *)cda,
1621                 };
1622
1623                 return netlink_dump_start(ctnl, skb, nlh, &c);
1624         }
1625
1626         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1627         if (err < 0)
1628                 return err;
1629
1630         if (cda[CTA_TUPLE_ORIG])
1631                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1632                                             u3, &zone);
1633         else if (cda[CTA_TUPLE_REPLY])
1634                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1635                                             u3, &zone);
1636         else
1637                 return -EINVAL;
1638
1639         if (err < 0)
1640                 return err;
1641
1642         h = nf_conntrack_find_get(net, &zone, &tuple);
1643         if (!h)
1644                 return -ENOENT;
1645
1646         ct = nf_ct_tuplehash_to_ctrack(h);
1647
1648         err = -ENOMEM;
1649         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1650         if (skb2 == NULL) {
1651                 nf_ct_put(ct);
1652                 return -ENOMEM;
1653         }
1654
1655         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1656                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct, true, 0);
1657         nf_ct_put(ct);
1658         if (err <= 0)
1659                 goto free;
1660
1661         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1662         if (err < 0)
1663                 goto out;
1664
1665         return 0;
1666
1667 free:
1668         kfree_skb(skb2);
1669 out:
1670         /* this avoids a loop in nfnetlink. */
1671         return err == -EAGAIN ? -ENOBUFS : err;
1672 }
1673
1674 static int ctnetlink_done_list(struct netlink_callback *cb)
1675 {
1676         if (cb->args[1])
1677                 nf_ct_put((struct nf_conn *)cb->args[1]);
1678         return 0;
1679 }
1680
1681 static int
1682 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1683 {
1684         struct nf_conn *ct, *last;
1685         struct nf_conntrack_tuple_hash *h;
1686         struct hlist_nulls_node *n;
1687         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1688         u_int8_t l3proto = nfmsg->nfgen_family;
1689         int res;
1690         int cpu;
1691         struct hlist_nulls_head *list;
1692         struct net *net = sock_net(skb->sk);
1693
1694         if (cb->args[2])
1695                 return 0;
1696
1697         last = (struct nf_conn *)cb->args[1];
1698
1699         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1700                 struct ct_pcpu *pcpu;
1701
1702                 if (!cpu_possible(cpu))
1703                         continue;
1704
1705                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1706                 spin_lock_bh(&pcpu->lock);
1707                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1708 restart:
1709                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1710                         ct = nf_ct_tuplehash_to_ctrack(h);
1711                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1712                                 continue;
1713                         if (cb->args[1]) {
1714                                 if (ct != last)
1715                                         continue;
1716                                 cb->args[1] = 0;
1717                         }
1718
1719                         /* We can't dump extension info for the unconfirmed
1720                          * list because unconfirmed conntracks can have
1721                          * ct->ext reallocated (and thus freed).
1722                          *
1723                          * In the dying list case ct->ext can't be free'd
1724                          * until after we drop pcpu->lock.
1725                          */
1726                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1727                                                   cb->nlh->nlmsg_seq,
1728                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1729                                                   ct, dying ? true : false, 0);
1730                         if (res < 0) {
1731                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1732                                         continue;
1733                                 cb->args[0] = cpu;
1734                                 cb->args[1] = (unsigned long)ct;
1735                                 spin_unlock_bh(&pcpu->lock);
1736                                 goto out;
1737                         }
1738                 }
1739                 if (cb->args[1]) {
1740                         cb->args[1] = 0;
1741                         goto restart;
1742                 }
1743                 spin_unlock_bh(&pcpu->lock);
1744         }
1745         cb->args[2] = 1;
1746 out:
1747         if (last)
1748                 nf_ct_put(last);
1749
1750         return skb->len;
1751 }
1752
1753 static int
1754 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1755 {
1756         return ctnetlink_dump_list(skb, cb, true);
1757 }
1758
1759 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1760                                   struct sk_buff *skb,
1761                                   const struct nlmsghdr *nlh,
1762                                   const struct nlattr * const cda[],
1763                                   struct netlink_ext_ack *extack)
1764 {
1765         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1766                 struct netlink_dump_control c = {
1767                         .dump = ctnetlink_dump_dying,
1768                         .done = ctnetlink_done_list,
1769                 };
1770                 return netlink_dump_start(ctnl, skb, nlh, &c);
1771         }
1772
1773         return -EOPNOTSUPP;
1774 }
1775
1776 static int
1777 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1778 {
1779         return ctnetlink_dump_list(skb, cb, false);
1780 }
1781
1782 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1783                                         struct sk_buff *skb,
1784                                         const struct nlmsghdr *nlh,
1785                                         const struct nlattr * const cda[],
1786                                         struct netlink_ext_ack *extack)
1787 {
1788         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1789                 struct netlink_dump_control c = {
1790                         .dump = ctnetlink_dump_unconfirmed,
1791                         .done = ctnetlink_done_list,
1792                 };
1793                 return netlink_dump_start(ctnl, skb, nlh, &c);
1794         }
1795
1796         return -EOPNOTSUPP;
1797 }
1798
1799 #if IS_ENABLED(CONFIG_NF_NAT)
1800 static int
1801 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1802                           enum nf_nat_manip_type manip,
1803                           const struct nlattr *attr)
1804         __must_hold(RCU)
1805 {
1806         struct nf_nat_hook *nat_hook;
1807         int err;
1808
1809         nat_hook = rcu_dereference(nf_nat_hook);
1810         if (!nat_hook) {
1811 #ifdef CONFIG_MODULES
1812                 rcu_read_unlock();
1813                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1814                 if (request_module("nf-nat") < 0) {
1815                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1816                         rcu_read_lock();
1817                         return -EOPNOTSUPP;
1818                 }
1819                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1820                 rcu_read_lock();
1821                 nat_hook = rcu_dereference(nf_nat_hook);
1822                 if (nat_hook)
1823                         return -EAGAIN;
1824 #endif
1825                 return -EOPNOTSUPP;
1826         }
1827
1828         err = nat_hook->parse_nat_setup(ct, manip, attr);
1829         if (err == -EAGAIN) {
1830 #ifdef CONFIG_MODULES
1831                 rcu_read_unlock();
1832                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1833                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1834                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1835                         rcu_read_lock();
1836                         return -EOPNOTSUPP;
1837                 }
1838                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1839                 rcu_read_lock();
1840 #else
1841                 err = -EOPNOTSUPP;
1842 #endif
1843         }
1844         return err;
1845 }
1846 #endif
1847
1848 static void
1849 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1850                           unsigned long off)
1851 {
1852         unsigned int bit;
1853
1854         /* Ignore these unchangable bits */
1855         on &= ~IPS_UNCHANGEABLE_MASK;
1856         off &= ~IPS_UNCHANGEABLE_MASK;
1857
1858         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1859                 if (on & (1 << bit))
1860                         set_bit(bit, &ct->status);
1861                 else if (off & (1 << bit))
1862                         clear_bit(bit, &ct->status);
1863         }
1864 }
1865
1866 static int
1867 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1868 {
1869         unsigned long d;
1870         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1871         d = ct->status ^ status;
1872
1873         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1874                 /* unchangeable */
1875                 return -EBUSY;
1876
1877         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1878                 /* SEEN_REPLY bit can only be set */
1879                 return -EBUSY;
1880
1881         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1882                 /* ASSURED bit can only be set */
1883                 return -EBUSY;
1884
1885         __ctnetlink_change_status(ct, status, 0);
1886         return 0;
1887 }
1888
1889 static int
1890 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1891 {
1892 #if IS_ENABLED(CONFIG_NF_NAT)
1893         int ret;
1894
1895         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1896                 return 0;
1897
1898         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1899                                         cda[CTA_NAT_DST]);
1900         if (ret < 0)
1901                 return ret;
1902
1903         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1904                                          cda[CTA_NAT_SRC]);
1905 #else
1906         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1907                 return 0;
1908         return -EOPNOTSUPP;
1909 #endif
1910 }
1911
1912 static int ctnetlink_change_helper(struct nf_conn *ct,
1913                                    const struct nlattr * const cda[])
1914 {
1915         struct nf_conntrack_helper *helper;
1916         struct nf_conn_help *help = nfct_help(ct);
1917         char *helpname = NULL;
1918         struct nlattr *helpinfo = NULL;
1919         int err;
1920
1921         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1922         if (err < 0)
1923                 return err;
1924
1925         /* don't change helper of sibling connections */
1926         if (ct->master) {
1927                 /* If we try to change the helper to the same thing twice,
1928                  * treat the second attempt as a no-op instead of returning
1929                  * an error.
