net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32 #include <linux/siphash.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netlink.h>
  36 #include <net/sock.h>
  37 #include <net/netfilter/nf_conntrack.h>
  38 #include <net/netfilter/nf_conntrack_core.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_seqadj.h>
  42 #include <net/netfilter/nf_conntrack_l3proto.h>
  43 #include <net/netfilter/nf_conntrack_l4proto.h>
  44 #include <net/netfilter/nf_conntrack_tuple.h>
  45 #include <net/netfilter/nf_conntrack_acct.h>
  46 #include <net/netfilter/nf_conntrack_zones.h>
  47 #include <net/netfilter/nf_conntrack_timestamp.h>
  48 #include <net/netfilter/nf_conntrack_labels.h>
  49 #include <net/netfilter/nf_conntrack_seqadj.h>
  50 #include <net/netfilter/nf_conntrack_synproxy.h>
  51 #ifdef CONFIG_NF_NAT_NEEDED
  52 #include <net/netfilter/nf_nat_core.h>
  53 #include <net/netfilter/nf_nat_l4proto.h>
  54 #include <net/netfilter/nf_nat_helper.h>
  55 #endif
  56
  57 #include <linux/netfilter/nfnetlink.h>
  58 #include <linux/netfilter/nfnetlink_conntrack.h>
  59
  60 MODULE_LICENSE("GPL");
  61
  62 static char __initdata version[] = "0.93";
  63
  64 static inline int
  65 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  66                             const struct nf_conntrack_tuple *tuple,
  67                             struct nf_conntrack_l4proto *l4proto)
  68 {
  69         int ret = 0;
  70         struct nlattr *nest_parms;
  71
  72         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  73         if (!nest_parms)
  74                 goto nla_put_failure;
  75         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  76                 goto nla_put_failure;
  77
  78         if (likely(l4proto->tuple_to_nlattr))
  79                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  80
  81         nla_nest_end(skb, nest_parms);
  82
  83         return ret;
  84
  85 nla_put_failure:
  86         return -1;
  87 }
  88
  89 static inline int
  90 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  91                          const struct nf_conntrack_tuple *tuple,
  92                          struct nf_conntrack_l3proto *l3proto)
  93 {
  94         int ret = 0;
  95         struct nlattr *nest_parms;
  96
  97         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  98         if (!nest_parms)
  99                 goto nla_put_failure;
 100
 101         if (likely(l3proto->tuple_to_nlattr))
 102                 ret = l3proto->tuple_to_nlattr(skb, tuple);
 103
 104         nla_nest_end(skb, nest_parms);
 105
 106         return ret;
 107
 108 nla_put_failure:
 109         return -1;
 110 }
 111
 112 static int
 113 ctnetlink_dump_tuples(struct sk_buff *skb,
 114                       const struct nf_conntrack_tuple *tuple)
 115 {
 116         int ret;
 117         struct nf_conntrack_l3proto *l3proto;
 118         struct nf_conntrack_l4proto *l4proto;
 119
 120         rcu_read_lock();
 121         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 122         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 123
 124         if (ret >= 0) {
 125                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 126                                                tuple->dst.protonum);
 127                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 128         }
 129         rcu_read_unlock();
 130         return ret;
 131 }
 132
 133 static inline int
 134 ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 135                        const struct nf_conntrack_zone *zone, int dir)
 136 {
 137         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 138                 return 0;
 139         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 140                 goto nla_put_failure;
 141         return 0;
 142
 143 nla_put_failure:
 144         return -1;
 145 }
 146
 147 static inline int
 148 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 149 {
 150         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 151                 goto nla_put_failure;
 152         return 0;
 153
 154 nla_put_failure:
 155         return -1;
 156 }
 157
 158 static inline int
 159 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 160 {
 161         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 162
 163         if (timeout < 0)
 164                 timeout = 0;
 165
 166         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 167                 goto nla_put_failure;
 168         return 0;
 169
 170 nla_put_failure:
 171         return -1;
 172 }
 173
 174 static inline int
 175 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 176 {
 177         struct nf_conntrack_l4proto *l4proto;
 178         struct nlattr *nest_proto;
 179         int ret;
 180
 181         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 182         if (!l4proto->to_nlattr)
 183                 return 0;
 184
 185         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 186         if (!nest_proto)
 187                 goto nla_put_failure;
 188
 189         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 190
 191         nla_nest_end(skb, nest_proto);
 192
 193         return ret;
 194
 195 nla_put_failure:
 196         return -1;
 197 }
 198
 199 static inline int
 200 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 201 {
 202         struct nlattr *nest_helper;
 203         const struct nf_conn_help *help = nfct_help(ct);
 204         struct nf_conntrack_helper *helper;
 205
 206         if (!help)
 207                 return 0;
 208
 209         helper = rcu_dereference(help->helper);
 210         if (!helper)
 211                 goto out;
 212
 213         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 214         if (!nest_helper)
 215                 goto nla_put_failure;
 216         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 217                 goto nla_put_failure;
 218
 219         if (helper->to_nlattr)
 220                 helper->to_nlattr(skb, ct);
 221
 222         nla_nest_end(skb, nest_helper);
 223 out:
 224         return 0;
 225
 226 nla_put_failure:
 227         return -1;
 228 }
 229
 230 static int
 231 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 232               enum ip_conntrack_dir dir, int type)
 233 {
 234         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 235         struct nf_conn_counter *counter = acct->counter;
 236         struct nlattr *nest_count;
 237         u64 pkts, bytes;
 238
 239         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 240                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 241                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 242         } else {
 243                 pkts = atomic64_read(&counter[dir].packets);
 244                 bytes = atomic64_read(&counter[dir].bytes);
 245         }
 246
 247         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 248         if (!nest_count)
 249                 goto nla_put_failure;
 250
 251         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 252             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 253                 goto nla_put_failure;
 254
 255         nla_nest_end(skb, nest_count);
 256
 257         return 0;
 258
 259 nla_put_failure:
 260         return -1;
 261 }
 262
 263 static int
 264 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 265 {
 266         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 267
 268         if (!acct)
 269                 return 0;
 270
 271         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 272                 return -1;
 273         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 274                 return -1;
 275
 276         return 0;
 277 }
 278
 279 static int
 280 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 281 {
 282         struct nlattr *nest_count;
 283         const struct nf_conn_tstamp *tstamp;
 284
 285         tstamp = nf_conn_tstamp_find(ct);
 286         if (!tstamp)
 287                 return 0;
 288
 289         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 290         if (!nest_count)
 291                 goto nla_put_failure;
 292
 293         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 294             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 295                                                cpu_to_be64(tstamp->stop))))
 296                 goto nla_put_failure;
 297         nla_nest_end(skb, nest_count);
 298
 299         return 0;
 300
 301 nla_put_failure:
 302         return -1;
 303 }
 304
 305 #ifdef CONFIG_NF_CONNTRACK_MARK
 306 static inline int
 307 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 308 {
 309         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 310                 goto nla_put_failure;
 311         return 0;
 312
 313 nla_put_failure:
 314         return -1;
 315 }
 316 #else
 317 #define ctnetlink_dump_mark(a, b) (0)
 318 #endif
 319
 320 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 321 static inline int
 322 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 323 {
 324         struct nlattr *nest_secctx;
 325         int len, ret;
 326         char *secctx;
 327
 328         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 329         if (ret)
 330                 return 0;
 331
 332         ret = -1;
 333         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 334         if (!nest_secctx)
 335                 goto nla_put_failure;
 336
 337         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 338                 goto nla_put_failure;
 339         nla_nest_end(skb, nest_secctx);
 340
 341         ret = 0;
 342 nla_put_failure:
 343         security_release_secctx(secctx, len);
 344         return ret;
 345 }
 346 #else
 347 #define ctnetlink_dump_secctx(a, b) (0)
 348 #endif
 349
 350 #ifdef CONFIG_NF_CONNTRACK_LABELS
 351 static int ctnetlink_label_size(const struct nf_conn *ct)
 352 {
 353         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 354
 355         if (!labels)
 356                 return 0;
 357         return nla_total_size(labels->words * sizeof(long));
 358 }
 359
 360 static int
 361 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 362 {
 363         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 364         unsigned int len, i;
 365
 366         if (!labels)
 367                 return 0;
 368
 369         len = labels->words * sizeof(long);
 370         i = 0;
 371         do {
 372                 if (labels->bits[i] != 0)
 373                         return nla_put(skb, CTA_LABELS, len, labels->bits);
 374                 i++;
 375         } while (i < labels->words);
 376
 377         return 0;
 378 }
 379 #else
 380 #define ctnetlink_dump_labels(a, b) (0)
 381 #define ctnetlink_label_size(a) (0)
 382 #endif
 383
 384 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 385
 386 static inline int
 387 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 388 {
 389         struct nlattr *nest_parms;
 390
 391         if (!(ct->status & IPS_EXPECTED))
 392                 return 0;
 393
 394         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 395         if (!nest_parms)
 396                 goto nla_put_failure;
 397         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 398                 goto nla_put_failure;
 399         nla_nest_end(skb, nest_parms);
 400
 401         return 0;
 402
 403 nla_put_failure:
 404         return -1;
 405 }
 406
 407 static int
 408 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 409 {
 410         struct nlattr *nest_parms;
 411
 412         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 413         if (!nest_parms)
 414                 goto nla_put_failure;
 415
 416         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 417                          htonl(seq->correction_pos)) ||
 418             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 419                          htonl(seq->offset_before)) ||
 420             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 421                          htonl(seq->offset_after)))
 422                 goto nla_put_failure;
 423
 424         nla_nest_end(skb, nest_parms);
 425
 426         return 0;
 427
 428 nla_put_failure:
 429         return -1;
 430 }
 431
 432 static inline int
 433 ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 434 {
 435         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 436         struct nf_ct_seqadj *seq;
 437
 438         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 439                 return 0;
 440
 441         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 442         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 443                 return -1;
 444
 445         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 446         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 447                 return -1;
 448
 449         return 0;
 450 }
 451
 452 static inline int
 453 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 454 {
 455         __be32 id = (__force __be32)nf_ct_get_id(ct);
 456
 457         if (nla_put_be32(skb, CTA_ID, id))
 458                 goto nla_put_failure;
 459         return 0;
 460
 461 nla_put_failure:
 462         return -1;
 463 }
 464
 465 static inline int
 466 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 467 {
 468         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 469                 goto nla_put_failure;
 470         return 0;
 471
 472 nla_put_failure:
 473         return -1;
 474 }
 475
 476 static int
 477 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 478                     struct nf_conn *ct)
 479 {
 480         const struct nf_conntrack_zone *zone;
 481         struct nlmsghdr *nlh;
 482         struct nfgenmsg *nfmsg;
 483         struct nlattr *nest_parms;
 484         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 485
 486         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 487         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 488         if (nlh == NULL)
