kernel/net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_seqadj.h>
  41 #include <net/netfilter/nf_conntrack_l3proto.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #ifdef CONFIG_NF_NAT_NEEDED
  49 #include <net/netfilter/nf_nat_core.h>
  50 #include <net/netfilter/nf_nat_l4proto.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 MODULE_LICENSE("GPL");
  58
  59 static char __initdata version[] = "0.93";
  60
  61 static inline int
  62 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  63                             const struct nf_conntrack_tuple *tuple,
  64                             struct nf_conntrack_l4proto *l4proto)
  65 {
  66         int ret = 0;
  67         struct nlattr *nest_parms;
  68
  69         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  70         if (!nest_parms)
  71                 goto nla_put_failure;
  72         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  73                 goto nla_put_failure;
  74
  75         if (likely(l4proto->tuple_to_nlattr))
  76                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  77
  78         nla_nest_end(skb, nest_parms);
  79
  80         return ret;
  81
  82 nla_put_failure:
  83         return -1;
  84 }
  85
  86 static inline int
  87 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  88                          const struct nf_conntrack_tuple *tuple,
  89                          struct nf_conntrack_l3proto *l3proto)
  90 {
  91         int ret = 0;
  92         struct nlattr *nest_parms;
  93
  94         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  95         if (!nest_parms)
  96                 goto nla_put_failure;
  97
  98         if (likely(l3proto->tuple_to_nlattr))
  99                 ret = l3proto->tuple_to_nlattr(skb, tuple);
 100
 101         nla_nest_end(skb, nest_parms);
 102
 103         return ret;
 104
 105 nla_put_failure:
 106         return -1;
 107 }
 108
 109 static int
 110 ctnetlink_dump_tuples(struct sk_buff *skb,
 111                       const struct nf_conntrack_tuple *tuple)
 112 {
 113         int ret;
 114         struct nf_conntrack_l3proto *l3proto;
 115         struct nf_conntrack_l4proto *l4proto;
 116
 117         rcu_read_lock();
 118         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 119         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 120
 121         if (ret >= 0) {
 122                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 123                                                tuple->dst.protonum);
 124                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 125         }
 126         rcu_read_unlock();
 127         return ret;
 128 }
 129
 130 static inline int
 131 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 132 {
 133         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 134                 goto nla_put_failure;
 135         return 0;
 136
 137 nla_put_failure:
 138         return -1;
 139 }
 140
 141 static inline int
 142 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 143 {
 144         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 145
 146         if (timeout < 0)
 147                 timeout = 0;
 148
 149         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 150                 goto nla_put_failure;
 151         return 0;
 152
 153 nla_put_failure:
 154         return -1;
 155 }
 156
 157 static inline int
 158 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 159 {
 160         struct nf_conntrack_l4proto *l4proto;
 161         struct nlattr *nest_proto;
 162         int ret;
 163
 164         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 165         if (!l4proto->to_nlattr)
 166                 return 0;
 167
 168         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 169         if (!nest_proto)
 170                 goto nla_put_failure;
 171
 172         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 173
 174         nla_nest_end(skb, nest_proto);
 175
 176         return ret;
 177
 178 nla_put_failure:
 179         return -1;
 180 }
 181
 182 static inline int
 183 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 184 {
 185         struct nlattr *nest_helper;
 186         const struct nf_conn_help *help = nfct_help(ct);
 187         struct nf_conntrack_helper *helper;
 188
 189         if (!help)
 190                 return 0;
 191
 192         helper = rcu_dereference(help->helper);
 193         if (!helper)
 194                 goto out;
 195
 196         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 197         if (!nest_helper)
 198                 goto nla_put_failure;
 199         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 200                 goto nla_put_failure;
 201
 202         if (helper->to_nlattr)
 203                 helper->to_nlattr(skb, ct);
 204
 205         nla_nest_end(skb, nest_helper);
 206 out:
 207         return 0;
 208
 209 nla_put_failure:
 210         return -1;
 211 }
 212
 213 static int
 214 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 215               enum ip_conntrack_dir dir, int type)
 216 {
 217         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 218         struct nf_conn_counter *counter = acct->counter;
 219         struct nlattr *nest_count;
 220         u64 pkts, bytes;
 221
 222         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 223                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 224                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 225         } else {
 226                 pkts = atomic64_read(&counter[dir].packets);
 227                 bytes = atomic64_read(&counter[dir].bytes);
 228         }
 229
 230         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 231         if (!nest_count)
 232                 goto nla_put_failure;
 233
 234         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 235             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 236                 goto nla_put_failure;
 237
 238         nla_nest_end(skb, nest_count);
 239
 240         return 0;
 241
 242 nla_put_failure:
 243         return -1;
 244 }
 245
 246 static int
 247 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 248 {
 249         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 250
 251         if (!acct)
 252                 return 0;
 253
 254         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 255                 return -1;
 256         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 257                 return -1;
 258
 259         return 0;
 260 }
 261
 262 static int
 263 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 264 {
 265         struct nlattr *nest_count;
 266         const struct nf_conn_tstamp *tstamp;
 267
 268         tstamp = nf_conn_tstamp_find(ct);
 269         if (!tstamp)
 270                 return 0;
 271
 272         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 273         if (!nest_count)
 274                 goto nla_put_failure;
 275
 276         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 277             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 278                                                cpu_to_be64(tstamp->stop))))
 279                 goto nla_put_failure;
 280         nla_nest_end(skb, nest_count);
 281
 282         return 0;
 283
 284 nla_put_failure:
 285         return -1;
 286 }
 287
 288 #ifdef CONFIG_NF_CONNTRACK_MARK
 289 static inline int
 290 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 291 {
 292         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 293                 goto nla_put_failure;
 294         return 0;
 295
 296 nla_put_failure:
 297         return -1;
 298 }
 299 #else
 300 #define ctnetlink_dump_mark(a, b) (0)
 301 #endif
 302
 303 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 304 static inline int
 305 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 306 {
 307         struct nlattr *nest_secctx;
 308         int len, ret;
 309         char *secctx;
 310
 311         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 312         if (ret)
 313                 return 0;
 314
 315         ret = -1;
 316         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 317         if (!nest_secctx)
 318                 goto nla_put_failure;
 319
 320         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 321                 goto nla_put_failure;
 322         nla_nest_end(skb, nest_secctx);
 323
 324         ret = 0;
 325 nla_put_failure:
 326         security_release_secctx(secctx, len);
 327         return ret;
 328 }
 329 #else
 330 #define ctnetlink_dump_secctx(a, b) (0)
 331 #endif
 332
 333 #ifdef CONFIG_NF_CONNTRACK_LABELS
 334 static int ctnetlink_label_size(const struct nf_conn *ct)
 335 {
 336         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 337
 338         if (!labels)
 339                 return 0;
 340         return nla_total_size(labels->words * sizeof(long));
 341 }
 342
 343 static int
 344 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 345 {
 346         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 347         unsigned int len, i;
 348
 349         if (!labels)
 350                 return 0;
 351
 352         len = labels->words * sizeof(long);
 353         i = 0;
 354         do {
 355                 if (labels->bits[i] != 0)
 356                         return nla_put(skb, CTA_LABELS, len, labels->bits);
 357                 i++;
 358         } while (i < labels->words);
 359
 360         return 0;
 361 }
 362 #else
 363 #define ctnetlink_dump_labels(a, b) (0)
 364 #define ctnetlink_label_size(a) (0)
 365 #endif
 366
 367 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 368
 369 static inline int
 370 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 371 {
 372         struct nlattr *nest_parms;
 373
 374         if (!(ct->status & IPS_EXPECTED))
 375                 return 0;
 376
 377         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 378         if (!nest_parms)
 379                 goto nla_put_failure;
 380         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 381                 goto nla_put_failure;
 382         nla_nest_end(skb, nest_parms);
 383
 384         return 0;
 385
 386 nla_put_failure:
 387         return -1;
 388 }
 389
 390 static int
 391 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 392 {
 393         struct nlattr *nest_parms;
 394
 395         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 396         if (!nest_parms)
 397                 goto nla_put_failure;
 398
 399         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 400                          htonl(seq->correction_pos)) ||
 401             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 402                          htonl(seq->offset_before)) ||
 403             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 404                          htonl(seq->offset_after)))
 405                 goto nla_put_failure;
 406
 407         nla_nest_end(skb, nest_parms);
 408
 409         return 0;
 410
 411 nla_put_failure:
 412         return -1;
 413 }
 414
 415 static inline int
 416 ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 417 {
 418         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 419         struct nf_ct_seqadj *seq;
 420
 421         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 422                 return 0;
 423
 424         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 425         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 426                 return -1;
 427
 428         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 429         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 430                 return -1;
 431
 432         return 0;
 433 }
 434
 435 static inline int
 436 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 437 {
 438         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 439                 goto nla_put_failure;
 440         return 0;
 441
 442 nla_put_failure:
 443         return -1;
 444 }
 445
 446 static inline int
 447 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 448 {
 449         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 450                 goto nla_put_failure;
 451         return 0;
 452
 453 nla_put_failure:
 454         return -1;
 455 }
 456
 457 static int
 458 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 459                     struct nf_conn *ct)
 460 {
 461         struct nlmsghdr *nlh;
 462         struct nfgenmsg *nfmsg;
 463         struct nlattr *nest_parms;
 464         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 465
 466         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 467         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 468         if (nlh == NULL)
 469                 goto nlmsg_failure;
 470
 471         nfmsg = nlmsg_data(nlh);
 472         nfmsg->nfgen_family = nf_ct_l3num(ct);
