blob: 34d094cadb11a3ef9b8351140e9c0ae4d60a4d1e [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 * ROUTE - implementation of the IP router.
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Alan Cox, <gw4pts@gw4pts.ampr.org>
11 * Linus Torvalds, <Linus.Torvalds@helsinki.fi>
12 * Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
13 *
14 * Fixes:
15 * Alan Cox : Verify area fixes.
16 * Alan Cox : cli() protects routing changes
17 * Rui Oliveira : ICMP routing table updates
18 * (rco@di.uminho.pt) Routing table insertion and update
19 * Linus Torvalds : Rewrote bits to be sensible
20 * Alan Cox : Added BSD route gw semantics
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090021 * Alan Cox : Super /proc >4K
Linus Torvalds1da177e2005-04-16 15:20:36 -070022 * Alan Cox : MTU in route table
23 * Alan Cox : MSS actually. Also added the window
24 * clamper.
25 * Sam Lantinga : Fixed route matching in rt_del()
26 * Alan Cox : Routing cache support.
27 * Alan Cox : Removed compatibility cruft.
28 * Alan Cox : RTF_REJECT support.
29 * Alan Cox : TCP irtt support.
30 * Jonathan Naylor : Added Metric support.
31 * Miquel van Smoorenburg : BSD API fixes.
32 * Miquel van Smoorenburg : Metrics.
33 * Alan Cox : Use __u32 properly
34 * Alan Cox : Aligned routing errors more closely with BSD
35 * our system is still very different.
36 * Alan Cox : Faster /proc handling
37 * Alexey Kuznetsov : Massive rework to support tree based routing,
38 * routing caches and better behaviour.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090039 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070040 * Olaf Erb : irtt wasn't being copied right.
41 * Bjorn Ekwall : Kerneld route support.
42 * Alan Cox : Multicast fixed (I hope)
43 * Pavel Krauz : Limited broadcast fixed
44 * Mike McLagan : Routing by source
45 * Alexey Kuznetsov : End of old history. Split to fib.c and
46 * route.c and rewritten from scratch.
47 * Andi Kleen : Load-limit warning messages.
48 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
49 * Vitaly E. Lavrov : Race condition in ip_route_input_slow.
50 * Tobias Ringstrom : Uninitialized res.type in ip_route_output_slow.
51 * Vladimir V. Ivanov : IP rule info (flowid) is really useful.
52 * Marc Boucher : routing by fwmark
53 * Robert Olsson : Added rt_cache statistics
54 * Arnaldo C. Melo : Convert proc stuff to seq_file
Eric Dumazetbb1d23b2005-07-05 15:00:32 -070055 * Eric Dumazet : hashed spinlocks and rt_check_expire() fixes.
Ilia Sotnikovcef26852006-03-25 01:38:55 -080056 * Ilia Sotnikov : Ignore TOS on PMTUD and Redirect
57 * Ilia Sotnikov : Removed TOS from hash calculations
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 *
59 * This program is free software; you can redistribute it and/or
60 * modify it under the terms of the GNU General Public License
61 * as published by the Free Software Foundation; either version
62 * 2 of the License, or (at your option) any later version.
63 */
64
Joe Perchesafd465032012-03-12 07:03:32 +000065#define pr_fmt(fmt) "IPv4: " fmt
66
Linus Torvalds1da177e2005-04-16 15:20:36 -070067#include <linux/module.h>
68#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070069#include <linux/bitops.h>
70#include <linux/types.h>
71#include <linux/kernel.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070072#include <linux/mm.h>
73#include <linux/string.h>
74#include <linux/socket.h>
75#include <linux/sockios.h>
76#include <linux/errno.h>
77#include <linux/in.h>
78#include <linux/inet.h>
79#include <linux/netdevice.h>
80#include <linux/proc_fs.h>
81#include <linux/init.h>
82#include <linux/skbuff.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070083#include <linux/inetdevice.h>
84#include <linux/igmp.h>
85#include <linux/pkt_sched.h>
86#include <linux/mroute.h>
87#include <linux/netfilter_ipv4.h>
88#include <linux/random.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/rcupdate.h>
90#include <linux/times.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090091#include <linux/slab.h>
Herbert Xu352e5122007-11-13 21:34:06 -080092#include <net/dst.h>
Eric W. Biederman457c4cb2007-09-12 12:01:34 +020093#include <net/net_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070094#include <net/protocol.h>
95#include <net/ip.h>
96#include <net/route.h>
97#include <net/inetpeer.h>
98#include <net/sock.h>
99#include <net/ip_fib.h>
100#include <net/arp.h>
101#include <net/tcp.h>
102#include <net/icmp.h>
103#include <net/xfrm.h>
Tom Tucker8d717402006-07-30 20:43:36 -0700104#include <net/netevent.h>
Thomas Graf63f34442007-03-22 11:55:17 -0700105#include <net/rtnetlink.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#ifdef CONFIG_SYSCTL
107#include <linux/sysctl.h>
Shan Wei7426a562012-04-18 18:05:46 +0000108#include <linux/kmemleak.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109#endif
David S. Miller6e5714e2011-08-03 20:50:44 -0700110#include <net/secure_seq.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
David S. Miller68a5e3d2011-03-11 20:07:33 -0500112#define RT_FL_TOS(oldflp4) \
Julian Anastasovf61759e2011-12-02 11:39:42 +0000113 ((oldflp4)->flowi4_tos & (IPTOS_RT_MASK | RTO_ONLINK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115#define RT_GC_TIMEOUT (300*HZ)
116
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117static int ip_rt_max_size;
Stephen Hemminger817bc4d2008-03-22 17:43:59 -0700118static int ip_rt_redirect_number __read_mostly = 9;
119static int ip_rt_redirect_load __read_mostly = HZ / 50;
120static int ip_rt_redirect_silence __read_mostly = ((HZ / 50) << (9 + 1));
121static int ip_rt_error_cost __read_mostly = HZ;
122static int ip_rt_error_burst __read_mostly = 5 * HZ;
Stephen Hemminger817bc4d2008-03-22 17:43:59 -0700123static int ip_rt_mtu_expires __read_mostly = 10 * 60 * HZ;
124static int ip_rt_min_pmtu __read_mostly = 512 + 20 + 20;
125static int ip_rt_min_advmss __read_mostly = 256;
Eric Dumazet9f28a2f2011-12-21 15:47:16 -0500126
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127/*
128 * Interface to generic destination cache.
129 */
130
131static struct dst_entry *ipv4_dst_check(struct dst_entry *dst, u32 cookie);
David S. Miller0dbaee32010-12-13 12:52:14 -0800132static unsigned int ipv4_default_advmss(const struct dst_entry *dst);
Steffen Klassertebb762f2011-11-23 02:12:51 +0000133static unsigned int ipv4_mtu(const struct dst_entry *dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134static struct dst_entry *ipv4_negative_advice(struct dst_entry *dst);
135static void ipv4_link_failure(struct sk_buff *skb);
David S. Miller6700c272012-07-17 03:29:28 -0700136static void ip_rt_update_pmtu(struct dst_entry *dst, struct sock *sk,
137 struct sk_buff *skb, u32 mtu);
138static void ip_do_redirect(struct dst_entry *dst, struct sock *sk,
139 struct sk_buff *skb);
David S. Millercaacf052012-07-31 15:06:50 -0700140static void ipv4_dst_destroy(struct dst_entry *dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
David S. Miller62fa8a82011-01-26 20:51:05 -0800142static u32 *ipv4_cow_metrics(struct dst_entry *dst, unsigned long old)
143{
David S. Miller31248732012-07-10 07:08:18 -0700144 WARN_ON(1);
145 return NULL;
David S. Miller62fa8a82011-01-26 20:51:05 -0800146}
147
David S. Millerf894cbf2012-07-02 21:52:24 -0700148static struct neighbour *ipv4_neigh_lookup(const struct dst_entry *dst,
149 struct sk_buff *skb,
150 const void *daddr);
David S. Millerd3aaeb32011-07-18 00:40:17 -0700151
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152static struct dst_ops ipv4_dst_ops = {
153 .family = AF_INET,
Harvey Harrison09640e62009-02-01 00:45:17 -0800154 .protocol = cpu_to_be16(ETH_P_IP),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 .check = ipv4_dst_check,
David S. Miller0dbaee32010-12-13 12:52:14 -0800156 .default_advmss = ipv4_default_advmss,
Steffen Klassertebb762f2011-11-23 02:12:51 +0000157 .mtu = ipv4_mtu,
David S. Miller62fa8a82011-01-26 20:51:05 -0800158 .cow_metrics = ipv4_cow_metrics,
David S. Millercaacf052012-07-31 15:06:50 -0700159 .destroy = ipv4_dst_destroy,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 .negative_advice = ipv4_negative_advice,
161 .link_failure = ipv4_link_failure,
162 .update_pmtu = ip_rt_update_pmtu,
David S. Millere47a1852012-07-11 20:55:47 -0700163 .redirect = ip_do_redirect,
Herbert Xu1ac06e02008-05-20 14:32:14 -0700164 .local_out = __ip_local_out,
David S. Millerd3aaeb32011-07-18 00:40:17 -0700165 .neigh_lookup = ipv4_neigh_lookup,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166};
167
168#define ECN_OR_COST(class) TC_PRIO_##class
169
Philippe De Muyter4839c522007-07-09 15:32:57 -0700170const __u8 ip_tos2prio[16] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 TC_PRIO_BESTEFFORT,
Dan Siemon4a2b9c32011-03-15 13:56:07 +0000172 ECN_OR_COST(BESTEFFORT),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 TC_PRIO_BESTEFFORT,
174 ECN_OR_COST(BESTEFFORT),
175 TC_PRIO_BULK,
176 ECN_OR_COST(BULK),
177 TC_PRIO_BULK,
178 ECN_OR_COST(BULK),
179 TC_PRIO_INTERACTIVE,
180 ECN_OR_COST(INTERACTIVE),
181 TC_PRIO_INTERACTIVE,
182 ECN_OR_COST(INTERACTIVE),
183 TC_PRIO_INTERACTIVE_BULK,
184 ECN_OR_COST(INTERACTIVE_BULK),
185 TC_PRIO_INTERACTIVE_BULK,
186 ECN_OR_COST(INTERACTIVE_BULK)
187};
Amir Vadaid4a96862012-04-04 21:33:28 +0000188EXPORT_SYMBOL(ip_tos2prio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189
Eric Dumazet2f970d82006-01-17 02:54:36 -0800190static DEFINE_PER_CPU(struct rt_cache_stat, rt_cache_stat);
Christoph Lameter3ed66e92014-04-07 15:39:40 -0700191#define RT_CACHE_STAT_INC(field) raw_cpu_inc(rt_cache_stat.field)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193#ifdef CONFIG_PROC_FS
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194static void *rt_cache_seq_start(struct seq_file *seq, loff_t *pos)
195{
Eric Dumazet29e75252008-01-31 17:05:09 -0800196 if (*pos)
David S. Miller89aef892012-07-17 11:00:09 -0700197 return NULL;
Eric Dumazet29e75252008-01-31 17:05:09 -0800198 return SEQ_START_TOKEN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199}
200
201static void *rt_cache_seq_next(struct seq_file *seq, void *v, loff_t *pos)
202{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 ++*pos;
David S. Miller89aef892012-07-17 11:00:09 -0700204 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205}
206
207static void rt_cache_seq_stop(struct seq_file *seq, void *v)
208{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209}
210
211static int rt_cache_seq_show(struct seq_file *seq, void *v)
212{
213 if (v == SEQ_START_TOKEN)
214 seq_printf(seq, "%-127s\n",
215 "Iface\tDestination\tGateway \tFlags\t\tRefCnt\tUse\t"
216 "Metric\tSource\t\tMTU\tWindow\tIRTT\tTOS\tHHRef\t"
217 "HHUptod\tSpecDst");
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900218 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219}
220
Stephen Hemmingerf6908082007-03-12 14:34:29 -0700221static const struct seq_operations rt_cache_seq_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 .start = rt_cache_seq_start,
223 .next = rt_cache_seq_next,
224 .stop = rt_cache_seq_stop,
225 .show = rt_cache_seq_show,
226};
227
228static int rt_cache_seq_open(struct inode *inode, struct file *file)
229{
David S. Miller89aef892012-07-17 11:00:09 -0700230 return seq_open(file, &rt_cache_seq_ops);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231}
232
Arjan van de Ven9a321442007-02-12 00:55:35 -0800233static const struct file_operations rt_cache_seq_fops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 .owner = THIS_MODULE,
235 .open = rt_cache_seq_open,
236 .read = seq_read,
237 .llseek = seq_lseek,
David S. Miller89aef892012-07-17 11:00:09 -0700238 .release = seq_release,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239};
240
241
242static void *rt_cpu_seq_start(struct seq_file *seq, loff_t *pos)
243{
244 int cpu;
245
246 if (*pos == 0)
247 return SEQ_START_TOKEN;
248
Rusty Russell0f23174a2008-12-29 12:23:42 +0000249 for (cpu = *pos-1; cpu < nr_cpu_ids; ++cpu) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250 if (!cpu_possible(cpu))
251 continue;
252 *pos = cpu+1;
Eric Dumazet2f970d82006-01-17 02:54:36 -0800253 return &per_cpu(rt_cache_stat, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 }
255 return NULL;
256}
257
258static void *rt_cpu_seq_next(struct seq_file *seq, void *v, loff_t *pos)
259{
260 int cpu;
261
Rusty Russell0f23174a2008-12-29 12:23:42 +0000262 for (cpu = *pos; cpu < nr_cpu_ids; ++cpu) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 if (!cpu_possible(cpu))
264 continue;
265 *pos = cpu+1;
Eric Dumazet2f970d82006-01-17 02:54:36 -0800266 return &per_cpu(rt_cache_stat, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 }
268 return NULL;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270}
271
272static void rt_cpu_seq_stop(struct seq_file *seq, void *v)
273{
274
275}
276
277static int rt_cpu_seq_show(struct seq_file *seq, void *v)
278{
279 struct rt_cache_stat *st = v;
280
281 if (v == SEQ_START_TOKEN) {
Olaf Rempel5bec0032005-04-28 12:16:08 -0700282 seq_printf(seq, "entries in_hit in_slow_tot in_slow_mc in_no_route in_brd in_martian_dst in_martian_src out_hit out_slow_tot out_slow_mc gc_total gc_ignored gc_goal_miss gc_dst_overflow in_hlist_search out_hlist_search\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 return 0;
284 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 seq_printf(seq,"%08x %08x %08x %08x %08x %08x %08x %08x "
287 " %08x %08x %08x %08x %08x %08x %08x %08x %08x \n",
Eric Dumazetfc66f952010-10-08 06:37:34 +0000288 dst_entries_get_slow(&ipv4_dst_ops),
Eric Dumazet0baf2b32013-10-16 02:49:04 -0700289 0, /* st->in_hit */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 st->in_slow_tot,
291 st->in_slow_mc,
292 st->in_no_route,
293 st->in_brd,
294 st->in_martian_dst,
295 st->in_martian_src,
296
Eric Dumazet0baf2b32013-10-16 02:49:04 -0700297 0, /* st->out_hit */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 st->out_slow_tot,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900299 st->out_slow_mc,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
Eric Dumazet0baf2b32013-10-16 02:49:04 -0700301 0, /* st->gc_total */
302 0, /* st->gc_ignored */
303 0, /* st->gc_goal_miss */
304 0, /* st->gc_dst_overflow */
305 0, /* st->in_hlist_search */
306 0 /* st->out_hlist_search */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 );
308 return 0;
309}
310
Stephen Hemmingerf6908082007-03-12 14:34:29 -0700311static const struct seq_operations rt_cpu_seq_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 .start = rt_cpu_seq_start,
313 .next = rt_cpu_seq_next,
314 .stop = rt_cpu_seq_stop,
315 .show = rt_cpu_seq_show,
316};
317
318
319static int rt_cpu_seq_open(struct inode *inode, struct file *file)
320{
321 return seq_open(file, &rt_cpu_seq_ops);
322}
323
Arjan van de Ven9a321442007-02-12 00:55:35 -0800324static const struct file_operations rt_cpu_seq_fops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 .owner = THIS_MODULE,
326 .open = rt_cpu_seq_open,
327 .read = seq_read,
328 .llseek = seq_lseek,
329 .release = seq_release,
330};
331
Patrick McHardyc7066f72011-01-14 13:36:42 +0100332#ifdef CONFIG_IP_ROUTE_CLASSID
Alexey Dobriyana661c412009-11-25 15:40:35 -0800333static int rt_acct_proc_show(struct seq_file *m, void *v)
Pavel Emelyanov78c686e2007-12-05 21:13:48 -0800334{
Alexey Dobriyana661c412009-11-25 15:40:35 -0800335 struct ip_rt_acct *dst, *src;
336 unsigned int i, j;
Pavel Emelyanov78c686e2007-12-05 21:13:48 -0800337
Alexey Dobriyana661c412009-11-25 15:40:35 -0800338 dst = kcalloc(256, sizeof(struct ip_rt_acct), GFP_KERNEL);
339 if (!dst)
340 return -ENOMEM;
Pavel Emelyanov78c686e2007-12-05 21:13:48 -0800341
Alexey Dobriyana661c412009-11-25 15:40:35 -0800342 for_each_possible_cpu(i) {
343 src = (struct ip_rt_acct *)per_cpu_ptr(ip_rt_acct, i);
344 for (j = 0; j < 256; j++) {
345 dst[j].o_bytes += src[j].o_bytes;
346 dst[j].o_packets += src[j].o_packets;
347 dst[j].i_bytes += src[j].i_bytes;
348 dst[j].i_packets += src[j].i_packets;
Pavel Emelyanov78c686e2007-12-05 21:13:48 -0800349 }
350 }
Alexey Dobriyana661c412009-11-25 15:40:35 -0800351
352 seq_write(m, dst, 256 * sizeof(struct ip_rt_acct));
353 kfree(dst);
354 return 0;
Pavel Emelyanov78c686e2007-12-05 21:13:48 -0800355}
Alexey Dobriyana661c412009-11-25 15:40:35 -0800356
357static int rt_acct_proc_open(struct inode *inode, struct file *file)
358{
359 return single_open(file, rt_acct_proc_show, NULL);
360}
361
362static const struct file_operations rt_acct_proc_fops = {
363 .owner = THIS_MODULE,
364 .open = rt_acct_proc_open,
365 .read = seq_read,
366 .llseek = seq_lseek,
367 .release = single_release,
368};
Pavel Emelyanov78c686e2007-12-05 21:13:48 -0800369#endif
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800370
Denis V. Lunev73b38712008-02-28 20:51:18 -0800371static int __net_init ip_rt_do_proc_init(struct net *net)
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800372{
373 struct proc_dir_entry *pde;
374
Gao fengd4beaa62013-02-18 01:34:54 +0000375 pde = proc_create("rt_cache", S_IRUGO, net->proc_net,
376 &rt_cache_seq_fops);
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800377 if (!pde)
378 goto err1;
379
Wang Chen77020722008-02-28 14:14:25 -0800380 pde = proc_create("rt_cache", S_IRUGO,
381 net->proc_net_stat, &rt_cpu_seq_fops);
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800382 if (!pde)
383 goto err2;
384
Patrick McHardyc7066f72011-01-14 13:36:42 +0100385#ifdef CONFIG_IP_ROUTE_CLASSID
Alexey Dobriyana661c412009-11-25 15:40:35 -0800386 pde = proc_create("rt_acct", 0, net->proc_net, &rt_acct_proc_fops);
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800387 if (!pde)
388 goto err3;
389#endif
390 return 0;
391
Patrick McHardyc7066f72011-01-14 13:36:42 +0100392#ifdef CONFIG_IP_ROUTE_CLASSID
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800393err3:
394 remove_proc_entry("rt_cache", net->proc_net_stat);
395#endif
396err2:
397 remove_proc_entry("rt_cache", net->proc_net);
398err1:
399 return -ENOMEM;
400}
Denis V. Lunev73b38712008-02-28 20:51:18 -0800401
402static void __net_exit ip_rt_do_proc_exit(struct net *net)
403{
404 remove_proc_entry("rt_cache", net->proc_net_stat);
405 remove_proc_entry("rt_cache", net->proc_net);
Patrick McHardyc7066f72011-01-14 13:36:42 +0100406#ifdef CONFIG_IP_ROUTE_CLASSID
Denis V. Lunev73b38712008-02-28 20:51:18 -0800407 remove_proc_entry("rt_acct", net->proc_net);
Alexey Dobriyan0a931ac2010-01-17 03:32:50 +0000408#endif
Denis V. Lunev73b38712008-02-28 20:51:18 -0800409}
410
411static struct pernet_operations ip_rt_proc_ops __net_initdata = {
412 .init = ip_rt_do_proc_init,
413 .exit = ip_rt_do_proc_exit,
414};
415
416static int __init ip_rt_proc_init(void)
417{
418 return register_pernet_subsys(&ip_rt_proc_ops);
419}
420
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800421#else
Denis V. Lunev73b38712008-02-28 20:51:18 -0800422static inline int ip_rt_proc_init(void)
Pavel Emelyanov107f1632007-12-05 21:14:28 -0800423{
424 return 0;
425}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426#endif /* CONFIG_PROC_FS */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900427
Eric Dumazet4331deb2012-07-25 05:11:23 +0000428static inline bool rt_is_expired(const struct rtable *rth)
Denis V. Luneve84f84f2008-07-05 19:04:32 -0700429{
fan.duca4c3fc2013-07-30 08:33:53 +0800430 return rth->rt_genid != rt_genid_ipv4(dev_net(rth->dst.dev));
Denis V. Luneve84f84f2008-07-05 19:04:32 -0700431}
432
Nicolas Dichtel4ccfe6d2012-09-07 00:45:29 +0000433void rt_cache_flush(struct net *net)
Eric Dumazet29e75252008-01-31 17:05:09 -0800434{
fan.duca4c3fc2013-07-30 08:33:53 +0800435 rt_genid_bump_ipv4(net);
Eric Dumazet98376382010-03-08 03:20:00 +0000436}
437
David S. Millerf894cbf2012-07-02 21:52:24 -0700438static struct neighbour *ipv4_neigh_lookup(const struct dst_entry *dst,
439 struct sk_buff *skb,
440 const void *daddr)
David Miller3769cff2011-07-11 22:44:24 +0000441{
David S. Millerd3aaeb32011-07-18 00:40:17 -0700442 struct net_device *dev = dst->dev;
443 const __be32 *pkey = daddr;
David S. Miller39232972012-01-26 15:22:32 -0500444 const struct rtable *rt;
David Miller3769cff2011-07-11 22:44:24 +0000445 struct neighbour *n;
446
David S. Miller39232972012-01-26 15:22:32 -0500447 rt = (const struct rtable *) dst;
David S. Millera263b302012-07-02 02:02:15 -0700448 if (rt->rt_gateway)
David S. Miller39232972012-01-26 15:22:32 -0500449 pkey = (const __be32 *) &rt->rt_gateway;
David S. Millerf894cbf2012-07-02 21:52:24 -0700450 else if (skb)
451 pkey = &ip_hdr(skb)->daddr;
David S. Millerd3aaeb32011-07-18 00:40:17 -0700452
David S. Miller80703d22012-02-15 17:48:35 -0500453 n = __ipv4_neigh_lookup(dev, *(__force u32 *)pkey);
David S. Millerd3aaeb32011-07-18 00:40:17 -0700454 if (n)
455 return n;
David Miller32092ec2011-07-25 00:01:41 +0000456 return neigh_create(&arp_tbl, pkey, dev);
David S. Millerd3aaeb32011-07-18 00:40:17 -0700457}
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459/*
460 * Peer allocation may fail only in serious out-of-memory conditions. However
461 * we still can generate some output.
