net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <[email protected]>
   5  * (C) 2002-2006 by Harald Welte <[email protected]>
   6  * (C) 2003 by Patrick Mchardy <[email protected]>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <[email protected]>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #ifdef CONFIG_NF_NAT_NEEDED
  49 #include <net/netfilter/nf_nat_core.h>
  50 #include <net/netfilter/nf_nat_l4proto.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 MODULE_LICENSE("GPL");
  58
  59 static char __initdata version[] = "0.93";
  60
  61 static inline int
  62 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  63                             const struct nf_conntrack_tuple *tuple,
  64                             struct nf_conntrack_l4proto *l4proto)
  65 {
  66         int ret = 0;
  67         struct nlattr *nest_parms;
  68
  69         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  70         if (!nest_parms)
  71                 goto nla_put_failure;
  72         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  73                 goto nla_put_failure;
  74
  75         if (likely(l4proto->tuple_to_nlattr))
  76                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  77
  78         nla_nest_end(skb, nest_parms);
  79
  80         return ret;
  81
  82 nla_put_failure:
  83         return -1;
  84 }
  85
  86 static inline int
  87 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  88                          const struct nf_conntrack_tuple *tuple,
  89                          struct nf_conntrack_l3proto *l3proto)
  90 {
  91         int ret = 0;
  92         struct nlattr *nest_parms;
  93
  94         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  95         if (!nest_parms)
  96                 goto nla_put_failure;
  97
  98         if (likely(l3proto->tuple_to_nlattr))
  99                 ret = l3proto->tuple_to_nlattr(skb, tuple);
 100
 101         nla_nest_end(skb, nest_parms);
 102
 103         return ret;
 104
 105 nla_put_failure:
 106         return -1;
 107 }
 108
 109 static int
 110 ctnetlink_dump_tuples(struct sk_buff *skb,
 111                       const struct nf_conntrack_tuple *tuple)
 112 {
 113         int ret;
 114         struct nf_conntrack_l3proto *l3proto;
 115         struct nf_conntrack_l4proto *l4proto;
 116
 117         rcu_read_lock();
 118         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 119         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 120
 121         if (ret >= 0) {
 122                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 123                                                tuple->dst.protonum);
 124                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 125         }
 126         rcu_read_unlock();
 127         return ret;
 128 }
 129
 130 static inline int
 131 ctnetlink_dump_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 #ifdef CONFIG_NF_CONNTRACK_ZONES
 601                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE */
 602 #endif
 603                + ctnetlink_proto_size(ct)
 604                + ctnetlink_label_size(ct)
 605                ;
 606 }
 607
 608 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 609 static int
 610 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 611 {
 612         struct net *net;
 613         struct nlmsghdr *nlh;
 614         struct nfgenmsg *nfmsg;
 615         struct nlattr *nest_parms;
 616         struct nf_conn *ct = item->ct;
 617         struct sk_buff *skb;
 618         unsigned int type;
 619         unsigned int flags = 0, group;
 620         int err;
 621
 622         /* ignore our fake conntrack entry */
 623         if (nf_ct_is_untracked(ct))
 624                 return 0;
 625
 626         if (events & (1 << IPCT_DESTROY)) {
 627                 type = IPCTNL_MSG_CT_DELETE;
 628                 group = NFNLGRP_CONNTRACK_DESTROY;
 629         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 630                 type = IPCTNL_MSG_CT_NEW;
 631                 flags = NLM_F_CREATE|NLM_F_EXCL;
 632                 group = NFNLGRP_CONNTRACK_NEW;
 633         } else  if (events) {
 634                 type = IPCTNL_MSG_CT_NEW;
 635                 group = NFNLGRP_CONNTRACK_UPDATE;
 636         } else
 637                 return 0;
 638
 639         net = nf_ct_net(ct);
 640         if (!item->report && !nfnetlink_has_listeners(net, group))
 641                 return 0;
 642
 643         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 644         if (skb == NULL)
 645                 goto errout;
 646
 647         type |= NFNL_SUBSYS_CTNETLINK << 8;
 648         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 649         if (nlh == NULL)
 650                 goto nlmsg_failure;
 651
 652         nfmsg = nlmsg_data(nlh);
 653         nfmsg->nfgen_family = nf_ct_l3num(ct);
 654         nfmsg->version  = NFNETLINK_V0;
 655         nfmsg->res_id   = 0;
 656
 657         rcu_read_lock();
 658         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 659         if (!nest_parms)
 660                 goto nla_put_failure;
 661         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 662                 goto nla_put_failure;
 663         nla_nest_end(skb, nest_parms);
 664
 665         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 666         if (!nest_parms)
 667                 goto nla_put_failure;
 668         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 669                 goto nla_put_failure;
 670         nla_nest_end(skb, nest_parms);
 671
 672         if (nf_ct_zone(ct) &&
 673             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 674                 goto nla_put_failure;
 675
 676         if (ctnetlink_dump_id(skb, ct) < 0)
 677                 goto nla_put_failure;
 678
 679         if (ctnetlink_dump_status(skb, ct) < 0)
 680                 goto nla_put_failure;
 681
 682         if (events & (1 << IPCT_DESTROY)) {
 683                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 684                     ctnetlink_dump_timestamp(skb, ct) < 0)
 685                         goto nla_put_failure;
 686         } else {
 687                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 688                         goto nla_put_failure;
 689
 690                 if (events & (1 << IPCT_PROTOINFO)
 691                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 692                         goto nla_put_failure;
 693
 694                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 695                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 696                         goto nla_put_failure;
 697
 698 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 699                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 700                     && ctnetlink_dump_secctx(skb, ct) < 0)
 701                         goto nla_put_failure;
 702 #endif
 703                 if (events & (1 << IPCT_LABEL) &&
 704                      ctnetlink_dump_labels(skb, ct) < 0)
 705                         goto nla_put_failure;
 706
 707                 if (events & (1 << IPCT_RELATED) &&
 708                     ctnetlink_dump_master(skb, ct) < 0)
 709                         goto nla_put_failure;
 710
 711                 if (events & (1 << IPCT_SEQADJ) &&
 712                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 713                         goto nla_put_failure;
 714         }
 715
 716 #ifdef CONFIG_NF_CONNTRACK_MARK
 717         if ((events & (1 << IPCT_MARK) || ct->mark)
 718             && ctnetlink_dump_mark(skb, ct) < 0)
 719                 goto nla_put_failure;
 720 #endif
 721         rcu_read_unlock();
 722
 723         nlmsg_end(skb, nlh);
 724         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 725                              GFP_ATOMIC);
 726         if (err == -ENOBUFS || err == -EAGAIN)
 727                 return -ENOBUFS;
 728
 729         return 0;
 730
 731 nla_put_failure:
 732         rcu_read_unlock();
 733         nlmsg_cancel(skb, nlh);
 734 nlmsg_failure:
 735         kfree_skb(skb);
 736 errout:
 737         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 738                 return -ENOBUFS;
 739
 740         return 0;
 741 }
 742 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 743
 744 static int ctnetlink_done(struct netlink_callback *cb)
 745 {
 746         if (cb->args[1])
 747                 nf_ct_put((struct nf_conn *)cb->args[1]);
 748         kfree(cb->data);
 749         return 0;
 750 }
 751
 752 struct ctnetlink_dump_filter {
 753         struct {
 754                 u_int32_t val;
 755                 u_int32_t mask;
 756         } mark;
 757 };
 758
 759 static int
 760 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 761 {
 762         struct net *net = sock_net(skb->sk);
 763         struct nf_conn *ct, *last;
 764         struct nf_conntrack_tuple_hash *h;
 765         struct hlist_nulls_node *n;
 766         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 767         u_int8_t l3proto = nfmsg->nfgen_family;
 768         int res;
 769         spinlock_t *lockp;
 770
 771 #ifdef CONFIG_NF_CONNTRACK_MARK
 772         const struct ctnetlink_dump_filter *filter = cb->data;
 773 #endif
 774
 775         last = (struct nf_conn *)cb->args[1];
 776
 777         local_bh_disable();
 778         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 779 restart:
 780                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 781                 spin_lock(lockp);
 782                 if (cb->args[0] >= net->ct.htable_size) {
 783                         spin_unlock(lockp);
 784                         goto out;
 785                 }
 786                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 787                                          hnnode) {
 788                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 789                                 continue;
 790                         ct = nf_ct_tuplehash_to_ctrack(h);
 791                         /* Dump entries of a given L3 protocol number.