1930                  */
1931                 err = -EBUSY;
1932                 if (help) {
1933                         rcu_read_lock();
1934                         helper = rcu_dereference(help->helper);
1935                         if (helper && !strcmp(helper->name, helpname))
1936                                 err = 0;
1937                         rcu_read_unlock();
1938                 }
1939
1940                 return err;
1941         }
1942
1943         if (!strcmp(helpname, "")) {
1944                 if (help && help->helper) {
1945                         /* we had a helper before ... */
1946                         nf_ct_remove_expectations(ct);
1947                         RCU_INIT_POINTER(help->helper, NULL);
1948                 }
1949
1950                 return 0;
1951         }
1952
1953         rcu_read_lock();
1954         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1955                                             nf_ct_protonum(ct));
1956         if (helper == NULL) {
1957                 rcu_read_unlock();
1958                 return -EOPNOTSUPP;
1959         }
1960
1961         if (help) {
1962                 if (help->helper == helper) {
1963                         /* update private helper data if allowed. */
1964                         if (helper->from_nlattr)
1965                                 helper->from_nlattr(helpinfo, ct);
1966                         err = 0;
1967                 } else
1968                         err = -EBUSY;
1969         } else {
1970                 /* we cannot set a helper for an existing conntrack */
1971                 err = -EOPNOTSUPP;
1972         }
1973
1974         rcu_read_unlock();
1975         return err;
1976 }
1977
1978 static int ctnetlink_change_timeout(struct nf_conn *ct,
1979                                     const struct nlattr * const cda[])
1980 {
1981         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1982
1983         if (timeout > INT_MAX)
1984                 timeout = INT_MAX;
1985         ct->timeout = nfct_time_stamp + (u32)timeout;
1986
1987         if (test_bit(IPS_DYING_BIT, &ct->status))
1988                 return -ETIME;
1989
1990         return 0;
1991 }
1992
1993 #if defined(CONFIG_NF_CONNTRACK_MARK)
1994 static void ctnetlink_change_mark(struct nf_conn *ct,
1995                                     const struct nlattr * const cda[])
1996 {
1997         u32 mark, newmark, mask = 0;
1998
1999         if (cda[CTA_MARK_MASK])
2000                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2001
2002         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2003         newmark = (ct->mark & mask) ^ mark;
2004         if (newmark != ct->mark)
2005                 ct->mark = newmark;
2006 }
2007 #endif
2008
2009 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2010         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
2011         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
2012         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
2013 };
2014
2015 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2016                                       const struct nlattr * const cda[])
2017 {
2018         const struct nlattr *attr = cda[CTA_PROTOINFO];
2019         const struct nf_conntrack_l4proto *l4proto;
2020         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2021         int err = 0;
2022
2023         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2024                                           protoinfo_policy, NULL);
2025         if (err < 0)
2026                 return err;
2027
2028         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2029         if (l4proto->from_nlattr)
2030                 err = l4proto->from_nlattr(tb, ct);
2031
2032         return err;
2033 }
2034
2035 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2036         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2037         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2038         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2039 };
2040
2041 static int change_seq_adj(struct nf_ct_seqadj *seq,
2042                           const struct nlattr * const attr)
2043 {
2044         int err;
2045         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2046
2047         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2048                                           seqadj_policy, NULL);
2049         if (err < 0)
2050                 return err;
2051
2052         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2053                 return -EINVAL;
2054
2055         seq->correction_pos =
2056                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2057
2058         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2059                 return -EINVAL;
2060
2061         seq->offset_before =
2062                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2063
2064         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2065                 return -EINVAL;
2066
2067         seq->offset_after =
2068                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2069
2070         return 0;
2071 }
2072
2073 static int
2074 ctnetlink_change_seq_adj(struct nf_conn *ct,
2075                          const struct nlattr * const cda[])
2076 {
2077         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2078         int ret = 0;
2079
2080         if (!seqadj)
2081                 return 0;
2082
2083         spin_lock_bh(&ct->lock);
2084         if (cda[CTA_SEQ_ADJ_ORIG]) {
2085                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2086                                      cda[CTA_SEQ_ADJ_ORIG]);
2087                 if (ret < 0)
2088                         goto err;
2089
2090                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2091         }
2092
2093         if (cda[CTA_SEQ_ADJ_REPLY]) {
2094                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2095                                      cda[CTA_SEQ_ADJ_REPLY]);
2096                 if (ret < 0)
2097                         goto err;
2098
2099                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2100         }
2101
2102         spin_unlock_bh(&ct->lock);
2103         return 0;
2104 err:
2105         spin_unlock_bh(&ct->lock);
2106         return ret;
2107 }
2108
2109 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2110         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2111         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2112         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2113 };
2114
2115 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2116                                      const struct nlattr * const cda[])
2117 {
2118         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2119         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2120         int err;
2121
2122         if (!synproxy)
2123                 return 0;
2124
2125         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2126                                           cda[CTA_SYNPROXY], synproxy_policy,
2127                                           NULL);
2128         if (err < 0)
2129                 return err;
2130
2131         if (!tb[CTA_SYNPROXY_ISN] ||
2132             !tb[CTA_SYNPROXY_ITS] ||
2133             !tb[CTA_SYNPROXY_TSOFF])
2134                 return -EINVAL;
2135
2136         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2137         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2138         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2139
2140         return 0;
2141 }
2142
2143 static int
2144 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2145 {
2146 #ifdef CONFIG_NF_CONNTRACK_LABELS
2147         size_t len = nla_len(cda[CTA_LABELS]);
2148         const void *mask = cda[CTA_LABELS_MASK];
2149
2150         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2151                 return -EINVAL;
2152
2153         if (mask) {
2154                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2155                     nla_len(cda[CTA_LABELS_MASK]) != len)
2156                         return -EINVAL;
2157                 mask = nla_data(cda[CTA_LABELS_MASK]);
2158         }
2159
2160         len /= sizeof(u32);
2161
2162         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2163 #else
2164         return -EOPNOTSUPP;
2165 #endif
2166 }
2167
2168 static int
2169 ctnetlink_change_conntrack(struct nf_conn *ct,
2170                            const struct nlattr * const cda[])
2171 {
2172         int err;
2173
2174         /* only allow NAT changes and master assignation for new conntracks */
2175         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2176                 return -EOPNOTSUPP;
2177
2178         if (cda[CTA_HELP]) {
2179                 err = ctnetlink_change_helper(ct, cda);
2180                 if (err < 0)
2181                         return err;
2182         }
2183
2184         if (cda[CTA_TIMEOUT]) {
2185                 err = ctnetlink_change_timeout(ct, cda);
2186                 if (err < 0)
2187                         return err;
2188         }
2189
2190         if (cda[CTA_STATUS]) {
2191                 err = ctnetlink_change_status(ct, cda);
2192                 if (err < 0)
2193                         return err;
2194         }
2195
2196         if (cda[CTA_PROTOINFO]) {
2197                 err = ctnetlink_change_protoinfo(ct, cda);
2198                 if (err < 0)
2199                         return err;
2200         }
2201
2202 #if defined(CONFIG_NF_CONNTRACK_MARK)
2203         if (cda[CTA_MARK])
2204                 ctnetlink_change_mark(ct, cda);
2205 #endif
2206
2207         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2208                 err = ctnetlink_change_seq_adj(ct, cda);
2209                 if (err < 0)
2210                         return err;
2211         }
2212
2213         if (cda[CTA_SYNPROXY]) {
2214                 err = ctnetlink_change_synproxy(ct, cda);
2215                 if (err < 0)
2216                         return err;
2217         }
2218
2219         if (cda[CTA_LABELS]) {
2220                 err = ctnetlink_attach_labels(ct, cda);
2221                 if (err < 0)
2222                         return err;
2223         }
2224
2225         return 0;
2226 }
2227
2228 static struct nf_conn *
2229 ctnetlink_create_conntrack(struct net *net,
2230                            const struct nf_conntrack_zone *zone,
2231                            const struct nlattr * const cda[],
2232                            struct nf_conntrack_tuple *otuple,
2233                            struct nf_conntrack_tuple *rtuple,
2234                            u8 u3)
2235 {
2236         struct nf_conn *ct;
2237         int err = -EINVAL;
2238         struct nf_conntrack_helper *helper;
2239         struct nf_conn_tstamp *tstamp;
2240         u64 timeout;
2241
2242         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2243         if (IS_ERR(ct))
2244                 return ERR_PTR(-ENOMEM);
2245
2246         if (!cda[CTA_TIMEOUT])
2247                 goto err1;
2248
2249         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2250         if (timeout > INT_MAX)
2251                 timeout = INT_MAX;
2252         ct->timeout = (u32)timeout + nfct_time_stamp;
2253
2254         rcu_read_lock();
2255         if (cda[CTA_HELP]) {