462 * Random ID selection looks a bit dangerous because we have no chances to
463 * select ID being unique in a reasonable period of time.
464 * But broken packet identifier may be better than no packet at all.
465 */
466static void ip_select_fb_ident(struct iphdr *iph)
467{
468 static DEFINE_SPINLOCK(ip_fb_id_lock);
469 static u32 ip_fallback_id;
470 u32 salt;
471
472 spin_lock_bh(&ip_fb_id_lock);
Al Viroe4485152006-09-26 22:15:01 -0700473 salt = secure_ip_id((__force __be32)ip_fallback_id ^ iph->daddr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 iph->id = htons(salt & 0xFFFF);
475 ip_fallback_id = salt;
476 spin_unlock_bh(&ip_fb_id_lock);
477}
478
479void __ip_select_ident(struct iphdr *iph, struct dst_entry *dst, int more)
480{
David S. Miller1d861aa2012-07-10 03:58:16 -0700481 struct net *net = dev_net(dst->dev);
482 struct inet_peer *peer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483
David S. Miller1d861aa2012-07-10 03:58:16 -0700484 peer = inet_getpeer_v4(net->ipv4.peers, iph->daddr, 1);
485 if (peer) {
486 iph->id = htons(inet_getid(peer, more));
487 inet_putpeer(peer);
488 return;
489 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490
491 ip_select_fb_ident(iph);
492}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000493EXPORT_SYMBOL(__ip_select_ident);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494
Eric Dumazet5abf7f72012-07-17 22:42:13 +0200495static void __build_flow_key(struct flowi4 *fl4, const struct sock *sk,
David S. Miller4895c772012-07-17 04:19:00 -0700496 const struct iphdr *iph,
497 int oif, u8 tos,
498 u8 prot, u32 mark, int flow_flags)
499{
500 if (sk) {
501 const struct inet_sock *inet = inet_sk(sk);
502
503 oif = sk->sk_bound_dev_if;
504 mark = sk->sk_mark;
505 tos = RT_CONN_FLAGS(sk);
506 prot = inet->hdrincl ? IPPROTO_RAW : sk->sk_protocol;
507 }
508 flowi4_init_output(fl4, oif, mark, tos,
509 RT_SCOPE_UNIVERSE, prot,
510 flow_flags,
511 iph->daddr, iph->saddr, 0, 0);
512}
513
Eric Dumazet5abf7f72012-07-17 22:42:13 +0200514static void build_skb_flow_key(struct flowi4 *fl4, const struct sk_buff *skb,
515 const struct sock *sk)
David S. Miller4895c772012-07-17 04:19:00 -0700516{
517 const struct iphdr *iph = ip_hdr(skb);
518 int oif = skb->dev->ifindex;
519 u8 tos = RT_TOS(iph->tos);
520 u8 prot = iph->protocol;
521 u32 mark = skb->mark;
522
523 __build_flow_key(fl4, sk, iph, oif, tos, prot, mark, 0);
524}
525
Eric Dumazet5abf7f72012-07-17 22:42:13 +0200526static void build_sk_flow_key(struct flowi4 *fl4, const struct sock *sk)
David S. Miller4895c772012-07-17 04:19:00 -0700527{
528 const struct inet_sock *inet = inet_sk(sk);
Eric Dumazet5abf7f72012-07-17 22:42:13 +0200529 const struct ip_options_rcu *inet_opt;
David S. Miller4895c772012-07-17 04:19:00 -0700530 __be32 daddr = inet->inet_daddr;
531
532 rcu_read_lock();
533 inet_opt = rcu_dereference(inet->inet_opt);
534 if (inet_opt && inet_opt->opt.srr)
535 daddr = inet_opt->opt.faddr;
536 flowi4_init_output(fl4, sk->sk_bound_dev_if, sk->sk_mark,
537 RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
538 inet->hdrincl ? IPPROTO_RAW : sk->sk_protocol,
539 inet_sk_flowi_flags(sk),
540 daddr, inet->inet_saddr, 0, 0);
541 rcu_read_unlock();
542}
543
Eric Dumazet5abf7f72012-07-17 22:42:13 +0200544static void ip_rt_build_flow_key(struct flowi4 *fl4, const struct sock *sk,
545 const struct sk_buff *skb)
David S. Miller4895c772012-07-17 04:19:00 -0700546{
547 if (skb)
548 build_skb_flow_key(fl4, skb, sk);
549 else
550 build_sk_flow_key(fl4, sk);
551}
552
David S. Millerc5038a82012-07-31 15:02:02 -0700553static inline void rt_free(struct rtable *rt)
554{
555 call_rcu(&rt->dst.rcu_head, dst_rcu_free);
556}
557
558static DEFINE_SPINLOCK(fnhe_lock);
David S. Miller4895c772012-07-17 04:19:00 -0700559
Timo Teräs2ffae992013-06-27 10:27:05 +0300560static void fnhe_flush_routes(struct fib_nh_exception *fnhe)
561{
562 struct rtable *rt;
563
564 rt = rcu_dereference(fnhe->fnhe_rth_input);
565 if (rt) {
566 RCU_INIT_POINTER(fnhe->fnhe_rth_input, NULL);
567 rt_free(rt);
568 }
569 rt = rcu_dereference(fnhe->fnhe_rth_output);
570 if (rt) {
571 RCU_INIT_POINTER(fnhe->fnhe_rth_output, NULL);
572 rt_free(rt);
573 }
574}
575
Julian Anastasovaee06da2012-07-18 10:15:35 +0000576static struct fib_nh_exception *fnhe_oldest(struct fnhe_hash_bucket *hash)
David S. Miller4895c772012-07-17 04:19:00 -0700577{
578 struct fib_nh_exception *fnhe, *oldest;
579
580 oldest = rcu_dereference(hash->chain);
581 for (fnhe = rcu_dereference(oldest->fnhe_next); fnhe;
582 fnhe = rcu_dereference(fnhe->fnhe_next)) {
583 if (time_before(fnhe->fnhe_stamp, oldest->fnhe_stamp))
584 oldest = fnhe;
585 }
Timo Teräs2ffae992013-06-27 10:27:05 +0300586 fnhe_flush_routes(oldest);
David S. Miller4895c772012-07-17 04:19:00 -0700587 return oldest;
588}
589
David S. Millerd3a25c92012-07-17 13:23:08 -0700590static inline u32 fnhe_hashfun(__be32 daddr)
591{
592 u32 hval;
593
594 hval = (__force u32) daddr;
595 hval ^= (hval >> 11) ^ (hval >> 22);
596
597 return hval & (FNHE_HASH_SIZE - 1);
598}
599
Timo Teräs387aa652013-05-27 20:46:31 +0000600static void fill_route_from_fnhe(struct rtable *rt, struct fib_nh_exception *fnhe)
601{
602 rt->rt_pmtu = fnhe->fnhe_pmtu;
603 rt->dst.expires = fnhe->fnhe_expires;
604
605 if (fnhe->fnhe_gw) {
606 rt->rt_flags |= RTCF_REDIRECTED;
607 rt->rt_gateway = fnhe->fnhe_gw;
608 rt->rt_uses_gateway = 1;
609 }
610}
611
Julian Anastasovaee06da2012-07-18 10:15:35 +0000612static void update_or_create_fnhe(struct fib_nh *nh, __be32 daddr, __be32 gw,
613 u32 pmtu, unsigned long expires)
David S. Miller4895c772012-07-17 04:19:00 -0700614{
Julian Anastasovaee06da2012-07-18 10:15:35 +0000615 struct fnhe_hash_bucket *hash;
David S. Miller4895c772012-07-17 04:19:00 -0700616 struct fib_nh_exception *fnhe;
Timo Teräs387aa652013-05-27 20:46:31 +0000617 struct rtable *rt;
618 unsigned int i;
David S. Miller4895c772012-07-17 04:19:00 -0700619 int depth;
Julian Anastasovaee06da2012-07-18 10:15:35 +0000620 u32 hval = fnhe_hashfun(daddr);
David S. Miller4895c772012-07-17 04:19:00 -0700621
David S. Millerc5038a82012-07-31 15:02:02 -0700622 spin_lock_bh(&fnhe_lock);
Julian Anastasovaee06da2012-07-18 10:15:35 +0000623
624 hash = nh->nh_exceptions;
David S. Miller4895c772012-07-17 04:19:00 -0700625 if (!hash) {
Julian Anastasovaee06da2012-07-18 10:15:35 +0000626 hash = kzalloc(FNHE_HASH_SIZE * sizeof(*hash), GFP_ATOMIC);
David S. Miller4895c772012-07-17 04:19:00 -0700627 if (!hash)
Julian Anastasovaee06da2012-07-18 10:15:35 +0000628 goto out_unlock;
629 nh->nh_exceptions = hash;
David S. Miller4895c772012-07-17 04:19:00 -0700630 }
631
David S. Miller4895c772012-07-17 04:19:00 -0700632 hash += hval;
633
634 depth = 0;
635 for (fnhe = rcu_dereference(hash->chain); fnhe;
636 fnhe = rcu_dereference(fnhe->fnhe_next)) {
637 if (fnhe->fnhe_daddr == daddr)
Julian Anastasovaee06da2012-07-18 10:15:35 +0000638 break;
David S. Miller4895c772012-07-17 04:19:00 -0700639 depth++;
640 }
641
Julian Anastasovaee06da2012-07-18 10:15:35 +0000642 if (fnhe) {
643 if (gw)
644 fnhe->fnhe_gw = gw;
645 if (pmtu) {
646 fnhe->fnhe_pmtu = pmtu;
Timo Teräs387aa652013-05-27 20:46:31 +0000647 fnhe->fnhe_expires = max(1UL, expires);
Julian Anastasovaee06da2012-07-18 10:15:35 +0000648 }
Timo Teräs387aa652013-05-27 20:46:31 +0000649 /* Update all cached dsts too */
Timo Teräs2ffae992013-06-27 10:27:05 +0300650 rt = rcu_dereference(fnhe->fnhe_rth_input);
651 if (rt)
652 fill_route_from_fnhe(rt, fnhe);
653 rt = rcu_dereference(fnhe->fnhe_rth_output);
Timo Teräs387aa652013-05-27 20:46:31 +0000654 if (rt)
655 fill_route_from_fnhe(rt, fnhe);
Julian Anastasovaee06da2012-07-18 10:15:35 +0000656 } else {
657 if (depth > FNHE_RECLAIM_DEPTH)
658 fnhe = fnhe_oldest(hash);
659 else {
660 fnhe = kzalloc(sizeof(*fnhe), GFP_ATOMIC);
661 if (!fnhe)
662 goto out_unlock;
663
664 fnhe->fnhe_next = hash->chain;
665 rcu_assign_pointer(hash->chain, fnhe);
666 }
Timo Teräs5aad1de2013-05-27 20:46:33 +0000667 fnhe->fnhe_genid = fnhe_genid(dev_net(nh->nh_dev));
Julian Anastasovaee06da2012-07-18 10:15:35 +0000668 fnhe->fnhe_daddr = daddr;
669 fnhe->fnhe_gw = gw;
670 fnhe->fnhe_pmtu = pmtu;
671 fnhe->fnhe_expires = expires;
Timo Teräs387aa652013-05-27 20:46:31 +0000672
673 /* Exception created; mark the cached routes for the nexthop
674 * stale, so anyone caching it rechecks if this exception
675 * applies to them.
676 */
Timo Teräs2ffae992013-06-27 10:27:05 +0300677 rt = rcu_dereference(nh->nh_rth_input);
678 if (rt)
679 rt->dst.obsolete = DST_OBSOLETE_KILL;
680
Timo Teräs387aa652013-05-27 20:46:31 +0000681 for_each_possible_cpu(i) {
682 struct rtable __rcu **prt;
683 prt = per_cpu_ptr(nh->nh_pcpu_rth_output, i);
684 rt = rcu_dereference(*prt);
685 if (rt)
686 rt->dst.obsolete = DST_OBSOLETE_KILL;
687 }
David S. Miller4895c772012-07-17 04:19:00 -0700688 }
David S. Miller4895c772012-07-17 04:19:00 -0700689
David S. Miller4895c772012-07-17 04:19:00 -0700690 fnhe->fnhe_stamp = jiffies;
Julian Anastasovaee06da2012-07-18 10:15:35 +0000691
692out_unlock:
David S. Millerc5038a82012-07-31 15:02:02 -0700693 spin_unlock_bh(&fnhe_lock);
David S. Miller4895c772012-07-17 04:19:00 -0700694}
695
David S. Millerceb33202012-07-17 11:31:28 -0700696static void __ip_do_redirect(struct rtable *rt, struct sk_buff *skb, struct flowi4 *fl4,
697 bool kill_route)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698{
David S. Millere47a1852012-07-11 20:55:47 -0700699 __be32 new_gw = icmp_hdr(skb)->un.gateway;
David S. Miller94206122012-07-11 20:38:08 -0700700 __be32 old_gw = ip_hdr(skb)->saddr;
David S. Millere47a1852012-07-11 20:55:47 -0700701 struct net_device *dev = skb->dev;
David S. Millere47a1852012-07-11 20:55:47 -0700702 struct in_device *in_dev;
David S. Miller4895c772012-07-17 04:19:00 -0700703 struct fib_result res;
David S. Millere47a1852012-07-11 20:55:47 -0700704 struct neighbour *n;
Denis V. Lunev317805b2008-02-28 20:50:06 -0800705 struct net *net;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706
David S. Miller94206122012-07-11 20:38:08 -0700707 switch (icmp_hdr(skb)->code & 7) {
708 case ICMP_REDIR_NET:
709 case ICMP_REDIR_NETTOS:
710 case ICMP_REDIR_HOST:
711 case ICMP_REDIR_HOSTTOS:
712 break;
713
714 default:
715 return;
716 }
717
David S. Millere47a1852012-07-11 20:55:47 -0700718 if (rt->rt_gateway != old_gw)
719 return;
720
721 in_dev = __in_dev_get_rcu(dev);
722 if (!in_dev)
723 return;
724
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900725 net = dev_net(dev);
Joe Perches9d4fb272009-11-23 10:41:23 -0800726 if (new_gw == old_gw || !IN_DEV_RX_REDIRECTS(in_dev) ||
727 ipv4_is_multicast(new_gw) || ipv4_is_lbcast(new_gw) ||
728 ipv4_is_zeronet(new_gw))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 goto reject_redirect;
730
731 if (!IN_DEV_SHARED_MEDIA(in_dev)) {
732 if (!inet_addr_onlink(in_dev, new_gw, old_gw))
733 goto reject_redirect;
734 if (IN_DEV_SEC_REDIRECTS(in_dev) && ip_fib_check_default(new_gw, dev))
735 goto reject_redirect;
736 } else {
Denis V. Lunev317805b2008-02-28 20:50:06 -0800737 if (inet_addr_type(net, new_gw) != RTN_UNICAST)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 goto reject_redirect;
739 }
740
David S. Miller4895c772012-07-17 04:19:00 -0700741 n = ipv4_neigh_lookup(&rt->dst, NULL, &new_gw);
David S. Millere47a1852012-07-11 20:55:47 -0700742 if (n) {
743 if (!(n->nud_state & NUD_VALID)) {
744 neigh_event_send(n, NULL);
745 } else {
David S. Miller4895c772012-07-17 04:19:00 -0700746 if (fib_lookup(net, fl4, &res) == 0) {
747 struct fib_nh *nh = &FIB_RES_NH(res);
David S. Miller4895c772012-07-17 04:19:00 -0700748
Julian Anastasovaee06da2012-07-18 10:15:35 +0000749 update_or_create_fnhe(nh, fl4->daddr, new_gw,
750 0, 0);
David S. Miller4895c772012-07-17 04:19:00 -0700751 }
David S. Millerceb33202012-07-17 11:31:28 -0700752 if (kill_route)
753 rt->dst.obsolete = DST_OBSOLETE_KILL;
David S. Millere47a1852012-07-11 20:55:47 -0700754 call_netevent_notifiers(NETEVENT_NEIGH_UPDATE, n);
755 }
756 neigh_release(n);
757 }
758 return;
759
760reject_redirect:
761#ifdef CONFIG_IP_ROUTE_VERBOSE
David S. Miller99ee0382012-07-12 07:40:05 -0700762 if (IN_DEV_LOG_MARTIANS(in_dev)) {
763 const struct iphdr *iph = (const struct iphdr *) skb->data;
764 __be32 daddr = iph->daddr;
765 __be32 saddr = iph->saddr;
766
David S. Millere47a1852012-07-11 20:55:47 -0700767 net_info_ratelimited("Redirect from %pI4 on %s about %pI4 ignored\n"
768 " Advised path = %pI4 -> %pI4\n",
769 &old_gw, dev->name, &new_gw,
770 &saddr, &daddr);
David S. Miller99ee0382012-07-12 07:40:05 -0700771 }
David S. Millere47a1852012-07-11 20:55:47 -0700772#endif
773 ;
774}
775
David S. Miller4895c772012-07-17 04:19:00 -0700776static void ip_do_redirect(struct dst_entry *dst, struct sock *sk, struct sk_buff *skb)
777{
778 struct rtable *rt;
779 struct flowi4 fl4;
Michal Kubecekf96ef982013-05-28 08:26:49 +0200780 const struct iphdr *iph = (const struct iphdr *) skb->data;
781 int oif = skb->dev->ifindex;
782 u8 tos = RT_TOS(iph->tos);
783 u8 prot = iph->protocol;
784 u32 mark = skb->mark;
David S. Miller4895c772012-07-17 04:19:00 -0700785
786 rt = (struct rtable *) dst;
787
Michal Kubecekf96ef982013-05-28 08:26:49 +0200788 __build_flow_key(&fl4, sk, iph, oif, tos, prot, mark, 0);
David S. Millerceb33202012-07-17 11:31:28 -0700789 __ip_do_redirect(rt, skb, &fl4, true);
David S. Miller4895c772012-07-17 04:19:00 -0700790}
791
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792static struct dst_entry *ipv4_negative_advice(struct dst_entry *dst)
793{
Eric Dumazetee6b9672008-03-05 18:30:47 -0800794 struct rtable *rt = (struct rtable *)dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 struct dst_entry *ret = dst;
796
797 if (rt) {
Timo Teräsd11a4dc2010-03-18 23:20:20 +0000798 if (dst->obsolete > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 ip_rt_put(rt);
800 ret = NULL;
David S. Miller59436342012-07-10 06:58:42 -0700801 } else if ((rt->rt_flags & RTCF_REDIRECTED) ||
802 rt->dst.expires) {
David S. Miller89aef892012-07-17 11:00:09 -0700803 ip_rt_put(rt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 ret = NULL;
805 }
806 }
807 return ret;
808}
809
810/*
811 * Algorithm:
812 * 1. The first ip_rt_redirect_number redirects are sent
813 * with exponential backoff, then we stop sending them at all,
814 * assuming that the host ignores our redirects.
