File Index
Files and directories
Sign in
FILE

nf_conntrack_netlink.c

File preview with syntax highlighting.

linux-6.19.14/net/netfilter/nf_conntrack_netlink.c C
   1/* Connection tracking via netlink socket. Allows for user space
   2 * protocol helpers and general trouble making from userspace.
   3 *
   4 * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5 * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6 * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7 * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8 *
   9 * Initial connection tracking via netlink development funded and
  10 * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11 *
  12 * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13 *
  14 * This software may be used and distributed according to the terms
  15 * of the GNU General Public License, incorporated herein by reference.
  16 */
  17
  18#include <linux/init.h>
  19#include <linux/module.h>
  20#include <linux/kernel.h>
  21#include <linux/rculist.h>
  22#include <linux/rculist_nulls.h>
  23#include <linux/types.h>
  24#include <linux/timer.h>
  25#include <linux/security.h>
  26#include <linux/skbuff.h>
  27#include <linux/errno.h>
  28#include <linux/netlink.h>
  29#include <linux/spinlock.h>
  30#include <linux/interrupt.h>
  31#include <linux/slab.h>
  32#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
  57#include "nf_internals.h"
  58
  59MODULE_LICENSE("GPL");
  60MODULE_DESCRIPTION("List and change connection tracking table");
  61
  62struct ctnetlink_list_dump_ctx {
  63	unsigned long last_id;
  64	unsigned int cpu;
  65	bool done;
  66};
  67
  68static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  69				const struct nf_conntrack_tuple *tuple,
  70				const struct nf_conntrack_l4proto *l4proto)
  71{
  72	int ret = 0;
  73	struct nlattr *nest_parms;
  74
  75	nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO);
  76	if (!nest_parms)
  77		goto nla_put_failure;
  78	if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  79		goto nla_put_failure;
  80
  81	if (likely(l4proto->tuple_to_nlattr))
  82		ret = l4proto->tuple_to_nlattr(skb, tuple);
  83
  84	nla_nest_end(skb, nest_parms);
  85
  86	return ret;
  87
  88nla_put_failure:
  89	return -1;
  90}
  91
  92static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  93				const struct nf_conntrack_tuple *tuple)
  94{
  95	if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  96	    nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  97		return -EMSGSIZE;
  98	return 0;
  99}
 100
 101static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
 102				const struct nf_conntrack_tuple *tuple)
 103{
 104	if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
 105	    nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
 106		return -EMSGSIZE;
 107	return 0;
 108}
 109
 110static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 111				    const struct nf_conntrack_tuple *tuple)
 112{
 113	int ret = 0;
 114	struct nlattr *nest_parms;
 115
 116	nest_parms = nla_nest_start(skb, CTA_TUPLE_IP);
 117	if (!nest_parms)
 118		goto nla_put_failure;
 119
 120	switch (tuple->src.l3num) {
 121	case NFPROTO_IPV4:
 122		ret = ipv4_tuple_to_nlattr(skb, tuple);
 123		break;
 124	case NFPROTO_IPV6:
 125		ret = ipv6_tuple_to_nlattr(skb, tuple);
 126		break;
 127	}
 128
 129	nla_nest_end(skb, nest_parms);
 130
 131	return ret;
 132
 133nla_put_failure:
 134	return -1;
 135}
 136
 137static int ctnetlink_dump_tuples(struct sk_buff *skb,
 138				 const struct nf_conntrack_tuple *tuple)
 139{
 140	const struct nf_conntrack_l4proto *l4proto;
 141	int ret;
 142
 143	rcu_read_lock();
 144	ret = ctnetlink_dump_tuples_ip(skb, tuple);
 145
 146	if (ret >= 0) {
 147		l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 148		ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 149	}
 150	rcu_read_unlock();
 151	return ret;
 152}
 153
 154static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 155				  const struct nf_conntrack_zone *zone, int dir)
 156{
 157	if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 158		return 0;
 159	if (nla_put_be16(skb, attrtype, htons(zone->id)))
 160		goto nla_put_failure;
 161	return 0;
 162
 163nla_put_failure:
 164	return -1;
 165}
 166
 167static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 168{
 169	if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 170		goto nla_put_failure;
 171	return 0;
 172
 173nla_put_failure:
 174	return -1;
 175}
 176
 177static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct,
 178				  bool skip_zero)
 179{
 180	long timeout;
 181
 182	if (nf_ct_is_confirmed(ct))
 183		timeout = nf_ct_expires(ct) / HZ;
 184	else
 185		timeout = ct->timeout / HZ;
 186
 187	if (skip_zero && timeout == 0)
 188		return 0;
 189
 190	if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 191		goto nla_put_failure;
 192	return 0;
 193
 194nla_put_failure:
 195	return -1;
 196}
 197
 198static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct,
 199				    bool destroy)
 200{
 201	const struct nf_conntrack_l4proto *l4proto;
 202	struct nlattr *nest_proto;
 203	int ret;
 204
 205	l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 206	if (!l4proto->to_nlattr)
 207		return 0;
 208
 209	nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 210	if (!nest_proto)
 211		goto nla_put_failure;
 212
 213	ret = l4proto->to_nlattr(skb, nest_proto, ct, destroy);
 214
 215	nla_nest_end(skb, nest_proto);
 216
 217	return ret;
 218
 219nla_put_failure:
 220	return -1;
 221}
 222
 223static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 224				   const struct nf_conn *ct)
 225{
 226	struct nlattr *nest_helper;
 227	const struct nf_conn_help *help = nfct_help(ct);
 228	struct nf_conntrack_helper *helper;
 229
 230	if (!help)
 231		return 0;
 232
 233	rcu_read_lock();
 234	helper = rcu_dereference(help->helper);
 235	if (!helper)
 236		goto out;
 237
 238	nest_helper = nla_nest_start(skb, CTA_HELP);
 239	if (!nest_helper)
 240		goto nla_put_failure;
 241	if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 242		goto nla_put_failure;
 243
 244	if (helper->to_nlattr)
 245		helper->to_nlattr(skb, ct);
 246
 247	nla_nest_end(skb, nest_helper);
 248out:
 249	rcu_read_unlock();
 250	return 0;
 251
 252nla_put_failure:
 253	rcu_read_unlock();
 254	return -1;
 255}
 256
 257static int
 258dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 259	      enum ip_conntrack_dir dir, int type)
 260{
 261	enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 262	struct nf_conn_counter *counter = acct->counter;
 263	struct nlattr *nest_count;
 264	u64 pkts, bytes;
 265
 266	if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 267		pkts = atomic64_xchg(&counter[dir].packets, 0);
 268		bytes = atomic64_xchg(&counter[dir].bytes, 0);
 269	} else {
 270		pkts = atomic64_read(&counter[dir].packets);
 271		bytes = atomic64_read(&counter[dir].bytes);
 272	}
 273
 274	nest_count = nla_nest_start(skb, attr);
 275	if (!nest_count)
 276		goto nla_put_failure;
 277
 278	if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 279			 CTA_COUNTERS_PAD) ||
 280	    nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 281			 CTA_COUNTERS_PAD))
 282		goto nla_put_failure;
 283
 284	nla_nest_end(skb, nest_count);
 285
 286	return 0;
 287
 288nla_put_failure:
 289	return -1;
 290}
 291
 292static int
 293ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 294{
 295	struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 296
 297	if (!acct)
 298		return 0;
 299
 300	if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 301		return -1;
 302	if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 303		return -1;
 304
 305	return 0;
 306}
 307
 308static int
 309ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 310{
 311	struct nlattr *nest_count;
 312	const struct nf_conn_tstamp *tstamp;
 313
 314	tstamp = nf_conn_tstamp_find(ct);
 315	if (!tstamp)
 316		return 0;
 317
 318	nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 319	if (!nest_count)
 320		goto nla_put_failure;
 321
 322	if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 323			 CTA_TIMESTAMP_PAD) ||
 324	    (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 325					       cpu_to_be64(tstamp->stop),
 326					       CTA_TIMESTAMP_PAD)))
 327		goto nla_put_failure;
 328	nla_nest_end(skb, nest_count);
 329
 330	return 0;
 331
 332nla_put_failure:
 333	return -1;
 334}
 335
 336#ifdef CONFIG_NF_CONNTRACK_MARK
 337static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct,
 338			       bool dump)
 339{
 340	u32 mark = READ_ONCE(ct->mark);
 341
 342	if (!mark && !dump)
 343		return 0;
 344
 345	if (nla_put_be32(skb, CTA_MARK, htonl(mark)))
 346		goto nla_put_failure;
 347	return 0;
 348
 349nla_put_failure:
 350	return -1;
 351}
 352#else
 353#define ctnetlink_dump_mark(a, b, c) (0)
 354#endif
 355
 356#ifdef CONFIG_NF_CONNTRACK_SECMARK
 357static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 358{
 359	struct nlattr *nest_secctx;
 360	struct lsm_context ctx;
 361	int ret;
 362
 363	ret = security_secid_to_secctx(ct->secmark, &ctx);
 364	if (ret < 0)
 365		return 0;
 366
 367	ret = -1;
 368	nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 369	if (!nest_secctx)
 370		goto nla_put_failure;
 371
 372	if (nla_put_string(skb, CTA_SECCTX_NAME, ctx.context))
 373		goto nla_put_failure;
 374	nla_nest_end(skb, nest_secctx);
 375
 376	ret = 0;
 377nla_put_failure:
 378	security_release_secctx(&ctx);
 379	return ret;
 380}
 381#else
 382#define ctnetlink_dump_secctx(a, b) (0)
 383#endif
 384
 385#ifdef CONFIG_NF_CONNTRACK_EVENTS
 386static int
 387ctnetlink_dump_event_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 388{
 389#ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 390	const struct nf_conntrack_ecache *e = nf_ct_ecache_find(ct);
 391
 392	if (e) {
 393		u64 ts = local64_read(&e->timestamp);
 394
 395		if (ts)
 396			return nla_put_be64(skb, CTA_TIMESTAMP_EVENT,
 397					    cpu_to_be64(ts), CTA_TIMESTAMP_PAD);
 398	}
 399#endif
 400	return 0;
 401}
 402
 403static inline int ctnetlink_label_size(const struct nf_conn *ct)
 404{
 405	struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 406
 407	if (!labels)
 408		return 0;
 409	return nla_total_size(sizeof(labels->bits));
 410}
 411#endif
 412
 413static int
 414ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 415{
 416	struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 417	unsigned int i;
 418
 419	if (!labels)
 420		return 0;
 421
 422	i = 0;
 423	do {
 424		if (labels->bits[i] != 0)
 425			return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 426				       labels->bits);
 427		i++;
 428	} while (i < ARRAY_SIZE(labels->bits));
 429
 430	return 0;
 431}
 432
 433#define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 434
 435static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 436{
 437	struct nlattr *nest_parms;
 438
 439	if (!(ct->status & IPS_EXPECTED))
 440		return 0;
 441
 442	nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 443	if (!nest_parms)
 444		goto nla_put_failure;
 445	if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 446		goto nla_put_failure;
 447	nla_nest_end(skb, nest_parms);
 448
 449	return 0;
 450
 451nla_put_failure:
 452	return -1;
 453}
 454
 455static int
 456dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 457{
 458	struct nlattr *nest_parms;
 459
 460	nest_parms = nla_nest_start(skb, type);
 461	if (!nest_parms)
 462		goto nla_put_failure;
 463
 464	if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 465			 htonl(seq->correction_pos)) ||
 466	    nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 467			 htonl(seq->offset_before)) ||
 468	    nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 469			 htonl(seq->offset_after)))
 470		goto nla_put_failure;
 471
 472	nla_nest_end(skb, nest_parms);
 473
 474	return 0;
 475
 476nla_put_failure:
 477	return -1;
 478}
 479
 480static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 481{
 482	struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 483	struct nf_ct_seqadj *seq;
 484
 485	if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 486		return 0;
 487
 488	spin_lock_bh(&ct->lock);
 489	seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 490	if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 491		goto err;
 492
 493	seq = &seqadj->seq[IP_CT_DIR_REPLY];
 494	if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 495		goto err;
 496
 497	spin_unlock_bh(&ct->lock);
 498	return 0;
 499err:
 500	spin_unlock_bh(&ct->lock);
 501	return -1;
 502}
 503
 504static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 505{
 506	struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 507	struct nlattr *nest_parms;
 508
 509	if (!synproxy)
 510		return 0;
 511
 512	nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 513	if (!nest_parms)
 514		goto nla_put_failure;
 515
 516	if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 517	    nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 518	    nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 519		goto nla_put_failure;
 520
 521	nla_nest_end(skb, nest_parms);
 522
 523	return 0;
 524
 525nla_put_failure:
 526	return -1;
 527}
 528
 529static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 530{
 531	__be32 id = (__force __be32)nf_ct_get_id(ct);
 532
 533	if (nla_put_be32(skb, CTA_ID, id))
 534		goto nla_put_failure;
 535	return 0;
 536
 537nla_put_failure:
 538	return -1;
 539}
 540
 541static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 542{
 543	if (nla_put_be32(skb, CTA_USE, htonl(refcount_read(&ct->ct_general.use))))
 544		goto nla_put_failure;
 545	return 0;
 546
 547nla_put_failure:
 548	return -1;
 549}
 550
 551/* all these functions access ct->ext. Caller must either hold a reference
 552 * on ct or prevent its deletion by holding either the bucket spinlock or
 553 * pcpu dying list lock.