2256                 char *helpname = NULL;
2257                 struct nlattr *helpinfo = NULL;
2258
2259                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2260                 if (err < 0)
2261                         goto err2;
2262
2263                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2264                                                     nf_ct_protonum(ct));
2265                 if (helper == NULL) {
2266                         rcu_read_unlock();
2267 #ifdef CONFIG_MODULES
2268                         if (request_module("nfct-helper-%s", helpname) < 0) {
2269                                 err = -EOPNOTSUPP;
2270                                 goto err1;
2271                         }
2272
2273                         rcu_read_lock();
2274                         helper = __nf_conntrack_helper_find(helpname,
2275                                                             nf_ct_l3num(ct),
2276                                                             nf_ct_protonum(ct));
2277                         if (helper) {
2278                                 err = -EAGAIN;
2279                                 goto err2;
2280                         }
2281                         rcu_read_unlock();
2282 #endif
2283                         err = -EOPNOTSUPP;
2284                         goto err1;
2285                 } else {
2286                         struct nf_conn_help *help;
2287
2288                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2289                         if (help == NULL) {
2290                                 err = -ENOMEM;
2291                                 goto err2;
2292                         }
2293                         /* set private helper data if allowed. */
2294                         if (helper->from_nlattr)
2295                                 helper->from_nlattr(helpinfo, ct);
2296
2297                         /* not in hash table yet so not strictly necessary */
2298                         RCU_INIT_POINTER(help->helper, helper);
2299                 }
2300         } else {
2301                 /* try an implicit helper assignation */
2302                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2303                 if (err < 0)
2304                         goto err2;
2305         }
2306
2307         err = ctnetlink_setup_nat(ct, cda);
2308         if (err < 0)
2309                 goto err2;
2310
2311         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2312         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2313         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2314         nf_ct_labels_ext_add(ct);
2315         nfct_seqadj_ext_add(ct);
2316         nfct_synproxy_ext_add(ct);
2317
2318         /* we must add conntrack extensions before confirmation. */
2319         ct->status |= IPS_CONFIRMED;
2320
2321         if (cda[CTA_STATUS]) {
2322                 err = ctnetlink_change_status(ct, cda);
2323                 if (err < 0)
2324                         goto err2;
2325         }
2326
2327         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2328                 err = ctnetlink_change_seq_adj(ct, cda);
2329                 if (err < 0)
2330                         goto err2;
2331         }
2332
2333         memset(&ct->proto, 0, sizeof(ct->proto));
2334         if (cda[CTA_PROTOINFO]) {
2335                 err = ctnetlink_change_protoinfo(ct, cda);
2336                 if (err < 0)
2337                         goto err2;
2338         }
2339
2340         if (cda[CTA_SYNPROXY]) {
2341                 err = ctnetlink_change_synproxy(ct, cda);
2342                 if (err < 0)
2343                         goto err2;
2344         }
2345
2346 #if defined(CONFIG_NF_CONNTRACK_MARK)
2347         if (cda[CTA_MARK])
2348                 ctnetlink_change_mark(ct, cda);
2349 #endif
2350
2351         /* setup master conntrack: this is a confirmed expectation */
2352         if (cda[CTA_TUPLE_MASTER]) {
2353                 struct nf_conntrack_tuple master;
2354                 struct nf_conntrack_tuple_hash *master_h;
2355                 struct nf_conn *master_ct;
2356
2357                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2358                                             u3, NULL);
2359                 if (err < 0)
2360                         goto err2;
2361
2362                 master_h = nf_conntrack_find_get(net, zone, &master);
2363                 if (master_h == NULL) {
2364                         err = -ENOENT;
2365                         goto err2;
2366                 }
2367                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2368                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2369                 ct->master = master_ct;
2370         }
2371         tstamp = nf_conn_tstamp_find(ct);
2372         if (tstamp)
2373                 tstamp->start = ktime_get_real_ns();
2374
2375         err = nf_conntrack_hash_check_insert(ct);
2376         if (err < 0)
2377                 goto err2;
2378
2379         rcu_read_unlock();
2380
2381         return ct;
2382
2383 err2:
2384         rcu_read_unlock();
2385 err1:
2386         nf_conntrack_free(ct);
2387         return ERR_PTR(err);
2388 }
2389
2390 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2391                                    struct sk_buff *skb,
2392                                    const struct nlmsghdr *nlh,
2393                                    const struct nlattr * const cda[],
2394                                    struct netlink_ext_ack *extack)
2395 {
2396         struct nf_conntrack_tuple otuple, rtuple;
2397         struct nf_conntrack_tuple_hash *h = NULL;
2398         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2399         struct nf_conn *ct;
2400         u_int8_t u3 = nfmsg->nfgen_family;
2401         struct nf_conntrack_zone zone;
2402         int err;
2403
2404         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2405         if (err < 0)
2406                 return err;
2407
2408         if (cda[CTA_TUPLE_ORIG]) {
2409                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2410                                             u3, &zone);
2411                 if (err < 0)
2412                         return err;
2413         }
2414
2415         if (cda[CTA_TUPLE_REPLY]) {
2416                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2417                                             u3, &zone);
2418                 if (err < 0)
2419                         return err;
2420         }
2421
2422         if (cda[CTA_TUPLE_ORIG])
2423                 h = nf_conntrack_find_get(net, &zone, &otuple);
2424         else if (cda[CTA_TUPLE_REPLY])
2425                 h = nf_conntrack_find_get(net, &zone, &rtuple);
2426
2427         if (h == NULL) {
2428                 err = -ENOENT;
2429                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2430                         enum ip_conntrack_events events;
2431
2432                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2433                                 return -EINVAL;
2434                         if (otuple.dst.protonum != rtuple.dst.protonum)
2435                                 return -EINVAL;
2436
2437                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2438                                                         &rtuple, u3);
2439                         if (IS_ERR(ct))
2440                                 return PTR_ERR(ct);
2441
2442                         err = 0;
2443                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2444                                 events = 1 << IPCT_RELATED;
2445                         else
2446                                 events = 1 << IPCT_NEW;
2447
2448                         if (cda[CTA_LABELS] &&
2449                             ctnetlink_attach_labels(ct, cda) == 0)
2450                                 events |= (1 << IPCT_LABEL);
2451
2452                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2453                                                       (1 << IPCT_ASSURED) |
2454                                                       (1 << IPCT_HELPER) |
2455                                                       (1 << IPCT_PROTOINFO) |
2456                                                       (1 << IPCT_SEQADJ) |
2457                                                       (1 << IPCT_MARK) |
2458                                                       (1 << IPCT_SYNPROXY) |
2459                                                       events,
2460                                                       ct, NETLINK_CB(skb).portid,
2461                                                       nlmsg_report(nlh));
2462                         nf_ct_put(ct);
2463                 }
2464
2465                 return err;
2466         }
2467         /* implicit 'else' */
2468
2469         err = -EEXIST;
2470         ct = nf_ct_tuplehash_to_ctrack(h);
2471         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2472                 err = ctnetlink_change_conntrack(ct, cda);
2473                 if (err == 0) {
2474                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2475                                                       (1 << IPCT_ASSURED) |
2476                                                       (1 << IPCT_HELPER) |
2477                                                       (1 << IPCT_LABEL) |
2478                                                       (1 << IPCT_PROTOINFO) |
2479                                                       (1 << IPCT_SEQADJ) |
2480                                                       (1 << IPCT_MARK) |
2481                                                       (1 << IPCT_SYNPROXY),
2482                                                       ct, NETLINK_CB(skb).portid,
2483                                                       nlmsg_report(nlh));
2484                 }
2485         }
2486
2487         nf_ct_put(ct);
2488         return err;
2489 }
2490
2491 static int
2492 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2493                                 __u16 cpu, const struct ip_conntrack_stat *st)
2494 {
2495         struct nlmsghdr *nlh;
2496         struct nfgenmsg *nfmsg;
2497         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2498
2499         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2500                               IPCTNL_MSG_CT_GET_STATS_CPU);
2501         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2502         if (nlh == NULL)
2503                 goto nlmsg_failure;
2504
2505         nfmsg = nlmsg_data(nlh);
2506         nfmsg->nfgen_family = AF_UNSPEC;
2507         nfmsg->version      = NFNETLINK_V0;
2508         nfmsg->res_id       = htons(cpu);
2509
2510         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2511             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2512             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2513             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2514             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2515                                 htonl(st->insert_failed)) ||
