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_status(struct sk_buff *skb, const struct nf_conn *ct)
 132 {
 133         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 134                 goto nla_put_failure;
 135         return 0;
 136
 137 nla_put_failure:
 138         return -1;
 139 }
 140
 141 static inline int
 142 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 143 {
 144         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 145
 146         if (timeout < 0)
 147                 timeout = 0;
 148
 149         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 150                 goto nla_put_failure;
 151         return 0;
 152
 153 nla_put_failure:
 154         return -1;
 155 }
 156
 157 static inline int
 158 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 159 {
 160         struct nf_conntrack_l4proto *l4proto;
 161         struct nlattr *nest_proto;
 162         int ret;
 163
 164         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 165         if (!l4proto->to_nlattr)
 166                 return 0;
 167
 168         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 169         if (!nest_proto)
 170                 goto nla_put_failure;
 171
 172         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 173
 174         nla_nest_end(skb, nest_proto);
 175
 176         return ret;
 177
 178 nla_put_failure:
 179         return -1;
 180 }
 181
 182 static inline int
 183 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 184 {
 185         struct nlattr *nest_helper;
 186         const struct nf_conn_help *help = nfct_help(ct);
 187         struct nf_conntrack_helper *helper;
 188
 189         if (!help)
 190                 return 0;
 191
 192         helper = rcu_dereference(help->helper);
 193         if (!helper)
 194                 goto out;
 195
 196         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 197         if (!nest_helper)
 198                 goto nla_put_failure;
 199         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 200                 goto nla_put_failure;
 201
 202         if (helper->to_nlattr)
 203                 helper->to_nlattr(skb, ct);
 204
 205         nla_nest_end(skb, nest_helper);
 206 out:
 207         return 0;
 208
 209 nla_put_failure:
 210         return -1;
 211 }
 212
 213 static int
 214 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 215               enum ip_conntrack_dir dir, int type)
 216 {
 217         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 218         struct nf_conn_counter *counter = acct->counter;
 219         struct nlattr *nest_count;
 220         u64 pkts, bytes;
 221
 222         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 223                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 224                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 225         } else {
 226                 pkts = atomic64_read(&counter[dir].packets);
 227                 bytes = atomic64_read(&counter[dir].bytes);
 228         }
 229
 230         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 231         if (!nest_count)
 232                 goto nla_put_failure;
 233
 234         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 235             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 236                 goto nla_put_failure;
 237
 238         nla_nest_end(skb, nest_count);
 239
 240         return 0;
 241
 242 nla_put_failure:
 243         return -1;
 244 }
 245
 246 static int
 247 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 248 {
 249         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 250
 251         if (!acct)
 252                 return 0;
 253
 254         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 255                 return -1;
 256         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 257                 return -1;
 258
 259         return 0;
 260 }
 261
 262 static int
 263 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 264 {
 265         struct nlattr *nest_count;
 266         const struct nf_conn_tstamp *tstamp;
 267
 268         tstamp = nf_conn_tstamp_find(ct);
 269         if (!tstamp)
 270                 return 0;
 271
 272         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 273         if (!nest_count)
 274                 goto nla_put_failure;
 275
 276         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 277             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 278                                                cpu_to_be64(tstamp->stop))))
 279                 goto nla_put_failure;
 280         nla_nest_end(skb, nest_count);
 281
 282         return 0;
 283
 284 nla_put_failure:
 285         return -1;
 286 }
 287
 288 #ifdef CONFIG_NF_CONNTRACK_MARK
 289 static inline int
 290 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 291 {
 292         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 293                 goto nla_put_failure;
 294         return 0;
 295
 296 nla_put_failure:
 297         return -1;
 298 }
 299 #else
 300 #define ctnetlink_dump_mark(a, b) (0)
 301 #endif
 302
 303 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 304 static inline int
 305 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 306 {
 307         struct nlattr *nest_secctx;
 308         int len, ret;
 309         char *secctx;
 310
 311         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 312         if (ret)
 313                 return 0;
 314
 315         ret = -1;
 316         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 317         if (!nest_secctx)
 318                 goto nla_put_failure;
 319
 320         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 321                 goto nla_put_failure;
 322         nla_nest_end(skb, nest_secctx);
 323
 324         ret = 0;
 325 nla_put_failure:
 326         security_release_secctx(secctx, len);
 327         return ret;
 328 }
 329 #else
 330 #define ctnetlink_dump_secctx(a, b) (0)
 331 #endif
 332
 333 #ifdef CONFIG_NF_CONNTRACK_LABELS
 334 static int ctnetlink_label_size(const struct nf_conn *ct)
 335 {
 336         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 337
 338         if (!labels)
 339                 return 0;
 340         return nla_total_size(labels->words * sizeof(long));
 341 }
 342
 343 static int
 344 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 345 {
 346         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 347         unsigned int len, i;
 348
 349         if (!labels)
 350                 return 0;
 351
 352         len = labels->words * sizeof(long);
 353         i = 0;
 354         do {
 355                 if (labels->bits[i] != 0)
 356                         return nla_put(skb, CTA_LABELS, len, labels->bits);
 357                 i++;
 358         } while (i < labels->words);
 359
 360         return 0;
 361 }
 362 #else
 363 #define ctnetlink_dump_labels(a, b) (0)
 364 #define ctnetlink_label_size(a) (0)
 365 #endif
 366
 367 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 368
 369 static inline int
 370 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 371 {
 372         struct nlattr *nest_parms;
 373
 374         if (!(ct->status & IPS_EXPECTED))
 375                 return 0;
 376
 377         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 378         if (!nest_parms)
 379                 goto nla_put_failure;
 380         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 381                 goto nla_put_failure;
 382         nla_nest_end(skb, nest_parms);
 383
 384         return 0;
 385
 386 nla_put_failure:
 387         return -1;
 388 }
 389
 390 static int
 391 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 392 {
 393         struct nlattr *nest_parms;
 394
 395         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 396         if (!nest_parms)
 397                 goto nla_put_failure;
 398
 399         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 400                          htonl(seq->correction_pos)) ||
 401             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 402                          htonl(seq->offset_before)) ||
 403             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 404                          htonl(seq->offset_after)))
 405                 goto nla_put_failure;
 406
 407         nla_nest_end(skb, nest_parms);
 408
 409         return 0;
 410
 411 nla_put_failure:
 412         return -1;
 413 }
 414
 415 static inline int
 416 ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 417 {
 418         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 419         struct nf_ct_seqadj *seq;
 420
 421         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 422                 return 0;
 423
 424         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 425         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 426                 return -1;
 427
 428         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 429         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 430                 return -1;
 431
 432         return 0;
 433 }
 434
 435 static inline int
 436 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 437 {
 438         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 439                 goto nla_put_failure;
 440         return 0;
 441
 442 nla_put_failure:
 443         return -1;
 444 }
 445
 446 static inline int
 447 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 448 {
 449         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 450                 goto nla_put_failure;
 451         return 0;
 452
 453 nla_put_failure:
 454         return -1;
 455 }
 456
 457 static int
 458 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 459                     struct nf_conn *ct)
 460 {
 461         struct nlmsghdr *nlh;
 462         struct nfgenmsg *nfmsg;
 463         struct nlattr *nest_parms;
 464         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 465
 466         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 467         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 468         if (nlh == NULL)
 469                 goto nlmsg_failure;
 470
 471         nfmsg = nlmsg_data(nlh);
 472         nfmsg->nfgen_family = nf_ct_l3num(ct);
 473         nfmsg->version      = NFNETLINK_V0;
 474         nfmsg->res_id       = 0;
 475
 476         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 477         if (!nest_parms)
 478                 goto nla_put_failure;
 479         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 480                 goto nla_put_failure;
 481         nla_nest_end(skb, nest_parms);
 482
 483         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 484         if (!nest_parms)
 485                 goto nla_put_failure;
 486         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 487                 goto nla_put_failure;
 488         nla_nest_end(skb, nest_parms);
 489
 490         if (nf_ct_zone(ct) &&
 491             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 492                 goto nla_put_failure;
 493
 494         if (ctnetlink_dump_status(skb, ct) < 0 ||
 495             ctnetlink_dump_timeout(skb, ct) < 0 ||
 496             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 497             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 498             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 499             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 500             ctnetlink_dump_mark(skb, ct) < 0 ||
 501             ctnetlink_dump_secctx(skb, ct) < 0 ||
 502             ctnetlink_dump_labels(skb, ct) < 0 ||
 503             ctnetlink_dump_id(skb, ct) < 0 ||
 504             ctnetlink_dump_use(skb, ct) < 0 ||
 505             ctnetlink_dump_master(skb, ct) < 0 ||
 506             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 507                 goto nla_put_failure;
 508
 509         nlmsg_end(skb, nlh);
 510         return skb->len;
 511
 512 nlmsg_failure:
 513 nla_put_failure:
 514         nlmsg_cancel(skb, nlh);
 515         return -1;
 516 }
 517
 518 static inline size_t
 519 ctnetlink_proto_size(const struct nf_conn *ct)
 520 {
 521         struct nf_conntrack_l3proto *l3proto;
 522         struct nf_conntrack_l4proto *l4proto;
 523         size_t len = 0;
 524
 525         rcu_read_lock();
 526         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 527         len += l3proto->nla_size;
 528
 529         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 530         len += l4proto->nla_size;
 531         rcu_read_unlock();
 532
 533         return len;
 534 }
 535
 536 static inline size_t
 537 ctnetlink_acct_size(const struct nf_conn *ct)
 538 {
 539         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 540                 return 0;
 541         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 542                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 543                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 544                ;
 545 }
 546
 547 static inline int
 548 ctnetlink_secctx_size(const struct nf_conn *ct)
 549 {
 550 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 551         int len, ret;
 552
 553         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 554         if (ret)
 555                 return 0;
 556
 557         return nla_total_size(0) /* CTA_SECCTX */
 558                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 559 #else
 560         return 0;
 561 #endif
 562 }
 563
 564 static inline size_t
 565 ctnetlink_timestamp_size(const struct nf_conn *ct)
 566 {
 567 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 568         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 569                 return 0;
 570         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 571 #else
 572         return 0;
 573 #endif
 574 }
 575
 576 static inline size_t
 577 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 578 {
 579         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 580                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 581                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 582                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 583                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 584                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 585                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 586                + ctnetlink_acct_size(ct)