 554 */
 555static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 556				  struct nf_conn *ct, u32 type)
 557{
 558	if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 559	    ctnetlink_dump_timestamp(skb, ct) < 0 ||
 560	    ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 561	    ctnetlink_dump_labels(skb, ct) < 0 ||
 562	    ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 563	    ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 564		return -1;
 565
 566	return 0;
 567}
 568
 569static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 570{
 571	if (ctnetlink_dump_status(skb, ct) < 0 ||
 572	    ctnetlink_dump_mark(skb, ct, true) < 0 ||
 573	    ctnetlink_dump_secctx(skb, ct) < 0 ||
 574	    ctnetlink_dump_id(skb, ct) < 0 ||
 575	    ctnetlink_dump_use(skb, ct) < 0 ||
 576	    ctnetlink_dump_master(skb, ct) < 0)
 577		return -1;
 578
 579	if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 580	    (ctnetlink_dump_timeout(skb, ct, false) < 0 ||
 581	     ctnetlink_dump_protoinfo(skb, ct, false) < 0))
 582		return -1;
 583
 584	return 0;
 585}
 586
 587static int
 588ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 589		    struct nf_conn *ct, bool extinfo, unsigned int flags)
 590{
 591	const struct nf_conntrack_zone *zone;
 592	struct nlmsghdr *nlh;
 593	struct nlattr *nest_parms;
 594	unsigned int event;
 595
 596	if (portid)
 597		flags |= NLM_F_MULTI;
 598	event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 599	nlh = nfnl_msg_put(skb, portid, seq, event, flags, nf_ct_l3num(ct),
 600			   NFNETLINK_V0, 0);
 601	if (!nlh)
 602		goto nlmsg_failure;
 603
 604	zone = nf_ct_zone(ct);
 605
 606	nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 607	if (!nest_parms)
 608		goto nla_put_failure;
 609	if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 610		goto nla_put_failure;
 611	if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 612				   NF_CT_ZONE_DIR_ORIG) < 0)
 613		goto nla_put_failure;
 614	nla_nest_end(skb, nest_parms);
 615
 616	nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 617	if (!nest_parms)
 618		goto nla_put_failure;
 619	if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 620		goto nla_put_failure;
 621	if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 622				   NF_CT_ZONE_DIR_REPL) < 0)
 623		goto nla_put_failure;
 624	nla_nest_end(skb, nest_parms);
 625
 626	if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 627				   NF_CT_DEFAULT_ZONE_DIR) < 0)
 628		goto nla_put_failure;
 629
 630	if (ctnetlink_dump_info(skb, ct) < 0)
 631		goto nla_put_failure;
 632	if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 633		goto nla_put_failure;
 634
 635	nlmsg_end(skb, nlh);
 636	return skb->len;
 637
 638nlmsg_failure:
 639nla_put_failure:
 640	nlmsg_cancel(skb, nlh);
 641	return -1;
 642}
 643
 644static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 645	[CTA_IP_V4_SRC]	= { .type = NLA_U32 },
 646	[CTA_IP_V4_DST]	= { .type = NLA_U32 },
 647	[CTA_IP_V6_SRC]	= { .len = sizeof(__be32) * 4 },
 648	[CTA_IP_V6_DST]	= { .len = sizeof(__be32) * 4 },
 649};
 650
 651#if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 652static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 653{
 654	const struct nf_conntrack_l4proto *l4proto;
 655	size_t len, len4 = 0;
 656
 657	len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 658	len *= 3u; /* ORIG, REPLY, MASTER */
 659
 660	l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 661	len += l4proto->nlattr_size;
 662	if (l4proto->nlattr_tuple_size) {
 663		len4 = l4proto->nlattr_tuple_size();
 664		len4 *= 3u; /* ORIG, REPLY, MASTER */
 665	}
 666
 667	return len + len4;
 668}
 669
 670static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 671{
 672	if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 673		return 0;
 674	return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 675	       + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 676	       + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 677	       ;
 678}
 679
 680static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 681{
 682#ifdef CONFIG_NF_CONNTRACK_SECMARK
 683	int ret;
 684
 685	ret = security_secid_to_secctx(ct->secmark, NULL);
 686	if (ret < 0)
 687		return 0;
 688
 689	return nla_total_size(0) /* CTA_SECCTX */
 690	       + nla_total_size(sizeof(char) * ret); /* CTA_SECCTX_NAME */
 691#else
 692	return 0;
 693#endif
 694}
 695
 696static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 697{
 698#ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 699	if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 700		return 0;
 701	return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 702#else
 703	return 0;
 704#endif
 705}
 706#endif
 707
 708#ifdef CONFIG_NF_CONNTRACK_EVENTS
 709static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 710{
 711	return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 712	       + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 713	       + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 714	       + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 715	       + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 716	       + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 717	       + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 718	       + ctnetlink_acct_size(ct)
 719	       + ctnetlink_timestamp_size(ct)
 720	       + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 721	       + nla_total_size(0) /* CTA_PROTOINFO */
 722	       + nla_total_size(0) /* CTA_HELP */
 723	       + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 724	       + ctnetlink_secctx_size(ct)
 725#if IS_ENABLED(CONFIG_NF_NAT)
 726	       + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 727	       + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 728#endif
 729#ifdef CONFIG_NF_CONNTRACK_MARK
 730	       + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 731#endif
 732#ifdef CONFIG_NF_CONNTRACK_ZONES
 733	       + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 734#endif
 735	       + ctnetlink_proto_size(ct)
 736	       + ctnetlink_label_size(ct)
 737#ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 738	       + nla_total_size(sizeof(u64)) /* CTA_TIMESTAMP_EVENT */
 739#endif
 740	       ;
 741}
 742
 743static int
 744ctnetlink_conntrack_event(unsigned int events, const struct nf_ct_event *item)
 745{
 746	const struct nf_conntrack_zone *zone;
 747	struct net *net;
 748	struct nlmsghdr *nlh;
 749	struct nlattr *nest_parms;
 750	struct nf_conn *ct = item->ct;
 751	struct sk_buff *skb;
 752	unsigned int type;
 753	unsigned int flags = 0, group;
 754	int err;
 755
 756	if (events & (1 << IPCT_DESTROY)) {
 757		type = IPCTNL_MSG_CT_DELETE;
 758		group = NFNLGRP_CONNTRACK_DESTROY;
 759	} else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 760		type = IPCTNL_MSG_CT_NEW;
 761		flags = NLM_F_CREATE|NLM_F_EXCL;
 762		group = NFNLGRP_CONNTRACK_NEW;
 763	} else if (events) {
 764		type = IPCTNL_MSG_CT_NEW;
 765		group = NFNLGRP_CONNTRACK_UPDATE;
 766	} else
 767		return 0;
 768
 769	net = nf_ct_net(ct);
 770	if (!item->report && !nfnetlink_has_listeners(net, group))
 771		return 0;
 772
 773	skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 774	if (skb == NULL)
 775		goto errout;
 776
 777	type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 778	nlh = nfnl_msg_put(skb, item->portid, 0, type, flags, nf_ct_l3num(ct),
 779			   NFNETLINK_V0, 0);
 780	if (!nlh)
 781		goto nlmsg_failure;
 782
 783	zone = nf_ct_zone(ct);
 784
 785	nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 786	if (!nest_parms)
 787		goto nla_put_failure;
 788	if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 789		goto nla_put_failure;
 790	if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 791				   NF_CT_ZONE_DIR_ORIG) < 0)
 792		goto nla_put_failure;
 793	nla_nest_end(skb, nest_parms);
 794
 795	nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 796	if (!nest_parms)
 797		goto nla_put_failure;
 798	if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 799		goto nla_put_failure;
 800	if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 801				   NF_CT_ZONE_DIR_REPL) < 0)
 802		goto nla_put_failure;
 803	nla_nest_end(skb, nest_parms);
 804
 805	if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 806				   NF_CT_DEFAULT_ZONE_DIR) < 0)
 807		goto nla_put_failure;
 808
 809	if (ctnetlink_dump_id(skb, ct) < 0)
 810		goto nla_put_failure;
 811
 812	if (ctnetlink_dump_status(skb, ct) < 0)
 813		goto nla_put_failure;
 814
 815	if (events & (1 << IPCT_DESTROY)) {
 816		if (ctnetlink_dump_timeout(skb, ct, true) < 0)
 817			goto nla_put_failure;
 818
 819		if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 820		    ctnetlink_dump_timestamp(skb, ct) < 0 ||
 821		    ctnetlink_dump_protoinfo(skb, ct, true) < 0)
 822			goto nla_put_failure;
 823	} else {
 824		if (ctnetlink_dump_timeout(skb, ct, false) < 0)
 825			goto nla_put_failure;
 826
 827		if (events & (1 << IPCT_PROTOINFO) &&
 828		    ctnetlink_dump_protoinfo(skb, ct, false) < 0)
 829			goto nla_put_failure;
 830
 831		if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 832		    && ctnetlink_dump_helpinfo(skb, ct) < 0)
 833			goto nla_put_failure;
 834
 835#ifdef CONFIG_NF_CONNTRACK_SECMARK
 836		if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 837		    && ctnetlink_dump_secctx(skb, ct) < 0)
 838			goto nla_put_failure;
 839#endif
 840		if (events & (1 << IPCT_LABEL) &&
 841		     ctnetlink_dump_labels(skb, ct) < 0)
 842			goto nla_put_failure;
 843
 844		if (events & (1 << IPCT_RELATED) &&
 845		    ctnetlink_dump_master(skb, ct) < 0)
 846			goto nla_put_failure;
 847
 848		if (events & (1 << IPCT_SEQADJ) &&
 849		    ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 850			goto nla_put_failure;
 851
 852		if (events & (1 << IPCT_SYNPROXY) &&
 853		    ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 854			goto nla_put_failure;
 855	}
 856
 857#ifdef CONFIG_NF_CONNTRACK_MARK
 858	if (ctnetlink_dump_mark(skb, ct, events & (1 << IPCT_MARK)))
 859		goto nla_put_failure;
 860#endif
 861
 862	if (ctnetlink_dump_event_timestamp(skb, ct))
 863		goto nla_put_failure;
 864
 865	nlmsg_end(skb, nlh);
 866	err = nfnetlink_send(skb, net, item->portid, group, item->report,
 867			     GFP_ATOMIC);
 868	if (err == -ENOBUFS || err == -EAGAIN)
 869		return -ENOBUFS;
 870
 871	return 0;
 872
 873nla_put_failure:
 874	nlmsg_cancel(skb, nlh);
 875nlmsg_failure:
 876	kfree_skb(skb);
 877errout:
 878	if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 879		return -ENOBUFS;
 880
 881	return 0;
 882}
 883#endif /* CONFIG_NF_CONNTRACK_EVENTS */
 884
 885static int ctnetlink_done(struct netlink_callback *cb)
 886{
 887	kfree(cb->data);
 888	return 0;
 889}
 890
 891struct ctnetlink_filter_u32 {
 892	u32 val;
 893	u32 mask;
 894};
 895
 896struct ctnetlink_filter {
 897	u8 family;
 898	bool zone_filter;
 899
 900	u_int32_t orig_flags;
 901	u_int32_t reply_flags;
 902
 903	struct nf_conntrack_tuple orig;
 904	struct nf_conntrack_tuple reply;
 905	struct nf_conntrack_zone zone;
 906
 907	struct ctnetlink_filter_u32 mark;
 908	struct ctnetlink_filter_u32 status;
 909};
 910
 911static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 912	[CTA_FILTER_ORIG_FLAGS]		= NLA_POLICY_MASK(NLA_U32, CTA_FILTER_F_ALL),
 913	[CTA_FILTER_REPLY_FLAGS]	= NLA_POLICY_MASK(NLA_U32, CTA_FILTER_F_ALL),
 914};
 915
 916static int ctnetlink_parse_filter(const struct nlattr *attr,
 917				  struct ctnetlink_filter *filter)
 918{
 919	struct nlattr *tb[CTA_FILTER_MAX + 1];
 920	int ret = 0;
 921
 922	ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 923			       NULL);
 924	if (ret)
 925		return ret;
 926
 927	if (tb[CTA_FILTER_ORIG_FLAGS])
 928		filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 929
 930	if (tb[CTA_FILTER_REPLY_FLAGS])
 931		filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 932
 933	return 0;
 934}
 935
 936static int ctnetlink_parse_zone(const struct nlattr *attr,
 937				struct nf_conntrack_zone *zone);
 938static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 939					 struct nf_conntrack_tuple *tuple,
 940					 u32 type, u_int8_t l3num,
 941					 struct nf_conntrack_zone *zone,
 942					 u_int32_t flags);
 943
 944static int ctnetlink_filter_parse_mark(struct ctnetlink_filter_u32 *mark,
 945				       const struct nlattr * const cda[])
 946{
 947#ifdef CONFIG_NF_CONNTRACK_MARK
 948	if (cda[CTA_MARK]) {
 949		mark->val = ntohl(nla_get_be32(cda[CTA_MARK]));
 950
 951		if (cda[CTA_MARK_MASK])
 952			mark->mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 953		else
 954			mark->mask = 0xffffffff;
 955	} else if (cda[CTA_MARK_MASK]) {
 956		return -EINVAL;
 957	}
 958#endif
 959	return 0;
 960}
 961
 962static int ctnetlink_filter_parse_status(struct ctnetlink_filter_u32 *status,
 963					 const struct nlattr * const cda[])
 964{
 965	if (cda[CTA_STATUS]) {
 966		status->val = ntohl(nla_get_be32(cda[CTA_STATUS]));
 967		if (cda[CTA_STATUS_MASK])
 968			status->mask = ntohl(nla_get_be32(cda[CTA_STATUS_MASK]));
 969		else
 970			status->mask = status->val;
 971
 972		/* status->val == 0? always true, else always false. */
 973		if (status->mask == 0)
 974			return -EINVAL;
 975	} else if (cda[CTA_STATUS_MASK]) {
 976		return -EINVAL;
 977	}
 978
 979	/* CTA_STATUS is NLA_U32, if this fires UAPI needs to be extended */
 980	BUILD_BUG_ON(__IPS_MAX_BIT >= 32);
 981	return 0;
 982}
 983
 984static struct ctnetlink_filter *
 985ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 986{
 987	struct ctnetlink_filter *filter;
 988	int err;
 989
 990#ifndef CONFIG_NF_CONNTRACK_MARK
 991	if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 992		return ERR_PTR(-EOPNOTSUPP);
 993#endif
 994
 995	filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 996	if (filter == NULL)
 997		return ERR_PTR(-ENOMEM);
 998
 999	filter->family = family;
1000
1001	err = ctnetlink_filter_parse_mark(&filter->mark, cda);
1002	if (err)
1003		goto err_filter;
1004
1005	err = ctnetlink_filter_parse_status(&filter->status, cda);
1006	if (err)
1007		goto err_filter;
1008
1009	if (cda[CTA_ZONE]) {
1010		err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
1011		if (err < 0)
1012			goto err_filter;
1013		filter->zone_filter = true;
1014	}
1015
1016	if (!cda[CTA_FILTER])
1017		return filter;
1018
1019	err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
1020	if (err < 0)
1021		goto err_filter;
1022
1023	if (filter->orig_flags) {
1024		if (!cda[CTA_TUPLE_ORIG]) {
1025			err = -EINVAL;
1026			goto err_filter;
1027		}
1028
1029		err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
1030						   CTA_TUPLE_ORIG,
1031						   filter->family,
1032						   &filter->zone,
1033						   filter->orig_flags);
1034		if (err < 0)
1035			goto err_filter;
1036	}
1037
1038	if (filter->reply_flags) {
1039		if (!cda[CTA_TUPLE_REPLY]) {
1040			err = -EINVAL;
1041			goto err_filter;
1042		}
1043
1044		err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
1045						   CTA_TUPLE_REPLY,
1046						   filter->family,
1047						   &filter->zone,
1048						   filter->reply_flags);
1049		if (err < 0)
1050			goto err_filter;
1051	}
1052
1053	return filter;
1054
1055err_filter:
1056	kfree(filter);
1057
1058	return ERR_PTR(err);
1059}
1060
1061static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
1062{
1063	return family || cda[CTA_MARK] || cda[CTA_FILTER] || cda[CTA_STATUS] || cda[CTA_ZONE];
1064}
1065
1066static int ctnetlink_start(struct netlink_callback *cb)
1067{
1068	const struct nlattr * const *cda = cb->data;
1069	struct ctnetlink_filter *filter = NULL;
1070	struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1071	u8 family = nfmsg->nfgen_family;
1072
1073	if (ctnetlink_needs_filter(family, cda)) {
1074		filter = ctnetlink_alloc_filter(cda, family);
1075		if (IS_ERR(filter))
1076			return PTR_ERR(filter);
1077	}
1078
1079	cb->data = filter;
1080	return 0;
1081}
1082
1083static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1084					struct nf_conntrack_tuple *ct_tuple,
1085					u_int32_t flags, int family)
1086{
1087	switch (family) {
1088	case NFPROTO_IPV4:
1089		if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1090		    filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1091			return  0;
1092
1093		if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1094		    filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1095			return  0;
1096		break;
1097	case NFPROTO_IPV6:
1098		if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1099		    !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1100				   &ct_tuple->src.u3.in6))
1101			return 0;
1102
1103		if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1104		    !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1105				   &ct_tuple->dst.u3.in6))
1106			return 0;
1107		break;
1108	}
1109
1110	if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1111	    filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1112		return 0;
1113
1114	switch (ct_tuple->dst.protonum) {
1115	case IPPROTO_TCP:
1116	case IPPROTO_UDP:
1117		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1118		    filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1119			return 0;
1120
1121		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1122		    filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1123			return 0;
1124		break;
1125	case IPPROTO_ICMP:
1126		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1127		    filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1128			return 0;
1129		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1130		    filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1131			return 0;
1132		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1133		    filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1134			return 0;
1135		break;
1136	case IPPROTO_ICMPV6:
1137		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1138		    filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1139			return 0;
1140		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1141		    filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1142			return 0;
1143		if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1144		    filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1145			return 0;
1146		break;
1147	}
1148
1149	return 1;
1150}
1151
1152static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1153{
1154	struct ctnetlink_filter *filter = data;
1155	struct nf_conntrack_tuple *tuple;
1156	u32 status;
1157
1158	if (filter == NULL)
1159		goto out;
1160
1161	/* Match entries of a given L3 protocol number.
