blob: eb6f69a8f27aff4db2de494389dd7d9584b93ece [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
Stephen Hemminger7f9b8052008-01-14 23:14:20 -080054 local_table = fib_hash_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
Stephen Hemminger7f9b8052008-01-14 23:14:20 -080058 main_table = fib_hash_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
85 tb = fib_hash_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. Lunev010278e2008-01-22 22:04:04 -0800117void fib_select_default(struct net *net,
118 const struct flowi *flp, struct fib_result *res)
Denis V. Lunev64c2d532008-01-22 22:03:33 -0800119{
120 struct fib_table *tb;
121 int table = RT_TABLE_MAIN;
122#ifdef CONFIG_IP_MULTIPLE_TABLES
123 if (res->r == NULL || res->r->action != FR_ACT_TO_TBL)
124 return;
125 table = res->r->table;
126#endif
Denis V. Lunev010278e2008-01-22 22:04:04 -0800127 tb = fib_get_table(net, table);
Denis V. Lunev64c2d532008-01-22 22:03:33 -0800128 if (FIB_RES_GW(*res) && FIB_RES_NH(*res).nh_scope == RT_SCOPE_LINK)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000129 fib_table_select_default(tb, flp, res);
Denis V. Lunev64c2d532008-01-22 22:03:33 -0800130}
131
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800132static void fib_flush(struct net *net)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133{
134 int flushed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700136 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800137 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700138 unsigned int h;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700140 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800141 head = &net->ipv4.fib_table_hash[h];
142 hlist_for_each_entry(tb, node, head, tb_hlist)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000143 flushed += fib_table_flush(tb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
146 if (flushed)
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700147 rt_cache_flush(net, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148}
149
Eric Dumazet82efee12010-09-30 03:31:56 +0000150/**
151 * __ip_dev_find - find the first device with a given source address.
152 * @net: the net namespace
153 * @addr: the source address
154 * @devref: if true, take a reference on the found device
155 *
Eric Dumazet9e917dc2010-10-19 00:39:18 +0000156 * If a caller uses devref=false, it should be protected by RCU, or RTNL
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 */
Eric Dumazet82efee12010-09-30 03:31:56 +0000158struct net_device *__ip_dev_find(struct net *net, __be32 addr, bool devref)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159{
Eric Dumazet82efee12010-09-30 03:31:56 +0000160 struct flowi fl = {
161 .nl_u = {
162 .ip4_u = {
163 .daddr = addr
164 }
165 },
166 .flags = FLOWI_FLAG_MATCH_ANY_IIF
167 };
168 struct fib_result res = { 0 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 struct net_device *dev = NULL;
170
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000171 rcu_read_lock();
172 if (fib_lookup(net, &fl, &res)) {
173 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 return NULL;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000175 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 if (res.type != RTN_LOCAL)
177 goto out;
178 dev = FIB_RES_DEV(res);
179
Eric Dumazet82efee12010-09-30 03:31:56 +0000180 if (dev && devref)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 dev_hold(dev);
182out:
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000183 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 return dev;
185}
Eric Dumazet82efee12010-09-30 03:31:56 +0000186EXPORT_SYMBOL(__ip_dev_find);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800188/*
189 * Find address type as if only "dev" was present in the system. If
190 * on_dev is NULL then all interfaces are taken into consideration.
191 */
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800192static inline unsigned __inet_dev_addr_type(struct net *net,
193 const struct net_device *dev,
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800194 __be32 addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195{
196 struct flowi fl = { .nl_u = { .ip4_u = { .daddr = addr } } };
197 struct fib_result res;
198 unsigned ret = RTN_BROADCAST;
Pavel Emelyanov03cf7862007-10-23 21:17:27 -0700199 struct fib_table *local_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200
Jan Engelhardt1e637c72008-01-21 03:18:08 -0800201 if (ipv4_is_zeronet(addr) || ipv4_is_lbcast(addr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 return RTN_BROADCAST;
Joe Perchesf97c1e02007-12-16 13:45:43 -0800203 if (ipv4_is_multicast(addr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 return RTN_MULTICAST;
205
206#ifdef CONFIG_IP_MULTIPLE_TABLES
207 res.r = NULL;
208#endif
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900209
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800210 local_table = fib_get_table(net, RT_TABLE_LOCAL);
Pavel Emelyanov03cf7862007-10-23 21:17:27 -0700211 if (local_table) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 ret = RTN_UNICAST;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000213 rcu_read_lock();
214 if (!fib_table_lookup(local_table, &fl, &res, FIB_LOOKUP_NOREF)) {
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800215 if (!dev || dev == res.fi->fib_dev)
216 ret = res.type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 }
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000218 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 }
220 return ret;
221}
222
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800223unsigned int inet_addr_type(struct net *net, __be32 addr)
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800224{
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800225 return __inet_dev_addr_type(net, NULL, addr);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800226}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000227EXPORT_SYMBOL(inet_addr_type);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800228
Eric W. Biederman6b175b22008-01-10 03:25:28 -0800229unsigned int inet_dev_addr_type(struct net *net, const struct net_device *dev,
230 __be32 addr)
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800231{
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000232 return __inet_dev_addr_type(net, dev, addr);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800233}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000234EXPORT_SYMBOL(inet_dev_addr_type);
Laszlo Attila Toth05538112007-12-04 23:28:46 -0800235
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236/* Given (packet source, input interface) and optional (dst, oif, tos):
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000237 * - (main) check, that source is valid i.e. not broadcast or our local
238 * address.
239 * - figure out what "logical" interface this packet arrived
240 * and calculate "specific destination" address.
241 * - check, that packet arrived from expected physical interface.
