blob: 622ac4c95026150a76d09e7df9a49aadf9e09420 [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: semantics.
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 <asm/uaccess.h>
17#include <asm/system.h>
18#include <linux/bitops.h>
19#include <linux/types.h>
20#include <linux/kernel.h>
21#include <linux/jiffies.h>
22#include <linux/mm.h>
23#include <linux/string.h>
24#include <linux/socket.h>
25#include <linux/sockios.h>
26#include <linux/errno.h>
27#include <linux/in.h>
28#include <linux/inet.h>
Arnaldo Carvalho de Melo14c85022005-12-27 02:43:12 -020029#include <linux/inetdevice.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/netdevice.h>
31#include <linux/if_arp.h>
32#include <linux/proc_fs.h>
33#include <linux/skbuff.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/init.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090035#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
Arnaldo Carvalho de Melo14c85022005-12-27 02:43:12 -020037#include <net/arp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <net/ip.h>
39#include <net/protocol.h>
40#include <net/route.h>
41#include <net/tcp.h>
42#include <net/sock.h>
43#include <net/ip_fib.h>
Thomas Graff21c7bc2006-08-15 00:34:17 -070044#include <net/netlink.h>
Thomas Graf4e902c52006-08-17 18:14:52 -070045#include <net/nexthop.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include "fib_lookup.h"
48
Stephen Hemminger832b4c52006-08-29 16:48:09 -070049static DEFINE_SPINLOCK(fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -070050static struct hlist_head *fib_info_hash;
51static struct hlist_head *fib_info_laddrhash;
David S. Miller123b9732011-02-01 15:34:21 -080052static unsigned int fib_info_hash_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -070053static unsigned int fib_info_cnt;
54
55#define DEVINDEX_HASHBITS 8
56#define DEVINDEX_HASHSIZE (1U << DEVINDEX_HASHBITS)
57static struct hlist_head fib_info_devhash[DEVINDEX_HASHSIZE];
58
59#ifdef CONFIG_IP_ROUTE_MULTIPATH
60
61static DEFINE_SPINLOCK(fib_multipath_lock);
62
Eric Dumazet6a31d2a2010-10-04 20:00:18 +000063#define for_nexthops(fi) { \
64 int nhsel; const struct fib_nh *nh; \
65 for (nhsel = 0, nh = (fi)->fib_nh; \
66 nhsel < (fi)->fib_nhs; \
67 nh++, nhsel++)
Linus Torvalds1da177e2005-04-16 15:20:36 -070068
Eric Dumazet6a31d2a2010-10-04 20:00:18 +000069#define change_nexthops(fi) { \
70 int nhsel; struct fib_nh *nexthop_nh; \
71 for (nhsel = 0, nexthop_nh = (struct fib_nh *)((fi)->fib_nh); \
72 nhsel < (fi)->fib_nhs; \
73 nexthop_nh++, nhsel++)
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
75#else /* CONFIG_IP_ROUTE_MULTIPATH */
76
77/* Hope, that gcc will optimize it to get rid of dummy loop */
78
Eric Dumazet6a31d2a2010-10-04 20:00:18 +000079#define for_nexthops(fi) { \
80 int nhsel; const struct fib_nh *nh = (fi)->fib_nh; \
81 for (nhsel = 0; nhsel < 1; nhsel++)
Linus Torvalds1da177e2005-04-16 15:20:36 -070082
Eric Dumazet6a31d2a2010-10-04 20:00:18 +000083#define change_nexthops(fi) { \
84 int nhsel; \
85 struct fib_nh *nexthop_nh = (struct fib_nh *)((fi)->fib_nh); \
86 for (nhsel = 0; nhsel < 1; nhsel++)
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88#endif /* CONFIG_IP_ROUTE_MULTIPATH */
89
90#define endfor_nexthops(fi) }
91
92
David S. Miller3be06862011-03-07 15:01:10 -080093const struct fib_prop fib_props[RTN_MAX + 1] = {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +000094 [RTN_UNSPEC] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 .error = 0,
96 .scope = RT_SCOPE_NOWHERE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +000097 },
98 [RTN_UNICAST] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -070099 .error = 0,
100 .scope = RT_SCOPE_UNIVERSE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000101 },
102 [RTN_LOCAL] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 .error = 0,
104 .scope = RT_SCOPE_HOST,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000105 },
106 [RTN_BROADCAST] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 .error = 0,
108 .scope = RT_SCOPE_LINK,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000109 },
110 [RTN_ANYCAST] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 .error = 0,
112 .scope = RT_SCOPE_LINK,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000113 },
114 [RTN_MULTICAST] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 .error = 0,
116 .scope = RT_SCOPE_UNIVERSE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000117 },
118 [RTN_BLACKHOLE] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 .error = -EINVAL,
120 .scope = RT_SCOPE_UNIVERSE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000121 },
122 [RTN_UNREACHABLE] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 .error = -EHOSTUNREACH,
124 .scope = RT_SCOPE_UNIVERSE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000125 },
126 [RTN_PROHIBIT] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 .error = -EACCES,
128 .scope = RT_SCOPE_UNIVERSE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000129 },
130 [RTN_THROW] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 .error = -EAGAIN,
132 .scope = RT_SCOPE_UNIVERSE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000133 },
134 [RTN_NAT] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 .error = -EINVAL,
136 .scope = RT_SCOPE_NOWHERE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000137 },
138 [RTN_XRESOLVE] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 .error = -EINVAL,
140 .scope = RT_SCOPE_NOWHERE,
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000141 },
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142};
143
144
145/* Release a nexthop info record */
146
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000147static void free_fib_info_rcu(struct rcu_head *head)
148{
149 struct fib_info *fi = container_of(head, struct fib_info, rcu);
150
David S. Miller725d1e12011-01-28 14:05:05 -0800151 if (fi->fib_metrics != (u32 *) dst_default_metrics)
152 kfree(fi->fib_metrics);
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000153 kfree(fi);
154}
155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156void free_fib_info(struct fib_info *fi)
157{
158 if (fi->fib_dead == 0) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000159 pr_warning("Freeing alive fib_info %p\n", fi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 return;
161 }
162 change_nexthops(fi) {
David S. Miller71fceff2010-01-15 01:16:40 -0800163 if (nexthop_nh->nh_dev)
164 dev_put(nexthop_nh->nh_dev);
165 nexthop_nh->nh_dev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 } endfor_nexthops(fi);
167 fib_info_cnt--;
Denis V. Lunev57d7a602008-04-16 02:00:50 -0700168 release_net(fi->fib_net);
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000169 call_rcu(&fi->rcu, free_fib_info_rcu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170}
171
172void fib_release_info(struct fib_info *fi)
173{
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700174 spin_lock_bh(&fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 if (fi && --fi->fib_treeref == 0) {
176 hlist_del(&fi->fib_hash);
177 if (fi->fib_prefsrc)
178 hlist_del(&fi->fib_lhash);
179 change_nexthops(fi) {