 473         nfmsg->version      = NFNETLINK_V0;
 474         nfmsg->res_id       = 0;
 475
 476         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 477         if (!nest_parms)
 478                 goto nla_put_failure;
 479         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 480                 goto nla_put_failure;
 481         nla_nest_end(skb, nest_parms);
 482
 483         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 484         if (!nest_parms)
 485                 goto nla_put_failure;
 486         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 487                 goto nla_put_failure;
 488         nla_nest_end(skb, nest_parms);
 489
 490         if (nf_ct_zone(ct) &&
 491             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 492                 goto nla_put_failure;
 493
 494         if (ctnetlink_dump_status(skb, ct) < 0 ||
 495             ctnetlink_dump_timeout(skb, ct) < 0 ||
 496             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 497             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 498             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 499             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 500             ctnetlink_dump_mark(skb, ct) < 0 ||
 501             ctnetlink_dump_secctx(skb, ct) < 0 ||
 502             ctnetlink_dump_labels(skb, ct) < 0 ||
 503             ctnetlink_dump_id(skb, ct) < 0 ||
 504             ctnetlink_dump_use(skb, ct) < 0 ||
 505             ctnetlink_dump_master(skb, ct) < 0 ||
 506             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 507                 goto nla_put_failure;
 508
 509         nlmsg_end(skb, nlh);
 510         return skb->len;
 511
 512 nlmsg_failure:
 513 nla_put_failure:
 514         nlmsg_cancel(skb, nlh);
 515         return -1;
 516 }
 517
 518 static inline size_t
 519 ctnetlink_proto_size(const struct nf_conn *ct)
 520 {
 521         struct nf_conntrack_l3proto *l3proto;
 522         struct nf_conntrack_l4proto *l4proto;
 523         size_t len = 0;
 524
 525         rcu_read_lock();
 526         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 527         len += l3proto->nla_size;
 528
 529         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 530         len += l4proto->nla_size;
 531         rcu_read_unlock();
 532
 533         return len;
 534 }
 535
 536 static inline size_t
 537 ctnetlink_acct_size(const struct nf_conn *ct)
 538 {
 539         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 540                 return 0;
 541         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 542                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 543                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 544                ;
 545 }
 546
 547 static inline int
 548 ctnetlink_secctx_size(const struct nf_conn *ct)
 549 {
 550 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 551         int len, ret;
 552
 553         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 554         if (ret)
 555                 return 0;
 556
 557         return nla_total_size(0) /* CTA_SECCTX */
 558                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 559 #else
 560         return 0;
 561 #endif
 562 }
 563
 564 static inline size_t
 565 ctnetlink_timestamp_size(const struct nf_conn *ct)
 566 {
 567 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 568         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 569                 return 0;
 570         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 571 #else
 572         return 0;
 573 #endif
 574 }
 575
 576 static inline size_t
 577 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 578 {
 579         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 580                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 581                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 582                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 583                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 584                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 585                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 586                + ctnetlink_acct_size(ct)
 587                + ctnetlink_timestamp_size(ct)
 588                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 589                + nla_total_size(0) /* CTA_PROTOINFO */
 590                + nla_total_size(0) /* CTA_HELP */
 591                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 592                + ctnetlink_secctx_size(ct)
 593 #ifdef CONFIG_NF_NAT_NEEDED
 594                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 595                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 596 #endif
 597 #ifdef CONFIG_NF_CONNTRACK_MARK
 598                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 599 #endif
 600 #ifdef CONFIG_NF_CONNTRACK_ZONES
 601                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE */
 602 #endif
 603                + ctnetlink_proto_size(ct)
 604                + ctnetlink_label_size(ct)
 605                ;
 606 }
 607
 608 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 609 static int
 610 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 611 {
 612         struct net *net;
 613         struct nlmsghdr *nlh;
 614         struct nfgenmsg *nfmsg;
 615         struct nlattr *nest_parms;
 616         struct nf_conn *ct = item->ct;
 617         struct sk_buff *skb;
 618         unsigned int type;
 619         unsigned int flags = 0, group;
 620         int err;
 621
 622         /* ignore our fake conntrack entry */
 623         if (nf_ct_is_untracked(ct))
 624                 return 0;
 625
 626         if (events & (1 << IPCT_DESTROY)) {
 627                 type = IPCTNL_MSG_CT_DELETE;
 628                 group = NFNLGRP_CONNTRACK_DESTROY;
 629         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 630                 type = IPCTNL_MSG_CT_NEW;
 631                 flags = NLM_F_CREATE|NLM_F_EXCL;
 632                 group = NFNLGRP_CONNTRACK_NEW;
 633         } else  if (events) {
 634                 type = IPCTNL_MSG_CT_NEW;
 635                 group = NFNLGRP_CONNTRACK_UPDATE;
 636         } else
 637                 return 0;
 638
 639         net = nf_ct_net(ct);
 640         if (!item->report && !nfnetlink_has_listeners(net, group))
 641                 return 0;
 642
 643         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 644         if (skb == NULL)
 645                 goto errout;
 646
 647         type |= NFNL_SUBSYS_CTNETLINK << 8;
 648         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 649         if (nlh == NULL)
 650                 goto nlmsg_failure;
 651
 652         nfmsg = nlmsg_data(nlh);
 653         nfmsg->nfgen_family = nf_ct_l3num(ct);
 654         nfmsg->version  = NFNETLINK_V0;
 655         nfmsg->res_id   = 0;
 656
 657         rcu_read_lock();
 658         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 659         if (!nest_parms)
 660                 goto nla_put_failure;
 661         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 662                 goto nla_put_failure;
 663         nla_nest_end(skb, nest_parms);
 664
 665         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 666         if (!nest_parms)
 667                 goto nla_put_failure;
 668         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 669                 goto nla_put_failure;
 670         nla_nest_end(skb, nest_parms);
 671
 672         if (nf_ct_zone(ct) &&
 673             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 674                 goto nla_put_failure;
 675
 676         if (ctnetlink_dump_id(skb, ct) < 0)
 677                 goto nla_put_failure;
 678
 679         if (ctnetlink_dump_status(skb, ct) < 0)
 680                 goto nla_put_failure;
 681
 682         if (events & (1 << IPCT_DESTROY)) {
 683                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 684                     ctnetlink_dump_timestamp(skb, ct) < 0)
 685                         goto nla_put_failure;
 686         } else {
 687                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 688                         goto nla_put_failure;
 689
 690                 if (events & (1 << IPCT_PROTOINFO)
 691                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 692                         goto nla_put_failure;
 693
 694                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 695                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 696                         goto nla_put_failure;
 697
 698 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 699                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 700                     && ctnetlink_dump_secctx(skb, ct) < 0)
 701                         goto nla_put_failure;
 702 #endif
 703                 if (events & (1 << IPCT_LABEL) &&
 704                      ctnetlink_dump_labels(skb, ct) < 0)
 705                         goto nla_put_failure;
 706
 707                 if (events & (1 << IPCT_RELATED) &&
 708                     ctnetlink_dump_master(skb, ct) < 0)
 709                         goto nla_put_failure;
 710
 711                 if (events & (1 << IPCT_SEQADJ) &&
 712                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 713                         goto nla_put_failure;
 714         }
 715
 716 #ifdef CONFIG_NF_CONNTRACK_MARK
 717         if ((events & (1 << IPCT_MARK) || ct->mark)
 718             && ctnetlink_dump_mark(skb, ct) < 0)
 719                 goto nla_put_failure;
 720 #endif
 721         rcu_read_unlock();
 722
 723         nlmsg_end(skb, nlh);
 724         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 725                              GFP_ATOMIC);
 726         if (err == -ENOBUFS || err == -EAGAIN)
 727                 return -ENOBUFS;
 728
 729         return 0;
 730
 731 nla_put_failure:
 732         rcu_read_unlock();
 733         nlmsg_cancel(skb, nlh);
 734 nlmsg_failure:
 735         kfree_skb(skb);
 736 errout:
 737         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 738                 return -ENOBUFS;
 739
 740         return 0;
 741 }
 742 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 743
 744 static int ctnetlink_done(struct netlink_callback *cb)
 745 {
 746         if (cb->args[1])
 747                 nf_ct_put((struct nf_conn *)cb->args[1]);
 748         kfree(cb->data);
 749         return 0;
 750 }
 751
 752 struct ctnetlink_filter {
 753         struct {
 754                 u_int32_t val;
 755                 u_int32_t mask;
 756         } mark;
 757 };
 758
 759 static struct ctnetlink_filter *
 760 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 761 {
 762 #ifdef CONFIG_NF_CONNTRACK_MARK
 763         struct ctnetlink_filter *filter;
 764
 765         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 766         if (filter == NULL)
 767                 return ERR_PTR(-ENOMEM);
 768
 769         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 770         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 771
 772         return filter;
 773 #else
 774         return ERR_PTR(-EOPNOTSUPP);
 775 #endif
 776 }
 777
 778 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 779 {
 780         struct ctnetlink_filter *filter = data;
 781
 782         if (filter == NULL)
 783                 return 1;
 784
 785 #ifdef CONFIG_NF_CONNTRACK_MARK
 786         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 787                 return 1;
 788 #endif
 789
 790         return 0;
 791 }
 792
 793 static int
 794 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 795 {
 796         struct net *net = sock_net(skb->sk);
 797         struct nf_conn *ct, *last;
 798         struct nf_conntrack_tuple_hash *h;
 799         struct hlist_nulls_node *n;
 800         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 801         u_int8_t l3proto = nfmsg->nfgen_family;
 802         int res;
 803         spinlock_t *lockp;
 804
 805         last = (struct nf_conn *)cb->args[1];
 806
 807         local_bh_disable();
 808         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 809 restart:
 810                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 811                 spin_lock(lockp);
 812                 if (cb->args[0] >= net->ct.htable_size) {
 813                         spin_unlock(lockp);
 814                         goto out;
 815                 }
 816                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 817                                          hnnode) {
 818                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 819                                 continue;
 820                         ct = nf_ct_tuplehash_to_ctrack(h);
 821                         /* Dump entries of a given L3 protocol number.
