Loading...
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 = ®_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}
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
115static struct mr_table *ipmr_mr_table_iter(struct net *net,
116 struct mr_table *mrt)
117{
118 struct mr_table *ret;
119
120 if (!mrt)
121 ret = list_entry_rcu(net->ipv4.mr_tables.next,
122 struct mr_table, list);
123 else
124 ret = list_entry_rcu(mrt->list.next,
125 struct mr_table, list);
126
127 if (&ret->list == &net->ipv4.mr_tables)
128 return NULL;
129 return ret;
130}
131
132static struct mr_table *ipmr_get_table(struct net *net, u32 id)
133{
134 struct mr_table *mrt;
135
136 ipmr_for_each_table(mrt, net) {
137 if (mrt->id == id)
138 return mrt;
139 }
140 return NULL;
141}
142
143static int ipmr_fib_lookup(struct net *net, struct flowi4 *flp4,
144 struct mr_table **mrt)
145{
146 int err;
147 struct ipmr_result res;
148 struct fib_lookup_arg arg = {
149 .result = &res,
150 .flags = FIB_LOOKUP_NOREF,
151 };
152
153 /* update flow if oif or iif point to device enslaved to l3mdev */
154 l3mdev_update_flow(net, flowi4_to_flowi(flp4));
155
156 err = fib_rules_lookup(net->ipv4.mr_rules_ops,
157 flowi4_to_flowi(flp4), 0, &arg);
158 if (err < 0)
159 return err;
160 *mrt = res.mrt;
161 return 0;
162}
163
164static int ipmr_rule_action(struct fib_rule *rule, struct flowi *flp,
165 int flags, struct fib_lookup_arg *arg)
166{
167 struct ipmr_result *res = arg->result;
168 struct mr_table *mrt;
169
170 switch (rule->action) {
171 case FR_ACT_TO_TBL:
172 break;
173 case FR_ACT_UNREACHABLE:
174 return -ENETUNREACH;
175 case FR_ACT_PROHIBIT:
176 return -EACCES;
177 case FR_ACT_BLACKHOLE:
178 default:
179 return -EINVAL;
180 }
181
182 arg->table = fib_rule_get_table(rule, arg);
183
184 mrt = ipmr_get_table(rule->fr_net, arg->table);
185 if (!mrt)
186 return -EAGAIN;
187 res->mrt = mrt;
188 return 0;
189}
190
191static int ipmr_rule_match(struct fib_rule *rule, struct flowi *fl, int flags)
192{
193 return 1;
194}
195
196static const struct nla_policy ipmr_rule_policy[FRA_MAX + 1] = {
197 FRA_GENERIC_POLICY,
198};
199
200static int ipmr_rule_configure(struct fib_rule *rule, struct sk_buff *skb,
201 struct fib_rule_hdr *frh, struct nlattr **tb,
202 struct netlink_ext_ack *extack)
203{
204 return 0;
205}
206
207static int ipmr_rule_compare(struct fib_rule *rule, struct fib_rule_hdr *frh,
208 struct nlattr **tb)
209{
210 return 1;
211}
212
213static int ipmr_rule_fill(struct fib_rule *rule, struct sk_buff *skb,
214 struct fib_rule_hdr *frh)
215{
216 frh->dst_len = 0;
217 frh->src_len = 0;
218 frh->tos = 0;
219 return 0;
220}
221
222static const struct fib_rules_ops __net_initconst ipmr_rules_ops_template = {
223 .family = RTNL_FAMILY_IPMR,
224 .rule_size = sizeof(struct ipmr_rule),
225 .addr_size = sizeof(u32),
226 .action = ipmr_rule_action,
227 .match = ipmr_rule_match,
228 .configure = ipmr_rule_configure,
229 .compare = ipmr_rule_compare,
230 .fill = ipmr_rule_fill,
231 .nlgroup = RTNLGRP_IPV4_RULE,
232 .policy = ipmr_rule_policy,
233 .owner = THIS_MODULE,
234};
235
236static int __net_init ipmr_rules_init(struct net *net)
237{
238 struct fib_rules_ops *ops;
239 struct mr_table *mrt;
240 int err;
241
242 ops = fib_rules_register(&ipmr_rules_ops_template, net);
243 if (IS_ERR(ops))
244 return PTR_ERR(ops);
245
246 INIT_LIST_HEAD(&net->ipv4.mr_tables);
247
248 mrt = ipmr_new_table(net, RT_TABLE_DEFAULT);
249 if (IS_ERR(mrt)) {
250 err = PTR_ERR(mrt);
251 goto err1;
252 }
253
254 err = fib_default_rule_add(ops, 0x7fff, RT_TABLE_DEFAULT, 0);
255 if (err < 0)
256 goto err2;
257
258 net->ipv4.mr_rules_ops = ops;
259 return 0;
260
261err2:
262 ipmr_free_table(mrt);
263err1:
264 fib_rules_unregister(ops);
265 return err;
266}
267
268static void __net_exit ipmr_rules_exit(struct net *net)
269{
270 struct mr_table *mrt, *next;
271
272 rtnl_lock();
273 list_for_each_entry_safe(mrt, next, &net->ipv4.mr_tables, list) {
274 list_del(&mrt->list);
275 ipmr_free_table(mrt);
276 }
277 fib_rules_unregister(net->ipv4.mr_rules_ops);
278 rtnl_unlock();
279}
280
281static int ipmr_rules_dump(struct net *net, struct notifier_block *nb)
282{
283 return fib_rules_dump(net, nb, RTNL_FAMILY_IPMR);
284}
285
286static unsigned int ipmr_rules_seq_read(struct net *net)
287{
288 return fib_rules_seq_read(net, RTNL_FAMILY_IPMR);
289}
290
291bool ipmr_rule_default(const struct fib_rule *rule)
292{
293 return fib_rule_matchall(rule) && rule->table == RT_TABLE_DEFAULT;
294}
295EXPORT_SYMBOL(ipmr_rule_default);
296#else
297#define ipmr_for_each_table(mrt, net) \
298 for (mrt = net->ipv4.mrt; mrt; mrt = NULL)
299
300static struct mr_table *ipmr_mr_table_iter(struct net *net,
301 struct mr_table *mrt)
302{
303 if (!mrt)
304 return net->ipv4.mrt;
305 return NULL;
306}
307
308static struct mr_table *ipmr_get_table(struct net *net, u32 id)
309{
310 return net->ipv4.mrt;
311}
312
313static int ipmr_fib_lookup(struct net *net, struct flowi4 *flp4,
314 struct mr_table **mrt)
315{
316 *mrt = net->ipv4.mrt;
317 return 0;
318}
319
320static int __net_init ipmr_rules_init(struct net *net)
321{
322 struct mr_table *mrt;
323
324 mrt = ipmr_new_table(net, RT_TABLE_DEFAULT);
325 if (IS_ERR(mrt))
326 return PTR_ERR(mrt);
327 net->ipv4.mrt = mrt;
328 return 0;
329}
330
331static void __net_exit ipmr_rules_exit(struct net *net)
332{
333 rtnl_lock();
334 ipmr_free_table(net->ipv4.mrt);
335 net->ipv4.mrt = NULL;
336 rtnl_unlock();
337}
338
339static int ipmr_rules_dump(struct net *net, struct notifier_block *nb)
340{
341 return 0;
342}
343
344static unsigned int ipmr_rules_seq_read(struct net *net)
345{
346 return 0;
347}
348
349bool ipmr_rule_default(const struct fib_rule *rule)
350{
351 return true;
352}
353EXPORT_SYMBOL(ipmr_rule_default);
354#endif
355
356static inline int ipmr_hash_cmp(struct rhashtable_compare_arg *arg,
357 const void *ptr)
358{
359 const struct mfc_cache_cmp_arg *cmparg = arg->key;
360 struct mfc_cache *c = (struct mfc_cache *)ptr;
361
362 return cmparg->mfc_mcastgrp != c->mfc_mcastgrp ||
363 cmparg->mfc_origin != c->mfc_origin;
364}
365
366static const struct rhashtable_params ipmr_rht_params = {
367 .head_offset = offsetof(struct mr_mfc, mnode),
368 .key_offset = offsetof(struct mfc_cache, cmparg),
369 .key_len = sizeof(struct mfc_cache_cmp_arg),
370 .nelem_hint = 3,
371 .obj_cmpfn = ipmr_hash_cmp,
372 .automatic_shrinking = true,
373};
374
375static void ipmr_new_table_set(struct mr_table *mrt,
376 struct net *net)
377{
378#ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
379 list_add_tail_rcu(&mrt->list, &net->ipv4.mr_tables);
380#endif
381}
382
383static struct mfc_cache_cmp_arg ipmr_mr_table_ops_cmparg_any = {
384 .mfc_mcastgrp = htonl(INADDR_ANY),
385 .mfc_origin = htonl(INADDR_ANY),
386};
387
388static struct mr_table_ops ipmr_mr_table_ops = {
389 .rht_params = &ipmr_rht_params,
390 .cmparg_any = &ipmr_mr_table_ops_cmparg_any,
391};
392
393static struct mr_table *ipmr_new_table(struct net *net, u32 id)
394{
395 struct mr_table *mrt;
396
397 /* "pimreg%u" should not exceed 16 bytes (IFNAMSIZ) */
398 if (id != RT_TABLE_DEFAULT && id >= 1000000000)
399 return ERR_PTR(-EINVAL);
400
401 mrt = ipmr_get_table(net, id);
402 if (mrt)
403 return mrt;
404
405 return mr_table_alloc(net, id, &ipmr_mr_table_ops,
406 ipmr_expire_process, ipmr_new_table_set);
407}
408
409static void ipmr_free_table(struct mr_table *mrt)
410{
411 del_timer_sync(&mrt->ipmr_expire_timer);
412 mroute_clean_tables(mrt, MRT_FLUSH_VIFS | MRT_FLUSH_VIFS_STATIC |
413 MRT_FLUSH_MFC | MRT_FLUSH_MFC_STATIC);
414 rhltable_destroy(&mrt->mfc_hash);
415 kfree(mrt);
416}
417
418/* Service routines creating virtual interfaces: DVMRP tunnels and PIMREG */
419
420static void ipmr_del_tunnel(struct net_device *dev, struct vifctl *v)
421{
422 struct net *net = dev_net(dev);
423
424 dev_close(dev);
425
426 dev = __dev_get_by_name(net, "tunl0");
427 if (dev) {
428 const struct net_device_ops *ops = dev->netdev_ops;
429 struct ifreq ifr;
430 struct ip_tunnel_parm p;
431
432 memset(&p, 0, sizeof(p));
433 p.iph.daddr = v->vifc_rmt_addr.s_addr;
434 p.iph.saddr = v->vifc_lcl_addr.s_addr;
435 p.iph.version = 4;
436 p.iph.ihl = 5;
437 p.iph.protocol = IPPROTO_IPIP;
438 sprintf(p.name, "dvmrp%d", v->vifc_vifi);
439 ifr.ifr_ifru.ifru_data = (__force void __user *)&p;
440
441 if (ops->ndo_do_ioctl) {
442 mm_segment_t oldfs = get_fs();
443
444 set_fs(KERNEL_DS);
445 ops->ndo_do_ioctl(dev, &ifr, SIOCDELTUNNEL);
446 set_fs(oldfs);
447 }
448 }
449}
450
451/* Initialize ipmr pimreg/tunnel in_device */
452static bool ipmr_init_vif_indev(const struct net_device *dev)
453{
454 struct in_device *in_dev;
455
456 ASSERT_RTNL();
457
458 in_dev = __in_dev_get_rtnl(dev);
459 if (!in_dev)
460 return false;
461 ipv4_devconf_setall(in_dev);
462 neigh_parms_data_state_setall(in_dev->arp_parms);
463 IPV4_DEVCONF(in_dev->cnf, RP_FILTER) = 0;
464
465 return true;
466}
467
468static struct net_device *ipmr_new_tunnel(struct net *net, struct vifctl *v)
469{
470 struct net_device *dev;
471
472 dev = __dev_get_by_name(net, "tunl0");
473
474 if (dev) {
475 const struct net_device_ops *ops = dev->netdev_ops;
476 int err;
477 struct ifreq ifr;
478 struct ip_tunnel_parm p;
479
480 memset(&p, 0, sizeof(p));
481 p.iph.daddr = v->vifc_rmt_addr.s_addr;
482 p.iph.saddr = v->vifc_lcl_addr.s_addr;
483 p.iph.version = 4;
484 p.iph.ihl = 5;
485 p.iph.protocol = IPPROTO_IPIP;
486 sprintf(p.name, "dvmrp%d", v->vifc_vifi);
487 ifr.ifr_ifru.ifru_data = (__force void __user *)&p;
488
489 if (ops->ndo_do_ioctl) {
490 mm_segment_t oldfs = get_fs();
491
492 set_fs(KERNEL_DS);
493 err = ops->ndo_do_ioctl(dev, &ifr, SIOCADDTUNNEL);
494 set_fs(oldfs);
495 } else {
496 err = -EOPNOTSUPP;
497 }
498 dev = NULL;
499
500 if (err == 0 &&
501 (dev = __dev_get_by_name(net, p.name)) != NULL) {
502 dev->flags |= IFF_MULTICAST;
503 if (!ipmr_init_vif_indev(dev))
504 goto failure;
505 if (dev_open(dev, NULL))
506 goto failure;
507 dev_hold(dev);
508 }
509 }
510 return dev;
511
512failure:
513 unregister_netdevice(dev);
514 return NULL;
515}
516
517#if defined(CONFIG_IP_PIMSM_V1) || defined(CONFIG_IP_PIMSM_V2)
518static netdev_tx_t reg_vif_xmit(struct sk_buff *skb, struct net_device *dev)
519{
520 struct net *net = dev_net(dev);
521 struct mr_table *mrt;
522 struct flowi4 fl4 = {
523 .flowi4_oif = dev->ifindex,
524 .flowi4_iif = skb->skb_iif ? : LOOPBACK_IFINDEX,
525 .flowi4_mark = skb->mark,
526 };
527 int err;
528
529 err = ipmr_fib_lookup(net, &fl4, &mrt);
530 if (err < 0) {
531 kfree_skb(skb);
532 return err;
533 }
534
535 read_lock(&mrt_lock);
536 dev->stats.tx_bytes += skb->len;
537 dev->stats.tx_packets++;
538 ipmr_cache_report(mrt, skb, mrt->mroute_reg_vif_num, IGMPMSG_WHOLEPKT);
539 read_unlock(&mrt_lock);
540 kfree_skb(skb);
541 return NETDEV_TX_OK;
542}
543
544static int reg_vif_get_iflink(const struct net_device *dev)
545{
546 return 0;
547}
548
549static const struct net_device_ops reg_vif_netdev_ops = {
550 .ndo_start_xmit = reg_vif_xmit,
551 .ndo_get_iflink = reg_vif_get_iflink,
552};
553
554static void reg_vif_setup(struct net_device *dev)
555{
556 dev->type = ARPHRD_PIMREG;
557 dev->mtu = ETH_DATA_LEN - sizeof(struct iphdr) - 8;
558 dev->flags = IFF_NOARP;
559 dev->netdev_ops = ®_vif_netdev_ops;
560 dev->needs_free_netdev = true;
561 dev->features |= NETIF_F_NETNS_LOCAL;
562}
563
564static struct net_device *ipmr_reg_vif(struct net *net, struct mr_table *mrt)
565{
566 struct net_device *dev;
567 char name[IFNAMSIZ];
568
569 if (mrt->id == RT_TABLE_DEFAULT)
570 sprintf(name, "pimreg");
571 else
572 sprintf(name, "pimreg%u", mrt->id);
573
574 dev = alloc_netdev(0, name, NET_NAME_UNKNOWN, reg_vif_setup);
575
576 if (!dev)
577 return NULL;
578
579 dev_net_set(dev, net);
580
581 if (register_netdevice(dev)) {
582 free_netdev(dev);
583 return NULL;
584 }
585
586 if (!ipmr_init_vif_indev(dev))
587 goto failure;
