Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 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 Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 8 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 9 | * 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 Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 21 | * Alan Cox : Super /proc >4K |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 22 | * 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 Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 39 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 40 | * 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 Dumazet | bb1d23b | 2005-07-05 15:00:32 -0700 | [diff] [blame] | 55 | * Eric Dumazet : hashed spinlocks and rt_check_expire() fixes. |
Ilia Sotnikov | cef2685 | 2006-03-25 01:38:55 -0800 | [diff] [blame] | 56 | * Ilia Sotnikov : Ignore TOS on PMTUD and Redirect |
| 57 | * Ilia Sotnikov : Removed TOS from hash calculations |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 58 | * |
| 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 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 65 | #include <linux/module.h> |
| 66 | #include <asm/uaccess.h> |
| 67 | #include <asm/system.h> |
| 68 | #include <linux/bitops.h> |
| 69 | #include <linux/types.h> |
| 70 | #include <linux/kernel.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 71 | #include <linux/mm.h> |
Eric Dumazet | 424c4b7 | 2005-07-05 14:58:19 -0700 | [diff] [blame] | 72 | #include <linux/bootmem.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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> |
Eric Dumazet | 39c90ec | 2007-09-15 10:55:54 -0700 | [diff] [blame] | 82 | #include <linux/workqueue.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 83 | #include <linux/skbuff.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 84 | #include <linux/inetdevice.h> |
| 85 | #include <linux/igmp.h> |
| 86 | #include <linux/pkt_sched.h> |
| 87 | #include <linux/mroute.h> |
| 88 | #include <linux/netfilter_ipv4.h> |
| 89 | #include <linux/random.h> |
| 90 | #include <linux/jhash.h> |
| 91 | #include <linux/rcupdate.h> |
| 92 | #include <linux/times.h> |
Herbert Xu | 352e512 | 2007-11-13 21:34:06 -0800 | [diff] [blame] | 93 | #include <net/dst.h> |
Eric W. Biederman | 457c4cb | 2007-09-12 12:01:34 +0200 | [diff] [blame] | 94 | #include <net/net_namespace.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 95 | #include <net/protocol.h> |
| 96 | #include <net/ip.h> |
| 97 | #include <net/route.h> |
| 98 | #include <net/inetpeer.h> |
| 99 | #include <net/sock.h> |
| 100 | #include <net/ip_fib.h> |
| 101 | #include <net/arp.h> |
| 102 | #include <net/tcp.h> |
| 103 | #include <net/icmp.h> |
| 104 | #include <net/xfrm.h> |
Tom Tucker | 8d71740 | 2006-07-30 20:43:36 -0700 | [diff] [blame] | 105 | #include <net/netevent.h> |
Thomas Graf | 63f3444 | 2007-03-22 11:55:17 -0700 | [diff] [blame] | 106 | #include <net/rtnetlink.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 107 | #ifdef CONFIG_SYSCTL |
| 108 | #include <linux/sysctl.h> |
| 109 | #endif |
| 110 | |
| 111 | #define RT_FL_TOS(oldflp) \ |
| 112 | ((u32)(oldflp->fl4_tos & (IPTOS_RT_MASK | RTO_ONLINK))) |
| 113 | |
| 114 | #define IP_MAX_MTU 0xFFF0 |
| 115 | |
| 116 | #define RT_GC_TIMEOUT (300*HZ) |
| 117 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 118 | static int ip_rt_max_size; |
Stephen Hemminger | 817bc4d | 2008-03-22 17:43:59 -0700 | [diff] [blame] | 119 | static int ip_rt_gc_timeout __read_mostly = RT_GC_TIMEOUT; |
| 120 | static int ip_rt_gc_interval __read_mostly = 60 * HZ; |
| 121 | static int ip_rt_gc_min_interval __read_mostly = HZ / 2; |
| 122 | static int ip_rt_redirect_number __read_mostly = 9; |
| 123 | static int ip_rt_redirect_load __read_mostly = HZ / 50; |
| 124 | static int ip_rt_redirect_silence __read_mostly = ((HZ / 50) << (9 + 1)); |
| 125 | static int ip_rt_error_cost __read_mostly = HZ; |
| 126 | static int ip_rt_error_burst __read_mostly = 5 * HZ; |
| 127 | static int ip_rt_gc_elasticity __read_mostly = 8; |
| 128 | static int ip_rt_mtu_expires __read_mostly = 10 * 60 * HZ; |
| 129 | static int ip_rt_min_pmtu __read_mostly = 512 + 20 + 20; |
| 130 | static int ip_rt_min_advmss __read_mostly = 256; |
| 131 | static int ip_rt_secret_interval __read_mostly = 10 * 60 * HZ; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 132 | static int rt_chain_length_max __read_mostly = 20; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 133 | |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 134 | static void rt_worker_func(struct work_struct *work); |
| 135 | static DECLARE_DELAYED_WORK(expires_work, rt_worker_func); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 136 | |
| 137 | /* |
| 138 | * Interface to generic destination cache. |
| 139 | */ |
| 140 | |
| 141 | static struct dst_entry *ipv4_dst_check(struct dst_entry *dst, u32 cookie); |
| 142 | static void ipv4_dst_destroy(struct dst_entry *dst); |
| 143 | static void ipv4_dst_ifdown(struct dst_entry *dst, |
| 144 | struct net_device *dev, int how); |
| 145 | static struct dst_entry *ipv4_negative_advice(struct dst_entry *dst); |
| 146 | static void ipv4_link_failure(struct sk_buff *skb); |
| 147 | static void ip_rt_update_pmtu(struct dst_entry *dst, u32 mtu); |
Daniel Lezcano | 569d364 | 2008-01-18 03:56:57 -0800 | [diff] [blame] | 148 | static int rt_garbage_collect(struct dst_ops *ops); |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 149 | static void rt_emergency_hash_rebuild(struct net *net); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 150 | |
| 151 | |
| 152 | static struct dst_ops ipv4_dst_ops = { |
| 153 | .family = AF_INET, |
| 154 | .protocol = __constant_htons(ETH_P_IP), |
| 155 | .gc = rt_garbage_collect, |
| 156 | .check = ipv4_dst_check, |
| 157 | .destroy = ipv4_dst_destroy, |
| 158 | .ifdown = ipv4_dst_ifdown, |
| 159 | .negative_advice = ipv4_negative_advice, |
| 160 | .link_failure = ipv4_link_failure, |
| 161 | .update_pmtu = ip_rt_update_pmtu, |
Herbert Xu | 1ac06e0 | 2008-05-20 14:32:14 -0700 | [diff] [blame] | 162 | .local_out = __ip_local_out, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 163 | .entry_size = sizeof(struct rtable), |
Eric Dumazet | e242297 | 2008-01-30 20:07:45 -0800 | [diff] [blame] | 164 | .entries = ATOMIC_INIT(0), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 165 | }; |
| 166 | |
| 167 | #define ECN_OR_COST(class) TC_PRIO_##class |
| 168 | |
Philippe De Muyter | 4839c52 | 2007-07-09 15:32:57 -0700 | [diff] [blame] | 169 | const __u8 ip_tos2prio[16] = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 170 | TC_PRIO_BESTEFFORT, |
| 171 | ECN_OR_COST(FILLER), |
| 172 | TC_PRIO_BESTEFFORT, |
| 173 | ECN_OR_COST(BESTEFFORT), |
| 174 | TC_PRIO_BULK, |
| 175 | ECN_OR_COST(BULK), |
| 176 | TC_PRIO_BULK, |
| 177 | ECN_OR_COST(BULK), |
| 178 | TC_PRIO_INTERACTIVE, |
| 179 | ECN_OR_COST(INTERACTIVE), |
| 180 | TC_PRIO_INTERACTIVE, |
| 181 | ECN_OR_COST(INTERACTIVE), |
| 182 | TC_PRIO_INTERACTIVE_BULK, |
| 183 | ECN_OR_COST(INTERACTIVE_BULK), |
| 184 | TC_PRIO_INTERACTIVE_BULK, |
| 185 | ECN_OR_COST(INTERACTIVE_BULK) |
| 186 | }; |
| 187 | |
| 188 | |
| 189 | /* |
| 190 | * Route cache. |
| 191 | */ |
| 192 | |
| 193 | /* The locking scheme is rather straight forward: |
| 194 | * |
| 195 | * 1) Read-Copy Update protects the buckets of the central route hash. |
| 196 | * 2) Only writers remove entries, and they hold the lock |
| 197 | * as they look at rtable reference counts. |
| 198 | * 3) Only readers acquire references to rtable entries, |
| 199 | * they do so with atomic increments and with the |
| 200 | * lock held. |
| 201 | */ |
| 202 | |
| 203 | struct rt_hash_bucket { |
| 204 | struct rtable *chain; |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 205 | }; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 206 | |
Ingo Molnar | 8a25d5d | 2006-07-03 00:24:54 -0700 | [diff] [blame] | 207 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \ |
| 208 | defined(CONFIG_PROVE_LOCKING) |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 209 | /* |
| 210 | * Instead of using one spinlock for each rt_hash_bucket, we use a table of spinlocks |
| 211 | * The size of this table is a power of two and depends on the number of CPUS. |
Ingo Molnar | 6205120 | 2006-07-03 00:24:59 -0700 | [diff] [blame] | 212 | * (on lockdep we have a quite big spinlock_t, so keep the size down there) |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 213 | */ |
Ingo Molnar | 6205120 | 2006-07-03 00:24:59 -0700 | [diff] [blame] | 214 | #ifdef CONFIG_LOCKDEP |
| 215 | # define RT_HASH_LOCK_SZ 256 |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 216 | #else |
Ingo Molnar | 6205120 | 2006-07-03 00:24:59 -0700 | [diff] [blame] | 217 | # if NR_CPUS >= 32 |
| 218 | # define RT_HASH_LOCK_SZ 4096 |
| 219 | # elif NR_CPUS >= 16 |
| 220 | # define RT_HASH_LOCK_SZ 2048 |
| 221 | # elif NR_CPUS >= 8 |
| 222 | # define RT_HASH_LOCK_SZ 1024 |
| 223 | # elif NR_CPUS >= 4 |
| 224 | # define RT_HASH_LOCK_SZ 512 |
| 225 | # else |
| 226 | # define RT_HASH_LOCK_SZ 256 |
| 227 | # endif |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 228 | #endif |
| 229 | |
| 230 | static spinlock_t *rt_hash_locks; |
| 231 | # define rt_hash_lock_addr(slot) &rt_hash_locks[(slot) & (RT_HASH_LOCK_SZ - 1)] |
Pavel Emelyanov | 1ff1cc2 | 2007-12-05 21:15:05 -0800 | [diff] [blame] | 232 | |
| 233 | static __init void rt_hash_lock_init(void) |
| 234 | { |
| 235 | int i; |
| 236 | |
| 237 | rt_hash_locks = kmalloc(sizeof(spinlock_t) * RT_HASH_LOCK_SZ, |
| 238 | GFP_KERNEL); |
| 239 | if (!rt_hash_locks) |
| 240 | panic("IP: failed to allocate rt_hash_locks\n"); |
| 241 | |
| 242 | for (i = 0; i < RT_HASH_LOCK_SZ; i++) |
| 243 | spin_lock_init(&rt_hash_locks[i]); |
| 244 | } |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 245 | #else |
| 246 | # define rt_hash_lock_addr(slot) NULL |
Pavel Emelyanov | 1ff1cc2 | 2007-12-05 21:15:05 -0800 | [diff] [blame] | 247 | |
| 248 | static inline void rt_hash_lock_init(void) |
| 249 | { |
| 250 | } |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 251 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 252 | |
Stephen Hemminger | 817bc4d | 2008-03-22 17:43:59 -0700 | [diff] [blame] | 253 | static struct rt_hash_bucket *rt_hash_table __read_mostly; |
| 254 | static unsigned rt_hash_mask __read_mostly; |
| 255 | static unsigned int rt_hash_log __read_mostly; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 256 | |
Eric Dumazet | 2f970d8 | 2006-01-17 02:54:36 -0800 | [diff] [blame] | 257 | static DEFINE_PER_CPU(struct rt_cache_stat, rt_cache_stat); |
Andrew Morton | dbd2915 | 2006-01-17 21:58:01 -0800 | [diff] [blame] | 258 | #define RT_CACHE_STAT_INC(field) \ |
Paul Mackerras | bfe5d83 | 2006-06-25 05:47:14 -0700 | [diff] [blame] | 259 | (__raw_get_cpu_var(rt_cache_stat).field++) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 260 | |
Denis V. Lunev | b00180d | 2008-07-05 19:04:09 -0700 | [diff] [blame] | 261 | static inline unsigned int rt_hash(__be32 daddr, __be32 saddr, int idx, |
| 262 | int genid) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 263 | { |
Stephen Hemminger | 1294fc4 | 2008-04-10 01:54:01 -0700 | [diff] [blame] | 264 | return jhash_3words((__force u32)(__be32)(daddr), |
| 265 | (__force u32)(__be32)(saddr), |
Denis V. Lunev | b00180d | 2008-07-05 19:04:09 -0700 | [diff] [blame] | 266 | idx, genid) |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 267 | & rt_hash_mask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 268 | } |
| 269 | |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 270 | static inline int rt_genid(struct net *net) |
| 271 | { |
| 272 | return atomic_read(&net->ipv4.rt_genid); |
| 273 | } |
| 274 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 275 | #ifdef CONFIG_PROC_FS |
| 276 | struct rt_cache_iter_state { |
Denis V. Lunev | a75e936 | 2008-02-28 20:50:55 -0800 | [diff] [blame] | 277 | struct seq_net_private p; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 278 | int bucket; |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 279 | int genid; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 280 | }; |
| 281 | |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 282 | static struct rtable *rt_cache_get_first(struct seq_file *seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 283 | { |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 284 | struct rt_cache_iter_state *st = seq->private; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 285 | struct rtable *r = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 286 | |
| 287 | for (st->bucket = rt_hash_mask; st->bucket >= 0; --st->bucket) { |
Eric Dumazet | a627266 | 2008-08-28 01:11:25 -0700 | [diff] [blame] | 288 | if (!rt_hash_table[st->bucket].chain) |
| 289 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 290 | rcu_read_lock_bh(); |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 291 | r = rcu_dereference(rt_hash_table[st->bucket].chain); |
| 292 | while (r) { |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 293 | if (dev_net(r->u.dst.dev) == seq_file_net(seq) && |
Denis V. Lunev | a75e936 | 2008-02-28 20:50:55 -0800 | [diff] [blame] | 294 | r->rt_genid == st->genid) |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 295 | return r; |
| 296 | r = rcu_dereference(r->u.dst.rt_next); |
| 297 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 298 | rcu_read_unlock_bh(); |
| 299 | } |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 300 | return r; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 301 | } |
| 302 | |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 303 | static struct rtable *__rt_cache_get_next(struct seq_file *seq, |
Denis V. Lunev | 642d631 | 2008-02-28 20:50:33 -0800 | [diff] [blame] | 304 | struct rtable *r) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 305 | { |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 306 | struct rt_cache_iter_state *st = seq->private; |
Eric Dumazet | a627266 | 2008-08-28 01:11:25 -0700 | [diff] [blame] | 307 | |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 308 | r = r->u.dst.rt_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 309 | while (!r) { |
| 310 | rcu_read_unlock_bh(); |
Eric Dumazet | a627266 | 2008-08-28 01:11:25 -0700 | [diff] [blame] | 311 | do { |
| 312 | if (--st->bucket < 0) |
| 313 | return NULL; |
| 314 | } while (!rt_hash_table[st->bucket].chain); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 315 | rcu_read_lock_bh(); |
| 316 | r = rt_hash_table[st->bucket].chain; |
| 317 | } |
Eric Dumazet | 0bccead | 2008-01-10 03:55:57 -0800 | [diff] [blame] | 318 | return rcu_dereference(r); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 319 | } |
| 320 | |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 321 | static struct rtable *rt_cache_get_next(struct seq_file *seq, |
Denis V. Lunev | 642d631 | 2008-02-28 20:50:33 -0800 | [diff] [blame] | 322 | struct rtable *r) |
| 323 | { |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 324 | struct rt_cache_iter_state *st = seq->private; |
| 325 | while ((r = __rt_cache_get_next(seq, r)) != NULL) { |
| 326 | if (dev_net(r->u.dst.dev) != seq_file_net(seq)) |
Denis V. Lunev | a75e936 | 2008-02-28 20:50:55 -0800 | [diff] [blame] | 327 | continue; |
Denis V. Lunev | 642d631 | 2008-02-28 20:50:33 -0800 | [diff] [blame] | 328 | if (r->rt_genid == st->genid) |
| 329 | break; |
| 330 | } |
| 331 | return r; |
| 332 | } |
| 333 | |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 334 | static struct rtable *rt_cache_get_idx(struct seq_file *seq, loff_t pos) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 335 | { |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 336 | struct rtable *r = rt_cache_get_first(seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 337 | |
| 338 | if (r) |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 339 | while (pos && (r = rt_cache_get_next(seq, r))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 340 | --pos; |
| 341 | return pos ? NULL : r; |
| 342 | } |
| 343 | |
| 344 | static void *rt_cache_seq_start(struct seq_file *seq, loff_t *pos) |
| 345 | { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 346 | struct rt_cache_iter_state *st = seq->private; |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 347 | if (*pos) |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 348 | return rt_cache_get_idx(seq, *pos - 1); |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 349 | st->genid = rt_genid(seq_file_net(seq)); |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 350 | return SEQ_START_TOKEN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 351 | } |
| 352 | |
| 353 | static void *rt_cache_seq_next(struct seq_file *seq, void *v, loff_t *pos) |
| 354 | { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 355 | struct rtable *r; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 356 | |
| 357 | if (v == SEQ_START_TOKEN) |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 358 | r = rt_cache_get_first(seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 359 | else |
YOSHIFUJI Hideaki | 1218854 | 2008-03-26 02:36:06 +0900 | [diff] [blame] | 360 | r = rt_cache_get_next(seq, v); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 361 | ++*pos; |
| 362 | return r; |
| 363 | } |
| 364 | |
| 365 | static void rt_cache_seq_stop(struct seq_file *seq, void *v) |
| 366 | { |
| 367 | if (v && v != SEQ_START_TOKEN) |
| 368 | rcu_read_unlock_bh(); |
| 369 | } |
| 370 | |
| 371 | static int rt_cache_seq_show(struct seq_file *seq, void *v) |
| 372 | { |
| 373 | if (v == SEQ_START_TOKEN) |
| 374 | seq_printf(seq, "%-127s\n", |
| 375 | "Iface\tDestination\tGateway \tFlags\t\tRefCnt\tUse\t" |
| 376 | "Metric\tSource\t\tMTU\tWindow\tIRTT\tTOS\tHHRef\t" |
| 377 | "HHUptod\tSpecDst"); |
| 378 | else { |
| 379 | struct rtable *r = v; |
Pavel Emelyanov | 5e659e4 | 2008-04-24 01:02:16 -0700 | [diff] [blame] | 380 | int len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 381 | |
Pavel Emelyanov | 5e659e4 | 2008-04-24 01:02:16 -0700 | [diff] [blame] | 382 | seq_printf(seq, "%s\t%08lX\t%08lX\t%8X\t%d\t%u\t%d\t" |
| 383 | "%08lX\t%d\t%u\t%u\t%02X\t%d\t%1d\t%08X%n", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 384 | r->u.dst.dev ? r->u.dst.dev->name : "*", |
| 385 | (unsigned long)r->rt_dst, (unsigned long)r->rt_gateway, |
| 386 | r->rt_flags, atomic_read(&r->u.dst.__refcnt), |
| 387 | r->u.dst.__use, 0, (unsigned long)r->rt_src, |
| 388 | (dst_metric(&r->u.dst, RTAX_ADVMSS) ? |
| 389 | (int)dst_metric(&r->u.dst, RTAX_ADVMSS) + 40 : 0), |
| 390 | dst_metric(&r->u.dst, RTAX_WINDOW), |
| 391 | (int)((dst_metric(&r->u.dst, RTAX_RTT) >> 3) + |
| 392 | dst_metric(&r->u.dst, RTAX_RTTVAR)), |
| 393 | r->fl.fl4_tos, |
| 394 | r->u.dst.hh ? atomic_read(&r->u.dst.hh->hh_refcnt) : -1, |