1162	 * If it is not specified, ie. l3proto == 0,
1163	 * then match everything.
1164	 */
1165	if (filter->family && nf_ct_l3num(ct) != filter->family)
1166		goto ignore_entry;
1167
1168	if (filter->zone_filter &&
1169	    !nf_ct_zone_equal_any(ct, &filter->zone))
1170		goto ignore_entry;
1171
1172	if (filter->orig_flags) {
1173		tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1174		if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1175						  filter->orig_flags,
1176						  filter->family))
1177			goto ignore_entry;
1178	}
1179
1180	if (filter->reply_flags) {
1181		tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1182		if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1183						  filter->reply_flags,
1184						  filter->family))
1185			goto ignore_entry;
1186	}
1187
1188#ifdef CONFIG_NF_CONNTRACK_MARK
1189	if ((READ_ONCE(ct->mark) & filter->mark.mask) != filter->mark.val)
1190		goto ignore_entry;
1191#endif
1192	status = (u32)READ_ONCE(ct->status);
1193	if ((status & filter->status.mask) != filter->status.val)
1194		goto ignore_entry;
1195
1196out:
1197	return 1;
1198
1199ignore_entry:
1200	return 0;
1201}
1202
1203static unsigned long ctnetlink_get_id(const struct nf_conn *ct)
1204{
1205	unsigned long id = nf_ct_get_id(ct);
1206
1207	return id ? id : 1;
1208}
1209
1210static int
1211ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1212{
1213	unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1214	struct net *net = sock_net(skb->sk);
1215	unsigned long last_id = cb->args[1];
1216	struct nf_conntrack_tuple_hash *h;
1217	struct hlist_nulls_node *n;
1218	struct nf_conn *nf_ct_evict[8];
1219	struct nf_conn *ct;
1220	int res, i;
1221	spinlock_t *lockp;
1222
1223	i = 0;
1224
1225	local_bh_disable();
1226	for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1227restart:
1228		while (i) {
1229			i--;
1230			if (nf_ct_should_gc(nf_ct_evict[i]))
1231				nf_ct_kill(nf_ct_evict[i]);
1232			nf_ct_put(nf_ct_evict[i]);
1233		}
1234
1235		lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1236		nf_conntrack_lock(lockp);
1237		if (cb->args[0] >= nf_conntrack_htable_size) {
1238			spin_unlock(lockp);
1239			goto out;
1240		}
1241		hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1242					   hnnode) {
1243			ct = nf_ct_tuplehash_to_ctrack(h);
1244			if (nf_ct_is_expired(ct)) {
1245				/* need to defer nf_ct_kill() until lock is released */
1246				if (i < ARRAY_SIZE(nf_ct_evict) &&
1247				    refcount_inc_not_zero(&ct->ct_general.use))
1248					nf_ct_evict[i++] = ct;
1249				continue;
1250			}
1251
1252			if (!net_eq(net, nf_ct_net(ct)))
1253				continue;
1254
1255			if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1256				continue;
1257
1258			if (cb->args[1]) {
1259				if (ctnetlink_get_id(ct) != last_id)
1260					continue;
1261				cb->args[1] = 0;
1262			}
1263			if (!ctnetlink_filter_match(ct, cb->data))
1264				continue;
1265
1266			res =
1267			ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1268					    cb->nlh->nlmsg_seq,
1269					    NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1270					    ct, true, flags);
1271			if (res < 0) {
1272				cb->args[1] = ctnetlink_get_id(ct);
1273				spin_unlock(lockp);
1274				goto out;
1275			}
1276		}
1277		spin_unlock(lockp);
1278		if (cb->args[1]) {
1279			cb->args[1] = 0;
1280			goto restart;
1281		}
1282	}
1283out:
1284	local_bh_enable();
1285	if (last_id) {
1286		/* nf ct hash resize happened, now clear the leftover. */
1287		if (cb->args[1] == last_id)
1288			cb->args[1] = 0;
1289	}
1290
1291	while (i) {
1292		i--;
1293		if (nf_ct_should_gc(nf_ct_evict[i]))
1294			nf_ct_kill(nf_ct_evict[i]);
1295		nf_ct_put(nf_ct_evict[i]);
1296	}
1297
1298	return skb->len;
1299}
1300
1301static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1302				struct nf_conntrack_tuple *t,
1303				u_int32_t flags)
1304{
1305	if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1306		if (!tb[CTA_IP_V4_SRC])
1307			return -EINVAL;
1308
1309		t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1310	}
1311
1312	if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1313		if (!tb[CTA_IP_V4_DST])
1314			return -EINVAL;
1315
1316		t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1317	}
1318
1319	return 0;
1320}
1321
1322static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1323				struct nf_conntrack_tuple *t,
1324				u_int32_t flags)
1325{
1326	if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1327		if (!tb[CTA_IP_V6_SRC])
1328			return -EINVAL;
1329
1330		t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1331	}
1332
1333	if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1334		if (!tb[CTA_IP_V6_DST])
1335			return -EINVAL;
1336
1337		t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1338	}
1339
1340	return 0;
1341}
1342
1343static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1344				    struct nf_conntrack_tuple *tuple,
1345				    u_int32_t flags)
1346{
1347	struct nlattr *tb[CTA_IP_MAX+1];
1348	int ret = 0;
1349
1350	ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr,
1351					  cta_ip_nla_policy, NULL);
1352	if (ret < 0)
1353		return ret;
1354
1355	switch (tuple->src.l3num) {
1356	case NFPROTO_IPV4:
1357		ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1358		break;
1359	case NFPROTO_IPV6:
1360		ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1361		break;
1362	}
1363
1364	return ret;
1365}
1366
1367static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1368	[CTA_PROTO_NUM]	= { .type = NLA_U8 },
1369};
1370
1371static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1372				       struct nf_conntrack_tuple *tuple,
1373				       u_int32_t flags)
1374{
1375	const struct nf_conntrack_l4proto *l4proto;
1376	struct nlattr *tb[CTA_PROTO_MAX+1];
1377	int ret = 0;
1378
1379	ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1380					  proto_nla_policy, NULL);
1381	if (ret < 0)
1382		return ret;
1383
1384	if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1385		return 0;
1386
1387	if (!tb[CTA_PROTO_NUM])
1388		return -EINVAL;
1389
1390	tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1391
1392	rcu_read_lock();
1393	l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1394
1395	if (likely(l4proto->nlattr_to_tuple)) {
1396		ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1397						     l4proto->nla_policy,
1398						     NULL);
1399		if (ret == 0)
1400			ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1401	}
1402
1403	rcu_read_unlock();
1404
1405	return ret;
1406}
1407
1408static int
1409ctnetlink_parse_zone(const struct nlattr *attr,
1410		     struct nf_conntrack_zone *zone)
1411{
1412	nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1413			NF_CT_DEFAULT_ZONE_DIR, 0);
1414#ifdef CONFIG_NF_CONNTRACK_ZONES
1415	if (attr)
1416		zone->id = ntohs(nla_get_be16(attr));
1417#else
1418	if (attr)
1419		return -EOPNOTSUPP;
1420#endif
1421	return 0;
1422}
1423
1424static int
1425ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1426			   struct nf_conntrack_zone *zone)
1427{
1428	int ret;
1429
1430	if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1431		return -EINVAL;
1432
1433	ret = ctnetlink_parse_zone(attr, zone);
1434	if (ret < 0)
1435		return ret;
1436
1437	if (type == CTA_TUPLE_REPLY)
1438		zone->dir = NF_CT_ZONE_DIR_REPL;
1439	else
1440		zone->dir = NF_CT_ZONE_DIR_ORIG;
1441
1442	return 0;
1443}
1444
1445static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1446	[CTA_TUPLE_IP]		= { .type = NLA_NESTED },
1447	[CTA_TUPLE_PROTO]	= { .type = NLA_NESTED },
1448	[CTA_TUPLE_ZONE]	= { .type = NLA_U16 },
1449};
1450
1451#define CTA_FILTER_F_ALL_CTA_PROTO \
1452  (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1453   CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1454   CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1455   CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1456   CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1457   CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1458   CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1459   CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1460
1461static int
1462ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1463			      struct nf_conntrack_tuple *tuple, u32 type,
1464			      u_int8_t l3num, struct nf_conntrack_zone *zone,
1465			      u_int32_t flags)
1466{
1467	struct nlattr *tb[CTA_TUPLE_MAX+1];
1468	int err;
1469
1470	memset(tuple, 0, sizeof(*tuple));
1471
1472	err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1473					  tuple_nla_policy, NULL);
1474	if (err < 0)
1475		return err;
1476
1477	if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1478		return -EOPNOTSUPP;
1479	tuple->src.l3num = l3num;
1480
1481	if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1482	    flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1483		if (!tb[CTA_TUPLE_IP])
1484			return -EINVAL;
1485
1486		err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1487		if (err < 0)
1488			return err;
1489	}
1490
1491	if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1492		if (!tb[CTA_TUPLE_PROTO])
1493			return -EINVAL;
1494
1495		err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1496		if (err < 0)
1497			return err;
1498	} else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1499		/* Can't manage proto flags without a protonum  */
1500		return -EINVAL;
1501	}
1502
1503	if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1504		if (!zone)
1505			return -EINVAL;
1506
1507		err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1508						 type, zone);
1509		if (err < 0)
1510			return err;
1511	}
1512
1513	/* orig and expect tuples get DIR_ORIGINAL */
1514	if (type == CTA_TUPLE_REPLY)
1515		tuple->dst.dir = IP_CT_DIR_REPLY;
1516	else
1517		tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1518
1519	return 0;
1520}
1521
1522static int
1523ctnetlink_parse_tuple(const struct nlattr * const cda[],
1524		      struct nf_conntrack_tuple *tuple, u32 type,
1525		      u_int8_t l3num, struct nf_conntrack_zone *zone)
1526{
1527	return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1528					    CTA_FILTER_FLAG(ALL));
1529}
1530
1531static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1532	[CTA_HELP_NAME]		= { .type = NLA_NUL_STRING,
1533				    .len = NF_CT_HELPER_NAME_LEN - 1 },
1534};
1535
1536static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1537				struct nlattr **helpinfo)
1538{
1539	int err;
1540	struct nlattr *tb[CTA_HELP_MAX+1];
1541
1542	err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1543					  help_nla_policy, NULL);
1544	if (err < 0)
1545		return err;
1546
1547	if (!tb[CTA_HELP_NAME])
1548		return -EINVAL;
1549
1550	*helper_name = nla_data(tb[CTA_HELP_NAME]);
1551
1552	if (tb[CTA_HELP_INFO])
1553		*helpinfo = tb[CTA_HELP_INFO];
1554
1555	return 0;
1556}
1557
1558static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1559	[CTA_TUPLE_ORIG]	= { .type = NLA_NESTED },