588 if (dev_open(dev, NULL))
589 goto failure;
590
591 dev_hold(dev);
592
593 return dev;
594
595failure:
596 unregister_netdevice(dev);
597 return NULL;
598}
599
600/* called with rcu_read_lock() */
601static int __pim_rcv(struct mr_table *mrt, struct sk_buff *skb,
602 unsigned int pimlen)
603{
604 struct net_device *reg_dev = NULL;
605 struct iphdr *encap;
606
607 encap = (struct iphdr *)(skb_transport_header(skb) + pimlen);
608 /* Check that:
609 * a. packet is really sent to a multicast group
610 * b. packet is not a NULL-REGISTER
611 * c. packet is not truncated
612 */
613 if (!ipv4_is_multicast(encap->daddr) ||
614 encap->tot_len == 0 ||
615 ntohs(encap->tot_len) + pimlen > skb->len)
616 return 1;
617
618 read_lock(&mrt_lock);
619 if (mrt->mroute_reg_vif_num >= 0)
620 reg_dev = mrt->vif_table[mrt->mroute_reg_vif_num].dev;
621 read_unlock(&mrt_lock);
622
623 if (!reg_dev)
624 return 1;
625
626 skb->mac_header = skb->network_header;
627 skb_pull(skb, (u8 *)encap - skb->data);
628 skb_reset_network_header(skb);
629 skb->protocol = htons(ETH_P_IP);
630 skb->ip_summed = CHECKSUM_NONE;
631
632 skb_tunnel_rx(skb, reg_dev, dev_net(reg_dev));
633
634 netif_rx(skb);
635
636 return NET_RX_SUCCESS;
637}
638#else
639static struct net_device *ipmr_reg_vif(struct net *net, struct mr_table *mrt)
640{
641 return NULL;
642}
643#endif
644
645static int call_ipmr_vif_entry_notifiers(struct net *net,
646 enum fib_event_type event_type,
647 struct vif_device *vif,
648 vifi_t vif_index, u32 tb_id)
649{
650 return mr_call_vif_notifiers(net, RTNL_FAMILY_IPMR, event_type,
651 vif, vif_index, tb_id,
652 &net->ipv4.ipmr_seq);
653}
654
655static int call_ipmr_mfc_entry_notifiers(struct net *net,
656 enum fib_event_type event_type,
657 struct mfc_cache *mfc, u32 tb_id)
658{
659 return mr_call_mfc_notifiers(net, RTNL_FAMILY_IPMR, event_type,
660 &mfc->_c, tb_id, &net->ipv4.ipmr_seq);
661}
662
663/**
664 * vif_delete - Delete a VIF entry
665 * @notify: Set to 1, if the caller is a notifier_call
666 */
667static int vif_delete(struct mr_table *mrt, int vifi, int notify,
668 struct list_head *head)
669{
670 struct net *net = read_pnet(&mrt->net);
671 struct vif_device *v;
672 struct net_device *dev;
673 struct in_device *in_dev;
674
675 if (vifi < 0 || vifi >= mrt->maxvif)
676 return -EADDRNOTAVAIL;
677
678 v = &mrt->vif_table[vifi];
679
680 if (VIF_EXISTS(mrt, vifi))
681 call_ipmr_vif_entry_notifiers(net, FIB_EVENT_VIF_DEL, v, vifi,
682 mrt->id);
683
684 write_lock_bh(&mrt_lock);
685 dev = v->dev;
686 v->dev = NULL;
687
688 if (!dev) {
689 write_unlock_bh(&mrt_lock);
690 return -EADDRNOTAVAIL;
691 }
692
693 if (vifi == mrt->mroute_reg_vif_num)
694 mrt->mroute_reg_vif_num = -1;
695
696 if (vifi + 1 == mrt->maxvif) {
697 int tmp;
698
699 for (tmp = vifi - 1; tmp >= 0; tmp--) {
700 if (VIF_EXISTS(mrt, tmp))
701 break;
702 }
703 mrt->maxvif = tmp+1;
704 }
705
706 write_unlock_bh(&mrt_lock);
707
708 dev_set_allmulti(dev, -1);
709
710 in_dev = __in_dev_get_rtnl(dev);
711 if (in_dev) {
712 IPV4_DEVCONF(in_dev->cnf, MC_FORWARDING)--;
713 inet_netconf_notify_devconf(dev_net(dev), RTM_NEWNETCONF,
714 NETCONFA_MC_FORWARDING,
715 dev->ifindex, &in_dev->cnf);
716 ip_rt_multicast_event(in_dev);
717 }
718
719 if (v->flags & (VIFF_TUNNEL | VIFF_REGISTER) && !notify)
720 unregister_netdevice_queue(dev, head);
721
722 dev_put(dev);
723 return 0;
724}
725
726static void ipmr_cache_free_rcu(struct rcu_head *head)
727{
728 struct mr_mfc *c = container_of(head, struct mr_mfc, rcu);
729
730 kmem_cache_free(mrt_cachep, (struct mfc_cache *)c);
731}
732
733static void ipmr_cache_free(struct mfc_cache *c)
734{
735 call_rcu(&c->_c.rcu, ipmr_cache_free_rcu);
736}
737
738/* Destroy an unresolved cache entry, killing queued skbs
739 * and reporting error to netlink readers.
740 */
741static void ipmr_destroy_unres(struct mr_table *mrt, struct mfc_cache *c)
742{
743 struct net *net = read_pnet(&mrt->net);
744 struct sk_buff *skb;
745 struct nlmsgerr *e;
746
747 atomic_dec(&mrt->cache_resolve_queue_len);
748
749 while ((skb = skb_dequeue(&c->_c.mfc_un.unres.unresolved))) {
750 if (ip_hdr(skb)->version == 0) {
751 struct nlmsghdr *nlh = skb_pull(skb,
752 sizeof(struct iphdr));
753 nlh->nlmsg_type = NLMSG_ERROR;
754 nlh->nlmsg_len = nlmsg_msg_size(sizeof(struct nlmsgerr));
755 skb_trim(skb, nlh->nlmsg_len);
756 e = nlmsg_data(nlh);
757 e->error = -ETIMEDOUT;
758 memset(&e->msg, 0, sizeof(e->msg));
759
760 rtnl_unicast(skb, net, NETLINK_CB(skb).portid);
761 } else {
762 kfree_skb(skb);
763 }
764 }
765
766 ipmr_cache_free(c);
767}
768
769/* Timer process for the unresolved queue. */
770static void ipmr_expire_process(struct timer_list *t)
771{
772 struct mr_table *mrt = from_timer(mrt, t, ipmr_expire_timer);
773 struct mr_mfc *c, *next;
774 unsigned long expires;
775 unsigned long now;
776
777 if (!spin_trylock(&mfc_unres_lock)) {
778 mod_timer(&mrt->ipmr_expire_timer, jiffies+HZ/10);
779 return;
780 }
781
782 if (list_empty(&mrt->mfc_unres_queue))
783 goto out;
784
785 now = jiffies;
786 expires = 10*HZ;
787
788 list_for_each_entry_safe(c, next, &mrt->mfc_unres_queue, list) {
789 if (time_after(c->mfc_un.unres.expires, now)) {
790 unsigned long interval = c->mfc_un.unres.expires - now;
791 if (interval < expires)
792 expires = interval;
793 continue;
794 }
795
796 list_del(&c->list);
797 mroute_netlink_event(mrt, (struct mfc_cache *)c, RTM_DELROUTE);
798 ipmr_destroy_unres(mrt, (struct mfc_cache *)c);
799 }
800
801 if (!list_empty(&mrt->mfc_unres_queue))
802 mod_timer(&mrt->ipmr_expire_timer, jiffies + expires);
803
804out:
805 spin_unlock(&mfc_unres_lock);
806}
807
808/* Fill oifs list. It is called under write locked mrt_lock. */
809static void ipmr_update_thresholds(struct mr_table *mrt, struct mr_mfc *cache,
810 unsigned char *ttls)
811{
812 int vifi;
813
814 cache->mfc_un.res.minvif = MAXVIFS;
815 cache->mfc_un.res.maxvif = 0;
816 memset(cache->mfc_un.res.ttls, 255, MAXVIFS);
817
818 for (vifi = 0; vifi < mrt->maxvif; vifi++) {
819 if (VIF_EXISTS(mrt, vifi) &&
820 ttls[vifi] && ttls[vifi] < 255) {
821 cache->mfc_un.res.ttls[vifi] = ttls[vifi];
822 if (cache->mfc_un.res.minvif > vifi)
823 cache->mfc_un.res.minvif = vifi;
824 if (cache->mfc_un.res.maxvif <= vifi)
825 cache->mfc_un.res.maxvif = vifi + 1;
826 }
827 }
828 cache->mfc_un.res.lastuse = jiffies;
829}
830
831static int vif_add(struct net *net, struct mr_table *mrt,
832 struct vifctl *vifc, int mrtsock)
833{
834 struct netdev_phys_item_id ppid = { };
835 int vifi = vifc->vifc_vifi;
836 struct vif_device *v = &mrt->vif_table[vifi];
837 struct net_device *dev;
838 struct in_device *in_dev;
839 int err;
840
841 /* Is vif busy ? */
842 if (VIF_EXISTS(mrt, vifi))
843 return -EADDRINUSE;
844
845 switch (vifc->vifc_flags) {
846 case VIFF_REGISTER:
847 if (!ipmr_pimsm_enabled())
848 return -EINVAL;
849 /* Special Purpose VIF in PIM
850 * All the packets will be sent to the daemon
851 */
852 if (mrt->mroute_reg_vif_num >= 0)
853 return -EADDRINUSE;
854 dev = ipmr_reg_vif(net, mrt);
855 if (!dev)
856 return -ENOBUFS;
857 err = dev_set_allmulti(dev, 1);
858 if (err) {
859 unregister_netdevice(dev);
860 dev_put(dev);
861 return err;
862 }
863 break;
864 case VIFF_TUNNEL:
865 dev = ipmr_new_tunnel(net, vifc);
866 if (!dev)
867 return -ENOBUFS;
868 err = dev_set_allmulti(dev, 1);
869 if (err) {
870 ipmr_del_tunnel(dev, vifc);
871 dev_put(dev);
872 return err;
873 }
874 break;
875 case VIFF_USE_IFINDEX:
876 case 0:
877 if (vifc->vifc_flags == VIFF_USE_IFINDEX) {
878 dev = dev_get_by_index(net, vifc->vifc_lcl_ifindex);
879 if (dev && !__in_dev_get_rtnl(dev)) {
880 dev_put(dev);
881 return -EADDRNOTAVAIL;
882 }
883 } else {
884 dev = ip_dev_find(net, vifc->vifc_lcl_addr.s_addr);
885 }
886 if (!dev)
887 return -EADDRNOTAVAIL;
888 err = dev_set_allmulti(dev, 1);
889 if (err) {
890 dev_put(dev);
891 return err;
892 }
893 break;
894 default:
895 return -EINVAL;
896 }
897
898 in_dev = __in_dev_get_rtnl(dev);
899 if (!in_dev) {
900 dev_put(dev);
901 return -EADDRNOTAVAIL;
902 }
903 IPV4_DEVCONF(in_dev->cnf, MC_FORWARDING)++;
904 inet_netconf_notify_devconf(net, RTM_NEWNETCONF, NETCONFA_MC_FORWARDING,
905 dev->ifindex, &in_dev->cnf);
906 ip_rt_multicast_event(in_dev);
907
908 /* Fill in the VIF structures */
909 vif_device_init(v, dev, vifc->vifc_rate_limit,
910 vifc->vifc_threshold,
911 vifc->vifc_flags | (!mrtsock ? VIFF_STATIC : 0),
912 (VIFF_TUNNEL | VIFF_REGISTER));
913
914 err = dev_get_port_parent_id(dev, &ppid, true);
915 if (err == 0) {
916 memcpy(v->dev_parent_id.id, ppid.id, ppid.id_len);
917 v->dev_parent_id.id_len = ppid.id_len;
918 } else {
919 v->dev_parent_id.id_len = 0;
920 }
921
922 v->local = vifc->vifc_lcl_addr.s_addr;
923 v->remote = vifc->vifc_rmt_addr.s_addr;
924
925 /* And finish update writing critical data */
926 write_lock_bh(&mrt_lock);
927 v->dev = dev;
928 if (v->flags & VIFF_REGISTER)
929 mrt->mroute_reg_vif_num = vifi;
930 if (vifi+1 > mrt->maxvif)
931 mrt->maxvif = vifi+1;
932 write_unlock_bh(&mrt_lock);
933 call_ipmr_vif_entry_notifiers(net, FIB_EVENT_VIF_ADD, v, vifi, mrt->id);
934 return 0;
935}
936
937/* called with rcu_read_lock() */
938static struct mfc_cache *ipmr_cache_find(struct mr_table *mrt,
939 __be32 origin,
940 __be32 mcastgrp)
941{
942 struct mfc_cache_cmp_arg arg = {
943 .mfc_mcastgrp = mcastgrp,
944 .mfc_origin = origin
945 };
946
947 return mr_mfc_find(mrt, &arg);
948}
949
950/* Look for a (*,G) entry */
951static struct mfc_cache *ipmr_cache_find_any(struct mr_table *mrt,
952 __be32 mcastgrp, int vifi)
953{
954 struct mfc_cache_cmp_arg arg = {
955 .mfc_mcastgrp = mcastgrp,
956 .mfc_origin = htonl(INADDR_ANY)
957 };
958
959 if (mcastgrp == htonl(INADDR_ANY))
960 return mr_mfc_find_any_parent(mrt, vifi);
961 return mr_mfc_find_any(mrt, vifi, &arg);
962}
963
964/* Look for a (S,G,iif) entry if parent != -1 */
965static struct mfc_cache *ipmr_cache_find_parent(struct mr_table *mrt,
966 __be32 origin, __be32 mcastgrp,
967 int parent)
968{
969 struct mfc_cache_cmp_arg arg = {
970 .mfc_mcastgrp = mcastgrp,
971 .mfc_origin = origin,
972 };
973
974 return mr_mfc_find_parent(mrt, &arg, parent);
975}
976
977/* Allocate a multicast cache entry */
978static struct mfc_cache *ipmr_cache_alloc(void)
979{
980 struct mfc_cache *c = kmem_cache_zalloc(mrt_cachep, GFP_KERNEL);
981
982 if (c) {
983 c->_c.mfc_un.res.last_assert = jiffies - MFC_ASSERT_THRESH - 1;
984 c->_c.mfc_un.res.minvif = MAXVIFS;
985 c->_c.free = ipmr_cache_free_rcu;
986 refcount_set(&c->_c.mfc_un.res.refcount, 1);
987 }
988 return c;
989}
990
991static struct mfc_cache *ipmr_cache_alloc_unres(void)
992{
993 struct mfc_cache *c = kmem_cache_zalloc(mrt_cachep, GFP_ATOMIC);
994
995 if (c) {
996 skb_queue_head_init(&c->_c.mfc_un.unres.unresolved);
997 c->_c.mfc_un.unres.expires = jiffies + 10 * HZ;
998 }
999 return c;
1000}
1001
1002/* A cache entry has gone into a resolved state from queued */
1003static void ipmr_cache_resolve(struct net *net, struct mr_table *mrt,
1004 struct mfc_cache *uc, struct mfc_cache *c)
1005{
1006 struct sk_buff *skb;
1007 struct nlmsgerr *e;
1008
1009 /* Play the pending entries through our router */
1010 while ((skb = __skb_dequeue(&uc->_c.mfc_un.unres.unresolved))) {
1011 if (ip_hdr(skb)->version == 0) {
1012 struct nlmsghdr *nlh = skb_pull(skb,
1013 sizeof(struct iphdr));
1014
1015 if (mr_fill_mroute(mrt, skb, &c->_c,
1016 nlmsg_data(nlh)) > 0) {
1017 nlh->nlmsg_len = skb_tail_pointer(skb) -
1018 (u8 *)nlh;
1019 } else {
1020 nlh->nlmsg_type = NLMSG_ERROR;
1021 nlh->nlmsg_len = nlmsg_msg_size(sizeof(struct nlmsgerr));
1022 skb_trim(skb, nlh->nlmsg_len);
1023 e = nlmsg_data(nlh);
1024 e->error = -EMSGSIZE;
1025 memset(&e->msg, 0, sizeof(e->msg));
1026 }
1027
1028 rtnl_unicast(skb, net, NETLINK_CB(skb).portid);
1029 } else {
1030 ip_mr_forward(net, mrt, skb->dev, skb, c, 0);
1031 }
1032 }
1033}
1034
1035/* Bounce a cache query up to mrouted and netlink.
