blob: be8abe73bb9f464a2e68679255acde3b708ce84b [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 * The IP forwarding functionality.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09007 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 * Authors: see ip.c
9 *
10 * Fixes:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090011 * Many : Split from ip.c , see ip_input.c for
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * history.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090013 * Dave Gregorich : NULL ip_rt_put fix for multicast
Linus Torvalds1da177e2005-04-16 15:20:36 -070014 * routing.
15 * Jos Vos : Add call_out_firewall before sending,
16 * use output device for accounting.
17 * Jos Vos : Call forward firewall after routing
18 * (always use output device).
19 * Mike McLagan : Routing by source
20 */
21
Linus Torvalds1da177e2005-04-16 15:20:36 -070022#include <linux/types.h>
23#include <linux/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/skbuff.h>
25#include <linux/ip.h>
26#include <linux/icmp.h>
27#include <linux/netdevice.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090028#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <net/sock.h>
30#include <net/ip.h>
31#include <net/tcp.h>
32#include <net/udp.h>
33#include <net/icmp.h>
34#include <linux/tcp.h>
35#include <linux/udp.h>
36#include <linux/netfilter_ipv4.h>
37#include <net/checksum.h>
38#include <linux/route.h>
39#include <net/route.h>
40#include <net/xfrm.h>
41
Florian Westphalfe6cc552014-02-13 23:09:12 +010042static bool ip_may_fragment(const struct sk_buff *skb)
43{
44 return unlikely((ip_hdr(skb)->frag_off & htons(IP_DF)) == 0) ||
45 !skb->local_df;
46}
47
48static bool ip_exceeds_mtu(const struct sk_buff *skb, unsigned int mtu)
49{
50 if (skb->len <= mtu || skb->local_df)
51 return false;
52
53 if (skb_is_gso(skb) && skb_gso_network_seglen(skb) <= mtu)
54 return false;
55
56 return true;
57}
58
59static bool ip_gso_exceeds_dst_mtu(const struct sk_buff *skb)
60{
61 unsigned int mtu;
62
63 if (skb->local_df || !skb_is_gso(skb))
64 return false;
65
66 mtu = ip_dst_mtu_maybe_forward(skb_dst(skb), true);
67
68 /* if seglen > mtu, do software segmentation for IP fragmentation on
69 * output. DF bit cannot be set since ip_forward would have sent
70 * icmp error.
71 */
72 return skb_gso_network_seglen(skb) > mtu;
73}
74
75/* called if GSO skb needs to be fragmented on forward */
76static int ip_forward_finish_gso(struct sk_buff *skb)
77{
78 struct dst_entry *dst = skb_dst(skb);
79 netdev_features_t features;
80 struct sk_buff *segs;
81 int ret = 0;
82
83 features = netif_skb_dev_features(skb, dst->dev);
84 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
85 if (IS_ERR(segs)) {
86 kfree_skb(skb);
87 return -ENOMEM;
88 }
89
90 consume_skb(skb);
91
92 do {
93 struct sk_buff *nskb = segs->next;
94 int err;
95
96 segs->next = NULL;
97 err = dst_output(segs);
98
99 if (err && ret == 0)
100 ret = err;
101 segs = nskb;
102 } while (segs);
103
104 return ret;
105}
106
Patrick McHardy861d0482007-10-15 01:48:39 -0700107static int ip_forward_finish(struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108{
Daniel Baluta5e73ea12012-04-15 01:34:41 +0000109 struct ip_options *opt = &(IPCB(skb)->opt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
Eric Dumazetadf30902009-06-02 05:19:30 +0000111 IP_INC_STATS_BH(dev_net(skb_dst(skb)->dev), IPSTATS_MIB_OUTFORWDATAGRAMS);
Vincent Bernat2d8dbb02012-06-05 03:41:42 +0000112 IP_ADD_STATS_BH(dev_net(skb_dst(skb)->dev), IPSTATS_MIB_OUTOCTETS, skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
114 if (unlikely(opt->optlen))
115 ip_forward_options(skb);
116
Florian Westphalfe6cc552014-02-13 23:09:12 +0100117 if (ip_gso_exceeds_dst_mtu(skb))
118 return ip_forward_finish_gso(skb);
119
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 return dst_output(skb);
121}
122
123int ip_forward(struct sk_buff *skb)
124{
Hannes Frederic Sowaf87c10a2014-01-09 10:01:15 +0100125 u32 mtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 struct iphdr *iph; /* Our header */
127 struct rtable *rt; /* Route we use */
Daniel Baluta5e73ea12012-04-15 01:34:41 +0000128 struct ip_options *opt = &(IPCB(skb)->opt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
Denis Kirjanovd4f2fa62014-02-13 08:58:56 +0400130 /* that should never happen */
131 if (skb->pkt_type != PACKET_HOST)
132 goto drop;
133
Ben Hutchings4497b072008-06-19 16:22:28 -0700134 if (skb_warn_if_lro(skb))
135 goto drop;
136
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 if (!xfrm4_policy_check(NULL, XFRM_POLICY_FWD, skb))
138 goto drop;
139
140 if (IPCB(skb)->opt.router_alert && ip_call_ra_chain(skb))
141 return NET_RX_SUCCESS;
142
Herbert Xu35fc92a2007-03-26 23:22:20 -0700143 skb_forward_csum(skb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900144
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 /*
146 * According to the RFC, we must first decrease the TTL field. If
147 * that reaches zero, we must reply an ICMP control message telling
148 * that the packet's lifetime expired.