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000242 * called with rcu_read_lock()
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 */
Al Virod9c9df82006-09-26 21:28:14 -0700244int fib_validate_source(__be32 src, __be32 dst, u8 tos, int oif,
jamalb0c110c2009-10-18 02:12:33 +0000245 struct net_device *dev, __be32 *spec_dst,
246 u32 *itag, u32 mark)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247{
248 struct in_device *in_dev;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000249 struct flowi fl = {
250 .nl_u = {
251 .ip4_u = {
252 .daddr = src,
253 .saddr = dst,
254 .tos = tos
255 }
256 },
257 .mark = mark,
258 .iif = oif
259 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 struct fib_result res;
Patrick McHardy8153a102009-12-03 01:25:58 +0000261 int no_addr, rpf, accept_local;
David S. Miller6f86b322010-09-06 22:36:19 -0700262 bool dev_match;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 int ret;
Denis V. Lunev5b707aa2008-01-21 17:33:15 -0800264 struct net *net;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265
Patrick McHardy8153a102009-12-03 01:25:58 +0000266 no_addr = rpf = accept_local = 0;
Herbert Xue5ed6392005-10-03 14:35:55 -0700267 in_dev = __in_dev_get_rcu(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 if (in_dev) {
269 no_addr = in_dev->ifa_list == NULL;
270 rpf = IN_DEV_RPFILTER(in_dev);
Patrick McHardy8153a102009-12-03 01:25:58 +0000271 accept_local = IN_DEV_ACCEPT_LOCAL(in_dev);
Jamal Hadi Salim28f6aee2009-12-25 17:30:22 -0800272 if (mark && !IN_DEV_SRC_VMARK(in_dev))
273 fl.mark = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275
276 if (in_dev == NULL)
277 goto e_inval;
278
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900279 net = dev_net(dev);
Denis V. Lunev5b707aa2008-01-21 17:33:15 -0800280 if (fib_lookup(net, &fl, &res))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 goto last_resort;
Patrick McHardy8153a102009-12-03 01:25:58 +0000282 if (res.type != RTN_UNICAST) {
283 if (res.type != RTN_LOCAL || !accept_local)
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000284 goto e_inval;
Patrick McHardy8153a102009-12-03 01:25:58 +0000285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 *spec_dst = FIB_RES_PREFSRC(res);
287 fib_combine_itag(itag, &res);
David S. Miller6f86b322010-09-06 22:36:19 -0700288 dev_match = false;
289
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290#ifdef CONFIG_IP_ROUTE_MULTIPATH
David S. Miller6f86b322010-09-06 22:36:19 -0700291 for (ret = 0; ret < res.fi->fib_nhs; ret++) {
292 struct fib_nh *nh = &res.fi->fib_nh[ret];
293
294 if (nh->nh_dev == dev) {
295 dev_match = true;
296 break;
297 }
298 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299#else
300 if (FIB_RES_DEV(res) == dev)
David S. Miller6f86b322010-09-06 22:36:19 -0700301 dev_match = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302#endif
David S. Miller6f86b322010-09-06 22:36:19 -0700303 if (dev_match) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 ret = FIB_RES_NH(res).nh_scope >= RT_SCOPE_HOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 return ret;
306 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 if (no_addr)
308 goto last_resort;
Stephen Hemmingerc1cf8422009-02-20 08:25:36 +0000309 if (rpf == 1)
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000310 goto e_rpf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 fl.oif = dev->ifindex;
312
313 ret = 0;
Denis V. Lunev5b707aa2008-01-21 17:33:15 -0800314 if (fib_lookup(net, &fl, &res) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 if (res.type == RTN_UNICAST) {
316 *spec_dst = FIB_RES_PREFSRC(res);
317 ret = FIB_RES_NH(res).nh_scope >= RT_SCOPE_HOST;
318 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319 }
320 return ret;
321
322last_resort:
323 if (rpf)
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000324 goto e_rpf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 *spec_dst = inet_select_addr(dev, 0, RT_SCOPE_UNIVERSE);
326 *itag = 0;
327 return 0;
328
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329e_inval:
330 return -EINVAL;
Eric Dumazetb5f7e752010-06-02 12:05:27 +0000331e_rpf:
332 return -EXDEV;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333}
334
Al Viro81f7bf62006-09-27 18:40:00 -0700335static inline __be32 sk_extract_addr(struct sockaddr *addr)
Thomas Graf4e902c52006-08-17 18:14:52 -0700336{
337 return ((struct sockaddr_in *) addr)->sin_addr.s_addr;
338}
339
340static int put_rtax(struct nlattr *mx, int len, int type, u32 value)
341{
342 struct nlattr *nla;
343
344 nla = (struct nlattr *) ((char *) mx + len);
345 nla->nla_type = type;
346 nla->nla_len = nla_attr_size(4);
347 *(u32 *) nla_data(nla) = value;
348
349 return len + nla_total_size(4);
350}
351
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800352static int rtentry_to_fib_config(struct net *net, int cmd, struct rtentry *rt,
Thomas Graf4e902c52006-08-17 18:14:52 -0700353 struct fib_config *cfg)
354{
Al Viro6d85c102006-09-26 22:15:46 -0700355 __be32 addr;
Thomas Graf4e902c52006-08-17 18:14:52 -0700356 int plen;
357
358 memset(cfg, 0, sizeof(*cfg));
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800359 cfg->fc_nlinfo.nl_net = net;
Thomas Graf4e902c52006-08-17 18:14:52 -0700360
361 if (rt->rt_dst.sa_family != AF_INET)
362 return -EAFNOSUPPORT;
363
364 /*
365 * Check mask for validity:
366 * a) it must be contiguous.