David S. Miller71fceff2010-01-15 01:16:40 -0800180 if (!nexthop_nh->nh_dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 continue;
David S. Miller71fceff2010-01-15 01:16:40 -0800182 hlist_del(&nexthop_nh->nh_hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 } endfor_nexthops(fi)
184 fi->fib_dead = 1;
185 fib_info_put(fi);
186 }
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700187 spin_unlock_bh(&fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188}
189
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000190static inline int nh_comp(const struct fib_info *fi, const struct fib_info *ofi)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191{
192 const struct fib_nh *onh = ofi->fib_nh;
193
194 for_nexthops(fi) {
195 if (nh->nh_oif != onh->nh_oif ||
196 nh->nh_gw != onh->nh_gw ||
197 nh->nh_scope != onh->nh_scope ||
198#ifdef CONFIG_IP_ROUTE_MULTIPATH
199 nh->nh_weight != onh->nh_weight ||
200#endif
Patrick McHardyc7066f72011-01-14 13:36:42 +0100201#ifdef CONFIG_IP_ROUTE_CLASSID
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 nh->nh_tclassid != onh->nh_tclassid ||
203#endif
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000204 ((nh->nh_flags ^ onh->nh_flags) & ~RTNH_F_DEAD))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 return -1;
206 onh++;
207 } endfor_nexthops(fi);
208 return 0;
209}
210
David S. Miller88ebc722008-01-12 21:49:01 -0800211static inline unsigned int fib_devindex_hashfn(unsigned int val)
212{
213 unsigned int mask = DEVINDEX_HASHSIZE - 1;
214
215 return (val ^
216 (val >> DEVINDEX_HASHBITS) ^
217 (val >> (DEVINDEX_HASHBITS * 2))) & mask;
218}
219
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220static inline unsigned int fib_info_hashfn(const struct fib_info *fi)
221{
David S. Miller123b9732011-02-01 15:34:21 -0800222 unsigned int mask = (fib_info_hash_size - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 unsigned int val = fi->fib_nhs;
224
225 val ^= fi->fib_protocol;
Al Viro81f7bf62006-09-27 18:40:00 -0700226 val ^= (__force u32)fi->fib_prefsrc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227 val ^= fi->fib_priority;
David S. Miller88ebc722008-01-12 21:49:01 -0800228 for_nexthops(fi) {
229 val ^= fib_devindex_hashfn(nh->nh_oif);
230 } endfor_nexthops(fi)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
232 return (val ^ (val >> 7) ^ (val >> 12)) & mask;
233}
234
235static struct fib_info *fib_find_info(const struct fib_info *nfi)
236{
237 struct hlist_head *head;
238 struct hlist_node *node;
239 struct fib_info *fi;
240 unsigned int hash;
241
242 hash = fib_info_hashfn(nfi);
243 head = &fib_info_hash[hash];
244
245 hlist_for_each_entry(fi, node, head, fib_hash) {
Octavian Purdila09ad9bc2009-11-25 15:14:13 -0800246 if (!net_eq(fi->fib_net, nfi->fib_net))
Denis V. Lunev4814bdb2008-01-31 18:50:07 -0800247 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 if (fi->fib_nhs != nfi->fib_nhs)
249 continue;
250 if (nfi->fib_protocol == fi->fib_protocol &&
251 nfi->fib_prefsrc == fi->fib_prefsrc &&
252 nfi->fib_priority == fi->fib_priority &&
253 memcmp(nfi->fib_metrics, fi->fib_metrics,
254 sizeof(fi->fib_metrics)) == 0 &&
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000255 ((nfi->fib_flags ^ fi->fib_flags) & ~RTNH_F_DEAD) == 0 &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 (nfi->fib_nhs == 0 || nh_comp(fi, nfi) == 0))
257 return fi;
258 }
259
260 return NULL;
261}
262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263/* Check, that the gateway is already configured.
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000264 * Used only by redirect accept routine.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 */
Al Virod878e72e2006-09-26 22:18:13 -0700266int ip_fib_check_default(__be32 gw, struct net_device *dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267{
268 struct hlist_head *head;
269 struct hlist_node *node;
270 struct fib_nh *nh;
271 unsigned int hash;
272
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700273 spin_lock(&fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274
275 hash = fib_devindex_hashfn(dev->ifindex);
276 head = &fib_info_devhash[hash];
277 hlist_for_each_entry(nh, node, head, nh_hash) {
278 if (nh->nh_dev == dev &&
279 nh->nh_gw == gw &&
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000280 !(nh->nh_flags & RTNH_F_DEAD)) {
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700281 spin_unlock(&fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 return 0;
283 }
284 }
285
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700286 spin_unlock(&fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287
288 return -1;
289}
290
Thomas Graf339bf982006-11-10 14:10:15 -0800291static inline size_t fib_nlmsg_size(struct fib_info *fi)
292{
293 size_t payload = NLMSG_ALIGN(sizeof(struct rtmsg))
294 + nla_total_size(4) /* RTA_TABLE */
295 + nla_total_size(4) /* RTA_DST */
296 + nla_total_size(4) /* RTA_PRIORITY */
297 + nla_total_size(4); /* RTA_PREFSRC */
298
299 /* space for nested metrics */
300 payload += nla_total_size((RTAX_MAX * nla_total_size(4)));
301
302 if (fi->fib_nhs) {
303 /* Also handles the special case fib_nhs == 1 */
304
305 /* each nexthop is packed in an attribute */
306 size_t nhsize = nla_total_size(sizeof(struct rtnexthop));
307
308 /* may contain flow and gateway attribute */
309 nhsize += 2 * nla_total_size(4);
310
311 /* all nexthops are packed in a nested attribute */
312 payload += nla_total_size(fi->fib_nhs * nhsize);
313 }
314
315 return payload;
316}
317
Al Viro81f7bf62006-09-27 18:40:00 -0700318void rtmsg_fib(int event, __be32 key, struct fib_alias *fa,
Milan Kocianb8f55832007-05-23 14:55:06 -0700319 int dst_len, u32 tb_id, struct nl_info *info,
320 unsigned int nlm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321{
322 struct sk_buff *skb;
Thomas Graf4e902c52006-08-17 18:14:52 -0700323 u32 seq = info->nlh ? info->nlh->nlmsg_seq : 0;
Thomas Graff21c7bc2006-08-15 00:34:17 -0700324 int err = -ENOBUFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325
Thomas Graf339bf982006-11-10 14:10:15 -0800326 skb = nlmsg_new(fib_nlmsg_size(fa->fa_info), GFP_KERNEL);
Thomas Graff21c7bc2006-08-15 00:34:17 -0700327 if (skb == NULL)
328 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329
Thomas Graf4e902c52006-08-17 18:14:52 -0700330 err = fib_dump_info(skb, info->pid, seq, event, tb_id,
Thomas Grafbe403ea2006-08-17 18:15:17 -0700331 fa->fa_type, fa->fa_scope, key, dst_len,
Milan Kocianb8f55832007-05-23 14:55:06 -0700332 fa->fa_tos, fa->fa_info, nlm_flags);
Patrick McHardy26932562007-01-31 23:16:40 -0800333 if (err < 0) {
334 /* -EMSGSIZE implies BUG in fib_nlmsg_size() */
335 WARN_ON(err == -EMSGSIZE);
336 kfree_skb(skb);
337 goto errout;
338 }
Pablo Neira Ayuso1ce85fe2009-02-24 23:18:28 -0800339 rtnl_notify(skb, info->nl_net, info->pid, RTNLGRP_IPV4_ROUTE,
340 info->nlh, GFP_KERNEL);
341 return;
Thomas Graff21c7bc2006-08-15 00:34:17 -0700342errout:
343 if (err < 0)
Denis V. Lunev4d1169c2008-01-10 03:26:13 -0800344 rtnl_set_sk_err(info->nl_net, RTNLGRP_IPV4_ROUTE, err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345}
346
347/* Return the first fib alias matching TOS with
348 * priority less than or equal to PRIO.
