net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #ifdef CONFIG_NF_NAT_NEEDED
  49 #include <net/netfilter/nf_nat_core.h>
  50 #include <net/netfilter/nf_nat_l4proto.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 MODULE_LICENSE("GPL");
  58
  59 static char __initdata version[] = "0.93";
  60
  61 static inline int
  62 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  63                             const struct nf_conntrack_tuple *tuple,
  64                             struct nf_conntrack_l4proto *l4proto)
  65 {
  66         int ret = 0;
  67         struct nlattr *nest_parms;
  68
  69         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  70         if (!nest_parms)
  71                 goto nla_put_failure;
  72         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  73                 goto nla_put_failure;
  74
  75         if (likely(l4proto->tuple_to_nlattr))
  76                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  77
  78         nla_nest_end(skb, nest_parms);
  79
  80         return ret;
  81
  82 nla_put_failure:
  83         return -1;
  84 }
  85
  86 static inline int
  87 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  88                          const struct nf_conntrack_tuple *tuple,
  89                          struct nf_conntrack_l3proto *l3proto)
  90 {
  91         int ret = 0;
  92         struct nlattr *nest_parms;
  93
  94         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  95         if (!nest_parms)
  96                 goto nla_put_failure;
  97
  98         if (likely(l3proto->tuple_to_nlattr))
  99                 ret = l3proto->tuple_to_nlattr(skb, tuple);
 100
 101         nla_nest_end(skb, nest_parms);
 102
 103         return ret;
 104
 105 nla_put_failure:
 106         return -1;
 107 }
 108
 109 static int
 110 ctnetlink_dump_tuples(struct sk_buff *skb,
 111                       const struct nf_conntrack_tuple *tuple)
 112 {
 113         int ret;
 114         struct nf_conntrack_l3proto *l3proto;
 115         struct nf_conntrack_l4proto *l4proto;
 116
 117         rcu_read_lock();
 118         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 119         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 120
 121         if (ret >= 0) {
 122                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 123                                                tuple->dst.protonum);
 124                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 125         }
 126         rcu_read_unlock();
 127         return ret;
 128 }
 129
 130 static inline int
 131 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 132 {
 133         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 134                 goto nla_put_failure;
 135         return 0;
 136
 137 nla_put_failure:
 138         return -1;
 139 }
 140
 141 static inline int
 142 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 143 {
 144         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 145
 146         if (timeout < 0)
 147                 timeout = 0;
 148
 149         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 150                 goto nla_put_failure;
 151         return 0;
 152
 153 nla_put_failure:
 154         return -1;
 155 }
 156
 157 static inline int
 158 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 159 {
 160         struct nf_conntrack_l4proto *l4proto;
 161         struct nlattr *nest_proto;
 162         int ret;
 163
 164         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 165         if (!l4proto->to_nlattr)
 166                 return 0;
 167
 168         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 169         if (!nest_proto)
 170                 goto nla_put_failure;
 171
 172         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 173
 174         nla_nest_end(skb, nest_proto);
 175
 176         return ret;
 177
 178 nla_put_failure:
 179         return -1;
 180 }
 181
 182 static inline int
 183 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 184 {
 185         struct nlattr *nest_helper;
 186         const struct nf_conn_help *help = nfct_help(ct);
 187         struct nf_conntrack_helper *helper;
 188
 189         if (!help)
 190                 return 0;
 191
 192         helper = rcu_dereference(help->helper);
 193         if (!helper)
 194                 goto out;
 195
 196         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 197         if (!nest_helper)
 198                 goto nla_put_failure;
 199         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 200                 goto nla_put_failure;
 201
 202         if (helper->to_nlattr)
 203                 helper->to_nlattr(skb, ct);
 204
 205         nla_nest_end(skb, nest_helper);
 206 out:
 207         return 0;
 208
 209 nla_put_failure:
 210         return -1;
 211 }
 212
 213 static int
 214 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 215               enum ip_conntrack_dir dir, int type)
 216 {
 217         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 218         struct nf_conn_counter *counter = acct->counter;
 219         struct nlattr *nest_count;
 220         u64 pkts, bytes;
 221
 222         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 223                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 224                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 225         } else {
 226                 pkts = atomic64_read(&counter[dir].packets);
 227                 bytes = atomic64_read(&counter[dir].bytes);
 228         }
 229
 230         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 231         if (!nest_count)
 232                 goto nla_put_failure;
 233
 234         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 235             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 236                 goto nla_put_failure;
 237
 238         nla_nest_end(skb, nest_count);
 239
 240         return 0;
 241
 242 nla_put_failure:
 243         return -1;
 244 }
 245
 246 static int
 247 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 248 {
 249         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 250
 251         if (!acct)
 252                 return 0;
 253
 254         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 255                 return -1;
 256         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 257                 return -1;
 258
 259         return 0;
 260 }
 261
 262 static int
 263 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 264 {
 265         struct nlattr *nest_count;
 266         const struct nf_conn_tstamp *tstamp;
 267
 268         tstamp = nf_conn_tstamp_find(ct);
 269         if (!tstamp)
 270                 return 0;
 271
 272         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 273         if (!nest_count)
 274                 goto nla_put_failure;
 275
 276         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 277             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 278                                                cpu_to_be64(tstamp->stop))))
 279                 goto nla_put_failure;
 280         nla_nest_end(skb, nest_count);
 281
 282         return 0;
 283
 284 nla_put_failure:
 285         return -1;
 286 }
 287
 288 #ifdef CONFIG_NF_CONNTRACK_MARK
 289 static inline int
 290 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 291 {
 292         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 293                 goto nla_put_failure;
 294         return 0;
 295
 296 nla_put_failure:
 297         return -1;
 298 }
 299 #else
 300 #define ctnetlink_dump_mark(a, b) (0)
 301 #endif
 302
 303 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 304 static inline int
 305 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 306 {
 307         struct nlattr *nest_secctx;
 308         int len, ret;
 309         char *secctx;
 310
 311         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 312         if (ret)
 313                 return 0;
 314
 315         ret = -1;
 316         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 317         if (!nest_secctx)
 318                 goto nla_put_failure;
 319
 320         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 321                 goto nla_put_failure;
 322         nla_nest_end(skb, nest_secctx);
 323
 324         ret = 0;
 325 nla_put_failure:
 326         security_release_secctx(secctx, len);
 327         return ret;
 328 }
 329 #else
 330 #define ctnetlink_dump_secctx(a, b) (0)
 331 #endif
 332
 333 #ifdef CONFIG_NF_CONNTRACK_LABELS
 334 static int ctnetlink_label_size(const struct nf_conn *ct)
 335 {
 336         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 337
 338         if (!labels)
 339                 return 0;
 340         return nla_total_size(labels->words * sizeof(long));
 341 }
 342
 343 static int
 344 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 345 {
 346         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 347         unsigned int len, i;
 348
 349         if (!labels)
 350                 return 0;
 351
 352         len = labels->words * sizeof(long);
 353         i = 0;
 354         do {
 355                 if (labels->bits[i] != 0)
 356                         return nla_put(skb, CTA_LABELS, len, labels->bits);
 357                 i++;
 358         } while (i < labels->words);
 359
 360         return 0;
 361 }
 362 #else
 363 #define ctnetlink_dump_labels(a, b) (0)
 364 #define ctnetlink_label_size(a) (0)
 365 #endif
 366
 367 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 368
 369 static inline int
 370 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 371 {
 372         struct nlattr *nest_parms;
 373
 374         if (!(ct->status & IPS_EXPECTED))
 375                 return 0;
 376
 377         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 378         if (!nest_parms)
 379                 goto nla_put_failure;
 380         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 381                 goto nla_put_failure;
 382         nla_nest_end(skb, nest_parms);
 383
 384         return 0;
 385
 386 nla_put_failure:
 387         return -1;
 388 }
 389
 390 static int
 391 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 392 {
 393         struct nlattr *nest_parms;
 394
 395         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 396         if (!nest_parms)
 397                 goto nla_put_failure;
 398
 399         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 400                          htonl(seq->correction_pos)) ||
 401             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 402                          htonl(seq->offset_before)) ||
 403             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 404                          htonl(seq->offset_after)))
 405                 goto nla_put_failure;
 406
 407         nla_nest_end(skb, nest_parms);
 408
 409         return 0;
 410
 411 nla_put_failure:
 412         return -1;
 413 }
 414
 415 static inline int
 416 ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 417 {
 418         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 419         struct nf_ct_seqadj *seq;
 420
 421         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 422                 return 0;
 423
 424         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 425         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 426                 return -1;
 427
 428         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 429         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 430                 return -1;
 431
 432         return 0;
 433 }
 434
 435 static inline int
 436 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 437 {
 438         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 439                 goto nla_put_failure;
 440         return 0;
 441
 442 nla_put_failure:
 443         return -1;
 444 }
 445
 446 static inline int
 447 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 448 {
 449         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 450                 goto nla_put_failure;
 451         return 0;
 452
 453 nla_put_failure:
 454         return -1;
 455 }
 456
 457 static int
 458 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 459                     struct nf_conn *ct)
 460 {
 461         struct nlmsghdr *nlh;
 462         struct nfgenmsg *nfmsg;