 587                + ctnetlink_timestamp_size(ct)
 588                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 589                + nla_total_size(0) /* CTA_PROTOINFO */
 590                + nla_total_size(0) /* CTA_HELP */
 591                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 592                + ctnetlink_secctx_size(ct)
 593 #ifdef CONFIG_NF_NAT_NEEDED
 594                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 595                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 596 #endif
 597 #ifdef CONFIG_NF_CONNTRACK_MARK
 598                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 599 #endif
 600                + ctnetlink_proto_size(ct)
 601                + ctnetlink_label_size(ct)
 602                ;
 603 }
 604
 605 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 606 static int
 607 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 608 {
 609         struct net *net;
 610         struct nlmsghdr *nlh;
 611         struct nfgenmsg *nfmsg;
 612         struct nlattr *nest_parms;
 613         struct nf_conn *ct = item->ct;
 614         struct sk_buff *skb;
 615         unsigned int type;
 616         unsigned int flags = 0, group;
 617         int err;
 618
 619         /* ignore our fake conntrack entry */
 620         if (nf_ct_is_untracked(ct))
 621                 return 0;
 622
 623         if (events & (1 << IPCT_DESTROY)) {
 624                 type = IPCTNL_MSG_CT_DELETE;
 625                 group = NFNLGRP_CONNTRACK_DESTROY;
 626         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 627                 type = IPCTNL_MSG_CT_NEW;
 628                 flags = NLM_F_CREATE|NLM_F_EXCL;
 629                 group = NFNLGRP_CONNTRACK_NEW;
 630         } else  if (events) {
 631                 type = IPCTNL_MSG_CT_NEW;
 632                 group = NFNLGRP_CONNTRACK_UPDATE;
 633         } else
 634                 return 0;
 635
 636         net = nf_ct_net(ct);
 637         if (!item->report && !nfnetlink_has_listeners(net, group))
 638                 return 0;
 639
 640         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 641         if (skb == NULL)
 642                 goto errout;
 643
 644         type |= NFNL_SUBSYS_CTNETLINK << 8;
 645         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 646         if (nlh == NULL)
 647                 goto nlmsg_failure;
 648
 649         nfmsg = nlmsg_data(nlh);
 650         nfmsg->nfgen_family = nf_ct_l3num(ct);
 651         nfmsg->version  = NFNETLINK_V0;
 652         nfmsg->res_id   = 0;
 653
 654         rcu_read_lock();
 655         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 656         if (!nest_parms)
 657                 goto nla_put_failure;
 658         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 659                 goto nla_put_failure;
 660         nla_nest_end(skb, nest_parms);
 661
 662         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 663         if (!nest_parms)
 664                 goto nla_put_failure;
 665         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 666                 goto nla_put_failure;
 667         nla_nest_end(skb, nest_parms);
 668
 669         if (nf_ct_zone(ct) &&
 670             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 671                 goto nla_put_failure;
 672
 673         if (ctnetlink_dump_id(skb, ct) < 0)
 674                 goto nla_put_failure;
 675
 676         if (ctnetlink_dump_status(skb, ct) < 0)
 677                 goto nla_put_failure;
 678
 679         if (events & (1 << IPCT_DESTROY)) {
 680                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 681                     ctnetlink_dump_timestamp(skb, ct) < 0)
 682                         goto nla_put_failure;
 683         } else {
 684                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 685                         goto nla_put_failure;
 686
 687                 if (events & (1 << IPCT_PROTOINFO)
 688                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 689                         goto nla_put_failure;
 690
 691                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 692                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 693                         goto nla_put_failure;
 694
 695 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 696                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 697                     && ctnetlink_dump_secctx(skb, ct) < 0)
 698                         goto nla_put_failure;
 699 #endif
 700                 if (events & (1 << IPCT_LABEL) &&
 701                      ctnetlink_dump_labels(skb, ct) < 0)
 702                         goto nla_put_failure;
 703
 704                 if (events & (1 << IPCT_RELATED) &&
 705                     ctnetlink_dump_master(skb, ct) < 0)
 706                         goto nla_put_failure;
 707
 708                 if (events & (1 << IPCT_SEQADJ) &&
 709                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 710                         goto nla_put_failure;
 711         }
 712
 713 #ifdef CONFIG_NF_CONNTRACK_MARK
 714         if ((events & (1 << IPCT_MARK) || ct->mark)
 715             && ctnetlink_dump_mark(skb, ct) < 0)
 716                 goto nla_put_failure;
 717 #endif
 718         rcu_read_unlock();
 719
 720         nlmsg_end(skb, nlh);
 721         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 722                              GFP_ATOMIC);
 723         if (err == -ENOBUFS || err == -EAGAIN)
 724                 return -ENOBUFS;
 725
 726         return 0;
 727
 728 nla_put_failure:
 729         rcu_read_unlock();
 730         nlmsg_cancel(skb, nlh);
 731 nlmsg_failure:
 732         kfree_skb(skb);
 733 errout:
 734         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 735                 return -ENOBUFS;
 736
 737         return 0;
 738 }
 739 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 740
 741 static int ctnetlink_done(struct netlink_callback *cb)
 742 {
 743         if (cb->args[1])
 744                 nf_ct_put((struct nf_conn *)cb->args[1]);
 745         if (cb->data)
 746                 kfree(cb->data);
 747         return 0;
 748 }
 749
 750 struct ctnetlink_dump_filter {
 751         struct {
 752                 u_int32_t val;
 753                 u_int32_t mask;
 754         } mark;
 755 };
 756
 757 static int
 758 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 759 {
 760         struct net *net = sock_net(skb->sk);
 761         struct nf_conn *ct, *last;
 762         struct nf_conntrack_tuple_hash *h;
 763         struct hlist_nulls_node *n;
 764         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 765         u_int8_t l3proto = nfmsg->nfgen_family;
 766         int res;
 767         spinlock_t *lockp;
 768
 769 #ifdef CONFIG_NF_CONNTRACK_MARK
 770         const struct ctnetlink_dump_filter *filter = cb->data;
 771 #endif
 772
 773         last = (struct nf_conn *)cb->args[1];
 774
 775         local_bh_disable();
 776         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 777 restart:
 778                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 779                 spin_lock(lockp);
 780                 if (cb->args[0] >= net->ct.htable_size) {
 781                         spin_unlock(lockp);
 782                         goto out;
 783                 }
 784                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 785                                          hnnode) {
 786                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 787                                 continue;
 788                         ct = nf_ct_tuplehash_to_ctrack(h);
 789                         /* Dump entries of a given L3 protocol number.
 790                          * If it is not specified, ie. l3proto == 0,
 791                          * then dump everything. */
 792                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 793                                 continue;
 794                         if (cb->args[1]) {
 795                                 if (ct != last)
 796                                         continue;
 797                                 cb->args[1] = 0;
 798                         }
 799 #ifdef CONFIG_NF_CONNTRACK_MARK
 800                         if (filter && !((ct->mark & filter->mark.mask) ==
 801                                         filter->mark.val)) {
 802                                 continue;
 803                         }
 804 #endif
 805                         rcu_read_lock();
 806                         res =
 807                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 808                                             cb->nlh->nlmsg_seq,
 809                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 810                                             ct);
 811                         rcu_read_unlock();
 812                         if (res < 0) {
 813                                 nf_conntrack_get(&ct->ct_general);
 814                                 cb->args[1] = (unsigned long)ct;
 815                                 spin_unlock(lockp);
 816                                 goto out;
 817                         }
 818                 }
 819                 spin_unlock(lockp);
 820                 if (cb->args[1]) {
 821                         cb->args[1] = 0;
 822                         goto restart;
 823                 }
 824         }
 825 out:
 826         local_bh_enable();
 827         if (last)
 828                 nf_ct_put(last);
 829
 830         return skb->len;
 831 }
 832
 833 static inline int
 834 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 835 {
 836         struct nlattr *tb[CTA_IP_MAX+1];
 837         struct nf_conntrack_l3proto *l3proto;
 838         int ret = 0;
 839
 840         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 841         if (ret < 0)
 842                 return ret;
 843
 844         rcu_read_lock();
 845         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 846
 847         if (likely(l3proto->nlattr_to_tuple)) {
 848                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 849                                           l3proto->nla_policy);
 850                 if (ret == 0)
 851                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 852         }
 853
 854         rcu_read_unlock();
 855
 856         return ret;
 857 }
 858
 859 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 860         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 861 };
 862
 863 static inline int
 864 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 865                             struct nf_conntrack_tuple *tuple)
 866 {
 867         struct nlattr *tb[CTA_PROTO_MAX+1];
 868         struct nf_conntrack_l4proto *l4proto;
 869         int ret = 0;
 870
 871         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 872         if (ret < 0)
 873                 return ret;
 874
 875         if (!tb[CTA_PROTO_NUM])
 876                 return -EINVAL;
 877         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 878
 879         rcu_read_lock();
 880         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 881
 882         if (likely(l4proto->nlattr_to_tuple)) {
 883                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 884                                           l4proto->nla_policy);
 885                 if (ret == 0)
 886                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 887         }
 888
 889         rcu_read_unlock();
 890
 891         return ret;
 892 }
 893
 894 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 895         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 896         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 897 };
 898
 899 static int
 900 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 901                       struct nf_conntrack_tuple *tuple,
 902                       enum ctattr_type type, u_int8_t l3num)
 903 {
 904         struct nlattr *tb[CTA_TUPLE_MAX+1];
 905         int err;
 906
 907         memset(tuple, 0, sizeof(*tuple));
 908
 909         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 910         if (err < 0)
 911                 return err;
 912
 913         if (!tb[CTA_TUPLE_IP])
 914                 return -EINVAL;
 915
 916         tuple->src.l3num = l3num;
 917
 918         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 919         if (err < 0)
 920                 return err;
 921
 922         if (!tb[CTA_TUPLE_PROTO])
 923                 return -EINVAL;
 924
 925         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 926         if (err < 0)
 927                 return err;
 928
 929         /* orig and expect tuples get DIR_ORIGINAL */
 930         if (type == CTA_TUPLE_REPLY)
 931                 tuple->dst.dir = IP_CT_DIR_REPLY;
 932         else
 933                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 934
 935         return 0;
 936 }
 937
 938 static int
 939 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 940 {
 941         if (attr)
 942 #ifdef CONFIG_NF_CONNTRACK_ZONES
 943                 *zone = ntohs(nla_get_be16(attr));
 944 #else
 945                 return -EOPNOTSUPP;
 946 #endif
 947         else
 948                 *zone = 0;
 949
 950         return 0;
 951 }
 952
 953 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 954         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 955                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 956 };
 957
 958 static inline int
 959 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 960                      struct nlattr **helpinfo)
 961 {
 962         int err;
 963         struct nlattr *tb[CTA_HELP_MAX+1];
 964
 965         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 966         if (err < 0)
 967                 return err;
 968
 969         if (!tb[CTA_HELP_NAME])
 970                 return -EINVAL;
 971
 972         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 973
 974         if (tb[CTA_HELP_INFO])
 975                 *helpinfo = tb[CTA_HELP_INFO];
