blob: a580349f0b8ab53c77b04b00d504de0bf1f83e09 [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
Linus Torvalds1da177e2005-04-16 15:20:36 -070018static struct xfrm_policy_afinfo xfrm4_policy_afinfo;
19
Alexey Dobriyanc5b3cf42008-11-25 17:51:25 -080020static struct dst_entry *xfrm4_dst_lookup(struct net *net, int tos,
21 xfrm_address_t *saddr,
Herbert Xu66cdb3c2007-11-13 21:37:28 -080022 xfrm_address_t *daddr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070023{
Herbert Xu66cdb3c2007-11-13 21:37:28 -080024 struct flowi fl = {
Patrick McHardya1e59ab2006-09-19 12:57:34 -070025 .nl_u = {
26 .ip4_u = {
Herbert Xu66cdb3c2007-11-13 21:37:28 -080027 .tos = tos,
Patrick McHardya1e59ab2006-09-19 12:57:34 -070028 .daddr = daddr->a4,
29 },
30 },
31 };
Herbert Xu66cdb3c2007-11-13 21:37:28 -080032 struct dst_entry *dst;
33 struct rtable *rt;
34 int err;
Patrick McHardya1e59ab2006-09-19 12:57:34 -070035
Herbert Xu66cdb3c2007-11-13 21:37:28 -080036 if (saddr)
37 fl.fl4_src = saddr->a4;
38
Alexey Dobriyanc5b3cf42008-11-25 17:51:25 -080039 err = __ip_route_output_key(net, &rt, &fl);
Changli Gaod8d1f302010-06-10 23:31:35 -070040 dst = &rt->dst;
Herbert Xu66cdb3c2007-11-13 21:37:28 -080041 if (err)
42 dst = ERR_PTR(err);
43 return dst;
44}
45
Alexey Dobriyanfbda33b2008-11-25 17:56:49 -080046static int xfrm4_get_saddr(struct net *net,
47 xfrm_address_t *saddr, xfrm_address_t *daddr)
Herbert Xu66cdb3c2007-11-13 21:37:28 -080048{
49 struct dst_entry *dst;
50 struct rtable *rt;
51
Alexey Dobriyanfbda33b2008-11-25 17:56:49 -080052 dst = xfrm4_dst_lookup(net, 0, NULL, daddr);
Herbert Xu66cdb3c2007-11-13 21:37:28 -080053 if (IS_ERR(dst))
54 return -EHOSTUNREACH;
55
56 rt = (struct rtable *)dst;
57 saddr->a4 = rt->rt_src;
58 dst_release(dst);
59 return 0;
Patrick McHardya1e59ab2006-09-19 12:57:34 -070060}
61
Herbert Xu25ee3282007-12-11 09:32:34 -080062static int xfrm4_get_tos(struct flowi *fl)
Linus Torvalds1da177e2005-04-16 15:20:36 -070063{
Ulrich Weber94e223892010-09-22 06:45:11 +000064 return IPTOS_RT_MASK & fl->fl4_tos; /* Strip ECN bits */
Herbert Xu25ee3282007-12-11 09:32:34 -080065}
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
Masahide NAKAMURAa1b05142007-12-20 20:41:12 -080067static int xfrm4_init_path(struct xfrm_dst *path, struct dst_entry *dst,
68 int nfheader_len)
69{
70 return 0;
71}
72
Herbert Xu87c1e122010-03-02 02:51:56 +000073static int xfrm4_fill_dst(struct xfrm_dst *xdst, struct net_device *dev,
74 struct flowi *fl)
Herbert Xu25ee3282007-12-11 09:32:34 -080075{
76 struct rtable *rt = (struct rtable *)xdst->route;
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Herbert Xu87c1e122010-03-02 02:51:56 +000078 xdst->u.rt.fl = *fl;
Linus Torvalds1da177e2005-04-16 15:20:36 -070079
Herbert Xu25ee3282007-12-11 09:32:34 -080080 xdst->u.dst.dev = dev;
81 dev_hold(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -070082
Herbert Xu25ee3282007-12-11 09:32:34 -080083 xdst->u.rt.idev = in_dev_get(dev);
84 if (!xdst->u.rt.idev)
85 return -ENODEV;
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
Herbert Xu25ee3282007-12-11 09:32:34 -080087 xdst->u.rt.peer = rt->peer;
88 if (rt->peer)
89 atomic_inc(&rt->peer->refcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
Herbert Xu25ee3282007-12-11 09:32:34 -080091 /* Sheit... I remember I did this right. Apparently,
92 * it was magically lost, so this code needs audit */
93 xdst->u.rt.rt_flags = rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST |
94 RTCF_LOCAL);
95 xdst->u.rt.rt_type = rt->rt_type;
96 xdst->u.rt.rt_src = rt->rt_src;
97 xdst->u.rt.rt_dst = rt->rt_dst;
98 xdst->u.rt.rt_gateway = rt->rt_gateway;
99 xdst->u.rt.rt_spec_dst = rt->rt_spec_dst;
Miika Komu43372262007-02-06 14:27:32 -0800100
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102}
103
104static void
Herbert Xud5422ef2007-12-12 10:44:16 -0800105_decode_session4(struct sk_buff *skb, struct flowi *fl, int reverse)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106{
Arnaldo Carvalho de Meloeddc9ec2007-04-20 22:47:35 -0700107 struct iphdr *iph = ip_hdr(skb);