367 * b) destination must have all host bits clear.
368 * c) if application forgot to set correct family (AF_INET),
369 * reject request unless it is absolutely clear i.e.
370 * both family and mask are zero.
371 */
372 plen = 32;
373 addr = sk_extract_addr(&rt->rt_dst);
374 if (!(rt->rt_flags & RTF_HOST)) {
Al Viro81f7bf62006-09-27 18:40:00 -0700375 __be32 mask = sk_extract_addr(&rt->rt_genmask);
Thomas Graf4e902c52006-08-17 18:14:52 -0700376
377 if (rt->rt_genmask.sa_family != AF_INET) {
378 if (mask || rt->rt_genmask.sa_family)
379 return -EAFNOSUPPORT;
380 }
381
382 if (bad_mask(mask, addr))
383 return -EINVAL;
384
385 plen = inet_mask_len(mask);
386 }
387
388 cfg->fc_dst_len = plen;
389 cfg->fc_dst = addr;
390
391 if (cmd != SIOCDELRT) {
392 cfg->fc_nlflags = NLM_F_CREATE;
393 cfg->fc_protocol = RTPROT_BOOT;
394 }
395
396 if (rt->rt_metric)
397 cfg->fc_priority = rt->rt_metric - 1;
398
399 if (rt->rt_flags & RTF_REJECT) {
400 cfg->fc_scope = RT_SCOPE_HOST;
401 cfg->fc_type = RTN_UNREACHABLE;
402 return 0;
403 }
404
405 cfg->fc_scope = RT_SCOPE_NOWHERE;
406 cfg->fc_type = RTN_UNICAST;
407
408 if (rt->rt_dev) {
409 char *colon;
410 struct net_device *dev;
411 char devname[IFNAMSIZ];
412
413 if (copy_from_user(devname, rt->rt_dev, IFNAMSIZ-1))
414 return -EFAULT;
415
416 devname[IFNAMSIZ-1] = 0;
417 colon = strchr(devname, ':');
418 if (colon)
419 *colon = 0;
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800420 dev = __dev_get_by_name(net, devname);
Thomas Graf4e902c52006-08-17 18:14:52 -0700421 if (!dev)
422 return -ENODEV;
423 cfg->fc_oif = dev->ifindex;
424 if (colon) {
425 struct in_ifaddr *ifa;
426 struct in_device *in_dev = __in_dev_get_rtnl(dev);
427 if (!in_dev)
428 return -ENODEV;
429 *colon = ':';
430 for (ifa = in_dev->ifa_list; ifa; ifa = ifa->ifa_next)
431 if (strcmp(ifa->ifa_label, devname) == 0)
432 break;
433 if (ifa == NULL)
434 return -ENODEV;
435 cfg->fc_prefsrc = ifa->ifa_local;
436 }
437 }
438
439 addr = sk_extract_addr(&rt->rt_gateway);
440 if (rt->rt_gateway.sa_family == AF_INET && addr) {
441 cfg->fc_gw = addr;
442 if (rt->rt_flags & RTF_GATEWAY &&
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800443 inet_addr_type(net, addr) == RTN_UNICAST)
Thomas Graf4e902c52006-08-17 18:14:52 -0700444 cfg->fc_scope = RT_SCOPE_UNIVERSE;
445 }
446
447 if (cmd == SIOCDELRT)
448 return 0;
449
450 if (rt->rt_flags & RTF_GATEWAY && !cfg->fc_gw)
451 return -EINVAL;
452
453 if (cfg->fc_scope == RT_SCOPE_NOWHERE)
454 cfg->fc_scope = RT_SCOPE_LINK;
455
456 if (rt->rt_flags & (RTF_MTU | RTF_WINDOW | RTF_IRTT)) {
457 struct nlattr *mx;
458 int len = 0;
459
460 mx = kzalloc(3 * nla_total_size(4), GFP_KERNEL);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900461 if (mx == NULL)
Thomas Graf4e902c52006-08-17 18:14:52 -0700462 return -ENOMEM;
463
464 if (rt->rt_flags & RTF_MTU)
465 len = put_rtax(mx, len, RTAX_ADVMSS, rt->rt_mtu - 40);
466
467 if (rt->rt_flags & RTF_WINDOW)
468 len = put_rtax(mx, len, RTAX_WINDOW, rt->rt_window);
469
470 if (rt->rt_flags & RTF_IRTT)
471 len = put_rtax(mx, len, RTAX_RTT, rt->rt_irtt << 3);
472
473 cfg->fc_mx = mx;
474 cfg->fc_mx_len = len;
475 }
476
477 return 0;
478}
479
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480/*
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000481 * Handle IP routing ioctl calls.