1560	[CTA_TUPLE_REPLY]	= { .type = NLA_NESTED },
1561	[CTA_STATUS] 		= { .type = NLA_U32 },
1562	[CTA_PROTOINFO]		= { .type = NLA_NESTED },
1563	[CTA_HELP]		= { .type = NLA_NESTED },
1564	[CTA_NAT_SRC]		= { .type = NLA_NESTED },
1565	[CTA_TIMEOUT] 		= { .type = NLA_U32 },
1566	[CTA_MARK]		= { .type = NLA_U32 },
1567	[CTA_ID]		= { .type = NLA_U32 },
1568	[CTA_NAT_DST]		= { .type = NLA_NESTED },
1569	[CTA_TUPLE_MASTER]	= { .type = NLA_NESTED },
1570	[CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1571	[CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1572	[CTA_ZONE]		= { .type = NLA_U16 },
1573	[CTA_MARK_MASK]		= { .type = NLA_U32 },
1574	[CTA_LABELS]		= { .type = NLA_BINARY,
1575				    .len = NF_CT_LABELS_MAX_SIZE },
1576	[CTA_LABELS_MASK]	= { .type = NLA_BINARY,
1577				    .len = NF_CT_LABELS_MAX_SIZE },
1578	[CTA_FILTER]		= { .type = NLA_NESTED },
1579	[CTA_STATUS_MASK]	= { .type = NLA_U32 },
1580	[CTA_TIMESTAMP_EVENT]	= { .type = NLA_REJECT },
1581};
1582
1583static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1584{
1585	return ctnetlink_filter_match(ct, data);
1586}
1587
1588static int ctnetlink_flush_conntrack(struct net *net,
1589				     const struct nlattr * const cda[],
1590				     u32 portid, int report, u8 family)
1591{
1592	struct ctnetlink_filter *filter = NULL;
1593	struct nf_ct_iter_data iter = {
1594		.net		= net,
1595		.portid		= portid,
1596		.report		= report,
1597	};
1598
1599	if (ctnetlink_needs_filter(family, cda)) {
1600		filter = ctnetlink_alloc_filter(cda, family);
1601		if (IS_ERR(filter))
1602			return PTR_ERR(filter);
1603
1604		iter.data = filter;
1605	}
1606
1607	nf_ct_iterate_cleanup_net(ctnetlink_flush_iterate, &iter);
1608	kfree(filter);
1609
1610	return 0;
1611}
1612
1613static int ctnetlink_del_conntrack(struct sk_buff *skb,
1614				   const struct nfnl_info *info,
1615				   const struct nlattr * const cda[])
1616{
1617	u8 family = info->nfmsg->nfgen_family;
1618	struct nf_conntrack_tuple_hash *h;
1619	struct nf_conntrack_tuple tuple;
1620	struct nf_conntrack_zone zone;
1621	struct nf_conn *ct;
1622	int err;
1623
1624	err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1625	if (err < 0)
1626		return err;
1627
1628	if (cda[CTA_TUPLE_ORIG] && !cda[CTA_FILTER])
1629		err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1630					    family, &zone);
1631	else if (cda[CTA_TUPLE_REPLY] && !cda[CTA_FILTER])
1632		err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1633					    family, &zone);
1634	else {
1635		u8 u3 = info->nfmsg->version || cda[CTA_FILTER] ? family : AF_UNSPEC;
1636
1637		return ctnetlink_flush_conntrack(info->net, cda,
1638						 NETLINK_CB(skb).portid,
1639						 nlmsg_report(info->nlh), u3);
1640	}
1641
1642	if (err < 0)
1643		return err;
1644
1645	h = nf_conntrack_find_get(info->net, &zone, &tuple);
1646	if (!h)
1647		return -ENOENT;
1648
1649	ct = nf_ct_tuplehash_to_ctrack(h);
1650
1651	if (cda[CTA_ID]) {
1652		__be32 id = nla_get_be32(cda[CTA_ID]);
1653
1654		if (id != (__force __be32)nf_ct_get_id(ct)) {
1655			nf_ct_put(ct);
1656			return -ENOENT;
1657		}
1658	}
1659
1660	nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(info->nlh));
1661	nf_ct_put(ct);
1662
1663	return 0;
1664}
1665
1666static int ctnetlink_get_conntrack(struct sk_buff *skb,
1667				   const struct nfnl_info *info,
1668				   const struct nlattr * const cda[])
1669{
1670	u_int8_t u3 = info->nfmsg->nfgen_family;
1671	struct nf_conntrack_tuple_hash *h;
1672	struct nf_conntrack_tuple tuple;
1673	struct nf_conntrack_zone zone;
1674	struct sk_buff *skb2;
1675	struct nf_conn *ct;
1676	int err;
1677
1678	if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1679		struct netlink_dump_control c = {
1680			.start = ctnetlink_start,
1681			.dump = ctnetlink_dump_table,
1682			.done = ctnetlink_done,
1683			.data = (void *)cda,
1684		};
1685
1686		return netlink_dump_start(info->sk, skb, info->nlh, &c);
1687	}
1688
1689	err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1690	if (err < 0)
1691		return err;
1692
1693	if (cda[CTA_TUPLE_ORIG])
1694		err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1695					    u3, &zone);
1696	else if (cda[CTA_TUPLE_REPLY])
1697		err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1698					    u3, &zone);
1699	else
1700		return -EINVAL;
1701
1702	if (err < 0)
1703		return err;
1704
1705	h = nf_conntrack_find_get(info->net, &zone, &tuple);
1706	if (!h)
1707		return -ENOENT;
1708
1709	ct = nf_ct_tuplehash_to_ctrack(h);
1710
1711	skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1712	if (!skb2) {
1713		nf_ct_put(ct);
1714		return -ENOMEM;
1715	}
1716
1717	err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid,
1718				  info->nlh->nlmsg_seq,
1719				  NFNL_MSG_TYPE(info->nlh->nlmsg_type), ct,
1720				  true, 0);
1721	nf_ct_put(ct);
1722	if (err <= 0) {
1723		kfree_skb(skb2);
1724		return -ENOMEM;
1725	}
1726
1727	return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
1728}
1729
1730#ifdef CONFIG_NF_CONNTRACK_EVENTS
1731static int ctnetlink_dump_one_entry(struct sk_buff *skb,
1732				    struct netlink_callback *cb,
1733				    struct nf_conn *ct,
1734				    bool dying)
1735{
1736	struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1737	struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1738	u8 l3proto = nfmsg->nfgen_family;
1739	int res;
1740
1741	if (l3proto && nf_ct_l3num(ct) != l3proto)
1742		return 0;
1743
1744	if (ctx->last_id) {
1745		if (ctnetlink_get_id(ct) != ctx->last_id)
1746			return 0;
1747
1748		ctx->last_id = 0;
1749	}
1750
1751	/* We can't dump extension info for the unconfirmed
1752	 * list because unconfirmed conntracks can have
1753	 * ct->ext reallocated (and thus freed).
1754	 *
1755	 * In the dying list case ct->ext can't be free'd
1756	 * until after we drop pcpu->lock.
1757	 */
1758	res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1759				  cb->nlh->nlmsg_seq,
1760				  NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1761				  ct, dying, 0);
1762	if (res < 0)
1763		ctx->last_id = ctnetlink_get_id(ct);
1764
1765	return res;
1766}
1767#endif
1768
1769static int
1770ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1771{
1772	return 0;
1773}
1774
1775static int
1776ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1777{
1778	struct ctnetlink_list_dump_ctx *ctx = (void *)cb->ctx;
1779#ifdef CONFIG_NF_CONNTRACK_EVENTS
1780	const struct net *net = sock_net(skb->sk);
1781	struct nf_conntrack_net_ecache *ecache_net;
1782	unsigned long last_id = ctx->last_id;
1783	struct nf_conntrack_tuple_hash *h;
1784	struct hlist_nulls_node *n;
1785#endif
1786
1787	if (ctx->done)
1788		return 0;
1789
1790	ctx->last_id = 0;
1791
1792#ifdef CONFIG_NF_CONNTRACK_EVENTS
1793	ecache_net = nf_conn_pernet_ecache(net);
1794	spin_lock_bh(&ecache_net->dying_lock);
1795
1796	hlist_nulls_for_each_entry(h, n, &ecache_net->dying_list, hnnode) {
1797		struct nf_conn *ct;
1798		int res;
1799
1800		ct = nf_ct_tuplehash_to_ctrack(h);
1801		if (last_id && last_id != ctnetlink_get_id(ct))
1802			continue;
1803
1804		res = ctnetlink_dump_one_entry(skb, cb, ct, true);
1805		if (res < 0) {
1806			spin_unlock_bh(&ecache_net->dying_lock);
1807			return skb->len;
1808		}
1809
1810		last_id = 0;
1811	}
1812
1813	spin_unlock_bh(&ecache_net->dying_lock);
1814#endif
1815	ctx->done = true;
1816
1817	return skb->len;
1818}
1819
1820static int ctnetlink_get_ct_dying(struct sk_buff *skb,
1821				  const struct nfnl_info *info,
1822				  const struct nlattr * const cda[])
1823{
1824	if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1825		struct netlink_dump_control c = {
1826			.dump = ctnetlink_dump_dying,
1827		};
1828		return netlink_dump_start(info->sk, skb, info->nlh, &c);
1829	}
1830
1831	return -EOPNOTSUPP;
1832}
1833
1834static int ctnetlink_get_ct_unconfirmed(struct sk_buff *skb,
1835					const struct nfnl_info *info,
1836					const struct nlattr * const cda[])
1837{
1838	if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1839		struct netlink_dump_control c = {
1840			.dump = ctnetlink_dump_unconfirmed,
1841		};
1842		return netlink_dump_start(info->sk, skb, info->nlh, &c);
1843	}
1844
1845	return -EOPNOTSUPP;
1846}
1847
1848#if IS_ENABLED(CONFIG_NF_NAT)
1849static int
1850ctnetlink_parse_nat_setup(struct nf_conn *ct,
1851			  enum nf_nat_manip_type manip,
1852			  const struct nlattr *attr)
1853	__must_hold(RCU)
1854{
1855	const struct nf_nat_hook *nat_hook;
1856	int err;
1857
1858	nat_hook = rcu_dereference(nf_nat_hook);
1859	if (!nat_hook) {
1860#ifdef CONFIG_MODULES
1861		rcu_read_unlock();
1862		nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1863		if (request_module("nf-nat") < 0) {
1864			nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1865			rcu_read_lock();
1866			return -EOPNOTSUPP;
1867		}
1868		nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1869		rcu_read_lock();
1870		nat_hook = rcu_dereference(nf_nat_hook);
1871		if (nat_hook)
1872			return -EAGAIN;
1873#endif
1874		return -EOPNOTSUPP;
1875	}
1876
1877	err = nat_hook->parse_nat_setup(ct, manip, attr);
1878	if (err == -EAGAIN) {
1879#ifdef CONFIG_MODULES
1880		rcu_read_unlock();
1881		nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1882		if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1883			nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1884			rcu_read_lock();
1885			return -EOPNOTSUPP;
1886		}
1887		nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1888		rcu_read_lock();
1889#else
1890		err = -EOPNOTSUPP;
1891#endif
1892	}
1893	return err;
1894}
1895#endif
1896
1897static int
1898ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1899{
1900	return nf_ct_change_status_common(ct, ntohl(nla_get_be32(cda[CTA_STATUS])));
1901}
1902
1903static int
1904ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1905{
1906#if IS_ENABLED(CONFIG_NF_NAT)
1907	int ret;
1908
1909	if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1910		return 0;
1911
1912	ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1913					cda[CTA_NAT_DST]);
1914	if (ret < 0)
1915		return ret;
1916
1917	return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1918					 cda[CTA_NAT_SRC]);
1919#else
1920	if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1921		return 0;
1922	return -EOPNOTSUPP;
1923#endif
1924}
1925
1926static int ctnetlink_change_helper(struct nf_conn *ct,
1927				   const struct nlattr * const cda[])
1928{
1929	struct nf_conntrack_helper *helper;
1930	struct nf_conn_help *help = nfct_help(ct);
1931	char *helpname = NULL;
1932	struct nlattr *helpinfo = NULL;
1933	int err;
1934
1935	err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1936	if (err < 0)
1937		return err;
1938
1939	/* don't change helper of sibling connections */
1940	if (ct->master) {
1941		/* If we try to change the helper to the same thing twice,
1942		 * treat the second attempt as a no-op instead of returning
1943		 * an error.