Arnaldo Carvalho de Melod56f90a2007-04-10 20:50:43 -0700108 u8 *xprth = skb_network_header(skb) + iph->ihl * 4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110 memset(fl, 0, sizeof(struct flowi));
Peter Kosyh44b451f2010-07-02 07:47:55 +0000111 fl->mark = skb->mark;
112
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 if (!(iph->frag_off & htons(IP_MF | IP_OFFSET))) {
114 switch (iph->protocol) {
115 case IPPROTO_UDP:
Gerrit Renkerba4e58e2006-11-27 11:10:57 -0800116 case IPPROTO_UDPLITE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 case IPPROTO_TCP:
118 case IPPROTO_SCTP:
Patrick McHardy9e999992005-12-19 14:03:46 -0800119 case IPPROTO_DCCP:
Wei Yongjunc615c9f2009-07-02 16:57:23 +0000120 if (xprth + 4 < skb->data ||
121 pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro8c689a62006-11-08 00:20:21 -0800122 __be16 *ports = (__be16 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
Herbert Xud5422ef2007-12-12 10:44:16 -0800124 fl->fl_ip_sport = ports[!!reverse];
125 fl->fl_ip_dport = ports[!reverse];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 }
127 break;
128
129 case IPPROTO_ICMP:
130 if (pskb_may_pull(skb, xprth + 2 - skb->data)) {
131 u8 *icmp = xprth;
132
133 fl->fl_icmp_type = icmp[0];
134 fl->fl_icmp_code = icmp[1];
135 }
136 break;
137
138 case IPPROTO_ESP:
139 if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700140 __be32 *ehdr = (__be32 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
142 fl->fl_ipsec_spi = ehdr[0];
143 }
144 break;
145
146 case IPPROTO_AH:
147 if (pskb_may_pull(skb, xprth + 8 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700148 __be32 *ah_hdr = (__be32*)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
150 fl->fl_ipsec_spi = ah_hdr[1];
151 }
152 break;
153
154 case IPPROTO_COMP:
155 if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
Al Viro4324a172006-09-27 18:49:07 -0700156 __be16 *ipcomp_hdr = (__be16 *)xprth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
Alexey Dobriyan4195f812006-05-22 16:53:22 -0700158 fl->fl_ipsec_spi = htonl(ntohs(ipcomp_hdr[1]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 }
160 break;
161 default:
162 fl->fl_ipsec_spi = 0;
163 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -0700164 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 }
166 fl->proto = iph->protocol;
Herbert Xud5422ef2007-12-12 10:44:16 -0800167 fl->fl4_dst = reverse ? iph->saddr : iph->daddr;
168 fl->fl4_src = reverse ? iph->daddr : iph->saddr;
Herbert Xu4da30892006-02-23 16:19:26 -0800169 fl->fl4_tos = iph->tos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170}
171
Daniel Lezcano569d3642008-01-18 03:56:57 -0800172static inline int xfrm4_garbage_collect(struct dst_ops *ops)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173{
Alexey Dobriyand7c75442010-01-24 22:47:53 -0800174 struct net *net = container_of(ops, struct net, xfrm.xfrm4_dst_ops);
175
176 xfrm4_policy_afinfo.garbage_collect(net);
177 return (atomic_read(&ops->entries) > ops->gc_thresh * 2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178}
179
180static void xfrm4_update_pmtu(struct dst_entry *dst, u32 mtu)
181{
182 struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
183 struct dst_entry *path = xdst->route;
184
185 path->ops->update_pmtu(path, mtu);
186}
187
Herbert Xuaabc9762005-05-03 16:27:10 -0700188static void xfrm4_dst_destroy(struct dst_entry *dst)
189{
190 struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
191
192 if (likely(xdst->u.rt.idev))
193 in_dev_put(xdst->u.rt.idev);
Herbert Xued3e37d2007-10-17 21:34:46 -0700194 if (likely(xdst->u.rt.peer))
David S. Miller26db1672006-12-06 23:45:15 -0800195 inet_putpeer(xdst->u.rt.peer);
Herbert Xuaabc9762005-05-03 16:27:10 -0700196 xfrm_dst_destroy(xdst);
197}
198
199static void xfrm4_dst_ifdown(struct dst_entry *dst, struct net_device *dev,
200 int unregister)
201{
202 struct xfrm_dst *xdst;
203
204 if (!unregister)
205 return;
206
207 xdst = (struct xfrm_dst *)dst;
208 if (xdst->u.rt.idev->dev == dev) {
Denis V. Lunev5a3e55d2007-12-07 00:38:10 -0800209 struct in_device *loopback_idev =
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900210 in_dev_get(dev_net(dev)->loopback_dev);