 976
 977         return 0;
 978 }
 979
 980 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 981         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 982         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 983         [CTA_STATUS]            = { .type = NLA_U32 },
 984         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 985         [CTA_HELP]              = { .type = NLA_NESTED },
 986         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 987         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 988         [CTA_MARK]              = { .type = NLA_U32 },
 989         [CTA_ID]                = { .type = NLA_U32 },
 990         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 991         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 992         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
 993         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
 994         [CTA_ZONE]              = { .type = NLA_U16 },
 995         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 996         [CTA_LABELS]            = { .type = NLA_BINARY,
 997                                     .len = NF_CT_LABELS_MAX_SIZE },
 998         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
 999                                     .len = NF_CT_LABELS_MAX_SIZE },
1000 };
1001
1002 static int
1003 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
1004                         const struct nlmsghdr *nlh,
1005                         const struct nlattr * const cda[])
1006 {
1007         struct net *net = sock_net(ctnl);
1008         struct nf_conntrack_tuple_hash *h;
1009         struct nf_conntrack_tuple tuple;
1010         struct nf_conn *ct;
1011         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1012         u_int8_t u3 = nfmsg->nfgen_family;
1013         u16 zone;
1014         int err;
1015
1016         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1017         if (err < 0)
1018                 return err;
1019
1020         if (cda[CTA_TUPLE_ORIG])
1021                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1022         else if (cda[CTA_TUPLE_REPLY])
1023                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1024         else {
1025                 /* Flush the whole table */
1026                 nf_conntrack_flush_report(net,
1027                                          NETLINK_CB(skb).portid,
1028                                          nlmsg_report(nlh));
1029                 return 0;
1030         }
1031
1032         if (err < 0)
1033                 return err;
1034
1035         h = nf_conntrack_find_get(net, zone, &tuple);
1036         if (!h)
1037                 return -ENOENT;
1038
1039         ct = nf_ct_tuplehash_to_ctrack(h);
1040
1041         if (cda[CTA_ID]) {
1042                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1043                 if (id != (u32)(unsigned long)ct) {
1044                         nf_ct_put(ct);
1045                         return -ENOENT;
1046                 }
1047         }
1048
1049         if (del_timer(&ct->timeout))
1050                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1051
1052         nf_ct_put(ct);
1053
1054         return 0;
1055 }
1056
1057 static int
1058 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1059                         const struct nlmsghdr *nlh,
1060                         const struct nlattr * const cda[])
1061 {
1062         struct net *net = sock_net(ctnl);
1063         struct nf_conntrack_tuple_hash *h;
1064         struct nf_conntrack_tuple tuple;
1065         struct nf_conn *ct;
1066         struct sk_buff *skb2 = NULL;
1067         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1068         u_int8_t u3 = nfmsg->nfgen_family;
1069         u16 zone;
1070         int err;
1071
1072         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1073                 struct netlink_dump_control c = {
1074                         .dump = ctnetlink_dump_table,
1075                         .done = ctnetlink_done,
1076                 };
1077 #ifdef CONFIG_NF_CONNTRACK_MARK
1078                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1079                         struct ctnetlink_dump_filter *filter;
1080
1081                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1082                                          GFP_ATOMIC);
1083                         if (filter == NULL)
1084                                 return -ENOMEM;
1085
1086                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1087                         filter->mark.mask =
1088                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1089                         c.data = filter;
1090                 }
1091 #endif
1092                 return netlink_dump_start(ctnl, skb, nlh, &c);
1093         }
1094
1095         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1096         if (err < 0)
1097                 return err;
1098
1099         if (cda[CTA_TUPLE_ORIG])
1100                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1101         else if (cda[CTA_TUPLE_REPLY])
1102                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1103         else
1104                 return -EINVAL;
1105
1106         if (err < 0)
1107                 return err;
1108
1109         h = nf_conntrack_find_get(net, zone, &tuple);
1110         if (!h)
1111                 return -ENOENT;
1112
1113         ct = nf_ct_tuplehash_to_ctrack(h);
1114
1115         err = -ENOMEM;
1116         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1117         if (skb2 == NULL) {
1118                 nf_ct_put(ct);
1119                 return -ENOMEM;
1120         }
1121
1122         rcu_read_lock();
1123         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1124                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1125         rcu_read_unlock();
1126         nf_ct_put(ct);
1127         if (err <= 0)
1128                 goto free;
1129
1130         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1131         if (err < 0)
1132                 goto out;
1133
1134         return 0;
1135
1136 free:
1137         kfree_skb(skb2);
1138 out:
1139         /* this avoids a loop in nfnetlink. */
1140         return err == -EAGAIN ? -ENOBUFS : err;
1141 }
1142
1143 static int ctnetlink_done_list(struct netlink_callback *cb)
1144 {
1145         if (cb->args[1])
1146                 nf_ct_put((struct nf_conn *)cb->args[1]);
1147         return 0;
1148 }
1149
1150 static int
1151 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1152 {
1153         struct nf_conn *ct, *last = NULL;
1154         struct nf_conntrack_tuple_hash *h;
1155         struct hlist_nulls_node *n;
1156         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1157         u_int8_t l3proto = nfmsg->nfgen_family;
1158         int res;
1159         int cpu;
1160         struct hlist_nulls_head *list;
1161         struct net *net = sock_net(skb->sk);
1162
1163         if (cb->args[2])
1164                 return 0;
1165
1166         if (cb->args[0] == nr_cpu_ids)
1167                 return 0;
1168
1169         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1170                 struct ct_pcpu *pcpu;
1171
1172                 if (!cpu_possible(cpu))
1173                         continue;
1174
1175                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1176                 spin_lock_bh(&pcpu->lock);
1177                 last = (struct nf_conn *)cb->args[1];
1178                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1179 restart:
1180                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1181                         ct = nf_ct_tuplehash_to_ctrack(h);
1182                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1183                                 continue;
1184                         if (cb->args[1]) {
1185                                 if (ct != last)
1186                                         continue;
1187                                 cb->args[1] = 0;
1188                         }
1189                         rcu_read_lock();
1190                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1191                                                   cb->nlh->nlmsg_seq,
1192                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1193                                                   ct);
1194                         rcu_read_unlock();
1195                         if (res < 0) {
1196                                 nf_conntrack_get(&ct->ct_general);
1197                                 cb->args[1] = (unsigned long)ct;
1198                                 spin_unlock_bh(&pcpu->lock);
1199                                 goto out;
1200                         }
1201                 }
1202                 if (cb->args[1]) {
1203                         cb->args[1] = 0;
1204                         goto restart;
1205                 } else
1206                         cb->args[2] = 1;
1207                 spin_unlock_bh(&pcpu->lock);
1208         }
1209 out:
1210         if (last)
1211                 nf_ct_put(last);
1212
1213         return skb->len;
1214 }
1215
1216 static int
1217 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1218 {
1219         return ctnetlink_dump_list(skb, cb, true);
1220 }
1221
1222 static int
1223 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1224                        const struct nlmsghdr *nlh,
1225                        const struct nlattr * const cda[])
1226 {
1227         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1228                 struct netlink_dump_control c = {
1229                         .dump = ctnetlink_dump_dying,
1230                         .done = ctnetlink_done_list,
1231                 };
1232                 return netlink_dump_start(ctnl, skb, nlh, &c);
1233         }
1234
1235         return -EOPNOTSUPP;
1236 }
1237
1238 static int
1239 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1240 {
1241         return ctnetlink_dump_list(skb, cb, false);
1242 }
1243
1244 static int
1245 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1246                              const struct nlmsghdr *nlh,
1247                              const struct nlattr * const cda[])
1248 {
1249         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1250                 struct netlink_dump_control c = {
1251                         .dump = ctnetlink_dump_unconfirmed,
1252                         .done = ctnetlink_done_list,
1253                 };
1254                 return netlink_dump_start(ctnl, skb, nlh, &c);
1255         }
1256
1257         return -EOPNOTSUPP;
1258 }
1259
1260 #ifdef CONFIG_NF_NAT_NEEDED
1261 static int
1262 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1263                           enum nf_nat_manip_type manip,
1264                           const struct nlattr *attr)
1265 {
1266         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1267         int err;
1268
1269         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1270         if (!parse_nat_setup) {
1271 #ifdef CONFIG_MODULES
1272                 rcu_read_unlock();
1273                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1274                 if (request_module("nf-nat") < 0) {
1275                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1276                         rcu_read_lock();
1277                         return -EOPNOTSUPP;
1278                 }
1279                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1280                 rcu_read_lock();
1281                 if (nfnetlink_parse_nat_setup_hook)
1282                         return -EAGAIN;
1283 #endif
1284                 return -EOPNOTSUPP;
1285         }
1286
1287         err = parse_nat_setup(ct, manip, attr);
1288         if (err == -EAGAIN) {
1289 #ifdef CONFIG_MODULES
1290                 rcu_read_unlock();
1291                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1292                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1293                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1294                         rcu_read_lock();
1295                         return -EOPNOTSUPP;
1296                 }
1297                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1298                 rcu_read_lock();
1299 #else
1300                 err = -EOPNOTSUPP;
1301 #endif
1302         }
1303         return err;
1304 }
1305 #endif
1306
1307 static int
1308 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1309 {
1310         unsigned long d;
1311         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1312         d = ct->status ^ status;
1313
1314         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1315                 /* unchangeable */
1316                 return -EBUSY;
1317
1318         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1319                 /* SEEN_REPLY bit can only be set */
1320                 return -EBUSY;
1321
1322         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1323                 /* ASSURED bit can only be set */
1324                 return -EBUSY;
1325
1326         /* Be careful here, modifying NAT bits can screw up things,
1327          * so don't let users modify them directly if they don't pass
1328          * nf_nat_range. */
1329         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1330         return 0;
1331 }
1332
1333 static int
1334 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1335 {
1336 #ifdef CONFIG_NF_NAT_NEEDED
1337         int ret;
1338
1339         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1340                                         cda[CTA_NAT_DST]);
1341         if (ret < 0)
1342                 return ret;
1343
1344         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1345                                         cda[CTA_NAT_SRC]);
1346         return ret;
1347 #else
1348         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1349                 return 0;
1350         return -EOPNOTSUPP;
1351 #endif
1352 }
1353
1354 static inline int
1355 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1356 {
1357         struct nf_conntrack_helper *helper;
1358         struct nf_conn_help *help = nfct_help(ct);
1359         char *helpname = NULL;
