linux/net/netfilter/nf_conntrack_netlink.c
<<
>>
Prefs
   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#include <linux/siphash.h>
  33
  34#include <linux/netfilter.h>
  35#include <net/netlink.h>
  36#include <net/sock.h>
  37#include <net/netfilter/nf_conntrack.h>
  38#include <net/netfilter/nf_conntrack_core.h>
  39#include <net/netfilter/nf_conntrack_expect.h>
  40#include <net/netfilter/nf_conntrack_helper.h>
  41#include <net/netfilter/nf_conntrack_seqadj.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#if IS_ENABLED(CONFIG_NF_NAT)
  50#include <net/netfilter/nf_nat.h>
  51#include <net/netfilter/nf_nat_helper.h>
  52#endif
  53
  54#include <linux/netfilter/nfnetlink.h>
  55#include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57MODULE_LICENSE("GPL");
  58
  59static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  60                                const struct nf_conntrack_tuple *tuple,
  61                                const struct nf_conntrack_l4proto *l4proto)
  62{
  63        int ret = 0;
  64        struct nlattr *nest_parms;
  65
  66        nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO);
  67        if (!nest_parms)
  68                goto nla_put_failure;
  69        if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  70                goto nla_put_failure;
  71
  72        if (likely(l4proto->tuple_to_nlattr))
  73                ret = l4proto->tuple_to_nlattr(skb, tuple);
  74
  75        nla_nest_end(skb, nest_parms);
  76
  77        return ret;
  78
  79nla_put_failure:
  80        return -1;
  81}
  82
  83static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  84                                const struct nf_conntrack_tuple *tuple)
  85{
  86        if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  87            nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  88                return -EMSGSIZE;
  89        return 0;
  90}
  91
  92static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
  93                                const struct nf_conntrack_tuple *tuple)
  94{
  95        if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
  96            nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
  97                return -EMSGSIZE;
  98        return 0;
  99}
 100
 101static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 102                                    const struct nf_conntrack_tuple *tuple)
 103{
 104        int ret = 0;
 105        struct nlattr *nest_parms;
 106
 107        nest_parms = nla_nest_start(skb, CTA_TUPLE_IP);
 108        if (!nest_parms)
 109                goto nla_put_failure;
 110
 111        switch (tuple->src.l3num) {
 112        case NFPROTO_IPV4:
 113                ret = ipv4_tuple_to_nlattr(skb, tuple);
 114                break;
 115        case NFPROTO_IPV6:
 116                ret = ipv6_tuple_to_nlattr(skb, tuple);
 117                break;
 118        }
 119
 120        nla_nest_end(skb, nest_parms);
 121
 122        return ret;
 123
 124nla_put_failure:
 125        return -1;
 126}
 127
 128static int ctnetlink_dump_tuples(struct sk_buff *skb,
 129                                 const struct nf_conntrack_tuple *tuple)
 130{
 131        const struct nf_conntrack_l4proto *l4proto;
 132        int ret;
 133
 134        rcu_read_lock();
 135        ret = ctnetlink_dump_tuples_ip(skb, tuple);
 136
 137        if (ret >= 0) {
 138                l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 139                ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 140        }
 141        rcu_read_unlock();
 142        return ret;
 143}
 144
 145static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 146                                  const struct nf_conntrack_zone *zone, int dir)
 147{
 148        if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 149                return 0;
 150        if (nla_put_be16(skb, attrtype, htons(zone->id)))
 151                goto nla_put_failure;
 152        return 0;
 153
 154nla_put_failure:
 155        return -1;
 156}
 157
 158static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 159{
 160        if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 161                goto nla_put_failure;
 162        return 0;
 163
 164nla_put_failure:
 165        return -1;
 166}
 167
 168static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 169{
 170        long timeout = nf_ct_expires(ct) / HZ;
 171
 172        if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 173                goto nla_put_failure;
 174        return 0;
 175
 176nla_put_failure:
 177        return -1;
 178}
 179
 180static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 181{
 182        const struct nf_conntrack_l4proto *l4proto;
 183        struct nlattr *nest_proto;
 184        int ret;
 185
 186        l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 187        if (!l4proto->to_nlattr)
 188                return 0;
 189
 190        nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 191        if (!nest_proto)
 192                goto nla_put_failure;
 193
 194        ret = l4proto->to_nlattr(skb, nest_proto, ct);
 195
 196        nla_nest_end(skb, nest_proto);
 197
 198        return ret;
 199
 200nla_put_failure:
 201        return -1;
 202}
 203
 204static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 205                                   const struct nf_conn *ct)
 206{
 207        struct nlattr *nest_helper;
 208        const struct nf_conn_help *help = nfct_help(ct);
 209        struct nf_conntrack_helper *helper;
 210
 211        if (!help)
 212                return 0;
 213
 214        helper = rcu_dereference(help->helper);
 215        if (!helper)
 216                goto out;
 217
 218        nest_helper = nla_nest_start(skb, CTA_HELP);
 219        if (!nest_helper)
 220                goto nla_put_failure;
 221        if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 222                goto nla_put_failure;
 223
 224        if (helper->to_nlattr)
 225                helper->to_nlattr(skb, ct);
 226
 227        nla_nest_end(skb, nest_helper);
 228out:
 229        return 0;
 230
 231nla_put_failure:
 232        return -1;
 233}
 234
 235static int
 236dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 237              enum ip_conntrack_dir dir, int type)
 238{
 239        enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 240        struct nf_conn_counter *counter = acct->counter;
 241        struct nlattr *nest_count;
 242        u64 pkts, bytes;
 243
 244        if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 245                pkts = atomic64_xchg(&counter[dir].packets, 0);
 246                bytes = atomic64_xchg(&counter[dir].bytes, 0);
 247        } else {
 248                pkts = atomic64_read(&counter[dir].packets);
 249                bytes = atomic64_read(&counter[dir].bytes);
 250        }
 251
 252        nest_count = nla_nest_start(skb, attr);
 253        if (!nest_count)
 254                goto nla_put_failure;
 255
 256        if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 257                         CTA_COUNTERS_PAD) ||
 258            nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 259                         CTA_COUNTERS_PAD))
 260                goto nla_put_failure;
 261
 262        nla_nest_end(skb, nest_count);
 263
 264        return 0;
 265
 266nla_put_failure:
 267        return -1;
 268}
 269
 270static int
 271ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 272{
 273        struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 274
 275        if (!acct)
 276                return 0;
 277
 278        if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 279                return -1;
 280        if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 281                return -1;
 282
 283        return 0;
 284}
 285
 286static int
 287ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 288{
 289        struct nlattr *nest_count;
 290        const struct nf_conn_tstamp *tstamp;
 291
 292        tstamp = nf_conn_tstamp_find(ct);
 293        if (!tstamp)
 294                return 0;
 295
 296        nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 297        if (!nest_count)
 298                goto nla_put_failure;
 299
 300        if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 301                         CTA_TIMESTAMP_PAD) ||
 302            (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 303                                               cpu_to_be64(tstamp->stop),
 304                                               CTA_TIMESTAMP_PAD)))
 305                goto nla_put_failure;
 306        nla_nest_end(skb, nest_count);
 307
 308        return 0;
 309
 310nla_put_failure:
 311        return -1;
 312}
 313
 314#ifdef CONFIG_NF_CONNTRACK_MARK
 315static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 316{
 317        if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 318                goto nla_put_failure;
 319        return 0;
 320
 321nla_put_failure:
 322        return -1;
 323}
 324#else
 325#define ctnetlink_dump_mark(a, b) (0)
 326#endif
 327
 328#ifdef CONFIG_NF_CONNTRACK_SECMARK
 329static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 330{
 331        struct nlattr *nest_secctx;
 332        int len, ret;
 333        char *secctx;
 334
 335        ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 336        if (ret)
 337                return 0;
 338
 339        ret = -1;
 340        nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 341        if (!nest_secctx)
 342                goto nla_put_failure;
 343
 344        if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 345                goto nla_put_failure;
 346        nla_nest_end(skb, nest_secctx);
 347
 348        ret = 0;
 349nla_put_failure:
 350        security_release_secctx(secctx, len);
 351        return ret;
 352}
 353#else
 354#define ctnetlink_dump_secctx(a, b) (0)
 355#endif
 356
 357#ifdef CONFIG_NF_CONNTRACK_LABELS
 358static inline int ctnetlink_label_size(const struct nf_conn *ct)
 359{
 360        struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 361
 362        if (!labels)
 363                return 0;
 364        return nla_total_size(sizeof(labels->bits));
 365}
 366
 367static int
 368ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 369{
 370        struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 371        unsigned int i;
 372
 373        if (!labels)
 374                return 0;
 375
 376        i = 0;
 377        do {
 378                if (labels->bits[i] != 0)
 379                        return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 380                                       labels->bits);
 381                i++;
 382        } while (i < ARRAY_SIZE(labels->bits));
 383
 384        return 0;
 385}
 386#else
 387#define ctnetlink_dump_labels(a, b) (0)
 388#define ctnetlink_label_size(a) (0)
 389#endif
 390
 391#define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 392
 393static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 394{
 395        struct nlattr *nest_parms;
 396
 397        if (!(ct->status & IPS_EXPECTED))
 398                return 0;
 399
 400        nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 401        if (!nest_parms)
 402                goto nla_put_failure;
 403        if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 404                goto nla_put_failure;
 405        nla_nest_end(skb, nest_parms);
 406
 407        return 0;
 408
 409nla_put_failure:
 410        return -1;
 411}
 412
 413static int
 414dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 415{
 416        struct nlattr *nest_parms;
 417
 418        nest_parms = nla_nest_start(skb, type);
 419        if (!nest_parms)
 420                goto nla_put_failure;
 421
 422        if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 423                         htonl(seq->correction_pos)) ||
 424            nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 425                         htonl(seq->offset_before)) ||
 426            nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 427                         htonl(seq->offset_after)))
 428                goto nla_put_failure;
 429
 430        nla_nest_end(skb, nest_parms);
 431
 432        return 0;
 433
 434nla_put_failure:
 435        return -1;
 436}
 437
 438static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 439{
 440        struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 441        struct nf_ct_seqadj *seq;
 442
 443        if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 444                return 0;
 445
 446        spin_lock_bh(&ct->lock);
 447        seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 448        if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 449                goto err;
 450
 451        seq = &seqadj->seq[IP_CT_DIR_REPLY];
 452        if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 453                goto err;
 454
 455        spin_unlock_bh(&ct->lock);
 456        return 0;
 457err:
 458        spin_unlock_bh(&ct->lock);
 459        return -1;
 460}
 461
 462static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 463{
 464        struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 465        struct nlattr *nest_parms;
 466
 467        if (!synproxy)
 468                return 0;
 469
 470        nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 471        if (!nest_parms)
 472                goto nla_put_failure;
 473
 474        if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 475            nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 476            nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 477                goto nla_put_failure;
 478
 479        nla_nest_end(skb, nest_parms);
 480
 481        return 0;
 482
 483nla_put_failure:
 484        return -1;
 485}
 486
 487static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 488{
 489        __be32 id = (__force __be32)nf_ct_get_id(ct);
 490
 491        if (nla_put_be32(skb, CTA_ID, id))
 492                goto nla_put_failure;
 493        return 0;
 494
 495nla_put_failure:
 496        return -1;
 497}
 498
 499static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 500{
 501        if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 502                goto nla_put_failure;