815 * 2. If we did not see packets requiring redirects
816 * during ip_rt_redirect_silence, we assume that the host
817 * forgot redirected route and start to send redirects again.
818 *
819 * This algorithm is much cheaper and more intelligent than dumb load limiting
820 * in icmp.c.
821 *
822 * NOTE. Do not forget to inhibit load limiting for redirects (redundant)
823 * and "frag. need" (breaks PMTU discovery) in icmp.c.
824 */
825
826void ip_rt_send_redirect(struct sk_buff *skb)
827{
Eric Dumazet511c3f92009-06-02 05:14:27 +0000828 struct rtable *rt = skb_rtable(skb);
Eric Dumazet30038fc2009-08-28 23:52:01 -0700829 struct in_device *in_dev;
David S. Miller92d86822011-02-04 15:55:25 -0800830 struct inet_peer *peer;
David S. Miller1d861aa2012-07-10 03:58:16 -0700831 struct net *net;
Eric Dumazet30038fc2009-08-28 23:52:01 -0700832 int log_martians;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833
Eric Dumazet30038fc2009-08-28 23:52:01 -0700834 rcu_read_lock();
Changli Gaod8d1f302010-06-10 23:31:35 -0700835 in_dev = __in_dev_get_rcu(rt->dst.dev);
Eric Dumazet30038fc2009-08-28 23:52:01 -0700836 if (!in_dev || !IN_DEV_TX_REDIRECTS(in_dev)) {
837 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 return;
Eric Dumazet30038fc2009-08-28 23:52:01 -0700839 }
840 log_martians = IN_DEV_LOG_MARTIANS(in_dev);
841 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
David S. Miller1d861aa2012-07-10 03:58:16 -0700843 net = dev_net(rt->dst.dev);
844 peer = inet_getpeer_v4(net->ipv4.peers, ip_hdr(skb)->saddr, 1);
David S. Miller92d86822011-02-04 15:55:25 -0800845 if (!peer) {
Julian Anastasove81da0e2012-10-08 11:41:15 +0000846 icmp_send(skb, ICMP_REDIRECT, ICMP_REDIR_HOST,
847 rt_nexthop(rt, ip_hdr(skb)->daddr));
David S. Miller92d86822011-02-04 15:55:25 -0800848 return;
849 }
850
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 /* No redirected packets during ip_rt_redirect_silence;
852 * reset the algorithm.
853 */
David S. Miller92d86822011-02-04 15:55:25 -0800854 if (time_after(jiffies, peer->rate_last + ip_rt_redirect_silence))
855 peer->rate_tokens = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856
857 /* Too many ignored redirects; do not send anything
Changli Gaod8d1f302010-06-10 23:31:35 -0700858 * set dst.rate_last to the last seen redirected packet.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 */
David S. Miller92d86822011-02-04 15:55:25 -0800860 if (peer->rate_tokens >= ip_rt_redirect_number) {
861 peer->rate_last = jiffies;
David S. Miller1d861aa2012-07-10 03:58:16 -0700862 goto out_put_peer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 }
864
865 /* Check for load limit; set rate_last to the latest sent
866 * redirect.
867 */
David S. Miller92d86822011-02-04 15:55:25 -0800868 if (peer->rate_tokens == 0 ||
Li Yewang14fb8a72006-12-18 00:26:35 -0800869 time_after(jiffies,
David S. Miller92d86822011-02-04 15:55:25 -0800870 (peer->rate_last +
871 (ip_rt_redirect_load << peer->rate_tokens)))) {
Julian Anastasove81da0e2012-10-08 11:41:15 +0000872 __be32 gw = rt_nexthop(rt, ip_hdr(skb)->daddr);
873
874 icmp_send(skb, ICMP_REDIRECT, ICMP_REDIR_HOST, gw);
David S. Miller92d86822011-02-04 15:55:25 -0800875 peer->rate_last = jiffies;
876 ++peer->rate_tokens;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877#ifdef CONFIG_IP_ROUTE_VERBOSE
Eric Dumazet30038fc2009-08-28 23:52:01 -0700878 if (log_martians &&
Joe Perchese87cc472012-05-13 21:56:26 +0000879 peer->rate_tokens == ip_rt_redirect_number)
880 net_warn_ratelimited("host %pI4/if%d ignores redirects for %pI4 to %pI4\n",
David S. Miller92101b32012-07-23 16:29:00 -0700881 &ip_hdr(skb)->saddr, inet_iif(skb),
Julian Anastasove81da0e2012-10-08 11:41:15 +0000882 &ip_hdr(skb)->daddr, &gw);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883#endif
884 }
David S. Miller1d861aa2012-07-10 03:58:16 -0700885out_put_peer:
886 inet_putpeer(peer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887}
888
889static int ip_error(struct sk_buff *skb)
890{
David S. Miller251da412012-06-26 16:27:09 -0700891 struct in_device *in_dev = __in_dev_get_rcu(skb->dev);
Eric Dumazet511c3f92009-06-02 05:14:27 +0000892 struct rtable *rt = skb_rtable(skb);
David S. Miller92d86822011-02-04 15:55:25 -0800893 struct inet_peer *peer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 unsigned long now;
David S. Miller251da412012-06-26 16:27:09 -0700895 struct net *net;
David S. Miller92d86822011-02-04 15:55:25 -0800896 bool send;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 int code;
898
David S. Miller251da412012-06-26 16:27:09 -0700899 net = dev_net(rt->dst.dev);
900 if (!IN_DEV_FORWARD(in_dev)) {
901 switch (rt->dst.error) {
902 case EHOSTUNREACH:
903 IP_INC_STATS_BH(net, IPSTATS_MIB_INADDRERRORS);
904 break;
905
906 case ENETUNREACH:
907 IP_INC_STATS_BH(net, IPSTATS_MIB_INNOROUTES);
908 break;
909 }
910 goto out;
911 }
912
Changli Gaod8d1f302010-06-10 23:31:35 -0700913 switch (rt->dst.error) {
Joe Perches4500ebf2011-07-01 09:43:07 +0000914 case EINVAL:
915 default:
916 goto out;
917 case EHOSTUNREACH:
918 code = ICMP_HOST_UNREACH;
919 break;
920 case ENETUNREACH:
921 code = ICMP_NET_UNREACH;
David S. Miller251da412012-06-26 16:27:09 -0700922 IP_INC_STATS_BH(net, IPSTATS_MIB_INNOROUTES);
Joe Perches4500ebf2011-07-01 09:43:07 +0000923 break;
924 case EACCES:
925 code = ICMP_PKT_FILTERED;
926 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 }
928
David S. Miller1d861aa2012-07-10 03:58:16 -0700929 peer = inet_getpeer_v4(net->ipv4.peers, ip_hdr(skb)->saddr, 1);
David S. Miller92d86822011-02-04 15:55:25 -0800930
931 send = true;
932 if (peer) {
933 now = jiffies;
934 peer->rate_tokens += now - peer->rate_last;
935 if (peer->rate_tokens > ip_rt_error_burst)
936 peer->rate_tokens = ip_rt_error_burst;
937 peer->rate_last = now;
938 if (peer->rate_tokens >= ip_rt_error_cost)
939 peer->rate_tokens -= ip_rt_error_cost;
940 else
941 send = false;
David S. Miller1d861aa2012-07-10 03:58:16 -0700942 inet_putpeer(peer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 }
David S. Miller92d86822011-02-04 15:55:25 -0800944 if (send)
945 icmp_send(skb, ICMP_DEST_UNREACH, code, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
947out: kfree_skb(skb);
948 return 0;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900949}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950
Steffen Klassertd851c122012-10-07 22:47:25 +0000951static void __ip_rt_update_pmtu(struct rtable *rt, struct flowi4 *fl4, u32 mtu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952{
Steffen Klassertd851c122012-10-07 22:47:25 +0000953 struct dst_entry *dst = &rt->dst;
David S. Miller4895c772012-07-17 04:19:00 -0700954 struct fib_result res;
David S. Miller2c8cec52011-02-09 20:42:07 -0800955
Steffen Klassertfa1e4922013-01-16 20:58:10 +0000956 if (dst_metric_locked(dst, RTAX_MTU))
957 return;
958
Steffen Klassert7f92d332012-10-07 22:48:18 +0000959 if (dst->dev->mtu < mtu)
960 return;
961
David S. Miller59436342012-07-10 06:58:42 -0700962 if (mtu < ip_rt_min_pmtu)
963 mtu = ip_rt_min_pmtu;
Eric Dumazetfe6fe792011-06-08 06:07:07 +0000964
Timo Teräsf0162292013-05-27 20:46:32 +0000965 if (rt->rt_pmtu == mtu &&
966 time_before(jiffies, dst->expires - ip_rt_mtu_expires / 2))
967 return;
968
Eric Dumazetc5ae7d42012-08-28 12:33:07 +0000969 rcu_read_lock();
Steffen Klassertd851c122012-10-07 22:47:25 +0000970 if (fib_lookup(dev_net(dst->dev), fl4, &res) == 0) {
David S. Miller4895c772012-07-17 04:19:00 -0700971 struct fib_nh *nh = &FIB_RES_NH(res);
David S. Miller4895c772012-07-17 04:19:00 -0700972
Julian Anastasovaee06da2012-07-18 10:15:35 +0000973 update_or_create_fnhe(nh, fl4->daddr, 0, mtu,
974 jiffies + ip_rt_mtu_expires);
David S. Miller4895c772012-07-17 04:19:00 -0700975 }
Eric Dumazetc5ae7d42012-08-28 12:33:07 +0000976 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977}
978
David S. Miller4895c772012-07-17 04:19:00 -0700979static void ip_rt_update_pmtu(struct dst_entry *dst, struct sock *sk,
980 struct sk_buff *skb, u32 mtu)
981{
982 struct rtable *rt = (struct rtable *) dst;
983 struct flowi4 fl4;
984
985 ip_rt_build_flow_key(&fl4, sk, skb);
Steffen Klassertd851c122012-10-07 22:47:25 +0000986 __ip_rt_update_pmtu(rt, &fl4, mtu);
David S. Miller4895c772012-07-17 04:19:00 -0700987}
988
David S. Miller36393392012-06-14 22:21:46 -0700989void ipv4_update_pmtu(struct sk_buff *skb, struct net *net, u32 mtu,
990 int oif, u32 mark, u8 protocol, int flow_flags)
991{
David S. Miller4895c772012-07-17 04:19:00 -0700992 const struct iphdr *iph = (const struct iphdr *) skb->data;
David S. Miller36393392012-06-14 22:21:46 -0700993 struct flowi4 fl4;
994 struct rtable *rt;
995
David S. Miller4895c772012-07-17 04:19:00 -0700996 __build_flow_key(&fl4, NULL, iph, oif,
997 RT_TOS(iph->tos), protocol, mark, flow_flags);
David S. Miller36393392012-06-14 22:21:46 -0700998 rt = __ip_route_output_key(net, &fl4);
999 if (!IS_ERR(rt)) {
David S. Miller4895c772012-07-17 04:19:00 -07001000 __ip_rt_update_pmtu(rt, &fl4, mtu);
David S. Miller36393392012-06-14 22:21:46 -07001001 ip_rt_put(rt);
1002 }
1003}
1004EXPORT_SYMBOL_GPL(ipv4_update_pmtu);
1005
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001006static void __ipv4_sk_update_pmtu(struct sk_buff *skb, struct sock *sk, u32 mtu)
David S. Miller36393392012-06-14 22:21:46 -07001007{
David S. Miller4895c772012-07-17 04:19:00 -07001008 const struct iphdr *iph = (const struct iphdr *) skb->data;
1009 struct flowi4 fl4;
1010 struct rtable *rt;
David S. Miller36393392012-06-14 22:21:46 -07001011
David S. Miller4895c772012-07-17 04:19:00 -07001012 __build_flow_key(&fl4, sk, iph, 0, 0, 0, 0, 0);
1013 rt = __ip_route_output_key(sock_net(sk), &fl4);
1014 if (!IS_ERR(rt)) {
1015 __ip_rt_update_pmtu(rt, &fl4, mtu);
1016 ip_rt_put(rt);
1017 }
David S. Miller36393392012-06-14 22:21:46 -07001018}
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001019
1020void ipv4_sk_update_pmtu(struct sk_buff *skb, struct sock *sk, u32 mtu)
1021{
1022 const struct iphdr *iph = (const struct iphdr *) skb->data;
1023 struct flowi4 fl4;
1024 struct rtable *rt;
1025 struct dst_entry *dst;
Steffen Klassertb44108d2013-01-22 00:01:28 +00001026 bool new = false;
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001027
1028 bh_lock_sock(sk);
Hannes Frederic Sowa482fc602013-11-05 02:24:17 +01001029
1030 if (!ip_sk_accept_pmtu(sk))
1031 goto out;
1032
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001033 rt = (struct rtable *) __sk_dst_get(sk);
1034
1035 if (sock_owned_by_user(sk) || !rt) {
1036 __ipv4_sk_update_pmtu(skb, sk, mtu);
1037 goto out;
1038 }
1039
1040 __build_flow_key(&fl4, sk, iph, 0, 0, 0, 0, 0);
1041
1042 if (!__sk_dst_check(sk, 0)) {
1043 rt = ip_route_output_flow(sock_net(sk), &fl4, sk);
1044 if (IS_ERR(rt))
1045 goto out;
Steffen Klassertb44108d2013-01-22 00:01:28 +00001046
1047 new = true;
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001048 }
1049
1050 __ip_rt_update_pmtu((struct rtable *) rt->dst.path, &fl4, mtu);
1051
1052 dst = dst_check(&rt->dst, 0);
1053 if (!dst) {
Steffen Klassertb44108d2013-01-22 00:01:28 +00001054 if (new)
1055 dst_release(&rt->dst);
1056
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001057 rt = ip_route_output_flow(sock_net(sk), &fl4, sk);
1058 if (IS_ERR(rt))
1059 goto out;
1060
Steffen Klassertb44108d2013-01-22 00:01:28 +00001061 new = true;
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001062 }
1063
Steffen Klassertb44108d2013-01-22 00:01:28 +00001064 if (new)
1065 __sk_dst_set(sk, &rt->dst);
Steffen Klassert9cb3a502013-01-21 01:59:11 +00001066
1067out:
1068 bh_unlock_sock(sk);
1069}
David S. Miller36393392012-06-14 22:21:46 -07001070EXPORT_SYMBOL_GPL(ipv4_sk_update_pmtu);
David S. Millerf39925d2011-02-09 22:00:16 -08001071
David S. Millerb42597e2012-07-11 21:25:45 -07001072void ipv4_redirect(struct sk_buff *skb, struct net *net,
1073 int oif, u32 mark, u8 protocol, int flow_flags)
1074{
David S. Miller4895c772012-07-17 04:19:00 -07001075 const struct iphdr *iph = (const struct iphdr *) skb->data;
David S. Millerb42597e2012-07-11 21:25:45 -07001076 struct flowi4 fl4;
1077 struct rtable *rt;
1078
David S. Miller4895c772012-07-17 04:19:00 -07001079 __build_flow_key(&fl4, NULL, iph, oif,
1080 RT_TOS(iph->tos), protocol, mark, flow_flags);
David S. Millerb42597e2012-07-11 21:25:45 -07001081 rt = __ip_route_output_key(net, &fl4);
1082 if (!IS_ERR(rt)) {
David S. Millerceb33202012-07-17 11:31:28 -07001083 __ip_do_redirect(rt, skb, &fl4, false);
David S. Millerb42597e2012-07-11 21:25:45 -07001084 ip_rt_put(rt);
1085 }
1086}
1087EXPORT_SYMBOL_GPL(ipv4_redirect);
1088
1089void ipv4_sk_redirect(struct sk_buff *skb, struct sock *sk)
1090{
David S. Miller4895c772012-07-17 04:19:00 -07001091 const struct iphdr *iph = (const struct iphdr *) skb->data;
1092 struct flowi4 fl4;
1093 struct rtable *rt;
David S. Millerb42597e2012-07-11 21:25:45 -07001094
David S. Miller4895c772012-07-17 04:19:00 -07001095 __build_flow_key(&fl4, sk, iph, 0, 0, 0, 0, 0);
1096 rt = __ip_route_output_key(sock_net(sk), &fl4);
1097 if (!IS_ERR(rt)) {
David S. Millerceb33202012-07-17 11:31:28 -07001098 __ip_do_redirect(rt, skb, &fl4, false);
David S. Miller4895c772012-07-17 04:19:00 -07001099 ip_rt_put(rt);
1100 }
David S. Millerb42597e2012-07-11 21:25:45 -07001101}
1102EXPORT_SYMBOL_GPL(ipv4_sk_redirect);
1103
David S. Millerefbc3682011-12-01 13:38:59 -05001104static struct dst_entry *ipv4_dst_check(struct dst_entry *dst, u32 cookie)
1105{
1106 struct rtable *rt = (struct rtable *) dst;
1107
David S. Millerceb33202012-07-17 11:31:28 -07001108 /* All IPV4 dsts are created with ->obsolete set to the value
1109 * DST_OBSOLETE_FORCE_CHK which forces validation calls down
1110 * into this function always.
