blob: a0b4c5da8d43a9ad0bac1f2614c375b61bd17aa0 [file] [log] [blame]
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * xfrm4_policy.c
3 *
4 * Changes:
5 * Kazunori MIYAZAWA @USAGI
6 * YOSHIFUJI Hideaki @USAGI
7 * Split up af-specific portion
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09008 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 */
10
Herbert Xu66cdb3c2007-11-13 21:37:28 -080011#include <linux/err.h>
12#include <linux/kernel.h>
Herbert Xuaabc9762005-05-03 16:27:10 -070013#include <linux/inetdevice.h>
Timo Teräscc9ff192010-11-03 04:41:38 +000014#include <linux/if_tunnel.h>
Herbert Xu45ff5a32007-11-13 21:35:32 -080015#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <net/xfrm.h>
17#include <net/ip.h>
18
Linus Torvalds1da177e2005-04-16 15:20:36 -070019static struct xfrm_policy_afinfo xfrm4_policy_afinfo;
20
David S. Miller8f01cb02011-05-09 15:13:28 -070021static struct dst_entry *__xfrm4_dst_lookup(struct net *net, struct flowi4 *fl4,
22 int tos,
23 const xfrm_address_t *saddr,
24 const xfrm_address_t *daddr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070025{
Herbert Xu66cdb3c2007-11-13 21:37:28 -080026 struct rtable *rt;
Patrick McHardya1e59ab2006-09-19 12:57:34 -070027
David S. Miller8f01cb02011-05-09 15:13:28 -070028 memset(fl4, 0, sizeof(*fl4));
29 fl4->daddr = daddr->a4;
30 fl4->flowi4_tos = tos;
Herbert Xu66cdb3c2007-11-13 21:37:28 -080031 if (saddr)
David S. Miller8f01cb02011-05-09 15:13:28 -070032 fl4->saddr = saddr->a4;
Herbert Xu66cdb3c2007-11-13 21:37:28 -080033
David S. Miller8f01cb02011-05-09 15:13:28 -070034 rt = __ip_route_output_key(net, fl4);
David S. Millerb23dd4f2011-03-02 14:31:35 -080035 if (!IS_ERR(rt))
36 return &rt->dst;
37
38 return ERR_CAST(rt);
Herbert Xu66cdb3c2007-11-13 21:37:28 -080039}
40
David S. Miller8f01cb02011-05-09 15:13:28 -070041static struct dst_entry *xfrm4_dst_lookup(struct net *net, int tos,
42 const xfrm_address_t *saddr,
43 const xfrm_address_t *daddr)
44{
45 struct flowi4 fl4;
46
47 return __xfrm4_dst_lookup(net, &fl4, tos, saddr, daddr);
48}
49
Alexey Dobriyanfbda33b2008-11-25 17:56:49 -080050static int xfrm4_get_saddr(struct net *net,
51 xfrm_address_t *saddr, xfrm_address_t *daddr)
Herbert Xu66cdb3c2007-11-13 21:37:28 -080052{
53 struct dst_entry *dst;
David S. Miller8f01cb02011-05-09 15:13:28 -070054 struct flowi4 fl4;
Herbert Xu66cdb3c2007-11-13 21:37:28 -080055
David S. Miller8f01cb02011-05-09 15:13:28 -070056 dst = __xfrm4_dst_lookup(net, &fl4, 0, NULL, daddr);
Herbert Xu66cdb3c2007-11-13 21:37:28 -080057 if (IS_ERR(dst))
58 return -EHOSTUNREACH;
59
David S. Miller8f01cb02011-05-09 15:13:28 -070060 saddr->a4 = fl4.saddr;
Herbert Xu66cdb3c2007-11-13 21:37:28 -080061 dst_release(dst);
62 return 0;
Patrick McHardya1e59ab2006-09-19 12:57:34 -070063}
64
David S. Miller05d84022011-02-22 17:47:10 -080065static int xfrm4_get_tos(const struct flowi *fl)
Linus Torvalds1da177e2005-04-16 15:20:36 -070066{
David S. Miller7e1dc7b2011-03-12 02:42:11 -050067 return IPTOS_RT_MASK & fl->u.ip4.flowi4_tos; /* Strip ECN bits */
Herbert Xu25ee3282007-12-11 09:32:34 -080068}
Linus Torvalds1da177e2005-04-16 15:20:36 -070069
Masahide NAKAMURAa1b05142007-12-20 20:41:12 -080070static int xfrm4_init_path(struct xfrm_dst *path, struct dst_entry *dst,
71 int nfheader_len)
72{
73 return 0;
74}
75
Herbert Xu87c1e122010-03-02 02:51:56 +000076static int xfrm4_fill_dst(struct xfrm_dst *xdst, struct net_device *dev,
David S. Miller0c7b3ee2011-02-22 17:48:57 -080077 const struct flowi *fl)
Herbert Xu25ee3282007-12-11 09:32:34 -080078{