 503        return 0;
 504
 505nla_put_failure:
 506        return -1;
 507}
 508
 509static int
 510ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 511                    struct nf_conn *ct)
 512{
 513        const struct nf_conntrack_zone *zone;
 514        struct nlmsghdr *nlh;
 515        struct nfgenmsg *nfmsg;
 516        struct nlattr *nest_parms;
 517        unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 518
 519        event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 520        nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 521        if (nlh == NULL)
 522                goto nlmsg_failure;
 523
 524        nfmsg = nlmsg_data(nlh);
 525        nfmsg->nfgen_family = nf_ct_l3num(ct);
 526        nfmsg->version      = NFNETLINK_V0;
 527        nfmsg->res_id       = 0;
 528
 529        zone = nf_ct_zone(ct);
 530
 531        nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 532        if (!nest_parms)
 533                goto nla_put_failure;
 534        if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 535                goto nla_put_failure;
 536        if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 537                                   NF_CT_ZONE_DIR_ORIG) < 0)
 538                goto nla_put_failure;
 539        nla_nest_end(skb, nest_parms);
 540
 541        nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 542        if (!nest_parms)
 543                goto nla_put_failure;
 544        if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 545                goto nla_put_failure;
 546        if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 547                                   NF_CT_ZONE_DIR_REPL) < 0)
 548                goto nla_put_failure;
 549        nla_nest_end(skb, nest_parms);
 550
 551        if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 552                                   NF_CT_DEFAULT_ZONE_DIR) < 0)
 553                goto nla_put_failure;
 554
 555        if (ctnetlink_dump_status(skb, ct) < 0 ||
 556            ctnetlink_dump_acct(skb, ct, type) < 0 ||
 557            ctnetlink_dump_timestamp(skb, ct) < 0 ||
 558            ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 559            ctnetlink_dump_mark(skb, ct) < 0 ||
 560            ctnetlink_dump_secctx(skb, ct) < 0 ||
 561            ctnetlink_dump_labels(skb, ct) < 0 ||
 562            ctnetlink_dump_id(skb, ct) < 0 ||
 563            ctnetlink_dump_use(skb, ct) < 0 ||
 564            ctnetlink_dump_master(skb, ct) < 0 ||
 565            ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 566            ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 567                goto nla_put_failure;
 568
 569        if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 570            (ctnetlink_dump_timeout(skb, ct) < 0 ||
 571             ctnetlink_dump_protoinfo(skb, ct) < 0))
 572                goto nla_put_failure;
 573
 574        nlmsg_end(skb, nlh);
 575        return skb->len;
 576
 577nlmsg_failure:
 578nla_put_failure:
 579        nlmsg_cancel(skb, nlh);
 580        return -1;
 581}
 582
 583static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 584        [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 585        [CTA_IP_V4_DST] = { .type = NLA_U32 },
 586        [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 587        [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 588};
 589
 590#if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 591static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 592{
 593        const struct nf_conntrack_l4proto *l4proto;
 594        size_t len, len4 = 0;
 595
 596        len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 597        len *= 3u; /* ORIG, REPLY, MASTER */
 598
 599        l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 600        len += l4proto->nlattr_size;
 601        if (l4proto->nlattr_tuple_size) {
 602                len4 = l4proto->nlattr_tuple_size();
 603                len4 *= 3u; /* ORIG, REPLY, MASTER */
 604        }
 605
 606        return len + len4;
 607}
 608#endif
 609
 610static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 611{
 612        if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 613                return 0;
 614        return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 615               + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 616               + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 617               ;
 618}
 619
 620static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 621{
 622#ifdef CONFIG_NF_CONNTRACK_SECMARK
 623        int len, ret;
 624
 625        ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 626        if (ret)
 627                return 0;
 628
 629        return nla_total_size(0) /* CTA_SECCTX */
 630               + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 631#else
 632        return 0;
 633#endif
 634}
 635
 636static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 637{
 638#ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 639        if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 640                return 0;
 641        return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 642#else
 643        return 0;
 644#endif
 645}
 646
 647#ifdef CONFIG_NF_CONNTRACK_EVENTS
 648static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 649{
 650        return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 651               + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 652               + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 653               + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 654               + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 655               + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 656               + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 657               + ctnetlink_acct_size(ct)
 658               + ctnetlink_timestamp_size(ct)
 659               + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 660               + nla_total_size(0) /* CTA_PROTOINFO */
 661               + nla_total_size(0) /* CTA_HELP */
 662               + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 663               + ctnetlink_secctx_size(ct)
 664#if IS_ENABLED(CONFIG_NF_NAT)
 665               + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 666               + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 667#endif
 668#ifdef CONFIG_NF_CONNTRACK_MARK
 669               + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 670#endif
 671#ifdef CONFIG_NF_CONNTRACK_ZONES
 672               + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 673#endif
 674               + ctnetlink_proto_size(ct)
 675               + ctnetlink_label_size(ct)
 676               ;
 677}
 678
 679static int
 680ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 681{
 682        const struct nf_conntrack_zone *zone;
 683        struct net *net;
 684        struct nlmsghdr *nlh;
 685        struct nfgenmsg *nfmsg;
 686        struct nlattr *nest_parms;
 687        struct nf_conn *ct = item->ct;
 688        struct sk_buff *skb;
 689        unsigned int type;
 690        unsigned int flags = 0, group;
 691        int err;
 692
 693        if (events & (1 << IPCT_DESTROY)) {
 694                type = IPCTNL_MSG_CT_DELETE;
 695                group = NFNLGRP_CONNTRACK_DESTROY;
 696        } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 697                type = IPCTNL_MSG_CT_NEW;
 698                flags = NLM_F_CREATE|NLM_F_EXCL;
 699                group = NFNLGRP_CONNTRACK_NEW;
 700        } else if (events) {
 701                type = IPCTNL_MSG_CT_NEW;
 702                group = NFNLGRP_CONNTRACK_UPDATE;
 703        } else
 704                return 0;
 705
 706        net = nf_ct_net(ct);
 707        if (!item->report && !nfnetlink_has_listeners(net, group))
 708                return 0;
 709
 710        skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 711        if (skb == NULL)
 712                goto errout;
 713
 714        type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 715        nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 716        if (nlh == NULL)
 717                goto nlmsg_failure;
 718
 719        nfmsg = nlmsg_data(nlh);
 720        nfmsg->nfgen_family = nf_ct_l3num(ct);
 721        nfmsg->version  = NFNETLINK_V0;
 722        nfmsg->res_id   = 0;
 723
 724        zone = nf_ct_zone(ct);
 725
 726        nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 727        if (!nest_parms)
 728                goto nla_put_failure;
 729        if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 730                goto nla_put_failure;
 731        if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 732                                   NF_CT_ZONE_DIR_ORIG) < 0)
 733                goto nla_put_failure;
 734        nla_nest_end(skb, nest_parms);
 735
 736        nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 737        if (!nest_parms)
 738                goto nla_put_failure;
 739        if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 740                goto nla_put_failure;
 741        if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 742                                   NF_CT_ZONE_DIR_REPL) < 0)
 743                goto nla_put_failure;
 744        nla_nest_end(skb, nest_parms);
 745
 746        if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 747                                   NF_CT_DEFAULT_ZONE_DIR) < 0)
 748                goto nla_put_failure;
 749
 750        if (ctnetlink_dump_id(skb, ct) < 0)
 751                goto nla_put_failure;
 752
 753        if (ctnetlink_dump_status(skb, ct) < 0)
 754                goto nla_put_failure;
 755
 756        if (events & (1 << IPCT_DESTROY)) {
 757                if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 758                    ctnetlink_dump_timestamp(skb, ct) < 0)
 759                        goto nla_put_failure;
 760        } else {
 761                if (ctnetlink_dump_timeout(skb, ct) < 0)
 762                        goto nla_put_failure;
 763
 764                if (events & (1 << IPCT_PROTOINFO)
 765                    && ctnetlink_dump_protoinfo(skb, ct) < 0)
 766                        goto nla_put_failure;
 767
 768                if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 769                    && ctnetlink_dump_helpinfo(skb, ct) < 0)
 770                        goto nla_put_failure;
 771
 772#ifdef CONFIG_NF_CONNTRACK_SECMARK
 773                if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 774                    && ctnetlink_dump_secctx(skb, ct) < 0)
 775                        goto nla_put_failure;
 776#endif
 777                if (events & (1 << IPCT_LABEL) &&
 778                     ctnetlink_dump_labels(skb, ct) < 0)
 779                        goto nla_put_failure;
 780
 781                if (events & (1 << IPCT_RELATED) &&
 782                    ctnetlink_dump_master(skb, ct) < 0)
 783                        goto nla_put_failure;
 784
 785                if (events & (1 << IPCT_SEQADJ) &&
 786                    ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 787                        goto nla_put_failure;
 788
 789                if (events & (1 << IPCT_SYNPROXY) &&
 790                    ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 791                        goto nla_put_failure;
 792        }
 793
 794#ifdef CONFIG_NF_CONNTRACK_MARK
 795        if ((events & (1 << IPCT_MARK) || ct->mark)
 796            && ctnetlink_dump_mark(skb, ct) < 0)
 797                goto nla_put_failure;
 798#endif
 799        nlmsg_end(skb, nlh);
 800        err = nfnetlink_send(skb, net, item->portid, group, item->report,
 801                             GFP_ATOMIC);
 802        if (err == -ENOBUFS || err == -EAGAIN)
 803                return -ENOBUFS;
 804
 805        return 0;
 806
 807nla_put_failure:
 808        nlmsg_cancel(skb, nlh);
 809nlmsg_failure:
 810        kfree_skb(skb);
 811errout:
 812        if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 813                return -ENOBUFS;
 814
 815        return 0;
 816}
 817#endif /* CONFIG_NF_CONNTRACK_EVENTS */
 818
 819static int ctnetlink_done(struct netlink_callback *cb)
 820{
 821        if (cb->args[1])
 822                nf_ct_put((struct nf_conn *)cb->args[1]);
 823        kfree(cb->data);
 824        return 0;
 825}
 826
 827struct ctnetlink_filter {
 828        u8 family;
 829        struct {
 830                u_int32_t val;
 831                u_int32_t mask;
 832        } mark;
 833};
 834
 835static struct ctnetlink_filter *
 836ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 837{
 838        struct ctnetlink_filter *filter;
 839
 840#ifndef CONFIG_NF_CONNTRACK_MARK
 841        if (cda[CTA_MARK] && cda[CTA_MARK_MASK])
 842                return ERR_PTR(-EOPNOTSUPP);
 843#endif
 844
 845        filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 846        if (filter == NULL)
 847                return ERR_PTR(-ENOMEM);
 848
 849        filter->family = family;
 850
 851#ifdef CONFIG_NF_CONNTRACK_MARK
 852        if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
 853                filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 854                filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 855        }
 856#endif
 857        return filter;
 858}
 859
 860static int ctnetlink_start(struct netlink_callback *cb)
 861{
 862        const struct nlattr * const *cda = cb->data;
 863        struct ctnetlink_filter *filter = NULL;
 864        struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 865        u8 family = nfmsg->nfgen_family;
 866
 867        if (family || (cda[CTA_MARK] && cda[CTA_MARK_MASK])) {
 868                filter = ctnetlink_alloc_filter(cda, family);
 869                if (IS_ERR(filter))
 870                        return PTR_ERR(filter);
 871        }
 872
 873        cb->data = filter;
 874        return 0;
 875}
 876
 877static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
 878{
 879        struct ctnetlink_filter *filter = data;
 880
 881        if (filter == NULL)
 882                goto out;
 883
 884        /* Match entries of a given L3 protocol number.
 885         * If it is not specified, ie. l3proto == 0,
 886         * then match everything.