1111 *
Timo Teräs387aa652013-05-27 20:46:31 +00001112 * When a PMTU/redirect information update invalidates a route,
1113 * this is indicated by setting obsolete to DST_OBSOLETE_KILL or
1114 * DST_OBSOLETE_DEAD by dst_free().
David S. Millerceb33202012-07-17 11:31:28 -07001115 */
Timo Teräs387aa652013-05-27 20:46:31 +00001116 if (dst->obsolete != DST_OBSOLETE_FORCE_CHK || rt_is_expired(rt))
David S. Millerefbc3682011-12-01 13:38:59 -05001117 return NULL;
Timo Teräsd11a4dc2010-03-18 23:20:20 +00001118 return dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119}
1120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121static void ipv4_link_failure(struct sk_buff *skb)
1122{
1123 struct rtable *rt;
1124
1125 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_HOST_UNREACH, 0);
1126
Eric Dumazet511c3f92009-06-02 05:14:27 +00001127 rt = skb_rtable(skb);
David S. Miller59436342012-07-10 06:58:42 -07001128 if (rt)
1129 dst_set_expires(&rt->dst, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130}
1131
1132static int ip_rt_bug(struct sk_buff *skb)
1133{
Joe Perches91df42b2012-05-15 14:11:54 +00001134 pr_debug("%s: %pI4 -> %pI4, %s\n",
1135 __func__, &ip_hdr(skb)->saddr, &ip_hdr(skb)->daddr,
1136 skb->dev ? skb->dev->name : "?");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 kfree_skb(skb);
Dave Jonesc378a9c2011-05-21 07:16:42 +00001138 WARN_ON(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 return 0;
1140}
1141
1142/*
1143 We do not cache source address of outgoing interface,
1144 because it is used only by IP RR, TS and SRR options,
1145 so that it out of fast path.
1146
1147 BTW remember: "addr" is allowed to be not aligned
1148 in IP options!
1149 */
1150
David S. Miller8e363602011-05-13 17:29:41 -04001151void ip_rt_get_source(u8 *addr, struct sk_buff *skb, struct rtable *rt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152{
Al Viroa61ced52006-09-26 21:27:54 -07001153 __be32 src;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154
David S. Millerc7537962010-11-11 17:07:48 -08001155 if (rt_is_output_route(rt))
David S. Millerc5be24f2011-05-13 18:01:21 -04001156 src = ip_hdr(skb)->saddr;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001157 else {
David S. Miller8e363602011-05-13 17:29:41 -04001158 struct fib_result res;
1159 struct flowi4 fl4;
1160 struct iphdr *iph;
1161
1162 iph = ip_hdr(skb);
1163
1164 memset(&fl4, 0, sizeof(fl4));
1165 fl4.daddr = iph->daddr;
1166 fl4.saddr = iph->saddr;
Julian Anastasovb0fe4a32011-07-23 02:00:41 +00001167 fl4.flowi4_tos = RT_TOS(iph->tos);
David S. Miller8e363602011-05-13 17:29:41 -04001168 fl4.flowi4_oif = rt->dst.dev->ifindex;
1169 fl4.flowi4_iif = skb->dev->ifindex;
1170 fl4.flowi4_mark = skb->mark;
David S. Miller5e2b61f2011-03-04 21:47:09 -08001171
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001172 rcu_read_lock();
David S. Miller68a5e3d2011-03-11 20:07:33 -05001173 if (fib_lookup(dev_net(rt->dst.dev), &fl4, &res) == 0)
David S. Miller436c3b62011-03-24 17:42:21 -07001174 src = FIB_RES_PREFSRC(dev_net(rt->dst.dev), res);
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001175 else
David S. Millerf8126f12012-07-13 05:03:45 -07001176 src = inet_select_addr(rt->dst.dev,
1177 rt_nexthop(rt, iph->daddr),
1178 RT_SCOPE_UNIVERSE);
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001179 rcu_read_unlock();
1180 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 memcpy(addr, &src, 4);
1182}
1183
Patrick McHardyc7066f72011-01-14 13:36:42 +01001184#ifdef CONFIG_IP_ROUTE_CLASSID
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185static void set_class_tag(struct rtable *rt, u32 tag)
1186{
Changli Gaod8d1f302010-06-10 23:31:35 -07001187 if (!(rt->dst.tclassid & 0xFFFF))
1188 rt->dst.tclassid |= tag & 0xFFFF;
1189 if (!(rt->dst.tclassid & 0xFFFF0000))
1190 rt->dst.tclassid |= tag & 0xFFFF0000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191}
1192#endif
1193
David S. Miller0dbaee32010-12-13 12:52:14 -08001194static unsigned int ipv4_default_advmss(const struct dst_entry *dst)
1195{
1196 unsigned int advmss = dst_metric_raw(dst, RTAX_ADVMSS);
1197
1198 if (advmss == 0) {
1199 advmss = max_t(unsigned int, dst->dev->mtu - 40,
1200 ip_rt_min_advmss);
1201 if (advmss > 65535 - 40)
1202 advmss = 65535 - 40;
1203 }
1204 return advmss;
1205}
1206
Steffen Klassertebb762f2011-11-23 02:12:51 +00001207static unsigned int ipv4_mtu(const struct dst_entry *dst)
David S. Millerd33e4552010-12-14 13:01:14 -08001208{
Steffen Klassert261663b2011-11-23 02:14:50 +00001209 const struct rtable *rt = (const struct rtable *) dst;
David S. Miller59436342012-07-10 06:58:42 -07001210 unsigned int mtu = rt->rt_pmtu;
1211
Alexander Duyck98d75c32012-08-27 06:30:01 +00001212 if (!mtu || time_after_eq(jiffies, rt->dst.expires))
David S. Miller59436342012-07-10 06:58:42 -07001213 mtu = dst_metric_raw(dst, RTAX_MTU);
Steffen Klassert618f9bc2011-11-23 02:13:31 +00001214
Steffen Klassert38d523e2013-01-16 20:55:01 +00001215 if (mtu)
Steffen Klassert618f9bc2011-11-23 02:13:31 +00001216 return mtu;
1217
1218 mtu = dst->dev->mtu;
David S. Millerd33e4552010-12-14 13:01:14 -08001219
1220 if (unlikely(dst_metric_locked(dst, RTAX_MTU))) {
Julian Anastasov155e8332012-10-08 11:41:18 +00001221 if (rt->rt_uses_gateway && mtu > 576)
David S. Millerd33e4552010-12-14 13:01:14 -08001222 mtu = 576;
1223 }
1224
Eric Dumazet734d2722013-08-18 19:08:07 -07001225 return min_t(unsigned int, mtu, IP_MAX_MTU);
David S. Millerd33e4552010-12-14 13:01:14 -08001226}
1227
David S. Millerf2bb4be2012-07-17 12:20:47 -07001228static struct fib_nh_exception *find_exception(struct fib_nh *nh, __be32 daddr)
David S. Miller4895c772012-07-17 04:19:00 -07001229{
1230 struct fnhe_hash_bucket *hash = nh->nh_exceptions;
1231 struct fib_nh_exception *fnhe;
1232 u32 hval;
1233
David S. Millerf2bb4be2012-07-17 12:20:47 -07001234 if (!hash)
1235 return NULL;
1236
David S. Millerd3a25c92012-07-17 13:23:08 -07001237 hval = fnhe_hashfun(daddr);
David S. Miller4895c772012-07-17 04:19:00 -07001238
1239 for (fnhe = rcu_dereference(hash[hval].chain); fnhe;
1240 fnhe = rcu_dereference(fnhe->fnhe_next)) {
David S. Millerf2bb4be2012-07-17 12:20:47 -07001241 if (fnhe->fnhe_daddr == daddr)
1242 return fnhe;
1243 }
1244 return NULL;
1245}
David S. Miller4895c772012-07-17 04:19:00 -07001246
David S. Millercaacf052012-07-31 15:06:50 -07001247static bool rt_bind_exception(struct rtable *rt, struct fib_nh_exception *fnhe,
David S. Millerf2bb4be2012-07-17 12:20:47 -07001248 __be32 daddr)
1249{
David S. Millercaacf052012-07-31 15:06:50 -07001250 bool ret = false;
1251
David S. Millerc5038a82012-07-31 15:02:02 -07001252 spin_lock_bh(&fnhe_lock);
Julian Anastasovaee06da2012-07-18 10:15:35 +00001253
David S. Millerc5038a82012-07-31 15:02:02 -07001254 if (daddr == fnhe->fnhe_daddr) {
Timo Teräs2ffae992013-06-27 10:27:05 +03001255 struct rtable __rcu **porig;
1256 struct rtable *orig;
Timo Teräs5aad1de2013-05-27 20:46:33 +00001257 int genid = fnhe_genid(dev_net(rt->dst.dev));
Timo Teräs2ffae992013-06-27 10:27:05 +03001258
1259 if (rt_is_input_route(rt))
1260 porig = &fnhe->fnhe_rth_input;
1261 else
1262 porig = &fnhe->fnhe_rth_output;
1263 orig = rcu_dereference(*porig);
Timo Teräs5aad1de2013-05-27 20:46:33 +00001264
1265 if (fnhe->fnhe_genid != genid) {
1266 fnhe->fnhe_genid = genid;
Steffen Klassert13d82bf2012-10-17 21:17:44 +00001267 fnhe->fnhe_gw = 0;
1268 fnhe->fnhe_pmtu = 0;
1269 fnhe->fnhe_expires = 0;
Timo Teräs2ffae992013-06-27 10:27:05 +03001270 fnhe_flush_routes(fnhe);
1271 orig = NULL;
Steffen Klassert13d82bf2012-10-17 21:17:44 +00001272 }
Timo Teräs387aa652013-05-27 20:46:31 +00001273 fill_route_from_fnhe(rt, fnhe);
1274 if (!rt->rt_gateway)
Julian Anastasov155e8332012-10-08 11:41:18 +00001275 rt->rt_gateway = daddr;
David S. Millerf2bb4be2012-07-17 12:20:47 -07001276
Timo Teräs2ffae992013-06-27 10:27:05 +03001277 if (!(rt->dst.flags & DST_NOCACHE)) {
1278 rcu_assign_pointer(*porig, rt);
1279 if (orig)
1280 rt_free(orig);
1281 ret = true;
1282 }
David S. Millerc5038a82012-07-31 15:02:02 -07001283
1284 fnhe->fnhe_stamp = jiffies;
David S. Millerc5038a82012-07-31 15:02:02 -07001285 }
1286 spin_unlock_bh(&fnhe_lock);
David S. Millercaacf052012-07-31 15:06:50 -07001287
1288 return ret;
Eric Dumazet54764bb2012-07-31 01:08:23 +00001289}
1290
David S. Millercaacf052012-07-31 15:06:50 -07001291static bool rt_cache_route(struct fib_nh *nh, struct rtable *rt)
David S. Millerf2bb4be2012-07-17 12:20:47 -07001292{
Eric Dumazetd26b3a72012-07-31 05:45:30 +00001293 struct rtable *orig, *prev, **p;
David S. Millercaacf052012-07-31 15:06:50 -07001294 bool ret = true;
David S. Millerf2bb4be2012-07-17 12:20:47 -07001295
Eric Dumazetd26b3a72012-07-31 05:45:30 +00001296 if (rt_is_input_route(rt)) {
Eric Dumazet54764bb2012-07-31 01:08:23 +00001297 p = (struct rtable **)&nh->nh_rth_input;
Eric Dumazetd26b3a72012-07-31 05:45:30 +00001298 } else {
Eric Dumazetd26b3a72012-07-31 05:45:30 +00001299 p = (struct rtable **)__this_cpu_ptr(nh->nh_pcpu_rth_output);
1300 }
David S. Millerf2bb4be2012-07-17 12:20:47 -07001301 orig = *p;
1302
1303 prev = cmpxchg(p, orig, rt);
1304 if (prev == orig) {
David S. Millerf2bb4be2012-07-17 12:20:47 -07001305 if (orig)
Eric Dumazet54764bb2012-07-31 01:08:23 +00001306 rt_free(orig);
Julian Anastasov155e8332012-10-08 11:41:18 +00001307 } else
David S. Millercaacf052012-07-31 15:06:50 -07001308 ret = false;
David S. Millercaacf052012-07-31 15:06:50 -07001309
1310 return ret;
1311}
1312
1313static DEFINE_SPINLOCK(rt_uncached_lock);
1314static LIST_HEAD(rt_uncached_list);
1315
1316static void rt_add_uncached_list(struct rtable *rt)
1317{
1318 spin_lock_bh(&rt_uncached_lock);
1319 list_add_tail(&rt->rt_uncached, &rt_uncached_list);
1320 spin_unlock_bh(&rt_uncached_lock);
1321}
1322
1323static void ipv4_dst_destroy(struct dst_entry *dst)
1324{
1325 struct rtable *rt = (struct rtable *) dst;
1326
Eric Dumazet78df76a2012-08-24 05:40:47 +00001327 if (!list_empty(&rt->rt_uncached)) {
David S. Millercaacf052012-07-31 15:06:50 -07001328 spin_lock_bh(&rt_uncached_lock);
1329 list_del(&rt->rt_uncached);
1330 spin_unlock_bh(&rt_uncached_lock);
1331 }
1332}
1333
1334void rt_flush_dev(struct net_device *dev)
1335{
1336 if (!list_empty(&rt_uncached_list)) {
1337 struct net *net = dev_net(dev);
1338 struct rtable *rt;
1339
1340 spin_lock_bh(&rt_uncached_lock);
1341 list_for_each_entry(rt, &rt_uncached_list, rt_uncached) {
1342 if (rt->dst.dev != dev)
1343 continue;
1344 rt->dst.dev = net->loopback_dev;
1345 dev_hold(rt->dst.dev);
1346 dev_put(dev);
1347 }
1348 spin_unlock_bh(&rt_uncached_lock);
David S. Miller4895c772012-07-17 04:19:00 -07001349 }
1350}
1351
Eric Dumazet4331deb2012-07-25 05:11:23 +00001352static bool rt_cache_valid(const struct rtable *rt)
David S. Millerd2d68ba2012-07-17 12:58:50 -07001353{
Eric Dumazet4331deb2012-07-25 05:11:23 +00001354 return rt &&
1355 rt->dst.obsolete == DST_OBSOLETE_FORCE_CHK &&
1356 !rt_is_expired(rt);
David S. Millerd2d68ba2012-07-17 12:58:50 -07001357}
1358
David S. Millerf2bb4be2012-07-17 12:20:47 -07001359static void rt_set_nexthop(struct rtable *rt, __be32 daddr,
David S. Miller5e2b61f2011-03-04 21:47:09 -08001360 const struct fib_result *res,
David S. Millerf2bb4be2012-07-17 12:20:47 -07001361 struct fib_nh_exception *fnhe,
David S. Miller982721f2011-02-16 21:44:24 -08001362 struct fib_info *fi, u16 type, u32 itag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363{
David S. Millercaacf052012-07-31 15:06:50 -07001364 bool cached = false;
1365
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 if (fi) {
David S. Miller4895c772012-07-17 04:19:00 -07001367 struct fib_nh *nh = &FIB_RES_NH(*res);
1368
Julian Anastasov155e8332012-10-08 11:41:18 +00001369 if (nh->nh_gw && nh->nh_scope == RT_SCOPE_LINK) {
David S. Miller4895c772012-07-17 04:19:00 -07001370 rt->rt_gateway = nh->nh_gw;
Julian Anastasov155e8332012-10-08 11:41:18 +00001371 rt->rt_uses_gateway = 1;
1372 }
David S. Miller28605832012-07-17 14:55:59 -07001373 dst_init_metrics(&rt->dst, fi->fib_metrics, true);
Patrick McHardyc7066f72011-01-14 13:36:42 +01001374#ifdef CONFIG_IP_ROUTE_CLASSID
David S. Millerf2bb4be2012-07-17 12:20:47 -07001375 rt->dst.tclassid = nh->nh_tclassid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376#endif
David S. Millerc5038a82012-07-31 15:02:02 -07001377 if (unlikely(fnhe))
David S. Millercaacf052012-07-31 15:06:50 -07001378 cached = rt_bind_exception(rt, fnhe, daddr);
David S. Millerc5038a82012-07-31 15:02:02 -07001379 else if (!(rt->dst.flags & DST_NOCACHE))
David S. Millercaacf052012-07-31 15:06:50 -07001380 cached = rt_cache_route(nh, rt);
Julian Anastasov155e8332012-10-08 11:41:18 +00001381 if (unlikely(!cached)) {
1382 /* Routes we intend to cache in nexthop exception or
1383 * FIB nexthop have the DST_NOCACHE bit clear.
1384 * However, if we are unsuccessful at storing this
1385 * route into the cache we really need to set it.