 822                          * If it is not specified, ie. l3proto == 0,
 823                          * then dump everything. */
 824                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 825                                 continue;
 826                         if (cb->args[1]) {
 827                                 if (ct != last)
 828                                         continue;
 829                                 cb->args[1] = 0;
 830                         }
 831                         if (!ctnetlink_filter_match(ct, cb->data))
 832                                 continue;
 833
 834                         rcu_read_lock();
 835                         res =
 836                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 837                                             cb->nlh->nlmsg_seq,
 838                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 839                                             ct);
 840                         rcu_read_unlock();
 841                         if (res < 0) {
 842                                 nf_conntrack_get(&ct->ct_general);
 843                                 cb->args[1] = (unsigned long)ct;
 844                                 spin_unlock(lockp);
 845                                 goto out;
 846                         }
 847                 }
 848                 spin_unlock(lockp);
 849                 if (cb->args[1]) {
 850                         cb->args[1] = 0;
 851                         goto restart;
 852                 }
 853         }
 854 out:
 855         local_bh_enable();
 856         if (last)
 857                 nf_ct_put(last);
 858
 859         return skb->len;
 860 }
 861
 862 static inline int
 863 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 864 {
 865         struct nlattr *tb[CTA_IP_MAX+1];
 866         struct nf_conntrack_l3proto *l3proto;
 867         int ret = 0;
 868
 869         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 870         if (ret < 0)
 871                 return ret;
 872
 873         rcu_read_lock();
 874         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 875
 876         if (likely(l3proto->nlattr_to_tuple)) {
 877                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 878                                           l3proto->nla_policy);
 879                 if (ret == 0)
 880                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 881         }
 882
 883         rcu_read_unlock();
 884
 885         return ret;
 886 }
 887
 888 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 889         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 890 };
 891
 892 static inline int
 893 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 894                             struct nf_conntrack_tuple *tuple)
 895 {
 896         struct nlattr *tb[CTA_PROTO_MAX+1];
 897         struct nf_conntrack_l4proto *l4proto;
 898         int ret = 0;
 899
 900         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 901         if (ret < 0)
 902                 return ret;
 903
 904         if (!tb[CTA_PROTO_NUM])
 905                 return -EINVAL;
 906         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 907
 908         rcu_read_lock();
 909         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 910
 911         if (likely(l4proto->nlattr_to_tuple)) {
 912                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 913                                           l4proto->nla_policy);
 914                 if (ret == 0)
 915                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 916         }
 917
 918         rcu_read_unlock();
 919
 920         return ret;
 921 }
 922
 923 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 924         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 925         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 926 };
 927
 928 static int
 929 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 930                       struct nf_conntrack_tuple *tuple,
 931                       enum ctattr_type type, u_int8_t l3num)
 932 {
 933         struct nlattr *tb[CTA_TUPLE_MAX+1];
 934         int err;
 935
 936         memset(tuple, 0, sizeof(*tuple));
 937
 938         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 939         if (err < 0)
 940                 return err;
 941
 942         if (!tb[CTA_TUPLE_IP])
 943                 return -EINVAL;
 944
 945         tuple->src.l3num = l3num;
 946
 947         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 948         if (err < 0)
 949                 return err;
 950
 951         if (!tb[CTA_TUPLE_PROTO])
 952                 return -EINVAL;
 953
 954         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 955         if (err < 0)
 956                 return err;
 957
 958         /* orig and expect tuples get DIR_ORIGINAL */
 959         if (type == CTA_TUPLE_REPLY)
 960                 tuple->dst.dir = IP_CT_DIR_REPLY;
 961         else
 962                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 963
 964         return 0;
 965 }
 966
 967 static int
 968 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 969 {
 970         if (attr)
 971 #ifdef CONFIG_NF_CONNTRACK_ZONES
 972                 *zone = ntohs(nla_get_be16(attr));
 973 #else
 974                 return -EOPNOTSUPP;
 975 #endif
 976         else
 977                 *zone = 0;
 978
 979         return 0;
 980 }
 981
 982 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 983         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
 984                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
 985 };
 986
 987 static inline int
 988 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 989                      struct nlattr **helpinfo)
 990 {
 991         int err;
 992         struct nlattr *tb[CTA_HELP_MAX+1];
 993
 994         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 995         if (err < 0)
 996                 return err;
 997
 998         if (!tb[CTA_HELP_NAME])
 999                 return -EINVAL;
1000
1001         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1002
1003         if (tb[CTA_HELP_INFO])
1004                 *helpinfo = tb[CTA_HELP_INFO];
1005
1006         return 0;
1007 }
1008
1009 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1010         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1011         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1012         [CTA_STATUS]            = { .type = NLA_U32 },
1013         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1014         [CTA_HELP]              = { .type = NLA_NESTED },
1015         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1016         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1017         [CTA_MARK]              = { .type = NLA_U32 },
1018         [CTA_ID]                = { .type = NLA_U32 },
1019         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1020         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1021         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1022         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1023         [CTA_ZONE]              = { .type = NLA_U16 },
1024         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1025         [CTA_LABELS]            = { .type = NLA_BINARY,
1026                                     .len = NF_CT_LABELS_MAX_SIZE },
1027         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1028                                     .len = NF_CT_LABELS_MAX_SIZE },
1029 };
1030
1031 static int ctnetlink_flush_conntrack(struct net *net,
1032                                      const struct nlattr * const cda[],
1033                                      u32 portid, int report)
1034 {
1035         struct ctnetlink_filter *filter = NULL;
1036
1037         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1038                 filter = ctnetlink_alloc_filter(cda);
1039                 if (IS_ERR(filter))
1040                         return PTR_ERR(filter);
1041         }
1042
1043         nf_ct_iterate_cleanup(net, ctnetlink_filter_match, filter,
1044                               portid, report);
1045         kfree(filter);
1046
1047         return 0;
1048 }
1049
1050 static int
1051 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
1052                         const struct nlmsghdr *nlh,
1053                         const struct nlattr * const cda[])
1054 {
1055         struct net *net = sock_net(ctnl);
1056         struct nf_conntrack_tuple_hash *h;
1057         struct nf_conntrack_tuple tuple;
1058         struct nf_conn *ct;
1059         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1060         u_int8_t u3 = nfmsg->nfgen_family;
1061         u16 zone;
1062         int err;
1063
1064         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1065         if (err < 0)
1066                 return err;
1067
1068         if (cda[CTA_TUPLE_ORIG])
1069                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1070         else if (cda[CTA_TUPLE_REPLY])
1071                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1072         else {
1073                 return ctnetlink_flush_conntrack(net, cda,
1074                                                  NETLINK_CB(skb).portid,
1075                                                  nlmsg_report(nlh));
1076         }
1077
1078         if (err < 0)
1079                 return err;
1080
1081         h = nf_conntrack_find_get(net, zone, &tuple);
1082         if (!h)
1083                 return -ENOENT;
1084
1085         ct = nf_ct_tuplehash_to_ctrack(h);
1086
1087         if (cda[CTA_ID]) {
1088                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1089                 if (id != (u32)(unsigned long)ct) {
1090                         nf_ct_put(ct);
1091                         return -ENOENT;
1092                 }
1093         }
1094
1095         if (del_timer(&ct->timeout))
1096                 nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1097
1098         nf_ct_put(ct);
1099
1100         return 0;
1101 }
1102
1103 static int
1104 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1105                         const struct nlmsghdr *nlh,
1106                         const struct nlattr * const cda[])
1107 {
1108         struct net *net = sock_net(ctnl);
1109         struct nf_conntrack_tuple_hash *h;
1110         struct nf_conntrack_tuple tuple;
1111         struct nf_conn *ct;
1112         struct sk_buff *skb2 = NULL;
1113         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1114         u_int8_t u3 = nfmsg->nfgen_family;
1115         u16 zone;
1116         int err;
1117
1118         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1119                 struct netlink_dump_control c = {
1120                         .dump = ctnetlink_dump_table,
1121                         .done = ctnetlink_done,
1122                 };
1123
1124                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1125                         struct ctnetlink_filter *filter;
1126
1127                         filter = ctnetlink_alloc_filter(cda);
1128                         if (IS_ERR(filter))
1129                                 return PTR_ERR(filter);
1130
1131                         c.data = filter;
1132                 }
1133                 return netlink_dump_start(ctnl, skb, nlh, &c);
1134         }
1135
1136         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1137         if (err < 0)
1138                 return err;
1139
1140         if (cda[CTA_TUPLE_ORIG])
1141                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1142         else if (cda[CTA_TUPLE_REPLY])
1143                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1144         else
1145                 return -EINVAL;
1146
1147         if (err < 0)
1148                 return err;
1149
1150         h = nf_conntrack_find_get(net, zone, &tuple);
1151         if (!h)
1152                 return -ENOENT;
1153
1154         ct = nf_ct_tuplehash_to_ctrack(h);
1155
1156         err = -ENOMEM;
1157         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1158         if (skb2 == NULL) {
1159                 nf_ct_put(ct);
1160                 return -ENOMEM;
1161         }
1162
1163         rcu_read_lock();
1164         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1165                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1166         rcu_read_unlock();
1167         nf_ct_put(ct);
1168         if (err <= 0)
1169                 goto free;
1170
1171         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1172         if (err < 0)
1173                 goto out;
1174
1175         return 0;
1176
1177 free:
1178         kfree_skb(skb2);
1179 out:
1180         /* this avoids a loop in nfnetlink. */
1181         return err == -EAGAIN ? -ENOBUFS : err;
1182 }
1183
1184 static int ctnetlink_done_list(struct netlink_callback *cb)
1185 {
1186         if (cb->args[1])
1187                 nf_ct_put((struct nf_conn *)cb->args[1]);
1188         return 0;
1189 }
1190
1191 static int
1192 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1193 {
1194         struct nf_conn *ct, *last;
1195         struct nf_conntrack_tuple_hash *h;
1196         struct hlist_nulls_node *n;
1197         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1198         u_int8_t l3proto = nfmsg->nfgen_family;
1199         int res;
1200         int cpu;
1201         struct hlist_nulls_head *list;
1202         struct net *net = sock_net(skb->sk);
1203
1204         if (cb->args[2])
1205                 return 0;
1206
1207         last = (struct nf_conn *)cb->args[1];
1208
1209         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1210                 struct ct_pcpu *pcpu;
1211
1212                 if (!cpu_possible(cpu))
1213                         continue;
1214
1215                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1216                 spin_lock_bh(&pcpu->lock);
1217                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1218 restart:
1219                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1220                         ct = nf_ct_tuplehash_to_ctrack(h);
1221                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1222                                 continue;
1223                         if (cb->args[1]) {
1224                                 if (ct != last)
1225                                         continue;
1226                                 cb->args[1] = 0;
1227                         }
1228                         rcu_read_lock();
1229                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1230                                                   cb->nlh->nlmsg_seq,
1231                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1232                                                   ct);
1233                         rcu_read_unlock();
1234                         if (res < 0) {
1235                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1236                                         continue;
1237                                 cb->args[0] = cpu;
1238                                 cb->args[1] = (unsigned long)ct;
1239                                 spin_unlock_bh(&pcpu->lock);
1240                                 goto out;
1241                         }
1242                 }
1243                 if (cb->args[1]) {
1244                         cb->args[1] = 0;