349 */
350struct fib_alias *fib_find_alias(struct list_head *fah, u8 tos, u32 prio)
351{
352 if (fah) {
353 struct fib_alias *fa;
354 list_for_each_entry(fa, fah, fa_list) {
355 if (fa->fa_tos > tos)
356 continue;
357 if (fa->fa_info->fib_priority >= prio ||
358 fa->fa_tos < tos)
359 return fa;
360 }
361 }
362 return NULL;
363}
364
365int fib_detect_death(struct fib_info *fi, int order,
Denis V. Lunevc17860a2007-12-08 00:22:13 -0800366 struct fib_info **last_resort, int *last_idx, int dflt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367{
368 struct neighbour *n;
369 int state = NUD_NONE;
370
371 n = neigh_lookup(&arp_tbl, &fi->fib_nh[0].nh_gw, fi->fib_dev);
372 if (n) {
373 state = n->nud_state;
374 neigh_release(n);
375 }
Jianjun Kongd9319102008-11-03 00:23:42 -0800376 if (state == NUD_REACHABLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 return 0;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000378 if ((state & NUD_VALID) && order != dflt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 return 0;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000380 if ((state & NUD_VALID) ||
381 (*last_idx < 0 && order > dflt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 *last_resort = fi;
383 *last_idx = order;
384 }
385 return 1;
386}
387
388#ifdef CONFIG_IP_ROUTE_MULTIPATH
389
Thomas Graf4e902c52006-08-17 18:14:52 -0700390static int fib_count_nexthops(struct rtnexthop *rtnh, int remaining)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391{
392 int nhs = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Thomas Graf4e902c52006-08-17 18:14:52 -0700394 while (rtnh_ok(rtnh, remaining)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 nhs++;
Thomas Graf4e902c52006-08-17 18:14:52 -0700396 rtnh = rtnh_next(rtnh, &remaining);
397 }
398
399 /* leftover implies invalid nexthop configuration, discard it */
400 return remaining > 0 ? 0 : nhs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401}
402
Thomas Graf4e902c52006-08-17 18:14:52 -0700403static int fib_get_nhs(struct fib_info *fi, struct rtnexthop *rtnh,
404 int remaining, struct fib_config *cfg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 change_nexthops(fi) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700407 int attrlen;
408
409 if (!rtnh_ok(rtnh, remaining))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 return -EINVAL;
Thomas Graf4e902c52006-08-17 18:14:52 -0700411
David S. Miller71fceff2010-01-15 01:16:40 -0800412 nexthop_nh->nh_flags =
413 (cfg->fc_flags & ~0xFF) | rtnh->rtnh_flags;
414 nexthop_nh->nh_oif = rtnh->rtnh_ifindex;
415 nexthop_nh->nh_weight = rtnh->rtnh_hops + 1;
Thomas Graf4e902c52006-08-17 18:14:52 -0700416
417 attrlen = rtnh_attrlen(rtnh);
418 if (attrlen > 0) {
419 struct nlattr *nla, *attrs = rtnh_attrs(rtnh);
420
421 nla = nla_find(attrs, attrlen, RTA_GATEWAY);
David S. Miller71fceff2010-01-15 01:16:40 -0800422 nexthop_nh->nh_gw = nla ? nla_get_be32(nla) : 0;
Patrick McHardyc7066f72011-01-14 13:36:42 +0100423#ifdef CONFIG_IP_ROUTE_CLASSID
Thomas Graf4e902c52006-08-17 18:14:52 -0700424 nla = nla_find(attrs, attrlen, RTA_FLOW);
David S. Miller71fceff2010-01-15 01:16:40 -0800425 nexthop_nh->nh_tclassid = nla ? nla_get_u32(nla) : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426#endif
427 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700428
429 rtnh = rtnh_next(rtnh, &remaining);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 } endfor_nexthops(fi);
Thomas Graf4e902c52006-08-17 18:14:52 -0700431
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 return 0;
433}
434
435#endif
436
Thomas Graf4e902c52006-08-17 18:14:52 -0700437int fib_nh_match(struct fib_config *cfg, struct fib_info *fi)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438{
439#ifdef CONFIG_IP_ROUTE_MULTIPATH
Thomas Graf4e902c52006-08-17 18:14:52 -0700440 struct rtnexthop *rtnh;
441 int remaining;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442#endif
443
Thomas Graf4e902c52006-08-17 18:14:52 -0700444 if (cfg->fc_priority && cfg->fc_priority != fi->fib_priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 return 1;
446
Thomas Graf4e902c52006-08-17 18:14:52 -0700447 if (cfg->fc_oif || cfg->fc_gw) {
448 if ((!cfg->fc_oif || cfg->fc_oif == fi->fib_nh->nh_oif) &&
449 (!cfg->fc_gw || cfg->fc_gw == fi->fib_nh->nh_gw))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 return 0;
451 return 1;
452 }
453
454#ifdef CONFIG_IP_ROUTE_MULTIPATH
Thomas Graf4e902c52006-08-17 18:14:52 -0700455 if (cfg->fc_mp == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 return 0;
Thomas Graf4e902c52006-08-17 18:14:52 -0700457
458 rtnh = cfg->fc_mp;
459 remaining = cfg->fc_mp_len;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900460
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 for_nexthops(fi) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700462 int attrlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
Thomas Graf4e902c52006-08-17 18:14:52 -0700464 if (!rtnh_ok(rtnh, remaining))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 return -EINVAL;
Thomas Graf4e902c52006-08-17 18:14:52 -0700466
467 if (rtnh->rtnh_ifindex && rtnh->rtnh_ifindex != nh->nh_oif)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 return 1;
Thomas Graf4e902c52006-08-17 18:14:52 -0700469
470 attrlen = rtnh_attrlen(rtnh);
471 if (attrlen < 0) {
472 struct nlattr *nla, *attrs = rtnh_attrs(rtnh);
473
474 nla = nla_find(attrs, attrlen, RTA_GATEWAY);
Al Viro17fb2c62006-09-26 22:15:25 -0700475 if (nla && nla_get_be32(nla) != nh->nh_gw)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 return 1;
Patrick McHardyc7066f72011-01-14 13:36:42 +0100477#ifdef CONFIG_IP_ROUTE_CLASSID
Thomas Graf4e902c52006-08-17 18:14:52 -0700478 nla = nla_find(attrs, attrlen, RTA_FLOW);
479 if (nla && nla_get_u32(nla) != nh->nh_tclassid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 return 1;
481#endif
482 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700483
484 rtnh = rtnh_next(rtnh, &remaining);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 } endfor_nexthops(fi);
486#endif
487 return 0;
488}
489
490
491/*
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000492 * Picture
493 * -------
494 *
495 * Semantics of nexthop is very messy by historical reasons.
496 * We have to take into account, that:
497 * a) gateway can be actually local interface address,
498 * so that gatewayed route is direct.
499 * b) gateway must be on-link address, possibly
500 * described not by an ifaddr, but also by a direct route.