 489                 goto nlmsg_failure;
 490
 491         nfmsg = nlmsg_data(nlh);
 492         nfmsg->nfgen_family = nf_ct_l3num(ct);
 493         nfmsg->version      = NFNETLINK_V0;
 494         nfmsg->res_id       = 0;
 495
 496         zone = nf_ct_zone(ct);
 497
 498         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 499         if (!nest_parms)
 500                 goto nla_put_failure;
 501         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 502                 goto nla_put_failure;
 503         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 504                                    NF_CT_ZONE_DIR_ORIG) < 0)
 505                 goto nla_put_failure;
 506         nla_nest_end(skb, nest_parms);
 507
 508         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 509         if (!nest_parms)
 510                 goto nla_put_failure;
 511         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 512                 goto nla_put_failure;
 513         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 514                                    NF_CT_ZONE_DIR_REPL) < 0)
 515                 goto nla_put_failure;
 516         nla_nest_end(skb, nest_parms);
 517
 518         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 519                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 520                 goto nla_put_failure;
 521
 522         if (ctnetlink_dump_status(skb, ct) < 0 ||
 523             ctnetlink_dump_timeout(skb, ct) < 0 ||
 524             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 525             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 526             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 527             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 528             ctnetlink_dump_mark(skb, ct) < 0 ||
 529             ctnetlink_dump_secctx(skb, ct) < 0 ||
 530             ctnetlink_dump_labels(skb, ct) < 0 ||
 531             ctnetlink_dump_id(skb, ct) < 0 ||
 532             ctnetlink_dump_use(skb, ct) < 0 ||
 533             ctnetlink_dump_master(skb, ct) < 0 ||
 534             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 535                 goto nla_put_failure;
 536
 537         nlmsg_end(skb, nlh);
 538         return skb->len;
 539
 540 nlmsg_failure:
 541 nla_put_failure:
 542         nlmsg_cancel(skb, nlh);
 543         return -1;
 544 }
 545
 546 static inline size_t
 547 ctnetlink_proto_size(const struct nf_conn *ct)
 548 {
 549         struct nf_conntrack_l3proto *l3proto;
 550         struct nf_conntrack_l4proto *l4proto;
 551         size_t len = 0;
 552
 553         rcu_read_lock();
 554         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 555         len += l3proto->nla_size;
 556
 557         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 558         len += l4proto->nla_size;
 559         rcu_read_unlock();
 560
 561         return len;
 562 }
 563
 564 static inline size_t
 565 ctnetlink_acct_size(const struct nf_conn *ct)
 566 {
 567         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 568                 return 0;
 569         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 570                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 571                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 572                ;
 573 }
 574
 575 static inline int
 576 ctnetlink_secctx_size(const struct nf_conn *ct)
 577 {
 578 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 579         int len, ret;
 580
 581         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 582         if (ret)
 583                 return 0;
 584
 585         return nla_total_size(0) /* CTA_SECCTX */
 586                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 587 #else
 588         return 0;
 589 #endif
 590 }
 591
 592 static inline size_t
 593 ctnetlink_timestamp_size(const struct nf_conn *ct)
 594 {
 595 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 596         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 597                 return 0;
 598         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 599 #else
 600         return 0;
 601 #endif
 602 }
 603
 604 static inline size_t
 605 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 606 {
 607         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 608                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 609                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 610                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 611                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 612                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 613                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 614                + ctnetlink_acct_size(ct)
 615                + ctnetlink_timestamp_size(ct)
 616                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 617                + nla_total_size(0) /* CTA_PROTOINFO */
 618                + nla_total_size(0) /* CTA_HELP */
 619                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 620                + ctnetlink_secctx_size(ct)
 621 #ifdef CONFIG_NF_NAT_NEEDED
 622                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 623                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 624 #endif
 625 #ifdef CONFIG_NF_CONNTRACK_MARK
 626                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 627 #endif
 628 #ifdef CONFIG_NF_CONNTRACK_ZONES
 629                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 630 #endif
 631                + ctnetlink_proto_size(ct)
 632                + ctnetlink_label_size(ct)
 633                ;
 634 }
 635
 636 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 637 static int
 638 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 639 {
 640         const struct nf_conntrack_zone *zone;
 641         struct net *net;
 642         struct nlmsghdr *nlh;
 643         struct nfgenmsg *nfmsg;
 644         struct nlattr *nest_parms;
 645         struct nf_conn *ct = item->ct;
 646         struct sk_buff *skb;
 647         unsigned int type;
 648         unsigned int flags = 0, group;
 649         int err;
 650
 651         /* ignore our fake conntrack entry */
 652         if (nf_ct_is_untracked(ct))
 653                 return 0;
 654
 655         if (events & (1 << IPCT_DESTROY)) {
 656                 type = IPCTNL_MSG_CT_DELETE;
 657                 group = NFNLGRP_CONNTRACK_DESTROY;
 658         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 659                 type = IPCTNL_MSG_CT_NEW;
 660                 flags = NLM_F_CREATE|NLM_F_EXCL;
 661                 group = NFNLGRP_CONNTRACK_NEW;
 662         } else  if (events) {
 663                 type = IPCTNL_MSG_CT_NEW;
 664                 group = NFNLGRP_CONNTRACK_UPDATE;
 665         } else
 666                 return 0;
 667
 668         net = nf_ct_net(ct);
 669         if (!item->report && !nfnetlink_has_listeners(net, group))
 670                 return 0;
 671
 672         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 673         if (skb == NULL)
 674                 goto errout;
 675
 676         type |= NFNL_SUBSYS_CTNETLINK << 8;
 677         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 678         if (nlh == NULL)
 679                 goto nlmsg_failure;
 680
 681         nfmsg = nlmsg_data(nlh);
 682         nfmsg->nfgen_family = nf_ct_l3num(ct);
 683         nfmsg->version  = NFNETLINK_V0;
 684         nfmsg->res_id   = 0;
 685
 686         rcu_read_lock();
 687         zone = nf_ct_zone(ct);
 688
 689         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 690         if (!nest_parms)
 691                 goto nla_put_failure;
 692         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 693                 goto nla_put_failure;
 694         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 695                                    NF_CT_ZONE_DIR_ORIG) < 0)
 696                 goto nla_put_failure;
 697         nla_nest_end(skb, nest_parms);
 698
 699         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 700         if (!nest_parms)
 701                 goto nla_put_failure;
 702         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 703                 goto nla_put_failure;
 704         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 705                                    NF_CT_ZONE_DIR_REPL) < 0)
 706                 goto nla_put_failure;
 707         nla_nest_end(skb, nest_parms);
 708
 709         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 710                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 711                 goto nla_put_failure;
 712
 713         if (ctnetlink_dump_id(skb, ct) < 0)
 714                 goto nla_put_failure;
 715
 716         if (ctnetlink_dump_status(skb, ct) < 0)
 717                 goto nla_put_failure;
 718
 719         if (events & (1 << IPCT_DESTROY)) {
 720                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 721                     ctnetlink_dump_timestamp(skb, ct) < 0)
 722                         goto nla_put_failure;
 723         } else {
 724                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 725                         goto nla_put_failure;
 726
 727                 if (events & (1 << IPCT_PROTOINFO)
 728                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 729                         goto nla_put_failure;
 730
 731                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 732                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 733                         goto nla_put_failure;
 734
 735 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 736                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 737                     && ctnetlink_dump_secctx(skb, ct) < 0)
 738                         goto nla_put_failure;
 739 #endif
 740                 if (events & (1 << IPCT_LABEL) &&
 741                      ctnetlink_dump_labels(skb, ct) < 0)
 742                         goto nla_put_failure;
 743
 744                 if (events & (1 << IPCT_RELATED) &&
 745                     ctnetlink_dump_master(skb, ct) < 0)
 746                         goto nla_put_failure;
 747
 748                 if (events & (1 << IPCT_SEQADJ) &&
 749                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 750                         goto nla_put_failure;
 751         }
 752
 753 #ifdef CONFIG_NF_CONNTRACK_MARK
 754         if ((events & (1 << IPCT_MARK) || ct->mark)
 755             && ctnetlink_dump_mark(skb, ct) < 0)
 756                 goto nla_put_failure;
 757 #endif
 758         rcu_read_unlock();
 759
 760         nlmsg_end(skb, nlh);
 761         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 762                              GFP_ATOMIC);
 763         if (err == -ENOBUFS || err == -EAGAIN)
 764                 return -ENOBUFS;
 765
 766         return 0;
 767
 768 nla_put_failure:
 769         rcu_read_unlock();
 770         nlmsg_cancel(skb, nlh);
 771 nlmsg_failure:
 772         kfree_skb(skb);
 773 errout:
 774         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 775                 return -ENOBUFS;
 776
 777         return 0;
 778 }
 779 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 780
 781 static int ctnetlink_done(struct netlink_callback *cb)
 782 {
 783         if (cb->args[1])
 784                 nf_ct_put((struct nf_conn *)cb->args[1]);
 785         kfree(cb->data);
 786         return 0;
 787 }
 788
 789 struct ctnetlink_filter {
 790         struct {
 791                 u_int32_t val;
 792                 u_int32_t mask;
 793         } mark;
 794 };
 795
 796 static struct ctnetlink_filter *
 797 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 798 {
 799 #ifdef CONFIG_NF_CONNTRACK_MARK
 800         struct ctnetlink_filter *filter;
 801
 802         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 803         if (filter == NULL)
 804                 return ERR_PTR(-ENOMEM);
 805
 806         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 807         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 808
 809         return filter;
 810 #else
 811         return ERR_PTR(-EOPNOTSUPP);
 812 #endif
 813 }
 814
 815 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 816 {
 817         struct ctnetlink_filter *filter = data;
 818
 819         if (filter == NULL)
 820                 return 1;
 821
 822 #ifdef CONFIG_NF_CONNTRACK_MARK
 823         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 824                 return 1;
 825 #endif
 826
 827         return 0;
 828 }
 829
 830 static int
 831 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 832 {
 833         struct net *net = sock_net(skb->sk);
 834         struct nf_conn *ct, *last;
 835         struct nf_conntrack_tuple_hash *h;
 836         struct hlist_nulls_node *n;
 837         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 838         u_int8_t l3proto = nfmsg->nfgen_family;
 839         int res;
 840         spinlock_t *lockp;
 841
 842         last = (struct nf_conn *)cb->args[1];
 843
 844         local_bh_disable();
 845         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 846 restart:
 847                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 848                 spin_lock(lockp);
 849                 if (cb->args[0] >= net->ct.htable_size) {
 850                         spin_unlock(lockp);
 851                         goto out;
 852                 }
 853                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 854                                          hnnode) {
 855                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 856                                 continue;
 857                         ct = nf_ct_tuplehash_to_ctrack(h);
 858                         /* Dump entries of a given L3 protocol number.
