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
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.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 #ifdef CONFIG_NF_NAT_NEEDED
  50 #include <net/netfilter/nf_nat_core.h>
  51 #include <net/netfilter/nf_nat_l4proto.h>
  52 #include <net/netfilter/nf_nat_helper.h>
  53 #endif
  54
  55 #include <linux/netfilter/nfnetlink.h>
  56 #include <linux/netfilter/nfnetlink_conntrack.h>
  57
  58 MODULE_LICENSE("GPL");
  59
  60 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  61                                 const struct nf_conntrack_tuple *tuple,
  62                                 const struct nf_conntrack_l4proto *l4proto)
  63 {
  64         int ret = 0;
  65         struct nlattr *nest_parms;
  66
  67         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  68         if (!nest_parms)
  69                 goto nla_put_failure;
  70         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  71                 goto nla_put_failure;
  72
  73         if (likely(l4proto->tuple_to_nlattr))
  74                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  75
  76         nla_nest_end(skb, nest_parms);
  77
  78         return ret;
  79
  80 nla_put_failure:
  81         return -1;
  82 }
  83
  84 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  85                                     const struct nf_conntrack_tuple *tuple,
  86                                     const struct nf_conntrack_l3proto *l3proto)
  87 {
  88         int ret = 0;
  89         struct nlattr *nest_parms;
  90
  91         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  92         if (!nest_parms)
  93                 goto nla_put_failure;
  94
  95         if (likely(l3proto->tuple_to_nlattr))
  96                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  97
  98         nla_nest_end(skb, nest_parms);
  99
 100         return ret;
 101
 102 nla_put_failure:
 103         return -1;
 104 }
 105
 106 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 107                                  const struct nf_conntrack_tuple *tuple)
 108 {
 109         const struct nf_conntrack_l3proto *l3proto;
 110         const struct nf_conntrack_l4proto *l4proto;
 111         int ret;
 112
 113         rcu_read_lock();
 114         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 115         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 116
 117         if (ret >= 0) {
 118                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 119                                                tuple->dst.protonum);
 120                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 121         }
 122         rcu_read_unlock();
 123         return ret;
 124 }
 125
 126 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 127                                   const struct nf_conntrack_zone *zone, int dir)
 128 {
 129         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 130                 return 0;
 131         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 132                 goto nla_put_failure;
 133         return 0;
 134
 135 nla_put_failure:
 136         return -1;
 137 }
 138
 139 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 140 {
 141         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 142                 goto nla_put_failure;
 143         return 0;
 144
 145 nla_put_failure:
 146         return -1;
 147 }
 148
 149 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 150 {
 151         long timeout = nf_ct_expires(ct) / HZ;
 152
 153         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 154                 goto nla_put_failure;
 155         return 0;
 156
 157 nla_put_failure:
 158         return -1;
 159 }
 160
 161 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 162 {
 163         const struct nf_conntrack_l4proto *l4proto;
 164         struct nlattr *nest_proto;
 165         int ret;
 166
 167         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 168         if (!l4proto->to_nlattr)
 169                 return 0;
 170
 171         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 172         if (!nest_proto)
 173                 goto nla_put_failure;
 174
 175         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 176
 177         nla_nest_end(skb, nest_proto);
 178
 179         return ret;
 180
 181 nla_put_failure:
 182         return -1;
 183 }
 184
 185 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 186                                    const struct nf_conn *ct)
 187 {
 188         struct nlattr *nest_helper;
 189         const struct nf_conn_help *help = nfct_help(ct);
 190         struct nf_conntrack_helper *helper;
 191
 192         if (!help)
 193                 return 0;
 194
 195         helper = rcu_dereference(help->helper);
 196         if (!helper)
 197                 goto out;
 198
 199         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 200         if (!nest_helper)
 201                 goto nla_put_failure;
 202         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 203                 goto nla_put_failure;
 204
 205         if (helper->to_nlattr)
 206                 helper->to_nlattr(skb, ct);
 207
 208         nla_nest_end(skb, nest_helper);
 209 out:
 210         return 0;
 211
 212 nla_put_failure:
 213         return -1;
 214 }
 215
 216 static int
 217 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 218               enum ip_conntrack_dir dir, int type)
 219 {
 220         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 221         struct nf_conn_counter *counter = acct->counter;
 222         struct nlattr *nest_count;
 223         u64 pkts, bytes;
 224
 225         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 226                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 227                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 228         } else {
 229                 pkts = atomic64_read(&counter[dir].packets);
 230                 bytes = atomic64_read(&counter[dir].bytes);
 231         }
 232
 233         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 234         if (!nest_count)
 235                 goto nla_put_failure;
 236
 237         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 238                          CTA_COUNTERS_PAD) ||
 239             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 240                          CTA_COUNTERS_PAD))
 241                 goto nla_put_failure;
 242
 243         nla_nest_end(skb, nest_count);
 244
 245         return 0;
 246
 247 nla_put_failure:
 248         return -1;
 249 }
 250
 251 static int
 252 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 253 {
 254         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 255
 256         if (!acct)
 257                 return 0;
 258
 259         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 260                 return -1;
 261         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 262                 return -1;
 263
 264         return 0;
 265 }
 266
 267 static int
 268 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 269 {
 270         struct nlattr *nest_count;
 271         const struct nf_conn_tstamp *tstamp;
 272
 273         tstamp = nf_conn_tstamp_find(ct);
 274         if (!tstamp)
 275                 return 0;
 276
 277         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 278         if (!nest_count)
 279                 goto nla_put_failure;
 280
 281         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 282                          CTA_TIMESTAMP_PAD) ||
 283             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 284                                                cpu_to_be64(tstamp->stop),
 285                                                CTA_TIMESTAMP_PAD)))
 286                 goto nla_put_failure;
 287         nla_nest_end(skb, nest_count);
 288
 289         return 0;
 290
 291 nla_put_failure:
 292         return -1;
 293 }
 294
 295 #ifdef CONFIG_NF_CONNTRACK_MARK
 296 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 297 {
 298         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 299                 goto nla_put_failure;
 300         return 0;
 301
 302 nla_put_failure:
 303         return -1;
 304 }
 305 #else
 306 #define ctnetlink_dump_mark(a, b) (0)
 307 #endif
 308
 309 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 310 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 311 {
 312         struct nlattr *nest_secctx;
 313         int len, ret;
 314         char *secctx;
 315
 316         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 317         if (ret)
 318                 return 0;
 319
 320         ret = -1;
 321         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 322         if (!nest_secctx)
 323                 goto nla_put_failure;
 324
 325         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 326                 goto nla_put_failure;
 327         nla_nest_end(skb, nest_secctx);
 328
 329         ret = 0;
 330 nla_put_failure:
 331         security_release_secctx(secctx, len);
 332         return ret;
 333 }
 334 #else
 335 #define ctnetlink_dump_secctx(a, b) (0)
 336 #endif
 337
 338 #ifdef CONFIG_NF_CONNTRACK_LABELS
 339 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 340 {
 341         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 342
 343         if (!labels)
 344                 return 0;
 345         return nla_total_size(sizeof(labels->bits));
 346 }
 347
 348 static int
 349 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 350 {
 351         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 352         unsigned int i;
 353
 354         if (!labels)
 355                 return 0;
 356
 357         i = 0;
 358         do {
 359                 if (labels->bits[i] != 0)
 360                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 361                                        labels->bits);
 362                 i++;
 363         } while (i < ARRAY_SIZE(labels->bits));
 364
 365         return 0;
 366 }
 367 #else
 368 #define ctnetlink_dump_labels(a, b) (0)
 369 #define ctnetlink_label_size(a) (0)
 370 #endif
 371
 372 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 373
 374 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 375 {
 376         struct nlattr *nest_parms;
 377
 378         if (!(ct->status & IPS_EXPECTED))
 379                 return 0;
 380
 381         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 382         if (!nest_parms)
 383                 goto nla_put_failure;
 384         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 385                 goto nla_put_failure;
 386         nla_nest_end(skb, nest_parms);
 387
 388         return 0;
 389
 390 nla_put_failure:
 391         return -1;
 392 }
 393
 394 static int
 395 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 396 {
 397         struct nlattr *nest_parms;
 398
 399         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 400         if (!nest_parms)
 401                 goto nla_put_failure;
 402
 403         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 404                          htonl(seq->correction_pos)) ||
 405             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 406                          htonl(seq->offset_before)) ||
 407             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 408                          htonl(seq->offset_after)))
 409                 goto nla_put_failure;
 410
 411         nla_nest_end(skb, nest_parms);
 412
 413         return 0;
 414
 415 nla_put_failure:
 416         return -1;
 417 }
 418
 419 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 420 {
 421         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 422         struct nf_ct_seqadj *seq;
 423
 424         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 425                 return 0;
 426
 427         spin_lock_bh(&ct->lock);
 428         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 429         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 430                 goto err;
 431
 432         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 433         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 434                 goto err;
 435
 436         spin_unlock_bh(&ct->lock);
 437         return 0;
 438 err:
 439         spin_unlock_bh(&ct->lock);
 440         return -1;
 441 }
 442
 443 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 444 {
 445         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 446         struct nlattr *nest_parms;
 447
 448         if (!synproxy)
 449                 return 0;
 450
 451         nest_parms = nla_nest_start(skb, CTA_SYNPROXY | NLA_F_NESTED);
 452         if (!nest_parms)
 453                 goto nla_put_failure;
 454
 455         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 456             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 457             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 458                 goto nla_put_failure;
 459
 460         nla_nest_end(skb, nest_parms);
 461
 462         return 0;
 463
 464 nla_put_failure:
 465         return -1;
 466 }
 467
 468 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 469 {
 470         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 471                 goto nla_put_failure;
 472         return 0;
 473
 474 nla_put_failure:
 475         return -1;
 476 }
 477
 478 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 479 {
 480         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 481                 goto nla_put_failure;
 482         return 0;
 483
 484 nla_put_failure:
 485         return -1;
 486 }
 487
 488 static int
 489 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 490                     struct nf_conn *ct)
 491 {
 492         const struct nf_conntrack_zone *zone;
 493         struct nlmsghdr *nlh;
 494         struct nfgenmsg *nfmsg;
 495         struct nlattr *nest_parms;
 496         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 497
 498         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 499         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 500         if (nlh == NULL)
 501                 goto nlmsg_failure;
 502
 503         nfmsg = nlmsg_data(nlh);
 504         nfmsg->nfgen_family = nf_ct_l3num(ct);
 505         nfmsg->version      = NFNETLINK_V0;
 506         nfmsg->res_id       = 0;
 507
 508         zone = nf_ct_zone(ct);
 509
 510         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 511         if (!nest_parms)