79 struct rtable *rt = (struct rtable *)xdst->route;
David S. Miller7e1dc7b2011-03-12 02:42:11 -050080 const struct flowi4 *fl4 = &fl->u.ip4;
Linus Torvalds1da177e2005-04-16 15:20:36 -070081
Yan, Zhengb7323392011-10-22 21:58:20 +000082 xdst->u.rt.rt_key_dst = fl4->daddr;
83 xdst->u.rt.rt_key_src = fl4->saddr;
84 xdst->u.rt.rt_key_tos = fl4->flowi4_tos;
85 xdst->u.rt.rt_route_iif = fl4->flowi4_iif;
86 xdst->u.rt.rt_iif = fl4->flowi4_iif;
87 xdst->u.rt.rt_oif = fl4->flowi4_oif;
88 xdst->u.rt.rt_mark = fl4->flowi4_mark;
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
Herbert Xu25ee3282007-12-11 09:32:34 -080090 xdst->u.dst.dev = dev;
91 dev_hold(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
Herbert Xu25ee3282007-12-11 09:32:34 -080093 xdst->u.rt.peer = rt->peer;
94 if (rt->peer)
95 atomic_inc(&rt->peer->refcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
Herbert Xu25ee3282007-12-11 09:32:34 -080097 /* Sheit... I remember I did this right. Apparently,
98 * it was magically lost, so this code needs audit */
99 xdst->u.rt.rt_flags = rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST |
100 RTCF_LOCAL);
101 xdst->u.rt.rt_type = rt->rt_type;
102 xdst->u.rt.rt_src = rt->rt_src;
103 xdst->u.rt.rt_dst = rt->rt_dst;
104 xdst->u.rt.rt_gateway = rt->rt_gateway;
105 xdst->u.rt.rt_spec_dst = rt->rt_spec_dst;
Miika Komu43372262007-02-06 14:27:32 -0800106
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108}
109
110static void
Herbert Xud5422ef2007-12-12 10:44:16 -0800111_decode_session4(struct sk_buff *skb, struct flowi *fl, int reverse)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112{
Eric Dumazetb71d1d42011-04-22 04:53:02 +0000113 const struct iphdr *iph = ip_hdr(skb);
Arnaldo Carvalho de Melod56f90a2007-04-10 20:50:43 -0700114 u8 *xprth = skb_network_header(skb) + iph->ihl * 4;
David S. Miller7e1dc7b2011-03-12 02:42:11 -0500115 struct flowi4 *fl4 = &fl->u.ip4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116
David S. Miller7e1dc7b2011-03-12 02:42:11 -0500117 memset(fl4, 0, sizeof(struct flowi4));
118 fl4->flowi4_mark = skb->mark;
Peter Kosyh44b451f2010-07-02 07:47:55 +0000119
Paul Gortmaker56f8a752011-06-21 20:33:34 -0700120 if (!ip_is_fragment(iph)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 switch (iph->protocol) {
122 case IPPROTO_UDP:
Gerrit Renkerba4e58e2006-11-27 11:10:57 -0800123 case IPPROTO_UDPLITE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 case IPPROTO_TCP:
125 case IPPROTO_SCTP:
Patrick McHardy9e999992005-12-19 14:03:46 -0800126 case IPPROTO_DCCP:
Wei Yongjunc615c9f2009-07-02 16:57:23 +0000127 if (xprth + 4 < skb->data ||
128 pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro8c689a62006-11-08 00:20:21 -0800129 __be16 *ports = (__be16 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
David S. Miller9cce96d2011-03-12 03:00:33 -0500131 fl4->fl4_sport = ports[!!reverse];
132 fl4->fl4_dport = ports[!reverse];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 }
134 break;
135
136 case IPPROTO_ICMP:
137 if (pskb_may_pull(skb, xprth + 2 - skb->data)) {
138 u8 *icmp = xprth;
139
David S. Miller9cce96d2011-03-12 03:00:33 -0500140 fl4->fl4_icmp_type = icmp[0];
141 fl4->fl4_icmp_code = icmp[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 }
143 break;
144
145 case IPPROTO_ESP:
146 if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700147 __be32 *ehdr = (__be32 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
David S. Miller9cce96d2011-03-12 03:00:33 -0500149 fl4->fl4_ipsec_spi = ehdr[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 }
151 break;
152
