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 #include <net/netfilter/nf_conntrack_synproxy.h>
  49 #ifdef CONFIG_NF_NAT_NEEDED
  50 #include <net/netfilter/nf_nat_core.h>
  51 #include <net/netfilter/nf_nat_l4proto.h>
  52 #include <net/netfilter/nf_nat_helper.h>
  53 #endif
  54
  55 #include <linux/netfilter/nfnetlink.h>
  56 #include <linux/netfilter/nfnetlink_conntrack.h>
  57
  58 MODULE_LICENSE("GPL");
  59
  60 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  61                                 const struct nf_conntrack_tuple *tuple,
  62                                 const struct nf_conntrack_l4proto *l4proto)
  63 {
  64         int ret = 0;
  65         struct nlattr *nest_parms;
  66
  67         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  68         if (!nest_parms)
  69                 goto nla_put_failure;
  70         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  71                 goto nla_put_failure;
  72
  73         if (likely(l4proto->tuple_to_nlattr))
  74                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  75
  76         nla_nest_end(skb, nest_parms);
  77
  78         return ret;
  79
  80 nla_put_failure:
  81         return -1;
  82 }
  83
  84 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  85                                     const struct nf_conntrack_tuple *tuple,
  86                                     const struct nf_conntrack_l3proto *l3proto)
  87 {
  88         int ret = 0;
  89         struct nlattr *nest_parms;
  90
  91         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  92         if (!nest_parms)
  93                 goto nla_put_failure;
  94
  95         if (likely(l3proto->tuple_to_nlattr))
  96                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  97
  98         nla_nest_end(skb, nest_parms);
  99
 100         return ret;
 101
 102 nla_put_failure:
 103         return -1;
 104 }
 105
 106 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 107                                  const struct nf_conntrack_tuple *tuple)
 108 {
 109         const struct nf_conntrack_l3proto *l3proto;
 110         const struct nf_conntrack_l4proto *l4proto;
 111         int ret;
 112
 113         rcu_read_lock();
 114         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 115         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 116
 117         if (ret >= 0) {
 118                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 119                                                tuple->dst.protonum);
 120                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 121         }
 122         rcu_read_unlock();
 123         return ret;
 124 }
 125
 126 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 127                                   const struct nf_conntrack_zone *zone, int dir)
 128 {
 129         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 130                 return 0;
 131         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 132                 goto nla_put_failure;
 133         return 0;
 134
 135 nla_put_failure:
 136         return -1;
 137 }
 138
 139 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 140 {
 141         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 142                 goto nla_put_failure;
 143         return 0;
 144
 145 nla_put_failure:
 146         return -1;
 147 }
 148
 149 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 150 {
 151         long timeout = nf_ct_expires(ct) / HZ;
 152
 153         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 154                 goto nla_put_failure;
 155         return 0;
 156
 157 nla_put_failure:
 158         return -1;
 159 }
 160
 161 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 162 {
 163         const struct nf_conntrack_l4proto *l4proto;
 164         struct nlattr *nest_proto;
 165         int ret;
 166
 167         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 168         if (!l4proto->to_nlattr)
 169                 return 0;
 170
 171         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 172         if (!nest_proto)
 173                 goto nla_put_failure;
 174
 175         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 176
 177         nla_nest_end(skb, nest_proto);
 178
 179         return ret;
 180
 181 nla_put_failure:
 182         return -1;
 183 }
 184
 185 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 186                                    const struct nf_conn *ct)
 187 {
 188         struct nlattr *nest_helper;
 189         const struct nf_conn_help *help = nfct_help(ct);
 190         struct nf_conntrack_helper *helper;
 191
 192         if (!help)
 193                 return 0;
 194
 195         helper = rcu_dereference(help->helper);
 196         if (!helper)
 197                 goto out;
 198
 199         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 200         if (!nest_helper)
 201                 goto nla_put_failure;
 202         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 203                 goto nla_put_failure;
 204
 205         if (helper->to_nlattr)
 206                 helper->to_nlattr(skb, ct);
 207
 208         nla_nest_end(skb, nest_helper);
 209 out:
 210         return 0;
 211
 212 nla_put_failure:
 213         return -1;
 214 }
 215
 216 static int
 217 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 218               enum ip_conntrack_dir dir, int type)
 219 {
 220         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 221         struct nf_conn_counter *counter = acct->counter;
 222         struct nlattr *nest_count;
 223         u64 pkts, bytes;
 224
 225         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 226                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 227                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 228         } else {
 229                 pkts = atomic64_read(&counter[dir].packets);
 230                 bytes = atomic64_read(&counter[dir].bytes);
 231         }
 232
 233         nest_count = nla_nest_start(skb, attr | NLA_F_NESTED);
 234         if (!nest_count)
 235                 goto nla_put_failure;
 236
 237         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 238                          CTA_COUNTERS_PAD) ||
 239             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 240                          CTA_COUNTERS_PAD))
 241                 goto nla_put_failure;
 242
 243         nla_nest_end(skb, nest_count);
 244
 245         return 0;
 246
 247 nla_put_failure:
 248         return -1;
 249 }
 250
 251 static int
 252 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 253 {
 254         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 255
 256         if (!acct)
 257                 return 0;
 258
 259         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 260                 return -1;
 261         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 262                 return -1;
 263
 264         return 0;
 265 }
 266
 267 static int
 268 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 269 {
 270         struct nlattr *nest_count;
 271         const struct nf_conn_tstamp *tstamp;
 272
 273         tstamp = nf_conn_tstamp_find(ct);
 274         if (!tstamp)
 275                 return 0;
 276
 277         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 278         if (!nest_count)
 279                 goto nla_put_failure;
 280
 281         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 282                          CTA_TIMESTAMP_PAD) ||
 283             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 284                                                cpu_to_be64(tstamp->stop),
 285                                                CTA_TIMESTAMP_PAD)))
 286                 goto nla_put_failure;
 287         nla_nest_end(skb, nest_count);
 288
 289         return 0;
 290
 291 nla_put_failure:
 292         return -1;
 293 }
 294
 295 #ifdef CONFIG_NF_CONNTRACK_MARK
 296 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 297 {
 298         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 299                 goto nla_put_failure;
 300         return 0;
 301
 302 nla_put_failure:
 303         return -1;
 304 }
 305 #else
 306 #define ctnetlink_dump_mark(a, b) (0)
 307 #endif
 308
 309 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 310 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 311 {
 312         struct nlattr *nest_secctx;
 313         int len, ret;
 314         char *secctx;
 315
 316         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 317         if (ret)
 318                 return 0;
 319
 320         ret = -1;
 321         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 322         if (!nest_secctx)
 323                 goto nla_put_failure;
 324
 325         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 326                 goto nla_put_failure;
 327         nla_nest_end(skb, nest_secctx);
 328
 329         ret = 0;
 330 nla_put_failure:
 331         security_release_secctx(secctx, len);
 332         return ret;
 333 }
 334 #else
 335 #define ctnetlink_dump_secctx(a, b) (0)
 336 #endif
 337
 338 #ifdef CONFIG_NF_CONNTRACK_LABELS
 339 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 340 {
 341         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 342
 343         if (!labels)
 344                 return 0;
 345         return nla_total_size(sizeof(labels->bits));
 346 }
 347
 348 static int
 349 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 350 {
 351         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 352         unsigned int i;
 353
 354         if (!labels)
 355                 return 0;
 356
 357         i = 0;
 358         do {
 359                 if (labels->bits[i] != 0)
 360                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 361                                        labels->bits);
 362                 i++;
 363         } while (i < ARRAY_SIZE(labels->bits));
 364
 365         return 0;
 366 }
 367 #else
 368 #define ctnetlink_dump_labels(a, b) (0)
 369 #define ctnetlink_label_size(a) (0)
 370 #endif
 371
 372 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 373
 374 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 375 {
 376         struct nlattr *nest_parms;
 377
 378         if (!(ct->status & IPS_EXPECTED))
 379                 return 0;
 380
 381         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 382         if (!nest_parms)
 383                 goto nla_put_failure;
 384         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 385                 goto nla_put_failure;
 386         nla_nest_end(skb, nest_parms);
 387
 388         return 0;
 389
 390 nla_put_failure:
 391         return -1;
 392 }
 393
 394 static int
 395 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 396 {
 397         struct nlattr *nest_parms;
 398
 399         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 400         if (!nest_parms)
 401                 goto nla_put_failure;
 402
 403         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 404                          htonl(seq->correction_pos)) ||
 405             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 406                          htonl(seq->offset_before)) ||
 407             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 408                          htonl(seq->offset_after)))
 409                 goto nla_put_failure;
 410
 411         nla_nest_end(skb, nest_parms);
 412
 413         return 0;
 414
 415 nla_put_failure:
 416         return -1;
 417 }
 418
 419 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 420 {
 421         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 422         struct nf_ct_seqadj *seq;
 423
 424         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 425                 return 0;
 426
 427         spin_lock_bh(&ct->lock);
 428         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 429         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 430                 goto err;
 431
 432         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 433         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 434                 goto err;
 435
 436         spin_unlock_bh(&ct->lock);
 437         return 0;
 438 err:
 439         spin_unlock_bh(&ct->lock);
 440         return -1;
 441 }
 442
 443 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 444 {
 445         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 446                 goto nla_put_failure;
 447         return 0;
 448
 449 nla_put_failure:
 450         return -1;
 451 }
 452
 453 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 454 {
 455         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 456                 goto nla_put_failure;
 457         return 0;
 458
 459 nla_put_failure:
 460         return -1;
 461 }
 462
 463 static int
 464 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 465                     struct nf_conn *ct)
 466 {
 467         const struct nf_conntrack_zone *zone;
 468         struct nlmsghdr *nlh;
 469         struct nfgenmsg *nfmsg;
 470         struct nlattr *nest_parms;
 471         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 472
 473         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 474         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 475         if (nlh == NULL)
 476                 goto nlmsg_failure;
 477
 478         nfmsg = nlmsg_data(nlh);
 479         nfmsg->nfgen_family = nf_ct_l3num(ct);
 480         nfmsg->version      = NFNETLINK_V0;
 481         nfmsg->res_id       = 0;
 482
 483         zone = nf_ct_zone(ct);
 484
 485         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 486         if (!nest_parms)
 487                 goto nla_put_failure;
 488         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 489                 goto nla_put_failure;
 490         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 491                                    NF_CT_ZONE_DIR_ORIG) < 0)