482 * These are used to manipulate the routing tables
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 */
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800484int ip_rt_ioctl(struct net *net, unsigned int cmd, void __user *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485{
Thomas Graf4e902c52006-08-17 18:14:52 -0700486 struct fib_config cfg;
487 struct rtentry rt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
490 switch (cmd) {
491 case SIOCADDRT: /* Add a route */
492 case SIOCDELRT: /* Delete a route */
493 if (!capable(CAP_NET_ADMIN))
494 return -EPERM;
Thomas Graf4e902c52006-08-17 18:14:52 -0700495
496 if (copy_from_user(&rt, arg, sizeof(rt)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 return -EFAULT;
Thomas Graf4e902c52006-08-17 18:14:52 -0700498
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 rtnl_lock();
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800500 err = rtentry_to_fib_config(net, cmd, &rt, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 if (err == 0) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700502 struct fib_table *tb;
503
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 if (cmd == SIOCDELRT) {
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800505 tb = fib_get_table(net, cfg.fc_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 if (tb)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000507 err = fib_table_delete(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700508 else
509 err = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 } else {
Denis V. Lunev1bad1182008-01-10 03:29:53 -0800511 tb = fib_new_table(net, cfg.fc_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 if (tb)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000513 err = fib_table_insert(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700514 else
515 err = -ENOBUFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700517
518 /* allocated by rtentry_to_fib_config() */
519 kfree(cfg.fc_mx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 }
521 rtnl_unlock();
522 return err;
523 }
524 return -EINVAL;
525}
526
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000527const struct nla_policy rtm_ipv4_policy[RTA_MAX + 1] = {
Thomas Graf4e902c52006-08-17 18:14:52 -0700528 [RTA_DST] = { .type = NLA_U32 },
529 [RTA_SRC] = { .type = NLA_U32 },
530 [RTA_IIF] = { .type = NLA_U32 },
531 [RTA_OIF] = { .type = NLA_U32 },
532 [RTA_GATEWAY] = { .type = NLA_U32 },
533 [RTA_PRIORITY] = { .type = NLA_U32 },
534 [RTA_PREFSRC] = { .type = NLA_U32 },
535 [RTA_METRICS] = { .type = NLA_NESTED },
Thomas Graf5176f912006-08-26 20:13:18 -0700536 [RTA_MULTIPATH] = { .len = sizeof(struct rtnexthop) },
Thomas Graf4e902c52006-08-17 18:14:52 -0700537 [RTA_FLOW] = { .type = NLA_U32 },
Thomas Graf4e902c52006-08-17 18:14:52 -0700538};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800540static int rtm_to_fib_config(struct net *net, struct sk_buff *skb,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000541 struct nlmsghdr *nlh, struct fib_config *cfg)
Thomas Graf4e902c52006-08-17 18:14:52 -0700542{
543 struct nlattr *attr;
544 int err, remaining;
545 struct rtmsg *rtm;
546
547 err = nlmsg_validate(nlh, sizeof(*rtm), RTA_MAX, rtm_ipv4_policy);
548 if (err < 0)
549 goto errout;
550
551 memset(cfg, 0, sizeof(*cfg));
552
553 rtm = nlmsg_data(nlh);
Thomas Graf4e902c52006-08-17 18:14:52 -0700554 cfg->fc_dst_len = rtm->rtm_dst_len;
Thomas Graf4e902c52006-08-17 18:14:52 -0700555 cfg->fc_tos = rtm->rtm_tos;
556 cfg->fc_table = rtm->rtm_table;
557 cfg->fc_protocol = rtm->rtm_protocol;
558 cfg->fc_scope = rtm->rtm_scope;
559 cfg->fc_type = rtm->rtm_type;
560 cfg->fc_flags = rtm->rtm_flags;
561 cfg->fc_nlflags = nlh->nlmsg_flags;
562
563 cfg->fc_nlinfo.pid = NETLINK_CB(skb).pid;
564 cfg->fc_nlinfo.nlh = nlh;
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800565 cfg->fc_nlinfo.nl_net = net;
Thomas Graf4e902c52006-08-17 18:14:52 -0700566
Thomas Grafa0ee18b2007-03-24 20:32:54 -0700567 if (cfg->fc_type > RTN_MAX) {
568 err = -EINVAL;
569 goto errout;
570 }
571
Thomas Graf4e902c52006-08-17 18:14:52 -0700572 nlmsg_for_each_attr(attr, nlh, sizeof(struct rtmsg), remaining) {
Thomas Graf8f4c1f92007-09-12 14:44:36 +0200573 switch (nla_type(attr)) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700574 case RTA_DST:
Al Viro17fb2c62006-09-26 22:15:25 -0700575 cfg->fc_dst = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700576 break;
Thomas Graf4e902c52006-08-17 18:14:52 -0700577 case RTA_OIF:
578 cfg->fc_oif = nla_get_u32(attr);
579 break;
580 case RTA_GATEWAY:
Al Viro17fb2c62006-09-26 22:15:25 -0700581 cfg->fc_gw = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700582 break;
583 case RTA_PRIORITY:
584 cfg->fc_priority = nla_get_u32(attr);
585 break;
586 case RTA_PREFSRC:
Al Viro17fb2c62006-09-26 22:15:25 -0700587 cfg->fc_prefsrc = nla_get_be32(attr);
Thomas Graf4e902c52006-08-17 18:14:52 -0700588 break;
589 case RTA_METRICS:
590 cfg->fc_mx = nla_data(attr);
591 cfg->fc_mx_len = nla_len(attr);
592 break;
593 case RTA_MULTIPATH:
594 cfg->fc_mp = nla_data(attr);
595 cfg->fc_mp_len = nla_len(attr);
596 break;
597 case RTA_FLOW:
598 cfg->fc_flow = nla_get_u32(attr);
599 break;
Thomas Graf4e902c52006-08-17 18:14:52 -0700600 case RTA_TABLE:
601 cfg->fc_table = nla_get_u32(attr);
602 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 }
604 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700605
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 return 0;
Thomas Graf4e902c52006-08-17 18:14:52 -0700607errout:
608 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609}
610
