net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #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                 spin_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
1117 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
1118                         const struct nlmsghdr *nlh,
1119                         const struct nlattr * const cda[])
1120 {
1121         struct net *net = sock_net(ctnl);
1122         struct nf_conntrack_tuple_hash *h;
1123         struct nf_conntrack_tuple tuple;
1124         struct nf_conn *ct;
1125         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1126         u_int8_t u3 = nfmsg->nfgen_family;
1127         struct nf_conntrack_zone zone;
1128         int err;
1129
1130         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1131         if (err < 0)
1132                 return err;
1133
1134         if (cda[CTA_TUPLE_ORIG])
1135                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1136                                             u3, &zone);
1137         else if (cda[CTA_TUPLE_REPLY])
1138                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1139                                             u3, &zone);
1140         else {
1141                 return ctnetlink_flush_conntrack(net, cda,
1142                                                  NETLINK_CB(skb).portid,
1143                                                  nlmsg_report(nlh));
1144         }
1145
1146         if (err < 0)
1147                 return err;
1148
1149         h = nf_conntrack_find_get(net, &zone, &tuple);
1150         if (!h)
1151                 return -ENOENT;
1152
1153         ct = nf_ct_tuplehash_to_ctrack(h);
1154
1155         if (cda[CTA_ID]) {
1156                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1157                 if (id != (u32)(unsigned long)ct) {
1158                         nf_ct_put(ct);
1159                         return -ENOENT;
1160                 }
1161         }
1162
1163         if (del_timer(&ct->timeout))
1164                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1165
1166         nf_ct_put(ct);
1167
1168         return 0;
1169 }
1170
1171 static int
1172 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1173                         const struct nlmsghdr *nlh,
1174                         const struct nlattr * const cda[])
1175 {
1176         struct net *net = sock_net(ctnl);
1177         struct nf_conntrack_tuple_hash *h;
1178         struct nf_conntrack_tuple tuple;
1179         struct nf_conn *ct;
1180         struct sk_buff *skb2 = NULL;
1181         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1182         u_int8_t u3 = nfmsg->nfgen_family;
1183         struct nf_conntrack_zone zone;
1184         int err;
1185
1186         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1187                 struct netlink_dump_control c = {
1188                         .dump = ctnetlink_dump_table,
1189                         .done = ctnetlink_done,
1190                 };
1191
1192                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1193                         struct ctnetlink_filter *filter;
1194
1195                         filter = ctnetlink_alloc_filter(cda);
1196                         if (IS_ERR(filter))
1197                                 return PTR_ERR(filter);
1198
1199                         c.data = filter;
1200                 }
1201                 return netlink_dump_start(ctnl, skb, nlh, &c);
1202         }
1203
1204         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1205         if (err < 0)
1206                 return err;
1207
1208         if (cda[CTA_TUPLE_ORIG])
1209                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1210                                             u3, &zone);
1211         else if (cda[CTA_TUPLE_REPLY])
1212                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1213                                             u3, &zone);
1214         else
1215                 return -EINVAL;
1216
1217         if (err < 0)
1218                 return err;
1219
1220         h = nf_conntrack_find_get(net, &zone, &tuple);
1221         if (!h)
1222                 return -ENOENT;
1223
1224         ct = nf_ct_tuplehash_to_ctrack(h);
1225
1226         err = -ENOMEM;
1227         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1228         if (skb2 == NULL) {
1229                 nf_ct_put(ct);
1230                 return -ENOMEM;
1231         }
1232
1233         rcu_read_lock();
1234         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1235                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1236         rcu_read_unlock();
1237         nf_ct_put(ct);
1238         if (err <= 0)
1239                 goto free;
1240
1241         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1242         if (err < 0)
1243                 goto out;
1244
1245         return 0;
1246
1247 free:
1248         kfree_skb(skb2);
1249 out:
1250         /* this avoids a loop in nfnetlink. */
1251         return err == -EAGAIN ? -ENOBUFS : err;
1252 }
1253
1254 static int ctnetlink_done_list(struct netlink_callback *cb)
1255 {
1256         if (cb->args[1])
1257                 nf_ct_put((struct nf_conn *)cb->args[1]);
1258         return 0;
1259 }
1260
1261 static int
1262 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1263 {
1264         struct nf_conn *ct, *last;
1265         struct nf_conntrack_tuple_hash *h;
1266         struct hlist_nulls_node *n;
1267         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1268         u_int8_t l3proto = nfmsg->nfgen_family;
1269         int res;
1270         int cpu;
1271         struct hlist_nulls_head *list;
1272         struct net *net = sock_net(skb->sk);
1273
1274         if (cb->args[2])
1275                 return 0;
1276
1277         last = (struct nf_conn *)cb->args[1];
1278
1279         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1280                 struct ct_pcpu *pcpu;
1281
1282                 if (!cpu_possible(cpu))
1283                         continue;
1284
1285                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1286                 spin_lock_bh(&pcpu->lock);
1287                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1288 restart:
1289                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1290                         ct = nf_ct_tuplehash_to_ctrack(h);
1291                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1292                                 continue;
1293                         if (cb->args[1]) {
1294                                 if (ct != last)
1295                                         continue;
1296                                 cb->args[1] = 0;
1297                         }
1298                         rcu_read_lock();
1299                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1300                                                   cb->nlh->nlmsg_seq,
1301                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1302                                                   ct);
1303                         rcu_read_unlock();
1304                         if (res < 0) {
1305                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1306                                         continue;
1307                                 cb->args[0] = cpu;
1308                                 cb->args[1] = (unsigned long)ct;
1309                                 spin_unlock_bh(&pcpu->lock);
1310                                 goto out;
1311                         }
1312                 }
1313                 if (cb->args[1]) {
1314                         cb->args[1] = 0;
1315                         goto restart;
1316                 }
1317                 spin_unlock_bh(&pcpu->lock);
1318         }
1319         cb->args[2] = 1;
1320 out:
1321         if (last)
1322                 nf_ct_put(last);
1323
1324         return skb->len;
1325 }
1326
1327 static int
1328 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1329 {
1330         return ctnetlink_dump_list(skb, cb, true);
1331 }
1332
1333 static int
1334 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1335                        const struct nlmsghdr *nlh,
1336                        const struct nlattr * const cda[])
1337 {
1338         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1339                 struct netlink_dump_control c = {
1340                         .dump = ctnetlink_dump_dying,
1341                         .done = ctnetlink_done_list,
1342                 };
1343                 return netlink_dump_start(ctnl, skb, nlh, &c);
1344         }
1345
1346         return -EOPNOTSUPP;
1347 }
1348
1349 static int
1350 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1351 {
1352         return ctnetlink_dump_list(skb, cb, false);
1353 }
1354
1355 static int
1356 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1357                              const struct nlmsghdr *nlh,
1358                              const struct nlattr * const cda[])
1359 {
1360         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1361                 struct netlink_dump_control c = {
1362                         .dump = ctnetlink_dump_unconfirmed,
1363                         .done = ctnetlink_done_list,
1364                 };
1365                 return netlink_dump_start(ctnl, skb, nlh, &c);
1366         }
1367
1368         return -EOPNOTSUPP;
1369 }
1370
1371 #ifdef CONFIG_NF_NAT_NEEDED
1372 static int
1373 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1374                           enum nf_nat_manip_type manip,
1375                           const struct nlattr *attr)
1376 {
1377         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1378         int err;
1379
1380         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1381         if (!parse_nat_setup) {
1382 #ifdef CONFIG_MODULES
1383                 rcu_read_unlock();
1384                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1385                 if (request_module("nf-nat") < 0) {
1386                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1387                         rcu_read_lock();
1388                         return -EOPNOTSUPP;
1389                 }
1390                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1391                 rcu_read_lock();
1392                 if (nfnetlink_parse_nat_setup_hook)
1393                         return -EAGAIN;
1394 #endif
1395                 return -EOPNOTSUPP;
1396         }
1397
1398         err = parse_nat_setup(ct, manip, attr);
1399         if (err == -EAGAIN) {
1400 #ifdef CONFIG_MODULES
1401                 rcu_read_unlock();
1402                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1403                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1404                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1405                         rcu_read_lock();