 859                          * If it is not specified, ie. l3proto == 0,
 860                          * then dump everything. */
 861                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 862                                 continue;
 863                         if (cb->args[1]) {
 864                                 if (ct != last)
 865                                         continue;
 866                                 cb->args[1] = 0;
 867                         }
 868                         if (!ctnetlink_filter_match(ct, cb->data))
 869                                 continue;
 870
 871                         rcu_read_lock();
 872                         res =
 873                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 874                                             cb->nlh->nlmsg_seq,
 875                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 876                                             ct);
 877                         rcu_read_unlock();
 878                         if (res < 0) {
 879                                 nf_conntrack_get(&ct->ct_general);
 880                                 cb->args[1] = (unsigned long)ct;
 881                                 spin_unlock(lockp);
 882                                 goto out;
 883                         }
 884                 }
 885                 spin_unlock(lockp);
 886                 if (cb->args[1]) {
 887                         cb->args[1] = 0;
 888                         goto restart;
 889                 }
 890         }
 891 out:
 892         local_bh_enable();
 893         if (last) {
 894                 /* nf ct hash resize happened, now clear the leftover. */
 895                 if ((struct nf_conn *)cb->args[1] == last)
 896                         cb->args[1] = 0;
 897
 898                 nf_ct_put(last);
 899         }
 900
 901         return skb->len;
 902 }
 903
 904 static inline int
 905 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 906 {
 907         struct nlattr *tb[CTA_IP_MAX+1];
 908         struct nf_conntrack_l3proto *l3proto;
 909         int ret = 0;
 910
 911         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 912         if (ret < 0)
 913                 return ret;
 914
 915         rcu_read_lock();
 916         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 917
 918         if (likely(l3proto->nlattr_to_tuple)) {
 919                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 920                                           l3proto->nla_policy);
 921                 if (ret == 0)
 922                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 923         }
 924
 925         rcu_read_unlock();
 926
 927         return ret;
 928 }
 929
 930 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 931         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 932 };
 933
 934 static inline int
 935 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 936                             struct nf_conntrack_tuple *tuple)
 937 {
 938         struct nlattr *tb[CTA_PROTO_MAX+1];
 939         struct nf_conntrack_l4proto *l4proto;
 940         int ret = 0;
 941
 942         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 943         if (ret < 0)
 944                 return ret;
 945
 946         if (!tb[CTA_PROTO_NUM])
 947                 return -EINVAL;
 948         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 949
 950         rcu_read_lock();
 951         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 952
 953         if (likely(l4proto->nlattr_to_tuple)) {
 954                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 955                                           l4proto->nla_policy);
 956                 if (ret == 0)
 957                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 958         }
 959
 960         rcu_read_unlock();
 961
 962         return ret;
 963 }
 964
 965 static int
 966 ctnetlink_parse_zone(const struct nlattr *attr,
 967                      struct nf_conntrack_zone *zone)
 968 {
 969         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
 970                         NF_CT_DEFAULT_ZONE_DIR, 0);
 971 #ifdef CONFIG_NF_CONNTRACK_ZONES
 972         if (attr)
 973                 zone->id = ntohs(nla_get_be16(attr));
 974 #else
 975         if (attr)
 976                 return -EOPNOTSUPP;
 977 #endif
 978         return 0;
 979 }
 980
 981 static int
 982 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
 983                            struct nf_conntrack_zone *zone)
 984 {
 985         int ret;
 986
 987         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
 988                 return -EINVAL;
 989
 990         ret = ctnetlink_parse_zone(attr, zone);
 991         if (ret < 0)
 992                 return ret;
 993
 994         if (type == CTA_TUPLE_REPLY)
 995                 zone->dir = NF_CT_ZONE_DIR_REPL;
 996         else
 997                 zone->dir = NF_CT_ZONE_DIR_ORIG;
 998
 999         return 0;
1000 }
1001
1002 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1003         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1004         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1005         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1006 };
1007
1008 static int
1009 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1010                       struct nf_conntrack_tuple *tuple, u32 type,
1011                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1012 {
1013         struct nlattr *tb[CTA_TUPLE_MAX+1];
1014         int err;
1015
1016         memset(tuple, 0, sizeof(*tuple));
1017
1018         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
1019         if (err < 0)
1020                 return err;
1021
1022         if (!tb[CTA_TUPLE_IP])
1023                 return -EINVAL;
1024
1025         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1026                 return -EOPNOTSUPP;
1027         tuple->src.l3num = l3num;
1028
1029         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1030         if (err < 0)
1031                 return err;
1032
1033         if (!tb[CTA_TUPLE_PROTO])
1034                 return -EINVAL;
1035
1036         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1037         if (err < 0)
1038                 return err;
1039
1040         if (tb[CTA_TUPLE_ZONE]) {
1041                 if (!zone)
1042                         return -EINVAL;
1043
1044                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1045                                                  type, zone);
1046                 if (err < 0)
1047                         return err;
1048         }
1049
1050         /* orig and expect tuples get DIR_ORIGINAL */
1051         if (type == CTA_TUPLE_REPLY)
1052                 tuple->dst.dir = IP_CT_DIR_REPLY;
1053         else
1054                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1055
1056         return 0;
1057 }
1058
1059 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1060         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1061                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1062 };
1063
1064 static inline int
1065 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1066                      struct nlattr **helpinfo)
1067 {
1068         int err;
1069         struct nlattr *tb[CTA_HELP_MAX+1];
1070
1071         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
1072         if (err < 0)
1073                 return err;
1074
1075         if (!tb[CTA_HELP_NAME])
1076                 return -EINVAL;
1077
1078         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1079
1080         if (tb[CTA_HELP_INFO])
1081                 *helpinfo = tb[CTA_HELP_INFO];
1082
1083         return 0;
1084 }
1085
1086 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1087         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1088         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1089         [CTA_STATUS]            = { .type = NLA_U32 },
1090         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1091         [CTA_HELP]              = { .type = NLA_NESTED },
1092         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1093         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1094         [CTA_MARK]              = { .type = NLA_U32 },
1095         [CTA_ID]                = { .type = NLA_U32 },
1096         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1097         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1098         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1099         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1100         [CTA_ZONE]              = { .type = NLA_U16 },
1101         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1102         [CTA_LABELS]            = { .type = NLA_BINARY,
1103                                     .len = NF_CT_LABELS_MAX_SIZE },
1104         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1105                                     .len = NF_CT_LABELS_MAX_SIZE },
1106 };
1107
1108 static int ctnetlink_flush_conntrack(struct net *net,
1109                                      const struct nlattr * const cda[],
1110                                      u32 portid, int report)
1111 {
1112         struct ctnetlink_filter *filter = NULL;
1113
1114         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1115                 filter = ctnetlink_alloc_filter(cda);
1116                 if (IS_ERR(filter))
1117                         return PTR_ERR(filter);
1118         }
1119
1120         nf_ct_iterate_cleanup(net, ctnetlink_filter_match, filter,
1121                               portid, report);
1122         kfree(filter);
1123
1124         return 0;
1125 }
1126
1127 static int
1128 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
1129                         const struct nlmsghdr *nlh,
1130                         const struct nlattr * const cda[])
1131 {
1132         struct net *net = sock_net(ctnl);
1133         struct nf_conntrack_tuple_hash *h;
1134         struct nf_conntrack_tuple tuple;
1135         struct nf_conn *ct;
1136         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1137         u_int8_t u3 = nfmsg->nfgen_family;
1138         struct nf_conntrack_zone zone;
1139         int err;
1140
1141         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1142         if (err < 0)
1143                 return err;
1144
1145         if (cda[CTA_TUPLE_ORIG])
1146                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1147                                             u3, &zone);
1148         else if (cda[CTA_TUPLE_REPLY])
1149                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1150                                             u3, &zone);
1151         else {
1152                 return ctnetlink_flush_conntrack(net, cda,
1153                                                  NETLINK_CB(skb).portid,
1154                                                  nlmsg_report(nlh));
1155         }
1156
1157         if (err < 0)
1158                 return err;
1159
1160         h = nf_conntrack_find_get(net, &zone, &tuple);
1161         if (!h)
1162                 return -ENOENT;
1163
1164         ct = nf_ct_tuplehash_to_ctrack(h);
1165
1166         if (cda[CTA_ID]) {
1167                 __be32 id = nla_get_be32(cda[CTA_ID]);
1168
1169                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1170                         nf_ct_put(ct);
1171                         return -ENOENT;
1172                 }
1173         }
1174
1175         if (del_timer(&ct->timeout))
1176                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1177
1178         nf_ct_put(ct);
1179
1180         return 0;
1181 }
1182
1183 static int
1184 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1185                         const struct nlmsghdr *nlh,
1186                         const struct nlattr * const cda[])
1187 {
1188         struct net *net = sock_net(ctnl);
1189         struct nf_conntrack_tuple_hash *h;
1190         struct nf_conntrack_tuple tuple;
1191         struct nf_conn *ct;
1192         struct sk_buff *skb2 = NULL;
1193         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1194         u_int8_t u3 = nfmsg->nfgen_family;
1195         struct nf_conntrack_zone zone;
1196         int err;
1197
1198         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1199                 struct netlink_dump_control c = {
1200                         .dump = ctnetlink_dump_table,
1201                         .done = ctnetlink_done,
1202                 };
1203
1204                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1205                         struct ctnetlink_filter *filter;
1206
1207                         filter = ctnetlink_alloc_filter(cda);
1208                         if (IS_ERR(filter))
1209                                 return PTR_ERR(filter);
1210
1211                         c.data = filter;
1212                 }
1213                 return netlink_dump_start(ctnl, skb, nlh, &c);
1214         }
1215
1216         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1217         if (err < 0)
1218                 return err;
1219
1220         if (cda[CTA_TUPLE_ORIG])
1221                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1222                                             u3, &zone);
1223         else if (cda[CTA_TUPLE_REPLY])
1224                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1225                                             u3, &zone);
1226         else
1227                 return -EINVAL;
1228
1229         if (err < 0)
1230                 return err;
1231
1232         h = nf_conntrack_find_get(net, &zone, &tuple);
1233         if (!h)
1234                 return -ENOENT;
1235
1236         ct = nf_ct_tuplehash_to_ctrack(h);
1237
1238         err = -ENOMEM;
1239         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1240         if (skb2 == NULL) {
1241                 nf_ct_put(ct);
1242                 return -ENOMEM;
1243         }
1244
1245         rcu_read_lock();
1246         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1247                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1248         rcu_read_unlock();
1249         nf_ct_put(ct);
1250         if (err <= 0)
1251                 goto free;
1252
1253         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1254         if (err < 0)
1255                 goto out;
1256
1257         return 0;
1258
1259 free:
1260         kfree_skb(skb2);
1261 out:
1262         /* this avoids a loop in nfnetlink. */
1263         return err == -EAGAIN ? -ENOBUFS : err;
1264 }
1265
1266 static int ctnetlink_done_list(struct netlink_callback *cb)
1267 {
1268         if (cb->args[1])
1269                 nf_ct_put((struct nf_conn *)cb->args[1]);
1270         return 0;
1271 }
1272
1273 static int
1274 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1275 {
1276         struct nf_conn *ct, *last;
1277         struct nf_conntrack_tuple_hash *h;
1278         struct hlist_nulls_node *n;
1279         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1280         u_int8_t l3proto = nfmsg->nfgen_family;
1281         int res;
1282         int cpu;
1283         struct hlist_nulls_head *list;
1284         struct net *net = sock_net(skb->sk);
1285
1286         if (cb->args[2])
1287                 return 0;
1288
1289         last = (struct nf_conn *)cb->args[1];