Jianjun Kong6ed25332008-11-03 00:25:16 -0800611static int inet_rtm_delroute(struct sk_buff *skb, struct nlmsghdr *nlh, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900613 struct net *net = sock_net(skb->sk);
Thomas Graf4e902c52006-08-17 18:14:52 -0700614 struct fib_config cfg;
615 struct fib_table *tb;
616 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800618 err = rtm_to_fib_config(net, skb, nlh, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700619 if (err < 0)
620 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
Denis V. Lunev8ad49422008-01-10 03:24:11 -0800622 tb = fib_get_table(net, cfg.fc_table);
Thomas Graf4e902c52006-08-17 18:14:52 -0700623 if (tb == NULL) {
624 err = -ESRCH;
625 goto errout;
626 }
627
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000628 err = fib_table_delete(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700629errout:
630 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631}
632
Jianjun Kong6ed25332008-11-03 00:25:16 -0800633static int inet_rtm_newroute(struct sk_buff *skb, struct nlmsghdr *nlh, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900635 struct net *net = sock_net(skb->sk);
Thomas Graf4e902c52006-08-17 18:14:52 -0700636 struct fib_config cfg;
637 struct fib_table *tb;
638 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800640 err = rtm_to_fib_config(net, skb, nlh, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700641 if (err < 0)
642 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
Denis V. Lunev226b0b4a52008-01-10 03:30:24 -0800644 tb = fib_new_table(net, cfg.fc_table);
Thomas Graf4e902c52006-08-17 18:14:52 -0700645 if (tb == NULL) {
646 err = -ENOBUFS;
647 goto errout;
648 }
649
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000650 err = fib_table_insert(tb, &cfg);
Thomas Graf4e902c52006-08-17 18:14:52 -0700651errout:
652 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653}
654
Thomas Graf63f34442007-03-22 11:55:17 -0700655static int inet_dump_fib(struct sk_buff *skb, struct netlink_callback *cb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900657 struct net *net = sock_net(skb->sk);
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700658 unsigned int h, s_h;
659 unsigned int e = 0, s_e;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 struct fib_table *tb;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700661 struct hlist_node *node;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800662 struct hlist_head *head;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700663 int dumped = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
Thomas Grafbe403ea2006-08-17 18:15:17 -0700665 if (nlmsg_len(cb->nlh) >= sizeof(struct rtmsg) &&
666 ((struct rtmsg *) nlmsg_data(cb->nlh))->rtm_flags & RTM_F_CLONED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 return ip_rt_dump(skb, cb);
668
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700669 s_h = cb->args[0];
670 s_e = cb->args[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700672 for (h = s_h; h < FIB_TABLE_HASHSZ; h++, s_e = 0) {
673 e = 0;
Denis V. Luneve4aef8a2008-01-10 03:28:24 -0800674 head = &net->ipv4.fib_table_hash[h];
675 hlist_for_each_entry(tb, node, head, tb_hlist) {
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700676 if (e < s_e)
677 goto next;
678 if (dumped)
679 memset(&cb->args[2], 0, sizeof(cb->args) -
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900680 2 * sizeof(cb->args[0]));
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000681 if (fib_table_dump(tb, skb, cb) < 0)
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700682 goto out;
683 dumped = 1;
684next:
685 e++;
686 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 }
Patrick McHardy1af5a8c2006-08-10 23:10:46 -0700688out:
689 cb->args[1] = e;
690 cb->args[0] = h;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
692 return skb->len;
693}
694
695/* Prepare and feed intra-kernel routing request.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000696 * Really, it should be netlink message, but :-( netlink
697 * can be not configured, so that we feed it directly
698 * to fib engine. It is legal, because all events occur
699 * only when netlink is already locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 */
Al Viro81f7bf62006-09-27 18:40:00 -0700701static void fib_magic(int cmd, int type, __be32 dst, int dst_len, struct in_ifaddr *ifa)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702{
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900703 struct net *net = dev_net(ifa->ifa_dev->dev);
Thomas Graf4e902c52006-08-17 18:14:52 -0700704 struct fib_table *tb;
705 struct fib_config cfg = {
706 .fc_protocol = RTPROT_KERNEL,
707 .fc_type = type,
708 .fc_dst = dst,
709 .fc_dst_len = dst_len,
710 .fc_prefsrc = ifa->ifa_local,
711 .fc_oif = ifa->ifa_dev->dev->ifindex,
712 .fc_nlflags = NLM_F_CREATE | NLM_F_APPEND,
Denis V. Lunev4d1169c2008-01-10 03:26:13 -0800713 .fc_nlinfo = {
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800714 .nl_net = net,
Denis V. Lunev4d1169c2008-01-10 03:26:13 -0800715 },
Thomas Graf4e902c52006-08-17 18:14:52 -0700716 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717
718 if (type == RTN_UNICAST)
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800719 tb = fib_new_table(net, RT_TABLE_MAIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 else
Denis V. Lunev4b5d47d2008-01-10 03:29:23 -0800721 tb = fib_new_table(net, RT_TABLE_LOCAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722
723 if (tb == NULL)
724 return;
725
Thomas Graf4e902c52006-08-17 18:14:52 -0700726 cfg.fc_table = tb->tb_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727