 463         struct nlattr *nest_parms;
 464         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 465
 466         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 467         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 468         if (nlh == NULL)
 469                 goto nlmsg_failure;
 470
 471         nfmsg = nlmsg_data(nlh);
 472         nfmsg->nfgen_family = nf_ct_l3num(ct);
 473         nfmsg->version      = NFNETLINK_V0;
 474         nfmsg->res_id       = 0;
 475
 476         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 477         if (!nest_parms)
 478                 goto nla_put_failure;
 479         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 480                 goto nla_put_failure;
 481         nla_nest_end(skb, nest_parms);
 482
 483         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 484         if (!nest_parms)
 485                 goto nla_put_failure;
 486         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 487                 goto nla_put_failure;
 488         nla_nest_end(skb, nest_parms);
 489
 490         if (nf_ct_zone(ct) &&
 491             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 492                 goto nla_put_failure;
 493
 494         if (ctnetlink_dump_status(skb, ct) < 0 ||
 495             ctnetlink_dump_timeout(skb, ct) < 0 ||
 496             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 497             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 498             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 499             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 500             ctnetlink_dump_mark(skb, ct) < 0 ||
 501             ctnetlink_dump_secctx(skb, ct) < 0 ||
 502             ctnetlink_dump_labels(skb, ct) < 0 ||
 503             ctnetlink_dump_id(skb, ct) < 0 ||
 504             ctnetlink_dump_use(skb, ct) < 0 ||
 505             ctnetlink_dump_master(skb, ct) < 0 ||
 506             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 507                 goto nla_put_failure;
 508
 509         nlmsg_end(skb, nlh);
 510         return skb->len;
 511
 512 nlmsg_failure:
 513 nla_put_failure:
 514         nlmsg_cancel(skb, nlh);
 515         return -1;
 516 }
 517
 518 static inline size_t
 519 ctnetlink_proto_size(const struct nf_conn *ct)
 520 {
 521         struct nf_conntrack_l3proto *l3proto;
 522         struct nf_conntrack_l4proto *l4proto;
 523         size_t len = 0;
 524
 525         rcu_read_lock();
 526         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 527         len += l3proto->nla_size;
 528
 529         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 530         len += l4proto->nla_size;
 531         rcu_read_unlock();
 532
 533         return len;
 534 }
 535
 536 static inline size_t
 537 ctnetlink_acct_size(const struct nf_conn *ct)
 538 {
 539         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 540                 return 0;
 541         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 542                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 543                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 544                ;
 545 }
 546
 547 static inline int
 548 ctnetlink_secctx_size(const struct nf_conn *ct)
 549 {
 550 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 551         int len, ret;
 552
 553         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 554         if (ret)
 555                 return 0;
 556
 557         return nla_total_size(0) /* CTA_SECCTX */
 558                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 559 #else
 560         return 0;
 561 #endif
 562 }
 563
 564 static inline size_t
 565 ctnetlink_timestamp_size(const struct nf_conn *ct)
 566 {
 567 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 568         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 569                 return 0;
 570         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 571 #else
 572         return 0;
 573 #endif
 574 }
 575
 576 static inline size_t
 577 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 578 {
 579         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 580                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 581                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 582                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 583                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 584                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 585                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 586                + ctnetlink_acct_size(ct)
 587                + ctnetlink_timestamp_size(ct)
 588                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 589                + nla_total_size(0) /* CTA_PROTOINFO */
 590                + nla_total_size(0) /* CTA_HELP */
 591                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 592                + ctnetlink_secctx_size(ct)
 593 #ifdef CONFIG_NF_NAT_NEEDED
 594                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 595                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 596 #endif
 597 #ifdef CONFIG_NF_CONNTRACK_MARK
 598                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 599 #endif
 600                + ctnetlink_proto_size(ct)
 601                + ctnetlink_label_size(ct)
 602                ;
 603 }
 604
 605 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 606 static int
 607 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 608 {
 609         struct net *net;
 610         struct nlmsghdr *nlh;
 611         struct nfgenmsg *nfmsg;
 612         struct nlattr *nest_parms;
 613         struct nf_conn *ct = item->ct;
 614         struct sk_buff *skb;
 615         unsigned int type;
 616         unsigned int flags = 0, group;
 617         int err;
 618
 619         /* ignore our fake conntrack entry */
 620         if (nf_ct_is_untracked(ct))
 621                 return 0;
 622
 623         if (events & (1 << IPCT_DESTROY)) {
 624                 type = IPCTNL_MSG_CT_DELETE;
 625                 group = NFNLGRP_CONNTRACK_DESTROY;
 626         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 627                 type = IPCTNL_MSG_CT_NEW;
 628                 flags = NLM_F_CREATE|NLM_F_EXCL;
 629                 group = NFNLGRP_CONNTRACK_NEW;
 630         } else  if (events) {
 631                 type = IPCTNL_MSG_CT_NEW;
 632                 group = NFNLGRP_CONNTRACK_UPDATE;
 633         } else
 634                 return 0;
 635
 636         net = nf_ct_net(ct);
 637         if (!item->report && !nfnetlink_has_listeners(net, group))
 638                 return 0;
 639
 640         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 641         if (skb == NULL)
 642                 goto errout;
 643
 644         type |= NFNL_SUBSYS_CTNETLINK << 8;
 645         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 646         if (nlh == NULL)
 647                 goto nlmsg_failure;
 648
 649         nfmsg = nlmsg_data(nlh);
 650         nfmsg->nfgen_family = nf_ct_l3num(ct);
 651         nfmsg->version  = NFNETLINK_V0;
 652         nfmsg->res_id   = 0;
 653
 654         rcu_read_lock();
 655         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 656         if (!nest_parms)
 657                 goto nla_put_failure;
 658         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 659                 goto nla_put_failure;
 660         nla_nest_end(skb, nest_parms);
 661
 662         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 663         if (!nest_parms)
 664                 goto nla_put_failure;
 665         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 666                 goto nla_put_failure;
 667         nla_nest_end(skb, nest_parms);
 668
 669         if (nf_ct_zone(ct) &&
 670             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 671                 goto nla_put_failure;
 672
 673         if (ctnetlink_dump_id(skb, ct) < 0)
 674                 goto nla_put_failure;
 675
 676         if (ctnetlink_dump_status(skb, ct) < 0)
 677                 goto nla_put_failure;
 678
 679         if (events & (1 << IPCT_DESTROY)) {
 680                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 681                     ctnetlink_dump_timestamp(skb, ct) < 0)
 682                         goto nla_put_failure;
 683         } else {
 684                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 685                         goto nla_put_failure;
 686
 687                 if (events & (1 << IPCT_PROTOINFO)
 688                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 689                         goto nla_put_failure;
 690
 691                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 692                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 693                         goto nla_put_failure;
 694
 695 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 696                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 697                     && ctnetlink_dump_secctx(skb, ct) < 0)
 698                         goto nla_put_failure;
 699 #endif
 700                 if (events & (1 << IPCT_LABEL) &&
 701                      ctnetlink_dump_labels(skb, ct) < 0)
 702                         goto nla_put_failure;
 703
 704                 if (events & (1 << IPCT_RELATED) &&
 705                     ctnetlink_dump_master(skb, ct) < 0)
 706                         goto nla_put_failure;
 707
 708                 if (events & (1 << IPCT_SEQADJ) &&
 709                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 710                         goto nla_put_failure;
 711         }
 712
 713 #ifdef CONFIG_NF_CONNTRACK_MARK
 714         if ((events & (1 << IPCT_MARK) || ct->mark)
 715             && ctnetlink_dump_mark(skb, ct) < 0)
 716                 goto nla_put_failure;
 717 #endif
 718         rcu_read_unlock();
 719
 720         nlmsg_end(skb, nlh);
 721         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 722                              GFP_ATOMIC);
 723         if (err == -ENOBUFS || err == -EAGAIN)
 724                 return -ENOBUFS;
 725
 726         return 0;
 727
 728 nla_put_failure:
 729         rcu_read_unlock();
 730         nlmsg_cancel(skb, nlh);
 731 nlmsg_failure:
 732         kfree_skb(skb);
 733 errout:
 734         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 735                 return -ENOBUFS;
 736
 737         return 0;
 738 }
 739 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 740
 741 static int ctnetlink_done(struct netlink_callback *cb)
 742 {
 743         if (cb->args[1])
 744                 nf_ct_put((struct nf_conn *)cb->args[1]);
 745         if (cb->data)
 746                 kfree(cb->data);
 747         return 0;
 748 }
 749
 750 struct ctnetlink_dump_filter {
 751         struct {
 752                 u_int32_t val;
 753                 u_int32_t mask;
 754         } mark;
 755 };
 756
 757 static int
 758 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 759 {
 760         struct net *net = sock_net(skb->sk);
 761         struct nf_conn *ct, *last;
 762         struct nf_conntrack_tuple_hash *h;
 763         struct hlist_nulls_node *n;
 764         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 765         u_int8_t l3proto = nfmsg->nfgen_family;
 766         int res;
 767         spinlock_t *lockp;
 768
 769 #ifdef CONFIG_NF_CONNTRACK_MARK
 770         const struct ctnetlink_dump_filter *filter = cb->data;
 771 #endif
 772
 773         last = (struct nf_conn *)cb->args[1];
 774
 775         local_bh_disable();
 776         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 777 restart:
 778                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 779                 spin_lock(lockp);
 780                 if (cb->args[0] >= net->ct.htable_size) {
 781                         spin_unlock(lockp);
 782                         goto out;
 783                 }
 784                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 785                                          hnnode) {
 786                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 787                                 continue;
 788                         ct = nf_ct_tuplehash_to_ctrack(h);
 789                         /* Dump entries of a given L3 protocol number.
