blob: 92fc5f69f5da1c36aff62213d0c6bca358e624f7 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * IPv4 Forwarding Information Base: FIB frontend.
7 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License
12 * as published by the Free Software Foundation; either version
13 * 2 of the License, or (at your option) any later version.
14 */
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/module.h>
17#include <asm/uaccess.h>
18#include <asm/system.h>
19#include <linux/bitops.h>
Randy Dunlap4fc268d2006-01-11 12:17:47 -080020#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/types.h>
22#include <linux/kernel.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/mm.h>
24#include <linux/string.h>
25#include <linux/socket.h>
26#include <linux/sockios.h>
27#include <linux/errno.h>
28#include <linux/in.h>
29#include <linux/inet.h>
Arnaldo Carvalho de Melo14c85022005-12-27 02:43:12 -020030#include <linux/inetdevice.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#include <linux/netdevice.h>
Thomas Graf18237302006-08-04 23:04:54 -070032#include <linux/if_addr.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070033#include <linux/if_arp.h>
34#include <linux/skbuff.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include <linux/init.h>
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070036#include <linux/list.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090037#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038
39#include <net/ip.h>
40#include <net/protocol.h>
41#include <net/route.h>
42#include <net/tcp.h>
43#include <net/sock.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044#include <net/arp.h>
45#include <net/ip_fib.h>
Thomas Graf63f34442007-03-22 11:55:17 -070046#include <net/rtnetlink.h>
Michael Smith990078a2011-04-07 04:51:51 +000047#include <net/xfrm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#ifndef CONFIG_IP_MULTIPLE_TABLES
50
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -080051static int __net_init fib4_rules_init(struct net *net)
Pavel Emelyanovc3e9a352007-11-06 23:34:04 -080052{
Denis V. Lunev93456b62008-01-10 03:23:38 -080053 struct fib_table *local_table, *main_table;
54
David S. Miller5348ba82011-02-01 15:30:56 -080055 local_table = fib_trie_table(RT_TABLE_LOCAL);
Denis V. Lunev93456b62008-01-10 03:23:38 -080056 if (local_table == NULL)
Denis V. Lunevdbb50162008-01-10 03:21:49 -080057 return -ENOMEM;
58
David S. Miller5348ba82011-02-01 15:30:56 -080059 main_table = fib_trie_table(RT_TABLE_MAIN);
Denis V. Lunev93456b62008-01-10 03:23:38 -080060 if (main_table == NULL)
Denis V. Lunevdbb50162008-01-10 03:21:49 -080061 goto fail;
62
Denis V. Lunev93456b62008-01-10 03:23:38 -080063 hlist_add_head_rcu(&local_table->tb_hlist,
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080064 &net->ipv4.fib_table_hash[TABLE_LOCAL_INDEX]);
Denis V. Lunev93456b62008-01-10 03:23:38 -080065 hlist_add_head_rcu(&main_table->tb_hlist,
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080066 &net->ipv4.fib_table_hash[TABLE_MAIN_INDEX]);
Denis V. Lunevdbb50162008-01-10 03:21:49 -080067 return 0;
68
69fail:
Denis V. Lunev93456b62008-01-10 03:23:38 -080070 kfree(local_table);
Denis V. Lunevdbb50162008-01-10 03:21:49 -080071 return -ENOMEM;
Pavel Emelyanovc3e9a352007-11-06 23:34:04 -080072}
Linus Torvalds1da177e2005-04-16 15:20:36 -070073#else
74
Denis V. Lunev8ad49422008-01-10 03:24:11 -080075struct fib_table *fib_new_table(struct net *net, u32 id)
Linus Torvalds1da177e2005-04-16 15:20:36 -070076{
77 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070078 unsigned int h;
Linus Torvalds1da177e2005-04-16 15:20:36 -070079
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070080 if (id == 0)
81 id = RT_TABLE_MAIN;
Denis V. Lunev8ad49422008-01-10 03:24:11 -080082 tb = fib_get_table(net, id);
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070083 if (tb)
84 return tb;
Stephen Hemminger7f9b8052008-01-14 23:14:20 -080085
David S. Miller5348ba82011-02-01 15:30:56 -080086 tb = fib_trie_table(id);
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 if (!tb)
88 return NULL;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070089 h = id & (FIB_TABLE_HASHSZ - 1);
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080090 hlist_add_head_rcu(&tb->tb_hlist, &net->ipv4.fib_table_hash[h]);
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 return tb;
92}
93
Denis V. Lunev8ad49422008-01-10 03:24:11 -080094struct fib_table *fib_get_table(struct net *net, u32 id)
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070095{
96 struct fib_table *tb;
97 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080098 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070099 unsigned int h;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700101 if (id == 0)
102 id = RT_TABLE_MAIN;
103 h = id & (FIB_TABLE_HASHSZ - 1);
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800104
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700105 rcu_read_lock();
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800106 head = &net->ipv4.fib_table_hash[h];
107 hlist_for_each_entry_rcu(tb, node, head, tb_hlist) {
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700108 if (tb->tb_id == id) {
109 rcu_read_unlock();
110 return tb;
111 }
112 }
113 rcu_read_unlock();
114 return NULL;
115}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116#endif /* CONFIG_IP_MULTIPLE_TABLES */
117
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800118static void fib_flush(struct net *net)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119{
120 int flushed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700122 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800123 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700124 unsigned int h;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700126 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800127 head = &net->ipv4.fib_table_hash[h];
128 hlist_for_each_entry(tb, node, head, tb_hlist)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000129 flushed += fib_table_flush(tb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131
132 if (flushed)
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700133 rt_cache_flush(net, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134}
135
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800136/*
137 * Find address type as if only "dev" was present in the system. If
138 * on_dev is NULL then all interfaces are taken into consideration.