 887         */
 888        if (filter->family && nf_ct_l3num(ct) != filter->family)
 889                goto ignore_entry;
 890
 891#ifdef CONFIG_NF_CONNTRACK_MARK
 892        if ((ct->mark & filter->mark.mask) != filter->mark.val)
 893                goto ignore_entry;
 894#endif
 895
 896out:
 897        return 1;
 898
 899ignore_entry:
 900        return 0;
 901}
 902
 903static int
 904ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 905{
 906        struct net *net = sock_net(skb->sk);
 907        struct nf_conn *ct, *last;
 908        struct nf_conntrack_tuple_hash *h;
 909        struct hlist_nulls_node *n;
 910        struct nf_conn *nf_ct_evict[8];
 911        int res, i;
 912        spinlock_t *lockp;
 913
 914        last = (struct nf_conn *)cb->args[1];
 915        i = 0;
 916
 917        local_bh_disable();
 918        for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
 919restart:
 920                while (i) {
 921                        i--;
 922                        if (nf_ct_should_gc(nf_ct_evict[i]))
 923                                nf_ct_kill(nf_ct_evict[i]);
 924                        nf_ct_put(nf_ct_evict[i]);
 925                }
 926
 927                lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
 928                nf_conntrack_lock(lockp);
 929                if (cb->args[0] >= nf_conntrack_htable_size) {
 930                        spin_unlock(lockp);
 931                        goto out;
 932                }
 933                hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
 934                                           hnnode) {
 935                        if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 936                                continue;
 937                        ct = nf_ct_tuplehash_to_ctrack(h);
 938                        if (nf_ct_is_expired(ct)) {
 939                                if (i < ARRAY_SIZE(nf_ct_evict) &&
 940                                    atomic_inc_not_zero(&ct->ct_general.use))
 941                                        nf_ct_evict[i++] = ct;
 942                                continue;
 943                        }
 944
 945                        if (!net_eq(net, nf_ct_net(ct)))
 946                                continue;
 947
 948                        if (cb->args[1]) {
 949                                if (ct != last)
 950                                        continue;
 951                                cb->args[1] = 0;
 952                        }
 953                        if (!ctnetlink_filter_match(ct, cb->data))
 954                                continue;
 955
 956                        rcu_read_lock();
 957                        res =
 958                        ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 959                                            cb->nlh->nlmsg_seq,
 960                                            NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 961                                            ct);
 962                        rcu_read_unlock();
 963                        if (res < 0) {
 964                                nf_conntrack_get(&ct->ct_general);
 965                                cb->args[1] = (unsigned long)ct;
 966                                spin_unlock(lockp);
 967                                goto out;
 968                        }
 969                }
 970                spin_unlock(lockp);
 971                if (cb->args[1]) {
 972                        cb->args[1] = 0;
 973                        goto restart;
 974                }
 975        }
 976out:
 977        local_bh_enable();
 978        if (last) {
 979                /* nf ct hash resize happened, now clear the leftover. */
 980                if ((struct nf_conn *)cb->args[1] == last)
 981                        cb->args[1] = 0;
 982
 983                nf_ct_put(last);
 984        }
 985
 986        while (i) {
 987                i--;
 988                if (nf_ct_should_gc(nf_ct_evict[i]))
 989                        nf_ct_kill(nf_ct_evict[i]);
 990                nf_ct_put(nf_ct_evict[i]);
 991        }
 992
 993        return skb->len;
 994}
 995
 996static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
 997                                struct nf_conntrack_tuple *t)
 998{
 999        if (!tb[CTA_IP_V4_SRC] || !tb[CTA_IP_V4_DST])
1000                return -EINVAL;
1001
1002        t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1003        t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1004
1005        return 0;
1006}
1007
1008static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1009                                struct nf_conntrack_tuple *t)
1010{
1011        if (!tb[CTA_IP_V6_SRC] || !tb[CTA_IP_V6_DST])
1012                return -EINVAL;
1013
1014        t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1015        t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1016
1017        return 0;
1018}
1019
1020static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1021                                    struct nf_conntrack_tuple *tuple)
1022{
1023        struct nlattr *tb[CTA_IP_MAX+1];
1024        int ret = 0;
1025
1026        ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1027        if (ret < 0)
1028                return ret;
1029
1030        ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1031                                             cta_ip_nla_policy, NULL);
1032        if (ret)
1033                return ret;
1034
1035        switch (tuple->src.l3num) {
1036        case NFPROTO_IPV4:
1037                ret = ipv4_nlattr_to_tuple(tb, tuple);
1038                break;
1039        case NFPROTO_IPV6:
1040                ret = ipv6_nlattr_to_tuple(tb, tuple);
1041                break;
1042        }
1043
1044        return ret;
1045}
1046
1047static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1048        [CTA_PROTO_NUM] = { .type = NLA_U8 },
1049};
1050
1051static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1052                                       struct nf_conntrack_tuple *tuple)
1053{
1054        const struct nf_conntrack_l4proto *l4proto;
1055        struct nlattr *tb[CTA_PROTO_MAX+1];
1056        int ret = 0;
1057
1058        ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1059                                          proto_nla_policy, NULL);
1060        if (ret < 0)
1061                return ret;
1062
1063        if (!tb[CTA_PROTO_NUM])
1064                return -EINVAL;
1065        tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1066
1067        rcu_read_lock();
1068        l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1069
1070        if (likely(l4proto->nlattr_to_tuple)) {
1071                ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1072                                                     l4proto->nla_policy,
1073                                                     NULL);
1074                if (ret == 0)
1075                        ret = l4proto->nlattr_to_tuple(tb, tuple);
1076        }
1077
1078        rcu_read_unlock();
1079
1080        return ret;
1081}
1082
1083static int
1084ctnetlink_parse_zone(const struct nlattr *attr,
1085                     struct nf_conntrack_zone *zone)
1086{
1087        nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1088                        NF_CT_DEFAULT_ZONE_DIR, 0);
1089#ifdef CONFIG_NF_CONNTRACK_ZONES
1090        if (attr)
1091                zone->id = ntohs(nla_get_be16(attr));
1092#else
1093        if (attr)
1094                return -EOPNOTSUPP;
1095#endif
1096        return 0;
1097}
1098
1099static int
1100ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1101                           struct nf_conntrack_zone *zone)
1102{
1103        int ret;
1104
1105        if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1106                return -EINVAL;
1107
1108        ret = ctnetlink_parse_zone(attr, zone);
1109        if (ret < 0)
1110                return ret;
1111
1112        if (type == CTA_TUPLE_REPLY)
1113                zone->dir = NF_CT_ZONE_DIR_REPL;
1114        else
1115                zone->dir = NF_CT_ZONE_DIR_ORIG;
1116
1117        return 0;
1118}
1119
1120static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1121        [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1122        [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1123        [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1124};
1125
1126static int
1127ctnetlink_parse_tuple(const struct nlattr * const cda[],
1128                      struct nf_conntrack_tuple *tuple, u32 type,
1129                      u_int8_t l3num, struct nf_conntrack_zone *zone)
1130{
1131        struct nlattr *tb[CTA_TUPLE_MAX+1];
1132        int err;
1133
1134        memset(tuple, 0, sizeof(*tuple));
1135
1136        err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1137                                          tuple_nla_policy, NULL);
1138        if (err < 0)
1139                return err;
1140
1141        if (!tb[CTA_TUPLE_IP])
1142                return -EINVAL;
1143
1144        tuple->src.l3num = l3num;
1145
1146        err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
1147        if (err < 0)
1148                return err;
1149
1150        if (!tb[CTA_TUPLE_PROTO])
1151                return -EINVAL;
1152
1153        err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
1154        if (err < 0)
1155                return err;
1156
1157        if (tb[CTA_TUPLE_ZONE]) {
1158                if (!zone)
1159                        return -EINVAL;
1160
1161                err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1162                                                 type, zone);
1163                if (err < 0)
1164                        return err;
1165        }
1166
1167        /* orig and expect tuples get DIR_ORIGINAL */
1168        if (type == CTA_TUPLE_REPLY)
1169                tuple->dst.dir = IP_CT_DIR_REPLY;
1170        else
1171                tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1172
1173        return 0;
1174}
1175
1176static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1177        [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1178                                    .len = NF_CT_HELPER_NAME_LEN - 1 },
1179};
1180
1181static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1182                                struct nlattr **helpinfo)
1183{
1184        int err;
1185        struct nlattr *tb[CTA_HELP_MAX+1];
1186
1187        err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1188                                          help_nla_policy, NULL);
1189        if (err < 0)
1190                return err;
1191
1192        if (!tb[CTA_HELP_NAME])
1193                return -EINVAL;
1194
1195        *helper_name = nla_data(tb[CTA_HELP_NAME]);
1196
1197        if (tb[CTA_HELP_INFO])
1198                *helpinfo = tb[CTA_HELP_INFO];
1199
1200        return 0;
1201}
1202
1203static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1204        [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1205        [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1206        [CTA_STATUS]            = { .type = NLA_U32 },
1207        [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1208        [CTA_HELP]              = { .type = NLA_NESTED },
1209        [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1210        [CTA_TIMEOUT]           = { .type = NLA_U32 },
1211        [CTA_MARK]              = { .type = NLA_U32 },
1212        [CTA_ID]                = { .type = NLA_U32 },
1213        [CTA_NAT_DST]           = { .type = NLA_NESTED },
1214        [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1215        [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1216        [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1217        [CTA_ZONE]              = { .type = NLA_U16 },
1218        [CTA_MARK_MASK]         = { .type = NLA_U32 },
1219        [CTA_LABELS]            = { .type = NLA_BINARY,
1220                                    .len = NF_CT_LABELS_MAX_SIZE },
1221        [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1222                                    .len = NF_CT_LABELS_MAX_SIZE },
1223};
1224
1225static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1226{
1227        if (test_bit(IPS_OFFLOAD_BIT, &ct->status))
1228                return 0;
1229
1230        return ctnetlink_filter_match(ct, data);
1231}
1232
1233static int ctnetlink_flush_conntrack(struct net *net,
1234                                     const struct nlattr * const cda[],
1235                                     u32 portid, int report, u8 family)
1236{
1237        struct ctnetlink_filter *filter = NULL;
1238
1239        if (family || (cda[CTA_MARK] && cda[CTA_MARK_MASK])) {
1240                filter = ctnetlink_alloc_filter(cda, family);
1241                if (IS_ERR(filter))
1242                        return PTR_ERR(filter);
1243        }
1244
1245        nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1246                                  portid, report);
1247        kfree(filter);
1248
1249        return 0;
1250}
1251
1252static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1253                                   struct sk_buff *skb,
1254                                   const struct nlmsghdr *nlh,
1255                                   const struct nlattr * const cda[],
1256                                   struct netlink_ext_ack *extack)
1257{
1258        struct nf_conntrack_tuple_hash *h;
1259        struct nf_conntrack_tuple tuple;
1260        struct nf_conn *ct;
1261        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1262        struct nf_conntrack_zone zone;
1263        int err;
1264
1265        err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1266        if (err < 0)
1267                return err;
1268
1269        if (cda[CTA_TUPLE_ORIG])
1270                err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1271                                            nfmsg->nfgen_family, &zone);
1272        else if (cda[CTA_TUPLE_REPLY])
1273                err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1274                                            nfmsg->nfgen_family, &zone);
1275        else {
1276                u_int8_t u3 = nfmsg->version ? nfmsg->nfgen_family : AF_UNSPEC;
1277
1278                return ctnetlink_flush_conntrack(net, cda,
1279                                                 NETLINK_CB(skb).portid,
1280                                                 nlmsg_report(nlh), u3);
1281        }
1282
1283        if (err < 0)
1284                return err;
1285
1286        h = nf_conntrack_find_get(net, &zone, &tuple);
1287        if (!h)
1288                return -ENOENT;
1289
1290        ct = nf_ct_tuplehash_to_ctrack(h);
1291
1292        if (test_bit(IPS_OFFLOAD_BIT, &ct->status)) {
1293                nf_ct_put(ct);
1294                return -EBUSY;
1295        }
1296
1297        if (cda[CTA_ID]) {
1298                __be32 id = nla_get_be32(cda[CTA_ID]);
1299
1300                if (id != (__force __be32)nf_ct_get_id(ct)) {
1301                        nf_ct_put(ct);
1302                        return -ENOENT;
1303                }
1304        }
1305
1306        nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1307        nf_ct_put(ct);
1308
1309        return 0;
1310}
1311
1312static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1313                                   struct sk_buff *skb,
1314                                   const struct nlmsghdr *nlh,
1315                                   const struct nlattr * const cda[],
1316                                   struct netlink_ext_ack *extack)
1317{
1318        struct nf_conntrack_tuple_hash *h;
1319        struct nf_conntrack_tuple tuple;
1320        struct nf_conn *ct;
1321        struct sk_buff *skb2 = NULL;
1322        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1323        u_int8_t u3 = nfmsg->nfgen_family;
1324        struct nf_conntrack_zone zone;
1325        int err;
1326
1327        if (nlh->nlmsg_flags & NLM_F_DUMP) {
1328                struct netlink_dump_control c = {
1329                        .start = ctnetlink_start,
1330                        .dump = ctnetlink_dump_table,
1331                        .done = ctnetlink_done,
1332                        .data = (void *)cda,
1333                };
1334
1335                return netlink_dump_start(ctnl, skb, nlh, &c);
1336        }
1337
1338        err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1339        if (err < 0)
1340                return err;
1341
1342        if (cda[CTA_TUPLE_ORIG])
1343                err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1344                                            u3, &zone);
1345        else if (cda[CTA_TUPLE_REPLY])
1346                err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1347                                            u3, &zone);
1348        else
1349                return -EINVAL;