Thomas Graf4e902c52006-08-17 18:14:52 -0700728 if (type != RTN_LOCAL)
729 cfg.fc_scope = RT_SCOPE_LINK;
730 else
731 cfg.fc_scope = RT_SCOPE_HOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732
733 if (cmd == RTM_NEWROUTE)
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000734 fib_table_insert(tb, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 else
Stephen Hemminger16c6cf82009-09-20 10:35:36 +0000736 fib_table_delete(tb, &cfg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737}
738
Jamal Hadi Salim0ff60a42005-11-22 14:47:37 -0800739void fib_add_ifaddr(struct in_ifaddr *ifa)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740{
741 struct in_device *in_dev = ifa->ifa_dev;
742 struct net_device *dev = in_dev->dev;
743 struct in_ifaddr *prim = ifa;
Al Viroa144ea42006-09-28 18:00:55 -0700744 __be32 mask = ifa->ifa_mask;
745 __be32 addr = ifa->ifa_local;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000746 __be32 prefix = ifa->ifa_address & mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000748 if (ifa->ifa_flags & IFA_F_SECONDARY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 prim = inet_ifa_byprefix(in_dev, prefix, mask);
750 if (prim == NULL) {
Stephen Hemmingera6db9012008-01-12 20:58:35 -0800751 printk(KERN_WARNING "fib_add_ifaddr: bug: prim == NULL\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 return;
753 }
754 }
755
756 fib_magic(RTM_NEWROUTE, RTN_LOCAL, addr, 32, prim);
757
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000758 if (!(dev->flags & IFF_UP))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 return;
760
761 /* Add broadcast address, if it is explicitly assigned. */
Al Viroa144ea42006-09-28 18:00:55 -0700762 if (ifa->ifa_broadcast && ifa->ifa_broadcast != htonl(0xFFFFFFFF))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, ifa->ifa_broadcast, 32, prim);
764
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000765 if (!ipv4_is_zeronet(prefix) && !(ifa->ifa_flags & IFA_F_SECONDARY) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766 (prefix != addr || ifa->ifa_prefixlen < 32)) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000767 fib_magic(RTM_NEWROUTE,
768 dev->flags & IFF_LOOPBACK ? RTN_LOCAL : RTN_UNICAST,
769 prefix, ifa->ifa_prefixlen, prim);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770
771 /* Add network specific broadcasts, when it takes a sense */
772 if (ifa->ifa_prefixlen < 31) {
773 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, prefix, 32, prim);
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000774 fib_magic(RTM_NEWROUTE, RTN_BROADCAST, prefix | ~mask,
775 32, prim);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 }
777 }
778}
779
780static void fib_del_ifaddr(struct in_ifaddr *ifa)
781{
782 struct in_device *in_dev = ifa->ifa_dev;
783 struct net_device *dev = in_dev->dev;
784 struct in_ifaddr *ifa1;
785 struct in_ifaddr *prim = ifa;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000786 __be32 brd = ifa->ifa_address | ~ifa->ifa_mask;
787 __be32 any = ifa->ifa_address & ifa->ifa_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788#define LOCAL_OK 1
789#define BRD_OK 2
790#define BRD0_OK 4
791#define BRD1_OK 8
792 unsigned ok = 0;
793
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000794 if (!(ifa->ifa_flags & IFA_F_SECONDARY))
795 fib_magic(RTM_DELROUTE,
796 dev->flags & IFF_LOOPBACK ? RTN_LOCAL : RTN_UNICAST,
797 any, ifa->ifa_prefixlen, prim);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 else {
799 prim = inet_ifa_byprefix(in_dev, any, ifa->ifa_mask);
800 if (prim == NULL) {
Stephen Hemmingera6db9012008-01-12 20:58:35 -0800801 printk(KERN_WARNING "fib_del_ifaddr: bug: prim == NULL\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 return;
803 }
804 }
805
806 /* Deletion is more complicated than add.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000807 * We should take care of not to delete too much :-)
808 *
809 * Scan address list to be sure that addresses are really gone.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 */
811
812 for (ifa1 = in_dev->ifa_list; ifa1; ifa1 = ifa1->ifa_next) {
813 if (ifa->ifa_local == ifa1->ifa_local)
814 ok |= LOCAL_OK;
815 if (ifa->ifa_broadcast == ifa1->ifa_broadcast)
816 ok |= BRD_OK;
817 if (brd == ifa1->ifa_broadcast)
818 ok |= BRD1_OK;
819 if (any == ifa1->ifa_broadcast)
820 ok |= BRD0_OK;
821 }
822
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000823 if (!(ok & BRD_OK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 fib_magic(RTM_DELROUTE, RTN_BROADCAST, ifa->ifa_broadcast, 32, prim);
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000825 if (!(ok & BRD1_OK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 fib_magic(RTM_DELROUTE, RTN_BROADCAST, brd, 32, prim);
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000827 if (!(ok & BRD0_OK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 fib_magic(RTM_DELROUTE, RTN_BROADCAST, any, 32, prim);
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000829 if (!(ok & LOCAL_OK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 fib_magic(RTM_DELROUTE, RTN_LOCAL, ifa->ifa_local, 32, prim);
831
832 /* Check, that this local address finally disappeared. */
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900833 if (inet_addr_type(dev_net(dev), ifa->ifa_local) != RTN_LOCAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 /* And the last, but not the least thing.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000835 * We must flush stray FIB entries.
836 *
837 * First of all, we scan fib_info list searching
838 * for stray nexthop entries, then ignite fib_flush.