| 395 | r->u.dst.hh ? (r->u.dst.hh->hh_output == |
| 396 | dev_queue_xmit) : 0, |
Pavel Emelyanov | 5e659e4 | 2008-04-24 01:02:16 -0700 | [diff] [blame] | 397 | r->rt_spec_dst, &len); |
| 398 | |
| 399 | seq_printf(seq, "%*s\n", 127 - len, ""); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 400 | } |
| 401 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 402 | } |
| 403 | |
Stephen Hemminger | f690808 | 2007-03-12 14:34:29 -0700 | [diff] [blame] | 404 | static const struct seq_operations rt_cache_seq_ops = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 405 | .start = rt_cache_seq_start, |
| 406 | .next = rt_cache_seq_next, |
| 407 | .stop = rt_cache_seq_stop, |
| 408 | .show = rt_cache_seq_show, |
| 409 | }; |
| 410 | |
| 411 | static int rt_cache_seq_open(struct inode *inode, struct file *file) |
| 412 | { |
Denis V. Lunev | a75e936 | 2008-02-28 20:50:55 -0800 | [diff] [blame] | 413 | return seq_open_net(inode, file, &rt_cache_seq_ops, |
Pavel Emelyanov | cf7732e | 2007-10-10 02:29:29 -0700 | [diff] [blame] | 414 | sizeof(struct rt_cache_iter_state)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 415 | } |
| 416 | |
Arjan van de Ven | 9a32144 | 2007-02-12 00:55:35 -0800 | [diff] [blame] | 417 | static const struct file_operations rt_cache_seq_fops = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 418 | .owner = THIS_MODULE, |
| 419 | .open = rt_cache_seq_open, |
| 420 | .read = seq_read, |
| 421 | .llseek = seq_lseek, |
Denis V. Lunev | a75e936 | 2008-02-28 20:50:55 -0800 | [diff] [blame] | 422 | .release = seq_release_net, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 423 | }; |
| 424 | |
| 425 | |
| 426 | static void *rt_cpu_seq_start(struct seq_file *seq, loff_t *pos) |
| 427 | { |
| 428 | int cpu; |
| 429 | |
| 430 | if (*pos == 0) |
| 431 | return SEQ_START_TOKEN; |
| 432 | |
| 433 | for (cpu = *pos-1; cpu < NR_CPUS; ++cpu) { |
| 434 | if (!cpu_possible(cpu)) |
| 435 | continue; |
| 436 | *pos = cpu+1; |
Eric Dumazet | 2f970d8 | 2006-01-17 02:54:36 -0800 | [diff] [blame] | 437 | return &per_cpu(rt_cache_stat, cpu); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 438 | } |
| 439 | return NULL; |
| 440 | } |
| 441 | |
| 442 | static void *rt_cpu_seq_next(struct seq_file *seq, void *v, loff_t *pos) |
| 443 | { |
| 444 | int cpu; |
| 445 | |
| 446 | for (cpu = *pos; cpu < NR_CPUS; ++cpu) { |
| 447 | if (!cpu_possible(cpu)) |
| 448 | continue; |
| 449 | *pos = cpu+1; |
Eric Dumazet | 2f970d8 | 2006-01-17 02:54:36 -0800 | [diff] [blame] | 450 | return &per_cpu(rt_cache_stat, cpu); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 451 | } |
| 452 | return NULL; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 453 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 454 | } |
| 455 | |
| 456 | static void rt_cpu_seq_stop(struct seq_file *seq, void *v) |
| 457 | { |
| 458 | |
| 459 | } |
| 460 | |
| 461 | static int rt_cpu_seq_show(struct seq_file *seq, void *v) |
| 462 | { |
| 463 | struct rt_cache_stat *st = v; |
| 464 | |
| 465 | if (v == SEQ_START_TOKEN) { |
Olaf Rempel | 5bec003 | 2005-04-28 12:16:08 -0700 | [diff] [blame] | 466 | 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 467 | return 0; |
| 468 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 469 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 470 | seq_printf(seq,"%08x %08x %08x %08x %08x %08x %08x %08x " |
| 471 | " %08x %08x %08x %08x %08x %08x %08x %08x %08x \n", |
| 472 | atomic_read(&ipv4_dst_ops.entries), |
| 473 | st->in_hit, |
| 474 | st->in_slow_tot, |
| 475 | st->in_slow_mc, |
| 476 | st->in_no_route, |
| 477 | st->in_brd, |
| 478 | st->in_martian_dst, |
| 479 | st->in_martian_src, |
| 480 | |
| 481 | st->out_hit, |
| 482 | st->out_slow_tot, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 483 | st->out_slow_mc, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 484 | |
| 485 | st->gc_total, |
| 486 | st->gc_ignored, |
| 487 | st->gc_goal_miss, |
| 488 | st->gc_dst_overflow, |
| 489 | st->in_hlist_search, |
| 490 | st->out_hlist_search |
| 491 | ); |
| 492 | return 0; |
| 493 | } |
| 494 | |
Stephen Hemminger | f690808 | 2007-03-12 14:34:29 -0700 | [diff] [blame] | 495 | static const struct seq_operations rt_cpu_seq_ops = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 496 | .start = rt_cpu_seq_start, |
| 497 | .next = rt_cpu_seq_next, |
| 498 | .stop = rt_cpu_seq_stop, |
| 499 | .show = rt_cpu_seq_show, |
| 500 | }; |
| 501 | |
| 502 | |
| 503 | static int rt_cpu_seq_open(struct inode *inode, struct file *file) |
| 504 | { |
| 505 | return seq_open(file, &rt_cpu_seq_ops); |
| 506 | } |
| 507 | |
Arjan van de Ven | 9a32144 | 2007-02-12 00:55:35 -0800 | [diff] [blame] | 508 | static const struct file_operations rt_cpu_seq_fops = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 509 | .owner = THIS_MODULE, |
| 510 | .open = rt_cpu_seq_open, |
| 511 | .read = seq_read, |
| 512 | .llseek = seq_lseek, |
| 513 | .release = seq_release, |
| 514 | }; |
| 515 | |
Pavel Emelyanov | 78c686e | 2007-12-05 21:13:48 -0800 | [diff] [blame] | 516 | #ifdef CONFIG_NET_CLS_ROUTE |
| 517 | static int ip_rt_acct_read(char *buffer, char **start, off_t offset, |
| 518 | int length, int *eof, void *data) |
| 519 | { |
| 520 | unsigned int i; |
| 521 | |
| 522 | if ((offset & 3) || (length & 3)) |
| 523 | return -EIO; |
| 524 | |
| 525 | if (offset >= sizeof(struct ip_rt_acct) * 256) { |
| 526 | *eof = 1; |
| 527 | return 0; |
| 528 | } |
| 529 | |
| 530 | if (offset + length >= sizeof(struct ip_rt_acct) * 256) { |
| 531 | length = sizeof(struct ip_rt_acct) * 256 - offset; |
| 532 | *eof = 1; |
| 533 | } |
| 534 | |
| 535 | offset /= sizeof(u32); |
| 536 | |
| 537 | if (length > 0) { |
| 538 | u32 *dst = (u32 *) buffer; |
| 539 | |
| 540 | *start = buffer; |
| 541 | memset(dst, 0, length); |
| 542 | |
| 543 | for_each_possible_cpu(i) { |
| 544 | unsigned int j; |
| 545 | u32 *src; |
| 546 | |
| 547 | src = ((u32 *) per_cpu_ptr(ip_rt_acct, i)) + offset; |
| 548 | for (j = 0; j < length/4; j++) |
| 549 | dst[j] += src[j]; |
| 550 | } |
| 551 | } |
| 552 | return length; |
| 553 | } |
| 554 | #endif |
Pavel Emelyanov | 107f163 | 2007-12-05 21:14:28 -0800 | [diff] [blame] | 555 | |
Denis V. Lunev | 73b3871 | 2008-02-28 20:51:18 -0800 | [diff] [blame] | 556 | static int __net_init ip_rt_do_proc_init(struct net *net) |
Pavel Emelyanov | 107f163 | 2007-12-05 21:14:28 -0800 | [diff] [blame] | 557 | { |
| 558 | struct proc_dir_entry *pde; |
| 559 | |
| 560 | pde = proc_net_fops_create(net, "rt_cache", S_IRUGO, |
| 561 | &rt_cache_seq_fops); |
| 562 | if (!pde) |
| 563 | goto err1; |
| 564 | |
Wang Chen | 7702072 | 2008-02-28 14:14:25 -0800 | [diff] [blame] | 565 | pde = proc_create("rt_cache", S_IRUGO, |
| 566 | net->proc_net_stat, &rt_cpu_seq_fops); |
Pavel Emelyanov | 107f163 | 2007-12-05 21:14:28 -0800 | [diff] [blame] | 567 | if (!pde) |
| 568 | goto err2; |
| 569 | |
Pavel Emelyanov | 107f163 | 2007-12-05 21:14:28 -0800 | [diff] [blame] | 570 | #ifdef CONFIG_NET_CLS_ROUTE |
| 571 | pde = create_proc_read_entry("rt_acct", 0, net->proc_net, |
| 572 | ip_rt_acct_read, NULL); |
| 573 | if (!pde) |
| 574 | goto err3; |
| 575 | #endif |
| 576 | return 0; |
| 577 | |
| 578 | #ifdef CONFIG_NET_CLS_ROUTE |
| 579 | err3: |
| 580 | remove_proc_entry("rt_cache", net->proc_net_stat); |
| 581 | #endif |
| 582 | err2: |
| 583 | remove_proc_entry("rt_cache", net->proc_net); |
| 584 | err1: |
| 585 | return -ENOMEM; |
| 586 | } |
Denis V. Lunev | 73b3871 | 2008-02-28 20:51:18 -0800 | [diff] [blame] | 587 | |
| 588 | static void __net_exit ip_rt_do_proc_exit(struct net *net) |
| 589 | { |
| 590 | remove_proc_entry("rt_cache", net->proc_net_stat); |
| 591 | remove_proc_entry("rt_cache", net->proc_net); |
| 592 | remove_proc_entry("rt_acct", net->proc_net); |
| 593 | } |
| 594 | |
| 595 | static struct pernet_operations ip_rt_proc_ops __net_initdata = { |
| 596 | .init = ip_rt_do_proc_init, |
| 597 | .exit = ip_rt_do_proc_exit, |
| 598 | }; |
| 599 | |
| 600 | static int __init ip_rt_proc_init(void) |
| 601 | { |
| 602 | return register_pernet_subsys(&ip_rt_proc_ops); |
| 603 | } |
| 604 | |
Pavel Emelyanov | 107f163 | 2007-12-05 21:14:28 -0800 | [diff] [blame] | 605 | #else |
Denis V. Lunev | 73b3871 | 2008-02-28 20:51:18 -0800 | [diff] [blame] | 606 | static inline int ip_rt_proc_init(void) |
Pavel Emelyanov | 107f163 | 2007-12-05 21:14:28 -0800 | [diff] [blame] | 607 | { |
| 608 | return 0; |
| 609 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 610 | #endif /* CONFIG_PROC_FS */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 611 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 612 | static inline void rt_free(struct rtable *rt) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 613 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 614 | call_rcu_bh(&rt->u.dst.rcu_head, dst_rcu_free); |
| 615 | } |
| 616 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 617 | static inline void rt_drop(struct rtable *rt) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 618 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 619 | ip_rt_put(rt); |
| 620 | call_rcu_bh(&rt->u.dst.rcu_head, dst_rcu_free); |
| 621 | } |
| 622 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 623 | static inline int rt_fast_clean(struct rtable *rth) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 624 | { |
| 625 | /* Kill broadcast/multicast entries very aggresively, if they |
| 626 | collide in hash table with more useful entries */ |
| 627 | return (rth->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST)) && |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 628 | rth->fl.iif && rth->u.dst.rt_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 629 | } |
| 630 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 631 | static inline int rt_valuable(struct rtable *rth) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 632 | { |
| 633 | return (rth->rt_flags & (RTCF_REDIRECTED | RTCF_NOTIFY)) || |
| 634 | rth->u.dst.expires; |
| 635 | } |
| 636 | |
| 637 | static int rt_may_expire(struct rtable *rth, unsigned long tmo1, unsigned long tmo2) |
| 638 | { |
| 639 | unsigned long age; |
| 640 | int ret = 0; |
| 641 | |
| 642 | if (atomic_read(&rth->u.dst.__refcnt)) |
| 643 | goto out; |
| 644 | |
| 645 | ret = 1; |
| 646 | if (rth->u.dst.expires && |
| 647 | time_after_eq(jiffies, rth->u.dst.expires)) |
| 648 | goto out; |
| 649 | |
| 650 | age = jiffies - rth->u.dst.lastuse; |
| 651 | ret = 0; |
| 652 | if ((age <= tmo1 && !rt_fast_clean(rth)) || |
| 653 | (age <= tmo2 && rt_valuable(rth))) |
| 654 | goto out; |
| 655 | ret = 1; |
| 656 | out: return ret; |
| 657 | } |
| 658 | |
| 659 | /* Bits of score are: |
| 660 | * 31: very valuable |
| 661 | * 30: not quite useless |
| 662 | * 29..0: usage counter |
| 663 | */ |
| 664 | static inline u32 rt_score(struct rtable *rt) |
| 665 | { |
| 666 | u32 score = jiffies - rt->u.dst.lastuse; |
| 667 | |
| 668 | score = ~score & ~(3<<30); |
| 669 | |
| 670 | if (rt_valuable(rt)) |
| 671 | score |= (1<<31); |
| 672 | |
| 673 | if (!rt->fl.iif || |
| 674 | !(rt->rt_flags & (RTCF_BROADCAST|RTCF_MULTICAST|RTCF_LOCAL))) |
| 675 | score |= (1<<30); |
| 676 | |
| 677 | return score; |
| 678 | } |
| 679 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 680 | static inline bool rt_caching(const struct net *net) |
| 681 | { |
| 682 | return net->ipv4.current_rt_cache_rebuild_count <= |
| 683 | net->ipv4.sysctl_rt_cache_rebuild_count; |
| 684 | } |
| 685 | |
| 686 | static inline bool compare_hash_inputs(const struct flowi *fl1, |
| 687 | const struct flowi *fl2) |
| 688 | { |
| 689 | return (__force u32)(((fl1->nl_u.ip4_u.daddr ^ fl2->nl_u.ip4_u.daddr) | |
| 690 | (fl1->nl_u.ip4_u.saddr ^ fl2->nl_u.ip4_u.saddr) | |
| 691 | (fl1->iif ^ fl2->iif)) == 0); |
| 692 | } |
| 693 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 694 | static inline int compare_keys(struct flowi *fl1, struct flowi *fl2) |
| 695 | { |
Al Viro | 714e85b | 2006-11-14 20:51:49 -0800 | [diff] [blame] | 696 | return ((__force u32)((fl1->nl_u.ip4_u.daddr ^ fl2->nl_u.ip4_u.daddr) | |
| 697 | (fl1->nl_u.ip4_u.saddr ^ fl2->nl_u.ip4_u.saddr)) | |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 698 | (fl1->mark ^ fl2->mark) | |
David S. Miller | 8238b21 | 2006-10-12 00:49:15 -0700 | [diff] [blame] | 699 | (*(u16 *)&fl1->nl_u.ip4_u.tos ^ |
| 700 | *(u16 *)&fl2->nl_u.ip4_u.tos) | |
| 701 | (fl1->oif ^ fl2->oif) | |
| 702 | (fl1->iif ^ fl2->iif)) == 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 703 | } |
| 704 | |
Denis V. Lunev | b592191 | 2008-01-22 23:50:25 -0800 | [diff] [blame] | 705 | static inline int compare_netns(struct rtable *rt1, struct rtable *rt2) |
| 706 | { |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 707 | return dev_net(rt1->u.dst.dev) == dev_net(rt2->u.dst.dev); |
Denis V. Lunev | b592191 | 2008-01-22 23:50:25 -0800 | [diff] [blame] | 708 | } |
| 709 | |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 710 | static inline int rt_is_expired(struct rtable *rth) |
| 711 | { |
| 712 | return rth->rt_genid != rt_genid(dev_net(rth->u.dst.dev)); |
| 713 | } |
| 714 | |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 715 | /* |
| 716 | * Perform a full scan of hash table and free all entries. |
| 717 | * Can be called by a softirq or a process. |
| 718 | * In the later case, we want to be reschedule if necessary |
| 719 | */ |
| 720 | static void rt_do_flush(int process_context) |
| 721 | { |
| 722 | unsigned int i; |
| 723 | struct rtable *rth, *next; |
Denis V. Lunev | 32cb5b4 | 2008-07-05 19:06:12 -0700 | [diff] [blame] | 724 | struct rtable * tail; |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 725 | |
| 726 | for (i = 0; i <= rt_hash_mask; i++) { |
| 727 | if (process_context && need_resched()) |
| 728 | cond_resched(); |
| 729 | rth = rt_hash_table[i].chain; |
| 730 | if (!rth) |
| 731 | continue; |
| 732 | |
| 733 | spin_lock_bh(rt_hash_lock_addr(i)); |
Denis V. Lunev | 32cb5b4 | 2008-07-05 19:06:12 -0700 | [diff] [blame] | 734 | #ifdef CONFIG_NET_NS |
| 735 | { |
| 736 | struct rtable ** prev, * p; |
| 737 | |
| 738 | rth = rt_hash_table[i].chain; |
| 739 | |
| 740 | /* defer releasing the head of the list after spin_unlock */ |
| 741 | for (tail = rth; tail; tail = tail->u.dst.rt_next) |
| 742 | if (!rt_is_expired(tail)) |
| 743 | break; |
| 744 | if (rth != tail) |
| 745 | rt_hash_table[i].chain = tail; |
| 746 | |
| 747 | /* call rt_free on entries after the tail requiring flush */ |
| 748 | prev = &rt_hash_table[i].chain; |
| 749 | for (p = *prev; p; p = next) { |
| 750 | next = p->u.dst.rt_next; |
| 751 | if (!rt_is_expired(p)) { |
| 752 | prev = &p->u.dst.rt_next; |
| 753 | } else { |
| 754 | *prev = next; |
| 755 | rt_free(p); |
| 756 | } |
| 757 | } |
| 758 | } |
| 759 | #else |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 760 | rth = rt_hash_table[i].chain; |
| 761 | rt_hash_table[i].chain = NULL; |
Denis V. Lunev | 32cb5b4 | 2008-07-05 19:06:12 -0700 | [diff] [blame] | 762 | tail = NULL; |
| 763 | #endif |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 764 | spin_unlock_bh(rt_hash_lock_addr(i)); |
| 765 | |
Denis V. Lunev | 32cb5b4 | 2008-07-05 19:06:12 -0700 | [diff] [blame] | 766 | for (; rth != tail; rth = next) { |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 767 | next = rth->u.dst.rt_next; |
| 768 | rt_free(rth); |
| 769 | } |
| 770 | } |
| 771 | } |
| 772 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 773 | /* |
| 774 | * While freeing expired entries, we compute average chain length |
| 775 | * and standard deviation, using fixed-point arithmetic. |
| 776 | * This to have an estimation of rt_chain_length_max |
| 777 | * rt_chain_length_max = max(elasticity, AVG + 4*SD) |
| 778 | * We use 3 bits for frational part, and 29 (or 61) for magnitude. |
| 779 | */ |
| 780 | |
| 781 | #define FRACT_BITS 3 |
| 782 | #define ONE (1UL << FRACT_BITS) |
| 783 | |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 784 | static void rt_check_expire(void) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 785 | { |
Eric Dumazet | bb1d23b | 2005-07-05 15:00:32 -0700 | [diff] [blame] | 786 | static unsigned int rover; |
| 787 | unsigned int i = rover, goal; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 788 | struct rtable *rth, **rthp; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 789 | unsigned long length = 0, samples = 0; |
| 790 | unsigned long sum = 0, sum2 = 0; |
Eric Dumazet | bb1d23b | 2005-07-05 15:00:32 -0700 | [diff] [blame] | 791 | u64 mult; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 792 | |
Eric Dumazet | bb1d23b | 2005-07-05 15:00:32 -0700 | [diff] [blame] | 793 | mult = ((u64)ip_rt_gc_interval) << rt_hash_log; |
| 794 | if (ip_rt_gc_timeout > 1) |
| 795 | do_div(mult, ip_rt_gc_timeout); |
| 796 | goal = (unsigned int)mult; |
Eric Dumazet | 39c90ec | 2007-09-15 10:55:54 -0700 | [diff] [blame] | 797 | if (goal > rt_hash_mask) |
| 798 | goal = rt_hash_mask + 1; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 799 | length = 0; |
Eric Dumazet | bb1d23b | 2005-07-05 15:00:32 -0700 | [diff] [blame] | 800 | for (; goal > 0; goal--) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 801 | unsigned long tmo = ip_rt_gc_timeout; |
| 802 | |
| 803 | i = (i + 1) & rt_hash_mask; |
| 804 | rthp = &rt_hash_table[i].chain; |
| 805 | |
Eric Dumazet | d90bf5a | 2007-11-14 16:14:05 -0800 | [diff] [blame] | 806 | if (need_resched()) |
| 807 | cond_resched(); |
| 808 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 809 | samples++; |
| 810 | |
Stephen Hemminger | cfcabdc | 2007-10-09 01:59:42 -0700 | [diff] [blame] | 811 | if (*rthp == NULL) |
Eric Dumazet | bb1d23b | 2005-07-05 15:00:32 -0700 | [diff] [blame] | 812 | continue; |
Eric Dumazet | 39c90ec | 2007-09-15 10:55:54 -0700 | [diff] [blame] | 813 | spin_lock_bh(rt_hash_lock_addr(i)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 814 | while ((rth = *rthp) != NULL) { |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 815 | if (rt_is_expired(rth)) { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 816 | *rthp = rth->u.dst.rt_next; |
| 817 | rt_free(rth); |
| 818 | continue; |
| 819 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 820 | if (rth->u.dst.expires) { |
| 821 | /* Entry is expired even if it is in use */ |
Eric Dumazet | 39c90ec | 2007-09-15 10:55:54 -0700 | [diff] [blame] | 822 | if (time_before_eq(jiffies, rth->u.dst.expires)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 823 | tmo >>= 1; |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 824 | rthp = &rth->u.dst.rt_next; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 825 | /* |
| 826 | * Only bump our length if the hash |
| 827 | * inputs on entries n and n+1 are not |
| 828 | * the same, we only count entries on |
| 829 | * a chain with equal hash inputs once |
| 830 | * so that entries for different QOS |