1386 */
1387 rt->dst.flags |= DST_NOCACHE;
1388 if (!rt->rt_gateway)
1389 rt->rt_gateway = daddr;
1390 rt_add_uncached_list(rt);
1391 }
1392 } else
David S. Millercaacf052012-07-31 15:06:50 -07001393 rt_add_uncached_list(rt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
Patrick McHardyc7066f72011-01-14 13:36:42 +01001395#ifdef CONFIG_IP_ROUTE_CLASSID
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396#ifdef CONFIG_IP_MULTIPLE_TABLES
David S. Miller85b91b02012-07-13 08:21:29 -07001397 set_class_tag(rt, res->tclassid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398#endif
1399 set_class_tag(rt, itag);
1400#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401}
1402
David S. Miller5c1e6aa2011-04-28 14:13:38 -07001403static struct rtable *rt_dst_alloc(struct net_device *dev,
David S. Millerf2bb4be2012-07-17 12:20:47 -07001404 bool nopolicy, bool noxfrm, bool will_cache)
David S. Miller0c4dcd52011-02-17 15:42:37 -08001405{
David S. Millerf5b0a872012-07-19 12:31:33 -07001406 return dst_alloc(&ipv4_dst_ops, dev, 1, DST_OBSOLETE_FORCE_CHK,
David S. Millerc6cffba2012-07-26 11:14:38 +00001407 (will_cache ? 0 : (DST_HOST | DST_NOCACHE)) |
David S. Miller5c1e6aa2011-04-28 14:13:38 -07001408 (nopolicy ? DST_NOPOLICY : 0) |
1409 (noxfrm ? DST_NOXFRM : 0));
David S. Miller0c4dcd52011-02-17 15:42:37 -08001410}
1411
Eric Dumazet96d36222010-06-02 19:21:31 +00001412/* called in rcu_read_lock() section */
Al Viro9e12bb22006-09-26 21:25:20 -07001413static int ip_route_input_mc(struct sk_buff *skb, __be32 daddr, __be32 saddr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 u8 tos, struct net_device *dev, int our)
1415{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416 struct rtable *rth;
Eric Dumazet96d36222010-06-02 19:21:31 +00001417 struct in_device *in_dev = __in_dev_get_rcu(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 u32 itag = 0;
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001419 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420
1421 /* Primary sanity checks. */
1422
1423 if (in_dev == NULL)
1424 return -EINVAL;
1425
Jan Engelhardt1e637c72008-01-21 03:18:08 -08001426 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
Thomas Grafd0daebc32012-06-12 00:44:01 +00001427 skb->protocol != htons(ETH_P_IP))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 goto e_inval;
1429
Thomas Grafd0daebc32012-06-12 00:44:01 +00001430 if (likely(!IN_DEV_ROUTE_LOCALNET(in_dev)))
1431 if (ipv4_is_loopback(saddr))
1432 goto e_inval;
1433
Joe Perchesf97c1e02007-12-16 13:45:43 -08001434 if (ipv4_is_zeronet(saddr)) {
1435 if (!ipv4_is_local_multicast(daddr))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 goto e_inval;
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001437 } else {
David S. Miller9e56e382012-06-28 18:54:02 -07001438 err = fib_validate_source(skb, saddr, 0, tos, 0, dev,
1439 in_dev, &itag);
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001440 if (err < 0)
1441 goto e_err;
1442 }
Benjamin LaHaise4e7b2f12012-03-27 15:55:32 +00001443 rth = rt_dst_alloc(dev_net(dev)->loopback_dev,
David S. Millerf2bb4be2012-07-17 12:20:47 -07001444 IN_DEV_CONF_GET(in_dev, NOPOLICY), false, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 if (!rth)
1446 goto e_nobufs;
1447
Patrick McHardyc7066f72011-01-14 13:36:42 +01001448#ifdef CONFIG_IP_ROUTE_CLASSID
Changli Gaod8d1f302010-06-10 23:31:35 -07001449 rth->dst.tclassid = itag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450#endif
David S. Millercf911662011-04-28 14:31:47 -07001451 rth->dst.output = ip_rt_bug;
1452
fan.duca4c3fc2013-07-30 08:33:53 +08001453 rth->rt_genid = rt_genid_ipv4(dev_net(dev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 rth->rt_flags = RTCF_MULTICAST;
Eric Dumazet29e75252008-01-31 17:05:09 -08001455 rth->rt_type = RTN_MULTICAST;
David S. Miller9917e1e82012-07-17 14:44:26 -07001456 rth->rt_is_input= 1;
David S. Miller13378ca2012-07-23 13:57:45 -07001457 rth->rt_iif = 0;
David S. Miller59436342012-07-10 06:58:42 -07001458 rth->rt_pmtu = 0;
David S. Millerf8126f12012-07-13 05:03:45 -07001459 rth->rt_gateway = 0;
Julian Anastasov155e8332012-10-08 11:41:18 +00001460 rth->rt_uses_gateway = 0;
David S. Millercaacf052012-07-31 15:06:50 -07001461 INIT_LIST_HEAD(&rth->rt_uncached);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 if (our) {
Changli Gaod8d1f302010-06-10 23:31:35 -07001463 rth->dst.input= ip_local_deliver;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 rth->rt_flags |= RTCF_LOCAL;
1465 }
1466
1467#ifdef CONFIG_IP_MROUTE
Joe Perchesf97c1e02007-12-16 13:45:43 -08001468 if (!ipv4_is_local_multicast(daddr) && IN_DEV_MFORWARD(in_dev))
Changli Gaod8d1f302010-06-10 23:31:35 -07001469 rth->dst.input = ip_mr_input;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470#endif
1471 RT_CACHE_STAT_INC(in_slow_mc);
1472
David S. Miller89aef892012-07-17 11:00:09 -07001473 skb_dst_set(skb, &rth->dst);
1474 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475
1476e_nobufs:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 return -ENOBUFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478e_inval:
Eric Dumazet96d36222010-06-02 19:21:31 +00001479 return -EINVAL;
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001480e_err:
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001481 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482}
1483
1484
1485static void ip_handle_martian_source(struct net_device *dev,
1486 struct in_device *in_dev,
1487 struct sk_buff *skb,
Al Viro9e12bb22006-09-26 21:25:20 -07001488 __be32 daddr,
1489 __be32 saddr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490{
1491 RT_CACHE_STAT_INC(in_martian_src);
1492#ifdef CONFIG_IP_ROUTE_VERBOSE
1493 if (IN_DEV_LOG_MARTIANS(in_dev) && net_ratelimit()) {
1494 /*
1495 * RFC1812 recommendation, if source is martian,
1496 * the only hint is MAC header.
1497 */
Joe Perches058bd4d2012-03-11 18:36:11 +00001498 pr_warn("martian source %pI4 from %pI4, on dev %s\n",
Harvey Harrison673d57e2008-10-31 00:53:57 -07001499 &daddr, &saddr, dev->name);
Arnaldo Carvalho de Melo98e399f2007-03-19 15:33:04 -07001500 if (dev->hard_header_len && skb_mac_header_was_set(skb)) {
Joe Perches058bd4d2012-03-11 18:36:11 +00001501 print_hex_dump(KERN_WARNING, "ll header: ",
1502 DUMP_PREFIX_OFFSET, 16, 1,
1503 skb_mac_header(skb),
1504 dev->hard_header_len, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 }
1506 }
1507#endif
1508}
1509
Eric Dumazet47360222010-06-03 04:13:21 +00001510/* called in rcu_read_lock() section */
Stephen Hemminger5969f712008-04-10 01:52:09 -07001511static int __mkroute_input(struct sk_buff *skb,
David S. Miller982721f2011-02-16 21:44:24 -08001512 const struct fib_result *res,
Stephen Hemminger5969f712008-04-10 01:52:09 -07001513 struct in_device *in_dev,
David S. Millerc6cffba2012-07-26 11:14:38 +00001514 __be32 daddr, __be32 saddr, u32 tos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515{
Timo Teräs2ffae992013-06-27 10:27:05 +03001516 struct fib_nh_exception *fnhe;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 struct rtable *rth;
1518 int err;
1519 struct in_device *out_dev;
Eric Dumazet47360222010-06-03 04:13:21 +00001520 unsigned int flags = 0;
David S. Millerd2d68ba2012-07-17 12:58:50 -07001521 bool do_cache;
Al Virod9c9df82006-09-26 21:28:14 -07001522 u32 itag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523
1524 /* get a working reference to the output device */
Eric Dumazet47360222010-06-03 04:13:21 +00001525 out_dev = __in_dev_get_rcu(FIB_RES_DEV(*res));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 if (out_dev == NULL) {
Joe Perchese87cc472012-05-13 21:56:26 +00001527 net_crit_ratelimited("Bug in ip_route_input_slow(). Please report.\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 return -EINVAL;
1529 }
1530
Michael Smith5c04c812011-04-07 04:51:50 +00001531 err = fib_validate_source(skb, saddr, daddr, tos, FIB_RES_OIF(*res),
David S. Miller9e56e382012-06-28 18:54:02 -07001532 in_dev->dev, in_dev, &itag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 if (err < 0) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001534 ip_handle_martian_source(in_dev->dev, in_dev, skb, daddr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535 saddr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001536
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 goto cleanup;
1538 }
1539
Julian Anastasove81da0e2012-10-08 11:41:15 +00001540 do_cache = res->fi && !itag;
1541 if (out_dev == in_dev && err && IN_DEV_TX_REDIRECTS(out_dev) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 (IN_DEV_SHARED_MEDIA(out_dev) ||
Julian Anastasove81da0e2012-10-08 11:41:15 +00001543 inet_addr_onlink(out_dev, saddr, FIB_RES_GW(*res)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 flags |= RTCF_DOREDIRECT;
Julian Anastasove81da0e2012-10-08 11:41:15 +00001545 do_cache = false;
1546 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547
1548 if (skb->protocol != htons(ETH_P_IP)) {
1549 /* Not IP (i.e. ARP). Do not create route, if it is
1550 * invalid for proxy arp. DNAT routes are always valid.
Jesper Dangaard Brouer65324142010-01-05 05:50:47 +00001551 *
1552 * Proxy arp feature have been extended to allow, ARP
1553 * replies back to the same interface, to support
1554 * Private VLAN switch technologies. See arp.c.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 */
Jesper Dangaard Brouer65324142010-01-05 05:50:47 +00001556 if (out_dev == in_dev &&
1557 IN_DEV_PROXY_ARP_PVLAN(in_dev) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 err = -EINVAL;
1559 goto cleanup;
1560 }
1561 }
1562
Timo Teräs2ffae992013-06-27 10:27:05 +03001563 fnhe = find_exception(&FIB_RES_NH(*res), daddr);
Julian Anastasove81da0e2012-10-08 11:41:15 +00001564 if (do_cache) {
Timo Teräs2ffae992013-06-27 10:27:05 +03001565 if (fnhe != NULL)
1566 rth = rcu_dereference(fnhe->fnhe_rth_input);
1567 else
1568 rth = rcu_dereference(FIB_RES_NH(*res).nh_rth_input);
1569
Julian Anastasove81da0e2012-10-08 11:41:15 +00001570 if (rt_cache_valid(rth)) {
1571 skb_dst_set_noref(skb, &rth->dst);
1572 goto out;
David S. Millerd2d68ba2012-07-17 12:58:50 -07001573 }
1574 }
David S. Millerf2bb4be2012-07-17 12:20:47 -07001575
David S. Miller5c1e6aa2011-04-28 14:13:38 -07001576 rth = rt_dst_alloc(out_dev->dev,
1577 IN_DEV_CONF_GET(in_dev, NOPOLICY),
David S. Millerd2d68ba2012-07-17 12:58:50 -07001578 IN_DEV_CONF_GET(out_dev, NOXFRM), do_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 if (!rth) {
1580 err = -ENOBUFS;
1581 goto cleanup;
1582 }
1583
fan.duca4c3fc2013-07-30 08:33:53 +08001584 rth->rt_genid = rt_genid_ipv4(dev_net(rth->dst.dev));
David S. Millercf911662011-04-28 14:31:47 -07001585 rth->rt_flags = flags;
1586 rth->rt_type = res->type;
David S. Miller9917e1e82012-07-17 14:44:26 -07001587 rth->rt_is_input = 1;
David S. Miller13378ca2012-07-23 13:57:45 -07001588 rth->rt_iif = 0;
David S. Miller59436342012-07-10 06:58:42 -07001589 rth->rt_pmtu = 0;
David S. Millerf8126f12012-07-13 05:03:45 -07001590 rth->rt_gateway = 0;
Julian Anastasov155e8332012-10-08 11:41:18 +00001591 rth->rt_uses_gateway = 0;
David S. Millercaacf052012-07-31 15:06:50 -07001592 INIT_LIST_HEAD(&rth->rt_uncached);
Duan Jionga6254862014-02-17 15:23:43 +08001593 RT_CACHE_STAT_INC(in_slow_tot);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594
Changli Gaod8d1f302010-06-10 23:31:35 -07001595 rth->dst.input = ip_forward;
1596 rth->dst.output = ip_output;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597
Timo Teräs2ffae992013-06-27 10:27:05 +03001598 rt_set_nexthop(rth, daddr, res, fnhe, res->fi, res->type, itag);
David S. Millerc6cffba2012-07-26 11:14:38 +00001599 skb_dst_set(skb, &rth->dst);
David S. Millerd2d68ba2012-07-17 12:58:50 -07001600out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 err = 0;
1602 cleanup:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 return err;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001604}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605
Stephen Hemminger5969f712008-04-10 01:52:09 -07001606static int ip_mkroute_input(struct sk_buff *skb,
1607 struct fib_result *res,
David S. Miller68a5e3d2011-03-11 20:07:33 -05001608 const struct flowi4 *fl4,
Stephen Hemminger5969f712008-04-10 01:52:09 -07001609 struct in_device *in_dev,
1610 __be32 daddr, __be32 saddr, u32 tos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612#ifdef CONFIG_IP_ROUTE_MULTIPATH
David S. Millerff3fccb2011-03-10 16:23:24 -08001613 if (res->fi && res->fi->fib_nhs > 1)
David S. Miller1b7fe5932011-03-10 17:01:16 -08001614 fib_select_multipath(res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615#endif
1616
1617 /* create a routing cache entry */
David S. Millerc6cffba2012-07-26 11:14:38 +00001618 return __mkroute_input(skb, res, in_dev, daddr, saddr, tos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619}
1620
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621/*
1622 * NOTE. We drop all the packets that has local source
1623 * addresses, because every properly looped back packet
1624 * must have correct destination already attached by output routine.
1625 *
1626 * Such approach solves two big problems:
1627 * 1. Not simplex devices are handled properly.
1628 * 2. IP spoofing attempts are filtered with 100% of guarantee.
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001629 * called with rcu_read_lock()
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 */
1631
Al Viro9e12bb22006-09-26 21:25:20 -07001632static int ip_route_input_slow(struct sk_buff *skb, __be32 daddr, __be32 saddr,
David S. Millerc10237e2012-06-27 17:05:06 -07001633 u8 tos, struct net_device *dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634{
1635 struct fib_result res;
Eric Dumazet96d36222010-06-02 19:21:31 +00001636 struct in_device *in_dev = __in_dev_get_rcu(dev);
David S. Miller68a5e3d2011-03-11 20:07:33 -05001637 struct flowi4 fl4;
Eric Dumazet95c96172012-04-15 05:58:06 +00001638 unsigned int flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 u32 itag = 0;
Eric Dumazet95c96172012-04-15 05:58:06 +00001640 struct rtable *rth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 int err = -EINVAL;
Daniel Baluta5e73ea12012-04-15 01:34:41 +00001642 struct net *net = dev_net(dev);
David S. Millerd2d68ba2012-07-17 12:58:50 -07001643 bool do_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644
1645 /* IP on this device is disabled. */
1646
1647 if (!in_dev)
1648 goto out;
1649
1650 /* Check for the most weird martians, which can be not detected
1651 by fib_lookup.
1652 */
1653
Thomas Grafd0daebc32012-06-12 00:44:01 +00001654 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 goto martian_source;
1656
David S. Millerd2d68ba2012-07-17 12:58:50 -07001657 res.fi = NULL;
Andy Walls27a954b2010-10-17 15:11:22 +00001658 if (ipv4_is_lbcast(daddr) || (saddr == 0 && daddr == 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 goto brd_input;
1660
1661 /* Accept zero addresses only to limited broadcast;
1662 * I even do not know to fix it or not. Waiting for complains :-)
1663 */
Joe Perchesf97c1e02007-12-16 13:45:43 -08001664 if (ipv4_is_zeronet(saddr))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 goto martian_source;
1666
Thomas Grafd0daebc32012-06-12 00:44:01 +00001667 if (ipv4_is_zeronet(daddr))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 goto martian_destination;
1669
Eric Dumazet9eb43e72012-08-03 21:27:25 +00001670 /* Following code try to avoid calling IN_DEV_NET_ROUTE_LOCALNET(),
1671 * and call it once if daddr or/and saddr are loopback addresses
1672 */
1673 if (ipv4_is_loopback(daddr)) {
1674 if (!IN_DEV_NET_ROUTE_LOCALNET(in_dev, net))
Thomas Grafd0daebc32012-06-12 00:44:01 +00001675 goto martian_destination;
Eric Dumazet9eb43e72012-08-03 21:27:25 +00001676 } else if (ipv4_is_loopback(saddr)) {
1677 if (!IN_DEV_NET_ROUTE_LOCALNET(in_dev, net))
Thomas Grafd0daebc32012-06-12 00:44:01 +00001678 goto martian_source;
1679 }
1680
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 /*
1682 * Now we are ready to route packet.