 512                 goto nla_put_failure;
 513         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 514                 goto nla_put_failure;
 515         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 516                                    NF_CT_ZONE_DIR_ORIG) < 0)
 517                 goto nla_put_failure;
 518         nla_nest_end(skb, nest_parms);
 519
 520         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 521         if (!nest_parms)
 522                 goto nla_put_failure;
 523         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 524                 goto nla_put_failure;
 525         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 526                                    NF_CT_ZONE_DIR_REPL) < 0)
 527                 goto nla_put_failure;
 528         nla_nest_end(skb, nest_parms);
 529
 530         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 531                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 532                 goto nla_put_failure;
 533
 534         if (ctnetlink_dump_status(skb, ct) < 0 ||
 535             ctnetlink_dump_timeout(skb, ct) < 0 ||
 536             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 537             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 538             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 539             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 540             ctnetlink_dump_mark(skb, ct) < 0 ||
 541             ctnetlink_dump_secctx(skb, ct) < 0 ||
 542             ctnetlink_dump_labels(skb, ct) < 0 ||
 543             ctnetlink_dump_id(skb, ct) < 0 ||
 544             ctnetlink_dump_use(skb, ct) < 0 ||
 545             ctnetlink_dump_master(skb, ct) < 0 ||
 546             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 547             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 548                 goto nla_put_failure;
 549
 550         nlmsg_end(skb, nlh);
 551         return skb->len;
 552
 553 nlmsg_failure:
 554 nla_put_failure:
 555         nlmsg_cancel(skb, nlh);
 556         return -1;
 557 }
 558
 559 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 560 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 561 {
 562         const struct nf_conntrack_l3proto *l3proto;
 563         const struct nf_conntrack_l4proto *l4proto;
 564         size_t len, len4 = 0;
 565
 566         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 567         len = l3proto->nla_size;
 568         len *= 3u; /* ORIG, REPLY, MASTER */
 569
 570         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 571         len += l4proto->nlattr_size;
 572         if (l4proto->nlattr_tuple_size) {
 573                 len4 = l4proto->nlattr_tuple_size();
 574                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 575         }
 576
 577         return len + len4;
 578 }
 579 #endif
 580
 581 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 582 {
 583         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 584                 return 0;
 585         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 586                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 587                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 588                ;
 589 }
 590
 591 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 592 {
 593 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 594         int len, ret;
 595
 596         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 597         if (ret)
 598                 return 0;
 599
 600         return nla_total_size(0) /* CTA_SECCTX */
 601                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 602 #else
 603         return 0;
 604 #endif
 605 }
 606
 607 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 608 {
 609 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 610         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 611                 return 0;
 612         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 613 #else
 614         return 0;
 615 #endif
 616 }
 617
 618 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 619 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 620 {
 621         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 622                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 623                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 624                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 625                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 626                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 627                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 628                + ctnetlink_acct_size(ct)
 629                + ctnetlink_timestamp_size(ct)
 630                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 631                + nla_total_size(0) /* CTA_PROTOINFO */
 632                + nla_total_size(0) /* CTA_HELP */
 633                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 634                + ctnetlink_secctx_size(ct)
 635 #ifdef CONFIG_NF_NAT_NEEDED
 636                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 637                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 638 #endif
 639 #ifdef CONFIG_NF_CONNTRACK_MARK
 640                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 641 #endif
 642 #ifdef CONFIG_NF_CONNTRACK_ZONES
 643                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 644 #endif
 645                + ctnetlink_proto_size(ct)
 646                + ctnetlink_label_size(ct)
 647                ;
 648 }
 649
 650 static int
 651 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 652 {
 653         const struct nf_conntrack_zone *zone;
 654         struct net *net;
 655         struct nlmsghdr *nlh;
 656         struct nfgenmsg *nfmsg;
 657         struct nlattr *nest_parms;
 658         struct nf_conn *ct = item->ct;
 659         struct sk_buff *skb;
 660         unsigned int type;
 661         unsigned int flags = 0, group;
 662         int err;
 663
 664         if (events & (1 << IPCT_DESTROY)) {
 665                 type = IPCTNL_MSG_CT_DELETE;
 666                 group = NFNLGRP_CONNTRACK_DESTROY;
 667         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 668                 type = IPCTNL_MSG_CT_NEW;
 669                 flags = NLM_F_CREATE|NLM_F_EXCL;
 670                 group = NFNLGRP_CONNTRACK_NEW;
 671         } else if (events) {
 672                 type = IPCTNL_MSG_CT_NEW;
 673                 group = NFNLGRP_CONNTRACK_UPDATE;
 674         } else
 675                 return 0;
 676
 677         net = nf_ct_net(ct);
 678         if (!item->report && !nfnetlink_has_listeners(net, group))
 679                 return 0;
 680
 681         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 682         if (skb == NULL)
 683                 goto errout;
 684
 685         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 686         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 687         if (nlh == NULL)
 688                 goto nlmsg_failure;
 689
 690         nfmsg = nlmsg_data(nlh);
 691         nfmsg->nfgen_family = nf_ct_l3num(ct);
 692         nfmsg->version  = NFNETLINK_V0;
 693         nfmsg->res_id   = 0;
 694
 695         zone = nf_ct_zone(ct);
 696
 697         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 698         if (!nest_parms)
 699                 goto nla_put_failure;
 700         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 701                 goto nla_put_failure;
 702         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 703                                    NF_CT_ZONE_DIR_ORIG) < 0)
 704                 goto nla_put_failure;
 705         nla_nest_end(skb, nest_parms);
 706
 707         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 708         if (!nest_parms)
 709                 goto nla_put_failure;
 710         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 711                 goto nla_put_failure;
 712         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 713                                    NF_CT_ZONE_DIR_REPL) < 0)
 714                 goto nla_put_failure;
 715         nla_nest_end(skb, nest_parms);
 716
 717         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 718                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 719                 goto nla_put_failure;
 720
 721         if (ctnetlink_dump_id(skb, ct) < 0)
 722                 goto nla_put_failure;
 723
 724         if (ctnetlink_dump_status(skb, ct) < 0)
 725                 goto nla_put_failure;
 726
 727         if (events & (1 << IPCT_DESTROY)) {
 728                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 729                     ctnetlink_dump_timestamp(skb, ct) < 0)
 730                         goto nla_put_failure;
 731         } else {
 732                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 733                         goto nla_put_failure;
 734
 735                 if (events & (1 << IPCT_PROTOINFO)
 736                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 737                         goto nla_put_failure;
 738
 739                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 740                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 741                         goto nla_put_failure;
 742
 743 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 744                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 745                     && ctnetlink_dump_secctx(skb, ct) < 0)
 746                         goto nla_put_failure;
 747 #endif
 748                 if (events & (1 << IPCT_LABEL) &&
 749                      ctnetlink_dump_labels(skb, ct) < 0)
 750                         goto nla_put_failure;
 751
 752                 if (events & (1 << IPCT_RELATED) &&
 753                     ctnetlink_dump_master(skb, ct) < 0)
 754                         goto nla_put_failure;
 755
 756                 if (events & (1 << IPCT_SEQADJ) &&
 757                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 758                         goto nla_put_failure;
 759
 760                 if (events & (1 << IPCT_SYNPROXY) &&
 761                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 762                         goto nla_put_failure;
 763         }
 764
 765 #ifdef CONFIG_NF_CONNTRACK_MARK
 766         if ((events & (1 << IPCT_MARK) || ct->mark)
 767             && ctnetlink_dump_mark(skb, ct) < 0)
 768                 goto nla_put_failure;
 769 #endif
 770         nlmsg_end(skb, nlh);
 771         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 772                              GFP_ATOMIC);
 773         if (err == -ENOBUFS || err == -EAGAIN)
 774                 return -ENOBUFS;
 775
 776         return 0;
 777
 778 nla_put_failure:
 779         nlmsg_cancel(skb, nlh);
 780 nlmsg_failure:
 781         kfree_skb(skb);
 782 errout:
 783         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 784                 return -ENOBUFS;
 785
 786         return 0;
 787 }
 788 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 789
 790 static int ctnetlink_done(struct netlink_callback *cb)
 791 {
 792         if (cb->args[1])
 793                 nf_ct_put((struct nf_conn *)cb->args[1]);
 794         kfree(cb->data);
 795         return 0;
 796 }
 797
 798 struct ctnetlink_filter {
 799         struct {
 800                 u_int32_t val;
 801                 u_int32_t mask;
 802         } mark;
 803 };
 804
 805 static struct ctnetlink_filter *
 806 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 807 {
 808 #ifdef CONFIG_NF_CONNTRACK_MARK
 809         struct ctnetlink_filter *filter;
 810
 811         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 812         if (filter == NULL)
 813                 return ERR_PTR(-ENOMEM);
 814
 815         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 816         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 817
 818         return filter;
 819 #else
 820         return ERR_PTR(-EOPNOTSUPP);
 821 #endif
 822 }
 823
 824 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 825 {
 826         struct ctnetlink_filter *filter = data;
 827
 828         if (filter == NULL)
 829                 return 1;
 830
 831 #ifdef CONFIG_NF_CONNTRACK_MARK
 832         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 833                 return 1;
 834 #endif
 835
 836         return 0;
 837 }
 838
 839 static int
 840 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 841 {
 842         struct net *net = sock_net(skb->sk);
 843         struct nf_conn *ct, *last;
 844         struct nf_conntrack_tuple_hash *h;
 845         struct hlist_nulls_node *n;
 846         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 847         u_int8_t l3proto = nfmsg->nfgen_family;
 848         struct nf_conn *nf_ct_evict[8];
 849         int res, i;
 850         spinlock_t *lockp;
 851
 852         last = (struct nf_conn *)cb->args[1];
 853         i = 0;
 854
 855         local_bh_disable();
 856         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 857 restart:
 858                 while (i) {
 859                         i--;
 860                         if (nf_ct_should_gc(nf_ct_evict[i]))
 861                                 nf_ct_kill(nf_ct_evict[i]);
 862                         nf_ct_put(nf_ct_evict[i]);
 863                 }
 864
 865                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 866                 nf_conntrack_lock(lockp);
 867                 if (cb->args[0] >= nf_conntrack_htable_size) {
 868                         spin_unlock(lockp);
 869                         goto out;
 870                 }
 871                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 872                                            hnnode) {
 873                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 874                                 continue;
 875                         ct = nf_ct_tuplehash_to_ctrack(h);
 876                         if (nf_ct_is_expired(ct)) {
 877                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 878                                     atomic_inc_not_zero(&ct->ct_general.use))
 879                                         nf_ct_evict[i++] = ct;
 880                                 continue;
 881                         }
 882
 883                         if (!net_eq(net, nf_ct_net(ct)))
 884                                 continue;
 885
 886                         /* Dump entries of a given L3 protocol number.