139 */
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800140static inline unsigned __inet_dev_addr_type(struct net *net,
141 const struct net_device *dev,
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800142 __be32 addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143{
David S. Miller9ade2282011-03-12 02:02:42 -0500144 struct flowi4 fl4 = { .daddr = addr };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 struct fib_result res;
146 unsigned ret = RTN_BROADCAST;
Pavel Emelyanov03cf7862007-10-23 21:17:27 -0700147 struct fib_table *local_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
Jan Engelhardt1e637c72008-01-21 03:18:08 -0800149 if (ipv4_is_zeronet(addr) || ipv4_is_lbcast(addr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 return RTN_BROADCAST;
Joe Perchesf97c1e02007-12-16 13:45:43 -0800151 if (ipv4_is_multicast(addr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 return RTN_MULTICAST;
153
154#ifdef CONFIG_IP_MULTIPLE_TABLES
155 res.r = NULL;
156#endif
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900157
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800158 local_table = fib_get_table(net, RT_TABLE_LOCAL);
Pavel Emelyanov03cf7862007-10-23 21:17:27 -0700159 if (local_table) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 ret = RTN_UNICAST;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000161 rcu_read_lock();
David S. Miller9ade2282011-03-12 02:02:42 -0500162 if (!fib_table_lookup(local_table, &fl4, &res, FIB_LOOKUP_NOREF)) {
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800163 if (!dev || dev == res.fi->fib_dev)
164 ret = res.type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 }
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000166 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 }
168 return ret;
169}
170
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800171unsigned int inet_addr_type(struct net *net, __be32 addr)
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800172{
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800173 return __inet_dev_addr_type(net, NULL, addr);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800174}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000175EXPORT_SYMBOL(inet_addr_type);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800176
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800177unsigned int inet_dev_addr_type(struct net *net, const struct net_device *dev,
178 __be32 addr)
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800179{
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000180 return __inet_dev_addr_type(net, dev, addr);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800181}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000182EXPORT_SYMBOL(inet_dev_addr_type);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800183
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184/* Given (packet source, input interface) and optional (dst, oif, tos):
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000185 * - (main) check, that source is valid i.e. not broadcast or our local
186 * address.
187 * - figure out what "logical" interface this packet arrived
188 * and calculate "specific destination" address.
189 * - check, that packet arrived from expected physical interface.
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000190 * called with rcu_read_lock()
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 */
Michael Smith5c04c812011-04-07 04:51:50 +0000192int fib_validate_source(struct sk_buff *skb, __be32 src, __be32 dst, u8 tos,
193 int oif, struct net_device *dev, __be32 *spec_dst,
194 u32 *itag)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195{
196 struct in_device *in_dev;
David S. Miller9ade2282011-03-12 02:02:42 -0500197 struct flowi4 fl4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 struct fib_result res;
Patrick McHardy8153a102009-12-03 01:25:58 +0000199 int no_addr, rpf, accept_local;
David S. Miller6f86b322010-09-06 22:36:19 -0700200 bool dev_match;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201 int ret;
Denis V. Lunev5b707aa2008-01-21 17:33:15 -0800202 struct net *net;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
David S. Miller9ade2282011-03-12 02:02:42 -0500204 fl4.flowi4_oif = 0;
205 fl4.flowi4_iif = oif;
David S. Miller9ade2282011-03-12 02:02:42 -0500206 fl4.daddr = src;
207 fl4.saddr = dst;
208 fl4.flowi4_tos = tos;
209 fl4.flowi4_scope = RT_SCOPE_UNIVERSE;
David S. Millercc7e17e2011-03-09 20:57:50 -0800210
Patrick McHardy8153a102009-12-03 01:25:58 +0000211 no_addr = rpf = accept_local = 0;
Herbert Xue5ed6392005-10-03 14:35:55 -0700212 in_dev = __in_dev_get_rcu(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 if (in_dev) {
214 no_addr = in_dev->ifa_list == NULL;
Michael Smith990078a2011-04-07 04:51:51 +0000215
216 /* Ignore rp_filter for packets protected by IPsec. */
217 rpf = secpath_exists(skb) ? 0 : IN_DEV_RPFILTER(in_dev);
218
Patrick McHardy8153a102009-12-03 01:25:58 +0000219 accept_local = IN_DEV_ACCEPT_LOCAL(in_dev);
Michael Smith5c04c812011-04-07 04:51:50 +0000220 fl4.flowi4_mark = IN_DEV_SRC_VMARK(in_dev) ? skb->mark : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222
223 if (in_dev == NULL)
224 goto e_inval;
225
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900226 net = dev_net(dev);
David S. Miller9ade2282011-03-12 02:02:42 -0500227 if (fib_lookup(net, &fl4, &res))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 goto last_resort;
Patrick McHardy8153a102009-12-03 01:25:58 +0000229 if (res.type != RTN_UNICAST) {
230 if (res.type != RTN_LOCAL || !accept_local)
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000231 goto e_inval;
Patrick McHardy8153a102009-12-03 01:25:58 +0000232 }
David S. Miller436c3b62011-03-24 17:42:21 -0700233 *spec_dst = FIB_RES_PREFSRC(net, res);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 fib_combine_itag(itag, &res);
David S. Miller6f86b322010-09-06 22:36:19 -0700235 dev_match = false;
236
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237#ifdef CONFIG_IP_ROUTE_MULTIPATH
David S. Miller6f86b322010-09-06 22:36:19 -0700238 for (ret = 0; ret < res.fi->fib_nhs; ret++) {
239 struct fib_nh *nh = &res.fi->fib_nh[ret];
240
241 if (nh->nh_dev == dev) {
242 dev_match = true;
243 break;
244 }
245 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246#else
247 if (FIB_RES_DEV(res) == dev)
David S. Miller6f86b322010-09-06 22:36:19 -0700248 dev_match = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249#endif
David S. Miller6f86b322010-09-06 22:36:19 -0700250 if (dev_match) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 ret = FIB_RES_NH(res).nh_scope >= RT_SCOPE_HOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 return ret;
253 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 if (no_addr)
255 goto last_resort;
Stephen Hemmingerc1cf8422009-02-20 08:25:36 +0000256 if (rpf == 1)
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000257 goto e_rpf;
David S. Miller9ade2282011-03-12 02:02:42 -0500258 fl4.flowi4_oif = dev->ifindex;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259
260 ret = 0;
David S. Miller9ade2282011-03-12 02:02:42 -0500261 if (fib_lookup(net, &fl4, &res) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 if (res.type == RTN_UNICAST) {
David S. Miller436c3b62011-03-24 17:42:21 -0700263 *spec_dst = FIB_RES_PREFSRC(net, res);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 ret = FIB_RES_NH(res).nh_scope >= RT_SCOPE_HOST;
265 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 }
267 return ret;
268
269last_resort:
270 if (rpf)
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000271 goto e_rpf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 *spec_dst = inet_select_addr(dev, 0, RT_SCOPE_UNIVERSE);
273 *itag = 0;
274 return 0;
275
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276e_inval:
277 return -EINVAL;
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000278e_rpf:
279 return -EXDEV;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280}
281
Al Viro81f7bf62006-09-27 18:40:00 -0700282static inline __be32 sk_extract_addr(struct sockaddr *addr)
Thomas Graf4e902c52006-08-17 18:14:52 -0700283{
284 return ((struct sockaddr_in *) addr)->sin_addr.s_addr;
285}
286
287static int put_rtax(struct nlattr *mx, int len, int type, u32 value)
288{
289 struct nlattr *nla;
290
291 nla = (struct nlattr *) ((char *) mx + len);
292 nla->nla_type = type;
293 nla->nla_len = nla_attr_size(4);
294 *(u32 *) nla_data(nla) = value;
295
296 return len + nla_total_size(4);
297}
298
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800299static int rtentry_to_fib_config(struct net *net, int cmd, struct rtentry *rt,
Thomas Graf4e902c52006-08-17 18:14:52 -0700300 struct fib_config *cfg)
301{
Al Viro6d85c102006-09-26 22:15:46 -0700302 __be32 addr;
Thomas Graf4e902c52006-08-17 18:14:52 -0700303 int plen;
304
305 memset(cfg, 0, sizeof(*cfg));
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800306 cfg->fc_nlinfo.nl_net = net;
Thomas Graf4e902c52006-08-17 18:14:52 -0700307
308 if (rt->rt_dst.sa_family != AF_INET)
309 return -EAFNOSUPPORT;
310
311 /*
312 * Check mask for validity:
313 * a) it must be contiguous.
