blob: f116ce8f1b466204f41279e06f57932e274fbdad [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Linus Torvalds1da177e2005-04-16 15:20:36 -070048#ifndef CONFIG_IP_MULTIPLE_TABLES
49
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -080050static int __net_init fib4_rules_init(struct net *net)
Pavel Emelyanovc3e9a352007-11-06 23:34:04 -080051{
Denis V. Lunev93456b62008-01-10 03:23:38 -080052 struct fib_table *local_table, *main_table;
53
David S. Miller5348ba82011-02-01 15:30:56 -080054 local_table = fib_trie_table(RT_TABLE_LOCAL);
Denis V. Lunev93456b62008-01-10 03:23:38 -080055 if (local_table == NULL)
Denis V. Lunevdbb50162008-01-10 03:21:49 -080056 return -ENOMEM;
57
David S. Miller5348ba82011-02-01 15:30:56 -080058 main_table = fib_trie_table(RT_TABLE_MAIN);
Denis V. Lunev93456b62008-01-10 03:23:38 -080059 if (main_table == NULL)
Denis V. Lunevdbb50162008-01-10 03:21:49 -080060 goto fail;
61
Denis V. Lunev93456b62008-01-10 03:23:38 -080062 hlist_add_head_rcu(&local_table->tb_hlist,
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080063 &net->ipv4.fib_table_hash[TABLE_LOCAL_INDEX]);
Denis V. Lunev93456b62008-01-10 03:23:38 -080064 hlist_add_head_rcu(&main_table->tb_hlist,
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080065 &net->ipv4.fib_table_hash[TABLE_MAIN_INDEX]);
Denis V. Lunevdbb50162008-01-10 03:21:49 -080066 return 0;
67
68fail:
Denis V. Lunev93456b62008-01-10 03:23:38 -080069 kfree(local_table);
Denis V. Lunevdbb50162008-01-10 03:21:49 -080070 return -ENOMEM;
Pavel Emelyanovc3e9a352007-11-06 23:34:04 -080071}
Linus Torvalds1da177e2005-04-16 15:20:36 -070072#else
73
Denis V. Lunev8ad49422008-01-10 03:24:11 -080074struct fib_table *fib_new_table(struct net *net, u32 id)
Linus Torvalds1da177e2005-04-16 15:20:36 -070075{
76 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070077 unsigned int h;
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070079 if (id == 0)
80 id = RT_TABLE_MAIN;
Denis V. Lunev8ad49422008-01-10 03:24:11 -080081 tb = fib_get_table(net, id);
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070082 if (tb)
83 return tb;
Stephen Hemminger7f9b8052008-01-14 23:14:20 -080084
David S. Miller5348ba82011-02-01 15:30:56 -080085 tb = fib_trie_table(id);
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 if (!tb)
87 return NULL;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070088 h = id & (FIB_TABLE_HASHSZ - 1);
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080089 hlist_add_head_rcu(&tb->tb_hlist, &net->ipv4.fib_table_hash[h]);
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 return tb;
91}
92
Denis V. Lunev8ad49422008-01-10 03:24:11 -080093struct fib_table *fib_get_table(struct net *net, u32 id)
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070094{
95 struct fib_table *tb;
96 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -080097 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -070098 unsigned int h;
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700100 if (id == 0)
101 id = RT_TABLE_MAIN;
102 h = id & (FIB_TABLE_HASHSZ - 1);
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800103
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700104 rcu_read_lock();
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800105 head = &net->ipv4.fib_table_hash[h];
106 hlist_for_each_entry_rcu(tb, node, head, tb_hlist) {
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700107 if (tb->tb_id == id) {
108 rcu_read_unlock();
109 return tb;
110 }
111 }
112 rcu_read_unlock();
113 return NULL;
114}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115#endif /* CONFIG_IP_MULTIPLE_TABLES */
116
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800117static void fib_flush(struct net *net)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118{
119 int flushed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700121 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800122 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700123 unsigned int h;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700125 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800126 head = &net->ipv4.fib_table_hash[h];
127 hlist_for_each_entry(tb, node, head, tb_hlist)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000128 flushed += fib_table_flush(tb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
131 if (flushed)
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700132 rt_cache_flush(net, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133}
134
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800135/*
136 * Find address type as if only "dev" was present in the system. If
137 * on_dev is NULL then all interfaces are taken into consideration.
138 */
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800139static inline unsigned __inet_dev_addr_type(struct net *net,
140 const struct net_device *dev,
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800141 __be32 addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142{
David S. Miller9ade2282011-03-12 02:02:42 -0500143 struct flowi4 fl4 = { .daddr = addr };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 struct fib_result res;
145 unsigned ret = RTN_BROADCAST;
Pavel Emelyanov03cf7862007-10-23 21:17:27 -0700146 struct fib_table *local_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
Jan Engelhardt1e637c72008-01-21 03:18:08 -0800148 if (ipv4_is_zeronet(addr) || ipv4_is_lbcast(addr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 return RTN_BROADCAST;
Joe Perchesf97c1e02007-12-16 13:45:43 -0800150 if (ipv4_is_multicast(addr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 return RTN_MULTICAST;
152
153#ifdef CONFIG_IP_MULTIPLE_TABLES
154 res.r = NULL;
155#endif
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900156
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800157 local_table = fib_get_table(net, RT_TABLE_LOCAL);
Pavel Emelyanov03cf7862007-10-23 21:17:27 -0700158 if (local_table) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 ret = RTN_UNICAST;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000160 rcu_read_lock();
David S. Miller9ade2282011-03-12 02:02:42 -0500161 if (!fib_table_lookup(local_table, &fl4, &res, FIB_LOOKUP_NOREF)) {
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800162 if (!dev || dev == res.fi->fib_dev)
163 ret = res.type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 }
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000165 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 }
167 return ret;
168}
169
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800170unsigned int inet_addr_type(struct net *net, __be32 addr)
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800171{
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800172 return __inet_dev_addr_type(net, NULL, addr);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800173}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000174EXPORT_SYMBOL(inet_addr_type);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800175
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800176unsigned int inet_dev_addr_type(struct net *net, const struct net_device *dev,
177 __be32 addr)
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800178{
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000179 return __inet_dev_addr_type(net, dev, addr);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800180}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000181EXPORT_SYMBOL(inet_dev_addr_type);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800182
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183/* Given (packet source, input interface) and optional (dst, oif, tos):
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000184 * - (main) check, that source is valid i.e. not broadcast or our local
185 * address.