1944		 */
1945		err = -EBUSY;
1946		if (help) {
1947			rcu_read_lock();
1948			helper = rcu_dereference(help->helper);
1949			if (helper && !strcmp(helper->name, helpname))
1950				err = 0;
1951			rcu_read_unlock();
1952		}
1953
1954		return err;
1955	}
1956
1957	if (!strcmp(helpname, "")) {
1958		if (help && help->helper) {
1959			/* we had a helper before ... */
1960			nf_ct_remove_expectations(ct);
1961			RCU_INIT_POINTER(help->helper, NULL);
1962		}
1963
1964		return 0;
1965	}
1966
1967	rcu_read_lock();
1968	helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1969					    nf_ct_protonum(ct));
1970	if (helper == NULL) {
1971		rcu_read_unlock();
1972		return -EOPNOTSUPP;
1973	}
1974
1975	if (help) {
1976		if (rcu_access_pointer(help->helper) == helper) {
1977			/* update private helper data if allowed. */
1978			if (helper->from_nlattr)
1979				helper->from_nlattr(helpinfo, ct);
1980			err = 0;
1981		} else
1982			err = -EBUSY;
1983	} else {
1984		/* we cannot set a helper for an existing conntrack */
1985		err = -EOPNOTSUPP;
1986	}
1987
1988	rcu_read_unlock();
1989	return err;
1990}
1991
1992static int ctnetlink_change_timeout(struct nf_conn *ct,
1993				    const struct nlattr * const cda[])
1994{
1995	return __nf_ct_change_timeout(ct, (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ);
1996}
1997
1998#if defined(CONFIG_NF_CONNTRACK_MARK)
1999static void ctnetlink_change_mark(struct nf_conn *ct,
2000				    const struct nlattr * const cda[])
2001{
2002	u32 mark, newmark, mask = 0;
2003
2004	if (cda[CTA_MARK_MASK])
2005		mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
2006
2007	mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2008	newmark = (READ_ONCE(ct->mark) & mask) ^ mark;
2009	if (newmark != READ_ONCE(ct->mark))
2010		WRITE_ONCE(ct->mark, newmark);
2011}
2012#endif
2013
2014static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
2015	[CTA_PROTOINFO_TCP]	= { .type = NLA_NESTED },
2016	[CTA_PROTOINFO_SCTP]	= { .type = NLA_NESTED },
2017};
2018
2019static int ctnetlink_change_protoinfo(struct nf_conn *ct,
2020				      const struct nlattr * const cda[])
2021{
2022	const struct nlattr *attr = cda[CTA_PROTOINFO];
2023	const struct nf_conntrack_l4proto *l4proto;
2024	struct nlattr *tb[CTA_PROTOINFO_MAX+1];
2025	int err = 0;
2026
2027	err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2028					  protoinfo_policy, NULL);
2029	if (err < 0)
2030		return err;
2031
2032	l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2033	if (l4proto->from_nlattr)
2034		err = l4proto->from_nlattr(tb, ct);
2035
2036	return err;
2037}
2038
2039static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2040	[CTA_SEQADJ_CORRECTION_POS]	= { .type = NLA_U32 },
2041	[CTA_SEQADJ_OFFSET_BEFORE]	= { .type = NLA_U32 },
2042	[CTA_SEQADJ_OFFSET_AFTER]	= { .type = NLA_U32 },
2043};
2044
2045static int change_seq_adj(struct nf_ct_seqadj *seq,
2046			  const struct nlattr * const attr)
2047{
2048	int err;
2049	struct nlattr *cda[CTA_SEQADJ_MAX+1];
2050
2051	err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2052					  seqadj_policy, NULL);
2053	if (err < 0)
2054		return err;
2055
2056	if (!cda[CTA_SEQADJ_CORRECTION_POS])
2057		return -EINVAL;
2058
2059	seq->correction_pos =
2060		ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2061
2062	if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2063		return -EINVAL;
2064
2065	seq->offset_before =
2066		ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2067
2068	if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2069		return -EINVAL;
2070
2071	seq->offset_after =
2072		ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2073
2074	return 0;
2075}
2076
2077static int
2078ctnetlink_change_seq_adj(struct nf_conn *ct,
2079			 const struct nlattr * const cda[])
2080{
2081	struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2082	int ret = 0;
2083
2084	if (!seqadj)
2085		return 0;
2086
2087	spin_lock_bh(&ct->lock);
2088	if (cda[CTA_SEQ_ADJ_ORIG]) {
2089		ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2090				     cda[CTA_SEQ_ADJ_ORIG]);
2091		if (ret < 0)
2092			goto err;
2093
2094		set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2095	}
2096
2097	if (cda[CTA_SEQ_ADJ_REPLY]) {
2098		ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2099				     cda[CTA_SEQ_ADJ_REPLY]);
2100		if (ret < 0)
2101			goto err;
2102
2103		set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2104	}
2105
2106	spin_unlock_bh(&ct->lock);
2107	return 0;
2108err:
2109	spin_unlock_bh(&ct->lock);
2110	return ret;
2111}
2112
2113static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2114	[CTA_SYNPROXY_ISN]	= { .type = NLA_U32 },
2115	[CTA_SYNPROXY_ITS]	= { .type = NLA_U32 },
2116	[CTA_SYNPROXY_TSOFF]	= { .type = NLA_U32 },
2117};
2118
2119static int ctnetlink_change_synproxy(struct nf_conn *ct,
2120				     const struct nlattr * const cda[])
2121{
2122	struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2123	struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2124	int err;
2125
2126	if (!synproxy)
2127		return 0;
2128
2129	err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2130					  cda[CTA_SYNPROXY], synproxy_policy,
2131					  NULL);
2132	if (err < 0)
2133		return err;
2134
2135	if (!tb[CTA_SYNPROXY_ISN] ||
2136	    !tb[CTA_SYNPROXY_ITS] ||
2137	    !tb[CTA_SYNPROXY_TSOFF])
2138		return -EINVAL;
2139
2140	synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2141	synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2142	synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2143
2144	return 0;
2145}
2146
2147static int
2148ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2149{
2150#ifdef CONFIG_NF_CONNTRACK_LABELS
2151	size_t len = nla_len(cda[CTA_LABELS]);
2152	const void *mask = cda[CTA_LABELS_MASK];
2153
2154	if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2155		return -EINVAL;
2156
2157	if (mask) {
2158		if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2159		    nla_len(cda[CTA_LABELS_MASK]) != len)
2160			return -EINVAL;
2161		mask = nla_data(cda[CTA_LABELS_MASK]);
2162	}
2163
2164	len /= sizeof(u32);
2165
2166	return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2167#else
2168	return -EOPNOTSUPP;
2169#endif
2170}
2171
2172static int
2173ctnetlink_change_conntrack(struct nf_conn *ct,
2174			   const struct nlattr * const cda[])
2175{
2176	int err;
2177
2178	/* only allow NAT changes and master assignation for new conntracks */
2179	if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2180		return -EOPNOTSUPP;
2181
2182	if (cda[CTA_HELP]) {
2183		err = ctnetlink_change_helper(ct, cda);
2184		if (err < 0)
2185			return err;
2186	}
2187
2188	if (cda[CTA_TIMEOUT]) {
2189		err = ctnetlink_change_timeout(ct, cda);
2190		if (err < 0)
2191			return err;
2192	}
2193
2194	if (cda[CTA_STATUS]) {
2195		err = ctnetlink_change_status(ct, cda);
2196		if (err < 0)
2197			return err;
2198	}
2199
2200	if (cda[CTA_PROTOINFO]) {
2201		err = ctnetlink_change_protoinfo(ct, cda);
2202		if (err < 0)
2203			return err;
2204	}
2205
2206#if defined(CONFIG_NF_CONNTRACK_MARK)
2207	if (cda[CTA_MARK])
2208		ctnetlink_change_mark(ct, cda);
2209#endif
2210
2211	if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2212		err = ctnetlink_change_seq_adj(ct, cda);
2213		if (err < 0)
2214			return err;
2215	}
2216
2217	if (cda[CTA_SYNPROXY]) {
2218		err = ctnetlink_change_synproxy(ct, cda);
2219		if (err < 0)
2220			return err;
2221	}
2222
2223	if (cda[CTA_LABELS]) {
2224		err = ctnetlink_attach_labels(ct, cda);
2225		if (err < 0)
2226			return err;
2227	}
2228
2229	return 0;
2230}
2231
2232static struct nf_conn *
2233ctnetlink_create_conntrack(struct net *net,
2234			   const struct nf_conntrack_zone *zone,
2235			   const struct nlattr * const cda[],
2236			   struct nf_conntrack_tuple *otuple,
2237			   struct nf_conntrack_tuple *rtuple,
2238			   u8 u3)
2239{
2240	struct nf_conn *ct;
2241	int err = -EINVAL;
2242	struct nf_conntrack_helper *helper;
2243	struct nf_conn_tstamp *tstamp;
2244	u64 timeout;
2245
2246	ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2247	if (IS_ERR(ct))
2248		return ERR_PTR(-ENOMEM);
2249
2250	if (!cda[CTA_TIMEOUT])
2251		goto err1;
2252
2253	rcu_read_lock();
2254 	if (cda[CTA_HELP]) {
2255		char *helpname = NULL;
2256		struct nlattr *helpinfo = NULL;
2257
2258		err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2259 		if (err < 0)
2260			goto err2;
2261
2262		helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2263						    nf_ct_protonum(ct));
2264		if (helper == NULL) {
2265			rcu_read_unlock();
2266#ifdef CONFIG_MODULES
2267			if (request_module("nfct-helper-%s", helpname) < 0) {
2268				err = -EOPNOTSUPP;
2269				goto err1;
2270			}
2271
2272			rcu_read_lock();
2273			helper = __nf_conntrack_helper_find(helpname,
2274							    nf_ct_l3num(ct),
2275							    nf_ct_protonum(ct));
2276			if (helper) {
2277				err = -EAGAIN;
2278				goto err2;
2279			}
2280			rcu_read_unlock();
2281#endif
2282			err = -EOPNOTSUPP;
2283			goto err1;
2284		} else {
2285			struct nf_conn_help *help;
2286
2287			help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2288			if (help == NULL) {
2289				err = -ENOMEM;
2290				goto err2;
2291			}
2292			/* set private helper data if allowed. */
2293			if (helper->from_nlattr)
2294				helper->from_nlattr(helpinfo, ct);
2295
2296			/* disable helper auto-assignment for this entry */
2297			ct->status |= IPS_HELPER;
2298			RCU_INIT_POINTER(help->helper, helper);
2299		}
2300	}
2301
2302	err = ctnetlink_setup_nat(ct, cda);
2303	if (err < 0)
2304		goto err2;
2305
2306	nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2307	nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2308	nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2309	nf_ct_labels_ext_add(ct);
2310	nfct_seqadj_ext_add(ct);
2311	nfct_synproxy_ext_add(ct);
2312
2313	/* we must add conntrack extensions before confirmation. */
2314	ct->status |= IPS_CONFIRMED;
2315
2316	timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2317	__nf_ct_set_timeout(ct, timeout);
2318
2319	if (cda[CTA_STATUS]) {
2320		err = ctnetlink_change_status(ct, cda);
2321		if (err < 0)
2322			goto err2;
2323	}
2324
2325	if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2326		err = ctnetlink_change_seq_adj(ct, cda);
2327		if (err < 0)
2328			goto err2;
2329	}
2330
2331	memset(&ct->proto, 0, sizeof(ct->proto));
2332	if (cda[CTA_PROTOINFO]) {
2333		err = ctnetlink_change_protoinfo(ct, cda);
2334		if (err < 0)
2335			goto err2;
2336	}
2337
2338	if (cda[CTA_SYNPROXY]) {
2339		err = ctnetlink_change_synproxy(ct, cda);
2340		if (err < 0)
2341			goto err2;
2342	}
2343
2344#if defined(CONFIG_NF_CONNTRACK_MARK)
2345	if (cda[CTA_MARK])
2346		ctnetlink_change_mark(ct, cda);
2347#endif
2348
2349	/* setup master conntrack: this is a confirmed expectation */
2350	if (cda[CTA_TUPLE_MASTER]) {
2351		struct nf_conntrack_tuple master;
2352		struct nf_conntrack_tuple_hash *master_h;
2353		struct nf_conn *master_ct;
2354
2355		err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2356					    u3, NULL);
2357		if (err < 0)
2358			goto err2;
2359
2360		master_h = nf_conntrack_find_get(net, zone, &master);
2361		if (master_h == NULL) {
2362			err = -ENOENT;
2363			goto err2;
2364		}
2365		master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2366		__set_bit(IPS_EXPECTED_BIT, &ct->status);
2367		ct->master = master_ct;
2368	}
2369	tstamp = nf_conn_tstamp_find(ct);
2370	if (tstamp)
2371		tstamp->start = ktime_get_real_ns();
2372
2373	err = nf_conntrack_hash_check_insert(ct);
2374	if (err < 0)
2375		goto err3;
2376
2377	rcu_read_unlock();
2378
2379	return ct;
2380
2381err3:
2382	if (ct->master)
2383		nf_ct_put(ct->master);
2384err2:
2385	rcu_read_unlock();
2386err1:
2387	nf_conntrack_free(ct);
2388	return ERR_PTR(err);
2389}
2390
2391static int ctnetlink_new_conntrack(struct sk_buff *skb,
2392				   const struct nfnl_info *info,
2393				   const struct nlattr * const cda[])
2394{
2395	struct nf_conntrack_tuple otuple, rtuple;
2396	struct nf_conntrack_tuple_hash *h = NULL;
2397	u_int8_t u3 = info->nfmsg->nfgen_family;
2398	struct nf_conntrack_zone zone;
2399	struct nf_conn *ct;
2400	int err;
2401
2402	err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2403	if (err < 0)
2404		return err;
2405
2406	if (cda[CTA_TUPLE_ORIG]) {
2407		err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2408					    u3, &zone);
2409		if (err < 0)
2410			return err;
2411	}
2412
2413	if (cda[CTA_TUPLE_REPLY]) {
2414		err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2415					    u3, &zone);
2416		if (err < 0)
2417			return err;
2418	}
2419
2420	if (cda[CTA_TUPLE_ORIG])
2421		h = nf_conntrack_find_get(info->net, &zone, &otuple);
2422	else if (cda[CTA_TUPLE_REPLY])
2423		h = nf_conntrack_find_get(info->net, &zone, &rtuple);
2424
2425	if (h == NULL) {
2426		err = -ENOENT;
2427		if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
2428			enum ip_conntrack_events events;
2429
2430			if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2431				return -EINVAL;
2432			if (otuple.dst.protonum != rtuple.dst.protonum)
2433				return -EINVAL;
2434
2435			ct = ctnetlink_create_conntrack(info->net, &zone, cda,
2436							&otuple, &rtuple, u3);
2437			if (IS_ERR(ct))
2438				return PTR_ERR(ct);
2439
2440			err = 0;
2441			if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2442				events = 1 << IPCT_RELATED;
2443			else
2444				events = 1 << IPCT_NEW;
2445
2446			if (cda[CTA_LABELS] &&
2447			    ctnetlink_attach_labels(ct, cda) == 0)
2448				events |= (1 << IPCT_LABEL);
2449
2450			nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2451						      (1 << IPCT_ASSURED) |
2452						      (1 << IPCT_HELPER) |
2453						      (1 << IPCT_PROTOINFO) |
2454						      (1 << IPCT_SEQADJ) |
2455						      (1 << IPCT_MARK) |
2456						      (1 << IPCT_SYNPROXY) |
2457						      events,
2458						      ct, NETLINK_CB(skb).portid,
2459						      nlmsg_report(info->nlh));
2460			nf_ct_put(ct);
2461		}
2462
2463		return err;
2464	}
2465	/* implicit 'else' */
2466
2467	err = -EEXIST;
2468	ct = nf_ct_tuplehash_to_ctrack(h);
2469	if (!(info->nlh->nlmsg_flags & NLM_F_EXCL)) {