1406                         return -EOPNOTSUPP;
1407                 }
1408                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1409                 rcu_read_lock();
1410 #else
1411                 err = -EOPNOTSUPP;
1412 #endif
1413         }
1414         return err;
1415 }
1416 #endif
1417
1418 static int
1419 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1420 {
1421         unsigned long d;
1422         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1423         d = ct->status ^ status;
1424
1425         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1426                 /* unchangeable */
1427                 return -EBUSY;
1428
1429         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1430                 /* SEEN_REPLY bit can only be set */
1431                 return -EBUSY;
1432
1433         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1434                 /* ASSURED bit can only be set */
1435                 return -EBUSY;
1436
1437         /* Be careful here, modifying NAT bits can screw up things,
1438          * so don't let users modify them directly if they don't pass
1439          * nf_nat_range. */
1440         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1441         return 0;
1442 }
1443
1444 static int
1445 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1446 {
1447 #ifdef CONFIG_NF_NAT_NEEDED
1448         int ret;
1449
1450         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1451                 return 0;
1452
1453         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1454                                         cda[CTA_NAT_DST]);
1455         if (ret < 0)
1456                 return ret;
1457
1458         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1459                                         cda[CTA_NAT_SRC]);
1460         return ret;
1461 #else
1462         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1463                 return 0;
1464         return -EOPNOTSUPP;
1465 #endif
1466 }
1467
1468 static inline int
1469 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1470 {
1471         struct nf_conntrack_helper *helper;
1472         struct nf_conn_help *help = nfct_help(ct);
1473         char *helpname = NULL;
1474         struct nlattr *helpinfo = NULL;
1475         int err;
1476
1477         /* don't change helper of sibling connections */
1478         if (ct->master)
1479                 return -EBUSY;
1480
1481         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1482         if (err < 0)
1483                 return err;
1484
1485         if (!strcmp(helpname, "")) {
1486                 if (help && help->helper) {
1487                         /* we had a helper before ... */
1488                         nf_ct_remove_expectations(ct);
1489                         RCU_INIT_POINTER(help->helper, NULL);
1490                 }
1491
1492                 return 0;
1493         }
1494
1495         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1496                                             nf_ct_protonum(ct));
1497         if (helper == NULL) {
1498 #ifdef CONFIG_MODULES
1499                 spin_unlock_bh(&nf_conntrack_expect_lock);
1500
1501                 if (request_module("nfct-helper-%s", helpname) < 0) {
1502                         spin_lock_bh(&nf_conntrack_expect_lock);
1503                         return -EOPNOTSUPP;
1504                 }
1505
1506                 spin_lock_bh(&nf_conntrack_expect_lock);
1507                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1508                                                     nf_ct_protonum(ct));
1509                 if (helper)
1510                         return -EAGAIN;
1511 #endif
1512                 return -EOPNOTSUPP;
1513         }
1514
1515         if (help) {
1516                 if (help->helper == helper) {
1517                         /* update private helper data if allowed. */
1518                         if (helper->from_nlattr)
1519                                 helper->from_nlattr(helpinfo, ct);
1520                         return 0;
1521                 } else
1522                         return -EBUSY;
1523         }
1524
1525         /* we cannot set a helper for an existing conntrack */
1526         return -EOPNOTSUPP;
1527 }
1528
1529 static inline int
1530 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1531 {
1532         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1533
1534         if (!del_timer(&ct->timeout))
1535                 return -ETIME;
1536
1537         ct->timeout.expires = jiffies + timeout * HZ;
1538         add_timer(&ct->timeout);
1539
1540         return 0;
1541 }
1542
1543 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1544         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1545         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1546         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1547 };
1548
1549 static inline int
1550 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1551 {
1552         const struct nlattr *attr = cda[CTA_PROTOINFO];
1553         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1554         struct nf_conntrack_l4proto *l4proto;
1555         int err = 0;
1556
1557         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1558         if (err < 0)
1559                 return err;
1560
1561         rcu_read_lock();
1562         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1563         if (l4proto->from_nlattr)
1564                 err = l4proto->from_nlattr(tb, ct);
1565         rcu_read_unlock();
1566
1567         return err;
1568 }
1569
1570 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1571         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1572         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1573         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1574 };
1575
1576 static inline int
1577 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1578 {
1579         int err;
1580         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1581
1582         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1583         if (err < 0)
1584                 return err;
1585
1586         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1587                 return -EINVAL;
1588
1589         seq->correction_pos =
1590                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1591
1592         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1593                 return -EINVAL;
1594
1595         seq->offset_before =
1596                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1597
1598         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1599                 return -EINVAL;
1600
1601         seq->offset_after =
1602                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1603
1604         return 0;
1605 }
1606
1607 static int
1608 ctnetlink_change_seq_adj(struct nf_conn *ct,
1609                          const struct nlattr * const cda[])
1610 {
1611         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1612         int ret = 0;
1613
1614         if (!seqadj)
1615                 return 0;
1616
1617         if (cda[CTA_SEQ_ADJ_ORIG]) {
1618                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1619                                      cda[CTA_SEQ_ADJ_ORIG]);
1620                 if (ret < 0)
1621                         return ret;
1622
1623                 ct->status |= IPS_SEQ_ADJUST;
1624         }
1625
1626         if (cda[CTA_SEQ_ADJ_REPLY]) {
1627                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1628                                      cda[CTA_SEQ_ADJ_REPLY]);
1629                 if (ret < 0)
1630                         return ret;
1631
1632                 ct->status |= IPS_SEQ_ADJUST;
1633         }
1634
1635         return 0;
1636 }
1637
1638 static int
1639 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1640 {
1641 #ifdef CONFIG_NF_CONNTRACK_LABELS
1642         size_t len = nla_len(cda[CTA_LABELS]);
1643         const void *mask = cda[CTA_LABELS_MASK];
1644
1645         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1646                 return -EINVAL;
1647
1648         if (mask) {
1649                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1650                     nla_len(cda[CTA_LABELS_MASK]) != len)
1651                         return -EINVAL;
1652                 mask = nla_data(cda[CTA_LABELS_MASK]);
1653         }
1654
1655         len /= sizeof(u32);
1656
1657         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1658 #else
1659         return -EOPNOTSUPP;
1660 #endif
1661 }
1662
1663 static int
1664 ctnetlink_change_conntrack(struct nf_conn *ct,
1665                            const struct nlattr * const cda[])
1666 {
1667         int err;
1668
1669         /* only allow NAT changes and master assignation for new conntracks */
1670         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1671                 return -EOPNOTSUPP;
1672
1673         if (cda[CTA_HELP]) {
1674                 err = ctnetlink_change_helper(ct, cda);
1675                 if (err < 0)
1676                         return err;
1677         }
1678
1679         if (cda[CTA_TIMEOUT]) {
1680                 err = ctnetlink_change_timeout(ct, cda);
1681                 if (err < 0)
1682                         return err;
1683         }
1684
1685         if (cda[CTA_STATUS]) {
1686                 err = ctnetlink_change_status(ct, cda);
1687                 if (err < 0)
1688                         return err;
1689         }
1690
1691         if (cda[CTA_PROTOINFO]) {
1692                 err = ctnetlink_change_protoinfo(ct, cda);
1693                 if (err < 0)
1694                         return err;
1695         }
1696
1697 #if defined(CONFIG_NF_CONNTRACK_MARK)
1698         if (cda[CTA_MARK])
1699                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1700 #endif
1701
1702         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1703                 err = ctnetlink_change_seq_adj(ct, cda);
1704                 if (err < 0)
1705                         return err;
1706         }
1707
1708         if (cda[CTA_LABELS]) {
1709                 err = ctnetlink_attach_labels(ct, cda);
1710                 if (err < 0)
1711                         return err;
1712         }
1713
1714         return 0;
1715 }
1716
1717 static struct nf_conn *
1718 ctnetlink_create_conntrack(struct net *net,
1719                            const struct nf_conntrack_zone *zone,
1720                            const struct nlattr * const cda[],
1721                            struct nf_conntrack_tuple *otuple,
1722                            struct nf_conntrack_tuple *rtuple,
1723                            u8 u3)
1724 {
1725         struct nf_conn *ct;
1726         int err = -EINVAL;
1727         struct nf_conntrack_helper *helper;
