kernel/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_GLUE_CT
2137 static size_t
2138 ctnetlink_glue_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 struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2166                                              enum ip_conntrack_info *ctinfo)
2167 {
2168         struct nf_conn *ct;
2169
2170         ct = nf_ct_get(skb, ctinfo);
2171         if (ct && nf_ct_is_untracked(ct))
2172                 ct = NULL;
2173
2174         return ct;
2175 }
2176
2177 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2178 {
2179         const struct nf_conntrack_zone *zone;
2180         struct nlattr *nest_parms;
2181
2182         rcu_read_lock();
2183         zone = nf_ct_zone(ct);
2184
2185         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2186         if (!nest_parms)
2187                 goto nla_put_failure;
2188         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2189                 goto nla_put_failure;
2190         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2191                                    NF_CT_ZONE_DIR_ORIG) < 0)
2192                 goto nla_put_failure;
2193         nla_nest_end(skb, nest_parms);
2194
2195         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2196         if (!nest_parms)
2197                 goto nla_put_failure;
2198         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2199                 goto nla_put_failure;
2200         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2201                                    NF_CT_ZONE_DIR_REPL) < 0)
2202                 goto nla_put_failure;
2203         nla_nest_end(skb, nest_parms);
2204
2205         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2206                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2207                 goto nla_put_failure;
2208
2209         if (ctnetlink_dump_id(skb, ct) < 0)
2210                 goto nla_put_failure;
2211
2212         if (ctnetlink_dump_status(skb, ct) < 0)
2213                 goto nla_put_failure;
2214
2215         if (ctnetlink_dump_timeout(skb, ct) < 0)
2216                 goto nla_put_failure;
2217
2218         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2219                 goto nla_put_failure;
2220
2221         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2222                 goto nla_put_failure;
2223
2224 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2225         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2226                 goto nla_put_failure;
2227 #endif
2228         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2229                 goto nla_put_failure;
2230
2231         if ((ct->status & IPS_SEQ_ADJUST) &&
2232             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2233                 goto nla_put_failure;
2234
2235 #ifdef CONFIG_NF_CONNTRACK_MARK
2236         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2237                 goto nla_put_failure;
2238 #endif
2239         if (ctnetlink_dump_labels(skb, ct) < 0)
2240                 goto nla_put_failure;
2241         rcu_read_unlock();
2242         return 0;
2243
2244 nla_put_failure:
2245         rcu_read_unlock();
2246         return -ENOSPC;
2247 }
2248
2249 static int
2250 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2251                      enum ip_conntrack_info ctinfo,
2252                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2253 {
2254         struct nlattr *nest_parms;
2255
2256         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2257         if (!nest_parms)
2258                 goto nla_put_failure;
2259
2260         if (__ctnetlink_glue_build(skb, ct) < 0)
2261                 goto nla_put_failure;
2262
2263         nla_nest_end(skb, nest_parms);
2264
2265         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2266                 goto nla_put_failure;
2267
2268         return 0;
2269
2270 nla_put_failure:
2271         return -ENOSPC;
2272 }
2273
2274 static int
2275 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2276 {
2277         int err;
2278
2279         if (cda[CTA_TIMEOUT]) {
2280                 err = ctnetlink_change_timeout(ct, cda);
2281                 if (err < 0)
2282                         return err;
2283         }
2284         if (cda[CTA_STATUS]) {
2285                 err = ctnetlink_change_status(ct, cda);
2286                 if (err < 0)
2287                         return err;
2288         }
2289         if (cda[CTA_HELP]) {
2290                 err = ctnetlink_change_helper(ct, cda);
2291                 if (err < 0)
2292                         return err;
2293         }
2294         if (cda[CTA_LABELS]) {
2295                 err = ctnetlink_attach_labels(ct, cda);
2296                 if (err < 0)
2297                         return err;
2298         }
2299 #if defined(CONFIG_NF_CONNTRACK_MARK)
2300         if (cda[CTA_MARK]) {
2301                 u32 mask = 0, mark, newmark;
2302                 if (cda[CTA_MARK_MASK])
2303                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2304
2305                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2306                 newmark = (ct->mark & mask) ^ mark;
2307                 if (newmark != ct->mark)
2308                         ct->mark = newmark;
2309         }
2310 #endif
2311         return 0;
2312 }
2313
2314 static int
2315 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2316 {
2317         struct nlattr *cda[CTA_MAX+1];
2318         int ret;
2319
2320         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2321         if (ret < 0)
2322                 return ret;
2323
2324         spin_lock_bh(&nf_conntrack_expect_lock);
2325         ret = ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2326         spin_unlock_bh(&nf_conntrack_expect_lock);
2327
2328         return ret;
2329 }
2330