 492                 goto nla_put_failure;
 493         nla_nest_end(skb, nest_parms);
 494
 495         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 496         if (!nest_parms)
 497                 goto nla_put_failure;
 498         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 499                 goto nla_put_failure;
 500         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 501                                    NF_CT_ZONE_DIR_REPL) < 0)
 502                 goto nla_put_failure;
 503         nla_nest_end(skb, nest_parms);
 504
 505         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 506                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 507                 goto nla_put_failure;
 508
 509         if (ctnetlink_dump_status(skb, ct) < 0 ||
 510             ctnetlink_dump_timeout(skb, ct) < 0 ||
 511             ctnetlink_dump_acct(skb, ct, type) < 0 ||
 512             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 513             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 514             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 515             ctnetlink_dump_mark(skb, ct) < 0 ||
 516             ctnetlink_dump_secctx(skb, ct) < 0 ||
 517             ctnetlink_dump_labels(skb, ct) < 0 ||
 518             ctnetlink_dump_id(skb, ct) < 0 ||
 519             ctnetlink_dump_use(skb, ct) < 0 ||
 520             ctnetlink_dump_master(skb, ct) < 0 ||
 521             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 522                 goto nla_put_failure;
 523
 524         nlmsg_end(skb, nlh);
 525         return skb->len;
 526
 527 nlmsg_failure:
 528 nla_put_failure:
 529         nlmsg_cancel(skb, nlh);
 530         return -1;
 531 }
 532
 533 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 534 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 535 {
 536         const struct nf_conntrack_l3proto *l3proto;
 537         const struct nf_conntrack_l4proto *l4proto;
 538         size_t len, len4 = 0;
 539
 540         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 541         len = l3proto->nla_size;
 542         len *= 3u; /* ORIG, REPLY, MASTER */
 543
 544         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 545         len += l4proto->nlattr_size;
 546         if (l4proto->nlattr_tuple_size) {
 547                 len4 = l4proto->nlattr_tuple_size();
 548                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 549         }
 550
 551         return len + len4;
 552 }
 553 #endif
 554
 555 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 556 {
 557         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 558                 return 0;
 559         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 560                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 561                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 562                ;
 563 }
 564
 565 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 566 {
 567 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 568         int len, ret;
 569
 570         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 571         if (ret)
 572                 return 0;
 573
 574         return nla_total_size(0) /* CTA_SECCTX */
 575                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 576 #else
 577         return 0;
 578 #endif
 579 }
 580
 581 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 582 {
 583 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 584         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 585                 return 0;
 586         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 587 #else
 588         return 0;
 589 #endif
 590 }
 591
 592 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 593 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 594 {
 595         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 596                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 597                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 598                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 599                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 600                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 601                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 602                + ctnetlink_acct_size(ct)
 603                + ctnetlink_timestamp_size(ct)
 604                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 605                + nla_total_size(0) /* CTA_PROTOINFO */
 606                + nla_total_size(0) /* CTA_HELP */
 607                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 608                + ctnetlink_secctx_size(ct)
 609 #ifdef CONFIG_NF_NAT_NEEDED
 610                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 611                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 612 #endif
 613 #ifdef CONFIG_NF_CONNTRACK_MARK
 614                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 615 #endif
 616 #ifdef CONFIG_NF_CONNTRACK_ZONES
 617                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 618 #endif
 619                + ctnetlink_proto_size(ct)
 620                + ctnetlink_label_size(ct)
 621                ;
 622 }
 623
 624 static int
 625 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 626 {
 627         const struct nf_conntrack_zone *zone;
 628         struct net *net;
 629         struct nlmsghdr *nlh;
 630         struct nfgenmsg *nfmsg;
 631         struct nlattr *nest_parms;
 632         struct nf_conn *ct = item->ct;
 633         struct sk_buff *skb;
 634         unsigned int type;
 635         unsigned int flags = 0, group;
 636         int err;
 637
 638         if (events & (1 << IPCT_DESTROY)) {
 639                 type = IPCTNL_MSG_CT_DELETE;
 640                 group = NFNLGRP_CONNTRACK_DESTROY;
 641         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 642                 type = IPCTNL_MSG_CT_NEW;
 643                 flags = NLM_F_CREATE|NLM_F_EXCL;
 644                 group = NFNLGRP_CONNTRACK_NEW;
 645         } else if (events) {
 646                 type = IPCTNL_MSG_CT_NEW;
 647                 group = NFNLGRP_CONNTRACK_UPDATE;
 648         } else
 649                 return 0;
 650
 651         net = nf_ct_net(ct);
 652         if (!item->report && !nfnetlink_has_listeners(net, group))
 653                 return 0;
 654
 655         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 656         if (skb == NULL)
 657                 goto errout;
 658
 659         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 660         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 661         if (nlh == NULL)
 662                 goto nlmsg_failure;
 663
 664         nfmsg = nlmsg_data(nlh);
 665         nfmsg->nfgen_family = nf_ct_l3num(ct);
 666         nfmsg->version  = NFNETLINK_V0;
 667         nfmsg->res_id   = 0;
 668
 669         zone = nf_ct_zone(ct);
 670
 671         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 672         if (!nest_parms)
 673                 goto nla_put_failure;
 674         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 675                 goto nla_put_failure;
 676         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 677                                    NF_CT_ZONE_DIR_ORIG) < 0)
 678                 goto nla_put_failure;
 679         nla_nest_end(skb, nest_parms);
 680
 681         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 682         if (!nest_parms)
 683                 goto nla_put_failure;
 684         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 685                 goto nla_put_failure;
 686         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 687                                    NF_CT_ZONE_DIR_REPL) < 0)
 688                 goto nla_put_failure;
 689         nla_nest_end(skb, nest_parms);
 690
 691         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 692                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 693                 goto nla_put_failure;
 694
 695         if (ctnetlink_dump_id(skb, ct) < 0)
 696                 goto nla_put_failure;
 697
 698         if (ctnetlink_dump_status(skb, ct) < 0)
 699                 goto nla_put_failure;
 700
 701         if (events & (1 << IPCT_DESTROY)) {
 702                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 703                     ctnetlink_dump_timestamp(skb, ct) < 0)
 704                         goto nla_put_failure;
 705         } else {
 706                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 707                         goto nla_put_failure;
 708
 709                 if (events & (1 << IPCT_PROTOINFO)
 710                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 711                         goto nla_put_failure;
 712
 713                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 714                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 715                         goto nla_put_failure;
 716
 717 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 718                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 719                     && ctnetlink_dump_secctx(skb, ct) < 0)
 720                         goto nla_put_failure;
 721 #endif
 722                 if (events & (1 << IPCT_LABEL) &&
 723                      ctnetlink_dump_labels(skb, ct) < 0)
 724                         goto nla_put_failure;
 725
 726                 if (events & (1 << IPCT_RELATED) &&
 727                     ctnetlink_dump_master(skb, ct) < 0)
 728                         goto nla_put_failure;
 729
 730                 if (events & (1 << IPCT_SEQADJ) &&
 731                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 732                         goto nla_put_failure;
 733         }
 734
 735 #ifdef CONFIG_NF_CONNTRACK_MARK
 736         if ((events & (1 << IPCT_MARK) || ct->mark)
 737             && ctnetlink_dump_mark(skb, ct) < 0)
 738                 goto nla_put_failure;
 739 #endif
 740         nlmsg_end(skb, nlh);
 741         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 742                              GFP_ATOMIC);
 743         if (err == -ENOBUFS || err == -EAGAIN)
 744                 return -ENOBUFS;
 745
 746         return 0;
 747
 748 nla_put_failure:
 749         nlmsg_cancel(skb, nlh);
 750 nlmsg_failure:
 751         kfree_skb(skb);
 752 errout:
 753         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 754                 return -ENOBUFS;
 755
 756         return 0;
 757 }
 758 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 759
 760 static int ctnetlink_done(struct netlink_callback *cb)
 761 {
 762         if (cb->args[1])
 763                 nf_ct_put((struct nf_conn *)cb->args[1]);
 764         kfree(cb->data);
 765         return 0;
 766 }
 767
 768 struct ctnetlink_filter {
 769         struct {
 770                 u_int32_t val;
 771                 u_int32_t mask;
 772         } mark;
 773 };
 774
 775 static struct ctnetlink_filter *
 776 ctnetlink_alloc_filter(const struct nlattr * const cda[])
 777 {
 778 #ifdef CONFIG_NF_CONNTRACK_MARK
 779         struct ctnetlink_filter *filter;
 780
 781         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 782         if (filter == NULL)
 783                 return ERR_PTR(-ENOMEM);
 784
 785         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 786         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 787
 788         return filter;
 789 #else
 790         return ERR_PTR(-EOPNOTSUPP);
 791 #endif
 792 }
 793
 794 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 795 {
 796         struct ctnetlink_filter *filter = data;
 797
 798         if (filter == NULL)
 799                 return 1;
 800
 801 #ifdef CONFIG_NF_CONNTRACK_MARK
 802         if ((ct->mark & filter->mark.mask) == filter->mark.val)
 803                 return 1;
 804 #endif
 805
 806         return 0;
 807 }
 808
 809 static int
 810 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 811 {
 812         struct net *net = sock_net(skb->sk);
 813         struct nf_conn *ct, *last;
 814         struct nf_conntrack_tuple_hash *h;
 815         struct hlist_nulls_node *n;
 816         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 817         u_int8_t l3proto = nfmsg->nfgen_family;
 818         struct nf_conn *nf_ct_evict[8];
 819         int res, i;
 820         spinlock_t *lockp;
 821
 822         last = (struct nf_conn *)cb->args[1];
 823         i = 0;
 824
 825         local_bh_disable();
 826         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 827 restart:
 828                 while (i) {
 829                         i--;
 830                         if (nf_ct_should_gc(nf_ct_evict[i]))
 831                                 nf_ct_kill(nf_ct_evict[i]);
 832                         nf_ct_put(nf_ct_evict[i]);
 833                 }
 834
 835                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 836                 nf_conntrack_lock(lockp);
 837                 if (cb->args[0] >= nf_conntrack_htable_size) {
 838                         spin_unlock(lockp);
 839                         goto out;
 840                 }
 841                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 842                                            hnnode) {
 843                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 844                                 continue;
 845                         ct = nf_ct_tuplehash_to_ctrack(h);
 846                         if (nf_ct_is_expired(ct)) {
 847                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
 848                                     atomic_inc_not_zero(&ct->ct_general.use))
 849                                         nf_ct_evict[i++] = ct;
 850                                 continue;
 851                         }
 852
 853                         if (!net_eq(net, nf_ct_net(ct)))
 854                                 continue;
 855
 856                         /* Dump entries of a given L3 protocol number.