1728         struct nf_conn_tstamp *tstamp;
1729
1730         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1731         if (IS_ERR(ct))
1732                 return ERR_PTR(-ENOMEM);
1733
1734         if (!cda[CTA_TIMEOUT])
1735                 goto err1;
1736         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1737
1738         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1739
1740         rcu_read_lock();
1741         if (cda[CTA_HELP]) {
1742                 char *helpname = NULL;
1743                 struct nlattr *helpinfo = NULL;
1744
1745                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1746                 if (err < 0)
1747                         goto err2;
1748
1749                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1750                                                     nf_ct_protonum(ct));
1751                 if (helper == NULL) {
1752                         rcu_read_unlock();
1753 #ifdef CONFIG_MODULES
1754                         if (request_module("nfct-helper-%s", helpname) < 0) {
1755                                 err = -EOPNOTSUPP;
1756                                 goto err1;
1757                         }
1758
1759                         rcu_read_lock();
1760                         helper = __nf_conntrack_helper_find(helpname,
1761                                                             nf_ct_l3num(ct),
1762                                                             nf_ct_protonum(ct));
1763                         if (helper) {
1764                                 err = -EAGAIN;
1765                                 goto err2;
1766                         }
1767                         rcu_read_unlock();
1768 #endif
1769                         err = -EOPNOTSUPP;
1770                         goto err1;
1771                 } else {
1772                         struct nf_conn_help *help;
1773
1774                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1775                         if (help == NULL) {
1776                                 err = -ENOMEM;
1777                                 goto err2;
1778                         }
1779                         /* set private helper data if allowed. */
1780                         if (helper->from_nlattr)
1781                                 helper->from_nlattr(helpinfo, ct);
1782
1783                         /* not in hash table yet so not strictly necessary */
1784                         RCU_INIT_POINTER(help->helper, helper);
1785                 }
1786         } else {
1787                 /* try an implicit helper assignation */
1788                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1789                 if (err < 0)
1790                         goto err2;
1791         }
1792
1793         err = ctnetlink_setup_nat(ct, cda);
1794         if (err < 0)
1795                 goto err2;
1796
1797         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1798         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1799         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1800         nf_ct_labels_ext_add(ct);
1801
1802         /* we must add conntrack extensions before confirmation. */
1803         ct->status |= IPS_CONFIRMED;
1804
1805         if (cda[CTA_STATUS]) {
1806                 err = ctnetlink_change_status(ct, cda);
1807                 if (err < 0)
1808                         goto err2;
1809         }
1810
1811         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1812                 err = ctnetlink_change_seq_adj(ct, cda);
1813                 if (err < 0)
1814                         goto err2;
1815         }
1816
1817         memset(&ct->proto, 0, sizeof(ct->proto));
1818         if (cda[CTA_PROTOINFO]) {
1819                 err = ctnetlink_change_protoinfo(ct, cda);
1820                 if (err < 0)
1821                         goto err2;
1822         }
1823
1824 #if defined(CONFIG_NF_CONNTRACK_MARK)
1825         if (cda[CTA_MARK])
1826                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1827 #endif
1828
1829         /* setup master conntrack: this is a confirmed expectation */
1830         if (cda[CTA_TUPLE_MASTER]) {
1831                 struct nf_conntrack_tuple master;
1832                 struct nf_conntrack_tuple_hash *master_h;
1833                 struct nf_conn *master_ct;
1834
1835                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1836                                             u3, NULL);
1837                 if (err < 0)
1838                         goto err2;
1839
1840                 master_h = nf_conntrack_find_get(net, zone, &master);
1841                 if (master_h == NULL) {
1842                         err = -ENOENT;
1843                         goto err2;
1844                 }
1845                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1846                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1847                 ct->master = master_ct;
1848         }
1849         tstamp = nf_conn_tstamp_find(ct);
1850         if (tstamp)
1851                 tstamp->start = ktime_get_real_ns();
1852
1853         err = nf_conntrack_hash_check_insert(ct);
1854         if (err < 0)
1855                 goto err2;
1856
1857         rcu_read_unlock();
1858
1859         return ct;
1860
1861 err2:
1862         rcu_read_unlock();
1863 err1:
1864         nf_conntrack_free(ct);
1865         return ERR_PTR(err);
1866 }
1867
1868 static int
1869 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1870                         const struct nlmsghdr *nlh,
1871                         const struct nlattr * const cda[])
1872 {
1873         struct net *net = sock_net(ctnl);
1874         struct nf_conntrack_tuple otuple, rtuple;
1875         struct nf_conntrack_tuple_hash *h = NULL;
1876         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1877         struct nf_conn *ct;
1878         u_int8_t u3 = nfmsg->nfgen_family;
1879         struct nf_conntrack_zone zone;
1880         int err;
1881
1882         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1883         if (err < 0)
1884                 return err;
1885
1886         if (cda[CTA_TUPLE_ORIG]) {
1887                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1888                                             u3, &zone);
1889                 if (err < 0)
1890                         return err;
1891         }
1892
1893         if (cda[CTA_TUPLE_REPLY]) {
1894                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1895                                             u3, &zone);
1896                 if (err < 0)
1897                         return err;
1898         }
1899
1900         if (cda[CTA_TUPLE_ORIG])
1901                 h = nf_conntrack_find_get(net, &zone, &otuple);
1902         else if (cda[CTA_TUPLE_REPLY])
1903                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1904
1905         if (h == NULL) {
1906                 err = -ENOENT;
1907                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1908                         enum ip_conntrack_events events;
1909
1910                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1911                                 return -EINVAL;
1912
1913                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1914                                                         &rtuple, u3);
1915                         if (IS_ERR(ct))
1916                                 return PTR_ERR(ct);
1917
1918                         err = 0;
1919                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1920                                 events = IPCT_RELATED;
1921                         else
1922                                 events = IPCT_NEW;
1923
1924                         if (cda[CTA_LABELS] &&
1925                             ctnetlink_attach_labels(ct, cda) == 0)
1926                                 events |= (1 << IPCT_LABEL);
1927
1928                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1929                                                       (1 << IPCT_ASSURED) |
1930                                                       (1 << IPCT_HELPER) |
1931                                                       (1 << IPCT_PROTOINFO) |
1932                                                       (1 << IPCT_SEQADJ) |
1933                                                       (1 << IPCT_MARK) | events,
1934                                                       ct, NETLINK_CB(skb).portid,
1935                                                       nlmsg_report(nlh));
1936                         nf_ct_put(ct);
1937                 }
1938
1939                 return err;
1940         }
1941         /* implicit 'else' */
1942
1943         err = -EEXIST;
1944         ct = nf_ct_tuplehash_to_ctrack(h);
1945         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1946                 spin_lock_bh(&nf_conntrack_expect_lock);
1947                 err = ctnetlink_change_conntrack(ct, cda);
1948                 spin_unlock_bh(&nf_conntrack_expect_lock);
1949                 if (err == 0) {
1950                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1951                                                       (1 << IPCT_ASSURED) |
1952                                                       (1 << IPCT_HELPER) |
1953                                                       (1 << IPCT_LABEL) |
1954                                                       (1 << IPCT_PROTOINFO) |
1955                                                       (1 << IPCT_SEQADJ) |
1956                                                       (1 << IPCT_MARK),
1957                                                       ct, NETLINK_CB(skb).portid,
1958                                                       nlmsg_report(nlh));
1959                 }
1960         }
1961
1962         nf_ct_put(ct);
1963         return err;
1964 }
1965
1966 static int
1967 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1968                                 __u16 cpu, const struct ip_conntrack_stat *st)
1969 {
1970         struct nlmsghdr *nlh;
1971         struct nfgenmsg *nfmsg;
1972         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1973
1974         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1975         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1976         if (nlh == NULL)
1977                 goto nlmsg_failure;
1978
1979         nfmsg = nlmsg_data(nlh);
1980         nfmsg->nfgen_family = AF_UNSPEC;
1981         nfmsg->version      = NFNETLINK_V0;
1982         nfmsg->res_id       = htons(cpu);
1983