314 * b) destination must have all host bits clear.
315 * c) if application forgot to set correct family (AF_INET),
316 * reject request unless it is absolutely clear i.e.
317 * both family and mask are zero.
318 */
319 plen = 32;
320 addr = sk_extract_addr(&rt->rt_dst);
321 if (!(rt->rt_flags & RTF_HOST)) {
Al Viro81f7bf62006-09-27 18:40:00 -0700322 __be32 mask = sk_extract_addr(&rt->rt_genmask);
Thomas Graf4e902c52006-08-17 18:14:52 -0700323
324 if (rt->rt_genmask.sa_family != AF_INET) {
325 if (mask || rt->rt_genmask.sa_family)
326 return -EAFNOSUPPORT;
327 }
328
329 if (bad_mask(mask, addr))
330 return -EINVAL;
331
332 plen = inet_mask_len(mask);
333 }
334
335 cfg->fc_dst_len = plen;
336 cfg->fc_dst = addr;
337
338 if (cmd != SIOCDELRT) {
339 cfg->fc_nlflags = NLM_F_CREATE;
340 cfg->fc_protocol = RTPROT_BOOT;
341 }
342
343 if (rt->rt_metric)
344 cfg->fc_priority = rt->rt_metric - 1;
345
346 if (rt->rt_flags & RTF_REJECT) {
347 cfg->fc_scope = RT_SCOPE_HOST;
348 cfg->fc_type = RTN_UNREACHABLE;
349 return 0;
350 }
351
352 cfg->fc_scope = RT_SCOPE_NOWHERE;
353 cfg->fc_type = RTN_UNICAST;
354
355 if (rt->rt_dev) {
356 char *colon;
357 struct net_device *dev;
358 char devname[IFNAMSIZ];
359
360 if (copy_from_user(devname, rt->rt_dev, IFNAMSIZ-1))
361 return -EFAULT;
362
363 devname[IFNAMSIZ-1] = 0;
364 colon = strchr(devname, ':');
365 if (colon)
366 *colon = 0;
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800367 dev = __dev_get_by_name(net, devname);
Thomas Graf4e902c52006-08-17 18:14:52 -0700368 if (!dev)
369 return -ENODEV;
370 cfg->fc_oif = dev->ifindex;
371 if (colon) {
372 struct in_ifaddr *ifa;
373 struct in_device *in_dev = __in_dev_get_rtnl(dev);
374 if (!in_dev)
375 return -ENODEV;
376 *colon = ':';
377 for (ifa = in_dev->ifa_list; ifa; ifa = ifa->ifa_next)
378 if (strcmp(ifa->ifa_label, devname) == 0)
379 break;
380 if (ifa == NULL)
381 return -ENODEV;
382 cfg->fc_prefsrc = ifa->ifa_local;
383 }
384 }
385
386 addr = sk_extract_addr(&rt->rt_gateway);
387 if (rt->rt_gateway.sa_family == AF_INET && addr) {
388 cfg->fc_gw = addr;
389 if (rt->rt_flags & RTF_GATEWAY &&
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800390 inet_addr_type(net, addr) == RTN_UNICAST)
Thomas Graf4e902c52006-08-17 18:14:52 -0700391 cfg->fc_scope = RT_SCOPE_UNIVERSE;
392 }
393
394 if (cmd == SIOCDELRT)
395 return 0;
396
397 if (rt->rt_flags & RTF_GATEWAY && !cfg->fc_gw)
398 return -EINVAL;
399
400 if (cfg->fc_scope == RT_SCOPE_NOWHERE)
401 cfg->fc_scope = RT_SCOPE_LINK;
402
403 if (rt->rt_flags & (RTF_MTU | RTF_WINDOW | RTF_IRTT)) {
404 struct nlattr *mx;
405 int len = 0;
406
407 mx = kzalloc(3 * nla_total_size(4), GFP_KERNEL);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900408 if (mx == NULL)
Thomas Graf4e902c52006-08-17 18:14:52 -0700409 return -ENOMEM;
410
411 if (rt->rt_flags & RTF_MTU)
412 len = put_rtax(mx, len, RTAX_ADVMSS, rt->rt_mtu - 40);
413
414 if (rt->rt_flags & RTF_WINDOW)
415 len = put_rtax(mx, len, RTAX_WINDOW, rt->rt_window);
416
417 if (rt->rt_flags & RTF_IRTT)
418 len = put_rtax(mx, len, RTAX_RTT, rt->rt_irtt << 3);
419
420 cfg->fc_mx = mx;
421 cfg->fc_mx_len = len;
422 }
423
424 return 0;
425}
426
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427/*
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000428 * Handle IP routing ioctl calls.