1036 *
1037 * Called under mrt_lock.
1038 */
1039static int ipmr_cache_report(struct mr_table *mrt,
1040 struct sk_buff *pkt, vifi_t vifi, int assert)
1041{
1042 const int ihl = ip_hdrlen(pkt);
1043 struct sock *mroute_sk;
1044 struct igmphdr *igmp;
1045 struct igmpmsg *msg;
1046 struct sk_buff *skb;
1047 int ret;
1048
1049 if (assert == IGMPMSG_WHOLEPKT || assert == IGMPMSG_WRVIFWHOLE)
1050 skb = skb_realloc_headroom(pkt, sizeof(struct iphdr));
1051 else
1052 skb = alloc_skb(128, GFP_ATOMIC);
1053
1054 if (!skb)
1055 return -ENOBUFS;
1056
1057 if (assert == IGMPMSG_WHOLEPKT || assert == IGMPMSG_WRVIFWHOLE) {
1058 /* Ugly, but we have no choice with this interface.
1059 * Duplicate old header, fix ihl, length etc.
1060 * And all this only to mangle msg->im_msgtype and
1061 * to set msg->im_mbz to "mbz" :-)
1062 */
1063 skb_push(skb, sizeof(struct iphdr));
1064 skb_reset_network_header(skb);
1065 skb_reset_transport_header(skb);
1066 msg = (struct igmpmsg *)skb_network_header(skb);
1067 memcpy(msg, skb_network_header(pkt), sizeof(struct iphdr));
1068 msg->im_msgtype = assert;
1069 msg->im_mbz = 0;
1070 if (assert == IGMPMSG_WRVIFWHOLE)
1071 msg->im_vif = vifi;
1072 else
1073 msg->im_vif = mrt->mroute_reg_vif_num;
1074 ip_hdr(skb)->ihl = sizeof(struct iphdr) >> 2;
1075 ip_hdr(skb)->tot_len = htons(ntohs(ip_hdr(pkt)->tot_len) +
1076 sizeof(struct iphdr));
1077 } else {
1078 /* Copy the IP header */
1079 skb_set_network_header(skb, skb->len);
1080 skb_put(skb, ihl);
1081 skb_copy_to_linear_data(skb, pkt->data, ihl);
1082 /* Flag to the kernel this is a route add */
1083 ip_hdr(skb)->protocol = 0;
1084 msg = (struct igmpmsg *)skb_network_header(skb);
1085 msg->im_vif = vifi;
1086 skb_dst_set(skb, dst_clone(skb_dst(pkt)));
1087 /* Add our header */
1088 igmp = skb_put(skb, sizeof(struct igmphdr));
1089 igmp->type = assert;
1090 msg->im_msgtype = assert;
1091 igmp->code = 0;
1092 ip_hdr(skb)->tot_len = htons(skb->len); /* Fix the length */
1093 skb->transport_header = skb->network_header;
1094 }
1095
1096 rcu_read_lock();
1097 mroute_sk = rcu_dereference(mrt->mroute_sk);
1098 if (!mroute_sk) {
1099 rcu_read_unlock();
1100 kfree_skb(skb);
1101 return -EINVAL;
1102 }
1103
1104 igmpmsg_netlink_event(mrt, skb);
1105
1106 /* Deliver to mrouted */
1107 ret = sock_queue_rcv_skb(mroute_sk, skb);
1108 rcu_read_unlock();
1109 if (ret < 0) {
1110 net_warn_ratelimited("mroute: pending queue full, dropping entries\n");
1111 kfree_skb(skb);
1112 }
1113
1114 return ret;
1115}
1116
1117/* Queue a packet for resolution. It gets locked cache entry! */
1118static int ipmr_cache_unresolved(struct mr_table *mrt, vifi_t vifi,
1119 struct sk_buff *skb, struct net_device *dev)
1120{
1121 const struct iphdr *iph = ip_hdr(skb);
1122 struct mfc_cache *c;
1123 bool found = false;
1124 int err;
1125
1126 spin_lock_bh(&mfc_unres_lock);
1127 list_for_each_entry(c, &mrt->mfc_unres_queue, _c.list) {
1128 if (c->mfc_mcastgrp == iph->daddr &&
1129 c->mfc_origin == iph->saddr) {
1130 found = true;
1131 break;
1132 }
1133 }
1134
1135 if (!found) {
1136 /* Create a new entry if allowable */
1137 c = ipmr_cache_alloc_unres();
1138 if (!c) {
1139 spin_unlock_bh(&mfc_unres_lock);
1140
1141 kfree_skb(skb);
1142 return -ENOBUFS;
1143 }
1144
1145 /* Fill in the new cache entry */
1146 c->_c.mfc_parent = -1;
1147 c->mfc_origin = iph->saddr;
1148 c->mfc_mcastgrp = iph->daddr;
1149
1150 /* Reflect first query at mrouted. */
1151 err = ipmr_cache_report(mrt, skb, vifi, IGMPMSG_NOCACHE);
1152
1153 if (err < 0) {
1154 /* If the report failed throw the cache entry
1155 out - Brad Parker
1156 */
1157 spin_unlock_bh(&mfc_unres_lock);
1158
1159 ipmr_cache_free(c);
1160 kfree_skb(skb);
1161 return err;
1162 }
1163
1164 atomic_inc(&mrt->cache_resolve_queue_len);
1165 list_add(&c->_c.list, &mrt->mfc_unres_queue);
1166 mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1167
1168 if (atomic_read(&mrt->cache_resolve_queue_len) == 1)
1169 mod_timer(&mrt->ipmr_expire_timer,
1170 c->_c.mfc_un.unres.expires);
1171 }
1172
1173 /* See if we can append the packet */
1174 if (c->_c.mfc_un.unres.unresolved.qlen > 3) {
1175 kfree_skb(skb);
1176 err = -ENOBUFS;
1177 } else {
1178 if (dev) {
1179 skb->dev = dev;
1180 skb->skb_iif = dev->ifindex;
1181 }
1182 skb_queue_tail(&c->_c.mfc_un.unres.unresolved, skb);
1183 err = 0;
1184 }
1185
1186 spin_unlock_bh(&mfc_unres_lock);
1187 return err;
1188}
1189
1190/* MFC cache manipulation by user space mroute daemon */
1191
1192static int ipmr_mfc_delete(struct mr_table *mrt, struct mfcctl *mfc, int parent)
1193{
1194 struct net *net = read_pnet(&mrt->net);
1195 struct mfc_cache *c;
1196
1197 /* The entries are added/deleted only under RTNL */
1198 rcu_read_lock();
1199 c = ipmr_cache_find_parent(mrt, mfc->mfcc_origin.s_addr,
1200 mfc->mfcc_mcastgrp.s_addr, parent);
1201 rcu_read_unlock();
1202 if (!c)
1203 return -ENOENT;
1204 rhltable_remove(&mrt->mfc_hash, &c->_c.mnode, ipmr_rht_params);
1205 list_del_rcu(&c->_c.list);
1206 call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, c, mrt->id);
1207 mroute_netlink_event(mrt, c, RTM_DELROUTE);
1208 mr_cache_put(&c->_c);
1209
1210 return 0;
1211}
1212
1213static int ipmr_mfc_add(struct net *net, struct mr_table *mrt,
1214 struct mfcctl *mfc, int mrtsock, int parent)
1215{
1216 struct mfc_cache *uc, *c;
1217 struct mr_mfc *_uc;
1218 bool found;
1219 int ret;
1220
1221 if (mfc->mfcc_parent >= MAXVIFS)
1222 return -ENFILE;
1223
1224 /* The entries are added/deleted only under RTNL */
1225 rcu_read_lock();
1226 c = ipmr_cache_find_parent(mrt, mfc->mfcc_origin.s_addr,
1227 mfc->mfcc_mcastgrp.s_addr, parent);
1228 rcu_read_unlock();
1229 if (c) {
1230 write_lock_bh(&mrt_lock);
1231 c->_c.mfc_parent = mfc->mfcc_parent;
1232 ipmr_update_thresholds(mrt, &c->_c, mfc->mfcc_ttls);
1233 if (!mrtsock)
1234 c->_c.mfc_flags |= MFC_STATIC;
1235 write_unlock_bh(&mrt_lock);
1236 call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_REPLACE, c,
1237 mrt->id);
1238 mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1239 return 0;
1240 }
1241
1242 if (mfc->mfcc_mcastgrp.s_addr != htonl(INADDR_ANY) &&
1243 !ipv4_is_multicast(mfc->mfcc_mcastgrp.s_addr))
1244 return -EINVAL;
1245
1246 c = ipmr_cache_alloc();
1247 if (!c)
1248 return -ENOMEM;
1249
1250 c->mfc_origin = mfc->mfcc_origin.s_addr;
1251 c->mfc_mcastgrp = mfc->mfcc_mcastgrp.s_addr;
1252 c->_c.mfc_parent = mfc->mfcc_parent;
1253 ipmr_update_thresholds(mrt, &c->_c, mfc->mfcc_ttls);
1254 if (!mrtsock)
1255 c->_c.mfc_flags |= MFC_STATIC;
1256
1257 ret = rhltable_insert_key(&mrt->mfc_hash, &c->cmparg, &c->_c.mnode,
1258 ipmr_rht_params);
1259 if (ret) {
1260 pr_err("ipmr: rhtable insert error %d\n", ret);
1261 ipmr_cache_free(c);
1262 return ret;
1263 }
1264 list_add_tail_rcu(&c->_c.list, &mrt->mfc_cache_list);
1265 /* Check to see if we resolved a queued list. If so we
1266 * need to send on the frames and tidy up.