501 * c) If both gateway and interface are specified, they should not
502 * contradict.
503 * d) If we use tunnel routes, gateway could be not on-link.
504 *
505 * Attempt to reconcile all of these (alas, self-contradictory) conditions
506 * results in pretty ugly and hairy code with obscure logic.
507 *
508 * I chose to generalized it instead, so that the size
509 * of code does not increase practically, but it becomes
510 * much more general.
511 * Every prefix is assigned a "scope" value: "host" is local address,
512 * "link" is direct route,
513 * [ ... "site" ... "interior" ... ]
514 * and "universe" is true gateway route with global meaning.
515 *
516 * Every prefix refers to a set of "nexthop"s (gw, oif),
517 * where gw must have narrower scope. This recursion stops
518 * when gw has LOCAL scope or if "nexthop" is declared ONLINK,
519 * which means that gw is forced to be on link.
520 *
521 * Code is still hairy, but now it is apparently logically
522 * consistent and very flexible. F.e. as by-product it allows
523 * to co-exists in peace independent exterior and interior
524 * routing processes.
525 *
526 * Normally it looks as following.
527 *
528 * {universe prefix} -> (gw, oif) [scope link]
529 * |
530 * |-> {link prefix} -> (gw, oif) [scope local]
531 * |
532 * |-> {local prefix} (terminal node)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 */
Thomas Graf4e902c52006-08-17 18:14:52 -0700534static int fib_check_nh(struct fib_config *cfg, struct fib_info *fi,
535 struct fib_nh *nh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536{
537 int err;
Denis V. Lunev86167a32008-01-21 17:34:00 -0800538 struct net *net;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000539 struct net_device *dev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540
Denis V. Lunev86167a32008-01-21 17:34:00 -0800541 net = cfg->fc_nlinfo.nl_net;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 if (nh->nh_gw) {
543 struct fib_result res;
544
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000545 if (nh->nh_flags & RTNH_F_ONLINK) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546
Thomas Graf4e902c52006-08-17 18:14:52 -0700547 if (cfg->fc_scope >= RT_SCOPE_LINK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 return -EINVAL;
Denis V. Lunev86167a32008-01-21 17:34:00 -0800549 if (inet_addr_type(net, nh->nh_gw) != RTN_UNICAST)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 return -EINVAL;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000551 dev = __dev_get_by_index(net, nh->nh_oif);
552 if (!dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 return -ENODEV;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000554 if (!(dev->flags & IFF_UP))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 return -ENETDOWN;
556 nh->nh_dev = dev;
557 dev_hold(dev);
558 nh->nh_scope = RT_SCOPE_LINK;
559 return 0;
560 }
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000561 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 {
David S. Miller9ade2282011-03-12 02:02:42 -0500563 struct flowi4 fl4 = {
564 .daddr = nh->nh_gw,
565 .flowi4_scope = cfg->fc_scope + 1,
566 .flowi4_oif = nh->nh_oif,
Thomas Graf4e902c52006-08-17 18:14:52 -0700567 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568
569 /* It is not necessary, but requires a bit of thinking */
David S. Miller9ade2282011-03-12 02:02:42 -0500570 if (fl4.flowi4_scope < RT_SCOPE_LINK)
571 fl4.flowi4_scope = RT_SCOPE_LINK;
572 err = fib_lookup(net, &fl4, &res);
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000573 if (err) {
574 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 return err;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +0000576 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 }
578 err = -EINVAL;
579 if (res.type != RTN_UNICAST && res.type != RTN_LOCAL)
580 goto out;
581 nh->nh_scope = res.scope;
582 nh->nh_oif = FIB_RES_OIF(res);
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000583 nh->nh_dev = dev = FIB_RES_DEV(res);
584 if (!dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 goto out;
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000586 dev_hold(dev);
Eric Dumazet8723e1b2010-10-19 00:39:26 +0000587 err = (dev->flags & IFF_UP) ? 0 : -ENETDOWN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 } else {
589 struct in_device *in_dev;
590
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000591 if (nh->nh_flags & (RTNH_F_PERVASIVE | RTNH_F_ONLINK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 return -EINVAL;
593
Eric Dumazet8723e1b2010-10-19 00:39:26 +0000594 rcu_read_lock();
595 err = -ENODEV;
Denis V. Lunev86167a32008-01-21 17:34:00 -0800596 in_dev = inetdev_by_index(net, nh->nh_oif);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 if (in_dev == NULL)
Eric Dumazet8723e1b2010-10-19 00:39:26 +0000598 goto out;
599 err = -ENETDOWN;
600 if (!(in_dev->dev->flags & IFF_UP))
601 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 nh->nh_dev = in_dev->dev;
603 dev_hold(nh->nh_dev);
604 nh->nh_scope = RT_SCOPE_HOST;
Eric Dumazet8723e1b2010-10-19 00:39:26 +0000605 err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 }
Eric Dumazet8723e1b2010-10-19 00:39:26 +0000607out:
608 rcu_read_unlock();
609 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610}
611
Al Viro81f7bf62006-09-27 18:40:00 -0700612static inline unsigned int fib_laddr_hashfn(__be32 val)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613{
David S. Miller123b9732011-02-01 15:34:21 -0800614 unsigned int mask = (fib_info_hash_size - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000616 return ((__force u32)val ^
617 ((__force u32)val >> 7) ^
618 ((__force u32)val >> 14)) & mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619}
620
David S. Miller123b9732011-02-01 15:34:21 -0800621static struct hlist_head *fib_info_hash_alloc(int bytes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622{
623 if (bytes <= PAGE_SIZE)
Joonwoo Park88f83492007-11-26 23:29:32 +0800624 return kzalloc(bytes, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625 else
626 return (struct hlist_head *)
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000627 __get_free_pages(GFP_KERNEL | __GFP_ZERO,
628 get_order(bytes));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629}
630
David S. Miller123b9732011-02-01 15:34:21 -0800631static void fib_info_hash_free(struct hlist_head *hash, int bytes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632{
633 if (!hash)
634 return;
635
636 if (bytes <= PAGE_SIZE)
637 kfree(hash);
638 else
639 free_pages((unsigned long) hash, get_order(bytes));