429 * These are used to manipulate the routing tables
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 */
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800431int ip_rt_ioctl(struct net *net, unsigned int cmd, void __user *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
Thomas Graf4e902c52006-08-17 18:14:52 -0700433 struct fib_config cfg;
434 struct rtentry rt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436
437 switch (cmd) {
438 case SIOCADDRT: /* Add a route */
439 case SIOCDELRT: /* Delete a route */
440 if (!capable(CAP_NET_ADMIN))
441 return -EPERM;
Thomas Graf4e902c52006-08-17 18:14:52 -0700442
443 if (copy_from_user(&rt, arg, sizeof(rt)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 return -EFAULT;
Thomas Graf4e902c52006-08-17 18:14:52 -0700445
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 rtnl_lock();
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800447 err = rtentry_to_fib_config(net, cmd, &rt, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 if (err == 0) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700449 struct fib_table *tb;
450
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 if (cmd == SIOCDELRT) {
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800452 tb = fib_get_table(net, cfg.fc_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 if (tb)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000454 err = fib_table_delete(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700455 else
456 err = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 } else {
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800458 tb = fib_new_table(net, cfg.fc_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 if (tb)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000460 err = fib_table_insert(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700461 else
462 err = -ENOBUFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700464
465 /* allocated by rtentry_to_fib_config() */
466 kfree(cfg.fc_mx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 }
468 rtnl_unlock();
469 return err;
470 }
471 return -EINVAL;
472}
473
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000474const struct nla_policy rtm_ipv4_policy[RTA_MAX + 1] = {
Thomas Graf4e902c52006-08-17 18:14:52 -0700475 [RTA_DST] = { .type = NLA_U32 },
476 [RTA_SRC] = { .type = NLA_U32 },
477 [RTA_IIF] = { .type = NLA_U32 },
478 [RTA_OIF] = { .type = NLA_U32 },
479 [RTA_GATEWAY] = { .type = NLA_U32 },
480 [RTA_PRIORITY] = { .type = NLA_U32 },
481 [RTA_PREFSRC] = { .type = NLA_U32 },
482 [RTA_METRICS] = { .type = NLA_NESTED },
Thomas Graf5176f912006-08-26 20:13:18 -0700483 [RTA_MULTIPATH] = { .len = sizeof(struct rtnexthop) },
Thomas Graf4e902c52006-08-17 18:14:52 -0700484 [RTA_FLOW] = { .type = NLA_U32 },
Thomas Graf4e902c52006-08-17 18:14:52 -0700485};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800487static int rtm_to_fib_config(struct net *net, struct sk_buff *skb,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000488 struct nlmsghdr *nlh, struct fib_config *cfg)
Thomas Graf4e902c52006-08-17 18:14:52 -0700489{
490 struct nlattr *attr;
491 int err, remaining;
492 struct rtmsg *rtm;
493
494 err = nlmsg_validate(nlh, sizeof(*rtm), RTA_MAX, rtm_ipv4_policy);
495 if (err < 0)
496 goto errout;
497
498 memset(cfg, 0, sizeof(*cfg));
499
500 rtm = nlmsg_data(nlh);
Thomas Graf4e902c52006-08-17 18:14:52 -0700501 cfg->fc_dst_len = rtm->rtm_dst_len;
Thomas Graf4e902c52006-08-17 18:14:52 -0700502 cfg->fc_tos = rtm->rtm_tos;
503 cfg->fc_table = rtm->rtm_table;
504 cfg->fc_protocol = rtm->rtm_protocol;
505 cfg->fc_scope = rtm->rtm_scope;
506 cfg->fc_type = rtm->rtm_type;
507 cfg->fc_flags = rtm->rtm_flags;
508 cfg->fc_nlflags = nlh->nlmsg_flags;
509
510 cfg->fc_nlinfo.pid = NETLINK_CB(skb).pid;
511 cfg->fc_nlinfo.nlh = nlh;
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800512 cfg->fc_nlinfo.nl_net = net;
Thomas Graf4e902c52006-08-17 18:14:52 -0700513
Thomas Grafa0ee18b2007-03-24 20:32:54 -0700514 if (cfg->fc_type > RTN_MAX) {
515 err = -EINVAL;
516 goto errout;
517 }
518
Thomas Graf4e902c52006-08-17 18:14:52 -0700519 nlmsg_for_each_attr(attr, nlh, sizeof(struct rtmsg), remaining) {
Thomas Graf8f4c1f92007-09-12 14:44:36 +0200520 switch (nla_type(attr)) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700521 case RTA_DST:
Al Viro17fb2c62006-09-26 22:15:25 -0700522 cfg->fc_dst = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700523 break;
Thomas Graf4e902c52006-08-17 18:14:52 -0700524 case RTA_OIF:
525 cfg->fc_oif = nla_get_u32(attr);
526 break;
527 case RTA_GATEWAY:
Al Viro17fb2c62006-09-26 22:15:25 -0700528 cfg->fc_gw = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700529 break;
530 case RTA_PRIORITY:
531 cfg->fc_priority = nla_get_u32(attr);
532 break;
533 case RTA_PREFSRC:
Al Viro17fb2c62006-09-26 22:15:25 -0700534 cfg->fc_prefsrc = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700535 break;
536 case RTA_METRICS:
537 cfg->fc_mx = nla_data(attr);
538 cfg->fc_mx_len = nla_len(attr);
539 break;
540 case RTA_MULTIPATH:
541 cfg->fc_mp = nla_data(attr);
542 cfg->fc_mp_len = nla_len(attr);
543 break;
544 case RTA_FLOW:
545 cfg->fc_flow = nla_get_u32(attr);
546 break;
Thomas Graf4e902c52006-08-17 18:14:52 -0700547 case RTA_TABLE:
548 cfg->fc_table = nla_get_u32(attr);
549 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 }
551 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700552
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 return 0;
Thomas Graf4e902c52006-08-17 18:14:52 -0700554errout:
555 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556}
557
Jianjun Kong6ed25332008-11-03 00:25:16 -0800558static int inet_rtm_delroute(struct sk_buff *skb, struct nlmsghdr *nlh, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900560 struct net *net = sock_net(skb->sk);
Thomas Graf4e902c52006-08-17 18:14:52 -0700561 struct fib_config cfg;
562 struct fib_table *tb;
563 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800565 err = rtm_to_fib_config(net, skb, nlh, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700566 if (err < 0)
567 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568
Denis V. Lunev8ad49422008-01-10 03:24:11 -0800569 tb = fib_get_table(net, cfg.fc_table);
Thomas Graf4e902c52006-08-17 18:14:52 -0700570 if (tb == NULL) {
571 err = -ESRCH;
572 goto errout;
573 }
574
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000575 err = fib_table_delete(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700576errout:
577 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578}
579
Jianjun Kong6ed25332008-11-03 00:25:16 -0800580static int inet_rtm_newroute(struct sk_buff *skb, struct nlmsghdr *nlh, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900582 struct net *net = sock_net(skb->sk);
Thomas Graf4e902c52006-08-17 18:14:52 -0700583 struct fib_config cfg;
584 struct fib_table *tb;
585 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800587 err = rtm_to_fib_config(net, skb, nlh, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700588 if (err < 0)