1245                         goto restart;
1246                 }
1247                 spin_unlock_bh(&pcpu->lock);
1248         }
1249         cb->args[2] = 1;
1250 out:
1251         if (last)
1252                 nf_ct_put(last);
1253
1254         return skb->len;
1255 }
1256
1257 static int
1258 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1259 {
1260         return ctnetlink_dump_list(skb, cb, true);
1261 }
1262
1263 static int
1264 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1265                        const struct nlmsghdr *nlh,
1266                        const struct nlattr * const cda[])
1267 {
1268         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1269                 struct netlink_dump_control c = {
1270                         .dump = ctnetlink_dump_dying,
1271                         .done = ctnetlink_done_list,
1272                 };
1273                 return netlink_dump_start(ctnl, skb, nlh, &c);
1274         }
1275
1276         return -EOPNOTSUPP;
1277 }
1278
1279 static int
1280 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1281 {
1282         return ctnetlink_dump_list(skb, cb, false);
1283 }
1284
1285 static int
1286 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1287                              const struct nlmsghdr *nlh,
1288                              const struct nlattr * const cda[])
1289 {
1290         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1291                 struct netlink_dump_control c = {
1292                         .dump = ctnetlink_dump_unconfirmed,
1293                         .done = ctnetlink_done_list,
1294                 };
1295                 return netlink_dump_start(ctnl, skb, nlh, &c);
1296         }
1297
1298         return -EOPNOTSUPP;
1299 }
1300
1301 #ifdef CONFIG_NF_NAT_NEEDED
1302 static int
1303 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1304                           enum nf_nat_manip_type manip,
1305                           const struct nlattr *attr)
1306 {
1307         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1308         int err;
1309
1310         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1311         if (!parse_nat_setup) {
1312 #ifdef CONFIG_MODULES
1313                 rcu_read_unlock();
1314                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1315                 if (request_module("nf-nat") < 0) {
1316                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1317                         rcu_read_lock();
1318                         return -EOPNOTSUPP;
1319                 }
1320                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1321                 rcu_read_lock();
1322                 if (nfnetlink_parse_nat_setup_hook)
1323                         return -EAGAIN;
1324 #endif
1325                 return -EOPNOTSUPP;
1326         }
1327
1328         err = parse_nat_setup(ct, manip, attr);
1329         if (err == -EAGAIN) {
1330 #ifdef CONFIG_MODULES
1331                 rcu_read_unlock();
1332                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1333                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1334                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1335                         rcu_read_lock();
1336                         return -EOPNOTSUPP;
1337                 }
1338                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1339                 rcu_read_lock();
1340 #else
1341                 err = -EOPNOTSUPP;
1342 #endif
1343         }
1344         return err;
1345 }
1346 #endif
1347
1348 static int
1349 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1350 {
1351         unsigned long d;
1352         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1353         d = ct->status ^ status;
1354
1355         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1356                 /* unchangeable */
1357                 return -EBUSY;
1358
1359         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1360                 /* SEEN_REPLY bit can only be set */
1361                 return -EBUSY;
1362
1363         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1364                 /* ASSURED bit can only be set */
1365                 return -EBUSY;
1366
1367         /* Be careful here, modifying NAT bits can screw up things,
1368          * so don't let users modify them directly if they don't pass
1369          * nf_nat_range. */
1370         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1371         return 0;
1372 }
1373
1374 static int
1375 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1376 {
1377 #ifdef CONFIG_NF_NAT_NEEDED
1378         int ret;
1379
1380         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1381                 return 0;
1382
1383         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1384                                         cda[CTA_NAT_DST]);
1385         if (ret < 0)
1386                 return ret;
1387
1388         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1389                                         cda[CTA_NAT_SRC]);
1390         return ret;
1391 #else
1392         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1393                 return 0;
1394         return -EOPNOTSUPP;
1395 #endif
1396 }
1397
1398 static inline int
1399 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1400 {
1401         struct nf_conntrack_helper *helper;
1402         struct nf_conn_help *help = nfct_help(ct);
1403         char *helpname = NULL;
1404         struct nlattr *helpinfo = NULL;
1405         int err;
1406
1407         /* don't change helper of sibling connections */
1408         if (ct->master)
1409                 return -EBUSY;
1410
1411         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1412         if (err < 0)
1413                 return err;
1414
1415         if (!strcmp(helpname, "")) {
1416                 if (help && help->helper) {
1417                         /* we had a helper before ... */
1418                         nf_ct_remove_expectations(ct);
1419                         RCU_INIT_POINTER(help->helper, NULL);
1420                 }
1421
1422                 return 0;
1423         }
1424
1425         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1426                                             nf_ct_protonum(ct));
1427         if (helper == NULL) {
1428 #ifdef CONFIG_MODULES
1429                 spin_unlock_bh(&nf_conntrack_expect_lock);
1430
1431                 if (request_module("nfct-helper-%s", helpname) < 0) {
1432                         spin_lock_bh(&nf_conntrack_expect_lock);
1433                         return -EOPNOTSUPP;
1434                 }
1435
1436                 spin_lock_bh(&nf_conntrack_expect_lock);
1437                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1438                                                     nf_ct_protonum(ct));
1439                 if (helper)
1440                         return -EAGAIN;
1441 #endif
1442                 return -EOPNOTSUPP;
1443         }
1444
1445         if (help) {
1446                 if (help->helper == helper) {
1447                         /* update private helper data if allowed. */
1448                         if (helper->from_nlattr)
1449                                 helper->from_nlattr(helpinfo, ct);
1450                         return 0;
1451                 } else
1452                         return -EBUSY;
1453         }
1454
1455         /* we cannot set a helper for an existing conntrack */
1456         return -EOPNOTSUPP;
1457 }
1458
1459 static inline int
1460 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1461 {
1462         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1463
1464         if (!del_timer(&ct->timeout))
1465                 return -ETIME;
1466
1467         ct->timeout.expires = jiffies + timeout * HZ;
1468         add_timer(&ct->timeout);
1469
1470         return 0;
1471 }
1472
1473 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1474         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1475         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1476         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1477 };
1478
1479 static inline int
1480 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1481 {
1482         const struct nlattr *attr = cda[CTA_PROTOINFO];
1483         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1484         struct nf_conntrack_l4proto *l4proto;
1485         int err = 0;
1486
1487         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1488         if (err < 0)
1489                 return err;
1490
1491         rcu_read_lock();
1492         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1493         if (l4proto->from_nlattr)
1494                 err = l4proto->from_nlattr(tb, ct);
1495         rcu_read_unlock();
1496
1497         return err;
1498 }
1499
1500 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1501         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1502         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1503         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1504 };
1505
1506 static inline int
1507 change_seq_adj(struct nf_ct_seqadj *seq, const struct nlattr * const attr)
1508 {
1509         int err;
1510         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1511
1512         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy);
1513         if (err < 0)
1514                 return err;
1515
1516         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1517                 return -EINVAL;
1518
1519         seq->correction_pos =
1520                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1521
1522         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1523                 return -EINVAL;
1524
1525         seq->offset_before =
1526                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1527
1528         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1529                 return -EINVAL;
1530
1531         seq->offset_after =
1532                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1533
1534         return 0;
1535 }
1536
1537 static int
1538 ctnetlink_change_seq_adj(struct nf_conn *ct,
1539                          const struct nlattr * const cda[])
1540 {
1541         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1542         int ret = 0;
1543
1544         if (!seqadj)
1545                 return 0;
1546
1547         if (cda[CTA_SEQ_ADJ_ORIG]) {
1548                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1549                                      cda[CTA_SEQ_ADJ_ORIG]);
1550                 if (ret < 0)
1551                         return ret;
1552
1553                 ct->status |= IPS_SEQ_ADJUST;
1554         }
1555
1556         if (cda[CTA_SEQ_ADJ_REPLY]) {
1557                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1558                                      cda[CTA_SEQ_ADJ_REPLY]);
1559                 if (ret < 0)
1560                         return ret;
1561
1562                 ct->status |= IPS_SEQ_ADJUST;
1563         }
1564
1565         return 0;
1566 }
1567
1568 static int
1569 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1570 {
1571 #ifdef CONFIG_NF_CONNTRACK_LABELS
1572         size_t len = nla_len(cda[CTA_LABELS]);
1573         const void *mask = cda[CTA_LABELS_MASK];
1574
1575         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1576                 return -EINVAL;
1577
1578         if (mask) {
1579                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1580                     nla_len(cda[CTA_LABELS_MASK]) != len)
1581                         return -EINVAL;
1582                 mask = nla_data(cda[CTA_LABELS_MASK]);
1583         }
1584
1585         len /= sizeof(u32);
1586
1587         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1588 #else
1589         return -EOPNOTSUPP;
1590 #endif
1591 }
1592
1593 static int
1594 ctnetlink_change_conntrack(struct nf_conn *ct,
1595                            const struct nlattr * const cda[])
1596 {
1597         int err;
1598
1599         /* only allow NAT changes and master assignation for new conntracks */
1600         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1601                 return -EOPNOTSUPP;
1602
1603         if (cda[CTA_HELP]) {
1604                 err = ctnetlink_change_helper(ct, cda);
1605                 if (err < 0)
1606                         return err;
1607         }
1608
1609         if (cda[CTA_TIMEOUT]) {
1610                 err = ctnetlink_change_timeout(ct, cda);
1611                 if (err < 0)
1612                         return err;
1613         }
1614
1615         if (cda[CTA_STATUS]) {
1616                 err = ctnetlink_change_status(ct, cda);
1617                 if (err < 0)
1618                         return err;
1619         }
1620
1621         if (cda[CTA_PROTOINFO]) {
1622                 err = ctnetlink_change_protoinfo(ct, cda);
1623                 if (err < 0)
1624                         return err;
1625         }
1626
1627 #if defined(CONFIG_NF_CONNTRACK_MARK)
1628         if (cda[CTA_MARK])
1629                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1630 #endif
1631
1632         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1633                 err = ctnetlink_change_seq_adj(ct, cda);
1634                 if (err < 0)
1635                         return err;
1636         }
1637
1638         if (cda[CTA_LABELS]) {
1639                 err = ctnetlink_attach_labels(ct, cda);
1640                 if (err < 0)
1641                         return err;
1642         }
1643
1644         return 0;
1645 }
1646
1647 static struct nf_conn *
1648 ctnetlink_create_conntrack(struct net *net, u16 zone,
1649                            const struct nlattr * const cda[],
1650                            struct nf_conntrack_tuple *otuple,
1651                            struct nf_conntrack_tuple *rtuple,
1652                            u8 u3)
1653 {
1654         struct nf_conn *ct;
1655         int err = -EINVAL;
1656         struct nf_conntrack_helper *helper;
1657         struct nf_conn_tstamp *tstamp;
1658
1659         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1660         if (IS_ERR(ct))
1661                 return ERR_PTR(-ENOMEM);
1662
1663         if (!cda[CTA_TIMEOUT])
1664                 goto err1;
1665         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1666
1667         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1668
1669         rcu_read_lock();
1670         if (cda[CTA_HELP]) {
1671                 char *helpname = NULL;
1672                 struct nlattr *helpinfo = NULL;
1673
1674                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1675                 if (err < 0)
1676                         goto err2;
1677
1678                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1679                                                     nf_ct_protonum(ct));
1680                 if (helper == NULL) {
1681                         rcu_read_unlock();
1682 #ifdef CONFIG_MODULES
1683                         if (request_module("nfct-helper-%s", helpname) < 0) {
1684                                 err = -EOPNOTSUPP;
1685                                 goto err1;
1686                         }
1687
1688                         rcu_read_lock();
1689                         helper = __nf_conntrack_helper_find(helpname,
