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