2470		err = ctnetlink_change_conntrack(ct, cda);
2471		if (err == 0) {
2472			nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2473						      (1 << IPCT_ASSURED) |
2474						      (1 << IPCT_HELPER) |
2475						      (1 << IPCT_LABEL) |
2476						      (1 << IPCT_PROTOINFO) |
2477						      (1 << IPCT_SEQADJ) |
2478						      (1 << IPCT_MARK) |
2479						      (1 << IPCT_SYNPROXY),
2480						      ct, NETLINK_CB(skb).portid,
2481						      nlmsg_report(info->nlh));
2482		}
2483	}
2484
2485	nf_ct_put(ct);
2486	return err;
2487}
2488
2489static int
2490ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2491				__u16 cpu, const struct ip_conntrack_stat *st)
2492{
2493	struct nlmsghdr *nlh;
2494	unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2495
2496	event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2497			      IPCTNL_MSG_CT_GET_STATS_CPU);
2498	nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2499			   NFNETLINK_V0, htons(cpu));
2500	if (!nlh)
2501		goto nlmsg_failure;
2502
2503	if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2504	    nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2505	    nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2506	    nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2507				htonl(st->insert_failed)) ||
2508	    nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2509	    nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2510	    nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2511	    nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2512				htonl(st->search_restart)) ||
2513	    nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2514				htonl(st->clash_resolve)) ||
2515	    nla_put_be32(skb, CTA_STATS_CHAIN_TOOLONG,
2516			 htonl(st->chaintoolong)))
2517		goto nla_put_failure;
2518
2519	nlmsg_end(skb, nlh);
2520	return skb->len;
2521
2522nla_put_failure:
2523nlmsg_failure:
2524	nlmsg_cancel(skb, nlh);
2525	return -1;
2526}
2527
2528static int
2529ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2530{
2531	int cpu;
2532	struct net *net = sock_net(skb->sk);
2533
2534	if (cb->args[0] == nr_cpu_ids)
2535		return 0;
2536
2537	for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2538		const struct ip_conntrack_stat *st;
2539
2540		if (!cpu_possible(cpu))
2541			continue;
2542
2543		st = per_cpu_ptr(net->ct.stat, cpu);
2544		if (ctnetlink_ct_stat_cpu_fill_info(skb,
2545						    NETLINK_CB(cb->skb).portid,
2546						    cb->nlh->nlmsg_seq,
2547						    cpu, st) < 0)
2548				break;
2549	}
2550	cb->args[0] = cpu;
2551
2552	return skb->len;
2553}
2554
2555static int ctnetlink_stat_ct_cpu(struct sk_buff *skb,
2556				 const struct nfnl_info *info,
2557				 const struct nlattr * const cda[])
2558{
2559	if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2560		struct netlink_dump_control c = {
2561			.dump = ctnetlink_ct_stat_cpu_dump,
2562		};
2563		return netlink_dump_start(info->sk, skb, info->nlh, &c);
2564	}
2565
2566	return 0;
2567}
2568
2569static int
2570ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2571			    struct net *net)
2572{
2573	unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2574	unsigned int nr_conntracks;
2575	struct nlmsghdr *nlh;
2576
2577	event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2578	nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2579			   NFNETLINK_V0, 0);
2580	if (!nlh)
2581		goto nlmsg_failure;
2582
2583	nr_conntracks = nf_conntrack_count(net);
2584	if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2585		goto nla_put_failure;
2586
2587	if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2588		goto nla_put_failure;
2589
2590	nlmsg_end(skb, nlh);
2591	return skb->len;
2592
2593nla_put_failure:
2594nlmsg_failure:
2595	nlmsg_cancel(skb, nlh);
2596	return -1;
2597}
2598
2599static int ctnetlink_stat_ct(struct sk_buff *skb, const struct nfnl_info *info,
2600			     const struct nlattr * const cda[])
2601{
2602	struct sk_buff *skb2;
2603	int err;
2604
2605	skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2606	if (skb2 == NULL)
2607		return -ENOMEM;
2608
2609	err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2610					  info->nlh->nlmsg_seq,
2611					  NFNL_MSG_TYPE(info->nlh->nlmsg_type),
2612					  sock_net(skb->sk));
2613	if (err <= 0) {
2614		kfree_skb(skb2);
2615		return -ENOMEM;
2616	}
2617
2618	return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
2619}
2620
2621static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2622	[CTA_EXPECT_MASTER]	= { .type = NLA_NESTED },
2623	[CTA_EXPECT_TUPLE]	= { .type = NLA_NESTED },
2624	[CTA_EXPECT_MASK]	= { .type = NLA_NESTED },
2625	[CTA_EXPECT_TIMEOUT]	= { .type = NLA_U32 },
2626	[CTA_EXPECT_ID]		= { .type = NLA_U32 },
2627	[CTA_EXPECT_HELP_NAME]	= { .type = NLA_NUL_STRING,
2628				    .len = NF_CT_HELPER_NAME_LEN - 1 },
2629	[CTA_EXPECT_ZONE]	= { .type = NLA_U16 },
2630	[CTA_EXPECT_FLAGS]	= NLA_POLICY_MASK(NLA_BE32, NF_CT_EXPECT_MASK),
2631	[CTA_EXPECT_CLASS]	= { .type = NLA_U32 },
2632	[CTA_EXPECT_NAT]	= { .type = NLA_NESTED },
2633	[CTA_EXPECT_FN]		= { .type = NLA_NUL_STRING },
2634};
2635
2636static struct nf_conntrack_expect *
2637ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2638		       struct nf_conntrack_tuple *tuple,
2639		       struct nf_conntrack_tuple *mask);
2640
2641#ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2642static size_t
2643ctnetlink_glue_build_size(const struct nf_conn *ct)
2644{
2645	return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2646	       + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2647	       + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2648	       + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2649	       + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2650	       + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2651	       + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2652	       + nla_total_size(0) /* CTA_PROTOINFO */
2653	       + nla_total_size(0) /* CTA_HELP */
2654	       + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2655	       + ctnetlink_secctx_size(ct)
2656	       + ctnetlink_acct_size(ct)
2657	       + ctnetlink_timestamp_size(ct)
2658#if IS_ENABLED(CONFIG_NF_NAT)
2659	       + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2660	       + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2661#endif
2662#ifdef CONFIG_NF_CONNTRACK_MARK
2663	       + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2664#endif
2665#ifdef CONFIG_NF_CONNTRACK_ZONES
2666	       + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2667#endif
2668	       + ctnetlink_proto_size(ct)
2669	       ;
2670}
2671
2672static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2673{
2674	const struct nf_conntrack_zone *zone;
2675	struct nlattr *nest_parms;
2676
2677	zone = nf_ct_zone(ct);
2678
2679	nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2680	if (!nest_parms)
2681		goto nla_put_failure;
2682	if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2683		goto nla_put_failure;
2684	if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2685				   NF_CT_ZONE_DIR_ORIG) < 0)
2686		goto nla_put_failure;
2687	nla_nest_end(skb, nest_parms);
2688
2689	nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2690	if (!nest_parms)
2691		goto nla_put_failure;
2692	if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2693		goto nla_put_failure;
2694	if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2695				   NF_CT_ZONE_DIR_REPL) < 0)
2696		goto nla_put_failure;
2697	nla_nest_end(skb, nest_parms);
2698
2699	if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2700				   NF_CT_DEFAULT_ZONE_DIR) < 0)
2701		goto nla_put_failure;
2702
2703	if (ctnetlink_dump_id(skb, ct) < 0)
2704		goto nla_put_failure;
2705
2706	if (ctnetlink_dump_status(skb, ct) < 0)
2707		goto nla_put_failure;
2708
2709	if (ctnetlink_dump_timeout(skb, ct, false) < 0)
2710		goto nla_put_failure;
2711
2712	if (ctnetlink_dump_protoinfo(skb, ct, false) < 0)
2713		goto nla_put_failure;
2714
2715	if (ctnetlink_dump_acct(skb, ct, IPCTNL_MSG_CT_GET) < 0 ||
2716	    ctnetlink_dump_timestamp(skb, ct) < 0)
2717		goto nla_put_failure;
2718
2719	if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2720		goto nla_put_failure;
2721
2722#ifdef CONFIG_NF_CONNTRACK_SECMARK
2723	if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2724		goto nla_put_failure;
2725#endif
2726	if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2727		goto nla_put_failure;
2728
2729	if ((ct->status & IPS_SEQ_ADJUST) &&
2730	    ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2731		goto nla_put_failure;
2732
2733	if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2734		goto nla_put_failure;
2735
2736#ifdef CONFIG_NF_CONNTRACK_MARK
2737	if (ctnetlink_dump_mark(skb, ct, true) < 0)
2738		goto nla_put_failure;
2739#endif
2740	if (ctnetlink_dump_labels(skb, ct) < 0)
2741		goto nla_put_failure;
2742	return 0;
2743
2744nla_put_failure:
2745	return -ENOSPC;
2746}
2747
2748static int
2749ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2750		     enum ip_conntrack_info ctinfo,
2751		     u_int16_t ct_attr, u_int16_t ct_info_attr)
2752{
2753	struct nlattr *nest_parms;
2754
2755	nest_parms = nla_nest_start(skb, ct_attr);
2756	if (!nest_parms)
2757		goto nla_put_failure;
2758
2759	if (__ctnetlink_glue_build(skb, ct) < 0)
2760		goto nla_put_failure;
2761
2762	nla_nest_end(skb, nest_parms);
2763
2764	if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2765		goto nla_put_failure;
2766
2767	return 0;
2768
2769nla_put_failure:
2770	return -ENOSPC;
2771}
2772
2773static int
2774ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2775{
2776	unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2777	unsigned long d = ct->status ^ status;
2778
2779	if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2780		/* SEEN_REPLY bit can only be set */
2781		return -EBUSY;
2782
2783	if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2784		/* ASSURED bit can only be set */
2785		return -EBUSY;
2786
2787	/* This check is less strict than ctnetlink_change_status()
2788	 * because callers often flip IPS_EXPECTED bits when sending
2789	 * an NFQA_CT attribute to the kernel.  So ignore the
2790	 * unchangeable bits but do not error out. Also user programs
2791	 * are allowed to clear the bits that they are allowed to change.
2792	 */
2793	__nf_ct_change_status(ct, status, ~status);
2794	return 0;
2795}
2796
2797static int
2798ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2799{
2800	int err;
2801
2802	if (cda[CTA_TIMEOUT]) {
2803		err = ctnetlink_change_timeout(ct, cda);
2804		if (err < 0)
2805			return err;
2806	}
2807	if (cda[CTA_STATUS]) {
2808		err = ctnetlink_update_status(ct, cda);
2809		if (err < 0)
2810			return err;
2811	}
2812	if (cda[CTA_HELP]) {
2813		err = ctnetlink_change_helper(ct, cda);
2814		if (err < 0)
2815			return err;
2816	}
2817	if (cda[CTA_LABELS]) {
2818		err = ctnetlink_attach_labels(ct, cda);
2819		if (err < 0)
2820			return err;
2821	}
2822#if defined(CONFIG_NF_CONNTRACK_MARK)
2823	if (cda[CTA_MARK]) {
2824		ctnetlink_change_mark(ct, cda);
2825	}
2826#endif
2827	return 0;
2828}
2829
2830static int
2831ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2832{
2833	struct nlattr *cda[CTA_MAX+1];
2834	int ret;
2835
2836	ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2837					  NULL);
2838	if (ret < 0)
2839		return ret;
2840
2841	return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2842}
2843
2844static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2845				    const struct nf_conn *ct,
2846				    struct nf_conntrack_tuple *tuple,
2847				    struct nf_conntrack_tuple *mask)
2848{
2849	int err;
2850
2851	err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2852				    nf_ct_l3num(ct), NULL);
2853	if (err < 0)
2854		return err;
2855
2856	return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2857				     nf_ct_l3num(ct), NULL);
2858}
2859
2860static int
2861ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2862			     u32 portid, u32 report)
2863{
2864	struct nlattr *cda[CTA_EXPECT_MAX+1];
2865	struct nf_conntrack_tuple tuple, mask;
2866	struct nf_conntrack_expect *exp;
2867	int err;
2868
2869	err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2870					  exp_nla_policy, NULL);
2871	if (err < 0)
2872		return err;
2873
2874	err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2875				       ct, &tuple, &mask);
2876	if (err < 0)
2877		return err;
2878
2879	exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2880				     &tuple, &mask);
2881	if (IS_ERR(exp))
2882		return PTR_ERR(exp);
2883
2884	err = nf_ct_expect_related_report(exp, portid, report, 0);
2885	nf_ct_expect_put(exp);
2886	return err;
2887}
2888
2889static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2890				  enum ip_conntrack_info ctinfo, int diff)
2891{
2892	if (!(ct->status & IPS_NAT_MASK))
2893		return;
2894
2895	nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2896}
2897
2898static const struct nfnl_ct_hook ctnetlink_glue_hook = {
2899	.build_size	= ctnetlink_glue_build_size,
2900	.build		= ctnetlink_glue_build,
2901	.parse		= ctnetlink_glue_parse,
2902	.attach_expect	= ctnetlink_glue_attach_expect,
2903	.seq_adjust	= ctnetlink_glue_seqadj,
2904};
2905#endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2906
2907/***********************************************************************
2908 * EXPECT
2909 ***********************************************************************/
2910
2911static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2912				    const struct nf_conntrack_tuple *tuple,
2913				    u32 type)
2914{
2915	struct nlattr *nest_parms;
2916
2917	nest_parms = nla_nest_start(skb, type);
2918	if (!nest_parms)
2919		goto nla_put_failure;
2920	if (ctnetlink_dump_tuples(skb, tuple) < 0)
2921		goto nla_put_failure;
2922	nla_nest_end(skb, nest_parms);
2923
2924	return 0;
2925
2926nla_put_failure:
2927	return -1;
2928}
2929
2930static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2931				   const struct nf_conntrack_tuple *tuple,
2932				   const struct nf_conntrack_tuple_mask *mask)
2933{
2934	const struct nf_conntrack_l4proto *l4proto;