2516             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2517             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2518             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2519             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2520                                 htonl(st->search_restart)))
2521                 goto nla_put_failure;
2522
2523         nlmsg_end(skb, nlh);
2524         return skb->len;
2525
2526 nla_put_failure:
2527 nlmsg_failure:
2528         nlmsg_cancel(skb, nlh);
2529         return -1;
2530 }
2531
2532 static int
2533 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2534 {
2535         int cpu;
2536         struct net *net = sock_net(skb->sk);
2537
2538         if (cb->args[0] == nr_cpu_ids)
2539                 return 0;
2540
2541         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2542                 const struct ip_conntrack_stat *st;
2543
2544                 if (!cpu_possible(cpu))
2545                         continue;
2546
2547                 st = per_cpu_ptr(net->ct.stat, cpu);
2548                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2549                                                     NETLINK_CB(cb->skb).portid,
2550                                                     cb->nlh->nlmsg_seq,
2551                                                     cpu, st) < 0)
2552                                 break;
2553         }
2554         cb->args[0] = cpu;
2555
2556         return skb->len;
2557 }
2558
2559 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2560                                  struct sk_buff *skb,
2561                                  const struct nlmsghdr *nlh,
2562                                  const struct nlattr * const cda[],
2563                                  struct netlink_ext_ack *extack)
2564 {
2565         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2566                 struct netlink_dump_control c = {
2567                         .dump = ctnetlink_ct_stat_cpu_dump,
2568                 };
2569                 return netlink_dump_start(ctnl, skb, nlh, &c);
2570         }
2571
2572         return 0;
2573 }
2574
2575 static int
2576 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2577                             struct net *net)
2578 {
2579         struct nlmsghdr *nlh;
2580         struct nfgenmsg *nfmsg;
2581         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2582         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2583
2584         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2585         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2586         if (nlh == NULL)
2587                 goto nlmsg_failure;
2588
2589         nfmsg = nlmsg_data(nlh);
2590         nfmsg->nfgen_family = AF_UNSPEC;
2591         nfmsg->version      = NFNETLINK_V0;
2592         nfmsg->res_id       = 0;
2593
2594         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2595                 goto nla_put_failure;
2596
2597         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2598                 goto nla_put_failure;
2599
2600         nlmsg_end(skb, nlh);
2601         return skb->len;
2602
2603 nla_put_failure:
2604 nlmsg_failure:
2605         nlmsg_cancel(skb, nlh);
2606         return -1;
2607 }
2608
2609 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2610                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2611                              const struct nlattr * const cda[],
2612                              struct netlink_ext_ack *extack)
2613 {
2614         struct sk_buff *skb2;
2615         int err;
2616
2617         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2618         if (skb2 == NULL)
2619                 return -ENOMEM;
2620
2621         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2622                                           nlh->nlmsg_seq,
2623                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2624                                           sock_net(skb->sk));
2625         if (err <= 0)
2626                 goto free;
2627
2628         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2629         if (err < 0)
2630                 goto out;
2631
2632         return 0;
2633
2634 free:
2635         kfree_skb(skb2);
2636 out:
2637         /* this avoids a loop in nfnetlink. */
2638         return err == -EAGAIN ? -ENOBUFS : err;
2639 }
2640
2641 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2642         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2643         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2644         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2645         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2646         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2647         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2648                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2649         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2650         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2651         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2652         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2653         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2654 };
2655
2656 static struct nf_conntrack_expect *
2657 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2658                        struct nf_conntrack_helper *helper,
2659                        struct nf_conntrack_tuple *tuple,
2660                        struct nf_conntrack_tuple *mask);
2661
2662 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2663 static size_t
2664 ctnetlink_glue_build_size(const struct nf_conn *ct)
2665 {
2666         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2667                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2668                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2669                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2670                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2671                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2672                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2673                + nla_total_size(0) /* CTA_PROTOINFO */
2674                + nla_total_size(0) /* CTA_HELP */
2675                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2676                + ctnetlink_secctx_size(ct)
2677 #if IS_ENABLED(CONFIG_NF_NAT)
2678                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2679                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2680 #endif
2681 #ifdef CONFIG_NF_CONNTRACK_MARK
2682                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2683 #endif
2684 #ifdef CONFIG_NF_CONNTRACK_ZONES
2685                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2686 #endif
2687                + ctnetlink_proto_size(ct)
2688                ;
2689 }
2690
2691 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2692                                              enum ip_conntrack_info *ctinfo)
2693 {
2694         return nf_ct_get(skb, ctinfo);
2695 }
2696
2697 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2698 {
2699         const struct nf_conntrack_zone *zone;
2700         struct nlattr *nest_parms;
2701
2702         zone = nf_ct_zone(ct);
2703
2704         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2705         if (!nest_parms)
2706                 goto nla_put_failure;
2707         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2708                 goto nla_put_failure;
2709         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2710                                    NF_CT_ZONE_DIR_ORIG) < 0)
2711                 goto nla_put_failure;
2712         nla_nest_end(skb, nest_parms);
2713
2714         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2715         if (!nest_parms)
2716                 goto nla_put_failure;
2717         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2718                 goto nla_put_failure;
2719         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2720                                    NF_CT_ZONE_DIR_REPL) < 0)
2721                 goto nla_put_failure;
2722         nla_nest_end(skb, nest_parms);
2723
2724         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2725                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2726                 goto nla_put_failure;
2727
2728         if (ctnetlink_dump_id(skb, ct) < 0)
2729                 goto nla_put_failure;
2730
2731         if (ctnetlink_dump_status(skb, ct) < 0)
2732                 goto nla_put_failure;
2733
2734         if (ctnetlink_dump_timeout(skb, ct) < 0)
2735                 goto nla_put_failure;
2736
2737         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2738                 goto nla_put_failure;
2739
2740         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2741                 goto nla_put_failure;
2742
2743 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2744         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2745                 goto nla_put_failure;
2746 #endif
2747         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2748                 goto nla_put_failure;
2749
2750         if ((ct->status & IPS_SEQ_ADJUST) &&
2751             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2752                 goto nla_put_failure;
2753
2754         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2755                 goto nla_put_failure;
2756
2757 #ifdef CONFIG_NF_CONNTRACK_MARK
2758         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2759                 goto nla_put_failure;
2760 #endif
2761         if (ctnetlink_dump_labels(skb, ct) < 0)
2762                 goto nla_put_failure;
2763         return 0;
2764
2765 nla_put_failure:
2766         return -ENOSPC;
2767 }
2768
2769 static int
2770 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2771                      enum ip_conntrack_info ctinfo,
2772                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2773 {
2774         struct nlattr *nest_parms;
2775
2776         nest_parms = nla_nest_start(skb, ct_attr);
2777         if (!nest_parms)
2778                 goto nla_put_failure;
2779
2780         if (__ctnetlink_glue_build(skb, ct) < 0)
2781                 goto nla_put_failure;
2782
2783         nla_nest_end(skb, nest_parms);
2784
2785         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2786                 goto nla_put_failure;
2787
2788         return 0;
2789
2790 nla_put_failure:
2791         return -ENOSPC;
2792 }
2793
2794 static int
2795 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2796 {
2797         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2798         unsigned long d = ct->status ^ status;
2799
2800         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2801                 /* SEEN_REPLY bit can only be set */
2802                 return -EBUSY;
2803
2804         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2805                 /* ASSURED bit can only be set */
2806                 return -EBUSY;
2807
2808         /* This check is less strict than ctnetlink_change_status()
2809          * because callers often flip IPS_EXPECTED bits when sending
2810          * an NFQA_CT attribute to the kernel.  So ignore the
2811          * unchangeable bits but do not error out. Also user programs
2812          * are allowed to clear the bits that they are allowed to change.