1350
1351        if (err < 0)
1352                return err;
1353
1354        h = nf_conntrack_find_get(net, &zone, &tuple);
1355        if (!h)
1356                return -ENOENT;
1357
1358        ct = nf_ct_tuplehash_to_ctrack(h);
1359
1360        err = -ENOMEM;
1361        skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1362        if (skb2 == NULL) {
1363                nf_ct_put(ct);
1364                return -ENOMEM;
1365        }
1366
1367        rcu_read_lock();
1368        err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1369                                  NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1370        rcu_read_unlock();
1371        nf_ct_put(ct);
1372        if (err <= 0)
1373                goto free;
1374
1375        err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1376        if (err < 0)
1377                goto out;
1378
1379        return 0;
1380
1381free:
1382        kfree_skb(skb2);
1383out:
1384        /* this avoids a loop in nfnetlink. */
1385        return err == -EAGAIN ? -ENOBUFS : err;
1386}
1387
1388static int ctnetlink_done_list(struct netlink_callback *cb)
1389{
1390        if (cb->args[1])
1391                nf_ct_put((struct nf_conn *)cb->args[1]);
1392        return 0;
1393}
1394
1395static int
1396ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1397{
1398        struct nf_conn *ct, *last;
1399        struct nf_conntrack_tuple_hash *h;
1400        struct hlist_nulls_node *n;
1401        struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1402        u_int8_t l3proto = nfmsg->nfgen_family;
1403        int res;
1404        int cpu;
1405        struct hlist_nulls_head *list;
1406        struct net *net = sock_net(skb->sk);
1407
1408        if (cb->args[2])
1409                return 0;
1410
1411        last = (struct nf_conn *)cb->args[1];
1412
1413        for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1414                struct ct_pcpu *pcpu;
1415
1416                if (!cpu_possible(cpu))
1417                        continue;
1418
1419                pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1420                spin_lock_bh(&pcpu->lock);
1421                list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1422restart:
1423                hlist_nulls_for_each_entry(h, n, list, hnnode) {
1424                        ct = nf_ct_tuplehash_to_ctrack(h);
1425                        if (l3proto && nf_ct_l3num(ct) != l3proto)
1426                                continue;
1427                        if (cb->args[1]) {
1428                                if (ct != last)
1429                                        continue;
1430                                cb->args[1] = 0;
1431                        }
1432                        rcu_read_lock();
1433                        res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1434                                                  cb->nlh->nlmsg_seq,
1435                                                  NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1436                                                  ct);
1437                        rcu_read_unlock();
1438                        if (res < 0) {
1439                                if (!atomic_inc_not_zero(&ct->ct_general.use))
1440                                        continue;
1441                                cb->args[0] = cpu;
1442                                cb->args[1] = (unsigned long)ct;
1443                                spin_unlock_bh(&pcpu->lock);
1444                                goto out;
1445                        }
1446                }
1447                if (cb->args[1]) {
1448                        cb->args[1] = 0;
1449                        goto restart;
1450                }
1451                spin_unlock_bh(&pcpu->lock);
1452        }
1453        cb->args[2] = 1;
1454out:
1455        if (last)
1456                nf_ct_put(last);
1457
1458        return skb->len;
1459}
1460
1461static int
1462ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1463{
1464        return ctnetlink_dump_list(skb, cb, true);
1465}
1466
1467static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1468                                  struct sk_buff *skb,
1469                                  const struct nlmsghdr *nlh,
1470                                  const struct nlattr * const cda[],
1471                                  struct netlink_ext_ack *extack)
1472{
1473        if (nlh->nlmsg_flags & NLM_F_DUMP) {
1474                struct netlink_dump_control c = {
1475                        .dump = ctnetlink_dump_dying,
1476                        .done = ctnetlink_done_list,
1477                };
1478                return netlink_dump_start(ctnl, skb, nlh, &c);
1479        }
1480
1481        return -EOPNOTSUPP;
1482}
1483
1484static int
1485ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1486{
1487        return ctnetlink_dump_list(skb, cb, false);
1488}
1489
1490static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1491                                        struct sk_buff *skb,
1492                                        const struct nlmsghdr *nlh,
1493                                        const struct nlattr * const cda[],
1494                                        struct netlink_ext_ack *extack)
1495{
1496        if (nlh->nlmsg_flags & NLM_F_DUMP) {
1497                struct netlink_dump_control c = {
1498                        .dump = ctnetlink_dump_unconfirmed,
1499                        .done = ctnetlink_done_list,
1500                };
1501                return netlink_dump_start(ctnl, skb, nlh, &c);
1502        }
1503
1504        return -EOPNOTSUPP;
1505}
1506
1507#if IS_ENABLED(CONFIG_NF_NAT)
1508static int
1509ctnetlink_parse_nat_setup(struct nf_conn *ct,
1510                          enum nf_nat_manip_type manip,
1511                          const struct nlattr *attr)
1512{
1513        struct nf_nat_hook *nat_hook;
1514        int err;
1515
1516        nat_hook = rcu_dereference(nf_nat_hook);
1517        if (!nat_hook) {
1518#ifdef CONFIG_MODULES
1519                rcu_read_unlock();
1520                nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1521                if (request_module("nf-nat") < 0) {
1522                        nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1523                        rcu_read_lock();
1524                        return -EOPNOTSUPP;
1525                }
1526                nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1527                rcu_read_lock();
1528                nat_hook = rcu_dereference(nf_nat_hook);
1529                if (nat_hook)
1530                        return -EAGAIN;
1531#endif
1532                return -EOPNOTSUPP;
1533        }
1534
1535        err = nat_hook->parse_nat_setup(ct, manip, attr);
1536        if (err == -EAGAIN) {
1537#ifdef CONFIG_MODULES
1538                rcu_read_unlock();
1539                nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1540                if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1541                        nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1542                        rcu_read_lock();
1543                        return -EOPNOTSUPP;
1544                }
1545                nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1546                rcu_read_lock();
1547#else
1548                err = -EOPNOTSUPP;
1549#endif
1550        }
1551        return err;
1552}
1553#endif
1554
1555static void
1556__ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1557                          unsigned long off)
1558{
1559        unsigned int bit;
1560
1561        /* Ignore these unchangable bits */
1562        on &= ~IPS_UNCHANGEABLE_MASK;
1563        off &= ~IPS_UNCHANGEABLE_MASK;
1564
1565        for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1566                if (on & (1 << bit))
1567                        set_bit(bit, &ct->status);
1568                else if (off & (1 << bit))
1569                        clear_bit(bit, &ct->status);
1570        }
1571}
1572
1573static int
1574ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1575{
1576        unsigned long d;
1577        unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1578        d = ct->status ^ status;
1579
1580        if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1581                /* unchangeable */
1582                return -EBUSY;
1583
1584        if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1585                /* SEEN_REPLY bit can only be set */
1586                return -EBUSY;
1587
1588        if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1589                /* ASSURED bit can only be set */
1590                return -EBUSY;
1591
1592        __ctnetlink_change_status(ct, status, 0);
1593        return 0;
1594}
1595
1596static int
1597ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1598{
1599#if IS_ENABLED(CONFIG_NF_NAT)
1600        int ret;
1601
1602        if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1603                return 0;
1604
1605        ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1606                                        cda[CTA_NAT_DST]);
1607        if (ret < 0)
1608                return ret;
1609
1610        return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1611                                         cda[CTA_NAT_SRC]);
1612#else
1613        if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1614                return 0;
1615        return -EOPNOTSUPP;
1616#endif
1617}
1618
1619static int ctnetlink_change_helper(struct nf_conn *ct,
1620                                   const struct nlattr * const cda[])
1621{
1622        struct nf_conntrack_helper *helper;
1623        struct nf_conn_help *help = nfct_help(ct);
1624        char *helpname = NULL;
1625        struct nlattr *helpinfo = NULL;
1626        int err;
1627
1628        err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1629        if (err < 0)
1630                return err;
1631
1632        /* don't change helper of sibling connections */
1633        if (ct->master) {
1634                /* If we try to change the helper to the same thing twice,
1635                 * treat the second attempt as a no-op instead of returning
1636                 * an error.
1637                 */
1638                err = -EBUSY;
1639                if (help) {
1640                        rcu_read_lock();
1641                        helper = rcu_dereference(help->helper);
1642                        if (helper && !strcmp(helper->name, helpname))
1643                                err = 0;
1644                        rcu_read_unlock();
1645                }
1646
1647                return err;
1648        }
1649
1650        if (!strcmp(helpname, "")) {
1651                if (help && help->helper) {
1652                        /* we had a helper before ... */
1653                        nf_ct_remove_expectations(ct);
1654                        RCU_INIT_POINTER(help->helper, NULL);
1655                }
1656
1657                return 0;
1658        }
1659
1660        rcu_read_lock();
1661        helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1662                                            nf_ct_protonum(ct));
1663        if (helper == NULL) {
1664                rcu_read_unlock();
1665                return -EOPNOTSUPP;
1666        }
1667
1668        if (help) {
1669                if (help->helper == helper) {
1670                        /* update private helper data if allowed. */
1671                        if (helper->from_nlattr)
1672                                helper->from_nlattr(helpinfo, ct);
1673                        err = 0;
1674                } else
1675                        err = -EBUSY;
1676        } else {
1677                /* we cannot set a helper for an existing conntrack */
1678                err = -EOPNOTSUPP;
1679        }
1680
1681        rcu_read_unlock();
1682        return err;
1683}
1684
1685static int ctnetlink_change_timeout(struct nf_conn *ct,
1686                                    const struct nlattr * const cda[])
1687{
1688        u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1689
1690        if (timeout > INT_MAX)
1691                timeout = INT_MAX;
1692        ct->timeout = nfct_time_stamp + (u32)timeout;
1693
1694        if (test_bit(IPS_DYING_BIT, &ct->status))
1695                return -ETIME;
1696
1697        return 0;
1698}
1699
1700#if defined(CONFIG_NF_CONNTRACK_MARK)
1701static void ctnetlink_change_mark(struct nf_conn *ct,
1702                                    const struct nlattr * const cda[])
1703{
1704        u32 mark, newmark, mask = 0;
1705
1706        if (cda[CTA_MARK_MASK])
1707                mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1708
1709        mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1710        newmark = (ct->mark & mask) ^ mark;
1711        if (newmark != ct->mark)
1712                ct->mark = newmark;
1713}
1714#endif
1715
1716static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1717        [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1718        [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1719        [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1720};
1721
1722static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1723                                      const struct nlattr * const cda[])
1724{
1725        const struct nlattr *attr = cda[CTA_PROTOINFO];
1726        const struct nf_conntrack_l4proto *l4proto;
1727        struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1728        int err = 0;
1729
1730        err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
1731                                          protoinfo_policy, NULL);
1732        if (err < 0)
1733                return err;
1734
1735        l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1736        if (l4proto->from_nlattr)
1737                err = l4proto->from_nlattr(tb, ct);
1738
1739        return err;
1740}
1741
1742static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
1743        [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
1744        [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
1745        [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
1746};
1747
1748static int change_seq_adj(struct nf_ct_seqadj *seq,
1749                          const struct nlattr * const attr)
1750{
1751        int err;
1752        struct nlattr *cda[CTA_SEQADJ_MAX+1];
1753
1754        err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
1755                                          seqadj_policy, NULL);
1756        if (err < 0)
1757                return err;
1758
1759        if (!cda[CTA_SEQADJ_CORRECTION_POS])
1760                return -EINVAL;
1761
1762        seq->correction_pos =
1763                ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
1764
1765        if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
1766                return -EINVAL;
1767
1768        seq->offset_before =
1769                ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
1770
1771        if (!cda[CTA_SEQADJ_OFFSET_AFTER])
1772                return -EINVAL;
1773
1774        seq->offset_after =
1775                ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
1776
1777        return 0;
1778}
1779
1780static int
1781ctnetlink_change_seq_adj(struct nf_conn *ct,
1782                         const struct nlattr * const cda[])
1783{
1784        struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
1785        int ret = 0;
1786
1787        if (!seqadj)
1788                return 0;
1789
1790        spin_lock_bh(&ct->lock);
1791        if (cda[CTA_SEQ_ADJ_ORIG]) {
1792                ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
1793                                     cda[CTA_SEQ_ADJ_ORIG]);
1794                if (ret < 0)
1795                        goto err;
1796
1797                set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1798        }
1799
1800        if (cda[CTA_SEQ_ADJ_REPLY]) {
1801                ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
1802                                     cda[CTA_SEQ_ADJ_REPLY]);
1803                if (ret < 0)
1804                        goto err;
1805
1806                set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
1807        }
1808
1809        spin_unlock_bh(&ct->lock);
1810        return 0;
1811err:
1812        spin_unlock_bh(&ct->lock);
1813        return ret;
1814}
1815
1816static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
1817        [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
1818        [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
1819        [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
1820};
1821
1822static int ctnetlink_change_synproxy(struct nf_conn *ct,
1823                                     const struct nlattr * const cda[])
1824{
1825        struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