 857                          * If it is not specified, ie. l3proto == 0,
 858                          * then dump everything. */
 859                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 860                                 continue;
 861                         if (cb->args[1]) {
 862                                 if (ct != last)
 863                                         continue;
 864                                 cb->args[1] = 0;
 865                         }
 866                         if (!ctnetlink_filter_match(ct, cb->data))
 867                                 continue;
 868
 869                         rcu_read_lock();
 870                         res =
 871                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 872                                             cb->nlh->nlmsg_seq,
 873                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 874                                             ct);
 875                         rcu_read_unlock();
 876                         if (res < 0) {
 877                                 nf_conntrack_get(&ct->ct_general);
 878                                 cb->args[1] = (unsigned long)ct;
 879                                 spin_unlock(lockp);
 880                                 goto out;
 881                         }
 882                 }
 883                 spin_unlock(lockp);
 884                 if (cb->args[1]) {
 885                         cb->args[1] = 0;
 886                         goto restart;
 887                 }
 888         }
 889 out:
 890         local_bh_enable();
 891         if (last) {
 892                 /* nf ct hash resize happened, now clear the leftover. */
 893                 if ((struct nf_conn *)cb->args[1] == last)
 894                         cb->args[1] = 0;
 895
 896                 nf_ct_put(last);
 897         }
 898
 899         while (i) {
 900                 i--;
 901                 if (nf_ct_should_gc(nf_ct_evict[i]))
 902                         nf_ct_kill(nf_ct_evict[i]);
 903                 nf_ct_put(nf_ct_evict[i]);
 904         }
 905
 906         return skb->len;
 907 }
 908
 909 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
 910                                     struct nf_conntrack_tuple *tuple)
 911 {
 912         struct nlattr *tb[CTA_IP_MAX+1];
 913         struct nf_conntrack_l3proto *l3proto;
 914         int ret = 0;
 915
 916         ret = nla_parse_nested(tb, CTA_IP_MAX, attr, NULL, NULL);
 917         if (ret < 0)
 918                 return ret;
 919
 920         rcu_read_lock();
 921         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 922
 923         if (likely(l3proto->nlattr_to_tuple)) {
 924                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 925                                           l3proto->nla_policy, NULL);
 926                 if (ret == 0)
 927                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 928         }
 929
 930         rcu_read_unlock();
 931
 932         return ret;
 933 }
 934
 935 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 936         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 937 };
 938
 939 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
 940                                        struct nf_conntrack_tuple *tuple)
 941 {
 942         const struct nf_conntrack_l4proto *l4proto;
 943         struct nlattr *tb[CTA_PROTO_MAX+1];
 944         int ret = 0;
 945
 946         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy,
 947                                NULL);
 948         if (ret < 0)
 949                 return ret;
 950
 951         if (!tb[CTA_PROTO_NUM])
 952                 return -EINVAL;
 953         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 954
 955         rcu_read_lock();
 956         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 957
 958         if (likely(l4proto->nlattr_to_tuple)) {
 959                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 960                                           l4proto->nla_policy, NULL);
 961                 if (ret == 0)
 962                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 963         }
 964
 965         rcu_read_unlock();
 966
 967         return ret;
 968 }
 969
 970 static int
 971 ctnetlink_parse_zone(const struct nlattr *attr,
 972                      struct nf_conntrack_zone *zone)
 973 {
 974         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
 975                         NF_CT_DEFAULT_ZONE_DIR, 0);
 976 #ifdef CONFIG_NF_CONNTRACK_ZONES
 977         if (attr)
 978                 zone->id = ntohs(nla_get_be16(attr));
 979 #else
 980         if (attr)
 981                 return -EOPNOTSUPP;
 982 #endif
 983         return 0;
 984 }
 985
 986 static int
 987 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
 988                            struct nf_conntrack_zone *zone)
 989 {
 990         int ret;
 991
 992         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
 993                 return -EINVAL;
 994
 995         ret = ctnetlink_parse_zone(attr, zone);
 996         if (ret < 0)
 997                 return ret;
 998
 999         if (type == CTA_TUPLE_REPLY)
1000                 zone->dir = NF_CT_ZONE_DIR_REPL;
1001         else
1002                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1003
1004         return 0;
1005 }
1006
1007 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1008         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1009         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1010         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1011 };
1012
1013 static int
1014 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1015                       struct nf_conntrack_tuple *tuple, u32 type,
1016                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1017 {
1018         struct nlattr *tb[CTA_TUPLE_MAX+1];
1019         int err;
1020
1021         memset(tuple, 0, sizeof(*tuple));
1022
1023         err = nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy,
1024                                NULL);
1025         if (err < 0)
1026                 return err;
1027
1028         if (!tb[CTA_TUPLE_IP])
1029                 return -EINVAL;
1030
1031         tuple->src.l3num = l3num;
1032
1033         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1034         if (err < 0)
1035                 return err;
1036
1037         if (!tb[CTA_TUPLE_PROTO])
1038                 return -EINVAL;
1039
1040         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1041         if (err < 0)
1042                 return err;
1043
1044         if (tb[CTA_TUPLE_ZONE]) {
1045                 if (!zone)
1046                         return -EINVAL;
1047
1048                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1049                                                  type, zone);
1050                 if (err < 0)
1051                         return err;
1052         }
1053
1054         /* orig and expect tuples get DIR_ORIGINAL */
1055         if (type == CTA_TUPLE_REPLY)
1056                 tuple->dst.dir = IP_CT_DIR_REPLY;
1057         else
1058                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1059
1060         return 0;
1061 }
1062
1063 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1064         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1065                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1066 };
1067
1068 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1069                                 struct nlattr **helpinfo)
1070 {
1071         int err;
1072         struct nlattr *tb[CTA_HELP_MAX+1];
1073
1074         err = nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy, NULL);
1075         if (err < 0)
1076                 return err;
1077
1078         if (!tb[CTA_HELP_NAME])
1079                 return -EINVAL;
1080
1081         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1082
1083         if (tb[CTA_HELP_INFO])
1084                 *helpinfo = tb[CTA_HELP_INFO];
1085
1086         return 0;
1087 }
1088
1089 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1090         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1091         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1092         [CTA_STATUS]            = { .type = NLA_U32 },
1093         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1094         [CTA_HELP]              = { .type = NLA_NESTED },
1095         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1096         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1097         [CTA_MARK]              = { .type = NLA_U32 },
1098         [CTA_ID]                = { .type = NLA_U32 },
1099         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1100         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1101         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1102         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1103         [CTA_ZONE]              = { .type = NLA_U16 },
1104         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1105         [CTA_LABELS]            = { .type = NLA_BINARY,
1106                                     .len = NF_CT_LABELS_MAX_SIZE },
1107         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1108                                     .len = NF_CT_LABELS_MAX_SIZE },
1109 };
1110
1111 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1112 {
1113         if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1114                 return 0;
1115
1116         return ctnetlink_filter_match(ct, data);
1117 }
1118
1119 static int ctnetlink_flush_conntrack(struct net *net,
1120                                      const struct nlattr * const cda[],
1121                                      u32 portid, int report)
1122 {
1123         struct ctnetlink_filter *filter = NULL;
1124
1125         if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1126                 filter = ctnetlink_alloc_filter(cda);
1127                 if (IS_ERR(filter))
1128                         return PTR_ERR(filter);
1129         }
1130
1131         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1132                                   portid, report);
1133         kfree(filter);
1134
1135         return 0;
1136 }
1137
1138 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1139                                    struct sk_buff *skb,
1140                                    const struct nlmsghdr *nlh,
1141                                    const struct nlattr * const cda[],
1142                                    struct netlink_ext_ack *extack)
1143 {
1144         struct nf_conntrack_tuple_hash *h;
1145         struct nf_conntrack_tuple tuple;
1146         struct nf_conn *ct;
1147         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1148         u_int8_t u3 = nfmsg->nfgen_family;
1149         struct nf_conntrack_zone zone;
1150         int err;
1151
1152         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1153         if (err < 0)
1154                 return err;
1155
1156         if (cda[CTA_TUPLE_ORIG])
1157                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1158                                             u3, &zone);
1159         else if (cda[CTA_TUPLE_REPLY])
1160                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1161                                             u3, &zone);
1162         else {
1163                 return ctnetlink_flush_conntrack(net, cda,
1164                                                  NETLINK_CB(skb).portid,
1165                                                  nlmsg_report(nlh));
1166         }
1167
1168         if (err < 0)
1169                 return err;
1170
1171         h = nf_conntrack_find_get(net, &zone, &tuple);
1172         if (!h)
1173                 return -ENOENT;
1174
1175         ct = nf_ct_tuplehash_to_ctrack(h);
1176
1177         if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1178                 nf_ct_put(ct);
1179                 return -EBUSY;
1180         }
1181
1182         if (cda[CTA_ID]) {
1183                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
1184                 if (id != (u32)(unsigned long)ct) {
1185                         nf_ct_put(ct);
1186                         return -ENOENT;
1187                 }
1188         }
1189
1190         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1191         nf_ct_put(ct);
1192
1193         return 0;
1194 }
1195
1196 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1197                                    struct sk_buff *skb,
1198                                    const struct nlmsghdr *nlh,
1199                                    const struct nlattr * const cda[],
1200                                    struct netlink_ext_ack *extack)
1201 {
1202         struct nf_conntrack_tuple_hash *h;
1203         struct nf_conntrack_tuple tuple;
1204         struct nf_conn *ct;
1205         struct sk_buff *skb2 = NULL;
1206         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1207         u_int8_t u3 = nfmsg->nfgen_family;
1208         struct nf_conntrack_zone zone;
1209         int err;
1210
1211         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1212                 struct netlink_dump_control c = {
1213                         .dump = ctnetlink_dump_table,
1214                         .done = ctnetlink_done,
1215                 };
1216
1217                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1218                         struct ctnetlink_filter *filter;
1219
1220                         filter = ctnetlink_alloc_filter(cda);
1221                         if (IS_ERR(filter))
1222                                 return PTR_ERR(filter);
1223
1224                         c.data = filter;
1225                 }
1226                 return netlink_dump_start(ctnl, skb, nlh, &c);
1227         }
1228
1229         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1230         if (err < 0)
1231                 return err;
1232
1233         if (cda[CTA_TUPLE_ORIG])
1234                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1235                                             u3, &zone);
1236         else if (cda[CTA_TUPLE_REPLY])
1237                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1238                                             u3, &zone);
1239         else
1240                 return -EINVAL;
1241
1242         if (err < 0)
1243                 return err;
1244
1245         h = nf_conntrack_find_get(net, &zone, &tuple);
1246         if (!h)
1247                 return -ENOENT;
1248
1249         ct = nf_ct_tuplehash_to_ctrack(h);
1250
1251         err = -ENOMEM;
1252         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1253         if (skb2 == NULL) {
1254                 nf_ct_put(ct);
1255                 return -ENOMEM;
1256         }
1257
1258         rcu_read_lock();
1259         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1260                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1261         rcu_read_unlock();
1262         nf_ct_put(ct);
1263         if (err <= 0)
1264                 goto free;
1265
1266         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1267         if (err < 0)
1268                 goto out;
1269
1270         return 0;
1271
1272 free:
1273         kfree_skb(skb2);
1274 out:
1275         /* this avoids a loop in nfnetlink. */
1276         return err == -EAGAIN ? -ENOBUFS : err;
1277 }
1278
1279 static int ctnetlink_done_list(struct netlink_callback *cb)
1280 {
1281         if (cb->args[1])
1282                 nf_ct_put((struct nf_conn *)cb->args[1]);
1283         return 0;
1284 }
1285
1286 static int
1287 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1288 {
1289         struct nf_conn *ct, *last;
1290         struct nf_conntrack_tuple_hash *h;
1291         struct hlist_nulls_node *n;
1292         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1293         u_int8_t l3proto = nfmsg->nfgen_family;
1294         int res;
1295         int cpu;
1296         struct hlist_nulls_head *list;
1297         struct net *net = sock_net(skb->sk);
1298
1299         if (cb->args[2])
1300                 return 0;
1301
1302         last = (struct nf_conn *)cb->args[1];
1303
1304         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1305                 struct ct_pcpu *pcpu;
1306
1307                 if (!cpu_possible(cpu))
1308                         continue;
1309
1310                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1311                 spin_lock_bh(&pcpu->lock);
1312                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1313 restart:
1314                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1315                         ct = nf_ct_tuplehash_to_ctrack(h);
1316                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1317                                 continue;
1318                         if (cb->args[1]) {
1319                                 if (ct != last)
1320                                         continue;
1321                                 cb->args[1] = 0;
1322                         }
1323                         rcu_read_lock();
1324                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1325                                                   cb->nlh->nlmsg_seq,
1326                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1327                                                   ct);
1328                         rcu_read_unlock();
1329                         if (res < 0) {
1330                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1331                                         continue;
1332                                 cb->args[0] = cpu;
1333                                 cb->args[1] = (unsigned long)ct;
1334                                 spin_unlock_bh(&pcpu->lock);
1335                                 goto out;
1336                         }
1337                 }
1338                 if (cb->args[1]) {
1339                         cb->args[1] = 0;
1340                         goto restart;
1341                 }
1342                 spin_unlock_bh(&pcpu->lock);
1343         }
1344         cb->args[2] = 1;
1345 out:
1346         if (last)
1347                 nf_ct_put(last);
1348
1349         return skb->len;
1350 }
1351
1352 static int
1353 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1354 {
1355         return ctnetlink_dump_list(skb, cb, true);
1356 }
1357
1358 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1359                                   struct sk_buff *skb,
1360                                   const struct nlmsghdr *nlh,
