blob: 1ba44742ebbf16ffef21d26218912ef98619a8f8 [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>
Herbert Xu45ff5a32007-11-13 21:35:32 -080014#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015#include <net/xfrm.h>
16#include <net/ip.h>
17
18static struct dst_ops xfrm4_dst_ops;
19static struct xfrm_policy_afinfo xfrm4_policy_afinfo;
20
Alexey Dobriyanc5b3cf42008-11-25 17:51:25 -080021static struct dst_entry *xfrm4_dst_lookup(struct net *net, int tos,
22 xfrm_address_t *saddr,
Herbert Xu66cdb3c2007-11-13 21:37:28 -080023 xfrm_address_t *daddr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070024{
Herbert Xu66cdb3c2007-11-13 21:37:28 -080025 struct flowi fl = {
Patrick McHardya1e59ab2006-09-19 12:57:34 -070026 .nl_u = {
27 .ip4_u = {
Herbert Xu66cdb3c2007-11-13 21:37:28 -080028 .tos = tos,
Patrick McHardya1e59ab2006-09-19 12:57:34 -070029 .daddr = daddr->a4,
30 },
31 },
32 };
Herbert Xu66cdb3c2007-11-13 21:37:28 -080033 struct dst_entry *dst;
34 struct rtable *rt;
35 int err;
Patrick McHardya1e59ab2006-09-19 12:57:34 -070036
Herbert Xu66cdb3c2007-11-13 21:37:28 -080037 if (saddr)
38 fl.fl4_src = saddr->a4;
39
Alexey Dobriyanc5b3cf42008-11-25 17:51:25 -080040 err = __ip_route_output_key(net, &rt, &fl);
Herbert Xu66cdb3c2007-11-13 21:37:28 -080041 dst = &rt->u.dst;
42 if (err)
43 dst = ERR_PTR(err);
44 return dst;
45}
46
Alexey Dobriyanfbda33b2008-11-25 17:56:49 -080047static int xfrm4_get_saddr(struct net *net,
48 xfrm_address_t *saddr, xfrm_address_t *daddr)
Herbert Xu66cdb3c2007-11-13 21:37:28 -080049{
50 struct dst_entry *dst;
51 struct rtable *rt;
52
Alexey Dobriyanfbda33b2008-11-25 17:56:49 -080053 dst = xfrm4_dst_lookup(net, 0, NULL, daddr);
Herbert Xu66cdb3c2007-11-13 21:37:28 -080054 if (IS_ERR(dst))
55 return -EHOSTUNREACH;
56
57 rt = (struct rtable *)dst;
58 saddr->a4 = rt->rt_src;
59 dst_release(dst);
60 return 0;
Patrick McHardya1e59ab2006-09-19 12:57:34 -070061}
62
Linus Torvalds1da177e2005-04-16 15:20:36 -070063static struct dst_entry *
64__xfrm4_find_bundle(struct flowi *fl, struct xfrm_policy *policy)
65{
66 struct dst_entry *dst;
67
68 read_lock_bh(&policy->lock);
69 for (dst = policy->bundles; dst; dst = dst->next) {
Jianjun Kong5a5f3a82008-11-03 00:24:34 -080070 struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -070071 if (xdst->u.rt.fl.oif == fl->oif && /*XXX*/
72 xdst->u.rt.fl.fl4_dst == fl->fl4_dst &&
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090073 xdst->u.rt.fl.fl4_src == fl->fl4_src &&
74 xdst->u.rt.fl.fl4_tos == fl->fl4_tos &&
Venkat Yekkirala5b368e62006-10-05 15:42:18 -050075 xfrm_bundle_ok(policy, xdst, fl, AF_INET, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 dst_clone(dst);
77 break;
78 }
79 }
80 read_unlock_bh(&policy->lock);
81 return dst;
82}
83
Herbert Xu25ee3282007-12-11 09:32:34 -080084static int xfrm4_get_tos(struct flowi *fl)
Linus Torvalds1da177e2005-04-16 15:20:36 -070085{
Herbert Xu25ee3282007-12-11 09:32:34 -080086 return fl->fl4_tos;
87}
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
Masahide NAKAMURAa1b05142007-12-20 20:41:12 -080089static int xfrm4_init_path(struct xfrm_dst *path, struct dst_entry *dst,
90 int nfheader_len)
91{
92 return 0;
93}
94
Herbert Xu25ee3282007-12-11 09:32:34 -080095static int xfrm4_fill_dst(struct xfrm_dst *xdst, struct net_device *dev)
96{
97 struct rtable *rt = (struct rtable *)xdst->route;
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
Herbert Xu25ee3282007-12-11 09:32:34 -080099 xdst->u.rt.fl = rt->fl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
Herbert Xu25ee3282007-12-11 09:32:34 -0800101 xdst->u.dst.dev = dev;
102 dev_hold(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Herbert Xu25ee3282007-12-11 09:32:34 -0800104 xdst->u.rt.idev = in_dev_get(dev);
105 if (!xdst->u.rt.idev)
106 return -ENODEV;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
Herbert Xu25ee3282007-12-11 09:32:34 -0800108 xdst->u.rt.peer = rt->peer;
109 if (rt->peer)
110 atomic_inc(&rt->peer->refcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
Herbert Xu25ee3282007-12-11 09:32:34 -0800112 /* Sheit... I remember I did this right. Apparently,
113 * it was magically lost, so this code needs audit */