1290
1291         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1292                 struct ct_pcpu *pcpu;
1293
1294                 if (!cpu_possible(cpu))
1295                         continue;
1296
1297                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1298                 spin_lock_bh(&pcpu->lock);
1299                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1300 restart:
1301                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1302                         ct = nf_ct_tuplehash_to_ctrack(h);
1303                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1304                                 continue;
1305                         if (cb->args[1]) {
1306                                 if (ct != last)
1307                                         continue;
1308                                 cb->args[1] = 0;
1309                         }
1310                         rcu_read_lock();
1311                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1312                                                   cb->nlh->nlmsg_seq,
1313                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1314                                                   ct);
1315                         rcu_read_unlock();
1316                         if (res < 0) {
1317                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1318                                         continue;
1319                                 cb->args[0] = cpu;
1320                                 cb->args[1] = (unsigned long)ct;
1321                                 spin_unlock_bh(&pcpu->lock);
1322                                 goto out;
1323                         }
1324                 }
1325                 if (cb->args[1]) {
1326                         cb->args[1] = 0;
1327                         goto restart;
1328                 }
1329                 spin_unlock_bh(&pcpu->lock);
1330         }
1331         cb->args[2] = 1;
1332 out:
1333         if (last)
1334                 nf_ct_put(last);
1335
1336         return skb->len;
1337 }
1338
1339 static int
1340 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1341 {
1342         return ctnetlink_dump_list(skb, cb, true);
1343 }
1344
1345 static int
1346 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1347                        const struct nlmsghdr *nlh,
1348                        const struct nlattr * const cda[])
1349 {
1350         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1351                 struct netlink_dump_control c = {
1352                         .dump = ctnetlink_dump_dying,
1353                         .done = ctnetlink_done_list,
1354                 };
1355                 return netlink_dump_start(ctnl, skb, nlh, &c);
1356         }
1357
1358         return -EOPNOTSUPP;
1359 }
1360
1361 static int
1362 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1363 {
1364         return ctnetlink_dump_list(skb, cb, false);
1365 }
1366
1367 static int
1368 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1369                              const struct nlmsghdr *nlh,
1370                              const struct nlattr * const cda[])
1371 {
1372         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1373                 struct netlink_dump_control c = {
1374                         .dump = ctnetlink_dump_unconfirmed,
1375                         .done = ctnetlink_done_list,
1376                 };
1377                 return netlink_dump_start(ctnl, skb, nlh, &c);
1378         }
1379
1380         return -EOPNOTSUPP;
1381 }
1382
1383 #ifdef CONFIG_NF_NAT_NEEDED
1384 static int
1385 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1386                           enum nf_nat_manip_type manip,
1387                           const struct nlattr *attr)
1388 {
1389         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1390         int err;
1391
1392         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1393         if (!parse_nat_setup) {
1394 #ifdef CONFIG_MODULES
1395                 rcu_read_unlock();
1396                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1397                 if (request_module("nf-nat") < 0) {
1398                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1399                         rcu_read_lock();
1400                         return -EOPNOTSUPP;
1401                 }
1402                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1403                 rcu_read_lock();
1404                 if (nfnetlink_parse_nat_setup_hook)
1405                         return -EAGAIN;
1406 #endif
1407                 return -EOPNOTSUPP;
1408         }
1409
1410         err = parse_nat_setup(ct, manip, attr);
1411         if (err == -EAGAIN) {
1412 #ifdef CONFIG_MODULES
1413                 rcu_read_unlock();
1414                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1415                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1416                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1417                         rcu_read_lock();
1418                         return -EOPNOTSUPP;
1419                 }
1420                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1421                 rcu_read_lock();
1422 #else
1423                 err = -EOPNOTSUPP;
1424 #endif
1425         }
1426         return err;
1427 }
1428 #endif
1429
1430 static int
1431 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1432 {
1433         unsigned long d;
1434         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1435         d = ct->status ^ status;
1436
1437         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1438                 /* unchangeable */
1439                 return -EBUSY;
1440
1441         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1442                 /* SEEN_REPLY bit can only be set */
1443                 return -EBUSY;
1444
1445         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1446                 /* ASSURED bit can only be set */
1447                 return -EBUSY;
1448
1449         /* Be careful here, modifying NAT bits can screw up things,
1450          * so don't let users modify them directly if they don't pass
1451          * nf_nat_range. */
1452         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1453         return 0;
1454 }
1455
1456 static int
1457 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1458 {
1459 #ifdef CONFIG_NF_NAT_NEEDED
1460         int ret;
1461
1462         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1463                 return 0;
1464
1465         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1466                                         cda[CTA_NAT_DST]);
1467         if (ret < 0)
1468                 return ret;
1469
1470         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1471                                         cda[CTA_NAT_SRC]);
1472         return ret;
1473 #else
1474         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1475                 return 0;
1476         return -EOPNOTSUPP;
1477 #endif
1478 }
1479
1480 static inline int
1481 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1482 {
1483         struct nf_conntrack_helper *helper;
1484         struct nf_conn_help *help = nfct_help(ct);
1485         char *helpname = NULL;
1486         struct nlattr *helpinfo = NULL;
1487         int err;
1488
1489         /* don't change helper of sibling connections */
1490         if (ct->master)
1491                 return -EBUSY;
1492
1493         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1494         if (err < 0)
1495                 return err;
1496
1497         if (!strcmp(helpname, "")) {
1498                 if (help && help->helper) {
1499                         /* we had a helper before ... */
1500                         nf_ct_remove_expectations(ct);
1501                         RCU_INIT_POINTER(help->helper, NULL);
1502                 }
1503
1504                 return 0;
1505         }
1506
1507         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1508                                             nf_ct_protonum(ct));
1509         if (helper == NULL) {
1510 #ifdef CONFIG_MODULES
1511                 spin_unlock_bh(&nf_conntrack_expect_lock);
1512
1513                 if (request_module("nfct-helper-%s", helpname) < 0) {
1514                         spin_lock_bh(&nf_conntrack_expect_lock);
1515                         return -EOPNOTSUPP;
1516                 }
1517
1518                 spin_lock_bh(&nf_conntrack_expect_lock);
1519                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1520                                                     nf_ct_protonum(ct));
1521                 if (helper)
1522                         return -EAGAIN;
1523 #endif
1524                 return -EOPNOTSUPP;
1525         }
1526
1527         if (help) {
1528                 if (help->helper == helper) {
1529                         /* update private helper data if allowed. */
1530                         if (helper->from_nlattr)
1531                                 helper->from_nlattr(helpinfo, ct);
1532                         return 0;
1533                 } else
1534                         return -EBUSY;
1535         }
1536
1537         /* we cannot set a helper for an existing conntrack */
1538         return -EOPNOTSUPP;
1539 }
1540
1541 static inline int
1542 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1543 {
1544         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1545
1546         if (!del_timer(&ct->timeout))
1547                 return -ETIME;
1548
1549         ct->timeout.expires = jiffies + timeout * HZ;
1550         add_timer(&ct->timeout);
1551
1552         return 0;
1553 }
1554
1555 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1556         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1557         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1558         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1559 };
1560
1561 static inline int
1562 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1563 {
1564         const struct nlattr *attr = cda[CTA_PROTOINFO];
1565         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1566         struct nf_conntrack_l4proto *l4proto;
1567         int err = 0;
1568
1569         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1570         if (err < 0)
1571                 return err;
1572
1573         rcu_read_lock();
1574         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1575         if (l4proto->from_nlattr)
1576                 err = l4proto->from_nlattr(tb, ct);
1577         rcu_read_unlock();
1578
1579         return err;
1580 }
1581
1582 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1583         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1584         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1585         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1586 };
1587
1588 static inline int
1589 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1590 {
1591         int err;
1592         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1593
1594         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1595         if (err < 0)
1596                 return err;
1597
1598         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1599                 return -EINVAL;
1600
1601         seq->correction_pos =
1602                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1603
1604         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1605                 return -EINVAL;
1606
1607         seq->offset_before =
1608                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1609
1610         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1611                 return -EINVAL;
1612
1613         seq->offset_after =
1614                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1615
1616         return 0;
1617 }
1618
1619 static int
1620 ctnetlink_change_seq_adj(struct nf_conn *ct,
1621                          const struct nlattr * const cda[])
1622 {
1623         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1624         int ret = 0;
1625
1626         if (!seqadj)
1627                 return 0;
1628
1629         if (cda[CTA_SEQ_ADJ_ORIG]) {
1630                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1631                                      cda[CTA_SEQ_ADJ_ORIG]);
1632                 if (ret < 0)
1633                         return ret;
1634
1635                 ct->status |= IPS_SEQ_ADJUST;
1636         }
1637
1638         if (cda[CTA_SEQ_ADJ_REPLY]) {
1639                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1640                                      cda[CTA_SEQ_ADJ_REPLY]);
1641                 if (ret < 0)
1642                         return ret;
1643
1644                 ct->status |= IPS_SEQ_ADJUST;
1645         }
1646
1647         return 0;
1648 }
1649
1650 static int
1651 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1652 {
1653 #ifdef CONFIG_NF_CONNTRACK_LABELS
1654         size_t len = nla_len(cda[CTA_LABELS]);
1655         const void *mask = cda[CTA_LABELS_MASK];
1656
1657         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1658                 return -EINVAL;
1659
1660         if (mask) {
1661                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1662                     nla_len(cda[CTA_LABELS_MASK]) != len)
1663                         return -EINVAL;
1664                 mask = nla_data(cda[CTA_LABELS_MASK]);
1665         }
1666
1667         len /= sizeof(u32);
1668
1669         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1670 #else
1671         return -EOPNOTSUPP;
1672 #endif
1673 }
1674
1675 static int
1676 ctnetlink_change_conntrack(struct nf_conn *ct,
1677                            const struct nlattr * const cda[])
1678 {
1679         int err;
1680
1681         /* only allow NAT changes and master assignation for new conntracks */
1682         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1683                 return -EOPNOTSUPP;
1684
1685         if (cda[CTA_HELP]) {
1686                 err = ctnetlink_change_helper(ct, cda);
1687                 if (err < 0)
1688                         return err;
1689         }
1690
1691         if (cda[CTA_TIMEOUT]) {
1692                 err = ctnetlink_change_timeout(ct, cda);
1693                 if (err < 0)
1694                         return err;
1695         }
1696
1697         if (cda[CTA_STATUS]) {
1698                 err = ctnetlink_change_status(ct, cda);
1699                 if (err < 0)
1700                         return err;
1701         }
1702
1703         if (cda[CTA_PROTOINFO]) {
1704                 err = ctnetlink_change_protoinfo(ct, cda);
1705                 if (err < 0)
1706                         return err;
1707         }
1708
1709 #if defined(CONFIG_NF_CONNTRACK_MARK)
1710         if (cda[CTA_MARK])
1711                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1712 #endif
1713
1714         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1715                 err = ctnetlink_change_seq_adj(ct, cda);
1716                 if (err < 0)
1717                         return err;
1718         }
1719
1720         if (cda[CTA_LABELS]) {
1721                 err = ctnetlink_attach_labels(ct, cda);
1722                 if (err < 0)
1723                         return err;
1724         }