1690                                                             nf_ct_l3num(ct),
1691                                                             nf_ct_protonum(ct));
1692                         if (helper) {
1693                                 err = -EAGAIN;
1694                                 goto err2;
1695                         }
1696                         rcu_read_unlock();
1697 #endif
1698                         err = -EOPNOTSUPP;
1699                         goto err1;
1700                 } else {
1701                         struct nf_conn_help *help;
1702
1703                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1704                         if (help == NULL) {
1705                                 err = -ENOMEM;
1706                                 goto err2;
1707                         }
1708                         /* set private helper data if allowed. */
1709                         if (helper->from_nlattr)
1710                                 helper->from_nlattr(helpinfo, ct);
1711
1712                         /* not in hash table yet so not strictly necessary */
1713                         RCU_INIT_POINTER(help->helper, helper);
1714                 }
1715         } else {
1716                 /* try an implicit helper assignation */
1717                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1718                 if (err < 0)
1719                         goto err2;
1720         }
1721
1722         err = ctnetlink_setup_nat(ct, cda);
1723         if (err < 0)
1724                 goto err2;
1725
1726         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1727         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1728         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1729         nf_ct_labels_ext_add(ct);
1730
1731         /* we must add conntrack extensions before confirmation. */
1732         ct->status |= IPS_CONFIRMED;
1733
1734         if (cda[CTA_STATUS]) {
1735                 err = ctnetlink_change_status(ct, cda);
1736                 if (err < 0)
1737                         goto err2;
1738         }
1739
1740         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1741                 err = ctnetlink_change_seq_adj(ct, cda);
1742                 if (err < 0)
1743                         goto err2;
1744         }
1745
1746         memset(&ct->proto, 0, sizeof(ct->proto));
1747         if (cda[CTA_PROTOINFO]) {
1748                 err = ctnetlink_change_protoinfo(ct, cda);
1749                 if (err < 0)
1750                         goto err2;
1751         }
1752
1753 #if defined(CONFIG_NF_CONNTRACK_MARK)
1754         if (cda[CTA_MARK])
1755                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1756 #endif
1757
1758         /* setup master conntrack: this is a confirmed expectation */
1759         if (cda[CTA_TUPLE_MASTER]) {
1760                 struct nf_conntrack_tuple master;
1761                 struct nf_conntrack_tuple_hash *master_h;
1762                 struct nf_conn *master_ct;
1763
1764                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1765                 if (err < 0)
1766                         goto err2;
1767
1768                 master_h = nf_conntrack_find_get(net, zone, &master);
1769                 if (master_h == NULL) {
1770                         err = -ENOENT;
1771                         goto err2;
1772                 }
1773                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1774                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1775                 ct->master = master_ct;
1776         }
1777         tstamp = nf_conn_tstamp_find(ct);
1778         if (tstamp)
1779                 tstamp->start = ktime_get_real_ns();
1780
1781         err = nf_conntrack_hash_check_insert(ct);
1782         if (err < 0)
1783                 goto err2;
1784
1785         rcu_read_unlock();
1786
1787         return ct;
1788
1789 err2:
1790         rcu_read_unlock();
1791 err1:
1792         nf_conntrack_free(ct);
1793         return ERR_PTR(err);
1794 }
1795
1796 static int
1797 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1798                         const struct nlmsghdr *nlh,
1799                         const struct nlattr * const cda[])
1800 {
1801         struct net *net = sock_net(ctnl);
1802         struct nf_conntrack_tuple otuple, rtuple;
1803         struct nf_conntrack_tuple_hash *h = NULL;
1804         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1805         struct nf_conn *ct;
1806         u_int8_t u3 = nfmsg->nfgen_family;
1807         u16 zone;
1808         int err;
1809
1810         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1811         if (err < 0)
1812                 return err;
1813
1814         if (cda[CTA_TUPLE_ORIG]) {
1815                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1816                 if (err < 0)
1817                         return err;
1818         }
1819
1820         if (cda[CTA_TUPLE_REPLY]) {
1821                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1822                 if (err < 0)
1823                         return err;
1824         }
1825
1826         if (cda[CTA_TUPLE_ORIG])
1827                 h = nf_conntrack_find_get(net, zone, &otuple);
1828         else if (cda[CTA_TUPLE_REPLY])
1829                 h = nf_conntrack_find_get(net, zone, &rtuple);
1830
1831         if (h == NULL) {
1832                 err = -ENOENT;
1833                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1834                         enum ip_conntrack_events events;
1835
1836                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1837                                 return -EINVAL;
1838
1839                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1840                                                         &rtuple, u3);
1841                         if (IS_ERR(ct))
1842                                 return PTR_ERR(ct);
1843
1844                         err = 0;
1845                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1846                                 events = IPCT_RELATED;
1847                         else
1848                                 events = IPCT_NEW;
1849
1850                         if (cda[CTA_LABELS] &&
1851                             ctnetlink_attach_labels(ct, cda) == 0)
1852                                 events |= (1 << IPCT_LABEL);
1853
1854                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1855                                                       (1 << IPCT_ASSURED) |
1856                                                       (1 << IPCT_HELPER) |
1857                                                       (1 << IPCT_PROTOINFO) |
1858                                                       (1 << IPCT_SEQADJ) |
1859                                                       (1 << IPCT_MARK) | events,
1860                                                       ct, NETLINK_CB(skb).portid,
1861                                                       nlmsg_report(nlh));
1862                         nf_ct_put(ct);
1863                 }
1864
1865                 return err;
1866         }
1867         /* implicit 'else' */
1868
1869         err = -EEXIST;
1870         ct = nf_ct_tuplehash_to_ctrack(h);
1871         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1872                 spin_lock_bh(&nf_conntrack_expect_lock);
1873                 err = ctnetlink_change_conntrack(ct, cda);
1874                 spin_unlock_bh(&nf_conntrack_expect_lock);
1875                 if (err == 0) {
1876                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1877                                                       (1 << IPCT_ASSURED) |
1878                                                       (1 << IPCT_HELPER) |
1879                                                       (1 << IPCT_LABEL) |
1880                                                       (1 << IPCT_PROTOINFO) |
1881                                                       (1 << IPCT_SEQADJ) |
1882                                                       (1 << IPCT_MARK),
1883                                                       ct, NETLINK_CB(skb).portid,
1884                                                       nlmsg_report(nlh));
1885                 }
1886         }
1887
1888         nf_ct_put(ct);
1889         return err;
1890 }
1891
1892 static int
1893 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1894                                 __u16 cpu, const struct ip_conntrack_stat *st)
1895 {
1896         struct nlmsghdr *nlh;
1897         struct nfgenmsg *nfmsg;
1898         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1899
1900         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1901         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1902         if (nlh == NULL)
1903                 goto nlmsg_failure;
1904
1905         nfmsg = nlmsg_data(nlh);
1906         nfmsg->nfgen_family = AF_UNSPEC;
1907         nfmsg->version      = NFNETLINK_V0;
1908         nfmsg->res_id       = htons(cpu);
1909
1910         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1911             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1912             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1913             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1914             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1915             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1916             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1917             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1918             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1919                                 htonl(st->insert_failed)) ||
1920             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1921             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1922             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1923             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1924                                 htonl(st->search_restart)))
1925                 goto nla_put_failure;
1926
1927         nlmsg_end(skb, nlh);
1928         return skb->len;
1929
1930 nla_put_failure:
1931 nlmsg_failure:
1932         nlmsg_cancel(skb, nlh);
1933         return -1;
1934 }
1935
1936 static int
1937 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1938 {
1939         int cpu;
1940         struct net *net = sock_net(skb->sk);
1941
1942         if (cb->args[0] == nr_cpu_ids)
1943                 return 0;
1944
1945         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1946                 const struct ip_conntrack_stat *st;
1947
1948                 if (!cpu_possible(cpu))
1949                         continue;
1950
1951                 st = per_cpu_ptr(net->ct.stat, cpu);
1952                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1953                                                     NETLINK_CB(cb->skb).portid,
1954                                                     cb->nlh->nlmsg_seq,
1955                                                     cpu, st) < 0)
1956                                 break;
1957         }
1958         cb->args[0] = cpu;
1959
1960         return skb->len;
1961 }
1962
1963 static int
1964 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1965                       const struct nlmsghdr *nlh,
1966                       const struct nlattr * const cda[])
1967 {
1968         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1969                 struct netlink_dump_control c = {
1970                         .dump = ctnetlink_ct_stat_cpu_dump,
1971                 };
1972                 return netlink_dump_start(ctnl, skb, nlh, &c);
1973         }
1974
1975         return 0;
1976 }
1977
1978 static int
1979 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1980                             struct net *net)
1981 {
1982         struct nlmsghdr *nlh;
1983         struct nfgenmsg *nfmsg;
1984         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1985         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1986
1987         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1988         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1989         if (nlh == NULL)
1990                 goto nlmsg_failure;
1991
1992         nfmsg = nlmsg_data(nlh);
1993         nfmsg->nfgen_family = AF_UNSPEC;
1994         nfmsg->version      = NFNETLINK_V0;
1995         nfmsg->res_id       = 0;
1996
1997         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1998                 goto nla_put_failure;
1999
2000         nlmsg_end(skb, nlh);
2001         return skb->len;
2002
2003 nla_put_failure:
2004 nlmsg_failure:
2005         nlmsg_cancel(skb, nlh);
2006         return -1;
2007 }
2008
2009 static int
2010 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
2011                   const struct nlmsghdr *nlh,
2012                   const struct nlattr * const cda[])
2013 {
2014         struct sk_buff *skb2;
2015         int err;
2016
2017         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2018         if (skb2 == NULL)
2019                 return -ENOMEM;
2020
2021         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2022                                           nlh->nlmsg_seq,
2023                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2024                                           sock_net(skb->sk));
2025         if (err <= 0)
2026                 goto free;
2027
2028         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2029         if (err < 0)
2030                 goto out;
2031
2032         return 0;
2033
2034 free:
2035         kfree_skb(skb2);
2036 out:
2037         /* this avoids a loop in nfnetlink. */
2038         return err == -EAGAIN ? -ENOBUFS : err;
2039 }
2040
2041 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2042         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2043         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2044         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2045         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2046         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2047         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2048                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2049         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2050         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2051         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2052         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2053         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2054 };
2055
2056 static struct nf_conntrack_expect *
2057 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2058                        struct nf_conntrack_helper *helper,
2059                        struct nf_conntrack_tuple *tuple,
2060                        struct nf_conntrack_tuple *mask);
2061
2062 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2063 static size_t
2064 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
2065 {
2066         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2067                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2068                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2069                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2070                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2071                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2072                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2073                + nla_total_size(0) /* CTA_PROTOINFO */