186 * - figure out what "logical" interface this packet arrived
187 * and calculate "specific destination" address.
188 * - check, that packet arrived from expected physical interface.
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000189 * called with rcu_read_lock()
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 */
Al Virod9c9df82006-09-26 21:28:14 -0700191int fib_validate_source(__be32 src, __be32 dst, u8 tos, int oif,
jamalb0c110c2009-10-18 02:12:33 +0000192 struct net_device *dev, __be32 *spec_dst,
193 u32 *itag, u32 mark)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194{
195 struct in_device *in_dev;
David S. Miller9ade2282011-03-12 02:02:42 -0500196 struct flowi4 fl4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 struct fib_result res;
Patrick McHardy8153a102009-12-03 01:25:58 +0000198 int no_addr, rpf, accept_local;
David S. Miller6f86b322010-09-06 22:36:19 -0700199 bool dev_match;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 int ret;
Denis V. Lunev5b707aa2008-01-21 17:33:15 -0800201 struct net *net;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202
David S. Miller9ade2282011-03-12 02:02:42 -0500203 fl4.flowi4_oif = 0;
204 fl4.flowi4_iif = oif;
205 fl4.flowi4_mark = mark;
206 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;
215 rpf = IN_DEV_RPFILTER(in_dev);
Patrick McHardy8153a102009-12-03 01:25:58 +0000216 accept_local = IN_DEV_ACCEPT_LOCAL(in_dev);
Jamal Hadi Salim28f6aee2009-12-25 17:30:22 -0800217 if (mark && !IN_DEV_SRC_VMARK(in_dev))
David S. Miller9ade2282011-03-12 02:02:42 -0500218 fl4.flowi4_mark = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220
221 if (in_dev == NULL)
222 goto e_inval;
223
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900224 net = dev_net(dev);
David S. Miller9ade2282011-03-12 02:02:42 -0500225 if (fib_lookup(net, &fl4, &res))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 goto last_resort;
Patrick McHardy8153a102009-12-03 01:25:58 +0000227 if (res.type != RTN_UNICAST) {
228 if (res.type != RTN_LOCAL || !accept_local)
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000229 goto e_inval;
Patrick McHardy8153a102009-12-03 01:25:58 +0000230 }
David S. Miller436c3b62011-03-24 17:42:21 -0700231 *spec_dst = FIB_RES_PREFSRC(net, res);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 fib_combine_itag(itag, &res);
David S. Miller6f86b322010-09-06 22:36:19 -0700233 dev_match = false;
234
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235#ifdef CONFIG_IP_ROUTE_MULTIPATH
David S. Miller6f86b322010-09-06 22:36:19 -0700236 for (ret = 0; ret < res.fi->fib_nhs; ret++) {
237 struct fib_nh *nh = &res.fi->fib_nh[ret];
238
239 if (nh->nh_dev == dev) {
240 dev_match = true;
241 break;
242 }
243 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244#else
245 if (FIB_RES_DEV(res) == dev)
David S. Miller6f86b322010-09-06 22:36:19 -0700246 dev_match = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247#endif
David S. Miller6f86b322010-09-06 22:36:19 -0700248 if (dev_match) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 ret = FIB_RES_NH(res).nh_scope >= RT_SCOPE_HOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250 return ret;
251 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 if (no_addr)
253 goto last_resort;
Stephen Hemmingerc1cf8422009-02-20 08:25:36 +0000254 if (rpf == 1)
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000255 goto e_rpf;
David S. Miller9ade2282011-03-12 02:02:42 -0500256 fl4.flowi4_oif = dev->ifindex;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257
258 ret = 0;
David S. Miller9ade2282011-03-12 02:02:42 -0500259 if (fib_lookup(net, &fl4, &res) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 if (res.type == RTN_UNICAST) {
David S. Miller436c3b62011-03-24 17:42:21 -0700261 *spec_dst = FIB_RES_PREFSRC(net, res);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 ret = FIB_RES_NH(res).nh_scope >= RT_SCOPE_HOST;
263 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 }
265 return ret;
266
267last_resort:
268 if (rpf)
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000269 goto e_rpf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 *spec_dst = inet_select_addr(dev, 0, RT_SCOPE_UNIVERSE);
271 *itag = 0;
272 return 0;
273
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274e_inval:
275 return -EINVAL;
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000276e_rpf:
277 return -EXDEV;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278}
279
Al Viro81f7bf62006-09-27 18:40:00 -0700280static inline __be32 sk_extract_addr(struct sockaddr *addr)
Thomas Graf4e902c52006-08-17 18:14:52 -0700281{
282 return ((struct sockaddr_in *) addr)->sin_addr.s_addr;
283}
284
285static int put_rtax(struct nlattr *mx, int len, int type, u32 value)
286{
287 struct nlattr *nla;
288
289 nla = (struct nlattr *) ((char *) mx + len);
290 nla->nla_type = type;
291 nla->nla_len = nla_attr_size(4);
292 *(u32 *) nla_data(nla) = value;
293
294 return len + nla_total_size(4);
295}
296
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800297static int rtentry_to_fib_config(struct net *net, int cmd, struct rtentry *rt,
Thomas Graf4e902c52006-08-17 18:14:52 -0700298 struct fib_config *cfg)
299{
Al Viro6d85c102006-09-26 22:15:46 -0700300 __be32 addr;
Thomas Graf4e902c52006-08-17 18:14:52 -0700301 int plen;
302
303 memset(cfg, 0, sizeof(*cfg));
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800304 cfg->fc_nlinfo.nl_net = net;
Thomas Graf4e902c52006-08-17 18:14:52 -0700305
306 if (rt->rt_dst.sa_family != AF_INET)
307 return -EAFNOSUPPORT;
308
309 /*
310 * Check mask for validity:
311 * a) it must be contiguous.
312 * b) destination must have all host bits clear.