1984         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1985             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1986             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1987             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1988             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1989             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1990             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1991             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1992             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1993                                 htonl(st->insert_failed)) ||
1994             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1995             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1996             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1997             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1998                                 htonl(st->search_restart)))
1999                 goto nla_put_failure;
2000
2001         nlmsg_end(skb, nlh);
2002         return skb->len;
2003
2004 nla_put_failure:
2005 nlmsg_failure:
2006         nlmsg_cancel(skb, nlh);
2007         return -1;
2008 }
2009
2010 static int
2011 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2012 {
2013         int cpu;
2014         struct net *net = sock_net(skb->sk);
2015
2016         if (cb->args[0] == nr_cpu_ids)
2017                 return 0;
2018
2019         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2020                 const struct ip_conntrack_stat *st;
2021
2022                 if (!cpu_possible(cpu))
2023                         continue;
2024
2025                 st = per_cpu_ptr(net->ct.stat, cpu);
2026                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2027                                                     NETLINK_CB(cb->skb).portid,
2028                                                     cb->nlh->nlmsg_seq,
2029                                                     cpu, st) < 0)
2030                                 break;
2031         }
2032         cb->args[0] = cpu;
2033
2034         return skb->len;
2035 }
2036
2037 static int
2038 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
2039                       const struct nlmsghdr *nlh,
2040                       const struct nlattr * const cda[])
2041 {
2042         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2043                 struct netlink_dump_control c = {
2044                         .dump = ctnetlink_ct_stat_cpu_dump,
2045                 };
2046                 return netlink_dump_start(ctnl, skb, nlh, &c);
2047         }
2048
2049         return 0;
2050 }
2051
2052 static int
2053 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2054                             struct net *net)
2055 {
2056         struct nlmsghdr *nlh;
2057         struct nfgenmsg *nfmsg;
2058         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2059         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2060
2061         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
2062         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2063         if (nlh == NULL)
2064                 goto nlmsg_failure;
2065
2066         nfmsg = nlmsg_data(nlh);
2067         nfmsg->nfgen_family = AF_UNSPEC;
2068         nfmsg->version      = NFNETLINK_V0;
2069         nfmsg->res_id       = 0;
2070
2071         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2072                 goto nla_put_failure;
2073
2074         nlmsg_end(skb, nlh);
2075         return skb->len;
2076
2077 nla_put_failure:
2078 nlmsg_failure:
2079         nlmsg_cancel(skb, nlh);
2080         return -1;
2081 }
2082
2083 static int
2084 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
2085                   const struct nlmsghdr *nlh,
2086                   const struct nlattr * const cda[])
2087 {
2088         struct sk_buff *skb2;
2089         int err;
2090
2091         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2092         if (skb2 == NULL)
2093                 return -ENOMEM;
2094
2095         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2096                                           nlh->nlmsg_seq,
2097                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2098                                           sock_net(skb->sk));
2099         if (err <= 0)
2100                 goto free;
2101
2102         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2103         if (err < 0)
2104                 goto out;
2105
2106         return 0;
2107
2108 free:
2109         kfree_skb(skb2);
2110 out:
2111         /* this avoids a loop in nfnetlink. */
2112         return err == -EAGAIN ? -ENOBUFS : err;
2113 }
2114
2115 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2116         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2117         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2118         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2119         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2120         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2121         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2122                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2123         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2124         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2125         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2126         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2127         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2128 };
2129
2130 static struct nf_conntrack_expect *
2131 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2132                        struct nf_conntrack_helper *helper,
2133                        struct nf_conntrack_tuple *tuple,
2134                        struct nf_conntrack_tuple *mask);
2135
2136 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2137 static size_t
2138 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2139 {
2140         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2141                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2142                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2143                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2144                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2145                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2146                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2147                + nla_total_size(0) /* CTA_PROTOINFO */
2148                + nla_total_size(0) /* CTA_HELP */
2149                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2150                + ctnetlink_secctx_size(ct)
2151 #ifdef CONFIG_NF_NAT_NEEDED
2152                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2153                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2154 #endif
2155 #ifdef CONFIG_NF_CONNTRACK_MARK
2156                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2157 #endif
2158 #ifdef CONFIG_NF_CONNTRACK_ZONES
2159                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2160 #endif
2161                + ctnetlink_proto_size(ct)
2162                ;
2163 }
2164
2165 static int
2166 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2167 {
2168         const struct nf_conntrack_zone *zone;
2169         struct nlattr *nest_parms;
2170
2171         rcu_read_lock();
2172         zone = nf_ct_zone(ct);
2173
2174         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2175         if (!nest_parms)
2176                 goto nla_put_failure;
2177         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2178                 goto nla_put_failure;
2179         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2180                                    NF_CT_ZONE_DIR_ORIG) < 0)
2181                 goto nla_put_failure;
2182         nla_nest_end(skb, nest_parms);
2183
2184         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2185         if (!nest_parms)
2186                 goto nla_put_failure;
2187         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2188                 goto nla_put_failure;
2189         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2190                                    NF_CT_ZONE_DIR_REPL) < 0)
2191                 goto nla_put_failure;
2192         nla_nest_end(skb, nest_parms);
2193
2194         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2195                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2196                 goto nla_put_failure;
2197
2198         if (ctnetlink_dump_id(skb, ct) < 0)
2199                 goto nla_put_failure;
2200
2201         if (ctnetlink_dump_status(skb, ct) < 0)
2202                 goto nla_put_failure;
2203
2204         if (ctnetlink_dump_timeout(skb, ct) < 0)
2205                 goto nla_put_failure;
2206
2207         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2208                 goto nla_put_failure;
2209
2210         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2211                 goto nla_put_failure;
2212
2213 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2214         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2215                 goto nla_put_failure;
2216 #endif
2217         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2218                 goto nla_put_failure;
2219
2220         if ((ct->status & IPS_SEQ_ADJUST) &&
2221             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2222                 goto nla_put_failure;
2223
2224 #ifdef CONFIG_NF_CONNTRACK_MARK
2225         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2226                 goto nla_put_failure;
2227 #endif
2228         if (ctnetlink_dump_labels(skb, ct) < 0)
2229                 goto nla_put_failure;
2230         rcu_read_unlock();
2231         return 0;
2232
2233 nla_put_failure:
2234         rcu_read_unlock();
2235         return -ENOSPC;
2236 }
2237
2238 static int
2239 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2240 {
2241         int err;
2242
2243         if (cda[CTA_TIMEOUT]) {
2244                 err = ctnetlink_change_timeout(ct, cda);
2245                 if (err < 0)
2246                         return err;
2247         }
2248         if (cda[CTA_STATUS]) {
2249                 err = ctnetlink_change_status(ct, cda);
2250                 if (err < 0)
2251                         return err;
2252         }
2253         if (cda[CTA_HELP]) {
2254                 err = ctnetlink_change_helper(ct, cda);
2255                 if (err < 0)
2256                         return err;
2257         }
2258         if (cda[CTA_LABELS]) {
2259                 err = ctnetlink_attach_labels(ct, cda);
2260                 if (err < 0)
2261                         return err;
2262         }
2263 #if defined(CONFIG_NF_CONNTRACK_MARK)
2264         if (cda[CTA_MARK]) {
2265                 u32 mask = 0, mark, newmark;
2266                 if (cda[CTA_MARK_MASK])
2267                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2268
2269                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2270                 newmark = (ct->mark & mask) ^ mark;