2074                + nla_total_size(0) /* CTA_HELP */
2075                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2076                + ctnetlink_secctx_size(ct)
2077 #ifdef CONFIG_NF_NAT_NEEDED
2078                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2079                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2080 #endif
2081 #ifdef CONFIG_NF_CONNTRACK_MARK
2082                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2083 #endif
2084 #ifdef CONFIG_NF_CONNTRACK_ZONES
2085                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE */
2086 #endif
2087                + ctnetlink_proto_size(ct)
2088                ;
2089 }
2090
2091 static int
2092 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
2093 {
2094         struct nlattr *nest_parms;
2095
2096         rcu_read_lock();
2097         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2098         if (!nest_parms)
2099                 goto nla_put_failure;
2100         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2101                 goto nla_put_failure;
2102         nla_nest_end(skb, nest_parms);
2103
2104         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2105         if (!nest_parms)
2106                 goto nla_put_failure;
2107         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2108                 goto nla_put_failure;
2109         nla_nest_end(skb, nest_parms);
2110
2111         if (nf_ct_zone(ct)) {
2112                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
2113                         goto nla_put_failure;
2114         }
2115
2116         if (ctnetlink_dump_id(skb, ct) < 0)
2117                 goto nla_put_failure;
2118
2119         if (ctnetlink_dump_status(skb, ct) < 0)
2120                 goto nla_put_failure;
2121
2122         if (ctnetlink_dump_timeout(skb, ct) < 0)
2123                 goto nla_put_failure;
2124
2125         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2126                 goto nla_put_failure;
2127
2128         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2129                 goto nla_put_failure;
2130
2131 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2132         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2133                 goto nla_put_failure;
2134 #endif
2135         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2136                 goto nla_put_failure;
2137
2138         if ((ct->status & IPS_SEQ_ADJUST) &&
2139             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2140                 goto nla_put_failure;
2141
2142 #ifdef CONFIG_NF_CONNTRACK_MARK
2143         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2144                 goto nla_put_failure;
2145 #endif
2146         if (ctnetlink_dump_labels(skb, ct) < 0)
2147                 goto nla_put_failure;
2148         rcu_read_unlock();
2149         return 0;
2150
2151 nla_put_failure:
2152         rcu_read_unlock();
2153         return -ENOSPC;
2154 }
2155
2156 static int
2157 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2158 {
2159         int err;
2160
2161         if (cda[CTA_TIMEOUT]) {
2162                 err = ctnetlink_change_timeout(ct, cda);
2163                 if (err < 0)
2164                         return err;
2165         }
2166         if (cda[CTA_STATUS]) {
2167                 err = ctnetlink_change_status(ct, cda);
2168                 if (err < 0)
2169                         return err;
2170         }
2171         if (cda[CTA_HELP]) {
2172                 err = ctnetlink_change_helper(ct, cda);
2173                 if (err < 0)
2174                         return err;
2175         }
2176         if (cda[CTA_LABELS]) {
2177                 err = ctnetlink_attach_labels(ct, cda);
2178                 if (err < 0)
2179                         return err;
2180         }
2181 #if defined(CONFIG_NF_CONNTRACK_MARK)
2182         if (cda[CTA_MARK]) {
2183                 u32 mask = 0, mark, newmark;
2184                 if (cda[CTA_MARK_MASK])
2185                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2186
2187                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2188                 newmark = (ct->mark & mask) ^ mark;
2189                 if (newmark != ct->mark)
2190                         ct->mark = newmark;
2191         }
2192 #endif
2193         return 0;
2194 }
2195
2196 static int
2197 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2198 {
2199         struct nlattr *cda[CTA_MAX+1];
2200         int ret;
2201
2202         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2203         if (ret < 0)
2204                 return ret;
2205
2206         spin_lock_bh(&nf_conntrack_expect_lock);
2207         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2208         spin_unlock_bh(&nf_conntrack_expect_lock);
2209
2210         return ret;
2211 }
2212
2213 static int ctnetlink_nfqueue_exp_parse(const struct nlattr * const *cda,
2214                                        const struct nf_conn *ct,
2215                                        struct nf_conntrack_tuple *tuple,
2216                                        struct nf_conntrack_tuple *mask)
2217 {
2218         int err;
2219
2220         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2221                                     nf_ct_l3num(ct));
2222         if (err < 0)
2223                 return err;
2224
2225         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2226                                      nf_ct_l3num(ct));
2227 }
2228
2229 static int
2230 ctnetlink_nfqueue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2231                                 u32 portid, u32 report)
2232 {
2233         struct nlattr *cda[CTA_EXPECT_MAX+1];
2234         struct nf_conntrack_tuple tuple, mask;
2235         struct nf_conntrack_helper *helper = NULL;
2236         struct nf_conntrack_expect *exp;
2237         int err;
2238
2239         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy);
2240         if (err < 0)
2241                 return err;
2242
2243         err = ctnetlink_nfqueue_exp_parse((const struct nlattr * const *)cda,
2244                                           ct, &tuple, &mask);
2245         if (err < 0)
2246                 return err;
2247
2248         if (cda[CTA_EXPECT_HELP_NAME]) {
2249                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2250
2251                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2252                                                     nf_ct_protonum(ct));
2253                 if (helper == NULL)
2254                         return -EOPNOTSUPP;
2255         }
2256
2257         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2258                                      helper, &tuple, &mask);
2259         if (IS_ERR(exp))
2260                 return PTR_ERR(exp);
2261
2262         err = nf_ct_expect_related_report(exp, portid, report);
2263         if (err < 0) {
2264                 nf_ct_expect_put(exp);
2265                 return err;
2266         }
2267
2268         return 0;
2269 }
2270
2271 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2272         .build_size     = ctnetlink_nfqueue_build_size,
2273         .build          = ctnetlink_nfqueue_build,
2274         .parse          = ctnetlink_nfqueue_parse,
2275         .attach_expect  = ctnetlink_nfqueue_attach_expect,
2276         .seq_adjust     = nf_ct_tcp_seqadj_set,
2277 };
2278 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2279
2280 /***********************************************************************
2281  * EXPECT
2282  ***********************************************************************/
2283
2284 static inline int
2285 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2286                          const struct nf_conntrack_tuple *tuple,
2287                          enum ctattr_expect type)
2288 {
2289         struct nlattr *nest_parms;
2290
2291         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2292         if (!nest_parms)
2293                 goto nla_put_failure;
2294         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2295                 goto nla_put_failure;
2296         nla_nest_end(skb, nest_parms);
2297
2298         return 0;
2299
2300 nla_put_failure:
2301         return -1;
2302 }
2303
2304 static inline int
2305 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2306                         const struct nf_conntrack_tuple *tuple,
2307                         const struct nf_conntrack_tuple_mask *mask)
2308 {
2309         int ret;
2310         struct nf_conntrack_l3proto *l3proto;
2311         struct nf_conntrack_l4proto *l4proto;
2312         struct nf_conntrack_tuple m;
2313         struct nlattr *nest_parms;
2314
2315         memset(&m, 0xFF, sizeof(m));
2316         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2317         m.src.u.all = mask->src.u.all;
2318         m.dst.protonum = tuple->dst.protonum;
2319
2320         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2321         if (!nest_parms)
2322                 goto nla_put_failure;
2323
2324         rcu_read_lock();
2325         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2326         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2327         if (ret >= 0) {
2328                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2329                                                tuple->dst.protonum);
2330         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2331         }
2332         rcu_read_unlock();
2333
2334         if (unlikely(ret < 0))
2335                 goto nla_put_failure;
2336
2337         nla_nest_end(skb, nest_parms);
2338
2339         return 0;
2340
2341 nla_put_failure:
2342         return -1;
2343 }
2344
2345 static const union nf_inet_addr any_addr;
2346
2347 static int
2348 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2349                           const struct nf_conntrack_expect *exp)
2350 {
2351         struct nf_conn *master = exp->master;
2352         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2353         struct nf_conn_help *help;
2354 #ifdef CONFIG_NF_NAT_NEEDED
2355         struct nlattr *nest_parms;
2356         struct nf_conntrack_tuple nat_tuple = {};
2357 #endif
2358         struct nf_ct_helper_expectfn *expfn;
2359
2360         if (timeout < 0)
2361                 timeout = 0;
2362
2363         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2364                 goto nla_put_failure;
2365         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2366                 goto nla_put_failure;
2367         if (ctnetlink_exp_dump_tuple(skb,
2368                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2369                                  CTA_EXPECT_MASTER) < 0)
2370                 goto nla_put_failure;
2371
2372 #ifdef CONFIG_NF_NAT_NEEDED
2373         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2374             exp->saved_proto.all) {
2375                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2376                 if (!nest_parms)
2377                         goto nla_put_failure;
2378
2379                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2380                         goto nla_put_failure;
2381
2382                 nat_tuple.src.l3num = nf_ct_l3num(master);
2383                 nat_tuple.src.u3 = exp->saved_addr;
2384                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2385                 nat_tuple.src.u = exp->saved_proto;
2386
2387                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2388                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2389                         goto nla_put_failure;
2390                 nla_nest_end(skb, nest_parms);
2391         }
2392 #endif
2393         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2394             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2395             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2396             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2397                 goto nla_put_failure;
2398         help = nfct_help(master);
2399         if (help) {
2400                 struct nf_conntrack_helper *helper;
2401
2402                 helper = rcu_dereference(help->helper);
2403                 if (helper &&
2404                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2405                         goto nla_put_failure;
2406         }
2407         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2408         if (expfn != NULL &&
2409             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2410                 goto nla_put_failure;
2411
2412         return 0;
2413
2414 nla_put_failure:
2415         return -1;
2416 }
2417
2418 static int
2419 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2420                         int event, const struct nf_conntrack_expect *exp)
2421 {
2422         struct nlmsghdr *nlh;
2423         struct nfgenmsg *nfmsg;
2424         unsigned int flags = portid ? NLM_F_MULTI : 0;
2425
2426         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2427         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2428         if (nlh == NULL)
2429                 goto nlmsg_failure;
2430
2431         nfmsg = nlmsg_data(nlh);
2432         nfmsg->nfgen_family = exp->tuple.src.l3num;
2433         nfmsg->version      = NFNETLINK_V0;
2434         nfmsg->res_id       = 0;
2435
2436         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2437                 goto nla_put_failure;
2438
2439         nlmsg_end(skb, nlh);
2440         return skb->len;
2441
2442 nlmsg_failure:
2443 nla_put_failure:
2444         nlmsg_cancel(skb, nlh);
2445         return -1;
2446 }
2447
2448 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2449 static int
2450 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2451 {
2452         struct nf_conntrack_expect *exp = item->exp;
2453         struct net *net = nf_ct_exp_net(exp);
2454         struct nlmsghdr *nlh;
2455         struct nfgenmsg *nfmsg;
2456         struct sk_buff *skb;
2457         unsigned int type, group;
2458         int flags = 0;
2459
2460         if (events & (1 << IPEXP_DESTROY)) {
2461                 type = IPCTNL_MSG_EXP_DELETE;
2462                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2463         } else if (events & (1 << IPEXP_NEW)) {
2464                 type = IPCTNL_MSG_EXP_NEW;
2465                 flags = NLM_F_CREATE|NLM_F_EXCL;
2466                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2467         } else
2468                 return 0;
2469
2470         if (!item->report && !nfnetlink_has_listeners(net, group))
2471                 return 0;
2472
2473         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2474         if (skb == NULL)
2475                 goto errout;
2476
2477         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2478         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2479         if (nlh == NULL)
2480                 goto nlmsg_failure;
2481
2482         nfmsg = nlmsg_data(nlh);
2483         nfmsg->nfgen_family = exp->tuple.src.l3num;
2484         nfmsg->version      = NFNETLINK_V0;
2485         nfmsg->res_id       = 0;
2486
2487         rcu_read_lock();
2488         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2489                 goto nla_put_failure;