1360         struct nlattr *helpinfo = NULL;
1361         int err;
1362
1363         /* don't change helper of sibling connections */
1364         if (ct->master)
1365                 return -EBUSY;
1366
1367         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1368         if (err < 0)
1369                 return err;
1370
1371         if (!strcmp(helpname, "")) {
1372                 if (help && help->helper) {
1373                         /* we had a helper before ... */
1374                         nf_ct_remove_expectations(ct);
1375                         RCU_INIT_POINTER(help->helper, NULL);
1376                 }
1377
1378                 return 0;
1379         }
1380
1381         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1382                                             nf_ct_protonum(ct));
1383         if (helper == NULL) {
1384 #ifdef CONFIG_MODULES
1385                 spin_unlock_bh(&nf_conntrack_expect_lock);
1386
1387                 if (request_module("nfct-helper-%s", helpname) < 0) {
1388                         spin_lock_bh(&nf_conntrack_expect_lock);
1389                         return -EOPNOTSUPP;
1390                 }
1391
1392                 spin_lock_bh(&nf_conntrack_expect_lock);
1393                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1394                                                     nf_ct_protonum(ct));
1395                 if (helper)
1396                         return -EAGAIN;
1397 #endif
1398                 return -EOPNOTSUPP;
1399         }
1400
1401         if (help) {
1402                 if (help->helper == helper) {
1403                         /* update private helper data if allowed. */
1404                         if (helper->from_nlattr)
1405                                 helper->from_nlattr(helpinfo, ct);
1406                         return 0;
1407                 } else
1408                         return -EBUSY;
1409         }
1410
1411         /* we cannot set a helper for an existing conntrack */
1412         return -EOPNOTSUPP;
1413 }
1414
1415 static inline int
1416 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1417 {
1418         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1419
1420         if (!del_timer(&ct->timeout))
1421                 return -ETIME;
1422
1423         ct->timeout.expires = jiffies + timeout * HZ;
1424         add_timer(&ct->timeout);
1425
1426         return 0;
1427 }
1428
1429 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1430         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1431         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1432         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1433 };
1434
1435 static inline int
1436 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1437 {
1438         const struct nlattr *attr = cda[CTA_PROTOINFO];
1439         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1440         struct nf_conntrack_l4proto *l4proto;
1441         int err = 0;
1442
1443         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1444         if (err < 0)
1445                 return err;
1446
1447         rcu_read_lock();
1448         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1449         if (l4proto->from_nlattr)
1450                 err = l4proto->from_nlattr(tb, ct);
1451         rcu_read_unlock();
1452
1453         return err;
1454 }
1455
1456 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1457         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1458         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1459         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1460 };
1461
1462 static inline int
1463 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1464 {
1465         int err;
1466         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1467
1468         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1469         if (err < 0)
1470                 return err;
1471
1472         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1473                 return -EINVAL;
1474
1475         seq->correction_pos =
1476                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1477
1478         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1479                 return -EINVAL;
1480
1481         seq->offset_before =
1482                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1483
1484         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1485                 return -EINVAL;
1486
1487         seq->offset_after =
1488                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1489
1490         return 0;
1491 }
1492
1493 static int
1494 ctnetlink_change_seq_adj(struct nf_conn *ct,
1495                          const struct nlattr * const cda[])
1496 {
1497         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1498         int ret = 0;
1499
1500         if (!seqadj)
1501                 return 0;
1502
1503         if (cda[CTA_SEQ_ADJ_ORIG]) {
1504                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1505                                      cda[CTA_SEQ_ADJ_ORIG]);
1506                 if (ret < 0)
1507                         return ret;
1508
1509                 ct->status |= IPS_SEQ_ADJUST;
1510         }
1511
1512         if (cda[CTA_SEQ_ADJ_REPLY]) {
1513                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1514                                      cda[CTA_SEQ_ADJ_REPLY]);
1515                 if (ret < 0)
1516                         return ret;
1517
1518                 ct->status |= IPS_SEQ_ADJUST;
1519         }
1520
1521         return 0;
1522 }
1523
1524 static int
1525 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1526 {
1527 #ifdef CONFIG_NF_CONNTRACK_LABELS
1528         size_t len = nla_len(cda[CTA_LABELS]);
1529         const void *mask = cda[CTA_LABELS_MASK];
1530
1531         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1532                 return -EINVAL;
1533
1534         if (mask) {
1535                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1536                     nla_len(cda[CTA_LABELS_MASK]) != len)
1537                         return -EINVAL;
1538                 mask = nla_data(cda[CTA_LABELS_MASK]);
1539         }
1540
1541         len /= sizeof(u32);
1542
1543         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1544 #else
1545         return -EOPNOTSUPP;
1546 #endif
1547 }
1548
1549 static int
1550 ctnetlink_change_conntrack(struct nf_conn *ct,
1551                            const struct nlattr * const cda[])
1552 {
1553         int err;
1554
1555         /* only allow NAT changes and master assignation for new conntracks */
1556         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1557                 return -EOPNOTSUPP;
1558
1559         if (cda[CTA_HELP]) {
1560                 err = ctnetlink_change_helper(ct, cda);
1561                 if (err < 0)
1562                         return err;
1563         }
1564
1565         if (cda[CTA_TIMEOUT]) {
1566                 err = ctnetlink_change_timeout(ct, cda);
1567                 if (err < 0)
1568                         return err;
1569         }
1570
1571         if (cda[CTA_STATUS]) {
1572                 err = ctnetlink_change_status(ct, cda);
1573                 if (err < 0)
1574                         return err;
1575         }
1576
1577         if (cda[CTA_PROTOINFO]) {
1578                 err = ctnetlink_change_protoinfo(ct, cda);
1579                 if (err < 0)
1580                         return err;
1581         }
1582
1583 #if defined(CONFIG_NF_CONNTRACK_MARK)
1584         if (cda[CTA_MARK])
1585                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1586 #endif
1587
1588         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1589                 err = ctnetlink_change_seq_adj(ct, cda);
1590                 if (err < 0)
1591                         return err;
1592         }
1593
1594         if (cda[CTA_LABELS]) {
1595                 err = ctnetlink_attach_labels(ct, cda);
1596                 if (err < 0)
1597                         return err;
1598         }
1599
1600         return 0;
1601 }
1602
1603 static struct nf_conn *
1604 ctnetlink_create_conntrack(struct net *net, u16 zone,
1605                            const struct nlattr * const cda[],
1606                            struct nf_conntrack_tuple *otuple,
1607                            struct nf_conntrack_tuple *rtuple,
1608                            u8 u3)
1609 {
1610         struct nf_conn *ct;
1611         int err = -EINVAL;
1612         struct nf_conntrack_helper *helper;
1613         struct nf_conn_tstamp *tstamp;
1614
1615         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1616         if (IS_ERR(ct))
1617                 return ERR_PTR(-ENOMEM);
1618
1619         if (!cda[CTA_TIMEOUT])
1620                 goto err1;
1621         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1622
1623         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1624
1625         rcu_read_lock();
1626         if (cda[CTA_HELP]) {
1627                 char *helpname = NULL;
1628                 struct nlattr *helpinfo = NULL;
1629
1630                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1631                 if (err < 0)
1632                         goto err2;
1633
1634                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1635                                                     nf_ct_protonum(ct));
1636                 if (helper == NULL) {
1637                         rcu_read_unlock();
1638 #ifdef CONFIG_MODULES
1639                         if (request_module("nfct-helper-%s", helpname) < 0) {
1640                                 err = -EOPNOTSUPP;
1641                                 goto err1;
1642                         }
1643
1644                         rcu_read_lock();
1645                         helper = __nf_conntrack_helper_find(helpname,
1646                                                             nf_ct_l3num(ct),
1647                                                             nf_ct_protonum(ct));
1648                         if (helper) {
1649                                 err = -EAGAIN;
1650                                 goto err2;
1651                         }
1652                         rcu_read_unlock();
1653 #endif
1654                         err = -EOPNOTSUPP;
1655                         goto err1;
1656                 } else {
1657                         struct nf_conn_help *help;
1658
1659                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1660                         if (help == NULL) {
1661                                 err = -ENOMEM;
1662                                 goto err2;
1663                         }
1664                         /* set private helper data if allowed. */
1665                         if (helper->from_nlattr)
1666                                 helper->from_nlattr(helpinfo, ct);
1667
1668                         /* not in hash table yet so not strictly necessary */
1669                         RCU_INIT_POINTER(help->helper, helper);
1670                 }
1671         } else {
1672                 /* try an implicit helper assignation */
1673                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1674                 if (err < 0)
1675                         goto err2;
1676         }
1677
1678         err = ctnetlink_setup_nat(ct, cda);
1679         if (err < 0)
1680                 goto err2;
1681
1682         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1683         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1684         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1685         nf_ct_labels_ext_add(ct);
1686
1687         /* we must add conntrack extensions before confirmation. */
1688         ct->status |= IPS_CONFIRMED;
1689
1690         if (cda[CTA_STATUS]) {
1691                 err = ctnetlink_change_status(ct, cda);
1692                 if (err < 0)
1693                         goto err2;
1694         }
1695
1696         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1697                 err = ctnetlink_change_seq_adj(ct, cda);
1698                 if (err < 0)
1699                         goto err2;
1700         }
1701
1702         memset(&ct->proto, 0, sizeof(ct->proto));
1703         if (cda[CTA_PROTOINFO]) {
1704                 err = ctnetlink_change_protoinfo(ct, cda);
1705                 if (err < 0)
1706                         goto err2;
1707         }
1708
1709 #if defined(CONFIG_NF_CONNTRACK_MARK)
1710         if (cda[CTA_MARK])
1711                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1712 #endif
1713
1714         /* setup master conntrack: this is a confirmed expectation */
1715         if (cda[CTA_TUPLE_MASTER]) {
1716                 struct nf_conntrack_tuple master;
1717                 struct nf_conntrack_tuple_hash *master_h;
1718                 struct nf_conn *master_ct;
1719
1720                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1721                 if (err < 0)
1722                         goto err2;
1723
1724                 master_h = nf_conntrack_find_get(net, zone, &master);
1725                 if (master_h == NULL) {
1726                         err = -ENOENT;
1727                         goto err2;
1728                 }
1729                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1730                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1731                 ct->master = master_ct;
1732         }
1733         tstamp = nf_conn_tstamp_find(ct);
1734         if (tstamp)
1735                 tstamp->start = ktime_to_ns(ktime_get_real());
1736
1737         err = nf_conntrack_hash_check_insert(ct);
1738         if (err < 0)
1739                 goto err2;
1740
1741         rcu_read_unlock();
1742
1743         return ct;
1744
1745 err2:
1746         rcu_read_unlock();
1747 err1:
1748         nf_conntrack_free(ct);
1749         return ERR_PTR(err);
1750 }
1751
1752 static int
1753 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1754                         const struct nlmsghdr *nlh,
1755                         const struct nlattr * const cda[])
1756 {
1757         struct net *net = sock_net(ctnl);
1758         struct nf_conntrack_tuple otuple, rtuple;
1759         struct nf_conntrack_tuple_hash *h = NULL;