640}
641
David S. Miller123b9732011-02-01 15:34:21 -0800642static void fib_info_hash_move(struct hlist_head *new_info_hash,
643 struct hlist_head *new_laddrhash,
644 unsigned int new_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
David S. Millerb7656e72005-08-05 04:12:48 -0700646 struct hlist_head *old_info_hash, *old_laddrhash;
David S. Miller123b9732011-02-01 15:34:21 -0800647 unsigned int old_size = fib_info_hash_size;
David S. Millerb7656e72005-08-05 04:12:48 -0700648 unsigned int i, bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700650 spin_lock_bh(&fib_info_lock);
David S. Millerb7656e72005-08-05 04:12:48 -0700651 old_info_hash = fib_info_hash;
652 old_laddrhash = fib_info_laddrhash;
David S. Miller123b9732011-02-01 15:34:21 -0800653 fib_info_hash_size = new_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
655 for (i = 0; i < old_size; i++) {
656 struct hlist_head *head = &fib_info_hash[i];
657 struct hlist_node *node, *n;
658 struct fib_info *fi;
659
660 hlist_for_each_entry_safe(fi, node, n, head, fib_hash) {
661 struct hlist_head *dest;
662 unsigned int new_hash;
663
664 hlist_del(&fi->fib_hash);
665
666 new_hash = fib_info_hashfn(fi);
667 dest = &new_info_hash[new_hash];
668 hlist_add_head(&fi->fib_hash, dest);
669 }
670 }
671 fib_info_hash = new_info_hash;
672
673 for (i = 0; i < old_size; i++) {
674 struct hlist_head *lhead = &fib_info_laddrhash[i];
675 struct hlist_node *node, *n;
676 struct fib_info *fi;
677
678 hlist_for_each_entry_safe(fi, node, n, lhead, fib_lhash) {
679 struct hlist_head *ldest;
680 unsigned int new_hash;
681
682 hlist_del(&fi->fib_lhash);
683
684 new_hash = fib_laddr_hashfn(fi->fib_prefsrc);
685 ldest = &new_laddrhash[new_hash];
686 hlist_add_head(&fi->fib_lhash, ldest);
687 }
688 }
689 fib_info_laddrhash = new_laddrhash;
690
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700691 spin_unlock_bh(&fib_info_lock);
David S. Millerb7656e72005-08-05 04:12:48 -0700692
693 bytes = old_size * sizeof(struct hlist_head *);
David S. Miller123b9732011-02-01 15:34:21 -0800694 fib_info_hash_free(old_info_hash, bytes);
695 fib_info_hash_free(old_laddrhash, bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696}
697
Thomas Graf4e902c52006-08-17 18:14:52 -0700698struct fib_info *fib_create_info(struct fib_config *cfg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699{
700 int err;
701 struct fib_info *fi = NULL;
702 struct fib_info *ofi;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 int nhs = 1;
Denis V. Lunev7462bd742008-01-31 18:49:32 -0800704 struct net *net = cfg->fc_nlinfo.nl_net;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
David S. Miller4c8237c2011-03-07 14:27:38 -0800706 if (cfg->fc_type > RTN_MAX)
707 goto err_inval;
708
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 /* Fast check to catch the most weird cases */
Thomas Graf4e902c52006-08-17 18:14:52 -0700710 if (fib_props[cfg->fc_type].scope > cfg->fc_scope)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 goto err_inval;
712
713#ifdef CONFIG_IP_ROUTE_MULTIPATH
Thomas Graf4e902c52006-08-17 18:14:52 -0700714 if (cfg->fc_mp) {
715 nhs = fib_count_nexthops(cfg->fc_mp, cfg->fc_mp_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 if (nhs == 0)
717 goto err_inval;
718 }
719#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720
721 err = -ENOBUFS;
David S. Miller123b9732011-02-01 15:34:21 -0800722 if (fib_info_cnt >= fib_info_hash_size) {
723 unsigned int new_size = fib_info_hash_size << 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 struct hlist_head *new_info_hash;
725 struct hlist_head *new_laddrhash;
726 unsigned int bytes;
727
728 if (!new_size)
729 new_size = 1;
730 bytes = new_size * sizeof(struct hlist_head *);
David S. Miller123b9732011-02-01 15:34:21 -0800731 new_info_hash = fib_info_hash_alloc(bytes);
732 new_laddrhash = fib_info_hash_alloc(bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 if (!new_info_hash || !new_laddrhash) {
David S. Miller123b9732011-02-01 15:34:21 -0800734 fib_info_hash_free(new_info_hash, bytes);
735 fib_info_hash_free(new_laddrhash, bytes);
Joonwoo Park88f83492007-11-26 23:29:32 +0800736 } else
David S. Miller123b9732011-02-01 15:34:21 -0800737 fib_info_hash_move(new_info_hash, new_laddrhash, new_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738
David S. Miller123b9732011-02-01 15:34:21 -0800739 if (!fib_info_hash_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 goto failure;
741 }
742
Panagiotis Issaris0da974f2006-07-21 14:51:30 -0700743 fi = kzalloc(sizeof(*fi)+nhs*sizeof(struct fib_nh), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 if (fi == NULL)
745 goto failure;
David S. Miller725d1e12011-01-28 14:05:05 -0800746 if (cfg->fc_mx) {
747 fi->fib_metrics = kzalloc(sizeof(u32) * RTAX_MAX, GFP_KERNEL);
748 if (!fi->fib_metrics)
749 goto failure;
750 } else
751 fi->fib_metrics = (u32 *) dst_default_metrics;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 fib_info_cnt++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753
Denis V. Lunev57d7a602008-04-16 02:00:50 -0700754 fi->fib_net = hold_net(net);
Thomas Graf4e902c52006-08-17 18:14:52 -0700755 fi->fib_protocol = cfg->fc_protocol;
756 fi->fib_flags = cfg->fc_flags;
757 fi->fib_priority = cfg->fc_priority;
758 fi->fib_prefsrc = cfg->fc_prefsrc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
760 fi->fib_nhs = nhs;
761 change_nexthops(fi) {
David S. Miller71fceff2010-01-15 01:16:40 -0800762 nexthop_nh->nh_parent = fi;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 } endfor_nexthops(fi)
764
Thomas Graf4e902c52006-08-17 18:14:52 -0700765 if (cfg->fc_mx) {
766 struct nlattr *nla;
767 int remaining;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768
Thomas Graf4e902c52006-08-17 18:14:52 -0700769 nla_for_each_attr(nla, cfg->fc_mx, cfg->fc_mx_len, remaining) {
Thomas Graf8f4c1f92007-09-12 14:44:36 +0200770 int type = nla_type(nla);
Thomas Graf4e902c52006-08-17 18:14:52 -0700771
772 if (type) {
773 if (type > RTAX_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 goto err_inval;
Thomas Graf4e902c52006-08-17 18:14:52 -0700775 fi->fib_metrics[type - 1] = nla_get_u32(nla);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 }
778 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779
Thomas Graf4e902c52006-08-17 18:14:52 -0700780 if (cfg->fc_mp) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781#ifdef CONFIG_IP_ROUTE_MULTIPATH
Thomas Graf4e902c52006-08-17 18:14:52 -0700782 err = fib_get_nhs(fi, cfg->fc_mp, cfg->fc_mp_len, cfg);
783 if (err != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 goto failure;