2271                 if (newmark != ct->mark)
2272                         ct->mark = newmark;
2273         }
2274 #endif
2275         return 0;
2276 }
2277
2278 static int
2279 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2280 {
2281         struct nlattr *cda[CTA_MAX+1];
2282         int ret;
2283
2284         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2285         if (ret < 0)
2286                 return ret;
2287
2288         spin_lock_bh(&nf_conntrack_expect_lock);
2289         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2290         spin_unlock_bh(&nf_conntrack_expect_lock);
2291
2292         return ret;
2293 }
2294
2295 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2296                                        const struct nf_conn *ct,
2297                                        struct nf_conntrack_tuple *tuple,
2298                                        struct nf_conntrack_tuple *mask)
2299 {
2300         int err;
2301
2302         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2303                                     nf_ct_l3num(ct), NULL);
2304         if (err < 0)
2305                 return err;
2306
2307         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2308                                      nf_ct_l3num(ct), NULL);
2309 }
2310
2311 static int
2312 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2313                                 u32 portid, u32 report)
2314 {
2315         struct nlattr *cda[CTA_EXPECT_MAX+1];
2316         struct nf_conntrack_tuple tuple, mask;
2317         struct nf_conntrack_helper *helper = NULL;
2318         struct nf_conntrack_expect *exp;
2319         int err;
2320
2321         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2322         if (err < 0)
2323                 return err;
2324
2325         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2326                                           ct, &tuple, &mask);
2327         if (err < 0)
2328                 return err;
2329
2330         if (cda[CTA_EXPECT_HELP_NAME]) {
2331                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2332
2333                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2334                                                     nf_ct_protonum(ct));
2335                 if (helper == NULL)
2336                         return -EOPNOTSUPP;
2337         }
2338
2339         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2340                                      helper, &tuple, &mask);
2341         if (IS_ERR(exp))
2342                 return PTR_ERR(exp);
2343
2344         err = nf_ct_expect_related_report(exp, portid, report);
2345         if (err < 0) {
2346                 nf_ct_expect_put(exp);
2347                 return err;
2348         }
2349
2350         return 0;
2351 }
2352
2353 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2354         .build_size     = ctnetlink_nfqueue_build_size,
2355         .build          = ctnetlink_nfqueue_build,
2356         .parse          = ctnetlink_nfqueue_parse,
2357         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2358         .seq_adjust     = nf_ct_tcp_seqadj_set,
2359 };
2360 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2361
2362 /***********************************************************************
2363  * EXPECT
2364  ***********************************************************************/
2365
2366 static inline int
2367 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2368                          const struct nf_conntrack_tuple *tuple,
2369                          enum ctattr_expect type)
2370 {
2371         struct nlattr *nest_parms;
2372
2373         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2374         if (!nest_parms)
2375                 goto nla_put_failure;
2376         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2377                 goto nla_put_failure;
2378         nla_nest_end(skb, nest_parms);
2379
2380         return 0;
2381
2382 nla_put_failure:
2383         return -1;
2384 }
2385
2386 static inline int
2387 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2388                         const struct nf_conntrack_tuple *tuple,
2389                         const struct nf_conntrack_tuple_mask *mask)
2390 {
2391         int ret;
2392         struct nf_conntrack_l3proto *l3proto;
2393         struct nf_conntrack_l4proto *l4proto;
2394         struct nf_conntrack_tuple m;
2395         struct nlattr *nest_parms;
2396
2397         memset(&m, 0xFF, sizeof(m));
2398         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2399         m.src.u.all = mask->src.u.all;
2400         m.dst.protonum = tuple->dst.protonum;
2401
2402         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2403         if (!nest_parms)
2404                 goto nla_put_failure;
2405
2406         rcu_read_lock();
2407         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2408         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2409         if (ret >= 0) {
2410                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2411                                                tuple->dst.protonum);
2412         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2413         }
2414         rcu_read_unlock();
2415
2416         if (unlikely(ret < 0))
2417                 goto nla_put_failure;
2418
2419         nla_nest_end(skb, nest_parms);
2420
2421         return 0;
2422
2423 nla_put_failure:
2424         return -1;
2425 }
2426
2427 static const union nf_inet_addr any_addr;
2428
2429 static int
2430 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2431                           const struct nf_conntrack_expect *exp)
2432 {
2433         struct nf_conn *master = exp->master;
2434         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2435         struct nf_conn_help *help;
2436 #ifdef CONFIG_NF_NAT_NEEDED
2437         struct nlattr *nest_parms;
2438         struct nf_conntrack_tuple nat_tuple = {};
2439 #endif
2440         struct nf_ct_helper_expectfn *expfn;
2441
2442         if (timeout < 0)
2443                 timeout = 0;
2444
2445         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2446                 goto nla_put_failure;
2447         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2448                 goto nla_put_failure;
2449         if (ctnetlink_exp_dump_tuple(skb,
2450                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2451                                  CTA_EXPECT_MASTER) < 0)
2452                 goto nla_put_failure;
2453
2454 #ifdef CONFIG_NF_NAT_NEEDED
2455         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2456             exp->saved_proto.all) {
2457                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2458                 if (!nest_parms)
2459                         goto nla_put_failure;
2460
2461                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2462                         goto nla_put_failure;
2463
2464                 nat_tuple.src.l3num = nf_ct_l3num(master);
2465                 nat_tuple.src.u3 = exp->saved_addr;
2466                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2467                 nat_tuple.src.u = exp->saved_proto;
2468
2469                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2470                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2471                         goto nla_put_failure;
2472                 nla_nest_end(skb, nest_parms);
2473         }
2474 #endif
2475         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2476             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2477             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2478             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2479                 goto nla_put_failure;
2480         help = nfct_help(master);
2481         if (help) {
2482                 struct nf_conntrack_helper *helper;
2483
2484                 helper = rcu_dereference(help->helper);
2485                 if (helper &&
2486                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2487                         goto nla_put_failure;
2488         }
2489         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2490         if (expfn != NULL &&
2491             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2492                 goto nla_put_failure;
2493
2494         return 0;
2495
2496 nla_put_failure:
2497         return -1;
2498 }
2499
2500 static int
2501 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2502                         int event, const struct nf_conntrack_expect *exp)
2503 {
2504         struct nlmsghdr *nlh;
2505         struct nfgenmsg *nfmsg;
2506         unsigned int flags = portid ? NLM_F_MULTI : 0;
2507
2508         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2509         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2510         if (nlh == NULL)
2511                 goto nlmsg_failure;
2512
2513         nfmsg = nlmsg_data(nlh);
2514         nfmsg->nfgen_family = exp->tuple.src.l3num;
2515         nfmsg->version      = NFNETLINK_V0;
2516         nfmsg->res_id       = 0;
2517
2518         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2519                 goto nla_put_failure;
2520
2521         nlmsg_end(skb, nlh);
2522         return skb->len;
2523
2524 nlmsg_failure:
2525 nla_put_failure:
2526         nlmsg_cancel(skb, nlh);
2527         return -1;
2528 }
2529
2530 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2531 static int
2532 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2533 {
2534         struct nf_conntrack_expect *exp = item->exp;
2535         struct net *net = nf_ct_exp_net(exp);
2536         struct nlmsghdr *nlh;
2537         struct nfgenmsg *nfmsg;
2538         struct sk_buff *skb;
2539         unsigned int type, group;
2540         int flags = 0;
2541
2542         if (events & (1 << IPEXP_DESTROY)) {
2543                 type = IPCTNL_MSG_EXP_DELETE;
2544                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2545         } else if (events & (1 << IPEXP_NEW)) {
2546                 type = IPCTNL_MSG_EXP_NEW;
2547                 flags = NLM_F_CREATE|NLM_F_EXCL;
2548                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2549         } else
2550                 return 0;
2551
2552         if (!item->report && !nfnetlink_has_listeners(net, group))
2553                 return 0;
2554
2555         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2556         if (skb == NULL)
2557                 goto errout;
2558
2559         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2560         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2561         if (nlh == NULL)
2562                 goto nlmsg_failure;
2563
2564         nfmsg = nlmsg_data(nlh);
2565         nfmsg->nfgen_family = exp->tuple.src.l3num;
2566         nfmsg->version      = NFNETLINK_V0;