1760         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1761         struct nf_conn *ct;
1762         u_int8_t u3 = nfmsg->nfgen_family;
1763         u16 zone;
1764         int err;
1765
1766         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1767         if (err < 0)
1768                 return err;
1769
1770         if (cda[CTA_TUPLE_ORIG]) {
1771                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1772                 if (err < 0)
1773                         return err;
1774         }
1775
1776         if (cda[CTA_TUPLE_REPLY]) {
1777                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1778                 if (err < 0)
1779                         return err;
1780         }
1781
1782         if (cda[CTA_TUPLE_ORIG])
1783                 h = nf_conntrack_find_get(net, zone, &otuple);
1784         else if (cda[CTA_TUPLE_REPLY])
1785                 h = nf_conntrack_find_get(net, zone, &rtuple);
1786
1787         if (h == NULL) {
1788                 err = -ENOENT;
1789                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1790                         enum ip_conntrack_events events;
1791
1792                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1793                                 return -EINVAL;
1794
1795                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1796                                                         &rtuple, u3);
1797                         if (IS_ERR(ct))
1798                                 return PTR_ERR(ct);
1799
1800                         err = 0;
1801                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1802                                 events = IPCT_RELATED;
1803                         else
1804                                 events = IPCT_NEW;
1805
1806                         if (cda[CTA_LABELS] &&
1807                             ctnetlink_attach_labels(ct, cda) == 0)
1808                                 events |= (1 << IPCT_LABEL);
1809
1810                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1811                                                       (1 << IPCT_ASSURED) |
1812                                                       (1 << IPCT_HELPER) |
1813                                                       (1 << IPCT_PROTOINFO) |
1814                                                       (1 << IPCT_SEQADJ) |
1815                                                       (1 << IPCT_MARK) | events,
1816                                                       ct, NETLINK_CB(skb).portid,
1817                                                       nlmsg_report(nlh));
1818                         nf_ct_put(ct);
1819                 }
1820
1821                 return err;
1822         }
1823         /* implicit 'else' */
1824
1825         err = -EEXIST;
1826         ct = nf_ct_tuplehash_to_ctrack(h);
1827         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1828                 spin_lock_bh(&nf_conntrack_expect_lock);
1829                 err = ctnetlink_change_conntrack(ct, cda);
1830                 spin_unlock_bh(&nf_conntrack_expect_lock);
1831                 if (err == 0) {
1832                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1833                                                       (1 << IPCT_ASSURED) |
1834                                                       (1 << IPCT_HELPER) |
1835                                                       (1 << IPCT_LABEL) |
1836                                                       (1 << IPCT_PROTOINFO) |
1837                                                       (1 << IPCT_SEQADJ) |
1838                                                       (1 << IPCT_MARK),
1839                                                       ct, NETLINK_CB(skb).portid,
1840                                                       nlmsg_report(nlh));
1841                 }
1842         }
1843
1844         nf_ct_put(ct);
1845         return err;
1846 }
1847
1848 static int
1849 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1850                                 __u16 cpu, const struct ip_conntrack_stat *st)
1851 {
1852         struct nlmsghdr *nlh;
1853         struct nfgenmsg *nfmsg;
1854         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1855
1856         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1857         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1858         if (nlh == NULL)
1859                 goto nlmsg_failure;
1860
1861         nfmsg = nlmsg_data(nlh);
1862         nfmsg->nfgen_family = AF_UNSPEC;
1863         nfmsg->version      = NFNETLINK_V0;
1864         nfmsg->res_id       = htons(cpu);
1865
1866         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1867             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1868             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1869             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1870             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1871             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1872             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1873             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1874             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1875                                 htonl(st->insert_failed)) ||
1876             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1877             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1878             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1879             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1880                                 htonl(st->search_restart)))
1881                 goto nla_put_failure;
1882
1883         nlmsg_end(skb, nlh);
1884         return skb->len;
1885
1886 nla_put_failure:
1887 nlmsg_failure:
1888         nlmsg_cancel(skb, nlh);
1889         return -1;
1890 }
1891
1892 static int
1893 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1894 {
1895         int cpu;
1896         struct net *net = sock_net(skb->sk);
1897
1898         if (cb->args[0] == nr_cpu_ids)
1899                 return 0;
1900
1901         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1902                 const struct ip_conntrack_stat *st;
1903
1904                 if (!cpu_possible(cpu))
1905                         continue;
1906
1907                 st = per_cpu_ptr(net->ct.stat, cpu);
1908                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1909                                                     NETLINK_CB(cb->skb).portid,
1910                                                     cb->nlh->nlmsg_seq,
1911                                                     cpu, st) < 0)
1912                                 break;
1913         }
1914         cb->args[0] = cpu;
1915
1916         return skb->len;
1917 }
1918
1919 static int
1920 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1921                       const struct nlmsghdr *nlh,
1922                       const struct nlattr * const cda[])
1923 {
1924         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1925                 struct netlink_dump_control c = {
1926                         .dump = ctnetlink_ct_stat_cpu_dump,
1927                 };
1928                 return netlink_dump_start(ctnl, skb, nlh, &c);
1929         }
1930
1931         return 0;
1932 }
1933
1934 static int
1935 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1936                             struct net *net)
1937 {
1938         struct nlmsghdr *nlh;
1939         struct nfgenmsg *nfmsg;
1940         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1941         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1942
1943         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1944         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1945         if (nlh == NULL)
1946                 goto nlmsg_failure;
1947
1948         nfmsg = nlmsg_data(nlh);
1949         nfmsg->nfgen_family = AF_UNSPEC;
1950         nfmsg->version      = NFNETLINK_V0;
1951         nfmsg->res_id       = 0;
1952
1953         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1954                 goto nla_put_failure;
1955
1956         nlmsg_end(skb, nlh);
1957         return skb->len;
1958
1959 nla_put_failure:
1960 nlmsg_failure:
1961         nlmsg_cancel(skb, nlh);
1962         return -1;
1963 }
1964
1965 static int
1966 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1967                   const struct nlmsghdr *nlh,
1968                   const struct nlattr * const cda[])
1969 {
1970         struct sk_buff *skb2;
1971         int err;
1972
1973         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1974         if (skb2 == NULL)
1975                 return -ENOMEM;
1976
1977         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1978                                           nlh->nlmsg_seq,
1979                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1980                                           sock_net(skb->sk));
1981         if (err <= 0)
1982                 goto free;
1983
1984         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1985         if (err < 0)
1986                 goto out;
1987
1988         return 0;
1989
1990 free:
1991         kfree_skb(skb2);
1992 out:
1993         /* this avoids a loop in nfnetlink. */
1994         return err == -EAGAIN ? -ENOBUFS : err;
1995 }
1996
1997 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
1998         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
1999         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2000         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2001         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2002         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2003         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2004                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2005         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2006         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2007         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2008         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2009         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2010 };
2011
2012 static struct nf_conntrack_expect *
2013 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2014                        struct nf_conntrack_helper *helper,
2015                        struct nf_conntrack_tuple *tuple,
2016                        struct nf_conntrack_tuple *mask);
2017
2018 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2019 static size_t
2020 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2021 {
2022         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2023                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2024                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2025                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2026                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2027                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2028                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2029                + nla_total_size(0) /* CTA_PROTOINFO */
2030                + nla_total_size(0) /* CTA_HELP */
2031                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2032                + ctnetlink_secctx_size(ct)
2033 #ifdef CONFIG_NF_NAT_NEEDED
2034                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2035                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2036 #endif
2037 #ifdef CONFIG_NF_CONNTRACK_MARK
2038                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2039 #endif
2040                + ctnetlink_proto_size(ct)
2041                ;
2042 }
2043
2044 static int
2045 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2046 {
2047         struct nlattr *nest_parms;
2048
2049         rcu_read_lock();
2050         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2051         if (!nest_parms)
2052                 goto nla_put_failure;
2053         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2054                 goto nla_put_failure;
2055         nla_nest_end(skb, nest_parms);
2056
2057         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2058         if (!nest_parms)
2059                 goto nla_put_failure;
2060         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2061                 goto nla_put_failure;
2062         nla_nest_end(skb, nest_parms);
2063
2064         if (nf_ct_zone(ct)) {
2065                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2066                         goto nla_put_failure;
2067         }
2068
2069         if (ctnetlink_dump_id(skb, ct) < 0)
2070                 goto nla_put_failure;
2071
2072         if (ctnetlink_dump_status(skb, ct) < 0)
2073                 goto nla_put_failure;
2074
2075         if (ctnetlink_dump_timeout(skb, ct) < 0)
2076                 goto nla_put_failure;
2077
2078         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2079                 goto nla_put_failure;
2080
2081         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2082                 goto nla_put_failure;
2083
2084 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2085         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2086                 goto nla_put_failure;
2087 #endif
2088         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2089                 goto nla_put_failure;
2090
2091         if ((ct->status & IPS_SEQ_ADJUST) &&
2092             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2093                 goto nla_put_failure;
2094
2095 #ifdef CONFIG_NF_CONNTRACK_MARK
2096         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2097                 goto nla_put_failure;
2098 #endif
2099         if (ctnetlink_dump_labels(skb, ct) < 0)
2100                 goto nla_put_failure;
2101         rcu_read_unlock();
2102         return 0;
2103
2104 nla_put_failure:
2105         rcu_read_unlock();
2106         return -ENOSPC;
2107 }
2108
2109 static int
2110 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2111 {
2112         int err;
2113
2114         if (cda[CTA_TIMEOUT]) {
2115                 err = ctnetlink_change_timeout(ct, cda);