313 * c) if application forgot to set correct family (AF_INET),
314 * reject request unless it is absolutely clear i.e.
315 * both family and mask are zero.
316 */
317 plen = 32;
318 addr = sk_extract_addr(&rt->rt_dst);
319 if (!(rt->rt_flags & RTF_HOST)) {
Al Viro81f7bf62006-09-27 18:40:00 -0700320 __be32 mask = sk_extract_addr(&rt->rt_genmask);
Thomas Graf4e902c52006-08-17 18:14:52 -0700321
322 if (rt->rt_genmask.sa_family != AF_INET) {
323 if (mask || rt->rt_genmask.sa_family)
324 return -EAFNOSUPPORT;
325 }
326
327 if (bad_mask(mask, addr))
328 return -EINVAL;
329
330 plen = inet_mask_len(mask);
331 }
332
333 cfg->fc_dst_len = plen;
334 cfg->fc_dst = addr;
335
336 if (cmd != SIOCDELRT) {
337 cfg->fc_nlflags = NLM_F_CREATE;
338 cfg->fc_protocol = RTPROT_BOOT;
339 }
340
341 if (rt->rt_metric)
342 cfg->fc_priority = rt->rt_metric - 1;
343
344 if (rt->rt_flags & RTF_REJECT) {
345 cfg->fc_scope = RT_SCOPE_HOST;
346 cfg->fc_type = RTN_UNREACHABLE;
347 return 0;
348 }
349
350 cfg->fc_scope = RT_SCOPE_NOWHERE;
351 cfg->fc_type = RTN_UNICAST;
352
353 if (rt->rt_dev) {
354 char *colon;
355 struct net_device *dev;
356 char devname[IFNAMSIZ];
357
358 if (copy_from_user(devname, rt->rt_dev, IFNAMSIZ-1))
359 return -EFAULT;
360
361 devname[IFNAMSIZ-1] = 0;
362 colon = strchr(devname, ':');
363 if (colon)
364 *colon = 0;
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800365 dev = __dev_get_by_name(net, devname);
Thomas Graf4e902c52006-08-17 18:14:52 -0700366 if (!dev)
367 return -ENODEV;
368 cfg->fc_oif = dev->ifindex;
369 if (colon) {
370 struct in_ifaddr *ifa;
371 struct in_device *in_dev = __in_dev_get_rtnl(dev);
372 if (!in_dev)
373 return -ENODEV;
374 *colon = ':';
375 for (ifa = in_dev->ifa_list; ifa; ifa = ifa->ifa_next)
376 if (strcmp(ifa->ifa_label, devname) == 0)
377 break;
378 if (ifa == NULL)
379 return -ENODEV;
380 cfg->fc_prefsrc = ifa->ifa_local;
381 }
382 }
383
384 addr = sk_extract_addr(&rt->rt_gateway);
385 if (rt->rt_gateway.sa_family == AF_INET && addr) {
386 cfg->fc_gw = addr;
387 if (rt->rt_flags & RTF_GATEWAY &&
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800388 inet_addr_type(net, addr) == RTN_UNICAST)
Thomas Graf4e902c52006-08-17 18:14:52 -0700389 cfg->fc_scope = RT_SCOPE_UNIVERSE;
390 }
391
392 if (cmd == SIOCDELRT)
393 return 0;
394
395 if (rt->rt_flags & RTF_GATEWAY && !cfg->fc_gw)
396 return -EINVAL;
397
398 if (cfg->fc_scope == RT_SCOPE_NOWHERE)
399 cfg->fc_scope = RT_SCOPE_LINK;
400
401 if (rt->rt_flags & (RTF_MTU | RTF_WINDOW | RTF_IRTT)) {
402 struct nlattr *mx;
403 int len = 0;
404
405 mx = kzalloc(3 * nla_total_size(4), GFP_KERNEL);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900406 if (mx == NULL)
Thomas Graf4e902c52006-08-17 18:14:52 -0700407 return -ENOMEM;
408
409 if (rt->rt_flags & RTF_MTU)
410 len = put_rtax(mx, len, RTAX_ADVMSS, rt->rt_mtu - 40);
411
412 if (rt->rt_flags & RTF_WINDOW)
413 len = put_rtax(mx, len, RTAX_WINDOW, rt->rt_window);
414
415 if (rt->rt_flags & RTF_IRTT)
416 len = put_rtax(mx, len, RTAX_RTT, rt->rt_irtt << 3);
417
418 cfg->fc_mx = mx;
419 cfg->fc_mx_len = len;
420 }
421
422 return 0;
423}
424
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425/*
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000426 * Handle IP routing ioctl calls.