2490         rcu_read_unlock();
2491
2492         nlmsg_end(skb, nlh);
2493         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2494         return 0;
2495
2496 nla_put_failure:
2497         rcu_read_unlock();
2498         nlmsg_cancel(skb, nlh);
2499 nlmsg_failure:
2500         kfree_skb(skb);
2501 errout:
2502         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2503         return 0;
2504 }
2505 #endif
2506 static int ctnetlink_exp_done(struct netlink_callback *cb)
2507 {
2508         if (cb->args[1])
2509                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2510         return 0;
2511 }
2512
2513 static int
2514 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2515 {
2516         struct net *net = sock_net(skb->sk);
2517         struct nf_conntrack_expect *exp, *last;
2518         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2519         u_int8_t l3proto = nfmsg->nfgen_family;
2520
2521         rcu_read_lock();
2522         last = (struct nf_conntrack_expect *)cb->args[1];
2523         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2524 restart:
2525                 hlist_for_each_entry(exp, &net->ct.expect_hash[cb->args[0]],
2526                                      hnode) {
2527                         if (l3proto && exp->tuple.src.l3num != l3proto)
2528                                 continue;
2529                         if (cb->args[1]) {
2530                                 if (exp != last)
2531                                         continue;
2532                                 cb->args[1] = 0;
2533                         }
2534                         if (ctnetlink_exp_fill_info(skb,
2535                                                     NETLINK_CB(cb->skb).portid,
2536                                                     cb->nlh->nlmsg_seq,
2537                                                     IPCTNL_MSG_EXP_NEW,
2538                                                     exp) < 0) {
2539                                 if (!atomic_inc_not_zero(&exp->use))
2540                                         continue;
2541                                 cb->args[1] = (unsigned long)exp;
2542                                 goto out;
2543                         }
2544                 }
2545                 if (cb->args[1]) {
2546                         cb->args[1] = 0;
2547                         goto restart;
2548                 }
2549         }
2550 out:
2551         rcu_read_unlock();
2552         if (last)
2553                 nf_ct_expect_put(last);
2554
2555         return skb->len;
2556 }
2557
2558 static int
2559 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2560 {
2561         struct nf_conntrack_expect *exp, *last;
2562         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2563         struct nf_conn *ct = cb->data;
2564         struct nf_conn_help *help = nfct_help(ct);
2565         u_int8_t l3proto = nfmsg->nfgen_family;
2566
2567         if (cb->args[0])
2568                 return 0;
2569
2570         rcu_read_lock();
2571         last = (struct nf_conntrack_expect *)cb->args[1];
2572 restart:
2573         hlist_for_each_entry(exp, &help->expectations, lnode) {
2574                 if (l3proto && exp->tuple.src.l3num != l3proto)
2575                         continue;
2576                 if (cb->args[1]) {
2577                         if (exp != last)
2578                                 continue;
2579                         cb->args[1] = 0;
2580                 }
2581                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2582                                             cb->nlh->nlmsg_seq,
2583                                             IPCTNL_MSG_EXP_NEW,
2584                                             exp) < 0) {
2585                         if (!atomic_inc_not_zero(&exp->use))
2586                                 continue;
2587                         cb->args[1] = (unsigned long)exp;
2588                         goto out;
2589                 }
2590         }
2591         if (cb->args[1]) {
2592                 cb->args[1] = 0;
2593                 goto restart;
2594         }
2595         cb->args[0] = 1;
2596 out:
2597         rcu_read_unlock();
2598         if (last)
2599                 nf_ct_expect_put(last);
2600
2601         return skb->len;
2602 }
2603
2604 static int ctnetlink_dump_exp_ct(struct sock *ctnl, struct sk_buff *skb,
2605                                  const struct nlmsghdr *nlh,
2606                                  const struct nlattr * const cda[])
2607 {
2608         int err;
2609         struct net *net = sock_net(ctnl);
2610         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2611         u_int8_t u3 = nfmsg->nfgen_family;
2612         struct nf_conntrack_tuple tuple;
2613         struct nf_conntrack_tuple_hash *h;
2614         struct nf_conn *ct;
2615         u16 zone = 0;
2616         struct netlink_dump_control c = {
2617                 .dump = ctnetlink_exp_ct_dump_table,
2618                 .done = ctnetlink_exp_done,
2619         };
2620
2621         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2622         if (err < 0)
2623                 return err;
2624
2625         if (cda[CTA_EXPECT_ZONE]) {
2626                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2627                 if (err < 0)
2628                         return err;
2629         }
2630
2631         h = nf_conntrack_find_get(net, zone, &tuple);
2632         if (!h)
2633                 return -ENOENT;
2634
2635         ct = nf_ct_tuplehash_to_ctrack(h);
2636         c.data = ct;
2637
2638         err = netlink_dump_start(ctnl, skb, nlh, &c);
2639         nf_ct_put(ct);
2640
2641         return err;
2642 }
2643
2644 static int
2645 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2646                      const struct nlmsghdr *nlh,
2647                      const struct nlattr * const cda[])
2648 {
2649         struct net *net = sock_net(ctnl);
2650         struct nf_conntrack_tuple tuple;
2651         struct nf_conntrack_expect *exp;
2652         struct sk_buff *skb2;
2653         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2654         u_int8_t u3 = nfmsg->nfgen_family;
2655         u16 zone;
2656         int err;
2657
2658         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2659                 if (cda[CTA_EXPECT_MASTER])
2660                         return ctnetlink_dump_exp_ct(ctnl, skb, nlh, cda);
2661                 else {
2662                         struct netlink_dump_control c = {
2663                                 .dump = ctnetlink_exp_dump_table,
2664                                 .done = ctnetlink_exp_done,
2665                         };
2666                         return netlink_dump_start(ctnl, skb, nlh, &c);
2667                 }
2668         }
2669
2670         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2671         if (err < 0)
2672                 return err;
2673
2674         if (cda[CTA_EXPECT_TUPLE])
2675                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2676         else if (cda[CTA_EXPECT_MASTER])
2677                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2678         else
2679                 return -EINVAL;
2680
2681         if (err < 0)
2682                 return err;
2683
2684         exp = nf_ct_expect_find_get(net, zone, &tuple);
2685         if (!exp)
2686                 return -ENOENT;
2687
2688         if (cda[CTA_EXPECT_ID]) {
2689                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2690                 if (ntohl(id) != (u32)(unsigned long)exp) {
2691                         nf_ct_expect_put(exp);
2692                         return -ENOENT;
2693                 }
2694         }
2695
2696         err = -ENOMEM;
2697         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2698         if (skb2 == NULL) {
2699                 nf_ct_expect_put(exp);
2700                 goto out;
2701         }
2702
2703         rcu_read_lock();
2704         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2705                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2706         rcu_read_unlock();
2707         nf_ct_expect_put(exp);
2708         if (err <= 0)
2709                 goto free;
2710
2711         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2712         if (err < 0)
2713                 goto out;
2714
2715         return 0;
2716
2717 free:
2718         kfree_skb(skb2);
2719 out:
2720         /* this avoids a loop in nfnetlink. */
2721         return err == -EAGAIN ? -ENOBUFS : err;
2722 }
2723
2724 static int
2725 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2726                      const struct nlmsghdr *nlh,
2727                      const struct nlattr * const cda[])
2728 {
2729         struct net *net = sock_net(ctnl);
2730         struct nf_conntrack_expect *exp;
2731         struct nf_conntrack_tuple tuple;
2732         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2733         struct hlist_node *next;
2734         u_int8_t u3 = nfmsg->nfgen_family;
2735         unsigned int i;
2736         u16 zone;
2737         int err;
2738
2739         if (cda[CTA_EXPECT_TUPLE]) {
2740                 /* delete a single expect by tuple */
2741                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2742                 if (err < 0)
2743                         return err;
2744
2745                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2746                 if (err < 0)
2747                         return err;
2748
2749                 /* bump usage count to 2 */
2750                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2751                 if (!exp)
2752                         return -ENOENT;
2753
2754                 if (cda[CTA_EXPECT_ID]) {
2755                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2756                         if (ntohl(id) != (u32)(unsigned long)exp) {
2757                                 nf_ct_expect_put(exp);
2758                                 return -ENOENT;
2759                         }
2760                 }
2761
2762                 /* after list removal, usage count == 1 */
2763                 spin_lock_bh(&nf_conntrack_expect_lock);
2764                 if (del_timer(&exp->timeout)) {
2765                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2766                                                    nlmsg_report(nlh));
2767                         nf_ct_expect_put(exp);
2768                 }
2769                 spin_unlock_bh(&nf_conntrack_expect_lock);
2770                 /* have to put what we 'get' above.
2771                  * after this line usage count == 0 */
2772                 nf_ct_expect_put(exp);
2773         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2774                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2775                 struct nf_conn_help *m_help;
2776
2777                 /* delete all expectations for this helper */
2778                 spin_lock_bh(&nf_conntrack_expect_lock);
2779                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2780                         hlist_for_each_entry_safe(exp, next,
2781                                                   &net->ct.expect_hash[i],
2782                                                   hnode) {
2783                                 m_help = nfct_help(exp->master);
2784                                 if (!strcmp(m_help->helper->name, name) &&
2785                                     del_timer(&exp->timeout)) {
2786                                         nf_ct_unlink_expect_report(exp,
2787                                                         NETLINK_CB(skb).portid,
2788                                                         nlmsg_report(nlh));
2789                                         nf_ct_expect_put(exp);
2790                                 }
2791                         }
2792                 }
2793                 spin_unlock_bh(&nf_conntrack_expect_lock);
2794         } else {
2795                 /* This basically means we have to flush everything*/
2796                 spin_lock_bh(&nf_conntrack_expect_lock);
2797                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2798                         hlist_for_each_entry_safe(exp, next,
2799                                                   &net->ct.expect_hash[i],
2800                                                   hnode) {
2801                                 if (del_timer(&exp->timeout)) {
2802                                         nf_ct_unlink_expect_report(exp,
2803                                                         NETLINK_CB(skb).portid,
2804                                                         nlmsg_report(nlh));
2805                                         nf_ct_expect_put(exp);
2806                                 }
2807                         }
2808                 }
2809                 spin_unlock_bh(&nf_conntrack_expect_lock);
2810         }
2811
2812         return 0;
2813 }
2814 static int
2815 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2816                         const struct nlattr * const cda[])
2817 {
2818         if (cda[CTA_EXPECT_TIMEOUT]) {
2819                 if (!del_timer(&x->timeout))
2820                         return -ETIME;
2821
2822                 x->timeout.expires = jiffies +
2823                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2824                 add_timer(&x->timeout);
2825         }
2826         return 0;
2827 }
2828
2829 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2830         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2831         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2832 };
2833
2834 static int
2835 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2836                            struct nf_conntrack_expect *exp,
2837                            u_int8_t u3)
2838 {
2839 #ifdef CONFIG_NF_NAT_NEEDED
2840         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2841         struct nf_conntrack_tuple nat_tuple = {};
2842         int err;
2843
2844         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2845         if (err < 0)
2846                 return err;
2847
2848         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2849                 return -EINVAL;
2850
2851         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2852                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2853         if (err < 0)
2854                 return err;
2855
2856         exp->saved_addr = nat_tuple.src.u3;
2857         exp->saved_proto = nat_tuple.src.u;
2858         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2859
2860         return 0;
2861 #else
2862         return -EOPNOTSUPP;
2863 #endif
2864 }
2865
2866 static struct nf_conntrack_expect *
2867 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
2868                        struct nf_conntrack_helper *helper,
2869                        struct nf_conntrack_tuple *tuple,
2870                        struct nf_conntrack_tuple *mask)
2871 {
2872         u_int32_t class = 0;
2873         struct nf_conntrack_expect *exp;
2874         struct nf_conn_help *help;
2875         int err;
2876
2877         if (cda[CTA_EXPECT_CLASS] && helper) {
2878                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2879                 if (class > helper->expect_class_max)
2880                         return ERR_PTR(-EINVAL);
2881         }
2882         exp = nf_ct_expect_alloc(ct);
2883         if (!exp)