2813          */
2814         __ctnetlink_change_status(ct, status, ~status);
2815         return 0;
2816 }
2817
2818 static int
2819 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2820 {
2821         int err;
2822
2823         if (cda[CTA_TIMEOUT]) {
2824                 err = ctnetlink_change_timeout(ct, cda);
2825                 if (err < 0)
2826                         return err;
2827         }
2828         if (cda[CTA_STATUS]) {
2829                 err = ctnetlink_update_status(ct, cda);
2830                 if (err < 0)
2831                         return err;
2832         }
2833         if (cda[CTA_HELP]) {
2834                 err = ctnetlink_change_helper(ct, cda);
2835                 if (err < 0)
2836                         return err;
2837         }
2838         if (cda[CTA_LABELS]) {
2839                 err = ctnetlink_attach_labels(ct, cda);
2840                 if (err < 0)
2841                         return err;
2842         }
2843 #if defined(CONFIG_NF_CONNTRACK_MARK)
2844         if (cda[CTA_MARK]) {
2845                 ctnetlink_change_mark(ct, cda);
2846         }
2847 #endif
2848         return 0;
2849 }
2850
2851 static int
2852 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2853 {
2854         struct nlattr *cda[CTA_MAX+1];
2855         int ret;
2856
2857         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2858                                           NULL);
2859         if (ret < 0)
2860                 return ret;
2861
2862         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2863 }
2864
2865 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2866                                     const struct nf_conn *ct,
2867                                     struct nf_conntrack_tuple *tuple,
2868                                     struct nf_conntrack_tuple *mask)
2869 {
2870         int err;
2871
2872         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2873                                     nf_ct_l3num(ct), NULL);
2874         if (err < 0)
2875                 return err;
2876
2877         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2878                                      nf_ct_l3num(ct), NULL);
2879 }
2880
2881 static int
2882 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2883                              u32 portid, u32 report)
2884 {
2885         struct nlattr *cda[CTA_EXPECT_MAX+1];
2886         struct nf_conntrack_tuple tuple, mask;
2887         struct nf_conntrack_helper *helper = NULL;
2888         struct nf_conntrack_expect *exp;
2889         int err;
2890
2891         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2892                                           exp_nla_policy, NULL);
2893         if (err < 0)
2894                 return err;
2895
2896         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2897                                        ct, &tuple, &mask);
2898         if (err < 0)
2899                 return err;
2900
2901         if (cda[CTA_EXPECT_HELP_NAME]) {
2902                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2903
2904                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2905                                                     nf_ct_protonum(ct));
2906                 if (helper == NULL)
2907                         return -EOPNOTSUPP;
2908         }
2909
2910         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2911                                      helper, &tuple, &mask);
2912         if (IS_ERR(exp))
2913                 return PTR_ERR(exp);
2914
2915         err = nf_ct_expect_related_report(exp, portid, report, 0);
2916         nf_ct_expect_put(exp);
2917         return err;
2918 }
2919
2920 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2921                                   enum ip_conntrack_info ctinfo, int diff)
2922 {
2923         if (!(ct->status & IPS_NAT_MASK))
2924                 return;
2925
2926         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2927 }
2928
2929 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2930         .get_ct         = ctnetlink_glue_get_ct,
2931         .build_size     = ctnetlink_glue_build_size,
2932         .build          = ctnetlink_glue_build,
2933         .parse          = ctnetlink_glue_parse,
2934         .attach_expect  = ctnetlink_glue_attach_expect,
2935         .seq_adjust     = ctnetlink_glue_seqadj,
2936 };
2937 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2938
2939 /***********************************************************************
2940  * EXPECT
2941  ***********************************************************************/
2942
2943 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2944                                     const struct nf_conntrack_tuple *tuple,
2945                                     u32 type)
2946 {
2947         struct nlattr *nest_parms;
2948
2949         nest_parms = nla_nest_start(skb, type);
2950         if (!nest_parms)
2951                 goto nla_put_failure;
2952         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2953                 goto nla_put_failure;
2954         nla_nest_end(skb, nest_parms);
2955
2956         return 0;
2957
2958 nla_put_failure:
2959         return -1;
2960 }
2961
2962 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2963                                    const struct nf_conntrack_tuple *tuple,
2964                                    const struct nf_conntrack_tuple_mask *mask)
2965 {
2966         const struct nf_conntrack_l4proto *l4proto;
2967         struct nf_conntrack_tuple m;
2968         struct nlattr *nest_parms;
2969         int ret;
2970
2971         memset(&m, 0xFF, sizeof(m));
2972         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2973         m.src.u.all = mask->src.u.all;
2974         m.dst.protonum = tuple->dst.protonum;
2975
2976         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2977         if (!nest_parms)
2978                 goto nla_put_failure;
2979
2980         rcu_read_lock();
2981         ret = ctnetlink_dump_tuples_ip(skb, &m);
2982         if (ret >= 0) {
2983                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2984                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2985         }
2986         rcu_read_unlock();
2987
2988         if (unlikely(ret < 0))
2989                 goto nla_put_failure;
2990
2991         nla_nest_end(skb, nest_parms);
2992
2993         return 0;
2994
2995 nla_put_failure:
2996         return -1;
2997 }
2998
2999 static const union nf_inet_addr any_addr;
3000
3001 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
3002 {
3003         static __read_mostly siphash_key_t exp_id_seed;
3004         unsigned long a, b, c, d;
3005
3006         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
3007
3008         a = (unsigned long)exp;
3009         b = (unsigned long)exp->helper;
3010         c = (unsigned long)exp->master;
3011         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
3012
3013 #ifdef CONFIG_64BIT
3014         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
3015 #else
3016         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
3017 #endif
3018 }
3019
3020 static int
3021 ctnetlink_exp_dump_expect(struct sk_buff *skb,
3022                           const struct nf_conntrack_expect *exp)
3023 {
3024         struct nf_conn *master = exp->master;
3025         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3026         struct nf_conn_help *help;
3027 #if IS_ENABLED(CONFIG_NF_NAT)
3028         struct nlattr *nest_parms;
3029         struct nf_conntrack_tuple nat_tuple = {};
3030 #endif
3031         struct nf_ct_helper_expectfn *expfn;
3032
3033         if (timeout < 0)
3034                 timeout = 0;
3035
3036         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3037                 goto nla_put_failure;
3038         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3039                 goto nla_put_failure;
3040         if (ctnetlink_exp_dump_tuple(skb,
3041                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3042                                  CTA_EXPECT_MASTER) < 0)
3043                 goto nla_put_failure;
3044
3045 #if IS_ENABLED(CONFIG_NF_NAT)
3046         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3047             exp->saved_proto.all) {
3048                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3049                 if (!nest_parms)
3050                         goto nla_put_failure;
3051
3052                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3053                         goto nla_put_failure;
3054
3055                 nat_tuple.src.l3num = nf_ct_l3num(master);
3056                 nat_tuple.src.u3 = exp->saved_addr;
3057                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3058                 nat_tuple.src.u = exp->saved_proto;
3059
3060                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3061                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3062                         goto nla_put_failure;
3063                 nla_nest_end(skb, nest_parms);
3064         }
3065 #endif
3066         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3067             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3068             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3069             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3070                 goto nla_put_failure;
3071         help = nfct_help(master);
3072         if (help) {
3073                 struct nf_conntrack_helper *helper;
3074
3075                 helper = rcu_dereference(help->helper);
3076                 if (helper &&
3077                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3078                         goto nla_put_failure;
3079         }
3080         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3081         if (expfn != NULL &&
3082             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3083                 goto nla_put_failure;
3084
3085         return 0;
3086
3087 nla_put_failure:
3088         return -1;
3089 }
3090
3091 static int
3092 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3093                         int event, const struct nf_conntrack_expect *exp)
3094 {
3095         struct nlmsghdr *nlh;
3096         struct nfgenmsg *nfmsg;
3097         unsigned int flags = portid ? NLM_F_MULTI : 0;
3098
3099         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3100         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3101         if (nlh == NULL)
3102                 goto nlmsg_failure;
3103
3104         nfmsg = nlmsg_data(nlh);
3105         nfmsg->nfgen_family = exp->tuple.src.l3num;
3106         nfmsg->version      = NFNETLINK_V0;
3107         nfmsg->res_id       = 0;
3108