589 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590
Denis V. Lunev226b0b42008-01-10 03:30:24 -0800591 tb = fib_new_table(net, cfg.fc_table);
Thomas Graf4e902c52006-08-17 18:14:52 -0700592 if (tb == NULL) {
593 err = -ENOBUFS;
594 goto errout;
595 }
596
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000597 err = fib_table_insert(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700598errout:
599 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600}
601
Thomas Graf63f34442007-03-22 11:55:17 -0700602static int inet_dump_fib(struct sk_buff *skb, struct netlink_callback *cb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900604 struct net *net = sock_net(skb->sk);
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700605 unsigned int h, s_h;
606 unsigned int e = 0, s_e;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700608 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800609 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700610 int dumped = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Thomas Grafbe403ea2006-08-17 18:15:17 -0700612 if (nlmsg_len(cb->nlh) >= sizeof(struct rtmsg) &&
613 ((struct rtmsg *) nlmsg_data(cb->nlh))->rtm_flags & RTM_F_CLONED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 return ip_rt_dump(skb, cb);
615
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700616 s_h = cb->args[0];
617 s_e = cb->args[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700619 for (h = s_h; h < FIB_TABLE_HASHSZ; h++, s_e = 0) {
620 e = 0;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800621 head = &net->ipv4.fib_table_hash[h];
622 hlist_for_each_entry(tb, node, head, tb_hlist) {
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700623 if (e < s_e)
624 goto next;
625 if (dumped)
626 memset(&cb->args[2], 0, sizeof(cb->args) -
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900627 2 * sizeof(cb->args[0]));
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000628 if (fib_table_dump(tb, skb, cb) < 0)
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700629 goto out;
630 dumped = 1;
631next:
632 e++;
633 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 }
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700635out:
636 cb->args[1] = e;
637 cb->args[0] = h;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638
639 return skb->len;
640}
641
642/* Prepare and feed intra-kernel routing request.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000643 * Really, it should be netlink message, but :-( netlink
644 * can be not configured, so that we feed it directly
645 * to fib engine. It is legal, because all events occur
646 * only when netlink is already locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 */
Al Viro81f7bf62006-09-27 18:40:00 -0700648static void fib_magic(int cmd, int type, __be32 dst, int dst_len, struct in_ifaddr *ifa)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649{
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900650 struct net *net = dev_net(ifa->ifa_dev->dev);
Thomas Graf4e902c52006-08-17 18:14:52 -0700651 struct fib_table *tb;
652 struct fib_config cfg = {
653 .fc_protocol = RTPROT_KERNEL,
654 .fc_type = type,
655 .fc_dst = dst,
656 .fc_dst_len = dst_len,
657 .fc_prefsrc = ifa->ifa_local,
658 .fc_oif = ifa->ifa_dev->dev->ifindex,
659 .fc_nlflags = NLM_F_CREATE | NLM_F_APPEND,
Denis V. Lunev4d1169c2008-01-10 03:26:13 -0800660 .fc_nlinfo = {
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800661 .nl_net = net,
Denis V. Lunev4d1169c2008-01-10 03:26:13 -0800662 },
Thomas Graf4e902c52006-08-17 18:14:52 -0700663 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665 if (type == RTN_UNICAST)
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800666 tb = fib_new_table(net, RT_TABLE_MAIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 else
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800668 tb = fib_new_table(net, RT_TABLE_LOCAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
670 if (tb == NULL)
671 return;
672
Thomas Graf4e902c52006-08-17 18:14:52 -0700673 cfg.fc_table = tb->tb_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674
Thomas Graf4e902c52006-08-17 18:14:52 -0700675 if (type != RTN_LOCAL)
676 cfg.fc_scope = RT_SCOPE_LINK;
677 else
678 cfg.fc_scope = RT_SCOPE_HOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679
680 if (cmd == RTM_NEWROUTE)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000681 fib_table_insert(tb, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 else
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000683 fib_table_delete(tb, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684}
685
Jamal Hadi Salim0ff60a42005-11-22 14:47:37 -0800686void fib_add_ifaddr(struct in_ifaddr *ifa)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687{
688 struct in_device *in_dev = ifa->ifa_dev;
689 struct net_device *dev = in_dev->dev;
690 struct in_ifaddr *prim = ifa;
Al Viroa144ea42006-09-28 18:00:55 -0700691 __be32 mask = ifa->ifa_mask;
692 __be32 addr = ifa->ifa_local;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000693 __be32 prefix = ifa->ifa_address & mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000695 if (ifa->ifa_flags & IFA_F_SECONDARY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 prim = inet_ifa_byprefix(in_dev, prefix, mask);
697 if (prim == NULL) {
Stephen Hemmingera6db9012008-01-12 20:58:35 -0800698 printk(KERN_WARNING "fib_add_ifaddr: bug: prim == NULL\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 return;
700 }
701 }
702
703 fib_magic(RTM_NEWROUTE, RTN_LOCAL, addr, 32, prim);
704
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000705 if (!(dev->flags & IFF_UP))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 return;
707
708 /* Add broadcast address, if it is explicitly assigned. */
Al Viroa144ea42006-09-28 18:00:55 -0700709 if (ifa->ifa_broadcast && ifa->ifa_broadcast != htonl(0xFFFFFFFF))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, ifa->ifa_broadcast, 32, prim);
711
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000712 if (!ipv4_is_zeronet(prefix) && !(ifa->ifa_flags & IFA_F_SECONDARY) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 (prefix != addr || ifa->ifa_prefixlen < 32)) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000714 fib_magic(RTM_NEWROUTE,
715 dev->flags & IFF_LOOPBACK ? RTN_LOCAL : RTN_UNICAST,
716 prefix, ifa->ifa_prefixlen, prim);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717
718 /* Add network specific broadcasts, when it takes a sense */
719 if (ifa->ifa_prefixlen < 31) {
720 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, prefix, 32, prim);
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000721 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, prefix | ~mask,
722 32, prim);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 }
724 }
725}
726
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000727/* Delete primary or secondary address.