 790                          * If it is not specified, ie. l3proto == 0,
 791                          * then dump everything. */
 792                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 793                                 continue;
 794                         if (cb->args[1]) {
 795                                 if (ct != last)
 796                                         continue;
 797                                 cb->args[1] = 0;
 798                         }
 799 #ifdef CONFIG_NF_CONNTRACK_MARK
 800                         if (filter && !((ct->mark & filter->mark.mask) ==
 801                                         filter->mark.val)) {
 802                                 continue;
 803                         }
 804 #endif
 805                         rcu_read_lock();
 806                         res =
 807                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 808                                             cb->nlh->nlmsg_seq,
 809                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 810                                             ct);
 811                         rcu_read_unlock();
 812                         if (res < 0) {
 813                                 nf_conntrack_get(&ct->ct_general);
 814                                 cb->args[1] = (unsigned long)ct;
 815                                 spin_unlock(lockp);
 816                                 goto out;
 817                         }
 818                 }
 819                 spin_unlock(lockp);
 820                 if (cb->args[1]) {
 821                         cb->args[1] = 0;
 822                         goto restart;
 823                 }
 824         }
 825 out:
 826         local_bh_enable();
 827         if (last)
 828                 nf_ct_put(last);
 829
 830         return skb->len;
 831 }
 832
 833 static inline int
 834 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 835 {
 836         struct nlattr *tb[CTA_IP_MAX+1];
 837         struct nf_conntrack_l3proto *l3proto;
 838         int ret = 0;
 839
 840         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 841         if (ret < 0)
 842                 return ret;
 843
 844         rcu_read_lock();
 845         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 846
 847         if (likely(l3proto->nlattr_to_tuple)) {
 848                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 849                                           l3proto->nla_policy);
 850                 if (ret == 0)
 851                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 852         }
 853
 854         rcu_read_unlock();
 855
 856         return ret;
 857 }
 858
 859 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 860         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 861 };
 862
 863 static inline int
 864 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 865                             struct nf_conntrack_tuple *tuple)
 866 {
 867         struct nlattr *tb[CTA_PROTO_MAX+1];
 868         struct nf_conntrack_l4proto *l4proto;
 869         int ret = 0;
 870
 871         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 872         if (ret < 0)
 873                 return ret;
 874
 875         if (!tb[CTA_PROTO_NUM])
 876                 return -EINVAL;
 877         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 878
 879         rcu_read_lock();
 880         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 881
 882         if (likely(l4proto->nlattr_to_tuple)) {
 883                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 884                                           l4proto->nla_policy);
 885                 if (ret == 0)
 886                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 887         }
 888
 889         rcu_read_unlock();
 890
 891         return ret;
 892 }
 893
 894 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 895         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 896         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 897 };
 898
 899 static int
 900 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 901                       struct nf_conntrack_tuple *tuple,
 902                       enum ctattr_type type, u_int8_t l3num)
 903 {
 904         struct nlattr *tb[CTA_TUPLE_MAX+1];
 905         int err;
 906
 907         memset(tuple, 0, sizeof(*tuple));
 908
 909         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 910         if (err < 0)
 911                 return err;
 912
 913         if (!tb[CTA_TUPLE_IP])
 914                 return -EINVAL;
 915
 916         tuple->src.l3num = l3num;
 917
 918         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 919         if (err < 0)
 920                 return err;
 921
 922         if (!tb[CTA_TUPLE_PROTO])
 923                 return -EINVAL;
 924
 925         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 926         if (err < 0)
 927                 return err;
 928
 929         /* orig and expect tuples get DIR_ORIGINAL */
 930         if (type == CTA_TUPLE_REPLY)
 931                 tuple->dst.dir = IP_CT_DIR_REPLY;
 932         else
 933                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 934
 935         return 0;
 936 }
 937
 938 static int
 939 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 940 {
 941         if (attr)
 942 #ifdef CONFIG_NF_CONNTRACK_ZONES
 943                 *zone = ntohs(nla_get_be16(attr));
 944 #else
 945                 return -EOPNOTSUPP;
 946 #endif
 947         else
 948                 *zone = 0;
 949
 950         return 0;
 951 }
 952
 953 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 954         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 955                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 956 };
 957
 958 static inline int
 959 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 960                      struct nlattr **helpinfo)
 961 {
 962         int err;
 963         struct nlattr *tb[CTA_HELP_MAX+1];
 964
 965         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 966         if (err < 0)
 967                 return err;
 968
 969         if (!tb[CTA_HELP_NAME])
 970                 return -EINVAL;
 971
 972         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 973
 974         if (tb[CTA_HELP_INFO])
 975                 *helpinfo = tb[CTA_HELP_INFO];
 976
 977         return 0;
 978 }
 979
 980 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 981         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 982         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 983         [CTA_STATUS]            = { .type = NLA_U32 },
 984         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 985         [CTA_HELP]              = { .type = NLA_NESTED },
 986         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 987         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 988         [CTA_MARK]              = { .type = NLA_U32 },
 989         [CTA_ID]                = { .type = NLA_U32 },
 990         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 991         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 992         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
 993         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
 994         [CTA_ZONE]              = { .type = NLA_U16 },
 995         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 996         [CTA_LABELS]            = { .type = NLA_BINARY,
 997                                     .len = NF_CT_LABELS_MAX_SIZE },
 998         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
 999                                     .len = NF_CT_LABELS_MAX_SIZE },
1000 };
1001
1002 static int
1003 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
1004                         const struct nlmsghdr *nlh,
1005                         const struct nlattr * const cda[])
1006 {
1007         struct net *net = sock_net(ctnl);
1008         struct nf_conntrack_tuple_hash *h;
1009         struct nf_conntrack_tuple tuple;
1010         struct nf_conn *ct;
1011         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1012         u_int8_t u3 = nfmsg->nfgen_family;
1013         u16 zone;
1014         int err;
1015
1016         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1017         if (err < 0)
1018                 return err;
1019
1020         if (cda[CTA_TUPLE_ORIG])
1021                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1022         else if (cda[CTA_TUPLE_REPLY])
1023                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1024         else {
1025                 /* Flush the whole table */
1026                 nf_conntrack_flush_report(net,
1027                                          NETLINK_CB(skb).portid,
1028                                          nlmsg_report(nlh));
1029                 return 0;
1030         }
1031
1032         if (err < 0)
1033                 return err;
1034
1035         h = nf_conntrack_find_get(net, zone, &tuple);
1036         if (!h)
1037                 return -ENOENT;
1038
1039         ct = nf_ct_tuplehash_to_ctrack(h);
1040
1041         if (cda[CTA_ID]) {
1042                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1043                 if (id != (u32)(unsigned long)ct) {
1044                         nf_ct_put(ct);
1045                         return -ENOENT;
1046                 }
1047         }
1048
1049         if (del_timer(&ct->timeout))
1050                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1051
1052         nf_ct_put(ct);
1053
1054         return 0;
1055 }
1056
1057 static int
1058 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1059                         const struct nlmsghdr *nlh,
1060                         const struct nlattr * const cda[])
1061 {
1062         struct net *net = sock_net(ctnl);
1063         struct nf_conntrack_tuple_hash *h;
1064         struct nf_conntrack_tuple tuple;
1065         struct nf_conn *ct;
1066         struct sk_buff *skb2 = NULL;
1067         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1068         u_int8_t u3 = nfmsg->nfgen_family;
1069         u16 zone;
1070         int err;
1071
1072         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1073                 struct netlink_dump_control c = {
1074                         .dump = ctnetlink_dump_table,
1075                         .done = ctnetlink_done,
1076                 };
1077 #ifdef CONFIG_NF_CONNTRACK_MARK
1078                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1079                         struct ctnetlink_dump_filter *filter;
1080
1081                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1082                                          GFP_ATOMIC);
1083                         if (filter == NULL)
1084                                 return -ENOMEM;
1085
1086                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1087                         filter->mark.mask =
1088                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1089                         c.data = filter;
1090                 }
1091 #endif
1092                 return netlink_dump_start(ctnl, skb, nlh, &c);
1093         }
1094
1095         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1096         if (err < 0)
1097                 return err;
1098
1099         if (cda[CTA_TUPLE_ORIG])
1100                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1101         else if (cda[CTA_TUPLE_REPLY])
1102                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1103         else
1104                 return -EINVAL;
1105
1106         if (err < 0)
1107                 return err;
1108
1109         h = nf_conntrack_find_get(net, zone, &tuple);
1110         if (!h)
1111                 return -ENOENT;
1112
1113         ct = nf_ct_tuplehash_to_ctrack(h);
1114
1115         err = -ENOMEM;
1116         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1117         if (skb2 == NULL) {
1118                 nf_ct_put(ct);
1119                 return -ENOMEM;
1120         }
1121
1122         rcu_read_lock();
1123         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1124                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1125         rcu_read_unlock();
1126         nf_ct_put(ct);
1127         if (err <= 0)
1128                 goto free;
1129
1130         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1131         if (err < 0)
1132                 goto out;
1133
1134         return 0;
1135
1136 free:
1137         kfree_skb(skb2);
1138 out:
1139         /* this avoids a loop in nfnetlink. */
1140         return err == -EAGAIN ? -ENOBUFS : err;
1141 }
1142
1143 static int ctnetlink_done_list(struct netlink_callback *cb)
1144 {
1145         if (cb->args[1])
1146                 nf_ct_put((struct nf_conn *)cb->args[1]);
1147         return 0;
1148 }
1149
1150 static int
1151 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1152 {
1153         struct nf_conn *ct, *last = NULL;
1154         struct nf_conntrack_tuple_hash *h;
1155         struct hlist_nulls_node *n;
1156         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1157         u_int8_t l3proto = nfmsg->nfgen_family;
1158         int res;
1159         int cpu;
1160         struct hlist_nulls_head *list;
1161         struct net *net = sock_net(skb->sk);
1162
1163         if (cb->args[2])
1164                 return 0;
1165
1166         if (cb->args[0] == nr_cpu_ids)
1167                 return 0;
1168
1169         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1170                 struct ct_pcpu *pcpu;
1171
1172                 if (!cpu_possible(cpu))
1173                         continue;
1174
1175                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1176                 spin_lock_bh(&pcpu->lock);
1177                 last = (struct nf_conn *)cb->args[1];
1178                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1179 restart:
1180                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1181                         ct = nf_ct_tuplehash_to_ctrack(h);
1182                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1183                                 continue;
1184                         if (cb->args[1]) {
1185                                 if (ct != last)
1186                                         continue;
1187                                 cb->args[1] = 0;
1188                         }
1189                         rcu_read_lock();
1190                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1191                                                   cb->nlh->nlmsg_seq,
1192                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1193                                                   ct);
1194                         rcu_read_unlock();
1195                         if (res < 0) {
1196                                 nf_conntrack_get(&ct->ct_general);
1197                                 cb->args[1] = (unsigned long)ct;
1198                                 spin_unlock_bh(&pcpu->lock);
1199                                 goto out;
1200                         }
1201                 }
1202                 if (cb->args[1]) {
1203                         cb->args[1] = 0;
1204                         goto restart;
1205                 } else
1206                         cb->args[2] = 1;
