Linux Audio

Check our new training course

Loading...
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 *	IP multicast routing support for mrouted 3.6/3.8
   4 *
   5 *		(c) 1995 Alan Cox, <alan@lxorguk.ukuu.org.uk>
   6 *	  Linux Consultancy and Custom Driver Development
   7 *
 
 
 
 
 
   8 *	Fixes:
   9 *	Michael Chastain	:	Incorrect size of copying.
  10 *	Alan Cox		:	Added the cache manager code
  11 *	Alan Cox		:	Fixed the clone/copy bug and device race.
  12 *	Mike McLagan		:	Routing by source
  13 *	Malcolm Beattie		:	Buffer handling fixes.
  14 *	Alexey Kuznetsov	:	Double buffer free and other fixes.
  15 *	SVR Anand		:	Fixed several multicast bugs and problems.
  16 *	Alexey Kuznetsov	:	Status, optimisations and more.
  17 *	Brad Parker		:	Better behaviour on mrouted upcall
  18 *					overflow.
  19 *      Carlos Picoto           :       PIMv1 Support
  20 *	Pavlin Ivanov Radoslavov:	PIMv2 Registers must checksum only PIM header
  21 *					Relax this requirement to work with older peers.
 
  22 */
  23
  24#include <linux/uaccess.h>
  25#include <linux/types.h>
  26#include <linux/cache.h>
  27#include <linux/capability.h>
  28#include <linux/errno.h>
 
  29#include <linux/mm.h>
  30#include <linux/kernel.h>
  31#include <linux/fcntl.h>
  32#include <linux/stat.h>
  33#include <linux/socket.h>
  34#include <linux/in.h>
  35#include <linux/inet.h>
  36#include <linux/netdevice.h>
  37#include <linux/inetdevice.h>
  38#include <linux/igmp.h>
  39#include <linux/proc_fs.h>
  40#include <linux/seq_file.h>
  41#include <linux/mroute.h>
  42#include <linux/init.h>
  43#include <linux/if_ether.h>
  44#include <linux/slab.h>
  45#include <net/net_namespace.h>
  46#include <net/ip.h>
  47#include <net/protocol.h>
  48#include <linux/skbuff.h>
  49#include <net/route.h>
 
  50#include <net/icmp.h>
  51#include <net/udp.h>
  52#include <net/raw.h>
  53#include <linux/notifier.h>
  54#include <linux/if_arp.h>
  55#include <linux/netfilter_ipv4.h>
  56#include <linux/compat.h>
  57#include <linux/export.h>
  58#include <linux/rhashtable.h>
  59#include <net/ip_tunnels.h>
  60#include <net/checksum.h>
  61#include <net/netlink.h>
  62#include <net/fib_rules.h>
  63#include <linux/netconf.h>
  64#include <net/rtnh.h>
  65
  66#include <linux/nospec.h>
  67
  68struct ipmr_rule {
  69	struct fib_rule		common;
  70};
  71
  72struct ipmr_result {
  73	struct mr_table		*mrt;
  74};
  75
  76/* Big lock, protecting vif table, mrt cache and mroute socket state.
  77 * Note that the changes are semaphored via rtnl_lock.
  78 */
  79
  80static DEFINE_RWLOCK(mrt_lock);
  81
  82/* Multicast router control variables */
  83
  84/* Special spinlock for queue of unresolved entries */
  85static DEFINE_SPINLOCK(mfc_unres_lock);
  86
  87/* We return to original Alan's scheme. Hash table of resolved
  88 * entries is changed only in process context and protected
  89 * with weak lock mrt_lock. Queue of unresolved entries is protected
  90 * with strong spinlock mfc_unres_lock.
  91 *
  92 * In this case data path is free of exclusive locks at all.
  93 */
  94
  95static struct kmem_cache *mrt_cachep __ro_after_init;
  96
  97static struct mr_table *ipmr_new_table(struct net *net, u32 id);
  98static void ipmr_free_table(struct mr_table *mrt);
  99
 100static void ip_mr_forward(struct net *net, struct mr_table *mrt,
 101			  struct net_device *dev, struct sk_buff *skb,
 102			  struct mfc_cache *cache, int local);
 103static int ipmr_cache_report(struct mr_table *mrt,
 104			     struct sk_buff *pkt, vifi_t vifi, int assert);
 
 
 105static void mroute_netlink_event(struct mr_table *mrt, struct mfc_cache *mfc,
 106				 int cmd);
 107static void igmpmsg_netlink_event(struct mr_table *mrt, struct sk_buff *pkt);
 108static void mroute_clean_tables(struct mr_table *mrt, int flags);
 109static void ipmr_expire_process(struct timer_list *t);
 110
 111#ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
 112#define ipmr_for_each_table(mrt, net)					\
 113	list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list,	\
 114				lockdep_rtnl_is_held() ||		\
 115				list_empty(&net->ipv4.mr_tables))
 116
 117static struct mr_table *ipmr_mr_table_iter(struct net *net,
 118					   struct mr_table *mrt)
 119{
 120	struct mr_table *ret;
 121
 122	if (!mrt)
 123		ret = list_entry_rcu(net->ipv4.mr_tables.next,
 124				     struct mr_table, list);
 125	else
 126		ret = list_entry_rcu(mrt->list.next,
 127				     struct mr_table, list);
 128
 129	if (&ret->list == &net->ipv4.mr_tables)
 130		return NULL;
 131	return ret;
 132}
 133
 134static struct mr_table *ipmr_get_table(struct net *net, u32 id)
 135{
 136	struct mr_table *mrt;
 137
 138	ipmr_for_each_table(mrt, net) {
 139		if (mrt->id == id)
 140			return mrt;
 141	}
 142	return NULL;
 143}
 144
 145static int ipmr_fib_lookup(struct net *net, struct flowi4 *flp4,
 146			   struct mr_table **mrt)
 147{
 148	int err;
 149	struct ipmr_result res;
 150	struct fib_lookup_arg arg = {
 151		.result = &res,
 152		.flags = FIB_LOOKUP_NOREF,
 153	};
 154
 155	/* update flow if oif or iif point to device enslaved to l3mdev */
 156	l3mdev_update_flow(net, flowi4_to_flowi(flp4));
 157
 158	err = fib_rules_lookup(net->ipv4.mr_rules_ops,
 159			       flowi4_to_flowi(flp4), 0, &arg);
 160	if (err < 0)
 161		return err;
 162	*mrt = res.mrt;
 163	return 0;
 164}
 165
 166static int ipmr_rule_action(struct fib_rule *rule, struct flowi *flp,
 167			    int flags, struct fib_lookup_arg *arg)
 168{
 169	struct ipmr_result *res = arg->result;
 170	struct mr_table *mrt;
 171
 172	switch (rule->action) {
 173	case FR_ACT_TO_TBL:
 174		break;
 175	case FR_ACT_UNREACHABLE:
 176		return -ENETUNREACH;
 177	case FR_ACT_PROHIBIT:
 178		return -EACCES;
 179	case FR_ACT_BLACKHOLE:
 180	default:
 181		return -EINVAL;
 182	}
 183
 184	arg->table = fib_rule_get_table(rule, arg);
 185
 186	mrt = ipmr_get_table(rule->fr_net, arg->table);
 187	if (!mrt)
 188		return -EAGAIN;
 189	res->mrt = mrt;
 190	return 0;
 191}
 192
 193static int ipmr_rule_match(struct fib_rule *rule, struct flowi *fl, int flags)
 194{
 195	return 1;
 196}
 197
 198static const struct nla_policy ipmr_rule_policy[FRA_MAX + 1] = {
 199	FRA_GENERIC_POLICY,
 200};
 201
 202static int ipmr_rule_configure(struct fib_rule *rule, struct sk_buff *skb,
 203			       struct fib_rule_hdr *frh, struct nlattr **tb,
 204			       struct netlink_ext_ack *extack)
 205{
 206	return 0;
 207}
 208
 209static int ipmr_rule_compare(struct fib_rule *rule, struct fib_rule_hdr *frh,
 210			     struct nlattr **tb)
 211{
 212	return 1;
 213}
 214
 215static int ipmr_rule_fill(struct fib_rule *rule, struct sk_buff *skb,
 216			  struct fib_rule_hdr *frh)
 217{
 218	frh->dst_len = 0;
 219	frh->src_len = 0;
 220	frh->tos     = 0;
 221	return 0;
 222}
 223
 224static const struct fib_rules_ops __net_initconst ipmr_rules_ops_template = {
 225	.family		= RTNL_FAMILY_IPMR,
 226	.rule_size	= sizeof(struct ipmr_rule),
 227	.addr_size	= sizeof(u32),
 228	.action		= ipmr_rule_action,
 229	.match		= ipmr_rule_match,
 230	.configure	= ipmr_rule_configure,
 231	.compare	= ipmr_rule_compare,
 232	.fill		= ipmr_rule_fill,
 233	.nlgroup	= RTNLGRP_IPV4_RULE,
 234	.policy		= ipmr_rule_policy,
 235	.owner		= THIS_MODULE,
 236};
 237
 238static int __net_init ipmr_rules_init(struct net *net)
 239{
 240	struct fib_rules_ops *ops;
 241	struct mr_table *mrt;
 242	int err;
 243
 244	ops = fib_rules_register(&ipmr_rules_ops_template, net);
 245	if (IS_ERR(ops))
 246		return PTR_ERR(ops);
 247
 248	INIT_LIST_HEAD(&net->ipv4.mr_tables);
 249
 250	mrt = ipmr_new_table(net, RT_TABLE_DEFAULT);
 251	if (IS_ERR(mrt)) {
 252		err = PTR_ERR(mrt);
 253		goto err1;
 254	}
 255
 256	err = fib_default_rule_add(ops, 0x7fff, RT_TABLE_DEFAULT, 0);
 257	if (err < 0)
 258		goto err2;
 259
 260	net->ipv4.mr_rules_ops = ops;
 261	return 0;
 262
 263err2:
 264	ipmr_free_table(mrt);
 265err1:
 266	fib_rules_unregister(ops);
 267	return err;
 268}
 269
 270static void __net_exit ipmr_rules_exit(struct net *net)
 271{
 272	struct mr_table *mrt, *next;
 273
 274	rtnl_lock();
 275	list_for_each_entry_safe(mrt, next, &net->ipv4.mr_tables, list) {
 276		list_del(&mrt->list);
 277		ipmr_free_table(mrt);
 278	}
 279	fib_rules_unregister(net->ipv4.mr_rules_ops);
 280	rtnl_unlock();
 281}
 282
 283static int ipmr_rules_dump(struct net *net, struct notifier_block *nb,
 284			   struct netlink_ext_ack *extack)
 285{
 286	return fib_rules_dump(net, nb, RTNL_FAMILY_IPMR, extack);
 287}
 288
 289static unsigned int ipmr_rules_seq_read(struct net *net)
 290{
 291	return fib_rules_seq_read(net, RTNL_FAMILY_IPMR);
 292}
 293
 294bool ipmr_rule_default(const struct fib_rule *rule)
 295{
 296	return fib_rule_matchall(rule) && rule->table == RT_TABLE_DEFAULT;
 297}
 298EXPORT_SYMBOL(ipmr_rule_default);
 299#else
 300#define ipmr_for_each_table(mrt, net) \
 301	for (mrt = net->ipv4.mrt; mrt; mrt = NULL)
 302
 303static struct mr_table *ipmr_mr_table_iter(struct net *net,
 304					   struct mr_table *mrt)
 305{
 306	if (!mrt)
 307		return net->ipv4.mrt;
 308	return NULL;
 309}
 310
 311static struct mr_table *ipmr_get_table(struct net *net, u32 id)
 312{
 313	return net->ipv4.mrt;
 314}
 315
 316static int ipmr_fib_lookup(struct net *net, struct flowi4 *flp4,
 317			   struct mr_table **mrt)
 318{
 319	*mrt = net->ipv4.mrt;
 320	return 0;
 321}
 322
 323static int __net_init ipmr_rules_init(struct net *net)
 324{
 325	struct mr_table *mrt;
 326
 327	mrt = ipmr_new_table(net, RT_TABLE_DEFAULT);
 328	if (IS_ERR(mrt))
 329		return PTR_ERR(mrt);
 330	net->ipv4.mrt = mrt;
 331	return 0;
 332}
 333
 334static void __net_exit ipmr_rules_exit(struct net *net)
 335{
 336	rtnl_lock();
 337	ipmr_free_table(net->ipv4.mrt);
 338	net->ipv4.mrt = NULL;
 339	rtnl_unlock();
 340}
 341
 342static int ipmr_rules_dump(struct net *net, struct notifier_block *nb,
 343			   struct netlink_ext_ack *extack)
 344{
 345	return 0;
 346}
 347
 348static unsigned int ipmr_rules_seq_read(struct net *net)
 349{
 350	return 0;
 351}
 352
 353bool ipmr_rule_default(const struct fib_rule *rule)
 354{
 355	return true;
 356}
 357EXPORT_SYMBOL(ipmr_rule_default);
 358#endif
 359
 360static inline int ipmr_hash_cmp(struct rhashtable_compare_arg *arg,
 361				const void *ptr)
 362{
 363	const struct mfc_cache_cmp_arg *cmparg = arg->key;
 364	struct mfc_cache *c = (struct mfc_cache *)ptr;
 365
 366	return cmparg->mfc_mcastgrp != c->mfc_mcastgrp ||
 367	       cmparg->mfc_origin != c->mfc_origin;
 368}
 369
 370static const struct rhashtable_params ipmr_rht_params = {
 371	.head_offset = offsetof(struct mr_mfc, mnode),
 372	.key_offset = offsetof(struct mfc_cache, cmparg),
 373	.key_len = sizeof(struct mfc_cache_cmp_arg),
 374	.nelem_hint = 3,
 375	.obj_cmpfn = ipmr_hash_cmp,
 376	.automatic_shrinking = true,
 377};
 378
 379static void ipmr_new_table_set(struct mr_table *mrt,
 380			       struct net *net)
 381{
 382#ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
 383	list_add_tail_rcu(&mrt->list, &net->ipv4.mr_tables);
 384#endif
 385}
 386
 387static struct mfc_cache_cmp_arg ipmr_mr_table_ops_cmparg_any = {
 388	.mfc_mcastgrp = htonl(INADDR_ANY),
 389	.mfc_origin = htonl(INADDR_ANY),
 390};
 391
 392static struct mr_table_ops ipmr_mr_table_ops = {
 393	.rht_params = &ipmr_rht_params,
 394	.cmparg_any = &ipmr_mr_table_ops_cmparg_any,
 395};
 396
 397static struct mr_table *ipmr_new_table(struct net *net, u32 id)
 398{
 399	struct mr_table *mrt;
 
 400
 401	/* "pimreg%u" should not exceed 16 bytes (IFNAMSIZ) */
 402	if (id != RT_TABLE_DEFAULT && id >= 1000000000)
 403		return ERR_PTR(-EINVAL);
 404
 405	mrt = ipmr_get_table(net, id);
 406	if (mrt)
 407		return mrt;
 408
 409	return mr_table_alloc(net, id, &ipmr_mr_table_ops,
 410			      ipmr_expire_process, ipmr_new_table_set);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 411}
 412
 413static void ipmr_free_table(struct mr_table *mrt)
 414{
 415	del_timer_sync(&mrt->ipmr_expire_timer);
 416	mroute_clean_tables(mrt, MRT_FLUSH_VIFS | MRT_FLUSH_VIFS_STATIC |
 417				 MRT_FLUSH_MFC | MRT_FLUSH_MFC_STATIC);
 418	rhltable_destroy(&mrt->mfc_hash);
 419	kfree(mrt);
 420}
 421
 422/* Service routines creating virtual interfaces: DVMRP tunnels and PIMREG */
 423
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 424/* Initialize ipmr pimreg/tunnel in_device */
 425static bool ipmr_init_vif_indev(const struct net_device *dev)
 426{
 427	struct in_device *in_dev;
 428
 429	ASSERT_RTNL();
 430
 431	in_dev = __in_dev_get_rtnl(dev);
 432	if (!in_dev)
 433		return false;
 434	ipv4_devconf_setall(in_dev);
 435	neigh_parms_data_state_setall(in_dev->arp_parms);
 436	IPV4_DEVCONF(in_dev->cnf, RP_FILTER) = 0;
 437
 438	return true;
 439}
 440
 441static struct net_device *ipmr_new_tunnel(struct net *net, struct vifctl *v)
 442{
 443	struct net_device *tunnel_dev, *new_dev;
 444	struct ip_tunnel_parm p = { };
 445	int err;
 446
 447	tunnel_dev = __dev_get_by_name(net, "tunl0");
 448	if (!tunnel_dev)
 449		goto out;
 450
 451	p.iph.daddr = v->vifc_rmt_addr.s_addr;
 452	p.iph.saddr = v->vifc_lcl_addr.s_addr;
 453	p.iph.version = 4;
 454	p.iph.ihl = 5;
 455	p.iph.protocol = IPPROTO_IPIP;
 456	sprintf(p.name, "dvmrp%d", v->vifc_vifi);
 457
 458	if (!tunnel_dev->netdev_ops->ndo_tunnel_ctl)
 459		goto out;
 460	err = tunnel_dev->netdev_ops->ndo_tunnel_ctl(tunnel_dev, &p,
 461			SIOCADDTUNNEL);
 462	if (err)
 463		goto out;
 464
 465	new_dev = __dev_get_by_name(net, p.name);
 466	if (!new_dev)
 467		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 468
 469	new_dev->flags |= IFF_MULTICAST;
 470	if (!ipmr_init_vif_indev(new_dev))
 471		goto out_unregister;
 472	if (dev_open(new_dev, NULL))
 473		goto out_unregister;
 474	dev_hold(new_dev);
 475	err = dev_set_allmulti(new_dev, 1);
 476	if (err) {
 477		dev_close(new_dev);
 478		tunnel_dev->netdev_ops->ndo_tunnel_ctl(tunnel_dev, &p,
 479				SIOCDELTUNNEL);
 480		dev_put(new_dev);
 481		new_dev = ERR_PTR(err);
 482	}
 483	return new_dev;
 484
 485out_unregister:
 486	unregister_netdevice(new_dev);
 487out:
 488	return ERR_PTR(-ENOBUFS);
 489}
 490
 491#if defined(CONFIG_IP_PIMSM_V1) || defined(CONFIG_IP_PIMSM_V2)
 492static netdev_tx_t reg_vif_xmit(struct sk_buff *skb, struct net_device *dev)
 493{
 494	struct net *net = dev_net(dev);
 495	struct mr_table *mrt;
 496	struct flowi4 fl4 = {
 497		.flowi4_oif	= dev->ifindex,
 498		.flowi4_iif	= skb->skb_iif ? : LOOPBACK_IFINDEX,
 499		.flowi4_mark	= skb->mark,
 500	};
 501	int err;
 502
 503	err = ipmr_fib_lookup(net, &fl4, &mrt);
 504	if (err < 0) {
 505		kfree_skb(skb);
 506		return err;
 507	}
 508
 509	read_lock(&mrt_lock);
 510	dev->stats.tx_bytes += skb->len;
 511	dev->stats.tx_packets++;
 512	ipmr_cache_report(mrt, skb, mrt->mroute_reg_vif_num, IGMPMSG_WHOLEPKT);
 513	read_unlock(&mrt_lock);
 514	kfree_skb(skb);
 515	return NETDEV_TX_OK;
 516}
 517
 518static int reg_vif_get_iflink(const struct net_device *dev)
 519{
 520	return 0;
 521}
 522
 523static const struct net_device_ops reg_vif_netdev_ops = {
 524	.ndo_start_xmit	= reg_vif_xmit,
 525	.ndo_get_iflink = reg_vif_get_iflink,
 526};
 527
 528static void reg_vif_setup(struct net_device *dev)
 529{
 530	dev->type		= ARPHRD_PIMREG;
 531	dev->mtu		= ETH_DATA_LEN - sizeof(struct iphdr) - 8;
 532	dev->flags		= IFF_NOARP;
 533	dev->netdev_ops		= &reg_vif_netdev_ops;
 534	dev->needs_free_netdev	= true;
 535	dev->features		|= NETIF_F_NETNS_LOCAL;
 536}
 537
 538static struct net_device *ipmr_reg_vif(struct net *net, struct mr_table *mrt)
 539{
 540	struct net_device *dev;
 541	char name[IFNAMSIZ];
 542
 543	if (mrt->id == RT_TABLE_DEFAULT)
 544		sprintf(name, "pimreg");
 545	else
 546		sprintf(name, "pimreg%u", mrt->id);
 547
 548	dev = alloc_netdev(0, name, NET_NAME_UNKNOWN, reg_vif_setup);
 549
 550	if (!dev)
 551		return NULL;
 552
 553	dev_net_set(dev, net);
 554
 555	if (register_netdevice(dev)) {
 556		free_netdev(dev);
 557		return NULL;
 558	}
 559
 560	if (!ipmr_init_vif_indev(dev))
 561		goto failure;
 562	if (dev_open(dev, NULL))
 563		goto failure;
 564
 565	dev_hold(dev);
 566
 567	return dev;
 568
 569failure:
 570	unregister_netdevice(dev);
 571	return NULL;
 572}
 573
 574/* called with rcu_read_lock() */
 575static int __pim_rcv(struct mr_table *mrt, struct sk_buff *skb,
 576		     unsigned int pimlen)
 577{
 578	struct net_device *reg_dev = NULL;
 579	struct iphdr *encap;
 580
 581	encap = (struct iphdr *)(skb_transport_header(skb) + pimlen);
 582	/* Check that:
 583	 * a. packet is really sent to a multicast group
 584	 * b. packet is not a NULL-REGISTER
 585	 * c. packet is not truncated
 586	 */
 587	if (!ipv4_is_multicast(encap->daddr) ||
 588	    encap->tot_len == 0 ||
 589	    ntohs(encap->tot_len) + pimlen > skb->len)
 590		return 1;
 591
 592	read_lock(&mrt_lock);
 593	if (mrt->mroute_reg_vif_num >= 0)
 594		reg_dev = mrt->vif_table[mrt->mroute_reg_vif_num].dev;
 595	read_unlock(&mrt_lock);
 596
 597	if (!reg_dev)
 598		return 1;
 599
 600	skb->mac_header = skb->network_header;
 601	skb_pull(skb, (u8 *)encap - skb->data);
 602	skb_reset_network_header(skb);
 603	skb->protocol = htons(ETH_P_IP);
 604	skb->ip_summed = CHECKSUM_NONE;
 605
 606	skb_tunnel_rx(skb, reg_dev, dev_net(reg_dev));
 607
 608	netif_rx(skb);
 609
 610	return NET_RX_SUCCESS;
 611}
 612#else
 613static struct net_device *ipmr_reg_vif(struct net *net, struct mr_table *mrt)
 614{
 615	return NULL;
 616}
 617#endif
 618
 619static int call_ipmr_vif_entry_notifiers(struct net *net,
 620					 enum fib_event_type event_type,
 621					 struct vif_device *vif,
 622					 vifi_t vif_index, u32 tb_id)
 623{
 624	return mr_call_vif_notifiers(net, RTNL_FAMILY_IPMR, event_type,
 625				     vif, vif_index, tb_id,
 626				     &net->ipv4.ipmr_seq);
 627}
 628
 629static int call_ipmr_mfc_entry_notifiers(struct net *net,
 630					 enum fib_event_type event_type,
 631					 struct mfc_cache *mfc, u32 tb_id)
 632{
 633	return mr_call_mfc_notifiers(net, RTNL_FAMILY_IPMR, event_type,
 634				     &mfc->_c, tb_id, &net->ipv4.ipmr_seq);
 635}
 636
 637/**
 638 *	vif_delete - Delete a VIF entry
 639 *	@mrt: Table to delete from
 640 *	@vifi: VIF identifier to delete
 641 *	@notify: Set to 1, if the caller is a notifier_call
 642 *	@head: if unregistering the VIF, place it on this queue
 643 */
 644static int vif_delete(struct mr_table *mrt, int vifi, int notify,
 645		      struct list_head *head)
 646{
 647	struct net *net = read_pnet(&mrt->net);
 648	struct vif_device *v;
 649	struct net_device *dev;
 650	struct in_device *in_dev;
 651
 652	if (vifi < 0 || vifi >= mrt->maxvif)
 653		return -EADDRNOTAVAIL;
 654
 655	v = &mrt->vif_table[vifi];
 656
 657	if (VIF_EXISTS(mrt, vifi))
 658		call_ipmr_vif_entry_notifiers(net, FIB_EVENT_VIF_DEL, v, vifi,
 659					      mrt->id);
 660
 661	write_lock_bh(&mrt_lock);
 662	dev = v->dev;
 663	v->dev = NULL;
 664
 665	if (!dev) {
 666		write_unlock_bh(&mrt_lock);
 667		return -EADDRNOTAVAIL;
 668	}
 669
 670	if (vifi == mrt->mroute_reg_vif_num)
 671		mrt->mroute_reg_vif_num = -1;
 672
 673	if (vifi + 1 == mrt->maxvif) {
 674		int tmp;
 675
 676		for (tmp = vifi - 1; tmp >= 0; tmp--) {
 677			if (VIF_EXISTS(mrt, tmp))
 678				break;
 679		}
 680		mrt->maxvif = tmp+1;
 681	}
 682
 683	write_unlock_bh(&mrt_lock);
 684
 685	dev_set_allmulti(dev, -1);
 686
 687	in_dev = __in_dev_get_rtnl(dev);
 688	if (in_dev) {
 689		IPV4_DEVCONF(in_dev->cnf, MC_FORWARDING)--;
 690		inet_netconf_notify_devconf(dev_net(dev), RTM_NEWNETCONF,
 691					    NETCONFA_MC_FORWARDING,
 692					    dev->ifindex, &in_dev->cnf);
 693		ip_rt_multicast_event(in_dev);
 694	}
 695
 696	if (v->flags & (VIFF_TUNNEL | VIFF_REGISTER) && !notify)
 697		unregister_netdevice_queue(dev, head);
 698
 699	dev_put(dev);
 700	return 0;
 701}
 702
 703static void ipmr_cache_free_rcu(struct rcu_head *head)
 704{
 705	struct mr_mfc *c = container_of(head, struct mr_mfc, rcu);
 706
 707	kmem_cache_free(mrt_cachep, (struct mfc_cache *)c);
 708}
 709
 710static void ipmr_cache_free(struct mfc_cache *c)
 711{
 712	call_rcu(&c->_c.rcu, ipmr_cache_free_rcu);
 713}
 714
 715/* Destroy an unresolved cache entry, killing queued skbs
 716 * and reporting error to netlink readers.
 717 */
 718static void ipmr_destroy_unres(struct mr_table *mrt, struct mfc_cache *c)
 719{
 720	struct net *net = read_pnet(&mrt->net);
 721	struct sk_buff *skb;
 722	struct nlmsgerr *e;
 723
 724	atomic_dec(&mrt->cache_resolve_queue_len);
 725
 726	while ((skb = skb_dequeue(&c->_c.mfc_un.unres.unresolved))) {
 727		if (ip_hdr(skb)->version == 0) {
 728			struct nlmsghdr *nlh = skb_pull(skb,
 729							sizeof(struct iphdr));
 730			nlh->nlmsg_type = NLMSG_ERROR;
 731			nlh->nlmsg_len = nlmsg_msg_size(sizeof(struct nlmsgerr));
 732			skb_trim(skb, nlh->nlmsg_len);
 733			e = nlmsg_data(nlh);
 734			e->error = -ETIMEDOUT;
 735			memset(&e->msg, 0, sizeof(e->msg));
 736
 737			rtnl_unicast(skb, net, NETLINK_CB(skb).portid);
 738		} else {
 739			kfree_skb(skb);
 740		}
 741	}
 742
 743	ipmr_cache_free(c);
 744}
 745
 746/* Timer process for the unresolved queue. */
 747static void ipmr_expire_process(struct timer_list *t)
 748{
 749	struct mr_table *mrt = from_timer(mrt, t, ipmr_expire_timer);
 750	struct mr_mfc *c, *next;
 751	unsigned long expires;
 752	unsigned long now;
 