2331 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2332                                     const struct nf_conn *ct,
2333                                     struct nf_conntrack_tuple *tuple,
2334                                     struct nf_conntrack_tuple *mask)
2335 {
2336         int err;
2337
2338         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2339                                     nf_ct_l3num(ct), NULL);
2340         if (err < 0)
2341                 return err;
2342
2343         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2344                                      nf_ct_l3num(ct), NULL);
2345 }
2346
2347 static int
2348 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2349                              u32 portid, u32 report)
2350 {
2351         struct nlattr *cda[CTA_EXPECT_MAX+1];
2352         struct nf_conntrack_tuple tuple, mask;
2353         struct nf_conntrack_helper *helper = NULL;
2354         struct nf_conntrack_expect *exp;
2355         int err;
2356
2357         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2358         if (err < 0)
2359                 return err;
2360
2361         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2362                                        ct, &tuple, &mask);
2363         if (err < 0)
2364                 return err;
2365
2366         if (cda[CTA_EXPECT_HELP_NAME]) {
2367                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2368
2369                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2370                                                     nf_ct_protonum(ct));
2371                 if (helper == NULL)
2372                         return -EOPNOTSUPP;
2373         }
2374
2375         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2376                                      helper, &tuple, &mask);
2377         if (IS_ERR(exp))
2378                 return PTR_ERR(exp);
2379
2380         err = nf_ct_expect_related_report(exp, portid, report);
2381         if (err < 0) {
2382                 nf_ct_expect_put(exp);
2383                 return err;
2384         }
2385
2386         return 0;
2387 }
2388
2389 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2390                                   enum ip_conntrack_info ctinfo, int diff)
2391 {
2392         if (!(ct->status & IPS_NAT_MASK))
2393                 return;
2394
2395         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2396 }
2397
2398 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2399         .get_ct         = ctnetlink_glue_get_ct,
2400         .build_size     = ctnetlink_glue_build_size,
2401         .build          = ctnetlink_glue_build,
2402         .parse          = ctnetlink_glue_parse,
2403         .attach_expect  = ctnetlink_glue_attach_expect,
2404         .seq_adjust     = ctnetlink_glue_seqadj,
2405 };
2406 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2407
2408 /***********************************************************************
2409  * EXPECT
2410  ***********************************************************************/
2411
2412 static inline int
2413 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2414                          const struct nf_conntrack_tuple *tuple,
2415                          enum ctattr_expect type)
2416 {
2417         struct nlattr *nest_parms;
2418
2419         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2420         if (!nest_parms)
2421                 goto nla_put_failure;
2422         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2423                 goto nla_put_failure;
2424         nla_nest_end(skb, nest_parms);
2425
2426         return 0;
2427
2428 nla_put_failure:
2429         return -1;
2430 }
2431
2432 static inline int
2433 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2434                         const struct nf_conntrack_tuple *tuple,
2435                         const struct nf_conntrack_tuple_mask *mask)
2436 {
2437         int ret;
2438         struct nf_conntrack_l3proto *l3proto;
2439         struct nf_conntrack_l4proto *l4proto;
2440         struct nf_conntrack_tuple m;
2441         struct nlattr *nest_parms;
2442
2443         memset(&m, 0xFF, sizeof(m));
2444         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2445         m.src.u.all = mask->src.u.all;
2446         m.dst.protonum = tuple->dst.protonum;
2447
2448         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2449         if (!nest_parms)
2450                 goto nla_put_failure;
2451
2452         rcu_read_lock();
2453         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2454         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2455         if (ret >= 0) {
2456                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2457                                                tuple->dst.protonum);
2458         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2459         }
2460         rcu_read_unlock();
2461
2462         if (unlikely(ret < 0))
2463                 goto nla_put_failure;
2464
2465         nla_nest_end(skb, nest_parms);
2466
2467         return 0;
2468
2469 nla_put_failure:
2470         return -1;
2471 }
2472
2473 static const union nf_inet_addr any_addr;
2474
2475 static int
2476 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2477                           const struct nf_conntrack_expect *exp)
2478 {
2479         struct nf_conn *master = exp->master;
2480         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2481         struct nf_conn_help *help;
2482 #ifdef CONFIG_NF_NAT_NEEDED
2483         struct nlattr *nest_parms;
2484         struct nf_conntrack_tuple nat_tuple = {};
2485 #endif
2486         struct nf_ct_helper_expectfn *expfn;
2487
2488         if (timeout < 0)
2489                 timeout = 0;
2490
2491         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2492                 goto nla_put_failure;
2493         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2494                 goto nla_put_failure;