1267 */
1268 found = false;
1269 spin_lock_bh(&mfc_unres_lock);
1270 list_for_each_entry(_uc, &mrt->mfc_unres_queue, list) {
1271 uc = (struct mfc_cache *)_uc;
1272 if (uc->mfc_origin == c->mfc_origin &&
1273 uc->mfc_mcastgrp == c->mfc_mcastgrp) {
1274 list_del(&_uc->list);
1275 atomic_dec(&mrt->cache_resolve_queue_len);
1276 found = true;
1277 break;
1278 }
1279 }
1280 if (list_empty(&mrt->mfc_unres_queue))
1281 del_timer(&mrt->ipmr_expire_timer);
1282 spin_unlock_bh(&mfc_unres_lock);
1283
1284 if (found) {
1285 ipmr_cache_resolve(net, mrt, uc, c);
1286 ipmr_cache_free(uc);
1287 }
1288 call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_ADD, c, mrt->id);
1289 mroute_netlink_event(mrt, c, RTM_NEWROUTE);
1290 return 0;
1291}
1292
1293/* Close the multicast socket, and clear the vif tables etc */
1294static void mroute_clean_tables(struct mr_table *mrt, int flags)
1295{
1296 struct net *net = read_pnet(&mrt->net);
1297 struct mr_mfc *c, *tmp;
1298 struct mfc_cache *cache;
1299 LIST_HEAD(list);
1300 int i;
1301
1302 /* Shut down all active vif entries */
1303 if (flags & (MRT_FLUSH_VIFS | MRT_FLUSH_VIFS_STATIC)) {
1304 for (i = 0; i < mrt->maxvif; i++) {
1305 if (((mrt->vif_table[i].flags & VIFF_STATIC) &&
1306 !(flags & MRT_FLUSH_VIFS_STATIC)) ||
1307 (!(mrt->vif_table[i].flags & VIFF_STATIC) && !(flags & MRT_FLUSH_VIFS)))
1308 continue;
1309 vif_delete(mrt, i, 0, &list);
1310 }
1311 unregister_netdevice_many(&list);
1312 }
1313
1314 /* Wipe the cache */
1315 if (flags & (MRT_FLUSH_MFC | MRT_FLUSH_MFC_STATIC)) {
1316 list_for_each_entry_safe(c, tmp, &mrt->mfc_cache_list, list) {
1317 if (((c->mfc_flags & MFC_STATIC) && !(flags & MRT_FLUSH_MFC_STATIC)) ||
1318 (!(c->mfc_flags & MFC_STATIC) && !(flags & MRT_FLUSH_MFC)))
1319 continue;
1320 rhltable_remove(&mrt->mfc_hash, &c->mnode, ipmr_rht_params);
1321 list_del_rcu(&c->list);
1322 cache = (struct mfc_cache *)c;
1323 call_ipmr_mfc_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, cache,
1324 mrt->id);
1325 mroute_netlink_event(mrt, cache, RTM_DELROUTE);
1326 mr_cache_put(c);
1327 }
1328 }
1329
1330 if (flags & MRT_FLUSH_MFC) {
1331 if (atomic_read(&mrt->cache_resolve_queue_len) != 0) {
1332 spin_lock_bh(&mfc_unres_lock);
1333 list_for_each_entry_safe(c, tmp, &mrt->mfc_unres_queue, list) {
1334 list_del(&c->list);
1335 cache = (struct mfc_cache *)c;
1336 mroute_netlink_event(mrt, cache, RTM_DELROUTE);
1337 ipmr_destroy_unres(mrt, cache);
1338 }
1339 spin_unlock_bh(&mfc_unres_lock);
1340 }
1341 }
1342}
1343
1344/* called from ip_ra_control(), before an RCU grace period,
1345 * we dont need to call synchronize_rcu() here
1346 */
1347static void mrtsock_destruct(struct sock *sk)
1348{
1349 struct net *net = sock_net(sk);
1350 struct mr_table *mrt;
1351
1352 rtnl_lock();
1353 ipmr_for_each_table(mrt, net) {
1354 if (sk == rtnl_dereference(mrt->mroute_sk)) {
1355 IPV4_DEVCONF_ALL(net, MC_FORWARDING)--;
1356 inet_netconf_notify_devconf(net, RTM_NEWNETCONF,
1357 NETCONFA_MC_FORWARDING,
1358 NETCONFA_IFINDEX_ALL,
1359 net->ipv4.devconf_all);
1360 RCU_INIT_POINTER(mrt->mroute_sk, NULL);
1361 mroute_clean_tables(mrt, MRT_FLUSH_VIFS | MRT_FLUSH_MFC);
1362 }
1363 }
1364 rtnl_unlock();
1365}
1366
1367/* Socket options and virtual interface manipulation. The whole
1368 * virtual interface system is a complete heap, but unfortunately
1369 * that's how BSD mrouted happens to think. Maybe one day with a proper
1370 * MOSPF/PIM router set up we can clean this up.
1371 */
1372
1373int ip_mroute_setsockopt(struct sock *sk, int optname, char __user *optval,
1374 unsigned int optlen)
1375{
1376 struct net *net = sock_net(sk);
1377 int val, ret = 0, parent = 0;
1378 struct mr_table *mrt;
1379 struct vifctl vif;
1380 struct mfcctl mfc;
1381 bool do_wrvifwhole;
1382 u32 uval;
1383
1384 /* There's one exception to the lock - MRT_DONE which needs to unlock */
1385 rtnl_lock();
1386 if (sk->sk_type != SOCK_RAW ||
1387 inet_sk(sk)->inet_num != IPPROTO_IGMP) {
1388 ret = -EOPNOTSUPP;
1389 goto out_unlock;
1390 }
1391
1392 mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1393 if (!mrt) {
1394 ret = -ENOENT;
1395 goto out_unlock;
1396 }
1397 if (optname != MRT_INIT) {
1398 if (sk != rcu_access_pointer(mrt->mroute_sk) &&
1399 !ns_capable(net->user_ns, CAP_NET_ADMIN)) {
1400 ret = -EACCES;
1401 goto out_unlock;
1402 }
1403 }
1404
1405 switch (optname) {
1406 case MRT_INIT:
1407 if (optlen != sizeof(int)) {
1408 ret = -EINVAL;
1409 break;
1410 }
1411 if (rtnl_dereference(mrt->mroute_sk)) {
1412 ret = -EADDRINUSE;
1413 break;
1414 }
1415
1416 ret = ip_ra_control(sk, 1, mrtsock_destruct);
1417 if (ret == 0) {
1418 rcu_assign_pointer(mrt->mroute_sk, sk);
1419 IPV4_DEVCONF_ALL(net, MC_FORWARDING)++;
1420 inet_netconf_notify_devconf(net, RTM_NEWNETCONF,
1421 NETCONFA_MC_FORWARDING,
1422 NETCONFA_IFINDEX_ALL,
1423 net->ipv4.devconf_all);
1424 }
1425 break;
1426 case MRT_DONE:
1427 if (sk != rcu_access_pointer(mrt->mroute_sk)) {
1428 ret = -EACCES;
1429 } else {
1430 /* We need to unlock here because mrtsock_destruct takes
1431 * care of rtnl itself and we can't change that due to
1432 * the IP_ROUTER_ALERT setsockopt which runs without it.
1433 */
1434 rtnl_unlock();
1435 ret = ip_ra_control(sk, 0, NULL);
1436 goto out;
1437 }
1438 break;
1439 case MRT_ADD_VIF:
1440 case MRT_DEL_VIF:
1441 if (optlen != sizeof(vif)) {
1442 ret = -EINVAL;
1443 break;
1444 }
1445 if (copy_from_user(&vif, optval, sizeof(vif))) {
1446 ret = -EFAULT;
1447 break;
1448 }
1449 if (vif.vifc_vifi >= MAXVIFS) {
1450 ret = -ENFILE;
1451 break;
1452 }
1453 if (optname == MRT_ADD_VIF) {
1454 ret = vif_add(net, mrt, &vif,
1455 sk == rtnl_dereference(mrt->mroute_sk));
1456 } else {
1457 ret = vif_delete(mrt, vif.vifc_vifi, 0, NULL);
1458 }
1459 break;
1460 /* Manipulate the forwarding caches. These live
1461 * in a sort of kernel/user symbiosis.
1462 */
1463 case MRT_ADD_MFC:
1464 case MRT_DEL_MFC:
1465 parent = -1;
1466 /* fall through */
1467 case MRT_ADD_MFC_PROXY:
1468 case MRT_DEL_MFC_PROXY:
1469 if (optlen != sizeof(mfc)) {
1470 ret = -EINVAL;
1471 break;
1472 }
1473 if (copy_from_user(&mfc, optval, sizeof(mfc))) {
1474 ret = -EFAULT;
1475 break;
1476 }
1477 if (parent == 0)
1478 parent = mfc.mfcc_parent;
1479 if (optname == MRT_DEL_MFC || optname == MRT_DEL_MFC_PROXY)
1480 ret = ipmr_mfc_delete(mrt, &mfc, parent);
1481 else
1482 ret = ipmr_mfc_add(net, mrt, &mfc,
1483 sk == rtnl_dereference(mrt->mroute_sk),
1484 parent);
1485 break;
1486 case MRT_FLUSH:
1487 if (optlen != sizeof(val)) {
1488 ret = -EINVAL;
1489 break;
1490 }
1491 if (get_user(val, (int __user *)optval)) {
1492 ret = -EFAULT;
1493 break;
1494 }
1495 mroute_clean_tables(mrt, val);
1496 break;
1497 /* Control PIM assert. */
1498 case MRT_ASSERT:
1499 if (optlen != sizeof(val)) {
1500 ret = -EINVAL;
1501 break;
1502 }
1503 if (get_user(val, (int __user *)optval)) {
1504 ret = -EFAULT;
1505 break;
1506 }
1507 mrt->mroute_do_assert = val;
1508 break;
1509 case MRT_PIM:
1510 if (!ipmr_pimsm_enabled()) {
1511 ret = -ENOPROTOOPT;
1512 break;
1513 }
1514 if (optlen != sizeof(val)) {
1515 ret = -EINVAL;
1516 break;
1517 }
1518 if (get_user(val, (int __user *)optval)) {
1519 ret = -EFAULT;
1520 break;
1521 }
1522
1523 do_wrvifwhole = (val == IGMPMSG_WRVIFWHOLE);
1524 val = !!val;
1525 if (val != mrt->mroute_do_pim) {
1526 mrt->mroute_do_pim = val;
1527 mrt->mroute_do_assert = val;
1528 mrt->mroute_do_wrvifwhole = do_wrvifwhole;
1529 }
1530 break;
1531 case MRT_TABLE:
1532 if (!IS_BUILTIN(CONFIG_IP_MROUTE_MULTIPLE_TABLES)) {
1533 ret = -ENOPROTOOPT;
1534 break;
1535 }
1536 if (optlen != sizeof(uval)) {
1537 ret = -EINVAL;
1538 break;
1539 }
1540 if (get_user(uval, (u32 __user *)optval)) {
1541 ret = -EFAULT;
1542 break;
1543 }
1544
1545 if (sk == rtnl_dereference(mrt->mroute_sk)) {
1546 ret = -EBUSY;
1547 } else {
1548 mrt = ipmr_new_table(net, uval);
1549 if (IS_ERR(mrt))
1550 ret = PTR_ERR(mrt);
1551 else
1552 raw_sk(sk)->ipmr_table = uval;
1553 }
1554 break;
1555 /* Spurious command, or MRT_VERSION which you cannot set. */
1556 default:
1557 ret = -ENOPROTOOPT;
1558 }
1559out_unlock:
1560 rtnl_unlock();
1561out:
1562 return ret;
1563}
1564
1565/* Getsock opt support for the multicast routing system. */
1566int ip_mroute_getsockopt(struct sock *sk, int optname, char __user *optval, int __user *optlen)
1567{
1568 int olr;
1569 int val;
1570 struct net *net = sock_net(sk);
1571 struct mr_table *mrt;
1572
1573 if (sk->sk_type != SOCK_RAW ||
1574 inet_sk(sk)->inet_num != IPPROTO_IGMP)
1575 return -EOPNOTSUPP;
1576
1577 mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1578 if (!mrt)
1579 return -ENOENT;
1580
1581 switch (optname) {
1582 case MRT_VERSION:
1583 val = 0x0305;
1584 break;
1585 case MRT_PIM:
1586 if (!ipmr_pimsm_enabled())
1587 return -ENOPROTOOPT;
1588 val = mrt->mroute_do_pim;
1589 break;
1590 case MRT_ASSERT:
1591 val = mrt->mroute_do_assert;
1592 break;
1593 default:
1594 return -ENOPROTOOPT;
1595 }
1596
1597 if (get_user(olr, optlen))
1598 return -EFAULT;
1599 olr = min_t(unsigned int, olr, sizeof(int));
1600 if (olr < 0)
1601 return -EINVAL;
1602 if (put_user(olr, optlen))
1603 return -EFAULT;
1604 if (copy_to_user(optval, &val, olr))
1605 return -EFAULT;
1606 return 0;
1607}
1608
1609/* The IP multicast ioctl support routines. */
1610int ipmr_ioctl(struct sock *sk, int cmd, void __user *arg)
1611{
1612 struct sioc_sg_req sr;
1613 struct sioc_vif_req vr;
1614 struct vif_device *vif;
1615 struct mfc_cache *c;
1616 struct net *net = sock_net(sk);
1617 struct mr_table *mrt;
1618
1619 mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1620 if (!mrt)
1621 return -ENOENT;
1622
1623 switch (cmd) {
1624 case SIOCGETVIFCNT:
1625 if (copy_from_user(&vr, arg, sizeof(vr)))
1626 return -EFAULT;
1627 if (vr.vifi >= mrt->maxvif)
1628 return -EINVAL;
1629 vr.vifi = array_index_nospec(vr.vifi, mrt->maxvif);
1630 read_lock(&mrt_lock);
1631 vif = &mrt->vif_table[vr.vifi];
1632 if (VIF_EXISTS(mrt, vr.vifi)) {
1633 vr.icount = vif->pkt_in;
1634 vr.ocount = vif->pkt_out;
1635 vr.ibytes = vif->bytes_in;
1636 vr.obytes = vif->bytes_out;
1637 read_unlock(&mrt_lock);
1638
1639 if (copy_to_user(arg, &vr, sizeof(vr)))
1640 return -EFAULT;
1641 return 0;
1642 }
1643 read_unlock(&mrt_lock);
1644 return -EADDRNOTAVAIL;
1645 case SIOCGETSGCNT:
1646 if (copy_from_user(&sr, arg, sizeof(sr)))
1647 return -EFAULT;
1648
1649 rcu_read_lock();
1650 c = ipmr_cache_find(mrt, sr.src.s_addr, sr.grp.s_addr);
1651 if (c) {
1652 sr.pktcnt = c->_c.mfc_un.res.pkt;
1653 sr.bytecnt = c->_c.mfc_un.res.bytes;
1654 sr.wrong_if = c->_c.mfc_un.res.wrong_if;
1655 rcu_read_unlock();
1656
1657 if (copy_to_user(arg, &sr, sizeof(sr)))
1658 return -EFAULT;
1659 return 0;
1660 }
1661 rcu_read_unlock();
1662 return -EADDRNOTAVAIL;
1663 default:
1664 return -ENOIOCTLCMD;
1665 }
1666}
1667
1668#ifdef CONFIG_COMPAT
1669struct compat_sioc_sg_req {
1670 struct in_addr src;
1671 struct in_addr grp;
1672 compat_ulong_t pktcnt;
1673 compat_ulong_t bytecnt;
1674 compat_ulong_t wrong_if;
1675};
1676
1677struct compat_sioc_vif_req {
1678 vifi_t vifi; /* Which iface */
1679 compat_ulong_t icount;
1680 compat_ulong_t ocount;
1681 compat_ulong_t ibytes;
1682 compat_ulong_t obytes;
1683};
1684
1685int ipmr_compat_ioctl(struct sock *sk, unsigned int cmd, void __user *arg)
1686{
1687 struct compat_sioc_sg_req sr;
1688 struct compat_sioc_vif_req vr;
1689 struct vif_device *vif;
1690 struct mfc_cache *c;
1691 struct net *net = sock_net(sk);
1692 struct mr_table *mrt;
1693
1694 mrt = ipmr_get_table(net, raw_sk(sk)->ipmr_table ? : RT_TABLE_DEFAULT);
1695 if (!mrt)
1696 return -ENOENT;
1697
1698 switch (cmd) {
1699 case SIOCGETVIFCNT:
1700 if (copy_from_user(&vr, arg, sizeof(vr)))
1701 return -EFAULT;
1702 if (vr.vifi >= mrt->maxvif)
1703 return -EINVAL;
1704 vr.vifi = array_index_nospec(vr.vifi, mrt->maxvif);
1705 read_lock(&mrt_lock);
1706 vif = &mrt->vif_table[vr.vifi];
1707 if (VIF_EXISTS(mrt, vr.vifi)) {
1708 vr.icount = vif->pkt_in;
1709 vr.ocount = vif->pkt_out;
1710 vr.ibytes = vif->bytes_in;
1711 vr.obytes = vif->bytes_out;
1712 read_unlock(&mrt_lock);
1713
1714 if (copy_to_user(arg, &vr, sizeof(vr)))
1715 return -EFAULT;
1716 return 0;
1717 }
1718 read_unlock(&mrt_lock);
1719 return -EADDRNOTAVAIL;
1720 case SIOCGETSGCNT:
1721 if (copy_from_user(&sr, arg, sizeof(sr)))
1722 return -EFAULT;
1723
1724 rcu_read_lock();
1725 c = ipmr_cache_find(mrt, sr.src.s_addr, sr.grp.s_addr);
1726 if (c) {
1727 sr.pktcnt = c->_c.mfc_un.res.pkt;
1728 sr.bytecnt = c->_c.mfc_un.res.bytes;
1729 sr.wrong_if = c->_c.mfc_un.res.wrong_if;
1730 rcu_read_unlock();
1731
1732 if (copy_to_user(arg, &sr, sizeof(sr)))
1733 return -EFAULT;
1734 return 0;
1735 }
1736 rcu_read_unlock();
1737 return -EADDRNOTAVAIL;
1738 default:
1739 return -ENOIOCTLCMD;
1740 }
1741}
1742#endif
1743
1744static int ipmr_device_event(struct notifier_block *this, unsigned long event, void *ptr)
1745{
1746 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1747 struct net *net = dev_net(dev);
1748 struct mr_table *mrt;
1749 struct vif_device *v;
1750 int ct;
1751
1752 if (event != NETDEV_UNREGISTER)
1753 return NOTIFY_DONE;
1754
1755 ipmr_for_each_table(mrt, net) {
1756 v = &mrt->vif_table[0];
1757 for (ct = 0; ct < mrt->maxvif; ct++, v++) {
1758 if (v->dev == dev)
1759 vif_delete(mrt, ct, 1, NULL);
1760 }
1761 }
1762 return NOTIFY_DONE;
1763}
1764
1765static struct notifier_block ip_mr_notifier = {
1766 .notifier_call = ipmr_device_event,
1767};
1768
1769/* Encapsulate a packet by attaching a valid IPIP header to it.