 
 753
 754	if (!spin_trylock(&mfc_unres_lock)) {
 755		mod_timer(&mrt->ipmr_expire_timer, jiffies+HZ/10);
 756		return;
 757	}
 758
 759	if (list_empty(&mrt->mfc_unres_queue))
 760		goto out;
 761
 762	now = jiffies;
 763	expires = 10*HZ;
 764
 765	list_for_each_entry_safe(c, next, &mrt->mfc_unres_queue, list) {
 766		if (time_after(c->mfc_un.unres.expires, now)) {
 767			unsigned long interval = c->mfc_un.unres.expires - now;
 768			if (interval < expires)
 769				expires = interval;
 770			continue;
 771		}
 772
 773		list_del(&c->list);
 774		mroute_netlink_event(mrt, (struct mfc_cache *)c, RTM_DELROUTE);
 775		ipmr_destroy_unres(mrt, (struct mfc_cache *)c);
 776	}
 777
 778	if (!list_empty(&mrt->mfc_unres_queue))
 779		mod_timer(&mrt->ipmr_expire_timer, jiffies + expires);
 780
 781out:
 782	spin_unlock(&mfc_unres_lock);
 783}
 784
 785/* Fill oifs list. It is called under write locked mrt_lock. */
 786static void ipmr_update_thresholds(struct mr_table *mrt, struct mr_mfc *cache,
 787				   unsigned char *ttls)
 788{
 789	int vifi;
 790
 791	cache->mfc_un.res.minvif = MAXVIFS;
 792	cache->mfc_un.res.maxvif = 0;
 793	memset(cache->mfc_un.res.ttls, 255, MAXVIFS);
 794
 795	for (vifi = 0; vifi < mrt->maxvif; vifi++) {
 796		if (VIF_EXISTS(mrt, vifi) &&
 797		    ttls[vifi] && ttls[vifi] < 255) {
 798			cache->mfc_un.res.ttls[vifi] = ttls[vifi];
 799			if (cache->mfc_un.res.minvif > vifi)
 800				cache->mfc_un.res.minvif = vifi;
 801			if (cache->mfc_un.res.maxvif <= vifi)
 802				cache->mfc_un.res.maxvif = vifi + 1;
 803		}
 804	}
 805	cache->mfc_un.res.lastuse = jiffies;
 806}
 807
 808static int vif_add(struct net *net, struct mr_table *mrt,
 809		   struct vifctl *vifc, int mrtsock)
 810{
 811	struct netdev_phys_item_id ppid = { };
 812	int vifi = vifc->vifc_vifi;
 813	struct vif_device *v = &mrt->vif_table[vifi];
 814	struct net_device *dev;
 815	struct in_device *in_dev;
 816	int err;
 817
 818	/* Is vif busy ? */
 819	if (VIF_EXISTS(mrt, vifi))
 820		return -EADDRINUSE;
 821
 822	switch (vifc->vifc_flags) {
 823	case VIFF_REGISTER:
 824		if (!ipmr_pimsm_enabled())
 825			return -EINVAL;
 826		/* Special Purpose VIF in PIM
 827		 * All the packets will be sent to the daemon
 828		 */
 829		if (mrt->mroute_reg_vif_num >= 0)
 830			return -EADDRINUSE;
 831		dev = ipmr_reg_vif(net, mrt);
 832		if (!dev)
 833			return -ENOBUFS;
 834		err = dev_set_allmulti(dev, 1);
 835		if (err) {
 836			unregister_netdevice(dev);
 837			dev_put(dev);
 838			return err;
 839		}
 840		break;
 841	case VIFF_TUNNEL:
 842		dev = ipmr_new_tunnel(net, vifc);
 843		if (IS_ERR(dev))
 844			return PTR_ERR(dev);
 
 
 
 
 
 
 845		break;
 846	case VIFF_USE_IFINDEX:
 847	case 0:
 848		if (vifc->vifc_flags == VIFF_USE_IFINDEX) {
 849			dev = dev_get_by_index(net, vifc->vifc_lcl_ifindex);
 850			if (dev && !__in_dev_get_rtnl(dev)) {
 851				dev_put(dev);
 852				return -EADDRNOTAVAIL;
 853			}
 854		} else {
 855			dev = ip_dev_find(net, vifc->vifc_lcl_addr.s_addr);
 856		}
 857		if (!dev)
 858			return -EADDRNOTAVAIL;
 859		err = dev_set_allmulti(dev, 1);
 860		if (err) {
 861			dev_put(dev);
 862			return err;
 863		}
 864		break;
 865	default:
 866		return -EINVAL;
 867	}
 868
 869	in_dev = __in_dev_get_rtnl(dev);
 870	if (!in_dev) {
 871		dev_put(dev);
 872		return -EADDRNOTAVAIL;
 873	}
 874	IPV4_DEVCONF(in_dev->cnf, MC_FORWARDING)++;
 875	inet_netconf_notify_devconf(net, RTM_NEWNETCONF, NETCONFA_MC_FORWARDING,
 876				    dev->ifindex, &in_dev->cnf);
 877	ip_rt_multicast_event(in_dev);
 878
 879	/* Fill in the VIF structures */
 880	vif_device_init(v, dev, vifc->vifc_rate_limit,
 881			vifc->vifc_threshold,
 882			vifc->vifc_flags | (!mrtsock ? VIFF_STATIC : 0),
 883			(VIFF_TUNNEL | VIFF_REGISTER));
 884
 885	err = dev_get_port_parent_id(dev, &ppid, true);
 886	if (err == 0) {
 887		memcpy(v->dev_parent_id.id, ppid.id, ppid.id_len);
 888		v->dev_parent_id.id_len = ppid.id_len;
 889	} else {
 890		v->dev_parent_id.id_len = 0;
 891	}
 892
 
 893	v->local = vifc->vifc_lcl_addr.s_addr;
 894	v->remote = vifc->vifc_rmt_addr.s_addr;
 
 
 
 
 
 
 
 
 
 
 
 895
 896	/* And finish update writing critical data */
 897	write_lock_bh(&mrt_lock);
 898	v->dev = dev;
 899	if (v->flags & VIFF_REGISTER)
 900		mrt->mroute_reg_vif_num = vifi;
 901	if (vifi+1 > mrt->maxvif)
 902		mrt->maxvif = vifi+1;
 903	write_unlock_bh(&mrt_lock);
 904	call_ipmr_vif_entry_notifiers(net, FIB_EVENT_VIF_ADD, v, vifi, mrt->id);
 905	return 0;
 906}
 907
 908/* called with rcu_read_lock() */
 909static struct mfc_cache *ipmr_cache_find(struct mr_table *mrt,
 910					 __be32 origin,
 911					 __be32 mcastgrp)
 912{
 913	struct mfc_cache_cmp_arg arg = {
 914			.mfc_mcastgrp = mcastgrp,
 915			.mfc_origin = origin
 916	};
 
 
 
 
 
 
 
 
 
 
 
 
 917
 918	return mr_mfc_find(mrt, &arg);
 
 
 
 
 
 
 919}
 920
 921/* Look for a (*,G) entry */
 922static struct mfc_cache *ipmr_cache_find_any(struct mr_table *mrt,
 923					     __be32 mcastgrp, int vifi)
 924{
 925	struct mfc_cache_cmp_arg arg = {
 926			.mfc_mcastgrp = mcastgrp,
 927			.mfc_origin = htonl(INADDR_ANY)
 928	};
 929
 930	if (mcastgrp == htonl(INADDR_ANY))
 931		return mr_mfc_find_any_parent(mrt, vifi);
 932	return mr_mfc_find_any(mrt, vifi, &arg);
 933}
 934
 935/* Look for a (S,G,iif) entry if parent != -1 */
 936static struct mfc_cache *ipmr_cache_find_parent(struct mr_table *mrt,
 937						__be32 origin, __be32 mcastgrp,
 938						int parent)
 939{
 940	struct mfc_cache_cmp_arg arg = {
 941			.mfc_mcastgrp = mcastgrp,
 942			.mfc_origin = origin,
 943	};
 
 
 
 944
 945	return mr_mfc_find_parent(mrt, &arg, parent);
 
 946}
 947
 948/* Allocate a multicast cache entry */
 949static struct mfc_cache *ipmr_cache_alloc(void)
 950{
 951	struct mfc_cache *c = kmem_cache_zalloc(mrt_cachep, GFP_KERNEL);
 952
 953	if (c) {
 954		c->_c.mfc_un.res.last_assert = jiffies - MFC_ASSERT_THRESH - 1;
 955		c->_c.mfc_un.res.minvif = MAXVIFS;
 956		c->_c.free = ipmr_cache_free_rcu;
 957		refcount_set(&c->_c.mfc_un.res.refcount, 1);
 958	}
 959	return c;
 960}
 961
 962static struct mfc_cache *ipmr_cache_alloc_unres(void)
 963{
 964	struct mfc_cache *c = kmem_cache_zalloc(mrt_cachep, GFP_ATOMIC);
 965
 966	if (c) {
 967		skb_queue_head_init(&c->_c.mfc_un.unres.unresolved);
 968		c->_c.mfc_un.unres.expires = jiffies + 10 * HZ;
 969	}
 970	return c;
 971}
 972
 973/* A cache entry has gone into a resolved state from queued */
 974static void ipmr_cache_resolve(struct net *net, struct mr_table *mrt,
 975			       struct mfc_cache *uc, struct mfc_cache *c)
 976{
 977	struct sk_buff *skb;
 978	struct nlmsgerr *e;
 979
 980	/* Play the pending entries through our router */
 981	while ((skb = __skb_dequeue(&uc->_c.mfc_un.unres.unresolved))) {
 982		if (ip_hdr(skb)->version == 0) {
 983			struct nlmsghdr *nlh = skb_pull(skb,
 984							sizeof(struct iphdr));
 985
 986			if (mr_fill_mroute(mrt, skb, &c->_c,
 987					   nlmsg_data(nlh)) > 0) {
 988				nlh->nlmsg_len = skb_tail_pointer(skb) -
 989						 (u8 *)nlh;
 990			} else {
 991				nlh->nlmsg_type = NLMSG_ERROR;
 992				nlh->nlmsg_len = nlmsg_msg_size(sizeof(struct nlmsgerr));
 993				skb_trim(skb, nlh->nlmsg_len);
 994				e = nlmsg_data(nlh);
 995				e->error = -EMSGSIZE;
 996				memset(&e->msg, 0, sizeof(e->msg));
 997			}
 998
 999			rtnl_unicast(skb, net, NETLINK_CB(skb).portid);
1000		} else {
1001			ip_mr_forward(net, mrt, skb->dev, skb, c, 0);
1002		}
1003	}
1004}
1005
1006/* Bounce a cache query up to mrouted and netlink.
 
1007 *
1008 * Called under mrt_lock.
1009 */
1010static int ipmr_cache_report(struct mr_table *mrt,
1011			     struct sk_buff *pkt, vifi_t vifi, int assert)
1012{
1013	const int ihl = ip_hdrlen(pkt);
1014	struct sock *mroute_sk;
1015	struct igmphdr *igmp;
1016	struct igmpmsg *msg;
1017	struct sk_buff *skb;
1018	int ret;
1019
1020	if (assert == IGMPMSG_WHOLEPKT || assert == IGMPMSG_WRVIFWHOLE)
1021		skb = skb_realloc_headroom(pkt, sizeof(struct iphdr));
1022	else
1023		skb = alloc_skb(128, GFP_ATOMIC);
1024
1025	if (!skb)
1026		return -ENOBUFS;
1027
1028	if (assert == IGMPMSG_WHOLEPKT || assert == IGMPMSG_WRVIFWHOLE) {
1029		/* Ugly, but we have no choice with this interface.
1030		 * Duplicate old header, fix ihl, length etc.
1031		 * And all this only to mangle msg->im_msgtype and
1032		 * to set msg->im_mbz to "mbz" :-)
1033		 */
1034		skb_push(skb, sizeof(struct iphdr));
1035		skb_reset_network_header(skb);
1036		skb_reset_transport_header(skb);
1037		msg = (struct igmpmsg *)skb_network_header(skb);
1038		memcpy(msg, skb_network_header(pkt), sizeof(struct iphdr));
1039		msg->im_msgtype = assert;
1040		msg->im_mbz = 0;
1041		if (assert == IGMPMSG_WRVIFWHOLE) {
1042			msg->im_vif = vifi;
1043			msg->im_vif_hi = vifi >> 8;
1044		} else {
1045			msg->im_vif = mrt->mroute_reg_vif_num;
1046			msg->im_vif_hi = mrt->mroute_reg_vif_num >> 8;
1047		}
1048		ip_hdr(skb)->ihl = sizeof(struct iphdr) >> 2;
1049		ip_hdr(skb)->tot_len = htons(ntohs(ip_hdr(pkt)->tot_len) +
1050					     sizeof(struct iphdr));
1051	} else {
1052		/* Copy the IP header */
1053		skb_set_network_header(skb, skb->len);
1054		skb_put(skb, ihl);
1055		skb_copy_to_linear_data(skb, pkt->data, ihl);
1056		/* Flag to the kernel this is a route add */
1057		ip_hdr(skb)->protocol = 0;
1058		msg = (struct igmpmsg *)skb_network_header(skb);
1059		msg->im_vif = vifi;
1060		msg->im_vif_hi = vifi >> 8;
1061		skb_dst_set(skb, dst_clone(skb_dst(pkt)));
1062		/* Add our header */
1063		igmp = skb_put(skb, sizeof(struct igmphdr));
1064		igmp->type = assert;
1065		msg->im_msgtype = assert;
1066		igmp->code = 0;
1067		ip_hdr(skb)->tot_len = htons(skb->len);	/* Fix the length */
1068		skb->transport_header = skb->network_header;
1069	}
1070
1071	rcu_read_lock();
1072	mroute_sk = rcu_dereference(mrt->mroute_sk);
1073	if (!mroute_sk) {
1074		rcu_read_unlock();
1075		kfree_skb(skb);
1076		return -EINVAL;
1077	}
1078
1079	igmpmsg_netlink_event(mrt, skb);
1080
1081	/* Deliver to mrouted */
1082	ret = sock_queue_rcv_skb(mroute_sk, skb);
1083	rcu_read_unlock();
1084	if (ret < 0) {
1085		net_warn_ratelimited("mroute: pending queue full, dropping entries\n");
1086		kfree_skb(skb);
1087	}
1088
1089	return ret;
1090}
1091
1092/* Queue a packet for resolution. It gets locked cache entry! */
1093static int ipmr_cache_unresolved(struct mr_table *mrt, vifi_t vifi,
1094				 struct sk_buff *skb, struct net_device *dev)
1095{
1096	const struct iphdr *iph = ip_hdr(skb);
1097	struct mfc_cache *c;
1098	bool found = false;
1099	int err;
 
 
1100
1101	spin_lock_bh(&mfc_unres_lock);
1102	list_for_each_entry(c, &mrt->mfc_unres_queue, _c.list) {
1103		if (c->mfc_mcastgrp == iph->daddr &&
1104		    c->mfc_origin == iph->saddr) {
1105			found = true;
1106			break;
1107		}
1108	}
1109
1110	if (!found) {
1111		/* Create a new entry if allowable */
1112		c = ipmr_cache_alloc_unres();
1113		if (!c) {
1114			spin_unlock_bh(&mfc_unres_lock);
1115
1116			kfree_skb(skb);
1117			return -ENOBUFS;
1118		}
1119
1120		/* Fill in the new cache entry */
1121		c->_c.mfc_parent = -1;
1122		c->mfc_origin	= iph->saddr;
1123		c->mfc_mcastgrp	= iph->daddr;
1124
1125		/* Reflect first query at mrouted. */
1126		err = ipmr_cache_report(mrt, skb, vifi, IGMPMSG_NOCACHE);
1127
1128		if (err < 0) {
1129			/* If the report failed throw the cache entry
1130			   out - Brad Parker
1131			 */
1132			spin_unlock_bh(&mfc_unres_lock);
1133
1134			ipmr_cache_free(c);
1135			kfree_skb(skb);
1136			return err;
1137		}
1138
1139		atomic_inc(&mrt->cache_resolve_queue_len);
1140		list_add(&c->_c.list, &mrt->mfc_unres_queue);
1141		mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1142
1143		if (atomic_read(&mrt->cache_resolve_queue_len) == 1)
1144			mod_timer(&mrt->ipmr_expire_timer,
1145				  c->_c.mfc_un.unres.expires);
1146	}
1147
1148	/* See if we can append the packet */
1149	if (c->_c.mfc_un.unres.unresolved.qlen > 3) {
1150		kfree_skb(skb);
1151		err = -ENOBUFS;
1152	} else {
1153		if (dev) {
1154			skb->dev = dev;
1155			skb->skb_iif = dev->ifindex;
1156		}
1157		skb_queue_tail(&c->_c.mfc_un.unres.unresolved, skb);
1158		err = 0;
1159	}
1160
1161	spin_unlock_bh(&mfc_unres_lock);
1162	return err;
1163}
1164
1165/* MFC cache manipulation by user space mroute daemon */
1166
1167static int ipmr_mfc_delete(struct mr_table *mrt, struct mfcctl *mfc, int parent)
1168{
1169	struct net *net = read_pnet(&mrt->net);
1170	struct mfc_cache *c;
1171
1172	/* The entries are added/deleted only under RTNL */
1173	rcu_read_lock();
1174	c = ipmr_cache_find_parent(mrt, mfc->mfcc_origin.s_addr,
1175				   mfc->mfcc_mcastgrp.s_addr, parent);
1176	rcu_read_unlock();
1177	if (!c)
1178		return -ENOENT;
1179	rhltable_remove(&mrt->mfc_hash, &c->_c.mnode, ipmr_rht_params);
1180	list_del_rcu(&c->_c.list);
1181	call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, c, mrt->id);
1182	mroute_netlink_event(mrt, c, RTM_DELROUTE);
1183	mr_cache_put(&c->_c);
1184
1185	return 0;
 
 
 
 
 
 
 
 
 
 
1186}
1187
1188static int ipmr_mfc_add(struct net *net, struct mr_table *mrt,
1189			struct mfcctl *mfc, int mrtsock, int parent)
1190{
 
 
1191	struct mfc_cache *uc, *c;
1192	struct mr_mfc *_uc;
1193	bool found;
1194	int ret;
1195
1196	if (mfc->mfcc_parent >= MAXVIFS)
1197		return -ENFILE;
1198
1199	/* The entries are added/deleted only under RTNL */
1200	rcu_read_lock();
1201	c = ipmr_cache_find_parent(mrt, mfc->mfcc_origin.s_addr,
1202				   mfc->mfcc_mcastgrp.s_addr, parent);
1203	rcu_read_unlock();
1204	if (c) {
 
 
 
 
 
 
1205		write_lock_bh(&mrt_lock);
1206		c->_c.mfc_parent = mfc->mfcc_parent;
1207		ipmr_update_thresholds(mrt, &c->_c, mfc->mfcc_ttls);
1208		if (!mrtsock)
1209			c->_c.mfc_flags |= MFC_STATIC;
1210		write_unlock_bh(&mrt_lock);
1211		call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_REPLACE, c,
1212					      mrt->id);
1213		mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1214		return 0;
1215	}
1216
1217	if (mfc->mfcc_mcastgrp.s_addr != htonl(INADDR_ANY) &&
1218	    !ipv4_is_multicast(mfc->mfcc_mcastgrp.s_addr))
1219		return -EINVAL;
1220
1221	c = ipmr_cache_alloc();
1222	if (!c)
1223		return -ENOMEM;
1224
1225	c->mfc_origin = mfc->mfcc_origin.s_addr;
1226	c->mfc_mcastgrp = mfc->mfcc_mcastgrp.s_addr;
1227	c->_c.mfc_parent = mfc->mfcc_parent;
1228	ipmr_update_thresholds(mrt, &c->_c, mfc->mfcc_ttls);
1229	if (!mrtsock)
1230		c->_c.mfc_flags |= MFC_STATIC;
 
 
1231
1232	ret = rhltable_insert_key(&mrt->mfc_hash, &c->cmparg, &c->_c.mnode,
1233				  ipmr_rht_params);
1234	if (ret) {
1235		pr_err("ipmr: rhtable insert error %d\n", ret);
1236		ipmr_cache_free(c);
1237		return ret;
1238	}
1239	list_add_tail_rcu(&c->_c.list, &mrt->mfc_cache_list);
1240	/* Check to see if we resolved a queued list. If so we
1241	 * need to send on the frames and tidy up.
1242	 */
1243	found = false;
1244	spin_lock_bh(&mfc_unres_lock);
1245	list_for_each_entry(_uc, &mrt->mfc_unres_queue, list) {
1246		uc = (struct mfc_cache *)_uc;
1247		if (uc->mfc_origin == c->mfc_origin &&
1248		    uc->mfc_mcastgrp == c->mfc_mcastgrp) {
1249			list_del(&_uc->list);
1250			atomic_dec(&mrt->cache_resolve_queue_len);
1251			found = true;
1252			break;
1253		}
1254	}
1255	if (list_empty(&mrt->mfc_unres_queue))
1256		del_timer(&mrt->ipmr_expire_timer);
1257	spin_unlock_bh(&mfc_unres_lock);
1258
1259	if (found) {
1260		ipmr_cache_resolve(net, mrt, uc, c);
1261		ipmr_cache_free(uc);
1262	}
1263	call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_ADD, c, mrt->id);
1264	mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1265	return 0;
1266}
1267
1268/* Close the multicast socket, and clear the vif tables etc */
1269static void mroute_clean_tables(struct mr_table *mrt, int flags)
1270{
1271	struct net *net = read_pnet(&mrt->net);
1272	struct mr_mfc *c, *tmp;
1273	struct mfc_cache *cache;
1274	LIST_HEAD(list);
1275	int i;
 
 
1276
1277	/* Shut down all active vif entries */
1278	if (flags & (MRT_FLUSH_VIFS | MRT_FLUSH_VIFS_STATIC)) {
1279		for (i = 0; i < mrt->maxvif; i++) {
1280			if (((mrt->vif_table[i].flags & VIFF_STATIC) &&
1281			     !(flags & MRT_FLUSH_VIFS_STATIC)) ||
1282			    (!(mrt->vif_table[i].flags & VIFF_STATIC) && !(flags & MRT_FLUSH_VIFS)))
1283				continue;
1284			vif_delete(mrt, i, 0, &list);
1285		}
1286		unregister_netdevice_many(&list);
1287	}
 
1288
1289	/* Wipe the cache */
1290	if (flags & (MRT_FLUSH_MFC | MRT_FLUSH_MFC_STATIC)) {
1291		list_for_each_entry_safe(c, tmp, &mrt->mfc_cache_list, list) {
1292			if (((c->mfc_flags & MFC_STATIC) && !(flags & MRT_FLUSH_MFC_STATIC)) ||
1293			    (!(c->mfc_flags & MFC_STATIC) && !(flags & MRT_FLUSH_MFC)))
1294				continue;
1295			rhltable_remove(&mrt->mfc_hash, &c->mnode, ipmr_rht_params);
1296			list_del_rcu(&c->list);
1297			cache = (struct mfc_cache *)c;
1298			call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, cache,
1299						      mrt->id);
1300			mroute_netlink_event(mrt, cache, RTM_DELROUTE);
1301			mr_cache_put(c);
1302		}
1303	}
1304
1305	if (flags & MRT_FLUSH_MFC) {
1306		if (atomic_read(&mrt->cache_resolve_queue_len) != 0) {
1307			spin_lock_bh(&mfc_unres_lock);
1308			list_for_each_entry_safe(c, tmp, &mrt->mfc_unres_queue, list) {
1309				list_del(&c->list);
1310				cache = (struct mfc_cache *)c;
1311				mroute_netlink_event(mrt, cache, RTM_DELROUTE);
1312				ipmr_destroy_unres(mrt, cache);
1313			}
1314			spin_unlock_bh(&mfc_unres_lock);
1315		}
 