2567         nfmsg->res_id       = 0;
2568
2569         rcu_read_lock();
2570         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2571                 goto nla_put_failure;
2572         rcu_read_unlock();
2573
2574         nlmsg_end(skb, nlh);
2575         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2576         return 0;
2577
2578 nla_put_failure:
2579         rcu_read_unlock();
2580         nlmsg_cancel(skb, nlh);
2581 nlmsg_failure:
2582         kfree_skb(skb);
2583 errout:
2584         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2585         return 0;
2586 }
2587 #endif
2588 static int ctnetlink_exp_done(struct netlink_callback *cb)
2589 {
2590         if (cb->args[1])
2591                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2592         return 0;
2593 }
2594
2595 static int
2596 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2597 {
2598         struct net *net = sock_net(skb->sk);
2599         struct nf_conntrack_expect *exp, *last;
2600         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2601         u_int8_t l3proto = nfmsg->nfgen_family;
2602
2603         rcu_read_lock();
2604         last = (struct nf_conntrack_expect *)cb->args[1];
2605         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2606 restart:
2607                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2608                                      hnode) {
2609                         if (l3proto && exp->tuple.src.l3num != l3proto)
2610                                 continue;
2611                         if (cb->args[1]) {
2612                                 if (exp != last)
2613                                         continue;
2614                                 cb->args[1] = 0;
2615                         }
2616                         if (ctnetlink_exp_fill_info(skb,
2617                                                     NETLINK_CB(cb->skb).portid,
2618                                                     cb->nlh->nlmsg_seq,
2619                                                     IPCTNL_MSG_EXP_NEW,
2620                                                     exp) < 0) {
2621                                 if (!atomic_inc_not_zero(&exp->use))
2622                                         continue;
2623                                 cb->args[1] = (unsigned long)exp;
2624                                 goto out;
2625                         }
2626                 }
2627                 if (cb->args[1]) {
2628                         cb->args[1] = 0;
2629                         goto restart;
2630                 }
2631         }
2632 out:
2633         rcu_read_unlock();
2634         if (last)
2635                 nf_ct_expect_put(last);
2636
2637         return skb->len;
2638 }
2639
2640 static int
2641 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2642 {
2643         struct nf_conntrack_expect *exp, *last;
2644         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2645         struct nf_conn *ct = cb->data;
2646         struct nf_conn_help *help = nfct_help(ct);
2647         u_int8_t l3proto = nfmsg->nfgen_family;
2648
2649         if (cb->args[0])
2650                 return 0;
2651
2652         rcu_read_lock();
2653         last = (struct nf_conntrack_expect *)cb->args[1];
2654 restart:
2655         hlist_for_each_entry(exp, &help->expectations, lnode) {
2656                 if (l3proto && exp->tuple.src.l3num != l3proto)
2657                         continue;
2658                 if (cb->args[1]) {
2659                         if (exp != last)
2660                                 continue;
2661                         cb->args[1] = 0;
2662                 }
2663                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2664                                             cb->nlh->nlmsg_seq,
2665                                             IPCTNL_MSG_EXP_NEW,
2666                                             exp) < 0) {
2667                         if (!atomic_inc_not_zero(&exp->use))
2668                                 continue;
2669                         cb->args[1] = (unsigned long)exp;
2670                         goto out;
2671                 }
2672         }
2673         if (cb->args[1]) {
2674                 cb->args[1] = 0;
2675                 goto restart;
2676         }
2677         cb->args[0] = 1;
2678 out:
2679         rcu_read_unlock();
2680         if (last)
2681                 nf_ct_expect_put(last);
2682
2683         return skb->len;
2684 }
2685
2686 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2687                                  const struct nlmsghdr *nlh,
2688                                  const struct nlattr * const cda[])
2689 {
2690         int err;
2691         struct net *net = sock_net(ctnl);
2692         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2693         u_int8_t u3 = nfmsg->nfgen_family;
2694         struct nf_conntrack_tuple tuple;
2695         struct nf_conntrack_tuple_hash *h;
2696         struct nf_conn *ct;
2697         struct nf_conntrack_zone zone;
2698         struct netlink_dump_control c = {
2699                 .dump = ctnetlink_exp_ct_dump_table,
2700                 .done = ctnetlink_exp_done,
2701         };
2702
2703         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2704                                     u3, NULL);
2705         if (err < 0)
2706                 return err;
2707
2708         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2709         if (err < 0)
2710                 return err;
2711
2712         h = nf_conntrack_find_get(net, &zone, &tuple);
2713         if (!h)
2714                 return -ENOENT;
2715
2716         ct = nf_ct_tuplehash_to_ctrack(h);
2717         c.data = ct;
2718
2719         err = netlink_dump_start(ctnl, skb, nlh, &c);
2720         nf_ct_put(ct);
2721
2722         return err;
2723 }
2724
2725 static int
2726 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2727                      const struct nlmsghdr *nlh,
2728                      const struct nlattr * const cda[])
2729 {
2730         struct net *net = sock_net(ctnl);
2731         struct nf_conntrack_tuple tuple;
2732         struct nf_conntrack_expect *exp;
2733         struct sk_buff *skb2;
2734         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2735         u_int8_t u3 = nfmsg->nfgen_family;
2736         struct nf_conntrack_zone zone;
2737         int err;
2738
2739         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2740                 if (cda[CTA_EXPECT_MASTER])
2741                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2742                 else {
2743                         struct netlink_dump_control c = {
2744                                 .dump = ctnetlink_exp_dump_table,
2745                                 .done = ctnetlink_exp_done,
2746                         };
2747                         return netlink_dump_start(ctnl, skb, nlh, &c);
2748                 }
2749         }
2750
2751         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2752         if (err < 0)
2753                 return err;
2754
2755         if (cda[CTA_EXPECT_TUPLE])
2756                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2757                                             u3, NULL);
2758         else if (cda[CTA_EXPECT_MASTER])
2759                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2760                                             u3, NULL);
2761         else
2762                 return -EINVAL;
2763
2764         if (err < 0)
2765                 return err;
2766
2767         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2768         if (!exp)
2769                 return -ENOENT;
2770
2771         if (cda[CTA_EXPECT_ID]) {
2772                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2773                 if (ntohl(id) != (u32)(unsigned long)exp) {
2774                         nf_ct_expect_put(exp);
2775                         return -ENOENT;
2776                 }
2777         }
2778
2779         err = -ENOMEM;
2780         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2781         if (skb2 == NULL) {
2782                 nf_ct_expect_put(exp);
2783                 goto out;
2784         }
2785
2786         rcu_read_lock();
2787         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2788                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2789         rcu_read_unlock();
2790         nf_ct_expect_put(exp);
2791         if (err <= 0)
2792                 goto free;
2793
2794         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2795         if (err < 0)
2796                 goto out;
2797
2798         return 0;
2799
2800 free:
2801         kfree_skb(skb2);
2802 out:
2803         /* this avoids a loop in nfnetlink. */
2804         return err == -EAGAIN ? -ENOBUFS : err;
2805 }
2806
2807 static int
2808 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2809                      const struct nlmsghdr *nlh,
2810                      const struct nlattr * const cda[])
2811 {
2812         struct net *net = sock_net(ctnl);
2813         struct nf_conntrack_expect *exp;
2814         struct nf_conntrack_tuple tuple;
2815         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2816         struct hlist_node *next;
2817         u_int8_t u3 = nfmsg->nfgen_family;
2818         struct nf_conntrack_zone zone;
2819         unsigned int i;
2820         int err;
2821
2822         if (cda[CTA_EXPECT_TUPLE]) {
2823                 /* delete a single expect by tuple */
2824                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2825                 if (err < 0)
2826                         return err;
2827
2828                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2829                                             u3, NULL);
2830                 if (err < 0)
2831                         return err;
2832
2833                 /* bump usage count to 2 */
2834                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2835                 if (!exp)
2836                         return -ENOENT;
2837
2838                 if (cda[CTA_EXPECT_ID]) {
2839                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2840                         if (ntohl(id) != (u32)(unsigned long)exp) {
2841                                 nf_ct_expect_put(exp);
2842                                 return -ENOENT;
2843                         }
2844                 }
2845
2846                 /* after list removal, usage count == 1 */
2847                 spin_lock_bh(&nf_conntrack_expect_lock);
2848                 if (del_timer(&exp->timeout)) {
2849                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2850                                                    nlmsg_report(nlh));
2851                         nf_ct_expect_put(exp);
2852                 }
2853                 spin_unlock_bh(&nf_conntrack_expect_lock);
2854                 /* have to put what we 'get' above.