 887                          * If it is not specified, ie. l3proto == 0,
 888                          * then dump everything. */
 889                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 890                                 continue;
 891                         if (cb->args[1]) {
 892                                 if (ct != last)
 893                                         continue;
 894                                 cb->args[1] = 0;
 895                         }
 896                         if (!ctnetlink_filter_match(ct, cb->data))
 897                                 continue;
 898
 899                         rcu_read_lock();
 900                         res =
 901                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 902                                             cb->nlh->nlmsg_seq,
 903                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 904                                             ct);
 905                         rcu_read_unlock();
 906                         if (res < 0) {
 907                                 nf_conntrack_get(&ct->ct_general);
 908                                 cb->args[1] = (unsigned long)ct;
 909                                 spin_unlock(lockp);
 910                                 goto out;
 911                         }
 912                 }
 913                 spin_unlock(lockp);
 914                 if (cb->args[1]) {
 915                         cb->args[1] = 0;
 916                         goto restart;
 917                 }
 918         }
 919 out:
 920         local_bh_enable();
 921         if (last) {
 922                 /* nf ct hash resize happened, now clear the leftover. */
 923                 if ((struct nf_conn *)cb->args[1] == last)
 924                         cb->args[1] = 0;
 925
 926                 nf_ct_put(last);
 927         }
 928
 929         while (i) {
 930                 i--;
 931                 if (nf_ct_should_gc(nf_ct_evict[i]))
 932                         nf_ct_kill(nf_ct_evict[i]);
 933                 nf_ct_put(nf_ct_evict[i]);
 934         }
 935
 936         return skb->len;
 937 }
 938
 939 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
 940                                     struct nf_conntrack_tuple *tuple)
 941 {
 942         struct nlattr *tb[CTA_IP_MAX+1];
 943         struct nf_conntrack_l3proto *l3proto;
 944         int ret = 0;
 945
 946         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL, NULL);
 947         if (ret < 0)
 948                 return ret;
 949
 950         rcu_read_lock();
 951         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 952
 953         if (likely(l3proto->nlattr_to_tuple)) {
 954                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 955                                           l3proto->nla_policy, NULL);
 956                 if (ret == 0)
 957                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 958         }
 959
 960         rcu_read_unlock();
 961
 962         return ret;
 963 }
 964
 965 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 966         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 967 };
 968
 969 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
 970                                        struct nf_conntrack_tuple *tuple)
 971 {
 972         const struct nf_conntrack_l4proto *l4proto;
 973         struct nlattr *tb[CTA_PROTO_MAX+1];
 974         int ret = 0;
 975
 976         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy,
 977                                NULL);
 978         if (ret < 0)
 979                 return ret;
 980
 981         if (!tb[CTA_PROTO_NUM])
 982                 return -EINVAL;
 983         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 984
 985         rcu_read_lock();
 986         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 987
 988         if (likely(l4proto->nlattr_to_tuple)) {
 989                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 990                                           l4proto->nla_policy, NULL);
 991                 if (ret == 0)
 992                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 993         }
 994
 995         rcu_read_unlock();
 996
 997         return ret;
 998 }
 999
1000 static int
1001 ctnetlink_parse_zone(const struct nlattr *attr,
1002                      struct nf_conntrack_zone *zone)
1003 {
1004         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1005                         NF_CT_DEFAULT_ZONE_DIR, 0);
1006 #ifdef CONFIG_NF_CONNTRACK_ZONES
1007         if (attr)
1008                 zone->id = ntohs(nla_get_be16(attr));
1009 #else
1010         if (attr)
1011                 return -EOPNOTSUPP;
1012 #endif
1013         return 0;
1014 }
1015
1016 static int
1017 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1018                            struct nf_conntrack_zone *zone)
1019 {
1020         int ret;
1021
1022         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1023                 return -EINVAL;
1024
1025         ret = ctnetlink_parse_zone(attr, zone);
1026         if (ret < 0)
1027                 return ret;
1028
1029         if (type == CTA_TUPLE_REPLY)
1030                 zone->dir = NF_CT_ZONE_DIR_REPL;
1031         else
1032                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1033
1034         return 0;
1035 }
1036
1037 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1038         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1039         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1040         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1041 };
1042
1043 static int
1044 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1045                       struct nf_conntrack_tuple *tuple, u32 type,
1046                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1047 {
1048         struct nlattr *tb[CTA_TUPLE_MAX+1];
1049         int err;
1050
1051         memset(tuple, 0, sizeof(*tuple));
1052
1053         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1054                                NULL);
1055         if (err < 0)
1056                 return err;
1057
1058         if (!tb[CTA_TUPLE_IP])
1059                 return -EINVAL;
1060
1061         tuple->src.l3num = l3num;
1062
1063         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1064         if (err < 0)
1065                 return err;
1066
1067         if (!tb[CTA_TUPLE_PROTO])
1068                 return -EINVAL;
1069
1070         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1071         if (err < 0)
1072                 return err;
1073
1074         if (tb[CTA_TUPLE_ZONE]) {
1075                 if (!zone)
1076                         return -EINVAL;
1077
1078                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1079                                                  type, zone);
1080                 if (err < 0)
1081                         return err;
1082         }
1083
1084         /* orig and expect tuples get DIR_ORIGINAL */
1085         if (type == CTA_TUPLE_REPLY)
1086                 tuple->dst.dir = IP_CT_DIR_REPLY;
1087         else
1088                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1089
1090         return 0;
1091 }
1092
1093 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1094         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1095                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1096 };
1097
1098 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1099                                 struct nlattr **helpinfo)
1100 {
1101         int err;
1102         struct nlattr *tb[CTA_HELP_MAX+1];
1103
1104         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1105         if (err < 0)
1106                 return err;
1107
1108         if (!tb[CTA_HELP_NAME])
1109                 return -EINVAL;
1110
1111         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1112
1113         if (tb[CTA_HELP_INFO])
1114                 *helpinfo = tb[CTA_HELP_INFO];
1115
1116         return 0;
1117 }
1118
1119 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1120         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1121         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1122         [CTA_STATUS]            = { .type = NLA_U32 },
1123         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1124         [CTA_HELP]              = { .type = NLA_NESTED },
1125         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1126         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1127         [CTA_MARK]              = { .type = NLA_U32 },
1128         [CTA_ID]                = { .type = NLA_U32 },
1129         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1130         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1131         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1132         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1133         [CTA_ZONE]              = { .type = NLA_U16 },
1134         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1135         [CTA_LABELS]            = { .type = NLA_BINARY,
1136                                     .len = NF_CT_LABELS_MAX_SIZE },
1137         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1138                                     .len = NF_CT_LABELS_MAX_SIZE },
1139 };
1140
1141 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1142 {
1143         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1144                 return 0;
1145
1146         return ctnetlink_filter_match(ct, data);
1147 }
1148
1149 static int ctnetlink_flush_conntrack(struct net *net,
1150                                      const struct nlattr * const cda[],
1151                                      u32 portid, int report)
1152 {
1153         struct ctnetlink_filter *filter = NULL;
1154
1155         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1156                 filter = ctnetlink_alloc_filter(cda);
1157                 if (IS_ERR(filter))
1158                         return PTR_ERR(filter);
1159         }
1160
1161         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1162                                   portid, report);
1163         kfree(filter);
1164
1165         return 0;
1166 }
1167
1168 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1169                                    struct sk_buff *skb,
1170                                    const struct nlmsghdr *nlh,
1171                                    const struct nlattr * const cda[],
1172                                    struct netlink_ext_ack *extack)
1173 {
1174         struct nf_conntrack_tuple_hash *h;
1175         struct nf_conntrack_tuple tuple;
1176         struct nf_conn *ct;
1177         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1178         u_int8_t u3 = nfmsg->nfgen_family;
1179         struct nf_conntrack_zone zone;
1180         int err;
1181
1182         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1183         if (err < 0)
1184                 return err;
1185
1186         if (cda[CTA_TUPLE_ORIG])
1187                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1188                                             u3, &zone);
1189         else if (cda[CTA_TUPLE_REPLY])
1190                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1191                                             u3, &zone);
1192         else {
1193                 return ctnetlink_flush_conntrack(net, cda,
1194                                                  NETLINK_CB(skb).portid,
1195                                                  nlmsg_report(nlh));
1196         }
1197
1198         if (err < 0)
1199                 return err;
1200
1201         h = nf_conntrack_find_get(net, &zone, &tuple);
1202         if (!h)
1203                 return -ENOENT;
1204
1205         ct = nf_ct_tuplehash_to_ctrack(h);
1206
1207         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1208                 nf_ct_put(ct);
1209                 return -EBUSY;
1210         }
1211
1212         if (cda[CTA_ID]) {
1213                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1214                 if (id != (u32)(unsigned long)ct) {
1215                         nf_ct_put(ct);
1216                         return -ENOENT;
1217                 }
1218         }
1219
1220         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1221         nf_ct_put(ct);
1222
1223         return 0;
1224 }
1225
1226 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1227                                    struct sk_buff *skb,
1228                                    const struct nlmsghdr *nlh,
1229                                    const struct nlattr * const cda[],
1230                                    struct netlink_ext_ack *extack)
1231 {
1232         struct nf_conntrack_tuple_hash *h;
1233         struct nf_conntrack_tuple tuple;
1234         struct nf_conn *ct;
1235         struct sk_buff *skb2 = NULL;
1236         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1237         u_int8_t u3 = nfmsg->nfgen_family;
1238         struct nf_conntrack_zone zone;
1239         int err;
1240
1241         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1242                 struct netlink_dump_control c = {
1243                         .dump = ctnetlink_dump_table,
1244                         .done = ctnetlink_done,
1245                 };
1246
1247                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1248                         struct ctnetlink_filter *filter;
1249
1250                         filter = ctnetlink_alloc_filter(cda);
1251                         if (IS_ERR(filter))
1252                                 return PTR_ERR(filter);
1253
1254                         c.data = filter;
1255                 }
1256                 return netlink_dump_start(ctnl, skb, nlh, &c);
1257         }
1258
1259         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1260         if (err < 0)
1261                 return err;
1262
1263         if (cda[CTA_TUPLE_ORIG])
1264                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1265                                             u3, &zone);
1266         else if (cda[CTA_TUPLE_REPLY])
1267                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1268                                             u3, &zone);
1269         else
1270                 return -EINVAL;
1271
1272         if (err < 0)
1273                 return err;
1274
1275         h = nf_conntrack_find_get(net, &zone, &tuple);
1276         if (!h)
1277                 return -ENOENT;
1278
1279         ct = nf_ct_tuplehash_to_ctrack(h);
1280
1281         err = -ENOMEM;
1282         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1283         if (skb2 == NULL) {
1284                 nf_ct_put(ct);
1285                 return -ENOMEM;
1286         }
1287
1288         rcu_read_lock();
1289         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1290                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1291         rcu_read_unlock();
1292         nf_ct_put(ct);
1293         if (err <= 0)
1294                 goto free;
1295
1296         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1297         if (err < 0)
1298                 goto out;
1299
1300         return 0;
1301
1302 free:
1303         kfree_skb(skb2);
1304 out:
1305         /* this avoids a loop in nfnetlink. */
1306         return err == -EAGAIN ? -ENOBUFS : err;
1307 }
1308
1309 static int ctnetlink_done_list(struct netlink_callback *cb)
1310 {
1311         if (cb->args[1])
1312                 nf_ct_put((struct nf_conn *)cb->args[1]);
1313         return 0;
1314 }
1315
1316 static int
1317 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1318 {
1319         struct nf_conn *ct, *last;
1320         struct nf_conntrack_tuple_hash *h;
1321         struct hlist_nulls_node *n;
1322         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1323         u_int8_t l3proto = nfmsg->nfgen_family;
1324         int res;
1325         int cpu;
1326         struct hlist_nulls_head *list;
1327         struct net *net = sock_net(skb->sk);
1328
1329         if (cb->args[2])
1330                 return 0;
1331
1332         last = (struct nf_conn *)cb->args[1];
1333
1334         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1335                 struct ct_pcpu *pcpu;
1336
1337                 if (!cpu_possible(cpu))
1338                         continue;
1339