1826        struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
1827        int err;
1828
1829        if (!synproxy)
1830                return 0;
1831
1832        err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
1833                                          cda[CTA_SYNPROXY], synproxy_policy,
1834                                          NULL);
1835        if (err < 0)
1836                return err;
1837
1838        if (!tb[CTA_SYNPROXY_ISN] ||
1839            !tb[CTA_SYNPROXY_ITS] ||
1840            !tb[CTA_SYNPROXY_TSOFF])
1841                return -EINVAL;
1842
1843        synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
1844        synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
1845        synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
1846
1847        return 0;
1848}
1849
1850static int
1851ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
1852{
1853#ifdef CONFIG_NF_CONNTRACK_LABELS
1854        size_t len = nla_len(cda[CTA_LABELS]);
1855        const void *mask = cda[CTA_LABELS_MASK];
1856
1857        if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
1858                return -EINVAL;
1859
1860        if (mask) {
1861                if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
1862                    nla_len(cda[CTA_LABELS_MASK]) != len)
1863                        return -EINVAL;
1864                mask = nla_data(cda[CTA_LABELS_MASK]);
1865        }
1866
1867        len /= sizeof(u32);
1868
1869        return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
1870#else
1871        return -EOPNOTSUPP;
1872#endif
1873}
1874
1875static int
1876ctnetlink_change_conntrack(struct nf_conn *ct,
1877                           const struct nlattr * const cda[])
1878{
1879        int err;
1880
1881        /* only allow NAT changes and master assignation for new conntracks */
1882        if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1883                return -EOPNOTSUPP;
1884
1885        if (cda[CTA_HELP]) {
1886                err = ctnetlink_change_helper(ct, cda);
1887                if (err < 0)
1888                        return err;
1889        }
1890
1891        if (cda[CTA_TIMEOUT]) {
1892                err = ctnetlink_change_timeout(ct, cda);
1893                if (err < 0)
1894                        return err;
1895        }
1896
1897        if (cda[CTA_STATUS]) {
1898                err = ctnetlink_change_status(ct, cda);
1899                if (err < 0)
1900                        return err;
1901        }
1902
1903        if (cda[CTA_PROTOINFO]) {
1904                err = ctnetlink_change_protoinfo(ct, cda);
1905                if (err < 0)
1906                        return err;
1907        }
1908
1909#if defined(CONFIG_NF_CONNTRACK_MARK)
1910        if (cda[CTA_MARK])
1911                ctnetlink_change_mark(ct, cda);
1912#endif
1913
1914        if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
1915                err = ctnetlink_change_seq_adj(ct, cda);
1916                if (err < 0)
1917                        return err;
1918        }
1919
1920        if (cda[CTA_SYNPROXY]) {
1921                err = ctnetlink_change_synproxy(ct, cda);
1922                if (err < 0)
1923                        return err;
1924        }
1925
1926        if (cda[CTA_LABELS]) {
1927                err = ctnetlink_attach_labels(ct, cda);
1928                if (err < 0)
1929                        return err;
1930        }
1931
1932        return 0;
1933}
1934
1935static struct nf_conn *
1936ctnetlink_create_conntrack(struct net *net,
1937                           const struct nf_conntrack_zone *zone,
1938                           const struct nlattr * const cda[],
1939                           struct nf_conntrack_tuple *otuple,
1940                           struct nf_conntrack_tuple *rtuple,
1941                           u8 u3)
1942{
1943        struct nf_conn *ct;
1944        int err = -EINVAL;
1945        struct nf_conntrack_helper *helper;
1946        struct nf_conn_tstamp *tstamp;
1947        u64 timeout;
1948
1949        ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1950        if (IS_ERR(ct))
1951                return ERR_PTR(-ENOMEM);
1952
1953        if (!cda[CTA_TIMEOUT])
1954                goto err1;
1955
1956        timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1957        if (timeout > INT_MAX)
1958                timeout = INT_MAX;
1959        ct->timeout = (u32)timeout + nfct_time_stamp;
1960
1961        rcu_read_lock();
1962        if (cda[CTA_HELP]) {
1963                char *helpname = NULL;
1964                struct nlattr *helpinfo = NULL;
1965
1966                err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1967                if (err < 0)
1968                        goto err2;
1969
1970                helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1971                                                    nf_ct_protonum(ct));
1972                if (helper == NULL) {
1973                        rcu_read_unlock();
1974#ifdef CONFIG_MODULES
1975                        if (request_module("nfct-helper-%s", helpname) < 0) {
1976                                err = -EOPNOTSUPP;
1977                                goto err1;
1978                        }
1979
1980                        rcu_read_lock();
1981                        helper = __nf_conntrack_helper_find(helpname,
1982                                                            nf_ct_l3num(ct),
1983                                                            nf_ct_protonum(ct));
1984                        if (helper) {
1985                                err = -EAGAIN;
1986                                goto err2;
1987                        }
1988                        rcu_read_unlock();
1989#endif
1990                        err = -EOPNOTSUPP;
1991                        goto err1;
1992                } else {
1993                        struct nf_conn_help *help;
1994
1995                        help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1996                        if (help == NULL) {
1997                                err = -ENOMEM;
1998                                goto err2;
1999                        }
2000                        /* set private helper data if allowed. */
2001                        if (helper->from_nlattr)
2002                                helper->from_nlattr(helpinfo, ct);
2003
2004                        /* not in hash table yet so not strictly necessary */
2005                        RCU_INIT_POINTER(help->helper, helper);
2006                }
2007        } else {
2008                /* try an implicit helper assignation */
2009                err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2010                if (err < 0)
2011                        goto err2;
2012        }
2013
2014        err = ctnetlink_setup_nat(ct, cda);
2015        if (err < 0)
2016                goto err2;
2017
2018        nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2019        nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2020        nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2021        nf_ct_labels_ext_add(ct);
2022        nfct_seqadj_ext_add(ct);
2023        nfct_synproxy_ext_add(ct);
2024
2025        /* we must add conntrack extensions before confirmation. */
2026        ct->status |= IPS_CONFIRMED;
2027
2028        if (cda[CTA_STATUS]) {
2029                err = ctnetlink_change_status(ct, cda);
2030                if (err < 0)
2031                        goto err2;
2032        }
2033
2034        if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2035                err = ctnetlink_change_seq_adj(ct, cda);
2036                if (err < 0)
2037                        goto err2;
2038        }
2039
2040        memset(&ct->proto, 0, sizeof(ct->proto));
2041        if (cda[CTA_PROTOINFO]) {
2042                err = ctnetlink_change_protoinfo(ct, cda);
2043                if (err < 0)
2044                        goto err2;
2045        }
2046
2047        if (cda[CTA_SYNPROXY]) {
2048                err = ctnetlink_change_synproxy(ct, cda);
2049                if (err < 0)
2050                        goto err2;
2051        }
2052
2053#if defined(CONFIG_NF_CONNTRACK_MARK)
2054        if (cda[CTA_MARK])
2055                ctnetlink_change_mark(ct, cda);
2056#endif
2057
2058        /* setup master conntrack: this is a confirmed expectation */
2059        if (cda[CTA_TUPLE_MASTER]) {
2060                struct nf_conntrack_tuple master;
2061                struct nf_conntrack_tuple_hash *master_h;
2062                struct nf_conn *master_ct;
2063
2064                err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2065                                            u3, NULL);
2066                if (err < 0)
2067                        goto err2;
2068
2069                master_h = nf_conntrack_find_get(net, zone, &master);
2070                if (master_h == NULL) {
2071                        err = -ENOENT;
2072                        goto err2;
2073                }
2074                master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2075                __set_bit(IPS_EXPECTED_BIT, &ct->status);
2076                ct->master = master_ct;
2077        }
2078        tstamp = nf_conn_tstamp_find(ct);
2079        if (tstamp)
2080                tstamp->start = ktime_get_real_ns();
2081
2082        err = nf_conntrack_hash_check_insert(ct);
2083        if (err < 0)
2084                goto err2;
2085
2086        rcu_read_unlock();
2087
2088        return ct;
2089
2090err2:
2091        rcu_read_unlock();
2092err1:
2093        nf_conntrack_free(ct);
2094        return ERR_PTR(err);
2095}
2096
2097static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2098                                   struct sk_buff *skb,
2099                                   const struct nlmsghdr *nlh,
2100                                   const struct nlattr * const cda[],
2101                                   struct netlink_ext_ack *extack)
2102{
2103        struct nf_conntrack_tuple otuple, rtuple;
2104        struct nf_conntrack_tuple_hash *h = NULL;
2105        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2106        struct nf_conn *ct;
2107        u_int8_t u3 = nfmsg->nfgen_family;
2108        struct nf_conntrack_zone zone;
2109        int err;
2110
2111        err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2112        if (err < 0)
2113                return err;
2114
2115        if (cda[CTA_TUPLE_ORIG]) {
2116                err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2117                                            u3, &zone);
2118                if (err < 0)
2119                        return err;
2120        }
2121
2122        if (cda[CTA_TUPLE_REPLY]) {
2123                err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2124                                            u3, &zone);
2125                if (err < 0)
2126                        return err;
2127        }
2128
2129        if (cda[CTA_TUPLE_ORIG])
2130                h = nf_conntrack_find_get(net, &zone, &otuple);
2131        else if (cda[CTA_TUPLE_REPLY])
2132                h = nf_conntrack_find_get(net, &zone, &rtuple);
2133
2134        if (h == NULL) {
2135                err = -ENOENT;
2136                if (nlh->nlmsg_flags & NLM_F_CREATE) {
2137                        enum ip_conntrack_events events;
2138
2139                        if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2140                                return -EINVAL;
2141                        if (otuple.dst.protonum != rtuple.dst.protonum)
2142                                return -EINVAL;
2143
2144                        ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2145                                                        &rtuple, u3);
2146                        if (IS_ERR(ct))
2147                                return PTR_ERR(ct);
2148
2149                        err = 0;
2150                        if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2151                                events = 1 << IPCT_RELATED;
2152                        else
2153                                events = 1 << IPCT_NEW;
2154
2155                        if (cda[CTA_LABELS] &&
2156                            ctnetlink_attach_labels(ct, cda) == 0)
2157                                events |= (1 << IPCT_LABEL);
2158
2159                        nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2160                                                      (1 << IPCT_ASSURED) |
2161                                                      (1 << IPCT_HELPER) |
2162                                                      (1 << IPCT_PROTOINFO) |
2163                                                      (1 << IPCT_SEQADJ) |
2164                                                      (1 << IPCT_MARK) |
2165                                                      (1 << IPCT_SYNPROXY) |
2166                                                      events,
2167                                                      ct, NETLINK_CB(skb).portid,
2168                                                      nlmsg_report(nlh));
2169                        nf_ct_put(ct);
2170                }
2171
2172                return err;
2173        }
2174        /* implicit 'else' */
2175
2176        err = -EEXIST;
2177        ct = nf_ct_tuplehash_to_ctrack(h);
2178        if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2179                err = ctnetlink_change_conntrack(ct, cda);
2180                if (err == 0) {
2181                        nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2182                                                      (1 << IPCT_ASSURED) |
2183                                                      (1 << IPCT_HELPER) |
2184                                                      (1 << IPCT_LABEL) |
2185                                                      (1 << IPCT_PROTOINFO) |
2186                                                      (1 << IPCT_SEQADJ) |
2187                                                      (1 << IPCT_MARK) |
2188                                                      (1 << IPCT_SYNPROXY),
2189                                                      ct, NETLINK_CB(skb).portid,
2190                                                      nlmsg_report(nlh));
2191                }
2192        }
2193
2194        nf_ct_put(ct);
2195        return err;
2196}
2197
2198static int
2199ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2200                                __u16 cpu, const struct ip_conntrack_stat *st)
2201{
2202        struct nlmsghdr *nlh;
2203        struct nfgenmsg *nfmsg;
2204        unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2205
2206        event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2207                              IPCTNL_MSG_CT_GET_STATS_CPU);
2208        nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2209        if (nlh == NULL)
2210                goto nlmsg_failure;
2211
2212        nfmsg = nlmsg_data(nlh);
2213        nfmsg->nfgen_family = AF_UNSPEC;
2214        nfmsg->version      = NFNETLINK_V0;
2215        nfmsg->res_id       = htons(cpu);
2216
2217        if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2218            nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2219            nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
2220            nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2221            nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2222                                htonl(st->insert_failed)) ||
2223            nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2224            nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2225            nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2226            nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2227                                htonl(st->search_restart)))
2228                goto nla_put_failure;
2229
2230        nlmsg_end(skb, nlh);
2231        return skb->len;
2232
2233nla_put_failure:
2234nlmsg_failure:
2235        nlmsg_cancel(skb, nlh);
2236        return -1;
2237}
2238
2239static int
2240ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2241{
2242        int cpu;
2243        struct net *net = sock_net(skb->sk);
2244
2245        if (cb->args[0] == nr_cpu_ids)
2246                return 0;
2247
2248        for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2249                const struct ip_conntrack_stat *st;
2250
2251                if (!cpu_possible(cpu))
2252                        continue;
2253
2254                st = per_cpu_ptr(net->ct.stat, cpu);
2255                if (ctnetlink_ct_stat_cpu_fill_info(skb,
2256                                                    NETLINK_CB(cb->skb).portid,
2257                                                    cb->nlh->nlmsg_seq,
2258                                                    cpu, st) < 0)
2259                                break;