1770 * This avoids tunnel drivers and other mess and gives us the speed so
1771 * important for multicast video.
1772 */
1773static void ip_encap(struct net *net, struct sk_buff *skb,
1774 __be32 saddr, __be32 daddr)
1775{
1776 struct iphdr *iph;
1777 const struct iphdr *old_iph = ip_hdr(skb);
1778
1779 skb_push(skb, sizeof(struct iphdr));
1780 skb->transport_header = skb->network_header;
1781 skb_reset_network_header(skb);
1782 iph = ip_hdr(skb);
1783
1784 iph->version = 4;
1785 iph->tos = old_iph->tos;
1786 iph->ttl = old_iph->ttl;
1787 iph->frag_off = 0;
1788 iph->daddr = daddr;
1789 iph->saddr = saddr;
1790 iph->protocol = IPPROTO_IPIP;
1791 iph->ihl = 5;
1792 iph->tot_len = htons(skb->len);
1793 ip_select_ident(net, skb, NULL);
1794 ip_send_check(iph);
1795
1796 memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
1797 nf_reset_ct(skb);
1798}
1799
1800static inline int ipmr_forward_finish(struct net *net, struct sock *sk,
1801 struct sk_buff *skb)
1802{
1803 struct ip_options *opt = &(IPCB(skb)->opt);
1804
1805 IP_INC_STATS(net, IPSTATS_MIB_OUTFORWDATAGRAMS);
1806 IP_ADD_STATS(net, IPSTATS_MIB_OUTOCTETS, skb->len);
1807
1808 if (unlikely(opt->optlen))
1809 ip_forward_options(skb);
1810
1811 return dst_output(net, sk, skb);
1812}
1813
1814#ifdef CONFIG_NET_SWITCHDEV
1815static bool ipmr_forward_offloaded(struct sk_buff *skb, struct mr_table *mrt,
1816 int in_vifi, int out_vifi)
1817{
1818 struct vif_device *out_vif = &mrt->vif_table[out_vifi];
1819 struct vif_device *in_vif = &mrt->vif_table[in_vifi];
1820
1821 if (!skb->offload_l3_fwd_mark)
1822 return false;
1823 if (!out_vif->dev_parent_id.id_len || !in_vif->dev_parent_id.id_len)
1824 return false;
1825 return netdev_phys_item_id_same(&out_vif->dev_parent_id,
1826 &in_vif->dev_parent_id);
1827}
1828#else
1829static bool ipmr_forward_offloaded(struct sk_buff *skb, struct mr_table *mrt,
1830 int in_vifi, int out_vifi)
1831{
1832 return false;
1833}
1834#endif
1835
1836/* Processing handlers for ipmr_forward */
1837
1838static void ipmr_queue_xmit(struct net *net, struct mr_table *mrt,
1839 int in_vifi, struct sk_buff *skb, int vifi)
1840{
1841 const struct iphdr *iph = ip_hdr(skb);
1842 struct vif_device *vif = &mrt->vif_table[vifi];
1843 struct net_device *dev;
1844 struct rtable *rt;
1845 struct flowi4 fl4;
1846 int encap = 0;
1847
1848 if (!vif->dev)
1849 goto out_free;
1850
1851 if (vif->flags & VIFF_REGISTER) {
1852 vif->pkt_out++;
1853 vif->bytes_out += skb->len;
1854 vif->dev->stats.tx_bytes += skb->len;
1855 vif->dev->stats.tx_packets++;
1856 ipmr_cache_report(mrt, skb, vifi, IGMPMSG_WHOLEPKT);
1857 goto out_free;
1858 }
1859
1860 if (ipmr_forward_offloaded(skb, mrt, in_vifi, vifi))
1861 goto out_free;
1862
1863 if (vif->flags & VIFF_TUNNEL) {
1864 rt = ip_route_output_ports(net, &fl4, NULL,
1865 vif->remote, vif->local,
1866 0, 0,
1867 IPPROTO_IPIP,
1868 RT_TOS(iph->tos), vif->link);
1869 if (IS_ERR(rt))
1870 goto out_free;
1871 encap = sizeof(struct iphdr);
1872 } else {
1873 rt = ip_route_output_ports(net, &fl4, NULL, iph->daddr, 0,
1874 0, 0,
1875 IPPROTO_IPIP,
1876 RT_TOS(iph->tos), vif->link);
1877 if (IS_ERR(rt))
1878 goto out_free;
1879 }
1880
1881 dev = rt->dst.dev;
1882
1883 if (skb->len+encap > dst_mtu(&rt->dst) && (ntohs(iph->frag_off) & IP_DF)) {
1884 /* Do not fragment multicasts. Alas, IPv4 does not
1885 * allow to send ICMP, so that packets will disappear
1886 * to blackhole.
1887 */
1888 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
1889 ip_rt_put(rt);
1890 goto out_free;
1891 }
1892
1893 encap += LL_RESERVED_SPACE(dev) + rt->dst.header_len;
1894
1895 if (skb_cow(skb, encap)) {
1896 ip_rt_put(rt);
1897 goto out_free;
1898 }
1899
1900 vif->pkt_out++;
1901 vif->bytes_out += skb->len;
1902
1903 skb_dst_drop(skb);
1904 skb_dst_set(skb, &rt->dst);
1905 ip_decrease_ttl(ip_hdr(skb));
1906
1907 /* FIXME: forward and output firewalls used to be called here.
1908 * What do we do with netfilter? -- RR
1909 */
1910 if (vif->flags & VIFF_TUNNEL) {
1911 ip_encap(net, skb, vif->local, vif->remote);
1912 /* FIXME: extra output firewall step used to be here. --RR */
1913 vif->dev->stats.tx_packets++;
1914 vif->dev->stats.tx_bytes += skb->len;
1915 }
1916
1917 IPCB(skb)->flags |= IPSKB_FORWARDED;
1918
1919 /* RFC1584 teaches, that DVMRP/PIM router must deliver packets locally
1920 * not only before forwarding, but after forwarding on all output
1921 * interfaces. It is clear, if mrouter runs a multicasting
1922 * program, it should receive packets not depending to what interface
1923 * program is joined.
1924 * If we will not make it, the program will have to join on all
1925 * interfaces. On the other hand, multihoming host (or router, but
1926 * not mrouter) cannot join to more than one interface - it will
1927 * result in receiving multiple packets.
1928 */
1929 NF_HOOK(NFPROTO_IPV4, NF_INET_FORWARD,
1930 net, NULL, skb, skb->dev, dev,
1931 ipmr_forward_finish);
1932 return;
1933
1934out_free:
1935 kfree_skb(skb);
1936}
1937
1938static int ipmr_find_vif(struct mr_table *mrt, struct net_device *dev)
1939{
1940 int ct;
1941
1942 for (ct = mrt->maxvif-1; ct >= 0; ct--) {
1943 if (mrt->vif_table[ct].dev == dev)
1944 break;
1945 }
1946 return ct;
1947}
1948
1949/* "local" means that we should preserve one skb (for local delivery) */
1950static void ip_mr_forward(struct net *net, struct mr_table *mrt,
1951 struct net_device *dev, struct sk_buff *skb,
1952 struct mfc_cache *c, int local)
1953{
1954 int true_vifi = ipmr_find_vif(mrt, dev);
1955 int psend = -1;
1956 int vif, ct;
1957
1958 vif = c->_c.mfc_parent;
1959 c->_c.mfc_un.res.pkt++;
1960 c->_c.mfc_un.res.bytes += skb->len;
1961 c->_c.mfc_un.res.lastuse = jiffies;
1962
1963 if (c->mfc_origin == htonl(INADDR_ANY) && true_vifi >= 0) {
1964 struct mfc_cache *cache_proxy;
1965
1966 /* For an (*,G) entry, we only check that the incomming
1967 * interface is part of the static tree.
1968 */
1969 cache_proxy = mr_mfc_find_any_parent(mrt, vif);
1970 if (cache_proxy &&
1971 cache_proxy->_c.mfc_un.res.ttls[true_vifi] < 255)
1972 goto forward;
1973 }
1974
1975 /* Wrong interface: drop packet and (maybe) send PIM assert. */
1976 if (mrt->vif_table[vif].dev != dev) {
1977 if (rt_is_output_route(skb_rtable(skb))) {
1978 /* It is our own packet, looped back.
1979 * Very complicated situation...
1980 *
1981 * The best workaround until routing daemons will be
1982 * fixed is not to redistribute packet, if it was
1983 * send through wrong interface. It means, that
1984 * multicast applications WILL NOT work for
1985 * (S,G), which have default multicast route pointing
1986 * to wrong oif. In any case, it is not a good
1987 * idea to use multicasting applications on router.
1988 */
1989 goto dont_forward;
1990 }
1991
1992 c->_c.mfc_un.res.wrong_if++;
1993
1994 if (true_vifi >= 0 && mrt->mroute_do_assert &&
1995 /* pimsm uses asserts, when switching from RPT to SPT,
1996 * so that we cannot check that packet arrived on an oif.
1997 * It is bad, but otherwise we would need to move pretty
1998 * large chunk of pimd to kernel. Ough... --ANK
1999 */
2000 (mrt->mroute_do_pim ||
2001 c->_c.mfc_un.res.ttls[true_vifi] < 255) &&
2002 time_after(jiffies,
2003 c->_c.mfc_un.res.last_assert +
2004 MFC_ASSERT_THRESH)) {
2005 c->_c.mfc_un.res.last_assert = jiffies;
2006 ipmr_cache_report(mrt, skb, true_vifi, IGMPMSG_WRONGVIF);
2007 if (mrt->mroute_do_wrvifwhole)
2008 ipmr_cache_report(mrt, skb, true_vifi,
2009 IGMPMSG_WRVIFWHOLE);
2010 }
2011 goto dont_forward;
2012 }
2013
2014forward:
2015 mrt->vif_table[vif].pkt_in++;
2016 mrt->vif_table[vif].bytes_in += skb->len;
2017
2018 /* Forward the frame */
2019 if (c->mfc_origin == htonl(INADDR_ANY) &&
2020 c->mfc_mcastgrp == htonl(INADDR_ANY)) {
2021 if (true_vifi >= 0 &&
2022 true_vifi != c->_c.mfc_parent &&
2023 ip_hdr(skb)->ttl >
2024 c->_c.mfc_un.res.ttls[c->_c.mfc_parent]) {
2025 /* It's an (*,*) entry and the packet is not coming from
2026 * the upstream: forward the packet to the upstream
2027 * only.
2028 */
2029 psend = c->_c.mfc_parent;
2030 goto last_forward;
2031 }
2032 goto dont_forward;
2033 }
2034 for (ct = c->_c.mfc_un.res.maxvif - 1;
2035 ct >= c->_c.mfc_un.res.minvif; ct--) {
2036 /* For (*,G) entry, don't forward to the incoming interface */
2037 if ((c->mfc_origin != htonl(INADDR_ANY) ||
2038 ct != true_vifi) &&
2039 ip_hdr(skb)->ttl > c->_c.mfc_un.res.ttls[ct]) {
2040 if (psend != -1) {
2041 struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
2042
2043 if (skb2)
2044 ipmr_queue_xmit(net, mrt, true_vifi,
2045 skb2, psend);
2046 }
2047 psend = ct;
2048 }
2049 }
2050last_forward:
2051 if (psend != -1) {
2052 if (local) {
2053 struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
2054
2055 if (skb2)
2056 ipmr_queue_xmit(net, mrt, true_vifi, skb2,
2057 psend);
2058 } else {
2059 ipmr_queue_xmit(net, mrt, true_vifi, skb, psend);
2060 return;
2061 }
2062 }
2063
2064dont_forward:
2065 if (!local)
2066 kfree_skb(skb);
2067}
2068
2069static struct mr_table *ipmr_rt_fib_lookup(struct net *net, struct sk_buff *skb)
2070{
2071 struct rtable *rt = skb_rtable(skb);
2072 struct iphdr *iph = ip_hdr(skb);
2073 struct flowi4 fl4 = {
2074 .daddr = iph->daddr,
2075 .saddr = iph->saddr,
2076 .flowi4_tos = RT_TOS(iph->tos),
2077 .flowi4_oif = (rt_is_output_route(rt) ?