1340                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1341                 spin_lock_bh(&pcpu->lock);
1342                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1343 restart:
1344                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1345                         ct = nf_ct_tuplehash_to_ctrack(h);
1346                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1347                                 continue;
1348                         if (cb->args[1]) {
1349                                 if (ct != last)
1350                                         continue;
1351                                 cb->args[1] = 0;
1352                         }
1353                         rcu_read_lock();
1354                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1355                                                   cb->nlh->nlmsg_seq,
1356                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1357                                                   ct);
1358                         rcu_read_unlock();
1359                         if (res < 0) {
1360                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1361                                         continue;
1362                                 cb->args[0] = cpu;
1363                                 cb->args[1] = (unsigned long)ct;
1364                                 spin_unlock_bh(&pcpu->lock);
1365                                 goto out;
1366                         }
1367                 }
1368                 if (cb->args[1]) {
1369                         cb->args[1] = 0;
1370                         goto restart;
1371                 }
1372                 spin_unlock_bh(&pcpu->lock);
1373         }
1374         cb->args[2] = 1;
1375 out:
1376         if (last)
1377                 nf_ct_put(last);
1378
1379         return skb->len;
1380 }
1381
1382 static int
1383 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1384 {
1385         return ctnetlink_dump_list(skb, cb, true);
1386 }
1387
1388 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1389                                   struct sk_buff *skb,
1390                                   const struct nlmsghdr *nlh,
1391                                   const struct nlattr * const cda[],
1392                                   struct netlink_ext_ack *extack)
1393 {
1394         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1395                 struct netlink_dump_control c = {
1396                         .dump = ctnetlink_dump_dying,
1397                         .done = ctnetlink_done_list,
1398                 };
1399                 return netlink_dump_start(ctnl, skb, nlh, &c);
1400         }
1401
1402         return -EOPNOTSUPP;
1403 }
1404
1405 static int
1406 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1407 {
1408         return ctnetlink_dump_list(skb, cb, false);
1409 }
1410
1411 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1412                                         struct sk_buff *skb,
1413                                         const struct nlmsghdr *nlh,
1414                                         const struct nlattr * const cda[],
1415                                         struct netlink_ext_ack *extack)
1416 {
1417         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1418                 struct netlink_dump_control c = {
1419                         .dump = ctnetlink_dump_unconfirmed,
1420                         .done = ctnetlink_done_list,
1421                 };
1422                 return netlink_dump_start(ctnl, skb, nlh, &c);
1423         }
1424
1425         return -EOPNOTSUPP;
1426 }
1427
1428 #ifdef CONFIG_NF_NAT_NEEDED
1429 static int
1430 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1431                           enum nf_nat_manip_type manip,
1432                           const struct nlattr *attr)
1433 {
1434         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1435         int err;
1436
1437         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1438         if (!parse_nat_setup) {
1439 #ifdef CONFIG_MODULES
1440                 rcu_read_unlock();
1441                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1442                 if (request_module("nf-nat") < 0) {
1443                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1444                         rcu_read_lock();
1445                         return -EOPNOTSUPP;
1446                 }
1447                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1448                 rcu_read_lock();
1449                 if (nfnetlink_parse_nat_setup_hook)
1450                         return -EAGAIN;
1451 #endif
1452                 return -EOPNOTSUPP;
1453         }
1454
1455         err = parse_nat_setup(ct, manip, attr);
1456         if (err == -EAGAIN) {
1457 #ifdef CONFIG_MODULES
1458                 rcu_read_unlock();
1459                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1460                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1461                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1462                         rcu_read_lock();
1463                         return -EOPNOTSUPP;
1464                 }
1465                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1466                 rcu_read_lock();
1467 #else
1468                 err = -EOPNOTSUPP;
1469 #endif
1470         }
1471         return err;
1472 }
1473 #endif
1474
1475 static void
1476 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1477                           unsigned long off)
1478 {
1479         unsigned int bit;
1480
1481         /* Ignore these unchangable bits */
1482         on &= ~IPS_UNCHANGEABLE_MASK;
1483         off &= ~IPS_UNCHANGEABLE_MASK;
1484
1485         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1486                 if (on & (1 << bit))
1487                         set_bit(bit, &ct->status);
1488                 else if (off & (1 << bit))
1489                         clear_bit(bit, &ct->status);
1490         }
1491 }
1492
1493 static int
1494 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1495 {
1496         unsigned long d;
1497         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1498         d = ct->status ^ status;
1499
1500         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1501                 /* unchangeable */
1502                 return -EBUSY;
1503
1504         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1505                 /* SEEN_REPLY bit can only be set */
1506                 return -EBUSY;
1507
1508         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1509                 /* ASSURED bit can only be set */
1510                 return -EBUSY;
1511
1512         __ctnetlink_change_status(ct, status, 0);
1513         return 0;
1514 }
1515
1516 static int
1517 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1518 {
1519 #ifdef CONFIG_NF_NAT_NEEDED
1520         int ret;
1521
1522         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1523                 return 0;
1524
1525         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1526                                         cda[CTA_NAT_DST]);
1527         if (ret < 0)
1528                 return ret;
1529
1530         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1531                                          cda[CTA_NAT_SRC]);
1532 #else
1533         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1534                 return 0;
1535         return -EOPNOTSUPP;
1536 #endif
1537 }
1538
1539 static int ctnetlink_change_helper(struct nf_conn *ct,
1540                                    const struct nlattr * const cda[])
1541 {
1542         struct nf_conntrack_helper *helper;
1543         struct nf_conn_help *help = nfct_help(ct);
1544         char *helpname = NULL;
1545         struct nlattr *helpinfo = NULL;
1546         int err;
1547
1548         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1549         if (err < 0)
1550                 return err;
1551
1552         /* don't change helper of sibling connections */
1553         if (ct->master) {
1554                 /* If we try to change the helper to the same thing twice,
1555                  * treat the second attempt as a no-op instead of returning
1556                  * an error.
1557                  */
1558                 err = -EBUSY;
1559                 if (help) {
1560                         rcu_read_lock();
1561                         helper = rcu_dereference(help->helper);
1562                         if (helper && !strcmp(helper->name, helpname))
1563                                 err = 0;
1564                         rcu_read_unlock();
1565                 }
1566
1567                 return err;
1568         }
1569
1570         if (!strcmp(helpname, "")) {
1571                 if (help && help->helper) {
1572                         /* we had a helper before ... */
1573                         nf_ct_remove_expectations(ct);
1574                         RCU_INIT_POINTER(help->helper, NULL);
1575                 }
1576
1577                 return 0;
1578         }
1579
1580         rcu_read_lock();
1581         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1582                                             nf_ct_protonum(ct));
1583         if (helper == NULL) {
1584                 rcu_read_unlock();
1585                 return -EOPNOTSUPP;
1586         }
1587
1588         if (help) {
1589                 if (help->helper == helper) {
1590                         /* update private helper data if allowed. */
1591                         if (helper->from_nlattr)
1592                                 helper->from_nlattr(helpinfo, ct);
1593                         err = 0;
1594                 } else
1595                         err = -EBUSY;
1596         } else {
1597                 /* we cannot set a helper for an existing conntrack */
1598                 err = -EOPNOTSUPP;
1599         }
1600
1601         rcu_read_unlock();
1602         return err;
1603 }
1604
1605 static int ctnetlink_change_timeout(struct nf_conn *ct,
1606                                     const struct nlattr * const cda[])
1607 {
1608         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1609
1610         if (timeout > INT_MAX)
1611                 timeout = INT_MAX;
1612         ct->timeout = nfct_time_stamp + (u32)timeout;
1613
1614         if (test_bit(IPS_DYING_BIT, &ct->status))
1615                 return -ETIME;
1616
1617         return 0;
1618 }
1619
1620 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1621         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1622         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1623         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1624 };
1625
1626 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1627                                       const struct nlattr * const cda[])
1628 {
1629         const struct nlattr *attr = cda[CTA_PROTOINFO];
1630         const struct nf_conntrack_l4proto *l4proto;
1631         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1632         int err = 0;
1633
1634         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1635                                NULL);
1636         if (err < 0)
1637                 return err;
1638
1639         rcu_read_lock();
1640         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1641         if (l4proto->from_nlattr)
1642                 err = l4proto->from_nlattr(tb, ct);
1643         rcu_read_unlock();
1644
1645         return err;
1646 }
1647
1648 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1649         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1650         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1651         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1652 };
1653
1654 static int change_seq_adj(struct nf_ct_seqadj *seq,
1655                           const struct nlattr * const attr)
1656 {
1657         int err;
1658         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1659
1660         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1661         if (err < 0)
1662                 return err;
1663
1664         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1665                 return -EINVAL;
1666
1667         seq->correction_pos =
1668                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1669
1670         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1671                 return -EINVAL;
1672
1673         seq->offset_before =
1674                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1675
1676         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1677                 return -EINVAL;
1678
1679         seq->offset_after =
1680                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1681
1682         return 0;
1683 }
1684
1685 static int
1686 ctnetlink_change_seq_adj(struct nf_conn *ct,
1687                          const struct nlattr * const cda[])
1688 {
1689         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1690         int ret = 0;
1691
1692         if (!seqadj)
1693                 return 0;
1694
1695         spin_lock_bh(&ct->lock);
1696         if (cda[CTA_SEQ_ADJ_ORIG]) {
1697                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1698                                      cda[CTA_SEQ_ADJ_ORIG]);
1699                 if (ret < 0)
1700                         goto err;
1701
1702                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1703         }
1704
1705         if (cda[CTA_SEQ_ADJ_REPLY]) {
1706                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1707                                      cda[CTA_SEQ_ADJ_REPLY]);
1708                 if (ret < 0)
1709                         goto err;
1710
1711                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1712         }
1713
1714         spin_unlock_bh(&ct->lock);
1715         return 0;
1716 err:
1717         spin_unlock_bh(&ct->lock);
1718         return ret;
1719 }
1720
1721 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
1722         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
1723         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
1724         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
1725 };
1726
1727 static int ctnetlink_change_synproxy(struct nf_conn *ct,
1728                                      const struct nlattr * const cda[])
1729 {
1730         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
1731         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
1732         int err;
1733
1734         if (!synproxy)
1735                 return 0;
1736
1737         err = nla_parse_nested(tb, CTA_SYNPROXY_MAX, cda[CTA_SYNPROXY],
1738                                synproxy_policy, NULL);
1739         if (err < 0)
1740                 return err;
1741
1742         if (!tb[CTA_SYNPROXY_ISN] ||
1743             !tb[CTA_SYNPROXY_ITS] ||
1744             !tb[CTA_SYNPROXY_TSOFF])
1745                 return -EINVAL;
1746
1747         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
1748         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
1749         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
1750
1751         return 0;
1752 }
1753
1754 static int
1755 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1756 {
1757 #ifdef CONFIG_NF_CONNTRACK_LABELS
1758         size_t len = nla_len(cda[CTA_LABELS]);
1759         const void *mask = cda[CTA_LABELS_MASK];
1760
1761         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1762                 return -EINVAL;
1763
1764         if (mask) {
1765                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1766                     nla_len(cda[CTA_LABELS_MASK]) != len)
1767                         return -EINVAL;
1768                 mask = nla_data(cda[CTA_LABELS_MASK]);
1769         }
1770
1771         len /= sizeof(u32);
1772
1773         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1774 #else
1775         return -EOPNOTSUPP;
1776 #endif
1777 }
1778
1779 static int
1780 ctnetlink_change_conntrack(struct nf_conn *ct,
1781                            const struct nlattr * const cda[])
1782 {
1783         int err;
1784
1785         /* only allow NAT changes and master assignation for new conntracks */
1786         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1787                 return -EOPNOTSUPP;
1788
1789         if (cda[CTA_HELP]) {