728 * Optionally, on secondary address promotion consider the addresses
729 * from subnet iprim as deleted, even if they are in device list.
730 * In this case the secondary ifa can be in device list.
731 */
732void fib_del_ifaddr(struct in_ifaddr *ifa, struct in_ifaddr *iprim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733{
734 struct in_device *in_dev = ifa->ifa_dev;
735 struct net_device *dev = in_dev->dev;
736 struct in_ifaddr *ifa1;
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000737 struct in_ifaddr *prim = ifa, *prim1 = NULL;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000738 __be32 brd = ifa->ifa_address | ~ifa->ifa_mask;
739 __be32 any = ifa->ifa_address & ifa->ifa_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740#define LOCAL_OK 1
741#define BRD_OK 2
742#define BRD0_OK 4
743#define BRD1_OK 8
744 unsigned ok = 0;
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000745 int subnet = 0; /* Primary network */
746 int gone = 1; /* Address is missing */
747 int same_prefsrc = 0; /* Another primary with same IP */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000749 if (ifa->ifa_flags & IFA_F_SECONDARY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 prim = inet_ifa_byprefix(in_dev, any, ifa->ifa_mask);
751 if (prim == NULL) {
Stephen Hemmingera6db9012008-01-12 20:58:35 -0800752 printk(KERN_WARNING "fib_del_ifaddr: bug: prim == NULL\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 return;
754 }
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000755 if (iprim && iprim != prim) {
756 printk(KERN_WARNING "fib_del_ifaddr: bug: iprim != prim\n");
757 return;
758 }
759 } else if (!ipv4_is_zeronet(any) &&
760 (any != ifa->ifa_local || ifa->ifa_prefixlen < 32)) {
761 fib_magic(RTM_DELROUTE,
762 dev->flags & IFF_LOOPBACK ? RTN_LOCAL : RTN_UNICAST,
763 any, ifa->ifa_prefixlen, prim);
764 subnet = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 }
766
767 /* Deletion is more complicated than add.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000768 * We should take care of not to delete too much :-)
769 *
770 * Scan address list to be sure that addresses are really gone.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 */
772
773 for (ifa1 = in_dev->ifa_list; ifa1; ifa1 = ifa1->ifa_next) {
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000774 if (ifa1 == ifa) {
775 /* promotion, keep the IP */
776 gone = 0;
777 continue;
778 }
779 /* Ignore IFAs from our subnet */
780 if (iprim && ifa1->ifa_mask == iprim->ifa_mask &&
781 inet_ifa_match(ifa1->ifa_address, iprim))
782 continue;
783
784 /* Ignore ifa1 if it uses different primary IP (prefsrc) */
785 if (ifa1->ifa_flags & IFA_F_SECONDARY) {
786 /* Another address from our subnet? */
787 if (ifa1->ifa_mask == prim->ifa_mask &&
788 inet_ifa_match(ifa1->ifa_address, prim))
789 prim1 = prim;
790 else {
791 /* We reached the secondaries, so
792 * same_prefsrc should be determined.
793 */
794 if (!same_prefsrc)
795 continue;
796 /* Search new prim1 if ifa1 is not
797 * using the current prim1
798 */
799 if (!prim1 ||
800 ifa1->ifa_mask != prim1->ifa_mask ||
801 !inet_ifa_match(ifa1->ifa_address, prim1))
802 prim1 = inet_ifa_byprefix(in_dev,
803 ifa1->ifa_address,
804 ifa1->ifa_mask);
805 if (!prim1)
806 continue;
807 if (prim1->ifa_local != prim->ifa_local)
808 continue;
809 }
810 } else {
811 if (prim->ifa_local != ifa1->ifa_local)
812 continue;
813 prim1 = ifa1;
814 if (prim != prim1)
815 same_prefsrc = 1;
816 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 if (ifa->ifa_local == ifa1->ifa_local)
818 ok |= LOCAL_OK;
819 if (ifa->ifa_broadcast == ifa1->ifa_broadcast)
820 ok |= BRD_OK;
821 if (brd == ifa1->ifa_broadcast)
822 ok |= BRD1_OK;
823 if (any == ifa1->ifa_broadcast)
824 ok |= BRD0_OK;
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000825 /* primary has network specific broadcasts */
826 if (prim1 == ifa1 && ifa1->ifa_prefixlen < 31) {
827 __be32 brd1 = ifa1->ifa_address | ~ifa1->ifa_mask;
828 __be32 any1 = ifa1->ifa_address & ifa1->ifa_mask;
829
830 if (!ipv4_is_zeronet(any1)) {
831 if (ifa->ifa_broadcast == brd1 ||
832 ifa->ifa_broadcast == any1)
833 ok |= BRD_OK;
834 if (brd == brd1 || brd == any1)
835 ok |= BRD1_OK;
836 if (any == brd1 || any == any1)
837 ok |= BRD0_OK;
838 }
839 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 }
841
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000842 if (!(ok & BRD_OK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 fib_magic(RTM_DELROUTE, RTN_BROADCAST, ifa->ifa_broadcast, 32, prim);
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000844 if (subnet && ifa->ifa_prefixlen < 31) {
845 if (!(ok & BRD1_OK))
846 fib_magic(RTM_DELROUTE, RTN_BROADCAST, brd, 32, prim);
847 if (!(ok & BRD0_OK))
848 fib_magic(RTM_DELROUTE, RTN_BROADCAST, any, 32, prim);
849 }
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000850 if (!(ok & LOCAL_OK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 fib_magic(RTM_DELROUTE, RTN_LOCAL, ifa->ifa_local, 32, prim);
852
853 /* Check, that this local address finally disappeared. */
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000854 if (gone &&
855 inet_addr_type(dev_net(dev), ifa->ifa_local) != RTN_LOCAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 /* And the last, but not the least thing.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000857 * We must flush stray FIB entries.
858 *
859 * First of all, we scan fib_info list searching
860 * for stray nexthop entries, then ignite fib_flush.