 792                          * If it is not specified, ie. l3proto == 0,
 793                          * then dump everything. */
 794                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 795                                 continue;
 796                         if (cb->args[1]) {
 797                                 if (ct != last)
 798                                         continue;
 799                                 cb->args[1] = 0;
 800                         }
 801 #ifdef CONFIG_NF_CONNTRACK_MARK
 802                         if (filter && !((ct->mark & filter->mark.mask) ==
 803                                         filter->mark.val)) {
 804                                 continue;
 805                         }
 806 #endif
 807                         rcu_read_lock();
 808                         res =
 809                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 810                                             cb->nlh->nlmsg_seq,
 811                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 812                                             ct);
 813                         rcu_read_unlock();
 814                         if (res < 0) {
 815                                 nf_conntrack_get(&ct->ct_general);
 816                                 cb->args[1] = (unsigned long)ct;
 817                                 spin_unlock(lockp);
 818                                 goto out;
 819                         }
 820                 }
 821                 spin_unlock(lockp);
 822                 if (cb->args[1]) {
 823                         cb->args[1] = 0;
 824                         goto restart;
 825                 }
 826         }
 827 out:
 828         local_bh_enable();
 829         if (last)
 830                 nf_ct_put(last);
 831
 832         return skb->len;
 833 }
 834
 835 static inline int
 836 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 837 {
 838         struct nlattr *tb[CTA_IP_MAX+1];
 839         struct nf_conntrack_l3proto *l3proto;
 840         int ret = 0;
 841
 842         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 843         if (ret < 0)
 844                 return ret;
 845
 846         rcu_read_lock();
 847         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 848
 849         if (likely(l3proto->nlattr_to_tuple)) {
 850                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 851                                           l3proto->nla_policy);
 852                 if (ret == 0)
 853                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 854         }
 855
 856         rcu_read_unlock();
 857
 858         return ret;
 859 }
 860
 861 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 862         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 863 };
 864
 865 static inline int
 866 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 867                             struct nf_conntrack_tuple *tuple)
 868 {
 869         struct nlattr *tb[CTA_PROTO_MAX+1];
 870         struct nf_conntrack_l4proto *l4proto;
 871         int ret = 0;
 872
 873         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 874         if (ret < 0)
 875                 return ret;
 876
 877         if (!tb[CTA_PROTO_NUM])
 878                 return -EINVAL;
 879         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 880
 881         rcu_read_lock();
 882         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 883
 884         if (likely(l4proto->nlattr_to_tuple)) {
 885                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 886                                           l4proto->nla_policy);
 887                 if (ret == 0)
 888                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 889         }
 890
 891         rcu_read_unlock();
 892
 893         return ret;
 894 }
 895
 896 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 897         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 898         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 899 };
 900
 901 static int
 902 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 903                       struct nf_conntrack_tuple *tuple,
 904                       enum ctattr_type type, u_int8_t l3num)
 905 {
 906         struct nlattr *tb[CTA_TUPLE_MAX+1];
 907         int err;
 908
 909         memset(tuple, 0, sizeof(*tuple));
 910
 911         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 912         if (err < 0)
 913                 return err;
 914
 915         if (!tb[CTA_TUPLE_IP])
 916                 return -EINVAL;
 917
 918         tuple->src.l3num = l3num;
 919
 920         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 921         if (err < 0)
 922                 return err;
 923
 924         if (!tb[CTA_TUPLE_PROTO])
 925                 return -EINVAL;
 926
 927         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 928         if (err < 0)
 929                 return err;
 930
 931         /* orig and expect tuples get DIR_ORIGINAL */
 932         if (type == CTA_TUPLE_REPLY)
 933                 tuple->dst.dir = IP_CT_DIR_REPLY;
 934         else
 935                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 936
 937         return 0;
 938 }
 939
 940 static int
 941 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 942 {
 943         if (attr)
 944 #ifdef CONFIG_NF_CONNTRACK_ZONES
 945                 *zone = ntohs(nla_get_be16(attr));
 946 #else
 947                 return -EOPNOTSUPP;
 948 #endif
 949         else
 950                 *zone = 0;
 951
 952         return 0;
 953 }
 954
 955 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 956         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 957                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 958 };
 959
 960 static inline int
 961 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 962                      struct nlattr **helpinfo)
 963 {
 964         int err;
 965         struct nlattr *tb[CTA_HELP_MAX+1];
 966
 967         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 968         if (err < 0)
 969                 return err;
 970
 971         if (!tb[CTA_HELP_NAME])
 972                 return -EINVAL;
 973
 974         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 975
 976         if (tb[CTA_HELP_INFO])
 977                 *helpinfo = tb[CTA_HELP_INFO];
 978
 979         return 0;
 980 }
 981
 982 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 983         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 984         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 985         [CTA_STATUS]            = { .type = NLA_U32 },
 986         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 987         [CTA_HELP]              = { .type = NLA_NESTED },
 988         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 989         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 990         [CTA_MARK]              = { .type = NLA_U32 },
 991         [CTA_ID]                = { .type = NLA_U32 },
 992         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 993         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 994         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
 995         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
 996         [CTA_ZONE]              = { .type = NLA_U16 },
 997         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 998         [CTA_LABELS]            = { .type = NLA_BINARY,
 999                                     .len = NF_CT_LABELS_MAX_SIZE },
1000         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1001                                     .len = NF_CT_LABELS_MAX_SIZE },
1002 };
1003
1004 static int
1005 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
1006                         const struct nlmsghdr *nlh,
1007                         const struct nlattr * const cda[])
1008 {
1009         struct net *net = sock_net(ctnl);
1010         struct nf_conntrack_tuple_hash *h;
1011         struct nf_conntrack_tuple tuple;
1012         struct nf_conn *ct;
1013         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1014         u_int8_t u3 = nfmsg->nfgen_family;
1015         u16 zone;
1016         int err;
1017
1018         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1019         if (err < 0)
1020                 return err;
1021
1022         if (cda[CTA_TUPLE_ORIG])
1023                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1024         else if (cda[CTA_TUPLE_REPLY])
1025                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1026         else {
1027                 /* Flush the whole table */
1028                 nf_conntrack_flush_report(net,
1029                                          NETLINK_CB(skb).portid,
1030                                          nlmsg_report(nlh));
1031                 return 0;
1032         }
1033
1034         if (err < 0)
1035                 return err;
1036
1037         h = nf_conntrack_find_get(net, zone, &tuple);
1038         if (!h)
1039                 return -ENOENT;
1040
1041         ct = nf_ct_tuplehash_to_ctrack(h);
1042
1043         if (cda[CTA_ID]) {
1044                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1045                 if (id != (u32)(unsigned long)ct) {
1046                         nf_ct_put(ct);
1047                         return -ENOENT;
1048                 }
1049         }
1050
1051         if (del_timer(&ct->timeout))
1052                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1053
1054         nf_ct_put(ct);
1055
1056         return 0;
1057 }
1058
1059 static int
1060 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1061                         const struct nlmsghdr *nlh,
1062                         const struct nlattr * const cda[])
1063 {
1064         struct net *net = sock_net(ctnl);
1065         struct nf_conntrack_tuple_hash *h;
1066         struct nf_conntrack_tuple tuple;
1067         struct nf_conn *ct;
1068         struct sk_buff *skb2 = NULL;
1069         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1070         u_int8_t u3 = nfmsg->nfgen_family;
1071         u16 zone;
1072         int err;
1073
1074         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1075                 struct netlink_dump_control c = {
1076                         .dump = ctnetlink_dump_table,
1077                         .done = ctnetlink_done,
1078                 };
1079 #ifdef CONFIG_NF_CONNTRACK_MARK
1080                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1081                         struct ctnetlink_dump_filter *filter;
1082
1083                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1084                                          GFP_ATOMIC);
1085                         if (filter == NULL)
1086                                 return -ENOMEM;
1087
1088                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1089                         filter->mark.mask =
1090                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1091                         c.data = filter;
1092                 }
1093 #endif
1094                 return netlink_dump_start(ctnl, skb, nlh, &c);
1095         }
1096
1097         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1098         if (err < 0)
1099                 return err;
1100
1101         if (cda[CTA_TUPLE_ORIG])
1102                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1103         else if (cda[CTA_TUPLE_REPLY])
1104                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1105         else
1106                 return -EINVAL;
1107
1108         if (err < 0)
1109                 return err;
1110
1111         h = nf_conntrack_find_get(net, zone, &tuple);
1112         if (!h)
1113                 return -ENOENT;
1114
1115         ct = nf_ct_tuplehash_to_ctrack(h);
1116
1117         err = -ENOMEM;
1118         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1119         if (skb2 == NULL) {
1120                 nf_ct_put(ct);
1121                 return -ENOMEM;
1122         }
1123
1124         rcu_read_lock();
1125         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1126                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1127         rcu_read_unlock();
1128         nf_ct_put(ct);
1129         if (err <= 0)
1130                 goto free;
1131
1132         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1133         if (err < 0)
1134                 goto out;
1135
1136         return 0;
1137
1138 free:
1139         kfree_skb(skb2);
1140 out:
1141         /* this avoids a loop in nfnetlink. */
1142         return err == -EAGAIN ? -ENOBUFS : err;
1143 }
1144
1145 static int ctnetlink_done_list(struct netlink_callback *cb)
1146 {
1147         if (cb->args[1])
1148                 nf_ct_put((struct nf_conn *)cb->args[1]);
1149         return 0;
1150 }
1151
1152 static int
1153 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1154 {
1155         struct nf_conn *ct, *last;
1156         struct nf_conntrack_tuple_hash *h;
1157         struct hlist_nulls_node *n;