1790                 err = ctnetlink_change_helper(ct, cda);
1791                 if (err < 0)
1792                         return err;
1793         }
1794
1795         if (cda[CTA_TIMEOUT]) {
1796                 err = ctnetlink_change_timeout(ct, cda);
1797                 if (err < 0)
1798                         return err;
1799         }
1800
1801         if (cda[CTA_STATUS]) {
1802                 err = ctnetlink_change_status(ct, cda);
1803                 if (err < 0)
1804                         return err;
1805         }
1806
1807         if (cda[CTA_PROTOINFO]) {
1808                 err = ctnetlink_change_protoinfo(ct, cda);
1809                 if (err < 0)
1810                         return err;
1811         }
1812
1813 #if defined(CONFIG_NF_CONNTRACK_MARK)
1814         if (cda[CTA_MARK])
1815                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1816 #endif
1817
1818         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1819                 err = ctnetlink_change_seq_adj(ct, cda);
1820                 if (err < 0)
1821                         return err;
1822         }
1823
1824         if (cda[CTA_SYNPROXY]) {
1825                 err = ctnetlink_change_synproxy(ct, cda);
1826                 if (err < 0)
1827                         return err;
1828         }
1829
1830         if (cda[CTA_LABELS]) {
1831                 err = ctnetlink_attach_labels(ct, cda);
1832                 if (err < 0)
1833                         return err;
1834         }
1835
1836         return 0;
1837 }
1838
1839 static struct nf_conn *
1840 ctnetlink_create_conntrack(struct net *net,
1841                            const struct nf_conntrack_zone *zone,
1842                            const struct nlattr * const cda[],
1843                            struct nf_conntrack_tuple *otuple,
1844                            struct nf_conntrack_tuple *rtuple,
1845                            u8 u3)
1846 {
1847         struct nf_conn *ct;
1848         int err = -EINVAL;
1849         struct nf_conntrack_helper *helper;
1850         struct nf_conn_tstamp *tstamp;
1851         u64 timeout;
1852
1853         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1854         if (IS_ERR(ct))
1855                 return ERR_PTR(-ENOMEM);
1856
1857         if (!cda[CTA_TIMEOUT])
1858                 goto err1;
1859
1860         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1861         if (timeout > INT_MAX)
1862                 timeout = INT_MAX;
1863         ct->timeout = (u32)timeout + nfct_time_stamp;
1864
1865         rcu_read_lock();
1866         if (cda[CTA_HELP]) {
1867                 char *helpname = NULL;
1868                 struct nlattr *helpinfo = NULL;
1869
1870                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1871                 if (err < 0)
1872                         goto err2;
1873
1874                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1875                                                     nf_ct_protonum(ct));
1876                 if (helper == NULL) {
1877                         rcu_read_unlock();
1878 #ifdef CONFIG_MODULES
1879                         if (request_module("nfct-helper-%s", helpname) < 0) {
1880                                 err = -EOPNOTSUPP;
1881                                 goto err1;
1882                         }
1883
1884                         rcu_read_lock();
1885                         helper = __nf_conntrack_helper_find(helpname,
1886                                                             nf_ct_l3num(ct),
1887                                                             nf_ct_protonum(ct));
1888                         if (helper) {
1889                                 err = -EAGAIN;
1890                                 goto err2;
1891                         }
1892                         rcu_read_unlock();
1893 #endif
1894                         err = -EOPNOTSUPP;
1895                         goto err1;
1896                 } else {
1897                         struct nf_conn_help *help;
1898
1899                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1900                         if (help == NULL) {
1901                                 err = -ENOMEM;
1902                                 goto err2;
1903                         }
1904                         /* set private helper data if allowed. */
1905                         if (helper->from_nlattr)
1906                                 helper->from_nlattr(helpinfo, ct);
1907
1908                         /* not in hash table yet so not strictly necessary */
1909                         RCU_INIT_POINTER(help->helper, helper);
1910                 }
1911         } else {
1912                 /* try an implicit helper assignation */
1913                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1914                 if (err < 0)
1915                         goto err2;
1916         }
1917
1918         err = ctnetlink_setup_nat(ct, cda);
1919         if (err < 0)
1920                 goto err2;
1921
1922         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1923         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1924         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1925         nf_ct_labels_ext_add(ct);
1926         nfct_seqadj_ext_add(ct);
1927         nfct_synproxy_ext_add(ct);
1928
1929         /* we must add conntrack extensions before confirmation. */
1930         ct->status |= IPS_CONFIRMED;
1931
1932         if (cda[CTA_STATUS]) {
1933                 err = ctnetlink_change_status(ct, cda);
1934                 if (err < 0)
1935                         goto err2;
1936         }
1937
1938         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1939                 err = ctnetlink_change_seq_adj(ct, cda);
1940                 if (err < 0)
1941                         goto err2;
1942         }
1943
1944         memset(&ct->proto, 0, sizeof(ct->proto));
1945         if (cda[CTA_PROTOINFO]) {
1946                 err = ctnetlink_change_protoinfo(ct, cda);
1947                 if (err < 0)
1948                         goto err2;
1949         }
1950
1951         if (cda[CTA_SYNPROXY]) {
1952                 err = ctnetlink_change_synproxy(ct, cda);
1953                 if (err < 0)
1954                         goto err2;
1955         }
1956
1957 #if defined(CONFIG_NF_CONNTRACK_MARK)
1958         if (cda[CTA_MARK])
1959                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1960 #endif
1961
1962         /* setup master conntrack: this is a confirmed expectation */
1963         if (cda[CTA_TUPLE_MASTER]) {
1964                 struct nf_conntrack_tuple master;
1965                 struct nf_conntrack_tuple_hash *master_h;
1966                 struct nf_conn *master_ct;
1967
1968                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1969                                             u3, NULL);
1970                 if (err < 0)
1971                         goto err2;
1972
1973                 master_h = nf_conntrack_find_get(net, zone, &master);
1974                 if (master_h == NULL) {
1975                         err = -ENOENT;
1976                         goto err2;
1977                 }
1978                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1979                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1980                 ct->master = master_ct;
1981         }
1982         tstamp = nf_conn_tstamp_find(ct);
1983         if (tstamp)
1984                 tstamp->start = ktime_get_real_ns();
1985
1986         err = nf_conntrack_hash_check_insert(ct);
1987         if (err < 0)
1988                 goto err2;
1989
1990         rcu_read_unlock();
1991
1992         return ct;
1993
1994 err2:
1995         rcu_read_unlock();
1996 err1:
1997         nf_conntrack_free(ct);
1998         return ERR_PTR(err);
1999 }
2000
2001 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2002                                    struct sk_buff *skb,
2003                                    const struct nlmsghdr *nlh,
2004                                    const struct nlattr * const cda[],
2005                                    struct netlink_ext_ack *extack)
2006 {
2007         struct nf_conntrack_tuple otuple, rtuple;
2008         struct nf_conntrack_tuple_hash *h = NULL;
2009         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2010         struct nf_conn *ct;
2011         u_int8_t u3 = nfmsg->nfgen_family;
2012         struct nf_conntrack_zone zone;
2013         int err;
2014
2015         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2016         if (err < 0)
2017                 return err;
2018
2019         if (cda[CTA_TUPLE_ORIG]) {
2020                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2021                                             u3, &zone);
2022                 if (err < 0)
2023                         return err;
2024         }
2025
2026         if (cda[CTA_TUPLE_REPLY]) {
2027                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2028                                             u3, &zone);
2029                 if (err < 0)
2030                         return err;
2031         }
2032
2033         if (cda[CTA_TUPLE_ORIG])
2034                 h = nf_conntrack_find_get(net, &zone, &otuple);
2035         else if (cda[CTA_TUPLE_REPLY])
2036                 h = nf_conntrack_find_get(net, &zone, &rtuple);
2037
2038         if (h == NULL) {
2039                 err = -ENOENT;
2040                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2041                         enum ip_conntrack_events events;
2042
2043                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2044                                 return -EINVAL;
2045                         if (otuple.dst.protonum != rtuple.dst.protonum)
2046                                 return -EINVAL;
2047
2048                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2049                                                         &rtuple, u3);
2050                         if (IS_ERR(ct))
2051                                 return PTR_ERR(ct);
2052
2053                         err = 0;
2054                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2055                                 events = 1 << IPCT_RELATED;
2056                         else
2057                                 events = 1 << IPCT_NEW;
2058
2059                         if (cda[CTA_LABELS] &&
2060                             ctnetlink_attach_labels(ct, cda) == 0)
2061                                 events |= (1 << IPCT_LABEL);
2062
2063                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2064                                                       (1 << IPCT_ASSURED) |
2065                                                       (1 << IPCT_HELPER) |
2066                                                       (1 << IPCT_PROTOINFO) |
2067                                                       (1 << IPCT_SEQADJ) |
2068                                                       (1 << IPCT_MARK) |
2069                                                       (1 << IPCT_SYNPROXY) |
2070                                                       events,
2071                                                       ct, NETLINK_CB(skb).portid,
2072                                                       nlmsg_report(nlh));
2073                         nf_ct_put(ct);
2074                 }
2075
2076                 return err;
2077         }
2078         /* implicit 'else' */
2079
2080         err = -EEXIST;
2081         ct = nf_ct_tuplehash_to_ctrack(h);
2082         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2083                 err = ctnetlink_change_conntrack(ct, cda);
2084                 if (err == 0) {
2085                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2086                                                       (1 << IPCT_ASSURED) |
2087                                                       (1 << IPCT_HELPER) |
2088                                                       (1 << IPCT_LABEL) |
2089                                                       (1 << IPCT_PROTOINFO) |
2090                                                       (1 << IPCT_SEQADJ) |
2091                                                       (1 << IPCT_MARK) |
2092                                                       (1 << IPCT_SYNPROXY),
2093                                                       ct, NETLINK_CB(skb).portid,
2094                                                       nlmsg_report(nlh));
2095                 }
2096         }
2097
2098         nf_ct_put(ct);
2099         return err;
2100 }
2101
2102 static int
2103 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2104                                 __u16 cpu, const struct ip_conntrack_stat *st)
2105 {
2106         struct nlmsghdr *nlh;
2107         struct nfgenmsg *nfmsg;
2108         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2109
2110         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2111                               IPCTNL_MSG_CT_GET_STATS_CPU);
2112         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2113         if (nlh == NULL)
2114                 goto nlmsg_failure;
2115
2116         nfmsg = nlmsg_data(nlh);
2117         nfmsg->nfgen_family = AF_UNSPEC;
2118         nfmsg->version      = NFNETLINK_V0;
2119         nfmsg->res_id       = htons(cpu);
2120
2121         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2122             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2123             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2124             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2125             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2126                                 htonl(st->insert_failed)) ||
2127             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2128             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2129             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2130             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2131                                 htonl(st->search_restart)))
2132                 goto nla_put_failure;
2133
2134         nlmsg_end(skb, nlh);
2135         return skb->len;
2136
2137 nla_put_failure:
2138 nlmsg_failure:
2139         nlmsg_cancel(skb, nlh);
2140         return -1;
2141 }
2142
2143 static int
2144 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2145 {
2146         int cpu;
2147         struct net *net = sock_net(skb->sk);
2148
2149         if (cb->args[0] == nr_cpu_ids)
2150                 return 0;
2151
2152         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2153                 const struct ip_conntrack_stat *st;
2154
2155                 if (!cpu_possible(cpu))
2156                         continue;
2157
2158                 st = per_cpu_ptr(net->ct.stat, cpu);
2159                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2160                                                     NETLINK_CB(cb->skb).portid,
2161                                                     cb->nlh->nlmsg_seq,
2162                                                     cpu, st) < 0)
2163                                 break;
2164         }
2165         cb->args[0] = cpu;
2166
2167         return skb->len;
2168 }
2169
2170 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2171                                  struct sk_buff *skb,
2172                                  const struct nlmsghdr *nlh,
2173                                  const struct nlattr * const cda[],
2174                                  struct netlink_ext_ack *extack)
2175 {
2176         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2177                 struct netlink_dump_control c = {
2178                         .dump = ctnetlink_ct_stat_cpu_dump,
2179                 };
2180                 return netlink_dump_start(ctnl, skb, nlh, &c);
2181         }
2182
2183         return 0;
2184 }
2185
2186 static int
2187 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2188                             struct net *net)
2189 {
2190         struct nlmsghdr *nlh;
2191         struct nfgenmsg *nfmsg;
2192         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2193         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2194
2195         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2196         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2197         if (nlh == NULL)