1361                                   const struct nlattr * const cda[],
1362                                   struct netlink_ext_ack *extack)
1363 {
1364         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1365                 struct netlink_dump_control c = {
1366                         .dump = ctnetlink_dump_dying,
1367                         .done = ctnetlink_done_list,
1368                 };
1369                 return netlink_dump_start(ctnl, skb, nlh, &c);
1370         }
1371
1372         return -EOPNOTSUPP;
1373 }
1374
1375 static int
1376 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1377 {
1378         return ctnetlink_dump_list(skb, cb, false);
1379 }
1380
1381 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1382                                         struct sk_buff *skb,
1383                                         const struct nlmsghdr *nlh,
1384                                         const struct nlattr * const cda[],
1385                                         struct netlink_ext_ack *extack)
1386 {
1387         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1388                 struct netlink_dump_control c = {
1389                         .dump = ctnetlink_dump_unconfirmed,
1390                         .done = ctnetlink_done_list,
1391                 };
1392                 return netlink_dump_start(ctnl, skb, nlh, &c);
1393         }
1394
1395         return -EOPNOTSUPP;
1396 }
1397
1398 #ifdef CONFIG_NF_NAT_NEEDED
1399 static int
1400 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1401                           enum nf_nat_manip_type manip,
1402                           const struct nlattr *attr)
1403 {
1404         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1405         int err;
1406
1407         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1408         if (!parse_nat_setup) {
1409 #ifdef CONFIG_MODULES
1410                 rcu_read_unlock();
1411                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1412                 if (request_module("nf-nat") < 0) {
1413                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1414                         rcu_read_lock();
1415                         return -EOPNOTSUPP;
1416                 }
1417                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1418                 rcu_read_lock();
1419                 if (nfnetlink_parse_nat_setup_hook)
1420                         return -EAGAIN;
1421 #endif
1422                 return -EOPNOTSUPP;
1423         }
1424
1425         err = parse_nat_setup(ct, manip, attr);
1426         if (err == -EAGAIN) {
1427 #ifdef CONFIG_MODULES
1428                 rcu_read_unlock();
1429                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1430                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1431                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1432                         rcu_read_lock();
1433                         return -EOPNOTSUPP;
1434                 }
1435                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1436                 rcu_read_lock();
1437 #else
1438                 err = -EOPNOTSUPP;
1439 #endif
1440         }
1441         return err;
1442 }
1443 #endif
1444
1445 static void
1446 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1447                           unsigned long off)
1448 {
1449         unsigned int bit;
1450
1451         /* Ignore these unchangable bits */
1452         on &= ~IPS_UNCHANGEABLE_MASK;
1453         off &= ~IPS_UNCHANGEABLE_MASK;
1454
1455         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1456                 if (on & (1 << bit))
1457                         set_bit(bit, &ct->status);
1458                 else if (off & (1 << bit))
1459                         clear_bit(bit, &ct->status);
1460         }
1461 }
1462
1463 static int
1464 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1465 {
1466         unsigned long d;
1467         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1468         d = ct->status ^ status;
1469
1470         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1471                 /* unchangeable */
1472                 return -EBUSY;
1473
1474         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1475                 /* SEEN_REPLY bit can only be set */
1476                 return -EBUSY;
1477
1478         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1479                 /* ASSURED bit can only be set */
1480                 return -EBUSY;
1481
1482         __ctnetlink_change_status(ct, status, 0);
1483         return 0;
1484 }
1485
1486 static int
1487 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1488 {
1489 #ifdef CONFIG_NF_NAT_NEEDED
1490         int ret;
1491
1492         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1493                 return 0;
1494
1495         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1496                                         cda[CTA_NAT_DST]);
1497         if (ret < 0)
1498                 return ret;
1499
1500         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1501                                         cda[CTA_NAT_SRC]);
1502         return ret;
1503 #else
1504         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1505                 return 0;
1506         return -EOPNOTSUPP;
1507 #endif
1508 }
1509
1510 static int ctnetlink_change_helper(struct nf_conn *ct,
1511                                    const struct nlattr * const cda[])
1512 {
1513         struct nf_conntrack_helper *helper;
1514         struct nf_conn_help *help = nfct_help(ct);
1515         char *helpname = NULL;
1516         struct nlattr *helpinfo = NULL;
1517         int err;
1518
1519         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1520         if (err < 0)
1521                 return err;
1522
1523         /* don't change helper of sibling connections */
1524         if (ct->master) {
1525                 /* If we try to change the helper to the same thing twice,
1526                  * treat the second attempt as a no-op instead of returning
1527                  * an error.
1528                  */
1529                 err = -EBUSY;
1530                 if (help) {
1531                         rcu_read_lock();
1532                         helper = rcu_dereference(help->helper);
1533                         if (helper && !strcmp(helper->name, helpname))
1534                                 err = 0;
1535                         rcu_read_unlock();
1536                 }
1537
1538                 return err;
1539         }
1540
1541         if (!strcmp(helpname, "")) {
1542                 if (help && help->helper) {
1543                         /* we had a helper before ... */
1544                         nf_ct_remove_expectations(ct);
1545                         RCU_INIT_POINTER(help->helper, NULL);
1546                 }
1547
1548                 return 0;
1549         }
1550
1551         rcu_read_lock();
1552         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1553                                             nf_ct_protonum(ct));
1554         if (helper == NULL) {
1555                 rcu_read_unlock();
1556                 return -EOPNOTSUPP;
1557         }
1558
1559         if (help) {
1560                 if (help->helper == helper) {
1561                         /* update private helper data if allowed. */
1562                         if (helper->from_nlattr)
1563                                 helper->from_nlattr(helpinfo, ct);
1564                         err = 0;
1565                 } else
1566                         err = -EBUSY;
1567         } else {
1568                 /* we cannot set a helper for an existing conntrack */
1569                 err = -EOPNOTSUPP;
1570         }
1571
1572         rcu_read_unlock();
1573         return err;
1574 }
1575
1576 static int ctnetlink_change_timeout(struct nf_conn *ct,
1577                                     const struct nlattr * const cda[])
1578 {
1579         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1580
1581         if (timeout > INT_MAX)
1582                 timeout = INT_MAX;
1583         ct->timeout = nfct_time_stamp + (u32)timeout;
1584
1585         if (test_bit(IPS_DYING_BIT, &ct->status))
1586                 return -ETIME;
1587
1588         return 0;
1589 }
1590
1591 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1592         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1593         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1594         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1595 };
1596
1597 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1598                                       const struct nlattr * const cda[])
1599 {
1600         const struct nlattr *attr = cda[CTA_PROTOINFO];
1601         const struct nf_conntrack_l4proto *l4proto;
1602         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1603         int err = 0;
1604
1605         err = nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy,
1606                                NULL);
1607         if (err < 0)
1608                 return err;
1609
1610         rcu_read_lock();
1611         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1612         if (l4proto->from_nlattr)
1613                 err = l4proto->from_nlattr(tb, ct);
1614         rcu_read_unlock();
1615
1616         return err;
1617 }
1618
1619 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1620         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1621         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1622         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1623 };
1624
1625 static int change_seq_adj(struct nf_ct_seqadj *seq,
1626                           const struct nlattr * const attr)
1627 {
1628         int err;
1629         struct nlattr *cda[CTA_SEQADJ_MAX+1];
1630
1631         err = nla_parse_nested(cda, CTA_SEQADJ_MAX, attr, seqadj_policy, NULL);
1632         if (err < 0)
1633                 return err;
1634
1635         if (!cda[CTA_SEQADJ_CORRECTION_POS])
1636                 return -EINVAL;
1637
1638         seq->correction_pos =
1639                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1640
1641         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1642                 return -EINVAL;
1643
1644         seq->offset_before =
1645                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1646
1647         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1648                 return -EINVAL;
1649
1650         seq->offset_after =
1651                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1652
1653         return 0;
1654 }
1655
1656 static int
1657 ctnetlink_change_seq_adj(struct nf_conn *ct,
1658                          const struct nlattr * const cda[])
1659 {
1660         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1661         int ret = 0;
1662
1663         if (!seqadj)
1664                 return 0;
1665
1666         spin_lock_bh(&ct->lock);
1667         if (cda[CTA_SEQ_ADJ_ORIG]) {
1668                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1669                                      cda[CTA_SEQ_ADJ_ORIG]);
1670                 if (ret < 0)
1671                         goto err;
1672
1673                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1674         }
1675
1676         if (cda[CTA_SEQ_ADJ_REPLY]) {
1677                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1678                                      cda[CTA_SEQ_ADJ_REPLY]);
1679                 if (ret < 0)
1680                         goto err;
1681
1682                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1683         }
1684
1685         spin_unlock_bh(&ct->lock);
1686         return 0;
1687 err:
1688         spin_unlock_bh(&ct->lock);
1689         return ret;
1690 }
1691
1692 static int
1693 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1694 {
1695 #ifdef CONFIG_NF_CONNTRACK_LABELS
1696         size_t len = nla_len(cda[CTA_LABELS]);
1697         const void *mask = cda[CTA_LABELS_MASK];
1698
1699         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1700                 return -EINVAL;
1701
1702         if (mask) {
1703                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1704                     nla_len(cda[CTA_LABELS_MASK]) != len)
1705                         return -EINVAL;
1706                 mask = nla_data(cda[CTA_LABELS_MASK]);
1707         }
1708
1709         len /= sizeof(u32);
1710
1711         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1712 #else
1713         return -EOPNOTSUPP;
1714 #endif
1715 }
1716
1717 static int
1718 ctnetlink_change_conntrack(struct nf_conn *ct,
1719                            const struct nlattr * const cda[])
1720 {
1721         int err;
1722
1723         /* only allow NAT changes and master assignation for new conntracks */
1724         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1725                 return -EOPNOTSUPP;
1726
1727         if (cda[CTA_HELP]) {
1728                 err = ctnetlink_change_helper(ct, cda);
1729                 if (err < 0)
1730                         return err;
1731         }
1732
1733         if (cda[CTA_TIMEOUT]) {
1734                 err = ctnetlink_change_timeout(ct, cda);
1735                 if (err < 0)
1736                         return err;
1737         }
1738
1739         if (cda[CTA_STATUS]) {
1740                 err = ctnetlink_change_status(ct, cda);
1741                 if (err < 0)
1742                         return err;
1743         }
1744
1745         if (cda[CTA_PROTOINFO]) {
1746                 err = ctnetlink_change_protoinfo(ct, cda);
1747                 if (err < 0)
1748                         return err;
1749         }
1750
1751 #if defined(CONFIG_NF_CONNTRACK_MARK)
1752         if (cda[CTA_MARK])
1753                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1754 #endif
1755
1756         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1757                 err = ctnetlink_change_seq_adj(ct, cda);
1758                 if (err < 0)
1759                         return err;
1760         }
1761
1762         if (cda[CTA_LABELS]) {
1763                 err = ctnetlink_attach_labels(ct, cda);
1764                 if (err < 0)
1765                         return err;
1766         }
1767
1768         return 0;
1769 }
1770
1771 static struct nf_conn *
1772 ctnetlink_create_conntrack(struct net *net,
1773                            const struct nf_conntrack_zone *zone,
1774                            const struct nlattr * const cda[],
1775                            struct nf_conntrack_tuple *otuple,
1776                            struct nf_conntrack_tuple *rtuple,
1777                            u8 u3)
1778 {
1779         struct nf_conn *ct;
1780         int err = -EINVAL;
1781         struct nf_conntrack_helper *helper;
1782         struct nf_conn_tstamp *tstamp;
1783         u64 timeout;
1784
1785         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1786         if (IS_ERR(ct))
1787                 return ERR_PTR(-ENOMEM);
1788
1789         if (!cda[CTA_TIMEOUT])
1790                 goto err1;
1791
1792         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1793         if (timeout > INT_MAX)
1794                 timeout = INT_MAX;
1795         ct->timeout = (u32)timeout + nfct_time_stamp;
1796
1797         rcu_read_lock();
1798         if (cda[CTA_HELP]) {
1799                 char *helpname = NULL;
1800                 struct nlattr *helpinfo = NULL;
1801
1802                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1803                 if (err < 0)
1804                         goto err2;
1805
1806                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1807                                                     nf_ct_protonum(ct));
1808                 if (helper == NULL) {
1809                         rcu_read_unlock();
1810 #ifdef CONFIG_MODULES
1811                         if (request_module("nfct-helper-%s", helpname) < 0) {
1812                                 err = -EOPNOTSUPP;
1813                                 goto err1;
1814                         }
1815
1816                         rcu_read_lock();
1817                         helper = __nf_conntrack_helper_find(helpname,
1818                                                             nf_ct_l3num(ct),
1819                                                             nf_ct_protonum(ct));
1820                         if (helper) {
1821                                 err = -EAGAIN;
1822                                 goto err2;
1823                         }
1824                         rcu_read_unlock();
1825 #endif
1826                         err = -EOPNOTSUPP;
1827                         goto err1;
1828                 } else {
1829                         struct nf_conn_help *help;
1830
1831                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1832                         if (help == NULL) {
1833                                 err = -ENOMEM;
1834                                 goto err2;
1835                         }
1836                         /* set private helper data if allowed. */
1837                         if (helper->from_nlattr)
1838                                 helper->from_nlattr(helpinfo, ct);
1839
1840                         /* not in hash table yet so not strictly necessary */
1841                         RCU_INIT_POINTER(help->helper, helper);
1842                 }
1843         } else {
1844                 /* try an implicit helper assignation */
1845                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1846                 if (err < 0)
1847                         goto err2;
1848         }
1849
1850         err = ctnetlink_setup_nat(ct, cda);
1851         if (err < 0)
1852                 goto err2;
1853
1854         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1855         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1856         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1857         nf_ct_labels_ext_add(ct);
1858         nfct_seqadj_ext_add(ct);
1859         nfct_synproxy_ext_add(ct);
1860
1861         /* we must add conntrack extensions before confirmation. */
1862         ct->status |= IPS_CONFIRMED;
1863
1864         if (cda[CTA_STATUS]) {
1865                 err = ctnetlink_change_status(ct, cda);
1866                 if (err < 0)
1867                         goto err2;
1868         }
1869
1870         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1871                 err = ctnetlink_change_seq_adj(ct, cda);
1872                 if (err < 0)
1873                         goto err2;
1874         }
1875
1876         memset(&ct->proto, 0, sizeof(ct->proto));
1877         if (cda[CTA_PROTOINFO]) {