427 * These are used to manipulate the routing tables
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 */
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800429int ip_rt_ioctl(struct net *net, unsigned int cmd, void __user *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430{
Thomas Graf4e902c52006-08-17 18:14:52 -0700431 struct fib_config cfg;
432 struct rtentry rt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
435 switch (cmd) {
436 case SIOCADDRT: /* Add a route */
437 case SIOCDELRT: /* Delete a route */
438 if (!capable(CAP_NET_ADMIN))
439 return -EPERM;
Thomas Graf4e902c52006-08-17 18:14:52 -0700440
441 if (copy_from_user(&rt, arg, sizeof(rt)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 return -EFAULT;
Thomas Graf4e902c52006-08-17 18:14:52 -0700443
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 rtnl_lock();
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800445 err = rtentry_to_fib_config(net, cmd, &rt, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 if (err == 0) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700447 struct fib_table *tb;
448
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 if (cmd == SIOCDELRT) {
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800450 tb = fib_get_table(net, cfg.fc_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 if (tb)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000452 err = fib_table_delete(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700453 else
454 err = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 } else {
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800456 tb = fib_new_table(net, cfg.fc_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 if (tb)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000458 err = fib_table_insert(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700459 else
460 err = -ENOBUFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700462
463 /* allocated by rtentry_to_fib_config() */
464 kfree(cfg.fc_mx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 }
466 rtnl_unlock();
467 return err;
468 }
469 return -EINVAL;
470}
471
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000472const struct nla_policy rtm_ipv4_policy[RTA_MAX + 1] = {
Thomas Graf4e902c52006-08-17 18:14:52 -0700473 [RTA_DST] = { .type = NLA_U32 },
474 [RTA_SRC] = { .type = NLA_U32 },
475 [RTA_IIF] = { .type = NLA_U32 },
476 [RTA_OIF] = { .type = NLA_U32 },
477 [RTA_GATEWAY] = { .type = NLA_U32 },
478 [RTA_PRIORITY] = { .type = NLA_U32 },
479 [RTA_PREFSRC] = { .type = NLA_U32 },
480 [RTA_METRICS] = { .type = NLA_NESTED },
Thomas Graf5176f912006-08-26 20:13:18 -0700481 [RTA_MULTIPATH] = { .len = sizeof(struct rtnexthop) },
Thomas Graf4e902c52006-08-17 18:14:52 -0700482 [RTA_FLOW] = { .type = NLA_U32 },
Thomas Graf4e902c52006-08-17 18:14:52 -0700483};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800485static int rtm_to_fib_config(struct net *net, struct sk_buff *skb,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000486 struct nlmsghdr *nlh, struct fib_config *cfg)
Thomas Graf4e902c52006-08-17 18:14:52 -0700487{
488 struct nlattr *attr;
489 int err, remaining;
490 struct rtmsg *rtm;
491
492 err = nlmsg_validate(nlh, sizeof(*rtm), RTA_MAX, rtm_ipv4_policy);
493 if (err < 0)
494 goto errout;
495
496 memset(cfg, 0, sizeof(*cfg));
497
498 rtm = nlmsg_data(nlh);
Thomas Graf4e902c52006-08-17 18:14:52 -0700499 cfg->fc_dst_len = rtm->rtm_dst_len;
Thomas Graf4e902c52006-08-17 18:14:52 -0700500 cfg->fc_tos = rtm->rtm_tos;
501 cfg->fc_table = rtm->rtm_table;
502 cfg->fc_protocol = rtm->rtm_protocol;
503 cfg->fc_scope = rtm->rtm_scope;
504 cfg->fc_type = rtm->rtm_type;
505 cfg->fc_flags = rtm->rtm_flags;
506 cfg->fc_nlflags = nlh->nlmsg_flags;
507
508 cfg->fc_nlinfo.pid = NETLINK_CB(skb).pid;
509 cfg->fc_nlinfo.nlh = nlh;
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800510 cfg->fc_nlinfo.nl_net = net;
Thomas Graf4e902c52006-08-17 18:14:52 -0700511
Thomas Grafa0ee18b2007-03-24 20:32:54 -0700512 if (cfg->fc_type > RTN_MAX) {
513 err = -EINVAL;
514 goto errout;
515 }
516
Thomas Graf4e902c52006-08-17 18:14:52 -0700517 nlmsg_for_each_attr(attr, nlh, sizeof(struct rtmsg), remaining) {
Thomas Graf8f4c1f92007-09-12 14:44:36 +0200518 switch (nla_type(attr)) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700519 case RTA_DST:
Al Viro17fb2c62006-09-26 22:15:25 -0700520 cfg->fc_dst = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700521 break;
Thomas Graf4e902c52006-08-17 18:14:52 -0700522 case RTA_OIF:
523 cfg->fc_oif = nla_get_u32(attr);
524 break;
525 case RTA_GATEWAY:
Al Viro17fb2c62006-09-26 22:15:25 -0700526 cfg->fc_gw = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700527 break;
528 case RTA_PRIORITY:
529 cfg->fc_priority = nla_get_u32(attr);
530 break;
531 case RTA_PREFSRC:
Al Viro17fb2c62006-09-26 22:15:25 -0700532 cfg->fc_prefsrc = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700533 break;
534 case RTA_METRICS:
535 cfg->fc_mx = nla_data(attr);
536 cfg->fc_mx_len = nla_len(attr);
537 break;
538 case RTA_MULTIPATH:
539 cfg->fc_mp = nla_data(attr);
540 cfg->fc_mp_len = nla_len(attr);
541 break;
542 case RTA_FLOW:
543 cfg->fc_flow = nla_get_u32(attr);
544 break;
Thomas Graf4e902c52006-08-17 18:14:52 -0700545 case RTA_TABLE:
546 cfg->fc_table = nla_get_u32(attr);
547 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 }
549 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700550
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 return 0;
Thomas Graf4e902c52006-08-17 18:14:52 -0700552errout:
553 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554}
555
Jianjun Kong6ed25332008-11-03 00:25:16 -0800556static int inet_rtm_delroute(struct sk_buff *skb, struct nlmsghdr *nlh, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900558 struct net *net = sock_net(skb->sk);
Thomas Graf4e902c52006-08-17 18:14:52 -0700559 struct fib_config cfg;
560 struct fib_table *tb;
561 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800563 err = rtm_to_fib_config(net, skb, nlh, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700564 if (err < 0)
565 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566
Denis V. Lunev8ad49422008-01-10 03:24:11 -0800567 tb = fib_get_table(net, cfg.fc_table);
Thomas Graf4e902c52006-08-17 18:14:52 -0700568 if (tb == NULL) {
569 err = -ESRCH;
570 goto errout;
571 }
572
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000573 err = fib_table_delete(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700574errout:
575 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576}
577
Jianjun Kong6ed25332008-11-03 00:25:16 -0800578static int inet_rtm_newroute(struct sk_buff *skb, struct nlmsghdr *nlh, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900580 struct net *net = sock_net(skb->sk);
Thomas Graf4e902c52006-08-17 18:14:52 -0700581 struct fib_config cfg;
582 struct fib_table *tb;
583 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800585 err = rtm_to_fib_config(net, skb, nlh, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700586 if (err < 0)
587 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588