| 831 | * levels, and other non-hash input |
| 832 | * attributes don't unfairly skew |
| 833 | * the length computation |
| 834 | */ |
| 835 | if ((*rthp == NULL) || |
| 836 | !compare_hash_inputs(&(*rthp)->fl, |
| 837 | &rth->fl)) |
| 838 | length += ONE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 839 | continue; |
| 840 | } |
| 841 | } else if (!rt_may_expire(rth, tmo, ip_rt_gc_timeout)) { |
| 842 | tmo >>= 1; |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 843 | rthp = &rth->u.dst.rt_next; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 844 | if ((*rthp == NULL) || |
| 845 | !compare_hash_inputs(&(*rthp)->fl, |
| 846 | &rth->fl)) |
| 847 | length += ONE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 848 | continue; |
| 849 | } |
| 850 | |
| 851 | /* Cleanup aged off entries. */ |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 852 | *rthp = rth->u.dst.rt_next; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 853 | rt_free(rth); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 854 | } |
Eric Dumazet | 39c90ec | 2007-09-15 10:55:54 -0700 | [diff] [blame] | 855 | spin_unlock_bh(rt_hash_lock_addr(i)); |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 856 | sum += length; |
| 857 | sum2 += length*length; |
| 858 | } |
| 859 | if (samples) { |
| 860 | unsigned long avg = sum / samples; |
| 861 | unsigned long sd = int_sqrt(sum2 / samples - avg*avg); |
| 862 | rt_chain_length_max = max_t(unsigned long, |
| 863 | ip_rt_gc_elasticity, |
| 864 | (avg + 4*sd) >> FRACT_BITS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 865 | } |
| 866 | rover = i; |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 867 | } |
| 868 | |
| 869 | /* |
| 870 | * rt_worker_func() is run in process context. |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 871 | * we call rt_check_expire() to scan part of the hash table |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 872 | */ |
| 873 | static void rt_worker_func(struct work_struct *work) |
| 874 | { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 875 | rt_check_expire(); |
Eric Dumazet | 39c90ec | 2007-09-15 10:55:54 -0700 | [diff] [blame] | 876 | schedule_delayed_work(&expires_work, ip_rt_gc_interval); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 877 | } |
| 878 | |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 879 | /* |
| 880 | * Pertubation of rt_genid by a small quantity [1..256] |
| 881 | * Using 8 bits of shuffling ensure we can call rt_cache_invalidate() |
| 882 | * many times (2^24) without giving recent rt_genid. |
| 883 | * Jenkins hash is strong enough that litle changes of rt_genid are OK. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 884 | */ |
Denis V. Lunev | 86c657f | 2008-07-05 19:03:31 -0700 | [diff] [blame] | 885 | static void rt_cache_invalidate(struct net *net) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 886 | { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 887 | unsigned char shuffle; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 888 | |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 889 | get_random_bytes(&shuffle, sizeof(shuffle)); |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 890 | atomic_add(shuffle + 1U, &net->ipv4.rt_genid); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 891 | } |
| 892 | |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 893 | /* |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 894 | * delay < 0 : invalidate cache (fast : entries will be deleted later) |
| 895 | * delay >= 0 : invalidate & flush cache (can be long) |
| 896 | */ |
Denis V. Lunev | 76e6ebf | 2008-07-05 19:00:44 -0700 | [diff] [blame] | 897 | void rt_cache_flush(struct net *net, int delay) |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 898 | { |
Denis V. Lunev | 86c657f | 2008-07-05 19:03:31 -0700 | [diff] [blame] | 899 | rt_cache_invalidate(net); |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 900 | if (delay >= 0) |
| 901 | rt_do_flush(!in_softirq()); |
| 902 | } |
| 903 | |
| 904 | /* |
| 905 | * We change rt_genid and let gc do the cleanup |
Eric Dumazet | beb659b | 2007-11-19 22:43:37 -0800 | [diff] [blame] | 906 | */ |
Denis V. Lunev | 9f5e97e | 2008-07-05 19:02:59 -0700 | [diff] [blame] | 907 | static void rt_secret_rebuild(unsigned long __net) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 908 | { |
Denis V. Lunev | 9f5e97e | 2008-07-05 19:02:59 -0700 | [diff] [blame] | 909 | struct net *net = (struct net *)__net; |
Denis V. Lunev | 86c657f | 2008-07-05 19:03:31 -0700 | [diff] [blame] | 910 | rt_cache_invalidate(net); |
Denis V. Lunev | 9f5e97e | 2008-07-05 19:02:59 -0700 | [diff] [blame] | 911 | mod_timer(&net->ipv4.rt_secret_timer, jiffies + ip_rt_secret_interval); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 912 | } |
| 913 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 914 | static void rt_secret_rebuild_oneshot(struct net *net) |
| 915 | { |
| 916 | del_timer_sync(&net->ipv4.rt_secret_timer); |
| 917 | rt_cache_invalidate(net); |
| 918 | if (ip_rt_secret_interval) { |
| 919 | net->ipv4.rt_secret_timer.expires += ip_rt_secret_interval; |
| 920 | add_timer(&net->ipv4.rt_secret_timer); |
| 921 | } |
| 922 | } |
| 923 | |
| 924 | static void rt_emergency_hash_rebuild(struct net *net) |
| 925 | { |
| 926 | if (net_ratelimit()) { |
| 927 | printk(KERN_WARNING "Route hash chain too long!\n"); |
| 928 | printk(KERN_WARNING "Adjust your secret_interval!\n"); |
| 929 | } |
| 930 | |
| 931 | rt_secret_rebuild_oneshot(net); |
| 932 | } |
| 933 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 934 | /* |
| 935 | Short description of GC goals. |
| 936 | |
| 937 | We want to build algorithm, which will keep routing cache |
| 938 | at some equilibrium point, when number of aged off entries |
| 939 | is kept approximately equal to newly generated ones. |
| 940 | |
| 941 | Current expiration strength is variable "expire". |
| 942 | We try to adjust it dynamically, so that if networking |
| 943 | is idle expires is large enough to keep enough of warm entries, |
| 944 | and when load increases it reduces to limit cache size. |
| 945 | */ |
| 946 | |
Daniel Lezcano | 569d364 | 2008-01-18 03:56:57 -0800 | [diff] [blame] | 947 | static int rt_garbage_collect(struct dst_ops *ops) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 948 | { |
| 949 | static unsigned long expire = RT_GC_TIMEOUT; |
| 950 | static unsigned long last_gc; |
| 951 | static int rover; |
| 952 | static int equilibrium; |
| 953 | struct rtable *rth, **rthp; |
| 954 | unsigned long now = jiffies; |
| 955 | int goal; |
| 956 | |
| 957 | /* |
| 958 | * Garbage collection is pretty expensive, |
| 959 | * do not make it too frequently. |
| 960 | */ |
| 961 | |
| 962 | RT_CACHE_STAT_INC(gc_total); |
| 963 | |
| 964 | if (now - last_gc < ip_rt_gc_min_interval && |
| 965 | atomic_read(&ipv4_dst_ops.entries) < ip_rt_max_size) { |
| 966 | RT_CACHE_STAT_INC(gc_ignored); |
| 967 | goto out; |
| 968 | } |
| 969 | |
| 970 | /* Calculate number of entries, which we want to expire now. */ |
| 971 | goal = atomic_read(&ipv4_dst_ops.entries) - |
| 972 | (ip_rt_gc_elasticity << rt_hash_log); |
| 973 | if (goal <= 0) { |
| 974 | if (equilibrium < ipv4_dst_ops.gc_thresh) |
| 975 | equilibrium = ipv4_dst_ops.gc_thresh; |
| 976 | goal = atomic_read(&ipv4_dst_ops.entries) - equilibrium; |
| 977 | if (goal > 0) { |
Eric Dumazet | b790ced | 2007-12-21 01:49:07 -0800 | [diff] [blame] | 978 | equilibrium += min_t(unsigned int, goal >> 1, rt_hash_mask + 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 979 | goal = atomic_read(&ipv4_dst_ops.entries) - equilibrium; |
| 980 | } |
| 981 | } else { |
| 982 | /* We are in dangerous area. Try to reduce cache really |
| 983 | * aggressively. |
| 984 | */ |
Eric Dumazet | b790ced | 2007-12-21 01:49:07 -0800 | [diff] [blame] | 985 | goal = max_t(unsigned int, goal >> 1, rt_hash_mask + 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 986 | equilibrium = atomic_read(&ipv4_dst_ops.entries) - goal; |
| 987 | } |
| 988 | |
| 989 | if (now - last_gc >= ip_rt_gc_min_interval) |
| 990 | last_gc = now; |
| 991 | |
| 992 | if (goal <= 0) { |
| 993 | equilibrium += goal; |
| 994 | goto work_done; |
| 995 | } |
| 996 | |
| 997 | do { |
| 998 | int i, k; |
| 999 | |
| 1000 | for (i = rt_hash_mask, k = rover; i >= 0; i--) { |
| 1001 | unsigned long tmo = expire; |
| 1002 | |
| 1003 | k = (k + 1) & rt_hash_mask; |
| 1004 | rthp = &rt_hash_table[k].chain; |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1005 | spin_lock_bh(rt_hash_lock_addr(k)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1006 | while ((rth = *rthp) != NULL) { |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1007 | if (!rt_is_expired(rth) && |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1008 | !rt_may_expire(rth, tmo, expire)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1009 | tmo >>= 1; |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1010 | rthp = &rth->u.dst.rt_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1011 | continue; |
| 1012 | } |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1013 | *rthp = rth->u.dst.rt_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1014 | rt_free(rth); |
| 1015 | goal--; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1016 | } |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1017 | spin_unlock_bh(rt_hash_lock_addr(k)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1018 | if (goal <= 0) |
| 1019 | break; |
| 1020 | } |
| 1021 | rover = k; |
| 1022 | |
| 1023 | if (goal <= 0) |
| 1024 | goto work_done; |
| 1025 | |
| 1026 | /* Goal is not achieved. We stop process if: |
| 1027 | |
| 1028 | - if expire reduced to zero. Otherwise, expire is halfed. |
| 1029 | - if table is not full. |
| 1030 | - if we are called from interrupt. |
| 1031 | - jiffies check is just fallback/debug loop breaker. |
| 1032 | We will not spin here for long time in any case. |
| 1033 | */ |
| 1034 | |
| 1035 | RT_CACHE_STAT_INC(gc_goal_miss); |
| 1036 | |
| 1037 | if (expire == 0) |
| 1038 | break; |
| 1039 | |
| 1040 | expire >>= 1; |
| 1041 | #if RT_CACHE_DEBUG >= 2 |
| 1042 | printk(KERN_DEBUG "expire>> %u %d %d %d\n", expire, |
| 1043 | atomic_read(&ipv4_dst_ops.entries), goal, i); |
| 1044 | #endif |
| 1045 | |
| 1046 | if (atomic_read(&ipv4_dst_ops.entries) < ip_rt_max_size) |
| 1047 | goto out; |
| 1048 | } while (!in_softirq() && time_before_eq(jiffies, now)); |
| 1049 | |
| 1050 | if (atomic_read(&ipv4_dst_ops.entries) < ip_rt_max_size) |
| 1051 | goto out; |
| 1052 | if (net_ratelimit()) |
| 1053 | printk(KERN_WARNING "dst cache overflow\n"); |
| 1054 | RT_CACHE_STAT_INC(gc_dst_overflow); |
| 1055 | return 1; |
| 1056 | |
| 1057 | work_done: |
| 1058 | expire += ip_rt_gc_min_interval; |
| 1059 | if (expire > ip_rt_gc_timeout || |
| 1060 | atomic_read(&ipv4_dst_ops.entries) < ipv4_dst_ops.gc_thresh) |
| 1061 | expire = ip_rt_gc_timeout; |
| 1062 | #if RT_CACHE_DEBUG >= 2 |
| 1063 | printk(KERN_DEBUG "expire++ %u %d %d %d\n", expire, |
| 1064 | atomic_read(&ipv4_dst_ops.entries), goal, rover); |
| 1065 | #endif |
| 1066 | out: return 0; |
| 1067 | } |
| 1068 | |
| 1069 | static int rt_intern_hash(unsigned hash, struct rtable *rt, struct rtable **rp) |
| 1070 | { |
| 1071 | struct rtable *rth, **rthp; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1072 | struct rtable *rthi; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1073 | unsigned long now; |
| 1074 | struct rtable *cand, **candp; |
| 1075 | u32 min_score; |
| 1076 | int chain_length; |
| 1077 | int attempts = !in_softirq(); |
| 1078 | |
| 1079 | restart: |
| 1080 | chain_length = 0; |
| 1081 | min_score = ~(u32)0; |
| 1082 | cand = NULL; |
| 1083 | candp = NULL; |
| 1084 | now = jiffies; |
| 1085 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1086 | if (!rt_caching(dev_net(rt->u.dst.dev))) { |
| 1087 | rt_drop(rt); |
| 1088 | return 0; |
| 1089 | } |
| 1090 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1091 | rthp = &rt_hash_table[hash].chain; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1092 | rthi = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1093 | |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1094 | spin_lock_bh(rt_hash_lock_addr(hash)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1095 | while ((rth = *rthp) != NULL) { |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1096 | if (rt_is_expired(rth)) { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1097 | *rthp = rth->u.dst.rt_next; |
| 1098 | rt_free(rth); |
| 1099 | continue; |
| 1100 | } |
Denis V. Lunev | b592191 | 2008-01-22 23:50:25 -0800 | [diff] [blame] | 1101 | if (compare_keys(&rth->fl, &rt->fl) && compare_netns(rth, rt)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1102 | /* Put it first */ |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1103 | *rthp = rth->u.dst.rt_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1104 | /* |
| 1105 | * Since lookup is lockfree, the deletion |
| 1106 | * must be visible to another weakly ordered CPU before |
| 1107 | * the insertion at the start of the hash chain. |
| 1108 | */ |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1109 | rcu_assign_pointer(rth->u.dst.rt_next, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1110 | rt_hash_table[hash].chain); |
| 1111 | /* |
| 1112 | * Since lookup is lockfree, the update writes |
| 1113 | * must be ordered for consistency on SMP. |
| 1114 | */ |
| 1115 | rcu_assign_pointer(rt_hash_table[hash].chain, rth); |
| 1116 | |
Pavel Emelyanov | 03f49f3 | 2007-11-10 21:28:34 -0800 | [diff] [blame] | 1117 | dst_use(&rth->u.dst, now); |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1118 | spin_unlock_bh(rt_hash_lock_addr(hash)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1119 | |
| 1120 | rt_drop(rt); |
| 1121 | *rp = rth; |
| 1122 | return 0; |
| 1123 | } |
| 1124 | |
| 1125 | if (!atomic_read(&rth->u.dst.__refcnt)) { |
| 1126 | u32 score = rt_score(rth); |
| 1127 | |
| 1128 | if (score <= min_score) { |
| 1129 | cand = rth; |
| 1130 | candp = rthp; |
| 1131 | min_score = score; |
| 1132 | } |
| 1133 | } |
| 1134 | |
| 1135 | chain_length++; |
| 1136 | |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1137 | rthp = &rth->u.dst.rt_next; |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1138 | |
| 1139 | /* |
| 1140 | * check to see if the next entry in the chain |
| 1141 | * contains the same hash input values as rt. If it does |
| 1142 | * This is where we will insert into the list, instead of |
| 1143 | * at the head. This groups entries that differ by aspects not |
| 1144 | * relvant to the hash function together, which we use to adjust |
| 1145 | * our chain length |
| 1146 | */ |
| 1147 | if (*rthp && compare_hash_inputs(&(*rthp)->fl, &rt->fl)) |
| 1148 | rthi = rth; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1149 | } |
| 1150 | |
| 1151 | if (cand) { |
| 1152 | /* ip_rt_gc_elasticity used to be average length of chain |
| 1153 | * length, when exceeded gc becomes really aggressive. |
| 1154 | * |
| 1155 | * The second limit is less certain. At the moment it allows |
| 1156 | * only 2 entries per bucket. We will see. |
| 1157 | */ |
| 1158 | if (chain_length > ip_rt_gc_elasticity) { |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1159 | *candp = cand->u.dst.rt_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1160 | rt_free(cand); |
| 1161 | } |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1162 | } else { |
| 1163 | if (chain_length > rt_chain_length_max) { |
| 1164 | struct net *net = dev_net(rt->u.dst.dev); |
| 1165 | int num = ++net->ipv4.current_rt_cache_rebuild_count; |
| 1166 | if (!rt_caching(dev_net(rt->u.dst.dev))) { |
| 1167 | printk(KERN_WARNING "%s: %d rebuilds is over limit, route caching disabled\n", |
| 1168 | rt->u.dst.dev->name, num); |
| 1169 | } |
| 1170 | rt_emergency_hash_rebuild(dev_net(rt->u.dst.dev)); |
| 1171 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1172 | } |
| 1173 | |
| 1174 | /* Try to bind route to arp only if it is output |
| 1175 | route or unicast forwarding path. |
| 1176 | */ |
| 1177 | if (rt->rt_type == RTN_UNICAST || rt->fl.iif == 0) { |
| 1178 | int err = arp_bind_neighbour(&rt->u.dst); |
| 1179 | if (err) { |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1180 | spin_unlock_bh(rt_hash_lock_addr(hash)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1181 | |
| 1182 | if (err != -ENOBUFS) { |
| 1183 | rt_drop(rt); |
| 1184 | return err; |
| 1185 | } |
| 1186 | |
| 1187 | /* Neighbour tables are full and nothing |
| 1188 | can be released. Try to shrink route cache, |
| 1189 | it is most likely it holds some neighbour records. |
| 1190 | */ |
| 1191 | if (attempts-- > 0) { |
| 1192 | int saved_elasticity = ip_rt_gc_elasticity; |
| 1193 | int saved_int = ip_rt_gc_min_interval; |
| 1194 | ip_rt_gc_elasticity = 1; |
| 1195 | ip_rt_gc_min_interval = 0; |
Daniel Lezcano | 569d364 | 2008-01-18 03:56:57 -0800 | [diff] [blame] | 1196 | rt_garbage_collect(&ipv4_dst_ops); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1197 | ip_rt_gc_min_interval = saved_int; |
| 1198 | ip_rt_gc_elasticity = saved_elasticity; |
| 1199 | goto restart; |
| 1200 | } |
| 1201 | |
| 1202 | if (net_ratelimit()) |
| 1203 | printk(KERN_WARNING "Neighbour table overflow.\n"); |
| 1204 | rt_drop(rt); |
| 1205 | return -ENOBUFS; |
| 1206 | } |
| 1207 | } |
| 1208 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1209 | if (rthi) |
| 1210 | rt->u.dst.rt_next = rthi->u.dst.rt_next; |
| 1211 | else |
| 1212 | rt->u.dst.rt_next = rt_hash_table[hash].chain; |
| 1213 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1214 | #if RT_CACHE_DEBUG >= 2 |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1215 | if (rt->u.dst.rt_next) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1216 | struct rtable *trt; |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 1217 | printk(KERN_DEBUG "rt_cache @%02x: " NIPQUAD_FMT, hash, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1218 | NIPQUAD(rt->rt_dst)); |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1219 | for (trt = rt->u.dst.rt_next; trt; trt = trt->u.dst.rt_next) |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 1220 | printk(" . " NIPQUAD_FMT, NIPQUAD(trt->rt_dst)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1221 | printk("\n"); |
| 1222 | } |
| 1223 | #endif |
Eric Dumazet | 00269b5 | 2008-10-16 14:18:29 -0700 | [diff] [blame] | 1224 | /* |
| 1225 | * Since lookup is lockfree, we must make sure |
| 1226 | * previous writes to rt are comitted to memory |
| 1227 | * before making rt visible to other CPUS. |
| 1228 | */ |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1229 | if (rthi) |
| 1230 | rcu_assign_pointer(rthi->u.dst.rt_next, rt); |
| 1231 | else |
| 1232 | rcu_assign_pointer(rt_hash_table[hash].chain, rt); |