153 case IPPROTO_AH:
154 if (pskb_may_pull(skb, xprth + 8 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700155 __be32 *ah_hdr = (__be32*)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
David S. Miller9cce96d2011-03-12 03:00:33 -0500157 fl4->fl4_ipsec_spi = ah_hdr[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 }
159 break;
160
161 case IPPROTO_COMP:
162 if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700163 __be16 *ipcomp_hdr = (__be16 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164
David S. Miller9cce96d2011-03-12 03:00:33 -0500165 fl4->fl4_ipsec_spi = htonl(ntohs(ipcomp_hdr[1]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 }
167 break;
Timo Teräscc9ff192010-11-03 04:41:38 +0000168
169 case IPPROTO_GRE:
170 if (pskb_may_pull(skb, xprth + 12 - skb->data)) {
171 __be16 *greflags = (__be16 *)xprth;
172 __be32 *gre_hdr = (__be32 *)xprth;
173
174 if (greflags[0] & GRE_KEY) {
175 if (greflags[0] & GRE_CSUM)
176 gre_hdr++;
David S. Miller9cce96d2011-03-12 03:00:33 -0500177 fl4->fl4_gre_key = gre_hdr[1];
Timo Teräscc9ff192010-11-03 04:41:38 +0000178 }
179 }
180 break;
181
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 default:
David S. Miller9cce96d2011-03-12 03:00:33 -0500183 fl4->fl4_ipsec_spi = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -0700185 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 }
David S. Miller7e1dc7b2011-03-12 02:42:11 -0500187 fl4->flowi4_proto = iph->protocol;
188 fl4->daddr = reverse ? iph->saddr : iph->daddr;
189 fl4->saddr = reverse ? iph->daddr : iph->saddr;
190 fl4->flowi4_tos = iph->tos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191}
192
Daniel Lezcano569d3642008-01-18 03:56:57 -0800193static inline int xfrm4_garbage_collect(struct dst_ops *ops)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194{
Alexey Dobriyand7c75442010-01-24 22:47:53 -0800195 struct net *net = container_of(ops, struct net, xfrm.xfrm4_dst_ops);
196
197 xfrm4_policy_afinfo.garbage_collect(net);
Eric Dumazetfc66f952010-10-08 06:37:34 +0000198 return (dst_entries_get_slow(ops) > ops->gc_thresh * 2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199}
200
201static void xfrm4_update_pmtu(struct dst_entry *dst, u32 mtu)
202{
203 struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
204 struct dst_entry *path = xdst->route;
205
206 path->ops->update_pmtu(path, mtu);
207}
208
Herbert Xuaabc9762005-05-03 16:27:10 -0700209static void xfrm4_dst_destroy(struct dst_entry *dst)
210{
211 struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
212
David S. Miller62fa8a82011-01-26 20:51:05 -0800213 dst_destroy_metrics_generic(dst);
214
Herbert Xued3e37d2007-10-17 21:34:46 -0700215 if (likely(xdst->u.rt.peer))
David S. Miller26db1672006-12-06 23:45:15 -0800216 inet_putpeer(xdst->u.rt.peer);
David S. Miller62fa8a82011-01-26 20:51:05 -0800217
Herbert Xuaabc9762005-05-03 16:27:10 -0700218 xfrm_dst_destroy(xdst);
219}
220
221static void xfrm4_dst_ifdown(struct dst_entry *dst, struct net_device *dev,
222 int unregister)
223{
Herbert Xuaabc9762005-05-03 16:27:10 -0700224 if (!unregister)
225 return;
226
Herbert Xuaabc9762005-05-03 16:27:10 -0700227 xfrm_dst_ifdown(dst, dev);
228}
229
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230static struct dst_ops xfrm4_dst_ops = {
231 .family = AF_INET,
Harvey Harrison09640e62009-02-01 00:45:17 -0800232 .protocol = cpu_to_be16(ETH_P_IP),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 .gc = xfrm4_garbage_collect,
234 .update_pmtu = xfrm4_update_pmtu,
David S. Miller62fa8a82011-01-26 20:51:05 -0800235 .cow_metrics = dst_cow_metrics_generic,