2116                 if (err < 0)
2117                         return err;
2118         }
2119         if (cda[CTA_STATUS]) {
2120                 err = ctnetlink_change_status(ct, cda);
2121                 if (err < 0)
2122                         return err;
2123         }
2124         if (cda[CTA_HELP]) {
2125                 err = ctnetlink_change_helper(ct, cda);
2126                 if (err < 0)
2127                         return err;
2128         }
2129         if (cda[CTA_LABELS]) {
2130                 err = ctnetlink_attach_labels(ct, cda);
2131                 if (err < 0)
2132                         return err;
2133         }
2134 #if defined(CONFIG_NF_CONNTRACK_MARK)
2135         if (cda[CTA_MARK]) {
2136                 u32 mask = 0, mark, newmark;
2137                 if (cda[CTA_MARK_MASK])
2138                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2139
2140                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2141                 newmark = (ct->mark & mask) ^ mark;
2142                 if (newmark != ct->mark)
2143                         ct->mark = newmark;
2144         }
2145 #endif
2146         return 0;
2147 }
2148
2149 static int
2150 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2151 {
2152         struct nlattr *cda[CTA_MAX+1];
2153         int ret;
2154
2155         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2156         if (ret < 0)
2157                 return ret;
2158
2159         spin_lock_bh(&nf_conntrack_expect_lock);
2160         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2161         spin_unlock_bh(&nf_conntrack_expect_lock);
2162
2163         return ret;
2164 }
2165
2166 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2167                                        const struct nf_conn *ct,
2168                                        struct nf_conntrack_tuple *tuple,
2169                                        struct nf_conntrack_tuple *mask)
2170 {
2171         int err;
2172
2173         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2174                                     nf_ct_l3num(ct));
2175         if (err < 0)
2176                 return err;
2177
2178         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2179                                      nf_ct_l3num(ct));
2180 }
2181
2182 static int
2183 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2184                                 u32 portid, u32 report)
2185 {
2186         struct nlattr *cda[CTA_EXPECT_MAX+1];
2187         struct nf_conntrack_tuple tuple, mask;
2188         struct nf_conntrack_helper *helper = NULL;
2189         struct nf_conntrack_expect *exp;
2190         int err;
2191
2192         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2193         if (err < 0)
2194                 return err;
2195
2196         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2197                                           ct, &tuple, &mask);
2198         if (err < 0)
2199                 return err;
2200
2201         if (cda[CTA_EXPECT_HELP_NAME]) {
2202                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2203
2204                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2205                                                     nf_ct_protonum(ct));
2206                 if (helper == NULL)
2207                         return -EOPNOTSUPP;
2208         }
2209
2210         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2211                                      helper, &tuple, &mask);
2212         if (IS_ERR(exp))
2213                 return PTR_ERR(exp);
2214
2215         err = nf_ct_expect_related_report(exp, portid, report);
2216         if (err < 0) {
2217                 nf_ct_expect_put(exp);
2218                 return err;
2219         }
2220
2221         return 0;
2222 }
2223
2224 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2225         .build_size     = ctnetlink_nfqueue_build_size,
2226         .build          = ctnetlink_nfqueue_build,
2227         .parse          = ctnetlink_nfqueue_parse,
2228         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2229         .seq_adjust     = nf_ct_tcp_seqadj_set,
2230 };
2231 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2232
2233 /***********************************************************************
2234  * EXPECT
2235  ***********************************************************************/
2236
2237 static inline int
2238 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2239                          const struct nf_conntrack_tuple *tuple,
2240                          enum ctattr_expect type)
2241 {
2242         struct nlattr *nest_parms;
2243
2244         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2245         if (!nest_parms)
2246                 goto nla_put_failure;
2247         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2248                 goto nla_put_failure;
2249         nla_nest_end(skb, nest_parms);
2250
2251         return 0;
2252
2253 nla_put_failure:
2254         return -1;
2255 }
2256
2257 static inline int
2258 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2259                         const struct nf_conntrack_tuple *tuple,
2260                         const struct nf_conntrack_tuple_mask *mask)
2261 {
2262         int ret;
2263         struct nf_conntrack_l3proto *l3proto;
2264         struct nf_conntrack_l4proto *l4proto;
2265         struct nf_conntrack_tuple m;
2266         struct nlattr *nest_parms;
2267
2268         memset(&m, 0xFF, sizeof(m));
2269         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2270         m.src.u.all = mask->src.u.all;
2271         m.dst.protonum = tuple->dst.protonum;
2272
2273         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2274         if (!nest_parms)
2275                 goto nla_put_failure;
2276
2277         rcu_read_lock();
2278         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2279         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2280         if (ret >= 0) {
2281                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2282                                                tuple->dst.protonum);
2283         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2284         }
2285         rcu_read_unlock();
2286
2287         if (unlikely(ret < 0))
2288                 goto nla_put_failure;
2289
2290         nla_nest_end(skb, nest_parms);
2291
2292         return 0;
2293
2294 nla_put_failure:
2295         return -1;
2296 }
2297
2298 static const union nf_inet_addr any_addr;
2299
2300 static int
2301 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2302                           const struct nf_conntrack_expect *exp)
2303 {
2304         struct nf_conn *master = exp->master;
2305         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2306         struct nf_conn_help *help;
2307 #ifdef CONFIG_NF_NAT_NEEDED
2308         struct nlattr *nest_parms;
2309         struct nf_conntrack_tuple nat_tuple = {};
2310 #endif
2311         struct nf_ct_helper_expectfn *expfn;
2312
2313         if (timeout < 0)
2314                 timeout = 0;
2315
2316         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2317                 goto nla_put_failure;
2318         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2319                 goto nla_put_failure;
2320         if (ctnetlink_exp_dump_tuple(skb,
2321                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2322                                  CTA_EXPECT_MASTER) < 0)
2323                 goto nla_put_failure;
2324
2325 #ifdef CONFIG_NF_NAT_NEEDED
2326         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2327             exp->saved_proto.all) {
2328                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2329                 if (!nest_parms)
2330                         goto nla_put_failure;
2331
2332                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2333                         goto nla_put_failure;
2334
2335                 nat_tuple.src.l3num = nf_ct_l3num(master);
2336                 nat_tuple.src.u3 = exp->saved_addr;
2337                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2338                 nat_tuple.src.u = exp->saved_proto;
2339
2340                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2341                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2342                         goto nla_put_failure;
2343                 nla_nest_end(skb, nest_parms);
2344         }
2345 #endif
2346         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2347             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2348             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2349             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2350                 goto nla_put_failure;
2351         help = nfct_help(master);
2352         if (help) {
2353                 struct nf_conntrack_helper *helper;
2354
2355                 helper = rcu_dereference(help->helper);
2356                 if (helper &&
2357                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2358                         goto nla_put_failure;
2359         }
2360         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2361         if (expfn != NULL &&
2362             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2363                 goto nla_put_failure;
2364
2365         return 0;
2366
2367 nla_put_failure:
2368         return -1;
2369 }
2370
2371 static int
2372 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2373                         int event, const struct nf_conntrack_expect *exp)
2374 {
2375         struct nlmsghdr *nlh;
2376         struct nfgenmsg *nfmsg;
2377         unsigned int flags = portid ? NLM_F_MULTI : 0;
2378
2379         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2380         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2381         if (nlh == NULL)
2382                 goto nlmsg_failure;
2383
2384         nfmsg = nlmsg_data(nlh);
2385         nfmsg->nfgen_family = exp->tuple.src.l3num;
2386         nfmsg->version      = NFNETLINK_V0;
2387         nfmsg->res_id       = 0;
2388
2389         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2390                 goto nla_put_failure;
2391
2392         nlmsg_end(skb, nlh);
2393         return skb->len;
2394
2395 nlmsg_failure:
2396 nla_put_failure:
2397         nlmsg_cancel(skb, nlh);
2398         return -1;
2399 }
2400
2401 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2402 static int
2403 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2404 {
2405         struct nf_conntrack_expect *exp = item->exp;
2406         struct net *net = nf_ct_exp_net(exp);
2407         struct nlmsghdr *nlh;
2408         struct nfgenmsg *nfmsg;
2409         struct sk_buff *skb;
2410         unsigned int type, group;
2411         int flags = 0;
2412
2413         if (events & (1 << IPEXP_DESTROY)) {
2414                 type = IPCTNL_MSG_EXP_DELETE;
2415                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2416         } else if (events & (1 << IPEXP_NEW)) {
2417                 type = IPCTNL_MSG_EXP_NEW;
2418                 flags = NLM_F_CREATE|NLM_F_EXCL;
2419                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2420         } else
2421                 return 0;
2422
2423         if (!item->report && !nfnetlink_has_listeners(net, group))
2424                 return 0;
2425
2426         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2427         if (skb == NULL)
2428                 goto errout;
2429
2430         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2431         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2432         if (nlh == NULL)
2433                 goto nlmsg_failure;
2434
2435         nfmsg = nlmsg_data(nlh);
2436         nfmsg->nfgen_family = exp->tuple.src.l3num;
2437         nfmsg->version      = NFNETLINK_V0;
2438         nfmsg->res_id       = 0;
2439
2440         rcu_read_lock();
2441         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2442                 goto nla_put_failure;
2443         rcu_read_unlock();
2444
2445         nlmsg_end(skb, nlh);
2446         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2447         return 0;
2448
2449 nla_put_failure:
2450         rcu_read_unlock();
2451         nlmsg_cancel(skb, nlh);
2452 nlmsg_failure:
2453         kfree_skb(skb);
2454 errout:
2455         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2456         return 0;
2457 }
2458 #endif
2459 static int ctnetlink_exp_done(struct netlink_callback *cb)
2460 {
2461         if (cb->args[1])
2462                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2463         return 0;
2464 }
2465
2466 static int
2467 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2468 {
2469         struct net *net = sock_net(skb->sk);
2470         struct nf_conntrack_expect *exp, *last;
2471         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2472         u_int8_t l3proto = nfmsg->nfgen_family;
2473
2474         rcu_read_lock();
2475         last = (struct nf_conntrack_expect *)cb->args[1];
2476         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2477 restart:
2478                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2479                                      hnode) {
2480                         if (l3proto && exp->tuple.src.l3num != l3proto)
2481                                 continue;
2482                         if (cb->args[1]) {
2483                                 if (exp != last)
2484                                         continue;
2485                                 cb->args[1] = 0;
2486                         }
2487                         if (ctnetlink_exp_fill_info(skb,
2488                                                     NETLINK_CB(cb->skb).portid,
2489                                                     cb->nlh->nlmsg_seq,
2490                                                     IPCTNL_MSG_EXP_NEW,
2491                                                     exp) < 0) {
2492                                 if (!atomic_inc_not_zero(&exp->use))
2493                                         continue;
2494                                 cb->args[1] = (unsigned long)exp;
2495                                 goto out;
2496                         }
2497                 }