1207                 spin_unlock_bh(&pcpu->lock);
1208         }
1209 out:
1210         if (last)
1211                 nf_ct_put(last);
1212
1213         return skb->len;
1214 }
1215
1216 static int
1217 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1218 {
1219         return ctnetlink_dump_list(skb, cb, true);
1220 }
1221
1222 static int
1223 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1224                        const struct nlmsghdr *nlh,
1225                        const struct nlattr * const cda[])
1226 {
1227         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1228                 struct netlink_dump_control c = {
1229                         .dump = ctnetlink_dump_dying,
1230                         .done = ctnetlink_done_list,
1231                 };
1232                 return netlink_dump_start(ctnl, skb, nlh, &c);
1233         }
1234
1235         return -EOPNOTSUPP;
1236 }
1237
1238 static int
1239 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1240 {
1241         return ctnetlink_dump_list(skb, cb, false);
1242 }
1243
1244 static int
1245 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1246                              const struct nlmsghdr *nlh,
1247                              const struct nlattr * const cda[])
1248 {
1249         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1250                 struct netlink_dump_control c = {
1251                         .dump = ctnetlink_dump_unconfirmed,
1252                         .done = ctnetlink_done_list,
1253                 };
1254                 return netlink_dump_start(ctnl, skb, nlh, &c);
1255         }
1256
1257         return -EOPNOTSUPP;
1258 }
1259
1260 #ifdef CONFIG_NF_NAT_NEEDED
1261 static int
1262 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1263                           enum nf_nat_manip_type manip,
1264                           const struct nlattr *attr)
1265 {
1266         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1267         int err;
1268
1269         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1270         if (!parse_nat_setup) {
1271 #ifdef CONFIG_MODULES
1272                 rcu_read_unlock();
1273                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1274                 if (request_module("nf-nat") < 0) {
1275                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1276                         rcu_read_lock();
1277                         return -EOPNOTSUPP;
1278                 }
1279                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1280                 rcu_read_lock();
1281                 if (nfnetlink_parse_nat_setup_hook)
1282                         return -EAGAIN;
1283 #endif
1284                 return -EOPNOTSUPP;
1285         }
1286
1287         err = parse_nat_setup(ct, manip, attr);
1288         if (err == -EAGAIN) {
1289 #ifdef CONFIG_MODULES
1290                 rcu_read_unlock();
1291                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1292                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1293                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1294                         rcu_read_lock();
1295                         return -EOPNOTSUPP;
1296                 }
1297                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1298                 rcu_read_lock();
1299 #else
1300                 err = -EOPNOTSUPP;
1301 #endif
1302         }
1303         return err;
1304 }
1305 #endif
1306
1307 static int
1308 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1309 {
1310         unsigned long d;
1311         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1312         d = ct->status ^ status;
1313
1314         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1315                 /* unchangeable */
1316                 return -EBUSY;
1317
1318         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1319                 /* SEEN_REPLY bit can only be set */
1320                 return -EBUSY;
1321
1322         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1323                 /* ASSURED bit can only be set */
1324                 return -EBUSY;
1325
1326         /* Be careful here, modifying NAT bits can screw up things,
1327          * so don't let users modify them directly if they don't pass
1328          * nf_nat_range. */
1329         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1330         return 0;
1331 }
1332
1333 static int
1334 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1335 {
1336 #ifdef CONFIG_NF_NAT_NEEDED
1337         int ret;
1338
1339         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1340                 return 0;
1341
1342         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1343                                         cda[CTA_NAT_DST]);
1344         if (ret < 0)
1345                 return ret;
1346
1347         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1348                                         cda[CTA_NAT_SRC]);
1349         return ret;
1350 #else
1351         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1352                 return 0;
1353         return -EOPNOTSUPP;
1354 #endif
1355 }
1356
1357 static inline int
1358 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1359 {
1360         struct nf_conntrack_helper *helper;
1361         struct nf_conn_help *help = nfct_help(ct);
1362         char *helpname = NULL;
1363         struct nlattr *helpinfo = NULL;
1364         int err;
1365
1366         /* don't change helper of sibling connections */
1367         if (ct->master)
1368                 return -EBUSY;
1369
1370         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1371         if (err < 0)
1372                 return err;
1373
1374         if (!strcmp(helpname, "")) {
1375                 if (help && help->helper) {
1376                         /* we had a helper before ... */
1377                         nf_ct_remove_expectations(ct);
1378                         RCU_INIT_POINTER(help->helper, NULL);
1379                 }
1380
1381                 return 0;
1382         }
1383
1384         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1385                                             nf_ct_protonum(ct));
1386         if (helper == NULL) {
1387 #ifdef CONFIG_MODULES
1388                 spin_unlock_bh(&nf_conntrack_expect_lock);
1389
1390                 if (request_module("nfct-helper-%s", helpname) < 0) {
1391                         spin_lock_bh(&nf_conntrack_expect_lock);
1392                         return -EOPNOTSUPP;
1393                 }
1394
1395                 spin_lock_bh(&nf_conntrack_expect_lock);
1396                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1397                                                     nf_ct_protonum(ct));
1398                 if (helper)
1399                         return -EAGAIN;
1400 #endif
1401                 return -EOPNOTSUPP;
1402         }
1403
1404         if (help) {
1405                 if (help->helper == helper) {
1406                         /* update private helper data if allowed. */
1407                         if (helper->from_nlattr)
1408                                 helper->from_nlattr(helpinfo, ct);
1409                         return 0;
1410                 } else
1411                         return -EBUSY;
1412         }
1413
1414         /* we cannot set a helper for an existing conntrack */
1415         return -EOPNOTSUPP;
1416 }
1417
1418 static inline int
1419 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1420 {
1421         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1422
1423         if (!del_timer(&ct->timeout))
1424                 return -ETIME;
1425
1426         ct->timeout.expires = jiffies + timeout * HZ;
1427         add_timer(&ct->timeout);
1428
1429         return 0;
1430 }
1431
1432 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1433         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1434         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1435         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1436 };
1437
1438 static inline int
1439 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1440 {
1441         const struct nlattr *attr = cda[CTA_PROTOINFO];
1442         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1443         struct nf_conntrack_l4proto *l4proto;
1444         int err = 0;
1445
1446         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1447         if (err < 0)
1448                 return err;
1449
1450         rcu_read_lock();
1451         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1452         if (l4proto->from_nlattr)
1453                 err = l4proto->from_nlattr(tb, ct);
1454         rcu_read_unlock();
1455
1456         return err;
1457 }
1458
1459 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1460         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1461         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1462         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1463 };
1464
1465 static inline int
1466 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1467 {
1468         int err;
1469         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1470
1471         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1472         if (err < 0)
1473                 return err;
1474
1475         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1476                 return -EINVAL;
1477
1478         seq->correction_pos =
1479                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1480
1481         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1482                 return -EINVAL;
1483
1484         seq->offset_before =
1485                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1486
1487         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1488                 return -EINVAL;
1489
1490         seq->offset_after =
1491                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1492
1493         return 0;
1494 }
1495
1496 static int
1497 ctnetlink_change_seq_adj(struct nf_conn *ct,
1498                          const struct nlattr * const cda[])
1499 {
1500         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1501         int ret = 0;
1502
1503         if (!seqadj)
1504                 return 0;
1505
1506         if (cda[CTA_SEQ_ADJ_ORIG]) {
1507                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1508                                      cda[CTA_SEQ_ADJ_ORIG]);
1509                 if (ret < 0)
1510                         return ret;
1511
1512                 ct->status |= IPS_SEQ_ADJUST;
1513         }
1514
1515         if (cda[CTA_SEQ_ADJ_REPLY]) {
1516                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1517                                      cda[CTA_SEQ_ADJ_REPLY]);
1518                 if (ret < 0)
1519                         return ret;
1520
1521                 ct->status |= IPS_SEQ_ADJUST;
1522         }
1523
1524         return 0;
1525 }
1526
1527 static int
1528 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1529 {
1530 #ifdef CONFIG_NF_CONNTRACK_LABELS
1531         size_t len = nla_len(cda[CTA_LABELS]);
1532         const void *mask = cda[CTA_LABELS_MASK];
1533
1534         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1535                 return -EINVAL;
1536
1537         if (mask) {
1538                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1539                     nla_len(cda[CTA_LABELS_MASK]) != len)
1540                         return -EINVAL;
1541                 mask = nla_data(cda[CTA_LABELS_MASK]);
1542         }
1543
1544         len /= sizeof(u32);
1545
1546         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1547 #else
1548         return -EOPNOTSUPP;
1549 #endif
1550 }
1551
1552 static int
1553 ctnetlink_change_conntrack(struct nf_conn *ct,
1554                            const struct nlattr * const cda[])
1555 {
1556         int err;
1557
1558         /* only allow NAT changes and master assignation for new conntracks */
1559         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1560                 return -EOPNOTSUPP;
1561
1562         if (cda[CTA_HELP]) {
1563                 err = ctnetlink_change_helper(ct, cda);
1564                 if (err < 0)
1565                         return err;
1566         }
1567
1568         if (cda[CTA_TIMEOUT]) {
1569                 err = ctnetlink_change_timeout(ct, cda);
1570                 if (err < 0)
1571                         return err;
1572         }
1573
1574         if (cda[CTA_STATUS]) {
1575                 err = ctnetlink_change_status(ct, cda);
1576                 if (err < 0)
1577                         return err;
1578         }
1579
1580         if (cda[CTA_PROTOINFO]) {
1581                 err = ctnetlink_change_protoinfo(ct, cda);
1582                 if (err < 0)
1583                         return err;
1584         }
1585
1586 #if defined(CONFIG_NF_CONNTRACK_MARK)
1587         if (cda[CTA_MARK])
1588                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1589 #endif
1590
1591         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1592                 err = ctnetlink_change_seq_adj(ct, cda);
1593                 if (err < 0)
1594                         return err;
1595         }
1596
1597         if (cda[CTA_LABELS]) {
1598                 err = ctnetlink_attach_labels(ct, cda);
1599                 if (err < 0)
1600                         return err;
1601         }
1602
1603         return 0;
1604 }
1605
1606 static struct nf_conn *
1607 ctnetlink_create_conntrack(struct net *net, u16 zone,
1608                            const struct nlattr * const cda[],
1609                            struct nf_conntrack_tuple *otuple,
1610                            struct nf_conntrack_tuple *rtuple,
1611                            u8 u3)
1612 {
1613         struct nf_conn *ct;
1614         int err = -EINVAL;
1615         struct nf_conntrack_helper *helper;
1616         struct nf_conn_tstamp *tstamp;
1617
1618         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1619         if (IS_ERR(ct))
1620                 return ERR_PTR(-ENOMEM);
1621
1622         if (!cda[CTA_TIMEOUT])
1623                 goto err1;
1624         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1625
1626         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1627
1628         rcu_read_lock();
1629         if (cda[CTA_HELP]) {
1630                 char *helpname = NULL;
1631                 struct nlattr *helpinfo = NULL;
1632
1633                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1634                 if (err < 0)
1635                         goto err2;
1636
1637                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1638                                                     nf_ct_protonum(ct));
1639                 if (helper == NULL) {
1640                         rcu_read_unlock();
1641 #ifdef CONFIG_MODULES
1642                         if (request_module("nfct-helper-%s", helpname) < 0) {
1643                                 err = -EOPNOTSUPP;
1644                                 goto err1;
1645                         }
1646
1647                         rcu_read_lock();
1648                         helper = __nf_conntrack_helper_find(helpname,
1649                                                             nf_ct_l3num(ct),