2495         if (ctnetlink_exp_dump_tuple(skb,
2496                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2497                                  CTA_EXPECT_MASTER) < 0)
2498                 goto nla_put_failure;
2499
2500 #ifdef CONFIG_NF_NAT_NEEDED
2501         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2502             exp->saved_proto.all) {
2503                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2504                 if (!nest_parms)
2505                         goto nla_put_failure;
2506
2507                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2508                         goto nla_put_failure;
2509
2510                 nat_tuple.src.l3num = nf_ct_l3num(master);
2511                 nat_tuple.src.u3 = exp->saved_addr;
2512                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2513                 nat_tuple.src.u = exp->saved_proto;
2514
2515                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2516                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2517                         goto nla_put_failure;
2518                 nla_nest_end(skb, nest_parms);
2519         }
2520 #endif
2521         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2522             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2523             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2524             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2525                 goto nla_put_failure;
2526         help = nfct_help(master);
2527         if (help) {
2528                 struct nf_conntrack_helper *helper;
2529
2530                 helper = rcu_dereference(help->helper);
2531                 if (helper &&
2532                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2533                         goto nla_put_failure;
2534         }
2535         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2536         if (expfn != NULL &&
2537             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2538                 goto nla_put_failure;
2539
2540         return 0;
2541
2542 nla_put_failure:
2543         return -1;
2544 }
2545
2546 static int
2547 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2548                         int event, const struct nf_conntrack_expect *exp)
2549 {
2550         struct nlmsghdr *nlh;
2551         struct nfgenmsg *nfmsg;
2552         unsigned int flags = portid ? NLM_F_MULTI : 0;
2553
2554         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2555         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2556         if (nlh == NULL)
2557                 goto nlmsg_failure;
2558
2559         nfmsg = nlmsg_data(nlh);
2560         nfmsg->nfgen_family = exp->tuple.src.l3num;
2561         nfmsg->version      = NFNETLINK_V0;
2562         nfmsg->res_id       = 0;
2563
2564         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2565                 goto nla_put_failure;
2566
2567         nlmsg_end(skb, nlh);
2568         return skb->len;
2569
2570 nlmsg_failure:
2571 nla_put_failure:
2572         nlmsg_cancel(skb, nlh);
2573         return -1;
2574 }
2575
2576 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2577 static int
2578 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2579 {
2580         struct nf_conntrack_expect *exp = item->exp;
2581         struct net *net = nf_ct_exp_net(exp);
2582         struct nlmsghdr *nlh;
2583         struct nfgenmsg *nfmsg;
2584         struct sk_buff *skb;
2585         unsigned int type, group;
2586         int flags = 0;
2587
2588         if (events & (1 << IPEXP_DESTROY)) {
2589                 type = IPCTNL_MSG_EXP_DELETE;
2590                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2591         } else if (events & (1 << IPEXP_NEW)) {
2592                 type = IPCTNL_MSG_EXP_NEW;
2593                 flags = NLM_F_CREATE|NLM_F_EXCL;
2594                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2595         } else
2596                 return 0;
2597
2598         if (!item->report && !nfnetlink_has_listeners(net, group))
2599                 return 0;
2600
2601         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2602         if (skb == NULL)
2603                 goto errout;
2604
2605         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2606         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2607         if (nlh == NULL)
2608                 goto nlmsg_failure;
2609
2610         nfmsg = nlmsg_data(nlh);
2611         nfmsg->nfgen_family = exp->tuple.src.l3num;
2612         nfmsg->version      = NFNETLINK_V0;
2613         nfmsg->res_id       = 0;
2614
2615         rcu_read_lock();
2616         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2617                 goto nla_put_failure;
2618         rcu_read_unlock();
2619
2620         nlmsg_end(skb, nlh);
2621         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2622         return 0;
2623
2624 nla_put_failure:
2625         rcu_read_unlock();
2626         nlmsg_cancel(skb, nlh);
2627 nlmsg_failure:
2628         kfree_skb(skb);
2629 errout:
2630         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2631         return 0;
2632 }
2633 #endif
2634 static int ctnetlink_exp_done(struct netlink_callback *cb)
2635 {
2636         if (cb->args[1])
2637                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2638         return 0;
2639 }
2640
2641 static int
2642 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2643 {
2644         struct net *net = sock_net(skb->sk);
2645         struct nf_conntrack_expect *exp, *last;
2646         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2647         u_int8_t l3proto = nfmsg->nfgen_family;
2648
2649         rcu_read_lock();
2650         last = (struct nf_conntrack_expect *)cb->args[1];
2651         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2652 restart:
2653                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2654                                      hnode) {
2655                         if (l3proto && exp->tuple.src.l3num != l3proto)
2656                                 continue;
2657                         if (cb->args[1]) {
2658                                 if (exp != last)
2659                                         continue;
2660                                 cb->args[1] = 0;
2661                         }
2662                         if (ctnetlink_exp_fill_info(skb,
2663                                                     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         }
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
2687 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2688 {
2689         struct nf_conntrack_expect *exp, *last;
2690         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2691         struct nf_conn *ct = cb->data;
2692         struct nf_conn_help *help = nfct_help(ct);
2693         u_int8_t l3proto = nfmsg->nfgen_family;
2694
2695         if (cb->args[0])
2696                 return 0;
2697
2698         rcu_read_lock();
2699         last = (struct nf_conntrack_expect *)cb->args[1];
2700 restart:
2701         hlist_for_each_entry(exp, &help->expectations, lnode) {
2702                 if (l3proto && exp->tuple.src.l3num != l3proto)
2703                         continue;
2704                 if (cb->args[1]) {
2705                         if (exp != last)
2706                                 continue;
2707                         cb->args[1] = 0;
2708                 }
2709                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2710                                             cb->nlh->nlmsg_seq,
2711                                             IPCTNL_MSG_EXP_NEW,
2712                                             exp) < 0) {
2713                         if (!atomic_inc_not_zero(&exp->use))
2714                                 continue;
2715                         cb->args[1] = (unsigned long)exp;
2716                         goto out;
2717                 }
2718         }
2719         if (cb->args[1]) {
2720                 cb->args[1] = 0;
2721                 goto restart;
2722         }
2723         cb->args[0] = 1;
2724 out:
2725         rcu_read_unlock();
2726         if (last)
2727                 nf_ct_expect_put(last);
2728
2729         return skb->len;
2730 }
2731
2732 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2733                                  const struct nlmsghdr *nlh,
2734                                  const struct nlattr * const cda[])
2735 {
2736         int err;
2737         struct net *net = sock_net(ctnl);
2738         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2739         u_int8_t u3 = nfmsg->nfgen_family;
2740         struct nf_conntrack_tuple tuple;
2741         struct nf_conntrack_tuple_hash *h;
2742         struct nf_conn *ct;
2743         struct nf_conntrack_zone zone;
2744         struct netlink_dump_control c = {
2745                 .dump = ctnetlink_exp_ct_dump_table,
2746                 .done = ctnetlink_exp_done,
2747         };
2748
2749         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2750                                     u3, NULL);
2751         if (err < 0)
2752                 return err;
2753
2754         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2755         if (err < 0)
2756                 return err;
2757
2758         h = nf_conntrack_find_get(net, &zone, &tuple);
2759         if (!h)
2760                 return -ENOENT;
2761
2762         ct = nf_ct_tuplehash_to_ctrack(h);
2763         c.data = ct;
2764
2765         err = netlink_dump_start(ctnl, skb, nlh, &c);
2766         nf_ct_put(ct);
2767
2768         return err;
2769 }
2770
2771 static int
2772 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2773                      const struct nlmsghdr *nlh,
2774                      const struct nlattr * const cda[])
2775 {
2776         struct net *net = sock_net(ctnl);
2777         struct nf_conntrack_tuple tuple;
2778         struct nf_conntrack_expect *exp;
2779         struct sk_buff *skb2;
2780         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2781         u_int8_t u3 = nfmsg->nfgen_family;
2782         struct nf_conntrack_zone zone;
2783         int err;
2784
2785         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2786                 if (cda[CTA_EXPECT_MASTER])
2787                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2788                 else {
2789                         struct netlink_dump_control c = {
2790                                 .dump = ctnetlink_exp_dump_table,
2791                                 .done = ctnetlink_exp_done,
2792                         };
2793                         return netlink_dump_start(ctnl, skb, nlh, &c);
2794                 }
2795         }
2796
2797         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2798         if (err < 0)
2799                 return err;
2800
2801         if (cda[CTA_EXPECT_TUPLE])
2802                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2803                                             u3, NULL);
2804         else if (cda[CTA_EXPECT_MASTER])
2805                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2806                                             u3, NULL);
2807         else
2808                 return -EINVAL;
2809
2810         if (err < 0)
2811                 return err;
2812
2813         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2814         if (!exp)
2815                 return -ENOENT;
2816
2817         if (cda[CTA_EXPECT_ID]) {
2818                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2819                 if (ntohl(id) != (u32)(unsigned long)exp) {
2820                         nf_ct_expect_put(exp);