2935	struct nf_conntrack_tuple m;
2936	struct nlattr *nest_parms;
2937	int ret;
2938
2939	memset(&m, 0xFF, sizeof(m));
2940	memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2941	m.src.u.all = mask->src.u.all;
2942	m.src.l3num = tuple->src.l3num;
2943	m.dst.protonum = tuple->dst.protonum;
2944
2945	nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2946	if (!nest_parms)
2947		goto nla_put_failure;
2948
2949	rcu_read_lock();
2950	ret = ctnetlink_dump_tuples_ip(skb, &m);
2951	if (ret >= 0) {
2952		l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2953		ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2954	}
2955	rcu_read_unlock();
2956
2957	if (unlikely(ret < 0))
2958		goto nla_put_failure;
2959
2960	nla_nest_end(skb, nest_parms);
2961
2962	return 0;
2963
2964nla_put_failure:
2965	return -1;
2966}
2967
2968#if IS_ENABLED(CONFIG_NF_NAT)
2969static const union nf_inet_addr any_addr;
2970#endif
2971
2972static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2973{
2974	static siphash_aligned_key_t exp_id_seed;
2975	unsigned long a, b, c, d;
2976
2977	net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2978
2979	a = (unsigned long)exp;
2980	b = (unsigned long)exp->helper;
2981	c = (unsigned long)exp->master;
2982	d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2983
2984#ifdef CONFIG_64BIT
2985	return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2986#else
2987	return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2988#endif
2989}
2990
2991static int
2992ctnetlink_exp_dump_expect(struct sk_buff *skb,
2993			  const struct nf_conntrack_expect *exp)
2994{
2995	struct nf_conn *master = exp->master;
2996	long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2997	struct nf_conntrack_helper *helper;
2998#if IS_ENABLED(CONFIG_NF_NAT)
2999	struct nlattr *nest_parms;
3000	struct nf_conntrack_tuple nat_tuple = {};
3001#endif
3002	struct nf_ct_helper_expectfn *expfn;
3003
3004	if (timeout < 0)
3005		timeout = 0;
3006
3007	if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3008		goto nla_put_failure;
3009	if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3010		goto nla_put_failure;
3011	if (ctnetlink_exp_dump_tuple(skb,
3012				 &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3013				 CTA_EXPECT_MASTER) < 0)
3014		goto nla_put_failure;
3015
3016#if IS_ENABLED(CONFIG_NF_NAT)
3017	if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3018	    exp->saved_proto.all) {
3019		nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3020		if (!nest_parms)
3021			goto nla_put_failure;
3022
3023		if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3024			goto nla_put_failure;
3025
3026		nat_tuple.src.l3num = nf_ct_l3num(master);
3027		nat_tuple.src.u3 = exp->saved_addr;
3028		nat_tuple.dst.protonum = nf_ct_protonum(master);
3029		nat_tuple.src.u = exp->saved_proto;
3030
3031		if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3032						CTA_EXPECT_NAT_TUPLE) < 0)
3033	                goto nla_put_failure;
3034	        nla_nest_end(skb, nest_parms);
3035	}
3036#endif
3037	if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3038	    nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3039	    nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3040	    nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3041		goto nla_put_failure;
3042
3043	helper = rcu_dereference(exp->helper);
3044	if (helper &&
3045	    nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3046		goto nla_put_failure;
3047
3048	expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3049	if (expfn != NULL &&
3050	    nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3051		goto nla_put_failure;
3052
3053	return 0;
3054
3055nla_put_failure:
3056	return -1;
3057}
3058
3059static int
3060ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3061			int event, const struct nf_conntrack_expect *exp)
3062{
3063	struct nlmsghdr *nlh;
3064	unsigned int flags = portid ? NLM_F_MULTI : 0;
3065
3066	event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3067	nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3068			   exp->tuple.src.l3num, NFNETLINK_V0, 0);
3069	if (!nlh)
3070		goto nlmsg_failure;
3071
3072	if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3073		goto nla_put_failure;
3074
3075	nlmsg_end(skb, nlh);
3076	return skb->len;
3077
3078nlmsg_failure:
3079nla_put_failure:
3080	nlmsg_cancel(skb, nlh);
3081	return -1;
3082}
3083
3084#ifdef CONFIG_NF_CONNTRACK_EVENTS
3085static int
3086ctnetlink_expect_event(unsigned int events, const struct nf_exp_event *item)
3087{
3088	struct nf_conntrack_expect *exp = item->exp;
3089	struct net *net = nf_ct_exp_net(exp);
3090	struct nlmsghdr *nlh;
3091	struct sk_buff *skb;
3092	unsigned int type, group;
3093	int flags = 0;
3094
3095	if (events & (1 << IPEXP_DESTROY)) {
3096		type = IPCTNL_MSG_EXP_DELETE;
3097		group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3098	} else if (events & (1 << IPEXP_NEW)) {
3099		type = IPCTNL_MSG_EXP_NEW;
3100		flags = NLM_F_CREATE|NLM_F_EXCL;
3101		group = NFNLGRP_CONNTRACK_EXP_NEW;
3102	} else
3103		return 0;
3104
3105	if (!item->report && !nfnetlink_has_listeners(net, group))
3106		return 0;
3107
3108	skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3109	if (skb == NULL)
3110		goto errout;
3111
3112	type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3113	nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3114			   exp->tuple.src.l3num, NFNETLINK_V0, 0);
3115	if (!nlh)
3116		goto nlmsg_failure;
3117
3118	if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3119		goto nla_put_failure;
3120
3121	nlmsg_end(skb, nlh);
3122	nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3123	return 0;
3124
3125nla_put_failure:
3126	nlmsg_cancel(skb, nlh);
3127nlmsg_failure:
3128	kfree_skb(skb);
3129errout:
3130	nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3131	return 0;
3132}
3133#endif
3134
3135static unsigned long ctnetlink_exp_id(const struct nf_conntrack_expect *exp)
3136{
3137	unsigned long id = (unsigned long)exp;
3138
3139	id += nf_ct_get_id(exp->master);
3140	id += exp->class;
3141
3142	return id ? id : 1;
3143}
3144
3145static int
3146ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3147{
3148	struct net *net = sock_net(skb->sk);
3149	struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3150	u_int8_t l3proto = nfmsg->nfgen_family;
3151	unsigned long last_id = cb->args[1];
3152	struct nf_conntrack_expect *exp;
3153
3154	rcu_read_lock();
3155	for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3156restart:
3157		hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3158					 hnode) {
3159			if (l3proto && exp->tuple.src.l3num != l3proto)
3160				continue;
3161
3162			if (!net_eq(nf_ct_net(exp->master), net))
3163				continue;
3164
3165			if (cb->args[1]) {
3166				if (ctnetlink_exp_id(exp) != last_id)
3167					continue;
3168				cb->args[1] = 0;
3169			}
3170			if (ctnetlink_exp_fill_info(skb,
3171						    NETLINK_CB(cb->skb).portid,
3172						    cb->nlh->nlmsg_seq,
3173						    IPCTNL_MSG_EXP_NEW,
3174						    exp) < 0) {
3175				cb->args[1] = ctnetlink_exp_id(exp);
3176				goto out;
3177			}
3178		}
3179		if (cb->args[1]) {
3180			cb->args[1] = 0;
3181			goto restart;
3182		}
3183	}
3184out:
3185	rcu_read_unlock();
3186	return skb->len;
3187}
3188
3189static int
3190ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3191{
3192	struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3193	struct nf_conn *ct = cb->data;
3194	struct nf_conn_help *help;
3195	u_int8_t l3proto = nfmsg->nfgen_family;
3196	unsigned long last_id = cb->args[1];
3197	struct nf_conntrack_expect *exp;
3198
3199	if (cb->args[0])
3200		return 0;
3201
3202	help = nfct_help(ct);
3203	if (!help)
3204		return 0;
3205
3206	rcu_read_lock();
3207
3208restart:
3209	hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3210		if (l3proto && exp->tuple.src.l3num != l3proto)
3211			continue;
3212		if (cb->args[1]) {
3213			if (ctnetlink_exp_id(exp) != last_id)
3214				continue;
3215			cb->args[1] = 0;
3216		}
3217		if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3218					    cb->nlh->nlmsg_seq,
3219					    IPCTNL_MSG_EXP_NEW,
3220					    exp) < 0) {
3221			cb->args[1] = ctnetlink_exp_id(exp);
3222			goto out;
3223		}
3224	}
3225	if (cb->args[1]) {
3226		cb->args[1] = 0;
3227		goto restart;
3228	}
3229	cb->args[0] = 1;
3230out:
3231	rcu_read_unlock();
3232	return skb->len;
3233}
3234
3235static int ctnetlink_dump_exp_ct_start(struct netlink_callback *cb)
3236{
3237	struct nf_conn *ct = cb->data;
3238
3239	if (!refcount_inc_not_zero(&ct->ct_general.use))
3240		return -ENOENT;
3241	return 0;
3242}
3243
3244static int ctnetlink_dump_exp_ct_done(struct netlink_callback *cb)
3245{
3246	struct nf_conn *ct = cb->data;
3247
3248	if (ct)
3249		nf_ct_put(ct);
3250	return 0;
3251}
3252
3253static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3254				 struct sk_buff *skb,
3255				 const struct nlmsghdr *nlh,
3256				 const struct nlattr * const cda[],
3257				 struct netlink_ext_ack *extack)
3258{
3259	int err;
3260	struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3261	u_int8_t u3 = nfmsg->nfgen_family;
3262	struct nf_conntrack_tuple tuple;
3263	struct nf_conntrack_tuple_hash *h;
3264	struct nf_conn *ct;
3265	struct nf_conntrack_zone zone;
3266	struct netlink_dump_control c = {
3267		.dump = ctnetlink_exp_ct_dump_table,
3268		.start = ctnetlink_dump_exp_ct_start,
3269		.done = ctnetlink_dump_exp_ct_done,
3270	};
3271
3272	err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3273				    u3, NULL);
3274	if (err < 0)
3275		return err;
3276
3277	err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3278	if (err < 0)
3279		return err;
3280
3281	h = nf_conntrack_find_get(net, &zone, &tuple);
3282	if (!h)
3283		return -ENOENT;
3284
3285	ct = nf_ct_tuplehash_to_ctrack(h);
3286	/* No expectation linked to this connection tracking. */
3287	if (!nfct_help(ct)) {
3288		nf_ct_put(ct);
3289		return 0;
3290	}
3291
3292	c.data = ct;
3293
3294	err = netlink_dump_start(ctnl, skb, nlh, &c);
3295	nf_ct_put(ct);
3296
3297	return err;
3298}
3299
3300static int ctnetlink_get_expect(struct sk_buff *skb,
3301				const struct nfnl_info *info,
3302				const struct nlattr * const cda[])
3303{
3304	u_int8_t u3 = info->nfmsg->nfgen_family;
3305	struct nf_conntrack_tuple tuple;
3306	struct nf_conntrack_expect *exp;
3307	struct nf_conntrack_zone zone;
3308	struct sk_buff *skb2;
3309	int err;
3310
3311	if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3312		if (cda[CTA_EXPECT_MASTER])
3313			return ctnetlink_dump_exp_ct(info->net, info->sk, skb,
3314						     info->nlh, cda,
3315						     info->extack);
3316		else {
3317			struct netlink_dump_control c = {
3318				.dump = ctnetlink_exp_dump_table,
3319			};
3320			return netlink_dump_start(info->sk, skb, info->nlh, &c);
3321		}
3322	}
3323
3324	err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3325	if (err < 0)
3326		return err;
3327
3328	if (cda[CTA_EXPECT_TUPLE])
3329		err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3330					    u3, NULL);
3331	else if (cda[CTA_EXPECT_MASTER])
3332		err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3333					    u3, NULL);
3334	else
3335		return -EINVAL;
3336
3337	if (err < 0)
3338		return err;
3339
3340	skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3341	if (!skb2)
3342		return -ENOMEM;
3343
3344	spin_lock_bh(&nf_conntrack_expect_lock);
3345	exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3346	if (!exp) {
3347		spin_unlock_bh(&nf_conntrack_expect_lock);
3348		kfree_skb(skb2);
3349		return -ENOENT;
3350	}
3351
3352	if (cda[CTA_EXPECT_ID]) {
3353		__be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3354
3355		if (id != nf_expect_get_id(exp)) {
3356			nf_ct_expect_put(exp);
3357			spin_unlock_bh(&nf_conntrack_expect_lock);
3358			kfree_skb(skb2);
3359			return -ENOENT;
3360		}
3361	}
3362
3363	rcu_read_lock();
3364	err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3365				      info->nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW,
3366				      exp);
3367	rcu_read_unlock();
3368	nf_ct_expect_put(exp);
3369	spin_unlock_bh(&nf_conntrack_expect_lock);
3370
3371	if (err <= 0) {
3372		kfree_skb(skb2);
3373		return -ENOMEM;
3374	}
3375
3376	return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
3377}
3378
3379static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3380{
3381	struct nf_conntrack_helper *helper;
3382	const char *name = data;
3383
3384	helper = rcu_dereference(exp->helper);
3385	if (!helper)
3386		return false;
3387
3388	return strcmp(helper->name, name) == 0;
3389}
3390
3391static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3392{
3393	return true;
3394}
3395
3396static int ctnetlink_del_expect(struct sk_buff *skb,
3397				const struct nfnl_info *info,
3398				const struct nlattr * const cda[])
3399{
3400	u_int8_t u3 = info->nfmsg->nfgen_family;
3401	struct nf_conntrack_expect *exp;
3402	struct nf_conntrack_tuple tuple;
3403	struct nf_conntrack_zone zone;
3404	int err;
3405
3406	if (cda[CTA_EXPECT_TUPLE]) {
3407		/* delete a single expect by tuple */
3408		err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3409		if (err < 0)
3410			return err;
3411
3412		err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3413					    u3, NULL);
3414		if (err < 0)
3415			return err;
3416
3417		spin_lock_bh(&nf_conntrack_expect_lock);
3418
3419		/* bump usage count to 2 */
3420		exp = nf_ct_expect_find_get(info->net, &zone, &tuple);
3421		if (!exp) {
3422			spin_unlock_bh(&nf_conntrack_expect_lock);
3423			return -ENOENT;
3424		}
3425
3426		if (cda[CTA_EXPECT_ID]) {
3427			__be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3428
3429			if (id != nf_expect_get_id(exp)) {
3430				nf_ct_expect_put(exp);
3431				spin_unlock_bh(&nf_conntrack_expect_lock);
3432				return -ENOENT;
3433			}
3434		}
3435
3436		/* after list removal, usage count == 1 */
3437		if (timer_delete(&exp->timeout)) {
3438			nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3439						   nlmsg_report(info->nlh));
3440			nf_ct_expect_put(exp);
3441		}
3442		spin_unlock_bh(&nf_conntrack_expect_lock);
3443		/* have to put what we 'get' above.