1650                                                             nf_ct_protonum(ct));
1651                         if (helper) {
1652                                 err = -EAGAIN;
1653                                 goto err2;
1654                         }
1655                         rcu_read_unlock();
1656 #endif
1657                         err = -EOPNOTSUPP;
1658                         goto err1;
1659                 } else {
1660                         struct nf_conn_help *help;
1661
1662                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1663                         if (help == NULL) {
1664                                 err = -ENOMEM;
1665                                 goto err2;
1666                         }
1667                         /* set private helper data if allowed. */
1668                         if (helper->from_nlattr)
1669                                 helper->from_nlattr(helpinfo, ct);
1670
1671                         /* not in hash table yet so not strictly necessary */
1672                         RCU_INIT_POINTER(help->helper, helper);
1673                 }
1674         } else {
1675                 /* try an implicit helper assignation */
1676                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1677                 if (err < 0)
1678                         goto err2;
1679         }
1680
1681         err = ctnetlink_setup_nat(ct, cda);
1682         if (err < 0)
1683                 goto err2;
1684
1685         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1686         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1687         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1688         nf_ct_labels_ext_add(ct);
1689
1690         /* we must add conntrack extensions before confirmation. */
1691         ct->status |= IPS_CONFIRMED;
1692
1693         if (cda[CTA_STATUS]) {
1694                 err = ctnetlink_change_status(ct, cda);
1695                 if (err < 0)
1696                         goto err2;
1697         }
1698
1699         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1700                 err = ctnetlink_change_seq_adj(ct, cda);
1701                 if (err < 0)
1702                         goto err2;
1703         }
1704
1705         memset(&ct->proto, 0, sizeof(ct->proto));
1706         if (cda[CTA_PROTOINFO]) {
1707                 err = ctnetlink_change_protoinfo(ct, cda);
1708                 if (err < 0)
1709                         goto err2;
1710         }
1711
1712 #if defined(CONFIG_NF_CONNTRACK_MARK)
1713         if (cda[CTA_MARK])
1714                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1715 #endif
1716
1717         /* setup master conntrack: this is a confirmed expectation */
1718         if (cda[CTA_TUPLE_MASTER]) {
1719                 struct nf_conntrack_tuple master;
1720                 struct nf_conntrack_tuple_hash *master_h;
1721                 struct nf_conn *master_ct;
1722
1723                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1724                 if (err < 0)
1725                         goto err2;
1726
1727                 master_h = nf_conntrack_find_get(net, zone, &master);
1728                 if (master_h == NULL) {
1729                         err = -ENOENT;
1730                         goto err2;
1731                 }
1732                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1733                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1734                 ct->master = master_ct;
1735         }
1736         tstamp = nf_conn_tstamp_find(ct);
1737         if (tstamp)
1738                 tstamp->start = ktime_to_ns(ktime_get_real());
1739
1740         err = nf_conntrack_hash_check_insert(ct);
1741         if (err < 0)
1742                 goto err2;
1743
1744         rcu_read_unlock();
1745
1746         return ct;
1747
1748 err2:
1749         rcu_read_unlock();
1750 err1:
1751         nf_conntrack_free(ct);
1752         return ERR_PTR(err);
1753 }
1754
1755 static int
1756 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1757                         const struct nlmsghdr *nlh,
1758                         const struct nlattr * const cda[])
1759 {
1760         struct net *net = sock_net(ctnl);
1761         struct nf_conntrack_tuple otuple, rtuple;
1762         struct nf_conntrack_tuple_hash *h = NULL;
1763         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1764         struct nf_conn *ct;
1765         u_int8_t u3 = nfmsg->nfgen_family;
1766         u16 zone;
1767         int err;
1768
1769         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1770         if (err < 0)
1771                 return err;
1772
1773         if (cda[CTA_TUPLE_ORIG]) {
1774                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1775                 if (err < 0)
1776                         return err;
1777         }
1778
1779         if (cda[CTA_TUPLE_REPLY]) {
1780                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1781                 if (err < 0)
1782                         return err;
1783         }
1784
1785         if (cda[CTA_TUPLE_ORIG])
1786                 h = nf_conntrack_find_get(net, zone, &otuple);
1787         else if (cda[CTA_TUPLE_REPLY])
1788                 h = nf_conntrack_find_get(net, zone, &rtuple);
1789
1790         if (h == NULL) {
1791                 err = -ENOENT;
1792                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1793                         enum ip_conntrack_events events;
1794
1795                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1796                                 return -EINVAL;
1797
1798                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1799                                                         &rtuple, u3);
1800                         if (IS_ERR(ct))
1801                                 return PTR_ERR(ct);
1802
1803                         err = 0;
1804                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1805                                 events = IPCT_RELATED;
1806                         else
1807                                 events = IPCT_NEW;
1808
1809                         if (cda[CTA_LABELS] &&
1810                             ctnetlink_attach_labels(ct, cda) == 0)
1811                                 events |= (1 << IPCT_LABEL);
1812
1813                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1814                                                       (1 << IPCT_ASSURED) |
1815                                                       (1 << IPCT_HELPER) |
1816                                                       (1 << IPCT_PROTOINFO) |
1817                                                       (1 << IPCT_SEQADJ) |
1818                                                       (1 << IPCT_MARK) | events,
1819                                                       ct, NETLINK_CB(skb).portid,
1820                                                       nlmsg_report(nlh));
1821                         nf_ct_put(ct);
1822                 }
1823
1824                 return err;
1825         }
1826         /* implicit 'else' */
1827
1828         err = -EEXIST;
1829         ct = nf_ct_tuplehash_to_ctrack(h);
1830         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1831                 spin_lock_bh(&nf_conntrack_expect_lock);
1832                 err = ctnetlink_change_conntrack(ct, cda);
1833                 spin_unlock_bh(&nf_conntrack_expect_lock);
1834                 if (err == 0) {
1835                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1836                                                       (1 << IPCT_ASSURED) |
1837                                                       (1 << IPCT_HELPER) |
1838                                                       (1 << IPCT_LABEL) |
1839                                                       (1 << IPCT_PROTOINFO) |
1840                                                       (1 << IPCT_SEQADJ) |
1841                                                       (1 << IPCT_MARK),
1842                                                       ct, NETLINK_CB(skb).portid,
1843                                                       nlmsg_report(nlh));
1844                 }
1845         }
1846
1847         nf_ct_put(ct);
1848         return err;
1849 }
1850
1851 static int
1852 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1853                                 __u16 cpu, const struct ip_conntrack_stat *st)
1854 {
1855         struct nlmsghdr *nlh;
1856         struct nfgenmsg *nfmsg;
1857         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1858
1859         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1860         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1861         if (nlh == NULL)
1862                 goto nlmsg_failure;
1863
1864         nfmsg = nlmsg_data(nlh);
1865         nfmsg->nfgen_family = AF_UNSPEC;
1866         nfmsg->version      = NFNETLINK_V0;
1867         nfmsg->res_id       = htons(cpu);
1868
1869         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1870             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1871             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1872             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1873             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1874             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1875             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1876             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1877             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1878                                 htonl(st->insert_failed)) ||
1879             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1880             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1881             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1882             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1883                                 htonl(st->search_restart)))
1884                 goto nla_put_failure;
1885
1886         nlmsg_end(skb, nlh);
1887         return skb->len;
1888
1889 nla_put_failure:
1890 nlmsg_failure:
1891         nlmsg_cancel(skb, nlh);
1892         return -1;
1893 }
1894
1895 static int
1896 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1897 {
1898         int cpu;
1899         struct net *net = sock_net(skb->sk);
1900
1901         if (cb->args[0] == nr_cpu_ids)
1902                 return 0;
1903
1904         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1905                 const struct ip_conntrack_stat *st;
1906
1907                 if (!cpu_possible(cpu))
1908                         continue;
1909
1910                 st = per_cpu_ptr(net->ct.stat, cpu);
1911                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1912                                                     NETLINK_CB(cb->skb).portid,
1913                                                     cb->nlh->nlmsg_seq,
1914                                                     cpu, st) < 0)
1915                                 break;
1916         }
1917         cb->args[0] = cpu;
1918
1919         return skb->len;
1920 }
1921
1922 static int
1923 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1924                       const struct nlmsghdr *nlh,
1925                       const struct nlattr * const cda[])
1926 {
1927         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1928                 struct netlink_dump_control c = {
1929                         .dump = ctnetlink_ct_stat_cpu_dump,
1930                 };
1931                 return netlink_dump_start(ctnl, skb, nlh, &c);
1932         }
1933
1934         return 0;
1935 }
1936
1937 static int
1938 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1939                             struct net *net)
1940 {
1941         struct nlmsghdr *nlh;
1942         struct nfgenmsg *nfmsg;
1943         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1944         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1945
1946         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1947         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1948         if (nlh == NULL)
1949                 goto nlmsg_failure;
1950
1951         nfmsg = nlmsg_data(nlh);
1952         nfmsg->nfgen_family = AF_UNSPEC;
1953         nfmsg->version      = NFNETLINK_V0;
1954         nfmsg->res_id       = 0;
1955
1956         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1957                 goto nla_put_failure;
1958
1959         nlmsg_end(skb, nlh);
1960         return skb->len;
1961
1962 nla_put_failure:
1963 nlmsg_failure:
1964         nlmsg_cancel(skb, nlh);
1965         return -1;
1966 }
1967
1968 static int
1969 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1970                   const struct nlmsghdr *nlh,
1971                   const struct nlattr * const cda[])
1972 {
1973         struct sk_buff *skb2;
1974         int err;
1975
1976         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1977         if (skb2 == NULL)
1978                 return -ENOMEM;
1979
1980         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1981                                           nlh->nlmsg_seq,
1982                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1983                                           sock_net(skb->sk));
1984         if (err <= 0)
1985                 goto free;
1986
1987         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1988         if (err < 0)
1989                 goto out;
1990
1991         return 0;
1992
1993 free:
1994         kfree_skb(skb2);
1995 out:
1996         /* this avoids a loop in nfnetlink. */
1997         return err == -EAGAIN ? -ENOBUFS : err;
1998 }
1999
2000 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2001         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2002         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2003         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2004         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2005         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2006         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2007                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2008         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2009         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2010         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2011         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2012         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2013 };
2014
2015 static struct nf_conntrack_expect *
2016 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2017                        struct nf_conntrack_helper *helper,
2018                        struct nf_conntrack_tuple *tuple,
2019                        struct nf_conntrack_tuple *mask);
2020
2021 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2022 static size_t
2023 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2024 {
2025         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2026                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2027                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2028                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2029                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2030                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2031                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2032                + nla_total_size(0) /* CTA_PROTOINFO */