1878                 err = ctnetlink_change_protoinfo(ct, cda);
1879                 if (err < 0)
1880                         goto err2;
1881         }
1882
1883 #if defined(CONFIG_NF_CONNTRACK_MARK)
1884         if (cda[CTA_MARK])
1885                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1886 #endif
1887
1888         /* setup master conntrack: this is a confirmed expectation */
1889         if (cda[CTA_TUPLE_MASTER]) {
1890                 struct nf_conntrack_tuple master;
1891                 struct nf_conntrack_tuple_hash *master_h;
1892                 struct nf_conn *master_ct;
1893
1894                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
1895                                             u3, NULL);
1896                 if (err < 0)
1897                         goto err2;
1898
1899                 master_h = nf_conntrack_find_get(net, zone, &master);
1900                 if (master_h == NULL) {
1901                         err = -ENOENT;
1902                         goto err2;
1903                 }
1904                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1905                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1906                 ct->master = master_ct;
1907         }
1908         tstamp = nf_conn_tstamp_find(ct);
1909         if (tstamp)
1910                 tstamp->start = ktime_get_real_ns();
1911
1912         err = nf_conntrack_hash_check_insert(ct);
1913         if (err < 0)
1914                 goto err2;
1915
1916         rcu_read_unlock();
1917
1918         return ct;
1919
1920 err2:
1921         rcu_read_unlock();
1922 err1:
1923         nf_conntrack_free(ct);
1924         return ERR_PTR(err);
1925 }
1926
1927 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
1928                                    struct sk_buff *skb,
1929                                    const struct nlmsghdr *nlh,
1930                                    const struct nlattr * const cda[],
1931                                    struct netlink_ext_ack *extack)
1932 {
1933         struct nf_conntrack_tuple otuple, rtuple;
1934         struct nf_conntrack_tuple_hash *h = NULL;
1935         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1936         struct nf_conn *ct;
1937         u_int8_t u3 = nfmsg->nfgen_family;
1938         struct nf_conntrack_zone zone;
1939         int err;
1940
1941         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1942         if (err < 0)
1943                 return err;
1944
1945         if (cda[CTA_TUPLE_ORIG]) {
1946                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
1947                                             u3, &zone);
1948                 if (err < 0)
1949                         return err;
1950         }
1951
1952         if (cda[CTA_TUPLE_REPLY]) {
1953                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
1954                                             u3, &zone);
1955                 if (err < 0)
1956                         return err;
1957         }
1958
1959         if (cda[CTA_TUPLE_ORIG])
1960                 h = nf_conntrack_find_get(net, &zone, &otuple);
1961         else if (cda[CTA_TUPLE_REPLY])
1962                 h = nf_conntrack_find_get(net, &zone, &rtuple);
1963
1964         if (h == NULL) {
1965                 err = -ENOENT;
1966                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1967                         enum ip_conntrack_events events;
1968
1969                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
1970                                 return -EINVAL;
1971                         if (otuple.dst.protonum != rtuple.dst.protonum)
1972                                 return -EINVAL;
1973
1974                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
1975                                                         &rtuple, u3);
1976                         if (IS_ERR(ct))
1977                                 return PTR_ERR(ct);
1978
1979                         err = 0;
1980                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1981                                 events = 1 << IPCT_RELATED;
1982                         else
1983                                 events = 1 << IPCT_NEW;
1984
1985                         if (cda[CTA_LABELS] &&
1986                             ctnetlink_attach_labels(ct, cda) == 0)
1987                                 events |= (1 << IPCT_LABEL);
1988
1989                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1990                                                       (1 << IPCT_ASSURED) |
1991                                                       (1 << IPCT_HELPER) |
1992                                                       (1 << IPCT_PROTOINFO) |
1993                                                       (1 << IPCT_SEQADJ) |
1994                                                       (1 << IPCT_MARK) | events,
1995                                                       ct, NETLINK_CB(skb).portid,
1996                                                       nlmsg_report(nlh));
1997                         nf_ct_put(ct);
1998                 }
1999
2000                 return err;
2001         }
2002         /* implicit 'else' */
2003
2004         err = -EEXIST;
2005         ct = nf_ct_tuplehash_to_ctrack(h);
2006         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2007                 err = ctnetlink_change_conntrack(ct, cda);
2008                 if (err == 0) {
2009                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2010                                                       (1 << IPCT_ASSURED) |
2011                                                       (1 << IPCT_HELPER) |
2012                                                       (1 << IPCT_LABEL) |
2013                                                       (1 << IPCT_PROTOINFO) |
2014                                                       (1 << IPCT_SEQADJ) |
2015                                                       (1 << IPCT_MARK),
2016                                                       ct, NETLINK_CB(skb).portid,
2017                                                       nlmsg_report(nlh));
2018                 }
2019         }
2020
2021         nf_ct_put(ct);
2022         return err;
2023 }
2024
2025 static int
2026 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2027                                 __u16 cpu, const struct ip_conntrack_stat *st)
2028 {
2029         struct nlmsghdr *nlh;
2030         struct nfgenmsg *nfmsg;
2031         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2032
2033         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2034                               IPCTNL_MSG_CT_GET_STATS_CPU);
2035         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2036         if (nlh == NULL)
2037                 goto nlmsg_failure;
2038
2039         nfmsg = nlmsg_data(nlh);
2040         nfmsg->nfgen_family = AF_UNSPEC;
2041         nfmsg->version      = NFNETLINK_V0;
2042         nfmsg->res_id       = htons(cpu);
2043
2044         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2045             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2046             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2047             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2048             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2049                                 htonl(st->insert_failed)) ||
2050             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2051             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2052             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2053             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2054                                 htonl(st->search_restart)))
2055                 goto nla_put_failure;
2056
2057         nlmsg_end(skb, nlh);
2058         return skb->len;
2059
2060 nla_put_failure:
2061 nlmsg_failure:
2062         nlmsg_cancel(skb, nlh);
2063         return -1;
2064 }
2065
2066 static int
2067 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2068 {
2069         int cpu;
2070         struct net *net = sock_net(skb->sk);
2071
2072         if (cb->args[0] == nr_cpu_ids)
2073                 return 0;
2074
2075         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2076                 const struct ip_conntrack_stat *st;
2077
2078                 if (!cpu_possible(cpu))
2079                         continue;
2080
2081                 st = per_cpu_ptr(net->ct.stat, cpu);
2082                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2083                                                     NETLINK_CB(cb->skb).portid,
2084                                                     cb->nlh->nlmsg_seq,
2085                                                     cpu, st) < 0)
2086                                 break;
2087         }
2088         cb->args[0] = cpu;
2089
2090         return skb->len;
2091 }
2092
2093 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2094                                  struct sk_buff *skb,
2095                                  const struct nlmsghdr *nlh,
2096                                  const struct nlattr * const cda[],
2097                                  struct netlink_ext_ack *extack)
2098 {
2099         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2100                 struct netlink_dump_control c = {
2101                         .dump = ctnetlink_ct_stat_cpu_dump,
2102                 };
2103                 return netlink_dump_start(ctnl, skb, nlh, &c);
2104         }
2105
2106         return 0;
2107 }
2108
2109 static int
2110 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2111                             struct net *net)
2112 {
2113         struct nlmsghdr *nlh;
2114         struct nfgenmsg *nfmsg;
2115         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2116         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2117
2118         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2119         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2120         if (nlh == NULL)
2121                 goto nlmsg_failure;
2122
2123         nfmsg = nlmsg_data(nlh);
2124         nfmsg->nfgen_family = AF_UNSPEC;
2125         nfmsg->version      = NFNETLINK_V0;
2126         nfmsg->res_id       = 0;
2127
2128         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2129                 goto nla_put_failure;
2130
2131         nlmsg_end(skb, nlh);
2132         return skb->len;
2133
2134 nla_put_failure:
2135 nlmsg_failure:
2136         nlmsg_cancel(skb, nlh);
2137         return -1;
2138 }
2139
2140 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2141                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2142                              const struct nlattr * const cda[],
2143                              struct netlink_ext_ack *extack)
2144 {
2145         struct sk_buff *skb2;
2146         int err;
2147
2148         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2149         if (skb2 == NULL)
2150                 return -ENOMEM;
2151
2152         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2153                                           nlh->nlmsg_seq,
2154                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2155                                           sock_net(skb->sk));
2156         if (err <= 0)
2157                 goto free;
2158
2159         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2160         if (err < 0)
2161                 goto out;
2162
2163         return 0;
2164
2165 free:
2166         kfree_skb(skb2);
2167 out:
2168         /* this avoids a loop in nfnetlink. */
2169         return err == -EAGAIN ? -ENOBUFS : err;
2170 }
2171
2172 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2173         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2174         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2175         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2176         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2177         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2178         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2179                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2180         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2181         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2182         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2183         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2184         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2185 };
2186
2187 static struct nf_conntrack_expect *
2188 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2189                        struct nf_conntrack_helper *helper,
2190                        struct nf_conntrack_tuple *tuple,
2191                        struct nf_conntrack_tuple *mask);
2192
2193 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2194 static size_t
2195 ctnetlink_glue_build_size(const struct nf_conn *ct)
2196 {
2197         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2198                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2199                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2200                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2201                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2202                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2203                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2204                + nla_total_size(0) /* CTA_PROTOINFO */
2205                + nla_total_size(0) /* CTA_HELP */
2206                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2207                + ctnetlink_secctx_size(ct)
2208 #ifdef CONFIG_NF_NAT_NEEDED
2209                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2210                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2211 #endif
2212 #ifdef CONFIG_NF_CONNTRACK_MARK
2213                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2214 #endif
2215 #ifdef CONFIG_NF_CONNTRACK_ZONES
2216                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2217 #endif
2218                + ctnetlink_proto_size(ct)
2219                ;
2220 }
2221
2222 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2223                                              enum ip_conntrack_info *ctinfo)
2224 {
2225         return nf_ct_get(skb, ctinfo);
2226 }
2227
2228 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2229 {
2230         const struct nf_conntrack_zone *zone;
2231         struct nlattr *nest_parms;
2232
2233         zone = nf_ct_zone(ct);
2234
2235         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
2236         if (!nest_parms)
2237                 goto nla_put_failure;
2238         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2239                 goto nla_put_failure;
2240         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2241                                    NF_CT_ZONE_DIR_ORIG) < 0)
2242                 goto nla_put_failure;
2243         nla_nest_end(skb, nest_parms);
2244
2245         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
2246         if (!nest_parms)
2247                 goto nla_put_failure;
2248         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2249                 goto nla_put_failure;
2250         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2251                                    NF_CT_ZONE_DIR_REPL) < 0)
2252                 goto nla_put_failure;
2253         nla_nest_end(skb, nest_parms);
2254
2255         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2256                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2257                 goto nla_put_failure;
2258
2259         if (ctnetlink_dump_id(skb, ct) < 0)
2260                 goto nla_put_failure;
2261
2262         if (ctnetlink_dump_status(skb, ct) < 0)
2263                 goto nla_put_failure;
2264
2265         if (ctnetlink_dump_timeout(skb, ct) < 0)
2266                 goto nla_put_failure;
2267
2268         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2269                 goto nla_put_failure;
2270
2271         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2272                 goto nla_put_failure;
2273
2274 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2275         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2276                 goto nla_put_failure;
2277 #endif
2278         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2279                 goto nla_put_failure;
2280
2281         if ((ct->status & IPS_SEQ_ADJUST) &&
2282             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2283                 goto nla_put_failure;
2284
2285 #ifdef CONFIG_NF_CONNTRACK_MARK
2286         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2287                 goto nla_put_failure;
2288 #endif
2289         if (ctnetlink_dump_labels(skb, ct) < 0)
2290                 goto nla_put_failure;
2291         return 0;
2292
2293 nla_put_failure:
2294         return -ENOSPC;
2295 }
2296
2297 static int
2298 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2299                      enum ip_conntrack_info ctinfo,
2300                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2301 {
2302         struct nlattr *nest_parms;
2303
2304         nest_parms = nla_nest_start(skb, ct_attr | NLA_F_NESTED);
2305         if (!nest_parms)
2306                 goto nla_put_failure;
2307
2308         if (__ctnetlink_glue_build(skb, ct) < 0)
2309                 goto nla_put_failure;
2310
2311         nla_nest_end(skb, nest_parms);
2312
2313         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2314                 goto nla_put_failure;
2315
2316         return 0;
2317
2318 nla_put_failure:
2319         return -ENOSPC;
2320 }
2321
2322 static int
2323 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2324 {
2325         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2326         unsigned long d = ct->status ^ status;
2327
2328         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2329                 /* SEEN_REPLY bit can only be set */
2330                 return -EBUSY;
2331
2332         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2333                 /* ASSURED bit can only be set */
2334                 return -EBUSY;
2335
2336         /* This check is less strict than ctnetlink_change_status()
2337          * because callers often flip IPS_EXPECTED bits when sending
2338          * an NFQA_CT attribute to the kernel.  So ignore the
2339          * unchangeable bits but do not error out. Also user programs
2340          * are allowed to clear the bits that they are allowed to change.