2821                         return -ENOENT;
2822                 }
2823         }
2824
2825         err = -ENOMEM;
2826         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2827         if (skb2 == NULL) {
2828                 nf_ct_expect_put(exp);
2829                 goto out;
2830         }
2831
2832         rcu_read_lock();
2833         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2834                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2835         rcu_read_unlock();
2836         nf_ct_expect_put(exp);
2837         if (err <= 0)
2838                 goto free;
2839
2840         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2841         if (err < 0)
2842                 goto out;
2843
2844         return 0;
2845
2846 free:
2847         kfree_skb(skb2);
2848 out:
2849         /* this avoids a loop in nfnetlink. */
2850         return err == -EAGAIN ? -ENOBUFS : err;
2851 }
2852
2853 static int
2854 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2855                      const struct nlmsghdr *nlh,
2856                      const struct nlattr * const cda[])
2857 {
2858         struct net *net = sock_net(ctnl);
2859         struct nf_conntrack_expect *exp;
2860         struct nf_conntrack_tuple tuple;
2861         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2862         struct hlist_node *next;
2863         u_int8_t u3 = nfmsg->nfgen_family;
2864         struct nf_conntrack_zone zone;
2865         unsigned int i;
2866         int err;
2867
2868         if (cda[CTA_EXPECT_TUPLE]) {
2869                 /* delete a single expect by tuple */
2870                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2871                 if (err < 0)
2872                         return err;
2873
2874                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2875                                             u3, NULL);
2876                 if (err < 0)
2877                         return err;
2878
2879                 /* bump usage count to 2 */
2880                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2881                 if (!exp)
2882                         return -ENOENT;
2883
2884                 if (cda[CTA_EXPECT_ID]) {
2885                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2886                         if (ntohl(id) != (u32)(unsigned long)exp) {
2887                                 nf_ct_expect_put(exp);
2888                                 return -ENOENT;
2889                         }
2890                 }
2891
2892                 /* after list removal, usage count == 1 */
2893                 spin_lock_bh(&nf_conntrack_expect_lock);
2894                 if (del_timer(&exp->timeout)) {
2895                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2896                                                    nlmsg_report(nlh));
2897                         nf_ct_expect_put(exp);
2898                 }
2899                 spin_unlock_bh(&nf_conntrack_expect_lock);
2900                 /* have to put what we 'get' above.
2901                  * after this line usage count == 0 */
2902                 nf_ct_expect_put(exp);
2903         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2904                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2905                 struct nf_conn_help *m_help;
2906
2907                 /* delete all expectations for this helper */
2908                 spin_lock_bh(&nf_conntrack_expect_lock);
2909                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2910                         hlist_for_each_entry_safe(exp, next,
2911                                                   &net->ct.expect_hash[i],
2912                                                   hnode) {
2913                                 m_help = nfct_help(exp->master);
2914                                 if (!strcmp(m_help->helper->name, name) &&
2915                                     del_timer(&exp->timeout)) {
2916                                         nf_ct_unlink_expect_report(exp,
2917                                                         NETLINK_CB(skb).portid,
2918                                                         nlmsg_report(nlh));
2919                                         nf_ct_expect_put(exp);
2920                                 }
2921                         }
2922                 }
2923                 spin_unlock_bh(&nf_conntrack_expect_lock);
2924         } else {
2925                 /* This basically means we have to flush everything*/
2926                 spin_lock_bh(&nf_conntrack_expect_lock);
2927                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2928                         hlist_for_each_entry_safe(exp, next,
2929                                                   &net->ct.expect_hash[i],
2930                                                   hnode) {
2931                                 if (del_timer(&exp->timeout)) {
2932                                         nf_ct_unlink_expect_report(exp,
2933                                                         NETLINK_CB(skb).portid,
2934                                                         nlmsg_report(nlh));
2935                                         nf_ct_expect_put(exp);
2936                                 }
2937                         }
2938                 }
2939                 spin_unlock_bh(&nf_conntrack_expect_lock);
2940         }
2941
2942         return 0;
2943 }
2944 static int
2945 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2946                         const struct nlattr * const cda[])
2947 {
2948         if (cda[CTA_EXPECT_TIMEOUT]) {
2949                 if (!del_timer(&x->timeout))
2950                         return -ETIME;
2951
2952                 x->timeout.expires = jiffies +
2953                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2954                 add_timer(&x->timeout);
2955         }