1158         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1159         u_int8_t l3proto = nfmsg->nfgen_family;
1160         int res;
1161         int cpu;
1162         struct hlist_nulls_head *list;
1163         struct net *net = sock_net(skb->sk);
1164
1165         if (cb->args[2])
1166                 return 0;
1167
1168         last = (struct nf_conn *)cb->args[1];
1169
1170         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1171                 struct ct_pcpu *pcpu;
1172
1173                 if (!cpu_possible(cpu))
1174                         continue;
1175
1176                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1177                 spin_lock_bh(&pcpu->lock);
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                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1197                                         continue;
1198                                 cb->args[0] = cpu;
1199                                 cb->args[1] = (unsigned long)ct;
1200                                 spin_unlock_bh(&pcpu->lock);
1201                                 goto out;
1202                         }
1203                 }
1204                 if (cb->args[1]) {
1205                         cb->args[1] = 0;
1206                         goto restart;
1207                 }
1208                 spin_unlock_bh(&pcpu->lock);
1209         }
1210         cb->args[2] = 1;
1211 out:
1212         if (last)
1213                 nf_ct_put(last);
1214
1215         return skb->len;
1216 }
1217
1218 static int
1219 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1220 {
1221         return ctnetlink_dump_list(skb, cb, true);
1222 }
1223
1224 static int
1225 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1226                        const struct nlmsghdr *nlh,
1227                        const struct nlattr * const cda[])
1228 {
1229         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1230                 struct netlink_dump_control c = {
1231                         .dump = ctnetlink_dump_dying,
1232                         .done = ctnetlink_done_list,
1233                 };
1234                 return netlink_dump_start(ctnl, skb, nlh, &c);
1235         }
1236
1237         return -EOPNOTSUPP;
1238 }
1239
1240 static int
1241 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1242 {
1243         return ctnetlink_dump_list(skb, cb, false);
1244 }
1245
1246 static int
1247 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1248                              const struct nlmsghdr *nlh,
1249                              const struct nlattr * const cda[])
1250 {
1251         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1252                 struct netlink_dump_control c = {
1253                         .dump = ctnetlink_dump_unconfirmed,
1254                         .done = ctnetlink_done_list,
1255                 };
1256                 return netlink_dump_start(ctnl, skb, nlh, &c);
1257         }
1258
1259         return -EOPNOTSUPP;
1260 }
1261
1262 #ifdef CONFIG_NF_NAT_NEEDED
1263 static int
1264 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1265                           enum nf_nat_manip_type manip,
1266                           const struct nlattr *attr)
1267 {
1268         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1269         int err;
1270
1271         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1272         if (!parse_nat_setup) {
1273 #ifdef CONFIG_MODULES
1274                 rcu_read_unlock();
1275                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1276                 if (request_module("nf-nat") < 0) {
1277                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1278                         rcu_read_lock();
1279                         return -EOPNOTSUPP;
1280                 }
1281                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1282                 rcu_read_lock();
1283                 if (nfnetlink_parse_nat_setup_hook)
1284                         return -EAGAIN;
1285 #endif
1286                 return -EOPNOTSUPP;
1287         }
1288
1289         err = parse_nat_setup(ct, manip, attr);
1290         if (err == -EAGAIN) {
1291 #ifdef CONFIG_MODULES
1292                 rcu_read_unlock();
1293                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1294                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1295                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1296                         rcu_read_lock();
1297                         return -EOPNOTSUPP;
1298                 }
1299                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1300                 rcu_read_lock();
1301 #else
1302                 err = -EOPNOTSUPP;
1303 #endif
1304         }
1305         return err;
1306 }
1307 #endif
1308
1309 static int
1310 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1311 {
1312         unsigned long d;
1313         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1314         d = ct->status ^ status;
1315
1316         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1317                 /* unchangeable */
1318                 return -EBUSY;
1319
1320         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1321                 /* SEEN_REPLY bit can only be set */
1322                 return -EBUSY;
1323
1324         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1325                 /* ASSURED bit can only be set */
1326                 return -EBUSY;
1327
1328         /* Be careful here, modifying NAT bits can screw up things,
1329          * so don't let users modify them directly if they don't pass
1330          * nf_nat_range. */
1331         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1332         return 0;
1333 }
1334
1335 static int
1336 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1337 {
1338 #ifdef CONFIG_NF_NAT_NEEDED
1339         int ret;
1340
1341         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1342                 return 0;
1343
1344         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1345                                         cda[CTA_NAT_DST]);
1346         if (ret < 0)
1347                 return ret;
1348
1349         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1350                                         cda[CTA_NAT_SRC]);
1351         return ret;
1352 #else
1353         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1354                 return 0;
1355         return -EOPNOTSUPP;
1356 #endif
1357 }
1358
1359 static inline int
1360 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1361 {
1362         struct nf_conntrack_helper *helper;
1363         struct nf_conn_help *help = nfct_help(ct);
1364         char *helpname = NULL;
1365         struct nlattr *helpinfo = NULL;
1366         int err;
1367
1368         /* don't change helper of sibling connections */
1369         if (ct->master)
1370                 return -EBUSY;
1371
1372         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1373         if (err < 0)
1374                 return err;
1375
1376         if (!strcmp(helpname, "")) {
1377                 if (help && help->helper) {
1378                         /* we had a helper before ... */
1379                         nf_ct_remove_expectations(ct);
1380                         RCU_INIT_POINTER(help->helper, NULL);
1381                 }
1382
1383                 return 0;
1384         }
1385
1386         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1387                                             nf_ct_protonum(ct));
1388         if (helper == NULL) {
1389 #ifdef CONFIG_MODULES
1390                 spin_unlock_bh(&nf_conntrack_expect_lock);
1391
1392                 if (request_module("nfct-helper-%s", helpname) < 0) {
1393                         spin_lock_bh(&nf_conntrack_expect_lock);
1394                         return -EOPNOTSUPP;
1395                 }
1396
1397                 spin_lock_bh(&nf_conntrack_expect_lock);
1398                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1399                                                     nf_ct_protonum(ct));
1400                 if (helper)
1401                         return -EAGAIN;
1402 #endif
1403                 return -EOPNOTSUPP;
1404         }
1405
1406         if (help) {
1407                 if (help->helper == helper) {
1408                         /* update private helper data if allowed. */
1409                         if (helper->from_nlattr)
1410                                 helper->from_nlattr(helpinfo, ct);
1411                         return 0;
1412                 } else
1413                         return -EBUSY;
1414         }
1415
1416         /* we cannot set a helper for an existing conntrack */
1417         return -EOPNOTSUPP;
1418 }
1419
1420 static inline int
1421 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1422 {
1423         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1424
1425         if (!del_timer(&ct->timeout))
1426                 return -ETIME;
1427
1428         ct->timeout.expires = jiffies + timeout * HZ;
1429         add_timer(&ct->timeout);
1430
1431         return 0;
1432 }
1433
1434 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1435         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1436         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1437         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1438 };
1439
1440 static inline int
1441 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1442 {
1443         const struct nlattr *attr = cda[CTA_PROTOINFO];
1444         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1445         struct nf_conntrack_l4proto *l4proto;
1446         int err = 0;
1447
1448         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1449         if (err < 0)
1450                 return err;
1451
1452         rcu_read_lock();
1453         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1454         if (l4proto->from_nlattr)
1455                 err = l4proto->from_nlattr(tb, ct);
1456         rcu_read_unlock();
1457
1458         return err;
1459 }
1460
1461 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1462         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1463         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1464         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1465 };
1466
1467 static inline int
1468 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1469 {
1470         int err;
1471         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1472
1473         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1474         if (err < 0)
1475                 return err;
1476
1477         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1478                 return -EINVAL;
1479
1480         seq->correction_pos =
1481                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1482
1483         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1484                 return -EINVAL;
1485
1486         seq->offset_before =
1487                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1488
1489         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1490                 return -EINVAL;
1491
1492         seq->offset_after =
1493                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1494
1495         return 0;
1496 }
1497
1498 static int
1499 ctnetlink_change_seq_adj(struct nf_conn *ct,
1500                          const struct nlattr * const cda[])
1501 {
1502         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1503         int ret = 0;
1504
1505         if (!seqadj)
1506                 return 0;
1507
1508         if (cda[CTA_SEQ_ADJ_ORIG]) {
1509                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1510                                      cda[CTA_SEQ_ADJ_ORIG]);
1511                 if (ret < 0)
1512                         return ret;
1513
1514                 ct->status |= IPS_SEQ_ADJUST;
1515         }
1516
1517         if (cda[CTA_SEQ_ADJ_REPLY]) {
1518                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1519                                      cda[CTA_SEQ_ADJ_REPLY]);
1520                 if (ret < 0)
1521                         return ret;
1522
1523                 ct->status |= IPS_SEQ_ADJUST;
1524         }
1525
1526         return 0;
1527 }
1528
1529 static int
1530 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1531 {
1532 #ifdef CONFIG_NF_CONNTRACK_LABELS
1533         size_t len = nla_len(cda[CTA_LABELS]);
1534         const void *mask = cda[CTA_LABELS_MASK];
1535
1536         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1537                 return -EINVAL;
1538
1539         if (mask) {
1540                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1541                     nla_len(cda[CTA_LABELS_MASK]) != len)
1542                         return -EINVAL;
1543                 mask = nla_data(cda[CTA_LABELS_MASK]);
1544         }
1545
1546         len /= sizeof(u32);
1547
1548         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1549 #else