2341          */
2342         __ctnetlink_change_status(ct, status, ~status);
2343         return 0;
2344 }
2345
2346 static int
2347 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2348 {
2349         int err;
2350
2351         if (cda[CTA_TIMEOUT]) {
2352                 err = ctnetlink_change_timeout(ct, cda);
2353                 if (err < 0)
2354                         return err;
2355         }
2356         if (cda[CTA_STATUS]) {
2357                 err = ctnetlink_update_status(ct, cda);
2358                 if (err < 0)
2359                         return err;
2360         }
2361         if (cda[CTA_HELP]) {
2362                 err = ctnetlink_change_helper(ct, cda);
2363                 if (err < 0)
2364                         return err;
2365         }
2366         if (cda[CTA_LABELS]) {
2367                 err = ctnetlink_attach_labels(ct, cda);
2368                 if (err < 0)
2369                         return err;
2370         }
2371 #if defined(CONFIG_NF_CONNTRACK_MARK)
2372         if (cda[CTA_MARK]) {
2373                 u32 mask = 0, mark, newmark;
2374                 if (cda[CTA_MARK_MASK])
2375                         mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2376
2377                 mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2378                 newmark = (ct->mark & mask) ^ mark;
2379                 if (newmark != ct->mark)
2380                         ct->mark = newmark;
2381         }
2382 #endif
2383         return 0;
2384 }
2385
2386 static int
2387 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2388 {
2389         struct nlattr *cda[CTA_MAX+1];
2390         int ret;
2391
2392         ret = nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy, NULL);
2393         if (ret < 0)
2394                 return ret;
2395
2396         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2397 }
2398
2399 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2400                                     const struct nf_conn *ct,
2401                                     struct nf_conntrack_tuple *tuple,
2402                                     struct nf_conntrack_tuple *mask)
2403 {
2404         int err;
2405
2406         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2407                                     nf_ct_l3num(ct), NULL);
2408         if (err < 0)
2409                 return err;
2410
2411         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2412                                      nf_ct_l3num(ct), NULL);
2413 }
2414
2415 static int
2416 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2417                              u32 portid, u32 report)
2418 {
2419         struct nlattr *cda[CTA_EXPECT_MAX+1];
2420         struct nf_conntrack_tuple tuple, mask;
2421         struct nf_conntrack_helper *helper = NULL;
2422         struct nf_conntrack_expect *exp;
2423         int err;
2424
2425         err = nla_parse_nested(cda, CTA_EXPECT_MAX, attr, exp_nla_policy,
2426                                NULL);
2427         if (err < 0)
2428                 return err;
2429
2430         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2431                                        ct, &tuple, &mask);
2432         if (err < 0)
2433                 return err;
2434
2435         if (cda[CTA_EXPECT_HELP_NAME]) {
2436                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2437
2438                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2439                                                     nf_ct_protonum(ct));
2440                 if (helper == NULL)
2441                         return -EOPNOTSUPP;
2442         }
2443
2444         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2445                                      helper, &tuple, &mask);
2446         if (IS_ERR(exp))
2447                 return PTR_ERR(exp);
2448
2449         err = nf_ct_expect_related_report(exp, portid, report);
2450         nf_ct_expect_put(exp);
2451         return err;
2452 }
2453
2454 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2455                                   enum ip_conntrack_info ctinfo, int diff)
2456 {
2457         if (!(ct->status & IPS_NAT_MASK))
2458                 return;
2459
2460         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2461 }
2462
2463 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2464         .get_ct         = ctnetlink_glue_get_ct,
2465         .build_size     = ctnetlink_glue_build_size,
2466         .build          = ctnetlink_glue_build,
2467         .parse          = ctnetlink_glue_parse,
2468         .attach_expect  = ctnetlink_glue_attach_expect,
2469         .seq_adjust     = ctnetlink_glue_seqadj,
2470 };
2471 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2472
2473 /***********************************************************************
2474  * EXPECT
2475  ***********************************************************************/
2476
2477 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2478                                     const struct nf_conntrack_tuple *tuple,
2479                                     u32 type)
2480 {
2481         struct nlattr *nest_parms;
2482
2483         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2484         if (!nest_parms)
2485                 goto nla_put_failure;
2486         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2487                 goto nla_put_failure;
2488         nla_nest_end(skb, nest_parms);
2489
2490         return 0;
2491
2492 nla_put_failure:
2493         return -1;
2494 }
2495
2496 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2497                                    const struct nf_conntrack_tuple *tuple,
2498                                    const struct nf_conntrack_tuple_mask *mask)
2499 {
2500         const struct nf_conntrack_l3proto *l3proto;
2501         const struct nf_conntrack_l4proto *l4proto;
2502         struct nf_conntrack_tuple m;
2503         struct nlattr *nest_parms;
2504         int ret;
2505
2506         memset(&m, 0xFF, sizeof(m));
2507         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2508         m.src.u.all = mask->src.u.all;
2509         m.dst.protonum = tuple->dst.protonum;
2510
2511         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2512         if (!nest_parms)
2513                 goto nla_put_failure;
2514
2515         rcu_read_lock();
2516         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2517         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2518         if (ret >= 0) {
2519                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2520                                                tuple->dst.protonum);
2521         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2522         }
2523         rcu_read_unlock();
2524
2525         if (unlikely(ret < 0))
2526                 goto nla_put_failure;
2527
2528         nla_nest_end(skb, nest_parms);
2529
2530         return 0;
2531
2532 nla_put_failure:
2533         return -1;
2534 }
2535
2536 static const union nf_inet_addr any_addr;
2537
2538 static int
2539 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2540                           const struct nf_conntrack_expect *exp)
2541 {
2542         struct nf_conn *master = exp->master;
2543         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2544         struct nf_conn_help *help;
2545 #ifdef CONFIG_NF_NAT_NEEDED
2546         struct nlattr *nest_parms;
2547         struct nf_conntrack_tuple nat_tuple = {};
2548 #endif
2549         struct nf_ct_helper_expectfn *expfn;
2550
2551         if (timeout < 0)
2552                 timeout = 0;
2553
2554         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2555                 goto nla_put_failure;
2556         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2557                 goto nla_put_failure;
2558         if (ctnetlink_exp_dump_tuple(skb,
2559                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2560                                  CTA_EXPECT_MASTER) < 0)
2561                 goto nla_put_failure;
2562
2563 #ifdef CONFIG_NF_NAT_NEEDED
2564         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2565             exp->saved_proto.all) {
2566                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2567                 if (!nest_parms)
2568                         goto nla_put_failure;
2569
2570                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2571                         goto nla_put_failure;
2572
2573                 nat_tuple.src.l3num = nf_ct_l3num(master);
2574                 nat_tuple.src.u3 = exp->saved_addr;
2575                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2576                 nat_tuple.src.u = exp->saved_proto;
2577
2578                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2579                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2580                         goto nla_put_failure;
2581                 nla_nest_end(skb, nest_parms);
2582         }
2583 #endif
2584         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2585             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2586             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2587             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2588                 goto nla_put_failure;
2589         help = nfct_help(master);
2590         if (help) {
2591                 struct nf_conntrack_helper *helper;
2592
2593                 helper = rcu_dereference(help->helper);
2594                 if (helper &&
2595                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2596                         goto nla_put_failure;
2597         }
2598         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2599         if (expfn != NULL &&
2600             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2601                 goto nla_put_failure;
2602
2603         return 0;
2604
2605 nla_put_failure:
2606         return -1;
2607 }
2608
2609 static int
2610 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2611                         int event, const struct nf_conntrack_expect *exp)
2612 {
2613         struct nlmsghdr *nlh;
2614         struct nfgenmsg *nfmsg;
2615         unsigned int flags = portid ? NLM_F_MULTI : 0;
2616
2617         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2618         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2619         if (nlh == NULL)
2620                 goto nlmsg_failure;
2621
2622         nfmsg = nlmsg_data(nlh);
2623         nfmsg->nfgen_family = exp->tuple.src.l3num;
2624         nfmsg->version      = NFNETLINK_V0;
2625         nfmsg->res_id       = 0;
2626
2627         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2628                 goto nla_put_failure;
2629
2630         nlmsg_end(skb, nlh);
2631         return skb->len;
2632
2633 nlmsg_failure:
2634 nla_put_failure:
2635         nlmsg_cancel(skb, nlh);
2636         return -1;
2637 }
2638
2639 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2640 static int
2641 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2642 {
2643         struct nf_conntrack_expect *exp = item->exp;
2644         struct net *net = nf_ct_exp_net(exp);
2645         struct nlmsghdr *nlh;
2646         struct nfgenmsg *nfmsg;
2647         struct sk_buff *skb;
2648         unsigned int type, group;
2649         int flags = 0;
2650
2651         if (events & (1 << IPEXP_DESTROY)) {
2652                 type = IPCTNL_MSG_EXP_DELETE;
2653                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2654         } else if (events & (1 << IPEXP_NEW)) {
2655                 type = IPCTNL_MSG_EXP_NEW;
2656                 flags = NLM_F_CREATE|NLM_F_EXCL;
2657                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2658         } else
2659                 return 0;
2660
2661         if (!item->report && !nfnetlink_has_listeners(net, group))
2662                 return 0;
2663
2664         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2665         if (skb == NULL)
2666                 goto errout;