2956         return 0;
2957 }
2958
2959 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2960         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2961         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2962 };
2963
2964 static int
2965 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2966                            struct nf_conntrack_expect *exp,
2967                            u_int8_t u3)
2968 {
2969 #ifdef CONFIG_NF_NAT_NEEDED
2970         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2971         struct nf_conntrack_tuple nat_tuple = {};
2972         int err;
2973
2974         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2975         if (err < 0)
2976                 return err;
2977
2978         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2979                 return -EINVAL;
2980
2981         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2982                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
2983                                     u3, NULL);
2984         if (err < 0)
2985                 return err;
2986
2987         exp->saved_addr = nat_tuple.src.u3;
2988         exp->saved_proto = nat_tuple.src.u;
2989         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2990
2991         return 0;
2992 #else
2993         return -EOPNOTSUPP;
2994 #endif
2995 }
2996
2997 static struct nf_conntrack_expect *
2998 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2999                        struct nf_conntrack_helper *helper,
3000                        struct nf_conntrack_tuple *tuple,
3001                        struct nf_conntrack_tuple *mask)
3002 {
3003         u_int32_t class = 0;
3004         struct nf_conntrack_expect *exp;
3005         struct nf_conn_help *help;
3006         int err;
3007
3008         if (cda[CTA_EXPECT_CLASS] && helper) {
3009                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3010                 if (class > helper->expect_class_max)
3011                         return ERR_PTR(-EINVAL);
3012         }
3013         exp = nf_ct_expect_alloc(ct);
3014         if (!exp)
3015                 return ERR_PTR(-ENOMEM);
3016
3017         help = nfct_help(ct);
3018         if (!help) {
3019                 if (!cda[CTA_EXPECT_TIMEOUT]) {
3020                         err = -EINVAL;
3021                         goto err_out;
3022                 }
3023                 exp->timeout.expires =
3024                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3025
3026                 exp->flags = NF_CT_EXPECT_USERSPACE;
3027                 if (cda[CTA_EXPECT_FLAGS]) {
3028                         exp->flags |=
3029                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3030                 }
3031         } else {
3032                 if (cda[CTA_EXPECT_FLAGS]) {
3033                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3034                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3035                 } else
3036                         exp->flags = 0;
3037         }
3038         if (cda[CTA_EXPECT_FN]) {
3039                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3040                 struct nf_ct_helper_expectfn *expfn;
3041
3042                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3043                 if (expfn == NULL) {
3044                         err = -EINVAL;
3045                         goto err_out;
3046                 }
3047                 exp->expectfn = expfn->expectfn;
3048         } else
3049                 exp->expectfn = NULL;
3050
3051         exp->class = class;
3052         exp->master = ct;
3053         exp->helper = helper;
3054         exp->tuple = *tuple;
3055         exp->mask.src.u3 = mask->src.u3;
3056         exp->mask.src.u.all = mask->src.u.all;
3057
3058         if (cda[CTA_EXPECT_NAT]) {
3059                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3060                                                  exp, nf_ct_l3num(ct));
3061                 if (err < 0)
3062                         goto err_out;
3063         }
3064         return exp;
3065 err_out:
3066         nf_ct_expect_put(exp);
3067         return ERR_PTR(err);
3068 }
3069
3070 static int
3071 ctnetlink_create_expect(struct net *net,
3072                         const struct nf_conntrack_zone *zone,
3073                         const struct nlattr * const cda[],
3074                         u_int8_t u3, u32 portid, int report)
3075 {
3076         struct nf_conntrack_tuple tuple, mask, master_tuple;
3077         struct nf_conntrack_tuple_hash *h = NULL;
3078         struct nf_conntrack_helper *helper = NULL;
3079         struct nf_conntrack_expect *exp;
3080         struct nf_conn *ct;
3081         int err;
3082
3083         /* caller guarantees that those three CTA_EXPECT_* exist */
3084         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3085                                     u3, NULL);
3086         if (err < 0)
3087                 return err;
3088         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3089                                     u3, NULL);
3090         if (err < 0)
3091                 return err;
3092         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3093                                     u3, NULL);
3094         if (err < 0)
3095                 return err;
3096
3097         /* Look for master conntrack of this expectation */
3098         h = nf_conntrack_find_get(net, zone, &master_tuple);
3099         if (!h)
3100                 return -ENOENT;
3101         ct = nf_ct_tuplehash_to_ctrack(h);
3102
3103         if (cda[CTA_EXPECT_HELP_NAME]) {
3104                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3105
3106                 helper = __nf_conntrack_helper_find(helpname, u3,
3107                                                     nf_ct_protonum(ct));
3108                 if (helper == NULL) {