1725
1726         return 0;
1727 }
1728
1729 static struct nf_conn *
1730 ctnetlink_create_conntrack(struct net *net,
1731                            const struct nf_conntrack_zone *zone,
1732                            const struct nlattr * const cda[],
1733                            struct nf_conntrack_tuple *otuple,
1734                            struct nf_conntrack_tuple *rtuple,
1735                            u8 u3)
1736 {
1737         struct nf_conn *ct;
1738         int err = -EINVAL;
1739         struct nf_conntrack_helper *helper;
1740         struct nf_conn_tstamp *tstamp;
1741
1742         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1743         if (IS_ERR(ct))
1744                 return ERR_PTR(-ENOMEM);
1745
1746         if (!cda[CTA_TIMEOUT])
1747                 goto err1;
1748         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1749
1750         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1751
1752         rcu_read_lock();
1753         if (cda[CTA_HELP]) {
1754                 char *helpname = NULL;
1755                 struct nlattr *helpinfo = NULL;
1756
1757                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1758                 if (err < 0)
1759                         goto err2;
1760
1761                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1762                                                     nf_ct_protonum(ct));
1763                 if (helper == NULL) {
1764                         rcu_read_unlock();
1765 #ifdef CONFIG_MODULES
1766                         if (request_module("nfct-helper-%s", helpname) < 0) {
1767                                 err = -EOPNOTSUPP;
1768                                 goto err1;
1769                         }
1770
1771                         rcu_read_lock();
1772                         helper = __nf_conntrack_helper_find(helpname,
1773                                                             nf_ct_l3num(ct),
1774                                                             nf_ct_protonum(ct));
1775                         if (helper) {
1776                                 err = -EAGAIN;
1777                                 goto err2;
1778                         }
1779                         rcu_read_unlock();
1780 #endif
1781                         err = -EOPNOTSUPP;
1782                         goto err1;
1783                 } else {
1784                         struct nf_conn_help *help;
1785
1786                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1787                         if (help == NULL) {
1788                                 err = -ENOMEM;
1789                                 goto err2;
1790                         }
1791                         /* set private helper data if allowed. */
1792                         if (helper->from_nlattr)
1793                                 helper->from_nlattr(helpinfo, ct);
1794
1795                         /* not in hash table yet so not strictly necessary */
1796                         RCU_INIT_POINTER(help->helper, helper);
1797                 }
1798         } else {
1799                 /* try an implicit helper assignation */
1800                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1801                 if (err < 0)
1802                         goto err2;
1803         }
1804
1805         err = ctnetlink_setup_nat(ct, cda);
1806         if (err < 0)
1807                 goto err2;
1808
1809         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1810         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1811         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1812         nf_ct_labels_ext_add(ct);
1813         nfct_seqadj_ext_add(ct);
1814         nfct_synproxy_ext_add(ct);
1815
1816         /* we must add conntrack extensions before confirmation. */
1817         ct->status |= IPS_CONFIRMED;
1818
1819         if (cda[CTA_STATUS]) {
1820                 err = ctnetlink_change_status(ct, cda);
1821                 if (err < 0)
1822                         goto err2;
1823         }
1824
1825         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1826                 err = ctnetlink_change_seq_adj(ct, cda);
1827                 if (err < 0)
1828                         goto err2;
1829         }
1830
1831         memset(&ct->proto, 0, sizeof(ct->proto));
1832         if (cda[CTA_PROTOINFO]) {
1833                 err = ctnetlink_change_protoinfo(ct, cda);
1834                 if (err < 0)
1835                         goto err2;
1836         }
1837
1838 #if defined(CONFIG_NF_CONNTRACK_MARK)
1839         if (cda[CTA_MARK])
1840                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1841 #endif
1842
1843         /* setup master conntrack: this is a confirmed expectation */
1844         if (cda[CTA_TUPLE_MASTER]) {
1845                 struct nf_conntrack_tuple master;
1846                 struct nf_conntrack_tuple_hash *master_h;
1847                 struct nf_conn *master_ct;
1848
1849                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1850                                             u3, NULL);
1851                 if (err < 0)
1852                         goto err2;
1853
1854                 master_h = nf_conntrack_find_get(net, zone, &master);
1855                 if (master_h == NULL) {
1856                         err = -ENOENT;
1857                         goto err2;
1858                 }
1859                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1860                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1861                 ct->master = master_ct;
1862         }
1863         tstamp = nf_conn_tstamp_find(ct);
1864         if (tstamp)
1865                 tstamp->start = ktime_get_real_ns();
1866
1867         err = nf_conntrack_hash_check_insert(ct);
1868         if (err < 0)
1869                 goto err2;
1870
1871         rcu_read_unlock();
1872
1873         return ct;
1874
1875 err2:
1876         rcu_read_unlock();
1877 err1:
1878         nf_conntrack_free(ct);
1879         return ERR_PTR(err);
1880 }
1881
1882 static int
1883 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1884                         const struct nlmsghdr *nlh,
1885                         const struct nlattr * const cda[])
1886 {
1887         struct net *net = sock_net(ctnl);
1888         struct nf_conntrack_tuple otuple, rtuple;
1889         struct nf_conntrack_tuple_hash *h = NULL;
1890         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1891         struct nf_conn *ct;
1892         u_int8_t u3 = nfmsg->nfgen_family;
1893         struct nf_conntrack_zone zone;
1894         int err;
1895
1896         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1897         if (err < 0)
1898                 return err;
1899
1900         if (cda[CTA_TUPLE_ORIG]) {
1901                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1902                                             u3, &zone);
1903                 if (err < 0)
1904                         return err;
1905         }
1906
1907         if (cda[CTA_TUPLE_REPLY]) {
1908                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1909                                             u3, &zone);
1910                 if (err < 0)
1911                         return err;
1912         }
1913
1914         if (cda[CTA_TUPLE_ORIG])
1915                 h = nf_conntrack_find_get(net, &zone, &otuple);
1916         else if (cda[CTA_TUPLE_REPLY])
1917                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1918
1919         if (h == NULL) {
1920                 err = -ENOENT;
1921                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1922                         enum ip_conntrack_events events;
1923
1924                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1925                                 return -EINVAL;
1926
1927                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1928                                                         &rtuple, u3);
1929                         if (IS_ERR(ct))
1930                                 return PTR_ERR(ct);
1931
1932                         err = 0;
1933                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1934                                 events = IPCT_RELATED;
1935                         else
1936                                 events = IPCT_NEW;
1937
1938                         if (cda[CTA_LABELS] &&
1939                             ctnetlink_attach_labels(ct, cda) == 0)
1940                                 events |= (1 << IPCT_LABEL);
1941
1942                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1943                                                       (1 << IPCT_ASSURED) |
1944                                                       (1 << IPCT_HELPER) |
1945                                                       (1 << IPCT_PROTOINFO) |
1946                                                       (1 << IPCT_SEQADJ) |
1947                                                       (1 << IPCT_MARK) | events,
1948                                                       ct, NETLINK_CB(skb).portid,
1949                                                       nlmsg_report(nlh));
1950                         nf_ct_put(ct);
1951                 }
1952
1953                 return err;
1954         }
1955         /* implicit 'else' */
1956
1957         err = -EEXIST;
1958         ct = nf_ct_tuplehash_to_ctrack(h);
1959         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1960                 spin_lock_bh(&nf_conntrack_expect_lock);
1961                 err = ctnetlink_change_conntrack(ct, cda);
1962                 spin_unlock_bh(&nf_conntrack_expect_lock);
1963                 if (err == 0) {
1964                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1965                                                       (1 << IPCT_ASSURED) |
1966                                                       (1 << IPCT_HELPER) |
1967                                                       (1 << IPCT_LABEL) |
1968                                                       (1 << IPCT_PROTOINFO) |
1969                                                       (1 << IPCT_SEQADJ) |
1970                                                       (1 << IPCT_MARK),
1971                                                       ct, NETLINK_CB(skb).portid,
1972                                                       nlmsg_report(nlh));
1973                 }
1974         }
1975
1976         nf_ct_put(ct);
1977         return err;
1978 }
1979
1980 static int
1981 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1982                                 __u16 cpu, const struct ip_conntrack_stat *st)
1983 {
1984         struct nlmsghdr *nlh;
1985         struct nfgenmsg *nfmsg;
1986         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1987
1988         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1989         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1990         if (nlh == NULL)
1991                 goto nlmsg_failure;
1992
1993         nfmsg = nlmsg_data(nlh);
1994         nfmsg->nfgen_family = AF_UNSPEC;
1995         nfmsg->version      = NFNETLINK_V0;
1996         nfmsg->res_id       = htons(cpu);
1997
1998         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1999             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2000             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
2001             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2002             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2003             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
2004             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
2005             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2006             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2007                                 htonl(st->insert_failed)) ||
2008             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2009             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2010             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2011             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2012                                 htonl(st->search_restart)))
2013                 goto nla_put_failure;
2014
2015         nlmsg_end(skb, nlh);
2016         return skb->len;
2017
2018 nla_put_failure:
2019 nlmsg_failure:
2020         nlmsg_cancel(skb, nlh);
2021         return -1;
2022 }
2023
2024 static int
2025 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2026 {
2027         int cpu;
2028         struct net *net = sock_net(skb->sk);
2029
2030         if (cb->args[0] == nr_cpu_ids)
2031                 return 0;
2032
2033         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2034                 const struct ip_conntrack_stat *st;
2035
2036                 if (!cpu_possible(cpu))
2037                         continue;
2038
2039                 st = per_cpu_ptr(net->ct.stat, cpu);
2040                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2041                                                     NETLINK_CB(cb->skb).portid,
2042                                                     cb->nlh->nlmsg_seq,
2043                                                     cpu, st) < 0)
2044                                 break;
2045         }
2046         cb->args[0] = cpu;
2047
2048         return skb->len;
2049 }
2050
2051 static int
2052 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
2053                       const struct nlmsghdr *nlh,
2054                       const struct nlattr * const cda[])
2055 {
2056         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2057                 struct netlink_dump_control c = {
2058                         .dump = ctnetlink_ct_stat_cpu_dump,
2059                 };
2060                 return netlink_dump_start(ctnl, skb, nlh, &c);
2061         }
2062
2063         return 0;
2064 }
2065
2066 static int
2067 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2068                             struct net *net)
2069 {
2070         struct nlmsghdr *nlh;
2071         struct nfgenmsg *nfmsg;
2072         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2073         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2074
2075         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
2076         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2077         if (nlh == NULL)
2078                 goto nlmsg_failure;
2079
2080         nfmsg = nlmsg_data(nlh);
2081         nfmsg->nfgen_family = AF_UNSPEC;
2082         nfmsg->version      = NFNETLINK_V0;
2083         nfmsg->res_id       = 0;
2084
2085         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2086                 goto nla_put_failure;
2087
2088         nlmsg_end(skb, nlh);
2089         return skb->len;
2090
2091 nla_put_failure:
2092 nlmsg_failure:
2093         nlmsg_cancel(skb, nlh);
2094         return -1;
2095 }
2096
2097 static int
2098 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
2099                   const struct nlmsghdr *nlh,
2100                   const struct nlattr * const cda[])
2101 {
2102         struct sk_buff *skb2;
2103         int err;
2104
2105         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2106         if (skb2 == NULL)
2107                 return -ENOMEM;
2108
2109         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2110                                           nlh->nlmsg_seq,