1316	}
1317}
1318
1319/* called from ip_ra_control(), before an RCU grace period,
1320 * we don't need to call synchronize_rcu() here
1321 */
1322static void mrtsock_destruct(struct sock *sk)
1323{
1324	struct net *net = sock_net(sk);
1325	struct mr_table *mrt;
1326
1327	rtnl_lock();
1328	ipmr_for_each_table(mrt, net) {
1329		if (sk == rtnl_dereference(mrt->mroute_sk)) {
1330			IPV4_DEVCONF_ALL(net, MC_FORWARDING)--;
1331			inet_netconf_notify_devconf(net, RTM_NEWNETCONF,
1332						    NETCONFA_MC_FORWARDING,
1333						    NETCONFA_IFINDEX_ALL,
1334						    net->ipv4.devconf_all);
1335			RCU_INIT_POINTER(mrt->mroute_sk, NULL);
1336			mroute_clean_tables(mrt, MRT_FLUSH_VIFS | MRT_FLUSH_MFC);
1337		}
1338	}
1339	rtnl_unlock();
1340}
1341
1342/* Socket options and virtual interface manipulation. The whole
1343 * virtual interface system is a complete heap, but unfortunately
1344 * that's how BSD mrouted happens to think. Maybe one day with a proper
1345 * MOSPF/PIM router set up we can clean this up.
1346 */
1347
1348int ip_mroute_setsockopt(struct sock *sk, int optname, sockptr_t optval,
1349			 unsigned int optlen)
1350{
1351	struct net *net = sock_net(sk);
1352	int val, ret = 0, parent = 0;
1353	struct mr_table *mrt;
1354	struct vifctl vif;
1355	struct mfcctl mfc;
1356	bool do_wrvifwhole;
1357	u32 uval;
1358
1359	/* There's one exception to the lock - MRT_DONE which needs to unlock */
1360	rtnl_lock();
1361	if (sk->sk_type != SOCK_RAW ||
1362	    inet_sk(sk)->inet_num != IPPROTO_IGMP) {
1363		ret = -EOPNOTSUPP;
1364		goto out_unlock;
1365	}
1366
1367	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1368	if (!mrt) {
1369		ret = -ENOENT;
1370		goto out_unlock;
1371	}
1372	if (optname != MRT_INIT) {
1373		if (sk != rcu_access_pointer(mrt->mroute_sk) &&
1374		    !ns_capable(net->user_ns, CAP_NET_ADMIN)) {
1375			ret = -EACCES;
1376			goto out_unlock;
1377		}
1378	}
1379
1380	switch (optname) {
1381	case MRT_INIT:
1382		if (optlen != sizeof(int)) {
1383			ret = -EINVAL;
1384			break;
1385		}
1386		if (rtnl_dereference(mrt->mroute_sk)) {
1387			ret = -EADDRINUSE;
1388			break;
1389		}
1390
1391		ret = ip_ra_control(sk, 1, mrtsock_destruct);
1392		if (ret == 0) {
1393			rcu_assign_pointer(mrt->mroute_sk, sk);
1394			IPV4_DEVCONF_ALL(net, MC_FORWARDING)++;
1395			inet_netconf_notify_devconf(net, RTM_NEWNETCONF,
1396						    NETCONFA_MC_FORWARDING,
1397						    NETCONFA_IFINDEX_ALL,
1398						    net->ipv4.devconf_all);
1399		}
1400		break;
1401	case MRT_DONE:
1402		if (sk != rcu_access_pointer(mrt->mroute_sk)) {
1403			ret = -EACCES;
1404		} else {
1405			/* We need to unlock here because mrtsock_destruct takes
1406			 * care of rtnl itself and we can't change that due to
1407			 * the IP_ROUTER_ALERT setsockopt which runs without it.
1408			 */
1409			rtnl_unlock();
1410			ret = ip_ra_control(sk, 0, NULL);
1411			goto out;
1412		}
1413		break;
1414	case MRT_ADD_VIF:
1415	case MRT_DEL_VIF:
1416		if (optlen != sizeof(vif)) {
1417			ret = -EINVAL;
1418			break;
1419		}
1420		if (copy_from_sockptr(&vif, optval, sizeof(vif))) {
1421			ret = -EFAULT;
1422			break;
1423		}
1424		if (vif.vifc_vifi >= MAXVIFS) {
1425			ret = -ENFILE;
1426			break;
1427		}
1428		if (optname == MRT_ADD_VIF) {
1429			ret = vif_add(net, mrt, &vif,
1430				      sk == rtnl_dereference(mrt->mroute_sk));
1431		} else {
1432			ret = vif_delete(mrt, vif.vifc_vifi, 0, NULL);
1433		}
1434		break;
1435	/* Manipulate the forwarding caches. These live
1436	 * in a sort of kernel/user symbiosis.
1437	 */
1438	case MRT_ADD_MFC:
1439	case MRT_DEL_MFC:
1440		parent = -1;
1441		fallthrough;
1442	case MRT_ADD_MFC_PROXY:
1443	case MRT_DEL_MFC_PROXY:
1444		if (optlen != sizeof(mfc)) {
1445			ret = -EINVAL;
1446			break;
1447		}
1448		if (copy_from_sockptr(&mfc, optval, sizeof(mfc))) {
1449			ret = -EFAULT;
1450			break;
1451		}
1452		if (parent == 0)
1453			parent = mfc.mfcc_parent;
1454		if (optname == MRT_DEL_MFC || optname == MRT_DEL_MFC_PROXY)
1455			ret = ipmr_mfc_delete(mrt, &mfc, parent);
1456		else
1457			ret = ipmr_mfc_add(net, mrt, &mfc,
1458					   sk == rtnl_dereference(mrt->mroute_sk),
1459					   parent);
1460		break;
1461	case MRT_FLUSH:
1462		if (optlen != sizeof(val)) {
1463			ret = -EINVAL;
1464			break;
1465		}
1466		if (copy_from_sockptr(&val, optval, sizeof(val))) {
1467			ret = -EFAULT;
1468			break;
1469		}
1470		mroute_clean_tables(mrt, val);
1471		break;
1472	/* Control PIM assert. */
1473	case MRT_ASSERT:
1474		if (optlen != sizeof(val)) {
1475			ret = -EINVAL;
1476			break;
1477		}
1478		if (copy_from_sockptr(&val, optval, sizeof(val))) {
1479			ret = -EFAULT;
1480			break;
1481		}
1482		mrt->mroute_do_assert = val;
1483		break;
1484	case MRT_PIM:
1485		if (!ipmr_pimsm_enabled()) {
1486			ret = -ENOPROTOOPT;
1487			break;
1488		}
1489		if (optlen != sizeof(val)) {
1490			ret = -EINVAL;
1491			break;
1492		}
1493		if (copy_from_sockptr(&val, optval, sizeof(val))) {
1494			ret = -EFAULT;
1495			break;
1496		}
1497
1498		do_wrvifwhole = (val == IGMPMSG_WRVIFWHOLE);
1499		val = !!val;
1500		if (val != mrt->mroute_do_pim) {
1501			mrt->mroute_do_pim = val;
1502			mrt->mroute_do_assert = val;
1503			mrt->mroute_do_wrvifwhole = do_wrvifwhole;
1504		}
1505		break;
1506	case MRT_TABLE:
1507		if (!IS_BUILTIN(CONFIG_IP_MROUTE_MULTIPLE_TABLES)) {
1508			ret = -ENOPROTOOPT;
1509			break;
1510		}
1511		if (optlen != sizeof(uval)) {
1512			ret = -EINVAL;
1513			break;
1514		}
1515		if (copy_from_sockptr(&uval, optval, sizeof(uval))) {
1516			ret = -EFAULT;
1517			break;
1518		}
1519
1520		if (sk == rtnl_dereference(mrt->mroute_sk)) {
1521			ret = -EBUSY;
1522		} else {
1523			mrt = ipmr_new_table(net, uval);
1524			if (IS_ERR(mrt))
1525				ret = PTR_ERR(mrt);
1526			else
1527				raw_sk(sk)->ipmr_table = uval;
1528		}
1529		break;
1530	/* Spurious command, or MRT_VERSION which you cannot set. */
1531	default:
1532		ret = -ENOPROTOOPT;
1533	}
1534out_unlock:
1535	rtnl_unlock();
1536out:
1537	return ret;
1538}
1539
1540/* Getsock opt support for the multicast routing system. */
1541int ip_mroute_getsockopt(struct sock *sk, int optname, char __user *optval, int __user *optlen)
1542{
1543	int olr;
1544	int val;
1545	struct net *net = sock_net(sk);
1546	struct mr_table *mrt;
1547
1548	if (sk->sk_type != SOCK_RAW ||
1549	    inet_sk(sk)->inet_num != IPPROTO_IGMP)
1550		return -EOPNOTSUPP;
1551
1552	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1553	if (!mrt)
1554		return -ENOENT;
1555
1556	switch (optname) {
1557	case MRT_VERSION:
1558		val = 0x0305;
1559		break;
1560	case MRT_PIM:
1561		if (!ipmr_pimsm_enabled())
1562			return -ENOPROTOOPT;
1563		val = mrt->mroute_do_pim;
1564		break;
1565	case MRT_ASSERT:
1566		val = mrt->mroute_do_assert;
1567		break;
1568	default:
1569		return -ENOPROTOOPT;
1570	}
1571
1572	if (get_user(olr, optlen))
1573		return -EFAULT;
1574	olr = min_t(unsigned int, olr, sizeof(int));
1575	if (olr < 0)
1576		return -EINVAL;
1577	if (put_user(olr, optlen))
1578		return -EFAULT;
1579	if (copy_to_user(optval, &val, olr))
1580		return -EFAULT;
1581	return 0;
1582}
1583
1584/* The IP multicast ioctl support routines. */
1585int ipmr_ioctl(struct sock *sk, int cmd, void __user *arg)
1586{
1587	struct sioc_sg_req sr;
1588	struct sioc_vif_req vr;
1589	struct vif_device *vif;
1590	struct mfc_cache *c;
1591	struct net *net = sock_net(sk);
1592	struct mr_table *mrt;
1593
1594	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1595	if (!mrt)
1596		return -ENOENT;
1597
1598	switch (cmd) {
1599	case SIOCGETVIFCNT:
1600		if (copy_from_user(&vr, arg, sizeof(vr)))
1601			return -EFAULT;
1602		if (vr.vifi >= mrt->maxvif)
1603			return -EINVAL;
1604		vr.vifi = array_index_nospec(vr.vifi, mrt->maxvif);
1605		read_lock(&mrt_lock);
1606		vif = &mrt->vif_table[vr.vifi];
1607		if (VIF_EXISTS(mrt, vr.vifi)) {
1608			vr.icount = vif->pkt_in;
1609			vr.ocount = vif->pkt_out;
1610			vr.ibytes = vif->bytes_in;
1611			vr.obytes = vif->bytes_out;
1612			read_unlock(&mrt_lock);
1613
1614			if (copy_to_user(arg, &vr, sizeof(vr)))
1615				return -EFAULT;
1616			return 0;
1617		}
1618		read_unlock(&mrt_lock);
1619		return -EADDRNOTAVAIL;
1620	case SIOCGETSGCNT:
1621		if (copy_from_user(&sr, arg, sizeof(sr)))
1622			return -EFAULT;
1623
1624		rcu_read_lock();
1625		c = ipmr_cache_find(mrt, sr.src.s_addr, sr.grp.s_addr);
1626		if (c) {
1627			sr.pktcnt = c->_c.mfc_un.res.pkt;
1628			sr.bytecnt = c->_c.mfc_un.res.bytes;
1629			sr.wrong_if = c->_c.mfc_un.res.wrong_if;
1630			rcu_read_unlock();
1631
1632			if (copy_to_user(arg, &sr, sizeof(sr)))
1633				return -EFAULT;
1634			return 0;
1635		}
1636		rcu_read_unlock();
1637		return -EADDRNOTAVAIL;
1638	default:
1639		return -ENOIOCTLCMD;
1640	}
1641}
1642
1643#ifdef CONFIG_COMPAT
1644struct compat_sioc_sg_req {
1645	struct in_addr src;
1646	struct in_addr grp;
1647	compat_ulong_t pktcnt;
1648	compat_ulong_t bytecnt;
1649	compat_ulong_t wrong_if;
1650};
1651
1652struct compat_sioc_vif_req {
1653	vifi_t	vifi;		/* Which iface */
1654	compat_ulong_t icount;
1655	compat_ulong_t ocount;
1656	compat_ulong_t ibytes;
1657	compat_ulong_t obytes;
1658};
1659
1660int ipmr_compat_ioctl(struct sock *sk, unsigned int cmd, void __user *arg)
1661{
1662	struct compat_sioc_sg_req sr;
1663	struct compat_sioc_vif_req vr;
1664	struct vif_device *vif;
1665	struct mfc_cache *c;
1666	struct net *net = sock_net(sk);
1667	struct mr_table *mrt;
1668
1669	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1670	if (!mrt)
1671		return -ENOENT;
1672
1673	switch (cmd) {
1674	case SIOCGETVIFCNT:
1675		if (copy_from_user(&vr, arg, sizeof(vr)))
1676			return -EFAULT;
1677		if (vr.vifi >= mrt->maxvif)
1678			return -EINVAL;
1679		vr.vifi = array_index_nospec(vr.vifi, mrt->maxvif);
1680		read_lock(&mrt_lock);
1681		vif = &mrt->vif_table[vr.vifi];
1682		if (VIF_EXISTS(mrt, vr.vifi)) {
1683			vr.icount = vif->pkt_in;
1684			vr.ocount = vif->pkt_out;
1685			vr.ibytes = vif->bytes_in;
1686			vr.obytes = vif->bytes_out;
1687			read_unlock(&mrt_lock);
1688
1689			if (copy_to_user(arg, &vr, sizeof(vr)))
1690				return -EFAULT;
1691			return 0;
1692		}
1693		read_unlock(&mrt_lock);
1694		return -EADDRNOTAVAIL;
1695	case SIOCGETSGCNT:
1696		if (copy_from_user(&sr, arg, sizeof(sr)))
1697			return -EFAULT;
1698
1699		rcu_read_lock();
1700		c = ipmr_cache_find(mrt, sr.src.s_addr, sr.grp.s_addr);
1701		if (c) {
1702			sr.pktcnt = c->_c.mfc_un.res.pkt;
1703			sr.bytecnt = c->_c.mfc_un.res.bytes;
1704			sr.wrong_if = c->_c.mfc_un.res.wrong_if;
1705			rcu_read_unlock();
1706
1707			if (copy_to_user(arg, &sr, sizeof(sr)))
1708				return -EFAULT;
1709			return 0;
1710		}
1711		rcu_read_unlock();
1712		return -EADDRNOTAVAIL;
1713	default:
1714		return -ENOIOCTLCMD;
1715	}
1716}
1717#endif
1718
1719static int ipmr_device_event(struct notifier_block *this, unsigned long event, void *ptr)
1720{
1721	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1722	struct net *net = dev_net(dev);
1723	struct mr_table *mrt;
1724	struct vif_device *v;
1725	int ct;
1726
1727	if (event != NETDEV_UNREGISTER)
1728		return NOTIFY_DONE;
1729
1730	ipmr_for_each_table(mrt, net) {
1731		v = &mrt->vif_table[0];
1732		for (ct = 0; ct < mrt->maxvif; ct++, v++) {
1733			if (v->dev == dev)
1734				vif_delete(mrt, ct, 1, NULL);
1735		}
1736	}
1737	return NOTIFY_DONE;
1738}
1739
1740static struct notifier_block ip_mr_notifier = {
1741	.notifier_call = ipmr_device_event,
1742};
1743
1744/* Encapsulate a packet by attaching a valid IPIP header to it.
1745 * This avoids tunnel drivers and other mess and gives us the speed so
1746 * important for multicast video.
1747 */
1748static void ip_encap(struct net *net, struct sk_buff *skb,
1749		     __be32 saddr, __be32 daddr)
1750{
1751	struct iphdr *iph;
1752	const struct iphdr *old_iph = ip_hdr(skb);
1753
1754	skb_push(skb, sizeof(struct iphdr));
1755	skb->transport_header = skb->network_header;
1756	skb_reset_network_header(skb);
1757	iph = ip_hdr(skb);
1758
1759	iph->version	=	4;
1760	iph->tos	=	old_iph->tos;
1761	iph->ttl	=	old_iph->ttl;
1762	iph->frag_off	=	0;
1763	iph->daddr	=	daddr;
1764	iph->saddr	=	saddr;
1765	iph->protocol	=	IPPROTO_IPIP;
1766	iph->ihl	=	5;
1767	iph->tot_len	=	htons(skb->len);
1768	ip_select_ident(net, skb, NULL);
1769	ip_send_check(iph);
1770
1771	memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
1772	nf_reset_ct(skb);
1773}
1774
1775static inline int ipmr_forward_finish(struct net *net, struct sock *sk,
1776				      struct sk_buff *skb)
1777{
1778	struct ip_options *opt = &(IPCB(skb)->opt);
1779
1780	IP_INC_STATS(net, IPSTATS_MIB_OUTFORWDATAGRAMS);
1781	IP_ADD_STATS(net, IPSTATS_MIB_OUTOCTETS, skb->len);
1782
1783	if (unlikely(opt->optlen))
1784		ip_forward_options(skb);
1785
1786	return dst_output(net, sk, skb);
1787}
1788
1789#ifdef CONFIG_NET_SWITCHDEV
1790static bool ipmr_forward_offloaded(struct sk_buff *skb, struct mr_table *mrt,
1791				   int in_vifi, int out_vifi)
1792{
1793	struct vif_device *out_vif = &mrt->vif_table[out_vifi];
1794	struct vif_device *in_vif = &mrt->vif_table[in_vifi];
1795
1796	if (!skb->offload_l3_fwd_mark)
1797		return false;
1798	if (!out_vif->dev_parent_id.id_len || !in_vif->dev_parent_id.id_len)
1799		return false;
1800	return netdev_phys_item_id_same(&out_vif->dev_parent_id,
1801					&in_vif->dev_parent_id);
1802}
1803#else
1804static bool ipmr_forward_offloaded(struct sk_buff *skb, struct mr_table *mrt,
1805				   int in_vifi, int out_vifi)
1806{
1807	return false;
1808}
1809#endif
1810
1811/* Processing handlers for ipmr_forward */
1812
1813static void ipmr_queue_xmit(struct net *net, struct mr_table *mrt,
1814			    int in_vifi, struct sk_buff *skb, int vifi)
1815{
1816	const struct iphdr *iph = ip_hdr(skb);
1817	struct vif_device *vif = &mrt->vif_table[vifi];
1818	struct net_device *dev;
1819	struct rtable *rt;
1820	struct flowi4 fl4;
1821	int    encap = 0;
1822
1823	if (!vif->dev)
1824		goto out_free;
1825
1826	if (vif->flags & VIFF_REGISTER) {
1827		vif->pkt_out++;
1828		vif->bytes_out += skb->len;
1829		vif->dev->stats.tx_bytes += skb->len;
1830		vif->dev->stats.tx_packets++;
1831		ipmr_cache_report(mrt, skb, vifi, IGMPMSG_WHOLEPKT);
1832		goto out_free;
1833	}
1834
1835	if (ipmr_forward_offloaded(skb, mrt, in_vifi, vifi))
1836		goto out_free;
1837
1838	if (vif->flags & VIFF_TUNNEL) {
1839		rt = ip_route_output_ports(net, &fl4, NULL,
1840					   vif->remote, vif->local,
1841					   0, 0,
1842					   IPPROTO_IPIP,
1843					   RT_TOS(iph->tos), vif->link);
1844		if (IS_ERR(rt))
1845			goto out_free;
1846		encap = sizeof(struct iphdr);
1847	} else {
1848		rt = ip_route_output_ports(net, &fl4, NULL, iph->daddr, 0,
1849					   0, 0,
1850					   IPPROTO_IPIP,
1851					   RT_TOS(iph->tos), vif->link);
1852		if (IS_ERR(rt))
1853			goto out_free;
1854	}
1855
1856	dev = rt->dst.dev;
1857
1858	if (skb->len+encap > dst_mtu(&rt->dst) && (ntohs(iph->frag_off) & IP_DF)) {
1859		/* Do not fragment multicasts. Alas, IPv4 does not
1860		 * allow to send ICMP, so that packets will disappear
1861		 * to blackhole.
1862		 */
1863		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
1864		ip_rt_put(rt);
1865		goto out_free;
1866	}
1867
1868	encap += LL_RESERVED_SPACE(dev) + rt->dst.header_len;
1869
1870	if (skb_cow(skb, encap)) {
1871		ip_rt_put(rt);
1872		goto out_free;
1873	}
1874
1875	vif->pkt_out++;
1876	vif->bytes_out += skb->len;
1877
1878	skb_dst_drop(skb);
1879	skb_dst_set(skb, &rt->dst);
1880	ip_decrease_ttl(ip_hdr(skb));
1881
1882	/* FIXME: forward and output firewalls used to be called here.
1883	 * What do we do with netfilter? -- RR
1884	 */
1885	if (vif->flags & VIFF_TUNNEL) {
1886		ip_encap(net, skb, vif->local, vif->remote);
1887		/* FIXME: extra output firewall step used to be here. --RR */
1888		vif->dev->stats.tx_packets++;
1889		vif->dev->stats.tx_bytes += skb->len;
1890	}
1891
1892	IPCB(skb)->flags |= IPSKB_FORWARDED;
1893
1894	/* RFC1584 teaches, that DVMRP/PIM router must deliver packets locally
1895	 * not only before forwarding, but after forwarding on all output
1896	 * interfaces. It is clear, if mrouter runs a multicasting
1897	 * program, it should receive packets not depending to what interface
1898	 * program is joined.
1899	 * If we will not make it, the program will have to join on all
1900	 * interfaces. On the other hand, multihoming host (or router, but
1901	 * not mrouter) cannot join to more than one interface - it will
1902	 * result in receiving multiple packets.
1903	 */
1904	NF_HOOK(NFPROTO_IPV4, NF_INET_FORWARD,
1905		net, NULL, skb, skb->dev, dev,
1906		ipmr_forward_finish);
1907	return;
1908
1909out_free:
1910	kfree_skb(skb);
1911}
1912
1913static int ipmr_find_vif(struct mr_table *mrt, struct net_device *dev)
1914{
1915	int ct;
1916
1917	for (ct = mrt->maxvif-1; ct >= 0; ct--) {
1918		if (mrt->vif_table[ct].dev == dev)
1919			break;
1920	}
1921	return ct;
1922}
1923
1924/* "local" means that we should preserve one skb (for local delivery) */
1925static void ip_mr_forward(struct net *net, struct mr_table *mrt,
1926			  struct net_device *dev, struct sk_buff *skb,
1927			  struct mfc_cache *c, int local)
1928{
1929	int true_vifi = ipmr_find_vif(mrt, dev);
1930	int psend = -1;
1931	int vif, ct;
 
1932
1933	vif = c->_c.mfc_parent;
1934	c->_c.mfc_un.res.pkt++;
1935	c->_c.mfc_un.res.bytes += skb->len;
1936	c->_c.mfc_un.res.lastuse = jiffies;
1937
1938	if (c->mfc_origin == htonl(INADDR_ANY) && true_vifi >= 0) {
1939		struct mfc_cache *cache_proxy;
1940
1941		/* For an (*,G) entry, we only check that the incoming
1942		 * interface is part of the static tree.
1943		 */
1944		cache_proxy = mr_mfc_find_any_parent(mrt, vif);
1945		if (cache_proxy &&
1946		    cache_proxy->_c.mfc_un.res.ttls[true_vifi] < 255)
1947			goto forward;
1948	}
1949
1950	/* Wrong interface: drop packet and (maybe) send PIM assert. */
1951	if (mrt->vif_table[vif].dev != dev) {
1952		if (rt_is_output_route(skb_rtable(skb))) {
1953			/* It is our own packet, looped back.
1954			 * Very complicated situation...
1955			 *
1956			 * The best workaround until routing daemons will be
1957			 * fixed is not to redistribute packet, if it was
1958			 * send through wrong interface. It means, that
1959			 * multicast applications WILL NOT work for
1960			 * (S,G), which have default multicast route pointing
1961			 * to wrong oif. In any case, it is not a good
1962			 * idea to use multicasting applications on router.
1963			 */
1964			goto dont_forward;
1965		}
1966
1967		c->_c.mfc_un.res.wrong_if++;
1968
1969		if (true_vifi >= 0 && mrt->mroute_do_assert &&
1970		    /* pimsm uses asserts, when switching from RPT to SPT,
1971		     * so that we cannot check that packet arrived on an oif.
1972		     * It is bad, but otherwise we would need to move pretty
1973		     * large chunk of pimd to kernel. Ough... --ANK
1974		     */
1975		    (mrt->mroute_do_pim ||
1976		     c->_c.mfc_un.res.ttls[true_vifi] < 255) &&
1977		    time_after(jiffies,
1978			       c->_c.mfc_un.res.last_assert +
1979			       MFC_ASSERT_THRESH)) {
1980			c->_c.mfc_un.res.last_assert = jiffies;
1981			ipmr_cache_report(mrt, skb, true_vifi, IGMPMSG_WRONGVIF);
1982			if (mrt->mroute_do_wrvifwhole)
1983				ipmr_cache_report(mrt, skb, true_vifi,
1984						  IGMPMSG_WRVIFWHOLE);
1985		}
1986		goto dont_forward;
1987	}
1988
1989forward:
1990	mrt->vif_table[vif].pkt_in++;
1991	mrt->vif_table[vif].bytes_in += skb->len;
1992
1993	/* Forward the frame */
1994	if (c->mfc_origin == htonl(INADDR_ANY) &&
1995	    c->mfc_mcastgrp == htonl(INADDR_ANY)) {
1996		if (true_vifi >= 0 &&
1997		    true_vifi != c->_c.mfc_parent &&
1998		    ip_hdr(skb)->ttl >
1999				c->_c.mfc_un.res.ttls[c->_c.mfc_parent]) {
2000			/* It's an (*,*) entry and the packet is not coming from
2001			 * the upstream: forward the packet to the upstream
2002			 * only.
2003			 */
2004			psend = c->_c.mfc_parent;
2005			goto last_forward;
2006		}
2007		goto dont_forward;
2008	}
2009	for (ct = c->_c.mfc_un.res.maxvif - 1;
2010	     ct >= c->_c.mfc_un.res.minvif; ct--) {
2011		/* For (*,G) entry, don't forward to the incoming interface */
2012		if ((c->mfc_origin != htonl(INADDR_ANY) ||
2013		     ct != true_vifi) &&
2014		    ip_hdr(skb)->ttl > c->_c.mfc_un.res.ttls[ct]) {
2015			if (psend != -1) {
2016				struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
2017
2018				if (skb2)
2019					ipmr_queue_xmit(net, mrt, true_vifi,
2020							skb2, psend);
2021			}
2022			psend = ct;
2023		}
2024	}
2025last_forward:
2026	if (psend != -1) {
2027		if (local) {
2028			struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
2029
2030			if (skb2)
2031				ipmr_queue_xmit(net, mrt, true_vifi, skb2,
2032						psend);
2033		} else {
2034			ipmr_queue_xmit(net, mrt, true_vifi, skb, psend);
2035			return;
2036		}
2037	}
2038
2039dont_forward:
2040	if (!local)
2041		kfree_skb(skb);
2042}
2043
2044static struct mr_table *ipmr_rt_fib_lookup(struct net *net, struct sk_buff *skb)
2045{
2046	struct rtable *rt = skb_rtable(skb);
2047	struct iphdr *iph = ip_hdr(skb);
2048	struct flowi4 fl4 = {
2049		.daddr = iph->daddr,
2050		.saddr = iph->saddr,
2051		.flowi4_tos = RT_TOS(iph->tos),
2052		.flowi4_oif = (rt_is_output_route(rt) ?
2053			       skb->dev->ifindex : 0),
2054		.flowi4_iif = (rt_is_output_route(rt) ?
2055			       LOOPBACK_IFINDEX :
2056			       skb->dev->ifindex),
2057		.flowi4_mark = skb->mark,
2058	};
2059	struct mr_table *mrt;
2060	int err;
2061
2062	err = ipmr_fib_lookup(net, &fl4, &mrt);
2063	if (err)
2064		return ERR_PTR(err);
2065	return mrt;
2066}
2067
2068/* Multicast packets for forwarding arrive here
2069 * Called with rcu_read_lock();
2070 */
2071int ip_mr_input(struct sk_buff *skb)
2072{
2073	struct mfc_cache *cache;
2074	struct net *net = dev_net(skb->dev);
2075	int local = skb_rtable(skb)->rt_flags & RTCF_LOCAL;
2076	struct mr_table *mrt;
2077	struct net_device *dev;
2078
2079	/* skb->dev passed in is the loX master dev for vrfs.
2080	 * As there are no vifs associated with loopback devices,
2081	 * get the proper interface that does have a vif associated with it.
2082	 */
2083	dev = skb->dev;
2084	if (netif_is_l3_master(skb->dev)) {
2085		dev = dev_get_by_index_rcu(net, IPCB(skb)->iif);
2086		if (!dev) {
2087			kfree_skb(skb);
2088			return -ENODEV;
2089		}
2090	}
2091
2092	/* Packet is looped back after forward, it should not be
2093	 * forwarded second time, but still can be delivered locally.
2094	 */
2095	if (IPCB(skb)->flags & IPSKB_FORWARDED)
2096		goto dont_forward;
2097
2098	mrt = ipmr_rt_fib_lookup(net, skb);
2099	if (IS_ERR(mrt)) {
2100		kfree_skb(skb);
2101		return PTR_ERR(mrt);
2102	}
2103	if (!local) {
2104		if (IPCB(skb)->opt.router_alert) {
2105			if (ip_call_ra_chain(skb))
2106				return 0;
2107		} else if (ip_hdr(skb)->protocol == IPPROTO_IGMP) {
2108			/* IGMPv1 (and broken IGMPv2 implementations sort of
2109			 * Cisco IOS <= 11.2(8)) do not put router alert
2110			 * option to IGMP packets destined to routable
2111			 * groups. It is very bad, because it means
2112			 * that we can forward NO IGMP messages.
2113			 */
2114			struct sock *mroute_sk;
2115
2116			mroute_sk = rcu_dereference(mrt->mroute_sk);
2117			if (mroute_sk) {
2118				nf_reset_ct(skb);
2119				raw_rcv(mroute_sk, skb);
2120				return 0;
2121			}
2122		}
2123	}
2124
2125	/* already under rcu_read_lock() */
2126	cache = ipmr_cache_find(mrt, ip_hdr(skb)->saddr, ip_hdr(skb)->daddr);
2127	if (!cache) {
2128		int vif = ipmr_find_vif(mrt, dev);
2129
2130		if (vif >= 0)
2131			cache = ipmr_cache_find_any(mrt, ip_hdr(skb)->daddr,
2132						    vif);
2133	}
2134
2135	/* No usable cache entry */
2136	if (!cache) {
2137		int vif;
2138
2139		if (local) {
2140			struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
2141			ip_local_deliver(skb);
2142			if (!skb2)
2143				return -ENOBUFS;
2144			skb = skb2;
2145		}
2146
2147		read_lock(&mrt_lock);
2148		vif = ipmr_find_vif(mrt, dev);
2149		if (vif >= 0) {
2150			int err2 = ipmr_cache_unresolved(mrt, vif, skb, dev);
2151			read_unlock(&mrt_lock);
2152
2153			return err2;
2154		}
2155		read_unlock(&mrt_lock);
2156		kfree_skb(skb);
2157		return -ENODEV;
2158	}
2159
2160	read_lock(&mrt_lock);
2161	ip_mr_forward(net, mrt, dev, skb, cache, local);
2162	read_unlock(&mrt_lock);
2163
2164	if (local)
2165		return ip_local_deliver(skb);
2166
2167	return 0;
2168
2169dont_forward:
2170	if (local)
2171		return ip_local_deliver(skb);
2172	kfree_skb(skb);
2173	return 0;
2174}
2175
2176#ifdef CONFIG_IP_PIMSM_V1
2177/* Handle IGMP messages of PIMv1 */
2178int pim_rcv_v1(struct sk_buff *skb)
2179{
2180	struct igmphdr *pim;
2181	struct net *net = dev_net(skb->dev);
2182	struct mr_table *mrt;
2183
2184	if (!pskb_may_pull(skb, sizeof(*pim) + sizeof(struct iphdr)))
2185		goto drop;
2186
2187	pim = igmp_hdr(skb);
2188
2189	mrt = ipmr_rt_fib_lookup(net, skb);
2190	if (IS_ERR(mrt))
2191		goto drop;
2192	if (!mrt->mroute_do_pim ||
2193	    pim->group != PIM_V1_VERSION || pim->code != PIM_V1_REGISTER)
2194		goto drop;
2195
2196	if (__pim_rcv(mrt, skb, sizeof(*pim))) {
2197drop:
2198		kfree_skb(skb);
2199	}
2200	return 0;
2201}
2202#endif
2203
2204#ifdef CONFIG_IP_PIMSM_V2
2205static int pim_rcv(struct sk_buff *skb)
2206{
2207	struct pimreghdr *pim;
2208	struct net *net = dev_net(skb->dev);
2209	struct mr_table *mrt;
2210
2211	if (!pskb_may_pull(skb, sizeof(*pim) + sizeof(struct iphdr)))
2212		goto drop;
2213
2214	pim = (struct pimreghdr *)skb_transport_header(skb);
2215	if (pim->type != ((PIM_VERSION << 4) | (PIM_TYPE_REGISTER)) ||
2216	    (pim->flags & PIM_NULL_REGISTER) ||
2217	    (ip_compute_csum((void *)pim, sizeof(*pim)) != 0 &&
2218	     csum_fold(skb_checksum(skb, 0, skb->len, 0))))
2219		goto drop;
2220
2221	mrt = ipmr_rt_fib_lookup(net, skb);
2222	if (IS_ERR(mrt))
2223		goto drop;
2224	if (__pim_rcv(mrt, skb, sizeof(*pim))) {
2225drop:
2226		kfree_skb(skb);
2227	}
2228	return 0;
2229}
2230#endif
2231
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2232int ipmr_get_route(struct net *net, struct sk_buff *skb,
2233		   __be32 saddr, __be32 daddr,
2234		   struct rtmsg *rtm, u32 portid)
2235{
2236	struct mfc_cache *cache;
2237	struct mr_table *mrt;
2238	int err;
2239
2240	mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2241	if (!mrt)
2242		return -ENOENT;
2243
2244	rcu_read_lock();
2245	cache = ipmr_cache_find(mrt, saddr, daddr);
2246	if (!cache && skb->dev) {
2247		int vif = ipmr_find_vif(mrt, skb->dev);
2248
2249		if (vif >= 0)
2250			cache = ipmr_cache_find_any(mrt, daddr, vif);
2251	}
2252	if (!cache) {
2253		struct sk_buff *skb2;
2254		struct iphdr *iph;
2255		struct net_device *dev;
2256		int vif = -1;
2257
 