3109 #ifdef CONFIG_MODULES
3110                         if (request_module("nfct-helper-%s", helpname) < 0) {
3111                                 err = -EOPNOTSUPP;
3112                                 goto err_ct;
3113                         }
3114                         helper = __nf_conntrack_helper_find(helpname, u3,
3115                                                             nf_ct_protonum(ct));
3116                         if (helper) {
3117                                 err = -EAGAIN;
3118                                 goto err_ct;
3119                         }
3120 #endif
3121                         err = -EOPNOTSUPP;
3122                         goto err_ct;
3123                 }
3124         }
3125
3126         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3127         if (IS_ERR(exp)) {
3128                 err = PTR_ERR(exp);
3129                 goto err_ct;
3130         }
3131
3132         err = nf_ct_expect_related_report(exp, portid, report);
3133         nf_ct_expect_put(exp);
3134 err_ct:
3135         nf_ct_put(ct);
3136         return err;
3137 }
3138
3139 static int
3140 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
3141                      const struct nlmsghdr *nlh,
3142                      const struct nlattr * const cda[])
3143 {
3144         struct net *net = sock_net(ctnl);
3145         struct nf_conntrack_tuple tuple;
3146         struct nf_conntrack_expect *exp;
3147         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3148         u_int8_t u3 = nfmsg->nfgen_family;
3149         struct nf_conntrack_zone zone;
3150         int err;
3151
3152         if (!cda[CTA_EXPECT_TUPLE]
3153             || !cda[CTA_EXPECT_MASK]
3154             || !cda[CTA_EXPECT_MASTER])
3155                 return -EINVAL;
3156
3157         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3158         if (err < 0)
3159                 return err;
3160
3161         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3162                                     u3, NULL);
3163         if (err < 0)
3164                 return err;
3165
3166         spin_lock_bh(&nf_conntrack_expect_lock);
3167         exp = __nf_ct_expect_find(net, &zone, &tuple);
3168         if (!exp) {
3169                 spin_unlock_bh(&nf_conntrack_expect_lock);
3170                 err = -ENOENT;
3171                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3172                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3173                                                       NETLINK_CB(skb).portid,
3174                                                       nlmsg_report(nlh));
3175                 }
3176                 return err;
3177         }
3178
3179         err = -EEXIST;
3180         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3181                 err = ctnetlink_change_expect(exp, cda);
3182         spin_unlock_bh(&nf_conntrack_expect_lock);
3183
3184         return err;
3185 }
3186
3187 static int
3188 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3189                              const struct ip_conntrack_stat *st)
3190 {
3191         struct nlmsghdr *nlh;
3192         struct nfgenmsg *nfmsg;
3193         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3194
3195         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3196         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3197         if (nlh == NULL)
3198                 goto nlmsg_failure;
3199
3200         nfmsg = nlmsg_data(nlh);
3201         nfmsg->nfgen_family = AF_UNSPEC;
3202         nfmsg->version      = NFNETLINK_V0;
3203         nfmsg->res_id       = htons(cpu);
3204
3205         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3206             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3207             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3208                 goto nla_put_failure;
3209
3210         nlmsg_end(skb, nlh);
3211         return skb->len;
3212
3213 nla_put_failure:
3214 nlmsg_failure:
3215         nlmsg_cancel(skb, nlh);
3216         return -1;
3217 }
3218
3219 static int
3220 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3221 {
3222         int cpu;
3223         struct net *net = sock_net(skb->sk);
3224
3225         if (cb->args[0] == nr_cpu_ids)
3226                 return 0;
3227
3228         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3229                 const struct ip_conntrack_stat *st;
3230
3231                 if (!cpu_possible(cpu))
3232                         continue;
3233
3234                 st = per_cpu_ptr(net->ct.stat, cpu);
3235                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3236                                                  cb->nlh->nlmsg_seq,
3237                                                  cpu, st) < 0)
3238                         break;
3239         }
3240         cb->args[0] = cpu;
3241
3242         return skb->len;
3243 }
3244
3245 static int
3246 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3247                        const struct nlmsghdr *nlh,
3248                        const struct nlattr * const cda[])
3249 {
3250         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3251                 struct netlink_dump_control c = {
3252                         .dump = ctnetlink_exp_stat_cpu_dump,
3253                 };
3254                 return netlink_dump_start(ctnl, skb, nlh, &c);
3255         }
3256
3257         return 0;
3258 }
3259
3260 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3261 static struct nf_ct_event_notifier ctnl_notifier = {
3262         .fcn = ctnetlink_conntrack_event,
3263 };
3264
3265 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3266         .fcn = ctnetlink_expect_event,
3267 };
3268 #endif
3269
3270 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3271         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3272                                             .attr_count = CTA_MAX,