| 1233 | |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1234 | spin_unlock_bh(rt_hash_lock_addr(hash)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1235 | *rp = rt; |
| 1236 | return 0; |
| 1237 | } |
| 1238 | |
| 1239 | void rt_bind_peer(struct rtable *rt, int create) |
| 1240 | { |
| 1241 | static DEFINE_SPINLOCK(rt_peer_lock); |
| 1242 | struct inet_peer *peer; |
| 1243 | |
| 1244 | peer = inet_getpeer(rt->rt_dst, create); |
| 1245 | |
| 1246 | spin_lock_bh(&rt_peer_lock); |
| 1247 | if (rt->peer == NULL) { |
| 1248 | rt->peer = peer; |
| 1249 | peer = NULL; |
| 1250 | } |
| 1251 | spin_unlock_bh(&rt_peer_lock); |
| 1252 | if (peer) |
| 1253 | inet_putpeer(peer); |
| 1254 | } |
| 1255 | |
| 1256 | /* |
| 1257 | * Peer allocation may fail only in serious out-of-memory conditions. However |
| 1258 | * we still can generate some output. |
| 1259 | * Random ID selection looks a bit dangerous because we have no chances to |
| 1260 | * select ID being unique in a reasonable period of time. |
| 1261 | * But broken packet identifier may be better than no packet at all. |
| 1262 | */ |
| 1263 | static void ip_select_fb_ident(struct iphdr *iph) |
| 1264 | { |
| 1265 | static DEFINE_SPINLOCK(ip_fb_id_lock); |
| 1266 | static u32 ip_fallback_id; |
| 1267 | u32 salt; |
| 1268 | |
| 1269 | spin_lock_bh(&ip_fb_id_lock); |
Al Viro | e448515 | 2006-09-26 22:15:01 -0700 | [diff] [blame] | 1270 | salt = secure_ip_id((__force __be32)ip_fallback_id ^ iph->daddr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1271 | iph->id = htons(salt & 0xFFFF); |
| 1272 | ip_fallback_id = salt; |
| 1273 | spin_unlock_bh(&ip_fb_id_lock); |
| 1274 | } |
| 1275 | |
| 1276 | void __ip_select_ident(struct iphdr *iph, struct dst_entry *dst, int more) |
| 1277 | { |
| 1278 | struct rtable *rt = (struct rtable *) dst; |
| 1279 | |
| 1280 | if (rt) { |
| 1281 | if (rt->peer == NULL) |
| 1282 | rt_bind_peer(rt, 1); |
| 1283 | |
| 1284 | /* If peer is attached to destination, it is never detached, |
| 1285 | so that we need not to grab a lock to dereference it. |
| 1286 | */ |
| 1287 | if (rt->peer) { |
| 1288 | iph->id = htons(inet_getid(rt->peer, more)); |
| 1289 | return; |
| 1290 | } |
| 1291 | } else |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1292 | printk(KERN_DEBUG "rt_bind_peer(0) @%p\n", |
Stephen Hemminger | 9c2b332 | 2005-04-19 22:39:42 -0700 | [diff] [blame] | 1293 | __builtin_return_address(0)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1294 | |
| 1295 | ip_select_fb_ident(iph); |
| 1296 | } |
| 1297 | |
| 1298 | static void rt_del(unsigned hash, struct rtable *rt) |
| 1299 | { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1300 | struct rtable **rthp, *aux; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1301 | |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1302 | rthp = &rt_hash_table[hash].chain; |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1303 | spin_lock_bh(rt_hash_lock_addr(hash)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1304 | ip_rt_put(rt); |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1305 | while ((aux = *rthp) != NULL) { |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1306 | if (aux == rt || rt_is_expired(aux)) { |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1307 | *rthp = aux->u.dst.rt_next; |
| 1308 | rt_free(aux); |
| 1309 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1310 | } |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1311 | rthp = &aux->u.dst.rt_next; |
| 1312 | } |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 1313 | spin_unlock_bh(rt_hash_lock_addr(hash)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1314 | } |
| 1315 | |
Al Viro | f765522 | 2006-09-26 21:25:43 -0700 | [diff] [blame] | 1316 | void ip_rt_redirect(__be32 old_gw, __be32 daddr, __be32 new_gw, |
| 1317 | __be32 saddr, struct net_device *dev) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1318 | { |
| 1319 | int i, k; |
| 1320 | struct in_device *in_dev = in_dev_get(dev); |
| 1321 | struct rtable *rth, **rthp; |
Al Viro | f765522 | 2006-09-26 21:25:43 -0700 | [diff] [blame] | 1322 | __be32 skeys[2] = { saddr, 0 }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1323 | int ikeys[2] = { dev->ifindex, 0 }; |
Tom Tucker | 8d71740 | 2006-07-30 20:43:36 -0700 | [diff] [blame] | 1324 | struct netevent_redirect netevent; |
Denis V. Lunev | 317805b | 2008-02-28 20:50:06 -0800 | [diff] [blame] | 1325 | struct net *net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1326 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1327 | if (!in_dev) |
| 1328 | return; |
| 1329 | |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 1330 | net = dev_net(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1331 | if (new_gw == old_gw || !IN_DEV_RX_REDIRECTS(in_dev) |
Jan Engelhardt | 1e637c7 | 2008-01-21 03:18:08 -0800 | [diff] [blame] | 1332 | || ipv4_is_multicast(new_gw) || ipv4_is_lbcast(new_gw) |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 1333 | || ipv4_is_zeronet(new_gw)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1334 | goto reject_redirect; |
| 1335 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 1336 | if (!rt_caching(net)) |
| 1337 | goto reject_redirect; |
| 1338 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1339 | if (!IN_DEV_SHARED_MEDIA(in_dev)) { |
| 1340 | if (!inet_addr_onlink(in_dev, new_gw, old_gw)) |
| 1341 | goto reject_redirect; |
| 1342 | if (IN_DEV_SEC_REDIRECTS(in_dev) && ip_fib_check_default(new_gw, dev)) |
| 1343 | goto reject_redirect; |
| 1344 | } else { |
Denis V. Lunev | 317805b | 2008-02-28 20:50:06 -0800 | [diff] [blame] | 1345 | if (inet_addr_type(net, new_gw) != RTN_UNICAST) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1346 | goto reject_redirect; |
| 1347 | } |
| 1348 | |
| 1349 | for (i = 0; i < 2; i++) { |
| 1350 | for (k = 0; k < 2; k++) { |
Denis V. Lunev | b00180d | 2008-07-05 19:04:09 -0700 | [diff] [blame] | 1351 | unsigned hash = rt_hash(daddr, skeys[i], ikeys[k], |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1352 | rt_genid(net)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1353 | |
| 1354 | rthp=&rt_hash_table[hash].chain; |
| 1355 | |
| 1356 | rcu_read_lock(); |
| 1357 | while ((rth = rcu_dereference(*rthp)) != NULL) { |
| 1358 | struct rtable *rt; |
| 1359 | |
| 1360 | if (rth->fl.fl4_dst != daddr || |
| 1361 | rth->fl.fl4_src != skeys[i] || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1362 | rth->fl.oif != ikeys[k] || |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1363 | rth->fl.iif != 0 || |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1364 | rt_is_expired(rth) || |
YOSHIFUJI Hideaki | 878628f | 2008-03-26 03:57:35 +0900 | [diff] [blame] | 1365 | !net_eq(dev_net(rth->u.dst.dev), net)) { |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 1366 | rthp = &rth->u.dst.rt_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1367 | continue; |
| 1368 | } |
| 1369 | |
| 1370 | if (rth->rt_dst != daddr || |
| 1371 | rth->rt_src != saddr || |
| 1372 | rth->u.dst.error || |
| 1373 | rth->rt_gateway != old_gw || |
| 1374 | rth->u.dst.dev != dev) |
| 1375 | break; |
| 1376 | |
| 1377 | dst_hold(&rth->u.dst); |
| 1378 | rcu_read_unlock(); |
| 1379 | |
| 1380 | rt = dst_alloc(&ipv4_dst_ops); |
| 1381 | if (rt == NULL) { |
| 1382 | ip_rt_put(rth); |
| 1383 | in_dev_put(in_dev); |
| 1384 | return; |
| 1385 | } |
| 1386 | |
| 1387 | /* Copy all the information. */ |
| 1388 | *rt = *rth; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1389 | INIT_RCU_HEAD(&rt->u.dst.rcu_head); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1390 | rt->u.dst.__use = 1; |
| 1391 | atomic_set(&rt->u.dst.__refcnt, 1); |
| 1392 | rt->u.dst.child = NULL; |
| 1393 | if (rt->u.dst.dev) |
| 1394 | dev_hold(rt->u.dst.dev); |
| 1395 | if (rt->idev) |
| 1396 | in_dev_hold(rt->idev); |
| 1397 | rt->u.dst.obsolete = 0; |
| 1398 | rt->u.dst.lastuse = jiffies; |
| 1399 | rt->u.dst.path = &rt->u.dst; |
| 1400 | rt->u.dst.neighbour = NULL; |
| 1401 | rt->u.dst.hh = NULL; |
Alexey Dobriyan | def8b4f | 2008-10-28 13:24:06 -0700 | [diff] [blame^] | 1402 | #ifdef CONFIG_XFRM |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1403 | rt->u.dst.xfrm = NULL; |
Alexey Dobriyan | def8b4f | 2008-10-28 13:24:06 -0700 | [diff] [blame^] | 1404 | #endif |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1405 | rt->rt_genid = rt_genid(net); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1406 | rt->rt_flags |= RTCF_REDIRECTED; |
| 1407 | |
| 1408 | /* Gateway is different ... */ |
| 1409 | rt->rt_gateway = new_gw; |
| 1410 | |
| 1411 | /* Redirect received -> path was valid */ |
| 1412 | dst_confirm(&rth->u.dst); |
| 1413 | |
| 1414 | if (rt->peer) |
| 1415 | atomic_inc(&rt->peer->refcnt); |
| 1416 | |
| 1417 | if (arp_bind_neighbour(&rt->u.dst) || |
| 1418 | !(rt->u.dst.neighbour->nud_state & |
| 1419 | NUD_VALID)) { |
| 1420 | if (rt->u.dst.neighbour) |
| 1421 | neigh_event_send(rt->u.dst.neighbour, NULL); |
| 1422 | ip_rt_put(rth); |
| 1423 | rt_drop(rt); |
| 1424 | goto do_next; |
| 1425 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1426 | |
Tom Tucker | 8d71740 | 2006-07-30 20:43:36 -0700 | [diff] [blame] | 1427 | netevent.old = &rth->u.dst; |
| 1428 | netevent.new = &rt->u.dst; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1429 | call_netevent_notifiers(NETEVENT_REDIRECT, |
| 1430 | &netevent); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1431 | |
| 1432 | rt_del(hash, rth); |
| 1433 | if (!rt_intern_hash(hash, rt, &rt)) |
| 1434 | ip_rt_put(rt); |
| 1435 | goto do_next; |
| 1436 | } |
| 1437 | rcu_read_unlock(); |
| 1438 | do_next: |
| 1439 | ; |
| 1440 | } |
| 1441 | } |
| 1442 | in_dev_put(in_dev); |
| 1443 | return; |
| 1444 | |
| 1445 | reject_redirect: |
| 1446 | #ifdef CONFIG_IP_ROUTE_VERBOSE |
| 1447 | if (IN_DEV_LOG_MARTIANS(in_dev) && net_ratelimit()) |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 1448 | printk(KERN_INFO "Redirect from " NIPQUAD_FMT " on %s about " |
| 1449 | NIPQUAD_FMT " ignored.\n" |
| 1450 | " Advised path = " NIPQUAD_FMT " -> " NIPQUAD_FMT "\n", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1451 | NIPQUAD(old_gw), dev->name, NIPQUAD(new_gw), |
Ilia Sotnikov | cef2685 | 2006-03-25 01:38:55 -0800 | [diff] [blame] | 1452 | NIPQUAD(saddr), NIPQUAD(daddr)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1453 | #endif |
| 1454 | in_dev_put(in_dev); |
| 1455 | } |
| 1456 | |
| 1457 | static struct dst_entry *ipv4_negative_advice(struct dst_entry *dst) |
| 1458 | { |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 1459 | struct rtable *rt = (struct rtable *)dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1460 | struct dst_entry *ret = dst; |
| 1461 | |
| 1462 | if (rt) { |
| 1463 | if (dst->obsolete) { |
| 1464 | ip_rt_put(rt); |
| 1465 | ret = NULL; |
| 1466 | } else if ((rt->rt_flags & RTCF_REDIRECTED) || |
| 1467 | rt->u.dst.expires) { |
Al Viro | 8c7bc84 | 2006-09-26 21:26:19 -0700 | [diff] [blame] | 1468 | unsigned hash = rt_hash(rt->fl.fl4_dst, rt->fl.fl4_src, |
Denis V. Lunev | b00180d | 2008-07-05 19:04:09 -0700 | [diff] [blame] | 1469 | rt->fl.oif, |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1470 | rt_genid(dev_net(dst->dev))); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1471 | #if RT_CACHE_DEBUG >= 1 |
Denis V. Lunev | 56c99d0 | 2007-12-06 02:19:07 -0800 | [diff] [blame] | 1472 | printk(KERN_DEBUG "ipv4_negative_advice: redirect to " |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 1473 | NIPQUAD_FMT "/%02x dropped\n", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1474 | NIPQUAD(rt->rt_dst), rt->fl.fl4_tos); |
| 1475 | #endif |
| 1476 | rt_del(hash, rt); |
| 1477 | ret = NULL; |
| 1478 | } |
| 1479 | } |
| 1480 | return ret; |
| 1481 | } |
| 1482 | |
| 1483 | /* |
| 1484 | * Algorithm: |
| 1485 | * 1. The first ip_rt_redirect_number redirects are sent |
| 1486 | * with exponential backoff, then we stop sending them at all, |
| 1487 | * assuming that the host ignores our redirects. |
| 1488 | * 2. If we did not see packets requiring redirects |
| 1489 | * during ip_rt_redirect_silence, we assume that the host |
| 1490 | * forgot redirected route and start to send redirects again. |
| 1491 | * |
| 1492 | * This algorithm is much cheaper and more intelligent than dumb load limiting |
| 1493 | * in icmp.c. |
| 1494 | * |
| 1495 | * NOTE. Do not forget to inhibit load limiting for redirects (redundant) |
| 1496 | * and "frag. need" (breaks PMTU discovery) in icmp.c. |
| 1497 | */ |
| 1498 | |
| 1499 | void ip_rt_send_redirect(struct sk_buff *skb) |
| 1500 | { |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 1501 | struct rtable *rt = skb->rtable; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1502 | struct in_device *in_dev = in_dev_get(rt->u.dst.dev); |
| 1503 | |
| 1504 | if (!in_dev) |
| 1505 | return; |
| 1506 | |
| 1507 | if (!IN_DEV_TX_REDIRECTS(in_dev)) |
| 1508 | goto out; |
| 1509 | |
| 1510 | /* No redirected packets during ip_rt_redirect_silence; |
| 1511 | * reset the algorithm. |
| 1512 | */ |
| 1513 | if (time_after(jiffies, rt->u.dst.rate_last + ip_rt_redirect_silence)) |
| 1514 | rt->u.dst.rate_tokens = 0; |
| 1515 | |
| 1516 | /* Too many ignored redirects; do not send anything |
| 1517 | * set u.dst.rate_last to the last seen redirected packet. |
| 1518 | */ |
| 1519 | if (rt->u.dst.rate_tokens >= ip_rt_redirect_number) { |
| 1520 | rt->u.dst.rate_last = jiffies; |
| 1521 | goto out; |
| 1522 | } |
| 1523 | |
| 1524 | /* Check for load limit; set rate_last to the latest sent |
| 1525 | * redirect. |
| 1526 | */ |
Li Yewang | 14fb8a7 | 2006-12-18 00:26:35 -0800 | [diff] [blame] | 1527 | if (rt->u.dst.rate_tokens == 0 || |
| 1528 | time_after(jiffies, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1529 | (rt->u.dst.rate_last + |
| 1530 | (ip_rt_redirect_load << rt->u.dst.rate_tokens)))) { |
| 1531 | icmp_send(skb, ICMP_REDIRECT, ICMP_REDIR_HOST, rt->rt_gateway); |
| 1532 | rt->u.dst.rate_last = jiffies; |
| 1533 | ++rt->u.dst.rate_tokens; |
| 1534 | #ifdef CONFIG_IP_ROUTE_VERBOSE |
| 1535 | if (IN_DEV_LOG_MARTIANS(in_dev) && |
| 1536 | rt->u.dst.rate_tokens == ip_rt_redirect_number && |
| 1537 | net_ratelimit()) |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 1538 | printk(KERN_WARNING "host " NIPQUAD_FMT "/if%d ignores " |
| 1539 | "redirects for " NIPQUAD_FMT " to " NIPQUAD_FMT ".\n", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1540 | NIPQUAD(rt->rt_src), rt->rt_iif, |
| 1541 | NIPQUAD(rt->rt_dst), NIPQUAD(rt->rt_gateway)); |
| 1542 | #endif |
| 1543 | } |
| 1544 | out: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1545 | in_dev_put(in_dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1546 | } |
| 1547 | |
| 1548 | static int ip_error(struct sk_buff *skb) |
| 1549 | { |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 1550 | struct rtable *rt = skb->rtable; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1551 | unsigned long now; |
| 1552 | int code; |
| 1553 | |
| 1554 | switch (rt->u.dst.error) { |
| 1555 | case EINVAL: |
| 1556 | default: |
| 1557 | goto out; |
| 1558 | case EHOSTUNREACH: |
| 1559 | code = ICMP_HOST_UNREACH; |
| 1560 | break; |
| 1561 | case ENETUNREACH: |
| 1562 | code = ICMP_NET_UNREACH; |
Pavel Emelyanov | 7c73a6f | 2008-07-16 20:20:11 -0700 | [diff] [blame] | 1563 | IP_INC_STATS_BH(dev_net(rt->u.dst.dev), |
| 1564 | IPSTATS_MIB_INNOROUTES); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1565 | break; |
| 1566 | case EACCES: |
| 1567 | code = ICMP_PKT_FILTERED; |
| 1568 | break; |
| 1569 | } |
| 1570 | |
| 1571 | now = jiffies; |
| 1572 | rt->u.dst.rate_tokens += now - rt->u.dst.rate_last; |
| 1573 | if (rt->u.dst.rate_tokens > ip_rt_error_burst) |
| 1574 | rt->u.dst.rate_tokens = ip_rt_error_burst; |
| 1575 | rt->u.dst.rate_last = now; |
| 1576 | if (rt->u.dst.rate_tokens >= ip_rt_error_cost) { |
| 1577 | rt->u.dst.rate_tokens -= ip_rt_error_cost; |
| 1578 | icmp_send(skb, ICMP_DEST_UNREACH, code, 0); |
| 1579 | } |
| 1580 | |
| 1581 | out: kfree_skb(skb); |
| 1582 | return 0; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1583 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1584 | |
| 1585 | /* |
| 1586 | * The last two values are not from the RFC but |
| 1587 | * are needed for AMPRnet AX.25 paths. |
| 1588 | */ |
| 1589 | |
Arjan van de Ven | 9b5b5cf | 2005-11-29 16:21:38 -0800 | [diff] [blame] | 1590 | static const unsigned short mtu_plateau[] = |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1591 | {32000, 17914, 8166, 4352, 2002, 1492, 576, 296, 216, 128 }; |
| 1592 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 1593 | static inline unsigned short guess_mtu(unsigned short old_mtu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1594 | { |
| 1595 | int i; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1596 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1597 | for (i = 0; i < ARRAY_SIZE(mtu_plateau); i++) |
| 1598 | if (old_mtu > mtu_plateau[i]) |
| 1599 | return mtu_plateau[i]; |
| 1600 | return 68; |
| 1601 | } |
| 1602 | |
Denis V. Lunev | b592191 | 2008-01-22 23:50:25 -0800 | [diff] [blame] | 1603 | unsigned short ip_rt_frag_needed(struct net *net, struct iphdr *iph, |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1604 | unsigned short new_mtu, |
| 1605 | struct net_device *dev) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1606 | { |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1607 | int i, k; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1608 | unsigned short old_mtu = ntohs(iph->tot_len); |
| 1609 | struct rtable *rth; |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1610 | int ikeys[2] = { dev->ifindex, 0 }; |
Al Viro | e448515 | 2006-09-26 22:15:01 -0700 | [diff] [blame] | 1611 | __be32 skeys[2] = { iph->saddr, 0, }; |
| 1612 | __be32 daddr = iph->daddr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1613 | unsigned short est_mtu = 0; |
| 1614 | |
| 1615 | if (ipv4_config.no_pmtu_disc) |
| 1616 | return 0; |
| 1617 | |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1618 | for (k = 0; k < 2; k++) { |
| 1619 | for (i = 0; i < 2; i++) { |
Denis V. Lunev | b00180d | 2008-07-05 19:04:09 -0700 | [diff] [blame] | 1620 | unsigned hash = rt_hash(daddr, skeys[i], ikeys[k], |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1621 | rt_genid(net)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1622 | |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1623 | rcu_read_lock(); |
| 1624 | for (rth = rcu_dereference(rt_hash_table[hash].chain); rth; |
| 1625 | rth = rcu_dereference(rth->u.dst.rt_next)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1626 | unsigned short mtu = new_mtu; |
| 1627 | |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1628 | if (rth->fl.fl4_dst != daddr || |
| 1629 | rth->fl.fl4_src != skeys[i] || |
| 1630 | rth->rt_dst != daddr || |
| 1631 | rth->rt_src != iph->saddr || |