1683 */
David S. Miller68a5e3d2011-03-11 20:07:33 -05001684 fl4.flowi4_oif = 0;
1685 fl4.flowi4_iif = dev->ifindex;
1686 fl4.flowi4_mark = skb->mark;
1687 fl4.flowi4_tos = tos;
1688 fl4.flowi4_scope = RT_SCOPE_UNIVERSE;
1689 fl4.daddr = daddr;
1690 fl4.saddr = saddr;
1691 err = fib_lookup(net, &fl4, &res);
Duan Jiongcd0f0b92014-02-14 18:26:22 +08001692 if (err != 0) {
1693 if (!IN_DEV_FORWARD(in_dev))
1694 err = -EHOSTUNREACH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 goto no_route;
Duan Jiongcd0f0b92014-02-14 18:26:22 +08001696 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697
1698 if (res.type == RTN_BROADCAST)
1699 goto brd_input;
1700
1701 if (res.type == RTN_LOCAL) {
Michael Smith5c04c812011-04-07 04:51:50 +00001702 err = fib_validate_source(skb, saddr, daddr, tos,
Pavel Emelyanov1fb94892012-08-08 21:53:36 +00001703 LOOPBACK_IFINDEX,
David S. Miller9e56e382012-06-28 18:54:02 -07001704 dev, in_dev, &itag);
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001705 if (err < 0)
1706 goto martian_source_keep_err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 goto local_input;
1708 }
1709
Duan Jiongcd0f0b92014-02-14 18:26:22 +08001710 if (!IN_DEV_FORWARD(in_dev)) {
1711 err = -EHOSTUNREACH;
David S. Miller251da412012-06-26 16:27:09 -07001712 goto no_route;
Duan Jiongcd0f0b92014-02-14 18:26:22 +08001713 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 if (res.type != RTN_UNICAST)
1715 goto martian_destination;
1716
David S. Miller68a5e3d2011-03-11 20:07:33 -05001717 err = ip_mkroute_input(skb, &res, &fl4, in_dev, daddr, saddr, tos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718out: return err;
1719
1720brd_input:
1721 if (skb->protocol != htons(ETH_P_IP))
1722 goto e_inval;
1723
David S. Miller41347dc2012-06-28 04:05:27 -07001724 if (!ipv4_is_zeronet(saddr)) {
David S. Miller9e56e382012-06-28 18:54:02 -07001725 err = fib_validate_source(skb, saddr, 0, tos, 0, dev,
1726 in_dev, &itag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 if (err < 0)
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001728 goto martian_source_keep_err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 }
1730 flags |= RTCF_BROADCAST;
1731 res.type = RTN_BROADCAST;
1732 RT_CACHE_STAT_INC(in_brd);
1733
1734local_input:
David S. Millerd2d68ba2012-07-17 12:58:50 -07001735 do_cache = false;
1736 if (res.fi) {
David S. Millerfe3edf42012-07-23 13:22:20 -07001737 if (!itag) {
Eric Dumazet54764bb2012-07-31 01:08:23 +00001738 rth = rcu_dereference(FIB_RES_NH(res).nh_rth_input);
David S. Millerd2d68ba2012-07-17 12:58:50 -07001739 if (rt_cache_valid(rth)) {
David S. Millerc6cffba2012-07-26 11:14:38 +00001740 skb_dst_set_noref(skb, &rth->dst);
1741 err = 0;
1742 goto out;
David S. Millerd2d68ba2012-07-17 12:58:50 -07001743 }
1744 do_cache = true;
1745 }
1746 }
1747
David S. Miller5c1e6aa2011-04-28 14:13:38 -07001748 rth = rt_dst_alloc(net->loopback_dev,
David S. Millerd2d68ba2012-07-17 12:58:50 -07001749 IN_DEV_CONF_GET(in_dev, NOPOLICY), false, do_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 if (!rth)
1751 goto e_nobufs;
1752
David S. Millercf911662011-04-28 14:31:47 -07001753 rth->dst.input= ip_local_deliver;
Changli Gaod8d1f302010-06-10 23:31:35 -07001754 rth->dst.output= ip_rt_bug;
David S. Millercf911662011-04-28 14:31:47 -07001755#ifdef CONFIG_IP_ROUTE_CLASSID
1756 rth->dst.tclassid = itag;
1757#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758
fan.duca4c3fc2013-07-30 08:33:53 +08001759 rth->rt_genid = rt_genid_ipv4(net);
David S. Millercf911662011-04-28 14:31:47 -07001760 rth->rt_flags = flags|RTCF_LOCAL;
1761 rth->rt_type = res.type;
David S. Miller9917e1e82012-07-17 14:44:26 -07001762 rth->rt_is_input = 1;
David S. Miller13378ca2012-07-23 13:57:45 -07001763 rth->rt_iif = 0;
David S. Miller59436342012-07-10 06:58:42 -07001764 rth->rt_pmtu = 0;
David S. Millerf8126f12012-07-13 05:03:45 -07001765 rth->rt_gateway = 0;
Julian Anastasov155e8332012-10-08 11:41:18 +00001766 rth->rt_uses_gateway = 0;
David S. Millercaacf052012-07-31 15:06:50 -07001767 INIT_LIST_HEAD(&rth->rt_uncached);
Duan Jionga6254862014-02-17 15:23:43 +08001768 RT_CACHE_STAT_INC(in_slow_tot);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 if (res.type == RTN_UNREACHABLE) {
Changli Gaod8d1f302010-06-10 23:31:35 -07001770 rth->dst.input= ip_error;
1771 rth->dst.error= -err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 rth->rt_flags &= ~RTCF_LOCAL;
1773 }
Alexei Starovoitovdcdfdf52013-11-19 19:12:34 -08001774 if (do_cache) {
1775 if (unlikely(!rt_cache_route(&FIB_RES_NH(res), rth))) {
1776 rth->dst.flags |= DST_NOCACHE;
1777 rt_add_uncached_list(rth);
1778 }
1779 }
David S. Miller89aef892012-07-17 11:00:09 -07001780 skb_dst_set(skb, &rth->dst);
David S. Millerb23dd4f2011-03-02 14:31:35 -08001781 err = 0;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001782 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783
1784no_route:
1785 RT_CACHE_STAT_INC(in_no_route);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 res.type = RTN_UNREACHABLE;
Mitsuru Chinen7f538782007-12-07 01:07:24 -08001787 if (err == -ESRCH)
1788 err = -ENETUNREACH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 goto local_input;
1790
1791 /*
1792 * Do not cache martian addresses: they should be logged (RFC1812)
1793 */
1794martian_destination:
1795 RT_CACHE_STAT_INC(in_martian_dst);
1796#ifdef CONFIG_IP_ROUTE_VERBOSE
Joe Perchese87cc472012-05-13 21:56:26 +00001797 if (IN_DEV_LOG_MARTIANS(in_dev))
1798 net_warn_ratelimited("martian destination %pI4 from %pI4, dev %s\n",
1799 &daddr, &saddr, dev->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800#endif
Dietmar Eggemann2c2910a2005-06-28 13:06:23 -07001801
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802e_inval:
1803 err = -EINVAL;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001804 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805
1806e_nobufs:
1807 err = -ENOBUFS;
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001808 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809
1810martian_source:
Eric Dumazetb5f7e752010-06-02 12:05:27 +00001811 err = -EINVAL;
1812martian_source_keep_err:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 ip_handle_martian_source(dev, in_dev, skb, daddr, saddr);
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001814 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815}
1816
David S. Millerc6cffba2012-07-26 11:14:38 +00001817int ip_route_input_noref(struct sk_buff *skb, __be32 daddr, __be32 saddr,
1818 u8 tos, struct net_device *dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819{
Eric Dumazet96d36222010-06-02 19:21:31 +00001820 int res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821
Eric Dumazet96d36222010-06-02 19:21:31 +00001822 rcu_read_lock();
1823
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 /* Multicast recognition logic is moved from route cache to here.
1825 The problem was that too many Ethernet cards have broken/missing
1826 hardware multicast filters :-( As result the host on multicasting
1827 network acquires a lot of useless route cache entries, sort of
1828 SDR messages from all the world. Now we try to get rid of them.
1829 Really, provided software IP multicast filter is organized
1830 reasonably (at least, hashed), it does not result in a slowdown
1831 comparing with route cache reject entries.
1832 Note, that multicast routers are not affected, because
1833 route cache entry is created eventually.
1834 */
Joe Perchesf97c1e02007-12-16 13:45:43 -08001835 if (ipv4_is_multicast(daddr)) {
Eric Dumazet96d36222010-06-02 19:21:31 +00001836 struct in_device *in_dev = __in_dev_get_rcu(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837
Eric Dumazet96d36222010-06-02 19:21:31 +00001838 if (in_dev) {
David S. Millerdbdd9a52011-03-10 16:34:38 -08001839 int our = ip_check_mc_rcu(in_dev, daddr, saddr,
1840 ip_hdr(skb)->protocol);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841 if (our
1842#ifdef CONFIG_IP_MROUTE
Joe Perches9d4fb272009-11-23 10:41:23 -08001843 ||
1844 (!ipv4_is_local_multicast(daddr) &&
1845 IN_DEV_MFORWARD(in_dev))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846#endif
Joe Perches9d4fb272009-11-23 10:41:23 -08001847 ) {
Eric Dumazet96d36222010-06-02 19:21:31 +00001848 int res = ip_route_input_mc(skb, daddr, saddr,
1849 tos, dev, our);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 rcu_read_unlock();
Eric Dumazet96d36222010-06-02 19:21:31 +00001851 return res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 }
1853 }
1854 rcu_read_unlock();
1855 return -EINVAL;
1856 }
David S. Millerc10237e2012-06-27 17:05:06 -07001857 res = ip_route_input_slow(skb, daddr, saddr, tos, dev);
Eric Dumazet96d36222010-06-02 19:21:31 +00001858 rcu_read_unlock();
1859 return res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860}
David S. Millerc6cffba2012-07-26 11:14:38 +00001861EXPORT_SYMBOL(ip_route_input_noref);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862
Eric Dumazetebc0ffa2010-10-05 10:41:36 +00001863/* called with rcu_read_lock() */
David S. Miller982721f2011-02-16 21:44:24 -08001864static struct rtable *__mkroute_output(const struct fib_result *res,
David Miller1a00fee2012-07-01 02:02:56 +00001865 const struct flowi4 *fl4, int orig_oif,
Julian Anastasovf61759e2011-12-02 11:39:42 +00001866 struct net_device *dev_out,
David S. Miller5ada5522011-02-17 15:29:00 -08001867 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868{
David S. Miller982721f2011-02-16 21:44:24 -08001869 struct fib_info *fi = res->fi;
David S. Millerf2bb4be2012-07-17 12:20:47 -07001870 struct fib_nh_exception *fnhe;
David S. Miller5ada5522011-02-17 15:29:00 -08001871 struct in_device *in_dev;
David S. Miller982721f2011-02-16 21:44:24 -08001872 u16 type = res->type;
David S. Miller5ada5522011-02-17 15:29:00 -08001873 struct rtable *rth;
Julian Anastasovc92b9652012-10-08 11:41:19 +00001874 bool do_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875
Thomas Grafd0daebc32012-06-12 00:44:01 +00001876 in_dev = __in_dev_get_rcu(dev_out);
1877 if (!in_dev)
David S. Miller5ada5522011-02-17 15:29:00 -08001878 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879
Thomas Grafd0daebc32012-06-12 00:44:01 +00001880 if (likely(!IN_DEV_ROUTE_LOCALNET(in_dev)))
1881 if (ipv4_is_loopback(fl4->saddr) && !(dev_out->flags & IFF_LOOPBACK))
1882 return ERR_PTR(-EINVAL);
1883
David S. Miller68a5e3d2011-03-11 20:07:33 -05001884 if (ipv4_is_lbcast(fl4->daddr))
David S. Miller982721f2011-02-16 21:44:24 -08001885 type = RTN_BROADCAST;
David S. Miller68a5e3d2011-03-11 20:07:33 -05001886 else if (ipv4_is_multicast(fl4->daddr))
David S. Miller982721f2011-02-16 21:44:24 -08001887 type = RTN_MULTICAST;
David S. Miller68a5e3d2011-03-11 20:07:33 -05001888 else if (ipv4_is_zeronet(fl4->daddr))
David S. Miller5ada5522011-02-17 15:29:00 -08001889 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890
1891 if (dev_out->flags & IFF_LOOPBACK)
1892 flags |= RTCF_LOCAL;
1893
Julian Anastasov63617422012-11-22 23:04:14 +02001894 do_cache = true;
David S. Miller982721f2011-02-16 21:44:24 -08001895 if (type == RTN_BROADCAST) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 flags |= RTCF_BROADCAST | RTCF_LOCAL;
David S. Miller982721f2011-02-16 21:44:24 -08001897 fi = NULL;
1898 } else if (type == RTN_MULTICAST) {
Eric Dumazetdd28d1a2010-09-29 11:53:50 +00001899 flags |= RTCF_MULTICAST | RTCF_LOCAL;
David S. Miller813b3b52011-04-28 14:48:42 -07001900 if (!ip_check_mc_rcu(in_dev, fl4->daddr, fl4->saddr,
1901 fl4->flowi4_proto))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 flags &= ~RTCF_LOCAL;
Julian Anastasov63617422012-11-22 23:04:14 +02001903 else
1904 do_cache = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 /* If multicast route do not exist use
Eric Dumazetdd28d1a2010-09-29 11:53:50 +00001906 * default one, but do not gateway in this case.
1907 * Yes, it is hack.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 */
David S. Miller982721f2011-02-16 21:44:24 -08001909 if (fi && res->prefixlen < 4)
1910 fi = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 }
1912
David S. Millerf2bb4be2012-07-17 12:20:47 -07001913 fnhe = NULL;
Julian Anastasov63617422012-11-22 23:04:14 +02001914 do_cache &= fi != NULL;
1915 if (do_cache) {
David S. Millerc5038a82012-07-31 15:02:02 -07001916 struct rtable __rcu **prth;
Julian Anastasovc92b9652012-10-08 11:41:19 +00001917 struct fib_nh *nh = &FIB_RES_NH(*res);
Eric Dumazetd26b3a72012-07-31 05:45:30 +00001918
Julian Anastasovc92b9652012-10-08 11:41:19 +00001919 fnhe = find_exception(nh, fl4->daddr);
David S. Millerc5038a82012-07-31 15:02:02 -07001920 if (fnhe)
Timo Teräs2ffae992013-06-27 10:27:05 +03001921 prth = &fnhe->fnhe_rth_output;
Julian Anastasovc92b9652012-10-08 11:41:19 +00001922 else {
1923 if (unlikely(fl4->flowi4_flags &
1924 FLOWI_FLAG_KNOWN_NH &&
1925 !(nh->nh_gw &&
1926 nh->nh_scope == RT_SCOPE_LINK))) {
1927 do_cache = false;
1928 goto add;
1929 }
1930 prth = __this_cpu_ptr(nh->nh_pcpu_rth_output);
1931 }
David S. Millerc5038a82012-07-31 15:02:02 -07001932 rth = rcu_dereference(*prth);
1933 if (rt_cache_valid(rth)) {
1934 dst_hold(&rth->dst);
1935 return rth;
David S. Millerf2bb4be2012-07-17 12:20:47 -07001936 }
1937 }
Julian Anastasovc92b9652012-10-08 11:41:19 +00001938
1939add:
David S. Miller5c1e6aa2011-04-28 14:13:38 -07001940 rth = rt_dst_alloc(dev_out,
1941 IN_DEV_CONF_GET(in_dev, NOPOLICY),
David S. Millerf2bb4be2012-07-17 12:20:47 -07001942 IN_DEV_CONF_GET(in_dev, NOXFRM),
Julian Anastasovc92b9652012-10-08 11:41:19 +00001943 do_cache);
Dimitris Michailidis8391d072010-10-07 14:48:38 +00001944 if (!rth)
David S. Miller5ada5522011-02-17 15:29:00 -08001945 return ERR_PTR(-ENOBUFS);
Dimitris Michailidis8391d072010-10-07 14:48:38 +00001946
David S. Millercf911662011-04-28 14:31:47 -07001947 rth->dst.output = ip_output;
1948
fan.duca4c3fc2013-07-30 08:33:53 +08001949 rth->rt_genid = rt_genid_ipv4(dev_net(dev_out));
David S. Millercf911662011-04-28 14:31:47 -07001950 rth->rt_flags = flags;
1951 rth->rt_type = type;
David S. Miller9917e1e82012-07-17 14:44:26 -07001952 rth->rt_is_input = 0;
David S. Miller13378ca2012-07-23 13:57:45 -07001953 rth->rt_iif = orig_oif ? : 0;
David S. Miller59436342012-07-10 06:58:42 -07001954 rth->rt_pmtu = 0;
David S. Millerf8126f12012-07-13 05:03:45 -07001955 rth->rt_gateway = 0;
Julian Anastasov155e8332012-10-08 11:41:18 +00001956 rth->rt_uses_gateway = 0;
David S. Millercaacf052012-07-31 15:06:50 -07001957 INIT_LIST_HEAD(&rth->rt_uncached);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958
1959 RT_CACHE_STAT_INC(out_slow_tot);
1960
David S. Miller41347dc2012-06-28 04:05:27 -07001961 if (flags & RTCF_LOCAL)
Changli Gaod8d1f302010-06-10 23:31:35 -07001962 rth->dst.input = ip_local_deliver;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 if (flags & (RTCF_BROADCAST | RTCF_MULTICAST)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001964 if (flags & RTCF_LOCAL &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 !(dev_out->flags & IFF_LOOPBACK)) {
Changli Gaod8d1f302010-06-10 23:31:35 -07001966 rth->dst.output = ip_mc_output;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967 RT_CACHE_STAT_INC(out_slow_mc);
1968 }
1969#ifdef CONFIG_IP_MROUTE
David S. Miller982721f2011-02-16 21:44:24 -08001970 if (type == RTN_MULTICAST) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 if (IN_DEV_MFORWARD(in_dev) &&
David S. Miller813b3b52011-04-28 14:48:42 -07001972 !ipv4_is_local_multicast(fl4->daddr)) {
Changli Gaod8d1f302010-06-10 23:31:35 -07001973 rth->dst.input = ip_mr_input;
1974 rth->dst.output = ip_mc_output;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 }
1976 }
1977#endif
1978 }
1979
David S. Millerf2bb4be2012-07-17 12:20:47 -07001980 rt_set_nexthop(rth, fl4->daddr, res, fnhe, fi, type, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981
David S. Miller5ada5522011-02-17 15:29:00 -08001982 return rth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983}
1984
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985/*
1986 * Major route resolver routine.
1987 */
1988
David S. Miller89aef892012-07-17 11:00:09 -07001989struct rtable *__ip_route_output_key(struct net *net, struct flowi4 *fl4)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 struct net_device *dev_out = NULL;
Julian Anastasovf61759e2011-12-02 11:39:42 +00001992 __u8 tos = RT_FL_TOS(fl4);
David S. Miller813b3b52011-04-28 14:48:42 -07001993 unsigned int flags = 0;
1994 struct fib_result res;
David S. Miller5ada5522011-02-17 15:29:00 -08001995 struct rtable *rth;
David S. Miller813b3b52011-04-28 14:48:42 -07001996 int orig_oif;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997
David S. Miller85b91b02012-07-13 08:21:29 -07001998 res.tclassid = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 res.fi = NULL;
David S. Miller8b96d222012-06-11 02:01:56 -07002000 res.table = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001
David S. Miller813b3b52011-04-28 14:48:42 -07002002 orig_oif = fl4->flowi4_oif;
2003
Pavel Emelyanov1fb94892012-08-08 21:53:36 +00002004 fl4->flowi4_iif = LOOPBACK_IFINDEX;
David S. Miller813b3b52011-04-28 14:48:42 -07002005 fl4->flowi4_tos = tos & IPTOS_RT_MASK;
2006 fl4->flowi4_scope = ((tos & RTO_ONLINK) ?
2007 RT_SCOPE_LINK : RT_SCOPE_UNIVERSE);
David S. Miller44713b62011-03-04 21:24:47 -08002008
David S. Miller010c2702011-02-17 15:37:09 -08002009 rcu_read_lock();
David S. Miller813b3b52011-04-28 14:48:42 -07002010 if (fl4->saddr) {
David S. Millerb23dd4f2011-03-02 14:31:35 -08002011 rth = ERR_PTR(-EINVAL);
David S. Miller813b3b52011-04-28 14:48:42 -07002012 if (ipv4_is_multicast(fl4->saddr) ||
2013 ipv4_is_lbcast(fl4->saddr) ||
2014 ipv4_is_zeronet(fl4->saddr))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 goto out;
2016
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017 /* I removed check for oif == dev_out->oif here.
2018 It was wrong for two reasons:
Denis V. Lunev1ab35272008-01-22 22:04:30 -08002019 1. ip_dev_find(net, saddr) can return wrong iface, if saddr
2020 is assigned to multiple interfaces.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 2. Moreover, we are allowed to send packets with saddr
2022 of another iface. --ANK
2023 */
2024
David S. Miller813b3b52011-04-28 14:48:42 -07002025 if (fl4->flowi4_oif == 0 &&
2026 (ipv4_is_multicast(fl4->daddr) ||
2027 ipv4_is_lbcast(fl4->daddr))) {
Julian Anastasova210d012008-10-01 07:28:28 -07002028 /* It is equivalent to inet_addr_type(saddr) == RTN_LOCAL */
David S. Miller813b3b52011-04-28 14:48:42 -07002029 dev_out = __ip_dev_find(net, fl4->saddr, false);
Julian Anastasova210d012008-10-01 07:28:28 -07002030 if (dev_out == NULL)
2031 goto out;
2032
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 /* Special hack: user can direct multicasts
2034 and limited broadcast via necessary interface
2035 without fiddling with IP_MULTICAST_IF or IP_PKTINFO.
2036 This hack is not just for fun, it allows
2037 vic,vat and friends to work.
2038 They bind socket to loopback, set ttl to zero
2039 and expect that it will work.
2040 From the viewpoint of routing cache they are broken,
2041 because we are not allowed to build multicast path
2042 with loopback source addr (look, routing cache
2043 cannot know, that ttl is zero, so that packet
2044 will not leave this host and route is valid).