Denis V. Lunev226b0b4a52008-01-10 03:30:24 -0800589 tb = fib_new_table(net, cfg.fc_table);
Thomas Graf4e902c52006-08-17 18:14:52 -0700590 if (tb == NULL) {
591 err = -ENOBUFS;
592 goto errout;
593 }
594
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000595 err = fib_table_insert(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700596errout:
597 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598}
599
Thomas Graf63f34442007-03-22 11:55:17 -0700600static int inet_dump_fib(struct sk_buff *skb, struct netlink_callback *cb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900602 struct net *net = sock_net(skb->sk);
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700603 unsigned int h, s_h;
604 unsigned int e = 0, s_e;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700606 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800607 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700608 int dumped = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609
Thomas Grafbe403ea2006-08-17 18:15:17 -0700610 if (nlmsg_len(cb->nlh) >= sizeof(struct rtmsg) &&
611 ((struct rtmsg *) nlmsg_data(cb->nlh))->rtm_flags & RTM_F_CLONED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 return ip_rt_dump(skb, cb);
613
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700614 s_h = cb->args[0];
615 s_e = cb->args[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700617 for (h = s_h; h < FIB_TABLE_HASHSZ; h++, s_e = 0) {
618 e = 0;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800619 head = &net->ipv4.fib_table_hash[h];
620 hlist_for_each_entry(tb, node, head, tb_hlist) {
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700621 if (e < s_e)
622 goto next;
623 if (dumped)
624 memset(&cb->args[2], 0, sizeof(cb->args) -
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900625 2 * sizeof(cb->args[0]));
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000626 if (fib_table_dump(tb, skb, cb) < 0)
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700627 goto out;
628 dumped = 1;
629next:
630 e++;
631 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 }
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700633out:
634 cb->args[1] = e;
635 cb->args[0] = h;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636
637 return skb->len;
638}
639
640/* Prepare and feed intra-kernel routing request.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000641 * Really, it should be netlink message, but :-( netlink
642 * can be not configured, so that we feed it directly
643 * to fib engine. It is legal, because all events occur
644 * only when netlink is already locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 */
Al Viro81f7bf62006-09-27 18:40:00 -0700646static void fib_magic(int cmd, int type, __be32 dst, int dst_len, struct in_ifaddr *ifa)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647{
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900648 struct net *net = dev_net(ifa->ifa_dev->dev);
Thomas Graf4e902c52006-08-17 18:14:52 -0700649 struct fib_table *tb;
650 struct fib_config cfg = {
651 .fc_protocol = RTPROT_KERNEL,
652 .fc_type = type,
653 .fc_dst = dst,
654 .fc_dst_len = dst_len,
655 .fc_prefsrc = ifa->ifa_local,
656 .fc_oif = ifa->ifa_dev->dev->ifindex,
657 .fc_nlflags = NLM_F_CREATE | NLM_F_APPEND,
Denis V. Lunev4d1169c2008-01-10 03:26:13 -0800658 .fc_nlinfo = {
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800659 .nl_net = net,
Denis V. Lunev4d1169c2008-01-10 03:26:13 -0800660 },
Thomas Graf4e902c52006-08-17 18:14:52 -0700661 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
663 if (type == RTN_UNICAST)
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800664 tb = fib_new_table(net, RT_TABLE_MAIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 else
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800666 tb = fib_new_table(net, RT_TABLE_LOCAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
668 if (tb == NULL)
669 return;
670
Thomas Graf4e902c52006-08-17 18:14:52 -0700671 cfg.fc_table = tb->tb_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672
Thomas Graf4e902c52006-08-17 18:14:52 -0700673 if (type != RTN_LOCAL)
674 cfg.fc_scope = RT_SCOPE_LINK;
675 else
676 cfg.fc_scope = RT_SCOPE_HOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677
678 if (cmd == RTM_NEWROUTE)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000679 fib_table_insert(tb, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 else
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000681 fib_table_delete(tb, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682}
683
Jamal Hadi Salim0ff60a42005-11-22 14:47:37 -0800684void fib_add_ifaddr(struct in_ifaddr *ifa)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685{
686 struct in_device *in_dev = ifa->ifa_dev;
687 struct net_device *dev = in_dev->dev;
688 struct in_ifaddr *prim = ifa;
Al Viroa144ea42006-09-28 18:00:55 -0700689 __be32 mask = ifa->ifa_mask;
690 __be32 addr = ifa->ifa_local;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000691 __be32 prefix = ifa->ifa_address & mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000693 if (ifa->ifa_flags & IFA_F_SECONDARY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 prim = inet_ifa_byprefix(in_dev, prefix, mask);
695 if (prim == NULL) {
Stephen Hemmingera6db9012008-01-12 20:58:35 -0800696 printk(KERN_WARNING "fib_add_ifaddr: bug: prim == NULL\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 return;
698 }
699 }
700
701 fib_magic(RTM_NEWROUTE, RTN_LOCAL, addr, 32, prim);
702
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000703 if (!(dev->flags & IFF_UP))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 return;
705
706 /* Add broadcast address, if it is explicitly assigned. */
Al Viroa144ea42006-09-28 18:00:55 -0700707 if (ifa->ifa_broadcast && ifa->ifa_broadcast != htonl(0xFFFFFFFF))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, ifa->ifa_broadcast, 32, prim);
709
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000710 if (!ipv4_is_zeronet(prefix) && !(ifa->ifa_flags & IFA_F_SECONDARY) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 (prefix != addr || ifa->ifa_prefixlen < 32)) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000712 fib_magic(RTM_NEWROUTE,
713 dev->flags & IFF_LOOPBACK ? RTN_LOCAL : RTN_UNICAST,
714 prefix, ifa->ifa_prefixlen, prim);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
716 /* Add network specific broadcasts, when it takes a sense */
717 if (ifa->ifa_prefixlen < 31) {
718 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, prefix, 32, prim);
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000719 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, prefix | ~mask,
720 32, prim);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 }
722 }
723}
724
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000725/* Delete primary or secondary address.
726 * Optionally, on secondary address promotion consider the addresses
727 * from subnet iprim as deleted, even if they are in device list.
728 * In this case the secondary ifa can be in device list.