839 */
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900840 if (fib_sync_down_addr(dev_net(dev), ifa->ifa_local))
841 fib_flush(dev_net(dev));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 }
843 }
844#undef LOCAL_OK
845#undef BRD_OK
846#undef BRD0_OK
847#undef BRD1_OK
848}
849
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000850static void nl_fib_lookup(struct fib_result_nl *frn, struct fib_table *tb)
Robert Olsson246955f2005-06-20 13:36:39 -0700851{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900852
Robert Olsson246955f2005-06-20 13:36:39 -0700853 struct fib_result res;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000854 struct flowi fl = {
855 .mark = frn->fl_mark,
856 .nl_u = {
857 .ip4_u = {
858 .daddr = frn->fl_addr,
859 .tos = frn->fl_tos,
860 .scope = frn->fl_scope
861 }
862 }
863 };
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700864
Sergey Vlasov912a41a2007-04-27 02:17:19 -0700865#ifdef CONFIG_IP_MULTIPLE_TABLES
866 res.r = NULL;
867#endif
868
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700869 frn->err = -ENOENT;
Robert Olsson246955f2005-06-20 13:36:39 -0700870 if (tb) {
871 local_bh_disable();
872
873 frn->tb_id = tb->tb_id;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000874 rcu_read_lock();
875 frn->err = fib_table_lookup(tb, &fl, &res, FIB_LOOKUP_NOREF);
Robert Olsson246955f2005-06-20 13:36:39 -0700876
877 if (!frn->err) {
878 frn->prefixlen = res.prefixlen;
879 frn->nh_sel = res.nh_sel;
880 frn->type = res.type;
881 frn->scope = res.scope;
882 }
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000883 rcu_read_unlock();
Robert Olsson246955f2005-06-20 13:36:39 -0700884 local_bh_enable();
885 }
886}
887
David S. Miller28f7b0362007-10-10 21:32:39 -0700888static void nl_fib_input(struct sk_buff *skb)
Robert Olsson246955f2005-06-20 13:36:39 -0700889{
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800890 struct net *net;
Robert Olsson246955f2005-06-20 13:36:39 -0700891 struct fib_result_nl *frn;
David S. Miller28f7b0362007-10-10 21:32:39 -0700892 struct nlmsghdr *nlh;
Robert Olsson246955f2005-06-20 13:36:39 -0700893 struct fib_table *tb;
David S. Miller28f7b0362007-10-10 21:32:39 -0700894 u32 pid;
Alexey Kuznetsov1194ed02007-04-25 13:07:28 -0700895
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900896 net = sock_net(skb->sk);
Arnaldo Carvalho de Melob529ccf2007-04-25 19:08:35 -0700897 nlh = nlmsg_hdr(skb);
Thomas Grafea865752005-12-01 14:30:00 -0800898 if (skb->len < NLMSG_SPACE(0) || skb->len < nlh->nlmsg_len ||
Denis V. Lunevd883a032007-12-21 02:01:53 -0800899 nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*frn)))
Thomas Grafea865752005-12-01 14:30:00 -0800900 return;
Denis V. Lunevd883a032007-12-21 02:01:53 -0800901
902 skb = skb_clone(skb, GFP_KERNEL);
903 if (skb == NULL)
904 return;
905 nlh = nlmsg_hdr(skb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900906
Robert Olsson246955f2005-06-20 13:36:39 -0700907 frn = (struct fib_result_nl *) NLMSG_DATA(nlh);
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800908 tb = fib_get_table(net, frn->tb_id_in);
Robert Olsson246955f2005-06-20 13:36:39 -0700909
910 nl_fib_lookup(frn, tb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900911
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000912 pid = NETLINK_CB(skb).pid; /* pid of sending process */
913 NETLINK_CB(skb).pid = 0; /* from kernel */
Patrick McHardyac6d4392005-08-14 19:29:52 -0700914 NETLINK_CB(skb).dst_group = 0; /* unicast */
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800915 netlink_unicast(net->ipv4.fibnl, skb, pid, MSG_DONTWAIT);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900916}
Robert Olsson246955f2005-06-20 13:36:39 -0700917
Alexey Dobriyan2c8c1e72010-01-17 03:35:32 +0000918static int __net_init nl_fib_lookup_init(struct net *net)
Robert Olsson246955f2005-06-20 13:36:39 -0700919{
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800920 struct sock *sk;
921 sk = netlink_kernel_create(net, NETLINK_FIB_LOOKUP, 0,
922 nl_fib_input, NULL, THIS_MODULE);
923 if (sk == NULL)
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -0800924 return -EAFNOSUPPORT;
Denis V. Lunev6bd48fc2008-01-10 03:28:55 -0800925 net->ipv4.fibnl = sk;
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -0800926 return 0;
927}
928
929static void nl_fib_lookup_exit(struct net *net)
930{
Denis V. Lunevb7c6ba62008-01-28 14:41:19 -0800931 netlink_kernel_release(net->ipv4.fibnl);
Denis V. Lunev775516b2008-01-18 23:55:19 -0800932 net->ipv4.fibnl = NULL;
Robert Olsson246955f2005-06-20 13:36:39 -0700933}
934
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000935static void fib_disable_ip(struct net_device *dev, int force, int delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936{
Denis V. Lunev85326fa2008-01-31 18:48:47 -0800937 if (fib_sync_down_dev(dev, force))
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900938 fib_flush(dev_net(dev));
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000939 rt_cache_flush(dev_net(dev), delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 arp_ifdown(dev);
941}
942
943static int fib_inetaddr_event(struct notifier_block *this, unsigned long event, void *ptr)
944{
Jianjun Kong6ed25332008-11-03 00:25:16 -0800945 struct in_ifaddr *ifa = (struct in_ifaddr *)ptr;
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700946 struct net_device *dev = ifa->ifa_dev->dev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
948 switch (event) {
949 case NETDEV_UP:
950 fib_add_ifaddr(ifa);
951#ifdef CONFIG_IP_ROUTE_MULTIPATH
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700952 fib_sync_up(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953#endif
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700954 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 break;
956 case NETDEV_DOWN:
957 fib_del_ifaddr(ifa);