Thomas Graf4e902c52006-08-17 18:14:52 -0700785 if (cfg->fc_oif && fi->fib_nh->nh_oif != cfg->fc_oif)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 goto err_inval;
Thomas Graf4e902c52006-08-17 18:14:52 -0700787 if (cfg->fc_gw && fi->fib_nh->nh_gw != cfg->fc_gw)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 goto err_inval;
Patrick McHardyc7066f72011-01-14 13:36:42 +0100789#ifdef CONFIG_IP_ROUTE_CLASSID
Thomas Graf4e902c52006-08-17 18:14:52 -0700790 if (cfg->fc_flow && fi->fib_nh->nh_tclassid != cfg->fc_flow)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 goto err_inval;
792#endif
793#else
794 goto err_inval;
795#endif
796 } else {
797 struct fib_nh *nh = fi->fib_nh;
Thomas Graf4e902c52006-08-17 18:14:52 -0700798
799 nh->nh_oif = cfg->fc_oif;
800 nh->nh_gw = cfg->fc_gw;
801 nh->nh_flags = cfg->fc_flags;
Patrick McHardyc7066f72011-01-14 13:36:42 +0100802#ifdef CONFIG_IP_ROUTE_CLASSID
Thomas Graf4e902c52006-08-17 18:14:52 -0700803 nh->nh_tclassid = cfg->fc_flow;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805#ifdef CONFIG_IP_ROUTE_MULTIPATH
806 nh->nh_weight = 1;
807#endif
808 }
809
Thomas Graf4e902c52006-08-17 18:14:52 -0700810 if (fib_props[cfg->fc_type].error) {
811 if (cfg->fc_gw || cfg->fc_oif || cfg->fc_mp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 goto err_inval;
813 goto link_it;
David S. Miller4c8237c2011-03-07 14:27:38 -0800814 } else {
815 switch (cfg->fc_type) {
816 case RTN_UNICAST:
817 case RTN_LOCAL:
818 case RTN_BROADCAST:
819 case RTN_ANYCAST:
820 case RTN_MULTICAST:
821 break;
822 default:
823 goto err_inval;
824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 }
826
Thomas Graf4e902c52006-08-17 18:14:52 -0700827 if (cfg->fc_scope > RT_SCOPE_HOST)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 goto err_inval;
829
Thomas Graf4e902c52006-08-17 18:14:52 -0700830 if (cfg->fc_scope == RT_SCOPE_HOST) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 struct fib_nh *nh = fi->fib_nh;
832
833 /* Local address is added. */
834 if (nhs != 1 || nh->nh_gw)
835 goto err_inval;
836 nh->nh_scope = RT_SCOPE_NOWHERE;
Denis V. Lunev7462bd742008-01-31 18:49:32 -0800837 nh->nh_dev = dev_get_by_index(net, fi->fib_nh->nh_oif);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 err = -ENODEV;
839 if (nh->nh_dev == NULL)
840 goto failure;
841 } else {
842 change_nexthops(fi) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000843 err = fib_check_nh(cfg, fi, nexthop_nh);
844 if (err != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 goto failure;
846 } endfor_nexthops(fi)
847 }
848
849 if (fi->fib_prefsrc) {
Thomas Graf4e902c52006-08-17 18:14:52 -0700850 if (cfg->fc_type != RTN_LOCAL || !cfg->fc_dst ||
851 fi->fib_prefsrc != cfg->fc_dst)
Denis V. Lunev7462bd742008-01-31 18:49:32 -0800852 if (inet_addr_type(net, fi->fib_prefsrc) != RTN_LOCAL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 goto err_inval;
854 }
855
David S. Miller1fc050a2011-03-07 20:54:48 -0800856 change_nexthops(fi) {
David S. Millera7ac8fc2011-03-08 11:03:21 -0800857 nexthop_nh->nh_cfg_scope = cfg->fc_scope;
David S. Miller1fc050a2011-03-07 20:54:48 -0800858 nexthop_nh->nh_saddr = inet_select_addr(nexthop_nh->nh_dev,
859 nexthop_nh->nh_gw,
David S. Millera7ac8fc2011-03-08 11:03:21 -0800860 nexthop_nh->nh_cfg_scope);
David S. Miller1fc050a2011-03-07 20:54:48 -0800861 } endfor_nexthops(fi)
862
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863link_it:
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000864 ofi = fib_find_info(fi);
865 if (ofi) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 fi->fib_dead = 1;
867 free_fib_info(fi);
868 ofi->fib_treeref++;
869 return ofi;
870 }
871
872 fi->fib_treeref++;
873 atomic_inc(&fi->fib_clntref);
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700874 spin_lock_bh(&fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875 hlist_add_head(&fi->fib_hash,
876 &fib_info_hash[fib_info_hashfn(fi)]);
877 if (fi->fib_prefsrc) {
878 struct hlist_head *head;
879
880 head = &fib_info_laddrhash[fib_laddr_hashfn(fi->fib_prefsrc)];
881 hlist_add_head(&fi->fib_lhash, head);
882 }
883 change_nexthops(fi) {
884 struct hlist_head *head;
885 unsigned int hash;
886
David S. Miller71fceff2010-01-15 01:16:40 -0800887 if (!nexthop_nh->nh_dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 continue;
David S. Miller71fceff2010-01-15 01:16:40 -0800889 hash = fib_devindex_hashfn(nexthop_nh->nh_dev->ifindex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 head = &fib_info_devhash[hash];
David S. Miller71fceff2010-01-15 01:16:40 -0800891 hlist_add_head(&nexthop_nh->nh_hash, head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 } endfor_nexthops(fi)
Stephen Hemminger832b4c52006-08-29 16:48:09 -0700893 spin_unlock_bh(&fib_info_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 return fi;
895
896err_inval:
897 err = -EINVAL;
898
899failure:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900900 if (fi) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 fi->fib_dead = 1;
902 free_fib_info(fi);
903 }
Thomas Graf4e902c52006-08-17 18:14:52 -0700904
905 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906}
907
Thomas Grafbe403ea2006-08-17 18:15:17 -0700908int fib_dump_info(struct sk_buff *skb, u32 pid, u32 seq, int event,
Al Viro81f7bf62006-09-27 18:40:00 -0700909 u32 tb_id, u8 type, u8 scope, __be32 dst, int dst_len, u8 tos,
Thomas Grafbe403ea2006-08-17 18:15:17 -0700910 struct fib_info *fi, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911{
Thomas Grafbe403ea2006-08-17 18:15:17 -0700912 struct nlmsghdr *nlh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 struct rtmsg *rtm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914
Thomas Grafbe403ea2006-08-17 18:15:17 -0700915 nlh = nlmsg_put(skb, pid, seq, event, sizeof(*rtm), flags);
916 if (nlh == NULL)
Patrick McHardy26932562007-01-31 23:16:40 -0800917 return -EMSGSIZE;
Thomas Grafbe403ea2006-08-17 18:15:17 -0700918
919 rtm = nlmsg_data(nlh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 rtm->rtm_family = AF_INET;
921 rtm->rtm_dst_len = dst_len;
922 rtm->rtm_src_len = 0;
923 rtm->rtm_tos = tos;
Krzysztof Piotr Oledzki709772e2008-06-10 15:44:49 -0700924 if (tb_id < 256)
925 rtm->rtm_table = tb_id;
926 else
927 rtm->rtm_table = RT_TABLE_COMPAT;
Thomas Grafbe403ea2006-08-17 18:15:17 -0700928 NLA_PUT_U32(skb, RTA_TABLE, tb_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 rtm->rtm_type = type;
930 rtm->rtm_flags = fi->fib_flags;
931 rtm->rtm_scope = scope;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 rtm->rtm_protocol = fi->fib_protocol;