| 1632 | rth->fl.oif != ikeys[k] || |
| 1633 | rth->fl.iif != 0 || |
| 1634 | dst_metric_locked(&rth->u.dst, RTAX_MTU) || |
| 1635 | !net_eq(dev_net(rth->u.dst.dev), net) || |
Hugh Dickins | 6c3b8fc | 2008-07-26 17:51:06 -0700 | [diff] [blame] | 1636 | rt_is_expired(rth)) |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1637 | continue; |
| 1638 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1639 | if (new_mtu < 68 || new_mtu >= old_mtu) { |
| 1640 | |
| 1641 | /* BSD 4.2 compatibility hack :-( */ |
| 1642 | if (mtu == 0 && |
Rami Rosen | 6d273f8 | 2008-08-06 02:33:49 -0700 | [diff] [blame] | 1643 | old_mtu >= dst_mtu(&rth->u.dst) && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1644 | old_mtu >= 68 + (iph->ihl << 2)) |
| 1645 | old_mtu -= iph->ihl << 2; |
| 1646 | |
| 1647 | mtu = guess_mtu(old_mtu); |
| 1648 | } |
Rami Rosen | 6d273f8 | 2008-08-06 02:33:49 -0700 | [diff] [blame] | 1649 | if (mtu <= dst_mtu(&rth->u.dst)) { |
| 1650 | if (mtu < dst_mtu(&rth->u.dst)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1651 | dst_confirm(&rth->u.dst); |
| 1652 | if (mtu < ip_rt_min_pmtu) { |
| 1653 | mtu = ip_rt_min_pmtu; |
| 1654 | rth->u.dst.metrics[RTAX_LOCK-1] |= |
| 1655 | (1 << RTAX_MTU); |
| 1656 | } |
| 1657 | rth->u.dst.metrics[RTAX_MTU-1] = mtu; |
| 1658 | dst_set_expires(&rth->u.dst, |
| 1659 | ip_rt_mtu_expires); |
| 1660 | } |
| 1661 | est_mtu = mtu; |
| 1662 | } |
| 1663 | } |
Timo Teras | 0010e46 | 2008-04-29 03:32:25 -0700 | [diff] [blame] | 1664 | rcu_read_unlock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1665 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1666 | } |
| 1667 | return est_mtu ? : new_mtu; |
| 1668 | } |
| 1669 | |
| 1670 | static void ip_rt_update_pmtu(struct dst_entry *dst, u32 mtu) |
| 1671 | { |
Rami Rosen | 6d273f8 | 2008-08-06 02:33:49 -0700 | [diff] [blame] | 1672 | if (dst_mtu(dst) > mtu && mtu >= 68 && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1673 | !(dst_metric_locked(dst, RTAX_MTU))) { |
| 1674 | if (mtu < ip_rt_min_pmtu) { |
| 1675 | mtu = ip_rt_min_pmtu; |
| 1676 | dst->metrics[RTAX_LOCK-1] |= (1 << RTAX_MTU); |
| 1677 | } |
| 1678 | dst->metrics[RTAX_MTU-1] = mtu; |
| 1679 | dst_set_expires(dst, ip_rt_mtu_expires); |
Tom Tucker | 8d71740 | 2006-07-30 20:43:36 -0700 | [diff] [blame] | 1680 | call_netevent_notifiers(NETEVENT_PMTU_UPDATE, dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1681 | } |
| 1682 | } |
| 1683 | |
| 1684 | static struct dst_entry *ipv4_dst_check(struct dst_entry *dst, u32 cookie) |
| 1685 | { |
| 1686 | return NULL; |
| 1687 | } |
| 1688 | |
| 1689 | static void ipv4_dst_destroy(struct dst_entry *dst) |
| 1690 | { |
| 1691 | struct rtable *rt = (struct rtable *) dst; |
| 1692 | struct inet_peer *peer = rt->peer; |
| 1693 | struct in_device *idev = rt->idev; |
| 1694 | |
| 1695 | if (peer) { |
| 1696 | rt->peer = NULL; |
| 1697 | inet_putpeer(peer); |
| 1698 | } |
| 1699 | |
| 1700 | if (idev) { |
| 1701 | rt->idev = NULL; |
| 1702 | in_dev_put(idev); |
| 1703 | } |
| 1704 | } |
| 1705 | |
| 1706 | static void ipv4_dst_ifdown(struct dst_entry *dst, struct net_device *dev, |
| 1707 | int how) |
| 1708 | { |
| 1709 | struct rtable *rt = (struct rtable *) dst; |
| 1710 | struct in_device *idev = rt->idev; |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 1711 | if (dev != dev_net(dev)->loopback_dev && idev && idev->dev == dev) { |
Denis V. Lunev | 5a3e55d | 2007-12-07 00:38:10 -0800 | [diff] [blame] | 1712 | struct in_device *loopback_idev = |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 1713 | in_dev_get(dev_net(dev)->loopback_dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1714 | if (loopback_idev) { |
| 1715 | rt->idev = loopback_idev; |
| 1716 | in_dev_put(idev); |
| 1717 | } |
| 1718 | } |
| 1719 | } |
| 1720 | |
| 1721 | static void ipv4_link_failure(struct sk_buff *skb) |
| 1722 | { |
| 1723 | struct rtable *rt; |
| 1724 | |
| 1725 | icmp_send(skb, ICMP_DEST_UNREACH, ICMP_HOST_UNREACH, 0); |
| 1726 | |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 1727 | rt = skb->rtable; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1728 | if (rt) |
| 1729 | dst_set_expires(&rt->u.dst, 0); |
| 1730 | } |
| 1731 | |
| 1732 | static int ip_rt_bug(struct sk_buff *skb) |
| 1733 | { |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 1734 | printk(KERN_DEBUG "ip_rt_bug: " NIPQUAD_FMT " -> " NIPQUAD_FMT ", %s\n", |
Arnaldo Carvalho de Melo | eddc9ec | 2007-04-20 22:47:35 -0700 | [diff] [blame] | 1735 | NIPQUAD(ip_hdr(skb)->saddr), NIPQUAD(ip_hdr(skb)->daddr), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1736 | skb->dev ? skb->dev->name : "?"); |
| 1737 | kfree_skb(skb); |
| 1738 | return 0; |
| 1739 | } |
| 1740 | |
| 1741 | /* |
| 1742 | We do not cache source address of outgoing interface, |
| 1743 | because it is used only by IP RR, TS and SRR options, |
| 1744 | so that it out of fast path. |
| 1745 | |
| 1746 | BTW remember: "addr" is allowed to be not aligned |
| 1747 | in IP options! |
| 1748 | */ |
| 1749 | |
| 1750 | void ip_rt_get_source(u8 *addr, struct rtable *rt) |
| 1751 | { |
Al Viro | a61ced5 | 2006-09-26 21:27:54 -0700 | [diff] [blame] | 1752 | __be32 src; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1753 | struct fib_result res; |
| 1754 | |
| 1755 | if (rt->fl.iif == 0) |
| 1756 | src = rt->rt_src; |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 1757 | else if (fib_lookup(dev_net(rt->u.dst.dev), &rt->fl, &res) == 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1758 | src = FIB_RES_PREFSRC(res); |
| 1759 | fib_res_put(&res); |
| 1760 | } else |
| 1761 | src = inet_select_addr(rt->u.dst.dev, rt->rt_gateway, |
| 1762 | RT_SCOPE_UNIVERSE); |
| 1763 | memcpy(addr, &src, 4); |
| 1764 | } |
| 1765 | |
| 1766 | #ifdef CONFIG_NET_CLS_ROUTE |
| 1767 | static void set_class_tag(struct rtable *rt, u32 tag) |
| 1768 | { |
| 1769 | if (!(rt->u.dst.tclassid & 0xFFFF)) |
| 1770 | rt->u.dst.tclassid |= tag & 0xFFFF; |
| 1771 | if (!(rt->u.dst.tclassid & 0xFFFF0000)) |
| 1772 | rt->u.dst.tclassid |= tag & 0xFFFF0000; |
| 1773 | } |
| 1774 | #endif |
| 1775 | |
| 1776 | static void rt_set_nexthop(struct rtable *rt, struct fib_result *res, u32 itag) |
| 1777 | { |
| 1778 | struct fib_info *fi = res->fi; |
| 1779 | |
| 1780 | if (fi) { |
| 1781 | if (FIB_RES_GW(*res) && |
| 1782 | FIB_RES_NH(*res).nh_scope == RT_SCOPE_LINK) |
| 1783 | rt->rt_gateway = FIB_RES_GW(*res); |
| 1784 | memcpy(rt->u.dst.metrics, fi->fib_metrics, |
| 1785 | sizeof(rt->u.dst.metrics)); |
| 1786 | if (fi->fib_mtu == 0) { |
| 1787 | rt->u.dst.metrics[RTAX_MTU-1] = rt->u.dst.dev->mtu; |
Satoru SATOH | 0bbeafd | 2008-05-04 22:12:43 -0700 | [diff] [blame] | 1788 | if (dst_metric_locked(&rt->u.dst, RTAX_MTU) && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1789 | rt->rt_gateway != rt->rt_dst && |
| 1790 | rt->u.dst.dev->mtu > 576) |
| 1791 | rt->u.dst.metrics[RTAX_MTU-1] = 576; |
| 1792 | } |
| 1793 | #ifdef CONFIG_NET_CLS_ROUTE |
| 1794 | rt->u.dst.tclassid = FIB_RES_NH(*res).nh_tclassid; |
| 1795 | #endif |
| 1796 | } else |
| 1797 | rt->u.dst.metrics[RTAX_MTU-1]= rt->u.dst.dev->mtu; |
| 1798 | |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 1799 | if (dst_metric(&rt->u.dst, RTAX_HOPLIMIT) == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1800 | rt->u.dst.metrics[RTAX_HOPLIMIT-1] = sysctl_ip_default_ttl; |
Rami Rosen | 6d273f8 | 2008-08-06 02:33:49 -0700 | [diff] [blame] | 1801 | if (dst_mtu(&rt->u.dst) > IP_MAX_MTU) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1802 | rt->u.dst.metrics[RTAX_MTU-1] = IP_MAX_MTU; |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 1803 | if (dst_metric(&rt->u.dst, RTAX_ADVMSS) == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1804 | rt->u.dst.metrics[RTAX_ADVMSS-1] = max_t(unsigned int, rt->u.dst.dev->mtu - 40, |
| 1805 | ip_rt_min_advmss); |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 1806 | if (dst_metric(&rt->u.dst, RTAX_ADVMSS) > 65535 - 40) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1807 | rt->u.dst.metrics[RTAX_ADVMSS-1] = 65535 - 40; |
| 1808 | |
| 1809 | #ifdef CONFIG_NET_CLS_ROUTE |
| 1810 | #ifdef CONFIG_IP_MULTIPLE_TABLES |
| 1811 | set_class_tag(rt, fib_rules_tclass(res)); |
| 1812 | #endif |
| 1813 | set_class_tag(rt, itag); |
| 1814 | #endif |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1815 | rt->rt_type = res->type; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1816 | } |
| 1817 | |
Al Viro | 9e12bb2 | 2006-09-26 21:25:20 -0700 | [diff] [blame] | 1818 | static int ip_route_input_mc(struct sk_buff *skb, __be32 daddr, __be32 saddr, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1819 | u8 tos, struct net_device *dev, int our) |
| 1820 | { |
| 1821 | unsigned hash; |
| 1822 | struct rtable *rth; |
Al Viro | a61ced5 | 2006-09-26 21:27:54 -0700 | [diff] [blame] | 1823 | __be32 spec_dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1824 | struct in_device *in_dev = in_dev_get(dev); |
| 1825 | u32 itag = 0; |
| 1826 | |
| 1827 | /* Primary sanity checks. */ |
| 1828 | |
| 1829 | if (in_dev == NULL) |
| 1830 | return -EINVAL; |
| 1831 | |
Jan Engelhardt | 1e637c7 | 2008-01-21 03:18:08 -0800 | [diff] [blame] | 1832 | if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) || |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 1833 | ipv4_is_loopback(saddr) || skb->protocol != htons(ETH_P_IP)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1834 | goto e_inval; |
| 1835 | |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 1836 | if (ipv4_is_zeronet(saddr)) { |
| 1837 | if (!ipv4_is_local_multicast(daddr)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1838 | goto e_inval; |
| 1839 | spec_dst = inet_select_addr(dev, 0, RT_SCOPE_LINK); |
| 1840 | } else if (fib_validate_source(saddr, 0, tos, 0, |
| 1841 | dev, &spec_dst, &itag) < 0) |
| 1842 | goto e_inval; |
| 1843 | |
| 1844 | rth = dst_alloc(&ipv4_dst_ops); |
| 1845 | if (!rth) |
| 1846 | goto e_nobufs; |
| 1847 | |
| 1848 | rth->u.dst.output= ip_rt_bug; |
| 1849 | |
| 1850 | atomic_set(&rth->u.dst.__refcnt, 1); |
| 1851 | rth->u.dst.flags= DST_HOST; |
Herbert Xu | 42f811b | 2007-06-04 23:34:44 -0700 | [diff] [blame] | 1852 | if (IN_DEV_CONF_GET(in_dev, NOPOLICY)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1853 | rth->u.dst.flags |= DST_NOPOLICY; |
| 1854 | rth->fl.fl4_dst = daddr; |
| 1855 | rth->rt_dst = daddr; |
| 1856 | rth->fl.fl4_tos = tos; |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 1857 | rth->fl.mark = skb->mark; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1858 | rth->fl.fl4_src = saddr; |
| 1859 | rth->rt_src = saddr; |
| 1860 | #ifdef CONFIG_NET_CLS_ROUTE |
| 1861 | rth->u.dst.tclassid = itag; |
| 1862 | #endif |
| 1863 | rth->rt_iif = |
| 1864 | rth->fl.iif = dev->ifindex; |
Eric W. Biederman | 2774c7a | 2007-09-26 22:10:56 -0700 | [diff] [blame] | 1865 | rth->u.dst.dev = init_net.loopback_dev; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1866 | dev_hold(rth->u.dst.dev); |
| 1867 | rth->idev = in_dev_get(rth->u.dst.dev); |
| 1868 | rth->fl.oif = 0; |
| 1869 | rth->rt_gateway = daddr; |
| 1870 | rth->rt_spec_dst= spec_dst; |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1871 | rth->rt_genid = rt_genid(dev_net(dev)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1872 | rth->rt_flags = RTCF_MULTICAST; |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 1873 | rth->rt_type = RTN_MULTICAST; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1874 | if (our) { |
| 1875 | rth->u.dst.input= ip_local_deliver; |
| 1876 | rth->rt_flags |= RTCF_LOCAL; |
| 1877 | } |
| 1878 | |
| 1879 | #ifdef CONFIG_IP_MROUTE |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 1880 | if (!ipv4_is_local_multicast(daddr) && IN_DEV_MFORWARD(in_dev)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1881 | rth->u.dst.input = ip_mr_input; |
| 1882 | #endif |
| 1883 | RT_CACHE_STAT_INC(in_slow_mc); |
| 1884 | |
| 1885 | in_dev_put(in_dev); |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 1886 | hash = rt_hash(daddr, saddr, dev->ifindex, rt_genid(dev_net(dev))); |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 1887 | return rt_intern_hash(hash, rth, &skb->rtable); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1888 | |
| 1889 | e_nobufs: |
| 1890 | in_dev_put(in_dev); |
| 1891 | return -ENOBUFS; |
| 1892 | |
| 1893 | e_inval: |
| 1894 | in_dev_put(in_dev); |
| 1895 | return -EINVAL; |
| 1896 | } |
| 1897 | |
| 1898 | |
| 1899 | static void ip_handle_martian_source(struct net_device *dev, |
| 1900 | struct in_device *in_dev, |
| 1901 | struct sk_buff *skb, |
Al Viro | 9e12bb2 | 2006-09-26 21:25:20 -0700 | [diff] [blame] | 1902 | __be32 daddr, |
| 1903 | __be32 saddr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1904 | { |
| 1905 | RT_CACHE_STAT_INC(in_martian_src); |
| 1906 | #ifdef CONFIG_IP_ROUTE_VERBOSE |
| 1907 | if (IN_DEV_LOG_MARTIANS(in_dev) && net_ratelimit()) { |
| 1908 | /* |
| 1909 | * RFC1812 recommendation, if source is martian, |
| 1910 | * the only hint is MAC header. |
| 1911 | */ |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 1912 | printk(KERN_WARNING "martian source " NIPQUAD_FMT " from " |
| 1913 | NIPQUAD_FMT", on dev %s\n", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1914 | NIPQUAD(daddr), NIPQUAD(saddr), dev->name); |
Arnaldo Carvalho de Melo | 98e399f | 2007-03-19 15:33:04 -0700 | [diff] [blame] | 1915 | if (dev->hard_header_len && skb_mac_header_was_set(skb)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1916 | int i; |
Arnaldo Carvalho de Melo | 98e399f | 2007-03-19 15:33:04 -0700 | [diff] [blame] | 1917 | const unsigned char *p = skb_mac_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1918 | printk(KERN_WARNING "ll header: "); |
| 1919 | for (i = 0; i < dev->hard_header_len; i++, p++) { |
| 1920 | printk("%02x", *p); |
| 1921 | if (i < (dev->hard_header_len - 1)) |
| 1922 | printk(":"); |
| 1923 | } |
| 1924 | printk("\n"); |
| 1925 | } |
| 1926 | } |
| 1927 | #endif |
| 1928 | } |
| 1929 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 1930 | static int __mkroute_input(struct sk_buff *skb, |
| 1931 | struct fib_result *res, |
| 1932 | struct in_device *in_dev, |
| 1933 | __be32 daddr, __be32 saddr, u32 tos, |
| 1934 | struct rtable **result) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1935 | { |
| 1936 | |
| 1937 | struct rtable *rth; |
| 1938 | int err; |
| 1939 | struct in_device *out_dev; |
| 1940 | unsigned flags = 0; |
Al Viro | d9c9df8 | 2006-09-26 21:28:14 -0700 | [diff] [blame] | 1941 | __be32 spec_dst; |
| 1942 | u32 itag; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1943 | |
| 1944 | /* get a working reference to the output device */ |
| 1945 | out_dev = in_dev_get(FIB_RES_DEV(*res)); |
| 1946 | if (out_dev == NULL) { |
| 1947 | if (net_ratelimit()) |
| 1948 | printk(KERN_CRIT "Bug in ip_route_input" \ |
| 1949 | "_slow(). Please, report\n"); |
| 1950 | return -EINVAL; |
| 1951 | } |
| 1952 | |
| 1953 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1954 | err = fib_validate_source(saddr, daddr, tos, FIB_RES_OIF(*res), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1955 | in_dev->dev, &spec_dst, &itag); |
| 1956 | if (err < 0) { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1957 | ip_handle_martian_source(in_dev->dev, in_dev, skb, daddr, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1958 | saddr); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1959 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1960 | err = -EINVAL; |
| 1961 | goto cleanup; |
| 1962 | } |
| 1963 | |
| 1964 | if (err) |
| 1965 | flags |= RTCF_DIRECTSRC; |
| 1966 | |
Thomas Graf | 51b77ca | 2008-06-03 16:36:01 -0700 | [diff] [blame] | 1967 | if (out_dev == in_dev && err && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1968 | (IN_DEV_SHARED_MEDIA(out_dev) || |
| 1969 | inet_addr_onlink(out_dev, saddr, FIB_RES_GW(*res)))) |
| 1970 | flags |= RTCF_DOREDIRECT; |
| 1971 | |
| 1972 | if (skb->protocol != htons(ETH_P_IP)) { |
| 1973 | /* Not IP (i.e. ARP). Do not create route, if it is |
| 1974 | * invalid for proxy arp. DNAT routes are always valid. |
| 1975 | */ |
Rami Rosen | cb7928a | 2008-01-09 00:18:24 -0800 | [diff] [blame] | 1976 | if (out_dev == in_dev) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1977 | err = -EINVAL; |
| 1978 | goto cleanup; |
| 1979 | } |
| 1980 | } |
| 1981 | |
| 1982 | |
| 1983 | rth = dst_alloc(&ipv4_dst_ops); |
| 1984 | if (!rth) { |
| 1985 | err = -ENOBUFS; |
| 1986 | goto cleanup; |
| 1987 | } |
| 1988 | |
Julian Anastasov | ce723d8 | 2005-09-08 13:34:47 -0700 | [diff] [blame] | 1989 | atomic_set(&rth->u.dst.__refcnt, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1990 | rth->u.dst.flags= DST_HOST; |
Herbert Xu | 42f811b | 2007-06-04 23:34:44 -0700 | [diff] [blame] | 1991 | if (IN_DEV_CONF_GET(in_dev, NOPOLICY)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1992 | rth->u.dst.flags |= DST_NOPOLICY; |
Herbert Xu | 42f811b | 2007-06-04 23:34:44 -0700 | [diff] [blame] | 1993 | if (IN_DEV_CONF_GET(out_dev, NOXFRM)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1994 | rth->u.dst.flags |= DST_NOXFRM; |
| 1995 | rth->fl.fl4_dst = daddr; |
| 1996 | rth->rt_dst = daddr; |
| 1997 | rth->fl.fl4_tos = tos; |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 1998 | rth->fl.mark = skb->mark; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1999 | rth->fl.fl4_src = saddr; |
| 2000 | rth->rt_src = saddr; |
| 2001 | rth->rt_gateway = daddr; |
| 2002 | rth->rt_iif = |
| 2003 | rth->fl.iif = in_dev->dev->ifindex; |
| 2004 | rth->u.dst.dev = (out_dev)->dev; |
| 2005 | dev_hold(rth->u.dst.dev); |
| 2006 | rth->idev = in_dev_get(rth->u.dst.dev); |
| 2007 | rth->fl.oif = 0; |
| 2008 | rth->rt_spec_dst= spec_dst; |
| 2009 | |
| 2010 | rth->u.dst.input = ip_forward; |
| 2011 | rth->u.dst.output = ip_output; |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2012 | rth->rt_genid = rt_genid(dev_net(rth->u.dst.dev)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2013 | |
| 2014 | rt_set_nexthop(rth, res, itag); |
| 2015 | |
| 2016 | rth->rt_flags = flags; |
| 2017 | |
| 2018 | *result = rth; |
| 2019 | err = 0; |
| 2020 | cleanup: |
| 2021 | /* release the working reference to the output device */ |
| 2022 | in_dev_put(out_dev); |
| 2023 | return err; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2024 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2025 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 2026 | static int ip_mkroute_input(struct sk_buff *skb, |
| 2027 | struct fib_result *res, |
| 2028 | const struct flowi *fl, |
| 2029 | struct in_device *in_dev, |
| 2030 | __be32 daddr, __be32 saddr, u32 tos) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2031 | { |
Chuck Short | 7abaa27 | 2005-06-22 22:10:23 -0700 | [diff] [blame] | 2032 | struct rtable* rth = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2033 | int err; |