 
 
 
 
2258		dev = skb->dev;
2259		read_lock(&mrt_lock);
2260		if (dev)
2261			vif = ipmr_find_vif(mrt, dev);
2262		if (vif < 0) {
2263			read_unlock(&mrt_lock);
2264			rcu_read_unlock();
2265			return -ENODEV;
2266		}
2267
2268		skb2 = skb_realloc_headroom(skb, sizeof(struct iphdr));
2269		if (!skb2) {
2270			read_unlock(&mrt_lock);
2271			rcu_read_unlock();
2272			return -ENOMEM;
2273		}
2274
2275		NETLINK_CB(skb2).portid = portid;
2276		skb_push(skb2, sizeof(struct iphdr));
2277		skb_reset_network_header(skb2);
2278		iph = ip_hdr(skb2);
2279		iph->ihl = sizeof(struct iphdr) >> 2;
2280		iph->saddr = saddr;
2281		iph->daddr = daddr;
2282		iph->version = 0;
2283		err = ipmr_cache_unresolved(mrt, vif, skb2, dev);
2284		read_unlock(&mrt_lock);
2285		rcu_read_unlock();
2286		return err;
2287	}
2288
2289	read_lock(&mrt_lock);
2290	err = mr_fill_mroute(mrt, skb, &cache->_c, rtm);
2291	read_unlock(&mrt_lock);
2292	rcu_read_unlock();
2293	return err;
2294}
2295
2296static int ipmr_fill_mroute(struct mr_table *mrt, struct sk_buff *skb,
2297			    u32 portid, u32 seq, struct mfc_cache *c, int cmd,
2298			    int flags)
2299{
2300	struct nlmsghdr *nlh;
2301	struct rtmsg *rtm;
2302	int err;
2303
2304	nlh = nlmsg_put(skb, portid, seq, cmd, sizeof(*rtm), flags);
2305	if (!nlh)
2306		return -EMSGSIZE;
2307
2308	rtm = nlmsg_data(nlh);
2309	rtm->rtm_family   = RTNL_FAMILY_IPMR;
2310	rtm->rtm_dst_len  = 32;
2311	rtm->rtm_src_len  = 32;
2312	rtm->rtm_tos      = 0;
2313	rtm->rtm_table    = mrt->id;
2314	if (nla_put_u32(skb, RTA_TABLE, mrt->id))
2315		goto nla_put_failure;
2316	rtm->rtm_type     = RTN_MULTICAST;
2317	rtm->rtm_scope    = RT_SCOPE_UNIVERSE;
2318	if (c->_c.mfc_flags & MFC_STATIC)
2319		rtm->rtm_protocol = RTPROT_STATIC;
2320	else
2321		rtm->rtm_protocol = RTPROT_MROUTED;
2322	rtm->rtm_flags    = 0;
2323
2324	if (nla_put_in_addr(skb, RTA_SRC, c->mfc_origin) ||
2325	    nla_put_in_addr(skb, RTA_DST, c->mfc_mcastgrp))
2326		goto nla_put_failure;
2327	err = mr_fill_mroute(mrt, skb, &c->_c, rtm);
2328	/* do not break the dump if cache is unresolved */
2329	if (err < 0 && err != -ENOENT)
2330		goto nla_put_failure;
2331
2332	nlmsg_end(skb, nlh);
2333	return 0;
2334
2335nla_put_failure:
2336	nlmsg_cancel(skb, nlh);
2337	return -EMSGSIZE;
2338}
2339
2340static int _ipmr_fill_mroute(struct mr_table *mrt, struct sk_buff *skb,
2341			     u32 portid, u32 seq, struct mr_mfc *c, int cmd,
2342			     int flags)
2343{
2344	return ipmr_fill_mroute(mrt, skb, portid, seq, (struct mfc_cache *)c,
2345				cmd, flags);
2346}
2347
2348static size_t mroute_msgsize(bool unresolved, int maxvif)
2349{
2350	size_t len =
2351		NLMSG_ALIGN(sizeof(struct rtmsg))
2352		+ nla_total_size(4)	/* RTA_TABLE */
2353		+ nla_total_size(4)	/* RTA_SRC */
2354		+ nla_total_size(4)	/* RTA_DST */
2355		;
2356
2357	if (!unresolved)
2358		len = len
2359		      + nla_total_size(4)	/* RTA_IIF */
2360		      + nla_total_size(0)	/* RTA_MULTIPATH */
2361		      + maxvif * NLA_ALIGN(sizeof(struct rtnexthop))
2362						/* RTA_MFC_STATS */
2363		      + nla_total_size_64bit(sizeof(struct rta_mfc_stats))
2364		;
2365
2366	return len;
2367}
2368
2369static void mroute_netlink_event(struct mr_table *mrt, struct mfc_cache *mfc,
2370				 int cmd)
2371{
2372	struct net *net = read_pnet(&mrt->net);
2373	struct sk_buff *skb;
2374	int err = -ENOBUFS;
2375
2376	skb = nlmsg_new(mroute_msgsize(mfc->_c.mfc_parent >= MAXVIFS,
2377				       mrt->maxvif),
2378			GFP_ATOMIC);
2379	if (!skb)
2380		goto errout;
2381
2382	err = ipmr_fill_mroute(mrt, skb, 0, 0, mfc, cmd, 0);
2383	if (err < 0)
2384		goto errout;
2385
2386	rtnl_notify(skb, net, 0, RTNLGRP_IPV4_MROUTE, NULL, GFP_ATOMIC);
2387	return;
2388
2389errout:
2390	kfree_skb(skb);
2391	if (err < 0)
2392		rtnl_set_sk_err(net, RTNLGRP_IPV4_MROUTE, err);
2393}
2394
2395static size_t igmpmsg_netlink_msgsize(size_t payloadlen)
2396{
2397	size_t len =
2398		NLMSG_ALIGN(sizeof(struct rtgenmsg))
2399		+ nla_total_size(1)	/* IPMRA_CREPORT_MSGTYPE */
2400		+ nla_total_size(4)	/* IPMRA_CREPORT_VIF_ID */
2401		+ nla_total_size(4)	/* IPMRA_CREPORT_SRC_ADDR */
2402		+ nla_total_size(4)	/* IPMRA_CREPORT_DST_ADDR */
2403		+ nla_total_size(4)	/* IPMRA_CREPORT_TABLE */
2404					/* IPMRA_CREPORT_PKT */
2405		+ nla_total_size(payloadlen)
2406		;
2407
2408	return len;
2409}
2410
2411static void igmpmsg_netlink_event(struct mr_table *mrt, struct sk_buff *pkt)
2412{
2413	struct net *net = read_pnet(&mrt->net);
2414	struct nlmsghdr *nlh;
2415	struct rtgenmsg *rtgenm;
2416	struct igmpmsg *msg;
2417	struct sk_buff *skb;
2418	struct nlattr *nla;
2419	int payloadlen;
2420
2421	payloadlen = pkt->len - sizeof(struct igmpmsg);
2422	msg = (struct igmpmsg *)skb_network_header(pkt);
2423
2424	skb = nlmsg_new(igmpmsg_netlink_msgsize(payloadlen), GFP_ATOMIC);
2425	if (!skb)
2426		goto errout;
2427
2428	nlh = nlmsg_put(skb, 0, 0, RTM_NEWCACHEREPORT,
2429			sizeof(struct rtgenmsg), 0);
2430	if (!nlh)
2431		goto errout;
2432	rtgenm = nlmsg_data(nlh);
2433	rtgenm->rtgen_family = RTNL_FAMILY_IPMR;
2434	if (nla_put_u8(skb, IPMRA_CREPORT_MSGTYPE, msg->im_msgtype) ||
2435	    nla_put_u32(skb, IPMRA_CREPORT_VIF_ID, msg->im_vif | (msg->im_vif_hi << 8)) ||
2436	    nla_put_in_addr(skb, IPMRA_CREPORT_SRC_ADDR,
2437			    msg->im_src.s_addr) ||
2438	    nla_put_in_addr(skb, IPMRA_CREPORT_DST_ADDR,
2439			    msg->im_dst.s_addr) ||
2440	    nla_put_u32(skb, IPMRA_CREPORT_TABLE, mrt->id))
2441		goto nla_put_failure;
2442
2443	nla = nla_reserve(skb, IPMRA_CREPORT_PKT, payloadlen);
2444	if (!nla || skb_copy_bits(pkt, sizeof(struct igmpmsg),
2445				  nla_data(nla), payloadlen))
2446		goto nla_put_failure;
2447
2448	nlmsg_end(skb, nlh);
2449
2450	rtnl_notify(skb, net, 0, RTNLGRP_IPV4_MROUTE_R, NULL, GFP_ATOMIC);
2451	return;
2452
2453nla_put_failure:
2454	nlmsg_cancel(skb, nlh);
2455errout:
2456	kfree_skb(skb);
2457	rtnl_set_sk_err(net, RTNLGRP_IPV4_MROUTE_R, -ENOBUFS);
2458}
2459
2460static int ipmr_rtm_valid_getroute_req(struct sk_buff *skb,
2461				       const struct nlmsghdr *nlh,
2462				       struct nlattr **tb,
2463				       struct netlink_ext_ack *extack)
2464{
2465	struct rtmsg *rtm;
2466	int i, err;
2467
2468	if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*rtm))) {
2469		NL_SET_ERR_MSG(extack, "ipv4: Invalid header for multicast route get request");
2470		return -EINVAL;
2471	}
2472
2473	if (!netlink_strict_get_check(skb))
2474		return nlmsg_parse_deprecated(nlh, sizeof(*rtm), tb, RTA_MAX,
2475					      rtm_ipv4_policy, extack);
2476
2477	rtm = nlmsg_data(nlh);
2478	if ((rtm->rtm_src_len && rtm->rtm_src_len != 32) ||
2479	    (rtm->rtm_dst_len && rtm->rtm_dst_len != 32) ||
2480	    rtm->rtm_tos || rtm->rtm_table || rtm->rtm_protocol ||
2481	    rtm->rtm_scope || rtm->rtm_type || rtm->rtm_flags) {
2482		NL_SET_ERR_MSG(extack, "ipv4: Invalid values in header for multicast route get request");
2483		return -EINVAL;
2484	}
2485
2486	err = nlmsg_parse_deprecated_strict(nlh, sizeof(*rtm), tb, RTA_MAX,
2487					    rtm_ipv4_policy, extack);
2488	if (err)
2489		return err;
2490
2491	if ((tb[RTA_SRC] && !rtm->rtm_src_len) ||
2492	    (tb[RTA_DST] && !rtm->rtm_dst_len)) {
2493		NL_SET_ERR_MSG(extack, "ipv4: rtm_src_len and rtm_dst_len must be 32 for IPv4");
2494		return -EINVAL;
2495	}
2496
2497	for (i = 0; i <= RTA_MAX; i++) {
2498		if (!tb[i])
2499			continue;
2500
2501		switch (i) {
2502		case RTA_SRC:
2503		case RTA_DST:
2504		case RTA_TABLE:
2505			break;
2506		default:
2507			NL_SET_ERR_MSG(extack, "ipv4: Unsupported attribute in multicast route get request");
2508			return -EINVAL;
2509		}
2510	}
2511
2512	return 0;
2513}
2514
2515static int ipmr_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr *nlh,
2516			     struct netlink_ext_ack *extack)
2517{
2518	struct net *net = sock_net(in_skb->sk);
2519	struct nlattr *tb[RTA_MAX + 1];
2520	struct sk_buff *skb = NULL;
2521	struct mfc_cache *cache;
2522	struct mr_table *mrt;
2523	__be32 src, grp;
2524	u32 tableid;
2525	int err;
2526
2527	err = ipmr_rtm_valid_getroute_req(in_skb, nlh, tb, extack);
2528	if (err < 0)
2529		goto errout;
2530
2531	src = tb[RTA_SRC] ? nla_get_in_addr(tb[RTA_SRC]) : 0;
2532	grp = tb[RTA_DST] ? nla_get_in_addr(tb[RTA_DST]) : 0;
2533	tableid = tb[RTA_TABLE] ? nla_get_u32(tb[RTA_TABLE]) : 0;
2534
2535	mrt = ipmr_get_table(net, tableid ? tableid : RT_TABLE_DEFAULT);
2536	if (!mrt) {
2537		err = -ENOENT;
2538		goto errout_free;
2539	}
2540
2541	/* entries are added/deleted only under RTNL */
2542	rcu_read_lock();
2543	cache = ipmr_cache_find(mrt, src, grp);
2544	rcu_read_unlock();
2545	if (!cache) {
2546		err = -ENOENT;
2547		goto errout_free;
2548	}
2549
2550	skb = nlmsg_new(mroute_msgsize(false, mrt->maxvif), GFP_KERNEL);
2551	if (!skb) {
2552		err = -ENOBUFS;
2553		goto errout_free;
2554	}
2555
2556	err = ipmr_fill_mroute(mrt, skb, NETLINK_CB(in_skb).portid,
2557			       nlh->nlmsg_seq, cache,
2558			       RTM_NEWROUTE, 0);
2559	if (err < 0)
2560		goto errout_free;
2561
2562	err = rtnl_unicast(skb, net, NETLINK_CB(in_skb).portid);
2563
2564errout:
2565	return err;
2566
2567errout_free:
2568	kfree_skb(skb);
2569	goto errout;
2570}
2571
2572static int ipmr_rtm_dumproute(struct sk_buff *skb, struct netlink_callback *cb)
2573{
2574	struct fib_dump_filter filter = {};
2575	int err;
2576
2577	if (cb->strict_check) {
2578		err = ip_valid_fib_dump_req(sock_net(skb->sk), cb->nlh,
2579					    &filter, cb);
2580		if (err < 0)
2581			return err;
2582	}
2583
2584	if (filter.table_id) {
2585		struct mr_table *mrt;
2586
2587		mrt = ipmr_get_table(sock_net(skb->sk), filter.table_id);
2588		if (!mrt) {
2589			if (rtnl_msg_family(cb->nlh) != RTNL_FAMILY_IPMR)
2590				return skb->len;
2591
2592			NL_SET_ERR_MSG(cb->extack, "ipv4: MR table does not exist");
2593			return -ENOENT;
2594		}
2595		err = mr_table_dump(mrt, skb, cb, _ipmr_fill_mroute,
2596				    &mfc_unres_lock, &filter);
2597		return skb->len ? : err;
 
 
2598	}
 
 
 
 
 
 
2599
2600	return mr_rtm_dumproute(skb, cb, ipmr_mr_table_iter,
2601				_ipmr_fill_mroute, &mfc_unres_lock, &filter);
2602}
2603
2604static const struct nla_policy rtm_ipmr_policy[RTA_MAX + 1] = {
2605	[RTA_SRC]	= { .type = NLA_U32 },
2606	[RTA_DST]	= { .type = NLA_U32 },
2607	[RTA_IIF]	= { .type = NLA_U32 },
2608	[RTA_TABLE]	= { .type = NLA_U32 },
2609	[RTA_MULTIPATH]	= { .len = sizeof(struct rtnexthop) },
2610};
2611
2612static bool ipmr_rtm_validate_proto(unsigned char rtm_protocol)
2613{
2614	switch (rtm_protocol) {
2615	case RTPROT_STATIC:
2616	case RTPROT_MROUTED:
2617		return true;
2618	}
2619	return false;
2620}
2621
2622static int ipmr_nla_get_ttls(const struct nlattr *nla, struct mfcctl *mfcc)
2623{
2624	struct rtnexthop *rtnh = nla_data(nla);
2625	int remaining = nla_len(nla), vifi = 0;
2626
2627	while (rtnh_ok(rtnh, remaining)) {
2628		mfcc->mfcc_ttls[vifi] = rtnh->rtnh_hops;
2629		if (++vifi == MAXVIFS)
2630			break;
2631		rtnh = rtnh_next(rtnh, &remaining);
2632	}
2633
2634	return remaining > 0 ? -EINVAL : vifi;
2635}
2636
2637/* returns < 0 on error, 0 for ADD_MFC and 1 for ADD_MFC_PROXY */
2638static int rtm_to_ipmr_mfcc(struct net *net, struct nlmsghdr *nlh,
2639			    struct mfcctl *mfcc, int *mrtsock,
2640			    struct mr_table **mrtret,
2641			    struct netlink_ext_ack *extack)
2642{
2643	struct net_device *dev = NULL;
2644	u32 tblid = RT_TABLE_DEFAULT;
2645	struct mr_table *mrt;
2646	struct nlattr *attr;
2647	struct rtmsg *rtm;
2648	int ret, rem;
2649
2650	ret = nlmsg_validate_deprecated(nlh, sizeof(*rtm), RTA_MAX,
2651					rtm_ipmr_policy, extack);
2652	if (ret < 0)
2653		goto out;
2654	rtm = nlmsg_data(nlh);
2655
2656	ret = -EINVAL;
2657	if (rtm->rtm_family != RTNL_FAMILY_IPMR || rtm->rtm_dst_len != 32 ||
2658	    rtm->rtm_type != RTN_MULTICAST ||
2659	    rtm->rtm_scope != RT_SCOPE_UNIVERSE ||
2660	    !ipmr_rtm_validate_proto(rtm->rtm_protocol))
2661		goto out;
2662
2663	memset(mfcc, 0, sizeof(*mfcc));
2664	mfcc->mfcc_parent = -1;
2665	ret = 0;
2666	nlmsg_for_each_attr(attr, nlh, sizeof(struct rtmsg), rem) {
2667		switch (nla_type(attr)) {
2668		case RTA_SRC:
2669			mfcc->mfcc_origin.s_addr = nla_get_be32(attr);
2670			break;
2671		case RTA_DST:
2672			mfcc->mfcc_mcastgrp.s_addr = nla_get_be32(attr);
2673			break;
2674		case RTA_IIF:
2675			dev = __dev_get_by_index(net, nla_get_u32(attr));
2676			if (!dev) {
2677				ret = -ENODEV;
2678				goto out;
2679			}
2680			break;
2681		case RTA_MULTIPATH:
2682			if (ipmr_nla_get_ttls(attr, mfcc) < 0) {
2683				ret = -EINVAL;
2684				goto out;
2685			}
2686			break;
2687		case RTA_PREFSRC:
2688			ret = 1;
2689			break;
2690		case RTA_TABLE:
2691			tblid = nla_get_u32(attr);
2692			break;
2693		}
2694	}
2695	mrt = ipmr_get_table(net, tblid);
2696	if (!mrt) {
2697		ret = -ENOENT;
2698		goto out;
2699	}
2700	*mrtret = mrt;
2701	*mrtsock = rtm->rtm_protocol == RTPROT_MROUTED ? 1 : 0;
2702	if (dev)
2703		mfcc->mfcc_parent = ipmr_find_vif(mrt, dev);
2704
2705out:
2706	return ret;
2707}
2708
2709/* takes care of both newroute and delroute */
2710static int ipmr_rtm_route(struct sk_buff *skb, struct nlmsghdr *nlh,
2711			  struct netlink_ext_ack *extack)
2712{
2713	struct net *net = sock_net(skb->sk);
2714	int ret, mrtsock, parent;
2715	struct mr_table *tbl;
2716	struct mfcctl mfcc;
2717
2718	mrtsock = 0;
2719	tbl = NULL;
2720	ret = rtm_to_ipmr_mfcc(net, nlh, &mfcc, &mrtsock, &tbl, extack);
2721	if (ret < 0)
2722		return ret;
2723
2724	parent = ret ? mfcc.mfcc_parent : -1;
2725	if (nlh->nlmsg_type == RTM_NEWROUTE)
2726		return ipmr_mfc_add(net, tbl, &mfcc, mrtsock, parent);
2727	else
2728		return ipmr_mfc_delete(tbl, &mfcc, parent);
2729}
2730
2731static bool ipmr_fill_table(struct mr_table *mrt, struct sk_buff *skb)
2732{
2733	u32 queue_len = atomic_read(&mrt->cache_resolve_queue_len);
2734
2735	if (nla_put_u32(skb, IPMRA_TABLE_ID, mrt->id) ||
2736	    nla_put_u32(skb, IPMRA_TABLE_CACHE_RES_QUEUE_LEN, queue_len) ||
2737	    nla_put_s32(skb, IPMRA_TABLE_MROUTE_REG_VIF_NUM,
2738			mrt->mroute_reg_vif_num) ||
2739	    nla_put_u8(skb, IPMRA_TABLE_MROUTE_DO_ASSERT,
2740		       mrt->mroute_do_assert) ||
2741	    nla_put_u8(skb, IPMRA_TABLE_MROUTE_DO_PIM, mrt->mroute_do_pim) ||
2742	    nla_put_u8(skb, IPMRA_TABLE_MROUTE_DO_WRVIFWHOLE,
2743		       mrt->mroute_do_wrvifwhole))
2744		return false;
2745
2746	return true;
2747}
2748
2749static bool ipmr_fill_vif(struct mr_table *mrt, u32 vifid, struct sk_buff *skb)
2750{
2751	struct nlattr *vif_nest;
2752	struct vif_device *vif;
2753
2754	/* if the VIF doesn't exist just continue */
2755	if (!VIF_EXISTS(mrt, vifid))
2756		return true;
2757
2758	vif = &mrt->vif_table[vifid];
2759	vif_nest = nla_nest_start_noflag(skb, IPMRA_VIF);
2760	if (!vif_nest)
2761		return false;
2762	if (nla_put_u32(skb, IPMRA_VIFA_IFINDEX, vif->dev->ifindex) ||
2763	    nla_put_u32(skb, IPMRA_VIFA_VIF_ID, vifid) ||
2764	    nla_put_u16(skb, IPMRA_VIFA_FLAGS, vif->flags) ||
2765	    nla_put_u64_64bit(skb, IPMRA_VIFA_BYTES_IN, vif->bytes_in,
2766			      IPMRA_VIFA_PAD) ||
2767	    nla_put_u64_64bit(skb, IPMRA_VIFA_BYTES_OUT, vif->bytes_out,
2768			      IPMRA_VIFA_PAD) ||
2769	    nla_put_u64_64bit(skb, IPMRA_VIFA_PACKETS_IN, vif->pkt_in,
2770			      IPMRA_VIFA_PAD) ||
2771	    nla_put_u64_64bit(skb, IPMRA_VIFA_PACKETS_OUT, vif->pkt_out,
2772			      IPMRA_VIFA_PAD) ||
2773	    nla_put_be32(skb, IPMRA_VIFA_LOCAL_ADDR, vif->local) ||
2774	    nla_put_be32(skb, IPMRA_VIFA_REMOTE_ADDR, vif->remote)) {
2775		nla_nest_cancel(skb, vif_nest);
2776		return false;
2777	}
2778	nla_nest_end(skb, vif_nest);
2779
2780	return true;
2781}
2782
2783static int ipmr_valid_dumplink(const struct nlmsghdr *nlh,
2784			       struct netlink_ext_ack *extack)
2785{
2786	struct ifinfomsg *ifm;
2787
2788	if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*ifm))) {
2789		NL_SET_ERR_MSG(extack, "ipv4: Invalid header for ipmr link dump");
2790		return -EINVAL;
2791	}
2792
2793	if (nlmsg_attrlen(nlh, sizeof(*ifm))) {
2794		NL_SET_ERR_MSG(extack, "Invalid data after header in ipmr link dump");
2795		return -EINVAL;
2796	}
2797
2798	ifm = nlmsg_data(nlh);
2799	if (ifm->__ifi_pad || ifm->ifi_type || ifm->ifi_flags ||
2800	    ifm->ifi_change || ifm->ifi_index) {
2801		NL_SET_ERR_MSG(extack, "Invalid values in header for ipmr link dump request");
2802		return -EINVAL;
2803	}
2804
2805	return 0;
2806}
2807
2808static int ipmr_rtm_dumplink(struct sk_buff *skb, struct netlink_callback *cb)
2809{
2810	struct net *net = sock_net(skb->sk);
2811	struct nlmsghdr *nlh = NULL;
2812	unsigned int t = 0, s_t;
2813	unsigned int e = 0, s_e;
2814	struct mr_table *mrt;
 
 
2815
2816	if (cb->strict_check) {
2817		int err = ipmr_valid_dumplink(cb->nlh, cb->extack);
2818
2819		if (err < 0)
2820			return err;
2821	}
2822
2823	s_t = cb->args[0];
2824	s_e = cb->args[1];
2825
2826	ipmr_for_each_table(mrt, net) {
2827		struct nlattr *vifs, *af;
2828		struct ifinfomsg *hdr;
2829		u32 i;
2830
2831		if (t < s_t)
2832			goto skip_table;
2833		nlh = nlmsg_put(skb, NETLINK_CB(cb->skb).portid,
2834				cb->nlh->nlmsg_seq, RTM_NEWLINK,
2835				sizeof(*hdr), NLM_F_MULTI);
2836		if (!nlh)
2837			break;
2838
2839		hdr = nlmsg_data(nlh);
2840		memset(hdr, 0, sizeof(*hdr));
2841		hdr->ifi_family = RTNL_FAMILY_IPMR;
2842
2843		af = nla_nest_start_noflag(skb, IFLA_AF_SPEC);
2844		if (!af) {
2845			nlmsg_cancel(skb, nlh);
2846			goto out;
2847		}
2848
2849		if (!ipmr_fill_table(mrt, skb)) {
2850			nlmsg_cancel(skb, nlh);
2851			goto out;
2852		}
2853
2854		vifs = nla_nest_start_noflag(skb, IPMRA_TABLE_VIFS);
2855		if (!vifs) {
2856			nla_nest_end(skb, af);
2857			nlmsg_end(skb, nlh);
2858			goto out;
2859		}
2860		for (i = 0; i < mrt->maxvif; i++) {
2861			if (e < s_e)
2862				goto skip_entry;
2863			if (!ipmr_fill_vif(mrt, i, skb)) {
2864				nla_nest_end(skb, vifs);
2865				nla_nest_end(skb, af);
2866				nlmsg_end(skb, nlh);
2867				goto out;
2868			}
2869skip_entry:
2870			e++;
2871		}
2872		s_e = 0;
2873		e = 0;
2874		nla_nest_end(skb, vifs);
2875		nla_nest_end(skb, af);
2876		nlmsg_end(skb, nlh);
2877skip_table:
2878		t++;
2879	}
2880
2881out:
2882	cb->args[1] = e;
2883	cb->args[0] = t;
2884
2885	return skb->len;
2886}
2887
2888#ifdef CONFIG_PROC_FS
2889/* The /proc interfaces to multicast routing :
2890 * /proc/net/ip_mr_cache & /proc/net/ip_mr_vif
2891 */
2892
2893static void *ipmr_vif_seq_start(struct seq_file *seq, loff_t *pos)
2894	__acquires(mrt_lock)
2895{
2896	struct mr_vif_iter *iter = seq->private;
2897	struct net *net = seq_file_net(seq);
2898	struct mr_table *mrt;
2899
2900	mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2901	if (!mrt)
2902		return ERR_PTR(-ENOENT);
2903
2904	iter->mrt = mrt;
2905
2906	read_lock(&mrt_lock);
2907	return mr_vif_seq_start(seq, pos);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2908}
2909
2910static void ipmr_vif_seq_stop(struct seq_file *seq, void *v)
2911	__releases(mrt_lock)
2912{
2913	read_unlock(&mrt_lock);
2914}
2915
2916static int ipmr_vif_seq_show(struct seq_file *seq, void *v)
2917{
2918	struct mr_vif_iter *iter = seq->private;
2919	struct mr_table *mrt = iter->mrt;
2920
2921	if (v == SEQ_START_TOKEN) {
2922		seq_puts(seq,
2923			 "Interface      BytesIn  PktsIn  BytesOut PktsOut Flags Local    Remote\n");
2924	} else {
2925		const struct vif_device *vif = v;
2926		const char *name =  vif->dev ?
2927				    vif->dev->name : "none";
2928
2929		seq_printf(seq,
2930			   "%2td %-10s %8ld %7ld  %8ld %7ld %05X %08X %08X\n",
2931			   vif - mrt->vif_table,
2932			   name, vif->bytes_in, vif->pkt_in,
2933			   vif->bytes_out, vif->pkt_out,
2934			   vif->flags, vif->local, vif->remote);
2935	}
2936	return 0;
2937}
2938
2939static const struct seq_operations ipmr_vif_seq_ops = {
2940	.start = ipmr_vif_seq_start,
2941	.next  = mr_vif_seq_next,
2942	.stop  = ipmr_vif_seq_stop,
2943	.show  = ipmr_vif_seq_show,
2944};
2945
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2946static void *ipmr_mfc_seq_start(struct seq_file *seq, loff_t *pos)
2947{
 