2078 skb->dev->ifindex : 0),
2079 .flowi4_iif = (rt_is_output_route(rt) ?
2080 LOOPBACK_IFINDEX :
2081 skb->dev->ifindex),
2082 .flowi4_mark = skb->mark,
2083 };
2084 struct mr_table *mrt;
2085 int err;
2086
2087 err = ipmr_fib_lookup(net, &fl4, &mrt);
2088 if (err)
2089 return ERR_PTR(err);
2090 return mrt;
2091}
2092
2093/* Multicast packets for forwarding arrive here
2094 * Called with rcu_read_lock();
2095 */
2096int ip_mr_input(struct sk_buff *skb)
2097{
2098 struct mfc_cache *cache;
2099 struct net *net = dev_net(skb->dev);
2100 int local = skb_rtable(skb)->rt_flags & RTCF_LOCAL;
2101 struct mr_table *mrt;
2102 struct net_device *dev;
2103
2104 /* skb->dev passed in is the loX master dev for vrfs.
2105 * As there are no vifs associated with loopback devices,
2106 * get the proper interface that does have a vif associated with it.
2107 */
2108 dev = skb->dev;
2109 if (netif_is_l3_master(skb->dev)) {
2110 dev = dev_get_by_index_rcu(net, IPCB(skb)->iif);
2111 if (!dev) {
2112 kfree_skb(skb);
2113 return -ENODEV;
2114 }
2115 }
2116
2117 /* Packet is looped back after forward, it should not be
2118 * forwarded second time, but still can be delivered locally.
2119 */
2120 if (IPCB(skb)->flags & IPSKB_FORWARDED)
2121 goto dont_forward;
2122
2123 mrt = ipmr_rt_fib_lookup(net, skb);
2124 if (IS_ERR(mrt)) {
2125 kfree_skb(skb);
2126 return PTR_ERR(mrt);
2127 }
2128 if (!local) {
2129 if (IPCB(skb)->opt.router_alert) {
2130 if (ip_call_ra_chain(skb))
2131 return 0;
2132 } else if (ip_hdr(skb)->protocol == IPPROTO_IGMP) {
2133 /* IGMPv1 (and broken IGMPv2 implementations sort of
2134 * Cisco IOS <= 11.2(8)) do not put router alert
2135 * option to IGMP packets destined to routable
2136 * groups. It is very bad, because it means
2137 * that we can forward NO IGMP messages.
2138 */
2139 struct sock *mroute_sk;
2140
2141 mroute_sk = rcu_dereference(mrt->mroute_sk);
2142 if (mroute_sk) {
2143 nf_reset_ct(skb);
2144 raw_rcv(mroute_sk, skb);
2145 return 0;
2146 }
2147 }
2148 }
2149
2150 /* already under rcu_read_lock() */
2151 cache = ipmr_cache_find(mrt, ip_hdr(skb)->saddr, ip_hdr(skb)->daddr);
2152 if (!cache) {
2153 int vif = ipmr_find_vif(mrt, dev);
2154
2155 if (vif >= 0)
2156 cache = ipmr_cache_find_any(mrt, ip_hdr(skb)->daddr,
2157 vif);
2158 }
2159
2160 /* No usable cache entry */
2161 if (!cache) {
2162 int vif;
2163
2164 if (local) {
2165 struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
2166 ip_local_deliver(skb);
2167 if (!skb2)
2168 return -ENOBUFS;
2169 skb = skb2;
2170 }
2171
2172 read_lock(&mrt_lock);
2173 vif = ipmr_find_vif(mrt, dev);
2174 if (vif >= 0) {
2175 int err2 = ipmr_cache_unresolved(mrt, vif, skb, dev);
2176 read_unlock(&mrt_lock);
2177
2178 return err2;
2179 }
2180 read_unlock(&mrt_lock);
2181 kfree_skb(skb);
2182 return -ENODEV;
2183 }
2184
2185 read_lock(&mrt_lock);
2186 ip_mr_forward(net, mrt, dev, skb, cache, local);
2187 read_unlock(&mrt_lock);
2188
2189 if (local)
2190 return ip_local_deliver(skb);
2191
2192 return 0;
2193
2194dont_forward:
2195 if (local)
2196 return ip_local_deliver(skb);
2197 kfree_skb(skb);
2198 return 0;
2199}
2200
2201#ifdef CONFIG_IP_PIMSM_V1
2202/* Handle IGMP messages of PIMv1 */
2203int pim_rcv_v1(struct sk_buff *skb)
2204{
2205 struct igmphdr *pim;
2206 struct net *net = dev_net(skb->dev);
2207 struct mr_table *mrt;
2208
2209 if (!pskb_may_pull(skb, sizeof(*pim) + sizeof(struct iphdr)))
2210 goto drop;
2211
2212 pim = igmp_hdr(skb);
2213
2214 mrt = ipmr_rt_fib_lookup(net, skb);
2215 if (IS_ERR(mrt))
2216 goto drop;
2217 if (!mrt->mroute_do_pim ||
2218 pim->group != PIM_V1_VERSION || pim->code != PIM_V1_REGISTER)
2219 goto drop;
2220
2221 if (__pim_rcv(mrt, skb, sizeof(*pim))) {
2222drop:
2223 kfree_skb(skb);
2224 }
2225 return 0;
2226}
2227#endif
2228
2229#ifdef CONFIG_IP_PIMSM_V2
2230static int pim_rcv(struct sk_buff *skb)
2231{
2232 struct pimreghdr *pim;
2233 struct net *net = dev_net(skb->dev);
2234 struct mr_table *mrt;
2235
2236 if (!pskb_may_pull(skb, sizeof(*pim) + sizeof(struct iphdr)))
2237 goto drop;
2238
2239 pim = (struct pimreghdr *)skb_transport_header(skb);
2240 if (pim->type != ((PIM_VERSION << 4) | (PIM_TYPE_REGISTER)) ||
2241 (pim->flags & PIM_NULL_REGISTER) ||
2242 (ip_compute_csum((void *)pim, sizeof(*pim)) != 0 &&
2243 csum_fold(skb_checksum(skb, 0, skb->len, 0))))
2244 goto drop;
2245
2246 mrt = ipmr_rt_fib_lookup(net, skb);
2247 if (IS_ERR(mrt))
2248 goto drop;
2249 if (__pim_rcv(mrt, skb, sizeof(*pim))) {
2250drop:
2251 kfree_skb(skb);
2252 }
2253 return 0;
2254}
2255#endif
2256
2257int ipmr_get_route(struct net *net, struct sk_buff *skb,
2258 __be32 saddr, __be32 daddr,
2259 struct rtmsg *rtm, u32 portid)
2260{
2261 struct mfc_cache *cache;
2262 struct mr_table *mrt;
2263 int err;
2264
2265 mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2266 if (!mrt)
2267 return -ENOENT;
2268
2269 rcu_read_lock();
2270 cache = ipmr_cache_find(mrt, saddr, daddr);
2271 if (!cache && skb->dev) {
2272 int vif = ipmr_find_vif(mrt, skb->dev);
2273
2274 if (vif >= 0)
2275 cache = ipmr_cache_find_any(mrt, daddr, vif);
2276 }
2277 if (!cache) {
2278 struct sk_buff *skb2;
2279 struct iphdr *iph;
2280 struct net_device *dev;
2281 int vif = -1;
2282
2283 dev = skb->dev;
2284 read_lock(&mrt_lock);
2285 if (dev)
2286 vif = ipmr_find_vif(mrt, dev);
2287 if (vif < 0) {
2288 read_unlock(&mrt_lock);
2289 rcu_read_unlock();
2290 return -ENODEV;
2291 }
2292
2293 skb2 = skb_realloc_headroom(skb, sizeof(struct iphdr));
2294 if (!skb2) {
2295 read_unlock(&mrt_lock);
2296 rcu_read_unlock();
2297 return -ENOMEM;
2298 }
2299
2300 NETLINK_CB(skb2).portid = portid;
2301 skb_push(skb2, sizeof(struct iphdr));
2302 skb_reset_network_header(skb2);
2303 iph = ip_hdr(skb2);
2304 iph->ihl = sizeof(struct iphdr) >> 2;
2305 iph->saddr = saddr;
2306 iph->daddr = daddr;
2307 iph->version = 0;
2308 err = ipmr_cache_unresolved(mrt, vif, skb2, dev);
2309 read_unlock(&mrt_lock);
2310 rcu_read_unlock();
2311 return err;
2312 }
2313
2314 read_lock(&mrt_lock);
2315 err = mr_fill_mroute(mrt, skb, &cache->_c, rtm);
2316 read_unlock(&mrt_lock);
2317 rcu_read_unlock();
2318 return err;
2319}
2320
2321static int ipmr_fill_mroute(struct mr_table *mrt, struct sk_buff *skb,
2322 u32 portid, u32 seq, struct mfc_cache *c, int cmd,
2323 int flags)
2324{
2325 struct nlmsghdr *nlh;
2326 struct rtmsg *rtm;
2327 int err;
2328
2329 nlh = nlmsg_put(skb, portid, seq, cmd, sizeof(*rtm), flags);
2330 if (!nlh)
2331 return -EMSGSIZE;
2332
2333 rtm = nlmsg_data(nlh);
2334 rtm->rtm_family = RTNL_FAMILY_IPMR;
2335 rtm->rtm_dst_len = 32;
2336 rtm->rtm_src_len = 32;
2337 rtm->rtm_tos = 0;
2338 rtm->rtm_table = mrt->id;
2339 if (nla_put_u32(skb, RTA_TABLE, mrt->id))
2340 goto nla_put_failure;
2341 rtm->rtm_type = RTN_MULTICAST;
2342 rtm->rtm_scope = RT_SCOPE_UNIVERSE;
2343 if (c->_c.mfc_flags & MFC_STATIC)
2344 rtm->rtm_protocol = RTPROT_STATIC;
2345 else
2346 rtm->rtm_protocol = RTPROT_MROUTED;
2347 rtm->rtm_flags = 0;
2348
2349 if (nla_put_in_addr(skb, RTA_SRC, c->mfc_origin) ||
2350 nla_put_in_addr(skb, RTA_DST, c->mfc_mcastgrp))
2351 goto nla_put_failure;
2352 err = mr_fill_mroute(mrt, skb, &c->_c, rtm);
2353 /* do not break the dump if cache is unresolved */
2354 if (err < 0 && err != -ENOENT)
2355 goto nla_put_failure;
2356
2357 nlmsg_end(skb, nlh);
2358 return 0;
2359
2360nla_put_failure:
2361 nlmsg_cancel(skb, nlh);
2362 return -EMSGSIZE;
2363}
2364
2365static int _ipmr_fill_mroute(struct mr_table *mrt, struct sk_buff *skb,
2366 u32 portid, u32 seq, struct mr_mfc *c, int cmd,
2367 int flags)
2368{
2369 return ipmr_fill_mroute(mrt, skb, portid, seq, (struct mfc_cache *)c,
2370 cmd, flags);
2371}
2372
2373static size_t mroute_msgsize(bool unresolved, int maxvif)
2374{
2375 size_t len =
2376 NLMSG_ALIGN(sizeof(struct rtmsg))
2377 + nla_total_size(4) /* RTA_TABLE */
2378 + nla_total_size(4) /* RTA_SRC */
2379 + nla_total_size(4) /* RTA_DST */
2380 ;
2381
2382 if (!unresolved)
2383 len = len
2384 + nla_total_size(4) /* RTA_IIF */
2385 + nla_total_size(0) /* RTA_MULTIPATH */
2386 + maxvif * NLA_ALIGN(sizeof(struct rtnexthop))
2387 /* RTA_MFC_STATS */
2388 + nla_total_size_64bit(sizeof(struct rta_mfc_stats))
2389 ;
2390
2391 return len;
2392}
2393
2394static void mroute_netlink_event(struct mr_table *mrt, struct mfc_cache *mfc,
2395 int cmd)
2396{
2397 struct net *net = read_pnet(&mrt->net);
2398 struct sk_buff *skb;
2399 int err = -ENOBUFS;
2400
2401 skb = nlmsg_new(mroute_msgsize(mfc->_c.mfc_parent >= MAXVIFS,
2402 mrt->maxvif),
2403 GFP_ATOMIC);
2404 if (!skb)
2405 goto errout;
2406
2407 err = ipmr_fill_mroute(mrt, skb, 0, 0, mfc, cmd, 0);
2408 if (err < 0)
2409 goto errout;
2410
2411 rtnl_notify(skb, net, 0, RTNLGRP_IPV4_MROUTE, NULL, GFP_ATOMIC);
2412 return;
2413
2414errout:
2415 kfree_skb(skb);
2416 if (err < 0)
2417 rtnl_set_sk_err(net, RTNLGRP_IPV4_MROUTE, err);
2418}
2419
2420static size_t igmpmsg_netlink_msgsize(size_t payloadlen)
2421{
2422 size_t len =
2423 NLMSG_ALIGN(sizeof(struct rtgenmsg))
2424 + nla_total_size(1) /* IPMRA_CREPORT_MSGTYPE */
2425 + nla_total_size(4) /* IPMRA_CREPORT_VIF_ID */
2426 + nla_total_size(4) /* IPMRA_CREPORT_SRC_ADDR */
2427 + nla_total_size(4) /* IPMRA_CREPORT_DST_ADDR */
2428 /* IPMRA_CREPORT_PKT */
2429 + nla_total_size(payloadlen)
2430 ;
2431
2432 return len;
2433}
2434
2435static void igmpmsg_netlink_event(struct mr_table *mrt, struct sk_buff *pkt)
2436{
2437 struct net *net = read_pnet(&mrt->net);
2438 struct nlmsghdr *nlh;
2439 struct rtgenmsg *rtgenm;
2440 struct igmpmsg *msg;
2441 struct sk_buff *skb;
2442 struct nlattr *nla;
2443 int payloadlen;
2444
2445 payloadlen = pkt->len - sizeof(struct igmpmsg);
2446 msg = (struct igmpmsg *)skb_network_header(pkt);
2447
2448 skb = nlmsg_new(igmpmsg_netlink_msgsize(payloadlen), GFP_ATOMIC);
2449 if (!skb)
2450 goto errout;
2451