2260        }
2261        cb->args[0] = cpu;
2262
2263        return skb->len;
2264}
2265
2266static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2267                                 struct sk_buff *skb,
2268                                 const struct nlmsghdr *nlh,
2269                                 const struct nlattr * const cda[],
2270                                 struct netlink_ext_ack *extack)
2271{
2272        if (nlh->nlmsg_flags & NLM_F_DUMP) {
2273                struct netlink_dump_control c = {
2274                        .dump = ctnetlink_ct_stat_cpu_dump,
2275                };
2276                return netlink_dump_start(ctnl, skb, nlh, &c);
2277        }
2278
2279        return 0;
2280}
2281
2282static int
2283ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2284                            struct net *net)
2285{
2286        struct nlmsghdr *nlh;
2287        struct nfgenmsg *nfmsg;
2288        unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2289        unsigned int nr_conntracks = atomic_read(&net->ct.count);
2290
2291        event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2292        nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2293        if (nlh == NULL)
2294                goto nlmsg_failure;
2295
2296        nfmsg = nlmsg_data(nlh);
2297        nfmsg->nfgen_family = AF_UNSPEC;
2298        nfmsg->version      = NFNETLINK_V0;
2299        nfmsg->res_id       = 0;
2300
2301        if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2302                goto nla_put_failure;
2303
2304        if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2305                goto nla_put_failure;
2306
2307        nlmsg_end(skb, nlh);
2308        return skb->len;
2309
2310nla_put_failure:
2311nlmsg_failure:
2312        nlmsg_cancel(skb, nlh);
2313        return -1;
2314}
2315
2316static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2317                             struct sk_buff *skb, const struct nlmsghdr *nlh,
2318                             const struct nlattr * const cda[],
2319                             struct netlink_ext_ack *extack)
2320{
2321        struct sk_buff *skb2;
2322        int err;
2323
2324        skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2325        if (skb2 == NULL)
2326                return -ENOMEM;
2327
2328        err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2329                                          nlh->nlmsg_seq,
2330                                          NFNL_MSG_TYPE(nlh->nlmsg_type),
2331                                          sock_net(skb->sk));
2332        if (err <= 0)
2333                goto free;
2334
2335        err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2336        if (err < 0)
2337                goto out;
2338
2339        return 0;
2340
2341free:
2342        kfree_skb(skb2);
2343out:
2344        /* this avoids a loop in nfnetlink. */
2345        return err == -EAGAIN ? -ENOBUFS : err;
2346}
2347
2348static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2349        [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2350        [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2351        [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2352        [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2353        [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2354        [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2355                                    .len = NF_CT_HELPER_NAME_LEN - 1 },
2356        [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2357        [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2358        [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2359        [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2360        [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2361};
2362
2363static struct nf_conntrack_expect *
2364ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2365                       struct nf_conntrack_helper *helper,
2366                       struct nf_conntrack_tuple *tuple,
2367                       struct nf_conntrack_tuple *mask);
2368
2369#ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2370static size_t
2371ctnetlink_glue_build_size(const struct nf_conn *ct)
2372{
2373        return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2374               + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2375               + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2376               + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2377               + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2378               + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2379               + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2380               + nla_total_size(0) /* CTA_PROTOINFO */
2381               + nla_total_size(0) /* CTA_HELP */
2382               + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2383               + ctnetlink_secctx_size(ct)
2384#if IS_ENABLED(CONFIG_NF_NAT)
2385               + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2386               + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2387#endif
2388#ifdef CONFIG_NF_CONNTRACK_MARK
2389               + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2390#endif
2391#ifdef CONFIG_NF_CONNTRACK_ZONES
2392               + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2393#endif
2394               + ctnetlink_proto_size(ct)
2395               ;
2396}
2397
2398static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2399                                             enum ip_conntrack_info *ctinfo)
2400{
2401        return nf_ct_get(skb, ctinfo);
2402}
2403
2404static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2405{
2406        const struct nf_conntrack_zone *zone;
2407        struct nlattr *nest_parms;
2408
2409        zone = nf_ct_zone(ct);
2410
2411        nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2412        if (!nest_parms)
2413                goto nla_put_failure;
2414        if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2415                goto nla_put_failure;
2416        if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2417                                   NF_CT_ZONE_DIR_ORIG) < 0)
2418                goto nla_put_failure;
2419        nla_nest_end(skb, nest_parms);
2420
2421        nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2422        if (!nest_parms)
2423                goto nla_put_failure;
2424        if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2425                goto nla_put_failure;
2426        if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2427                                   NF_CT_ZONE_DIR_REPL) < 0)
2428                goto nla_put_failure;
2429        nla_nest_end(skb, nest_parms);
2430
2431        if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2432                                   NF_CT_DEFAULT_ZONE_DIR) < 0)
2433                goto nla_put_failure;
2434
2435        if (ctnetlink_dump_id(skb, ct) < 0)
2436                goto nla_put_failure;
2437
2438        if (ctnetlink_dump_status(skb, ct) < 0)
2439                goto nla_put_failure;
2440
2441        if (ctnetlink_dump_timeout(skb, ct) < 0)
2442                goto nla_put_failure;
2443
2444        if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2445                goto nla_put_failure;
2446
2447        if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2448                goto nla_put_failure;
2449
2450#ifdef CONFIG_NF_CONNTRACK_SECMARK
2451        if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2452                goto nla_put_failure;
2453#endif
2454        if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2455                goto nla_put_failure;
2456
2457        if ((ct->status & IPS_SEQ_ADJUST) &&
2458            ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2459                goto nla_put_failure;
2460
2461        if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2462                goto nla_put_failure;
2463
2464#ifdef CONFIG_NF_CONNTRACK_MARK
2465        if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
2466                goto nla_put_failure;
2467#endif
2468        if (ctnetlink_dump_labels(skb, ct) < 0)
2469                goto nla_put_failure;
2470        return 0;
2471
2472nla_put_failure:
2473        return -ENOSPC;
2474}
2475
2476static int
2477ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2478                     enum ip_conntrack_info ctinfo,
2479                     u_int16_t ct_attr, u_int16_t ct_info_attr)
2480{
2481        struct nlattr *nest_parms;
2482
2483        nest_parms = nla_nest_start(skb, ct_attr);
2484        if (!nest_parms)
2485                goto nla_put_failure;
2486
2487        if (__ctnetlink_glue_build(skb, ct) < 0)
2488                goto nla_put_failure;
2489
2490        nla_nest_end(skb, nest_parms);
2491
2492        if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2493                goto nla_put_failure;
2494
2495        return 0;
2496
2497nla_put_failure:
2498        return -ENOSPC;
2499}
2500
2501static int
2502ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2503{
2504        unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2505        unsigned long d = ct->status ^ status;
2506
2507        if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2508                /* SEEN_REPLY bit can only be set */
2509                return -EBUSY;
2510
2511        if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2512                /* ASSURED bit can only be set */
2513                return -EBUSY;
2514
2515        /* This check is less strict than ctnetlink_change_status()
2516         * because callers often flip IPS_EXPECTED bits when sending
2517         * an NFQA_CT attribute to the kernel.  So ignore the
2518         * unchangeable bits but do not error out. Also user programs
2519         * are allowed to clear the bits that they are allowed to change.
2520         */
2521        __ctnetlink_change_status(ct, status, ~status);
2522        return 0;
2523}
2524
2525static int
2526ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2527{
2528        int err;
2529
2530        if (cda[CTA_TIMEOUT]) {
2531                err = ctnetlink_change_timeout(ct, cda);
2532                if (err < 0)
2533                        return err;
2534        }
2535        if (cda[CTA_STATUS]) {
2536                err = ctnetlink_update_status(ct, cda);
2537                if (err < 0)
2538                        return err;
2539        }
2540        if (cda[CTA_HELP]) {
2541                err = ctnetlink_change_helper(ct, cda);
2542                if (err < 0)
2543                        return err;
2544        }
2545        if (cda[CTA_LABELS]) {
2546                err = ctnetlink_attach_labels(ct, cda);
2547                if (err < 0)
2548                        return err;
2549        }
2550#if defined(CONFIG_NF_CONNTRACK_MARK)
2551        if (cda[CTA_MARK]) {
2552                ctnetlink_change_mark(ct, cda);
2553        }
2554#endif
2555        return 0;
2556}
2557
2558static int
2559ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2560{
2561        struct nlattr *cda[CTA_MAX+1];
2562        int ret;
2563
2564        ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2565                                          NULL);
2566        if (ret < 0)
2567                return ret;
2568
2569        return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2570}
2571
2572static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2573                                    const struct nf_conn *ct,
2574                                    struct nf_conntrack_tuple *tuple,
2575                                    struct nf_conntrack_tuple *mask)
2576{
2577        int err;
2578
2579        err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2580                                    nf_ct_l3num(ct), NULL);
2581        if (err < 0)
2582                return err;
2583
2584        return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2585                                     nf_ct_l3num(ct), NULL);
2586}
2587
2588static int
2589ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2590                             u32 portid, u32 report)
2591{
2592        struct nlattr *cda[CTA_EXPECT_MAX+1];
2593        struct nf_conntrack_tuple tuple, mask;
2594        struct nf_conntrack_helper *helper = NULL;
2595        struct nf_conntrack_expect *exp;
2596        int err;
2597
2598        err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2599                                          exp_nla_policy, NULL);
2600        if (err < 0)
2601                return err;
2602
2603        err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2604                                       ct, &tuple, &mask);
2605        if (err < 0)
2606                return err;
2607
2608        if (cda[CTA_EXPECT_HELP_NAME]) {
2609                const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2610
2611                helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2612                                                    nf_ct_protonum(ct));
2613                if (helper == NULL)
2614                        return -EOPNOTSUPP;
2615        }
2616
2617        exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2618                                     helper, &tuple, &mask);
2619        if (IS_ERR(exp))
2620                return PTR_ERR(exp);
2621
2622        err = nf_ct_expect_related_report(exp, portid, report, 0);
2623        nf_ct_expect_put(exp);
2624        return err;
2625}
2626
2627static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2628                                  enum ip_conntrack_info ctinfo, int diff)
2629{
2630        if (!(ct->status & IPS_NAT_MASK))
2631                return;
2632
2633        nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2634}
2635
2636static struct nfnl_ct_hook ctnetlink_glue_hook = {
2637        .get_ct         = ctnetlink_glue_get_ct,
2638        .build_size     = ctnetlink_glue_build_size,
2639        .build          = ctnetlink_glue_build,
2640        .parse          = ctnetlink_glue_parse,
2641        .attach_expect  = ctnetlink_glue_attach_expect,
2642        .seq_adjust     = ctnetlink_glue_seqadj,
2643};
2644#endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2645
2646/***********************************************************************
2647 * EXPECT
2648 ***********************************************************************/
2649
2650static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2651                                    const struct nf_conntrack_tuple *tuple,
2652                                    u32 type)
2653{
2654        struct nlattr *nest_parms;
2655
2656        nest_parms = nla_nest_start(skb, type);
2657        if (!nest_parms)
2658                goto nla_put_failure;
2659        if (ctnetlink_dump_tuples(skb, tuple) < 0)
2660                goto nla_put_failure;
2661        nla_nest_end(skb, nest_parms);
2662
2663        return 0;
2664
2665nla_put_failure:
2666        return -1;
2667}
2668
2669static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2670                                   const struct nf_conntrack_tuple *tuple,
2671                                   const struct nf_conntrack_tuple_mask *mask)
2672{
2673        const struct nf_conntrack_l4proto *l4proto;
2674        struct nf_conntrack_tuple m;
2675        struct nlattr *nest_parms;
2676        int ret;
2677
2678        memset(&m, 0xFF, sizeof(m));
2679        memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2680        m.src.u.all = mask->src.u.all;
2681        m.dst.protonum = tuple->dst.protonum;
2682
2683        nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2684        if (!nest_parms)
2685                goto nla_put_failure;
2686
2687        rcu_read_lock();
2688        ret = ctnetlink_dump_tuples_ip(skb, &m);
2689        if (ret >= 0) {
2690                l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2691                ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2692        }
2693        rcu_read_unlock();
2694
2695        if (unlikely(ret < 0))
2696                goto nla_put_failure;
2697
2698        nla_nest_end(skb, nest_parms);
2699
2700        return 0;
2701
2702nla_put_failure:
2703        return -1;
2704}
2705
2706static const union nf_inet_addr any_addr;
2707
2708static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2709{
2710        static __read_mostly siphash_key_t exp_id_seed;
2711        unsigned long a, b, c, d;
2712
2713        net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2714
2715        a = (unsigned long)exp;
2716        b = (unsigned long)exp->helper;
2717        c = (unsigned long)exp->master;
2718        d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2719
2720#ifdef CONFIG_64BIT
2721        return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2722#else