1550         return -EOPNOTSUPP;
1551 #endif
1552 }
1553
1554 static int
1555 ctnetlink_change_conntrack(struct nf_conn *ct,
1556                            const struct nlattr * const cda[])
1557 {
1558         int err;
1559
1560         /* only allow NAT changes and master assignation for new conntracks */
1561         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1562                 return -EOPNOTSUPP;
1563
1564         if (cda[CTA_HELP]) {
1565                 err = ctnetlink_change_helper(ct, cda);
1566                 if (err < 0)
1567                         return err;
1568         }
1569
1570         if (cda[CTA_TIMEOUT]) {
1571                 err = ctnetlink_change_timeout(ct, cda);
1572                 if (err < 0)
1573                         return err;
1574         }
1575
1576         if (cda[CTA_STATUS]) {
1577                 err = ctnetlink_change_status(ct, cda);
1578                 if (err < 0)
1579                         return err;
1580         }
1581
1582         if (cda[CTA_PROTOINFO]) {
1583                 err = ctnetlink_change_protoinfo(ct, cda);
1584                 if (err < 0)
1585                         return err;
1586         }
1587
1588 #if defined(CONFIG_NF_CONNTRACK_MARK)
1589         if (cda[CTA_MARK])
1590                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1591 #endif
1592
1593         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1594                 err = ctnetlink_change_seq_adj(ct, cda);
1595                 if (err < 0)
1596                         return err;
1597         }
1598
1599         if (cda[CTA_LABELS]) {
1600                 err = ctnetlink_attach_labels(ct, cda);
1601                 if (err < 0)
1602                         return err;
1603         }
1604
1605         return 0;
1606 }
1607
1608 static struct nf_conn *
1609 ctnetlink_create_conntrack(struct net *net, u16 zone,
1610                            const struct nlattr * const cda[],
1611                            struct nf_conntrack_tuple *otuple,
1612                            struct nf_conntrack_tuple *rtuple,
1613                            u8 u3)
1614 {
1615         struct nf_conn *ct;
1616         int err = -EINVAL;
1617         struct nf_conntrack_helper *helper;
1618         struct nf_conn_tstamp *tstamp;
1619
1620         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1621         if (IS_ERR(ct))
1622                 return ERR_PTR(-ENOMEM);
1623
1624         if (!cda[CTA_TIMEOUT])
1625                 goto err1;
1626         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1627
1628         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1629
1630         rcu_read_lock();
1631         if (cda[CTA_HELP]) {
1632                 char *helpname = NULL;
1633                 struct nlattr *helpinfo = NULL;
1634
1635                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1636                 if (err < 0)
1637                         goto err2;
1638
1639                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1640                                                     nf_ct_protonum(ct));
1641                 if (helper == NULL) {
1642                         rcu_read_unlock();
1643 #ifdef CONFIG_MODULES
1644                         if (request_module("nfct-helper-%s", helpname) < 0) {
1645                                 err = -EOPNOTSUPP;
1646                                 goto err1;
1647                         }
1648
1649                         rcu_read_lock();
1650                         helper = __nf_conntrack_helper_find(helpname,
1651                                                             nf_ct_l3num(ct),
1652                                                             nf_ct_protonum(ct));
1653                         if (helper) {
1654                                 err = -EAGAIN;
1655                                 goto err2;
1656                         }
1657                         rcu_read_unlock();
1658 #endif
1659                         err = -EOPNOTSUPP;
1660                         goto err1;
1661                 } else {
1662                         struct nf_conn_help *help;
1663
1664                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1665                         if (help == NULL) {
1666                                 err = -ENOMEM;
1667                                 goto err2;
1668                         }
1669                         /* set private helper data if allowed. */
1670                         if (helper->from_nlattr)
1671                                 helper->from_nlattr(helpinfo, ct);
1672
1673                         /* not in hash table yet so not strictly necessary */
1674                         RCU_INIT_POINTER(help->helper, helper);
1675                 }
1676         } else {
1677                 /* try an implicit helper assignation */
1678                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1679                 if (err < 0)
1680                         goto err2;
1681         }
1682
1683         err = ctnetlink_setup_nat(ct, cda);
1684         if (err < 0)
1685                 goto err2;
1686
1687         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1688         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1689         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1690         nf_ct_labels_ext_add(ct);
1691
1692         /* we must add conntrack extensions before confirmation. */
1693         ct->status |= IPS_CONFIRMED;
1694
1695         if (cda[CTA_STATUS]) {
1696                 err = ctnetlink_change_status(ct, cda);
1697                 if (err < 0)
1698                         goto err2;
1699         }
1700
1701         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1702                 err = ctnetlink_change_seq_adj(ct, cda);
1703                 if (err < 0)
1704                         goto err2;
1705         }
1706
1707         memset(&ct->proto, 0, sizeof(ct->proto));
1708         if (cda[CTA_PROTOINFO]) {
1709                 err = ctnetlink_change_protoinfo(ct, cda);
1710                 if (err < 0)
1711                         goto err2;
1712         }
1713
1714 #if defined(CONFIG_NF_CONNTRACK_MARK)
1715         if (cda[CTA_MARK])
1716                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1717 #endif
1718
1719         /* setup master conntrack: this is a confirmed expectation */
1720         if (cda[CTA_TUPLE_MASTER]) {
1721                 struct nf_conntrack_tuple master;
1722                 struct nf_conntrack_tuple_hash *master_h;
1723                 struct nf_conn *master_ct;
1724
1725                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1726                 if (err < 0)
1727                         goto err2;
1728
1729                 master_h = nf_conntrack_find_get(net, zone, &master);
1730                 if (master_h == NULL) {
1731                         err = -ENOENT;
1732                         goto err2;
1733                 }
1734                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1735                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1736                 ct->master = master_ct;
1737         }
1738         tstamp = nf_conn_tstamp_find(ct);
1739         if (tstamp)
1740                 tstamp->start = ktime_get_real_ns();
1741
1742         err = nf_conntrack_hash_check_insert(ct);
1743         if (err < 0)
1744                 goto err2;
1745
1746         rcu_read_unlock();
1747
1748         return ct;
1749
1750 err2:
1751         rcu_read_unlock();
1752 err1:
1753         nf_conntrack_free(ct);
1754         return ERR_PTR(err);
1755 }
1756
1757 static int
1758 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1759                         const struct nlmsghdr *nlh,
1760                         const struct nlattr * const cda[])
1761 {
1762         struct net *net = sock_net(ctnl);
1763         struct nf_conntrack_tuple otuple, rtuple;
1764         struct nf_conntrack_tuple_hash *h = NULL;
1765         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1766         struct nf_conn *ct;
1767         u_int8_t u3 = nfmsg->nfgen_family;
1768         u16 zone;
1769         int err;
1770
1771         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1772         if (err < 0)
1773                 return err;
1774
1775         if (cda[CTA_TUPLE_ORIG]) {
1776                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1777                 if (err < 0)
1778                         return err;
1779         }
1780
1781         if (cda[CTA_TUPLE_REPLY]) {
1782                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1783                 if (err < 0)
1784                         return err;
1785         }
1786
1787         if (cda[CTA_TUPLE_ORIG])
1788                 h = nf_conntrack_find_get(net, zone, &otuple);
1789         else if (cda[CTA_TUPLE_REPLY])
1790                 h = nf_conntrack_find_get(net, zone, &rtuple);
1791
1792         if (h == NULL) {
1793                 err = -ENOENT;
1794                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1795                         enum ip_conntrack_events events;
1796
1797                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1798                                 return -EINVAL;
1799
1800                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1801                                                         &rtuple, u3);
1802                         if (IS_ERR(ct))
1803                                 return PTR_ERR(ct);
1804
1805                         err = 0;
1806                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1807                                 events = IPCT_RELATED;
1808                         else
1809                                 events = IPCT_NEW;
1810
1811                         if (cda[CTA_LABELS] &&
1812                             ctnetlink_attach_labels(ct, cda) == 0)
1813                                 events |= (1 << IPCT_LABEL);
1814
1815                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1816                                                       (1 << IPCT_ASSURED) |
1817                                                       (1 << IPCT_HELPER) |
1818                                                       (1 << IPCT_PROTOINFO) |
1819                                                       (1 << IPCT_SEQADJ) |
1820                                                       (1 << IPCT_MARK) | events,
1821                                                       ct, NETLINK_CB(skb).portid,
1822                                                       nlmsg_report(nlh));
1823                         nf_ct_put(ct);
1824                 }
1825
1826                 return err;
1827         }
1828         /* implicit 'else' */
1829
1830         err = -EEXIST;
1831         ct = nf_ct_tuplehash_to_ctrack(h);
1832         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1833                 spin_lock_bh(&nf_conntrack_expect_lock);
1834                 err = ctnetlink_change_conntrack(ct, cda);
1835                 spin_unlock_bh(&nf_conntrack_expect_lock);
1836                 if (err == 0) {
1837                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1838                                                       (1 << IPCT_ASSURED) |
1839                                                       (1 << IPCT_HELPER) |
1840                                                       (1 << IPCT_LABEL) |
1841                                                       (1 << IPCT_PROTOINFO) |
1842                                                       (1 << IPCT_SEQADJ) |
1843                                                       (1 << IPCT_MARK),
1844                                                       ct, NETLINK_CB(skb).portid,
1845                                                       nlmsg_report(nlh));
1846                 }
1847         }
1848
1849         nf_ct_put(ct);
1850         return err;
1851 }
1852
1853 static int
1854 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1855                                 __u16 cpu, const struct ip_conntrack_stat *st)
1856 {
1857         struct nlmsghdr *nlh;
1858         struct nfgenmsg *nfmsg;
1859         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1860
1861         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1862         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1863         if (nlh == NULL)
1864                 goto nlmsg_failure;
1865
1866         nfmsg = nlmsg_data(nlh);
1867         nfmsg->nfgen_family = AF_UNSPEC;
1868         nfmsg->version      = NFNETLINK_V0;
1869         nfmsg->res_id       = htons(cpu);
1870
1871         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1872             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1873             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1874             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1875             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1876             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1877             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1878             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1879             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1880                                 htonl(st->insert_failed)) ||