2855                  * after this line usage count == 0 */
2856                 nf_ct_expect_put(exp);
2857         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2858                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2859                 struct nf_conn_help *m_help;
2860
2861                 /* delete all expectations for this helper */
2862                 spin_lock_bh(&nf_conntrack_expect_lock);
2863                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2864                         hlist_for_each_entry_safe(exp, next,
2865                                                   &net->ct.expect_hash[i],
2866                                                   hnode) {
2867                                 m_help = nfct_help(exp->master);
2868                                 if (!strcmp(m_help->helper->name, name) &&
2869                                     del_timer(&exp->timeout)) {
2870                                         nf_ct_unlink_expect_report(exp,
2871                                                         NETLINK_CB(skb).portid,
2872                                                         nlmsg_report(nlh));
2873                                         nf_ct_expect_put(exp);
2874                                 }
2875                         }
2876                 }
2877                 spin_unlock_bh(&nf_conntrack_expect_lock);
2878         } else {
2879                 /* This basically means we have to flush everything*/
2880                 spin_lock_bh(&nf_conntrack_expect_lock);
2881                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2882                         hlist_for_each_entry_safe(exp, next,
2883                                                   &net->ct.expect_hash[i],
2884                                                   hnode) {
2885                                 if (del_timer(&exp->timeout)) {
2886                                         nf_ct_unlink_expect_report(exp,
2887                                                         NETLINK_CB(skb).portid,
2888                                                         nlmsg_report(nlh));
2889                                         nf_ct_expect_put(exp);
2890                                 }
2891                         }
2892                 }
2893                 spin_unlock_bh(&nf_conntrack_expect_lock);
2894         }
2895
2896         return 0;
2897 }
2898 static int
2899 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2900                         const struct nlattr * const cda[])
2901 {
2902         if (cda[CTA_EXPECT_TIMEOUT]) {
2903                 if (!del_timer(&x->timeout))
2904                         return -ETIME;
2905
2906                 x->timeout.expires = jiffies +
2907                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2908                 add_timer(&x->timeout);
2909         }
2910         return 0;
2911 }
2912
2913 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2914         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2915         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2916 };
2917
2918 static int
2919 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2920                            struct nf_conntrack_expect *exp,
2921                            u_int8_t u3)
2922 {
2923 #ifdef CONFIG_NF_NAT_NEEDED
2924         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2925         struct nf_conntrack_tuple nat_tuple = {};
2926         int err;
2927
2928         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2929         if (err < 0)
2930                 return err;
2931
2932         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2933                 return -EINVAL;
2934
2935         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2936                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
2937                                     u3, NULL);
2938         if (err < 0)
2939                 return err;
2940
2941         exp->saved_addr = nat_tuple.src.u3;
2942         exp->saved_proto = nat_tuple.src.u;
2943         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2944
2945         return 0;
2946 #else
2947         return -EOPNOTSUPP;
2948 #endif
2949 }
2950
2951 static struct nf_conntrack_expect *
2952 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2953                        struct nf_conntrack_helper *helper,
2954                        struct nf_conntrack_tuple *tuple,
2955                        struct nf_conntrack_tuple *mask)
2956 {
2957         u_int32_t class = 0;
2958         struct nf_conntrack_expect *exp;
2959         struct nf_conn_help *help;
2960         int err;
2961
2962         if (cda[CTA_EXPECT_CLASS] && helper) {
2963                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2964                 if (class > helper->expect_class_max)
2965                         return ERR_PTR(-EINVAL);
2966         }
2967         exp = nf_ct_expect_alloc(ct);
2968         if (!exp)
2969                 return ERR_PTR(-ENOMEM);
2970
2971         help = nfct_help(ct);
2972         if (!help) {
2973                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2974                         err = -EINVAL;
2975                         goto err_out;
2976                 }
2977                 exp->timeout.expires =
2978                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2979
2980                 exp->flags = NF_CT_EXPECT_USERSPACE;
2981                 if (cda[CTA_EXPECT_FLAGS]) {
2982                         exp->flags |=
2983                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2984                 }
2985         } else {
2986                 if (cda[CTA_EXPECT_FLAGS]) {
2987                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2988                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2989                 } else
2990                         exp->flags = 0;
2991         }
2992         if (cda[CTA_EXPECT_FN]) {
2993                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2994                 struct nf_ct_helper_expectfn *expfn;
2995
2996                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2997                 if (expfn == NULL) {
2998                         err = -EINVAL;
2999                         goto err_out;
3000                 }
3001                 exp->expectfn = expfn->expectfn;
3002         } else
3003                 exp->expectfn = NULL;
3004
3005         exp->class = class;
3006         exp->master = ct;
3007         exp->helper = helper;
3008         exp->tuple = *tuple;
3009         exp->mask.src.u3 = mask->src.u3;
3010         exp->mask.src.u.all = mask->src.u.all;
3011
3012         if (cda[CTA_EXPECT_NAT]) {
3013                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3014                                                  exp, nf_ct_l3num(ct));
3015                 if (err < 0)
3016                         goto err_out;
3017         }
3018         return exp;
3019 err_out:
3020         nf_ct_expect_put(exp);
3021         return ERR_PTR(err);
3022 }
3023
3024 static int
3025 ctnetlink_create_expect(struct net *net,
3026                         const struct nf_conntrack_zone *zone,
3027                         const struct nlattr * const cda[],
3028                         u_int8_t u3, u32 portid, int report)
3029 {
3030         struct nf_conntrack_tuple tuple, mask, master_tuple;
3031         struct nf_conntrack_tuple_hash *h = NULL;
3032         struct nf_conntrack_helper *helper = NULL;
3033         struct nf_conntrack_expect *exp;
3034         struct nf_conn *ct;
3035         int err;
3036
3037         /* caller guarantees that those three CTA_EXPECT_* exist */
3038         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3039                                     u3, NULL);
3040         if (err < 0)
3041                 return err;
3042         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3043                                     u3, NULL);
3044         if (err < 0)
3045                 return err;
3046         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3047                                     u3, NULL);
3048         if (err < 0)
3049                 return err;
3050
3051         /* Look for master conntrack of this expectation */
3052         h = nf_conntrack_find_get(net, zone, &master_tuple);
3053         if (!h)
3054                 return -ENOENT;
3055         ct = nf_ct_tuplehash_to_ctrack(h);
3056
3057         if (cda[CTA_EXPECT_HELP_NAME]) {
3058                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3059
3060                 helper = __nf_conntrack_helper_find(helpname, u3,
3061                                                     nf_ct_protonum(ct));
3062                 if (helper == NULL) {
3063 #ifdef CONFIG_MODULES
3064                         if (request_module("nfct-helper-%s", helpname) < 0) {
3065                                 err = -EOPNOTSUPP;
3066                                 goto err_ct;
3067                         }
3068                         helper = __nf_conntrack_helper_find(helpname, u3,
3069                                                             nf_ct_protonum(ct));
3070                         if (helper) {
3071                                 err = -EAGAIN;
3072                                 goto err_ct;
3073                         }
3074 #endif
3075                         err = -EOPNOTSUPP;
3076                         goto err_ct;
3077                 }
3078         }
3079
3080         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3081         if (IS_ERR(exp)) {
3082                 err = PTR_ERR(exp);
3083                 goto err_ct;
3084         }
3085
3086         err = nf_ct_expect_related_report(exp, portid, report);
3087         nf_ct_expect_put(exp);
3088 err_ct:
3089         nf_ct_put(ct);
3090         return err;
3091 }
3092
3093 static int
3094 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
3095                      const struct nlmsghdr *nlh,
3096                      const struct nlattr * const cda[])
3097 {
3098         struct net *net = sock_net(ctnl);
3099         struct nf_conntrack_tuple tuple;
3100         struct nf_conntrack_expect *exp;
3101         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3102         u_int8_t u3 = nfmsg->nfgen_family;