729 */
730void fib_del_ifaddr(struct in_ifaddr *ifa, struct in_ifaddr *iprim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731{
732 struct in_device *in_dev = ifa->ifa_dev;
733 struct net_device *dev = in_dev->dev;
734 struct in_ifaddr *ifa1;
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000735 struct in_ifaddr *prim = ifa, *prim1 = NULL;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000736 __be32 brd = ifa->ifa_address | ~ifa->ifa_mask;
737 __be32 any = ifa->ifa_address & ifa->ifa_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738#define LOCAL_OK 1
739#define BRD_OK 2
740#define BRD0_OK 4
741#define BRD1_OK 8
742 unsigned ok = 0;
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000743 int subnet = 0; /* Primary network */
744 int gone = 1; /* Address is missing */
745 int same_prefsrc = 0; /* Another primary with same IP */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000747 if (ifa->ifa_flags & IFA_F_SECONDARY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 prim = inet_ifa_byprefix(in_dev, any, ifa->ifa_mask);
749 if (prim == NULL) {
Stephen Hemmingera6db9012008-01-12 20:58:35 -0800750 printk(KERN_WARNING "fib_del_ifaddr: bug: prim == NULL\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 return;
752 }
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000753 if (iprim && iprim != prim) {
754 printk(KERN_WARNING "fib_del_ifaddr: bug: iprim != prim\n");
755 return;
756 }
757 } else if (!ipv4_is_zeronet(any) &&
758 (any != ifa->ifa_local || ifa->ifa_prefixlen < 32)) {
759 fib_magic(RTM_DELROUTE,
760 dev->flags & IFF_LOOPBACK ? RTN_LOCAL : RTN_UNICAST,
761 any, ifa->ifa_prefixlen, prim);
762 subnet = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 }
764
765 /* Deletion is more complicated than add.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000766 * We should take care of not to delete too much :-)
767 *
768 * Scan address list to be sure that addresses are really gone.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 */
770
771 for (ifa1 = in_dev->ifa_list; ifa1; ifa1 = ifa1->ifa_next) {
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000772 if (ifa1 == ifa) {
773 /* promotion, keep the IP */
774 gone = 0;
775 continue;
776 }
777 /* Ignore IFAs from our subnet */
778 if (iprim && ifa1->ifa_mask == iprim->ifa_mask &&
779 inet_ifa_match(ifa1->ifa_address, iprim))
780 continue;
781
782 /* Ignore ifa1 if it uses different primary IP (prefsrc) */
783 if (ifa1->ifa_flags & IFA_F_SECONDARY) {
784 /* Another address from our subnet? */
785 if (ifa1->ifa_mask == prim->ifa_mask &&
786 inet_ifa_match(ifa1->ifa_address, prim))
787 prim1 = prim;
788 else {
789 /* We reached the secondaries, so
790 * same_prefsrc should be determined.
791 */
792 if (!same_prefsrc)
793 continue;
794 /* Search new prim1 if ifa1 is not
795 * using the current prim1
796 */
797 if (!prim1 ||
798 ifa1->ifa_mask != prim1->ifa_mask ||
799 !inet_ifa_match(ifa1->ifa_address, prim1))
800 prim1 = inet_ifa_byprefix(in_dev,
801 ifa1->ifa_address,
802 ifa1->ifa_mask);
803 if (!prim1)
804 continue;
805 if (prim1->ifa_local != prim->ifa_local)
806 continue;
807 }
808 } else {
809 if (prim->ifa_local != ifa1->ifa_local)
810 continue;
811 prim1 = ifa1;
812 if (prim != prim1)
813 same_prefsrc = 1;
814 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 if (ifa->ifa_local == ifa1->ifa_local)
816 ok |= LOCAL_OK;
817 if (ifa->ifa_broadcast == ifa1->ifa_broadcast)
818 ok |= BRD_OK;
819 if (brd == ifa1->ifa_broadcast)
820 ok |= BRD1_OK;
821 if (any == ifa1->ifa_broadcast)
822 ok |= BRD0_OK;
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000823 /* primary has network specific broadcasts */
824 if (prim1 == ifa1 && ifa1->ifa_prefixlen < 31) {
825 __be32 brd1 = ifa1->ifa_address | ~ifa1->ifa_mask;
826 __be32 any1 = ifa1->ifa_address & ifa1->ifa_mask;
827
828 if (!ipv4_is_zeronet(any1)) {
829 if (ifa->ifa_broadcast == brd1 ||
830 ifa->ifa_broadcast == any1)
831 ok |= BRD_OK;
832 if (brd == brd1 || brd == any1)
833 ok |= BRD1_OK;
834 if (any == brd1 || any == any1)
835 ok |= BRD0_OK;
836 }
837 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 }
839
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000840 if (!(ok & BRD_OK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 fib_magic(RTM_DELROUTE, RTN_BROADCAST, ifa->ifa_broadcast, 32, prim);
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000842 if (subnet && ifa->ifa_prefixlen < 31) {
843 if (!(ok & BRD1_OK))
844 fib_magic(RTM_DELROUTE, RTN_BROADCAST, brd, 32, prim);
845 if (!(ok & BRD0_OK))
846 fib_magic(RTM_DELROUTE, RTN_BROADCAST, any, 32, prim);
847 }
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000848 if (!(ok & LOCAL_OK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 fib_magic(RTM_DELROUTE, RTN_LOCAL, ifa->ifa_local, 32, prim);
850
851 /* Check, that this local address finally disappeared. */
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000852 if (gone &&
853 inet_addr_type(dev_net(dev), ifa->ifa_local) != RTN_LOCAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 /* And the last, but not the least thing.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000855 * We must flush stray FIB entries.
856 *
857 * First of all, we scan fib_info list searching
858 * for stray nexthop entries, then ignite fib_flush.