2111                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2112                                           sock_net(skb->sk));
2113         if (err <= 0)
2114                 goto free;
2115
2116         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2117         if (err < 0)
2118                 goto out;
2119
2120         return 0;
2121
2122 free:
2123         kfree_skb(skb2);
2124 out:
2125         /* this avoids a loop in nfnetlink. */
2126         return err == -EAGAIN ? -ENOBUFS : err;
2127 }
2128
2129 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2130         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2131         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2132         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2133         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2134         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2135         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2136                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2137         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2138         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2139         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2140         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2141         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2142 };
2143
2144 static struct nf_conntrack_expect *
2145 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2146                        struct nf_conntrack_helper *helper,
2147                        struct nf_conntrack_tuple *tuple,
2148                        struct nf_conntrack_tuple *mask);
2149
2150 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2151 static size_t
2152 ctnetlink_glue_build_size(const struct nf_conn *ct)
2153 {
2154         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2155                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2156                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2157                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2158                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2159                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2160                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2161                + nla_total_size(0) /* CTA_PROTOINFO */
2162                + nla_total_size(0) /* CTA_HELP */
2163                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2164                + ctnetlink_secctx_size(ct)
2165 #ifdef CONFIG_NF_NAT_NEEDED
2166                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2167                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2168 #endif
2169 #ifdef CONFIG_NF_CONNTRACK_MARK
2170                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2171 #endif
2172 #ifdef CONFIG_NF_CONNTRACK_ZONES
2173                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2174 #endif
2175                + ctnetlink_proto_size(ct)
2176                ;
2177 }
2178
2179 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2180                                              enum ip_conntrack_info *ctinfo)
2181 {
2182         struct nf_conn *ct;
2183
2184         ct = nf_ct_get(skb, ctinfo);
2185         if (ct && nf_ct_is_untracked(ct))
2186                 ct = NULL;
2187
2188         return ct;
2189 }
2190
2191 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2192 {
2193         const struct nf_conntrack_zone *zone;
2194         struct nlattr *nest_parms;
2195
2196         rcu_read_lock();
2197         zone = nf_ct_zone(ct);
2198
2199         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2200         if (!nest_parms)
2201                 goto nla_put_failure;
2202         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2203                 goto nla_put_failure;
2204         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2205                                    NF_CT_ZONE_DIR_ORIG) < 0)
2206                 goto nla_put_failure;
2207         nla_nest_end(skb, nest_parms);
2208
2209         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2210         if (!nest_parms)
2211                 goto nla_put_failure;
2212         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2213                 goto nla_put_failure;
2214         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2215                                    NF_CT_ZONE_DIR_REPL) < 0)
2216                 goto nla_put_failure;
2217         nla_nest_end(skb, nest_parms);
2218
2219         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2220                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2221                 goto nla_put_failure;
2222
2223         if (ctnetlink_dump_id(skb, ct) < 0)
2224                 goto nla_put_failure;
2225
2226         if (ctnetlink_dump_status(skb, ct) < 0)
2227                 goto nla_put_failure;
2228
2229         if (ctnetlink_dump_timeout(skb, ct) < 0)
2230                 goto nla_put_failure;
2231
2232         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2233                 goto nla_put_failure;
2234
2235         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2236                 goto nla_put_failure;
2237
2238 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2239         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2240                 goto nla_put_failure;
2241 #endif
2242         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2243                 goto nla_put_failure;
2244
2245         if ((ct->status & IPS_SEQ_ADJUST) &&
2246             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2247                 goto nla_put_failure;
2248
2249 #ifdef CONFIG_NF_CONNTRACK_MARK
2250         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2251                 goto nla_put_failure;
2252 #endif
2253         if (ctnetlink_dump_labels(skb, ct) < 0)
2254                 goto nla_put_failure;
2255         rcu_read_unlock();
2256         return 0;
2257
2258 nla_put_failure:
2259         rcu_read_unlock();
2260         return -ENOSPC;
2261 }
2262
2263 static int
2264 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2265                      enum ip_conntrack_info ctinfo,
2266                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2267 {
2268         struct nlattr *nest_parms;
2269
2270         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2271         if (!nest_parms)
2272                 goto nla_put_failure;
2273
2274         if (__ctnetlink_glue_build(skb, ct) < 0)
2275                 goto nla_put_failure;
2276
2277         nla_nest_end(skb, nest_parms);
2278
2279         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2280                 goto nla_put_failure;
2281
2282         return 0;
2283
2284 nla_put_failure:
2285         return -ENOSPC;
2286 }
2287
2288 static int
2289 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2290 {
2291         int err;
2292
2293         if (cda[CTA_TIMEOUT]) {
2294                 err = ctnetlink_change_timeout(ct, cda);
2295                 if (err < 0)
2296                         return err;
2297         }
2298         if (cda[CTA_STATUS]) {
2299                 err = ctnetlink_change_status(ct, cda);
2300                 if (err < 0)
2301                         return err;
2302         }
2303         if (cda[CTA_HELP]) {
2304                 err = ctnetlink_change_helper(ct, cda);
2305                 if (err < 0)
2306                         return err;
2307         }
2308         if (cda[CTA_LABELS]) {
2309                 err = ctnetlink_attach_labels(ct, cda);
2310                 if (err < 0)
2311                         return err;
2312         }
2313 #if defined(CONFIG_NF_CONNTRACK_MARK)
2314         if (cda[CTA_MARK]) {
2315                 u32 mask = 0, mark, newmark;
2316                 if (cda[CTA_MARK_MASK])
2317                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2318
2319                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2320                 newmark = (ct->mark & mask) ^ mark;
2321                 if (newmark != ct->mark)
2322                         ct->mark = newmark;
2323         }
2324 #endif
2325         return 0;
2326 }
2327
2328 static int
2329 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2330 {
2331         struct nlattr *cda[CTA_MAX+1];
2332         int ret;
2333
2334         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2335         if (ret < 0)
2336                 return ret;
2337
2338         spin_lock_bh(&nf_conntrack_expect_lock);
2339         ret = ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2340         spin_unlock_bh(&nf_conntrack_expect_lock);
2341
2342         return ret;
2343 }
2344
2345 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2346                                     const struct nf_conn *ct,
2347                                     struct nf_conntrack_tuple *tuple,
2348                                     struct nf_conntrack_tuple *mask)
2349 {
2350         int err;
2351
2352         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2353                                     nf_ct_l3num(ct), NULL);
2354         if (err < 0)
2355                 return err;
2356
2357         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2358                                      nf_ct_l3num(ct), NULL);
2359 }
2360
2361 static int
2362 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2363                              u32 portid, u32 report)
2364 {
2365         struct nlattr *cda[CTA_EXPECT_MAX+1];
2366         struct nf_conntrack_tuple tuple, mask;
2367         struct nf_conntrack_helper *helper = NULL;
2368         struct nf_conntrack_expect *exp;
2369         int err;
2370
2371         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2372         if (err < 0)
2373                 return err;
2374
2375         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2376                                        ct, &tuple, &mask);
2377         if (err < 0)
2378                 return err;
2379
2380         if (cda[CTA_EXPECT_HELP_NAME]) {
2381                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2382
2383                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2384                                                     nf_ct_protonum(ct));
2385                 if (helper == NULL)
2386                         return -EOPNOTSUPP;
2387         }
2388
2389         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2390                                      helper, &tuple, &mask);
2391         if (IS_ERR(exp))
2392                 return PTR_ERR(exp);
2393
2394         err = nf_ct_expect_related_report(exp, portid, report);
2395         if (err < 0) {
2396                 nf_ct_expect_put(exp);
2397                 return err;
2398         }
2399
2400         return 0;
2401 }
2402
2403 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2404                                   enum ip_conntrack_info ctinfo, int diff)
2405 {
2406         if (!(ct->status & IPS_NAT_MASK))
2407                 return;
2408
2409         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2410 }
2411
2412 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2413         .get_ct         = ctnetlink_glue_get_ct,
2414         .build_size     = ctnetlink_glue_build_size,
2415         .build          = ctnetlink_glue_build,
2416         .parse          = ctnetlink_glue_parse,
2417         .attach_expect  = ctnetlink_glue_attach_expect,
2418         .seq_adjust     = ctnetlink_glue_seqadj,
2419 };
2420 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2421
2422 /***********************************************************************
2423  * EXPECT
2424  ***********************************************************************/
2425
2426 static inline int
2427 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2428                          const struct nf_conntrack_tuple *tuple,
2429                          u32 type)
2430 {
2431         struct nlattr *nest_parms;
2432
2433         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2434         if (!nest_parms)
2435                 goto nla_put_failure;
2436         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2437                 goto nla_put_failure;
2438         nla_nest_end(skb, nest_parms);
2439
2440         return 0;
2441
2442 nla_put_failure:
2443         return -1;
2444 }
2445
2446 static inline int
2447 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2448                         const struct nf_conntrack_tuple *tuple,
2449                         const struct nf_conntrack_tuple_mask *mask)
2450 {
2451         int ret;
2452         struct nf_conntrack_l3proto *l3proto;
2453         struct nf_conntrack_l4proto *l4proto;
2454         struct nf_conntrack_tuple m;
2455         struct nlattr *nest_parms;
2456
2457         memset(&m, 0xFF, sizeof(m));
2458         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2459         m.src.u.all = mask->src.u.all;
2460         m.dst.protonum = tuple->dst.protonum;
2461
2462         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2463         if (!nest_parms)
2464                 goto nla_put_failure;
2465
2466         rcu_read_lock();
2467         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2468         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2469         if (ret >= 0) {
2470                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2471                                                tuple->dst.protonum);
2472         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2473         }
2474         rcu_read_unlock();
2475
2476         if (unlikely(ret < 0))
2477                 goto nla_put_failure;
2478
2479         nla_nest_end(skb, nest_parms);
2480
2481         return 0;
2482
2483 nla_put_failure:
2484         return -1;
2485 }
2486
2487 static const union nf_inet_addr any_addr;
2488
2489 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2490 {
2491         static __read_mostly siphash_key_t exp_id_seed;
2492         unsigned long a, b, c, d;
2493
2494         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2495
2496         a = (unsigned long)exp;
2497         b = (unsigned long)exp->helper;
2498         c = (unsigned long)exp->master;
2499         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2500
2501 #ifdef CONFIG_64BIT
2502         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2503 #else
2504         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2505 #endif
2506 }
2507
2508 static int
2509 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2510                           const struct nf_conntrack_expect *exp)
2511 {
2512         struct nf_conn *master = exp->master;
2513         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2514         struct nf_conn_help *help;
2515 #ifdef CONFIG_NF_NAT_NEEDED
2516         struct nlattr *nest_parms;
2517         struct nf_conntrack_tuple nat_tuple = {};
2518 #endif
2519         struct nf_ct_helper_expectfn *expfn;
2520
2521         if (timeout < 0)
2522                 timeout = 0;
2523
2524         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2525                 goto nla_put_failure;
2526         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2527                 goto nla_put_failure;
2528         if (ctnetlink_exp_dump_tuple(skb,