3103         struct nf_conntrack_zone zone;
3104         int err;
3105
3106         if (!cda[CTA_EXPECT_TUPLE]
3107             || !cda[CTA_EXPECT_MASK]
3108             || !cda[CTA_EXPECT_MASTER])
3109                 return -EINVAL;
3110
3111         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3112         if (err < 0)
3113                 return err;
3114
3115         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3116                                     u3, NULL);
3117         if (err < 0)
3118                 return err;
3119
3120         spin_lock_bh(&nf_conntrack_expect_lock);
3121         exp = __nf_ct_expect_find(net, &zone, &tuple);
3122         if (!exp) {
3123                 spin_unlock_bh(&nf_conntrack_expect_lock);
3124                 err = -ENOENT;
3125                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3126                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3127                                                       NETLINK_CB(skb).portid,
3128                                                       nlmsg_report(nlh));
3129                 }
3130                 return err;
3131         }
3132
3133         err = -EEXIST;
3134         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3135                 err = ctnetlink_change_expect(exp, cda);
3136         spin_unlock_bh(&nf_conntrack_expect_lock);
3137
3138         return err;
3139 }
3140
3141 static int
3142 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3143                              const struct ip_conntrack_stat *st)
3144 {
3145         struct nlmsghdr *nlh;
3146         struct nfgenmsg *nfmsg;
3147         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3148
3149         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3150         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3151         if (nlh == NULL)
3152                 goto nlmsg_failure;
3153
3154         nfmsg = nlmsg_data(nlh);
3155         nfmsg->nfgen_family = AF_UNSPEC;
3156         nfmsg->version      = NFNETLINK_V0;
3157         nfmsg->res_id       = htons(cpu);
3158
3159         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3160             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3161             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3162                 goto nla_put_failure;
3163
3164         nlmsg_end(skb, nlh);
3165         return skb->len;
3166
3167 nla_put_failure:
3168 nlmsg_failure:
3169         nlmsg_cancel(skb, nlh);
3170         return -1;
3171 }
3172
3173 static int
3174 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3175 {
3176         int cpu;
3177         struct net *net = sock_net(skb->sk);
3178
3179         if (cb->args[0] == nr_cpu_ids)
3180                 return 0;
3181
3182         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3183                 const struct ip_conntrack_stat *st;
3184
3185                 if (!cpu_possible(cpu))
3186                         continue;
3187
3188                 st = per_cpu_ptr(net->ct.stat, cpu);
3189                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3190                                                  cb->nlh->nlmsg_seq,
3191                                                  cpu, st) < 0)
3192                         break;
3193         }
3194         cb->args[0] = cpu;
3195
3196         return skb->len;
3197 }
3198
3199 static int
3200 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3201                        const struct nlmsghdr *nlh,
3202                        const struct nlattr * const cda[])
3203 {
3204         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3205                 struct netlink_dump_control c = {
3206                         .dump = ctnetlink_exp_stat_cpu_dump,
3207                 };
3208                 return netlink_dump_start(ctnl, skb, nlh, &c);
3209         }
3210
3211         return 0;
3212 }
3213
3214 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3215 static struct nf_ct_event_notifier ctnl_notifier = {
3216         .fcn = ctnetlink_conntrack_event,
3217 };
3218
3219 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3220         .fcn = ctnetlink_expect_event,
3221 };
3222 #endif
3223
3224 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3225         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3226                                             .attr_count = CTA_MAX,
3227                                             .policy = ct_nla_policy },
3228         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3229                                             .attr_count = CTA_MAX,
3230                                             .policy = ct_nla_policy },
3231         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3232                                             .attr_count = CTA_MAX,
3233                                             .policy = ct_nla_policy },
3234         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3235                                             .attr_count = CTA_MAX,
3236                                             .policy = ct_nla_policy },
3237         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3238         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3239         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3240         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3241 };
3242
3243 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3244         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3245                                             .attr_count = CTA_EXPECT_MAX,
3246                                             .policy = exp_nla_policy },
3247         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3248                                             .attr_count = CTA_EXPECT_MAX,
3249                                             .policy = exp_nla_policy },
3250         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3251                                             .attr_count = CTA_EXPECT_MAX,
3252                                             .policy = exp_nla_policy },
3253         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3254 };
3255
3256 static const struct nfnetlink_subsystem ctnl_subsys = {
3257         .name                           = "conntrack",
3258         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3259         .cb_count                       = IPCTNL_MSG_MAX,
3260         .cb                             = ctnl_cb,
3261 };
3262
3263 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3264         .name                           = "conntrack_expect",
3265         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3266         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3267         .cb                             = ctnl_exp_cb,
3268 };
3269
3270 MODULE_ALIAS("ip_conntrack_netlink");
3271 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3272 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3273
3274 static int __net_init ctnetlink_net_init(struct net *net)
3275 {
3276 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3277         int ret;
3278
3279         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3280         if (ret < 0) {
3281                 pr_err("ctnetlink_init: cannot register notifier.\n");
3282                 goto err_out;
3283         }
3284
3285         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3286         if (ret < 0) {
3287                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3288                 goto err_unreg_notifier;
3289         }
3290 #endif
3291         return 0;
3292
3293 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3294 err_unreg_notifier:
3295         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3296 err_out:
3297         return ret;
3298 #endif
3299 }
3300
3301 static void ctnetlink_net_exit(struct net *net)
3302 {
3303 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3304         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3305         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3306 #endif
3307 }
3308
3309 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3310 {
3311         struct net *net;
3312
3313         list_for_each_entry(net, net_exit_list, exit_list)
3314                 ctnetlink_net_exit(net);
3315 }
3316
3317 static struct pernet_operations ctnetlink_net_ops = {
3318         .init           = ctnetlink_net_init,
3319         .exit_batch     = ctnetlink_net_exit_batch,
3320 };
3321
3322 static int __init ctnetlink_init(void)
3323 {
3324         int ret;
3325
3326         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3327         ret = nfnetlink_subsys_register(&ctnl_subsys);
3328         if (ret < 0) {
3329                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3330                 goto err_out;
3331         }
3332
3333         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3334         if (ret < 0) {
3335                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3336                 goto err_unreg_subsys;
3337         }
3338
3339         ret = register_pernet_subsys(&ctnetlink_net_ops);
3340         if (ret < 0) {
3341                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3342                 goto err_unreg_exp_subsys;
3343         }
3344 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3345         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3346         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3347 #endif
3348         return 0;
3349
3350 err_unreg_exp_subsys:
3351         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3352 err_unreg_subsys:
3353         nfnetlink_subsys_unregister(&ctnl_subsys);
3354 err_out:
3355         return ret;
3356 }
3357
3358 static void __exit ctnetlink_exit(void)
3359 {
3360         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3361
3362         unregister_pernet_subsys(&ctnetlink_net_ops);
3363         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3364         nfnetlink_subsys_unregister(&ctnl_subsys);
3365 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3366         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3367 #endif
3368 }
3369
3370 module_init(ctnetlink_init);
3371 module_exit(ctnetlink_exit);