114 xdst->u.rt.rt_flags = rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST |
115 RTCF_LOCAL);
116 xdst->u.rt.rt_type = rt->rt_type;
117 xdst->u.rt.rt_src = rt->rt_src;
118 xdst->u.rt.rt_dst = rt->rt_dst;
119 xdst->u.rt.rt_gateway = rt->rt_gateway;
120 xdst->u.rt.rt_spec_dst = rt->rt_spec_dst;
Miika Komu43372262007-02-06 14:27:32 -0800121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123}
124
125static void
Herbert Xud5422ef2007-12-12 10:44:16 -0800126_decode_session4(struct sk_buff *skb, struct flowi *fl, int reverse)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127{
Arnaldo Carvalho de Meloeddc9ec2007-04-20 22:47:35 -0700128 struct iphdr *iph = ip_hdr(skb);
Arnaldo Carvalho de Melod56f90a2007-04-10 20:50:43 -0700129 u8 *xprth = skb_network_header(skb) + iph->ihl * 4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
131 memset(fl, 0, sizeof(struct flowi));
132 if (!(iph->frag_off & htons(IP_MF | IP_OFFSET))) {
133 switch (iph->protocol) {
134 case IPPROTO_UDP:
Gerrit Renkerba4e58e2006-11-27 11:10:57 -0800135 case IPPROTO_UDPLITE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 case IPPROTO_TCP:
137 case IPPROTO_SCTP:
Patrick McHardy9e999992005-12-19 14:03:46 -0800138 case IPPROTO_DCCP:
Wei Yongjunc615c9f2009-07-02 16:57:23 +0000139 if (xprth + 4 < skb->data ||
140 pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro8c689a62006-11-08 00:20:21 -0800141 __be16 *ports = (__be16 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142
Herbert Xud5422ef2007-12-12 10:44:16 -0800143 fl->fl_ip_sport = ports[!!reverse];
144 fl->fl_ip_dport = ports[!reverse];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 }
146 break;
147
148 case IPPROTO_ICMP:
149 if (pskb_may_pull(skb, xprth + 2 - skb->data)) {
150 u8 *icmp = xprth;
151
152 fl->fl_icmp_type = icmp[0];
153 fl->fl_icmp_code = icmp[1];
154 }
155 break;
156
157 case IPPROTO_ESP:
158 if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700159 __be32 *ehdr = (__be32 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
161 fl->fl_ipsec_spi = ehdr[0];
162 }
163 break;
164
165 case IPPROTO_AH:
166 if (pskb_may_pull(skb, xprth + 8 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700167 __be32 *ah_hdr = (__be32*)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
169 fl->fl_ipsec_spi = ah_hdr[1];
170 }
171 break;
172
173 case IPPROTO_COMP:
174 if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700175 __be16 *ipcomp_hdr = (__be16 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176
Alexey Dobriyan4195f812006-05-22 16:53:22 -0700177 fl->fl_ipsec_spi = htonl(ntohs(ipcomp_hdr[1]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 }
179 break;
180 default:
181 fl->fl_ipsec_spi = 0;
182 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -0700183 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 }
185 fl->proto = iph->protocol;
Herbert Xud5422ef2007-12-12 10:44:16 -0800186 fl->fl4_dst = reverse ? iph->saddr : iph->daddr;
187 fl->fl4_src = reverse ? iph->daddr : iph->saddr;
Herbert Xu4da30892006-02-23 16:19:26 -0800188 fl->fl4_tos = iph->tos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189}
190
Daniel Lezcano569d3642008-01-18 03:56:57 -0800191static inline int xfrm4_garbage_collect(struct dst_ops *ops)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192{
Alexey Dobriyanddcfd792008-11-25 17:37:23 -0800193 xfrm4_policy_afinfo.garbage_collect(&init_net);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 return (atomic_read(&xfrm4_dst_ops.entries) > xfrm4_dst_ops.gc_thresh*2);
195}
196
197static void xfrm4_update_pmtu(struct dst_entry *dst, u32 mtu)
198{
199 struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
200 struct dst_entry *path = xdst->route;
201
202 path->ops->update_pmtu(path, mtu);
203}
204
Herbert Xuaabc9762005-05-03 16:27:10 -0700205static void xfrm4_dst_destroy(struct dst_entry *dst)
206{
207 struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
208
209 if (likely(xdst->u.rt.idev))
210 in_dev_put(xdst->u.rt.idev);
Herbert Xued3e37d2007-10-17 21:34:46 -0700211 if (likely(xdst->u.rt.peer))
David S. Miller26db1672006-12-06 23:45:15 -0800212 inet_putpeer(xdst->u.rt.peer);