3273                                             .policy = ct_nla_policy },
3274         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3275                                             .attr_count = CTA_MAX,
3276                                             .policy = ct_nla_policy },
3277         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3278                                             .attr_count = CTA_MAX,
3279                                             .policy = ct_nla_policy },
3280         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3281                                             .attr_count = CTA_MAX,
3282                                             .policy = ct_nla_policy },
3283         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3284         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3285         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3286         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3287 };
3288
3289 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3290         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3291                                             .attr_count = CTA_EXPECT_MAX,
3292                                             .policy = exp_nla_policy },
3293         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3294                                             .attr_count = CTA_EXPECT_MAX,
3295                                             .policy = exp_nla_policy },
3296         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3297                                             .attr_count = CTA_EXPECT_MAX,
3298                                             .policy = exp_nla_policy },
3299         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3300 };
3301
3302 static const struct nfnetlink_subsystem ctnl_subsys = {
3303         .name                           = "conntrack",
3304         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3305         .cb_count                       = IPCTNL_MSG_MAX,
3306         .cb                             = ctnl_cb,
3307 };
3308
3309 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3310         .name                           = "conntrack_expect",
3311         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3312         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3313         .cb                             = ctnl_exp_cb,
3314 };
3315
3316 MODULE_ALIAS("ip_conntrack_netlink");
3317 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3318 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3319
3320 static int __net_init ctnetlink_net_init(struct net *net)
3321 {
3322 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3323         int ret;
3324
3325         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3326         if (ret < 0) {
3327                 pr_err("ctnetlink_init: cannot register notifier.\n");
3328                 goto err_out;
3329         }
3330
3331         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3332         if (ret < 0) {
3333                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3334                 goto err_unreg_notifier;
3335         }
3336 #endif
3337         return 0;
3338
3339 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3340 err_unreg_notifier:
3341         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3342 err_out:
3343         return ret;
3344 #endif
3345 }
3346
3347 static void ctnetlink_net_exit(struct net *net)
3348 {
3349 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3350         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3351         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3352 #endif
3353 }
3354
3355 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3356 {
3357         struct net *net;
3358
3359         list_for_each_entry(net, net_exit_list, exit_list)
3360                 ctnetlink_net_exit(net);
3361 }
3362
3363 static struct pernet_operations ctnetlink_net_ops = {
3364         .init           = ctnetlink_net_init,
3365         .exit_batch     = ctnetlink_net_exit_batch,
3366 };
3367
3368 static int __init ctnetlink_init(void)
3369 {
3370         int ret;
3371
3372         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3373         ret = nfnetlink_subsys_register(&ctnl_subsys);
3374         if (ret < 0) {
3375                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3376                 goto err_out;
3377         }
3378
3379         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3380         if (ret < 0) {
3381                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3382                 goto err_unreg_subsys;
3383         }
3384
3385         ret = register_pernet_subsys(&ctnetlink_net_ops);
3386         if (ret < 0) {
3387                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3388                 goto err_unreg_exp_subsys;
3389         }
3390 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3391         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3392         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3393 #endif
3394         return 0;
3395
3396 err_unreg_exp_subsys:
3397         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3398 err_unreg_subsys:
3399         nfnetlink_subsys_unregister(&ctnl_subsys);
3400 err_out:
3401         return ret;
3402 }
3403
3404 static void __exit ctnetlink_exit(void)
3405 {
3406         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3407
3408         unregister_pernet_subsys(&ctnetlink_net_ops);
3409         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3410         nfnetlink_subsys_unregister(&ctnl_subsys);
3411 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3412         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3413 #endif
3414 }
3415
3416 module_init(ctnetlink_init);
3417 module_exit(ctnetlink_exit);