2723        return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2724#endif
2725}
2726
2727static int
2728ctnetlink_exp_dump_expect(struct sk_buff *skb,
2729                          const struct nf_conntrack_expect *exp)
2730{
2731        struct nf_conn *master = exp->master;
2732        long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2733        struct nf_conn_help *help;
2734#if IS_ENABLED(CONFIG_NF_NAT)
2735        struct nlattr *nest_parms;
2736        struct nf_conntrack_tuple nat_tuple = {};
2737#endif
2738        struct nf_ct_helper_expectfn *expfn;
2739
2740        if (timeout < 0)
2741                timeout = 0;
2742
2743        if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2744                goto nla_put_failure;
2745        if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2746                goto nla_put_failure;
2747        if (ctnetlink_exp_dump_tuple(skb,
2748                                 &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2749                                 CTA_EXPECT_MASTER) < 0)
2750                goto nla_put_failure;
2751
2752#if IS_ENABLED(CONFIG_NF_NAT)
2753        if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2754            exp->saved_proto.all) {
2755                nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
2756                if (!nest_parms)
2757                        goto nla_put_failure;
2758
2759                if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2760                        goto nla_put_failure;
2761
2762                nat_tuple.src.l3num = nf_ct_l3num(master);
2763                nat_tuple.src.u3 = exp->saved_addr;
2764                nat_tuple.dst.protonum = nf_ct_protonum(master);
2765                nat_tuple.src.u = exp->saved_proto;
2766
2767                if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2768                                                CTA_EXPECT_NAT_TUPLE) < 0)
2769                        goto nla_put_failure;
2770                nla_nest_end(skb, nest_parms);
2771        }
2772#endif
2773        if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2774            nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
2775            nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2776            nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2777                goto nla_put_failure;
2778        help = nfct_help(master);
2779        if (help) {
2780                struct nf_conntrack_helper *helper;
2781
2782                helper = rcu_dereference(help->helper);
2783                if (helper &&
2784                    nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2785                        goto nla_put_failure;
2786        }
2787        expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2788        if (expfn != NULL &&
2789            nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2790                goto nla_put_failure;
2791
2792        return 0;
2793
2794nla_put_failure:
2795        return -1;
2796}
2797
2798static int
2799ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2800                        int event, const struct nf_conntrack_expect *exp)
2801{
2802        struct nlmsghdr *nlh;
2803        struct nfgenmsg *nfmsg;
2804        unsigned int flags = portid ? NLM_F_MULTI : 0;
2805
2806        event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
2807        nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2808        if (nlh == NULL)
2809                goto nlmsg_failure;
2810
2811        nfmsg = nlmsg_data(nlh);
2812        nfmsg->nfgen_family = exp->tuple.src.l3num;
2813        nfmsg->version      = NFNETLINK_V0;
2814        nfmsg->res_id       = 0;
2815
2816        if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2817                goto nla_put_failure;
2818
2819        nlmsg_end(skb, nlh);
2820        return skb->len;
2821
2822nlmsg_failure:
2823nla_put_failure:
2824        nlmsg_cancel(skb, nlh);
2825        return -1;
2826}
2827
2828#ifdef CONFIG_NF_CONNTRACK_EVENTS
2829static int
2830ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2831{
2832        struct nf_conntrack_expect *exp = item->exp;
2833        struct net *net = nf_ct_exp_net(exp);
2834        struct nlmsghdr *nlh;
2835        struct nfgenmsg *nfmsg;
2836        struct sk_buff *skb;
2837        unsigned int type, group;
2838        int flags = 0;
2839
2840        if (events & (1 << IPEXP_DESTROY)) {
2841                type = IPCTNL_MSG_EXP_DELETE;
2842                group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2843        } else if (events & (1 << IPEXP_NEW)) {
2844                type = IPCTNL_MSG_EXP_NEW;
2845                flags = NLM_F_CREATE|NLM_F_EXCL;
2846                group = NFNLGRP_CONNTRACK_EXP_NEW;
2847        } else
2848                return 0;
2849
2850        if (!item->report && !nfnetlink_has_listeners(net, group))
2851                return 0;
2852
2853        skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2854        if (skb == NULL)
2855                goto errout;
2856
2857        type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
2858        nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2859        if (nlh == NULL)
2860                goto nlmsg_failure;
2861
2862        nfmsg = nlmsg_data(nlh);
2863        nfmsg->nfgen_family = exp->tuple.src.l3num;
2864        nfmsg->version      = NFNETLINK_V0;
2865        nfmsg->res_id       = 0;
2866
2867        if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2868                goto nla_put_failure;
2869
2870        nlmsg_end(skb, nlh);
2871        nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2872        return 0;
2873
2874nla_put_failure:
2875        nlmsg_cancel(skb, nlh);
2876nlmsg_failure:
2877        kfree_skb(skb);
2878errout:
2879        nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2880        return 0;
2881}
2882#endif
2883static int ctnetlink_exp_done(struct netlink_callback *cb)
2884{
2885        if (cb->args[1])
2886                nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2887        return 0;
2888}
2889
2890static int
2891ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2892{
2893        struct net *net = sock_net(skb->sk);
2894        struct nf_conntrack_expect *exp, *last;
2895        struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2896        u_int8_t l3proto = nfmsg->nfgen_family;
2897
2898        rcu_read_lock();
2899        last = (struct nf_conntrack_expect *)cb->args[1];
2900        for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2901restart:
2902                hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
2903                                         hnode) {
2904                        if (l3proto && exp->tuple.src.l3num != l3proto)
2905                                continue;
2906
2907                        if (!net_eq(nf_ct_net(exp->master), net))
2908                                continue;
2909
2910                        if (cb->args[1]) {
2911                                if (exp != last)
2912                                        continue;
2913                                cb->args[1] = 0;
2914                        }
2915                        if (ctnetlink_exp_fill_info(skb,
2916                                                    NETLINK_CB(cb->skb).portid,
2917                                                    cb->nlh->nlmsg_seq,
2918                                                    IPCTNL_MSG_EXP_NEW,
2919                                                    exp) < 0) {
2920                                if (!refcount_inc_not_zero(&exp->use))
2921                                        continue;
2922                                cb->args[1] = (unsigned long)exp;
2923                                goto out;
2924                        }
2925                }
2926                if (cb->args[1]) {
2927                        cb->args[1] = 0;
2928                        goto restart;
2929                }
2930        }
2931out:
2932        rcu_read_unlock();
2933        if (last)
2934                nf_ct_expect_put(last);
2935
2936        return skb->len;
2937}
2938
2939static int
2940ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2941{
2942        struct nf_conntrack_expect *exp, *last;
2943        struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2944        struct nf_conn *ct = cb->data;
2945        struct nf_conn_help *help = nfct_help(ct);
2946        u_int8_t l3proto = nfmsg->nfgen_family;
2947
2948        if (cb->args[0])
2949                return 0;
2950
2951        rcu_read_lock();
2952        last = (struct nf_conntrack_expect *)cb->args[1];
2953restart:
2954        hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
2955                if (l3proto && exp->tuple.src.l3num != l3proto)
2956                        continue;
2957                if (cb->args[1]) {
2958                        if (exp != last)
2959                                continue;
2960                        cb->args[1] = 0;
2961                }
2962                if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
2963                                            cb->nlh->nlmsg_seq,
2964                                            IPCTNL_MSG_EXP_NEW,
2965                                            exp) < 0) {
2966                        if (!refcount_inc_not_zero(&exp->use))
2967                                continue;
2968                        cb->args[1] = (unsigned long)exp;
2969                        goto out;
2970                }
2971        }
2972        if (cb->args[1]) {
2973                cb->args[1] = 0;
2974                goto restart;
2975        }
2976        cb->args[0] = 1;
2977out:
2978        rcu_read_unlock();
2979        if (last)
2980                nf_ct_expect_put(last);
2981
2982        return skb->len;
2983}
2984
2985static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
2986                                 struct sk_buff *skb,
2987                                 const struct nlmsghdr *nlh,
2988                                 const struct nlattr * const cda[],
2989                                 struct netlink_ext_ack *extack)
2990{
2991        int err;
2992        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2993        u_int8_t u3 = nfmsg->nfgen_family;
2994        struct nf_conntrack_tuple tuple;
2995        struct nf_conntrack_tuple_hash *h;
2996        struct nf_conn *ct;
2997        struct nf_conntrack_zone zone;
2998        struct netlink_dump_control c = {
2999                .dump = ctnetlink_exp_ct_dump_table,
3000                .done = ctnetlink_exp_done,
3001        };
3002
3003        err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3004                                    u3, NULL);
3005        if (err < 0)
3006                return err;
3007
3008        err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3009        if (err < 0)
3010                return err;
3011
3012        h = nf_conntrack_find_get(net, &zone, &tuple);
3013        if (!h)
3014                return -ENOENT;
3015
3016        ct = nf_ct_tuplehash_to_ctrack(h);
3017        /* No expectation linked to this connection tracking. */
3018        if (!nfct_help(ct)) {
3019                nf_ct_put(ct);
3020                return 0;
3021        }
3022
3023        c.data = ct;
3024
3025        err = netlink_dump_start(ctnl, skb, nlh, &c);
3026        nf_ct_put(ct);
3027
3028        return err;
3029}
3030
3031static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
3032                                struct sk_buff *skb, const struct nlmsghdr *nlh,
3033                                const struct nlattr * const cda[],
3034                                struct netlink_ext_ack *extack)
3035{
3036        struct nf_conntrack_tuple tuple;
3037        struct nf_conntrack_expect *exp;
3038        struct sk_buff *skb2;
3039        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3040        u_int8_t u3 = nfmsg->nfgen_family;
3041        struct nf_conntrack_zone zone;
3042        int err;
3043
3044        if (nlh->nlmsg_flags & NLM_F_DUMP) {
3045                if (cda[CTA_EXPECT_MASTER])
3046                        return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
3047                                                     extack);
3048                else {
3049                        struct netlink_dump_control c = {
3050                                .dump = ctnetlink_exp_dump_table,
3051                                .done = ctnetlink_exp_done,
3052                        };
3053                        return netlink_dump_start(ctnl, skb, nlh, &c);
3054                }
3055        }
3056
3057        err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3058        if (err < 0)
3059                return err;
3060
3061        if (cda[CTA_EXPECT_TUPLE])
3062                err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3063                                            u3, NULL);
3064        else if (cda[CTA_EXPECT_MASTER])
3065                err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3066                                            u3, NULL);
3067        else
3068                return -EINVAL;
3069
3070        if (err < 0)
3071                return err;
3072
3073        exp = nf_ct_expect_find_get(net, &zone, &tuple);
3074        if (!exp)
3075                return -ENOENT;
3076
3077        if (cda[CTA_EXPECT_ID]) {
3078                __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3079
3080                if (id != nf_expect_get_id(exp)) {
3081                        nf_ct_expect_put(exp);
3082                        return -ENOENT;
3083                }
3084        }
3085
3086        err = -ENOMEM;
3087        skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3088        if (skb2 == NULL) {
3089                nf_ct_expect_put(exp);
3090                goto out;
3091        }
3092
3093        rcu_read_lock();
3094        err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3095                                      nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
3096        rcu_read_unlock();
3097        nf_ct_expect_put(exp);
3098        if (err <= 0)
3099                goto free;
3100
3101        err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
3102        if (err < 0)
3103                goto out;
3104
3105        return 0;
3106
3107free:
3108        kfree_skb(skb2);
3109out:
3110        /* this avoids a loop in nfnetlink. */
3111        return err == -EAGAIN ? -ENOBUFS : err;
3112}
3113
3114static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3115{
3116        const struct nf_conn_help *m_help;
3117        const char *name = data;
3118
3119        m_help = nfct_help(exp->master);
3120
3121        return strcmp(m_help->helper->name, name) == 0;
3122}
3123
3124static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3125{
3126        return true;
3127}
3128
3129static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3130                                struct sk_buff *skb, const struct nlmsghdr *nlh,
3131                                const struct nlattr * const cda[],
3132                                struct netlink_ext_ack *extack)
3133{
3134        struct nf_conntrack_expect *exp;
3135        struct nf_conntrack_tuple tuple;
3136        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3137        u_int8_t u3 = nfmsg->nfgen_family;
3138        struct nf_conntrack_zone zone;
3139        int err;
3140
3141        if (cda[CTA_EXPECT_TUPLE]) {
3142                /* delete a single expect by tuple */
3143                err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3144                if (err < 0)
3145                        return err;
3146
3147                err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3148                                            u3, NULL);
3149                if (err < 0)
3150                        return err;
3151
3152                /* bump usage count to 2 */
3153                exp = nf_ct_expect_find_get(net, &zone, &tuple);
3154                if (!exp)
3155                        return -ENOENT;
3156
3157                if (cda[CTA_EXPECT_ID]) {
3158                        __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3159                        if (ntohl(id) != (u32)(unsigned long)exp) {
3160                                nf_ct_expect_put(exp);
3161                                return -ENOENT;
3162                        }
3163                }
3164
3165                /* after list removal, usage count == 1 */
3166                spin_lock_bh(&nf_conntrack_expect_lock);
3167                if (del_timer(&exp->timeout)) {
3168                        nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3169                                                   nlmsg_report(nlh));
3170                        nf_ct_expect_put(exp);
3171                }
3172                spin_unlock_bh(&nf_conntrack_expect_lock);
3173                /* have to put what we 'get' above.