2948	struct net *net = seq_file_net(seq);
2949	struct mr_table *mrt;
2950
2951	mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2952	if (!mrt)
2953		return ERR_PTR(-ENOENT);
2954
2955	return mr_mfc_seq_start(seq, pos, mrt, &mfc_unres_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2956}
2957
2958static int ipmr_mfc_seq_show(struct seq_file *seq, void *v)
2959{
2960	int n;
2961
2962	if (v == SEQ_START_TOKEN) {
2963		seq_puts(seq,
2964		 "Group    Origin   Iif     Pkts    Bytes    Wrong Oifs\n");
2965	} else {
2966		const struct mfc_cache *mfc = v;
2967		const struct mr_mfc_iter *it = seq->private;
2968		const struct mr_table *mrt = it->mrt;
2969
2970		seq_printf(seq, "%08X %08X %-3hd",
2971			   (__force u32) mfc->mfc_mcastgrp,
2972			   (__force u32) mfc->mfc_origin,
2973			   mfc->_c.mfc_parent);
2974
2975		if (it->cache != &mrt->mfc_unres_queue) {
2976			seq_printf(seq, " %8lu %8lu %8lu",
2977				   mfc->_c.mfc_un.res.pkt,
2978				   mfc->_c.mfc_un.res.bytes,
2979				   mfc->_c.mfc_un.res.wrong_if);
2980			for (n = mfc->_c.mfc_un.res.minvif;
2981			     n < mfc->_c.mfc_un.res.maxvif; n++) {
2982				if (VIF_EXISTS(mrt, n) &&
2983				    mfc->_c.mfc_un.res.ttls[n] < 255)
2984					seq_printf(seq,
2985					   " %2d:%-3d",
2986					   n, mfc->_c.mfc_un.res.ttls[n]);
2987			}
2988		} else {
2989			/* unresolved mfc_caches don't contain
2990			 * pkt, bytes and wrong_if values
2991			 */
2992			seq_printf(seq, " %8lu %8lu %8lu", 0ul, 0ul, 0ul);
2993		}
2994		seq_putc(seq, '\n');
2995	}
2996	return 0;
2997}
2998
2999static const struct seq_operations ipmr_mfc_seq_ops = {
3000	.start = ipmr_mfc_seq_start,
3001	.next  = mr_mfc_seq_next,
3002	.stop  = mr_mfc_seq_stop,
3003	.show  = ipmr_mfc_seq_show,
3004};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3005#endif
3006
3007#ifdef CONFIG_IP_PIMSM_V2
3008static const struct net_protocol pim_protocol = {
3009	.handler	=	pim_rcv,
 
3010};
3011#endif
3012
3013static unsigned int ipmr_seq_read(struct net *net)
3014{
3015	ASSERT_RTNL();
3016
3017	return net->ipv4.ipmr_seq + ipmr_rules_seq_read(net);
3018}
3019
3020static int ipmr_dump(struct net *net, struct notifier_block *nb,
3021		     struct netlink_ext_ack *extack)
3022{
3023	return mr_dump(net, nb, RTNL_FAMILY_IPMR, ipmr_rules_dump,
3024		       ipmr_mr_table_iter, &mrt_lock, extack);
3025}
3026
3027static const struct fib_notifier_ops ipmr_notifier_ops_template = {
3028	.family		= RTNL_FAMILY_IPMR,
3029	.fib_seq_read	= ipmr_seq_read,
3030	.fib_dump	= ipmr_dump,
3031	.owner		= THIS_MODULE,
3032};
3033
3034static int __net_init ipmr_notifier_init(struct net *net)
3035{
3036	struct fib_notifier_ops *ops;
3037
3038	net->ipv4.ipmr_seq = 0;
3039
3040	ops = fib_notifier_ops_register(&ipmr_notifier_ops_template, net);
3041	if (IS_ERR(ops))
3042		return PTR_ERR(ops);
3043	net->ipv4.ipmr_notifier_ops = ops;
3044
3045	return 0;
3046}
3047
3048static void __net_exit ipmr_notifier_exit(struct net *net)
3049{
3050	fib_notifier_ops_unregister(net->ipv4.ipmr_notifier_ops);
3051	net->ipv4.ipmr_notifier_ops = NULL;
3052}
3053
3054/* Setup for IP multicast routing */
3055static int __net_init ipmr_net_init(struct net *net)
3056{
3057	int err;
3058
3059	err = ipmr_notifier_init(net);
3060	if (err)
3061		goto ipmr_notifier_fail;
3062
3063	err = ipmr_rules_init(net);
3064	if (err < 0)
3065		goto ipmr_rules_fail;
3066
3067#ifdef CONFIG_PROC_FS
3068	err = -ENOMEM;
3069	if (!proc_create_net("ip_mr_vif", 0, net->proc_net, &ipmr_vif_seq_ops,
3070			sizeof(struct mr_vif_iter)))
3071		goto proc_vif_fail;
3072	if (!proc_create_net("ip_mr_cache", 0, net->proc_net, &ipmr_mfc_seq_ops,
3073			sizeof(struct mr_mfc_iter)))
3074		goto proc_cache_fail;
3075#endif
3076	return 0;
3077
3078#ifdef CONFIG_PROC_FS
3079proc_cache_fail:
3080	remove_proc_entry("ip_mr_vif", net->proc_net);
3081proc_vif_fail:
3082	ipmr_rules_exit(net);
3083#endif
3084ipmr_rules_fail:
3085	ipmr_notifier_exit(net);
3086ipmr_notifier_fail:
3087	return err;
3088}
3089
3090static void __net_exit ipmr_net_exit(struct net *net)
3091{
3092#ifdef CONFIG_PROC_FS
3093	remove_proc_entry("ip_mr_cache", net->proc_net);
3094	remove_proc_entry("ip_mr_vif", net->proc_net);
3095#endif
3096	ipmr_notifier_exit(net);
3097	ipmr_rules_exit(net);
3098}
3099
3100static struct pernet_operations ipmr_net_ops = {
3101	.init = ipmr_net_init,
3102	.exit = ipmr_net_exit,
3103};
3104
3105int __init ip_mr_init(void)
3106{
3107	int err;
3108
3109	mrt_cachep = kmem_cache_create("ip_mrt_cache",
3110				       sizeof(struct mfc_cache),
3111				       0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
3112				       NULL);
3113
3114	err = register_pernet_subsys(&ipmr_net_ops);
3115	if (err)
3116		goto reg_pernet_fail;
3117
3118	err = register_netdevice_notifier(&ip_mr_notifier);
3119	if (err)
3120		goto reg_notif_fail;
3121#ifdef CONFIG_IP_PIMSM_V2
3122	if (inet_add_protocol(&pim_protocol, IPPROTO_PIM) < 0) {
3123		pr_err("%s: can't add PIM protocol\n", __func__);
3124		err = -EAGAIN;
3125		goto add_proto_fail;
3126	}
3127#endif
3128	rtnl_register(RTNL_FAMILY_IPMR, RTM_GETROUTE,
3129		      ipmr_rtm_getroute, ipmr_rtm_dumproute, 0);
3130	rtnl_register(RTNL_FAMILY_IPMR, RTM_NEWROUTE,
3131		      ipmr_rtm_route, NULL, 0);
3132	rtnl_register(RTNL_FAMILY_IPMR, RTM_DELROUTE,
3133		      ipmr_rtm_route, NULL, 0);
3134
3135	rtnl_register(RTNL_FAMILY_IPMR, RTM_GETLINK,
3136		      NULL, ipmr_rtm_dumplink, 0);
3137	return 0;
3138
3139#ifdef CONFIG_IP_PIMSM_V2
3140add_proto_fail:
3141	unregister_netdevice_notifier(&ip_mr_notifier);
3142#endif
3143reg_notif_fail:
3144	unregister_pernet_subsys(&ipmr_net_ops);
3145reg_pernet_fail:
3146	kmem_cache_destroy(mrt_cachep);
3147	return err;
3148}
v4.6
 
   1/*
   2 *	IP multicast routing support for mrouted 3.6/3.8
   3 *
   4 *		(c) 1995 Alan Cox, <alan@lxorguk.ukuu.org.uk>
   5 *	  Linux Consultancy and Custom Driver Development
   6 *
   7 *	This program is free software; you can redistribute it and/or
   8 *	modify it under the terms of the GNU General Public License
   9 *	as published by the Free Software Foundation; either version
  10 *	2 of the License, or (at your option) any later version.
  11 *
  12 *	Fixes:
  13 *	Michael Chastain	:	Incorrect size of copying.
  14 *	Alan Cox		:	Added the cache manager code
  15 *	Alan Cox		:	Fixed the clone/copy bug and device race.
  16 *	Mike McLagan		:	Routing by source
  17 *	Malcolm Beattie		:	Buffer handling fixes.
  18 *	Alexey Kuznetsov	:	Double buffer free and other fixes.
  19 *	SVR Anand		:	Fixed several multicast bugs and problems.
  20 *	Alexey Kuznetsov	:	Status, optimisations and more.
  21 *	Brad Parker		:	Better behaviour on mrouted upcall
  22 *					overflow.
  23 *      Carlos Picoto           :       PIMv1 Support
  24 *	Pavlin Ivanov Radoslavov:	PIMv2 Registers must checksum only PIM header
  25 *					Relax this requirement to work with older peers.
  26 *
  27 */
  28
  29#include <asm/uaccess.h>
  30#include <linux/types.h>
 
  31#include <linux/capability.h>
  32#include <linux/errno.h>
  33#include <linux/timer.h>
  34#include <linux/mm.h>
  35#include <linux/kernel.h>
  36#include <linux/fcntl.h>
  37#include <linux/stat.h>
  38#include <linux/socket.h>
  39#include <linux/in.h>
  40#include <linux/inet.h>
  41#include <linux/netdevice.h>
  42#include <linux/inetdevice.h>
  43#include <linux/igmp.h>
  44#include <linux/proc_fs.h>
  45#include <linux/seq_file.h>
  46#include <linux/mroute.h>
  47#include <linux/init.h>
  48#include <linux/if_ether.h>
  49#include <linux/slab.h>
  50#include <net/net_namespace.h>
  51#include <net/ip.h>
  52#include <net/protocol.h>
  53#include <linux/skbuff.h>
  54#include <net/route.h>
  55#include <net/sock.h>
  56#include <net/icmp.h>
  57#include <net/udp.h>
  58#include <net/raw.h>
  59#include <linux/notifier.h>
  60#include <linux/if_arp.h>
  61#include <linux/netfilter_ipv4.h>
  62#include <linux/compat.h>
  63#include <linux/export.h>
 
  64#include <net/ip_tunnels.h>
  65#include <net/checksum.h>
  66#include <net/netlink.h>
  67#include <net/fib_rules.h>
  68#include <linux/netconf.h>
  69#include <net/nexthop.h>
 
 
  70
  71struct ipmr_rule {
  72	struct fib_rule		common;
  73};
  74
  75struct ipmr_result {
  76	struct mr_table		*mrt;
  77};
  78
  79/* Big lock, protecting vif table, mrt cache and mroute socket state.
  80 * Note that the changes are semaphored via rtnl_lock.
  81 */
  82
  83static DEFINE_RWLOCK(mrt_lock);
  84
  85/* Multicast router control variables */
  86
  87/* Special spinlock for queue of unresolved entries */
  88static DEFINE_SPINLOCK(mfc_unres_lock);
  89
  90/* We return to original Alan's scheme. Hash table of resolved
  91 * entries is changed only in process context and protected
  92 * with weak lock mrt_lock. Queue of unresolved entries is protected
  93 * with strong spinlock mfc_unres_lock.
  94 *
  95 * In this case data path is free of exclusive locks at all.
  96 */
  97
  98static struct kmem_cache *mrt_cachep __read_mostly;
  99
 100static struct mr_table *ipmr_new_table(struct net *net, u32 id);
 101static void ipmr_free_table(struct mr_table *mrt);
 102
 103static void ip_mr_forward(struct net *net, struct mr_table *mrt,
 104			  struct sk_buff *skb, struct mfc_cache *cache,
 105			  int local);
 106static int ipmr_cache_report(struct mr_table *mrt,
 107			     struct sk_buff *pkt, vifi_t vifi, int assert);
 108static int __ipmr_fill_mroute(struct mr_table *mrt, struct sk_buff *skb,
 109			      struct mfc_cache *c, struct rtmsg *rtm);
 110static void mroute_netlink_event(struct mr_table *mrt, struct mfc_cache *mfc,
 111				 int cmd);
 112static void mroute_clean_tables(struct mr_table *mrt, bool all);
 113static void ipmr_expire_process(unsigned long arg);
 
 114
 115#ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
 116#define ipmr_for_each_table(mrt, net) \
 117	list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 118
 119static struct mr_table *ipmr_get_table(struct net *net, u32 id)
 120{
 121	struct mr_table *mrt;
 122
 123	ipmr_for_each_table(mrt, net) {
 124		if (mrt->id == id)
 125			return mrt;
 126	}
 127	return NULL;
 128}
 129
 130static int ipmr_fib_lookup(struct net *net, struct flowi4 *flp4,
 131			   struct mr_table **mrt)
 132{
 133	int err;
 134	struct ipmr_result res;
 135	struct fib_lookup_arg arg = {
 136		.result = &res,
 137		.flags = FIB_LOOKUP_NOREF,
 138	};
 139
 
 
 
 140	err = fib_rules_lookup(net->ipv4.mr_rules_ops,
 141			       flowi4_to_flowi(flp4), 0, &arg);
 142	if (err < 0)
 143		return err;
 144	*mrt = res.mrt;
 145	return 0;
 146}
 147
 148static int ipmr_rule_action(struct fib_rule *rule, struct flowi *flp,
 149			    int flags, struct fib_lookup_arg *arg)
 150{
 151	struct ipmr_result *res = arg->result;
 152	struct mr_table *mrt;
 153
 154	switch (rule->action) {
 155	case FR_ACT_TO_TBL:
 156		break;
 157	case FR_ACT_UNREACHABLE:
 158		return -ENETUNREACH;
 159	case FR_ACT_PROHIBIT:
 160		return -EACCES;
 161	case FR_ACT_BLACKHOLE:
 162	default:
 163		return -EINVAL;
 164	}
 165
 166	mrt = ipmr_get_table(rule->fr_net, rule->table);
 
 
 167	if (!mrt)
 168		return -EAGAIN;
 169	res->mrt = mrt;
 170	return 0;
 171}
 172
 173static int ipmr_rule_match(struct fib_rule *rule, struct flowi *fl, int flags)
 174{
 175	return 1;
 176}
 177
 178static const struct nla_policy ipmr_rule_policy[FRA_MAX + 1] = {
 179	FRA_GENERIC_POLICY,
 180};
 181
 182static int ipmr_rule_configure(struct fib_rule *rule, struct sk_buff *skb,
 183			       struct fib_rule_hdr *frh, struct nlattr **tb)
 
 184{
 185	return 0;
 186}
 187
 188static int ipmr_rule_compare(struct fib_rule *rule, struct fib_rule_hdr *frh,
 189			     struct nlattr **tb)
 190{
 191	return 1;
 192}
 193
 194static int ipmr_rule_fill(struct fib_rule *rule, struct sk_buff *skb,
 195			  struct fib_rule_hdr *frh)
 196{
 197	frh->dst_len = 0;
 198	frh->src_len = 0;
 199	frh->tos     = 0;
 200	return 0;
 201}
 202
 203static const struct fib_rules_ops __net_initconst ipmr_rules_ops_template = {
 204	.family		= RTNL_FAMILY_IPMR,
 205	.rule_size	= sizeof(struct ipmr_rule),
 206	.addr_size	= sizeof(u32),
 207	.action		= ipmr_rule_action,
 208	.match		= ipmr_rule_match,
 209	.configure	= ipmr_rule_configure,
 210	.compare	= ipmr_rule_compare,
 211	.fill		= ipmr_rule_fill,
 212	.nlgroup	= RTNLGRP_IPV4_RULE,
 213	.policy		= ipmr_rule_policy,
 214	.owner		= THIS_MODULE,
 215};
 216
 217static int __net_init ipmr_rules_init(struct net *net)
 218{
 219	struct fib_rules_ops *ops;
 220	struct mr_table *mrt;
 221	int err;
 222
 223	ops = fib_rules_register(&ipmr_rules_ops_template, net);
 224	if (IS_ERR(ops))
 225		return PTR_ERR(ops);
 226
 227	INIT_LIST_HEAD(&net->ipv4.mr_tables);
 228
 229	mrt = ipmr_new_table(net, RT_TABLE_DEFAULT);
 230	if (IS_ERR(mrt)) {
 231		err = PTR_ERR(mrt);
 232		goto err1;
 233	}
 234
 235	err = fib_default_rule_add(ops, 0x7fff, RT_TABLE_DEFAULT, 0);
 236	if (err < 0)
 237		goto err2;
 238
 239	net->ipv4.mr_rules_ops = ops;
 240	return 0;
 241
 242err2:
 243	ipmr_free_table(mrt);
 244err1:
 245	fib_rules_unregister(ops);
 246	return err;
 247}
 248
 249static void __net_exit ipmr_rules_exit(struct net *net)
 250{
 251	struct mr_table *mrt, *next;
 252
 253	rtnl_lock();
 254	list_for_each_entry_safe(mrt, next, &net->ipv4.mr_tables, list) {
 255		list_del(&mrt->list);
 256		ipmr_free_table(mrt);
 257	}
 258	fib_rules_unregister(net->ipv4.mr_rules_ops);
 259	rtnl_unlock();
 260}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 261#else
 262#define ipmr_for_each_table(mrt, net) \
 263	for (mrt = net->ipv4.mrt; mrt; mrt = NULL)
 264
 
 
 
 
 
 
 
 
 265static struct mr_table *ipmr_get_table(struct net *net, u32 id)
 266{
 267	return net->ipv4.mrt;
 268}
 269
 270static int ipmr_fib_lookup(struct net *net, struct flowi4 *flp4,
 271			   struct mr_table **mrt)
 272{
 273	*mrt = net->ipv4.mrt;
 274	return 0;
 275}
 276
 277static int __net_init ipmr_rules_init(struct net *net)
 278{
 279	struct mr_table *mrt;
 280
 281	mrt = ipmr_new_table(net, RT_TABLE_DEFAULT);
 282	if (IS_ERR(mrt))
 283		return PTR_ERR(mrt);
 284	net->ipv4.mrt = mrt;
 285	return 0;
 286}
 287
 288static void __net_exit ipmr_rules_exit(struct net *net)
 289{
 290	rtnl_lock();
 291	ipmr_free_table(net->ipv4.mrt);
 292	net->ipv4.mrt = NULL;
 293	rtnl_unlock();
 294}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 295#endif
 
 
 
 
 
 
 
 
 
 
 
 296
 297static struct mr_table *ipmr_new_table(struct net *net, u32 id)
 298{
 299	struct mr_table *mrt;
 300	unsigned int i;
 301
 302	/* "pimreg%u" should not exceed 16 bytes (IFNAMSIZ) */
 303	if (id != RT_TABLE_DEFAULT && id >= 1000000000)
 304		return ERR_PTR(-EINVAL);
 305
 306	mrt = ipmr_get_table(net, id);
 307	if (mrt)
 308		return mrt;
 309
 310	mrt = kzalloc(sizeof(*mrt), GFP_KERNEL);
 311	if (!mrt)
 312		return ERR_PTR(-ENOMEM);
 313	write_pnet(&mrt->net, net);
 314	mrt->id = id;
 315
 316	/* Forwarding cache */
 317	for (i = 0; i < MFC_LINES; i++)
 318		INIT_LIST_HEAD(&mrt->mfc_cache_array[i]);
 319
 320	INIT_LIST_HEAD(&mrt->mfc_unres_queue);
 321
 322	setup_timer(&mrt->ipmr_expire_timer, ipmr_expire_process,
 323		    (unsigned long)mrt);
 324
 325	mrt->mroute_reg_vif_num = -1;
 326#ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
 327	list_add_tail_rcu(&mrt->list, &net->ipv4.mr_tables);
 328#endif
 329	return mrt;
 330}
 331
 332static void ipmr_free_table(struct mr_table *mrt)
 333{
 334	del_timer_sync(&mrt->ipmr_expire_timer);
 335	mroute_clean_tables(mrt, true);
 
 
 336	kfree(mrt);
 337}
 338
 339/* Service routines creating virtual interfaces: DVMRP tunnels and PIMREG */
 340
 341static void ipmr_del_tunnel(struct net_device *dev, struct vifctl *v)
 342{
 343	struct net *net = dev_net(dev);
 344
 345	dev_close(dev);
 346
 347	dev = __dev_get_by_name(net, "tunl0");
 348	if (dev) {
 349		const struct net_device_ops *ops = dev->netdev_ops;
 350		struct ifreq ifr;
 351		struct ip_tunnel_parm p;
 352
 353		memset(&p, 0, sizeof(p));
 354		p.iph.daddr = v->vifc_rmt_addr.s_addr;
 355		p.iph.saddr = v->vifc_lcl_addr.s_addr;
 356		p.iph.version = 4;
 357		p.iph.ihl = 5;
 358		p.iph.protocol = IPPROTO_IPIP;
 359		sprintf(p.name, "dvmrp%d", v->vifc_vifi);
 360		ifr.ifr_ifru.ifru_data = (__force void __user *)&p;
 361
 362		if (ops->ndo_do_ioctl) {
 363			mm_segment_t oldfs = get_fs();
 364
 365			set_fs(KERNEL_DS);
 366			ops->ndo_do_ioctl(dev, &ifr, SIOCDELTUNNEL);
 367			set_fs(oldfs);
 368		}
 369	}
 370}
 371
 372/* Initialize ipmr pimreg/tunnel in_device */
 373static bool ipmr_init_vif_indev(const struct net_device *dev)
 374{
 375	struct in_device *in_dev;
 376
 377	ASSERT_RTNL();
 378
 379	in_dev = __in_dev_get_rtnl(dev);
 380	if (!in_dev)
 381		return false;
 382	ipv4_devconf_setall(in_dev);
 383	neigh_parms_data_state_setall(in_dev->arp_parms);
 384	IPV4_DEVCONF(in_dev->cnf, RP_FILTER) = 0;
 385
 386	return true;
 387}
 388
 389static struct net_device *ipmr_new_tunnel(struct net *net, struct vifctl *v)
 390{
 391	struct net_device  *dev;
 
 
 
 
 
 
 392
 393	dev = __dev_get_by_name(net, "tunl0");
 
 
 
 
 
 
 
 
 
 
 
 
 394
 395	if (dev) {
 396		const struct net_device_ops *ops = dev->netdev_ops;
 397		int err;
 398		struct ifreq ifr;
 399		struct ip_tunnel_parm p;
 400
 401		memset(&p, 0, sizeof(p));
 402		p.iph.daddr = v->vifc_rmt_addr.s_addr;
 403		p.iph.saddr = v->vifc_lcl_addr.s_addr;
 404		p.iph.version = 4;
 405		p.iph.ihl = 5;
 406		p.iph.protocol = IPPROTO_IPIP;
 407		sprintf(p.name, "dvmrp%d", v->vifc_vifi);
 408		ifr.ifr_ifru.ifru_data = (__force void __user *)&p;
 409
 410		if (ops->ndo_do_ioctl) {
 411			mm_segment_t oldfs = get_fs();
 412
 413			set_fs(KERNEL_DS);
 414			err = ops->ndo_do_ioctl(dev, &ifr, SIOCADDTUNNEL);
 415			set_fs(oldfs);
 416		} else {
 417			err = -EOPNOTSUPP;
 418		}
 419		dev = NULL;
 420
 421		if (err == 0 &&
 422		    (dev = __dev_get_by_name(net, p.name)) != NULL) {
 423			dev->flags |= IFF_MULTICAST;
 424			if (!ipmr_init_vif_indev(dev))
 425				goto failure;
 426			if (dev_open(dev))
 427				goto failure;
 428			dev_hold(dev);
 429		}
 
 
 
 
 430	}
 431	return dev;
 432
 433failure:
 434	unregister_netdevice(dev);
 435	return NULL;
 
 436}
 437
 438#if defined(CONFIG_IP_PIMSM_V1) || defined(CONFIG_IP_PIMSM_V2)
 439static netdev_tx_t reg_vif_xmit(struct sk_buff *skb, struct net_device *dev)
 440{
 441	struct net *net = dev_net(dev);
 442	struct mr_table *mrt;
 443	struct flowi4 fl4 = {
 444		.flowi4_oif	= dev->ifindex,
 445		.flowi4_iif	= skb->skb_iif ? : LOOPBACK_IFINDEX,
 446		.flowi4_mark	= skb->mark,
 447	};
 448	int err;
 449
 450	err = ipmr_fib_lookup(net, &fl4, &mrt);
 451	if (err < 0) {
 452		kfree_skb(skb);
 453		return err;
 454	}
 455
 456	read_lock(&mrt_lock);
 457	dev->stats.tx_bytes += skb->len;
 458	dev->stats.tx_packets++;
 459	ipmr_cache_report(mrt, skb, mrt->mroute_reg_vif_num, IGMPMSG_WHOLEPKT);
 460	read_unlock(&mrt_lock);
 461	kfree_skb(skb);
 462	return NETDEV_TX_OK;
 463}
 464
 465static int reg_vif_get_iflink(const struct net_device *dev)
 466{
 467	return 0;
 468}
 469
 470static const struct net_device_ops reg_vif_netdev_ops = {
 471	.ndo_start_xmit	= reg_vif_xmit,
 472	.ndo_get_iflink = reg_vif_get_iflink,
 473};
 474
 475static void reg_vif_setup(struct net_device *dev)
 476{
 477	dev->type		= ARPHRD_PIMREG;
 478	dev->mtu		= ETH_DATA_LEN - sizeof(struct iphdr) - 8;
 479	dev->flags		= IFF_NOARP;
 480	dev->netdev_ops		= &reg_vif_netdev_ops;
 481	dev->destructor		= free_netdev;
 482	dev->features		|= NETIF_F_NETNS_LOCAL;
 483}
 484
 485static struct net_device *ipmr_reg_vif(struct net *net, struct mr_table *mrt)
 486{
 487	struct net_device *dev;
 488	char name[IFNAMSIZ];
 489
 490	if (mrt->id == RT_TABLE_DEFAULT)
 491		sprintf(name, "pimreg");
 492	else
 493		sprintf(name, "pimreg%u", mrt->id);
 494
 495	dev = alloc_netdev(0, name, NET_NAME_UNKNOWN, reg_vif_setup);
 496
 497	if (!dev)
 498		return NULL;
 499
 500	dev_net_set(dev, net);
 501
 502	if (register_netdevice(dev)) {
 503		free_netdev(dev);
 504		return NULL;
 505	}
 506
 507	if (!ipmr_init_vif_indev(dev))
 508		goto failure;
 509	if (dev_open(dev))
 510		goto failure;
 511
 512	dev_hold(dev);
 513
 514	return dev;
 515
 516failure:
 517	unregister_netdevice(dev);
 518	return NULL;
 519}
 520
 521/* called with rcu_read_lock() */
 522static int __pim_rcv(struct mr_table *mrt, struct sk_buff *skb,
 523		     unsigned int pimlen)
 524{
 525	struct net_device *reg_dev = NULL;
 526	struct iphdr *encap;
 527
 528	encap = (struct iphdr *)(skb_transport_header(skb) + pimlen);
 529	/* Check that:
 530	 * a. packet is really sent to a multicast group
 531	 * b. packet is not a NULL-REGISTER
 532	 * c. packet is not truncated
 533	 */
 534	if (!ipv4_is_multicast(encap->daddr) ||
 535	    encap->tot_len == 0 ||
 536	    ntohs(encap->tot_len) + pimlen > skb->len)
 537		return 1;
 538
 539	read_lock(&mrt_lock);
 540	if (mrt->mroute_reg_vif_num >= 0)
 541		reg_dev = mrt->vif_table[mrt->mroute_reg_vif_num].dev;
 542	read_unlock(&mrt_lock);
 543
 544	if (!reg_dev)
 545		return 1;
 546
 547	skb->mac_header = skb->network_header;
 548	skb_pull(skb, (u8 *)encap - skb->data);
 549	skb_reset_network_header(skb);
 550	skb->protocol = htons(ETH_P_IP);
 551	skb->ip_summed = CHECKSUM_NONE;
 552
 553	skb_tunnel_rx(skb, reg_dev, dev_net(reg_dev));
 554
 555	netif_rx(skb);
 556
 557	return NET_RX_SUCCESS;
 558}
 559#else
 560static struct net_device *ipmr_reg_vif(struct net *net, struct mr_table *mrt)
 561{
 562	return NULL;
 563}
 564#endif
 565
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 566/**
 567 *	vif_delete - Delete a VIF entry
 
 
 568 *	@notify: Set to 1, if the caller is a notifier_call
 
 569 */
 570static int vif_delete(struct mr_table *mrt, int vifi, int notify,
 571		      struct list_head *head)
 572{
 
 573	struct vif_device *v;
 574	struct net_device *dev;
 575	struct in_device *in_dev;
 576
 577	if (vifi < 0 || vifi >= mrt->maxvif)
 578		return -EADDRNOTAVAIL;
 579
 580	v = &mrt->vif_table[vifi];
 581
 