Thomas Grafbe403ea2006-08-17 18:15:17 -0700933
934 if (rtm->rtm_dst_len)
Al Viro17fb2c62006-09-26 22:15:25 -0700935 NLA_PUT_BE32(skb, RTA_DST, dst);
Thomas Grafbe403ea2006-08-17 18:15:17 -0700936
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 if (fi->fib_priority)
Thomas Grafbe403ea2006-08-17 18:15:17 -0700938 NLA_PUT_U32(skb, RTA_PRIORITY, fi->fib_priority);
939
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 if (rtnetlink_put_metrics(skb, fi->fib_metrics) < 0)
Thomas Grafbe403ea2006-08-17 18:15:17 -0700941 goto nla_put_failure;
942
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (fi->fib_prefsrc)
Al Viro17fb2c62006-09-26 22:15:25 -0700944 NLA_PUT_BE32(skb, RTA_PREFSRC, fi->fib_prefsrc);
Thomas Grafbe403ea2006-08-17 18:15:17 -0700945
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 if (fi->fib_nhs == 1) {
947 if (fi->fib_nh->nh_gw)
Al Viro17fb2c62006-09-26 22:15:25 -0700948 NLA_PUT_BE32(skb, RTA_GATEWAY, fi->fib_nh->nh_gw);
Thomas Grafbe403ea2006-08-17 18:15:17 -0700949
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 if (fi->fib_nh->nh_oif)
Thomas Grafbe403ea2006-08-17 18:15:17 -0700951 NLA_PUT_U32(skb, RTA_OIF, fi->fib_nh->nh_oif);
Patrick McHardyc7066f72011-01-14 13:36:42 +0100952#ifdef CONFIG_IP_ROUTE_CLASSID
Patrick McHardy8265abc2006-07-21 15:09:55 -0700953 if (fi->fib_nh[0].nh_tclassid)
Thomas Grafbe403ea2006-08-17 18:15:17 -0700954 NLA_PUT_U32(skb, RTA_FLOW, fi->fib_nh[0].nh_tclassid);
Patrick McHardy8265abc2006-07-21 15:09:55 -0700955#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 }
957#ifdef CONFIG_IP_ROUTE_MULTIPATH
958 if (fi->fib_nhs > 1) {
Thomas Grafbe403ea2006-08-17 18:15:17 -0700959 struct rtnexthop *rtnh;
960 struct nlattr *mp;
961
962 mp = nla_nest_start(skb, RTA_MULTIPATH);
963 if (mp == NULL)
964 goto nla_put_failure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
966 for_nexthops(fi) {
Thomas Grafbe403ea2006-08-17 18:15:17 -0700967 rtnh = nla_reserve_nohdr(skb, sizeof(*rtnh));
968 if (rtnh == NULL)
969 goto nla_put_failure;
970
971 rtnh->rtnh_flags = nh->nh_flags & 0xFF;
972 rtnh->rtnh_hops = nh->nh_weight - 1;
973 rtnh->rtnh_ifindex = nh->nh_oif;
974
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 if (nh->nh_gw)
Al Viro17fb2c62006-09-26 22:15:25 -0700976 NLA_PUT_BE32(skb, RTA_GATEWAY, nh->nh_gw);
Patrick McHardyc7066f72011-01-14 13:36:42 +0100977#ifdef CONFIG_IP_ROUTE_CLASSID
Patrick McHardy8265abc2006-07-21 15:09:55 -0700978 if (nh->nh_tclassid)
Thomas Grafbe403ea2006-08-17 18:15:17 -0700979 NLA_PUT_U32(skb, RTA_FLOW, nh->nh_tclassid);
Patrick McHardy8265abc2006-07-21 15:09:55 -0700980#endif
Thomas Grafbe403ea2006-08-17 18:15:17 -0700981 /* length of rtnetlink header + attributes */
982 rtnh->rtnh_len = nlmsg_get_pos(skb) - (void *) rtnh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 } endfor_nexthops(fi);
Thomas Grafbe403ea2006-08-17 18:15:17 -0700984
985 nla_nest_end(skb, mp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 }
987#endif
Thomas Grafbe403ea2006-08-17 18:15:17 -0700988 return nlmsg_end(skb, nlh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989
Thomas Grafbe403ea2006-08-17 18:15:17 -0700990nla_put_failure:
Patrick McHardy26932562007-01-31 23:16:40 -0800991 nlmsg_cancel(skb, nlh);
992 return -EMSGSIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993}
994
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995/*
Eric Dumazet6a31d2a2010-10-04 20:00:18 +0000996 * Update FIB if:
997 * - local address disappeared -> we must delete all the entries
998 * referring to it.
999 * - device went down -> we must shutdown all nexthops going via it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 */
Denis V. Lunev4814bdb2008-01-31 18:50:07 -08001001int fib_sync_down_addr(struct net *net, __be32 local)
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001002{
1003 int ret = 0;
1004 unsigned int hash = fib_laddr_hashfn(local);
1005 struct hlist_head *head = &fib_info_laddrhash[hash];
1006 struct hlist_node *node;
1007 struct fib_info *fi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001009 if (fib_info_laddrhash == NULL || local == 0)
1010 return 0;
1011
1012 hlist_for_each_entry(fi, node, head, fib_lhash) {
Octavian Purdila09ad9bc2009-11-25 15:14:13 -08001013 if (!net_eq(fi->fib_net, net))
Denis V. Lunev4814bdb2008-01-31 18:50:07 -08001014 continue;
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001015 if (fi->fib_prefsrc == local) {
1016 fi->fib_flags |= RTNH_F_DEAD;
1017 ret++;
1018 }
1019 }
1020 return ret;
1021}
1022
1023int fib_sync_down_dev(struct net_device *dev, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024{
1025 int ret = 0;
1026 int scope = RT_SCOPE_NOWHERE;
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001027 struct fib_info *prev_fi = NULL;
1028 unsigned int hash = fib_devindex_hashfn(dev->ifindex);
1029 struct hlist_head *head = &fib_info_devhash[hash];
1030 struct hlist_node *node;
1031 struct fib_nh *nh;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001032
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 if (force)
1034 scope = -1;
1035
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001036 hlist_for_each_entry(nh, node, head, nh_hash) {
1037 struct fib_info *fi = nh->nh_parent;
1038 int dead;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001040 BUG_ON(!fi->fib_nhs);
1041 if (nh->nh_dev != dev || fi == prev_fi)
1042 continue;
1043 prev_fi = fi;
1044 dead = 0;
1045 change_nexthops(fi) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001046 if (nexthop_nh->nh_flags & RTNH_F_DEAD)
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001047 dead++;
David S. Miller71fceff2010-01-15 01:16:40 -08001048 else if (nexthop_nh->nh_dev == dev &&
1049 nexthop_nh->nh_scope != scope) {
1050 nexthop_nh->nh_flags |= RTNH_F_DEAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051#ifdef CONFIG_IP_ROUTE_MULTIPATH
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001052 spin_lock_bh(&fib_multipath_lock);
David S. Miller71fceff2010-01-15 01:16:40 -08001053 fi->fib_power -= nexthop_nh->nh_power;
1054 nexthop_nh->nh_power = 0;
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001055 spin_unlock_bh(&fib_multipath_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056#endif
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001057 dead++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058 }
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001059#ifdef CONFIG_IP_ROUTE_MULTIPATH
David S. Miller71fceff2010-01-15 01:16:40 -08001060 if (force > 1 && nexthop_nh->nh_dev == dev) {