2198                 goto nlmsg_failure;
2199
2200         nfmsg = nlmsg_data(nlh);
2201         nfmsg->nfgen_family = AF_UNSPEC;
2202         nfmsg->version      = NFNETLINK_V0;
2203         nfmsg->res_id       = 0;
2204
2205         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2206                 goto nla_put_failure;
2207
2208         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2209                 goto nla_put_failure;
2210
2211         nlmsg_end(skb, nlh);
2212         return skb->len;
2213
2214 nla_put_failure:
2215 nlmsg_failure:
2216         nlmsg_cancel(skb, nlh);
2217         return -1;
2218 }
2219
2220 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2221                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2222                              const struct nlattr * const cda[],
2223                              struct netlink_ext_ack *extack)
2224 {
2225         struct sk_buff *skb2;
2226         int err;
2227
2228         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2229         if (skb2 == NULL)
2230                 return -ENOMEM;
2231
2232         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2233                                           nlh->nlmsg_seq,
2234                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2235                                           sock_net(skb->sk));
2236         if (err <= 0)
2237                 goto free;
2238
2239         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2240         if (err < 0)
2241                 goto out;
2242
2243         return 0;
2244
2245 free:
2246         kfree_skb(skb2);
2247 out:
2248         /* this avoids a loop in nfnetlink. */
2249         return err == -EAGAIN ? -ENOBUFS : err;
2250 }
2251
2252 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2253         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2254         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2255         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2256         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2257         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2258         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2259                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2260         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2261         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2262         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2263         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2264         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2265 };
2266
2267 static struct nf_conntrack_expect *
2268 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2269                        struct nf_conntrack_helper *helper,
2270                        struct nf_conntrack_tuple *tuple,
2271                        struct nf_conntrack_tuple *mask);
2272
2273 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2274 static size_t
2275 ctnetlink_glue_build_size(const struct nf_conn *ct)
2276 {
2277         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2278                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2279                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2280                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2281                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2282                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2283                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2284                + nla_total_size(0) /* CTA_PROTOINFO */
2285                + nla_total_size(0) /* CTA_HELP */
2286                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2287                + ctnetlink_secctx_size(ct)
2288 #ifdef CONFIG_NF_NAT_NEEDED
2289                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2290                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2291 #endif
2292 #ifdef CONFIG_NF_CONNTRACK_MARK
2293                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2294 #endif
2295 #ifdef CONFIG_NF_CONNTRACK_ZONES
2296                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2297 #endif
2298                + ctnetlink_proto_size(ct)
2299                ;
2300 }
2301
2302 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2303                                              enum ip_conntrack_info *ctinfo)
2304 {
2305         return nf_ct_get(skb, ctinfo);
2306 }
2307
2308 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2309 {
2310         const struct nf_conntrack_zone *zone;
2311         struct nlattr *nest_parms;
2312
2313         zone = nf_ct_zone(ct);
2314
2315         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2316         if (!nest_parms)
2317                 goto nla_put_failure;
2318         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2319                 goto nla_put_failure;
2320         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2321                                    NF_CT_ZONE_DIR_ORIG) < 0)
2322                 goto nla_put_failure;
2323         nla_nest_end(skb, nest_parms);
2324
2325         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2326         if (!nest_parms)
2327                 goto nla_put_failure;
2328         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2329                 goto nla_put_failure;
2330         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2331                                    NF_CT_ZONE_DIR_REPL) < 0)
2332                 goto nla_put_failure;
2333         nla_nest_end(skb, nest_parms);
2334
2335         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2336                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2337                 goto nla_put_failure;
2338
2339         if (ctnetlink_dump_id(skb, ct) < 0)
2340                 goto nla_put_failure;
2341
2342         if (ctnetlink_dump_status(skb, ct) < 0)
2343                 goto nla_put_failure;
2344
2345         if (ctnetlink_dump_timeout(skb, ct) < 0)
2346                 goto nla_put_failure;
2347
2348         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2349                 goto nla_put_failure;
2350
2351         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2352                 goto nla_put_failure;
2353
2354 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2355         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2356                 goto nla_put_failure;
2357 #endif
2358         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2359                 goto nla_put_failure;
2360
2361         if ((ct->status & IPS_SEQ_ADJUST) &&
2362             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2363                 goto nla_put_failure;
2364
2365         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2366                 goto nla_put_failure;
2367
2368 #ifdef CONFIG_NF_CONNTRACK_MARK
2369         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2370                 goto nla_put_failure;
2371 #endif
2372         if (ctnetlink_dump_labels(skb, ct) < 0)
2373                 goto nla_put_failure;
2374         return 0;
2375
2376 nla_put_failure:
2377         return -ENOSPC;
2378 }
2379
2380 static int
2381 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2382                      enum ip_conntrack_info ctinfo,
2383                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2384 {
2385         struct nlattr *nest_parms;
2386
2387         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2388         if (!nest_parms)
2389                 goto nla_put_failure;
2390
2391         if (__ctnetlink_glue_build(skb, ct) < 0)
2392                 goto nla_put_failure;
2393
2394         nla_nest_end(skb, nest_parms);
2395
2396         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2397                 goto nla_put_failure;
2398
2399         return 0;
2400
2401 nla_put_failure:
2402         return -ENOSPC;
2403 }
2404
2405 static int
2406 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2407 {
2408         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2409         unsigned long d = ct->status ^ status;
2410
2411         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2412                 /* SEEN_REPLY bit can only be set */
2413                 return -EBUSY;
2414
2415         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2416                 /* ASSURED bit can only be set */
2417                 return -EBUSY;
2418
2419         /* This check is less strict than ctnetlink_change_status()
2420          * because callers often flip IPS_EXPECTED bits when sending
2421          * an NFQA_CT attribute to the kernel.  So ignore the
2422          * unchangeable bits but do not error out. Also user programs
2423          * are allowed to clear the bits that they are allowed to change.
2424          */
2425         __ctnetlink_change_status(ct, status, ~status);
2426         return 0;
2427 }
2428
2429 static int
2430 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2431 {
2432         int err;
2433
2434         if (cda[CTA_TIMEOUT]) {
2435                 err = ctnetlink_change_timeout(ct, cda);
2436                 if (err < 0)
2437                         return err;
2438         }
2439         if (cda[CTA_STATUS]) {
2440                 err = ctnetlink_update_status(ct, cda);
2441                 if (err < 0)
2442                         return err;
2443         }
2444         if (cda[CTA_HELP]) {
2445                 err = ctnetlink_change_helper(ct, cda);
2446                 if (err < 0)
2447                         return err;
2448         }
2449         if (cda[CTA_LABELS]) {
2450                 err = ctnetlink_attach_labels(ct, cda);
2451                 if (err < 0)
2452                         return err;
2453         }
2454 #if defined(CONFIG_NF_CONNTRACK_MARK)
2455         if (cda[CTA_MARK]) {
2456                 u32 mask = 0, mark, newmark;
2457                 if (cda[CTA_MARK_MASK])
2458                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2459
2460                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2461                 newmark = (ct->mark & mask) ^ mark;
2462                 if (newmark != ct->mark)
2463                         ct->mark = newmark;
2464         }
2465 #endif
2466         return 0;
2467 }
2468
2469 static int
2470 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2471 {
2472         struct nlattr *cda[CTA_MAX+1];
2473         int ret;
2474
2475         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2476         if (ret < 0)
2477                 return ret;
2478
2479         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2480 }
2481
2482 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2483                                     const struct nf_conn *ct,
2484                                     struct nf_conntrack_tuple *tuple,
2485                                     struct nf_conntrack_tuple *mask)
2486 {
2487         int err;
2488
2489         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2490                                     nf_ct_l3num(ct), NULL);
2491         if (err < 0)
2492                 return err;
2493
2494         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2495                                      nf_ct_l3num(ct), NULL);
2496 }
2497
2498 static int
2499 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2500                              u32 portid, u32 report)
2501 {
2502         struct nlattr *cda[CTA_EXPECT_MAX+1];
2503         struct nf_conntrack_tuple tuple, mask;
2504         struct nf_conntrack_helper *helper = NULL;
2505         struct nf_conntrack_expect *exp;
2506         int err;
2507
2508         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2509                                NULL);
2510         if (err < 0)
2511                 return err;
2512
2513         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2514                                        ct, &tuple, &mask);
2515         if (err < 0)
2516                 return err;
2517
2518         if (cda[CTA_EXPECT_HELP_NAME]) {
2519                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2520
2521                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2522                                                     nf_ct_protonum(ct));
2523                 if (helper == NULL)
2524                         return -EOPNOTSUPP;
2525         }
2526
2527         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2528                                      helper, &tuple, &mask);
2529         if (IS_ERR(exp))
2530                 return PTR_ERR(exp);
2531
2532         err = nf_ct_expect_related_report(exp, portid, report);
2533         nf_ct_expect_put(exp);
2534         return err;
2535 }
2536
2537 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2538                                   enum ip_conntrack_info ctinfo, int diff)
2539 {
2540         if (!(ct->status & IPS_NAT_MASK))
2541                 return;
2542
2543         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2544 }
2545
2546 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2547         .get_ct         = ctnetlink_glue_get_ct,
2548         .build_size     = ctnetlink_glue_build_size,
2549         .build          = ctnetlink_glue_build,
2550         .parse          = ctnetlink_glue_parse,
2551         .attach_expect  = ctnetlink_glue_attach_expect,
2552         .seq_adjust     = ctnetlink_glue_seqadj,
2553 };
2554 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2555
2556 /***********************************************************************
2557  * EXPECT
2558  ***********************************************************************/
2559
2560 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2561                                     const struct nf_conntrack_tuple *tuple,
2562                                     u32 type)
2563 {
2564         struct nlattr *nest_parms;
2565
2566         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2567         if (!nest_parms)
2568                 goto nla_put_failure;
2569         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2570                 goto nla_put_failure;
2571         nla_nest_end(skb, nest_parms);
2572
2573         return 0;
2574
2575 nla_put_failure:
2576         return -1;
2577 }
2578
2579 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2580                                    const struct nf_conntrack_tuple *tuple,
2581                                    const struct nf_conntrack_tuple_mask *mask)
2582 {
2583         const struct nf_conntrack_l3proto *l3proto;
2584         const struct nf_conntrack_l4proto *l4proto;
2585         struct nf_conntrack_tuple m;
2586         struct nlattr *nest_parms;
2587         int ret;
2588
2589         memset(&m, 0xFF, sizeof(m));
2590         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2591         m.src.u.all = mask->src.u.all;
2592         m.dst.protonum = tuple->dst.protonum;
2593
2594         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2595         if (!nest_parms)
2596                 goto nla_put_failure;
2597
2598         rcu_read_lock();
2599         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2600         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2601         if (ret >= 0) {
2602                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2603                                                tuple->dst.protonum);
2604         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2605         }
2606         rcu_read_unlock();
2607
2608         if (unlikely(ret < 0))
2609                 goto nla_put_failure;
2610
2611         nla_nest_end(skb, nest_parms);
2612
2613         return 0;
2614
2615 nla_put_failure:
2616         return -1;
2617 }
2618
2619 static const union nf_inet_addr any_addr;
2620
2621 static int
2622 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2623                           const struct nf_conntrack_expect *exp)
2624 {
2625         struct nf_conn *master = exp->master;
2626         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2627         struct nf_conn_help *help;
2628 #ifdef CONFIG_NF_NAT_NEEDED
2629         struct nlattr *nest_parms;
2630         struct nf_conntrack_tuple nat_tuple = {};
2631 #endif
2632         struct nf_ct_helper_expectfn *expfn;
2633
2634         if (timeout < 0)
2635                 timeout = 0;
2636