1881             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1882             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1883             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1884             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1885                                 htonl(st->search_restart)))
1886                 goto nla_put_failure;
1887
1888         nlmsg_end(skb, nlh);
1889         return skb->len;
1890
1891 nla_put_failure:
1892 nlmsg_failure:
1893         nlmsg_cancel(skb, nlh);
1894         return -1;
1895 }
1896
1897 static int
1898 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1899 {
1900         int cpu;
1901         struct net *net = sock_net(skb->sk);
1902
1903         if (cb->args[0] == nr_cpu_ids)
1904                 return 0;
1905
1906         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1907                 const struct ip_conntrack_stat *st;
1908
1909                 if (!cpu_possible(cpu))
1910                         continue;
1911
1912                 st = per_cpu_ptr(net->ct.stat, cpu);
1913                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1914                                                     NETLINK_CB(cb->skb).portid,
1915                                                     cb->nlh->nlmsg_seq,
1916                                                     cpu, st) < 0)
1917                                 break;
1918         }
1919         cb->args[0] = cpu;
1920
1921         return skb->len;
1922 }
1923
1924 static int
1925 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1926                       const struct nlmsghdr *nlh,
1927                       const struct nlattr * const cda[])
1928 {
1929         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1930                 struct netlink_dump_control c = {
1931                         .dump = ctnetlink_ct_stat_cpu_dump,
1932                 };
1933                 return netlink_dump_start(ctnl, skb, nlh, &c);
1934         }
1935
1936         return 0;
1937 }
1938
1939 static int
1940 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1941                             struct net *net)
1942 {
1943         struct nlmsghdr *nlh;
1944         struct nfgenmsg *nfmsg;
1945         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1946         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1947
1948         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1949         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1950         if (nlh == NULL)
1951                 goto nlmsg_failure;
1952
1953         nfmsg = nlmsg_data(nlh);
1954         nfmsg->nfgen_family = AF_UNSPEC;
1955         nfmsg->version      = NFNETLINK_V0;
1956         nfmsg->res_id       = 0;
1957
1958         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1959                 goto nla_put_failure;
1960
1961         nlmsg_end(skb, nlh);
1962         return skb->len;
1963
1964 nla_put_failure:
1965 nlmsg_failure:
1966         nlmsg_cancel(skb, nlh);
1967         return -1;
1968 }
1969
1970 static int
1971 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1972                   const struct nlmsghdr *nlh,
1973                   const struct nlattr * const cda[])
1974 {
1975         struct sk_buff *skb2;
1976         int err;
1977
1978         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1979         if (skb2 == NULL)
1980                 return -ENOMEM;
1981
1982         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1983                                           nlh->nlmsg_seq,
1984                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1985                                           sock_net(skb->sk));
1986         if (err <= 0)
1987                 goto free;
1988
1989         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1990         if (err < 0)
1991                 goto out;
1992
1993         return 0;
1994
1995 free:
1996         kfree_skb(skb2);
1997 out:
1998         /* this avoids a loop in nfnetlink. */
1999         return err == -EAGAIN ? -ENOBUFS : err;
2000 }
2001
2002 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2003         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2004         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2005         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2006         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2007         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2008         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2009                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2010         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2011         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2012         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2013         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2014         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2015 };
2016
2017 static struct nf_conntrack_expect *
2018 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2019                        struct nf_conntrack_helper *helper,
2020                        struct nf_conntrack_tuple *tuple,
2021                        struct nf_conntrack_tuple *mask);
2022
2023 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2024 static size_t
2025 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2026 {
2027         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2028                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2029                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2030                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2031                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2032                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2033                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2034                + nla_total_size(0) /* CTA_PROTOINFO */
2035                + nla_total_size(0) /* CTA_HELP */
2036                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2037                + ctnetlink_secctx_size(ct)
2038 #ifdef CONFIG_NF_NAT_NEEDED
2039                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2040                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2041 #endif
2042 #ifdef CONFIG_NF_CONNTRACK_MARK
2043                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2044 #endif
2045 #ifdef CONFIG_NF_CONNTRACK_ZONES
2046                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE */
2047 #endif
2048                + ctnetlink_proto_size(ct)
2049                ;
2050 }
2051
2052 static int
2053 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2054 {
2055         struct nlattr *nest_parms;
2056
2057         rcu_read_lock();
2058         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2059         if (!nest_parms)
2060                 goto nla_put_failure;
2061         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2062                 goto nla_put_failure;
2063         nla_nest_end(skb, nest_parms);
2064
2065         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2066         if (!nest_parms)
2067                 goto nla_put_failure;
2068         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2069                 goto nla_put_failure;
2070         nla_nest_end(skb, nest_parms);
2071
2072         if (nf_ct_zone(ct)) {
2073                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2074                         goto nla_put_failure;
2075         }
2076
2077         if (ctnetlink_dump_id(skb, ct) < 0)
2078                 goto nla_put_failure;
2079
2080         if (ctnetlink_dump_status(skb, ct) < 0)
2081                 goto nla_put_failure;
2082
2083         if (ctnetlink_dump_timeout(skb, ct) < 0)
2084                 goto nla_put_failure;
2085
2086         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2087                 goto nla_put_failure;
2088
2089         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2090                 goto nla_put_failure;
2091
2092 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2093         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2094                 goto nla_put_failure;
2095 #endif
2096         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2097                 goto nla_put_failure;
2098
2099         if ((ct->status & IPS_SEQ_ADJUST) &&
2100             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2101                 goto nla_put_failure;
2102
2103 #ifdef CONFIG_NF_CONNTRACK_MARK
2104         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2105                 goto nla_put_failure;
2106 #endif
2107         if (ctnetlink_dump_labels(skb, ct) < 0)
2108                 goto nla_put_failure;
2109         rcu_read_unlock();
2110         return 0;
2111
2112 nla_put_failure:
2113         rcu_read_unlock();
2114         return -ENOSPC;
2115 }
2116
2117 static int
2118 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2119 {
2120         int err;
2121
2122         if (cda[CTA_TIMEOUT]) {
2123                 err = ctnetlink_change_timeout(ct, cda);
2124                 if (err < 0)
2125                         return err;
2126         }
2127         if (cda[CTA_STATUS]) {
2128                 err = ctnetlink_change_status(ct, cda);
2129                 if (err < 0)
2130                         return err;
2131         }
2132         if (cda[CTA_HELP]) {
2133                 err = ctnetlink_change_helper(ct, cda);
2134                 if (err < 0)
2135                         return err;
2136         }
2137         if (cda[CTA_LABELS]) {
2138                 err = ctnetlink_attach_labels(ct, cda);
2139                 if (err < 0)
2140                         return err;
2141         }
2142 #if defined(CONFIG_NF_CONNTRACK_MARK)
2143         if (cda[CTA_MARK]) {
2144                 u32 mask = 0, mark, newmark;
2145                 if (cda[CTA_MARK_MASK])
2146                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2147
2148                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2149                 newmark = (ct->mark & mask) ^ mark;
2150                 if (newmark != ct->mark)
2151                         ct->mark = newmark;
2152         }
2153 #endif
2154         return 0;
2155 }
2156
2157 static int
2158 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2159 {
2160         struct nlattr *cda[CTA_MAX+1];
2161         int ret;
2162
2163         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2164         if (ret < 0)
2165                 return ret;
2166
2167         spin_lock_bh(&nf_conntrack_expect_lock);
2168         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2169         spin_unlock_bh(&nf_conntrack_expect_lock);
2170
2171         return ret;
2172 }
2173
2174 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2175                                        const struct nf_conn *ct,
2176                                        struct nf_conntrack_tuple *tuple,
2177                                        struct nf_conntrack_tuple *mask)
2178 {
2179         int err;
2180
2181         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2182                                     nf_ct_l3num(ct));
2183         if (err < 0)
2184                 return err;
2185
2186         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2187                                      nf_ct_l3num(ct));
2188 }
2189
2190 static int
2191 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2192                                 u32 portid, u32 report)
2193 {
2194         struct nlattr *cda[CTA_EXPECT_MAX+1];
2195         struct nf_conntrack_tuple tuple, mask;
2196         struct nf_conntrack_helper *helper = NULL;
2197         struct nf_conntrack_expect *exp;
2198         int err;
2199
2200         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2201         if (err < 0)
2202                 return err;
2203
2204         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2205                                           ct, &tuple, &mask);
2206         if (err < 0)
2207                 return err;
2208
2209         if (cda[CTA_EXPECT_HELP_NAME]) {
2210                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2211
2212                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2213                                                     nf_ct_protonum(ct));
2214                 if (helper == NULL)
2215                         return -EOPNOTSUPP;
2216         }
2217
2218         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2219                                      helper, &tuple, &mask);
2220         if (IS_ERR(exp))
2221                 return PTR_ERR(exp);
2222
2223         err = nf_ct_expect_related_report(exp, portid, report);
2224         if (err < 0) {
2225                 nf_ct_expect_put(exp);