2498                 if (cb->args[1]) {
2499                         cb->args[1] = 0;
2500                         goto restart;
2501                 }
2502         }
2503 out:
2504         rcu_read_unlock();
2505         if (last)
2506                 nf_ct_expect_put(last);
2507
2508         return skb->len;
2509 }
2510
2511 static int
2512 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2513 {
2514         struct nf_conntrack_expect *exp, *last;
2515         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2516         struct nf_conn *ct = cb->data;
2517         struct nf_conn_help *help = nfct_help(ct);
2518         u_int8_t l3proto = nfmsg->nfgen_family;
2519
2520         if (cb->args[0])
2521                 return 0;
2522
2523         rcu_read_lock();
2524         last = (struct nf_conntrack_expect *)cb->args[1];
2525 restart:
2526         hlist_for_each_entry(exp, &help->expectations, lnode) {
2527                 if (l3proto && exp->tuple.src.l3num != l3proto)
2528                         continue;
2529                 if (cb->args[1]) {
2530                         if (exp != last)
2531                                 continue;
2532                         cb->args[1] = 0;
2533                 }
2534                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2535                                             cb->nlh->nlmsg_seq,
2536                                             IPCTNL_MSG_EXP_NEW,
2537                                             exp) < 0) {
2538                         if (!atomic_inc_not_zero(&exp->use))
2539                                 continue;
2540                         cb->args[1] = (unsigned long)exp;
2541                         goto out;
2542                 }
2543         }
2544         if (cb->args[1]) {
2545                 cb->args[1] = 0;
2546                 goto restart;
2547         }
2548         cb->args[0] = 1;
2549 out:
2550         rcu_read_unlock();
2551         if (last)
2552                 nf_ct_expect_put(last);
2553
2554         return skb->len;
2555 }
2556
2557 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2558                                  const struct nlmsghdr *nlh,
2559                                  const struct nlattr * const cda[])
2560 {
2561         int err;
2562         struct net *net = sock_net(ctnl);
2563         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2564         u_int8_t u3 = nfmsg->nfgen_family;
2565         struct nf_conntrack_tuple tuple;
2566         struct nf_conntrack_tuple_hash *h;
2567         struct nf_conn *ct;
2568         u16 zone = 0;
2569         struct netlink_dump_control c = {
2570                 .dump = ctnetlink_exp_ct_dump_table,
2571                 .done = ctnetlink_exp_done,
2572         };
2573
2574         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2575         if (err < 0)
2576                 return err;
2577
2578         if (cda[CTA_EXPECT_ZONE]) {
2579                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2580                 if (err < 0)
2581                         return err;
2582         }
2583
2584         h = nf_conntrack_find_get(net, zone, &tuple);
2585         if (!h)
2586                 return -ENOENT;
2587
2588         ct = nf_ct_tuplehash_to_ctrack(h);
2589         c.data = ct;
2590
2591         err = netlink_dump_start(ctnl, skb, nlh, &c);
2592         nf_ct_put(ct);
2593
2594         return err;
2595 }
2596
2597 static int
2598 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2599                      const struct nlmsghdr *nlh,
2600                      const struct nlattr * const cda[])
2601 {
2602         struct net *net = sock_net(ctnl);
2603         struct nf_conntrack_tuple tuple;
2604         struct nf_conntrack_expect *exp;
2605         struct sk_buff *skb2;
2606         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2607         u_int8_t u3 = nfmsg->nfgen_family;
2608         u16 zone;
2609         int err;
2610
2611         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2612                 if (cda[CTA_EXPECT_MASTER])
2613                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2614                 else {
2615                         struct netlink_dump_control c = {
2616                                 .dump = ctnetlink_exp_dump_table,
2617                                 .done = ctnetlink_exp_done,
2618                         };
2619                         return netlink_dump_start(ctnl, skb, nlh, &c);
2620                 }
2621         }
2622
2623         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2624         if (err < 0)
2625                 return err;
2626
2627         if (cda[CTA_EXPECT_TUPLE])
2628                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2629         else if (cda[CTA_EXPECT_MASTER])
2630                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2631         else
2632                 return -EINVAL;
2633
2634         if (err < 0)
2635                 return err;
2636
2637         exp = nf_ct_expect_find_get(net, zone, &tuple);
2638         if (!exp)
2639                 return -ENOENT;
2640
2641         if (cda[CTA_EXPECT_ID]) {
2642                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2643                 if (ntohl(id) != (u32)(unsigned long)exp) {
2644                         nf_ct_expect_put(exp);
2645                         return -ENOENT;
2646                 }
2647         }
2648
2649         err = -ENOMEM;
2650         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2651         if (skb2 == NULL) {
2652                 nf_ct_expect_put(exp);
2653                 goto out;
2654         }
2655
2656         rcu_read_lock();
2657         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2658                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2659         rcu_read_unlock();
2660         nf_ct_expect_put(exp);
2661         if (err <= 0)
2662                 goto free;
2663
2664         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2665         if (err < 0)
2666                 goto out;
2667
2668         return 0;
2669
2670 free:
2671         kfree_skb(skb2);
2672 out:
2673         /* this avoids a loop in nfnetlink. */
2674         return err == -EAGAIN ? -ENOBUFS : err;
2675 }
2676
2677 static int
2678 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2679                      const struct nlmsghdr *nlh,
2680                      const struct nlattr * const cda[])
2681 {
2682         struct net *net = sock_net(ctnl);
2683         struct nf_conntrack_expect *exp;
2684         struct nf_conntrack_tuple tuple;
2685         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2686         struct hlist_node *next;
2687         u_int8_t u3 = nfmsg->nfgen_family;
2688         unsigned int i;
2689         u16 zone;
2690         int err;
2691
2692         if (cda[CTA_EXPECT_TUPLE]) {
2693                 /* delete a single expect by tuple */
2694                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2695                 if (err < 0)
2696                         return err;
2697
2698                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2699                 if (err < 0)
2700                         return err;
2701
2702                 /* bump usage count to 2 */
2703                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2704                 if (!exp)
2705                         return -ENOENT;
2706
2707                 if (cda[CTA_EXPECT_ID]) {
2708                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2709                         if (ntohl(id) != (u32)(unsigned long)exp) {
2710                                 nf_ct_expect_put(exp);
2711                                 return -ENOENT;
2712                         }
2713                 }
2714
2715                 /* after list removal, usage count == 1 */
2716                 spin_lock_bh(&nf_conntrack_expect_lock);
2717                 if (del_timer(&exp->timeout)) {
2718                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2719                                                    nlmsg_report(nlh));
2720                         nf_ct_expect_put(exp);
2721                 }
2722                 spin_unlock_bh(&nf_conntrack_expect_lock);
2723                 /* have to put what we 'get' above.
2724                  * after this line usage count == 0 */
2725                 nf_ct_expect_put(exp);
2726         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2727                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2728                 struct nf_conn_help *m_help;
2729
2730                 /* delete all expectations for this helper */
2731                 spin_lock_bh(&nf_conntrack_expect_lock);
2732                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2733                         hlist_for_each_entry_safe(exp, next,
2734                                                   &net->ct.expect_hash[i],
2735                                                   hnode) {
2736                                 m_help = nfct_help(exp->master);
2737                                 if (!strcmp(m_help->helper->name, name) &&
2738                                     del_timer(&exp->timeout)) {
2739                                         nf_ct_unlink_expect_report(exp,
2740                                                         NETLINK_CB(skb).portid,
2741                                                         nlmsg_report(nlh));
2742                                         nf_ct_expect_put(exp);
2743                                 }
2744                         }
2745                 }
2746                 spin_unlock_bh(&nf_conntrack_expect_lock);
2747         } else {
2748                 /* This basically means we have to flush everything*/
2749                 spin_lock_bh(&nf_conntrack_expect_lock);
2750                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2751                         hlist_for_each_entry_safe(exp, next,
2752                                                   &net->ct.expect_hash[i],
2753                                                   hnode) {
2754                                 if (del_timer(&exp->timeout)) {
2755                                         nf_ct_unlink_expect_report(exp,
2756                                                         NETLINK_CB(skb).portid,
2757                                                         nlmsg_report(nlh));
2758                                         nf_ct_expect_put(exp);
2759                                 }
2760                         }
2761                 }
2762                 spin_unlock_bh(&nf_conntrack_expect_lock);
2763         }
2764
2765         return 0;
2766 }
2767 static int
2768 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2769                         const struct nlattr * const cda[])
2770 {
2771         if (cda[CTA_EXPECT_TIMEOUT]) {
2772                 if (!del_timer(&x->timeout))
2773                         return -ETIME;
2774
2775                 x->timeout.expires = jiffies +
2776                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2777                 add_timer(&x->timeout);
2778         }
2779         return 0;
2780 }
2781
2782 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2783         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2784         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2785 };
2786
2787 static int
2788 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2789                            struct nf_conntrack_expect *exp,
2790                            u_int8_t u3)
2791 {
2792 #ifdef CONFIG_NF_NAT_NEEDED
2793         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2794         struct nf_conntrack_tuple nat_tuple = {};
2795         int err;
2796
2797         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2798         if (err < 0)
2799                 return err;
2800
2801         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2802                 return -EINVAL;
2803
2804         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2805                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2806         if (err < 0)
2807                 return err;
2808
2809         exp->saved_addr = nat_tuple.src.u3;
2810         exp->saved_proto = nat_tuple.src.u;
2811         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2812
2813         return 0;
2814 #else
2815         return -EOPNOTSUPP;
2816 #endif
2817 }
2818
2819 static struct nf_conntrack_expect *
2820 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2821                        struct nf_conntrack_helper *helper,
2822                        struct nf_conntrack_tuple *tuple,
2823                        struct nf_conntrack_tuple *mask)
2824 {
2825         u_int32_t class = 0;
2826         struct nf_conntrack_expect *exp;
2827         struct nf_conn_help *help;
2828         int err;
2829
2830         if (cda[CTA_EXPECT_CLASS] && helper) {
2831                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2832                 if (class > helper->expect_class_max)
2833                         return ERR_PTR(-EINVAL);
2834         }
2835         exp = nf_ct_expect_alloc(ct);
2836         if (!exp)
2837                 return ERR_PTR(-ENOMEM);
2838
2839         help = nfct_help(ct);
2840         if (!help) {
2841                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2842                         err = -EINVAL;
2843                         goto err_out;
2844                 }
2845                 exp->timeout.expires =
2846                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2847
2848                 exp->flags = NF_CT_EXPECT_USERSPACE;
2849                 if (cda[CTA_EXPECT_FLAGS]) {
2850                         exp->flags |=
2851                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2852                 }
2853         } else {
2854                 if (cda[CTA_EXPECT_FLAGS]) {
2855                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2856                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2857                 } else
2858                         exp->flags = 0;
2859         }
2860         if (cda[CTA_EXPECT_FN]) {
2861                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2862                 struct nf_ct_helper_expectfn *expfn;
2863