2667
2668         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2669         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2670         if (nlh == NULL)
2671                 goto nlmsg_failure;
2672
2673         nfmsg = nlmsg_data(nlh);
2674         nfmsg->nfgen_family = exp->tuple.src.l3num;
2675         nfmsg->version      = NFNETLINK_V0;
2676         nfmsg->res_id       = 0;
2677
2678         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2679                 goto nla_put_failure;
2680
2681         nlmsg_end(skb, nlh);
2682         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2683         return 0;
2684
2685 nla_put_failure:
2686         nlmsg_cancel(skb, nlh);
2687 nlmsg_failure:
2688         kfree_skb(skb);
2689 errout:
2690         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2691         return 0;
2692 }
2693 #endif
2694 static int ctnetlink_exp_done(struct netlink_callback *cb)
2695 {
2696         if (cb->args[1])
2697                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2698         return 0;
2699 }
2700
2701 static int
2702 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2703 {
2704         struct net *net = sock_net(skb->sk);
2705         struct nf_conntrack_expect *exp, *last;
2706         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2707         u_int8_t l3proto = nfmsg->nfgen_family;
2708
2709         rcu_read_lock();
2710         last = (struct nf_conntrack_expect *)cb->args[1];
2711         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2712 restart:
2713                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2714                                          hnode) {
2715                         if (l3proto && exp->tuple.src.l3num != l3proto)
2716                                 continue;
2717
2718                         if (!net_eq(nf_ct_net(exp->master), net))
2719                                 continue;
2720
2721                         if (cb->args[1]) {
2722                                 if (exp != last)
2723                                         continue;
2724                                 cb->args[1] = 0;
2725                         }
2726                         if (ctnetlink_exp_fill_info(skb,
2727                                                     NETLINK_CB(cb->skb).portid,
2728                                                     cb->nlh->nlmsg_seq,
2729                                                     IPCTNL_MSG_EXP_NEW,
2730                                                     exp) < 0) {
2731                                 if (!refcount_inc_not_zero(&exp->use))
2732                                         continue;
2733                                 cb->args[1] = (unsigned long)exp;
2734                                 goto out;
2735                         }
2736                 }
2737                 if (cb->args[1]) {
2738                         cb->args[1] = 0;
2739                         goto restart;
2740                 }
2741         }
2742 out:
2743         rcu_read_unlock();
2744         if (last)
2745                 nf_ct_expect_put(last);
2746
2747         return skb->len;
2748 }
2749
2750 static int
2751 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2752 {
2753         struct nf_conntrack_expect *exp, *last;
2754         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2755         struct nf_conn *ct = cb->data;
2756         struct nf_conn_help *help = nfct_help(ct);
2757         u_int8_t l3proto = nfmsg->nfgen_family;
2758
2759         if (cb->args[0])
2760                 return 0;
2761
2762         rcu_read_lock();
2763         last = (struct nf_conntrack_expect *)cb->args[1];
2764 restart:
2765         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2766                 if (l3proto && exp->tuple.src.l3num != l3proto)
2767                         continue;
2768                 if (cb->args[1]) {
2769                         if (exp != last)
2770                                 continue;
2771                         cb->args[1] = 0;
2772                 }
2773                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2774                                             cb->nlh->nlmsg_seq,
2775                                             IPCTNL_MSG_EXP_NEW,
2776                                             exp) < 0) {
2777                         if (!refcount_inc_not_zero(&exp->use))
2778                                 continue;
2779                         cb->args[1] = (unsigned long)exp;
2780                         goto out;
2781                 }
2782         }
2783         if (cb->args[1]) {
2784                 cb->args[1] = 0;
2785                 goto restart;
2786         }
2787         cb->args[0] = 1;
2788 out:
2789         rcu_read_unlock();
2790         if (last)
2791                 nf_ct_expect_put(last);
2792
2793         return skb->len;
2794 }
2795
2796 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2797                                  struct sk_buff *skb,
2798                                  const struct nlmsghdr *nlh,
2799                                  const struct nlattr * const cda[],
2800                                  struct netlink_ext_ack *extack)
2801 {
2802         int err;
2803         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2804         u_int8_t u3 = nfmsg->nfgen_family;
2805         struct nf_conntrack_tuple tuple;
2806         struct nf_conntrack_tuple_hash *h;
2807         struct nf_conn *ct;
2808         struct nf_conntrack_zone zone;
2809         struct netlink_dump_control c = {
2810                 .dump = ctnetlink_exp_ct_dump_table,
2811                 .done = ctnetlink_exp_done,
2812         };
2813
2814         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2815                                     u3, NULL);
2816         if (err < 0)
2817                 return err;
2818
2819         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2820         if (err < 0)
2821                 return err;
2822
2823         h = nf_conntrack_find_get(net, &zone, &tuple);
2824         if (!h)
2825                 return -ENOENT;
2826
2827         ct = nf_ct_tuplehash_to_ctrack(h);
2828         /* No expectation linked to this connection tracking. */
2829         if (!nfct_help(ct)) {
2830                 nf_ct_put(ct);
2831                 return 0;
2832         }
2833
2834         c.data = ct;
2835
2836         err = netlink_dump_start(ctnl, skb, nlh, &c);
2837         nf_ct_put(ct);
2838
2839         return err;
2840 }
2841
2842 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
2843                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2844                                 const struct nlattr * const cda[],
2845                                 struct netlink_ext_ack *extack)
2846 {
2847         struct nf_conntrack_tuple tuple;
2848         struct nf_conntrack_expect *exp;
2849         struct sk_buff *skb2;
2850         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2851         u_int8_t u3 = nfmsg->nfgen_family;
2852         struct nf_conntrack_zone zone;
2853         int err;
2854
2855         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2856                 if (cda[CTA_EXPECT_MASTER])
2857                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
2858                                                      extack);
2859                 else {
2860                         struct netlink_dump_control c = {
2861                                 .dump = ctnetlink_exp_dump_table,
2862                                 .done = ctnetlink_exp_done,
2863                         };
2864                         return netlink_dump_start(ctnl, skb, nlh, &c);
2865                 }
2866         }
2867
2868         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2869         if (err < 0)
2870                 return err;
2871
2872         if (cda[CTA_EXPECT_TUPLE])
2873                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2874                                             u3, NULL);
2875         else if (cda[CTA_EXPECT_MASTER])
2876                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
2877                                             u3, NULL);
2878         else
2879                 return -EINVAL;
2880
2881         if (err < 0)
2882                 return err;
2883
2884         exp = nf_ct_expect_find_get(net, &zone, &tuple);
2885         if (!exp)
2886                 return -ENOENT;
2887
2888         if (cda[CTA_EXPECT_ID]) {
2889                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2890                 if (ntohl(id) != (u32)(unsigned long)exp) {
2891                         nf_ct_expect_put(exp);
2892                         return -ENOENT;
2893                 }
2894         }
2895
2896         err = -ENOMEM;
2897         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2898         if (skb2 == NULL) {
2899                 nf_ct_expect_put(exp);
2900                 goto out;
2901         }
2902
2903         rcu_read_lock();
2904         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2905                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2906         rcu_read_unlock();
2907         nf_ct_expect_put(exp);
2908         if (err <= 0)
2909                 goto free;
2910
2911         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2912         if (err < 0)
2913                 goto out;
2914
2915         return 0;
2916
2917 free:
2918         kfree_skb(skb2);
2919 out:
2920         /* this avoids a loop in nfnetlink. */
2921         return err == -EAGAIN ? -ENOBUFS : err;
2922 }
2923
2924 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
2925 {
2926         const struct nf_conn_help *m_help;
2927         const char *name = data;
2928
2929         m_help = nfct_help(exp->master);
2930
2931         return strcmp(m_help->helper->name, name) == 0;
2932 }
2933
2934 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
2935 {
2936         return true;
2937 }
2938
2939 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
2940                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
2941                                 const struct nlattr * const cda[],
2942                                 struct netlink_ext_ack *extack)
2943 {
2944         struct nf_conntrack_expect *exp;
2945         struct nf_conntrack_tuple tuple;
2946         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2947         u_int8_t u3 = nfmsg->nfgen_family;
2948         struct nf_conntrack_zone zone;
2949         int err;
2950
2951         if (cda[CTA_EXPECT_TUPLE]) {
2952                 /* delete a single expect by tuple */
2953                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2954                 if (err < 0)
2955                         return err;
2956
2957                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
2958                                             u3, NULL);
2959                 if (err < 0)
2960                         return err;
2961
2962                 /* bump usage count to 2 */
2963                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
2964                 if (!exp)
2965                         return -ENOENT;
2966
2967                 if (cda[CTA_EXPECT_ID]) {
2968                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2969                         if (ntohl(id) != (u32)(unsigned long)exp) {
2970                                 nf_ct_expect_put(exp);
2971                                 return -ENOENT;
2972                         }
2973                 }
2974
2975                 /* after list removal, usage count == 1 */
2976                 spin_lock_bh(&nf_conntrack_expect_lock);
2977                 if (del_timer(&exp->timeout)) {
2978                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2979                                                    nlmsg_report(nlh));
2980                         nf_ct_expect_put(exp);
2981                 }
2982                 spin_unlock_bh(&nf_conntrack_expect_lock);
2983                 /* have to put what we 'get' above.