 
 
 
 582	write_lock_bh(&mrt_lock);
 583	dev = v->dev;
 584	v->dev = NULL;
 585
 586	if (!dev) {
 587		write_unlock_bh(&mrt_lock);
 588		return -EADDRNOTAVAIL;
 589	}
 590
 591	if (vifi == mrt->mroute_reg_vif_num)
 592		mrt->mroute_reg_vif_num = -1;
 593
 594	if (vifi + 1 == mrt->maxvif) {
 595		int tmp;
 596
 597		for (tmp = vifi - 1; tmp >= 0; tmp--) {
 598			if (VIF_EXISTS(mrt, tmp))
 599				break;
 600		}
 601		mrt->maxvif = tmp+1;
 602	}
 603
 604	write_unlock_bh(&mrt_lock);
 605
 606	dev_set_allmulti(dev, -1);
 607
 608	in_dev = __in_dev_get_rtnl(dev);
 609	if (in_dev) {
 610		IPV4_DEVCONF(in_dev->cnf, MC_FORWARDING)--;
 611		inet_netconf_notify_devconf(dev_net(dev),
 612					    NETCONFA_MC_FORWARDING,
 613					    dev->ifindex, &in_dev->cnf);
 614		ip_rt_multicast_event(in_dev);
 615	}
 616
 617	if (v->flags & (VIFF_TUNNEL | VIFF_REGISTER) && !notify)
 618		unregister_netdevice_queue(dev, head);
 619
 620	dev_put(dev);
 621	return 0;
 622}
 623
 624static void ipmr_cache_free_rcu(struct rcu_head *head)
 625{
 626	struct mfc_cache *c = container_of(head, struct mfc_cache, rcu);
 627
 628	kmem_cache_free(mrt_cachep, c);
 629}
 630
 631static inline void ipmr_cache_free(struct mfc_cache *c)
 632{
 633	call_rcu(&c->rcu, ipmr_cache_free_rcu);
 634}
 635
 636/* Destroy an unresolved cache entry, killing queued skbs
 637 * and reporting error to netlink readers.
 638 */
 639static void ipmr_destroy_unres(struct mr_table *mrt, struct mfc_cache *c)
 640{
 641	struct net *net = read_pnet(&mrt->net);
 642	struct sk_buff *skb;
 643	struct nlmsgerr *e;
 644
 645	atomic_dec(&mrt->cache_resolve_queue_len);
 646
 647	while ((skb = skb_dequeue(&c->mfc_un.unres.unresolved))) {
 648		if (ip_hdr(skb)->version == 0) {
 649			struct nlmsghdr *nlh = (struct nlmsghdr *)skb_pull(skb, sizeof(struct iphdr));
 
 650			nlh->nlmsg_type = NLMSG_ERROR;
 651			nlh->nlmsg_len = nlmsg_msg_size(sizeof(struct nlmsgerr));
 652			skb_trim(skb, nlh->nlmsg_len);
 653			e = nlmsg_data(nlh);
 654			e->error = -ETIMEDOUT;
 655			memset(&e->msg, 0, sizeof(e->msg));
 656
 657			rtnl_unicast(skb, net, NETLINK_CB(skb).portid);
 658		} else {
 659			kfree_skb(skb);
 660		}
 661	}
 662
 663	ipmr_cache_free(c);
 664}
 665
 666/* Timer process for the unresolved queue. */
 667static void ipmr_expire_process(unsigned long arg)
 668{
 669	struct mr_table *mrt = (struct mr_table *)arg;
 
 
 670	unsigned long now;
 671	unsigned long expires;
 672	struct mfc_cache *c, *next;
 673
 674	if (!spin_trylock(&mfc_unres_lock)) {
 675		mod_timer(&mrt->ipmr_expire_timer, jiffies+HZ/10);
 676		return;
 677	}
 678
 679	if (list_empty(&mrt->mfc_unres_queue))
 680		goto out;
 681
 682	now = jiffies;
 683	expires = 10*HZ;
 684
 685	list_for_each_entry_safe(c, next, &mrt->mfc_unres_queue, list) {
 686		if (time_after(c->mfc_un.unres.expires, now)) {
 687			unsigned long interval = c->mfc_un.unres.expires - now;
 688			if (interval < expires)
 689				expires = interval;
 690			continue;
 691		}
 692
 693		list_del(&c->list);
 694		mroute_netlink_event(mrt, c, RTM_DELROUTE);
 695		ipmr_destroy_unres(mrt, c);
 696	}
 697
 698	if (!list_empty(&mrt->mfc_unres_queue))
 699		mod_timer(&mrt->ipmr_expire_timer, jiffies + expires);
 700
 701out:
 702	spin_unlock(&mfc_unres_lock);
 703}
 704
 705/* Fill oifs list. It is called under write locked mrt_lock. */
 706static void ipmr_update_thresholds(struct mr_table *mrt, struct mfc_cache *cache,
 707				   unsigned char *ttls)
 708{
 709	int vifi;
 710
 711	cache->mfc_un.res.minvif = MAXVIFS;
 712	cache->mfc_un.res.maxvif = 0;
 713	memset(cache->mfc_un.res.ttls, 255, MAXVIFS);
 714
 715	for (vifi = 0; vifi < mrt->maxvif; vifi++) {
 716		if (VIF_EXISTS(mrt, vifi) &&
 717		    ttls[vifi] && ttls[vifi] < 255) {
 718			cache->mfc_un.res.ttls[vifi] = ttls[vifi];
 719			if (cache->mfc_un.res.minvif > vifi)
 720				cache->mfc_un.res.minvif = vifi;
 721			if (cache->mfc_un.res.maxvif <= vifi)
 722				cache->mfc_un.res.maxvif = vifi + 1;
 723		}
 724	}
 
 725}
 726
 727static int vif_add(struct net *net, struct mr_table *mrt,
 728		   struct vifctl *vifc, int mrtsock)
 729{
 
 730	int vifi = vifc->vifc_vifi;
 731	struct vif_device *v = &mrt->vif_table[vifi];
 732	struct net_device *dev;
 733	struct in_device *in_dev;
 734	int err;
 735
 736	/* Is vif busy ? */
 737	if (VIF_EXISTS(mrt, vifi))
 738		return -EADDRINUSE;
 739
 740	switch (vifc->vifc_flags) {
 741	case VIFF_REGISTER:
 742		if (!ipmr_pimsm_enabled())
 743			return -EINVAL;
 744		/* Special Purpose VIF in PIM
 745		 * All the packets will be sent to the daemon
 746		 */
 747		if (mrt->mroute_reg_vif_num >= 0)
 748			return -EADDRINUSE;
 749		dev = ipmr_reg_vif(net, mrt);
 750		if (!dev)
 751			return -ENOBUFS;
 752		err = dev_set_allmulti(dev, 1);
 753		if (err) {
 754			unregister_netdevice(dev);
 755			dev_put(dev);
 756			return err;
 757		}
 758		break;
 759	case VIFF_TUNNEL:
 760		dev = ipmr_new_tunnel(net, vifc);
 761		if (!dev)
 762			return -ENOBUFS;
 763		err = dev_set_allmulti(dev, 1);
 764		if (err) {
 765			ipmr_del_tunnel(dev, vifc);
 766			dev_put(dev);
 767			return err;
 768		}
 769		break;
 770	case VIFF_USE_IFINDEX:
 771	case 0:
 772		if (vifc->vifc_flags == VIFF_USE_IFINDEX) {
 773			dev = dev_get_by_index(net, vifc->vifc_lcl_ifindex);
 774			if (dev && !__in_dev_get_rtnl(dev)) {
 775				dev_put(dev);
 776				return -EADDRNOTAVAIL;
 777			}
 778		} else {
 779			dev = ip_dev_find(net, vifc->vifc_lcl_addr.s_addr);
 780		}
 781		if (!dev)
 782			return -EADDRNOTAVAIL;
 783		err = dev_set_allmulti(dev, 1);
 784		if (err) {
 785			dev_put(dev);
 786			return err;
 787		}
 788		break;
 789	default:
 790		return -EINVAL;
 791	}
 792
 793	in_dev = __in_dev_get_rtnl(dev);
 794	if (!in_dev) {
 795		dev_put(dev);
 796		return -EADDRNOTAVAIL;
 797	}
 798	IPV4_DEVCONF(in_dev->cnf, MC_FORWARDING)++;
 799	inet_netconf_notify_devconf(net, NETCONFA_MC_FORWARDING, dev->ifindex,
 800				    &in_dev->cnf);
 801	ip_rt_multicast_event(in_dev);
 802
 803	/* Fill in the VIF structures */
 
 
 
 
 
 
 
 
 
 
 
 
 804
 805	v->rate_limit = vifc->vifc_rate_limit;
 806	v->local = vifc->vifc_lcl_addr.s_addr;
 807	v->remote = vifc->vifc_rmt_addr.s_addr;
 808	v->flags = vifc->vifc_flags;
 809	if (!mrtsock)
 810		v->flags |= VIFF_STATIC;
 811	v->threshold = vifc->vifc_threshold;
 812	v->bytes_in = 0;
 813	v->bytes_out = 0;
 814	v->pkt_in = 0;
 815	v->pkt_out = 0;
 816	v->link = dev->ifindex;
 817	if (v->flags & (VIFF_TUNNEL | VIFF_REGISTER))
 818		v->link = dev_get_iflink(dev);
 819
 820	/* And finish update writing critical data */
 821	write_lock_bh(&mrt_lock);
 822	v->dev = dev;
 823	if (v->flags & VIFF_REGISTER)
 824		mrt->mroute_reg_vif_num = vifi;
 825	if (vifi+1 > mrt->maxvif)
 826		mrt->maxvif = vifi+1;
 827	write_unlock_bh(&mrt_lock);
 
 828	return 0;
 829}
 830
 831/* called with rcu_read_lock() */
 832static struct mfc_cache *ipmr_cache_find(struct mr_table *mrt,
 833					 __be32 origin,
 834					 __be32 mcastgrp)
 835{
 836	int line = MFC_HASH(mcastgrp, origin);
 837	struct mfc_cache *c;
 838
 839	list_for_each_entry_rcu(c, &mrt->mfc_cache_array[line], list) {
 840		if (c->mfc_origin == origin && c->mfc_mcastgrp == mcastgrp)
 841			return c;
 842	}
 843	return NULL;
 844}
 845
 846/* Look for a (*,*,oif) entry */
 847static struct mfc_cache *ipmr_cache_find_any_parent(struct mr_table *mrt,
 848						    int vifi)
 849{
 850	int line = MFC_HASH(htonl(INADDR_ANY), htonl(INADDR_ANY));
 851	struct mfc_cache *c;
 852
 853	list_for_each_entry_rcu(c, &mrt->mfc_cache_array[line], list)
 854		if (c->mfc_origin == htonl(INADDR_ANY) &&
 855		    c->mfc_mcastgrp == htonl(INADDR_ANY) &&
 856		    c->mfc_un.res.ttls[vifi] < 255)
 857			return c;
 858
 859	return NULL;
 860}
 861
 862/* Look for a (*,G) entry */
 863static struct mfc_cache *ipmr_cache_find_any(struct mr_table *mrt,
 864					     __be32 mcastgrp, int vifi)
 865{
 866	int line = MFC_HASH(mcastgrp, htonl(INADDR_ANY));
 867	struct mfc_cache *c, *proxy;
 
 
 868
 869	if (mcastgrp == htonl(INADDR_ANY))
 870		goto skip;
 
 
 871
 872	list_for_each_entry_rcu(c, &mrt->mfc_cache_array[line], list)
 873		if (c->mfc_origin == htonl(INADDR_ANY) &&
 874		    c->mfc_mcastgrp == mcastgrp) {
 875			if (c->mfc_un.res.ttls[vifi] < 255)
 876				return c;
 877
 878			/* It's ok if the vifi is part of the static tree */
 879			proxy = ipmr_cache_find_any_parent(mrt,
 880							   c->mfc_parent);
 881			if (proxy && proxy->mfc_un.res.ttls[vifi] < 255)
 882				return c;
 883		}
 884
 885skip:
 886	return ipmr_cache_find_any_parent(mrt, vifi);
 887}
 888
 889/* Allocate a multicast cache entry */
 890static struct mfc_cache *ipmr_cache_alloc(void)
 891{
 892	struct mfc_cache *c = kmem_cache_zalloc(mrt_cachep, GFP_KERNEL);
 893
 894	if (c)
 895		c->mfc_un.res.minvif = MAXVIFS;
 
 
 
 
 896	return c;
 897}
 898
 899static struct mfc_cache *ipmr_cache_alloc_unres(void)
 900{
 901	struct mfc_cache *c = kmem_cache_zalloc(mrt_cachep, GFP_ATOMIC);
 902
 903	if (c) {
 904		skb_queue_head_init(&c->mfc_un.unres.unresolved);
 905		c->mfc_un.unres.expires = jiffies + 10*HZ;
 906	}
 907	return c;
 908}
 909
 910/* A cache entry has gone into a resolved state from queued */
 911static void ipmr_cache_resolve(struct net *net, struct mr_table *mrt,
 912			       struct mfc_cache *uc, struct mfc_cache *c)
 913{
 914	struct sk_buff *skb;
 915	struct nlmsgerr *e;
 916
 917	/* Play the pending entries through our router */
 918	while ((skb = __skb_dequeue(&uc->mfc_un.unres.unresolved))) {
 919		if (ip_hdr(skb)->version == 0) {
 920			struct nlmsghdr *nlh = (struct nlmsghdr *)skb_pull(skb, sizeof(struct iphdr));
 
 921
 922			if (__ipmr_fill_mroute(mrt, skb, c, nlmsg_data(nlh)) > 0) {
 
 923				nlh->nlmsg_len = skb_tail_pointer(skb) -
 924						 (u8 *)nlh;
 925			} else {
 926				nlh->nlmsg_type = NLMSG_ERROR;
 927				nlh->nlmsg_len = nlmsg_msg_size(sizeof(struct nlmsgerr));
 928				skb_trim(skb, nlh->nlmsg_len);
 929				e = nlmsg_data(nlh);
 930				e->error = -EMSGSIZE;
 931				memset(&e->msg, 0, sizeof(e->msg));
 932			}
 933
 934			rtnl_unicast(skb, net, NETLINK_CB(skb).portid);
 935		} else {
 936			ip_mr_forward(net, mrt, skb, c, 0);
 937		}
 938	}
 939}
 940
 941/* Bounce a cache query up to mrouted. We could use netlink for this but mrouted
 942 * expects the following bizarre scheme.
 943 *
 944 * Called under mrt_lock.
 945 */
 946static int ipmr_cache_report(struct mr_table *mrt,
 947			     struct sk_buff *pkt, vifi_t vifi, int assert)
 948{
 949	const int ihl = ip_hdrlen(pkt);
 950	struct sock *mroute_sk;
 951	struct igmphdr *igmp;
 952	struct igmpmsg *msg;
 953	struct sk_buff *skb;
 954	int ret;
 955
 956	if (assert == IGMPMSG_WHOLEPKT)
 957		skb = skb_realloc_headroom(pkt, sizeof(struct iphdr));
 958	else
 959		skb = alloc_skb(128, GFP_ATOMIC);
 960
 961	if (!skb)
 962		return -ENOBUFS;
 963
 964	if (assert == IGMPMSG_WHOLEPKT) {
 965		/* Ugly, but we have no choice with this interface.
 966		 * Duplicate old header, fix ihl, length etc.
 967		 * And all this only to mangle msg->im_msgtype and
 968		 * to set msg->im_mbz to "mbz" :-)
 969		 */
 970		skb_push(skb, sizeof(struct iphdr));
 971		skb_reset_network_header(skb);
 972		skb_reset_transport_header(skb);
 973		msg = (struct igmpmsg *)skb_network_header(skb);
 974		memcpy(msg, skb_network_header(pkt), sizeof(struct iphdr));
 975		msg->im_msgtype = IGMPMSG_WHOLEPKT;
 976		msg->im_mbz = 0;
 977		msg->im_vif = mrt->mroute_reg_vif_num;
 
 
 
 
 
 
 978		ip_hdr(skb)->ihl = sizeof(struct iphdr) >> 2;
 979		ip_hdr(skb)->tot_len = htons(ntohs(ip_hdr(pkt)->tot_len) +
 980					     sizeof(struct iphdr));
 981	} else {
 982		/* Copy the IP header */
 983		skb_set_network_header(skb, skb->len);
 984		skb_put(skb, ihl);
 985		skb_copy_to_linear_data(skb, pkt->data, ihl);
 986		/* Flag to the kernel this is a route add */
 987		ip_hdr(skb)->protocol = 0;
 988		msg = (struct igmpmsg *)skb_network_header(skb);
 989		msg->im_vif = vifi;
 
 990		skb_dst_set(skb, dst_clone(skb_dst(pkt)));
 991		/* Add our header */
 992		igmp = (struct igmphdr *)skb_put(skb, sizeof(struct igmphdr));
 993		igmp->type = assert;
 994		msg->im_msgtype = assert;
 995		igmp->code = 0;
 996		ip_hdr(skb)->tot_len = htons(skb->len);	/* Fix the length */
 997		skb->transport_header = skb->network_header;
 998	}
 999
1000	rcu_read_lock();
1001	mroute_sk = rcu_dereference(mrt->mroute_sk);
1002	if (!mroute_sk) {
1003		rcu_read_unlock();
1004		kfree_skb(skb);
1005		return -EINVAL;
1006	}
1007
 
 
1008	/* Deliver to mrouted */
1009	ret = sock_queue_rcv_skb(mroute_sk, skb);
1010	rcu_read_unlock();
1011	if (ret < 0) {
1012		net_warn_ratelimited("mroute: pending queue full, dropping entries\n");
1013		kfree_skb(skb);
1014	}
1015
1016	return ret;
1017}
1018
1019/* Queue a packet for resolution. It gets locked cache entry! */
1020static int ipmr_cache_unresolved(struct mr_table *mrt, vifi_t vifi,
1021				 struct sk_buff *skb)
1022{
 
 
1023	bool found = false;
1024	int err;
1025	struct mfc_cache *c;
1026	const struct iphdr *iph = ip_hdr(skb);
1027
1028	spin_lock_bh(&mfc_unres_lock);
1029	list_for_each_entry(c, &mrt->mfc_unres_queue, list) {
1030		if (c->mfc_mcastgrp == iph->daddr &&
1031		    c->mfc_origin == iph->saddr) {
1032			found = true;
1033			break;
1034		}
1035	}
1036
1037	if (!found) {
1038		/* Create a new entry if allowable */
1039		if (atomic_read(&mrt->cache_resolve_queue_len) >= 10 ||
1040		    (c = ipmr_cache_alloc_unres()) == NULL) {
1041			spin_unlock_bh(&mfc_unres_lock);
1042
1043			kfree_skb(skb);
1044			return -ENOBUFS;
1045		}
1046
1047		/* Fill in the new cache entry */
1048		c->mfc_parent	= -1;
1049		c->mfc_origin	= iph->saddr;
1050		c->mfc_mcastgrp	= iph->daddr;
1051
1052		/* Reflect first query at mrouted. */
1053		err = ipmr_cache_report(mrt, skb, vifi, IGMPMSG_NOCACHE);
 
1054		if (err < 0) {
1055			/* If the report failed throw the cache entry
1056			   out - Brad Parker
1057			 */
1058			spin_unlock_bh(&mfc_unres_lock);
1059
1060			ipmr_cache_free(c);
1061			kfree_skb(skb);
1062			return err;
1063		}
1064
1065		atomic_inc(&mrt->cache_resolve_queue_len);
1066		list_add(&c->list, &mrt->mfc_unres_queue);
1067		mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1068
1069		if (atomic_read(&mrt->cache_resolve_queue_len) == 1)
1070			mod_timer(&mrt->ipmr_expire_timer, c->mfc_un.unres.expires);
 
1071	}
1072
1073	/* See if we can append the packet */
1074	if (c->mfc_un.unres.unresolved.qlen > 3) {
1075		kfree_skb(skb);
1076		err = -ENOBUFS;
1077	} else {
1078		skb_queue_tail(&c->mfc_un.unres.unresolved, skb);
 
 
 
 
1079		err = 0;
1080	}
1081
1082	spin_unlock_bh(&mfc_unres_lock);
1083	return err;
1084}
1085
1086/* MFC cache manipulation by user space mroute daemon */
1087
1088static int ipmr_mfc_delete(struct mr_table *mrt, struct mfcctl *mfc, int parent)
1089{
1090	int line;
1091	struct mfc_cache *c, *next;
1092
1093	line = MFC_HASH(mfc->mfcc_mcastgrp.s_addr, mfc->mfcc_origin.s_addr);
 
 
 
 
 
 
 
 
 
 
 
1094
1095	list_for_each_entry_safe(c, next, &mrt->mfc_cache_array[line], list) {
1096		if (c->mfc_origin == mfc->mfcc_origin.s_addr &&
1097		    c->mfc_mcastgrp == mfc->mfcc_mcastgrp.s_addr &&
1098		    (parent == -1 || parent == c->mfc_parent)) {
1099			list_del_rcu(&c->list);
1100			mroute_netlink_event(mrt, c, RTM_DELROUTE);
1101			ipmr_cache_free(c);
1102			return 0;
1103		}
1104	}
1105	return -ENOENT;
1106}
1107
1108static int ipmr_mfc_add(struct net *net, struct mr_table *mrt,
1109			struct mfcctl *mfc, int mrtsock, int parent)
1110{
1111	bool found = false;
1112	int line;
1113	struct mfc_cache *uc, *c;
 
 
 
1114
1115	if (mfc->mfcc_parent >= MAXVIFS)
1116		return -ENFILE;
1117
1118	line = MFC_HASH(mfc->mfcc_mcastgrp.s_addr, mfc->mfcc_origin.s_addr);
1119
1120	list_for_each_entry(c, &mrt->mfc_cache_array[line], list) {
1121		if (c->mfc_origin == mfc->mfcc_origin.s_addr &&
1122		    c->mfc_mcastgrp == mfc->mfcc_mcastgrp.s_addr &&
1123		    (parent == -1 || parent == c->mfc_parent)) {
1124			found = true;
1125			break;
1126		}
1127	}
1128
1129	if (found) {
1130		write_lock_bh(&mrt_lock);
1131		c->mfc_parent = mfc->mfcc_parent;
1132		ipmr_update_thresholds(mrt, c, mfc->mfcc_ttls);
1133		if (!mrtsock)
1134			c->mfc_flags |= MFC_STATIC;
1135		write_unlock_bh(&mrt_lock);
 
 
1136		mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1137		return 0;
1138	}
1139
1140	if (mfc->mfcc_mcastgrp.s_addr != htonl(INADDR_ANY) &&
1141	    !ipv4_is_multicast(mfc->mfcc_mcastgrp.s_addr))
1142		return -EINVAL;
1143
1144	c = ipmr_cache_alloc();
1145	if (!c)
1146		return -ENOMEM;
1147
1148	c->mfc_origin = mfc->mfcc_origin.s_addr;
1149	c->mfc_mcastgrp = mfc->mfcc_mcastgrp.s_addr;
1150	c->mfc_parent = mfc->mfcc_parent;
1151	ipmr_update_thresholds(mrt, c, mfc->mfcc_ttls);
1152	if (!mrtsock)
1153		c->mfc_flags |= MFC_STATIC;
1154
1155	list_add_rcu(&c->list, &mrt->mfc_cache_array[line]);
1156
 
 
 
 
 
 
 
 
1157	/* Check to see if we resolved a queued list. If so we
1158	 * need to send on the frames and tidy up.
1159	 */
1160	found = false;
1161	spin_lock_bh(&mfc_unres_lock);
1162	list_for_each_entry(uc, &mrt->mfc_unres_queue, list) {
 
1163		if (uc->mfc_origin == c->mfc_origin &&
1164		    uc->mfc_mcastgrp == c->mfc_mcastgrp) {
1165			list_del(&uc->list);
1166			atomic_dec(&mrt->cache_resolve_queue_len);
1167			found = true;
1168			break;
1169		}
1170	}
1171	if (list_empty(&mrt->mfc_unres_queue))
1172		del_timer(&mrt->ipmr_expire_timer);
1173	spin_unlock_bh(&mfc_unres_lock);
1174
1175	if (found) {
1176		ipmr_cache_resolve(net, mrt, uc, c);
1177		ipmr_cache_free(uc);
1178	}
 
1179	mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1180	return 0;
1181}
1182
1183/* Close the multicast socket, and clear the vif tables etc */
1184static void mroute_clean_tables(struct mr_table *mrt, bool all)
1185{
 
 
 
 
1186	int i;
1187	LIST_HEAD(list);
1188	struct mfc_cache *c, *next;
1189
1190	/* Shut down all active vif entries */
1191	for (i = 0; i < mrt->maxvif; i++) {
1192		if (!all && (mrt->vif_table[i].flags & VIFF_STATIC))
1193			continue;
1194		vif_delete(mrt, i, 0, &list);
 
 
 
 
 
1195	}
1196	unregister_netdevice_many(&list);
1197
1198	/* Wipe the cache */
1199	for (i = 0; i < MFC_LINES; i++) {
1200		list_for_each_entry_safe(c, next, &mrt->mfc_cache_array[i], list) {
1201			if (!all && (c->mfc_flags & MFC_STATIC))
 
1202				continue;
 
1203			list_del_rcu(&c->list);
1204			mroute_netlink_event(mrt, c, RTM_DELROUTE);
1205			ipmr_cache_free(c);
 
 
 
1206		}
1207	}
1208
1209	if (atomic_read(&mrt->cache_resolve_queue_len) != 0) {
1210		spin_lock_bh(&mfc_unres_lock);
1211		list_for_each_entry_safe(c, next, &mrt->mfc_unres_queue, list) {
1212			list_del(&c->list);
1213			mroute_netlink_event(mrt, c, RTM_DELROUTE);
1214			ipmr_destroy_unres(mrt, c);
 
 
 
 
1215		}
1216		spin_unlock_bh(&mfc_unres_lock);
1217	}
1218}
1219
1220/* called from ip_ra_control(), before an RCU grace period,
1221 * we dont need to call synchronize_rcu() here
1222 */
1223static void mrtsock_destruct(struct sock *sk)
1224{
1225	struct net *net = sock_net(sk);
1226	struct mr_table *mrt;
1227
1228	rtnl_lock();
1229	ipmr_for_each_table(mrt, net) {
1230		if (sk == rtnl_dereference(mrt->mroute_sk)) {
1231			IPV4_DEVCONF_ALL(net, MC_FORWARDING)--;
1232			inet_netconf_notify_devconf(net, NETCONFA_MC_FORWARDING,
 
1233						    NETCONFA_IFINDEX_ALL,
1234						    net->ipv4.devconf_all);
1235			RCU_INIT_POINTER(mrt->mroute_sk, NULL);
1236			mroute_clean_tables(mrt, false);
1237		}
1238	}
1239	rtnl_unlock();
1240}
1241
1242/* Socket options and virtual interface manipulation. The whole
1243 * virtual interface system is a complete heap, but unfortunately
1244 * that's how BSD mrouted happens to think. Maybe one day with a proper
1245 * MOSPF/PIM router set up we can clean this up.
1246 */
1247
1248int ip_mroute_setsockopt(struct sock *sk, int optname, char __user *optval,
1249			 unsigned int optlen)
1250{
1251	struct net *net = sock_net(sk);
1252	int val, ret = 0, parent = 0;
1253	struct mr_table *mrt;
1254	struct vifctl vif;
1255	struct mfcctl mfc;
 
1256	u32 uval;
1257
1258	/* There's one exception to the lock - MRT_DONE which needs to unlock */
1259	rtnl_lock();
1260	if (sk->sk_type != SOCK_RAW ||
1261	    inet_sk(sk)->inet_num != IPPROTO_IGMP) {
1262		ret = -EOPNOTSUPP;
1263		goto out_unlock;
1264	}
1265
1266	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1267	if (!mrt) {
1268		ret = -ENOENT;
1269		goto out_unlock;
1270	}
1271	if (optname != MRT_INIT) {
1272		if (sk != rcu_access_pointer(mrt->mroute_sk) &&
1273		    !ns_capable(net->user_ns, CAP_NET_ADMIN)) {
1274			ret = -EACCES;
1275			goto out_unlock;
1276		}
1277	}
1278
1279	switch (optname) {
1280	case MRT_INIT:
1281		if (optlen != sizeof(int)) {
1282			ret = -EINVAL;
1283			break;
1284		}
1285		if (rtnl_dereference(mrt->mroute_sk)) {
1286			ret = -EADDRINUSE;
1287			break;
1288		}
1289
1290		ret = ip_ra_control(sk, 1, mrtsock_destruct);
1291		if (ret == 0) {
1292			rcu_assign_pointer(mrt->mroute_sk, sk);
1293			IPV4_DEVCONF_ALL(net, MC_FORWARDING)++;
1294			inet_netconf_notify_devconf(net, NETCONFA_MC_FORWARDING,
 