2884                 return ERR_PTR(-ENOMEM);
2885
2886         help = nfct_help(ct);
2887         if (!help) {
2888                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2889                         err = -EINVAL;
2890                         goto err_out;
2891                 }
2892                 exp->timeout.expires =
2893                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2894
2895                 exp->flags = NF_CT_EXPECT_USERSPACE;
2896                 if (cda[CTA_EXPECT_FLAGS]) {
2897                         exp->flags |=
2898                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2899                 }
2900         } else {
2901                 if (cda[CTA_EXPECT_FLAGS]) {
2902                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2903                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2904                 } else
2905                         exp->flags = 0;
2906         }
2907         if (cda[CTA_EXPECT_FN]) {
2908                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2909                 struct nf_ct_helper_expectfn *expfn;
2910
2911                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2912                 if (expfn == NULL) {
2913                         err = -EINVAL;
2914                         goto err_out;
2915                 }
2916                 exp->expectfn = expfn->expectfn;
2917         } else
2918                 exp->expectfn = NULL;
2919
2920         exp->class = class;
2921         exp->master = ct;
2922         exp->helper = helper;
2923         exp->tuple = *tuple;
2924         exp->mask.src.u3 = mask->src.u3;
2925         exp->mask.src.u.all = mask->src.u.all;
2926
2927         if (cda[CTA_EXPECT_NAT]) {
2928                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2929                                                  exp, nf_ct_l3num(ct));
2930                 if (err < 0)
2931                         goto err_out;
2932         }
2933         return exp;
2934 err_out:
2935         nf_ct_expect_put(exp);
2936         return ERR_PTR(err);
2937 }
2938
2939 static int
2940 ctnetlink_create_expect(struct net *net, u16 zone,
2941                         const struct nlattr * const cda[],
2942                         u_int8_t u3, u32 portid, int report)
2943 {
2944         struct nf_conntrack_tuple tuple, mask, master_tuple;
2945         struct nf_conntrack_tuple_hash *h = NULL;
2946         struct nf_conntrack_helper *helper = NULL;
2947         struct nf_conntrack_expect *exp;
2948         struct nf_conn *ct;
2949         int err;
2950
2951         /* caller guarantees that those three CTA_EXPECT_* exist */
2952         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2953         if (err < 0)
2954                 return err;
2955         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2956         if (err < 0)
2957                 return err;
2958         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2959         if (err < 0)
2960                 return err;
2961
2962         /* Look for master conntrack of this expectation */
2963         h = nf_conntrack_find_get(net, zone, &master_tuple);
2964         if (!h)
2965                 return -ENOENT;
2966         ct = nf_ct_tuplehash_to_ctrack(h);
2967
2968         if (cda[CTA_EXPECT_HELP_NAME]) {
2969                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2970
2971                 helper = __nf_conntrack_helper_find(helpname, u3,
2972                                                     nf_ct_protonum(ct));
2973                 if (helper == NULL) {
2974 #ifdef CONFIG_MODULES
2975                         if (request_module("nfct-helper-%s", helpname) < 0) {
2976                                 err = -EOPNOTSUPP;
2977                                 goto err_ct;
2978                         }
2979                         helper = __nf_conntrack_helper_find(helpname, u3,
2980                                                             nf_ct_protonum(ct));
2981                         if (helper) {
2982                                 err = -EAGAIN;
2983                                 goto err_ct;
2984                         }
2985 #endif
2986                         err = -EOPNOTSUPP;
2987                         goto err_ct;
2988                 }
2989         }
2990
2991         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
2992         if (IS_ERR(exp)) {
2993                 err = PTR_ERR(exp);
2994                 goto err_ct;
2995         }
2996
2997         err = nf_ct_expect_related_report(exp, portid, report);
2998         if (err < 0)
2999                 goto err_exp;
3000
3001         return 0;
3002 err_exp:
3003         nf_ct_expect_put(exp);
3004 err_ct:
3005         nf_ct_put(ct);
3006         return err;
3007 }
3008
3009 static int
3010 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
3011                      const struct nlmsghdr *nlh,
3012                      const struct nlattr * const cda[])
3013 {
3014         struct net *net = sock_net(ctnl);
3015         struct nf_conntrack_tuple tuple;
3016         struct nf_conntrack_expect *exp;
3017         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3018         u_int8_t u3 = nfmsg->nfgen_family;
3019         u16 zone;
3020         int err;
3021
3022         if (!cda[CTA_EXPECT_TUPLE]
3023             || !cda[CTA_EXPECT_MASK]
3024             || !cda[CTA_EXPECT_MASTER])
3025                 return -EINVAL;
3026
3027         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3028         if (err < 0)
3029                 return err;
3030
3031         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
3032         if (err < 0)
3033                 return err;
3034
3035         spin_lock_bh(&nf_conntrack_expect_lock);
3036         exp = __nf_ct_expect_find(net, zone, &tuple);
3037
3038         if (!exp) {
3039                 spin_unlock_bh(&nf_conntrack_expect_lock);
3040                 err = -ENOENT;
3041                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3042                         err = ctnetlink_create_expect(net, zone, cda,
3043                                                       u3,
3044                                                       NETLINK_CB(skb).portid,
3045                                                       nlmsg_report(nlh));
3046                 }
3047                 return err;
3048         }
3049
3050         err = -EEXIST;
3051         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3052                 err = ctnetlink_change_expect(exp, cda);
3053         spin_unlock_bh(&nf_conntrack_expect_lock);
3054
3055         return err;
3056 }
3057
3058 static int
3059 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3060                              const struct ip_conntrack_stat *st)
3061 {
3062         struct nlmsghdr *nlh;
3063         struct nfgenmsg *nfmsg;
3064         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3065
3066         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
3067         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3068         if (nlh == NULL)
3069                 goto nlmsg_failure;
3070
3071         nfmsg = nlmsg_data(nlh);
3072         nfmsg->nfgen_family = AF_UNSPEC;
3073         nfmsg->version      = NFNETLINK_V0;
3074         nfmsg->res_id       = htons(cpu);
3075
3076         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3077             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3078             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3079                 goto nla_put_failure;
3080
3081         nlmsg_end(skb, nlh);
3082         return skb->len;
3083
3084 nla_put_failure:
3085 nlmsg_failure:
3086         nlmsg_cancel(skb, nlh);
3087         return -1;
3088 }
3089
3090 static int
3091 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3092 {
3093         int cpu;
3094         struct net *net = sock_net(skb->sk);
3095
3096         if (cb->args[0] == nr_cpu_ids)
3097                 return 0;
3098
3099         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3100                 const struct ip_conntrack_stat *st;
3101
3102                 if (!cpu_possible(cpu))
3103                         continue;
3104
3105                 st = per_cpu_ptr(net->ct.stat, cpu);
3106                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3107                                                  cb->nlh->nlmsg_seq,
3108                                                  cpu, st) < 0)
3109                         break;
3110         }
3111         cb->args[0] = cpu;
3112
3113         return skb->len;
3114 }
3115
3116 static int
3117 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
3118                        const struct nlmsghdr *nlh,
3119                        const struct nlattr * const cda[])
3120 {
3121         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3122                 struct netlink_dump_control c = {
3123                         .dump = ctnetlink_exp_stat_cpu_dump,
3124                 };
3125                 return netlink_dump_start(ctnl, skb, nlh, &c);
3126         }
3127
3128         return 0;
3129 }
3130
3131 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3132 static struct nf_ct_event_notifier ctnl_notifier = {
3133         .fcn = ctnetlink_conntrack_event,
3134 };
3135
3136 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3137         .fcn = ctnetlink_expect_event,
3138 };
3139 #endif
3140
3141 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3142         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3143                                             .attr_count = CTA_MAX,
3144                                             .policy = ct_nla_policy },
3145         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3146                                             .attr_count = CTA_MAX,
3147                                             .policy = ct_nla_policy },
3148         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3149                                             .attr_count = CTA_MAX,
3150                                             .policy = ct_nla_policy },
3151         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3152                                             .attr_count = CTA_MAX,
3153                                             .policy = ct_nla_policy },
3154         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3155         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3156         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3157         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3158 };
3159
3160 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3161         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3162                                             .attr_count = CTA_EXPECT_MAX,
3163                                             .policy = exp_nla_policy },
3164         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3165                                             .attr_count = CTA_EXPECT_MAX,
3166                                             .policy = exp_nla_policy },
3167         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3168                                             .attr_count = CTA_EXPECT_MAX,
3169                                             .policy = exp_nla_policy },
3170         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3171 };
3172
3173 static const struct nfnetlink_subsystem ctnl_subsys = {
3174         .name                           = "conntrack",
3175         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3176         .cb_count                       = IPCTNL_MSG_MAX,
3177         .cb                             = ctnl_cb,
3178 };
3179
3180 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3181         .name                           = "conntrack_expect",
3182         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3183         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3184         .cb                             = ctnl_exp_cb,
3185 };
3186
3187 MODULE_ALIAS("ip_conntrack_netlink");
3188 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3189 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3190
3191 static int __net_init ctnetlink_net_init(struct net *net)
3192 {
3193 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3194         int ret;
3195
3196         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3197         if (ret < 0) {
3198                 pr_err("ctnetlink_init: cannot register notifier.\n");
3199                 goto err_out;
3200         }
3201
3202         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3203         if (ret < 0) {
3204                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3205                 goto err_unreg_notifier;
3206         }
3207 #endif
3208         return 0;
3209
3210 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3211 err_unreg_notifier:
3212         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3213 err_out:
3214         return ret;
3215 #endif
3216 }
3217
3218 static void ctnetlink_net_exit(struct net *net)
3219 {
3220 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3221         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3222         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3223 #endif
3224 }
3225
3226 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3227 {
3228         struct net *net;
3229
3230         list_for_each_entry(net, net_exit_list, exit_list)
3231                 ctnetlink_net_exit(net);
3232 }
3233
3234 static struct pernet_operations ctnetlink_net_ops = {
3235         .init           = ctnetlink_net_init,
3236         .exit_batch     = ctnetlink_net_exit_batch,
3237 };
3238
3239 static int __init ctnetlink_init(void)
3240 {
3241         int ret;
3242
3243         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
3244         ret = nfnetlink_subsys_register(&ctnl_subsys);
3245         if (ret < 0) {
3246                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3247                 goto err_out;
3248         }
3249
3250         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3251         if (ret < 0) {
3252                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3253                 goto err_unreg_subsys;
3254         }
3255
3256         ret = register_pernet_subsys(&ctnetlink_net_ops);
3257         if (ret < 0) {
3258                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3259                 goto err_unreg_exp_subsys;
3260         }
3261 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3262         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3263         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
3264 #endif
3265         return 0;
3266
3267 err_unreg_exp_subsys:
3268         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3269 err_unreg_subsys:
3270         nfnetlink_subsys_unregister(&ctnl_subsys);
3271 err_out:
3272         return ret;
3273 }
3274
3275 static void __exit ctnetlink_exit(void)
3276 {
3277         pr_info("ctnetlink: unregistering from nfnetlink.\n");
3278
3279         unregister_pernet_subsys(&ctnetlink_net_ops);
3280         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3281         nfnetlink_subsys_unregister(&ctnl_subsys);
3282 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
3283         RCU_INIT_POINTER(nfq_ct_hook, NULL);
3284 #endif
3285 }
3286
3287 module_init(ctnetlink_init);
3288 module_exit(ctnetlink_exit);