859 */
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900860 if (fib_sync_down_addr(dev_net(dev), ifa->ifa_local))
861 fib_flush(dev_net(dev));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 }
863 }
864#undef LOCAL_OK
865#undef BRD_OK
866#undef BRD0_OK
867#undef BRD1_OK
868}
869
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000870static void nl_fib_lookup(struct fib_result_nl *frn, struct fib_table *tb)
Robert Olsson246955f2005-06-20 13:36:39 -0700871{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900872
Robert Olsson246955f2005-06-20 13:36:39 -0700873 struct fib_result res;
David S. Miller9ade2282011-03-12 02:02:42 -0500874 struct flowi4 fl4 = {
875 .flowi4_mark = frn->fl_mark,
876 .daddr = frn->fl_addr,
877 .flowi4_tos = frn->fl_tos,
878 .flowi4_scope = frn->fl_scope,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000879 };
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700880
Sergey Vlasov912a41a2007-04-27 02:17:19 -0700881#ifdef CONFIG_IP_MULTIPLE_TABLES
882 res.r = NULL;
883#endif
884
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700885 frn->err = -ENOENT;
Robert Olsson246955f2005-06-20 13:36:39 -0700886 if (tb) {
887 local_bh_disable();
888
889 frn->tb_id = tb->tb_id;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000890 rcu_read_lock();
David S. Miller9ade2282011-03-12 02:02:42 -0500891 frn->err = fib_table_lookup(tb, &fl4, &res, FIB_LOOKUP_NOREF);
Robert Olsson246955f2005-06-20 13:36:39 -0700892
893 if (!frn->err) {
894 frn->prefixlen = res.prefixlen;
895 frn->nh_sel = res.nh_sel;
896 frn->type = res.type;
897 frn->scope = res.scope;
898 }
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000899 rcu_read_unlock();
Robert Olsson246955f2005-06-20 13:36:39 -0700900 local_bh_enable();
901 }
902}
903
David S. Miller28f7b0362007-10-10 21:32:39 -0700904static void nl_fib_input(struct sk_buff *skb)
Robert Olsson246955f2005-06-20 13:36:39 -0700905{
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800906 struct net *net;
Robert Olsson246955f2005-06-20 13:36:39 -0700907 struct fib_result_nl *frn;
David S. Miller28f7b0362007-10-10 21:32:39 -0700908 struct nlmsghdr *nlh;
Robert Olsson246955f2005-06-20 13:36:39 -0700909 struct fib_table *tb;
David S. Miller28f7b0362007-10-10 21:32:39 -0700910 u32 pid;
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700911
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900912 net = sock_net(skb->sk);
Arnaldo Carvalho de Melob529ccf2007-04-25 19:08:35 -0700913 nlh = nlmsg_hdr(skb);
Thomas Grafea865752005-12-01 14:30:00 -0800914 if (skb->len < NLMSG_SPACE(0) || skb->len < nlh->nlmsg_len ||
Denis V. Lunevd883a032007-12-21 02:01:53 -0800915 nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*frn)))
Thomas Grafea865752005-12-01 14:30:00 -0800916 return;
Denis V. Lunevd883a032007-12-21 02:01:53 -0800917
918 skb = skb_clone(skb, GFP_KERNEL);
919 if (skb == NULL)
920 return;
921 nlh = nlmsg_hdr(skb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900922
Robert Olsson246955f2005-06-20 13:36:39 -0700923 frn = (struct fib_result_nl *) NLMSG_DATA(nlh);
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800924 tb = fib_get_table(net, frn->tb_id_in);
Robert Olsson246955f2005-06-20 13:36:39 -0700925
926 nl_fib_lookup(frn, tb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900927
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000928 pid = NETLINK_CB(skb).pid; /* pid of sending process */
929 NETLINK_CB(skb).pid = 0; /* from kernel */
Patrick McHardyac6d4392005-08-14 19:29:52 -0700930 NETLINK_CB(skb).dst_group = 0; /* unicast */
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800931 netlink_unicast(net->ipv4.fibnl, skb, pid, MSG_DONTWAIT);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900932}
Robert Olsson246955f2005-06-20 13:36:39 -0700933
Alexey Dobriyan2c8c1e72010-01-17 03:35:32 +0000934static int __net_init nl_fib_lookup_init(struct net *net)
Robert Olsson246955f2005-06-20 13:36:39 -0700935{
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800936 struct sock *sk;
937 sk = netlink_kernel_create(net, NETLINK_FIB_LOOKUP, 0,
938 nl_fib_input, NULL, THIS_MODULE);
939 if (sk == NULL)
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -0800940 return -EAFNOSUPPORT;
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800941 net->ipv4.fibnl = sk;
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -0800942 return 0;
943}
944
945static void nl_fib_lookup_exit(struct net *net)
946{
Denis V. Lunevb7c6ba62008-01-28 14:41:19 -0800947 netlink_kernel_release(net->ipv4.fibnl);
Denis V. Lunev775516b2008-01-18 23:55:19 -0800948 net->ipv4.fibnl = NULL;
Robert Olsson246955f2005-06-20 13:36:39 -0700949}
950
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000951static void fib_disable_ip(struct net_device *dev, int force, int delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952{
Denis V. Lunev85326fa2008-01-31 18:48:47 -0800953 if (fib_sync_down_dev(dev, force))
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900954 fib_flush(dev_net(dev));
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000955 rt_cache_flush(dev_net(dev), delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 arp_ifdown(dev);
957}
958
959static int fib_inetaddr_event(struct notifier_block *this, unsigned long event, void *ptr)
960{
Jianjun Kong6ed25332008-11-03 00:25:16 -0800961 struct in_ifaddr *ifa = (struct in_ifaddr *)ptr;
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700962 struct net_device *dev = ifa->ifa_dev->dev;
David S. Miller436c3b62011-03-24 17:42:21 -0700963 struct net *net = dev_net(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
965 switch (event) {
966 case NETDEV_UP:
967 fib_add_ifaddr(ifa);
968#ifdef CONFIG_IP_ROUTE_MULTIPATH
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700969 fib_sync_up(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970#endif
David S. Miller436c3b62011-03-24 17:42:21 -0700971 atomic_inc(&net->ipv4.dev_addr_genid);
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700972 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 break;
974 case NETDEV_DOWN:
Julian Anastasove6abbaa2011-03-19 12:13:49 +0000975 fib_del_ifaddr(ifa, NULL);
David S. Miller436c3b62011-03-24 17:42:21 -0700976 atomic_inc(&net->ipv4.dev_addr_genid);
Jayachandran C9fcc2e82005-10-27 15:10:01 -0700977 if (ifa->ifa_dev->ifa_list == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 /* Last address was deleted from this interface.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000979 * Disable IP.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 */
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000981 fib_disable_ip(dev, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 } else {
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700983 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 }
985 break;
986 }
987 return NOTIFY_DONE;
988}
989
990static int fib_netdev_event(struct notifier_block *this, unsigned long event, void *ptr)
991{
992 struct net_device *dev = ptr;
Herbert Xue5ed6392005-10-03 14:35:55 -0700993 struct in_device *in_dev = __in_dev_get_rtnl(dev);
David S. Miller436c3b62011-03-24 17:42:21 -0700994 struct net *net = dev_net(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995
996 if (event == NETDEV_UNREGISTER) {
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000997 fib_disable_ip(dev, 2, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 return NOTIFY_DONE;
999 }
1000
1001 if (!in_dev)
1002 return NOTIFY_DONE;
1003
1004 switch (event) {
1005 case NETDEV_UP:
1006 for_ifa(in_dev) {
1007 fib_add_ifaddr(ifa);
1008 } endfor_ifa(in_dev);
1009#ifdef CONFIG_IP_ROUTE_MULTIPATH
1010 fib_sync_up(dev);
1011#endif
David S. Miller436c3b62011-03-24 17:42:21 -07001012 atomic_inc(&net->ipv4.dev_addr_genid);
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -07001013 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 break;
1015 case NETDEV_DOWN:
Octavian Purdilae2ce1462009-11-16 13:49:49 +00001016 fib_disable_ip(dev, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 break;
1018 case NETDEV_CHANGEMTU:
1019 case NETDEV_CHANGE:
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -07001020 rt_cache_flush(dev_net(dev), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 break;
Eric W. Biedermana5ee1552009-11-29 15:45:58 +00001022 case NETDEV_UNREGISTER_BATCH:
David S. Miller6561a3b2010-12-19 21:11:20 -08001023 /* The batch unregister is only called on the first
1024 * device in the list of devices being unregistered.