1295						    NETCONFA_IFINDEX_ALL,
1296						    net->ipv4.devconf_all);
1297		}
1298		break;
1299	case MRT_DONE:
1300		if (sk != rcu_access_pointer(mrt->mroute_sk)) {
1301			ret = -EACCES;
1302		} else {
1303			/* We need to unlock here because mrtsock_destruct takes
1304			 * care of rtnl itself and we can't change that due to
1305			 * the IP_ROUTER_ALERT setsockopt which runs without it.
1306			 */
1307			rtnl_unlock();
1308			ret = ip_ra_control(sk, 0, NULL);
1309			goto out;
1310		}
1311		break;
1312	case MRT_ADD_VIF:
1313	case MRT_DEL_VIF:
1314		if (optlen != sizeof(vif)) {
1315			ret = -EINVAL;
1316			break;
1317		}
1318		if (copy_from_user(&vif, optval, sizeof(vif))) {
1319			ret = -EFAULT;
1320			break;
1321		}
1322		if (vif.vifc_vifi >= MAXVIFS) {
1323			ret = -ENFILE;
1324			break;
1325		}
1326		if (optname == MRT_ADD_VIF) {
1327			ret = vif_add(net, mrt, &vif,
1328				      sk == rtnl_dereference(mrt->mroute_sk));
1329		} else {
1330			ret = vif_delete(mrt, vif.vifc_vifi, 0, NULL);
1331		}
1332		break;
1333	/* Manipulate the forwarding caches. These live
1334	 * in a sort of kernel/user symbiosis.
1335	 */
1336	case MRT_ADD_MFC:
1337	case MRT_DEL_MFC:
1338		parent = -1;
 
1339	case MRT_ADD_MFC_PROXY:
1340	case MRT_DEL_MFC_PROXY:
1341		if (optlen != sizeof(mfc)) {
1342			ret = -EINVAL;
1343			break;
1344		}
1345		if (copy_from_user(&mfc, optval, sizeof(mfc))) {
1346			ret = -EFAULT;
1347			break;
1348		}
1349		if (parent == 0)
1350			parent = mfc.mfcc_parent;
1351		if (optname == MRT_DEL_MFC || optname == MRT_DEL_MFC_PROXY)
1352			ret = ipmr_mfc_delete(mrt, &mfc, parent);
1353		else
1354			ret = ipmr_mfc_add(net, mrt, &mfc,
1355					   sk == rtnl_dereference(mrt->mroute_sk),
1356					   parent);
1357		break;
 
 
 
 
 
 
 
 
 
 
 
1358	/* Control PIM assert. */
1359	case MRT_ASSERT:
1360		if (optlen != sizeof(val)) {
1361			ret = -EINVAL;
1362			break;
1363		}
1364		if (get_user(val, (int __user *)optval)) {
1365			ret = -EFAULT;
1366			break;
1367		}
1368		mrt->mroute_do_assert = val;
1369		break;
1370	case MRT_PIM:
1371		if (!ipmr_pimsm_enabled()) {
1372			ret = -ENOPROTOOPT;
1373			break;
1374		}
1375		if (optlen != sizeof(val)) {
1376			ret = -EINVAL;
1377			break;
1378		}
1379		if (get_user(val, (int __user *)optval)) {
1380			ret = -EFAULT;
1381			break;
1382		}
1383
 
1384		val = !!val;
1385		if (val != mrt->mroute_do_pim) {
1386			mrt->mroute_do_pim = val;
1387			mrt->mroute_do_assert = val;
 
1388		}
1389		break;
1390	case MRT_TABLE:
1391		if (!IS_BUILTIN(CONFIG_IP_MROUTE_MULTIPLE_TABLES)) {
1392			ret = -ENOPROTOOPT;
1393			break;
1394		}
1395		if (optlen != sizeof(uval)) {
1396			ret = -EINVAL;
1397			break;
1398		}
1399		if (get_user(uval, (u32 __user *)optval)) {
1400			ret = -EFAULT;
1401			break;
1402		}
1403
1404		if (sk == rtnl_dereference(mrt->mroute_sk)) {
1405			ret = -EBUSY;
1406		} else {
1407			mrt = ipmr_new_table(net, uval);
1408			if (IS_ERR(mrt))
1409				ret = PTR_ERR(mrt);
1410			else
1411				raw_sk(sk)->ipmr_table = uval;
1412		}
1413		break;
1414	/* Spurious command, or MRT_VERSION which you cannot set. */
1415	default:
1416		ret = -ENOPROTOOPT;
1417	}
1418out_unlock:
1419	rtnl_unlock();
1420out:
1421	return ret;
1422}
1423
1424/* Getsock opt support for the multicast routing system. */
1425int ip_mroute_getsockopt(struct sock *sk, int optname, char __user *optval, int __user *optlen)
1426{
1427	int olr;
1428	int val;
1429	struct net *net = sock_net(sk);
1430	struct mr_table *mrt;
1431
1432	if (sk->sk_type != SOCK_RAW ||
1433	    inet_sk(sk)->inet_num != IPPROTO_IGMP)
1434		return -EOPNOTSUPP;
1435
1436	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1437	if (!mrt)
1438		return -ENOENT;
1439
1440	switch (optname) {
1441	case MRT_VERSION:
1442		val = 0x0305;
1443		break;
1444	case MRT_PIM:
1445		if (!ipmr_pimsm_enabled())
1446			return -ENOPROTOOPT;
1447		val = mrt->mroute_do_pim;
1448		break;
1449	case MRT_ASSERT:
1450		val = mrt->mroute_do_assert;
1451		break;
1452	default:
1453		return -ENOPROTOOPT;
1454	}
1455
1456	if (get_user(olr, optlen))
1457		return -EFAULT;
1458	olr = min_t(unsigned int, olr, sizeof(int));
1459	if (olr < 0)
1460		return -EINVAL;
1461	if (put_user(olr, optlen))
1462		return -EFAULT;
1463	if (copy_to_user(optval, &val, olr))
1464		return -EFAULT;
1465	return 0;
1466}
1467
1468/* The IP multicast ioctl support routines. */
1469int ipmr_ioctl(struct sock *sk, int cmd, void __user *arg)
1470{
1471	struct sioc_sg_req sr;
1472	struct sioc_vif_req vr;
1473	struct vif_device *vif;
1474	struct mfc_cache *c;
1475	struct net *net = sock_net(sk);
1476	struct mr_table *mrt;
1477
1478	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1479	if (!mrt)
1480		return -ENOENT;
1481
1482	switch (cmd) {
1483	case SIOCGETVIFCNT:
1484		if (copy_from_user(&vr, arg, sizeof(vr)))
1485			return -EFAULT;
1486		if (vr.vifi >= mrt->maxvif)
1487			return -EINVAL;
 
1488		read_lock(&mrt_lock);
1489		vif = &mrt->vif_table[vr.vifi];
1490		if (VIF_EXISTS(mrt, vr.vifi)) {
1491			vr.icount = vif->pkt_in;
1492			vr.ocount = vif->pkt_out;
1493			vr.ibytes = vif->bytes_in;
1494			vr.obytes = vif->bytes_out;
1495			read_unlock(&mrt_lock);
1496
1497			if (copy_to_user(arg, &vr, sizeof(vr)))
1498				return -EFAULT;
1499			return 0;
1500		}
1501		read_unlock(&mrt_lock);
1502		return -EADDRNOTAVAIL;
1503	case SIOCGETSGCNT:
1504		if (copy_from_user(&sr, arg, sizeof(sr)))
1505			return -EFAULT;
1506
1507		rcu_read_lock();
1508		c = ipmr_cache_find(mrt, sr.src.s_addr, sr.grp.s_addr);
1509		if (c) {
1510			sr.pktcnt = c->mfc_un.res.pkt;
1511			sr.bytecnt = c->mfc_un.res.bytes;
1512			sr.wrong_if = c->mfc_un.res.wrong_if;
1513			rcu_read_unlock();
1514
1515			if (copy_to_user(arg, &sr, sizeof(sr)))
1516				return -EFAULT;
1517			return 0;
1518		}
1519		rcu_read_unlock();
1520		return -EADDRNOTAVAIL;
1521	default:
1522		return -ENOIOCTLCMD;
1523	}
1524}
1525
1526#ifdef CONFIG_COMPAT
1527struct compat_sioc_sg_req {
1528	struct in_addr src;
1529	struct in_addr grp;
1530	compat_ulong_t pktcnt;
1531	compat_ulong_t bytecnt;
1532	compat_ulong_t wrong_if;
1533};
1534
1535struct compat_sioc_vif_req {
1536	vifi_t	vifi;		/* Which iface */
1537	compat_ulong_t icount;
1538	compat_ulong_t ocount;
1539	compat_ulong_t ibytes;
1540	compat_ulong_t obytes;
1541};
1542
1543int ipmr_compat_ioctl(struct sock *sk, unsigned int cmd, void __user *arg)
1544{
1545	struct compat_sioc_sg_req sr;
1546	struct compat_sioc_vif_req vr;
1547	struct vif_device *vif;
1548	struct mfc_cache *c;
1549	struct net *net = sock_net(sk);
1550	struct mr_table *mrt;
1551
1552	mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1553	if (!mrt)
1554		return -ENOENT;
1555
1556	switch (cmd) {
1557	case SIOCGETVIFCNT:
1558		if (copy_from_user(&vr, arg, sizeof(vr)))
1559			return -EFAULT;
1560		if (vr.vifi >= mrt->maxvif)
1561			return -EINVAL;
 
1562		read_lock(&mrt_lock);
1563		vif = &mrt->vif_table[vr.vifi];
1564		if (VIF_EXISTS(mrt, vr.vifi)) {
1565			vr.icount = vif->pkt_in;
1566			vr.ocount = vif->pkt_out;
1567			vr.ibytes = vif->bytes_in;
1568			vr.obytes = vif->bytes_out;
1569			read_unlock(&mrt_lock);
1570
1571			if (copy_to_user(arg, &vr, sizeof(vr)))
1572				return -EFAULT;
1573			return 0;
1574		}
1575		read_unlock(&mrt_lock);
1576		return -EADDRNOTAVAIL;
1577	case SIOCGETSGCNT:
1578		if (copy_from_user(&sr, arg, sizeof(sr)))
1579			return -EFAULT;
1580
1581		rcu_read_lock();
1582		c = ipmr_cache_find(mrt, sr.src.s_addr, sr.grp.s_addr);
1583		if (c) {
1584			sr.pktcnt = c->mfc_un.res.pkt;
1585			sr.bytecnt = c->mfc_un.res.bytes;
1586			sr.wrong_if = c->mfc_un.res.wrong_if;
1587			rcu_read_unlock();
1588
1589			if (copy_to_user(arg, &sr, sizeof(sr)))
1590				return -EFAULT;
1591			return 0;
1592		}
1593		rcu_read_unlock();
1594		return -EADDRNOTAVAIL;
1595	default:
1596		return -ENOIOCTLCMD;
1597	}
1598}
1599#endif
1600
1601static int ipmr_device_event(struct notifier_block *this, unsigned long event, void *ptr)
1602{
1603	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1604	struct net *net = dev_net(dev);
1605	struct mr_table *mrt;
1606	struct vif_device *v;
1607	int ct;
1608
1609	if (event != NETDEV_UNREGISTER)
1610		return NOTIFY_DONE;
1611
1612	ipmr_for_each_table(mrt, net) {
1613		v = &mrt->vif_table[0];
1614		for (ct = 0; ct < mrt->maxvif; ct++, v++) {
1615			if (v->dev == dev)
1616				vif_delete(mrt, ct, 1, NULL);
1617		}
1618	}
1619	return NOTIFY_DONE;
1620}
1621
1622static struct notifier_block ip_mr_notifier = {
1623	.notifier_call = ipmr_device_event,
1624};
1625
1626/* Encapsulate a packet by attaching a valid IPIP header to it.
1627 * This avoids tunnel drivers and other mess and gives us the speed so
1628 * important for multicast video.
1629 */
1630static void ip_encap(struct net *net, struct sk_buff *skb,
1631		     __be32 saddr, __be32 daddr)
1632{
1633	struct iphdr *iph;
1634	const struct iphdr *old_iph = ip_hdr(skb);
1635
1636	skb_push(skb, sizeof(struct iphdr));
1637	skb->transport_header = skb->network_header;
1638	skb_reset_network_header(skb);
1639	iph = ip_hdr(skb);
1640
1641	iph->version	=	4;
1642	iph->tos	=	old_iph->tos;
1643	iph->ttl	=	old_iph->ttl;
1644	iph->frag_off	=	0;
1645	iph->daddr	=	daddr;
1646	iph->saddr	=	saddr;
1647	iph->protocol	=	IPPROTO_IPIP;
1648	iph->ihl	=	5;
1649	iph->tot_len	=	htons(skb->len);
1650	ip_select_ident(net, skb, NULL);
1651	ip_send_check(iph);
1652
1653	memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
1654	nf_reset(skb);
1655}
1656
1657static inline int ipmr_forward_finish(struct net *net, struct sock *sk,
1658				      struct sk_buff *skb)
1659{
1660	struct ip_options *opt = &(IPCB(skb)->opt);
1661
1662	IP_INC_STATS(net, IPSTATS_MIB_OUTFORWDATAGRAMS);
1663	IP_ADD_STATS(net, IPSTATS_MIB_OUTOCTETS, skb->len);
1664
1665	if (unlikely(opt->optlen))
1666		ip_forward_options(skb);
1667
1668	return dst_output(net, sk, skb);
1669}
1670
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1671/* Processing handlers for ipmr_forward */
1672
1673static void ipmr_queue_xmit(struct net *net, struct mr_table *mrt,
1674			    struct sk_buff *skb, struct mfc_cache *c, int vifi)
1675{
1676	const struct iphdr *iph = ip_hdr(skb);
1677	struct vif_device *vif = &mrt->vif_table[vifi];
1678	struct net_device *dev;
1679	struct rtable *rt;
1680	struct flowi4 fl4;
1681	int    encap = 0;
1682
1683	if (!vif->dev)
1684		goto out_free;
1685
1686	if (vif->flags & VIFF_REGISTER) {
1687		vif->pkt_out++;
1688		vif->bytes_out += skb->len;
1689		vif->dev->stats.tx_bytes += skb->len;
1690		vif->dev->stats.tx_packets++;
1691		ipmr_cache_report(mrt, skb, vifi, IGMPMSG_WHOLEPKT);
1692		goto out_free;
1693	}
1694
 
 
 
1695	if (vif->flags & VIFF_TUNNEL) {
1696		rt = ip_route_output_ports(net, &fl4, NULL,
1697					   vif->remote, vif->local,
1698					   0, 0,
1699					   IPPROTO_IPIP,
1700					   RT_TOS(iph->tos), vif->link);
1701		if (IS_ERR(rt))
1702			goto out_free;
1703		encap = sizeof(struct iphdr);
1704	} else {
1705		rt = ip_route_output_ports(net, &fl4, NULL, iph->daddr, 0,
1706					   0, 0,
1707					   IPPROTO_IPIP,
1708					   RT_TOS(iph->tos), vif->link);
1709		if (IS_ERR(rt))
1710			goto out_free;
1711	}
1712
1713	dev = rt->dst.dev;
1714
1715	if (skb->len+encap > dst_mtu(&rt->dst) && (ntohs(iph->frag_off) & IP_DF)) {
1716		/* Do not fragment multicasts. Alas, IPv4 does not
1717		 * allow to send ICMP, so that packets will disappear
1718		 * to blackhole.
1719		 */
1720		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
1721		ip_rt_put(rt);
1722		goto out_free;
1723	}
1724
1725	encap += LL_RESERVED_SPACE(dev) + rt->dst.header_len;
1726
1727	if (skb_cow(skb, encap)) {
1728		ip_rt_put(rt);
1729		goto out_free;
1730	}
1731
1732	vif->pkt_out++;
1733	vif->bytes_out += skb->len;
1734
1735	skb_dst_drop(skb);
1736	skb_dst_set(skb, &rt->dst);
1737	ip_decrease_ttl(ip_hdr(skb));
1738
1739	/* FIXME: forward and output firewalls used to be called here.
1740	 * What do we do with netfilter? -- RR
1741	 */
1742	if (vif->flags & VIFF_TUNNEL) {
1743		ip_encap(net, skb, vif->local, vif->remote);
1744		/* FIXME: extra output firewall step used to be here. --RR */
1745		vif->dev->stats.tx_packets++;
1746		vif->dev->stats.tx_bytes += skb->len;
1747	}
1748
1749	IPCB(skb)->flags |= IPSKB_FORWARDED;
1750
1751	/* RFC1584 teaches, that DVMRP/PIM router must deliver packets locally
1752	 * not only before forwarding, but after forwarding on all output
1753	 * interfaces. It is clear, if mrouter runs a multicasting
1754	 * program, it should receive packets not depending to what interface
1755	 * program is joined.
1756	 * If we will not make it, the program will have to join on all
1757	 * interfaces. On the other hand, multihoming host (or router, but
1758	 * not mrouter) cannot join to more than one interface - it will
1759	 * result in receiving multiple packets.
1760	 */
1761	NF_HOOK(NFPROTO_IPV4, NF_INET_FORWARD,
1762		net, NULL, skb, skb->dev, dev,
1763		ipmr_forward_finish);
1764	return;
1765
1766out_free:
1767	kfree_skb(skb);
1768}
1769
1770static int ipmr_find_vif(struct mr_table *mrt, struct net_device *dev)
1771{
1772	int ct;
1773
1774	for (ct = mrt->maxvif-1; ct >= 0; ct--) {
1775		if (mrt->vif_table[ct].dev == dev)
1776			break;
1777	}
1778	return ct;
1779}
1780
1781/* "local" means that we should preserve one skb (for local delivery) */
1782static void ip_mr_forward(struct net *net, struct mr_table *mrt,
1783			  struct sk_buff *skb, struct mfc_cache *cache,
1784			  int local)
1785{
 
1786	int psend = -1;
1787	int vif, ct;
1788	int true_vifi = ipmr_find_vif(mrt, skb->dev);
1789
1790	vif = cache->mfc_parent;
1791	cache->mfc_un.res.pkt++;
1792	cache->mfc_un.res.bytes += skb->len;
 
1793
1794	if (cache->mfc_origin == htonl(INADDR_ANY) && true_vifi >= 0) {
1795		struct mfc_cache *cache_proxy;
1796
1797		/* For an (*,G) entry, we only check that the incomming
1798		 * interface is part of the static tree.
1799		 */
1800		cache_proxy = ipmr_cache_find_any_parent(mrt, vif);
1801		if (cache_proxy &&
1802		    cache_proxy->mfc_un.res.ttls[true_vifi] < 255)
1803			goto forward;
1804	}
1805
1806	/* Wrong interface: drop packet and (maybe) send PIM assert. */
1807	if (mrt->vif_table[vif].dev != skb->dev) {
1808		if (rt_is_output_route(skb_rtable(skb))) {
1809			/* It is our own packet, looped back.
1810			 * Very complicated situation...
1811			 *
1812			 * The best workaround until routing daemons will be
1813			 * fixed is not to redistribute packet, if it was
1814			 * send through wrong interface. It means, that
1815			 * multicast applications WILL NOT work for
1816			 * (S,G), which have default multicast route pointing
1817			 * to wrong oif. In any case, it is not a good
1818			 * idea to use multicasting applications on router.
1819			 */
1820			goto dont_forward;
1821		}
1822
1823		cache->mfc_un.res.wrong_if++;
1824
1825		if (true_vifi >= 0 && mrt->mroute_do_assert &&
1826		    /* pimsm uses asserts, when switching from RPT to SPT,
1827		     * so that we cannot check that packet arrived on an oif.
1828		     * It is bad, but otherwise we would need to move pretty
1829		     * large chunk of pimd to kernel. Ough... --ANK
1830		     */
1831		    (mrt->mroute_do_pim ||
1832		     cache->mfc_un.res.ttls[true_vifi] < 255) &&
1833		    time_after(jiffies,
1834			       cache->mfc_un.res.last_assert + MFC_ASSERT_THRESH)) {
1835			cache->mfc_un.res.last_assert = jiffies;
 
1836			ipmr_cache_report(mrt, skb, true_vifi, IGMPMSG_WRONGVIF);
 
 
 
1837		}
1838		goto dont_forward;
1839	}
1840
1841forward:
1842	mrt->vif_table[vif].pkt_in++;
1843	mrt->vif_table[vif].bytes_in += skb->len;
1844
1845	/* Forward the frame */
1846	if (cache->mfc_origin == htonl(INADDR_ANY) &&
1847	    cache->mfc_mcastgrp == htonl(INADDR_ANY)) {
1848		if (true_vifi >= 0 &&
1849		    true_vifi != cache->mfc_parent &&
1850		    ip_hdr(skb)->ttl >
1851				cache->mfc_un.res.ttls[cache->mfc_parent]) {
1852			/* It's an (*,*) entry and the packet is not coming from
1853			 * the upstream: forward the packet to the upstream
1854			 * only.
1855			 */
1856			psend = cache->mfc_parent;
1857			goto last_forward;
1858		}
1859		goto dont_forward;
1860	}
1861	for (ct = cache->mfc_un.res.maxvif - 1;
1862	     ct >= cache->mfc_un.res.minvif; ct--) {
1863		/* For (*,G) entry, don't forward to the incoming interface */
1864		if ((cache->mfc_origin != htonl(INADDR_ANY) ||
1865		     ct != true_vifi) &&
1866		    ip_hdr(skb)->ttl > cache->mfc_un.res.ttls[ct]) {
1867			if (psend != -1) {
1868				struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
1869
1870				if (skb2)
1871					ipmr_queue_xmit(net, mrt, skb2, cache,
1872							psend);
1873			}
1874			psend = ct;
1875		}
1876	}
1877last_forward:
1878	if (psend != -1) {
1879		if (local) {
1880			struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
1881
1882			if (skb2)
1883				ipmr_queue_xmit(net, mrt, skb2, cache, psend);
 
1884		} else {
1885			ipmr_queue_xmit(net, mrt, skb, cache, psend);
1886			return;
1887		}
1888	}
1889
1890dont_forward:
1891	if (!local)
1892		kfree_skb(skb);
1893}
1894
1895static struct mr_table *ipmr_rt_fib_lookup(struct net *net, struct sk_buff *skb)
1896{
1897	struct rtable *rt = skb_rtable(skb);
1898	struct iphdr *iph = ip_hdr(skb);
1899	struct flowi4 fl4 = {
1900		.daddr = iph->daddr,
1901		.saddr = iph->saddr,
1902		.flowi4_tos = RT_TOS(iph->tos),
1903		.flowi4_oif = (rt_is_output_route(rt) ?
1904			       skb->dev->ifindex : 0),
1905		.flowi4_iif = (rt_is_output_route(rt) ?
1906			       LOOPBACK_IFINDEX :
1907			       skb->dev->ifindex),
1908		.flowi4_mark = skb->mark,
1909	};
1910	struct mr_table *mrt;
1911	int err;
1912
1913	err = ipmr_fib_lookup(net, &fl4, &mrt);
1914	if (err)
1915		return ERR_PTR(err);
1916	return mrt;
1917}
1918
1919/* Multicast packets for forwarding arrive here
1920 * Called with rcu_read_lock();
1921 */
1922int ip_mr_input(struct sk_buff *skb)
1923{
1924	struct mfc_cache *cache;
1925	struct net *net = dev_net(skb->dev);
1926	int local = skb_rtable(skb)->rt_flags & RTCF_LOCAL;
1927	struct mr_table *mrt;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1928
1929	/* Packet is looped back after forward, it should not be
1930	 * forwarded second time, but still can be delivered locally.
1931	 */
1932	if (IPCB(skb)->flags & IPSKB_FORWARDED)
1933		goto dont_forward;
1934
1935	mrt = ipmr_rt_fib_lookup(net, skb);
1936	if (IS_ERR(mrt)) {
1937		kfree_skb(skb);
1938		return PTR_ERR(mrt);
1939	}
1940	if (!local) {
1941		if (IPCB(skb)->opt.router_alert) {
1942			if (ip_call_ra_chain(skb))
1943				return 0;
1944		} else if (ip_hdr(skb)->protocol == IPPROTO_IGMP) {
1945			/* IGMPv1 (and broken IGMPv2 implementations sort of
1946			 * Cisco IOS <= 11.2(8)) do not put router alert
1947			 * option to IGMP packets destined to routable
1948			 * groups. It is very bad, because it means
1949			 * that we can forward NO IGMP messages.
1950			 */
1951			struct sock *mroute_sk;
1952
1953			mroute_sk = rcu_dereference(mrt->mroute_sk);
1954			if (mroute_sk) {
1955				nf_reset(skb);
1956				raw_rcv(mroute_sk, skb);
1957				return 0;
1958			}
1959		    }
1960	}
1961
1962	/* already under rcu_read_lock() */
1963	cache = ipmr_cache_find(mrt, ip_hdr(skb)->saddr, ip_hdr(skb)->daddr);
1964	if (!cache) {
1965		int vif = ipmr_find_vif(mrt, skb->dev);
1966
1967		if (vif >= 0)
1968			cache = ipmr_cache_find_any(mrt, ip_hdr(skb)->daddr,
1969						    vif);
1970	}
1971
1972	/* No usable cache entry */
1973	if (!cache) {
1974		int vif;
1975
1976		if (local) {
1977			struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
1978			ip_local_deliver(skb);
1979			if (!skb2)
1980				return -ENOBUFS;
1981			skb = skb2;
1982		}
1983
1984		read_lock(&mrt_lock);
1985		vif = ipmr_find_vif(mrt, skb->dev);
1986		if (vif >= 0) {
1987			int err2 = ipmr_cache_unresolved(mrt, vif, skb);
1988			read_unlock(&mrt_lock);
1989
1990			return err2;
1991		}
1992		read_unlock(&mrt_lock);
1993		kfree_skb(skb);
1994		return -ENODEV;
1995	}
1996
1997	read_lock(&mrt_lock);
1998	ip_mr_forward(net, mrt, skb, cache, local);
1999	read_unlock(&mrt_lock);
2000
2001	if (local)
2002		return ip_local_deliver(skb);
2003
2004	return 0;
2005
2006dont_forward:
2007	if (local)
2008		return ip_local_deliver(skb);
2009	kfree_skb(skb);
2010	return 0;
2011}
2012
2013#ifdef CONFIG_IP_PIMSM_V1
2014/* Handle IGMP messages of PIMv1 */
2015int pim_rcv_v1(struct sk_buff *skb)
2016{
2017	struct igmphdr *pim;
2018	struct net *net = dev_net(skb->dev);
2019	struct mr_table *mrt;
2020
2021	if (!pskb_may_pull(skb, sizeof(*pim) + sizeof(struct iphdr)))
2022		goto drop;
2023
2024	pim = igmp_hdr(skb);
2025
2026	mrt = ipmr_rt_fib_lookup(net, skb);
2027	if (IS_ERR(mrt))
2028		goto drop;
2029	if (!mrt->mroute_do_pim ||
2030	    pim->group != PIM_V1_VERSION || pim->code != PIM_V1_REGISTER)
2031		goto drop;
2032
2033	if (__pim_rcv(mrt, skb, sizeof(*pim))) {
2034drop:
2035		kfree_skb(skb);
2036	}
2037	return 0;
2038}
2039#endif
2040
2041#ifdef CONFIG_IP_PIMSM_V2
2042static int pim_rcv(struct sk_buff *skb)
2043{
2044	struct pimreghdr *pim;
2045	struct net *net = dev_net(skb->dev);
2046	struct mr_table *mrt;
2047
2048	if (!pskb_may_pull(skb, sizeof(*pim) + sizeof(struct iphdr)))
2049		goto drop;
2050
2051	pim = (struct pimreghdr *)skb_transport_header(skb);
2052	if (pim->type != ((PIM_VERSION << 4) | (PIM_REGISTER)) ||
2053	    (pim->flags & PIM_NULL_REGISTER) ||
2054	    (ip_compute_csum((void *)pim, sizeof(*pim)) != 0 &&
2055	     csum_fold(skb_checksum(skb, 0, skb->len, 0))))
2056		goto drop;
2057
2058	mrt = ipmr_rt_fib_lookup(net, skb);
2059	if (IS_ERR(mrt))
2060		goto drop;
2061	if (__pim_rcv(mrt, skb, sizeof(*pim))) {
2062drop:
2063		kfree_skb(skb);
2064	}
2065	return 0;
2066}
2067#endif
2068
2069static int __ipmr_fill_mroute(struct mr_table *mrt, struct sk_buff *skb,
2070			      struct mfc_cache *c, struct rtmsg *rtm)
2071{
2072	int ct;
2073	struct rtnexthop *nhp;
2074	struct nlattr *mp_attr;
2075	struct rta_mfc_stats mfcs;
2076
2077	/* If cache is unresolved, don't try to parse IIF and OIF */
2078	if (c->mfc_parent >= MAXVIFS)
2079		return -ENOENT;
2080
2081	if (VIF_EXISTS(mrt, c->mfc_parent) &&
2082	    nla_put_u32(skb, RTA_IIF, mrt->vif_table[c->mfc_parent].dev->ifindex) < 0)
2083		return -EMSGSIZE;
2084
2085	if (!(mp_attr = nla_nest_start(skb, RTA_MULTIPATH)))
2086		return -EMSGSIZE;
2087
2088	for (ct = c->mfc_un.res.minvif; ct < c->mfc_un.res.maxvif; ct++) {
2089		if (VIF_EXISTS(mrt, ct) && c->mfc_un.res.ttls[ct] < 255) {
2090			if (!(nhp = nla_reserve_nohdr(skb, sizeof(*nhp)))) {
2091				nla_nest_cancel(skb, mp_attr);
2092				return -EMSGSIZE;
2093			}
2094
2095			nhp->rtnh_flags = 0;
2096			nhp->rtnh_hops = c->mfc_un.res.ttls[ct];
2097			nhp->rtnh_ifindex = mrt->vif_table[ct].dev->ifindex;
2098			nhp->rtnh_len = sizeof(*nhp);
2099		}
2100	}
2101
2102	nla_nest_end(skb, mp_attr);
2103
2104	mfcs.mfcs_packets = c->mfc_un.res.pkt;
2105	mfcs.mfcs_bytes = c->mfc_un.res.bytes;
2106	mfcs.mfcs_wrong_if = c->mfc_un.res.wrong_if;
2107	if (nla_put(skb, RTA_MFC_STATS, sizeof(mfcs), &mfcs) < 0)
2108		return -EMSGSIZE;
2109
2110	rtm->rtm_type = RTN_MULTICAST;
2111	return 1;
2112}
2113
2114int ipmr_get_route(struct net *net, struct sk_buff *skb,
2115		   __be32 saddr, __be32 daddr,
2116		   struct rtmsg *rtm, int nowait)
2117{
2118	struct mfc_cache *cache;
2119	struct mr_table *mrt;
2120	int err;
2121
2122	mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2123	if (!mrt)
2124		return -ENOENT;
2125
2126	rcu_read_lock();
2127	cache = ipmr_cache_find(mrt, saddr, daddr);
2128	if (!cache && skb->dev) {
2129		int vif = ipmr_find_vif(mrt, skb->dev);
2130
2131		if (vif >= 0)
2132			cache = ipmr_cache_find_any(mrt, daddr, vif);
2133	}
2134	if (!cache) {
2135		struct sk_buff *skb2;
2136		struct iphdr *iph;
2137		struct net_device *dev;
2138		int vif = -1;
2139
2140		if (nowait) {
2141			rcu_read_unlock();
2142			return -EAGAIN;
2143		}
2144
2145		dev = skb->dev;
2146		read_lock(&mrt_lock);
2147		if (dev)
2148			vif = ipmr_find_vif(mrt, dev);
2149		if (vif < 0) {
2150			read_unlock(&mrt_lock);
2151			rcu_read_unlock();
2152			return -ENODEV;
2153		}
2154		skb2 = skb_clone(skb, GFP_ATOMIC);
 