2984                  * after this line usage count == 0 */
2985                 nf_ct_expect_put(exp);
2986         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2987                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2988
2989                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
2990                                          NETLINK_CB(skb).portid,
2991                                          nlmsg_report(nlh));
2992         } else {
2993                 /* This basically means we have to flush everything*/
2994                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
2995                                          NETLINK_CB(skb).portid,
2996                                          nlmsg_report(nlh));
2997         }
2998
2999         return 0;
3000 }
3001 static int
3002 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3003                         const struct nlattr * const cda[])
3004 {
3005         if (cda[CTA_EXPECT_TIMEOUT]) {
3006                 if (!del_timer(&x->timeout))
3007                         return -ETIME;
3008
3009                 x->timeout.expires = jiffies +
3010                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3011                 add_timer(&x->timeout);
3012         }
3013         return 0;
3014 }
3015
3016 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3017         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3018         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3019 };
3020
3021 static int
3022 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3023                            struct nf_conntrack_expect *exp,
3024                            u_int8_t u3)
3025 {
3026 #ifdef CONFIG_NF_NAT_NEEDED
3027         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3028         struct nf_conntrack_tuple nat_tuple = {};
3029         int err;
3030
3031         err = nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr,
3032                                exp_nat_nla_policy, NULL);
3033         if (err < 0)
3034                 return err;
3035
3036         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3037                 return -EINVAL;
3038
3039         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3040                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3041                                     u3, NULL);
3042         if (err < 0)
3043                 return err;
3044
3045         exp->saved_addr = nat_tuple.src.u3;
3046         exp->saved_proto = nat_tuple.src.u;
3047         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3048
3049         return 0;
3050 #else
3051         return -EOPNOTSUPP;
3052 #endif
3053 }
3054
3055 static struct nf_conntrack_expect *
3056 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3057                        struct nf_conntrack_helper *helper,
3058                        struct nf_conntrack_tuple *tuple,
3059                        struct nf_conntrack_tuple *mask)
3060 {
3061         u_int32_t class = 0;
3062         struct nf_conntrack_expect *exp;
3063         struct nf_conn_help *help;
3064         int err;
3065
3066         help = nfct_help(ct);
3067         if (!help)
3068                 return ERR_PTR(-EOPNOTSUPP);
3069
3070         if (cda[CTA_EXPECT_CLASS] && helper) {
3071                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3072                 if (class > helper->expect_class_max)
3073                         return ERR_PTR(-EINVAL);
3074         }
3075         exp = nf_ct_expect_alloc(ct);
3076         if (!exp)
3077                 return ERR_PTR(-ENOMEM);
3078
3079         if (cda[CTA_EXPECT_FLAGS]) {
3080                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3081                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3082         } else {
3083                 exp->flags = 0;
3084         }
3085         if (cda[CTA_EXPECT_FN]) {
3086                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3087                 struct nf_ct_helper_expectfn *expfn;
3088
3089                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3090                 if (expfn == NULL) {
3091                         err = -EINVAL;
3092                         goto err_out;
3093                 }
3094                 exp->expectfn = expfn->expectfn;
3095         } else
3096                 exp->expectfn = NULL;
3097
3098         exp->class = class;
3099         exp->master = ct;
3100         exp->helper = helper;
3101         exp->tuple = *tuple;
3102         exp->mask.src.u3 = mask->src.u3;
3103         exp->mask.src.u.all = mask->src.u.all;
3104
3105         if (cda[CTA_EXPECT_NAT]) {
3106                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3107                                                  exp, nf_ct_l3num(ct));
3108                 if (err < 0)
3109                         goto err_out;
3110         }
3111         return exp;
3112 err_out:
3113         nf_ct_expect_put(exp);
3114         return ERR_PTR(err);
3115 }
3116
3117 static int
3118 ctnetlink_create_expect(struct net *net,
3119                         const struct nf_conntrack_zone *zone,
3120                         const struct nlattr * const cda[],
3121                         u_int8_t u3, u32 portid, int report)
3122 {
3123         struct nf_conntrack_tuple tuple, mask, master_tuple;
3124         struct nf_conntrack_tuple_hash *h = NULL;
3125         struct nf_conntrack_helper *helper = NULL;
3126         struct nf_conntrack_expect *exp;
3127         struct nf_conn *ct;
3128         int err;
3129
3130         /* caller guarantees that those three CTA_EXPECT_* exist */
3131         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3132                                     u3, NULL);
3133         if (err < 0)
3134                 return err;
3135         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3136                                     u3, NULL);
3137         if (err < 0)
3138                 return err;
3139         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3140                                     u3, NULL);
3141         if (err < 0)
3142                 return err;
3143
3144         /* Look for master conntrack of this expectation */
3145         h = nf_conntrack_find_get(net, zone, &master_tuple);
3146         if (!h)
3147                 return -ENOENT;
3148         ct = nf_ct_tuplehash_to_ctrack(h);
3149
3150         rcu_read_lock();
3151         if (cda[CTA_EXPECT_HELP_NAME]) {
3152                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3153
3154                 helper = __nf_conntrack_helper_find(helpname, u3,
3155                                                     nf_ct_protonum(ct));
3156                 if (helper == NULL) {
3157                         rcu_read_unlock();
3158 #ifdef CONFIG_MODULES
3159                         if (request_module("nfct-helper-%s", helpname) < 0) {
3160                                 err = -EOPNOTSUPP;
3161                                 goto err_ct;
3162                         }
3163                         rcu_read_lock();
3164                         helper = __nf_conntrack_helper_find(helpname, u3,
3165                                                             nf_ct_protonum(ct));
3166                         if (helper) {
3167                                 err = -EAGAIN;
3168                                 goto err_rcu;
3169                         }
3170                         rcu_read_unlock();
3171 #endif
3172                         err = -EOPNOTSUPP;
3173                         goto err_ct;
3174                 }
3175         }
3176
3177         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3178         if (IS_ERR(exp)) {
3179                 err = PTR_ERR(exp);
3180                 goto err_rcu;
3181         }
3182
3183         err = nf_ct_expect_related_report(exp, portid, report);
3184         nf_ct_expect_put(exp);
3185 err_rcu:
3186         rcu_read_unlock();
3187 err_ct:
3188         nf_ct_put(ct);
3189         return err;
3190 }
3191
3192 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3193                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3194                                 const struct nlattr * const cda[],
3195                                 struct netlink_ext_ack *extack)
3196 {
3197         struct nf_conntrack_tuple tuple;
3198         struct nf_conntrack_expect *exp;
3199         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3200         u_int8_t u3 = nfmsg->nfgen_family;
3201         struct nf_conntrack_zone zone;
3202         int err;
3203
3204         if (!cda[CTA_EXPECT_TUPLE]
3205             || !cda[CTA_EXPECT_MASK]
3206             || !cda[CTA_EXPECT_MASTER])
3207                 return -EINVAL;
3208
3209         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3210         if (err < 0)
3211                 return err;
3212
3213         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3214                                     u3, NULL);
3215         if (err < 0)
3216                 return err;
3217
3218         spin_lock_bh(&nf_conntrack_expect_lock);
3219         exp = __nf_ct_expect_find(net, &zone, &tuple);
3220         if (!exp) {
3221                 spin_unlock_bh(&nf_conntrack_expect_lock);
3222                 err = -ENOENT;
3223                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3224                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3225                                                       NETLINK_CB(skb).portid,
3226                                                       nlmsg_report(nlh));
3227                 }
3228                 return err;
3229         }
3230
3231         err = -EEXIST;
3232         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3233                 err = ctnetlink_change_expect(exp, cda);
3234         spin_unlock_bh(&nf_conntrack_expect_lock);
3235
3236         return err;
3237 }
3238
3239 static int
3240 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3241                              const struct ip_conntrack_stat *st)
3242 {
3243         struct nlmsghdr *nlh;
3244         struct nfgenmsg *nfmsg;
3245         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3246
3247         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3248                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3249         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3250         if (nlh == NULL)
3251                 goto nlmsg_failure;
3252
3253         nfmsg = nlmsg_data(nlh);
3254         nfmsg->nfgen_family = AF_UNSPEC;
3255         nfmsg->version      = NFNETLINK_V0;
3256         nfmsg->res_id       = htons(cpu);
3257
3258         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3259             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3260             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3261                 goto nla_put_failure;
3262
3263         nlmsg_end(skb, nlh);
3264         return skb->len;
3265
3266 nla_put_failure:
3267 nlmsg_failure:
3268         nlmsg_cancel(skb, nlh);
3269         return -1;
3270 }
3271
3272 static int
3273 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3274 {
3275         int cpu;
3276         struct net *net = sock_net(skb->sk);
3277
3278         if (cb->args[0] == nr_cpu_ids)
3279                 return 0;
3280
3281         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3282                 const struct ip_conntrack_stat *st;
3283
3284                 if (!cpu_possible(cpu))
3285                         continue;
3286
3287                 st = per_cpu_ptr(net->ct.stat, cpu);
3288                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3289                                                  cb->nlh->nlmsg_seq,
3290                                                  cpu, st) < 0)
3291                         break;
3292         }
3293         cb->args[0] = cpu;
3294
3295         return skb->len;
3296 }
3297
3298 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3299                                   struct sk_buff *skb,
3300                                   const struct nlmsghdr *nlh,
3301                                   const struct nlattr * const cda[],
3302                                   struct netlink_ext_ack *extack)
3303 {
3304         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3305                 struct netlink_dump_control c = {
3306                         .dump = ctnetlink_exp_stat_cpu_dump,
3307                 };
3308                 return netlink_dump_start(ctnl, skb, nlh, &c);
3309         }
3310
3311         return 0;
3312 }
3313
3314 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3315 static struct nf_ct_event_notifier ctnl_notifier = {
3316         .fcn = ctnetlink_conntrack_event,
3317 };
3318
3319 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3320         .fcn = ctnetlink_expect_event,
3321 };
3322 #endif
3323
3324 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3325         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3326                                             .attr_count = CTA_MAX,
3327                                             .policy = ct_nla_policy },
3328         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3329                                             .attr_count = CTA_MAX,
3330                                             .policy = ct_nla_policy },
3331         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3332                                             .attr_count = CTA_MAX,
3333                                             .policy = ct_nla_policy },
3334         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3335                                             .attr_count = CTA_MAX,
3336                                             .policy = ct_nla_policy },
3337         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3338         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3339         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3340         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3341 };
3342
3343 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3344         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3345                                             .attr_count = CTA_EXPECT_MAX,
3346                                             .policy = exp_nla_policy },
3347         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3348                                             .attr_count = CTA_EXPECT_MAX,
3349                                             .policy = exp_nla_policy },
3350         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3351                                             .attr_count = CTA_EXPECT_MAX,
3352                                             .policy = exp_nla_policy },
3353         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3354 };
3355
3356 static const struct nfnetlink_subsystem ctnl_subsys = {
3357         .name                           = "conntrack",
3358         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3359         .cb_count                       = IPCTNL_MSG_MAX,
3360         .cb                             = ctnl_cb,
3361 };
3362
3363 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3364         .name                           = "conntrack_expect",
3365         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3366         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3367         .cb                             = ctnl_exp_cb,
3368 };
3369
3370 MODULE_ALIAS("ip_conntrack_netlink");
3371 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3372 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3373
3374 static int __net_init ctnetlink_net_init(struct net *net)
3375 {
3376 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3377         int ret;
3378
3379         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3380         if (ret < 0) {
3381                 pr_err("ctnetlink_init: cannot register notifier.\n");
3382                 goto err_out;
3383         }
3384
3385         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3386         if (ret < 0) {
3387                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3388                 goto err_unreg_notifier;
3389         }
3390 #endif
3391         return 0;
3392
3393 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3394 err_unreg_notifier:
3395         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3396 err_out:
3397         return ret;
3398 #endif
3399 }
3400
3401 static void ctnetlink_net_exit(struct net *net)
3402 {
3403 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3404         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3405         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3406 #endif
3407 }
3408
3409 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3410 {
3411         struct net *net;
3412
3413         list_for_each_entry(net, net_exit_list, exit_list)
3414                 ctnetlink_net_exit(net);
3415 }
3416
3417 static struct pernet_operations ctnetlink_net_ops = {
3418         .init           = ctnetlink_net_init,
3419         .exit_batch     = ctnetlink_net_exit_batch,
3420 };
3421
3422 static int __init ctnetlink_init(void)
3423 {
3424         int ret;
3425
3426         ret = nfnetlink_subsys_register(&ctnl_subsys);
3427         if (ret < 0) {
3428                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3429                 goto err_out;
3430         }
3431
3432         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3433         if (ret < 0) {
3434                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3435                 goto err_unreg_subsys;
3436         }
3437
3438         ret = register_pernet_subsys(&ctnetlink_net_ops);
3439         if (ret < 0) {
3440                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3441                 goto err_unreg_exp_subsys;
3442         }
3443 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3444         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3445         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3446 #endif
3447         return 0;
3448
3449 err_unreg_exp_subsys:
3450         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3451 err_unreg_subsys:
3452         nfnetlink_subsys_unregister(&ctnl_subsys);
3453 err_out:
3454         return ret;
3455 }
3456
3457 static void __exit ctnetlink_exit(void)
3458 {
3459         unregister_pernet_subsys(&ctnetlink_net_ops);
3460         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3461         nfnetlink_subsys_unregister(&ctnl_subsys);
3462 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3463         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3464 #endif
3465         synchronize_rcu();
3466 }
3467
3468 module_init(ctnetlink_init);
3469 module_exit(ctnetlink_exit);