2864                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2865                 if (expfn == NULL) {
2866                         err = -EINVAL;
2867                         goto err_out;
2868                 }
2869                 exp->expectfn = expfn->expectfn;
2870         } else
2871                 exp->expectfn = NULL;
2872
2873         exp->class = class;
2874         exp->master = ct;
2875         exp->helper = helper;
2876         exp->tuple = *tuple;
2877         exp->mask.src.u3 = mask->src.u3;
2878         exp->mask.src.u.all = mask->src.u.all;
2879
2880         if (cda[CTA_EXPECT_NAT]) {
2881                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2882                                                  exp, nf_ct_l3num(ct));
2883                 if (err < 0)
2884                         goto err_out;
2885         }
2886         return exp;
2887 err_out:
2888         nf_ct_expect_put(exp);
2889         return ERR_PTR(err);
2890 }
2891
2892 static int
2893 ctnetlink_create_expect(struct net *net, u16 zone,
2894                         const struct nlattr * const cda[],
2895                         u_int8_t u3, u32 portid, int report)
2896 {
2897         struct nf_conntrack_tuple tuple, mask, master_tuple;
2898         struct nf_conntrack_tuple_hash *h = NULL;
2899         struct nf_conntrack_helper *helper = NULL;
2900         struct nf_conntrack_expect *exp;
2901         struct nf_conn *ct;
2902         int err;
2903
2904         /* caller guarantees that those three CTA_EXPECT_* exist */
2905         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2906         if (err < 0)
2907                 return err;
2908         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2909         if (err < 0)
2910                 return err;
2911         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2912         if (err < 0)
2913                 return err;
2914
2915         /* Look for master conntrack of this expectation */
2916         h = nf_conntrack_find_get(net, zone, &master_tuple);
2917         if (!h)
2918                 return -ENOENT;
2919         ct = nf_ct_tuplehash_to_ctrack(h);
2920
2921         if (cda[CTA_EXPECT_HELP_NAME]) {
2922                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2923
2924                 helper = __nf_conntrack_helper_find(helpname, u3,
2925                                                     nf_ct_protonum(ct));
2926                 if (helper == NULL) {
2927 #ifdef CONFIG_MODULES
2928                         if (request_module("nfct-helper-%s", helpname) < 0) {
2929                                 err = -EOPNOTSUPP;
2930                                 goto err_ct;
2931                         }
2932                         helper = __nf_conntrack_helper_find(helpname, u3,
2933                                                             nf_ct_protonum(ct));
2934                         if (helper) {
2935                                 err = -EAGAIN;
2936                                 goto err_ct;
2937                         }
2938 #endif
2939                         err = -EOPNOTSUPP;
2940                         goto err_ct;
2941                 }
2942         }
2943
2944         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
2945         if (IS_ERR(exp)) {
2946                 err = PTR_ERR(exp);
2947                 goto err_ct;
2948         }
2949
2950         err = nf_ct_expect_related_report(exp, portid, report);
2951         if (err < 0)
2952                 goto err_exp;
2953
2954         return 0;
2955 err_exp:
2956         nf_ct_expect_put(exp);
2957 err_ct:
2958         nf_ct_put(ct);
2959         return err;
2960 }
2961
2962 static int
2963 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2964                      const struct nlmsghdr *nlh,
2965                      const struct nlattr * const cda[])
2966 {
2967         struct net *net = sock_net(ctnl);
2968         struct nf_conntrack_tuple tuple;
2969         struct nf_conntrack_expect *exp;
2970         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2971         u_int8_t u3 = nfmsg->nfgen_family;
2972         u16 zone;
2973         int err;
2974
2975         if (!cda[CTA_EXPECT_TUPLE]
2976             || !cda[CTA_EXPECT_MASK]
2977             || !cda[CTA_EXPECT_MASTER])
2978                 return -EINVAL;
2979
2980         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2981         if (err < 0)
2982                 return err;
2983
2984         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2985         if (err < 0)
2986                 return err;
2987
2988         spin_lock_bh(&nf_conntrack_expect_lock);
2989         exp = __nf_ct_expect_find(net, zone, &tuple);
2990
2991         if (!exp) {
2992                 spin_unlock_bh(&nf_conntrack_expect_lock);
2993                 err = -ENOENT;
2994                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2995                         err = ctnetlink_create_expect(net, zone, cda,
2996                                                       u3,
2997                                                       NETLINK_CB(skb).portid,
2998                                                       nlmsg_report(nlh));
2999                 }
3000                 return err;
3001         }
3002
3003         err = -EEXIST;
3004         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3005                 err = ctnetlink_change_expect(exp, cda);
3006         spin_unlock_bh(&nf_conntrack_expect_lock);
3007
3008         return err;
3009 }
3010
3011 static int
3012 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3013                              const struct ip_conntrack_stat *st)
3014 {
3015         struct nlmsghdr *nlh;
3016         struct nfgenmsg *nfmsg;
3017         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3018
3019         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3020         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3021         if (nlh == NULL)
3022                 goto nlmsg_failure;
3023
3024         nfmsg = nlmsg_data(nlh);
3025         nfmsg->nfgen_family = AF_UNSPEC;
3026         nfmsg->version      = NFNETLINK_V0;
3027         nfmsg->res_id       = htons(cpu);
3028
3029         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3030             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3031             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3032                 goto nla_put_failure;
3033
3034         nlmsg_end(skb, nlh);
3035         return skb->len;
3036
3037 nla_put_failure:
3038 nlmsg_failure:
3039         nlmsg_cancel(skb, nlh);
3040         return -1;
3041 }
3042
3043 static int
3044 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3045 {
3046         int cpu;
3047         struct net *net = sock_net(skb->sk);
3048
3049         if (cb->args[0] == nr_cpu_ids)
3050                 return 0;
3051
3052         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3053                 const struct ip_conntrack_stat *st;
3054
3055                 if (!cpu_possible(cpu))
3056                         continue;
3057
3058                 st = per_cpu_ptr(net->ct.stat, cpu);
3059                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3060                                                  cb->nlh->nlmsg_seq,
3061                                                  cpu, st) < 0)
3062                         break;
3063         }
3064         cb->args[0] = cpu;
3065
3066         return skb->len;
3067 }
3068
3069 static int
3070 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3071                        const struct nlmsghdr *nlh,
3072                        const struct nlattr * const cda[])
3073 {
3074         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3075                 struct netlink_dump_control c = {
3076                         .dump = ctnetlink_exp_stat_cpu_dump,
3077                 };
3078                 return netlink_dump_start(ctnl, skb, nlh, &c);
3079         }
3080
3081         return 0;
3082 }
3083
3084 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3085 static struct nf_ct_event_notifier ctnl_notifier = {
3086         .fcn = ctnetlink_conntrack_event,
3087 };
3088
3089 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3090         .fcn = ctnetlink_expect_event,
3091 };
3092 #endif
3093
3094 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3095         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3096                                             .attr_count = CTA_MAX,
3097                                             .policy = ct_nla_policy },
3098         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3099                                             .attr_count = CTA_MAX,
3100                                             .policy = ct_nla_policy },
3101         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3102                                             .attr_count = CTA_MAX,
3103                                             .policy = ct_nla_policy },
3104         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3105                                             .attr_count = CTA_MAX,
3106                                             .policy = ct_nla_policy },
3107         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3108         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3109         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3110         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3111 };
3112
3113 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3114         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3115                                             .attr_count = CTA_EXPECT_MAX,
3116                                             .policy = exp_nla_policy },
3117         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3118                                             .attr_count = CTA_EXPECT_MAX,
3119                                             .policy = exp_nla_policy },
3120         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3121                                             .attr_count = CTA_EXPECT_MAX,
3122                                             .policy = exp_nla_policy },
3123         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3124 };
3125
3126 static const struct nfnetlink_subsystem ctnl_subsys = {
3127         .name                           = "conntrack",
3128         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3129         .cb_count                       = IPCTNL_MSG_MAX,
3130         .cb                             = ctnl_cb,
3131 };
3132
3133 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3134         .name                           = "conntrack_expect",
3135         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3136         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3137         .cb                             = ctnl_exp_cb,
3138 };
3139
3140 MODULE_ALIAS("ip_conntrack_netlink");
3141 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3142 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3143
3144 static int __net_init ctnetlink_net_init(struct net *net)
3145 {
3146 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3147         int ret;
3148
3149         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3150         if (ret < 0) {
3151                 pr_err("ctnetlink_init: cannot register notifier.\n");
3152                 goto err_out;
3153         }
3154
3155         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3156         if (ret < 0) {
3157                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3158                 goto err_unreg_notifier;
3159         }
3160 #endif
3161         return 0;
3162
3163 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3164 err_unreg_notifier:
3165         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3166 err_out:
3167         return ret;
3168 #endif
3169 }
3170
3171 static void ctnetlink_net_exit(struct net *net)
3172 {
3173 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3174         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3175         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3176 #endif
3177 }
3178
3179 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3180 {
3181         struct net *net;
3182
3183         list_for_each_entry(net, net_exit_list, exit_list)
3184                 ctnetlink_net_exit(net);
3185 }
3186
3187 static struct pernet_operations ctnetlink_net_ops = {
3188         .init           = ctnetlink_net_init,
3189         .exit_batch     = ctnetlink_net_exit_batch,
3190 };
3191
3192 static int __init ctnetlink_init(void)
3193 {
3194         int ret;
3195
3196         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3197         ret = nfnetlink_subsys_register(&ctnl_subsys);
3198         if (ret < 0) {
3199                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3200                 goto err_out;
3201         }
3202
3203         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3204         if (ret < 0) {
3205                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3206                 goto err_unreg_subsys;
3207         }
3208
3209         ret = register_pernet_subsys(&ctnetlink_net_ops);
3210         if (ret < 0) {
3211                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3212                 goto err_unreg_exp_subsys;
3213         }
3214 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3215         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3216         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3217 #endif
3218         return 0;
3219
3220 err_unreg_exp_subsys:
3221         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3222 err_unreg_subsys:
3223         nfnetlink_subsys_unregister(&ctnl_subsys);
3224 err_out:
3225         return ret;
3226 }
3227
3228 static void __exit ctnetlink_exit(void)
3229 {
3230         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3231
3232         unregister_pernet_subsys(&ctnetlink_net_ops);
3233         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3234         nfnetlink_subsys_unregister(&ctnl_subsys);
3235 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3236         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3237 #endif
3238 }
3239
3240 module_init(ctnetlink_init);
3241 module_exit(ctnetlink_exit);