| 2034 | unsigned hash; |
| 2035 | |
| 2036 | #ifdef CONFIG_IP_ROUTE_MULTIPATH |
| 2037 | if (res->fi && res->fi->fib_nhs > 1 && fl->oif == 0) |
| 2038 | fib_select_multipath(fl, res); |
| 2039 | #endif |
| 2040 | |
| 2041 | /* create a routing cache entry */ |
| 2042 | err = __mkroute_input(skb, res, in_dev, daddr, saddr, tos, &rth); |
| 2043 | if (err) |
| 2044 | return err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2045 | |
| 2046 | /* put it into the cache */ |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2047 | hash = rt_hash(daddr, saddr, fl->iif, |
| 2048 | rt_genid(dev_net(rth->u.dst.dev))); |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 2049 | return rt_intern_hash(hash, rth, &skb->rtable); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2050 | } |
| 2051 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2052 | /* |
| 2053 | * NOTE. We drop all the packets that has local source |
| 2054 | * addresses, because every properly looped back packet |
| 2055 | * must have correct destination already attached by output routine. |
| 2056 | * |
| 2057 | * Such approach solves two big problems: |
| 2058 | * 1. Not simplex devices are handled properly. |
| 2059 | * 2. IP spoofing attempts are filtered with 100% of guarantee. |
| 2060 | */ |
| 2061 | |
Al Viro | 9e12bb2 | 2006-09-26 21:25:20 -0700 | [diff] [blame] | 2062 | static int ip_route_input_slow(struct sk_buff *skb, __be32 daddr, __be32 saddr, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2063 | u8 tos, struct net_device *dev) |
| 2064 | { |
| 2065 | struct fib_result res; |
| 2066 | struct in_device *in_dev = in_dev_get(dev); |
| 2067 | struct flowi fl = { .nl_u = { .ip4_u = |
| 2068 | { .daddr = daddr, |
| 2069 | .saddr = saddr, |
| 2070 | .tos = tos, |
| 2071 | .scope = RT_SCOPE_UNIVERSE, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2072 | } }, |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 2073 | .mark = skb->mark, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2074 | .iif = dev->ifindex }; |
| 2075 | unsigned flags = 0; |
| 2076 | u32 itag = 0; |
| 2077 | struct rtable * rth; |
| 2078 | unsigned hash; |
Al Viro | 9e12bb2 | 2006-09-26 21:25:20 -0700 | [diff] [blame] | 2079 | __be32 spec_dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2080 | int err = -EINVAL; |
| 2081 | int free_res = 0; |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 2082 | struct net * net = dev_net(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2083 | |
| 2084 | /* IP on this device is disabled. */ |
| 2085 | |
| 2086 | if (!in_dev) |
| 2087 | goto out; |
| 2088 | |
| 2089 | /* Check for the most weird martians, which can be not detected |
| 2090 | by fib_lookup. |
| 2091 | */ |
| 2092 | |
Jan Engelhardt | 1e637c7 | 2008-01-21 03:18:08 -0800 | [diff] [blame] | 2093 | if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) || |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2094 | ipv4_is_loopback(saddr)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2095 | goto martian_source; |
| 2096 | |
Al Viro | e448515 | 2006-09-26 22:15:01 -0700 | [diff] [blame] | 2097 | if (daddr == htonl(0xFFFFFFFF) || (saddr == 0 && daddr == 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2098 | goto brd_input; |
| 2099 | |
| 2100 | /* Accept zero addresses only to limited broadcast; |
| 2101 | * I even do not know to fix it or not. Waiting for complains :-) |
| 2102 | */ |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2103 | if (ipv4_is_zeronet(saddr)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2104 | goto martian_source; |
| 2105 | |
Jan Engelhardt | 1e637c7 | 2008-01-21 03:18:08 -0800 | [diff] [blame] | 2106 | if (ipv4_is_lbcast(daddr) || ipv4_is_zeronet(daddr) || |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2107 | ipv4_is_loopback(daddr)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2108 | goto martian_destination; |
| 2109 | |
| 2110 | /* |
| 2111 | * Now we are ready to route packet. |
| 2112 | */ |
Denis V. Lunev | 84a885f | 2008-01-21 17:34:35 -0800 | [diff] [blame] | 2113 | if ((err = fib_lookup(net, &fl, &res)) != 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2114 | if (!IN_DEV_FORWARD(in_dev)) |
Dietmar Eggemann | 2c2910a | 2005-06-28 13:06:23 -0700 | [diff] [blame] | 2115 | goto e_hostunreach; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2116 | goto no_route; |
| 2117 | } |
| 2118 | free_res = 1; |
| 2119 | |
| 2120 | RT_CACHE_STAT_INC(in_slow_tot); |
| 2121 | |
| 2122 | if (res.type == RTN_BROADCAST) |
| 2123 | goto brd_input; |
| 2124 | |
| 2125 | if (res.type == RTN_LOCAL) { |
| 2126 | int result; |
| 2127 | result = fib_validate_source(saddr, daddr, tos, |
Denis V. Lunev | 84a885f | 2008-01-21 17:34:35 -0800 | [diff] [blame] | 2128 | net->loopback_dev->ifindex, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2129 | dev, &spec_dst, &itag); |
| 2130 | if (result < 0) |
| 2131 | goto martian_source; |
| 2132 | if (result) |
| 2133 | flags |= RTCF_DIRECTSRC; |
| 2134 | spec_dst = daddr; |
| 2135 | goto local_input; |
| 2136 | } |
| 2137 | |
| 2138 | if (!IN_DEV_FORWARD(in_dev)) |
Dietmar Eggemann | 2c2910a | 2005-06-28 13:06:23 -0700 | [diff] [blame] | 2139 | goto e_hostunreach; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2140 | if (res.type != RTN_UNICAST) |
| 2141 | goto martian_destination; |
| 2142 | |
| 2143 | err = ip_mkroute_input(skb, &res, &fl, in_dev, daddr, saddr, tos); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2144 | done: |
| 2145 | in_dev_put(in_dev); |
| 2146 | if (free_res) |
| 2147 | fib_res_put(&res); |
| 2148 | out: return err; |
| 2149 | |
| 2150 | brd_input: |
| 2151 | if (skb->protocol != htons(ETH_P_IP)) |
| 2152 | goto e_inval; |
| 2153 | |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2154 | if (ipv4_is_zeronet(saddr)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2155 | spec_dst = inet_select_addr(dev, 0, RT_SCOPE_LINK); |
| 2156 | else { |
| 2157 | err = fib_validate_source(saddr, 0, tos, 0, dev, &spec_dst, |
| 2158 | &itag); |
| 2159 | if (err < 0) |
| 2160 | goto martian_source; |
| 2161 | if (err) |
| 2162 | flags |= RTCF_DIRECTSRC; |
| 2163 | } |
| 2164 | flags |= RTCF_BROADCAST; |
| 2165 | res.type = RTN_BROADCAST; |
| 2166 | RT_CACHE_STAT_INC(in_brd); |
| 2167 | |
| 2168 | local_input: |
| 2169 | rth = dst_alloc(&ipv4_dst_ops); |
| 2170 | if (!rth) |
| 2171 | goto e_nobufs; |
| 2172 | |
| 2173 | rth->u.dst.output= ip_rt_bug; |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2174 | rth->rt_genid = rt_genid(net); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2175 | |
| 2176 | atomic_set(&rth->u.dst.__refcnt, 1); |
| 2177 | rth->u.dst.flags= DST_HOST; |
Herbert Xu | 42f811b | 2007-06-04 23:34:44 -0700 | [diff] [blame] | 2178 | if (IN_DEV_CONF_GET(in_dev, NOPOLICY)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2179 | rth->u.dst.flags |= DST_NOPOLICY; |
| 2180 | rth->fl.fl4_dst = daddr; |
| 2181 | rth->rt_dst = daddr; |
| 2182 | rth->fl.fl4_tos = tos; |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 2183 | rth->fl.mark = skb->mark; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2184 | rth->fl.fl4_src = saddr; |
| 2185 | rth->rt_src = saddr; |
| 2186 | #ifdef CONFIG_NET_CLS_ROUTE |
| 2187 | rth->u.dst.tclassid = itag; |
| 2188 | #endif |
| 2189 | rth->rt_iif = |
| 2190 | rth->fl.iif = dev->ifindex; |
Denis V. Lunev | 84a885f | 2008-01-21 17:34:35 -0800 | [diff] [blame] | 2191 | rth->u.dst.dev = net->loopback_dev; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2192 | dev_hold(rth->u.dst.dev); |
| 2193 | rth->idev = in_dev_get(rth->u.dst.dev); |
| 2194 | rth->rt_gateway = daddr; |
| 2195 | rth->rt_spec_dst= spec_dst; |
| 2196 | rth->u.dst.input= ip_local_deliver; |
| 2197 | rth->rt_flags = flags|RTCF_LOCAL; |
| 2198 | if (res.type == RTN_UNREACHABLE) { |
| 2199 | rth->u.dst.input= ip_error; |
| 2200 | rth->u.dst.error= -err; |
| 2201 | rth->rt_flags &= ~RTCF_LOCAL; |
| 2202 | } |
| 2203 | rth->rt_type = res.type; |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2204 | hash = rt_hash(daddr, saddr, fl.iif, rt_genid(net)); |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 2205 | err = rt_intern_hash(hash, rth, &skb->rtable); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2206 | goto done; |
| 2207 | |
| 2208 | no_route: |
| 2209 | RT_CACHE_STAT_INC(in_no_route); |
| 2210 | spec_dst = inet_select_addr(dev, 0, RT_SCOPE_UNIVERSE); |
| 2211 | res.type = RTN_UNREACHABLE; |
Mitsuru Chinen | 7f53878 | 2007-12-07 01:07:24 -0800 | [diff] [blame] | 2212 | if (err == -ESRCH) |
| 2213 | err = -ENETUNREACH; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2214 | goto local_input; |
| 2215 | |
| 2216 | /* |
| 2217 | * Do not cache martian addresses: they should be logged (RFC1812) |
| 2218 | */ |
| 2219 | martian_destination: |
| 2220 | RT_CACHE_STAT_INC(in_martian_dst); |
| 2221 | #ifdef CONFIG_IP_ROUTE_VERBOSE |
| 2222 | if (IN_DEV_LOG_MARTIANS(in_dev) && net_ratelimit()) |
YOSHIFUJI Hideaki | a7d632b | 2008-04-14 04:09:00 -0700 | [diff] [blame] | 2223 | printk(KERN_WARNING "martian destination " NIPQUAD_FMT " from " |
| 2224 | NIPQUAD_FMT ", dev %s\n", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2225 | NIPQUAD(daddr), NIPQUAD(saddr), dev->name); |
| 2226 | #endif |
Dietmar Eggemann | 2c2910a | 2005-06-28 13:06:23 -0700 | [diff] [blame] | 2227 | |
| 2228 | e_hostunreach: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2229 | err = -EHOSTUNREACH; |
| 2230 | goto done; |
Dietmar Eggemann | 2c2910a | 2005-06-28 13:06:23 -0700 | [diff] [blame] | 2231 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2232 | e_inval: |
| 2233 | err = -EINVAL; |
| 2234 | goto done; |
| 2235 | |
| 2236 | e_nobufs: |
| 2237 | err = -ENOBUFS; |
| 2238 | goto done; |
| 2239 | |
| 2240 | martian_source: |
| 2241 | ip_handle_martian_source(dev, in_dev, skb, daddr, saddr); |
| 2242 | goto e_inval; |
| 2243 | } |
| 2244 | |
Al Viro | 9e12bb2 | 2006-09-26 21:25:20 -0700 | [diff] [blame] | 2245 | int ip_route_input(struct sk_buff *skb, __be32 daddr, __be32 saddr, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2246 | u8 tos, struct net_device *dev) |
| 2247 | { |
| 2248 | struct rtable * rth; |
| 2249 | unsigned hash; |
| 2250 | int iif = dev->ifindex; |
Denis V. Lunev | b592191 | 2008-01-22 23:50:25 -0800 | [diff] [blame] | 2251 | struct net *net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2252 | |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 2253 | net = dev_net(dev); |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 2254 | |
| 2255 | if (!rt_caching(net)) |
| 2256 | goto skip_cache; |
| 2257 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2258 | tos &= IPTOS_RT_MASK; |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2259 | hash = rt_hash(daddr, saddr, iif, rt_genid(net)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2260 | |
| 2261 | rcu_read_lock(); |
| 2262 | for (rth = rcu_dereference(rt_hash_table[hash].chain); rth; |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 2263 | rth = rcu_dereference(rth->u.dst.rt_next)) { |
Stephen Hemminger | c0b8c32 | 2008-04-10 04:00:28 -0700 | [diff] [blame] | 2264 | if (((rth->fl.fl4_dst ^ daddr) | |
| 2265 | (rth->fl.fl4_src ^ saddr) | |
| 2266 | (rth->fl.iif ^ iif) | |
| 2267 | rth->fl.oif | |
| 2268 | (rth->fl.fl4_tos ^ tos)) == 0 && |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 2269 | rth->fl.mark == skb->mark && |
YOSHIFUJI Hideaki | 878628f | 2008-03-26 03:57:35 +0900 | [diff] [blame] | 2270 | net_eq(dev_net(rth->u.dst.dev), net) && |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2271 | !rt_is_expired(rth)) { |
Pavel Emelyanov | 03f49f3 | 2007-11-10 21:28:34 -0800 | [diff] [blame] | 2272 | dst_use(&rth->u.dst, jiffies); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2273 | RT_CACHE_STAT_INC(in_hit); |
| 2274 | rcu_read_unlock(); |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 2275 | skb->rtable = rth; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2276 | return 0; |
| 2277 | } |
| 2278 | RT_CACHE_STAT_INC(in_hlist_search); |
| 2279 | } |
| 2280 | rcu_read_unlock(); |
| 2281 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 2282 | skip_cache: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2283 | /* Multicast recognition logic is moved from route cache to here. |
| 2284 | The problem was that too many Ethernet cards have broken/missing |
| 2285 | hardware multicast filters :-( As result the host on multicasting |
| 2286 | network acquires a lot of useless route cache entries, sort of |
| 2287 | SDR messages from all the world. Now we try to get rid of them. |
| 2288 | Really, provided software IP multicast filter is organized |
| 2289 | reasonably (at least, hashed), it does not result in a slowdown |
| 2290 | comparing with route cache reject entries. |
| 2291 | Note, that multicast routers are not affected, because |
| 2292 | route cache entry is created eventually. |
| 2293 | */ |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2294 | if (ipv4_is_multicast(daddr)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2295 | struct in_device *in_dev; |
| 2296 | |
| 2297 | rcu_read_lock(); |
Herbert Xu | e5ed639 | 2005-10-03 14:35:55 -0700 | [diff] [blame] | 2298 | if ((in_dev = __in_dev_get_rcu(dev)) != NULL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2299 | int our = ip_check_mc(in_dev, daddr, saddr, |
Arnaldo Carvalho de Melo | eddc9ec | 2007-04-20 22:47:35 -0700 | [diff] [blame] | 2300 | ip_hdr(skb)->protocol); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2301 | if (our |
| 2302 | #ifdef CONFIG_IP_MROUTE |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2303 | || (!ipv4_is_local_multicast(daddr) && |
| 2304 | IN_DEV_MFORWARD(in_dev)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2305 | #endif |
| 2306 | ) { |
| 2307 | rcu_read_unlock(); |
| 2308 | return ip_route_input_mc(skb, daddr, saddr, |
| 2309 | tos, dev, our); |
| 2310 | } |
| 2311 | } |
| 2312 | rcu_read_unlock(); |
| 2313 | return -EINVAL; |
| 2314 | } |
| 2315 | return ip_route_input_slow(skb, daddr, saddr, tos, dev); |
| 2316 | } |
| 2317 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 2318 | static int __mkroute_output(struct rtable **result, |
| 2319 | struct fib_result *res, |
| 2320 | const struct flowi *fl, |
| 2321 | const struct flowi *oldflp, |
| 2322 | struct net_device *dev_out, |
| 2323 | unsigned flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2324 | { |
| 2325 | struct rtable *rth; |
| 2326 | struct in_device *in_dev; |
| 2327 | u32 tos = RT_FL_TOS(oldflp); |
| 2328 | int err = 0; |
| 2329 | |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2330 | if (ipv4_is_loopback(fl->fl4_src) && !(dev_out->flags&IFF_LOOPBACK)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2331 | return -EINVAL; |
| 2332 | |
Al Viro | e448515 | 2006-09-26 22:15:01 -0700 | [diff] [blame] | 2333 | if (fl->fl4_dst == htonl(0xFFFFFFFF)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2334 | res->type = RTN_BROADCAST; |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2335 | else if (ipv4_is_multicast(fl->fl4_dst)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2336 | res->type = RTN_MULTICAST; |
Jan Engelhardt | 1e637c7 | 2008-01-21 03:18:08 -0800 | [diff] [blame] | 2337 | else if (ipv4_is_lbcast(fl->fl4_dst) || ipv4_is_zeronet(fl->fl4_dst)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2338 | return -EINVAL; |
| 2339 | |
| 2340 | if (dev_out->flags & IFF_LOOPBACK) |
| 2341 | flags |= RTCF_LOCAL; |
| 2342 | |
| 2343 | /* get work reference to inet device */ |
| 2344 | in_dev = in_dev_get(dev_out); |
| 2345 | if (!in_dev) |
| 2346 | return -EINVAL; |
| 2347 | |
| 2348 | if (res->type == RTN_BROADCAST) { |
| 2349 | flags |= RTCF_BROADCAST | RTCF_LOCAL; |
| 2350 | if (res->fi) { |
| 2351 | fib_info_put(res->fi); |
| 2352 | res->fi = NULL; |
| 2353 | } |
| 2354 | } else if (res->type == RTN_MULTICAST) { |
| 2355 | flags |= RTCF_MULTICAST|RTCF_LOCAL; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2356 | if (!ip_check_mc(in_dev, oldflp->fl4_dst, oldflp->fl4_src, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2357 | oldflp->proto)) |
| 2358 | flags &= ~RTCF_LOCAL; |
| 2359 | /* If multicast route do not exist use |
| 2360 | default one, but do not gateway in this case. |
| 2361 | Yes, it is hack. |
| 2362 | */ |
| 2363 | if (res->fi && res->prefixlen < 4) { |
| 2364 | fib_info_put(res->fi); |
| 2365 | res->fi = NULL; |
| 2366 | } |
| 2367 | } |
| 2368 | |
| 2369 | |
| 2370 | rth = dst_alloc(&ipv4_dst_ops); |
| 2371 | if (!rth) { |
| 2372 | err = -ENOBUFS; |
| 2373 | goto cleanup; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2374 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2375 | |
Julian Anastasov | ce723d8 | 2005-09-08 13:34:47 -0700 | [diff] [blame] | 2376 | atomic_set(&rth->u.dst.__refcnt, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2377 | rth->u.dst.flags= DST_HOST; |
Herbert Xu | 42f811b | 2007-06-04 23:34:44 -0700 | [diff] [blame] | 2378 | if (IN_DEV_CONF_GET(in_dev, NOXFRM)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2379 | rth->u.dst.flags |= DST_NOXFRM; |
Herbert Xu | 42f811b | 2007-06-04 23:34:44 -0700 | [diff] [blame] | 2380 | if (IN_DEV_CONF_GET(in_dev, NOPOLICY)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2381 | rth->u.dst.flags |= DST_NOPOLICY; |
| 2382 | |
| 2383 | rth->fl.fl4_dst = oldflp->fl4_dst; |
| 2384 | rth->fl.fl4_tos = tos; |
| 2385 | rth->fl.fl4_src = oldflp->fl4_src; |
| 2386 | rth->fl.oif = oldflp->oif; |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 2387 | rth->fl.mark = oldflp->mark; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2388 | rth->rt_dst = fl->fl4_dst; |
| 2389 | rth->rt_src = fl->fl4_src; |
| 2390 | rth->rt_iif = oldflp->oif ? : dev_out->ifindex; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2391 | /* get references to the devices that are to be hold by the routing |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2392 | cache entry */ |
| 2393 | rth->u.dst.dev = dev_out; |
| 2394 | dev_hold(dev_out); |
| 2395 | rth->idev = in_dev_get(dev_out); |
| 2396 | rth->rt_gateway = fl->fl4_dst; |
| 2397 | rth->rt_spec_dst= fl->fl4_src; |
| 2398 | |
| 2399 | rth->u.dst.output=ip_output; |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2400 | rth->rt_genid = rt_genid(dev_net(dev_out)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2401 | |
| 2402 | RT_CACHE_STAT_INC(out_slow_tot); |
| 2403 | |
| 2404 | if (flags & RTCF_LOCAL) { |
| 2405 | rth->u.dst.input = ip_local_deliver; |
| 2406 | rth->rt_spec_dst = fl->fl4_dst; |
| 2407 | } |
| 2408 | if (flags & (RTCF_BROADCAST | RTCF_MULTICAST)) { |
| 2409 | rth->rt_spec_dst = fl->fl4_src; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2410 | if (flags & RTCF_LOCAL && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2411 | !(dev_out->flags & IFF_LOOPBACK)) { |
| 2412 | rth->u.dst.output = ip_mc_output; |