Denis V. Lunev85326fa2008-01-31 18:48:47 -08001061 dead = fi->fib_nhs;
1062 break;
1063 }
1064#endif
1065 } endfor_nexthops(fi)
1066 if (dead == fi->fib_nhs) {
1067 fi->fib_flags |= RTNH_F_DEAD;
1068 ret++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 }
1070 }
1071
1072 return ret;
1073}
1074
David S. Miller0c838ff2011-01-31 16:16:50 -08001075/* Must be invoked inside of an RCU protected region. */
1076void fib_select_default(struct fib_result *res)
1077{
1078 struct fib_info *fi = NULL, *last_resort = NULL;
1079 struct list_head *fa_head = res->fa_head;
1080 struct fib_table *tb = res->table;
1081 int order = -1, last_idx = -1;
1082 struct fib_alias *fa;
1083
1084 list_for_each_entry_rcu(fa, fa_head, fa_list) {
1085 struct fib_info *next_fi = fa->fa_info;
1086
1087 if (fa->fa_scope != res->scope ||
1088 fa->fa_type != RTN_UNICAST)
1089 continue;
1090
1091 if (next_fi->fib_priority > res->fi->fib_priority)
1092 break;
1093 if (!next_fi->fib_nh[0].nh_gw ||
1094 next_fi->fib_nh[0].nh_scope != RT_SCOPE_LINK)
1095 continue;
1096
1097 fib_alias_accessed(fa);
1098
1099 if (fi == NULL) {
1100 if (next_fi != res->fi)
1101 break;
1102 } else if (!fib_detect_death(fi, order, &last_resort,
1103 &last_idx, tb->tb_default)) {
1104 fib_result_assign(res, fi);
1105 tb->tb_default = order;
1106 goto out;
1107 }
1108 fi = next_fi;
1109 order++;
1110 }
1111
1112 if (order <= 0 || fi == NULL) {
1113 tb->tb_default = -1;
1114 goto out;
1115 }
1116
1117 if (!fib_detect_death(fi, order, &last_resort, &last_idx,
1118 tb->tb_default)) {
1119 fib_result_assign(res, fi);
1120 tb->tb_default = order;
1121 goto out;
1122 }
1123
1124 if (last_idx >= 0)
1125 fib_result_assign(res, last_resort);
1126 tb->tb_default = last_idx;
1127out:
Eric Dumazet31d40932011-02-14 11:23:04 -08001128 return;
David S. Miller0c838ff2011-01-31 16:16:50 -08001129}
1130
David S. Miller1fc050a2011-03-07 20:54:48 -08001131void fib_update_nh_saddrs(struct net_device *dev)
1132{
1133 struct hlist_head *head;
1134 struct hlist_node *node;
1135 struct fib_nh *nh;
1136 unsigned int hash;
1137
1138 hash = fib_devindex_hashfn(dev->ifindex);
1139 head = &fib_info_devhash[hash];
1140 hlist_for_each_entry(nh, node, head, nh_hash) {
1141 if (nh->nh_dev != dev)
1142 continue;
1143 nh->nh_saddr = inet_select_addr(nh->nh_dev,
1144 nh->nh_gw,
David S. Millera7ac8fc2011-03-08 11:03:21 -08001145 nh->nh_cfg_scope);
David S. Miller1fc050a2011-03-07 20:54:48 -08001146 }
1147}
1148
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149#ifdef CONFIG_IP_ROUTE_MULTIPATH
1150
1151/*
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001152 * Dead device goes up. We wake up dead nexthops.
1153 * It takes sense only on multipath routes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155int fib_sync_up(struct net_device *dev)
1156{
1157 struct fib_info *prev_fi;
1158 unsigned int hash;
1159 struct hlist_head *head;
1160 struct hlist_node *node;
1161 struct fib_nh *nh;
1162 int ret;
1163
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001164 if (!(dev->flags & IFF_UP))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 return 0;
1166
1167 prev_fi = NULL;
1168 hash = fib_devindex_hashfn(dev->ifindex);
1169 head = &fib_info_devhash[hash];
1170 ret = 0;
1171
1172 hlist_for_each_entry(nh, node, head, nh_hash) {
1173 struct fib_info *fi = nh->nh_parent;
1174 int alive;
1175
1176 BUG_ON(!fi->fib_nhs);
1177 if (nh->nh_dev != dev || fi == prev_fi)
1178 continue;
1179
1180 prev_fi = fi;
1181 alive = 0;
1182 change_nexthops(fi) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001183 if (!(nexthop_nh->nh_flags & RTNH_F_DEAD)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 alive++;
1185 continue;
1186 }
David S. Miller71fceff2010-01-15 01:16:40 -08001187 if (nexthop_nh->nh_dev == NULL ||
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001188 !(nexthop_nh->nh_dev->flags & IFF_UP))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 continue;
David S. Miller71fceff2010-01-15 01:16:40 -08001190 if (nexthop_nh->nh_dev != dev ||
1191 !__in_dev_get_rtnl(dev))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 continue;
1193 alive++;
1194 spin_lock_bh(&fib_multipath_lock);
David S. Miller71fceff2010-01-15 01:16:40 -08001195 nexthop_nh->nh_power = 0;
1196 nexthop_nh->nh_flags &= ~RTNH_F_DEAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 spin_unlock_bh(&fib_multipath_lock);
1198 } endfor_nexthops(fi)
1199
1200 if (alive > 0) {
1201 fi->fib_flags &= ~RTNH_F_DEAD;
1202 ret++;
1203 }
1204 }
1205
1206 return ret;
1207}
1208
1209/*
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001210 * The algorithm is suboptimal, but it provides really
1211 * fair weighted route distribution.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 */
David S. Miller1b7fe5932011-03-10 17:01:16 -08001213void fib_select_multipath(struct fib_result *res)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214{
1215 struct fib_info *fi = res->fi;
1216 int w;
1217
1218 spin_lock_bh(&fib_multipath_lock);
1219 if (fi->fib_power <= 0) {
1220 int power = 0;
1221 change_nexthops(fi) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001222 if (!(nexthop_nh->nh_flags & RTNH_F_DEAD)) {
David S. Miller71fceff2010-01-15 01:16:40 -08001223 power += nexthop_nh->nh_weight;
1224 nexthop_nh->nh_power = nexthop_nh->nh_weight;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 }
1226 } endfor_nexthops(fi);
1227 fi->fib_power = power;
1228 if (power <= 0) {
1229 spin_unlock_bh(&fib_multipath_lock);
1230 /* Race condition: route has just become dead. */
1231 res->nh_sel = 0;
1232 return;
1233 }
1234 }
1235
1236
1237 /* w should be random number [0..fi->fib_power-1],
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001238 * it is pretty bad approximation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 */
1240
1241 w = jiffies % fi->fib_power;
1242
1243 change_nexthops(fi) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001244 if (!(nexthop_nh->nh_flags & RTNH_F_DEAD) &&
David S. Miller71fceff2010-01-15 01:16:40 -08001245 nexthop_nh->nh_power) {
Eric Dumazet6a31d2a2010-10-04 20:00:18 +00001246 w -= nexthop_nh->nh_power;
1247 if (w <= 0) {
David S. Miller71fceff2010-01-15 01:16:40 -08001248 nexthop_nh->nh_power--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 fi->fib_power--;
1250 res->nh_sel = nhsel;
1251 spin_unlock_bh(&fib_multipath_lock);
1252 return;
1253 }
1254 }
1255 } endfor_nexthops(fi);
1256
1257 /* Race condition: route has just become dead. */
1258 res->nh_sel = 0;
1259 spin_unlock_bh(&fib_multipath_lock);
1260}
1261#endif