2045 Luckily, this hack is good workaround.
2046 */
2047
David S. Miller813b3b52011-04-28 14:48:42 -07002048 fl4->flowi4_oif = dev_out->ifindex;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 goto make_route;
2050 }
Julian Anastasova210d012008-10-01 07:28:28 -07002051
David S. Miller813b3b52011-04-28 14:48:42 -07002052 if (!(fl4->flowi4_flags & FLOWI_FLAG_ANYSRC)) {
Julian Anastasova210d012008-10-01 07:28:28 -07002053 /* It is equivalent to inet_addr_type(saddr) == RTN_LOCAL */
David S. Miller813b3b52011-04-28 14:48:42 -07002054 if (!__ip_dev_find(net, fl4->saddr, false))
Julian Anastasova210d012008-10-01 07:28:28 -07002055 goto out;
Julian Anastasova210d012008-10-01 07:28:28 -07002056 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 }
2058
2059
David S. Miller813b3b52011-04-28 14:48:42 -07002060 if (fl4->flowi4_oif) {
2061 dev_out = dev_get_by_index_rcu(net, fl4->flowi4_oif);
David S. Millerb23dd4f2011-03-02 14:31:35 -08002062 rth = ERR_PTR(-ENODEV);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 if (dev_out == NULL)
2064 goto out;
Herbert Xue5ed6392005-10-03 14:35:55 -07002065
2066 /* RACE: Check return value of inet_select_addr instead. */
Eric Dumazetfc75fc82010-12-22 04:39:39 +00002067 if (!(dev_out->flags & IFF_UP) || !__in_dev_get_rcu(dev_out)) {
David S. Millerb23dd4f2011-03-02 14:31:35 -08002068 rth = ERR_PTR(-ENETUNREACH);
Eric Dumazetfc75fc82010-12-22 04:39:39 +00002069 goto out;
2070 }
David S. Miller813b3b52011-04-28 14:48:42 -07002071 if (ipv4_is_local_multicast(fl4->daddr) ||
2072 ipv4_is_lbcast(fl4->daddr)) {
2073 if (!fl4->saddr)
2074 fl4->saddr = inet_select_addr(dev_out, 0,
2075 RT_SCOPE_LINK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 goto make_route;
2077 }
Jiri Benc0a7e2262013-10-04 17:04:48 +02002078 if (!fl4->saddr) {
David S. Miller813b3b52011-04-28 14:48:42 -07002079 if (ipv4_is_multicast(fl4->daddr))
2080 fl4->saddr = inet_select_addr(dev_out, 0,
2081 fl4->flowi4_scope);
2082 else if (!fl4->daddr)
2083 fl4->saddr = inet_select_addr(dev_out, 0,
2084 RT_SCOPE_HOST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 }
2086 }
2087
David S. Miller813b3b52011-04-28 14:48:42 -07002088 if (!fl4->daddr) {
2089 fl4->daddr = fl4->saddr;
2090 if (!fl4->daddr)
2091 fl4->daddr = fl4->saddr = htonl(INADDR_LOOPBACK);
Denis V. Lunevb40afd02008-01-22 22:06:19 -08002092 dev_out = net->loopback_dev;
Pavel Emelyanov1fb94892012-08-08 21:53:36 +00002093 fl4->flowi4_oif = LOOPBACK_IFINDEX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 res.type = RTN_LOCAL;
2095 flags |= RTCF_LOCAL;
2096 goto make_route;
2097 }
2098
David S. Miller813b3b52011-04-28 14:48:42 -07002099 if (fib_lookup(net, fl4, &res)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 res.fi = NULL;
David S. Miller8b96d222012-06-11 02:01:56 -07002101 res.table = NULL;
David S. Miller813b3b52011-04-28 14:48:42 -07002102 if (fl4->flowi4_oif) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 /* Apparently, routing tables are wrong. Assume,
2104 that the destination is on link.
2105
2106 WHY? DW.
2107 Because we are allowed to send to iface
2108 even if it has NO routes and NO assigned
2109 addresses. When oif is specified, routing
2110 tables are looked up with only one purpose:
2111 to catch if destination is gatewayed, rather than
2112 direct. Moreover, if MSG_DONTROUTE is set,
2113 we send packet, ignoring both routing tables
2114 and ifaddr state. --ANK
2115
2116
2117 We could make it even if oif is unknown,
2118 likely IPv6, but we do not.
2119 */
2120
David S. Miller813b3b52011-04-28 14:48:42 -07002121 if (fl4->saddr == 0)
2122 fl4->saddr = inet_select_addr(dev_out, 0,
2123 RT_SCOPE_LINK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 res.type = RTN_UNICAST;
2125 goto make_route;
2126 }
David S. Millerb23dd4f2011-03-02 14:31:35 -08002127 rth = ERR_PTR(-ENETUNREACH);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 goto out;
2129 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130
2131 if (res.type == RTN_LOCAL) {
David S. Miller813b3b52011-04-28 14:48:42 -07002132 if (!fl4->saddr) {
Joel Sing9fc3bbb2011-01-03 20:24:20 +00002133 if (res.fi->fib_prefsrc)
David S. Miller813b3b52011-04-28 14:48:42 -07002134 fl4->saddr = res.fi->fib_prefsrc;
Joel Sing9fc3bbb2011-01-03 20:24:20 +00002135 else
David S. Miller813b3b52011-04-28 14:48:42 -07002136 fl4->saddr = fl4->daddr;
Joel Sing9fc3bbb2011-01-03 20:24:20 +00002137 }
Denis V. Lunevb40afd02008-01-22 22:06:19 -08002138 dev_out = net->loopback_dev;
David S. Miller813b3b52011-04-28 14:48:42 -07002139 fl4->flowi4_oif = dev_out->ifindex;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 flags |= RTCF_LOCAL;
2141 goto make_route;
2142 }
2143
2144#ifdef CONFIG_IP_ROUTE_MULTIPATH
David S. Miller813b3b52011-04-28 14:48:42 -07002145 if (res.fi->fib_nhs > 1 && fl4->flowi4_oif == 0)
David S. Miller1b7fe5932011-03-10 17:01:16 -08002146 fib_select_multipath(&res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 else
2148#endif
David S. Miller21d8c492011-04-14 14:49:37 -07002149 if (!res.prefixlen &&
2150 res.table->tb_num_default > 1 &&
David S. Miller813b3b52011-04-28 14:48:42 -07002151 res.type == RTN_UNICAST && !fl4->flowi4_oif)
David S. Miller0c838ff2011-01-31 16:16:50 -08002152 fib_select_default(&res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153
David S. Miller813b3b52011-04-28 14:48:42 -07002154 if (!fl4->saddr)
2155 fl4->saddr = FIB_RES_PREFSRC(net, res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 dev_out = FIB_RES_DEV(res);
David S. Miller813b3b52011-04-28 14:48:42 -07002158 fl4->flowi4_oif = dev_out->ifindex;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159
2160
2161make_route:
David Miller1a00fee2012-07-01 02:02:56 +00002162 rth = __mkroute_output(&res, fl4, orig_oif, dev_out, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163
David S. Miller010c2702011-02-17 15:37:09 -08002164out:
2165 rcu_read_unlock();
David S. Millerb23dd4f2011-03-02 14:31:35 -08002166 return rth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167}
Arnaldo Carvalho de Melod8c97a92005-08-09 20:12:12 -07002168EXPORT_SYMBOL_GPL(__ip_route_output_key);
2169
Jianzhao Wangae2688d2010-09-08 14:35:43 -07002170static struct dst_entry *ipv4_blackhole_dst_check(struct dst_entry *dst, u32 cookie)
2171{
2172 return NULL;
2173}
2174
Steffen Klassertebb762f2011-11-23 02:12:51 +00002175static unsigned int ipv4_blackhole_mtu(const struct dst_entry *dst)
Roland Dreierec831ea2011-01-31 13:16:00 -08002176{
Steffen Klassert618f9bc2011-11-23 02:13:31 +00002177 unsigned int mtu = dst_metric_raw(dst, RTAX_MTU);
2178
2179 return mtu ? : dst->dev->mtu;
Roland Dreierec831ea2011-01-31 13:16:00 -08002180}
2181
David S. Miller6700c272012-07-17 03:29:28 -07002182static void ipv4_rt_blackhole_update_pmtu(struct dst_entry *dst, struct sock *sk,
2183 struct sk_buff *skb, u32 mtu)
David S. Miller14e50e52007-05-24 18:17:54 -07002184{
2185}
2186
David S. Miller6700c272012-07-17 03:29:28 -07002187static void ipv4_rt_blackhole_redirect(struct dst_entry *dst, struct sock *sk,
2188 struct sk_buff *skb)
David S. Millerb587ee32012-07-12 00:39:24 -07002189{
2190}
2191
Held Bernhard0972ddb2011-04-24 22:07:32 +00002192static u32 *ipv4_rt_blackhole_cow_metrics(struct dst_entry *dst,
2193 unsigned long old)
2194{
2195 return NULL;
2196}
2197
David S. Miller14e50e52007-05-24 18:17:54 -07002198static struct dst_ops ipv4_dst_blackhole_ops = {
2199 .family = AF_INET,
Harvey Harrison09640e62009-02-01 00:45:17 -08002200 .protocol = cpu_to_be16(ETH_P_IP),
Jianzhao Wangae2688d2010-09-08 14:35:43 -07002201 .check = ipv4_blackhole_dst_check,
Steffen Klassertebb762f2011-11-23 02:12:51 +00002202 .mtu = ipv4_blackhole_mtu,
Eric Dumazet214f45c2011-02-18 11:39:01 -08002203 .default_advmss = ipv4_default_advmss,
David S. Miller14e50e52007-05-24 18:17:54 -07002204 .update_pmtu = ipv4_rt_blackhole_update_pmtu,
David S. Millerb587ee32012-07-12 00:39:24 -07002205 .redirect = ipv4_rt_blackhole_redirect,
Held Bernhard0972ddb2011-04-24 22:07:32 +00002206 .cow_metrics = ipv4_rt_blackhole_cow_metrics,
David S. Millerd3aaeb32011-07-18 00:40:17 -07002207 .neigh_lookup = ipv4_neigh_lookup,
David S. Miller14e50e52007-05-24 18:17:54 -07002208};
2209
David S. Miller2774c132011-03-01 14:59:04 -08002210struct dst_entry *ipv4_blackhole_route(struct net *net, struct dst_entry *dst_orig)
David S. Miller14e50e52007-05-24 18:17:54 -07002211{
David S. Miller2774c132011-03-01 14:59:04 -08002212 struct rtable *ort = (struct rtable *) dst_orig;
David S. Millerf5b0a872012-07-19 12:31:33 -07002213 struct rtable *rt;
David S. Miller14e50e52007-05-24 18:17:54 -07002214
David S. Millerf5b0a872012-07-19 12:31:33 -07002215 rt = dst_alloc(&ipv4_dst_blackhole_ops, NULL, 1, DST_OBSOLETE_NONE, 0);
David S. Miller14e50e52007-05-24 18:17:54 -07002216 if (rt) {
Changli Gaod8d1f302010-06-10 23:31:35 -07002217 struct dst_entry *new = &rt->dst;
David S. Miller14e50e52007-05-24 18:17:54 -07002218
David S. Miller14e50e52007-05-24 18:17:54 -07002219 new->__use = 1;
Herbert Xu352e5122007-11-13 21:34:06 -08002220 new->input = dst_discard;
2221 new->output = dst_discard;
David S. Miller14e50e52007-05-24 18:17:54 -07002222
Changli Gaod8d1f302010-06-10 23:31:35 -07002223 new->dev = ort->dst.dev;
David S. Miller14e50e52007-05-24 18:17:54 -07002224 if (new->dev)
2225 dev_hold(new->dev);
2226
David S. Miller9917e1e82012-07-17 14:44:26 -07002227 rt->rt_is_input = ort->rt_is_input;
David S. Miller5e2b61f2011-03-04 21:47:09 -08002228 rt->rt_iif = ort->rt_iif;
David S. Miller59436342012-07-10 06:58:42 -07002229 rt->rt_pmtu = ort->rt_pmtu;
David S. Miller14e50e52007-05-24 18:17:54 -07002230
fan.duca4c3fc2013-07-30 08:33:53 +08002231 rt->rt_genid = rt_genid_ipv4(net);
David S. Miller14e50e52007-05-24 18:17:54 -07002232 rt->rt_flags = ort->rt_flags;
2233 rt->rt_type = ort->rt_type;
David S. Miller14e50e52007-05-24 18:17:54 -07002234 rt->rt_gateway = ort->rt_gateway;
Julian Anastasov155e8332012-10-08 11:41:18 +00002235 rt->rt_uses_gateway = ort->rt_uses_gateway;
David S. Miller14e50e52007-05-24 18:17:54 -07002236
David S. Millercaacf052012-07-31 15:06:50 -07002237 INIT_LIST_HEAD(&rt->rt_uncached);
2238
David S. Miller14e50e52007-05-24 18:17:54 -07002239 dst_free(new);
2240 }
2241
David S. Miller2774c132011-03-01 14:59:04 -08002242 dst_release(dst_orig);
2243
2244 return rt ? &rt->dst : ERR_PTR(-ENOMEM);
David S. Miller14e50e52007-05-24 18:17:54 -07002245}
2246
David S. Miller9d6ec932011-03-12 01:12:47 -05002247struct rtable *ip_route_output_flow(struct net *net, struct flowi4 *flp4,
David S. Millerb23dd4f2011-03-02 14:31:35 -08002248 struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249{
David S. Miller9d6ec932011-03-12 01:12:47 -05002250 struct rtable *rt = __ip_route_output_key(net, flp4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251
David S. Millerb23dd4f2011-03-02 14:31:35 -08002252 if (IS_ERR(rt))
2253 return rt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254
David S. Miller56157872011-05-02 14:37:45 -07002255 if (flp4->flowi4_proto)
David S. Miller9d6ec932011-03-12 01:12:47 -05002256 rt = (struct rtable *) xfrm_lookup(net, &rt->dst,
2257 flowi4_to_flowi(flp4),
2258 sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
David S. Millerb23dd4f2011-03-02 14:31:35 -08002260 return rt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261}
Arnaldo Carvalho de Melod8c97a92005-08-09 20:12:12 -07002262EXPORT_SYMBOL_GPL(ip_route_output_flow);
2263
David S. Millerf1ce3062012-07-12 10:10:17 -07002264static int rt_fill_info(struct net *net, __be32 dst, __be32 src,
Eric W. Biederman15e47302012-09-07 20:12:54 +00002265 struct flowi4 *fl4, struct sk_buff *skb, u32 portid,
David S. Millerf1ce3062012-07-12 10:10:17 -07002266 u32 seq, int event, int nowait, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267{
Eric Dumazet511c3f92009-06-02 05:14:27 +00002268 struct rtable *rt = skb_rtable(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 struct rtmsg *r;
Thomas Grafbe403ea2006-08-17 18:15:17 -07002270 struct nlmsghdr *nlh;
Steffen Klassert2bc8ca42011-10-11 01:12:02 +00002271 unsigned long expires = 0;
David S. Millerf1850712012-07-10 07:26:01 -07002272 u32 error;
Julian Anastasov521f5492012-07-20 12:02:08 +03002273 u32 metrics[RTAX_MAX];
Thomas Grafbe403ea2006-08-17 18:15:17 -07002274
Eric W. Biederman15e47302012-09-07 20:12:54 +00002275 nlh = nlmsg_put(skb, portid, seq, event, sizeof(*r), flags);
Thomas Grafbe403ea2006-08-17 18:15:17 -07002276 if (nlh == NULL)
Patrick McHardy26932562007-01-31 23:16:40 -08002277 return -EMSGSIZE;
Thomas Grafbe403ea2006-08-17 18:15:17 -07002278
2279 r = nlmsg_data(nlh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 r->rtm_family = AF_INET;
2281 r->rtm_dst_len = 32;
2282 r->rtm_src_len = 0;
David Millerd6c0a4f2012-07-01 02:02:59 +00002283 r->rtm_tos = fl4->flowi4_tos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 r->rtm_table = RT_TABLE_MAIN;
David S. Millerf3756b72012-04-01 20:39:02 -04002285 if (nla_put_u32(skb, RTA_TABLE, RT_TABLE_MAIN))
2286 goto nla_put_failure;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 r->rtm_type = rt->rt_type;
2288 r->rtm_scope = RT_SCOPE_UNIVERSE;
2289 r->rtm_protocol = RTPROT_UNSPEC;
2290 r->rtm_flags = (rt->rt_flags & ~0xFFFF) | RTM_F_CLONED;
2291 if (rt->rt_flags & RTCF_NOTIFY)
2292 r->rtm_flags |= RTM_F_NOTIFY;
Thomas Grafbe403ea2006-08-17 18:15:17 -07002293
David S. Millerf1ce3062012-07-12 10:10:17 -07002294 if (nla_put_be32(skb, RTA_DST, dst))
David S. Millerf3756b72012-04-01 20:39:02 -04002295 goto nla_put_failure;
David Miller1a00fee2012-07-01 02:02:56 +00002296 if (src) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 r->rtm_src_len = 32;
David Miller1a00fee2012-07-01 02:02:56 +00002298 if (nla_put_be32(skb, RTA_SRC, src))
David S. Millerf3756b72012-04-01 20:39:02 -04002299 goto nla_put_failure;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 }
David S. Millerf3756b72012-04-01 20:39:02 -04002301 if (rt->dst.dev &&
2302 nla_put_u32(skb, RTA_OIF, rt->dst.dev->ifindex))
2303 goto nla_put_failure;
Patrick McHardyc7066f72011-01-14 13:36:42 +01002304#ifdef CONFIG_IP_ROUTE_CLASSID
David S. Millerf3756b72012-04-01 20:39:02 -04002305 if (rt->dst.tclassid &&
2306 nla_put_u32(skb, RTA_FLOW, rt->dst.tclassid))
2307 goto nla_put_failure;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308#endif
David S. Miller41347dc2012-06-28 04:05:27 -07002309 if (!rt_is_input_route(rt) &&
David Millerd6c0a4f2012-07-01 02:02:59 +00002310 fl4->saddr != src) {
2311 if (nla_put_be32(skb, RTA_PREFSRC, fl4->saddr))
David S. Millerf3756b72012-04-01 20:39:02 -04002312 goto nla_put_failure;
2313 }
Julian Anastasov155e8332012-10-08 11:41:18 +00002314 if (rt->rt_uses_gateway &&
David S. Millerf3756b72012-04-01 20:39:02 -04002315 nla_put_be32(skb, RTA_GATEWAY, rt->rt_gateway))
2316 goto nla_put_failure;
Thomas Grafbe403ea2006-08-17 18:15:17 -07002317
Steffen Klassertee9a8f72012-10-08 00:56:54 +00002318 expires = rt->dst.expires;
2319 if (expires) {
2320 unsigned long now = jiffies;
2321
2322 if (time_before(now, expires))
2323 expires -= now;
2324 else
2325 expires = 0;
2326 }
2327
Julian Anastasov521f5492012-07-20 12:02:08 +03002328 memcpy(metrics, dst_metrics_ptr(&rt->dst), sizeof(metrics));
Steffen Klassertee9a8f72012-10-08 00:56:54 +00002329 if (rt->rt_pmtu && expires)
Julian Anastasov521f5492012-07-20 12:02:08 +03002330 metrics[RTAX_MTU - 1] = rt->rt_pmtu;
2331 if (rtnetlink_put_metrics(skb, metrics) < 0)
Thomas Grafbe403ea2006-08-17 18:15:17 -07002332 goto nla_put_failure;
2333
David Millerb4869882012-07-01 02:03:01 +00002334 if (fl4->flowi4_mark &&
stephen hemminger68aaed52012-10-10 08:27:25 +00002335 nla_put_u32(skb, RTA_MARK, fl4->flowi4_mark))
David S. Millerf3756b72012-04-01 20:39:02 -04002336 goto nla_put_failure;
Eric Dumazet963bfee2010-07-20 22:03:14 +00002337
Changli Gaod8d1f302010-06-10 23:31:35 -07002338 error = rt->dst.error;
Thomas Grafbe403ea2006-08-17 18:15:17 -07002339
David S. Millerc7537962010-11-11 17:07:48 -08002340 if (rt_is_input_route(rt)) {
Nicolas Dichtel8caaf7b2012-12-04 01:03:07 +00002341#ifdef CONFIG_IP_MROUTE
2342 if (ipv4_is_multicast(dst) && !ipv4_is_local_multicast(dst) &&
2343 IPV4_DEVCONF_ALL(net, MC_FORWARDING)) {
2344 int err = ipmr_get_route(net, skb,
2345 fl4->saddr, fl4->daddr,
2346 r, nowait);
2347 if (err <= 0) {
2348 if (!nowait) {
2349 if (err == 0)
2350 return 0;
2351 goto nla_put_failure;
2352 } else {
2353 if (err == -EMSGSIZE)
2354 goto nla_put_failure;
2355 error = err;
2356 }
2357 }
2358 } else
2359#endif
2360 if (nla_put_u32(skb, RTA_IIF, rt->rt_iif))
2361 goto nla_put_failure;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 }
2363
David S. Millerf1850712012-07-10 07:26:01 -07002364 if (rtnl_put_cacheinfo(skb, &rt->dst, 0, expires, error) < 0)
Thomas Grafe3703b32006-11-27 09:27:07 -08002365 goto nla_put_failure;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366
Thomas Grafbe403ea2006-08-17 18:15:17 -07002367 return nlmsg_end(skb, nlh);
2368
2369nla_put_failure:
Patrick McHardy26932562007-01-31 23:16:40 -08002370 nlmsg_cancel(skb, nlh);
2371 return -EMSGSIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372}
2373
Thomas Graf661d2962013-03-21 07:45:29 +00002374static int inet_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr *nlh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +09002376 struct net *net = sock_net(in_skb->sk);
Thomas Grafd889ce32006-08-17 18:15:44 -07002377 struct rtmsg *rtm;
2378 struct nlattr *tb[RTA_MAX+1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 struct rtable *rt = NULL;
David Millerd6c0a4f2012-07-01 02:02:59 +00002380 struct flowi4 fl4;
Al Viro9e12bb22006-09-26 21:25:20 -07002381 __be32 dst = 0;
2382 __be32 src = 0;
2383 u32 iif;
Thomas Grafd889ce32006-08-17 18:15:44 -07002384 int err;
Eric Dumazet963bfee2010-07-20 22:03:14 +00002385 int mark;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 struct sk_buff *skb;
2387
Thomas Grafd889ce32006-08-17 18:15:44 -07002388 err = nlmsg_parse(nlh, sizeof(*rtm), tb, RTA_MAX, rtm_ipv4_policy);
2389 if (err < 0)
2390 goto errout;
2391
2392 rtm = nlmsg_data(nlh);
2393
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
Thomas Grafd889ce32006-08-17 18:15:44 -07002395 if (skb == NULL) {
2396 err = -ENOBUFS;
2397 goto errout;
2398 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399
2400 /* Reserve room for dummy headers, this skb can pass
2401 through good chunk of routing engine.