2033                + nla_total_size(0) /* CTA_HELP */
2034                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2035                + ctnetlink_secctx_size(ct)
2036 #ifdef CONFIG_NF_NAT_NEEDED
2037                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2038                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2039 #endif
2040 #ifdef CONFIG_NF_CONNTRACK_MARK
2041                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2042 #endif
2043                + ctnetlink_proto_size(ct)
2044                ;
2045 }
2046
2047 static int
2048 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2049 {
2050         struct nlattr *nest_parms;
2051
2052         rcu_read_lock();
2053         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2054         if (!nest_parms)
2055                 goto nla_put_failure;
2056         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2057                 goto nla_put_failure;
2058         nla_nest_end(skb, nest_parms);
2059
2060         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2061         if (!nest_parms)
2062                 goto nla_put_failure;
2063         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2064                 goto nla_put_failure;
2065         nla_nest_end(skb, nest_parms);
2066
2067         if (nf_ct_zone(ct)) {
2068                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2069                         goto nla_put_failure;
2070         }
2071
2072         if (ctnetlink_dump_id(skb, ct) < 0)
2073                 goto nla_put_failure;
2074
2075         if (ctnetlink_dump_status(skb, ct) < 0)
2076                 goto nla_put_failure;
2077
2078         if (ctnetlink_dump_timeout(skb, ct) < 0)
2079                 goto nla_put_failure;
2080
2081         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2082                 goto nla_put_failure;
2083
2084         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2085                 goto nla_put_failure;
2086
2087 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2088         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2089                 goto nla_put_failure;
2090 #endif
2091         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2092                 goto nla_put_failure;
2093
2094         if ((ct->status & IPS_SEQ_ADJUST) &&
2095             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2096                 goto nla_put_failure;
2097
2098 #ifdef CONFIG_NF_CONNTRACK_MARK
2099         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2100                 goto nla_put_failure;
2101 #endif
2102         if (ctnetlink_dump_labels(skb, ct) < 0)
2103                 goto nla_put_failure;
2104         rcu_read_unlock();
2105         return 0;
2106
2107 nla_put_failure:
2108         rcu_read_unlock();
2109         return -ENOSPC;
2110 }
2111
2112 static int
2113 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2114 {
2115         int err;
2116
2117         if (cda[CTA_TIMEOUT]) {
2118                 err = ctnetlink_change_timeout(ct, cda);
2119                 if (err < 0)
2120                         return err;
2121         }
2122         if (cda[CTA_STATUS]) {
2123                 err = ctnetlink_change_status(ct, cda);
2124                 if (err < 0)
2125                         return err;
2126         }
2127         if (cda[CTA_HELP]) {
2128                 err = ctnetlink_change_helper(ct, cda);
2129                 if (err < 0)
2130                         return err;
2131         }
2132         if (cda[CTA_LABELS]) {
2133                 err = ctnetlink_attach_labels(ct, cda);
2134                 if (err < 0)
2135                         return err;
2136         }
2137 #if defined(CONFIG_NF_CONNTRACK_MARK)
2138         if (cda[CTA_MARK]) {
2139                 u32 mask = 0, mark, newmark;
2140                 if (cda[CTA_MARK_MASK])
2141                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2142
2143                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2144                 newmark = (ct->mark & mask) ^ mark;
2145                 if (newmark != ct->mark)
2146                         ct->mark = newmark;
2147         }
2148 #endif
2149         return 0;
2150 }
2151
2152 static int
2153 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2154 {
2155         struct nlattr *cda[CTA_MAX+1];
2156         int ret;
2157
2158         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2159         if (ret < 0)
2160                 return ret;
2161
2162         spin_lock_bh(&nf_conntrack_expect_lock);
2163         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2164         spin_unlock_bh(&nf_conntrack_expect_lock);
2165
2166         return ret;
2167 }
2168
2169 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2170                                        const struct nf_conn *ct,
2171                                        struct nf_conntrack_tuple *tuple,
2172                                        struct nf_conntrack_tuple *mask)
2173 {
2174         int err;
2175
2176         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2177                                     nf_ct_l3num(ct));
2178         if (err < 0)
2179                 return err;
2180
2181         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2182                                      nf_ct_l3num(ct));
2183 }
2184
2185 static int
2186 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2187                                 u32 portid, u32 report)
2188 {
2189         struct nlattr *cda[CTA_EXPECT_MAX+1];
2190         struct nf_conntrack_tuple tuple, mask;
2191         struct nf_conntrack_helper *helper = NULL;
2192         struct nf_conntrack_expect *exp;
2193         int err;
2194
2195         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2196         if (err < 0)
2197                 return err;
2198
2199         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2200                                           ct, &tuple, &mask);
2201         if (err < 0)
2202                 return err;
2203
2204         if (cda[CTA_EXPECT_HELP_NAME]) {
2205                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2206
2207                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2208                                                     nf_ct_protonum(ct));
2209                 if (helper == NULL)
2210                         return -EOPNOTSUPP;
2211         }
2212
2213         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2214                                      helper, &tuple, &mask);
2215         if (IS_ERR(exp))
2216                 return PTR_ERR(exp);
2217
2218         err = nf_ct_expect_related_report(exp, portid, report);
2219         if (err < 0) {
2220                 nf_ct_expect_put(exp);
2221                 return err;
2222         }
2223
2224         return 0;
2225 }
2226
2227 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2228         .build_size     = ctnetlink_nfqueue_build_size,
2229         .build          = ctnetlink_nfqueue_build,
2230         .parse          = ctnetlink_nfqueue_parse,
2231         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2232         .seq_adjust     = nf_ct_tcp_seqadj_set,
2233 };
2234 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2235
2236 /***********************************************************************
2237  * EXPECT
2238  ***********************************************************************/
2239
2240 static inline int
2241 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2242                          const struct nf_conntrack_tuple *tuple,
2243                          enum ctattr_expect type)
2244 {
2245         struct nlattr *nest_parms;
2246
2247         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2248         if (!nest_parms)
2249                 goto nla_put_failure;
2250         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2251                 goto nla_put_failure;
2252         nla_nest_end(skb, nest_parms);
2253
2254         return 0;
2255
2256 nla_put_failure:
2257         return -1;
2258 }
2259
2260 static inline int
2261 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2262                         const struct nf_conntrack_tuple *tuple,
2263                         const struct nf_conntrack_tuple_mask *mask)
2264 {
2265         int ret;
2266         struct nf_conntrack_l3proto *l3proto;
2267         struct nf_conntrack_l4proto *l4proto;
2268         struct nf_conntrack_tuple m;
2269         struct nlattr *nest_parms;
2270
2271         memset(&m, 0xFF, sizeof(m));
2272         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2273         m.src.u.all = mask->src.u.all;
2274         m.dst.protonum = tuple->dst.protonum;
2275
2276         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2277         if (!nest_parms)
2278                 goto nla_put_failure;
2279
2280         rcu_read_lock();
2281         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2282         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2283         if (ret >= 0) {
2284                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2285                                                tuple->dst.protonum);
2286         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2287         }
2288         rcu_read_unlock();
2289
2290         if (unlikely(ret < 0))
2291                 goto nla_put_failure;
2292
2293         nla_nest_end(skb, nest_parms);
2294
2295         return 0;
2296
2297 nla_put_failure:
2298         return -1;
2299 }
2300
2301 static const union nf_inet_addr any_addr;
2302
2303 static int
2304 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2305                           const struct nf_conntrack_expect *exp)
2306 {
2307         struct nf_conn *master = exp->master;
2308         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2309         struct nf_conn_help *help;
2310 #ifdef CONFIG_NF_NAT_NEEDED
2311         struct nlattr *nest_parms;
2312         struct nf_conntrack_tuple nat_tuple = {};
2313 #endif
2314         struct nf_ct_helper_expectfn *expfn;
2315
2316         if (timeout < 0)
2317                 timeout = 0;
2318
2319         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2320                 goto nla_put_failure;
2321         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2322                 goto nla_put_failure;
2323         if (ctnetlink_exp_dump_tuple(skb,
2324                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2325                                  CTA_EXPECT_MASTER) < 0)
2326                 goto nla_put_failure;
2327
2328 #ifdef CONFIG_NF_NAT_NEEDED
2329         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2330             exp->saved_proto.all) {
2331                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2332                 if (!nest_parms)
2333                         goto nla_put_failure;
2334
2335                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2336                         goto nla_put_failure;
2337
2338                 nat_tuple.src.l3num = nf_ct_l3num(master);
2339                 nat_tuple.src.u3 = exp->saved_addr;
2340                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2341                 nat_tuple.src.u = exp->saved_proto;
2342
2343                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2344                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2345                         goto nla_put_failure;
2346                 nla_nest_end(skb, nest_parms);
2347         }
2348 #endif
2349         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2350             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2351             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2352             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2353                 goto nla_put_failure;
2354         help = nfct_help(master);
2355         if (help) {
2356                 struct nf_conntrack_helper *helper;
2357
2358                 helper = rcu_dereference(help->helper);
2359                 if (helper &&
2360                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2361                         goto nla_put_failure;
2362         }
2363         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2364         if (expfn != NULL &&
2365             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2366                 goto nla_put_failure;
2367
2368         return 0;
2369
2370 nla_put_failure:
2371         return -1;
2372 }
2373
2374 static int
2375 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2376                         int event, const struct nf_conntrack_expect *exp)
2377 {
2378         struct nlmsghdr *nlh;
2379         struct nfgenmsg *nfmsg;
2380         unsigned int flags = portid ? NLM_F_MULTI : 0;
2381
2382         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2383         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2384         if (nlh == NULL)
2385                 goto nlmsg_failure;
2386
2387         nfmsg = nlmsg_data(nlh);
2388         nfmsg->nfgen_family = exp->tuple.src.l3num;
2389         nfmsg->version      = NFNETLINK_V0;
2390         nfmsg->res_id       = 0;
2391
2392         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2393                 goto nla_put_failure;
2394
2395         nlmsg_end(skb, nlh);
2396         return skb->len;
2397
2398 nlmsg_failure:
2399 nla_put_failure:
2400         nlmsg_cancel(skb, nlh);
2401         return -1;
2402 }
2403
2404 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2405 static int
2406 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2407 {
2408         struct nf_conntrack_expect *exp = item->exp;
2409         struct net *net = nf_ct_exp_net(exp);
2410         struct nlmsghdr *nlh;
2411         struct nfgenmsg *nfmsg;
2412         struct sk_buff *skb;
2413         unsigned int type, group;
2414         int flags = 0;
2415
2416         if (events & (1 << IPEXP_DESTROY)) {
2417                 type = IPCTNL_MSG_EXP_DELETE;
2418                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2419         } else if (events & (1 << IPEXP_NEW)) {
2420                 type = IPCTNL_MSG_EXP_NEW;
2421                 flags = NLM_F_CREATE|NLM_F_EXCL;
2422                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2423         } else
2424                 return 0;
2425
2426         if (!item->report && !nfnetlink_has_listeners(net, group))
2427                 return 0;
2428
2429         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2430         if (skb == NULL)
2431                 goto errout;
2432
2433         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2434         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2435         if (nlh == NULL)
2436                 goto nlmsg_failure;
2437
2438         nfmsg = nlmsg_data(nlh);
2439         nfmsg->nfgen_family = exp->tuple.src.l3num;
2440         nfmsg->version      = NFNETLINK_V0;
2441         nfmsg->res_id       = 0;
2442
2443         rcu_read_lock();
2444         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2445                 goto nla_put_failure;
2446         rcu_read_unlock();
2447
2448         nlmsg_end(skb, nlh);