861 */
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900862 if (fib_sync_down_addr(dev_net(dev), ifa->ifa_local))
863 fib_flush(dev_net(dev));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 }
865 }
866#undef LOCAL_OK
867#undef BRD_OK
868#undef BRD0_OK
869#undef BRD1_OK
870}
871
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000872static void nl_fib_lookup(struct fib_result_nl *frn, struct fib_table *tb)
Robert Olsson246955f2005-06-20 13:36:39 -0700873{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900874
Robert Olsson246955f2005-06-20 13:36:39 -0700875 struct fib_result res;
David S. Miller9ade2282011-03-12 02:02:42 -0500876 struct flowi4 fl4 = {
877 .flowi4_mark = frn->fl_mark,
878 .daddr = frn->fl_addr,
879 .flowi4_tos = frn->fl_tos,
880 .flowi4_scope = frn->fl_scope,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000881 };
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700882
Sergey Vlasov912a41a2007-04-27 02:17:19 -0700883#ifdef CONFIG_IP_MULTIPLE_TABLES
884 res.r = NULL;
885#endif
886
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700887 frn->err = -ENOENT;
Robert Olsson246955f2005-06-20 13:36:39 -0700888 if (tb) {
889 local_bh_disable();
890
891 frn->tb_id = tb->tb_id;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000892 rcu_read_lock();
David S. Miller9ade2282011-03-12 02:02:42 -0500893 frn->err = fib_table_lookup(tb, &fl4, &res, FIB_LOOKUP_NOREF);
Robert Olsson246955f2005-06-20 13:36:39 -0700894
895 if (!frn->err) {
896 frn->prefixlen = res.prefixlen;
897 frn->nh_sel = res.nh_sel;
898 frn->type = res.type;
899 frn->scope = res.scope;
900 }
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000901 rcu_read_unlock();
Robert Olsson246955f2005-06-20 13:36:39 -0700902 local_bh_enable();
903 }
904}
905
David S. Miller28f7b032007-10-10 21:32:39 -0700906static void nl_fib_input(struct sk_buff *skb)
Robert Olsson246955f2005-06-20 13:36:39 -0700907{
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800908 struct net *net;
Robert Olsson246955f2005-06-20 13:36:39 -0700909 struct fib_result_nl *frn;
David S. Miller28f7b032007-10-10 21:32:39 -0700910 struct nlmsghdr *nlh;
Robert Olsson246955f2005-06-20 13:36:39 -0700911 struct fib_table *tb;
David S. Miller28f7b032007-10-10 21:32:39 -0700912 u32 pid;
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700913
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900914 net = sock_net(skb->sk);
Arnaldo Carvalho de Melob529ccf2007-04-25 19:08:35 -0700915 nlh = nlmsg_hdr(skb);
Thomas Grafea865752005-12-01 14:30:00 -0800916 if (skb->len < NLMSG_SPACE(0) || skb->len < nlh->nlmsg_len ||
Denis V. Lunevd883a032007-12-21 02:01:53 -0800917 nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*frn)))
Thomas Grafea865752005-12-01 14:30:00 -0800918 return;
Denis V. Lunevd883a032007-12-21 02:01:53 -0800919
920 skb = skb_clone(skb, GFP_KERNEL);
921 if (skb == NULL)
922 return;
923 nlh = nlmsg_hdr(skb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900924
Robert Olsson246955f2005-06-20 13:36:39 -0700925 frn = (struct fib_result_nl *) NLMSG_DATA(nlh);
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800926 tb = fib_get_table(net, frn->tb_id_in);
Robert Olsson246955f2005-06-20 13:36:39 -0700927
928 nl_fib_lookup(frn, tb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900929
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000930 pid = NETLINK_CB(skb).pid; /* pid of sending process */
931 NETLINK_CB(skb).pid = 0; /* from kernel */
Patrick McHardyac6d4392005-08-14 19:29:52 -0700932 NETLINK_CB(skb).dst_group = 0; /* unicast */
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800933 netlink_unicast(net->ipv4.fibnl, skb, pid, MSG_DONTWAIT);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900934}
Robert Olsson246955f2005-06-20 13:36:39 -0700935
Alexey Dobriyan2c8c1e72010-01-17 03:35:32 +0000936static int __net_init nl_fib_lookup_init(struct net *net)
Robert Olsson246955f2005-06-20 13:36:39 -0700937{
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800938 struct sock *sk;
939 sk = netlink_kernel_create(net, NETLINK_FIB_LOOKUP, 0,
940 nl_fib_input, NULL, THIS_MODULE);
941 if (sk == NULL)
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -0800942 return -EAFNOSUPPORT;
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800943 net->ipv4.fibnl = sk;
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -0800944 return 0;
945}
946
947static void nl_fib_lookup_exit(struct net *net)
948{
Denis V. Lunevb7c6ba62008-01-28 14:41:19 -0800949 netlink_kernel_release(net->ipv4.fibnl);
Denis V. Lunev775516b2008-01-18 23:55:19 -0800950 net->ipv4.fibnl = NULL;
Robert Olsson246955f2005-06-20 13:36:39 -0700951}
952
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000953static void fib_disable_ip(struct net_device *dev, int force, int delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954{
Denis V. Lunev85326fa2008-01-31 18:48:47 -0800955 if (fib_sync_down_dev(dev, force))
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900956 fib_flush(dev_net(dev));
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000957 rt_cache_flush(dev_net(dev), delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 arp_ifdown(dev);
959}
960
961static int fib_inetaddr_event(struct notifier_block *this, unsigned long event, void *ptr)
962{
Jianjun Kong6ed25332008-11-03 00:25:16 -0800963 struct in_ifaddr *ifa = (struct in_ifaddr *)ptr;
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700964 struct net_device *dev = ifa->ifa_dev->dev;
David S. Miller436c3b62011-03-24 17:42:21 -0700965 struct net *net = dev_net(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966
967 switch (event) {
968 case NETDEV_UP:
969 fib_add_ifaddr(ifa);
970#ifdef CONFIG_IP_ROUTE_MULTIPATH
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700971 fib_sync_up(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972#endif
David S. Miller436c3b62011-03-24 17:42:21 -0700973 atomic_inc(&net->ipv4.dev_addr_genid);
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700974 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 break;
976 case NETDEV_DOWN:
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000977 fib_del_ifaddr(ifa, NULL);
David S. Miller436c3b62011-03-24 17:42:21 -0700978 atomic_inc(&net->ipv4.dev_addr_genid);
Jayachandran C9fcc2e82005-10-27 15:10:01 -0700979 if (ifa->ifa_dev->ifa_list == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 /* Last address was deleted from this interface.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000981 * Disable IP.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 */
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000983 fib_disable_ip(dev, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 } else {
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700985 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 }
987 break;
988 }
989 return NOTIFY_DONE;
990}
991
992static int fib_netdev_event(struct notifier_block *this, unsigned long event, void *ptr)
993{
994 struct net_device *dev = ptr;
Herbert Xue5ed6392005-10-03 14:35:55 -0700995 struct in_device *in_dev = __in_dev_get_rtnl(dev);
David S. Miller436c3b62011-03-24 17:42:21 -0700996 struct net *net = dev_net(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997
998 if (event == NETDEV_UNREGISTER) {
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000999 fib_disable_ip(dev, 2, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 return NOTIFY_DONE;
1001 }
1002
1003 if (!in_dev)
1004 return NOTIFY_DONE;
1005
1006 switch (event) {
1007 case NETDEV_UP:
1008 for_ifa(in_dev) {
1009 fib_add_ifaddr(ifa);
1010 } endfor_ifa(in_dev);
1011#ifdef CONFIG_IP_ROUTE_MULTIPATH
1012 fib_sync_up(dev);
1013#endif
David S. Miller436c3b62011-03-24 17:42:21 -07001014 atomic_inc(&net->ipv4.dev_addr_genid);
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -07001015 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016 break;
1017 case NETDEV_DOWN:
Octavian Purdilae2ce1462009-11-16 13:49:49 +00001018 fib_disable_ip(dev, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 break;
1020 case NETDEV_CHANGEMTU:
1021 case NETDEV_CHANGE:
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -07001022 rt_cache_flush(dev_net(dev), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 break;
Eric W. Biedermana5ee1552009-11-29 15:45:58 +00001024 case NETDEV_UNREGISTER_BATCH:
David S. Miller6561a3b2010-12-19 21:11:20 -08001025 /* The batch unregister is only called on the first
1026 * device in the list of devices being unregistered.