Herbert Xuaabc9762005-05-03 16:27:10 -0700211 BUG_ON(!loopback_idev);
212
213 do {
214 in_dev_put(xdst->u.rt.idev);
215 xdst->u.rt.idev = loopback_idev;
216 in_dev_hold(loopback_idev);
217 xdst = (struct xfrm_dst *)xdst->u.dst.child;
218 } while (xdst->u.dst.xfrm);
219
220 __in_dev_put(loopback_idev);
221 }
222
223 xfrm_dst_ifdown(dst, dev);
224}
225
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226static struct dst_ops xfrm4_dst_ops = {
227 .family = AF_INET,
Harvey Harrison09640e62009-02-01 00:45:17 -0800228 .protocol = cpu_to_be16(ETH_P_IP),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 .gc = xfrm4_garbage_collect,
230 .update_pmtu = xfrm4_update_pmtu,
Herbert Xuaabc9762005-05-03 16:27:10 -0700231 .destroy = xfrm4_dst_destroy,
232 .ifdown = xfrm4_dst_ifdown,
Herbert Xu862b82c2007-11-13 21:43:11 -0800233 .local_out = __ip_local_out,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 .gc_thresh = 1024,
Eric Dumazete2422972008-01-30 20:07:45 -0800235 .entries = ATOMIC_INIT(0),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236};
237
238static struct xfrm_policy_afinfo xfrm4_policy_afinfo = {
239 .family = AF_INET,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 .dst_ops = &xfrm4_dst_ops,
241 .dst_lookup = xfrm4_dst_lookup,
Patrick McHardya1e59ab2006-09-19 12:57:34 -0700242 .get_saddr = xfrm4_get_saddr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 .decode_session = _decode_session4,
Herbert Xu25ee3282007-12-11 09:32:34 -0800244 .get_tos = xfrm4_get_tos,
Masahide NAKAMURAa1b05142007-12-20 20:41:12 -0800245 .init_path = xfrm4_init_path,
Herbert Xu25ee3282007-12-11 09:32:34 -0800246 .fill_dst = xfrm4_fill_dst,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247};
248
Randy Dunlapf8167002009-08-04 20:18:33 -0700249#ifdef CONFIG_SYSCTL
Neil Hormana44a4a02009-07-27 08:22:46 +0000250static struct ctl_table xfrm4_policy_table[] = {
251 {
Neil Hormana44a4a02009-07-27 08:22:46 +0000252 .procname = "xfrm4_gc_thresh",
Alexey Dobriyand7c75442010-01-24 22:47:53 -0800253 .data = &init_net.xfrm.xfrm4_dst_ops.gc_thresh,
Neil Hormana44a4a02009-07-27 08:22:46 +0000254 .maxlen = sizeof(int),
255 .mode = 0644,
256 .proc_handler = proc_dointvec,
257 },
258 { }
259};
260
261static struct ctl_table_header *sysctl_hdr;
Randy Dunlapf8167002009-08-04 20:18:33 -0700262#endif
Neil Hormana44a4a02009-07-27 08:22:46 +0000263
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264static void __init xfrm4_policy_init(void)
265{
266 xfrm_policy_register_afinfo(&xfrm4_policy_afinfo);
267}
268
269static void __exit xfrm4_policy_fini(void)
270{
Randy Dunlapf8167002009-08-04 20:18:33 -0700271#ifdef CONFIG_SYSCTL
Neil Hormana44a4a02009-07-27 08:22:46 +0000272 if (sysctl_hdr)
273 unregister_net_sysctl_table(sysctl_hdr);
Randy Dunlapf8167002009-08-04 20:18:33 -0700274#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 xfrm_policy_unregister_afinfo(&xfrm4_policy_afinfo);
276}
277
Neil Hormana33bc5c2009-07-30 18:52:15 -0700278void __init xfrm4_init(int rt_max_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279{
Neil Hormana33bc5c2009-07-30 18:52:15 -0700280 /*
281 * Select a default value for the gc_thresh based on the main route
282 * table hash size. It seems to me the worst case scenario is when
283 * we have ipsec operating in transport mode, in which we create a
284 * dst_entry per socket. The xfrm gc algorithm starts trying to remove
285 * entries at gc_thresh, and prevents new allocations as 2*gc_thresh
286 * so lets set an initial xfrm gc_thresh value at the rt_max_size/2.
287 * That will let us store an ipsec connection per route table entry,
288 * and start cleaning when were 1/2 full
289 */
290 xfrm4_dst_ops.gc_thresh = rt_max_size/2;
Alexey Dobriyand7c75442010-01-24 22:47:53 -0800291
292 xfrm4_state_init();
293 xfrm4_policy_init();
Randy Dunlapf8167002009-08-04 20:18:33 -0700294#ifdef CONFIG_SYSCTL
Neil Hormana44a4a02009-07-27 08:22:46 +0000295 sysctl_hdr = register_net_sysctl_table(&init_net, net_ipv4_ctl_path,
296 xfrm4_policy_table);
Randy Dunlapf8167002009-08-04 20:18:33 -0700297#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298}
299