149 */
Arnaldo Carvalho de Meloeddc9ec2007-04-20 22:47:35 -0700150 if (ip_hdr(skb)->ttl <= 1)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900151 goto too_many_hops;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
153 if (!xfrm4_route_forward(skb))
154 goto drop;
155
Eric Dumazet511c3f92009-06-02 05:14:27 +0000156 rt = skb_rtable(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
Julian Anastasov155e8332012-10-08 11:41:18 +0000158 if (opt->is_strictroute && rt->rt_uses_gateway)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 goto sr_failed;
160
Hannes Frederic Sowaf87c10a2014-01-09 10:01:15 +0100161 IPCB(skb)->flags |= IPSKB_FORWARDED;
162 mtu = ip_dst_mtu_maybe_forward(&rt->dst, true);
Florian Westphalfe6cc552014-02-13 23:09:12 +0100163 if (!ip_may_fragment(skb) && ip_exceeds_mtu(skb, mtu)) {
Changli Gaod8d1f302010-06-10 23:31:35 -0700164 IP_INC_STATS(dev_net(rt->dst.dev), IPSTATS_MIB_FRAGFAILS);
John Heffner9af39122007-03-25 23:32:29 -0700165 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
Hannes Frederic Sowaf87c10a2014-01-09 10:01:15 +0100166 htonl(mtu));
John Heffner9af39122007-03-25 23:32:29 -0700167 goto drop;
168 }
169
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 /* We are about to mangle packet. Copy it! */
Changli Gaod8d1f302010-06-10 23:31:35 -0700171 if (skb_cow(skb, LL_RESERVED_SPACE(rt->dst.dev)+rt->dst.header_len))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 goto drop;
Arnaldo Carvalho de Meloeddc9ec2007-04-20 22:47:35 -0700173 iph = ip_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174
175 /* Decrease ttl after skb cow done */
176 ip_decrease_ttl(iph);
177
178 /*
179 * We now generate an ICMP HOST REDIRECT giving the route
180 * we calculated.
181 */
Alexey Dobriyandef8b4f2008-10-28 13:24:06 -0700182 if (rt->rt_flags&RTCF_DOREDIRECT && !opt->srr && !skb_sec_path(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 ip_rt_send_redirect(skb);
184
185 skb->priority = rt_tos2priority(iph->tos);
186
Jan Engelhardt9bbc7682010-03-23 04:07:29 +0100187 return NF_HOOK(NFPROTO_IPV4, NF_INET_FORWARD, skb, skb->dev,
Changli Gaod8d1f302010-06-10 23:31:35 -0700188 rt->dst.dev, ip_forward_finish);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189
190sr_failed:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900191 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 * Strict routing permits no gatewaying
193 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900194 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_SR_FAILED, 0);
195 goto drop;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196
197too_many_hops:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900198 /* Tell the sender its packet died... */
Eric Dumazetadf30902009-06-02 05:19:30 +0000199 IP_INC_STATS_BH(dev_net(skb_dst(skb)->dev), IPSTATS_MIB_INHDRERRORS);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900200 icmp_send(skb, ICMP_TIME_EXCEEDED, ICMP_EXC_TTL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201drop:
202 kfree_skb(skb);
203 return NET_RX_DROP;
204}