3174                 * after this line usage count == 0 */
3175                nf_ct_expect_put(exp);
3176        } else if (cda[CTA_EXPECT_HELP_NAME]) {
3177                char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3178
3179                nf_ct_expect_iterate_net(net, expect_iter_name, name,
3180                                         NETLINK_CB(skb).portid,
3181                                         nlmsg_report(nlh));
3182        } else {
3183                /* This basically means we have to flush everything*/
3184                nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3185                                         NETLINK_CB(skb).portid,
3186                                         nlmsg_report(nlh));
3187        }
3188
3189        return 0;
3190}
3191static int
3192ctnetlink_change_expect(struct nf_conntrack_expect *x,
3193                        const struct nlattr * const cda[])
3194{
3195        if (cda[CTA_EXPECT_TIMEOUT]) {
3196                if (!del_timer(&x->timeout))
3197                        return -ETIME;
3198
3199                x->timeout.expires = jiffies +
3200                        ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3201                add_timer(&x->timeout);
3202        }
3203        return 0;
3204}
3205
3206static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3207        [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3208        [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3209};
3210
3211static int
3212ctnetlink_parse_expect_nat(const struct nlattr *attr,
3213                           struct nf_conntrack_expect *exp,
3214                           u_int8_t u3)
3215{
3216#if IS_ENABLED(CONFIG_NF_NAT)
3217        struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3218        struct nf_conntrack_tuple nat_tuple = {};
3219        int err;
3220
3221        err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3222                                          exp_nat_nla_policy, NULL);
3223        if (err < 0)
3224                return err;
3225
3226        if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3227                return -EINVAL;
3228
3229        err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3230                                    &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3231                                    u3, NULL);
3232        if (err < 0)
3233                return err;
3234
3235        exp->saved_addr = nat_tuple.src.u3;
3236        exp->saved_proto = nat_tuple.src.u;
3237        exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3238
3239        return 0;
3240#else
3241        return -EOPNOTSUPP;
3242#endif
3243}
3244
3245static struct nf_conntrack_expect *
3246ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3247                       struct nf_conntrack_helper *helper,
3248                       struct nf_conntrack_tuple *tuple,
3249                       struct nf_conntrack_tuple *mask)
3250{
3251        u_int32_t class = 0;
3252        struct nf_conntrack_expect *exp;
3253        struct nf_conn_help *help;
3254        int err;
3255
3256        help = nfct_help(ct);
3257        if (!help)
3258                return ERR_PTR(-EOPNOTSUPP);
3259
3260        if (cda[CTA_EXPECT_CLASS] && helper) {
3261                class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3262                if (class > helper->expect_class_max)
3263                        return ERR_PTR(-EINVAL);
3264        }
3265        exp = nf_ct_expect_alloc(ct);
3266        if (!exp)
3267                return ERR_PTR(-ENOMEM);
3268
3269        if (cda[CTA_EXPECT_FLAGS]) {
3270                exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3271                exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3272        } else {
3273                exp->flags = 0;
3274        }
3275        if (cda[CTA_EXPECT_FN]) {
3276                const char *name = nla_data(cda[CTA_EXPECT_FN]);
3277                struct nf_ct_helper_expectfn *expfn;
3278
3279                expfn = nf_ct_helper_expectfn_find_by_name(name);
3280                if (expfn == NULL) {
3281                        err = -EINVAL;
3282                        goto err_out;
3283                }
3284                exp->expectfn = expfn->expectfn;
3285        } else
3286                exp->expectfn = NULL;
3287
3288        exp->class = class;
3289        exp->master = ct;
3290        exp->helper = helper;
3291        exp->tuple = *tuple;
3292        exp->mask.src.u3 = mask->src.u3;
3293        exp->mask.src.u.all = mask->src.u.all;
3294
3295        if (cda[CTA_EXPECT_NAT]) {
3296                err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3297                                                 exp, nf_ct_l3num(ct));
3298                if (err < 0)
3299                        goto err_out;
3300        }
3301        return exp;
3302err_out:
3303        nf_ct_expect_put(exp);
3304        return ERR_PTR(err);
3305}
3306
3307static int
3308ctnetlink_create_expect(struct net *net,
3309                        const struct nf_conntrack_zone *zone,
3310                        const struct nlattr * const cda[],
3311                        u_int8_t u3, u32 portid, int report)
3312{
3313        struct nf_conntrack_tuple tuple, mask, master_tuple;
3314        struct nf_conntrack_tuple_hash *h = NULL;
3315        struct nf_conntrack_helper *helper = NULL;
3316        struct nf_conntrack_expect *exp;
3317        struct nf_conn *ct;
3318        int err;
3319
3320        /* caller guarantees that those three CTA_EXPECT_* exist */
3321        err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3322                                    u3, NULL);
3323        if (err < 0)
3324                return err;
3325        err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3326                                    u3, NULL);
3327        if (err < 0)
3328                return err;
3329        err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3330                                    u3, NULL);
3331        if (err < 0)
3332                return err;
3333
3334        /* Look for master conntrack of this expectation */
3335        h = nf_conntrack_find_get(net, zone, &master_tuple);
3336        if (!h)
3337                return -ENOENT;
3338        ct = nf_ct_tuplehash_to_ctrack(h);
3339
3340        rcu_read_lock();
3341        if (cda[CTA_EXPECT_HELP_NAME]) {
3342                const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3343
3344                helper = __nf_conntrack_helper_find(helpname, u3,
3345                                                    nf_ct_protonum(ct));
3346                if (helper == NULL) {
3347                        rcu_read_unlock();
3348#ifdef CONFIG_MODULES
3349                        if (request_module("nfct-helper-%s", helpname) < 0) {
3350                                err = -EOPNOTSUPP;
3351                                goto err_ct;
3352                        }
3353                        rcu_read_lock();
3354                        helper = __nf_conntrack_helper_find(helpname, u3,
3355                                                            nf_ct_protonum(ct));
3356                        if (helper) {
3357                                err = -EAGAIN;
3358                                goto err_rcu;
3359                        }
3360                        rcu_read_unlock();
3361#endif
3362                        err = -EOPNOTSUPP;
3363                        goto err_ct;
3364                }
3365        }
3366
3367        exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3368        if (IS_ERR(exp)) {
3369                err = PTR_ERR(exp);
3370                goto err_rcu;
3371        }
3372
3373        err = nf_ct_expect_related_report(exp, portid, report, 0);
3374        nf_ct_expect_put(exp);
3375err_rcu:
3376        rcu_read_unlock();
3377err_ct:
3378        nf_ct_put(ct);
3379        return err;
3380}
3381
3382static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3383                                struct sk_buff *skb, const struct nlmsghdr *nlh,
3384                                const struct nlattr * const cda[],
3385                                struct netlink_ext_ack *extack)
3386{
3387        struct nf_conntrack_tuple tuple;
3388        struct nf_conntrack_expect *exp;
3389        struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3390        u_int8_t u3 = nfmsg->nfgen_family;
3391        struct nf_conntrack_zone zone;
3392        int err;
3393
3394        if (!cda[CTA_EXPECT_TUPLE]
3395            || !cda[CTA_EXPECT_MASK]
3396            || !cda[CTA_EXPECT_MASTER])
3397                return -EINVAL;
3398
3399        err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3400        if (err < 0)
3401                return err;
3402
3403        err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3404                                    u3, NULL);
3405        if (err < 0)
3406                return err;
3407
3408        spin_lock_bh(&nf_conntrack_expect_lock);
3409        exp = __nf_ct_expect_find(net, &zone, &tuple);
3410        if (!exp) {
3411                spin_unlock_bh(&nf_conntrack_expect_lock);
3412                err = -ENOENT;
3413                if (nlh->nlmsg_flags & NLM_F_CREATE) {
3414                        err = ctnetlink_create_expect(net, &zone, cda, u3,
3415                                                      NETLINK_CB(skb).portid,
3416                                                      nlmsg_report(nlh));
3417                }
3418                return err;
3419        }
3420
3421        err = -EEXIST;
3422        if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3423                err = ctnetlink_change_expect(exp, cda);
3424        spin_unlock_bh(&nf_conntrack_expect_lock);
3425
3426        return err;
3427}
3428
3429static int
3430ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3431                             const struct ip_conntrack_stat *st)
3432{
3433        struct nlmsghdr *nlh;
3434        struct nfgenmsg *nfmsg;
3435        unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3436
3437        event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3438                              IPCTNL_MSG_EXP_GET_STATS_CPU);
3439        nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
3440        if (nlh == NULL)
3441                goto nlmsg_failure;
3442
3443        nfmsg = nlmsg_data(nlh);
3444        nfmsg->nfgen_family = AF_UNSPEC;
3445        nfmsg->version      = NFNETLINK_V0;
3446        nfmsg->res_id       = htons(cpu);
3447
3448        if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3449            nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3450            nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3451                goto nla_put_failure;
3452
3453        nlmsg_end(skb, nlh);
3454        return skb->len;
3455
3456nla_put_failure:
3457nlmsg_failure:
3458        nlmsg_cancel(skb, nlh);
3459        return -1;
3460}
3461
3462static int
3463ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3464{
3465        int cpu;
3466        struct net *net = sock_net(skb->sk);
3467
3468        if (cb->args[0] == nr_cpu_ids)
3469                return 0;
3470
3471        for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3472                const struct ip_conntrack_stat *st;
3473
3474                if (!cpu_possible(cpu))
3475                        continue;
3476
3477                st = per_cpu_ptr(net->ct.stat, cpu);
3478                if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3479                                                 cb->nlh->nlmsg_seq,
3480                                                 cpu, st) < 0)
3481                        break;
3482        }
3483        cb->args[0] = cpu;
3484
3485        return skb->len;
3486}
3487
3488static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3489                                  struct sk_buff *skb,
3490                                  const struct nlmsghdr *nlh,
3491                                  const struct nlattr * const cda[],
3492                                  struct netlink_ext_ack *extack)
3493{
3494        if (nlh->nlmsg_flags & NLM_F_DUMP) {
3495                struct netlink_dump_control c = {
3496                        .dump = ctnetlink_exp_stat_cpu_dump,
3497                };
3498                return netlink_dump_start(ctnl, skb, nlh, &c);
3499        }
3500
3501        return 0;
3502}
3503
3504#ifdef CONFIG_NF_CONNTRACK_EVENTS
3505static struct nf_ct_event_notifier ctnl_notifier = {
3506        .fcn = ctnetlink_conntrack_event,
3507};
3508
3509static struct nf_exp_event_notifier ctnl_notifier_exp = {
3510        .fcn = ctnetlink_expect_event,
3511};
3512#endif
3513
3514static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3515        [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3516                                            .attr_count = CTA_MAX,
3517                                            .policy = ct_nla_policy },
3518        [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3519                                            .attr_count = CTA_MAX,
3520                                            .policy = ct_nla_policy },
3521        [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3522                                            .attr_count = CTA_MAX,
3523                                            .policy = ct_nla_policy },
3524        [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3525                                            .attr_count = CTA_MAX,
3526                                            .policy = ct_nla_policy },
3527        [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3528        [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3529        [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3530        [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3531};
3532
3533static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3534        [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3535                                            .attr_count = CTA_EXPECT_MAX,
3536                                            .policy = exp_nla_policy },
3537        [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3538                                            .attr_count = CTA_EXPECT_MAX,
3539                                            .policy = exp_nla_policy },
3540        [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3541                                            .attr_count = CTA_EXPECT_MAX,
3542                                            .policy = exp_nla_policy },
3543        [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3544};
3545
3546static const struct nfnetlink_subsystem ctnl_subsys = {
3547        .name                           = "conntrack",
3548        .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3549        .cb_count                       = IPCTNL_MSG_MAX,
3550        .cb                             = ctnl_cb,
3551};
3552
3553static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3554        .name                           = "conntrack_expect",
3555        .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3556        .cb_count                       = IPCTNL_MSG_EXP_MAX,
3557        .cb                             = ctnl_exp_cb,
3558};
3559
3560MODULE_ALIAS("ip_conntrack_netlink");
3561MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3562MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3563
3564static int __net_init ctnetlink_net_init(struct net *net)
3565{
3566#ifdef CONFIG_NF_CONNTRACK_EVENTS
3567        int ret;
3568
3569        ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3570        if (ret < 0) {
3571                pr_err("ctnetlink_init: cannot register notifier.\n");
3572                goto err_out;
3573        }
3574
3575        ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3576        if (ret < 0) {
3577                pr_err("ctnetlink_init: cannot expect register notifier.\n");
3578                goto err_unreg_notifier;
3579        }
3580#endif
3581        return 0;
3582
3583#ifdef CONFIG_NF_CONNTRACK_EVENTS
3584err_unreg_notifier:
3585        nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3586err_out:
3587        return ret;
3588#endif
3589}
3590
3591static void ctnetlink_net_exit(struct net *net)
3592{
3593#ifdef CONFIG_NF_CONNTRACK_EVENTS
3594        nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3595        nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3596#endif
3597}
3598
3599static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3600{
3601        struct net *net;
3602
3603        list_for_each_entry(net, net_exit_list, exit_list)
3604                ctnetlink_net_exit(net);
3605}
3606
3607static struct pernet_operations ctnetlink_net_ops = {
3608        .init           = ctnetlink_net_init,
3609        .exit_batch     = ctnetlink_net_exit_batch,
3610};
3611
3612static int __init ctnetlink_init(void)
3613{
3614        int ret;
3615
3616        ret = nfnetlink_subsys_register(&ctnl_subsys);
3617        if (ret < 0) {
3618                pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3619                goto err_out;
3620        }
3621
3622        ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3623        if (ret < 0) {
3624                pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3625                goto err_unreg_subsys;
3626        }
3627
3628        ret = register_pernet_subsys(&ctnetlink_net_ops);
3629        if (ret < 0) {
3630                pr_err("ctnetlink_init: cannot register pernet operations\n");
3631                goto err_unreg_exp_subsys;
3632        }
3633#ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3634        /* setup interaction between nf_queue and nf_conntrack_netlink. */
3635        RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3636#endif
3637        return 0;
3638
3639err_unreg_exp_subsys:
3640        nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3641err_unreg_subsys:
3642        nfnetlink_subsys_unregister(&ctnl_subsys);
3643err_out:
3644        return ret;
3645}
3646
3647static void __exit ctnetlink_exit(void)
3648{
3649        unregister_pernet_subsys(&ctnetlink_net_ops);
3650        nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3651        nfnetlink_subsys_unregister(&ctnl_subsys);
3652#ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3653        RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3654#endif
3655        synchronize_rcu();
3656}
3657
3658module_init(ctnetlink_init);
3659module_exit(ctnetlink_exit);
3660