1027 * Therefore we should not pass dev_net(dev) in here.
1028 */
1029 rt_cache_flush_batch(NULL);
Eric W. Biedermana5ee1552009-11-29 15:45:58 +00001030 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 }
1032 return NOTIFY_DONE;
1033}
1034
1035static struct notifier_block fib_inetaddr_notifier = {
Jianjun Kong6ed25332008-11-03 00:25:16 -08001036 .notifier_call = fib_inetaddr_event,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037};
1038
1039static struct notifier_block fib_netdev_notifier = {
Jianjun Kong6ed25332008-11-03 00:25:16 -08001040 .notifier_call = fib_netdev_event,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041};
1042
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001043static int __net_init ip_fib_net_init(struct net *net)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044{
Denis V. Lunevdce5cbe2008-01-31 18:44:53 -08001045 int err;
Eric Dumazet10da66f2010-10-13 08:22:03 +00001046 size_t size = sizeof(struct hlist_head) * FIB_TABLE_HASHSZ;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -07001047
Eric Dumazet10da66f2010-10-13 08:22:03 +00001048 /* Avoid false sharing : Use at least a full cache line */
1049 size = max_t(size_t, size, L1_CACHE_BYTES);
1050
1051 net->ipv4.fib_table_hash = kzalloc(size, GFP_KERNEL);
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001052 if (net->ipv4.fib_table_hash == NULL)
1053 return -ENOMEM;
1054
Denis V. Lunevdce5cbe2008-01-31 18:44:53 -08001055 err = fib4_rules_init(net);
1056 if (err < 0)
1057 goto fail;
1058 return 0;
1059
1060fail:
1061 kfree(net->ipv4.fib_table_hash);
1062 return err;
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001063}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064
Alexey Dobriyan2c8c1e72010-01-17 03:35:32 +00001065static void ip_fib_net_exit(struct net *net)
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001066{
1067 unsigned int i;
Thomas Graf63f34442007-03-22 11:55:17 -07001068
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001069#ifdef CONFIG_IP_MULTIPLE_TABLES
1070 fib4_rules_exit(net);
1071#endif
1072
Eric Dumazete2666f82011-03-30 16:57:46 -07001073 rtnl_lock();
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001074 for (i = 0; i < FIB_TABLE_HASHSZ; i++) {
1075 struct fib_table *tb;
1076 struct hlist_head *head;
1077 struct hlist_node *node, *tmp;
1078
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001079 head = &net->ipv4.fib_table_hash[i];
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001080 hlist_for_each_entry_safe(tb, node, tmp, head, tb_hlist) {
1081 hlist_del(node);
Stephen Hemminger16c6cf82009-09-20 10:35:36 +00001082 fib_table_flush(tb);
Pavel Emelyanov4aa2c462010-10-28 02:00:43 +00001083 fib_free_table(tb);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001084 }
1085 }
Eric Dumazete2666f82011-03-30 16:57:46 -07001086 rtnl_unlock();
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001087 kfree(net->ipv4.fib_table_hash);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001088}
1089
1090static int __net_init fib_net_init(struct net *net)
1091{
1092 int error;
1093
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001094 error = ip_fib_net_init(net);
1095 if (error < 0)
1096 goto out;
1097 error = nl_fib_lookup_init(net);
1098 if (error < 0)
1099 goto out_nlfl;
1100 error = fib_proc_init(net);
1101 if (error < 0)
1102 goto out_proc;
1103out:
1104 return error;
1105
1106out_proc:
1107 nl_fib_lookup_exit(net);
1108out_nlfl:
1109 ip_fib_net_exit(net);
1110 goto out;
1111}
1112
1113static void __net_exit fib_net_exit(struct net *net)
1114{
1115 fib_proc_exit(net);
1116 nl_fib_lookup_exit(net);
1117 ip_fib_net_exit(net);
1118}
1119
1120static struct pernet_operations fib_net_ops = {
1121 .init = fib_net_init,
1122 .exit = fib_net_exit,
1123};
1124
1125void __init ip_fib_init(void)
1126{
Greg Rosec7ac8672011-06-10 01:27:09 +00001127 rtnl_register(PF_INET, RTM_NEWROUTE, inet_rtm_newroute, NULL, NULL);
1128 rtnl_register(PF_INET, RTM_DELROUTE, inet_rtm_delroute, NULL, NULL);
1129 rtnl_register(PF_INET, RTM_GETROUTE, NULL, inet_dump_fib, NULL);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001130
1131 register_pernet_subsys(&fib_net_ops);
1132 register_netdevice_notifier(&fib_netdev_notifier);
1133 register_inetaddr_notifier(&fib_inetaddr_notifier);
Stephen Hemminger7f9b8052008-01-14 23:14:20 -08001134
David S. Miller5348ba82011-02-01 15:30:56 -08001135 fib_trie_init();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136}