2226                 return err;
2227         }
2228
2229         return 0;
2230 }
2231
2232 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2233         .build_size     = ctnetlink_nfqueue_build_size,
2234         .build          = ctnetlink_nfqueue_build,
2235         .parse          = ctnetlink_nfqueue_parse,
2236         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2237         .seq_adjust     = nf_ct_tcp_seqadj_set,
2238 };
2239 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2240
2241 /***********************************************************************
2242  * EXPECT
2243  ***********************************************************************/
2244
2245 static inline int
2246 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2247                          const struct nf_conntrack_tuple *tuple,
2248                          enum ctattr_expect type)
2249 {
2250         struct nlattr *nest_parms;
2251
2252         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2253         if (!nest_parms)
2254                 goto nla_put_failure;
2255         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2256                 goto nla_put_failure;
2257         nla_nest_end(skb, nest_parms);
2258
2259         return 0;
2260
2261 nla_put_failure:
2262         return -1;
2263 }
2264
2265 static inline int
2266 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2267                         const struct nf_conntrack_tuple *tuple,
2268                         const struct nf_conntrack_tuple_mask *mask)
2269 {
2270         int ret;
2271         struct nf_conntrack_l3proto *l3proto;
2272         struct nf_conntrack_l4proto *l4proto;
2273         struct nf_conntrack_tuple m;
2274         struct nlattr *nest_parms;
2275
2276         memset(&m, 0xFF, sizeof(m));
2277         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2278         m.src.u.all = mask->src.u.all;
2279         m.dst.protonum = tuple->dst.protonum;
2280
2281         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2282         if (!nest_parms)
2283                 goto nla_put_failure;
2284
2285         rcu_read_lock();
2286         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2287         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2288         if (ret >= 0) {
2289                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2290                                                tuple->dst.protonum);
2291         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2292         }
2293         rcu_read_unlock();
2294
2295         if (unlikely(ret < 0))
2296                 goto nla_put_failure;
2297
2298         nla_nest_end(skb, nest_parms);
2299
2300         return 0;
2301
2302 nla_put_failure:
2303         return -1;
2304 }
2305
2306 static const union nf_inet_addr any_addr;
2307
2308 static int
2309 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2310                           const struct nf_conntrack_expect *exp)
2311 {
2312         struct nf_conn *master = exp->master;
2313         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2314         struct nf_conn_help *help;
2315 #ifdef CONFIG_NF_NAT_NEEDED
2316         struct nlattr *nest_parms;
2317         struct nf_conntrack_tuple nat_tuple = {};
2318 #endif
2319         struct nf_ct_helper_expectfn *expfn;
2320
2321         if (timeout < 0)
2322                 timeout = 0;
2323
2324         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2325                 goto nla_put_failure;
2326         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2327                 goto nla_put_failure;
2328         if (ctnetlink_exp_dump_tuple(skb,
2329                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2330                                  CTA_EXPECT_MASTER) < 0)
2331                 goto nla_put_failure;
2332
2333 #ifdef CONFIG_NF_NAT_NEEDED
2334         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2335             exp->saved_proto.all) {
2336                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2337                 if (!nest_parms)
2338                         goto nla_put_failure;
2339
2340                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2341                         goto nla_put_failure;
2342
2343                 nat_tuple.src.l3num = nf_ct_l3num(master);
2344                 nat_tuple.src.u3 = exp->saved_addr;
2345                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2346                 nat_tuple.src.u = exp->saved_proto;
2347
2348                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2349                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2350                         goto nla_put_failure;
2351                 nla_nest_end(skb, nest_parms);
2352         }
2353 #endif
2354         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2355             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2356             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2357             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2358                 goto nla_put_failure;
2359         help = nfct_help(master);
2360         if (help) {
2361                 struct nf_conntrack_helper *helper;
2362
2363                 helper = rcu_dereference(help->helper);
2364                 if (helper &&
2365                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2366                         goto nla_put_failure;
2367         }
2368         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2369         if (expfn != NULL &&
2370             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2371                 goto nla_put_failure;
2372
2373         return 0;
2374
2375 nla_put_failure:
2376         return -1;
2377 }
2378
2379 static int
2380 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2381                         int event, const struct nf_conntrack_expect *exp)
2382 {
2383         struct nlmsghdr *nlh;
2384         struct nfgenmsg *nfmsg;
2385         unsigned int flags = portid ? NLM_F_MULTI : 0;
2386
2387         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2388         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2389         if (nlh == NULL)
2390                 goto nlmsg_failure;
2391
2392         nfmsg = nlmsg_data(nlh);
2393         nfmsg->nfgen_family = exp->tuple.src.l3num;
2394         nfmsg->version      = NFNETLINK_V0;
2395         nfmsg->res_id       = 0;
2396
2397         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2398                 goto nla_put_failure;
2399
2400         nlmsg_end(skb, nlh);
2401         return skb->len;
2402
2403 nlmsg_failure:
2404 nla_put_failure:
2405         nlmsg_cancel(skb, nlh);
2406         return -1;
2407 }
2408
2409 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2410 static int
2411 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2412 {
2413         struct nf_conntrack_expect *exp = item->exp;
2414         struct net *net = nf_ct_exp_net(exp);
2415         struct nlmsghdr *nlh;
2416         struct nfgenmsg *nfmsg;
2417         struct sk_buff *skb;
2418         unsigned int type, group;
2419         int flags = 0;
2420
2421         if (events & (1 << IPEXP_DESTROY)) {
2422                 type = IPCTNL_MSG_EXP_DELETE;
2423                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2424         } else if (events & (1 << IPEXP_NEW)) {
2425                 type = IPCTNL_MSG_EXP_NEW;
2426                 flags = NLM_F_CREATE|NLM_F_EXCL;
2427                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2428         } else
2429                 return 0;
2430
2431         if (!item->report && !nfnetlink_has_listeners(net, group))
2432                 return 0;
2433
2434         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2435         if (skb == NULL)
2436                 goto errout;
2437
2438         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2439         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2440         if (nlh == NULL)
2441                 goto nlmsg_failure;
2442
2443         nfmsg = nlmsg_data(nlh);
2444         nfmsg->nfgen_family = exp->tuple.src.l3num;
2445         nfmsg->version      = NFNETLINK_V0;
2446         nfmsg->res_id       = 0;
2447
2448         rcu_read_lock();
2449         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2450                 goto nla_put_failure;
2451         rcu_read_unlock();
2452
2453         nlmsg_end(skb, nlh);
2454         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2455         return 0;
2456
2457 nla_put_failure:
2458         rcu_read_unlock();
2459         nlmsg_cancel(skb, nlh);
2460 nlmsg_failure:
2461         kfree_skb(skb);
2462 errout:
2463         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2464         return 0;
2465 }
2466 #endif
2467 static int ctnetlink_exp_done(struct netlink_callback *cb)
2468 {
2469         if (cb->args[1])
2470                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2471         return 0;
2472 }
2473
2474 static int
2475 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2476 {
2477         struct net *net = sock_net(skb->sk);
2478         struct nf_conntrack_expect *exp, *last;
2479         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2480         u_int8_t l3proto = nfmsg->nfgen_family;
2481
2482         rcu_read_lock();
2483         last = (struct nf_conntrack_expect *)cb->args[1];
2484         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2485 restart:
2486                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2487                                      hnode) {
2488                         if (l3proto && exp->tuple.src.l3num != l3proto)
2489                                 continue;
2490                         if (cb->args[1]) {
2491                                 if (exp != last)
2492                                         continue;
2493                                 cb->args[1] = 0;
2494                         }
2495                         if (ctnetlink_exp_fill_info(skb,
2496                                                     NETLINK_CB(cb->skb).portid,
2497                                                     cb->nlh->nlmsg_seq,
2498                                                     IPCTNL_MSG_EXP_NEW,
2499                                                     exp) < 0) {
2500                                 if (!atomic_inc_not_zero(&exp->use))
2501                                         continue;
2502                                 cb->args[1] = (unsigned long)exp;
2503                                 goto out;
2504                         }
2505                 }
2506                 if (cb->args[1]) {
2507                         cb->args[1] = 0;
2508                         goto restart;
2509                 }
2510         }
2511 out:
2512         rcu_read_unlock();
2513         if (last)
2514                 nf_ct_expect_put(last);
2515
2516         return skb->len;
2517 }
2518
2519 static int
2520 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2521 {
2522         struct nf_conntrack_expect *exp, *last;
2523         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2524         struct nf_conn *ct = cb->data;
2525         struct nf_conn_help *help = nfct_help(ct);
2526         u_int8_t l3proto = nfmsg->nfgen_family;
2527
2528         if (cb->args[0])
2529                 return 0;
2530
2531         rcu_read_lock();
2532         last = (struct nf_conntrack_expect *)cb->args[1];
2533 restart:
2534         hlist_for_each_entry(exp, &help->expectations, lnode) {
2535                 if (l3proto && exp->tuple.src.l3num != l3proto)
2536                         continue;
2537                 if (cb->args[1]) {
2538                         if (exp != last)
2539                                 continue;
2540                         cb->args[1] = 0;
2541                 }
2542                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2543                                             cb->nlh->nlmsg_seq,
2544                                             IPCTNL_MSG_EXP_NEW,
2545                                             exp) < 0) {
2546                         if (!atomic_inc_not_zero(&exp->use))
2547                                 continue;
2548                         cb->args[1] = (unsigned long)exp;
2549                         goto out;
2550                 }
2551         }
2552         if (cb->args[1]) {
2553                 cb->args[1] = 0;
2554                 goto restart;
2555         }
2556         cb->args[0] = 1;
2557 out:
2558         rcu_read_unlock();
2559         if (last)
2560                 nf_ct_expect_put(last);
2561
2562         return skb->len;
2563 }
2564
2565 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2566                                  const struct nlmsghdr *nlh,
2567                                  const struct nlattr * const cda[])
2568 {
2569         int err;
2570         struct net *net = sock_net(ctnl);
2571         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2572         u_int8_t u3 = nfmsg->nfgen_family;
2573         struct nf_conntrack_tuple tuple;
2574         struct nf_conntrack_tuple_hash *h;