| 2413 | RT_CACHE_STAT_INC(out_slow_mc); |
| 2414 | } |
| 2415 | #ifdef CONFIG_IP_MROUTE |
| 2416 | if (res->type == RTN_MULTICAST) { |
| 2417 | if (IN_DEV_MFORWARD(in_dev) && |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2418 | !ipv4_is_local_multicast(oldflp->fl4_dst)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2419 | rth->u.dst.input = ip_mr_input; |
| 2420 | rth->u.dst.output = ip_mc_output; |
| 2421 | } |
| 2422 | } |
| 2423 | #endif |
| 2424 | } |
| 2425 | |
| 2426 | rt_set_nexthop(rth, res, 0); |
| 2427 | |
| 2428 | rth->rt_flags = flags; |
| 2429 | |
| 2430 | *result = rth; |
| 2431 | cleanup: |
| 2432 | /* release work reference to inet device */ |
| 2433 | in_dev_put(in_dev); |
| 2434 | |
| 2435 | return err; |
| 2436 | } |
| 2437 | |
Stephen Hemminger | 5969f71 | 2008-04-10 01:52:09 -0700 | [diff] [blame] | 2438 | static int ip_mkroute_output(struct rtable **rp, |
| 2439 | struct fib_result *res, |
| 2440 | const struct flowi *fl, |
| 2441 | const struct flowi *oldflp, |
| 2442 | struct net_device *dev_out, |
| 2443 | unsigned flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2444 | { |
Chuck Short | 7abaa27 | 2005-06-22 22:10:23 -0700 | [diff] [blame] | 2445 | struct rtable *rth = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2446 | int err = __mkroute_output(&rth, res, fl, oldflp, dev_out, flags); |
| 2447 | unsigned hash; |
| 2448 | if (err == 0) { |
Denis V. Lunev | b00180d | 2008-07-05 19:04:09 -0700 | [diff] [blame] | 2449 | hash = rt_hash(oldflp->fl4_dst, oldflp->fl4_src, oldflp->oif, |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2450 | rt_genid(dev_net(dev_out))); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2451 | err = rt_intern_hash(hash, rth, rp); |
| 2452 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2453 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2454 | return err; |
| 2455 | } |
| 2456 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2457 | /* |
| 2458 | * Major route resolver routine. |
| 2459 | */ |
| 2460 | |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2461 | static int ip_route_output_slow(struct net *net, struct rtable **rp, |
| 2462 | const struct flowi *oldflp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2463 | { |
| 2464 | u32 tos = RT_FL_TOS(oldflp); |
| 2465 | struct flowi fl = { .nl_u = { .ip4_u = |
| 2466 | { .daddr = oldflp->fl4_dst, |
| 2467 | .saddr = oldflp->fl4_src, |
| 2468 | .tos = tos & IPTOS_RT_MASK, |
| 2469 | .scope = ((tos & RTO_ONLINK) ? |
| 2470 | RT_SCOPE_LINK : |
| 2471 | RT_SCOPE_UNIVERSE), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2472 | } }, |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 2473 | .mark = oldflp->mark, |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2474 | .iif = net->loopback_dev->ifindex, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2475 | .oif = oldflp->oif }; |
| 2476 | struct fib_result res; |
| 2477 | unsigned flags = 0; |
| 2478 | struct net_device *dev_out = NULL; |
| 2479 | int free_res = 0; |
| 2480 | int err; |
| 2481 | |
| 2482 | |
| 2483 | res.fi = NULL; |
| 2484 | #ifdef CONFIG_IP_MULTIPLE_TABLES |
| 2485 | res.r = NULL; |
| 2486 | #endif |
| 2487 | |
| 2488 | if (oldflp->fl4_src) { |
| 2489 | err = -EINVAL; |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2490 | if (ipv4_is_multicast(oldflp->fl4_src) || |
Jan Engelhardt | 1e637c7 | 2008-01-21 03:18:08 -0800 | [diff] [blame] | 2491 | ipv4_is_lbcast(oldflp->fl4_src) || |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2492 | ipv4_is_zeronet(oldflp->fl4_src)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2493 | goto out; |
| 2494 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2495 | /* I removed check for oif == dev_out->oif here. |
| 2496 | It was wrong for two reasons: |
Denis V. Lunev | 1ab3527 | 2008-01-22 22:04:30 -0800 | [diff] [blame] | 2497 | 1. ip_dev_find(net, saddr) can return wrong iface, if saddr |
| 2498 | is assigned to multiple interfaces. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2499 | 2. Moreover, we are allowed to send packets with saddr |
| 2500 | of another iface. --ANK |
| 2501 | */ |
| 2502 | |
David S. Miller | f6c5d73 | 2007-05-18 02:07:50 -0700 | [diff] [blame] | 2503 | if (oldflp->oif == 0 |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2504 | && (ipv4_is_multicast(oldflp->fl4_dst) || |
| 2505 | oldflp->fl4_dst == htonl(0xFFFFFFFF))) { |
Julian Anastasov | a210d01 | 2008-10-01 07:28:28 -0700 | [diff] [blame] | 2506 | /* It is equivalent to inet_addr_type(saddr) == RTN_LOCAL */ |
| 2507 | dev_out = ip_dev_find(net, oldflp->fl4_src); |
| 2508 | if (dev_out == NULL) |
| 2509 | goto out; |
| 2510 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2511 | /* Special hack: user can direct multicasts |
| 2512 | and limited broadcast via necessary interface |
| 2513 | without fiddling with IP_MULTICAST_IF or IP_PKTINFO. |
| 2514 | This hack is not just for fun, it allows |
| 2515 | vic,vat and friends to work. |
| 2516 | They bind socket to loopback, set ttl to zero |
| 2517 | and expect that it will work. |
| 2518 | From the viewpoint of routing cache they are broken, |
| 2519 | because we are not allowed to build multicast path |
| 2520 | with loopback source addr (look, routing cache |
| 2521 | cannot know, that ttl is zero, so that packet |
| 2522 | will not leave this host and route is valid). |
| 2523 | Luckily, this hack is good workaround. |
| 2524 | */ |
| 2525 | |
| 2526 | fl.oif = dev_out->ifindex; |
| 2527 | goto make_route; |
| 2528 | } |
Julian Anastasov | a210d01 | 2008-10-01 07:28:28 -0700 | [diff] [blame] | 2529 | |
| 2530 | if (!(oldflp->flags & FLOWI_FLAG_ANYSRC)) { |
| 2531 | /* It is equivalent to inet_addr_type(saddr) == RTN_LOCAL */ |
| 2532 | dev_out = ip_dev_find(net, oldflp->fl4_src); |
| 2533 | if (dev_out == NULL) |
| 2534 | goto out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2535 | dev_put(dev_out); |
Julian Anastasov | a210d01 | 2008-10-01 07:28:28 -0700 | [diff] [blame] | 2536 | dev_out = NULL; |
| 2537 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2538 | } |
| 2539 | |
| 2540 | |
| 2541 | if (oldflp->oif) { |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2542 | dev_out = dev_get_by_index(net, oldflp->oif); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2543 | err = -ENODEV; |
| 2544 | if (dev_out == NULL) |
| 2545 | goto out; |
Herbert Xu | e5ed639 | 2005-10-03 14:35:55 -0700 | [diff] [blame] | 2546 | |
| 2547 | /* RACE: Check return value of inet_select_addr instead. */ |
| 2548 | if (__in_dev_get_rtnl(dev_out) == NULL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2549 | dev_put(dev_out); |
| 2550 | goto out; /* Wrong error code */ |
| 2551 | } |
| 2552 | |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2553 | if (ipv4_is_local_multicast(oldflp->fl4_dst) || |
| 2554 | oldflp->fl4_dst == htonl(0xFFFFFFFF)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2555 | if (!fl.fl4_src) |
| 2556 | fl.fl4_src = inet_select_addr(dev_out, 0, |
| 2557 | RT_SCOPE_LINK); |
| 2558 | goto make_route; |
| 2559 | } |
| 2560 | if (!fl.fl4_src) { |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2561 | if (ipv4_is_multicast(oldflp->fl4_dst)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2562 | fl.fl4_src = inet_select_addr(dev_out, 0, |
| 2563 | fl.fl4_scope); |
| 2564 | else if (!oldflp->fl4_dst) |
| 2565 | fl.fl4_src = inet_select_addr(dev_out, 0, |
| 2566 | RT_SCOPE_HOST); |
| 2567 | } |
| 2568 | } |
| 2569 | |
| 2570 | if (!fl.fl4_dst) { |
| 2571 | fl.fl4_dst = fl.fl4_src; |
| 2572 | if (!fl.fl4_dst) |
| 2573 | fl.fl4_dst = fl.fl4_src = htonl(INADDR_LOOPBACK); |
| 2574 | if (dev_out) |
| 2575 | dev_put(dev_out); |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2576 | dev_out = net->loopback_dev; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2577 | dev_hold(dev_out); |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2578 | fl.oif = net->loopback_dev->ifindex; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2579 | res.type = RTN_LOCAL; |
| 2580 | flags |= RTCF_LOCAL; |
| 2581 | goto make_route; |
| 2582 | } |
| 2583 | |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2584 | if (fib_lookup(net, &fl, &res)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2585 | res.fi = NULL; |
| 2586 | if (oldflp->oif) { |
| 2587 | /* Apparently, routing tables are wrong. Assume, |
| 2588 | that the destination is on link. |
| 2589 | |
| 2590 | WHY? DW. |
| 2591 | Because we are allowed to send to iface |
| 2592 | even if it has NO routes and NO assigned |
| 2593 | addresses. When oif is specified, routing |
| 2594 | tables are looked up with only one purpose: |
| 2595 | to catch if destination is gatewayed, rather than |
| 2596 | direct. Moreover, if MSG_DONTROUTE is set, |
| 2597 | we send packet, ignoring both routing tables |
| 2598 | and ifaddr state. --ANK |
| 2599 | |
| 2600 | |
| 2601 | We could make it even if oif is unknown, |
| 2602 | likely IPv6, but we do not. |
| 2603 | */ |
| 2604 | |
| 2605 | if (fl.fl4_src == 0) |
| 2606 | fl.fl4_src = inet_select_addr(dev_out, 0, |
| 2607 | RT_SCOPE_LINK); |
| 2608 | res.type = RTN_UNICAST; |
| 2609 | goto make_route; |
| 2610 | } |
| 2611 | if (dev_out) |
| 2612 | dev_put(dev_out); |
| 2613 | err = -ENETUNREACH; |
| 2614 | goto out; |
| 2615 | } |
| 2616 | free_res = 1; |
| 2617 | |
| 2618 | if (res.type == RTN_LOCAL) { |
| 2619 | if (!fl.fl4_src) |
| 2620 | fl.fl4_src = fl.fl4_dst; |
| 2621 | if (dev_out) |
| 2622 | dev_put(dev_out); |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2623 | dev_out = net->loopback_dev; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2624 | dev_hold(dev_out); |
| 2625 | fl.oif = dev_out->ifindex; |
| 2626 | if (res.fi) |
| 2627 | fib_info_put(res.fi); |
| 2628 | res.fi = NULL; |
| 2629 | flags |= RTCF_LOCAL; |
| 2630 | goto make_route; |
| 2631 | } |
| 2632 | |
| 2633 | #ifdef CONFIG_IP_ROUTE_MULTIPATH |
| 2634 | if (res.fi->fib_nhs > 1 && fl.oif == 0) |
| 2635 | fib_select_multipath(&fl, &res); |
| 2636 | else |
| 2637 | #endif |
| 2638 | if (!res.prefixlen && res.type == RTN_UNICAST && !fl.oif) |
Denis V. Lunev | b40afd0 | 2008-01-22 22:06:19 -0800 | [diff] [blame] | 2639 | fib_select_default(net, &fl, &res); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2640 | |
| 2641 | if (!fl.fl4_src) |
| 2642 | fl.fl4_src = FIB_RES_PREFSRC(res); |
| 2643 | |
| 2644 | if (dev_out) |
| 2645 | dev_put(dev_out); |
| 2646 | dev_out = FIB_RES_DEV(res); |
| 2647 | dev_hold(dev_out); |
| 2648 | fl.oif = dev_out->ifindex; |
| 2649 | |
| 2650 | |
| 2651 | make_route: |
| 2652 | err = ip_mkroute_output(rp, &res, &fl, oldflp, dev_out, flags); |
| 2653 | |
| 2654 | |
| 2655 | if (free_res) |
| 2656 | fib_res_put(&res); |
| 2657 | if (dev_out) |
| 2658 | dev_put(dev_out); |
| 2659 | out: return err; |
| 2660 | } |
| 2661 | |
Denis V. Lunev | 611c183 | 2008-01-22 22:06:48 -0800 | [diff] [blame] | 2662 | int __ip_route_output_key(struct net *net, struct rtable **rp, |
| 2663 | const struct flowi *flp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2664 | { |
| 2665 | unsigned hash; |
| 2666 | struct rtable *rth; |
| 2667 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 2668 | if (!rt_caching(net)) |
| 2669 | goto slow_output; |
| 2670 | |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2671 | hash = rt_hash(flp->fl4_dst, flp->fl4_src, flp->oif, rt_genid(net)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2672 | |
| 2673 | rcu_read_lock_bh(); |
| 2674 | for (rth = rcu_dereference(rt_hash_table[hash].chain); rth; |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 2675 | rth = rcu_dereference(rth->u.dst.rt_next)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2676 | if (rth->fl.fl4_dst == flp->fl4_dst && |
| 2677 | rth->fl.fl4_src == flp->fl4_src && |
| 2678 | rth->fl.iif == 0 && |
| 2679 | rth->fl.oif == flp->oif && |
Thomas Graf | 47dcf0c | 2006-11-09 15:20:38 -0800 | [diff] [blame] | 2680 | rth->fl.mark == flp->mark && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2681 | !((rth->fl.fl4_tos ^ flp->fl4_tos) & |
Denis V. Lunev | b592191 | 2008-01-22 23:50:25 -0800 | [diff] [blame] | 2682 | (IPTOS_RT_MASK | RTO_ONLINK)) && |
YOSHIFUJI Hideaki | 878628f | 2008-03-26 03:57:35 +0900 | [diff] [blame] | 2683 | net_eq(dev_net(rth->u.dst.dev), net) && |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2684 | !rt_is_expired(rth)) { |
Pavel Emelyanov | 03f49f3 | 2007-11-10 21:28:34 -0800 | [diff] [blame] | 2685 | dst_use(&rth->u.dst, jiffies); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2686 | RT_CACHE_STAT_INC(out_hit); |
| 2687 | rcu_read_unlock_bh(); |
| 2688 | *rp = rth; |
| 2689 | return 0; |
| 2690 | } |
| 2691 | RT_CACHE_STAT_INC(out_hlist_search); |
| 2692 | } |
| 2693 | rcu_read_unlock_bh(); |
| 2694 | |
Neil Horman | 1080d70 | 2008-10-27 12:28:25 -0700 | [diff] [blame] | 2695 | slow_output: |
Denis V. Lunev | 611c183 | 2008-01-22 22:06:48 -0800 | [diff] [blame] | 2696 | return ip_route_output_slow(net, rp, flp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2697 | } |
| 2698 | |
Arnaldo Carvalho de Melo | d8c97a9 | 2005-08-09 20:12:12 -0700 | [diff] [blame] | 2699 | EXPORT_SYMBOL_GPL(__ip_route_output_key); |
| 2700 | |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 2701 | static void ipv4_rt_blackhole_update_pmtu(struct dst_entry *dst, u32 mtu) |
| 2702 | { |
| 2703 | } |
| 2704 | |
| 2705 | static struct dst_ops ipv4_dst_blackhole_ops = { |
| 2706 | .family = AF_INET, |
| 2707 | .protocol = __constant_htons(ETH_P_IP), |
| 2708 | .destroy = ipv4_dst_destroy, |
| 2709 | .check = ipv4_dst_check, |
| 2710 | .update_pmtu = ipv4_rt_blackhole_update_pmtu, |
| 2711 | .entry_size = sizeof(struct rtable), |
Eric Dumazet | e242297 | 2008-01-30 20:07:45 -0800 | [diff] [blame] | 2712 | .entries = ATOMIC_INIT(0), |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 2713 | }; |
| 2714 | |
| 2715 | |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2716 | static int ipv4_dst_blackhole(struct net *net, struct rtable **rp, struct flowi *flp) |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 2717 | { |
| 2718 | struct rtable *ort = *rp; |
| 2719 | struct rtable *rt = (struct rtable *) |
| 2720 | dst_alloc(&ipv4_dst_blackhole_ops); |
| 2721 | |
| 2722 | if (rt) { |
| 2723 | struct dst_entry *new = &rt->u.dst; |
| 2724 | |
| 2725 | atomic_set(&new->__refcnt, 1); |
| 2726 | new->__use = 1; |
Herbert Xu | 352e512 | 2007-11-13 21:34:06 -0800 | [diff] [blame] | 2727 | new->input = dst_discard; |
| 2728 | new->output = dst_discard; |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 2729 | memcpy(new->metrics, ort->u.dst.metrics, RTAX_MAX*sizeof(u32)); |
| 2730 | |
| 2731 | new->dev = ort->u.dst.dev; |
| 2732 | if (new->dev) |
| 2733 | dev_hold(new->dev); |
| 2734 | |
| 2735 | rt->fl = ort->fl; |
| 2736 | |
| 2737 | rt->idev = ort->idev; |
| 2738 | if (rt->idev) |
| 2739 | in_dev_hold(rt->idev); |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2740 | rt->rt_genid = rt_genid(net); |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 2741 | rt->rt_flags = ort->rt_flags; |
| 2742 | rt->rt_type = ort->rt_type; |
| 2743 | rt->rt_dst = ort->rt_dst; |
| 2744 | rt->rt_src = ort->rt_src; |
| 2745 | rt->rt_iif = ort->rt_iif; |
| 2746 | rt->rt_gateway = ort->rt_gateway; |
| 2747 | rt->rt_spec_dst = ort->rt_spec_dst; |
| 2748 | rt->peer = ort->peer; |
| 2749 | if (rt->peer) |
| 2750 | atomic_inc(&rt->peer->refcnt); |
| 2751 | |
| 2752 | dst_free(new); |
| 2753 | } |
| 2754 | |
| 2755 | dst_release(&(*rp)->u.dst); |
| 2756 | *rp = rt; |
| 2757 | return (rt ? 0 : -ENOMEM); |
| 2758 | } |
| 2759 | |
Denis V. Lunev | f1b050b | 2008-01-22 22:07:10 -0800 | [diff] [blame] | 2760 | int ip_route_output_flow(struct net *net, struct rtable **rp, struct flowi *flp, |
| 2761 | struct sock *sk, int flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2762 | { |
| 2763 | int err; |
| 2764 | |
Denis V. Lunev | f1b050b | 2008-01-22 22:07:10 -0800 | [diff] [blame] | 2765 | if ((err = __ip_route_output_key(net, rp, flp)) != 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2766 | return err; |
| 2767 | |
| 2768 | if (flp->proto) { |
| 2769 | if (!flp->fl4_src) |
| 2770 | flp->fl4_src = (*rp)->rt_src; |
| 2771 | if (!flp->fl4_dst) |
| 2772 | flp->fl4_dst = (*rp)->rt_dst; |
Herbert Xu | bb72845 | 2007-12-12 18:48:58 -0800 | [diff] [blame] | 2773 | err = __xfrm_lookup((struct dst_entry **)rp, flp, sk, |
| 2774 | flags ? XFRM_LOOKUP_WAIT : 0); |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 2775 | if (err == -EREMOTE) |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2776 | err = ipv4_dst_blackhole(net, rp, flp); |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 2777 | |
| 2778 | return err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2779 | } |
| 2780 | |
| 2781 | return 0; |
| 2782 | } |
| 2783 | |
Arnaldo Carvalho de Melo | d8c97a9 | 2005-08-09 20:12:12 -0700 | [diff] [blame] | 2784 | EXPORT_SYMBOL_GPL(ip_route_output_flow); |
| 2785 | |
Denis V. Lunev | f206351 | 2008-01-22 22:07:34 -0800 | [diff] [blame] | 2786 | int ip_route_output_key(struct net *net, struct rtable **rp, struct flowi *flp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2787 | { |
Denis V. Lunev | f206351 | 2008-01-22 22:07:34 -0800 | [diff] [blame] | 2788 | return ip_route_output_flow(net, rp, flp, NULL, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2789 | } |
| 2790 | |
| 2791 | static int rt_fill_info(struct sk_buff *skb, u32 pid, u32 seq, int event, |
Jamal Hadi Salim | b6544c0 | 2005-06-18 22:54:12 -0700 | [diff] [blame] | 2792 | int nowait, unsigned int flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2793 | { |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 2794 | struct rtable *rt = skb->rtable; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2795 | struct rtmsg *r; |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2796 | struct nlmsghdr *nlh; |
Thomas Graf | e3703b3 | 2006-11-27 09:27:07 -0800 | [diff] [blame] | 2797 | long expires; |
| 2798 | u32 id = 0, ts = 0, tsage = 0, error; |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2799 | |
| 2800 | nlh = nlmsg_put(skb, pid, seq, event, sizeof(*r), flags); |
| 2801 | if (nlh == NULL) |
Patrick McHardy | 2693256 | 2007-01-31 23:16:40 -0800 | [diff] [blame] | 2802 | return -EMSGSIZE; |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2803 | |
| 2804 | r = nlmsg_data(nlh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2805 | r->rtm_family = AF_INET; |
| 2806 | r->rtm_dst_len = 32; |
| 2807 | r->rtm_src_len = 0; |
| 2808 | r->rtm_tos = rt->fl.fl4_tos; |
| 2809 | r->rtm_table = RT_TABLE_MAIN; |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2810 | NLA_PUT_U32(skb, RTA_TABLE, RT_TABLE_MAIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2811 | r->rtm_type = rt->rt_type; |
| 2812 | r->rtm_scope = RT_SCOPE_UNIVERSE; |