2452 nlh = nlmsg_put(skb, 0, 0, RTM_NEWCACHEREPORT,
2453 sizeof(struct rtgenmsg), 0);
2454 if (!nlh)
2455 goto errout;
2456 rtgenm = nlmsg_data(nlh);
2457 rtgenm->rtgen_family = RTNL_FAMILY_IPMR;
2458 if (nla_put_u8(skb, IPMRA_CREPORT_MSGTYPE, msg->im_msgtype) ||
2459 nla_put_u32(skb, IPMRA_CREPORT_VIF_ID, msg->im_vif) ||
2460 nla_put_in_addr(skb, IPMRA_CREPORT_SRC_ADDR,
2461 msg->im_src.s_addr) ||
2462 nla_put_in_addr(skb, IPMRA_CREPORT_DST_ADDR,
2463 msg->im_dst.s_addr))
2464 goto nla_put_failure;
2465
2466 nla = nla_reserve(skb, IPMRA_CREPORT_PKT, payloadlen);
2467 if (!nla || skb_copy_bits(pkt, sizeof(struct igmpmsg),
2468 nla_data(nla), payloadlen))
2469 goto nla_put_failure;
2470
2471 nlmsg_end(skb, nlh);
2472
2473 rtnl_notify(skb, net, 0, RTNLGRP_IPV4_MROUTE_R, NULL, GFP_ATOMIC);
2474 return;
2475
2476nla_put_failure:
2477 nlmsg_cancel(skb, nlh);
2478errout:
2479 kfree_skb(skb);
2480 rtnl_set_sk_err(net, RTNLGRP_IPV4_MROUTE_R, -ENOBUFS);
2481}
2482
2483static int ipmr_rtm_valid_getroute_req(struct sk_buff *skb,
2484 const struct nlmsghdr *nlh,
2485 struct nlattr **tb,
2486 struct netlink_ext_ack *extack)
2487{
2488 struct rtmsg *rtm;
2489 int i, err;
2490
2491 if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*rtm))) {
2492 NL_SET_ERR_MSG(extack, "ipv4: Invalid header for multicast route get request");
2493 return -EINVAL;
2494 }
2495
2496 if (!netlink_strict_get_check(skb))
2497 return nlmsg_parse_deprecated(nlh, sizeof(*rtm), tb, RTA_MAX,
2498 rtm_ipv4_policy, extack);
2499
2500 rtm = nlmsg_data(nlh);
2501 if ((rtm->rtm_src_len && rtm->rtm_src_len != 32) ||
2502 (rtm->rtm_dst_len && rtm->rtm_dst_len != 32) ||
2503 rtm->rtm_tos || rtm->rtm_table || rtm->rtm_protocol ||
2504 rtm->rtm_scope || rtm->rtm_type || rtm->rtm_flags) {
2505 NL_SET_ERR_MSG(extack, "ipv4: Invalid values in header for multicast route get request");
2506 return -EINVAL;
2507 }
2508
2509 err = nlmsg_parse_deprecated_strict(nlh, sizeof(*rtm), tb, RTA_MAX,
2510 rtm_ipv4_policy, extack);
2511 if (err)
2512 return err;
2513
2514 if ((tb[RTA_SRC] && !rtm->rtm_src_len) ||
2515 (tb[RTA_DST] && !rtm->rtm_dst_len)) {
2516 NL_SET_ERR_MSG(extack, "ipv4: rtm_src_len and rtm_dst_len must be 32 for IPv4");
2517 return -EINVAL;
2518 }
2519
2520 for (i = 0; i <= RTA_MAX; i++) {
2521 if (!tb[i])
2522 continue;
2523
2524 switch (i) {
2525 case RTA_SRC:
2526 case RTA_DST:
2527 case RTA_TABLE:
2528 break;
2529 default:
2530 NL_SET_ERR_MSG(extack, "ipv4: Unsupported attribute in multicast route get request");
2531 return -EINVAL;
2532 }
2533 }
2534
2535 return 0;
2536}
2537
2538static int ipmr_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr *nlh,
2539 struct netlink_ext_ack *extack)
2540{
2541 struct net *net = sock_net(in_skb->sk);
2542 struct nlattr *tb[RTA_MAX + 1];
2543 struct sk_buff *skb = NULL;
2544 struct mfc_cache *cache;
2545 struct mr_table *mrt;
2546 __be32 src, grp;
2547 u32 tableid;
2548 int err;
2549
2550 err = ipmr_rtm_valid_getroute_req(in_skb, nlh, tb, extack);
2551 if (err < 0)
2552 goto errout;
2553
2554 src = tb[RTA_SRC] ? nla_get_in_addr(tb[RTA_SRC]) : 0;
2555 grp = tb[RTA_DST] ? nla_get_in_addr(tb[RTA_DST]) : 0;
2556 tableid = tb[RTA_TABLE] ? nla_get_u32(tb[RTA_TABLE]) : 0;
2557
2558 mrt = ipmr_get_table(net, tableid ? tableid : RT_TABLE_DEFAULT);
2559 if (!mrt) {
2560 err = -ENOENT;
2561 goto errout_free;
2562 }
2563
2564 /* entries are added/deleted only under RTNL */
2565 rcu_read_lock();
2566 cache = ipmr_cache_find(mrt, src, grp);
2567 rcu_read_unlock();
2568 if (!cache) {
2569 err = -ENOENT;
2570 goto errout_free;
2571 }
2572
2573 skb = nlmsg_new(mroute_msgsize(false, mrt->maxvif), GFP_KERNEL);
2574 if (!skb) {
2575 err = -ENOBUFS;
2576 goto errout_free;
2577 }
2578
2579 err = ipmr_fill_mroute(mrt, skb, NETLINK_CB(in_skb).portid,
2580 nlh->nlmsg_seq, cache,
2581 RTM_NEWROUTE, 0);
2582 if (err < 0)
2583 goto errout_free;
2584
2585 err = rtnl_unicast(skb, net, NETLINK_CB(in_skb).portid);
2586
2587errout:
2588 return err;
2589
2590errout_free:
2591 kfree_skb(skb);
2592 goto errout;
2593}
2594
2595static int ipmr_rtm_dumproute(struct sk_buff *skb, struct netlink_callback *cb)
2596{
2597 struct fib_dump_filter filter = {};
2598 int err;
2599
2600 if (cb->strict_check) {
2601 err = ip_valid_fib_dump_req(sock_net(skb->sk), cb->nlh,
2602 &filter, cb);
2603 if (err < 0)
2604 return err;
2605 }
2606
2607 if (filter.table_id) {
2608 struct mr_table *mrt;
2609
2610 mrt = ipmr_get_table(sock_net(skb->sk), filter.table_id);
2611 if (!mrt) {
2612 if (filter.dump_all_families)
2613 return skb->len;
2614
2615 NL_SET_ERR_MSG(cb->extack, "ipv4: MR table does not exist");
2616 return -ENOENT;
2617 }
2618 err = mr_table_dump(mrt, skb, cb, _ipmr_fill_mroute,
2619 &mfc_unres_lock, &filter);
2620 return skb->len ? : err;
2621 }
2622
2623 return mr_rtm_dumproute(skb, cb, ipmr_mr_table_iter,
2624 _ipmr_fill_mroute, &mfc_unres_lock, &filter);
2625}
2626
2627static const struct nla_policy rtm_ipmr_policy[RTA_MAX + 1] = {
2628 [RTA_SRC] = { .type = NLA_U32 },
2629 [RTA_DST] = { .type = NLA_U32 },
2630 [RTA_IIF] = { .type = NLA_U32 },
2631 [RTA_TABLE] = { .type = NLA_U32 },
2632 [RTA_MULTIPATH] = { .len = sizeof(struct rtnexthop) },
2633};
2634
2635static bool ipmr_rtm_validate_proto(unsigned char rtm_protocol)
2636{
2637 switch (rtm_protocol) {
2638 case RTPROT_STATIC:
2639 case RTPROT_MROUTED:
2640 return true;
2641 }
2642 return false;
2643}
2644
2645static int ipmr_nla_get_ttls(const struct nlattr *nla, struct mfcctl *mfcc)
2646{
2647 struct rtnexthop *rtnh = nla_data(nla);
2648 int remaining = nla_len(nla), vifi = 0;
2649
2650 while (rtnh_ok(rtnh, remaining)) {
2651 mfcc->mfcc_ttls[vifi] = rtnh->rtnh_hops;
2652 if (++vifi == MAXVIFS)
2653 break;
2654 rtnh = rtnh_next(rtnh, &remaining);
2655 }
2656
2657 return remaining > 0 ? -EINVAL : vifi;
2658}
2659
2660/* returns < 0 on error, 0 for ADD_MFC and 1 for ADD_MFC_PROXY */
2661static int rtm_to_ipmr_mfcc(struct net *net, struct nlmsghdr *nlh,
2662 struct mfcctl *mfcc, int *mrtsock,
2663 struct mr_table **mrtret,
2664 struct netlink_ext_ack *extack)
2665{
2666 struct net_device *dev = NULL;
2667 u32 tblid = RT_TABLE_DEFAULT;
2668 struct mr_table *mrt;
2669 struct nlattr *attr;
2670 struct rtmsg *rtm;
2671 int ret, rem;
2672
2673 ret = nlmsg_validate_deprecated(nlh, sizeof(*rtm), RTA_MAX,
2674 rtm_ipmr_policy, extack);
2675 if (ret < 0)
2676 goto out;
2677 rtm = nlmsg_data(nlh);
2678
2679 ret = -EINVAL;
2680 if (rtm->rtm_family != RTNL_FAMILY_IPMR || rtm->rtm_dst_len != 32 ||
2681 rtm->rtm_type != RTN_MULTICAST ||
2682 rtm->rtm_scope != RT_SCOPE_UNIVERSE ||
2683 !ipmr_rtm_validate_proto(rtm->rtm_protocol))
2684 goto out;
2685
2686 memset(mfcc, 0, sizeof(*mfcc));
2687 mfcc->mfcc_parent = -1;
2688 ret = 0;
2689 nlmsg_for_each_attr(attr, nlh, sizeof(struct rtmsg), rem) {
2690 switch (nla_type(attr)) {
2691 case RTA_SRC:
2692 mfcc->mfcc_origin.s_addr = nla_get_be32(attr);
2693 break;
2694 case RTA_DST:
2695 mfcc->mfcc_mcastgrp.s_addr = nla_get_be32(attr);
2696 break;
2697 case RTA_IIF:
2698 dev = __dev_get_by_index(net, nla_get_u32(attr));
2699 if (!dev) {
2700 ret = -ENODEV;
2701 goto out;
2702 }
2703 break;
2704 case RTA_MULTIPATH:
2705 if (ipmr_nla_get_ttls(attr, mfcc) < 0) {
2706 ret = -EINVAL;
2707 goto out;
2708 }
2709 break;
2710 case RTA_PREFSRC:
2711 ret = 1;
2712 break;
2713 case RTA_TABLE:
2714 tblid = nla_get_u32(attr);
2715 break;
2716 }
2717 }
2718 mrt = ipmr_get_table(net, tblid);
2719 if (!mrt) {
2720 ret = -ENOENT;
2721 goto out;
2722 }
2723 *mrtret = mrt;
2724 *mrtsock = rtm->rtm_protocol == RTPROT_MROUTED ? 1 : 0;
2725 if (dev)
2726 mfcc->mfcc_parent = ipmr_find_vif(mrt, dev);
2727
2728out:
2729 return ret;
2730}
2731
2732/* takes care of both newroute and delroute */
2733static int ipmr_rtm_route(struct sk_buff *skb, struct nlmsghdr *nlh,
2734 struct netlink_ext_ack *extack)
2735{
2736 struct net *net = sock_net(skb->sk);
2737 int ret, mrtsock, parent;
2738 struct mr_table *tbl;
2739 struct mfcctl mfcc;
2740
2741 mrtsock = 0;
2742 tbl = NULL;
2743 ret = rtm_to_ipmr_mfcc(net, nlh, &mfcc, &mrtsock, &tbl, extack);
2744 if (ret < 0)
2745 return ret;
2746
2747 parent = ret ? mfcc.mfcc_parent : -1;
2748 if (nlh->nlmsg_type == RTM_NEWROUTE)
2749 return ipmr_mfc_add(net, tbl, &mfcc, mrtsock, parent);
2750 else
2751 return ipmr_mfc_delete(tbl, &mfcc, parent);
2752}
2753
2754static bool ipmr_fill_table(struct mr_table *mrt, struct sk_buff *skb)
2755{
2756 u32 queue_len = atomic_read(&mrt->cache_resolve_queue_len);
2757
2758 if (nla_put_u32(skb, IPMRA_TABLE_ID, mrt->id) ||
2759 nla_put_u32(skb, IPMRA_TABLE_CACHE_RES_QUEUE_LEN, queue_len) ||
2760 nla_put_s32(skb, IPMRA_TABLE_MROUTE_REG_VIF_NUM,
2761 mrt->mroute_reg_vif_num) ||
2762 nla_put_u8(skb, IPMRA_TABLE_MROUTE_DO_ASSERT,
2763 mrt->mroute_do_assert) ||
2764 nla_put_u8(skb, IPMRA_TABLE_MROUTE_DO_PIM, mrt->mroute_do_pim) ||
2765 nla_put_u8(skb, IPMRA_TABLE_MROUTE_DO_WRVIFWHOLE,
2766 mrt->mroute_do_wrvifwhole))
2767 return false;
2768
2769 return true;
2770}
2771
2772static bool ipmr_fill_vif(struct mr_table *mrt, u32 vifid, struct sk_buff *skb)
2773{
2774 struct nlattr *vif_nest;
2775 struct vif_device *vif;
2776
2777 /* if the VIF doesn't exist just continue */
2778 if (!VIF_EXISTS(mrt, vifid))
2779 return true;
2780
2781 vif = &mrt->vif_table[vifid];
2782 vif_nest = nla_nest_start_noflag(skb, IPMRA_VIF);
2783 if (!vif_nest)
2784 return false;
2785 if (nla_put_u32(skb, IPMRA_VIFA_IFINDEX, vif->dev->ifindex) ||
2786 nla_put_u32(skb, IPMRA_VIFA_VIF_ID, vifid) ||
2787 nla_put_u16(skb, IPMRA_VIFA_FLAGS, vif->flags) ||
2788 nla_put_u64_64bit(skb, IPMRA_VIFA_BYTES_IN, vif->bytes_in,
2789 IPMRA_VIFA_PAD) ||
2790 nla_put_u64_64bit(skb, IPMRA_VIFA_BYTES_OUT, vif->bytes_out,
2791 IPMRA_VIFA_PAD) ||
2792 nla_put_u64_64bit(skb, IPMRA_VIFA_PACKETS_IN, vif->pkt_in,
2793 IPMRA_VIFA_PAD) ||