3444		 * after this line usage count == 0 */
3445		nf_ct_expect_put(exp);
3446	} else if (cda[CTA_EXPECT_HELP_NAME]) {
3447		char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3448
3449		nf_ct_expect_iterate_net(info->net, expect_iter_name, name,
3450					 NETLINK_CB(skb).portid,
3451					 nlmsg_report(info->nlh));
3452	} else {
3453		/* This basically means we have to flush everything*/
3454		nf_ct_expect_iterate_net(info->net, expect_iter_all, NULL,
3455					 NETLINK_CB(skb).portid,
3456					 nlmsg_report(info->nlh));
3457	}
3458
3459	return 0;
3460}
3461static int
3462ctnetlink_change_expect(struct nf_conntrack_expect *x,
3463			const struct nlattr * const cda[])
3464{
3465	if (cda[CTA_EXPECT_TIMEOUT]) {
3466		if (!timer_delete(&x->timeout))
3467			return -ETIME;
3468
3469		x->timeout.expires = jiffies +
3470			ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3471		add_timer(&x->timeout);
3472	}
3473	return 0;
3474}
3475
3476#if IS_ENABLED(CONFIG_NF_NAT)
3477static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3478	[CTA_EXPECT_NAT_DIR]	= NLA_POLICY_MAX(NLA_BE32, IP_CT_DIR_REPLY),
3479	[CTA_EXPECT_NAT_TUPLE]	= { .type = NLA_NESTED },
3480};
3481#endif
3482
3483static int
3484ctnetlink_parse_expect_nat(const struct nlattr *attr,
3485			   struct nf_conntrack_expect *exp,
3486			   u_int8_t u3)
3487{
3488#if IS_ENABLED(CONFIG_NF_NAT)
3489	struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3490	struct nf_conntrack_tuple nat_tuple = {};
3491	int err;
3492
3493	err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3494					  exp_nat_nla_policy, NULL);
3495	if (err < 0)
3496		return err;
3497
3498	if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3499		return -EINVAL;
3500
3501	err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3502				    &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3503				    u3, NULL);
3504	if (err < 0)
3505		return err;
3506
3507	exp->saved_addr = nat_tuple.src.u3;
3508	exp->saved_proto = nat_tuple.src.u;
3509	exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3510
3511	return 0;
3512#else
3513	return -EOPNOTSUPP;
3514#endif
3515}
3516
3517static struct nf_conntrack_expect *
3518ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3519		       struct nf_conntrack_tuple *tuple,
3520		       struct nf_conntrack_tuple *mask)
3521{
3522	struct net *net = read_pnet(&ct->ct_net);
3523	struct nf_conntrack_helper *helper;
3524	struct nf_conntrack_expect *exp;
3525	struct nf_conn_help *help;
3526	u32 class = 0;
3527	int err;
3528
3529	help = nfct_help(ct);
3530	if (!help)
3531		return ERR_PTR(-EOPNOTSUPP);
3532
3533	helper = rcu_dereference(help->helper);
3534	if (!helper)
3535		return ERR_PTR(-EOPNOTSUPP);
3536
3537	if (cda[CTA_EXPECT_CLASS]) {
3538		class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3539		if (class > helper->expect_class_max)
3540			return ERR_PTR(-EINVAL);
3541	}
3542	exp = nf_ct_expect_alloc(ct);
3543	if (!exp)
3544		return ERR_PTR(-ENOMEM);
3545
3546	if (cda[CTA_EXPECT_FLAGS]) {
3547		exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3548		exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3549	} else {
3550		exp->flags = 0;
3551	}
3552	if (cda[CTA_EXPECT_FN]) {
3553		const char *name = nla_data(cda[CTA_EXPECT_FN]);
3554		struct nf_ct_helper_expectfn *expfn;
3555
3556		expfn = nf_ct_helper_expectfn_find_by_name(name);
3557		if (expfn == NULL) {
3558			err = -EINVAL;
3559			goto err_out;
3560		}
3561		exp->expectfn = expfn->expectfn;
3562	} else
3563		exp->expectfn = NULL;
3564
3565	exp->class = class;
3566	exp->master = ct;
3567	write_pnet(&exp->net, net);
3568#ifdef CONFIG_NF_CONNTRACK_ZONES
3569	exp->zone = ct->zone;
3570#endif
3571	rcu_assign_pointer(exp->helper, helper);
3572	exp->tuple = *tuple;
3573	exp->mask.src.u3 = mask->src.u3;
3574	exp->mask.src.u.all = mask->src.u.all;
3575
3576	if (cda[CTA_EXPECT_NAT]) {
3577		err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3578						 exp, nf_ct_l3num(ct));
3579		if (err < 0)
3580			goto err_out;
3581#if IS_ENABLED(CONFIG_NF_NAT)
3582	} else {
3583		memset(&exp->saved_addr, 0, sizeof(exp->saved_addr));
3584		memset(&exp->saved_proto, 0, sizeof(exp->saved_proto));
3585		exp->dir = 0;
3586#endif
3587	}
3588	return exp;
3589err_out:
3590	nf_ct_expect_put(exp);
3591	return ERR_PTR(err);
3592}
3593
3594static int
3595ctnetlink_create_expect(struct net *net,
3596			const struct nf_conntrack_zone *zone,
3597			const struct nlattr * const cda[],
3598			u_int8_t u3, u32 portid, int report)
3599{
3600	struct nf_conntrack_tuple tuple, mask, master_tuple;
3601	struct nf_conntrack_tuple_hash *h = NULL;
3602	struct nf_conntrack_expect *exp;
3603	struct nf_conn *ct;
3604	int err;
3605
3606	/* caller guarantees that those three CTA_EXPECT_* exist */
3607	err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3608				    u3, NULL);
3609	if (err < 0)
3610		return err;
3611	err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3612				    u3, NULL);
3613	if (err < 0)
3614		return err;
3615	err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3616				    u3, NULL);
3617	if (err < 0)
3618		return err;
3619
3620	/* Look for master conntrack of this expectation */
3621	h = nf_conntrack_find_get(net, zone, &master_tuple);
3622	if (!h)
3623		return -ENOENT;
3624	ct = nf_ct_tuplehash_to_ctrack(h);
3625
3626	rcu_read_lock();
3627	exp = ctnetlink_alloc_expect(cda, ct, &tuple, &mask);
3628	if (IS_ERR(exp)) {
3629		err = PTR_ERR(exp);
3630		goto err_rcu;
3631	}
3632
3633	err = nf_ct_expect_related_report(exp, portid, report, 0);
3634	nf_ct_expect_put(exp);
3635err_rcu:
3636	rcu_read_unlock();
3637	nf_ct_put(ct);
3638
3639	return err;
3640}
3641
3642static int ctnetlink_new_expect(struct sk_buff *skb,
3643				const struct nfnl_info *info,
3644				const struct nlattr * const cda[])
3645{
3646	u_int8_t u3 = info->nfmsg->nfgen_family;
3647	struct nf_conntrack_tuple tuple;
3648	struct nf_conntrack_expect *exp;
3649	struct nf_conntrack_zone zone;
3650	int err;
3651
3652	if (!cda[CTA_EXPECT_TUPLE]
3653	    || !cda[CTA_EXPECT_MASK]
3654	    || !cda[CTA_EXPECT_MASTER])
3655		return -EINVAL;
3656
3657	err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3658	if (err < 0)
3659		return err;
3660
3661	err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3662				    u3, NULL);
3663	if (err < 0)
3664		return err;
3665
3666	spin_lock_bh(&nf_conntrack_expect_lock);
3667	exp = __nf_ct_expect_find(info->net, &zone, &tuple);
3668	if (!exp) {
3669		spin_unlock_bh(&nf_conntrack_expect_lock);
3670		err = -ENOENT;
3671		if (info->nlh->nlmsg_flags & NLM_F_CREATE) {
3672			err = ctnetlink_create_expect(info->net, &zone, cda, u3,
3673						      NETLINK_CB(skb).portid,
3674						      nlmsg_report(info->nlh));
3675		}
3676		return err;
3677	}
3678
3679	err = -EEXIST;
3680	if (!(info->nlh->nlmsg_flags & NLM_F_EXCL))
3681		err = ctnetlink_change_expect(exp, cda);
3682	spin_unlock_bh(&nf_conntrack_expect_lock);
3683
3684	return err;
3685}
3686
3687static int
3688ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3689			     const struct ip_conntrack_stat *st)
3690{
3691	struct nlmsghdr *nlh;
3692	unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3693
3694	event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3695			      IPCTNL_MSG_EXP_GET_STATS_CPU);
3696	nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3697			   NFNETLINK_V0, htons(cpu));
3698	if (!nlh)
3699		goto nlmsg_failure;
3700
3701	if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3702	    nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3703	    nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3704		goto nla_put_failure;
3705
3706	nlmsg_end(skb, nlh);
3707	return skb->len;
3708
3709nla_put_failure:
3710nlmsg_failure:
3711	nlmsg_cancel(skb, nlh);
3712	return -1;
3713}
3714
3715static int
3716ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3717{
3718	int cpu;
3719	struct net *net = sock_net(skb->sk);
3720
3721	if (cb->args[0] == nr_cpu_ids)
3722		return 0;
3723
3724	for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3725		const struct ip_conntrack_stat *st;
3726
3727		if (!cpu_possible(cpu))
3728			continue;
3729
3730		st = per_cpu_ptr(net->ct.stat, cpu);
3731		if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3732						 cb->nlh->nlmsg_seq,
3733						 cpu, st) < 0)
3734			break;
3735	}
3736	cb->args[0] = cpu;
3737
3738	return skb->len;
3739}
3740
3741static int ctnetlink_stat_exp_cpu(struct sk_buff *skb,
3742				  const struct nfnl_info *info,
3743				  const struct nlattr * const cda[])
3744{
3745	if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3746		struct netlink_dump_control c = {
3747			.dump = ctnetlink_exp_stat_cpu_dump,
3748		};
3749		return netlink_dump_start(info->sk, skb, info->nlh, &c);
3750	}
3751
3752	return 0;
3753}
3754
3755#ifdef CONFIG_NF_CONNTRACK_EVENTS
3756static struct nf_ct_event_notifier ctnl_notifier = {
3757	.ct_event = ctnetlink_conntrack_event,
3758	.exp_event = ctnetlink_expect_event,
3759};
3760#endif
3761
3762static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3763	[IPCTNL_MSG_CT_NEW]	= {
3764		.call		= ctnetlink_new_conntrack,
3765		.type		= NFNL_CB_MUTEX,
3766		.attr_count	= CTA_MAX,
3767		.policy		= ct_nla_policy
3768	},
3769	[IPCTNL_MSG_CT_GET]	= {
3770		.call		= ctnetlink_get_conntrack,
3771		.type		= NFNL_CB_MUTEX,
3772		.attr_count	= CTA_MAX,
3773		.policy		= ct_nla_policy
3774	},
3775	[IPCTNL_MSG_CT_DELETE]	= {
3776		.call		= ctnetlink_del_conntrack,
3777		.type		= NFNL_CB_MUTEX,
3778		.attr_count	= CTA_MAX,
3779		.policy		= ct_nla_policy
3780	},
3781	[IPCTNL_MSG_CT_GET_CTRZERO] = {
3782		.call		= ctnetlink_get_conntrack,
3783		.type		= NFNL_CB_MUTEX,
3784		.attr_count	= CTA_MAX,
3785		.policy		= ct_nla_policy
3786	},
3787	[IPCTNL_MSG_CT_GET_STATS_CPU] = {
3788		.call		= ctnetlink_stat_ct_cpu,
3789		.type		= NFNL_CB_MUTEX,
3790	},
3791	[IPCTNL_MSG_CT_GET_STATS] = {
3792		.call		= ctnetlink_stat_ct,
3793		.type		= NFNL_CB_MUTEX,
3794	},
3795	[IPCTNL_MSG_CT_GET_DYING] = {
3796		.call		= ctnetlink_get_ct_dying,
3797		.type		= NFNL_CB_MUTEX,
3798	},
3799	[IPCTNL_MSG_CT_GET_UNCONFIRMED]	= {
3800		.call		= ctnetlink_get_ct_unconfirmed,
3801		.type		= NFNL_CB_MUTEX,
3802	},
3803};
3804
3805static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3806	[IPCTNL_MSG_EXP_GET] = {
3807		.call		= ctnetlink_get_expect,
3808		.type		= NFNL_CB_MUTEX,
3809		.attr_count	= CTA_EXPECT_MAX,
3810		.policy		= exp_nla_policy
3811	},
3812	[IPCTNL_MSG_EXP_NEW] = {
3813		.call		= ctnetlink_new_expect,
3814		.type		= NFNL_CB_MUTEX,
3815		.attr_count	= CTA_EXPECT_MAX,
3816		.policy		= exp_nla_policy
3817	},
3818	[IPCTNL_MSG_EXP_DELETE] = {
3819		.call		= ctnetlink_del_expect,
3820		.type		= NFNL_CB_MUTEX,
3821		.attr_count	= CTA_EXPECT_MAX,
3822		.policy		= exp_nla_policy
3823	},
3824	[IPCTNL_MSG_EXP_GET_STATS_CPU] = {
3825		.call		= ctnetlink_stat_exp_cpu,
3826		.type		= NFNL_CB_MUTEX,
3827	},
3828};
3829
3830static const struct nfnetlink_subsystem ctnl_subsys = {
3831	.name				= "conntrack",
3832	.subsys_id			= NFNL_SUBSYS_CTNETLINK,
3833	.cb_count			= IPCTNL_MSG_MAX,
3834	.cb				= ctnl_cb,
3835};
3836
3837static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3838	.name				= "conntrack_expect",
3839	.subsys_id			= NFNL_SUBSYS_CTNETLINK_EXP,
3840	.cb_count			= IPCTNL_MSG_EXP_MAX,
3841	.cb				= ctnl_exp_cb,
3842};
3843
3844MODULE_ALIAS("ip_conntrack_netlink");
3845MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3846MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3847
3848static int __net_init ctnetlink_net_init(struct net *net)
3849{
3850#ifdef CONFIG_NF_CONNTRACK_EVENTS
3851	nf_conntrack_register_notifier(net, &ctnl_notifier);
3852#endif
3853	return 0;
3854}
3855
3856static void ctnetlink_net_pre_exit(struct net *net)
3857{
3858#ifdef CONFIG_NF_CONNTRACK_EVENTS
3859	nf_conntrack_unregister_notifier(net);
3860#endif
3861}
3862
3863static struct pernet_operations ctnetlink_net_ops = {
3864	.init		= ctnetlink_net_init,
3865	.pre_exit	= ctnetlink_net_pre_exit,
3866};
3867
3868static int __init ctnetlink_init(void)
3869{
3870	int ret;
3871
3872	NL_ASSERT_CTX_FITS(struct ctnetlink_list_dump_ctx);
3873
3874	ret = nfnetlink_subsys_register(&ctnl_subsys);
3875	if (ret < 0) {
3876		pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3877		goto err_out;
3878	}
3879
3880	ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3881	if (ret < 0) {
3882		pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3883		goto err_unreg_subsys;
3884	}
3885
3886	ret = register_pernet_subsys(&ctnetlink_net_ops);
3887	if (ret < 0) {
3888		pr_err("ctnetlink_init: cannot register pernet operations\n");
3889		goto err_unreg_exp_subsys;
3890	}
3891#ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3892	/* setup interaction between nf_queue and nf_conntrack_netlink. */
3893	RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3894#endif
3895	return 0;
3896
3897err_unreg_exp_subsys:
3898	nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3899err_unreg_subsys:
3900	nfnetlink_subsys_unregister(&ctnl_subsys);
3901err_out:
3902	return ret;
3903}
3904
3905static void __exit ctnetlink_exit(void)
3906{
3907	unregister_pernet_subsys(&ctnetlink_net_ops);
3908	nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3909	nfnetlink_subsys_unregister(&ctnl_subsys);
3910#ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3911	RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3912#endif
3913	synchronize_rcu();
3914}
3915
3916module_init(ctnetlink_init);
3917module_exit(ctnetlink_exit);
~/ Read-only file preview.