2637         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2638                 goto nla_put_failure;
2639         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2640                 goto nla_put_failure;
2641         if (ctnetlink_exp_dump_tuple(skb,
2642                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2643                                  CTA_EXPECT_MASTER) < 0)
2644                 goto nla_put_failure;
2645
2646 #ifdef CONFIG_NF_NAT_NEEDED
2647         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2648             exp->saved_proto.all) {
2649                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2650                 if (!nest_parms)
2651                         goto nla_put_failure;
2652
2653                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2654                         goto nla_put_failure;
2655
2656                 nat_tuple.src.l3num = nf_ct_l3num(master);
2657                 nat_tuple.src.u3 = exp->saved_addr;
2658                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2659                 nat_tuple.src.u = exp->saved_proto;
2660
2661                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2662                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2663                         goto nla_put_failure;
2664                 nla_nest_end(skb, nest_parms);
2665         }
2666 #endif
2667         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2668             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2669             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2670             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2671                 goto nla_put_failure;
2672         help = nfct_help(master);
2673         if (help) {
2674                 struct nf_conntrack_helper *helper;
2675
2676                 helper = rcu_dereference(help->helper);
2677                 if (helper &&
2678                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2679                         goto nla_put_failure;
2680         }
2681         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2682         if (expfn != NULL &&
2683             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2684                 goto nla_put_failure;
2685
2686         return 0;
2687
2688 nla_put_failure:
2689         return -1;
2690 }
2691
2692 static int
2693 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2694                         int event, const struct nf_conntrack_expect *exp)
2695 {
2696         struct nlmsghdr *nlh;
2697         struct nfgenmsg *nfmsg;
2698         unsigned int flags = portid ? NLM_F_MULTI : 0;
2699
2700         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2701         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2702         if (nlh == NULL)
2703                 goto nlmsg_failure;
2704
2705         nfmsg = nlmsg_data(nlh);
2706         nfmsg->nfgen_family = exp->tuple.src.l3num;
2707         nfmsg->version      = NFNETLINK_V0;
2708         nfmsg->res_id       = 0;
2709
2710         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2711                 goto nla_put_failure;
2712
2713         nlmsg_end(skb, nlh);
2714         return skb->len;
2715
2716 nlmsg_failure:
2717 nla_put_failure:
2718         nlmsg_cancel(skb, nlh);
2719         return -1;
2720 }
2721
2722 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2723 static int
2724 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2725 {
2726         struct nf_conntrack_expect *exp = item->exp;
2727         struct net *net = nf_ct_exp_net(exp);
2728         struct nlmsghdr *nlh;
2729         struct nfgenmsg *nfmsg;
2730         struct sk_buff *skb;
2731         unsigned int type, group;
2732         int flags = 0;
2733
2734         if (events & (1 << IPEXP_DESTROY)) {
2735                 type = IPCTNL_MSG_EXP_DELETE;
2736                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2737         } else if (events & (1 << IPEXP_NEW)) {
2738                 type = IPCTNL_MSG_EXP_NEW;
2739                 flags = NLM_F_CREATE|NLM_F_EXCL;
2740                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2741         } else
2742                 return 0;
2743
2744         if (!item->report && !nfnetlink_has_listeners(net, group))
2745                 return 0;
2746
2747         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2748         if (skb == NULL)
2749                 goto errout;
2750
2751         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2752         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2753         if (nlh == NULL)
2754                 goto nlmsg_failure;
2755
2756         nfmsg = nlmsg_data(nlh);
2757         nfmsg->nfgen_family = exp->tuple.src.l3num;
2758         nfmsg->version      = NFNETLINK_V0;
2759         nfmsg->res_id       = 0;
2760
2761         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2762                 goto nla_put_failure;
2763
2764         nlmsg_end(skb, nlh);
2765         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2766         return 0;
2767
2768 nla_put_failure:
2769         nlmsg_cancel(skb, nlh);
2770 nlmsg_failure:
2771         kfree_skb(skb);
2772 errout:
2773         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2774         return 0;
2775 }
2776 #endif
2777 static int ctnetlink_exp_done(struct netlink_callback *cb)
2778 {
2779         if (cb->args[1])
2780                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2781         return 0;
2782 }
2783
2784 static int
2785 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2786 {
2787         struct net *net = sock_net(skb->sk);
2788         struct nf_conntrack_expect *exp, *last;
2789         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2790         u_int8_t l3proto = nfmsg->nfgen_family;
2791
2792         rcu_read_lock();
2793         last = (struct nf_conntrack_expect *)cb->args[1];
2794         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2795 restart:
2796                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2797                                          hnode) {
2798                         if (l3proto && exp->tuple.src.l3num != l3proto)
2799                                 continue;
2800
2801                         if (!net_eq(nf_ct_net(exp->master), net))
2802                                 continue;
2803
2804                         if (cb->args[1]) {
2805                                 if (exp != last)
2806                                         continue;
2807                                 cb->args[1] = 0;
2808                         }
2809                         if (ctnetlink_exp_fill_info(skb,
2810                                                     NETLINK_CB(cb->skb).portid,
2811                                                     cb->nlh->nlmsg_seq,
2812                                                     IPCTNL_MSG_EXP_NEW,
2813                                                     exp) < 0) {
2814                                 if (!refcount_inc_not_zero(&exp->use))
2815                                         continue;
2816                                 cb->args[1] = (unsigned long)exp;
2817                                 goto out;
2818                         }
2819                 }
2820                 if (cb->args[1]) {
2821                         cb->args[1] = 0;
2822                         goto restart;
2823                 }
2824         }
2825 out:
2826         rcu_read_unlock();
2827         if (last)
2828                 nf_ct_expect_put(last);
2829
2830         return skb->len;
2831 }
2832
2833 static int
2834 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2835 {
2836         struct nf_conntrack_expect *exp, *last;
2837         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2838         struct nf_conn *ct = cb->data;
2839         struct nf_conn_help *help = nfct_help(ct);
2840         u_int8_t l3proto = nfmsg->nfgen_family;
2841
2842         if (cb->args[0])
2843                 return 0;
2844
2845         rcu_read_lock();
2846         last = (struct nf_conntrack_expect *)cb->args[1];
2847 restart:
2848         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2849                 if (l3proto && exp->tuple.src.l3num != l3proto)
2850                         continue;
2851                 if (cb->args[1]) {
2852                         if (exp != last)
2853                                 continue;
2854                         cb->args[1] = 0;
2855                 }
2856                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2857                                             cb->nlh->nlmsg_seq,
2858                                             IPCTNL_MSG_EXP_NEW,
2859                                             exp) < 0) {
2860                         if (!refcount_inc_not_zero(&exp->use))
2861                                 continue;
2862                         cb->args[1] = (unsigned long)exp;
2863                         goto out;
2864                 }
2865         }
2866         if (cb->args[1]) {
2867                 cb->args[1] = 0;
2868                 goto restart;
2869         }
2870         cb->args[0] = 1;
2871 out:
2872         rcu_read_unlock();
2873         if (last)
2874                 nf_ct_expect_put(last);
2875
2876         return skb->len;
2877 }
2878
2879 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2880                                  struct sk_buff *skb,
2881                                  const struct nlmsghdr *nlh,
2882                                  const struct nlattr * const cda[],
2883                                  struct netlink_ext_ack *extack)
2884 {
2885         int err;
2886         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2887         u_int8_t u3 = nfmsg->nfgen_family;
2888         struct nf_conntrack_tuple tuple;
2889         struct nf_conntrack_tuple_hash *h;
2890         struct nf_conn *ct;
2891         struct nf_conntrack_zone zone;
2892         struct netlink_dump_control c = {
2893                 .dump = ctnetlink_exp_ct_dump_table,
2894                 .done = ctnetlink_exp_done,
2895         };
2896
2897         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2898                                     u3, NULL);
2899         if (err < 0)
2900                 return err;
2901
2902         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2903         if (err < 0)
2904                 return err;
2905
2906         h = nf_conntrack_find_get(net, &zone, &tuple);
2907         if (!h)
2908                 return -ENOENT;
2909
2910         ct = nf_ct_tuplehash_to_ctrack(h);
2911         /* No expectation linked to this connection tracking. */
2912         if (!nfct_help(ct)) {
2913                 nf_ct_put(ct);
2914                 return 0;
2915         }
2916
2917         c.data = ct;
2918
2919         err = netlink_dump_start(ctnl, skb, nlh, &c);
2920         nf_ct_put(ct);
2921
2922         return err;
2923 }
2924
2925 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
2926                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2927                                 const struct nlattr * const cda[],
2928                                 struct netlink_ext_ack *extack)
2929 {
2930         struct nf_conntrack_tuple tuple;
2931         struct nf_conntrack_expect *exp;
2932         struct sk_buff *skb2;
2933         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2934         u_int8_t u3 = nfmsg->nfgen_family;
2935         struct nf_conntrack_zone zone;
2936         int err;
2937
2938         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2939                 if (cda[CTA_EXPECT_MASTER])
2940                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
2941                                                      extack);
2942                 else {
2943                         struct netlink_dump_control c = {
2944                                 .dump = ctnetlink_exp_dump_table,
2945                                 .done = ctnetlink_exp_done,
2946                         };
2947                         return netlink_dump_start(ctnl, skb, nlh, &c);
2948                 }
2949         }
2950
2951         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2952         if (err < 0)
2953                 return err;
2954
2955         if (cda[CTA_EXPECT_TUPLE])
2956                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2957                                             u3, NULL);
2958         else if (cda[CTA_EXPECT_MASTER])
2959                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2960                                             u3, NULL);
2961         else
2962                 return -EINVAL;
2963
2964         if (err < 0)
2965                 return err;
2966
2967         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2968         if (!exp)
2969                 return -ENOENT;
2970
2971         if (cda[CTA_EXPECT_ID]) {
2972                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2973                 if (ntohl(id) != (u32)(unsigned long)exp) {
2974                         nf_ct_expect_put(exp);
2975                         return -ENOENT;
2976                 }
2977         }
2978
2979         err = -ENOMEM;
2980         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2981         if (skb2 == NULL) {
2982                 nf_ct_expect_put(exp);
2983                 goto out;
2984         }
2985
2986         rcu_read_lock();
2987         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2988                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2989         rcu_read_unlock();
2990         nf_ct_expect_put(exp);
2991         if (err <= 0)
2992                 goto free;
2993
2994         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2995         if (err < 0)
2996                 goto out;
2997
2998         return 0;
2999
3000 free:
3001         kfree_skb(skb2);
3002 out:
3003         /* this avoids a loop in nfnetlink. */
3004         return err == -EAGAIN ? -ENOBUFS : err;
3005 }
3006
3007 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3008 {
3009         const struct nf_conn_help *m_help;
3010         const char *name = data;
3011
3012         m_help = nfct_help(exp->master);
3013
3014         return strcmp(m_help->helper->name, name) == 0;
3015 }
3016
3017 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3018 {
3019         return true;
3020 }
3021
3022 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3023                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3024                                 const struct nlattr * const cda[],
3025                                 struct netlink_ext_ack *extack)
3026 {
3027         struct nf_conntrack_expect *exp;
3028         struct nf_conntrack_tuple tuple;
3029         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3030         u_int8_t u3 = nfmsg->nfgen_family;
3031         struct nf_conntrack_zone zone;
3032         int err;
3033
3034         if (cda[CTA_EXPECT_TUPLE]) {
3035                 /* delete a single expect by tuple */
3036                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3037                 if (err < 0)
3038                         return err;
3039
3040                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3041                                             u3, NULL);
3042                 if (err < 0)
3043                         return err;
3044
3045                 /* bump usage count to 2 */
3046                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
3047                 if (!exp)
3048                         return -ENOENT;
3049
3050                 if (cda[CTA_EXPECT_ID]) {
3051                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3052                         if (ntohl(id) != (u32)(unsigned long)exp) {
3053                                 nf_ct_expect_put(exp);
3054                                 return -ENOENT;
3055                         }
3056                 }
3057
3058                 /* after list removal, usage count == 1 */
3059                 spin_lock_bh(&nf_conntrack_expect_lock);
3060                 if (del_timer(&exp->timeout)) {
3061                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3062                                                    nlmsg_report(nlh));
3063                         nf_ct_expect_put(exp);
3064                 }
3065                 spin_unlock_bh(&nf_conntrack_expect_lock);
3066                 /* have to put what we 'get' above.