2402 */
Arnaldo Carvalho de Melo459a98e2007-03-19 15:30:44 -07002403 skb_reset_mac_header(skb);
Arnaldo Carvalho de Meloc1d2bbe2007-04-10 20:45:18 -07002404 skb_reset_network_header(skb);
Stephen Hemmingerd2c962b2006-04-17 17:27:11 -07002405
2406 /* Bugfix: need to give ip_route_input enough of an IP header to not gag. */
Arnaldo Carvalho de Meloeddc9ec2007-04-20 22:47:35 -07002407 ip_hdr(skb)->protocol = IPPROTO_ICMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 skb_reserve(skb, MAX_HEADER + sizeof(struct iphdr));
2409
Al Viro17fb2c62006-09-26 22:15:25 -07002410 src = tb[RTA_SRC] ? nla_get_be32(tb[RTA_SRC]) : 0;
2411 dst = tb[RTA_DST] ? nla_get_be32(tb[RTA_DST]) : 0;
Thomas Grafd889ce32006-08-17 18:15:44 -07002412 iif = tb[RTA_IIF] ? nla_get_u32(tb[RTA_IIF]) : 0;
Eric Dumazet963bfee2010-07-20 22:03:14 +00002413 mark = tb[RTA_MARK] ? nla_get_u32(tb[RTA_MARK]) : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414
David Millerd6c0a4f2012-07-01 02:02:59 +00002415 memset(&fl4, 0, sizeof(fl4));
2416 fl4.daddr = dst;
2417 fl4.saddr = src;
2418 fl4.flowi4_tos = rtm->rtm_tos;
2419 fl4.flowi4_oif = tb[RTA_OIF] ? nla_get_u32(tb[RTA_OIF]) : 0;
2420 fl4.flowi4_mark = mark;
2421
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 if (iif) {
Thomas Grafd889ce32006-08-17 18:15:44 -07002423 struct net_device *dev;
2424
Denis V. Lunev19375042008-02-28 20:52:04 -08002425 dev = __dev_get_by_index(net, iif);
Thomas Grafd889ce32006-08-17 18:15:44 -07002426 if (dev == NULL) {
2427 err = -ENODEV;
2428 goto errout_free;
2429 }
2430
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 skb->protocol = htons(ETH_P_IP);
2432 skb->dev = dev;
Eric Dumazet963bfee2010-07-20 22:03:14 +00002433 skb->mark = mark;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 local_bh_disable();
2435 err = ip_route_input(skb, dst, src, rtm->rtm_tos, dev);
2436 local_bh_enable();
Thomas Grafd889ce32006-08-17 18:15:44 -07002437
Eric Dumazet511c3f92009-06-02 05:14:27 +00002438 rt = skb_rtable(skb);
Changli Gaod8d1f302010-06-10 23:31:35 -07002439 if (err == 0 && rt->dst.error)
2440 err = -rt->dst.error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 } else {
David S. Miller9d6ec932011-03-12 01:12:47 -05002442 rt = ip_route_output_key(net, &fl4);
David S. Millerb23dd4f2011-03-02 14:31:35 -08002443
2444 err = 0;
2445 if (IS_ERR(rt))
2446 err = PTR_ERR(rt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 }
Thomas Grafd889ce32006-08-17 18:15:44 -07002448
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 if (err)
Thomas Grafd889ce32006-08-17 18:15:44 -07002450 goto errout_free;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451
Changli Gaod8d1f302010-06-10 23:31:35 -07002452 skb_dst_set(skb, &rt->dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 if (rtm->rtm_flags & RTM_F_NOTIFY)
2454 rt->rt_flags |= RTCF_NOTIFY;
2455
David S. Millerf1ce3062012-07-12 10:10:17 -07002456 err = rt_fill_info(net, dst, src, &fl4, skb,
Eric W. Biederman15e47302012-09-07 20:12:54 +00002457 NETLINK_CB(in_skb).portid, nlh->nlmsg_seq,
Denis V. Lunev19375042008-02-28 20:52:04 -08002458 RTM_NEWROUTE, 0, 0);
Thomas Grafd889ce32006-08-17 18:15:44 -07002459 if (err <= 0)
2460 goto errout_free;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461
Eric W. Biederman15e47302012-09-07 20:12:54 +00002462 err = rtnl_unicast(skb, net, NETLINK_CB(in_skb).portid);
Thomas Grafd889ce32006-08-17 18:15:44 -07002463errout:
Thomas Graf2942e902006-08-15 00:30:25 -07002464 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
Thomas Grafd889ce32006-08-17 18:15:44 -07002466errout_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 kfree_skb(skb);
Thomas Grafd889ce32006-08-17 18:15:44 -07002468 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469}
2470
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471void ip_rt_multicast_event(struct in_device *in_dev)
2472{
Nicolas Dichtel4ccfe6d2012-09-07 00:45:29 +00002473 rt_cache_flush(dev_net(in_dev->dev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474}
2475
2476#ifdef CONFIG_SYSCTL
Gao feng082c7ca2013-02-19 00:43:12 +00002477static int ip_rt_gc_timeout __read_mostly = RT_GC_TIMEOUT;
2478static int ip_rt_gc_interval __read_mostly = 60 * HZ;
2479static int ip_rt_gc_min_interval __read_mostly = HZ / 2;
2480static int ip_rt_gc_elasticity __read_mostly = 8;
2481
Joe Perchesfe2c6332013-06-11 23:04:25 -07002482static int ipv4_sysctl_rtcache_flush(struct ctl_table *__ctl, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002483 void __user *buffer,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 size_t *lenp, loff_t *ppos)
2485{
Timo Teräs5aad1de2013-05-27 20:46:33 +00002486 struct net *net = (struct net *)__ctl->extra1;
2487
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 if (write) {
Timo Teräs5aad1de2013-05-27 20:46:33 +00002489 rt_cache_flush(net);
2490 fnhe_genid_bump(net);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 return 0;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002492 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493
2494 return -EINVAL;
2495}
2496
Joe Perchesfe2c6332013-06-11 23:04:25 -07002497static struct ctl_table ipv4_route_table[] = {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002498 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 .procname = "gc_thresh",
2500 .data = &ipv4_dst_ops.gc_thresh,
2501 .maxlen = sizeof(int),
2502 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002503 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504 },
2505 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 .procname = "max_size",
2507 .data = &ip_rt_max_size,
2508 .maxlen = sizeof(int),
2509 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002510 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 },
2512 {
2513 /* Deprecated. Use gc_min_interval_ms */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002514
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 .procname = "gc_min_interval",
2516 .data = &ip_rt_gc_min_interval,
2517 .maxlen = sizeof(int),
2518 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002519 .proc_handler = proc_dointvec_jiffies,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 },
2521 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 .procname = "gc_min_interval_ms",
2523 .data = &ip_rt_gc_min_interval,
2524 .maxlen = sizeof(int),
2525 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002526 .proc_handler = proc_dointvec_ms_jiffies,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 },
2528 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 .procname = "gc_timeout",
2530 .data = &ip_rt_gc_timeout,
2531 .maxlen = sizeof(int),
2532 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002533 .proc_handler = proc_dointvec_jiffies,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 },
2535 {
Eric Dumazet9f28a2f2011-12-21 15:47:16 -05002536 .procname = "gc_interval",
2537 .data = &ip_rt_gc_interval,
2538 .maxlen = sizeof(int),
2539 .mode = 0644,
2540 .proc_handler = proc_dointvec_jiffies,
2541 },
2542 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543 .procname = "redirect_load",
2544 .data = &ip_rt_redirect_load,
2545 .maxlen = sizeof(int),
2546 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002547 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 },
2549 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 .procname = "redirect_number",
2551 .data = &ip_rt_redirect_number,
2552 .maxlen = sizeof(int),
2553 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002554 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 },
2556 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 .procname = "redirect_silence",
2558 .data = &ip_rt_redirect_silence,
2559 .maxlen = sizeof(int),
2560 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002561 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 },
2563 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 .procname = "error_cost",
2565 .data = &ip_rt_error_cost,
2566 .maxlen = sizeof(int),
2567 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002568 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 },
2570 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 .procname = "error_burst",
2572 .data = &ip_rt_error_burst,
2573 .maxlen = sizeof(int),
2574 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002575 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 },
2577 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 .procname = "gc_elasticity",
2579 .data = &ip_rt_gc_elasticity,
2580 .maxlen = sizeof(int),
2581 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002582 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 },
2584 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585 .procname = "mtu_expires",
2586 .data = &ip_rt_mtu_expires,
2587 .maxlen = sizeof(int),
2588 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002589 .proc_handler = proc_dointvec_jiffies,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 },
2591 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 .procname = "min_pmtu",
2593 .data = &ip_rt_min_pmtu,
2594 .maxlen = sizeof(int),
2595 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002596 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 },
2598 {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 .procname = "min_adv_mss",
2600 .data = &ip_rt_min_advmss,
2601 .maxlen = sizeof(int),
2602 .mode = 0644,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002603 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 },
Eric W. Biedermanf8572d82009-11-05 13:32:03 -08002605 { }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606};
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002607
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002608static struct ctl_table ipv4_route_flush_table[] = {
2609 {
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002610 .procname = "flush",
2611 .maxlen = sizeof(int),
2612 .mode = 0200,
Alexey Dobriyan6d9f2392008-11-03 18:21:05 -08002613 .proc_handler = ipv4_sysctl_rtcache_flush,
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002614 },
Eric W. Biedermanf8572d82009-11-05 13:32:03 -08002615 { },
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002616};
2617
2618static __net_init int sysctl_route_net_init(struct net *net)
2619{
2620 struct ctl_table *tbl;
2621
2622 tbl = ipv4_route_flush_table;
Octavian Purdila09ad9bc2009-11-25 15:14:13 -08002623 if (!net_eq(net, &init_net)) {
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002624 tbl = kmemdup(tbl, sizeof(ipv4_route_flush_table), GFP_KERNEL);
2625 if (tbl == NULL)
2626 goto err_dup;
Eric W. Biederman464dc802012-11-16 03:02:59 +00002627
2628 /* Don't export sysctls to unprivileged users */
2629 if (net->user_ns != &init_user_ns)
2630 tbl[0].procname = NULL;
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002631 }
2632 tbl[0].extra1 = net;
2633
Eric W. Biedermanec8f23c2012-04-19 13:44:49 +00002634 net->ipv4.route_hdr = register_net_sysctl(net, "net/ipv4/route", tbl);
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002635 if (net->ipv4.route_hdr == NULL)
2636 goto err_reg;
2637 return 0;
2638
2639err_reg:
2640 if (tbl != ipv4_route_flush_table)
2641 kfree(tbl);
2642err_dup:
2643 return -ENOMEM;
2644}
2645
2646static __net_exit void sysctl_route_net_exit(struct net *net)
2647{
2648 struct ctl_table *tbl;
2649
2650 tbl = net->ipv4.route_hdr->ctl_table_arg;
2651 unregister_net_sysctl_table(net->ipv4.route_hdr);
2652 BUG_ON(tbl == ipv4_route_flush_table);
2653 kfree(tbl);
2654}
2655
2656static __net_initdata struct pernet_operations sysctl_route_ops = {
2657 .init = sysctl_route_net_init,
2658 .exit = sysctl_route_net_exit,
2659};
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660#endif
2661
Neil Horman3ee94372010-05-08 01:57:52 -07002662static __net_init int rt_genid_init(struct net *net)
Denis V. Lunev9f5e97e2008-07-05 19:02:59 -07002663{
fan.duca4c3fc2013-07-30 08:33:53 +08002664 atomic_set(&net->ipv4.rt_genid, 0);
Timo Teräs5aad1de2013-05-27 20:46:33 +00002665 atomic_set(&net->fnhe_genid, 0);
David S. Miller436c3b62011-03-24 17:42:21 -07002666 get_random_bytes(&net->ipv4.dev_addr_genid,
2667 sizeof(net->ipv4.dev_addr_genid));
Denis V. Lunev9f5e97e2008-07-05 19:02:59 -07002668 return 0;
2669}
2670
Neil Horman3ee94372010-05-08 01:57:52 -07002671static __net_initdata struct pernet_operations rt_genid_ops = {
2672 .init = rt_genid_init,
Denis V. Lunev9f5e97e2008-07-05 19:02:59 -07002673};
2674
David S. Millerc3426b42012-06-09 16:27:05 -07002675static int __net_init ipv4_inetpeer_init(struct net *net)
2676{
2677 struct inet_peer_base *bp = kmalloc(sizeof(*bp), GFP_KERNEL);
2678
2679 if (!bp)
2680 return -ENOMEM;
2681 inet_peer_base_init(bp);
2682 net->ipv4.peers = bp;
2683 return 0;
2684}
2685
2686static void __net_exit ipv4_inetpeer_exit(struct net *net)
2687{
2688 struct inet_peer_base *bp = net->ipv4.peers;
2689
2690 net->ipv4.peers = NULL;
David S. Miller56a6b242012-06-09 16:32:41 -07002691 inetpeer_invalidate_tree(bp);
David S. Millerc3426b42012-06-09 16:27:05 -07002692 kfree(bp);
2693}
2694
2695static __net_initdata struct pernet_operations ipv4_inetpeer_ops = {
2696 .init = ipv4_inetpeer_init,
2697 .exit = ipv4_inetpeer_exit,
2698};
Denis V. Lunev9f5e97e2008-07-05 19:02:59 -07002699
Patrick McHardyc7066f72011-01-14 13:36:42 +01002700#ifdef CONFIG_IP_ROUTE_CLASSID
Tejun Heo7d720c32010-02-16 15:20:26 +00002701struct ip_rt_acct __percpu *ip_rt_acct __read_mostly;
Patrick McHardyc7066f72011-01-14 13:36:42 +01002702#endif /* CONFIG_IP_ROUTE_CLASSID */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704int __init ip_rt_init(void)
2705{
Eric Dumazet424c4b72005-07-05 14:58:19 -07002706 int rc = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707
Patrick McHardyc7066f72011-01-14 13:36:42 +01002708#ifdef CONFIG_IP_ROUTE_CLASSID
Ingo Molnar0dcec8c2009-02-25 14:07:33 +01002709 ip_rt_acct = __alloc_percpu(256 * sizeof(struct ip_rt_acct), __alignof__(struct ip_rt_acct));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 if (!ip_rt_acct)
2711 panic("IP: failed to allocate ip_rt_acct\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712#endif
2713
Alexey Dobriyane5d679f332006-08-26 19:25:52 -07002714 ipv4_dst_ops.kmem_cachep =
2715 kmem_cache_create("ip_dst_cache", sizeof(struct rtable), 0,
Paul Mundt20c2df82007-07-20 10:11:58 +09002716 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717
David S. Miller14e50e52007-05-24 18:17:54 -07002718 ipv4_dst_blackhole_ops.kmem_cachep = ipv4_dst_ops.kmem_cachep;
2719
Eric Dumazetfc66f952010-10-08 06:37:34 +00002720 if (dst_entries_init(&ipv4_dst_ops) < 0)
2721 panic("IP: failed to allocate ipv4_dst_ops counter\n");
2722
2723 if (dst_entries_init(&ipv4_dst_blackhole_ops) < 0)
2724 panic("IP: failed to allocate ipv4_dst_blackhole_ops counter\n");
2725
David S. Miller89aef892012-07-17 11:00:09 -07002726 ipv4_dst_ops.gc_thresh = ~0;
2727 ip_rt_max_size = INT_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 devinet_init();
2730 ip_fib_init();
2731
Denis V. Lunev73b38712008-02-28 20:51:18 -08002732 if (ip_rt_proc_init())
Joe Perches058bd4d2012-03-11 18:36:11 +00002733 pr_err("Unable to create route proc files\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734#ifdef CONFIG_XFRM
2735 xfrm_init();
Steffen Klassert703fb942012-11-13 08:52:24 +01002736 xfrm4_init();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737#endif
Greg Rosec7ac8672011-06-10 01:27:09 +00002738 rtnl_register(PF_INET, RTM_GETROUTE, inet_rtm_getroute, NULL, NULL);
Thomas Graf63f34442007-03-22 11:55:17 -07002739
Denis V. Lunev39a23e72008-07-05 19:02:33 -07002740#ifdef CONFIG_SYSCTL
2741 register_pernet_subsys(&sysctl_route_ops);
2742#endif
Neil Horman3ee94372010-05-08 01:57:52 -07002743 register_pernet_subsys(&rt_genid_ops);
David S. Millerc3426b42012-06-09 16:27:05 -07002744 register_pernet_subsys(&ipv4_inetpeer_ops);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 return rc;
2746}
2747
Al Viroa1bc6eb2008-07-30 06:32:52 -04002748#ifdef CONFIG_SYSCTL
Al Viroeeb61f72008-07-27 08:59:33 +01002749/*
2750 * We really need to sanitize the damn ipv4 init order, then all
2751 * this nonsense will go away.
2752 */
2753void __init ip_static_sysctl_init(void)
2754{
Eric W. Biederman4e5ca782012-04-19 13:32:39 +00002755 register_net_sysctl(&init_net, "net/ipv4/route", ipv4_route_table);
Al Viroeeb61f72008-07-27 08:59:33 +01002756}
Al Viroa1bc6eb2008-07-30 06:32:52 -04002757#endif