2449         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2450         return 0;
2451
2452 nla_put_failure:
2453         rcu_read_unlock();
2454         nlmsg_cancel(skb, nlh);
2455 nlmsg_failure:
2456         kfree_skb(skb);
2457 errout:
2458         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2459         return 0;
2460 }
2461 #endif
2462 static int ctnetlink_exp_done(struct netlink_callback *cb)
2463 {
2464         if (cb->args[1])
2465                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2466         return 0;
2467 }
2468
2469 static int
2470 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2471 {
2472         struct net *net = sock_net(skb->sk);
2473         struct nf_conntrack_expect *exp, *last;
2474         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2475         u_int8_t l3proto = nfmsg->nfgen_family;
2476
2477         rcu_read_lock();
2478         last = (struct nf_conntrack_expect *)cb->args[1];
2479         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2480 restart:
2481                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2482                                      hnode) {
2483                         if (l3proto && exp->tuple.src.l3num != l3proto)
2484                                 continue;
2485                         if (cb->args[1]) {
2486                                 if (exp != last)
2487                                         continue;
2488                                 cb->args[1] = 0;
2489                         }
2490                         if (ctnetlink_exp_fill_info(skb,
2491                                                     NETLINK_CB(cb->skb).portid,
2492                                                     cb->nlh->nlmsg_seq,
2493                                                     IPCTNL_MSG_EXP_NEW,
2494                                                     exp) < 0) {
2495                                 if (!atomic_inc_not_zero(&exp->use))
2496                                         continue;
2497                                 cb->args[1] = (unsigned long)exp;
2498                                 goto out;
2499                         }
2500                 }
2501                 if (cb->args[1]) {
2502                         cb->args[1] = 0;
2503                         goto restart;
2504                 }
2505         }
2506 out:
2507         rcu_read_unlock();
2508         if (last)
2509                 nf_ct_expect_put(last);
2510
2511         return skb->len;
2512 }
2513
2514 static int
2515 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2516 {
2517         struct nf_conntrack_expect *exp, *last;
2518         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2519         struct nf_conn *ct = cb->data;
2520         struct nf_conn_help *help = nfct_help(ct);
2521         u_int8_t l3proto = nfmsg->nfgen_family;
2522
2523         if (cb->args[0])
2524                 return 0;
2525
2526         rcu_read_lock();
2527         last = (struct nf_conntrack_expect *)cb->args[1];
2528 restart:
2529         hlist_for_each_entry(exp, &help->expectations, lnode) {
2530                 if (l3proto && exp->tuple.src.l3num != l3proto)
2531                         continue;
2532                 if (cb->args[1]) {
2533                         if (exp != last)
2534                                 continue;
2535                         cb->args[1] = 0;
2536                 }
2537                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2538                                             cb->nlh->nlmsg_seq,
2539                                             IPCTNL_MSG_EXP_NEW,
2540                                             exp) < 0) {
2541                         if (!atomic_inc_not_zero(&exp->use))
2542                                 continue;
2543                         cb->args[1] = (unsigned long)exp;
2544                         goto out;
2545                 }
2546         }
2547         if (cb->args[1]) {
2548                 cb->args[1] = 0;
2549                 goto restart;
2550         }
2551         cb->args[0] = 1;
2552 out:
2553         rcu_read_unlock();
2554         if (last)
2555                 nf_ct_expect_put(last);
2556
2557         return skb->len;
2558 }
2559
2560 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2561                                  const struct nlmsghdr *nlh,
2562                                  const struct nlattr * const cda[])
2563 {
2564         int err;
2565         struct net *net = sock_net(ctnl);
2566         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2567         u_int8_t u3 = nfmsg->nfgen_family;
2568         struct nf_conntrack_tuple tuple;
2569         struct nf_conntrack_tuple_hash *h;
2570         struct nf_conn *ct;
2571         u16 zone = 0;
2572         struct netlink_dump_control c = {
2573                 .dump = ctnetlink_exp_ct_dump_table,
2574                 .done = ctnetlink_exp_done,
2575         };
2576
2577         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2578         if (err < 0)
2579                 return err;
2580
2581         if (cda[CTA_EXPECT_ZONE]) {
2582                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2583                 if (err < 0)
2584                         return err;
2585         }
2586
2587         h = nf_conntrack_find_get(net, zone, &tuple);
2588         if (!h)
2589                 return -ENOENT;
2590
2591         ct = nf_ct_tuplehash_to_ctrack(h);
2592         c.data = ct;
2593
2594         err = netlink_dump_start(ctnl, skb, nlh, &c);
2595         nf_ct_put(ct);
2596
2597         return err;
2598 }
2599
2600 static int
2601 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2602                      const struct nlmsghdr *nlh,
2603                      const struct nlattr * const cda[])
2604 {
2605         struct net *net = sock_net(ctnl);
2606         struct nf_conntrack_tuple tuple;
2607         struct nf_conntrack_expect *exp;
2608         struct sk_buff *skb2;
2609         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2610         u_int8_t u3 = nfmsg->nfgen_family;
2611         u16 zone;
2612         int err;
2613
2614         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2615                 if (cda[CTA_EXPECT_MASTER])
2616                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2617                 else {
2618                         struct netlink_dump_control c = {
2619                                 .dump = ctnetlink_exp_dump_table,
2620                                 .done = ctnetlink_exp_done,
2621                         };
2622                         return netlink_dump_start(ctnl, skb, nlh, &c);
2623                 }
2624         }
2625
2626         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2627         if (err < 0)
2628                 return err;
2629
2630         if (cda[CTA_EXPECT_TUPLE])
2631                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2632         else if (cda[CTA_EXPECT_MASTER])
2633                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2634         else
2635                 return -EINVAL;
2636
2637         if (err < 0)
2638                 return err;
2639
2640         exp = nf_ct_expect_find_get(net, zone, &tuple);
2641         if (!exp)
2642                 return -ENOENT;
2643
2644         if (cda[CTA_EXPECT_ID]) {
2645                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2646                 if (ntohl(id) != (u32)(unsigned long)exp) {
2647                         nf_ct_expect_put(exp);
2648                         return -ENOENT;
2649                 }
2650         }
2651
2652         err = -ENOMEM;
2653         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2654         if (skb2 == NULL) {
2655                 nf_ct_expect_put(exp);
2656                 goto out;
2657         }
2658
2659         rcu_read_lock();
2660         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2661                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2662         rcu_read_unlock();
2663         nf_ct_expect_put(exp);
2664         if (err <= 0)
2665                 goto free;
2666
2667         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2668         if (err < 0)
2669                 goto out;
2670
2671         return 0;
2672
2673 free:
2674         kfree_skb(skb2);
2675 out:
2676         /* this avoids a loop in nfnetlink. */
2677         return err == -EAGAIN ? -ENOBUFS : err;
2678 }
2679
2680 static int
2681 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2682                      const struct nlmsghdr *nlh,
2683                      const struct nlattr * const cda[])
2684 {
2685         struct net *net = sock_net(ctnl);
2686         struct nf_conntrack_expect *exp;
2687         struct nf_conntrack_tuple tuple;
2688         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2689         struct hlist_node *next;
2690         u_int8_t u3 = nfmsg->nfgen_family;
2691         unsigned int i;
2692         u16 zone;
2693         int err;
2694
2695         if (cda[CTA_EXPECT_TUPLE]) {
2696                 /* delete a single expect by tuple */
2697                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2698                 if (err < 0)
2699                         return err;
2700
2701                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2702                 if (err < 0)
2703                         return err;
2704
2705                 /* bump usage count to 2 */
2706                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2707                 if (!exp)
2708                         return -ENOENT;
2709
2710                 if (cda[CTA_EXPECT_ID]) {
2711                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2712                         if (ntohl(id) != (u32)(unsigned long)exp) {
2713                                 nf_ct_expect_put(exp);
2714                                 return -ENOENT;
2715                         }
2716                 }
2717
2718                 /* after list removal, usage count == 1 */
2719                 spin_lock_bh(&nf_conntrack_expect_lock);
2720                 if (del_timer(&exp->timeout)) {
2721                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2722                                                    nlmsg_report(nlh));
2723                         nf_ct_expect_put(exp);
2724                 }
2725                 spin_unlock_bh(&nf_conntrack_expect_lock);
2726                 /* have to put what we 'get' above.
2727                  * after this line usage count == 0 */
2728                 nf_ct_expect_put(exp);
2729         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2730                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2731                 struct nf_conn_help *m_help;
2732
2733                 /* delete all expectations for this helper */
2734                 spin_lock_bh(&nf_conntrack_expect_lock);
2735                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2736                         hlist_for_each_entry_safe(exp, next,
2737                                                   &net->ct.expect_hash[i],
2738                                                   hnode) {
2739                                 m_help = nfct_help(exp->master);
2740                                 if (!strcmp(m_help->helper->name, name) &&
2741                                     del_timer(&exp->timeout)) {
2742                                         nf_ct_unlink_expect_report(exp,
2743                                                         NETLINK_CB(skb).portid,
2744                                                         nlmsg_report(nlh));
2745                                         nf_ct_expect_put(exp);
2746                                 }
2747                         }
2748                 }
2749                 spin_unlock_bh(&nf_conntrack_expect_lock);
2750         } else {
2751                 /* This basically means we have to flush everything*/
2752                 spin_lock_bh(&nf_conntrack_expect_lock);
2753                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2754                         hlist_for_each_entry_safe(exp, next,
2755                                                   &net->ct.expect_hash[i],
2756                                                   hnode) {
2757                                 if (del_timer(&exp->timeout)) {
2758                                         nf_ct_unlink_expect_report(exp,
2759                                                         NETLINK_CB(skb).portid,
2760                                                         nlmsg_report(nlh));
2761                                         nf_ct_expect_put(exp);
2762                                 }
2763                         }
2764                 }
2765                 spin_unlock_bh(&nf_conntrack_expect_lock);
2766         }
2767
2768         return 0;
2769 }
2770 static int
2771 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2772                         const struct nlattr * const cda[])
2773 {
2774         if (cda[CTA_EXPECT_TIMEOUT]) {
2775                 if (!del_timer(&x->timeout))
2776                         return -ETIME;
2777
2778                 x->timeout.expires = jiffies +
2779                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2780                 add_timer(&x->timeout);
2781         }
2782         return 0;
2783 }
2784
2785 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2786         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2787         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2788 };
2789
2790 static int
2791 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2792                            struct nf_conntrack_expect *exp,
2793                            u_int8_t u3)
2794 {
2795 #ifdef CONFIG_NF_NAT_NEEDED
2796         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2797         struct nf_conntrack_tuple nat_tuple = {};
2798         int err;
2799
2800         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2801         if (err < 0)
2802                 return err;
2803
2804         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2805                 return -EINVAL;
2806
2807         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2808                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2809         if (err < 0)
2810                 return err;
2811
2812         exp->saved_addr = nat_tuple.src.u3;
2813         exp->saved_proto = nat_tuple.src.u;
2814         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2815
2816         return 0;
2817 #else
2818         return -EOPNOTSUPP;
2819 #endif
2820 }
2821
2822 static struct nf_conntrack_expect *
2823 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2824                        struct nf_conntrack_helper *helper,
2825                        struct nf_conntrack_tuple *tuple,
2826                        struct nf_conntrack_tuple *mask)
2827 {
2828         u_int32_t class = 0;
2829         struct nf_conntrack_expect *exp;
2830         struct nf_conn_help *help;
2831         int err;
2832
2833         if (cda[CTA_EXPECT_CLASS] && helper) {
2834                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2835                 if (class > helper->expect_class_max)
2836                         return ERR_PTR(-EINVAL);
2837         }
2838         exp = nf_ct_expect_alloc(ct);
2839         if (!exp)
2840                 return ERR_PTR(-ENOMEM);
2841
2842         help = nfct_help(ct);