1025 * Therefore we should not pass dev_net(dev) in here.
1026 */
1027 rt_cache_flush_batch(NULL);
Eric W. Biedermana5ee1552009-11-29 15:45:58 +00001028 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 }
1030 return NOTIFY_DONE;
1031}
1032
1033static struct notifier_block fib_inetaddr_notifier = {
Jianjun Kong6ed25332008-11-03 00:25:16 -08001034 .notifier_call = fib_inetaddr_event,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035};
1036
1037static struct notifier_block fib_netdev_notifier = {
Jianjun Kong6ed25332008-11-03 00:25:16 -08001038 .notifier_call = fib_netdev_event,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039};
1040
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001041static int __net_init ip_fib_net_init(struct net *net)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042{
Denis V. Lunevdce5cbe2008-01-31 18:44:53 -08001043 int err;
Eric Dumazet10da66f2010-10-13 08:22:03 +00001044 size_t size = sizeof(struct hlist_head) * FIB_TABLE_HASHSZ;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -07001045
Eric Dumazet10da66f2010-10-13 08:22:03 +00001046 /* Avoid false sharing : Use at least a full cache line */
1047 size = max_t(size_t, size, L1_CACHE_BYTES);
1048
1049 net->ipv4.fib_table_hash = kzalloc(size, GFP_KERNEL);
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001050 if (net->ipv4.fib_table_hash == NULL)
1051 return -ENOMEM;
1052
Denis V. Lunevdce5cbe2008-01-31 18:44:53 -08001053 err = fib4_rules_init(net);
1054 if (err < 0)
1055 goto fail;
1056 return 0;
1057
1058fail:
1059 kfree(net->ipv4.fib_table_hash);
1060 return err;
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001061}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062
Alexey Dobriyan2c8c1e72010-01-17 03:35:32 +00001063static void ip_fib_net_exit(struct net *net)
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001064{
1065 unsigned int i;
Thomas Graf63f34442007-03-22 11:55:17 -07001066
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001067#ifdef CONFIG_IP_MULTIPLE_TABLES
1068 fib4_rules_exit(net);
1069#endif
1070
1071 for (i = 0; i < FIB_TABLE_HASHSZ; i++) {
1072 struct fib_table *tb;
1073 struct hlist_head *head;
1074 struct hlist_node *node, *tmp;
1075
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001076 head = &net->ipv4.fib_table_hash[i];
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001077 hlist_for_each_entry_safe(tb, node, tmp, head, tb_hlist) {
1078 hlist_del(node);
Stephen Hemminger16c6cf82009-09-20 10:35:36 +00001079 fib_table_flush(tb);
Pavel Emelyanov4aa2c462010-10-28 02:00:43 +00001080 fib_free_table(tb);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001081 }
1082 }
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001083 kfree(net->ipv4.fib_table_hash);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001084}
1085
1086static int __net_init fib_net_init(struct net *net)
1087{
1088 int error;
1089
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001090 error = ip_fib_net_init(net);
1091 if (error < 0)
1092 goto out;
1093 error = nl_fib_lookup_init(net);
1094 if (error < 0)
1095 goto out_nlfl;
1096 error = fib_proc_init(net);
1097 if (error < 0)
1098 goto out_proc;
1099out:
1100 return error;
1101
1102out_proc:
1103 nl_fib_lookup_exit(net);
1104out_nlfl:
1105 ip_fib_net_exit(net);
1106 goto out;
1107}
1108
1109static void __net_exit fib_net_exit(struct net *net)
1110{
1111 fib_proc_exit(net);
1112 nl_fib_lookup_exit(net);
1113 ip_fib_net_exit(net);
1114}
1115
1116static struct pernet_operations fib_net_ops = {
1117 .init = fib_net_init,
1118 .exit = fib_net_exit,
1119};
1120
1121void __init ip_fib_init(void)
1122{
Thomas Graf63f34442007-03-22 11:55:17 -07001123 rtnl_register(PF_INET, RTM_NEWROUTE, inet_rtm_newroute, NULL);
1124 rtnl_register(PF_INET, RTM_DELROUTE, inet_rtm_delroute, NULL);
1125 rtnl_register(PF_INET, RTM_GETROUTE, NULL, inet_dump_fib);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001126
1127 register_pernet_subsys(&fib_net_ops);
1128 register_netdevice_notifier(&fib_netdev_notifier);
1129 register_inetaddr_notifier(&fib_inetaddr_notifier);
Stephen Hemminger7f9b8052008-01-14 23:14:20 -08001130
David S. Miller5348ba82011-02-01 15:30:56 -08001131 fib_trie_init();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132}