2529                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2530                                  CTA_EXPECT_MASTER) < 0)
2531                 goto nla_put_failure;
2532
2533 #ifdef CONFIG_NF_NAT_NEEDED
2534         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2535             exp->saved_proto.all) {
2536                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2537                 if (!nest_parms)
2538                         goto nla_put_failure;
2539
2540                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2541                         goto nla_put_failure;
2542
2543                 nat_tuple.src.l3num = nf_ct_l3num(master);
2544                 nat_tuple.src.u3 = exp->saved_addr;
2545                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2546                 nat_tuple.src.u = exp->saved_proto;
2547
2548                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2549                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2550                         goto nla_put_failure;
2551                 nla_nest_end(skb, nest_parms);
2552         }
2553 #endif
2554         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2555             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
2556             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2557             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2558                 goto nla_put_failure;
2559         help = nfct_help(master);
2560         if (help) {
2561                 struct nf_conntrack_helper *helper;
2562
2563                 helper = rcu_dereference(help->helper);
2564                 if (helper &&
2565                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2566                         goto nla_put_failure;
2567         }
2568         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2569         if (expfn != NULL &&
2570             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2571                 goto nla_put_failure;
2572
2573         return 0;
2574
2575 nla_put_failure:
2576         return -1;
2577 }
2578
2579 static int
2580 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2581                         int event, const struct nf_conntrack_expect *exp)
2582 {
2583         struct nlmsghdr *nlh;
2584         struct nfgenmsg *nfmsg;
2585         unsigned int flags = portid ? NLM_F_MULTI : 0;
2586
2587         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2588         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2589         if (nlh == NULL)
2590                 goto nlmsg_failure;
2591
2592         nfmsg = nlmsg_data(nlh);
2593         nfmsg->nfgen_family = exp->tuple.src.l3num;
2594         nfmsg->version      = NFNETLINK_V0;
2595         nfmsg->res_id       = 0;
2596
2597         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2598                 goto nla_put_failure;
2599
2600         nlmsg_end(skb, nlh);
2601         return skb->len;
2602
2603 nlmsg_failure:
2604 nla_put_failure:
2605         nlmsg_cancel(skb, nlh);
2606         return -1;
2607 }
2608
2609 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2610 static int
2611 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2612 {
2613         struct nf_conntrack_expect *exp = item->exp;
2614         struct net *net = nf_ct_exp_net(exp);
2615         struct nlmsghdr *nlh;
2616         struct nfgenmsg *nfmsg;
2617         struct sk_buff *skb;
2618         unsigned int type, group;
2619         int flags = 0;
2620
2621         if (events & (1 << IPEXP_DESTROY)) {
2622                 type = IPCTNL_MSG_EXP_DELETE;
2623                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2624         } else if (events & (1 << IPEXP_NEW)) {
2625                 type = IPCTNL_MSG_EXP_NEW;
2626                 flags = NLM_F_CREATE|NLM_F_EXCL;
2627                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2628         } else
2629                 return 0;
2630
2631         if (!item->report && !nfnetlink_has_listeners(net, group))
2632                 return 0;
2633
2634         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2635         if (skb == NULL)
2636                 goto errout;
2637
2638         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2639         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2640         if (nlh == NULL)
2641                 goto nlmsg_failure;
2642
2643         nfmsg = nlmsg_data(nlh);
2644         nfmsg->nfgen_family = exp->tuple.src.l3num;
2645         nfmsg->version      = NFNETLINK_V0;
2646         nfmsg->res_id       = 0;
2647
2648         rcu_read_lock();
2649         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2650                 goto nla_put_failure;
2651         rcu_read_unlock();
2652
2653         nlmsg_end(skb, nlh);
2654         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2655         return 0;
2656
2657 nla_put_failure:
2658         rcu_read_unlock();
2659         nlmsg_cancel(skb, nlh);
2660 nlmsg_failure:
2661         kfree_skb(skb);
2662 errout:
2663         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2664         return 0;
2665 }
2666 #endif
2667 static int ctnetlink_exp_done(struct netlink_callback *cb)
2668 {
2669         if (cb->args[1])
2670                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2671         return 0;
2672 }
2673
2674 static int
2675 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2676 {
2677         struct net *net = sock_net(skb->sk);
2678         struct nf_conntrack_expect *exp, *last;
2679         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2680         u_int8_t l3proto = nfmsg->nfgen_family;
2681
2682         rcu_read_lock();
2683         last = (struct nf_conntrack_expect *)cb->args[1];
2684         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2685 restart:
2686                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2687                                      hnode) {
2688                         if (l3proto && exp->tuple.src.l3num != l3proto)
2689                                 continue;
2690                         if (cb->args[1]) {
2691                                 if (exp != last)
2692                                         continue;
2693                                 cb->args[1] = 0;
2694                         }
2695                         if (ctnetlink_exp_fill_info(skb,
2696                                                     NETLINK_CB(cb->skb).portid,
2697                                                     cb->nlh->nlmsg_seq,
2698                                                     IPCTNL_MSG_EXP_NEW,
2699                                                     exp) < 0) {
2700                                 if (!atomic_inc_not_zero(&exp->use))
2701                                         continue;
2702                                 cb->args[1] = (unsigned long)exp;
2703                                 goto out;
2704                         }
2705                 }
2706                 if (cb->args[1]) {
2707                         cb->args[1] = 0;
2708                         goto restart;
2709                 }
2710         }
2711 out:
2712         rcu_read_unlock();
2713         if (last)
2714                 nf_ct_expect_put(last);
2715
2716         return skb->len;
2717 }
2718
2719 static int
2720 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2721 {
2722         struct nf_conntrack_expect *exp, *last;
2723         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2724         struct nf_conn *ct = cb->data;
2725         struct nf_conn_help *help = nfct_help(ct);
2726         u_int8_t l3proto = nfmsg->nfgen_family;
2727
2728         if (cb->args[0])
2729                 return 0;
2730
2731         rcu_read_lock();
2732         last = (struct nf_conntrack_expect *)cb->args[1];
2733 restart:
2734         hlist_for_each_entry(exp, &help->expectations, lnode) {
2735                 if (l3proto && exp->tuple.src.l3num != l3proto)
2736                         continue;
2737                 if (cb->args[1]) {
2738                         if (exp != last)
2739                                 continue;
2740                         cb->args[1] = 0;
2741                 }
2742                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2743                                             cb->nlh->nlmsg_seq,
2744                                             IPCTNL_MSG_EXP_NEW,
2745                                             exp) < 0) {
2746                         if (!atomic_inc_not_zero(&exp->use))
2747                                 continue;
2748                         cb->args[1] = (unsigned long)exp;
2749                         goto out;
2750                 }
2751         }
2752         if (cb->args[1]) {
2753                 cb->args[1] = 0;
2754                 goto restart;
2755         }
2756         cb->args[0] = 1;
2757 out:
2758         rcu_read_unlock();
2759         if (last)
2760                 nf_ct_expect_put(last);
2761
2762         return skb->len;
2763 }
2764
2765 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2766                                  const struct nlmsghdr *nlh,
2767                                  const struct nlattr * const cda[])
2768 {
2769         int err;
2770         struct net *net = sock_net(ctnl);
2771         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2772         u_int8_t u3 = nfmsg->nfgen_family;
2773         struct nf_conntrack_tuple tuple;
2774         struct nf_conntrack_tuple_hash *h;
2775         struct nf_conn *ct;
2776         struct nf_conntrack_zone zone;
2777         struct netlink_dump_control c = {
2778                 .dump = ctnetlink_exp_ct_dump_table,
2779                 .done = ctnetlink_exp_done,
2780         };
2781
2782         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2783                                     u3, NULL);
2784         if (err < 0)
2785                 return err;
2786
2787         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2788         if (err < 0)
2789                 return err;
2790
2791         h = nf_conntrack_find_get(net, &zone, &tuple);
2792         if (!h)
2793                 return -ENOENT;
2794
2795         ct = nf_ct_tuplehash_to_ctrack(h);
2796         c.data = ct;
2797
2798         err = netlink_dump_start(ctnl, skb, nlh, &c);
2799         nf_ct_put(ct);
2800
2801         return err;
2802 }
2803
2804 static int
2805 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2806                      const struct nlmsghdr *nlh,
2807                      const struct nlattr * const cda[])
2808 {
2809         struct net *net = sock_net(ctnl);
2810         struct nf_conntrack_tuple tuple;
2811         struct nf_conntrack_expect *exp;
2812         struct sk_buff *skb2;
2813         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2814         u_int8_t u3 = nfmsg->nfgen_family;
2815         struct nf_conntrack_zone zone;
2816         int err;
2817
2818         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2819                 if (cda[CTA_EXPECT_MASTER])
2820                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2821                 else {
2822                         struct netlink_dump_control c = {
2823                                 .dump = ctnetlink_exp_dump_table,
2824                                 .done = ctnetlink_exp_done,
2825                         };
2826                         return netlink_dump_start(ctnl, skb, nlh, &c);
2827                 }
2828         }
2829
2830         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2831         if (err < 0)
2832                 return err;
2833
2834         if (cda[CTA_EXPECT_TUPLE])
2835                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2836                                             u3, NULL);
2837         else if (cda[CTA_EXPECT_MASTER])
2838                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2839                                             u3, NULL);
2840         else
2841                 return -EINVAL;
2842
2843         if (err < 0)
2844                 return err;
2845
2846         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2847         if (!exp)
2848                 return -ENOENT;
2849
2850         if (cda[CTA_EXPECT_ID]) {
2851                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2852
2853                 if (id != nf_expect_get_id(exp)) {
2854                         nf_ct_expect_put(exp);
2855                         return -ENOENT;
2856                 }
2857         }
2858
2859         err = -ENOMEM;
2860         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2861         if (skb2 == NULL) {
2862                 nf_ct_expect_put(exp);
2863                 goto out;
2864         }
2865
2866         rcu_read_lock();
2867         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2868                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2869         rcu_read_unlock();
2870         nf_ct_expect_put(exp);
2871         if (err <= 0)
2872                 goto free;
2873
2874         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2875         if (err < 0)
2876                 goto out;
2877
2878         return 0;
2879
2880 free:
2881         kfree_skb(skb2);
2882 out:
2883         /* this avoids a loop in nfnetlink. */
2884         return err == -EAGAIN ? -ENOBUFS : err;
2885 }
2886
2887 static int
2888 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2889                      const struct nlmsghdr *nlh,
2890                      const struct nlattr * const cda[])
2891 {
2892         struct net *net = sock_net(ctnl);
2893         struct nf_conntrack_expect *exp;
2894         struct nf_conntrack_tuple tuple;
2895         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2896         struct hlist_node *next;
2897         u_int8_t u3 = nfmsg->nfgen_family;
2898         struct nf_conntrack_zone zone;
2899         unsigned int i;
2900         int err;
2901
2902         if (cda[CTA_EXPECT_TUPLE]) {
2903                 /* delete a single expect by tuple */
2904                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2905                 if (err < 0)
2906                         return err;
2907
2908                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2909                                             u3, NULL);
2910                 if (err < 0)
2911                         return err;
2912
2913                 /* bump usage count to 2 */
2914                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2915                 if (!exp)
2916                         return -ENOENT;
2917
2918                 if (cda[CTA_EXPECT_ID]) {
2919                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2920                         if (ntohl(id) != (u32)(unsigned long)exp) {
2921                                 nf_ct_expect_put(exp);
2922                                 return -ENOENT;
2923                         }
2924                 }
2925
2926                 /* after list removal, usage count == 1 */
2927                 spin_lock_bh(&nf_conntrack_expect_lock);
2928                 if (del_timer(&exp->timeout)) {
2929                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2930                                                    nlmsg_report(nlh));
2931                         nf_ct_expect_put(exp);
2932                 }
2933                 spin_unlock_bh(&nf_conntrack_expect_lock);
2934                 /* have to put what we 'get' above.