Herbert Xuaabc9762005-05-03 16:27:10 -0700236 .destroy = xfrm4_dst_destroy,
237 .ifdown = xfrm4_dst_ifdown,
Herbert Xu862b82c2007-11-13 21:43:11 -0800238 .local_out = __ip_local_out,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 .gc_thresh = 1024,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240};
241
242static struct xfrm_policy_afinfo xfrm4_policy_afinfo = {
243 .family = AF_INET,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 .dst_ops = &xfrm4_dst_ops,
245 .dst_lookup = xfrm4_dst_lookup,
Patrick McHardya1e59ab2006-09-19 12:57:34 -0700246 .get_saddr = xfrm4_get_saddr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 .decode_session = _decode_session4,
Herbert Xu25ee3282007-12-11 09:32:34 -0800248 .get_tos = xfrm4_get_tos,
Masahide NAKAMURAa1b05142007-12-20 20:41:12 -0800249 .init_path = xfrm4_init_path,
Herbert Xu25ee3282007-12-11 09:32:34 -0800250 .fill_dst = xfrm4_fill_dst,
David S. Miller2774c132011-03-01 14:59:04 -0800251 .blackhole_route = ipv4_blackhole_route,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252};
253
Randy Dunlapf8167002009-08-04 20:18:33 -0700254#ifdef CONFIG_SYSCTL
Neil Hormana44a4a02009-07-27 08:22:46 +0000255static struct ctl_table xfrm4_policy_table[] = {
256 {
Neil Hormana44a4a02009-07-27 08:22:46 +0000257 .procname = "xfrm4_gc_thresh",
Alexey Dobriyand7c75442010-01-24 22:47:53 -0800258 .data = &init_net.xfrm.xfrm4_dst_ops.gc_thresh,
Neil Hormana44a4a02009-07-27 08:22:46 +0000259 .maxlen = sizeof(int),
260 .mode = 0644,
261 .proc_handler = proc_dointvec,
262 },
263 { }
264};
265
266static struct ctl_table_header *sysctl_hdr;
Randy Dunlapf8167002009-08-04 20:18:33 -0700267#endif
Neil Hormana44a4a02009-07-27 08:22:46 +0000268
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269static void __init xfrm4_policy_init(void)
270{
271 xfrm_policy_register_afinfo(&xfrm4_policy_afinfo);
272}
273
274static void __exit xfrm4_policy_fini(void)
275{
Randy Dunlapf8167002009-08-04 20:18:33 -0700276#ifdef CONFIG_SYSCTL
Neil Hormana44a4a02009-07-27 08:22:46 +0000277 if (sysctl_hdr)
278 unregister_net_sysctl_table(sysctl_hdr);
Randy Dunlapf8167002009-08-04 20:18:33 -0700279#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 xfrm_policy_unregister_afinfo(&xfrm4_policy_afinfo);
281}
282
Neil Hormana33bc5c2009-07-30 18:52:15 -0700283void __init xfrm4_init(int rt_max_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284{
Neil Hormana33bc5c2009-07-30 18:52:15 -0700285 /*
286 * Select a default value for the gc_thresh based on the main route
287 * table hash size. It seems to me the worst case scenario is when
288 * we have ipsec operating in transport mode, in which we create a
289 * dst_entry per socket. The xfrm gc algorithm starts trying to remove
290 * entries at gc_thresh, and prevents new allocations as 2*gc_thresh
291 * so lets set an initial xfrm gc_thresh value at the rt_max_size/2.
292 * That will let us store an ipsec connection per route table entry,
293 * and start cleaning when were 1/2 full
294 */
295 xfrm4_dst_ops.gc_thresh = rt_max_size/2;
Eric Dumazetfc66f952010-10-08 06:37:34 +0000296 dst_entries_init(&xfrm4_dst_ops);
Alexey Dobriyand7c75442010-01-24 22:47:53 -0800297
298 xfrm4_state_init();
299 xfrm4_policy_init();
Randy Dunlapf8167002009-08-04 20:18:33 -0700300#ifdef CONFIG_SYSCTL
Neil Hormana44a4a02009-07-27 08:22:46 +0000301 sysctl_hdr = register_net_sysctl_table(&init_net, net_ipv4_ctl_path,
302 xfrm4_policy_table);
Randy Dunlapf8167002009-08-04 20:18:33 -0700303#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304}
305