3109         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3110                 goto nla_put_failure;
3111
3112         nlmsg_end(skb, nlh);
3113         return skb->len;
3114
3115 nlmsg_failure:
3116 nla_put_failure:
3117         nlmsg_cancel(skb, nlh);
3118         return -1;
3119 }
3120
3121 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3122 static int
3123 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
3124 {
3125         struct nf_conntrack_expect *exp = item->exp;
3126         struct net *net = nf_ct_exp_net(exp);
3127         struct nlmsghdr *nlh;
3128         struct nfgenmsg *nfmsg;
3129         struct sk_buff *skb;
3130         unsigned int type, group;
3131         int flags = 0;
3132
3133         if (events & (1 << IPEXP_DESTROY)) {
3134                 type = IPCTNL_MSG_EXP_DELETE;
3135                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3136         } else if (events & (1 << IPEXP_NEW)) {
3137                 type = IPCTNL_MSG_EXP_NEW;
3138                 flags = NLM_F_CREATE|NLM_F_EXCL;
3139                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3140         } else
3141                 return 0;
3142
3143         if (!item->report && !nfnetlink_has_listeners(net, group))
3144                 return 0;
3145
3146         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3147         if (skb == NULL)
3148                 goto errout;
3149
3150         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3151         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
3152         if (nlh == NULL)
3153                 goto nlmsg_failure;
3154
3155         nfmsg = nlmsg_data(nlh);
3156         nfmsg->nfgen_family = exp->tuple.src.l3num;
3157         nfmsg->version      = NFNETLINK_V0;
3158         nfmsg->res_id       = 0;
3159
3160         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3161                 goto nla_put_failure;
3162
3163         nlmsg_end(skb, nlh);
3164         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3165         return 0;
3166
3167 nla_put_failure:
3168         nlmsg_cancel(skb, nlh);
3169 nlmsg_failure:
3170         kfree_skb(skb);
3171 errout:
3172         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3173         return 0;
3174 }
3175 #endif
3176 static int ctnetlink_exp_done(struct netlink_callback *cb)
3177 {
3178         if (cb->args[1])
3179                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3180         return 0;
3181 }
3182
3183 static int
3184 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3185 {
3186         struct net *net = sock_net(skb->sk);
3187         struct nf_conntrack_expect *exp, *last;
3188         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3189         u_int8_t l3proto = nfmsg->nfgen_family;
3190
3191         rcu_read_lock();
3192         last = (struct nf_conntrack_expect *)cb->args[1];
3193         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3194 restart:
3195                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3196                                          hnode) {
3197                         if (l3proto && exp->tuple.src.l3num != l3proto)
3198                                 continue;
3199
3200                         if (!net_eq(nf_ct_net(exp->master), net))
3201                                 continue;
3202
3203                         if (cb->args[1]) {
3204                                 if (exp != last)
3205                                         continue;
3206                                 cb->args[1] = 0;
3207                         }
3208                         if (ctnetlink_exp_fill_info(skb,
3209                                                     NETLINK_CB(cb->skb).portid,
3210                                                     cb->nlh->nlmsg_seq,
3211                                                     IPCTNL_MSG_EXP_NEW,
3212                                                     exp) < 0) {
3213                                 if (!refcount_inc_not_zero(&exp->use))
3214                                         continue;
3215                                 cb->args[1] = (unsigned long)exp;
3216                                 goto out;
3217                         }
3218                 }
3219                 if (cb->args[1]) {
3220                         cb->args[1] = 0;
3221                         goto restart;
3222                 }
3223         }
3224 out:
3225         rcu_read_unlock();
3226         if (last)
3227                 nf_ct_expect_put(last);
3228
3229         return skb->len;
3230 }
3231
3232 static int
3233 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3234 {
3235         struct nf_conntrack_expect *exp, *last;
3236         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3237         struct nf_conn *ct = cb->data;
3238         struct nf_conn_help *help = nfct_help(ct);
3239         u_int8_t l3proto = nfmsg->nfgen_family;
3240
3241         if (cb->args[0])
3242                 return 0;
3243
3244         rcu_read_lock();
3245         last = (struct nf_conntrack_expect *)cb->args[1];
3246 restart:
3247         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3248                 if (l3proto && exp->tuple.src.l3num != l3proto)
3249                         continue;
3250                 if (cb->args[1]) {
3251                         if (exp != last)
3252                                 continue;
3253                         cb->args[1] = 0;
3254                 }
3255                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3256                                             cb->nlh->nlmsg_seq,
3257                                             IPCTNL_MSG_EXP_NEW,
3258                                             exp) < 0) {
3259                         if (!refcount_inc_not_zero(&exp->use))
3260                                 continue;
3261                         cb->args[1] = (unsigned long)exp;
3262                         goto out;
3263                 }
3264         }
3265         if (cb->args[1]) {
3266                 cb->args[1] = 0;
3267                 goto restart;
3268         }
3269         cb->args[0] = 1;
3270 out:
3271         rcu_read_unlock();
3272         if (last)
3273                 nf_ct_expect_put(last);
3274
3275         return skb->len;
3276 }
3277
3278 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3279                                  struct sk_buff *skb,
3280                                  const struct nlmsghdr *nlh,
3281                                  const struct nlattr * const cda[],
3282                                  struct netlink_ext_ack *extack)
3283 {
3284         int err;
3285         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3286         u_int8_t u3 = nfmsg->nfgen_family;
3287         struct nf_conntrack_tuple tuple;
3288         struct nf_conntrack_tuple_hash *h;
3289         struct nf_conn *ct;
3290         struct nf_conntrack_zone zone;
3291         struct netlink_dump_control c = {
3292                 .dump = ctnetlink_exp_ct_dump_table,
3293                 .done = ctnetlink_exp_done,
3294         };
3295
3296         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3297                                     u3, NULL);
3298         if (err < 0)
3299                 return err;
3300
3301         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3302         if (err < 0)
3303                 return err;
3304
3305         h = nf_conntrack_find_get(net, &zone, &tuple);
3306         if (!h)
3307                 return -ENOENT;
3308
3309         ct = nf_ct_tuplehash_to_ctrack(h);
3310         /* No expectation linked to this connection tracking. */
3311         if (!nfct_help(ct)) {
3312                 nf_ct_put(ct);
3313                 return 0;
3314         }
3315
3316         c.data = ct;
3317
3318         err = netlink_dump_start(ctnl, skb, nlh, &c);
3319         nf_ct_put(ct);
3320
3321         return err;
3322 }
3323
3324 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
3325                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3326                                 const struct nlattr * const cda[],
3327                                 struct netlink_ext_ack *extack)
3328 {
3329         struct nf_conntrack_tuple tuple;
3330         struct nf_conntrack_expect *exp;
3331         struct sk_buff *skb2;
3332         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3333         u_int8_t u3 = nfmsg->nfgen_family;
3334         struct nf_conntrack_zone zone;
3335         int err;
3336
3337         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3338                 if (cda[CTA_EXPECT_MASTER])
3339                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
3340                                                      extack);
3341                 else {
3342                         struct netlink_dump_control c = {
3343                                 .dump = ctnetlink_exp_dump_table,
3344                                 .done = ctnetlink_exp_done,
3345                         };
3346                         return netlink_dump_start(ctnl, skb, nlh, &c);
3347                 }
3348         }
3349
3350         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3351         if (err < 0)
3352                 return err;
3353
3354         if (cda[CTA_EXPECT_TUPLE])
3355                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3356                                             u3, NULL);
3357         else if (cda[CTA_EXPECT_MASTER])
3358                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3359                                             u3, NULL);
3360         else
3361                 return -EINVAL;
3362
3363         if (err < 0)
3364                 return err;
3365
3366         exp = nf_ct_expect_find_get(net, &zone, &tuple);
3367         if (!exp)
3368                 return -ENOENT;
3369
3370         if (cda[CTA_EXPECT_ID]) {
3371                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3372
3373                 if (id != nf_expect_get_id(exp)) {
3374                         nf_ct_expect_put(exp);
3375                         return -ENOENT;
3376                 }
3377         }
3378
3379         err = -ENOMEM;
3380         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3381         if (skb2 == NULL) {
3382                 nf_ct_expect_put(exp);
3383                 goto out;
3384         }
3385
3386         rcu_read_lock();
3387         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3388                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
3389         rcu_read_unlock();
3390         nf_ct_expect_put(exp);
3391         if (err <= 0)
3392                 goto free;
3393
3394         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
3395         if (err < 0)
3396                 goto out;
3397
3398         return 0;
3399
3400 free:
3401         kfree_skb(skb2);
3402 out:
3403         /* this avoids a loop in nfnetlink. */
3404         return err == -EAGAIN ? -ENOBUFS : err;
3405 }
3406
3407 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3408 {
3409         const struct nf_conn_help *m_help;
3410         const char *name = data;
3411
3412         m_help = nfct_help(exp->master);
3413
3414         return strcmp(m_help->helper->name, name) == 0;
3415 }
3416
3417 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3418 {
3419         return true;
3420 }
3421
3422 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3423                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3424                                 const struct nlattr * const cda[],
3425                                 struct netlink_ext_ack *extack)
3426 {
3427         struct nf_conntrack_expect *exp;
3428         struct nf_conntrack_tuple tuple;
3429         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3430         u_int8_t u3 = nfmsg->nfgen_family;
3431         struct nf_conntrack_zone zone;
3432         int err;
3433
3434         if (cda[CTA_EXPECT_TUPLE]) {
3435                 /* delete a single expect by tuple */
3436                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3437                 if (err < 0)
3438                         return err;
3439
3440                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3441                                             u3, NULL);
3442                 if (err < 0)
3443                         return err;
3444
3445                 /* bump usage count to 2 */
3446                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
3447                 if (!exp)
3448                         return -ENOENT;
3449
3450                 if (cda[CTA_EXPECT_ID]) {
3451                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3452                         if (ntohl(id) != (u32)(unsigned long)exp) {
3453                                 nf_ct_expect_put(exp);
3454                                 return -ENOENT;
3455                         }
3456                 }
3457
3458                 /* after list removal, usage count == 1 */
3459                 spin_lock_bh(&nf_conntrack_expect_lock);
3460                 if (del_timer(&exp->timeout)) {
3461                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3462                                                    nlmsg_report(nlh));
3463                         nf_ct_expect_put(exp);
3464                 }
3465                 spin_unlock_bh(&nf_conntrack_expect_lock);
3466                 /* have to put what we 'get' above.
3467                  * after this line usage count == 0 */
3468                 nf_ct_expect_put(exp);
3469         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3470                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3471
3472                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
3473                                          NETLINK_CB(skb).portid,
3474                                          nlmsg_report(nlh));
3475         } else {
3476                 /* This basically means we have to flush everything*/
3477                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3478                                          NETLINK_CB(skb).portid,
3479                                          nlmsg_report(nlh));
3480         }
3481
3482         return 0;
3483 }
3484 static int
3485 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3486                         const struct nlattr * const cda[])
3487 {
3488         if (cda[CTA_EXPECT_TIMEOUT]) {
3489                 if (!del_timer(&x->timeout))
3490                         return -ETIME;
3491
3492                 x->timeout.expires = jiffies +
3493                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3494                 add_timer(&x->timeout);
3495         }
3496         return 0;
3497 }
3498
3499 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3500         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3501         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3502 };
3503
3504 static int
3505 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3506                            struct nf_conntrack_expect *exp,
3507                            u_int8_t u3)
3508 {
3509 #if IS_ENABLED(CONFIG_NF_NAT)
3510         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3511         struct nf_conntrack_tuple nat_tuple = {};
3512         int err;
3513
3514         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3515                                           exp_nat_nla_policy, NULL);
3516         if (err < 0)
3517                 return err;
3518
3519         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3520                 return -EINVAL;
3521
3522         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3523                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3524                                     u3, NULL);
3525         if (err < 0)
3526                 return err;
3527
3528         exp->saved_addr = nat_tuple.src.u3;
3529         exp->saved_proto = nat_tuple.src.u;
3530         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3531
3532         return 0;
3533 #else
3534         return -EOPNOTSUPP;
3535 #endif
3536 }
3537
3538 static struct nf_conntrack_expect *
3539 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3540                        struct nf_conntrack_helper *helper,
3541                        struct nf_conntrack_tuple *tuple,
3542                        struct nf_conntrack_tuple *mask)
3543 {
3544         u_int32_t class = 0;
3545         struct nf_conntrack_expect *exp;
3546         struct nf_conn_help *help;
3547         int err;
3548
3549         help = nfct_help(ct);
3550         if (!help)
3551                 return ERR_PTR(-EOPNOTSUPP);
3552
3553         if (cda[CTA_EXPECT_CLASS] && helper) {
3554                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3555                 if (class > helper->expect_class_max)
3556                         return ERR_PTR(-EINVAL);
3557         }
3558         exp = nf_ct_expect_alloc(ct);
3559         if (!exp)
3560                 return ERR_PTR(-ENOMEM);
3561
3562         if (cda[CTA_EXPECT_FLAGS]) {
3563                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3564                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3565         } else {
3566                 exp->flags = 0;
3567         }
3568         if (cda[CTA_EXPECT_FN]) {
3569                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3570                 struct nf_ct_helper_expectfn *expfn;
3571
3572                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3573                 if (expfn == NULL) {
3574                         err = -EINVAL;
3575                         goto err_out;
3576                 }
3577                 exp->expectfn = expfn->expectfn;
3578         } else
3579                 exp->expectfn = NULL;
3580
3581         exp->class = class;
3582         exp->master = ct;
3583         exp->helper = helper;
3584         exp->tuple = *tuple;
3585         exp->mask.src.u3 = mask->src.u3;
3586         exp->mask.src.u.all = mask->src.u.all;
3587
3588         if (cda[CTA_EXPECT_NAT]) {
3589                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3590                                                  exp, nf_ct_l3num(ct));
3591                 if (err < 0)
3592                         goto err_out;
3593         }
3594         return exp;
3595 err_out:
3596         nf_ct_expect_put(exp);
3597         return ERR_PTR(err);
3598 }
3599
3600 static int
3601 ctnetlink_create_expect(struct net *net,
3602                         const struct nf_conntrack_zone *zone,
3603                         const struct nlattr * const cda[],
3604                         u_int8_t u3, u32 portid, int report)
3605 {
3606         struct nf_conntrack_tuple tuple, mask, master_tuple;
3607         struct nf_conntrack_tuple_hash *h = NULL;
3608         struct nf_conntrack_helper *helper = NULL;
3609         struct nf_conntrack_expect *exp;
3610         struct nf_conn *ct;
3611         int err;
3612
3613         /* caller guarantees that those three CTA_EXPECT_* exist */
3614         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3615                                     u3, NULL);
3616         if (err < 0)
3617                 return err;
3618         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3619                                     u3, NULL);
3620         if (err < 0)
3621                 return err;
3622         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3623                                     u3, NULL);
3624         if (err < 0)
3625                 return err;
3626
3627         /* Look for master conntrack of this expectation */
3628         h = nf_conntrack_find_get(net, zone, &master_tuple);
3629         if (!h)
3630                 return -ENOENT;
3631         ct = nf_ct_tuplehash_to_ctrack(h);
3632
3633         rcu_read_lock();
3634         if (cda[CTA_EXPECT_HELP_NAME]) {
3635                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3636
3637                 helper = __nf_conntrack_helper_find(helpname, u3,
3638                                                     nf_ct_protonum(ct));
3639                 if (helper == NULL) {
3640                         rcu_read_unlock();
3641 #ifdef CONFIG_MODULES
3642                         if (request_module("nfct-helper-%s", helpname) < 0) {
3643                                 err = -EOPNOTSUPP;
3644                                 goto err_ct;
3645                         }
3646                         rcu_read_lock();
3647                         helper = __nf_conntrack_helper_find(helpname, u3,
3648                                                             nf_ct_protonum(ct));
3649                         if (helper) {
3650                                 err = -EAGAIN;
3651                                 goto err_rcu;
3652                         }
3653                         rcu_read_unlock();
3654 #endif
3655                         err = -EOPNOTSUPP;
3656                         goto err_ct;
3657                 }
3658         }
3659
3660         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3661         if (IS_ERR(exp)) {
3662                 err = PTR_ERR(exp);
3663                 goto err_rcu;
3664         }
3665
3666         err = nf_ct_expect_related_report(exp, portid, report, 0);
3667         nf_ct_expect_put(exp);
3668 err_rcu:
3669         rcu_read_unlock();
3670 err_ct:
3671         nf_ct_put(ct);
3672         return err;
3673 }
3674
3675 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3676                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3677                                 const struct nlattr * const cda[],