2935                  * after this line usage count == 0 */
2936                 nf_ct_expect_put(exp);
2937         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2938                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2939                 struct nf_conn_help *m_help;
2940
2941                 /* delete all expectations for this helper */
2942                 spin_lock_bh(&nf_conntrack_expect_lock);
2943                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2944                         hlist_for_each_entry_safe(exp, next,
2945                                                   &net->ct.expect_hash[i],
2946                                                   hnode) {
2947                                 m_help = nfct_help(exp->master);
2948                                 if (!strcmp(m_help->helper->name, name) &&
2949                                     del_timer(&exp->timeout)) {
2950                                         nf_ct_unlink_expect_report(exp,
2951                                                         NETLINK_CB(skb).portid,
2952                                                         nlmsg_report(nlh));
2953                                         nf_ct_expect_put(exp);
2954                                 }
2955                         }
2956                 }
2957                 spin_unlock_bh(&nf_conntrack_expect_lock);
2958         } else {
2959                 /* This basically means we have to flush everything*/
2960                 spin_lock_bh(&nf_conntrack_expect_lock);
2961                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2962                         hlist_for_each_entry_safe(exp, next,
2963                                                   &net->ct.expect_hash[i],
2964                                                   hnode) {
2965                                 if (del_timer(&exp->timeout)) {
2966                                         nf_ct_unlink_expect_report(exp,
2967                                                         NETLINK_CB(skb).portid,
2968                                                         nlmsg_report(nlh));
2969                                         nf_ct_expect_put(exp);
2970                                 }
2971                         }
2972                 }
2973                 spin_unlock_bh(&nf_conntrack_expect_lock);
2974         }
2975
2976         return 0;
2977 }
2978 static int
2979 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2980                         const struct nlattr * const cda[])
2981 {
2982         if (cda[CTA_EXPECT_TIMEOUT]) {
2983                 if (!del_timer(&x->timeout))
2984                         return -ETIME;
2985
2986                 x->timeout.expires = jiffies +
2987                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2988                 add_timer(&x->timeout);
2989         }
2990         return 0;
2991 }
2992
2993 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2994         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2995         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2996 };
2997
2998 static int
2999 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3000                            struct nf_conntrack_expect *exp,
3001                            u_int8_t u3)
3002 {
3003 #ifdef CONFIG_NF_NAT_NEEDED
3004         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3005         struct nf_conntrack_tuple nat_tuple = {};
3006         int err;
3007
3008         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
3009         if (err < 0)
3010                 return err;
3011
3012         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3013                 return -EINVAL;
3014
3015         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3016                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3017                                     u3, NULL);
3018         if (err < 0)
3019                 return err;
3020
3021         exp->saved_addr = nat_tuple.src.u3;
3022         exp->saved_proto = nat_tuple.src.u;
3023         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3024
3025         return 0;
3026 #else
3027         return -EOPNOTSUPP;
3028 #endif
3029 }
3030
3031 static struct nf_conntrack_expect *
3032 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3033                        struct nf_conntrack_helper *helper,
3034                        struct nf_conntrack_tuple *tuple,
3035                        struct nf_conntrack_tuple *mask)
3036 {
3037         u_int32_t class = 0;
3038         struct nf_conntrack_expect *exp;
3039         struct nf_conn_help *help;
3040         int err;
3041
3042         if (cda[CTA_EXPECT_CLASS] && helper) {
3043                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3044                 if (class > helper->expect_class_max)
3045                         return ERR_PTR(-EINVAL);
3046         }
3047         exp = nf_ct_expect_alloc(ct);
3048         if (!exp)
3049                 return ERR_PTR(-ENOMEM);
3050
3051         help = nfct_help(ct);
3052         if (!help) {
3053                 if (!cda[CTA_EXPECT_TIMEOUT]) {
3054                         err = -EINVAL;
3055                         goto err_out;
3056                 }
3057                 exp->timeout.expires =
3058                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3059
3060                 exp->flags = NF_CT_EXPECT_USERSPACE;
3061                 if (cda[CTA_EXPECT_FLAGS]) {
3062                         exp->flags |=
3063                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3064                 }
3065         } else {
3066                 if (cda[CTA_EXPECT_FLAGS]) {
3067                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3068                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3069                 } else
3070                         exp->flags = 0;
3071         }
3072         if (cda[CTA_EXPECT_FN]) {
3073                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3074                 struct nf_ct_helper_expectfn *expfn;
3075
3076                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3077                 if (expfn == NULL) {
3078                         err = -EINVAL;
3079                         goto err_out;
3080                 }
3081                 exp->expectfn = expfn->expectfn;
3082         } else
3083                 exp->expectfn = NULL;
3084
3085         exp->class = class;
3086         exp->master = ct;
3087         exp->helper = helper;
3088         exp->tuple = *tuple;
3089         exp->mask.src.u3 = mask->src.u3;
3090         exp->mask.src.u.all = mask->src.u.all;
3091
3092         if (cda[CTA_EXPECT_NAT]) {
3093                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3094                                                  exp, nf_ct_l3num(ct));
3095                 if (err < 0)
3096                         goto err_out;
3097         }
3098         return exp;
3099 err_out:
3100         nf_ct_expect_put(exp);
3101         return ERR_PTR(err);
3102 }
3103
3104 static int
3105 ctnetlink_create_expect(struct net *net,
3106                         const struct nf_conntrack_zone *zone,
3107                         const struct nlattr * const cda[],
3108                         u_int8_t u3, u32 portid, int report)
3109 {
3110         struct nf_conntrack_tuple tuple, mask, master_tuple;
3111         struct nf_conntrack_tuple_hash *h = NULL;
3112         struct nf_conntrack_helper *helper = NULL;
3113         struct nf_conntrack_expect *exp;
3114         struct nf_conn *ct;
3115         int err;
3116
3117         /* caller guarantees that those three CTA_EXPECT_* exist */
3118         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3119                                     u3, NULL);
3120         if (err < 0)
3121                 return err;
3122         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3123                                     u3, NULL);
3124         if (err < 0)
3125                 return err;
3126         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3127                                     u3, NULL);
3128         if (err < 0)
3129                 return err;
3130
3131         /* Look for master conntrack of this expectation */
3132         h = nf_conntrack_find_get(net, zone, &master_tuple);
3133         if (!h)
3134                 return -ENOENT;
3135         ct = nf_ct_tuplehash_to_ctrack(h);
3136
3137         if (cda[CTA_EXPECT_HELP_NAME]) {
3138                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3139
3140                 helper = __nf_conntrack_helper_find(helpname, u3,
3141                                                     nf_ct_protonum(ct));
3142                 if (helper == NULL) {
3143 #ifdef CONFIG_MODULES
3144                         if (request_module("nfct-helper-%s", helpname) < 0) {
3145                                 err = -EOPNOTSUPP;
3146                                 goto err_ct;
3147                         }
3148                         helper = __nf_conntrack_helper_find(helpname, u3,
3149                                                             nf_ct_protonum(ct));
3150                         if (helper) {
3151                                 err = -EAGAIN;
3152                                 goto err_ct;
3153                         }
3154 #endif
3155                         err = -EOPNOTSUPP;
3156                         goto err_ct;
3157                 }
3158         }
3159
3160         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3161         if (IS_ERR(exp)) {
3162                 err = PTR_ERR(exp);
3163                 goto err_ct;
3164         }
3165
3166         err = nf_ct_expect_related_report(exp, portid, report);
3167         nf_ct_expect_put(exp);
3168 err_ct:
3169         nf_ct_put(ct);
3170         return err;
3171 }
3172
3173 static int
3174 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
3175                      const struct nlmsghdr *nlh,
3176                      const struct nlattr * const cda[])
3177 {
3178         struct net *net = sock_net(ctnl);
3179         struct nf_conntrack_tuple tuple;
3180         struct nf_conntrack_expect *exp;
3181         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3182         u_int8_t u3 = nfmsg->nfgen_family;
3183         struct nf_conntrack_zone zone;
3184         int err;
3185
3186         if (!cda[CTA_EXPECT_TUPLE]
3187             || !cda[CTA_EXPECT_MASK]
3188             || !cda[CTA_EXPECT_MASTER])
3189                 return -EINVAL;
3190
3191         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3192         if (err < 0)
3193                 return err;
3194
3195         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3196                                     u3, NULL);
3197         if (err < 0)
3198                 return err;
3199
3200         spin_lock_bh(&nf_conntrack_expect_lock);
3201         exp = __nf_ct_expect_find(net, &zone, &tuple);
3202         if (!exp) {
3203                 spin_unlock_bh(&nf_conntrack_expect_lock);
3204                 err = -ENOENT;
3205                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3206                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3207                                                       NETLINK_CB(skb).portid,
3208                                                       nlmsg_report(nlh));
3209                 }
3210                 return err;
3211         }
3212
3213         err = -EEXIST;
3214         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3215                 err = ctnetlink_change_expect(exp, cda);
3216         spin_unlock_bh(&nf_conntrack_expect_lock);
3217
3218         return err;
3219 }
3220
3221 static int
3222 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3223                              const struct ip_conntrack_stat *st)
3224 {
3225         struct nlmsghdr *nlh;
3226         struct nfgenmsg *nfmsg;
3227         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3228
3229         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3230         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3231         if (nlh == NULL)
3232                 goto nlmsg_failure;
3233
3234         nfmsg = nlmsg_data(nlh);
3235         nfmsg->nfgen_family = AF_UNSPEC;
3236         nfmsg->version      = NFNETLINK_V0;
3237         nfmsg->res_id       = htons(cpu);
3238
3239         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3240             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3241             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3242                 goto nla_put_failure;
3243
3244         nlmsg_end(skb, nlh);
3245         return skb->len;
3246
3247 nla_put_failure:
3248 nlmsg_failure:
3249         nlmsg_cancel(skb, nlh);
3250         return -1;
3251 }
3252
3253 static int
3254 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3255 {
3256         int cpu;
3257         struct net *net = sock_net(skb->sk);
3258
3259         if (cb->args[0] == nr_cpu_ids)
3260                 return 0;
3261
3262         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3263                 const struct ip_conntrack_stat *st;
3264
3265                 if (!cpu_possible(cpu))
3266                         continue;
3267
3268                 st = per_cpu_ptr(net->ct.stat, cpu);
3269                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3270                                                  cb->nlh->nlmsg_seq,
3271                                                  cpu, st) < 0)
3272                         break;
3273         }
3274         cb->args[0] = cpu;
3275
3276         return skb->len;
3277 }
3278
3279 static int
3280 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3281                        const struct nlmsghdr *nlh,
3282                        const struct nlattr * const cda[])
3283 {
3284         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3285                 struct netlink_dump_control c = {
3286                         .dump = ctnetlink_exp_stat_cpu_dump,
3287                 };
3288                 return netlink_dump_start(ctnl, skb, nlh, &c);
3289         }
3290
3291         return 0;
3292 }
3293
3294 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3295 static struct nf_ct_event_notifier ctnl_notifier = {
3296         .fcn = ctnetlink_conntrack_event,
3297 };
3298
3299 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3300         .fcn = ctnetlink_expect_event,
3301 };
3302 #endif
3303
3304 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3305         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3306                                             .attr_count = CTA_MAX,
3307                                             .policy = ct_nla_policy },
3308         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3309                                             .attr_count = CTA_MAX,
3310                                             .policy = ct_nla_policy },
3311         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3312                                             .attr_count = CTA_MAX,
3313                                             .policy = ct_nla_policy },
3314         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3315                                             .attr_count = CTA_MAX,
3316                                             .policy = ct_nla_policy },
3317         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3318         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3319         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3320         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3321 };
3322
3323 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3324         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3325                                             .attr_count = CTA_EXPECT_MAX,
3326                                             .policy = exp_nla_policy },
3327         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3328                                             .attr_count = CTA_EXPECT_MAX,
3329                                             .policy = exp_nla_policy },
3330         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3331                                             .attr_count = CTA_EXPECT_MAX,
3332                                             .policy = exp_nla_policy },
3333         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3334 };
3335
3336 static const struct nfnetlink_subsystem ctnl_subsys = {
3337         .name                           = "conntrack",
3338         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3339         .cb_count                       = IPCTNL_MSG_MAX,
3340         .cb                             = ctnl_cb,
3341 };
3342
3343 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3344         .name                           = "conntrack_expect",
3345         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3346         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3347         .cb                             = ctnl_exp_cb,
3348 };
3349
3350 MODULE_ALIAS("ip_conntrack_netlink");
3351 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3352 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3353
3354 static int __net_init ctnetlink_net_init(struct net *net)
3355 {
3356 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3357         int ret;
3358
3359         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3360         if (ret < 0) {
3361                 pr_err("ctnetlink_init: cannot register notifier.\n");
3362                 goto err_out;
3363         }
3364
3365         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3366         if (ret < 0) {
3367                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3368                 goto err_unreg_notifier;
3369         }
3370 #endif
3371         return 0;
3372
3373 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3374 err_unreg_notifier:
3375         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3376 err_out:
3377         return ret;
3378 #endif
3379 }
3380
3381 static void ctnetlink_net_exit(struct net *net)
3382 {
3383 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3384         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3385         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3386 #endif
3387 }
3388
3389 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3390 {
3391         struct net *net;
3392
3393         list_for_each_entry(net, net_exit_list, exit_list)
3394                 ctnetlink_net_exit(net);
3395
3396         /* wait for other cpus until they are done with ctnl_notifiers */
3397         synchronize_rcu();
3398 }
3399
3400 static struct pernet_operations ctnetlink_net_ops = {
3401         .init           = ctnetlink_net_init,
3402         .exit_batch     = ctnetlink_net_exit_batch,
3403 };
3404
3405 static int __init ctnetlink_init(void)
3406 {
3407         int ret;
3408
3409         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3410         ret = nfnetlink_subsys_register(&ctnl_subsys);
3411         if (ret < 0) {
3412                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3413                 goto err_out;
3414         }
3415
3416         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3417         if (ret < 0) {
3418                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3419                 goto err_unreg_subsys;
3420         }
3421
3422         ret = register_pernet_subsys(&ctnetlink_net_ops);
3423         if (ret < 0) {
3424                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3425                 goto err_unreg_exp_subsys;
3426         }
3427 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3428         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3429         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3430 #endif
3431         return 0;
3432
3433 err_unreg_exp_subsys:
3434         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3435 err_unreg_subsys:
3436         nfnetlink_subsys_unregister(&ctnl_subsys);
3437 err_out:
3438         return ret;
3439 }
3440
3441 static void __exit ctnetlink_exit(void)
3442 {
3443         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3444
3445         unregister_pernet_subsys(&ctnetlink_net_ops);
3446         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3447         nfnetlink_subsys_unregister(&ctnl_subsys);
3448 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3449         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3450 #endif
3451         synchronize_rcu();
3452 }
3453
3454 module_init(ctnetlink_init);
3455 module_exit(ctnetlink_exit);