2575         struct nf_conn *ct;
2576         u16 zone = 0;
2577         struct netlink_dump_control c = {
2578                 .dump = ctnetlink_exp_ct_dump_table,
2579                 .done = ctnetlink_exp_done,
2580         };
2581
2582         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2583         if (err < 0)
2584                 return err;
2585
2586         if (cda[CTA_EXPECT_ZONE]) {
2587                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2588                 if (err < 0)
2589                         return err;
2590         }
2591
2592         h = nf_conntrack_find_get(net, zone, &tuple);
2593         if (!h)
2594                 return -ENOENT;
2595
2596         ct = nf_ct_tuplehash_to_ctrack(h);
2597         c.data = ct;
2598
2599         err = netlink_dump_start(ctnl, skb, nlh, &c);
2600         nf_ct_put(ct);
2601
2602         return err;
2603 }
2604
2605 static int
2606 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2607                      const struct nlmsghdr *nlh,
2608                      const struct nlattr * const cda[])
2609 {
2610         struct net *net = sock_net(ctnl);
2611         struct nf_conntrack_tuple tuple;
2612         struct nf_conntrack_expect *exp;
2613         struct sk_buff *skb2;
2614         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2615         u_int8_t u3 = nfmsg->nfgen_family;
2616         u16 zone;
2617         int err;
2618
2619         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2620                 if (cda[CTA_EXPECT_MASTER])
2621                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2622                 else {
2623                         struct netlink_dump_control c = {
2624                                 .dump = ctnetlink_exp_dump_table,
2625                                 .done = ctnetlink_exp_done,
2626                         };
2627                         return netlink_dump_start(ctnl, skb, nlh, &c);
2628                 }
2629         }
2630
2631         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2632         if (err < 0)
2633                 return err;
2634
2635         if (cda[CTA_EXPECT_TUPLE])
2636                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2637         else if (cda[CTA_EXPECT_MASTER])
2638                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2639         else
2640                 return -EINVAL;
2641
2642         if (err < 0)
2643                 return err;
2644
2645         exp = nf_ct_expect_find_get(net, zone, &tuple);
2646         if (!exp)
2647                 return -ENOENT;
2648
2649         if (cda[CTA_EXPECT_ID]) {
2650                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2651                 if (ntohl(id) != (u32)(unsigned long)exp) {
2652                         nf_ct_expect_put(exp);
2653                         return -ENOENT;
2654                 }
2655         }
2656
2657         err = -ENOMEM;
2658         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2659         if (skb2 == NULL) {
2660                 nf_ct_expect_put(exp);
2661                 goto out;
2662         }
2663
2664         rcu_read_lock();
2665         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2666                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2667         rcu_read_unlock();
2668         nf_ct_expect_put(exp);
2669         if (err <= 0)
2670                 goto free;
2671
2672         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2673         if (err < 0)
2674                 goto out;
2675
2676         return 0;
2677
2678 free:
2679         kfree_skb(skb2);
2680 out:
2681         /* this avoids a loop in nfnetlink. */
2682         return err == -EAGAIN ? -ENOBUFS : err;
2683 }
2684
2685 static int
2686 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2687                      const struct nlmsghdr *nlh,
2688                      const struct nlattr * const cda[])
2689 {
2690         struct net *net = sock_net(ctnl);
2691         struct nf_conntrack_expect *exp;
2692         struct nf_conntrack_tuple tuple;
2693         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2694         struct hlist_node *next;
2695         u_int8_t u3 = nfmsg->nfgen_family;
2696         unsigned int i;
2697         u16 zone;
2698         int err;
2699
2700         if (cda[CTA_EXPECT_TUPLE]) {
2701                 /* delete a single expect by tuple */
2702                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2703                 if (err < 0)
2704                         return err;
2705
2706                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2707                 if (err < 0)
2708                         return err;
2709
2710                 /* bump usage count to 2 */
2711                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2712                 if (!exp)
2713                         return -ENOENT;
2714
2715                 if (cda[CTA_EXPECT_ID]) {
2716                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2717                         if (ntohl(id) != (u32)(unsigned long)exp) {
2718                                 nf_ct_expect_put(exp);
2719                                 return -ENOENT;
2720                         }
2721                 }
2722
2723                 /* after list removal, usage count == 1 */
2724                 spin_lock_bh(&nf_conntrack_expect_lock);
2725                 if (del_timer(&exp->timeout)) {
2726                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2727                                                    nlmsg_report(nlh));
2728                         nf_ct_expect_put(exp);
2729                 }
2730                 spin_unlock_bh(&nf_conntrack_expect_lock);
2731                 /* have to put what we 'get' above.
2732                  * after this line usage count == 0 */
2733                 nf_ct_expect_put(exp);
2734         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2735                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2736                 struct nf_conn_help *m_help;
2737
2738                 /* delete all expectations for this helper */
2739                 spin_lock_bh(&nf_conntrack_expect_lock);
2740                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2741                         hlist_for_each_entry_safe(exp, next,
2742                                                   &net->ct.expect_hash[i],
2743                                                   hnode) {
2744                                 m_help = nfct_help(exp->master);
2745                                 if (!strcmp(m_help->helper->name, name) &&
2746                                     del_timer(&exp->timeout)) {
2747                                         nf_ct_unlink_expect_report(exp,
2748                                                         NETLINK_CB(skb).portid,
2749                                                         nlmsg_report(nlh));
2750                                         nf_ct_expect_put(exp);
2751                                 }
2752                         }
2753                 }
2754                 spin_unlock_bh(&nf_conntrack_expect_lock);
2755         } else {
2756                 /* This basically means we have to flush everything*/
2757                 spin_lock_bh(&nf_conntrack_expect_lock);
2758                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2759                         hlist_for_each_entry_safe(exp, next,
2760                                                   &net->ct.expect_hash[i],
2761                                                   hnode) {
2762                                 if (del_timer(&exp->timeout)) {
2763                                         nf_ct_unlink_expect_report(exp,
2764                                                         NETLINK_CB(skb).portid,
2765                                                         nlmsg_report(nlh));
2766                                         nf_ct_expect_put(exp);
2767                                 }
2768                         }
2769                 }
2770                 spin_unlock_bh(&nf_conntrack_expect_lock);
2771         }
2772
2773         return 0;
2774 }
2775 static int
2776 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2777                         const struct nlattr * const cda[])
2778 {
2779         if (cda[CTA_EXPECT_TIMEOUT]) {
2780                 if (!del_timer(&x->timeout))
2781                         return -ETIME;
2782
2783                 x->timeout.expires = jiffies +
2784                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2785                 add_timer(&x->timeout);
2786         }
2787         return 0;
2788 }
2789
2790 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2791         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2792         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2793 };
2794
2795 static int
2796 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2797                            struct nf_conntrack_expect *exp,
2798                            u_int8_t u3)
2799 {
2800 #ifdef CONFIG_NF_NAT_NEEDED
2801         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2802         struct nf_conntrack_tuple nat_tuple = {};
2803         int err;
2804
2805         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2806         if (err < 0)
2807                 return err;
2808
2809         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2810                 return -EINVAL;
2811
2812         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2813                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2814         if (err < 0)
2815                 return err;
2816
2817         exp->saved_addr = nat_tuple.src.u3;
2818         exp->saved_proto = nat_tuple.src.u;
2819         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2820
2821         return 0;
2822 #else
2823         return -EOPNOTSUPP;
2824 #endif
2825 }
2826
2827 static struct nf_conntrack_expect *
2828 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2829                        struct nf_conntrack_helper *helper,
2830                        struct nf_conntrack_tuple *tuple,
2831                        struct nf_conntrack_tuple *mask)
2832 {
2833         u_int32_t class = 0;
2834         struct nf_conntrack_expect *exp;
2835         struct nf_conn_help *help;
2836         int err;
2837
2838         if (cda[CTA_EXPECT_CLASS] && helper) {
2839                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2840                 if (class > helper->expect_class_max)
2841                         return ERR_PTR(-EINVAL);
2842         }
2843         exp = nf_ct_expect_alloc(ct);
2844         if (!exp)
2845                 return ERR_PTR(-ENOMEM);
2846
2847         help = nfct_help(ct);
2848         if (!help) {
2849                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2850                         err = -EINVAL;
2851                         goto err_out;
2852                 }
2853                 exp->timeout.expires =
2854                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2855
2856                 exp->flags = NF_CT_EXPECT_USERSPACE;
2857                 if (cda[CTA_EXPECT_FLAGS]) {
2858                         exp->flags |=
2859                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2860                 }
2861         } else {
2862                 if (cda[CTA_EXPECT_FLAGS]) {
2863                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2864                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2865                 } else
2866                         exp->flags = 0;
2867         }
2868         if (cda[CTA_EXPECT_FN]) {
2869                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2870                 struct nf_ct_helper_expectfn *expfn;
2871
2872                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2873                 if (expfn == NULL) {
2874                         err = -EINVAL;
2875                         goto err_out;
2876                 }
2877                 exp->expectfn = expfn->expectfn;
2878         } else
2879                 exp->expectfn = NULL;
2880
2881         exp->class = class;
2882         exp->master = ct;
2883         exp->helper = helper;
2884         exp->tuple = *tuple;
2885         exp->mask.src.u3 = mask->src.u3;
2886         exp->mask.src.u.all = mask->src.u.all;
2887
2888         if (cda[CTA_EXPECT_NAT]) {
2889                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2890                                                  exp, nf_ct_l3num(ct));
2891                 if (err < 0)
2892                         goto err_out;
2893         }
2894         return exp;
2895 err_out:
2896         nf_ct_expect_put(exp);
2897         return ERR_PTR(err);
2898 }
2899
2900 static int
2901 ctnetlink_create_expect(struct net *net, u16 zone,
2902                         const struct nlattr * const cda[],
2903                         u_int8_t u3, u32 portid, int report)
2904 {
2905         struct nf_conntrack_tuple tuple, mask, master_tuple;