2843         if (!help) {
2844                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2845                         err = -EINVAL;
2846                         goto err_out;
2847                 }
2848                 exp->timeout.expires =
2849                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2850
2851                 exp->flags = NF_CT_EXPECT_USERSPACE;
2852                 if (cda[CTA_EXPECT_FLAGS]) {
2853                         exp->flags |=
2854                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2855                 }
2856         } else {
2857                 if (cda[CTA_EXPECT_FLAGS]) {
2858                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2859                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2860                 } else
2861                         exp->flags = 0;
2862         }
2863         if (cda[CTA_EXPECT_FN]) {
2864                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2865                 struct nf_ct_helper_expectfn *expfn;
2866
2867                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2868                 if (expfn == NULL) {
2869                         err = -EINVAL;
2870                         goto err_out;
2871                 }
2872                 exp->expectfn = expfn->expectfn;
2873         } else
2874                 exp->expectfn = NULL;
2875
2876         exp->class = class;
2877         exp->master = ct;
2878         exp->helper = helper;
2879         exp->tuple = *tuple;
2880         exp->mask.src.u3 = mask->src.u3;
2881         exp->mask.src.u.all = mask->src.u.all;
2882
2883         if (cda[CTA_EXPECT_NAT]) {
2884                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2885                                                  exp, nf_ct_l3num(ct));
2886                 if (err < 0)
2887                         goto err_out;
2888         }
2889         return exp;
2890 err_out:
2891         nf_ct_expect_put(exp);
2892         return ERR_PTR(err);
2893 }
2894
2895 static int
2896 ctnetlink_create_expect(struct net *net, u16 zone,
2897                         const struct nlattr * const cda[],
2898                         u_int8_t u3, u32 portid, int report)
2899 {
2900         struct nf_conntrack_tuple tuple, mask, master_tuple;
2901         struct nf_conntrack_tuple_hash *h = NULL;
2902         struct nf_conntrack_helper *helper = NULL;
2903         struct nf_conntrack_expect *exp;
2904         struct nf_conn *ct;
2905         int err;
2906
2907         /* caller guarantees that those three CTA_EXPECT_* exist */
2908         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2909         if (err < 0)
2910                 return err;
2911         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2912         if (err < 0)
2913                 return err;
2914         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2915         if (err < 0)
2916                 return err;
2917
2918         /* Look for master conntrack of this expectation */
2919         h = nf_conntrack_find_get(net, zone, &master_tuple);
2920         if (!h)
2921                 return -ENOENT;
2922         ct = nf_ct_tuplehash_to_ctrack(h);
2923
2924         if (cda[CTA_EXPECT_HELP_NAME]) {
2925                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2926
2927                 helper = __nf_conntrack_helper_find(helpname, u3,
2928                                                     nf_ct_protonum(ct));
2929                 if (helper == NULL) {
2930 #ifdef CONFIG_MODULES
2931                         if (request_module("nfct-helper-%s", helpname) < 0) {
2932                                 err = -EOPNOTSUPP;
2933                                 goto err_ct;
2934                         }
2935                         helper = __nf_conntrack_helper_find(helpname, u3,
2936                                                             nf_ct_protonum(ct));
2937                         if (helper) {
2938                                 err = -EAGAIN;
2939                                 goto err_ct;
2940                         }
2941 #endif
2942                         err = -EOPNOTSUPP;
2943                         goto err_ct;
2944                 }
2945         }
2946
2947         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
2948         if (IS_ERR(exp)) {
2949                 err = PTR_ERR(exp);
2950                 goto err_ct;
2951         }
2952
2953         err = nf_ct_expect_related_report(exp, portid, report);
2954         if (err < 0)
2955                 goto err_exp;
2956
2957         return 0;
2958 err_exp:
2959         nf_ct_expect_put(exp);
2960 err_ct:
2961         nf_ct_put(ct);
2962         return err;
2963 }
2964
2965 static int
2966 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2967                      const struct nlmsghdr *nlh,
2968                      const struct nlattr * const cda[])
2969 {
2970         struct net *net = sock_net(ctnl);
2971         struct nf_conntrack_tuple tuple;
2972         struct nf_conntrack_expect *exp;
2973         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2974         u_int8_t u3 = nfmsg->nfgen_family;
2975         u16 zone;
2976         int err;
2977
2978         if (!cda[CTA_EXPECT_TUPLE]
2979             || !cda[CTA_EXPECT_MASK]
2980             || !cda[CTA_EXPECT_MASTER])
2981                 return -EINVAL;
2982
2983         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2984         if (err < 0)
2985                 return err;
2986
2987         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2988         if (err < 0)
2989                 return err;
2990
2991         spin_lock_bh(&nf_conntrack_expect_lock);
2992         exp = __nf_ct_expect_find(net, zone, &tuple);
2993
2994         if (!exp) {
2995                 spin_unlock_bh(&nf_conntrack_expect_lock);
2996                 err = -ENOENT;
2997                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2998                         err = ctnetlink_create_expect(net, zone, cda,
2999                                                       u3,
3000                                                       NETLINK_CB(skb).portid,
3001                                                       nlmsg_report(nlh));
3002                 }
3003                 return err;
3004         }
3005
3006         err = -EEXIST;
3007         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3008                 err = ctnetlink_change_expect(exp, cda);
3009         spin_unlock_bh(&nf_conntrack_expect_lock);
3010
3011         return err;
3012 }
3013
3014 static int
3015 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3016                              const struct ip_conntrack_stat *st)
3017 {
3018         struct nlmsghdr *nlh;
3019         struct nfgenmsg *nfmsg;
3020         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3021
3022         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3023         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3024         if (nlh == NULL)
3025                 goto nlmsg_failure;
3026
3027         nfmsg = nlmsg_data(nlh);
3028         nfmsg->nfgen_family = AF_UNSPEC;
3029         nfmsg->version      = NFNETLINK_V0;
3030         nfmsg->res_id       = htons(cpu);
3031
3032         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3033             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3034             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3035                 goto nla_put_failure;
3036
3037         nlmsg_end(skb, nlh);
3038         return skb->len;
3039
3040 nla_put_failure:
3041 nlmsg_failure:
3042         nlmsg_cancel(skb, nlh);
3043         return -1;
3044 }
3045
3046 static int
3047 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3048 {
3049         int cpu;
3050         struct net *net = sock_net(skb->sk);
3051
3052         if (cb->args[0] == nr_cpu_ids)
3053                 return 0;
3054
3055         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3056                 const struct ip_conntrack_stat *st;
3057
3058                 if (!cpu_possible(cpu))
3059                         continue;
3060
3061                 st = per_cpu_ptr(net->ct.stat, cpu);
3062                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3063                                                  cb->nlh->nlmsg_seq,
3064                                                  cpu, st) < 0)
3065                         break;
3066         }
3067         cb->args[0] = cpu;
3068
3069         return skb->len;
3070 }
3071
3072 static int
3073 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3074                        const struct nlmsghdr *nlh,
3075                        const struct nlattr * const cda[])
3076 {
3077         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3078                 struct netlink_dump_control c = {
3079                         .dump = ctnetlink_exp_stat_cpu_dump,
3080                 };
3081                 return netlink_dump_start(ctnl, skb, nlh, &c);
3082         }
3083
3084         return 0;
3085 }
3086
3087 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3088 static struct nf_ct_event_notifier ctnl_notifier = {
3089         .fcn = ctnetlink_conntrack_event,
3090 };
3091
3092 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3093         .fcn = ctnetlink_expect_event,
3094 };
3095 #endif
3096
3097 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3098         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3099                                             .attr_count = CTA_MAX,
3100                                             .policy = ct_nla_policy },
3101         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3102                                             .attr_count = CTA_MAX,
3103                                             .policy = ct_nla_policy },
3104         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3105                                             .attr_count = CTA_MAX,
3106                                             .policy = ct_nla_policy },
3107         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3108                                             .attr_count = CTA_MAX,
3109                                             .policy = ct_nla_policy },
3110         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3111         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3112         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3113         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3114 };
3115
3116 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3117         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3118                                             .attr_count = CTA_EXPECT_MAX,
3119                                             .policy = exp_nla_policy },
3120         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3121                                             .attr_count = CTA_EXPECT_MAX,
3122                                             .policy = exp_nla_policy },
3123         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3124                                             .attr_count = CTA_EXPECT_MAX,
3125                                             .policy = exp_nla_policy },
3126         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3127 };
3128
3129 static const struct nfnetlink_subsystem ctnl_subsys = {
3130         .name                           = "conntrack",
3131         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3132         .cb_count                       = IPCTNL_MSG_MAX,
3133         .cb                             = ctnl_cb,
3134 };
3135
3136 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3137         .name                           = "conntrack_expect",
3138         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3139         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3140         .cb                             = ctnl_exp_cb,
3141 };
3142
3143 MODULE_ALIAS("ip_conntrack_netlink");
3144 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3145 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3146
3147 static int __net_init ctnetlink_net_init(struct net *net)
3148 {
3149 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3150         int ret;
3151
3152         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3153         if (ret < 0) {
3154                 pr_err("ctnetlink_init: cannot register notifier.\n");
3155                 goto err_out;
3156         }
3157
3158         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3159         if (ret < 0) {
3160                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3161                 goto err_unreg_notifier;
3162         }
3163 #endif
3164         return 0;
3165
3166 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3167 err_unreg_notifier:
3168         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3169 err_out:
3170         return ret;
3171 #endif
3172 }
3173
3174 static void ctnetlink_net_exit(struct net *net)
3175 {
3176 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3177         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3178         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3179 #endif
3180 }
3181
3182 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3183 {
3184         struct net *net;
3185
3186         list_for_each_entry(net, net_exit_list, exit_list)
3187                 ctnetlink_net_exit(net);
3188 }
3189
3190 static struct pernet_operations ctnetlink_net_ops = {
3191         .init           = ctnetlink_net_init,
3192         .exit_batch     = ctnetlink_net_exit_batch,
3193 };
3194
3195 static int __init ctnetlink_init(void)
3196 {
3197         int ret;
3198
3199         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3200         ret = nfnetlink_subsys_register(&ctnl_subsys);
3201         if (ret < 0) {
3202                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3203                 goto err_out;
3204         }
3205
3206         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3207         if (ret < 0) {
3208                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3209                 goto err_unreg_subsys;
3210         }
3211
3212         ret = register_pernet_subsys(&ctnetlink_net_ops);
3213         if (ret < 0) {
3214                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3215                 goto err_unreg_exp_subsys;
3216         }
3217 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3218         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3219         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3220 #endif
3221         return 0;
3222
3223 err_unreg_exp_subsys:
3224         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3225 err_unreg_subsys:
3226         nfnetlink_subsys_unregister(&ctnl_subsys);
3227 err_out:
3228         return ret;
3229 }
3230
3231 static void __exit ctnetlink_exit(void)
3232 {
3233         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3234
3235         unregister_pernet_subsys(&ctnetlink_net_ops);
3236         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3237         nfnetlink_subsys_unregister(&ctnl_subsys);
3238 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3239         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3240 #endif
3241 }
3242
3243 module_init(ctnetlink_init);
3244 module_exit(ctnetlink_exit);