2155		if (!skb2) {
2156			read_unlock(&mrt_lock);
2157			rcu_read_unlock();
2158			return -ENOMEM;
2159		}
2160
 
2161		skb_push(skb2, sizeof(struct iphdr));
2162		skb_reset_network_header(skb2);
2163		iph = ip_hdr(skb2);
2164		iph->ihl = sizeof(struct iphdr) >> 2;
2165		iph->saddr = saddr;
2166		iph->daddr = daddr;
2167		iph->version = 0;
2168		err = ipmr_cache_unresolved(mrt, vif, skb2);
2169		read_unlock(&mrt_lock);
2170		rcu_read_unlock();
2171		return err;
2172	}
2173
2174	read_lock(&mrt_lock);
2175	err = __ipmr_fill_mroute(mrt, skb, cache, rtm);
2176	read_unlock(&mrt_lock);
2177	rcu_read_unlock();
2178	return err;
2179}
2180
2181static int ipmr_fill_mroute(struct mr_table *mrt, struct sk_buff *skb,
2182			    u32 portid, u32 seq, struct mfc_cache *c, int cmd,
2183			    int flags)
2184{
2185	struct nlmsghdr *nlh;
2186	struct rtmsg *rtm;
2187	int err;
2188
2189	nlh = nlmsg_put(skb, portid, seq, cmd, sizeof(*rtm), flags);
2190	if (!nlh)
2191		return -EMSGSIZE;
2192
2193	rtm = nlmsg_data(nlh);
2194	rtm->rtm_family   = RTNL_FAMILY_IPMR;
2195	rtm->rtm_dst_len  = 32;
2196	rtm->rtm_src_len  = 32;
2197	rtm->rtm_tos      = 0;
2198	rtm->rtm_table    = mrt->id;
2199	if (nla_put_u32(skb, RTA_TABLE, mrt->id))
2200		goto nla_put_failure;
2201	rtm->rtm_type     = RTN_MULTICAST;
2202	rtm->rtm_scope    = RT_SCOPE_UNIVERSE;
2203	if (c->mfc_flags & MFC_STATIC)
2204		rtm->rtm_protocol = RTPROT_STATIC;
2205	else
2206		rtm->rtm_protocol = RTPROT_MROUTED;
2207	rtm->rtm_flags    = 0;
2208
2209	if (nla_put_in_addr(skb, RTA_SRC, c->mfc_origin) ||
2210	    nla_put_in_addr(skb, RTA_DST, c->mfc_mcastgrp))
2211		goto nla_put_failure;
2212	err = __ipmr_fill_mroute(mrt, skb, c, rtm);
2213	/* do not break the dump if cache is unresolved */
2214	if (err < 0 && err != -ENOENT)
2215		goto nla_put_failure;
2216
2217	nlmsg_end(skb, nlh);
2218	return 0;
2219
2220nla_put_failure:
2221	nlmsg_cancel(skb, nlh);
2222	return -EMSGSIZE;
2223}
2224
 
 
 
 
 
 
 
 
2225static size_t mroute_msgsize(bool unresolved, int maxvif)
2226{
2227	size_t len =
2228		NLMSG_ALIGN(sizeof(struct rtmsg))
2229		+ nla_total_size(4)	/* RTA_TABLE */
2230		+ nla_total_size(4)	/* RTA_SRC */
2231		+ nla_total_size(4)	/* RTA_DST */
2232		;
2233
2234	if (!unresolved)
2235		len = len
2236		      + nla_total_size(4)	/* RTA_IIF */
2237		      + nla_total_size(0)	/* RTA_MULTIPATH */
2238		      + maxvif * NLA_ALIGN(sizeof(struct rtnexthop))
2239						/* RTA_MFC_STATS */
2240		      + nla_total_size(sizeof(struct rta_mfc_stats))
2241		;
2242
2243	return len;
2244}
2245
2246static void mroute_netlink_event(struct mr_table *mrt, struct mfc_cache *mfc,
2247				 int cmd)
2248{
2249	struct net *net = read_pnet(&mrt->net);
2250	struct sk_buff *skb;
2251	int err = -ENOBUFS;
2252
2253	skb = nlmsg_new(mroute_msgsize(mfc->mfc_parent >= MAXVIFS, mrt->maxvif),
 
2254			GFP_ATOMIC);
2255	if (!skb)
2256		goto errout;
2257
2258	err = ipmr_fill_mroute(mrt, skb, 0, 0, mfc, cmd, 0);
2259	if (err < 0)
2260		goto errout;
2261
2262	rtnl_notify(skb, net, 0, RTNLGRP_IPV4_MROUTE, NULL, GFP_ATOMIC);
2263	return;
2264
2265errout:
2266	kfree_skb(skb);
2267	if (err < 0)
2268		rtnl_set_sk_err(net, RTNLGRP_IPV4_MROUTE, err);
2269}
2270
2271static int ipmr_rtm_dumproute(struct sk_buff *skb, struct netlink_callback *cb)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2272{
2273	struct net *net = sock_net(skb->sk);
 
 
 
2274	struct mr_table *mrt;
2275	struct mfc_cache *mfc;
2276	unsigned int t = 0, s_t;
2277	unsigned int h = 0, s_h;
2278	unsigned int e = 0, s_e;
 
 
 
 
 
 
 
2279
2280	s_t = cb->args[0];
2281	s_h = cb->args[1];
2282	s_e = cb->args[2];
 
 
2283
 
2284	rcu_read_lock();
2285	ipmr_for_each_table(mrt, net) {
2286		if (t < s_t)
2287			goto next_table;
2288		if (t > s_t)
2289			s_h = 0;
2290		for (h = s_h; h < MFC_LINES; h++) {
2291			list_for_each_entry_rcu(mfc, &mrt->mfc_cache_array[h], list) {
2292				if (e < s_e)
2293					goto next_entry;
2294				if (ipmr_fill_mroute(mrt, skb,
2295						     NETLINK_CB(cb->skb).portid,
2296						     cb->nlh->nlmsg_seq,
2297						     mfc, RTM_NEWROUTE,
2298						     NLM_F_MULTI) < 0)
2299					goto done;
2300next_entry:
2301				e++;
2302			}
2303			e = s_e = 0;
2304		}
2305		spin_lock_bh(&mfc_unres_lock);
2306		list_for_each_entry(mfc, &mrt->mfc_unres_queue, list) {
2307			if (e < s_e)
2308				goto next_entry2;
2309			if (ipmr_fill_mroute(mrt, skb,
2310					     NETLINK_CB(cb->skb).portid,
2311					     cb->nlh->nlmsg_seq,
2312					     mfc, RTM_NEWROUTE,
2313					     NLM_F_MULTI) < 0) {
2314				spin_unlock_bh(&mfc_unres_lock);
2315				goto done;
2316			}
2317next_entry2:
2318			e++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2319		}
2320		spin_unlock_bh(&mfc_unres_lock);
2321		e = s_e = 0;
2322		s_h = 0;
2323next_table:
2324		t++;
2325	}
2326done:
2327	rcu_read_unlock();
2328
2329	cb->args[2] = e;
2330	cb->args[1] = h;
2331	cb->args[0] = t;
2332
2333	return skb->len;
 
2334}
2335
2336static const struct nla_policy rtm_ipmr_policy[RTA_MAX + 1] = {
2337	[RTA_SRC]	= { .type = NLA_U32 },
2338	[RTA_DST]	= { .type = NLA_U32 },
2339	[RTA_IIF]	= { .type = NLA_U32 },
2340	[RTA_TABLE]	= { .type = NLA_U32 },
2341	[RTA_MULTIPATH]	= { .len = sizeof(struct rtnexthop) },
2342};
2343
2344static bool ipmr_rtm_validate_proto(unsigned char rtm_protocol)
2345{
2346	switch (rtm_protocol) {
2347	case RTPROT_STATIC:
2348	case RTPROT_MROUTED:
2349		return true;
2350	}
2351	return false;
2352}
2353
2354static int ipmr_nla_get_ttls(const struct nlattr *nla, struct mfcctl *mfcc)
2355{
2356	struct rtnexthop *rtnh = nla_data(nla);
2357	int remaining = nla_len(nla), vifi = 0;
2358
2359	while (rtnh_ok(rtnh, remaining)) {
2360		mfcc->mfcc_ttls[vifi] = rtnh->rtnh_hops;
2361		if (++vifi == MAXVIFS)
2362			break;
2363		rtnh = rtnh_next(rtnh, &remaining);
2364	}
2365
2366	return remaining > 0 ? -EINVAL : vifi;
2367}
2368
2369/* returns < 0 on error, 0 for ADD_MFC and 1 for ADD_MFC_PROXY */
2370static int rtm_to_ipmr_mfcc(struct net *net, struct nlmsghdr *nlh,
2371			    struct mfcctl *mfcc, int *mrtsock,
2372			    struct mr_table **mrtret)
 
2373{
2374	struct net_device *dev = NULL;
2375	u32 tblid = RT_TABLE_DEFAULT;
2376	struct mr_table *mrt;
2377	struct nlattr *attr;
2378	struct rtmsg *rtm;
2379	int ret, rem;
2380
2381	ret = nlmsg_validate(nlh, sizeof(*rtm), RTA_MAX, rtm_ipmr_policy);
 
2382	if (ret < 0)
2383		goto out;
2384	rtm = nlmsg_data(nlh);
2385
2386	ret = -EINVAL;
2387	if (rtm->rtm_family != RTNL_FAMILY_IPMR || rtm->rtm_dst_len != 32 ||
2388	    rtm->rtm_type != RTN_MULTICAST ||
2389	    rtm->rtm_scope != RT_SCOPE_UNIVERSE ||
2390	    !ipmr_rtm_validate_proto(rtm->rtm_protocol))
2391		goto out;
2392
2393	memset(mfcc, 0, sizeof(*mfcc));
2394	mfcc->mfcc_parent = -1;
2395	ret = 0;
2396	nlmsg_for_each_attr(attr, nlh, sizeof(struct rtmsg), rem) {
2397		switch (nla_type(attr)) {
2398		case RTA_SRC:
2399			mfcc->mfcc_origin.s_addr = nla_get_be32(attr);
2400			break;
2401		case RTA_DST:
2402			mfcc->mfcc_mcastgrp.s_addr = nla_get_be32(attr);
2403			break;
2404		case RTA_IIF:
2405			dev = __dev_get_by_index(net, nla_get_u32(attr));
2406			if (!dev) {
2407				ret = -ENODEV;
2408				goto out;
2409			}
2410			break;
2411		case RTA_MULTIPATH:
2412			if (ipmr_nla_get_ttls(attr, mfcc) < 0) {
2413				ret = -EINVAL;
2414				goto out;
2415			}
2416			break;
2417		case RTA_PREFSRC:
2418			ret = 1;
2419			break;
2420		case RTA_TABLE:
2421			tblid = nla_get_u32(attr);
2422			break;
2423		}
2424	}
2425	mrt = ipmr_get_table(net, tblid);
2426	if (!mrt) {
2427		ret = -ENOENT;
2428		goto out;
2429	}
2430	*mrtret = mrt;
2431	*mrtsock = rtm->rtm_protocol == RTPROT_MROUTED ? 1 : 0;
2432	if (dev)
2433		mfcc->mfcc_parent = ipmr_find_vif(mrt, dev);
2434
2435out:
2436	return ret;
2437}
2438
2439/* takes care of both newroute and delroute */
2440static int ipmr_rtm_route(struct sk_buff *skb, struct nlmsghdr *nlh)
 
2441{
2442	struct net *net = sock_net(skb->sk);
2443	int ret, mrtsock, parent;
2444	struct mr_table *tbl;
2445	struct mfcctl mfcc;
2446
2447	mrtsock = 0;
2448	tbl = NULL;
2449	ret = rtm_to_ipmr_mfcc(net, nlh, &mfcc, &mrtsock, &tbl);
2450	if (ret < 0)
2451		return ret;
2452
2453	parent = ret ? mfcc.mfcc_parent : -1;
2454	if (nlh->nlmsg_type == RTM_NEWROUTE)
2455		return ipmr_mfc_add(net, tbl, &mfcc, mrtsock, parent);
2456	else
2457		return ipmr_mfc_delete(tbl, &mfcc, parent);
2458}
2459
2460#ifdef CONFIG_PROC_FS
2461/* The /proc interfaces to multicast routing :
2462 * /proc/net/ip_mr_cache & /proc/net/ip_mr_vif
2463 */
2464struct ipmr_vif_iter {
2465	struct seq_net_private p;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2466	struct mr_table *mrt;
2467	int ct;
2468};
2469
2470static struct vif_device *ipmr_vif_seq_idx(struct net *net,
2471					   struct ipmr_vif_iter *iter,
2472					   loff_t pos)
2473{
2474	struct mr_table *mrt = iter->mrt;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2475
2476	for (iter->ct = 0; iter->ct < mrt->maxvif; ++iter->ct) {
2477		if (!VIF_EXISTS(mrt, iter->ct))
2478			continue;
2479		if (pos-- == 0)
2480			return &mrt->vif_table[iter->ct];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2481	}
2482	return NULL;
 
 
 
 
 
2483}
2484
 
 
 
 
 
2485static void *ipmr_vif_seq_start(struct seq_file *seq, loff_t *pos)
2486	__acquires(mrt_lock)
2487{
2488	struct ipmr_vif_iter *iter = seq->private;
2489	struct net *net = seq_file_net(seq);
2490	struct mr_table *mrt;
2491
2492	mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2493	if (!mrt)
2494		return ERR_PTR(-ENOENT);
2495
2496	iter->mrt = mrt;
2497
2498	read_lock(&mrt_lock);
2499	return *pos ? ipmr_vif_seq_idx(net, seq->private, *pos - 1)
2500		: SEQ_START_TOKEN;
2501}
2502
2503static void *ipmr_vif_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2504{
2505	struct ipmr_vif_iter *iter = seq->private;
2506	struct net *net = seq_file_net(seq);
2507	struct mr_table *mrt = iter->mrt;
2508
2509	++*pos;
2510	if (v == SEQ_START_TOKEN)
2511		return ipmr_vif_seq_idx(net, iter, 0);
2512
2513	while (++iter->ct < mrt->maxvif) {
2514		if (!VIF_EXISTS(mrt, iter->ct))
2515			continue;
2516		return &mrt->vif_table[iter->ct];
2517	}
2518	return NULL;
2519}
2520
2521static void ipmr_vif_seq_stop(struct seq_file *seq, void *v)
2522	__releases(mrt_lock)
2523{
2524	read_unlock(&mrt_lock);
2525}
2526
2527static int ipmr_vif_seq_show(struct seq_file *seq, void *v)
2528{
2529	struct ipmr_vif_iter *iter = seq->private;
2530	struct mr_table *mrt = iter->mrt;
2531
2532	if (v == SEQ_START_TOKEN) {
2533		seq_puts(seq,
2534			 "Interface      BytesIn  PktsIn  BytesOut PktsOut Flags Local    Remote\n");
2535	} else {
2536		const struct vif_device *vif = v;
2537		const char *name =  vif->dev ? vif->dev->name : "none";
 
2538
2539		seq_printf(seq,
2540			   "%2Zd %-10s %8ld %7ld  %8ld %7ld %05X %08X %08X\n",
2541			   vif - mrt->vif_table,
2542			   name, vif->bytes_in, vif->pkt_in,
2543			   vif->bytes_out, vif->pkt_out,
2544			   vif->flags, vif->local, vif->remote);
2545	}
2546	return 0;
2547}
2548
2549static const struct seq_operations ipmr_vif_seq_ops = {
2550	.start = ipmr_vif_seq_start,
2551	.next  = ipmr_vif_seq_next,
2552	.stop  = ipmr_vif_seq_stop,
2553	.show  = ipmr_vif_seq_show,
2554};
2555
2556static int ipmr_vif_open(struct inode *inode, struct file *file)
2557{
2558	return seq_open_net(inode, file, &ipmr_vif_seq_ops,
2559			    sizeof(struct ipmr_vif_iter));
2560}
2561
2562static const struct file_operations ipmr_vif_fops = {
2563	.owner	 = THIS_MODULE,
2564	.open    = ipmr_vif_open,
2565	.read    = seq_read,
2566	.llseek  = seq_lseek,
2567	.release = seq_release_net,
2568};
2569
2570struct ipmr_mfc_iter {
2571	struct seq_net_private p;
2572	struct mr_table *mrt;
2573	struct list_head *cache;
2574	int ct;
2575};
2576
2577
2578static struct mfc_cache *ipmr_mfc_seq_idx(struct net *net,
2579					  struct ipmr_mfc_iter *it, loff_t pos)
2580{
2581	struct mr_table *mrt = it->mrt;
2582	struct mfc_cache *mfc;
2583
2584	rcu_read_lock();
2585	for (it->ct = 0; it->ct < MFC_LINES; it->ct++) {
2586		it->cache = &mrt->mfc_cache_array[it->ct];
2587		list_for_each_entry_rcu(mfc, it->cache, list)
2588			if (pos-- == 0)
2589				return mfc;
2590	}
2591	rcu_read_unlock();
2592
2593	spin_lock_bh(&mfc_unres_lock);
2594	it->cache = &mrt->mfc_unres_queue;
2595	list_for_each_entry(mfc, it->cache, list)
2596		if (pos-- == 0)
2597			return mfc;
2598	spin_unlock_bh(&mfc_unres_lock);
2599
2600	it->cache = NULL;
2601	return NULL;
2602}
2603
2604
2605static void *ipmr_mfc_seq_start(struct seq_file *seq, loff_t *pos)
2606{
2607	struct ipmr_mfc_iter *it = seq->private;
2608	struct net *net = seq_file_net(seq);
2609	struct mr_table *mrt;
2610
2611	mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2612	if (!mrt)
2613		return ERR_PTR(-ENOENT);
2614
2615	it->mrt = mrt;
2616	it->cache = NULL;
2617	it->ct = 0;
2618	return *pos ? ipmr_mfc_seq_idx(net, seq->private, *pos - 1)
2619		: SEQ_START_TOKEN;
2620}
2621
2622static void *ipmr_mfc_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2623{
2624	struct mfc_cache *mfc = v;
2625	struct ipmr_mfc_iter *it = seq->private;
2626	struct net *net = seq_file_net(seq);
2627	struct mr_table *mrt = it->mrt;
2628
2629	++*pos;
2630
2631	if (v == SEQ_START_TOKEN)
2632		return ipmr_mfc_seq_idx(net, seq->private, 0);
2633
2634	if (mfc->list.next != it->cache)
2635		return list_entry(mfc->list.next, struct mfc_cache, list);
2636
2637	if (it->cache == &mrt->mfc_unres_queue)
2638		goto end_of_list;
2639
2640	BUG_ON(it->cache != &mrt->mfc_cache_array[it->ct]);
2641
2642	while (++it->ct < MFC_LINES) {
2643		it->cache = &mrt->mfc_cache_array[it->ct];
2644		if (list_empty(it->cache))
2645			continue;
2646		return list_first_entry(it->cache, struct mfc_cache, list);
2647	}
2648
2649	/* exhausted cache_array, show unresolved */
2650	rcu_read_unlock();
2651	it->cache = &mrt->mfc_unres_queue;
2652	it->ct = 0;
2653
2654	spin_lock_bh(&mfc_unres_lock);
2655	if (!list_empty(it->cache))
2656		return list_first_entry(it->cache, struct mfc_cache, list);
2657
2658end_of_list:
2659	spin_unlock_bh(&mfc_unres_lock);
2660	it->cache = NULL;
2661
2662	return NULL;
2663}
2664
2665static void ipmr_mfc_seq_stop(struct seq_file *seq, void *v)
2666{
2667	struct ipmr_mfc_iter *it = seq->private;
2668	struct mr_table *mrt = it->mrt;
2669
2670	if (it->cache == &mrt->mfc_unres_queue)
2671		spin_unlock_bh(&mfc_unres_lock);
2672	else if (it->cache == &mrt->mfc_cache_array[it->ct])
2673		rcu_read_unlock();
2674}
2675
2676static int ipmr_mfc_seq_show(struct seq_file *seq, void *v)
2677{
2678	int n;
2679
2680	if (v == SEQ_START_TOKEN) {
2681		seq_puts(seq,
2682		 "Group    Origin   Iif     Pkts    Bytes    Wrong Oifs\n");
2683	} else {
2684		const struct mfc_cache *mfc = v;
2685		const struct ipmr_mfc_iter *it = seq->private;
2686		const struct mr_table *mrt = it->mrt;
2687
2688		seq_printf(seq, "%08X %08X %-3hd",
2689			   (__force u32) mfc->mfc_mcastgrp,
2690			   (__force u32) mfc->mfc_origin,
2691			   mfc->mfc_parent);
2692
2693		if (it->cache != &mrt->mfc_unres_queue) {
2694			seq_printf(seq, " %8lu %8lu %8lu",
2695				   mfc->mfc_un.res.pkt,
2696				   mfc->mfc_un.res.bytes,
2697				   mfc->mfc_un.res.wrong_if);
2698			for (n = mfc->mfc_un.res.minvif;
2699			     n < mfc->mfc_un.res.maxvif; n++) {
2700				if (VIF_EXISTS(mrt, n) &&
2701				    mfc->mfc_un.res.ttls[n] < 255)
2702					seq_printf(seq,
2703					   " %2d:%-3d",
2704					   n, mfc->mfc_un.res.ttls[n]);
2705			}
2706		} else {
2707			/* unresolved mfc_caches don't contain
2708			 * pkt, bytes and wrong_if values
2709			 */
2710			seq_printf(seq, " %8lu %8lu %8lu", 0ul, 0ul, 0ul);
2711		}
2712		seq_putc(seq, '\n');
2713	}
2714	return 0;
2715}
2716
2717static const struct seq_operations ipmr_mfc_seq_ops = {
2718	.start = ipmr_mfc_seq_start,
2719	.next  = ipmr_mfc_seq_next,
2720	.stop  = ipmr_mfc_seq_stop,
2721	.show  = ipmr_mfc_seq_show,
2722};
2723
2724static int ipmr_mfc_open(struct inode *inode, struct file *file)
2725{
2726	return seq_open_net(inode, file, &ipmr_mfc_seq_ops,
2727			    sizeof(struct ipmr_mfc_iter));
2728}
2729
2730static const struct file_operations ipmr_mfc_fops = {
2731	.owner	 = THIS_MODULE,
2732	.open    = ipmr_mfc_open,
2733	.read    = seq_read,
2734	.llseek  = seq_lseek,
2735	.release = seq_release_net,
2736};
2737#endif
2738
2739#ifdef CONFIG_IP_PIMSM_V2
2740static const struct net_protocol pim_protocol = {
2741	.handler	=	pim_rcv,
2742	.netns_ok	=	1,
2743};
2744#endif
2745
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2746/* Setup for IP multicast routing */
2747static int __net_init ipmr_net_init(struct net *net)
2748{
2749	int err;
2750
 
 
 
 
2751	err = ipmr_rules_init(net);
2752	if (err < 0)
2753		goto fail;
2754
2755#ifdef CONFIG_PROC_FS
2756	err = -ENOMEM;
2757	if (!proc_create("ip_mr_vif", 0, net->proc_net, &ipmr_vif_fops))
 
2758		goto proc_vif_fail;
2759	if (!proc_create("ip_mr_cache", 0, net->proc_net, &ipmr_mfc_fops))
 
2760		goto proc_cache_fail;
2761#endif
2762	return 0;
2763
2764#ifdef CONFIG_PROC_FS
2765proc_cache_fail:
2766	remove_proc_entry("ip_mr_vif", net->proc_net);
2767proc_vif_fail:
2768	ipmr_rules_exit(net);
2769#endif
2770fail:
 
 
2771	return err;
2772}
2773
2774static void __net_exit ipmr_net_exit(struct net *net)
2775{
2776#ifdef CONFIG_PROC_FS
2777	remove_proc_entry("ip_mr_cache", net->proc_net);
2778	remove_proc_entry("ip_mr_vif", net->proc_net);
2779#endif
 
2780	ipmr_rules_exit(net);
2781}
2782
2783static struct pernet_operations ipmr_net_ops = {
2784	.init = ipmr_net_init,
2785	.exit = ipmr_net_exit,
2786};
2787
2788int __init ip_mr_init(void)
2789{
2790	int err;
2791
2792	mrt_cachep = kmem_cache_create("ip_mrt_cache",
2793				       sizeof(struct mfc_cache),
2794				       0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
2795				       NULL);
2796
2797	err = register_pernet_subsys(&ipmr_net_ops);
2798	if (err)
2799		goto reg_pernet_fail;
2800
2801	err = register_netdevice_notifier(&ip_mr_notifier);
2802	if (err)
2803		goto reg_notif_fail;
2804#ifdef CONFIG_IP_PIMSM_V2
2805	if (inet_add_protocol(&pim_protocol, IPPROTO_PIM) < 0) {
2806		pr_err("%s: can't add PIM protocol\n", __func__);
2807		err = -EAGAIN;
2808		goto add_proto_fail;
2809	}
2810#endif
2811	rtnl_register(RTNL_FAMILY_IPMR, RTM_GETROUTE,
2812		      NULL, ipmr_rtm_dumproute, NULL);
2813	rtnl_register(RTNL_FAMILY_IPMR, RTM_NEWROUTE,
2814		      ipmr_rtm_route, NULL, NULL);
2815	rtnl_register(RTNL_FAMILY_IPMR, RTM_DELROUTE,
2816		      ipmr_rtm_route, NULL, NULL);
 
 
 
2817	return 0;
2818
2819#ifdef CONFIG_IP_PIMSM_V2
2820add_proto_fail:
2821	unregister_netdevice_notifier(&ip_mr_notifier);
2822#endif
2823reg_notif_fail:
2824	unregister_pernet_subsys(&ipmr_net_ops);
2825reg_pernet_fail:
2826	kmem_cache_destroy(mrt_cachep);
2827	return err;
2828}