3067                  * after this line usage count == 0 */
3068                 nf_ct_expect_put(exp);
3069         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3070                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3071
3072                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
3073                                          NETLINK_CB(skb).portid,
3074                                          nlmsg_report(nlh));
3075         } else {
3076                 /* This basically means we have to flush everything*/
3077                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3078                                          NETLINK_CB(skb).portid,
3079                                          nlmsg_report(nlh));
3080         }
3081
3082         return 0;
3083 }
3084 static int
3085 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3086                         const struct nlattr * const cda[])
3087 {
3088         if (cda[CTA_EXPECT_TIMEOUT]) {
3089                 if (!del_timer(&x->timeout))
3090                         return -ETIME;
3091
3092                 x->timeout.expires = jiffies +
3093                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3094                 add_timer(&x->timeout);
3095         }
3096         return 0;
3097 }
3098
3099 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3100         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3101         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3102 };
3103
3104 static int
3105 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3106                            struct nf_conntrack_expect *exp,
3107                            u_int8_t u3)
3108 {
3109 #ifdef CONFIG_NF_NAT_NEEDED
3110         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3111         struct nf_conntrack_tuple nat_tuple = {};
3112         int err;
3113
3114         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3115                                exp_nat_nla_policy, NULL);
3116         if (err < 0)
3117                 return err;
3118
3119         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3120                 return -EINVAL;
3121
3122         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3123                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3124                                     u3, NULL);
3125         if (err < 0)
3126                 return err;
3127
3128         exp->saved_addr = nat_tuple.src.u3;
3129         exp->saved_proto = nat_tuple.src.u;
3130         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3131
3132         return 0;
3133 #else
3134         return -EOPNOTSUPP;
3135 #endif
3136 }
3137
3138 static struct nf_conntrack_expect *
3139 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3140                        struct nf_conntrack_helper *helper,
3141                        struct nf_conntrack_tuple *tuple,
3142                        struct nf_conntrack_tuple *mask)
3143 {
3144         u_int32_t class = 0;
3145         struct nf_conntrack_expect *exp;
3146         struct nf_conn_help *help;
3147         int err;
3148
3149         help = nfct_help(ct);
3150         if (!help)
3151                 return ERR_PTR(-EOPNOTSUPP);
3152
3153         if (cda[CTA_EXPECT_CLASS] && helper) {
3154                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3155                 if (class > helper->expect_class_max)
3156                         return ERR_PTR(-EINVAL);
3157         }
3158         exp = nf_ct_expect_alloc(ct);
3159         if (!exp)
3160                 return ERR_PTR(-ENOMEM);
3161
3162         if (cda[CTA_EXPECT_FLAGS]) {
3163                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3164                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3165         } else {
3166                 exp->flags = 0;
3167         }
3168         if (cda[CTA_EXPECT_FN]) {
3169                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3170                 struct nf_ct_helper_expectfn *expfn;
3171
3172                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3173                 if (expfn == NULL) {
3174                         err = -EINVAL;
3175                         goto err_out;
3176                 }
3177                 exp->expectfn = expfn->expectfn;
3178         } else
3179                 exp->expectfn = NULL;
3180
3181         exp->class = class;
3182         exp->master = ct;
3183         exp->helper = helper;
3184         exp->tuple = *tuple;
3185         exp->mask.src.u3 = mask->src.u3;
3186         exp->mask.src.u.all = mask->src.u.all;
3187
3188         if (cda[CTA_EXPECT_NAT]) {
3189                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3190                                                  exp, nf_ct_l3num(ct));
3191                 if (err < 0)
3192                         goto err_out;
3193         }
3194         return exp;
3195 err_out:
3196         nf_ct_expect_put(exp);
3197         return ERR_PTR(err);
3198 }
3199
3200 static int
3201 ctnetlink_create_expect(struct net *net,
3202                         const struct nf_conntrack_zone *zone,
3203                         const struct nlattr * const cda[],
3204                         u_int8_t u3, u32 portid, int report)
3205 {
3206         struct nf_conntrack_tuple tuple, mask, master_tuple;
3207         struct nf_conntrack_tuple_hash *h = NULL;
3208         struct nf_conntrack_helper *helper = NULL;
3209         struct nf_conntrack_expect *exp;
3210         struct nf_conn *ct;
3211         int err;
3212
3213         /* caller guarantees that those three CTA_EXPECT_* exist */
3214         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3215                                     u3, NULL);
3216         if (err < 0)
3217                 return err;
3218         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3219                                     u3, NULL);
3220         if (err < 0)
3221                 return err;
3222         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3223                                     u3, NULL);
3224         if (err < 0)
3225                 return err;
3226
3227         /* Look for master conntrack of this expectation */
3228         h = nf_conntrack_find_get(net, zone, &master_tuple);
3229         if (!h)
3230                 return -ENOENT;
3231         ct = nf_ct_tuplehash_to_ctrack(h);
3232
3233         rcu_read_lock();
3234         if (cda[CTA_EXPECT_HELP_NAME]) {
3235                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3236
3237                 helper = __nf_conntrack_helper_find(helpname, u3,
3238                                                     nf_ct_protonum(ct));
3239                 if (helper == NULL) {
3240                         rcu_read_unlock();
3241 #ifdef CONFIG_MODULES
3242                         if (request_module("nfct-helper-%s", helpname) < 0) {
3243                                 err = -EOPNOTSUPP;
3244                                 goto err_ct;
3245                         }
3246                         rcu_read_lock();
3247                         helper = __nf_conntrack_helper_find(helpname, u3,
3248                                                             nf_ct_protonum(ct));
3249                         if (helper) {
3250                                 err = -EAGAIN;
3251                                 goto err_rcu;
3252                         }
3253                         rcu_read_unlock();
3254 #endif
3255                         err = -EOPNOTSUPP;
3256                         goto err_ct;
3257                 }
3258         }
3259
3260         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3261         if (IS_ERR(exp)) {
3262                 err = PTR_ERR(exp);
3263                 goto err_rcu;
3264         }
3265
3266         err = nf_ct_expect_related_report(exp, portid, report);
3267         nf_ct_expect_put(exp);
3268 err_rcu:
3269         rcu_read_unlock();
3270 err_ct:
3271         nf_ct_put(ct);
3272         return err;
3273 }
3274
3275 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3276                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3277                                 const struct nlattr * const cda[],
3278                                 struct netlink_ext_ack *extack)
3279 {
3280         struct nf_conntrack_tuple tuple;
3281         struct nf_conntrack_expect *exp;
3282         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3283         u_int8_t u3 = nfmsg->nfgen_family;
3284         struct nf_conntrack_zone zone;
3285         int err;
3286
3287         if (!cda[CTA_EXPECT_TUPLE]
3288             || !cda[CTA_EXPECT_MASK]
3289             || !cda[CTA_EXPECT_MASTER])
3290                 return -EINVAL;
3291
3292         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3293         if (err < 0)
3294                 return err;
3295
3296         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3297                                     u3, NULL);
3298         if (err < 0)
3299                 return err;
3300
3301         spin_lock_bh(&nf_conntrack_expect_lock);
3302         exp = __nf_ct_expect_find(net, &zone, &tuple);
3303         if (!exp) {
3304                 spin_unlock_bh(&nf_conntrack_expect_lock);
3305                 err = -ENOENT;
3306                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3307                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3308                                                       NETLINK_CB(skb).portid,
3309                                                       nlmsg_report(nlh));
3310                 }
3311                 return err;
3312         }
3313
3314         err = -EEXIST;
3315         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3316                 err = ctnetlink_change_expect(exp, cda);
3317         spin_unlock_bh(&nf_conntrack_expect_lock);
3318
3319         return err;
3320 }
3321
3322 static int
3323 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3324                              const struct ip_conntrack_stat *st)
3325 {
3326         struct nlmsghdr *nlh;
3327         struct nfgenmsg *nfmsg;
3328         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3329
3330         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3331                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3332         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3333         if (nlh == NULL)
3334                 goto nlmsg_failure;
3335
3336         nfmsg = nlmsg_data(nlh);
3337         nfmsg->nfgen_family = AF_UNSPEC;
3338         nfmsg->version      = NFNETLINK_V0;
3339         nfmsg->res_id       = htons(cpu);
3340
3341         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3342             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3343             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3344                 goto nla_put_failure;
3345
3346         nlmsg_end(skb, nlh);
3347         return skb->len;
3348
3349 nla_put_failure:
3350 nlmsg_failure:
3351         nlmsg_cancel(skb, nlh);
3352         return -1;
3353 }
3354
3355 static int
3356 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3357 {
3358         int cpu;
3359         struct net *net = sock_net(skb->sk);
3360
3361         if (cb->args[0] == nr_cpu_ids)
3362                 return 0;
3363
3364         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3365                 const struct ip_conntrack_stat *st;
3366
3367                 if (!cpu_possible(cpu))
3368                         continue;
3369
3370                 st = per_cpu_ptr(net->ct.stat, cpu);
3371                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3372                                                  cb->nlh->nlmsg_seq,
3373                                                  cpu, st) < 0)
3374                         break;
3375         }
3376         cb->args[0] = cpu;
3377
3378         return skb->len;
3379 }
3380
3381 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3382                                   struct sk_buff *skb,
3383                                   const struct nlmsghdr *nlh,
3384                                   const struct nlattr * const cda[],
3385                                   struct netlink_ext_ack *extack)
3386 {
3387         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3388                 struct netlink_dump_control c = {
3389                         .dump = ctnetlink_exp_stat_cpu_dump,
3390                 };
3391                 return netlink_dump_start(ctnl, skb, nlh, &c);
3392         }
3393
3394         return 0;
3395 }
3396
3397 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3398 static struct nf_ct_event_notifier ctnl_notifier = {
3399         .fcn = ctnetlink_conntrack_event,
3400 };
3401
3402 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3403         .fcn = ctnetlink_expect_event,
3404 };
3405 #endif
3406
3407 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3408         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3409                                             .attr_count = CTA_MAX,
3410                                             .policy = ct_nla_policy },
3411         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3412                                             .attr_count = CTA_MAX,
3413                                             .policy = ct_nla_policy },
3414         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3415                                             .attr_count = CTA_MAX,
3416                                             .policy = ct_nla_policy },
3417         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3418                                             .attr_count = CTA_MAX,
3419                                             .policy = ct_nla_policy },
3420         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3421         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3422         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3423         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3424 };
3425
3426 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3427         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3428                                             .attr_count = CTA_EXPECT_MAX,
3429                                             .policy = exp_nla_policy },
3430         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3431                                             .attr_count = CTA_EXPECT_MAX,
3432                                             .policy = exp_nla_policy },
3433         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3434                                             .attr_count = CTA_EXPECT_MAX,
3435                                             .policy = exp_nla_policy },
3436         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3437 };
3438
3439 static const struct nfnetlink_subsystem ctnl_subsys = {
3440         .name                           = "conntrack",
3441         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3442         .cb_count                       = IPCTNL_MSG_MAX,
3443         .cb                             = ctnl_cb,
3444 };
3445
3446 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3447         .name                           = "conntrack_expect",
3448         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3449         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3450         .cb                             = ctnl_exp_cb,
3451 };
3452
3453 MODULE_ALIAS("ip_conntrack_netlink");
3454 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3455 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3456
3457 static int __net_init ctnetlink_net_init(struct net *net)
3458 {
3459 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3460         int ret;
3461
3462         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3463         if (ret < 0) {
3464                 pr_err("ctnetlink_init: cannot register notifier.\n");
3465                 goto err_out;
3466         }
3467
3468         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3469         if (ret < 0) {
3470                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3471                 goto err_unreg_notifier;
3472         }
3473 #endif
3474         return 0;
3475
3476 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3477 err_unreg_notifier:
3478         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3479 err_out:
3480         return ret;
3481 #endif
3482 }
3483
3484 static void ctnetlink_net_exit(struct net *net)
3485 {
3486 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3487         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3488         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3489 #endif
3490 }
3491
3492 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3493 {
3494         struct net *net;
3495
3496         list_for_each_entry(net, net_exit_list, exit_list)
3497                 ctnetlink_net_exit(net);
3498 }
3499
3500 static struct pernet_operations ctnetlink_net_ops = {
3501         .init           = ctnetlink_net_init,
3502         .exit_batch     = ctnetlink_net_exit_batch,
3503 };
3504
3505 static int __init ctnetlink_init(void)
3506 {
3507         int ret;
3508
3509         ret = nfnetlink_subsys_register(&ctnl_subsys);
3510         if (ret < 0) {
3511                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3512                 goto err_out;
3513         }
3514
3515         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3516         if (ret < 0) {
3517                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3518                 goto err_unreg_subsys;
3519         }
3520
3521         ret = register_pernet_subsys(&ctnetlink_net_ops);
3522         if (ret < 0) {
3523                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3524                 goto err_unreg_exp_subsys;
3525         }
3526 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3527         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3528         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3529 #endif
3530         return 0;
3531
3532 err_unreg_exp_subsys:
3533         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3534 err_unreg_subsys:
3535         nfnetlink_subsys_unregister(&ctnl_subsys);
3536 err_out:
3537         return ret;
3538 }
3539
3540 static void __exit ctnetlink_exit(void)
3541 {
3542         unregister_pernet_subsys(&ctnetlink_net_ops);
3543         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3544         nfnetlink_subsys_unregister(&ctnl_subsys);
3545 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3546         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3547 #endif
3548         synchronize_rcu();
3549 }
3550
3551 module_init(ctnetlink_init);
3552 module_exit(ctnetlink_exit);