Herbert Xuaabc9762005-05-03 16:27:10 -0700213 xfrm_dst_destroy(xdst);
214}
215
216static void xfrm4_dst_ifdown(struct dst_entry *dst, struct net_device *dev,
217 int unregister)
218{
219 struct xfrm_dst *xdst;
220
221 if (!unregister)
222 return;
223
224 xdst = (struct xfrm_dst *)dst;
225 if (xdst->u.rt.idev->dev == dev) {
Denis V. Lunev5a3e55d2007-12-07 00:38:10 -0800226 struct in_device *loopback_idev =
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900227 in_dev_get(dev_net(dev)->loopback_dev);
Herbert Xuaabc9762005-05-03 16:27:10 -0700228 BUG_ON(!loopback_idev);
229
230 do {
231 in_dev_put(xdst->u.rt.idev);
232 xdst->u.rt.idev = loopback_idev;
233 in_dev_hold(loopback_idev);
234 xdst = (struct xfrm_dst *)xdst->u.dst.child;
235 } while (xdst->u.dst.xfrm);
236
237 __in_dev_put(loopback_idev);
238 }
239
240 xfrm_dst_ifdown(dst, dev);
241}
242
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243static struct dst_ops xfrm4_dst_ops = {
244 .family = AF_INET,
Harvey Harrison09640e62009-02-01 00:45:17 -0800245 .protocol = cpu_to_be16(ETH_P_IP),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 .gc = xfrm4_garbage_collect,
247 .update_pmtu = xfrm4_update_pmtu,
Herbert Xuaabc9762005-05-03 16:27:10 -0700248 .destroy = xfrm4_dst_destroy,
249 .ifdown = xfrm4_dst_ifdown,
Herbert Xu862b82c2007-11-13 21:43:11 -0800250 .local_out = __ip_local_out,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 .gc_thresh = 1024,
Eric Dumazete2422972008-01-30 20:07:45 -0800252 .entries = ATOMIC_INIT(0),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253};
254
255static struct xfrm_policy_afinfo xfrm4_policy_afinfo = {
256 .family = AF_INET,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 .dst_ops = &xfrm4_dst_ops,
258 .dst_lookup = xfrm4_dst_lookup,
Patrick McHardya1e59ab2006-09-19 12:57:34 -0700259 .get_saddr = xfrm4_get_saddr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 .find_bundle = __xfrm4_find_bundle,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 .decode_session = _decode_session4,
Herbert Xu25ee3282007-12-11 09:32:34 -0800262 .get_tos = xfrm4_get_tos,
Masahide NAKAMURAa1b05142007-12-20 20:41:12 -0800263 .init_path = xfrm4_init_path,
Herbert Xu25ee3282007-12-11 09:32:34 -0800264 .fill_dst = xfrm4_fill_dst,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265};
266
Neil Hormana44a4a02009-07-27 08:22:46 +0000267static struct ctl_table xfrm4_policy_table[] = {
268 {
269 .ctl_name = CTL_UNNUMBERED,
270 .procname = "xfrm4_gc_thresh",
271 .data = &xfrm4_dst_ops.gc_thresh,
272 .maxlen = sizeof(int),
273 .mode = 0644,
274 .proc_handler = proc_dointvec,
275 },
276 { }
277};
278
279static struct ctl_table_header *sysctl_hdr;
280
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281static void __init xfrm4_policy_init(void)
282{
283 xfrm_policy_register_afinfo(&xfrm4_policy_afinfo);
284}
285
286static void __exit xfrm4_policy_fini(void)
287{
Neil Hormana44a4a02009-07-27 08:22:46 +0000288 if (sysctl_hdr)
289 unregister_net_sysctl_table(sysctl_hdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 xfrm_policy_unregister_afinfo(&xfrm4_policy_afinfo);
291}
292
Neil Hormana33bc5c2009-07-30 18:52:15 -0700293void __init xfrm4_init(int rt_max_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
295 xfrm4_state_init();
296 xfrm4_policy_init();
Neil Hormana33bc5c2009-07-30 18:52:15 -0700297 /*
298 * Select a default value for the gc_thresh based on the main route
299 * table hash size. It seems to me the worst case scenario is when
300 * we have ipsec operating in transport mode, in which we create a
301 * dst_entry per socket. The xfrm gc algorithm starts trying to remove
302 * entries at gc_thresh, and prevents new allocations as 2*gc_thresh
303 * so lets set an initial xfrm gc_thresh value at the rt_max_size/2.
304 * That will let us store an ipsec connection per route table entry,
305 * and start cleaning when were 1/2 full
306 */
307 xfrm4_dst_ops.gc_thresh = rt_max_size/2;
Neil Hormana44a4a02009-07-27 08:22:46 +0000308 sysctl_hdr = register_net_sysctl_table(&init_net, net_ipv4_ctl_path,
309 xfrm4_policy_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310}
311