3678                                 struct netlink_ext_ack *extack)
3679 {
3680         struct nf_conntrack_tuple tuple;
3681         struct nf_conntrack_expect *exp;
3682         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3683         u_int8_t u3 = nfmsg->nfgen_family;
3684         struct nf_conntrack_zone zone;
3685         int err;
3686
3687         if (!cda[CTA_EXPECT_TUPLE]
3688             || !cda[CTA_EXPECT_MASK]
3689             || !cda[CTA_EXPECT_MASTER])
3690                 return -EINVAL;
3691
3692         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3693         if (err < 0)
3694                 return err;
3695
3696         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3697                                     u3, NULL);
3698         if (err < 0)
3699                 return err;
3700
3701         spin_lock_bh(&nf_conntrack_expect_lock);
3702         exp = __nf_ct_expect_find(net, &zone, &tuple);
3703         if (!exp) {
3704                 spin_unlock_bh(&nf_conntrack_expect_lock);
3705                 err = -ENOENT;
3706                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3707                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3708                                                       NETLINK_CB(skb).portid,
3709                                                       nlmsg_report(nlh));
3710                 }
3711                 return err;
3712         }
3713
3714         err = -EEXIST;
3715         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3716                 err = ctnetlink_change_expect(exp, cda);
3717         spin_unlock_bh(&nf_conntrack_expect_lock);
3718
3719         return err;
3720 }
3721
3722 static int
3723 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3724                              const struct ip_conntrack_stat *st)
3725 {
3726         struct nlmsghdr *nlh;
3727         struct nfgenmsg *nfmsg;
3728         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3729
3730         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3731                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3732         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3733         if (nlh == NULL)
3734                 goto nlmsg_failure;
3735
3736         nfmsg = nlmsg_data(nlh);
3737         nfmsg->nfgen_family = AF_UNSPEC;
3738         nfmsg->version      = NFNETLINK_V0;
3739         nfmsg->res_id       = htons(cpu);
3740
3741         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3742             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3743             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3744                 goto nla_put_failure;
3745
3746         nlmsg_end(skb, nlh);
3747         return skb->len;
3748
3749 nla_put_failure:
3750 nlmsg_failure:
3751         nlmsg_cancel(skb, nlh);
3752         return -1;
3753 }
3754
3755 static int
3756 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3757 {
3758         int cpu;
3759         struct net *net = sock_net(skb->sk);
3760
3761         if (cb->args[0] == nr_cpu_ids)
3762                 return 0;
3763
3764         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3765                 const struct ip_conntrack_stat *st;
3766
3767                 if (!cpu_possible(cpu))
3768                         continue;
3769
3770                 st = per_cpu_ptr(net->ct.stat, cpu);
3771                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3772                                                  cb->nlh->nlmsg_seq,
3773                                                  cpu, st) < 0)
3774                         break;
3775         }
3776         cb->args[0] = cpu;
3777
3778         return skb->len;
3779 }
3780
3781 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3782                                   struct sk_buff *skb,
3783                                   const struct nlmsghdr *nlh,
3784                                   const struct nlattr * const cda[],
3785                                   struct netlink_ext_ack *extack)
3786 {
3787         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3788                 struct netlink_dump_control c = {
3789                         .dump = ctnetlink_exp_stat_cpu_dump,
3790                 };
3791                 return netlink_dump_start(ctnl, skb, nlh, &c);
3792         }
3793
3794         return 0;
3795 }
3796
3797 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3798 static struct nf_ct_event_notifier ctnl_notifier = {
3799         .fcn = ctnetlink_conntrack_event,
3800 };
3801
3802 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3803         .fcn = ctnetlink_expect_event,
3804 };
3805 #endif
3806
3807 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3808         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3809                                             .attr_count = CTA_MAX,
3810                                             .policy = ct_nla_policy },
3811         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3812                                             .attr_count = CTA_MAX,
3813                                             .policy = ct_nla_policy },
3814         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3815                                             .attr_count = CTA_MAX,
3816                                             .policy = ct_nla_policy },
3817         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3818                                             .attr_count = CTA_MAX,
3819                                             .policy = ct_nla_policy },
3820         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3821         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3822         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3823         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3824 };
3825
3826 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3827         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3828                                             .attr_count = CTA_EXPECT_MAX,
3829                                             .policy = exp_nla_policy },
3830         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3831                                             .attr_count = CTA_EXPECT_MAX,
3832                                             .policy = exp_nla_policy },
3833         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3834                                             .attr_count = CTA_EXPECT_MAX,
3835                                             .policy = exp_nla_policy },
3836         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3837 };
3838
3839 static const struct nfnetlink_subsystem ctnl_subsys = {
3840         .name                           = "conntrack",
3841         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3842         .cb_count                       = IPCTNL_MSG_MAX,
3843         .cb                             = ctnl_cb,
3844 };
3845
3846 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3847         .name                           = "conntrack_expect",
3848         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3849         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3850         .cb                             = ctnl_exp_cb,
3851 };
3852
3853 MODULE_ALIAS("ip_conntrack_netlink");
3854 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3855 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3856
3857 static int __net_init ctnetlink_net_init(struct net *net)
3858 {
3859 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3860         int ret;
3861
3862         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3863         if (ret < 0) {
3864                 pr_err("ctnetlink_init: cannot register notifier.\n");
3865                 goto err_out;
3866         }
3867
3868         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3869         if (ret < 0) {
3870                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3871                 goto err_unreg_notifier;
3872         }
3873 #endif
3874         return 0;
3875
3876 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3877 err_unreg_notifier:
3878         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3879 err_out:
3880         return ret;
3881 #endif
3882 }
3883
3884 static void ctnetlink_net_exit(struct net *net)
3885 {
3886 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3887         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3888         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3889 #endif
3890 }
3891
3892 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3893 {
3894         struct net *net;
3895
3896         list_for_each_entry(net, net_exit_list, exit_list)
3897                 ctnetlink_net_exit(net);
3898
3899         /* wait for other cpus until they are done with ctnl_notifiers */
3900         synchronize_rcu();
3901 }
3902
3903 static struct pernet_operations ctnetlink_net_ops = {
3904         .init           = ctnetlink_net_init,
3905         .exit_batch     = ctnetlink_net_exit_batch,
3906 };
3907
3908 static int __init ctnetlink_init(void)
3909 {
3910         int ret;
3911
3912         ret = nfnetlink_subsys_register(&ctnl_subsys);
3913         if (ret < 0) {
3914                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3915                 goto err_out;
3916         }
3917
3918         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3919         if (ret < 0) {
3920                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3921                 goto err_unreg_subsys;
3922         }
3923
3924         ret = register_pernet_subsys(&ctnetlink_net_ops);
3925         if (ret < 0) {
3926                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3927                 goto err_unreg_exp_subsys;
3928         }
3929 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3930         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3931         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3932 #endif
3933         return 0;
3934
3935 err_unreg_exp_subsys:
3936         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3937 err_unreg_subsys:
3938         nfnetlink_subsys_unregister(&ctnl_subsys);
3939 err_out:
3940         return ret;
3941 }
3942
3943 static void __exit ctnetlink_exit(void)
3944 {
3945         unregister_pernet_subsys(&ctnetlink_net_ops);
3946         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3947         nfnetlink_subsys_unregister(&ctnl_subsys);
3948 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3949         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3950 #endif
3951         synchronize_rcu();
3952 }
3953
3954 module_init(ctnetlink_init);
3955 module_exit(ctnetlink_exit);