2906         struct nf_conntrack_tuple_hash *h = NULL;
2907         struct nf_conntrack_helper *helper = NULL;
2908         struct nf_conntrack_expect *exp;
2909         struct nf_conn *ct;
2910         int err;
2911
2912         /* caller guarantees that those three CTA_EXPECT_* exist */
2913         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2914         if (err < 0)
2915                 return err;
2916         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2917         if (err < 0)
2918                 return err;
2919         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2920         if (err < 0)
2921                 return err;
2922
2923         /* Look for master conntrack of this expectation */
2924         h = nf_conntrack_find_get(net, zone, &master_tuple);
2925         if (!h)
2926                 return -ENOENT;
2927         ct = nf_ct_tuplehash_to_ctrack(h);
2928
2929         if (cda[CTA_EXPECT_HELP_NAME]) {
2930                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2931
2932                 helper = __nf_conntrack_helper_find(helpname, u3,
2933                                                     nf_ct_protonum(ct));
2934                 if (helper == NULL) {
2935 #ifdef CONFIG_MODULES
2936                         if (request_module("nfct-helper-%s", helpname) < 0) {
2937                                 err = -EOPNOTSUPP;
2938                                 goto err_ct;
2939                         }
2940                         helper = __nf_conntrack_helper_find(helpname, u3,
2941                                                             nf_ct_protonum(ct));
2942                         if (helper) {
2943                                 err = -EAGAIN;
2944                                 goto err_ct;
2945                         }
2946 #endif
2947                         err = -EOPNOTSUPP;
2948                         goto err_ct;
2949                 }
2950         }
2951
2952         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
2953         if (IS_ERR(exp)) {
2954                 err = PTR_ERR(exp);
2955                 goto err_ct;
2956         }
2957
2958         err = nf_ct_expect_related_report(exp, portid, report);
2959         if (err < 0)
2960                 goto err_exp;
2961
2962         return 0;
2963 err_exp:
2964         nf_ct_expect_put(exp);
2965 err_ct:
2966         nf_ct_put(ct);
2967         return err;
2968 }
2969
2970 static int
2971 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2972                      const struct nlmsghdr *nlh,
2973                      const struct nlattr * const cda[])
2974 {
2975         struct net *net = sock_net(ctnl);
2976         struct nf_conntrack_tuple tuple;
2977         struct nf_conntrack_expect *exp;
2978         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2979         u_int8_t u3 = nfmsg->nfgen_family;
2980         u16 zone;
2981         int err;
2982
2983         if (!cda[CTA_EXPECT_TUPLE]
2984             || !cda[CTA_EXPECT_MASK]
2985             || !cda[CTA_EXPECT_MASTER])
2986                 return -EINVAL;
2987
2988         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2989         if (err < 0)
2990                 return err;
2991
2992         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2993         if (err < 0)
2994                 return err;
2995
2996         spin_lock_bh(&nf_conntrack_expect_lock);
2997         exp = __nf_ct_expect_find(net, zone, &tuple);
2998
2999         if (!exp) {
3000                 spin_unlock_bh(&nf_conntrack_expect_lock);
3001                 err = -ENOENT;
3002                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3003                         err = ctnetlink_create_expect(net, zone, cda,
3004                                                       u3,
3005                                                       NETLINK_CB(skb).portid,
3006                                                       nlmsg_report(nlh));
3007                 }
3008                 return err;
3009         }
3010
3011         err = -EEXIST;
3012         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3013                 err = ctnetlink_change_expect(exp, cda);
3014         spin_unlock_bh(&nf_conntrack_expect_lock);
3015
3016         return err;
3017 }
3018
3019 static int
3020 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3021                              const struct ip_conntrack_stat *st)
3022 {
3023         struct nlmsghdr *nlh;
3024         struct nfgenmsg *nfmsg;
3025         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3026
3027         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3028         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3029         if (nlh == NULL)
3030                 goto nlmsg_failure;
3031
3032         nfmsg = nlmsg_data(nlh);
3033         nfmsg->nfgen_family = AF_UNSPEC;
3034         nfmsg->version      = NFNETLINK_V0;
3035         nfmsg->res_id       = htons(cpu);
3036
3037         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3038             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3039             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3040                 goto nla_put_failure;
3041
3042         nlmsg_end(skb, nlh);
3043         return skb->len;
3044
3045 nla_put_failure:
3046 nlmsg_failure:
3047         nlmsg_cancel(skb, nlh);
3048         return -1;
3049 }
3050
3051 static int
3052 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3053 {
3054         int cpu;
3055         struct net *net = sock_net(skb->sk);
3056
3057         if (cb->args[0] == nr_cpu_ids)
3058                 return 0;
3059
3060         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3061                 const struct ip_conntrack_stat *st;
3062
3063                 if (!cpu_possible(cpu))
3064                         continue;
3065
3066                 st = per_cpu_ptr(net->ct.stat, cpu);
3067                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3068                                                  cb->nlh->nlmsg_seq,
3069                                                  cpu, st) < 0)
3070                         break;
3071         }
3072         cb->args[0] = cpu;
3073
3074         return skb->len;
3075 }
3076
3077 static int
3078 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3079                        const struct nlmsghdr *nlh,
3080                        const struct nlattr * const cda[])
3081 {
3082         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3083                 struct netlink_dump_control c = {
3084                         .dump = ctnetlink_exp_stat_cpu_dump,
3085                 };
3086                 return netlink_dump_start(ctnl, skb, nlh, &c);
3087         }
3088
3089         return 0;
3090 }
3091
3092 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3093 static struct nf_ct_event_notifier ctnl_notifier = {
3094         .fcn = ctnetlink_conntrack_event,
3095 };
3096
3097 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3098         .fcn = ctnetlink_expect_event,
3099 };
3100 #endif
3101
3102 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3103         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3104                                             .attr_count = CTA_MAX,
3105                                             .policy = ct_nla_policy },
3106         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3107                                             .attr_count = CTA_MAX,
3108                                             .policy = ct_nla_policy },
3109         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3110                                             .attr_count = CTA_MAX,
3111                                             .policy = ct_nla_policy },
3112         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3113                                             .attr_count = CTA_MAX,
3114                                             .policy = ct_nla_policy },
3115         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3116         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3117         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3118         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3119 };
3120
3121 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3122         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3123                                             .attr_count = CTA_EXPECT_MAX,
3124                                             .policy = exp_nla_policy },
3125         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3126                                             .attr_count = CTA_EXPECT_MAX,
3127                                             .policy = exp_nla_policy },
3128         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3129                                             .attr_count = CTA_EXPECT_MAX,
3130                                             .policy = exp_nla_policy },
3131         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3132 };
3133
3134 static const struct nfnetlink_subsystem ctnl_subsys = {
3135         .name                           = "conntrack",
3136         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3137         .cb_count                       = IPCTNL_MSG_MAX,
3138         .cb                             = ctnl_cb,
3139 };
3140
3141 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3142         .name                           = "conntrack_expect",
3143         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3144         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3145         .cb                             = ctnl_exp_cb,
3146 };
3147
3148 MODULE_ALIAS("ip_conntrack_netlink");
3149 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3150 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3151
3152 static int __net_init ctnetlink_net_init(struct net *net)
3153 {
3154 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3155         int ret;
3156
3157         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3158         if (ret < 0) {
3159                 pr_err("ctnetlink_init: cannot register notifier.\n");
3160                 goto err_out;
3161         }
3162
3163         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3164         if (ret < 0) {
3165                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3166                 goto err_unreg_notifier;
3167         }
3168 #endif
3169         return 0;
3170
3171 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3172 err_unreg_notifier:
3173         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3174 err_out:
3175         return ret;
3176 #endif
3177 }
3178
3179 static void ctnetlink_net_exit(struct net *net)
3180 {
3181 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3182         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3183         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3184 #endif
3185 }
3186
3187 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3188 {
3189         struct net *net;
3190
3191         list_for_each_entry(net, net_exit_list, exit_list)
3192                 ctnetlink_net_exit(net);
3193 }
3194
3195 static struct pernet_operations ctnetlink_net_ops = {
3196         .init           = ctnetlink_net_init,
3197         .exit_batch     = ctnetlink_net_exit_batch,
3198 };
3199
3200 static int __init ctnetlink_init(void)
3201 {
3202         int ret;
3203
3204         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3205         ret = nfnetlink_subsys_register(&ctnl_subsys);
3206         if (ret < 0) {
3207                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3208                 goto err_out;
3209         }
3210
3211         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3212         if (ret < 0) {
3213                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3214                 goto err_unreg_subsys;
3215         }
3216
3217         ret = register_pernet_subsys(&ctnetlink_net_ops);
3218         if (ret < 0) {
3219                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3220                 goto err_unreg_exp_subsys;
3221         }
3222 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3223         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3224         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3225 #endif
3226         return 0;
3227
3228 err_unreg_exp_subsys:
3229         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3230 err_unreg_subsys:
3231         nfnetlink_subsys_unregister(&ctnl_subsys);
3232 err_out:
3233         return ret;
3234 }
3235
3236 static void __exit ctnetlink_exit(void)
3237 {
3238         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3239
3240         unregister_pernet_subsys(&ctnetlink_net_ops);
3241         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3242         nfnetlink_subsys_unregister(&ctnl_subsys);
3243 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3244         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3245 #endif
3246 }
3247
3248 module_init(ctnetlink_init);
3249 module_exit(ctnetlink_exit);