2794 nla_put_u64_64bit(skb, IPMRA_VIFA_PACKETS_OUT, vif->pkt_out,
2795 IPMRA_VIFA_PAD) ||
2796 nla_put_be32(skb, IPMRA_VIFA_LOCAL_ADDR, vif->local) ||
2797 nla_put_be32(skb, IPMRA_VIFA_REMOTE_ADDR, vif->remote)) {
2798 nla_nest_cancel(skb, vif_nest);
2799 return false;
2800 }
2801 nla_nest_end(skb, vif_nest);
2802
2803 return true;
2804}
2805
2806static int ipmr_valid_dumplink(const struct nlmsghdr *nlh,
2807 struct netlink_ext_ack *extack)
2808{
2809 struct ifinfomsg *ifm;
2810
2811 if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*ifm))) {
2812 NL_SET_ERR_MSG(extack, "ipv4: Invalid header for ipmr link dump");
2813 return -EINVAL;
2814 }
2815
2816 if (nlmsg_attrlen(nlh, sizeof(*ifm))) {
2817 NL_SET_ERR_MSG(extack, "Invalid data after header in ipmr link dump");
2818 return -EINVAL;
2819 }
2820
2821 ifm = nlmsg_data(nlh);
2822 if (ifm->__ifi_pad || ifm->ifi_type || ifm->ifi_flags ||
2823 ifm->ifi_change || ifm->ifi_index) {
2824 NL_SET_ERR_MSG(extack, "Invalid values in header for ipmr link dump request");
2825 return -EINVAL;
2826 }
2827
2828 return 0;
2829}
2830
2831static int ipmr_rtm_dumplink(struct sk_buff *skb, struct netlink_callback *cb)
2832{
2833 struct net *net = sock_net(skb->sk);
2834 struct nlmsghdr *nlh = NULL;
2835 unsigned int t = 0, s_t;
2836 unsigned int e = 0, s_e;
2837 struct mr_table *mrt;
2838
2839 if (cb->strict_check) {
2840 int err = ipmr_valid_dumplink(cb->nlh, cb->extack);
2841
2842 if (err < 0)
2843 return err;
2844 }
2845
2846 s_t = cb->args[0];
2847 s_e = cb->args[1];
2848
2849 ipmr_for_each_table(mrt, net) {
2850 struct nlattr *vifs, *af;
2851 struct ifinfomsg *hdr;
2852 u32 i;
2853
2854 if (t < s_t)
2855 goto skip_table;
2856 nlh = nlmsg_put(skb, NETLINK_CB(cb->skb).portid,
2857 cb->nlh->nlmsg_seq, RTM_NEWLINK,
2858 sizeof(*hdr), NLM_F_MULTI);
2859 if (!nlh)
2860 break;
2861
2862 hdr = nlmsg_data(nlh);
2863 memset(hdr, 0, sizeof(*hdr));
2864 hdr->ifi_family = RTNL_FAMILY_IPMR;
2865
2866 af = nla_nest_start_noflag(skb, IFLA_AF_SPEC);
2867 if (!af) {
2868 nlmsg_cancel(skb, nlh);
2869 goto out;
2870 }
2871
2872 if (!ipmr_fill_table(mrt, skb)) {
2873 nlmsg_cancel(skb, nlh);
2874 goto out;
2875 }
2876
2877 vifs = nla_nest_start_noflag(skb, IPMRA_TABLE_VIFS);
2878 if (!vifs) {
2879 nla_nest_end(skb, af);
2880 nlmsg_end(skb, nlh);
2881 goto out;
2882 }
2883 for (i = 0; i < mrt->maxvif; i++) {
2884 if (e < s_e)
2885 goto skip_entry;
2886 if (!ipmr_fill_vif(mrt, i, skb)) {
2887 nla_nest_end(skb, vifs);
2888 nla_nest_end(skb, af);
2889 nlmsg_end(skb, nlh);
2890 goto out;
2891 }
2892skip_entry:
2893 e++;
2894 }
2895 s_e = 0;
2896 e = 0;
2897 nla_nest_end(skb, vifs);
2898 nla_nest_end(skb, af);
2899 nlmsg_end(skb, nlh);
2900skip_table:
2901 t++;
2902 }
2903
2904out:
2905 cb->args[1] = e;
2906 cb->args[0] = t;
2907
2908 return skb->len;
2909}
2910
2911#ifdef CONFIG_PROC_FS
2912/* The /proc interfaces to multicast routing :
2913 * /proc/net/ip_mr_cache & /proc/net/ip_mr_vif
2914 */
2915
2916static void *ipmr_vif_seq_start(struct seq_file *seq, loff_t *pos)
2917 __acquires(mrt_lock)
2918{
2919 struct mr_vif_iter *iter = seq->private;
2920 struct net *net = seq_file_net(seq);
2921 struct mr_table *mrt;
2922
2923 mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2924 if (!mrt)
2925 return ERR_PTR(-ENOENT);
2926
2927 iter->mrt = mrt;
2928
2929 read_lock(&mrt_lock);
2930 return mr_vif_seq_start(seq, pos);
2931}
2932
2933static void ipmr_vif_seq_stop(struct seq_file *seq, void *v)
2934 __releases(mrt_lock)
2935{
2936 read_unlock(&mrt_lock);
2937}
2938
2939static int ipmr_vif_seq_show(struct seq_file *seq, void *v)
2940{
2941 struct mr_vif_iter *iter = seq->private;
2942 struct mr_table *mrt = iter->mrt;
2943
2944 if (v == SEQ_START_TOKEN) {
2945 seq_puts(seq,
2946 "Interface BytesIn PktsIn BytesOut PktsOut Flags Local Remote\n");
2947 } else {
2948 const struct vif_device *vif = v;
2949 const char *name = vif->dev ?
2950 vif->dev->name : "none";
2951
2952 seq_printf(seq,
2953 "%2td %-10s %8ld %7ld %8ld %7ld %05X %08X %08X\n",
2954 vif - mrt->vif_table,
2955 name, vif->bytes_in, vif->pkt_in,
2956 vif->bytes_out, vif->pkt_out,
2957 vif->flags, vif->local, vif->remote);
2958 }
2959 return 0;
2960}
2961
2962static const struct seq_operations ipmr_vif_seq_ops = {
2963 .start = ipmr_vif_seq_start,
2964 .next = mr_vif_seq_next,
2965 .stop = ipmr_vif_seq_stop,
2966 .show = ipmr_vif_seq_show,
2967};
2968
2969static void *ipmr_mfc_seq_start(struct seq_file *seq, loff_t *pos)
2970{
2971 struct net *net = seq_file_net(seq);
2972 struct mr_table *mrt;
2973
2974 mrt = ipmr_get_table(net, RT_TABLE_DEFAULT);
2975 if (!mrt)
2976 return ERR_PTR(-ENOENT);
2977
2978 return mr_mfc_seq_start(seq, pos, mrt, &mfc_unres_lock);
2979}
2980
2981static int ipmr_mfc_seq_show(struct seq_file *seq, void *v)
2982{
2983 int n;
2984
2985 if (v == SEQ_START_TOKEN) {
2986 seq_puts(seq,
2987 "Group Origin Iif Pkts Bytes Wrong Oifs\n");
2988 } else {
2989 const struct mfc_cache *mfc = v;
2990 const struct mr_mfc_iter *it = seq->private;
2991 const struct mr_table *mrt = it->mrt;
2992
2993 seq_printf(seq, "%08X %08X %-3hd",
2994 (__force u32) mfc->mfc_mcastgrp,
2995 (__force u32) mfc->mfc_origin,
2996 mfc->_c.mfc_parent);
2997
2998 if (it->cache != &mrt->mfc_unres_queue) {
2999 seq_printf(seq, " %8lu %8lu %8lu",
3000 mfc->_c.mfc_un.res.pkt,
3001 mfc->_c.mfc_un.res.bytes,
3002 mfc->_c.mfc_un.res.wrong_if);
3003 for (n = mfc->_c.mfc_un.res.minvif;
3004 n < mfc->_c.mfc_un.res.maxvif; n++) {
3005 if (VIF_EXISTS(mrt, n) &&
3006 mfc->_c.mfc_un.res.ttls[n] < 255)
3007 seq_printf(seq,
3008 " %2d:%-3d",
3009 n, mfc->_c.mfc_un.res.ttls[n]);
3010 }
3011 } else {
3012 /* unresolved mfc_caches don't contain
3013 * pkt, bytes and wrong_if values
3014 */
3015 seq_printf(seq, " %8lu %8lu %8lu", 0ul, 0ul, 0ul);
3016 }
3017 seq_putc(seq, '\n');
3018 }
3019 return 0;
3020}
3021
3022static const struct seq_operations ipmr_mfc_seq_ops = {
3023 .start = ipmr_mfc_seq_start,
3024 .next = mr_mfc_seq_next,
3025 .stop = mr_mfc_seq_stop,
3026 .show = ipmr_mfc_seq_show,
3027};
3028#endif
3029
3030#ifdef CONFIG_IP_PIMSM_V2
3031static const struct net_protocol pim_protocol = {
3032 .handler = pim_rcv,
3033 .netns_ok = 1,
3034};
3035#endif
3036
3037static unsigned int ipmr_seq_read(struct net *net)
3038{
3039 ASSERT_RTNL();
3040
3041 return net->ipv4.ipmr_seq + ipmr_rules_seq_read(net);
3042}
3043
3044static int ipmr_dump(struct net *net, struct notifier_block *nb)
3045{
3046 return mr_dump(net, nb, RTNL_FAMILY_IPMR, ipmr_rules_dump,
3047 ipmr_mr_table_iter, &mrt_lock);
3048}
3049
3050static const struct fib_notifier_ops ipmr_notifier_ops_template = {
3051 .family = RTNL_FAMILY_IPMR,
3052 .fib_seq_read = ipmr_seq_read,
3053 .fib_dump = ipmr_dump,
3054 .owner = THIS_MODULE,
3055};
3056
3057static int __net_init ipmr_notifier_init(struct net *net)
3058{
3059 struct fib_notifier_ops *ops;
3060
3061 net->ipv4.ipmr_seq = 0;
3062
3063 ops = fib_notifier_ops_register(&ipmr_notifier_ops_template, net);
3064 if (IS_ERR(ops))
3065 return PTR_ERR(ops);
3066 net->ipv4.ipmr_notifier_ops = ops;
3067
3068 return 0;
3069}
3070
3071static void __net_exit ipmr_notifier_exit(struct net *net)
3072{
3073 fib_notifier_ops_unregister(net->ipv4.ipmr_notifier_ops);
3074 net->ipv4.ipmr_notifier_ops = NULL;
3075}
3076
3077/* Setup for IP multicast routing */
3078static int __net_init ipmr_net_init(struct net *net)
3079{
3080 int err;
3081
3082 err = ipmr_notifier_init(net);
3083 if (err)
3084 goto ipmr_notifier_fail;
3085
3086 err = ipmr_rules_init(net);
3087 if (err < 0)
3088 goto ipmr_rules_fail;
3089
3090#ifdef CONFIG_PROC_FS
3091 err = -ENOMEM;
3092 if (!proc_create_net("ip_mr_vif", 0, net->proc_net, &ipmr_vif_seq_ops,
3093 sizeof(struct mr_vif_iter)))
3094 goto proc_vif_fail;
3095 if (!proc_create_net("ip_mr_cache", 0, net->proc_net, &ipmr_mfc_seq_ops,
3096 sizeof(struct mr_mfc_iter)))
3097 goto proc_cache_fail;
3098#endif
3099 return 0;
3100
3101#ifdef CONFIG_PROC_FS
3102proc_cache_fail:
3103 remove_proc_entry("ip_mr_vif", net->proc_net);
3104proc_vif_fail:
3105 ipmr_rules_exit(net);
3106#endif
3107ipmr_rules_fail:
3108 ipmr_notifier_exit(net);
3109ipmr_notifier_fail:
3110 return err;
3111}
3112
3113static void __net_exit ipmr_net_exit(struct net *net)
3114{
3115#ifdef CONFIG_PROC_FS
3116 remove_proc_entry("ip_mr_cache", net->proc_net);
3117 remove_proc_entry("ip_mr_vif", net->proc_net);
3118#endif
3119 ipmr_notifier_exit(net);
3120 ipmr_rules_exit(net);
3121}
3122
3123static struct pernet_operations ipmr_net_ops = {
3124 .init = ipmr_net_init,
3125 .exit = ipmr_net_exit,
3126};
3127
3128int __init ip_mr_init(void)
3129{
3130 int err;
3131
3132 mrt_cachep = kmem_cache_create("ip_mrt_cache",
3133 sizeof(struct mfc_cache),
3134 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
3135 NULL);
3136
3137 err = register_pernet_subsys(&ipmr_net_ops);
3138 if (err)
3139 goto reg_pernet_fail;
3140
3141 err = register_netdevice_notifier(&ip_mr_notifier);
3142 if (err)
3143 goto reg_notif_fail;
3144#ifdef CONFIG_IP_PIMSM_V2
3145 if (inet_add_protocol(&pim_protocol, IPPROTO_PIM) < 0) {
3146 pr_err("%s: can't add PIM protocol\n", __func__);
3147 err = -EAGAIN;
3148 goto add_proto_fail;
3149 }
3150#endif
3151 rtnl_register(RTNL_FAMILY_IPMR, RTM_GETROUTE,
3152 ipmr_rtm_getroute, ipmr_rtm_dumproute, 0);
3153 rtnl_register(RTNL_FAMILY_IPMR, RTM_NEWROUTE,
3154 ipmr_rtm_route, NULL, 0);
3155 rtnl_register(RTNL_FAMILY_IPMR, RTM_DELROUTE,
3156 ipmr_rtm_route, NULL, 0);
3157
3158 rtnl_register(RTNL_FAMILY_IPMR, RTM_GETLINK,
3159 NULL, ipmr_rtm_dumplink, 0);
3160 return 0;
3161
3162#ifdef CONFIG_IP_PIMSM_V2
3163add_proto_fail:
3164 unregister_netdevice_notifier(&ip_mr_notifier);
3165#endif
3166reg_notif_fail:
3167 unregister_pernet_subsys(&ipmr_net_ops);
3168reg_pernet_fail:
3169 kmem_cache_destroy(mrt_cachep);
3170 return err;
3171}