Jayachandran C9fcc2e82005-10-27 15:10:01 -0700958 if (ifa->ifa_dev->ifa_list == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 /* Last address was deleted from this interface.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000960 * Disable IP.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 */
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000962 fib_disable_ip(dev, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 } else {
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700964 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 }
966 break;
967 }
968 return NOTIFY_DONE;
969}
970
971static int fib_netdev_event(struct notifier_block *this, unsigned long event, void *ptr)
972{
973 struct net_device *dev = ptr;
Herbert Xue5ed6392005-10-03 14:35:55 -0700974 struct in_device *in_dev = __in_dev_get_rtnl(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
976 if (event == NETDEV_UNREGISTER) {
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000977 fib_disable_ip(dev, 2, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 return NOTIFY_DONE;
979 }
980
981 if (!in_dev)
982 return NOTIFY_DONE;
983
984 switch (event) {
985 case NETDEV_UP:
986 for_ifa(in_dev) {
987 fib_add_ifaddr(ifa);
988 } endfor_ifa(in_dev);
989#ifdef CONFIG_IP_ROUTE_MULTIPATH
990 fib_sync_up(dev);
991#endif
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700992 rt_cache_flush(dev_net(dev), -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 break;
994 case NETDEV_DOWN:
Octavian Purdilae2ce1462009-11-16 13:49:49 +0000995 fib_disable_ip(dev, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 break;
997 case NETDEV_CHANGEMTU:
998 case NETDEV_CHANGE:
Denis V. Lunev76e6ebf2008-07-05 19:00:44 -0700999 rt_cache_flush(dev_net(dev), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 break;
Eric W. Biedermana5ee1552009-11-29 15:45:58 +00001001 case NETDEV_UNREGISTER_BATCH:
1002 rt_cache_flush_batch();
1003 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 }
1005 return NOTIFY_DONE;
1006}
1007
1008static struct notifier_block fib_inetaddr_notifier = {
Jianjun Kong6ed25332008-11-03 00:25:16 -08001009 .notifier_call = fib_inetaddr_event,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010};
1011
1012static struct notifier_block fib_netdev_notifier = {
Jianjun Kong6ed25332008-11-03 00:25:16 -08001013 .notifier_call = fib_netdev_event,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014};
1015
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001016static int __net_init ip_fib_net_init(struct net *net)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017{
Denis V. Lunevdce5cbe2008-01-31 18:44:53 -08001018 int err;
Eric Dumazet10da66f2010-10-13 08:22:03 +00001019 size_t size = sizeof(struct hlist_head) * FIB_TABLE_HASHSZ;
Patrick McHardy1af5a8c2006-08-10 23:10:46 -07001020
Eric Dumazet10da66f2010-10-13 08:22:03 +00001021 /* Avoid false sharing : Use at least a full cache line */
1022 size = max_t(size_t, size, L1_CACHE_BYTES);
1023
1024 net->ipv4.fib_table_hash = kzalloc(size, GFP_KERNEL);
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001025 if (net->ipv4.fib_table_hash == NULL)
1026 return -ENOMEM;
1027
Denis V. Lunevdce5cbe2008-01-31 18:44:53 -08001028 err = fib4_rules_init(net);
1029 if (err < 0)
1030 goto fail;
1031 return 0;
1032
1033fail:
1034 kfree(net->ipv4.fib_table_hash);
1035 return err;
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001036}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037
Alexey Dobriyan2c8c1e72010-01-17 03:35:32 +00001038static void ip_fib_net_exit(struct net *net)
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001039{
1040 unsigned int i;
Thomas Graf63f34442007-03-22 11:55:17 -07001041
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001042#ifdef CONFIG_IP_MULTIPLE_TABLES
1043 fib4_rules_exit(net);
1044#endif
1045
1046 for (i = 0; i < FIB_TABLE_HASHSZ; i++) {
1047 struct fib_table *tb;
1048 struct hlist_head *head;
1049 struct hlist_node *node, *tmp;
1050
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001051 head = &net->ipv4.fib_table_hash[i];
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001052 hlist_for_each_entry_safe(tb, node, tmp, head, tb_hlist) {
1053 hlist_del(node);
Stephen Hemminger16c6cf82009-09-20 10:35:36 +00001054 fib_table_flush(tb);
Pavel Emelyanov4aa2c462010-10-28 02:00:43 +00001055 fib_free_table(tb);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001056 }
1057 }
Denis V. Luneve4aef8a2008-01-10 03:28:24 -08001058 kfree(net->ipv4.fib_table_hash);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001059}
1060
1061static int __net_init fib_net_init(struct net *net)
1062{
1063 int error;
1064
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001065 error = ip_fib_net_init(net);
1066 if (error < 0)
1067 goto out;
1068 error = nl_fib_lookup_init(net);
1069 if (error < 0)
1070 goto out_nlfl;
1071 error = fib_proc_init(net);
1072 if (error < 0)
1073 goto out_proc;
1074out:
1075 return error;
1076
1077out_proc:
1078 nl_fib_lookup_exit(net);
1079out_nlfl:
1080 ip_fib_net_exit(net);
1081 goto out;
1082}
1083
1084static void __net_exit fib_net_exit(struct net *net)
1085{
1086 fib_proc_exit(net);
1087 nl_fib_lookup_exit(net);
1088 ip_fib_net_exit(net);
1089}
1090
1091static struct pernet_operations fib_net_ops = {
1092 .init = fib_net_init,
1093 .exit = fib_net_exit,
1094};
1095
1096void __init ip_fib_init(void)
1097{
Thomas Graf63f34442007-03-22 11:55:17 -07001098 rtnl_register(PF_INET, RTM_NEWROUTE, inet_rtm_newroute, NULL);
1099 rtnl_register(PF_INET, RTM_DELROUTE, inet_rtm_delroute, NULL);
1100 rtnl_register(PF_INET, RTM_GETROUTE, NULL, inet_dump_fib);
Denis V. Lunev7b1a74f2008-01-10 03:22:17 -08001101
1102 register_pernet_subsys(&fib_net_ops);
1103 register_netdevice_notifier(&fib_netdev_notifier);
1104 register_inetaddr_notifier(&fib_inetaddr_notifier);
Stephen Hemminger7f9b8052008-01-14 23:14:20 -08001105
1106 fib_hash_init();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107}