| 2813 | r->rtm_protocol = RTPROT_UNSPEC; |
| 2814 | r->rtm_flags = (rt->rt_flags & ~0xFFFF) | RTM_F_CLONED; |
| 2815 | if (rt->rt_flags & RTCF_NOTIFY) |
| 2816 | r->rtm_flags |= RTM_F_NOTIFY; |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2817 | |
Al Viro | 17fb2c6 | 2006-09-26 22:15:25 -0700 | [diff] [blame] | 2818 | NLA_PUT_BE32(skb, RTA_DST, rt->rt_dst); |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2819 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2820 | if (rt->fl.fl4_src) { |
| 2821 | r->rtm_src_len = 32; |
Al Viro | 17fb2c6 | 2006-09-26 22:15:25 -0700 | [diff] [blame] | 2822 | NLA_PUT_BE32(skb, RTA_SRC, rt->fl.fl4_src); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2823 | } |
| 2824 | if (rt->u.dst.dev) |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2825 | NLA_PUT_U32(skb, RTA_OIF, rt->u.dst.dev->ifindex); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2826 | #ifdef CONFIG_NET_CLS_ROUTE |
| 2827 | if (rt->u.dst.tclassid) |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2828 | NLA_PUT_U32(skb, RTA_FLOW, rt->u.dst.tclassid); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2829 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2830 | if (rt->fl.iif) |
Al Viro | 17fb2c6 | 2006-09-26 22:15:25 -0700 | [diff] [blame] | 2831 | NLA_PUT_BE32(skb, RTA_PREFSRC, rt->rt_spec_dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2832 | else if (rt->rt_src != rt->fl.fl4_src) |
Al Viro | 17fb2c6 | 2006-09-26 22:15:25 -0700 | [diff] [blame] | 2833 | NLA_PUT_BE32(skb, RTA_PREFSRC, rt->rt_src); |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2834 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2835 | if (rt->rt_dst != rt->rt_gateway) |
Al Viro | 17fb2c6 | 2006-09-26 22:15:25 -0700 | [diff] [blame] | 2836 | NLA_PUT_BE32(skb, RTA_GATEWAY, rt->rt_gateway); |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2837 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2838 | if (rtnetlink_put_metrics(skb, rt->u.dst.metrics) < 0) |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2839 | goto nla_put_failure; |
| 2840 | |
Thomas Graf | e3703b3 | 2006-11-27 09:27:07 -0800 | [diff] [blame] | 2841 | error = rt->u.dst.error; |
| 2842 | expires = rt->u.dst.expires ? rt->u.dst.expires - jiffies : 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2843 | if (rt->peer) { |
Thomas Graf | e3703b3 | 2006-11-27 09:27:07 -0800 | [diff] [blame] | 2844 | id = rt->peer->ip_id_count; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2845 | if (rt->peer->tcp_ts_stamp) { |
Thomas Graf | e3703b3 | 2006-11-27 09:27:07 -0800 | [diff] [blame] | 2846 | ts = rt->peer->tcp_ts; |
James Morris | 9d729f7 | 2007-03-04 16:12:44 -0800 | [diff] [blame] | 2847 | tsage = get_seconds() - rt->peer->tcp_ts_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2848 | } |
| 2849 | } |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2850 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2851 | if (rt->fl.iif) { |
| 2852 | #ifdef CONFIG_IP_MROUTE |
Al Viro | e448515 | 2006-09-26 22:15:01 -0700 | [diff] [blame] | 2853 | __be32 dst = rt->rt_dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2854 | |
Joe Perches | f97c1e0 | 2007-12-16 13:45:43 -0800 | [diff] [blame] | 2855 | if (ipv4_is_multicast(dst) && !ipv4_is_local_multicast(dst) && |
Pavel Emelyanov | 586f121 | 2007-12-16 13:32:48 -0800 | [diff] [blame] | 2856 | IPV4_DEVCONF_ALL(&init_net, MC_FORWARDING)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2857 | int err = ipmr_get_route(skb, r, nowait); |
| 2858 | if (err <= 0) { |
| 2859 | if (!nowait) { |
| 2860 | if (err == 0) |
| 2861 | return 0; |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2862 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2863 | } else { |
| 2864 | if (err == -EMSGSIZE) |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2865 | goto nla_put_failure; |
Thomas Graf | e3703b3 | 2006-11-27 09:27:07 -0800 | [diff] [blame] | 2866 | error = err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2867 | } |
| 2868 | } |
| 2869 | } else |
| 2870 | #endif |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2871 | NLA_PUT_U32(skb, RTA_IIF, rt->fl.iif); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2872 | } |
| 2873 | |
Thomas Graf | e3703b3 | 2006-11-27 09:27:07 -0800 | [diff] [blame] | 2874 | if (rtnl_put_cacheinfo(skb, &rt->u.dst, id, ts, tsage, |
| 2875 | expires, error) < 0) |
| 2876 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2877 | |
Thomas Graf | be403ea | 2006-08-17 18:15:17 -0700 | [diff] [blame] | 2878 | return nlmsg_end(skb, nlh); |
| 2879 | |
| 2880 | nla_put_failure: |
Patrick McHardy | 2693256 | 2007-01-31 23:16:40 -0800 | [diff] [blame] | 2881 | nlmsg_cancel(skb, nlh); |
| 2882 | return -EMSGSIZE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2883 | } |
| 2884 | |
Thomas Graf | 63f3444 | 2007-03-22 11:55:17 -0700 | [diff] [blame] | 2885 | static int inet_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr* nlh, void *arg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2886 | { |
YOSHIFUJI Hideaki | 3b1e0a6 | 2008-03-26 02:26:21 +0900 | [diff] [blame] | 2887 | struct net *net = sock_net(in_skb->sk); |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2888 | struct rtmsg *rtm; |
| 2889 | struct nlattr *tb[RTA_MAX+1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2890 | struct rtable *rt = NULL; |
Al Viro | 9e12bb2 | 2006-09-26 21:25:20 -0700 | [diff] [blame] | 2891 | __be32 dst = 0; |
| 2892 | __be32 src = 0; |
| 2893 | u32 iif; |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2894 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2895 | struct sk_buff *skb; |
| 2896 | |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2897 | err = nlmsg_parse(nlh, sizeof(*rtm), tb, RTA_MAX, rtm_ipv4_policy); |
| 2898 | if (err < 0) |
| 2899 | goto errout; |
| 2900 | |
| 2901 | rtm = nlmsg_data(nlh); |
| 2902 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2903 | skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL); |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2904 | if (skb == NULL) { |
| 2905 | err = -ENOBUFS; |
| 2906 | goto errout; |
| 2907 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2908 | |
| 2909 | /* Reserve room for dummy headers, this skb can pass |
| 2910 | through good chunk of routing engine. |
| 2911 | */ |
Arnaldo Carvalho de Melo | 459a98e | 2007-03-19 15:30:44 -0700 | [diff] [blame] | 2912 | skb_reset_mac_header(skb); |
Arnaldo Carvalho de Melo | c1d2bbe | 2007-04-10 20:45:18 -0700 | [diff] [blame] | 2913 | skb_reset_network_header(skb); |
Stephen Hemminger | d2c962b | 2006-04-17 17:27:11 -0700 | [diff] [blame] | 2914 | |
| 2915 | /* Bugfix: need to give ip_route_input enough of an IP header to not gag. */ |
Arnaldo Carvalho de Melo | eddc9ec | 2007-04-20 22:47:35 -0700 | [diff] [blame] | 2916 | ip_hdr(skb)->protocol = IPPROTO_ICMP; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2917 | skb_reserve(skb, MAX_HEADER + sizeof(struct iphdr)); |
| 2918 | |
Al Viro | 17fb2c6 | 2006-09-26 22:15:25 -0700 | [diff] [blame] | 2919 | src = tb[RTA_SRC] ? nla_get_be32(tb[RTA_SRC]) : 0; |
| 2920 | dst = tb[RTA_DST] ? nla_get_be32(tb[RTA_DST]) : 0; |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2921 | iif = tb[RTA_IIF] ? nla_get_u32(tb[RTA_IIF]) : 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2922 | |
| 2923 | if (iif) { |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2924 | struct net_device *dev; |
| 2925 | |
Denis V. Lunev | 1937504 | 2008-02-28 20:52:04 -0800 | [diff] [blame] | 2926 | dev = __dev_get_by_index(net, iif); |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2927 | if (dev == NULL) { |
| 2928 | err = -ENODEV; |
| 2929 | goto errout_free; |
| 2930 | } |
| 2931 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2932 | skb->protocol = htons(ETH_P_IP); |
| 2933 | skb->dev = dev; |
| 2934 | local_bh_disable(); |
| 2935 | err = ip_route_input(skb, dst, src, rtm->rtm_tos, dev); |
| 2936 | local_bh_enable(); |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2937 | |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 2938 | rt = skb->rtable; |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2939 | if (err == 0 && rt->u.dst.error) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2940 | err = -rt->u.dst.error; |
| 2941 | } else { |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2942 | struct flowi fl = { |
| 2943 | .nl_u = { |
| 2944 | .ip4_u = { |
| 2945 | .daddr = dst, |
| 2946 | .saddr = src, |
| 2947 | .tos = rtm->rtm_tos, |
| 2948 | }, |
| 2949 | }, |
| 2950 | .oif = tb[RTA_OIF] ? nla_get_u32(tb[RTA_OIF]) : 0, |
| 2951 | }; |
Denis V. Lunev | 1937504 | 2008-02-28 20:52:04 -0800 | [diff] [blame] | 2952 | err = ip_route_output_key(net, &rt, &fl); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2953 | } |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2954 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2955 | if (err) |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2956 | goto errout_free; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2957 | |
Eric Dumazet | ee6b967 | 2008-03-05 18:30:47 -0800 | [diff] [blame] | 2958 | skb->rtable = rt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2959 | if (rtm->rtm_flags & RTM_F_NOTIFY) |
| 2960 | rt->rt_flags |= RTCF_NOTIFY; |
| 2961 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2962 | err = rt_fill_info(skb, NETLINK_CB(in_skb).pid, nlh->nlmsg_seq, |
Denis V. Lunev | 1937504 | 2008-02-28 20:52:04 -0800 | [diff] [blame] | 2963 | RTM_NEWROUTE, 0, 0); |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2964 | if (err <= 0) |
| 2965 | goto errout_free; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2966 | |
Denis V. Lunev | 1937504 | 2008-02-28 20:52:04 -0800 | [diff] [blame] | 2967 | err = rtnl_unicast(skb, net, NETLINK_CB(in_skb).pid); |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2968 | errout: |
Thomas Graf | 2942e90 | 2006-08-15 00:30:25 -0700 | [diff] [blame] | 2969 | return err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2970 | |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2971 | errout_free: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2972 | kfree_skb(skb); |
Thomas Graf | d889ce3 | 2006-08-17 18:15:44 -0700 | [diff] [blame] | 2973 | goto errout; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2974 | } |
| 2975 | |
| 2976 | int ip_rt_dump(struct sk_buff *skb, struct netlink_callback *cb) |
| 2977 | { |
| 2978 | struct rtable *rt; |
| 2979 | int h, s_h; |
| 2980 | int idx, s_idx; |
Denis V. Lunev | 1937504 | 2008-02-28 20:52:04 -0800 | [diff] [blame] | 2981 | struct net *net; |
| 2982 | |
YOSHIFUJI Hideaki | 3b1e0a6 | 2008-03-26 02:26:21 +0900 | [diff] [blame] | 2983 | net = sock_net(skb->sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2984 | |
| 2985 | s_h = cb->args[0]; |
Eric Dumazet | d8c9283 | 2008-01-07 21:52:14 -0800 | [diff] [blame] | 2986 | if (s_h < 0) |
| 2987 | s_h = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2988 | s_idx = idx = cb->args[1]; |
Eric Dumazet | a627266 | 2008-08-28 01:11:25 -0700 | [diff] [blame] | 2989 | for (h = s_h; h <= rt_hash_mask; h++, s_idx = 0) { |
| 2990 | if (!rt_hash_table[h].chain) |
| 2991 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2992 | rcu_read_lock_bh(); |
| 2993 | for (rt = rcu_dereference(rt_hash_table[h].chain), idx = 0; rt; |
Eric Dumazet | 093c2ca | 2007-02-09 16:19:26 -0800 | [diff] [blame] | 2994 | rt = rcu_dereference(rt->u.dst.rt_next), idx++) { |
YOSHIFUJI Hideaki | 878628f | 2008-03-26 03:57:35 +0900 | [diff] [blame] | 2995 | if (!net_eq(dev_net(rt->u.dst.dev), net) || idx < s_idx) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2996 | continue; |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 2997 | if (rt_is_expired(rt)) |
Eric Dumazet | 29e7525 | 2008-01-31 17:05:09 -0800 | [diff] [blame] | 2998 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2999 | skb->dst = dst_clone(&rt->u.dst); |
| 3000 | if (rt_fill_info(skb, NETLINK_CB(cb->skb).pid, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3001 | cb->nlh->nlmsg_seq, RTM_NEWROUTE, |
Jamal Hadi Salim | b6544c0 | 2005-06-18 22:54:12 -0700 | [diff] [blame] | 3002 | 1, NLM_F_MULTI) <= 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3003 | dst_release(xchg(&skb->dst, NULL)); |
| 3004 | rcu_read_unlock_bh(); |
| 3005 | goto done; |
| 3006 | } |
| 3007 | dst_release(xchg(&skb->dst, NULL)); |
| 3008 | } |
| 3009 | rcu_read_unlock_bh(); |
| 3010 | } |
| 3011 | |
| 3012 | done: |
| 3013 | cb->args[0] = h; |
| 3014 | cb->args[1] = idx; |
| 3015 | return skb->len; |
| 3016 | } |
| 3017 | |
| 3018 | void ip_rt_multicast_event(struct in_device *in_dev) |
| 3019 | { |
Denis V. Lunev | 76e6ebf | 2008-07-05 19:00:44 -0700 | [diff] [blame] | 3020 | rt_cache_flush(dev_net(in_dev->dev), 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3021 | } |
| 3022 | |
| 3023 | #ifdef CONFIG_SYSCTL |
Denis V. Lunev | 81c684d | 2008-07-08 03:05:28 -0700 | [diff] [blame] | 3024 | static int ipv4_sysctl_rtcache_flush(ctl_table *__ctl, int write, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3025 | struct file *filp, void __user *buffer, |
| 3026 | size_t *lenp, loff_t *ppos) |
| 3027 | { |
| 3028 | if (write) { |
Denis V. Lunev | 639e104 | 2008-07-05 19:02:06 -0700 | [diff] [blame] | 3029 | int flush_delay; |
Denis V. Lunev | 81c684d | 2008-07-08 03:05:28 -0700 | [diff] [blame] | 3030 | ctl_table ctl; |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3031 | struct net *net; |
Denis V. Lunev | 639e104 | 2008-07-05 19:02:06 -0700 | [diff] [blame] | 3032 | |
Denis V. Lunev | 81c684d | 2008-07-08 03:05:28 -0700 | [diff] [blame] | 3033 | memcpy(&ctl, __ctl, sizeof(ctl)); |
| 3034 | ctl.data = &flush_delay; |
| 3035 | proc_dointvec(&ctl, write, filp, buffer, lenp, ppos); |
Denis V. Lunev | 639e104 | 2008-07-05 19:02:06 -0700 | [diff] [blame] | 3036 | |
Denis V. Lunev | 81c684d | 2008-07-08 03:05:28 -0700 | [diff] [blame] | 3037 | net = (struct net *)__ctl->extra1; |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3038 | rt_cache_flush(net, flush_delay); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3039 | return 0; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3040 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3041 | |
| 3042 | return -EINVAL; |
| 3043 | } |
| 3044 | |
| 3045 | static int ipv4_sysctl_rtcache_flush_strategy(ctl_table *table, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3046 | void __user *oldval, |
| 3047 | size_t __user *oldlenp, |
| 3048 | void __user *newval, |
Alexey Dobriyan | 1f29bcd | 2006-12-10 02:19:10 -0800 | [diff] [blame] | 3049 | size_t newlen) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3050 | { |
| 3051 | int delay; |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3052 | struct net *net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3053 | if (newlen != sizeof(int)) |
| 3054 | return -EINVAL; |
| 3055 | if (get_user(delay, (int __user *)newval)) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3056 | return -EFAULT; |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3057 | net = (struct net *)table->extra1; |
| 3058 | rt_cache_flush(net, delay); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3059 | return 0; |
| 3060 | } |
| 3061 | |
Herbert Xu | c6153b5 | 2008-08-15 13:44:31 -0700 | [diff] [blame] | 3062 | static void rt_secret_reschedule(int old) |
| 3063 | { |
| 3064 | struct net *net; |
| 3065 | int new = ip_rt_secret_interval; |
| 3066 | int diff = new - old; |
| 3067 | |
| 3068 | if (!diff) |
| 3069 | return; |
| 3070 | |
| 3071 | rtnl_lock(); |
| 3072 | for_each_net(net) { |
| 3073 | int deleted = del_timer_sync(&net->ipv4.rt_secret_timer); |
| 3074 | |
| 3075 | if (!new) |
| 3076 | continue; |
| 3077 | |
| 3078 | if (deleted) { |
| 3079 | long time = net->ipv4.rt_secret_timer.expires - jiffies; |
| 3080 | |
| 3081 | if (time <= 0 || (time += diff) <= 0) |
| 3082 | time = 0; |
| 3083 | |
| 3084 | net->ipv4.rt_secret_timer.expires = time; |
| 3085 | } else |
| 3086 | net->ipv4.rt_secret_timer.expires = new; |
| 3087 | |
| 3088 | net->ipv4.rt_secret_timer.expires += jiffies; |
| 3089 | add_timer(&net->ipv4.rt_secret_timer); |
| 3090 | } |
| 3091 | rtnl_unlock(); |
| 3092 | } |
| 3093 | |
| 3094 | static int ipv4_sysctl_rt_secret_interval(ctl_table *ctl, int write, |
| 3095 | struct file *filp, |
| 3096 | void __user *buffer, size_t *lenp, |
| 3097 | loff_t *ppos) |
| 3098 | { |
| 3099 | int old = ip_rt_secret_interval; |
| 3100 | int ret = proc_dointvec_jiffies(ctl, write, filp, buffer, lenp, ppos); |
| 3101 | |
| 3102 | rt_secret_reschedule(old); |
| 3103 | |
| 3104 | return ret; |
| 3105 | } |
| 3106 | |
| 3107 | static int ipv4_sysctl_rt_secret_interval_strategy(ctl_table *table, |
Herbert Xu | c6153b5 | 2008-08-15 13:44:31 -0700 | [diff] [blame] | 3108 | void __user *oldval, |
| 3109 | size_t __user *oldlenp, |
| 3110 | void __user *newval, |
| 3111 | size_t newlen) |
| 3112 | { |
| 3113 | int old = ip_rt_secret_interval; |
Alexey Dobriyan | f221e72 | 2008-10-15 22:04:23 -0700 | [diff] [blame] | 3114 | int ret = sysctl_jiffies(table, oldval, oldlenp, newval, newlen); |
Herbert Xu | c6153b5 | 2008-08-15 13:44:31 -0700 | [diff] [blame] | 3115 | |
| 3116 | rt_secret_reschedule(old); |
| 3117 | |
| 3118 | return ret; |
| 3119 | } |
| 3120 | |
Al Viro | eeb61f7 | 2008-07-27 08:59:33 +0100 | [diff] [blame] | 3121 | static ctl_table ipv4_route_table[] = { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3122 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3123 | .ctl_name = NET_IPV4_ROUTE_GC_THRESH, |
| 3124 | .procname = "gc_thresh", |
| 3125 | .data = &ipv4_dst_ops.gc_thresh, |
| 3126 | .maxlen = sizeof(int), |
| 3127 | .mode = 0644, |
| 3128 | .proc_handler = &proc_dointvec, |
| 3129 | }, |
| 3130 | { |
| 3131 | .ctl_name = NET_IPV4_ROUTE_MAX_SIZE, |
| 3132 | .procname = "max_size", |
| 3133 | .data = &ip_rt_max_size, |
| 3134 | .maxlen = sizeof(int), |
| 3135 | .mode = 0644, |
| 3136 | .proc_handler = &proc_dointvec, |
| 3137 | }, |
| 3138 | { |
| 3139 | /* Deprecated. Use gc_min_interval_ms */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3140 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3141 | .ctl_name = NET_IPV4_ROUTE_GC_MIN_INTERVAL, |
| 3142 | .procname = "gc_min_interval", |
| 3143 | .data = &ip_rt_gc_min_interval, |
| 3144 | .maxlen = sizeof(int), |
| 3145 | .mode = 0644, |
| 3146 | .proc_handler = &proc_dointvec_jiffies, |
| 3147 | .strategy = &sysctl_jiffies, |
| 3148 | }, |
| 3149 | { |
| 3150 | .ctl_name = NET_IPV4_ROUTE_GC_MIN_INTERVAL_MS, |
| 3151 | .procname = "gc_min_interval_ms", |
| 3152 | .data = &ip_rt_gc_min_interval, |
| 3153 | .maxlen = sizeof(int), |
| 3154 | .mode = 0644, |
| 3155 | .proc_handler = &proc_dointvec_ms_jiffies, |
| 3156 | .strategy = &sysctl_ms_jiffies, |
| 3157 | }, |
| 3158 | { |
| 3159 | .ctl_name = NET_IPV4_ROUTE_GC_TIMEOUT, |
| 3160 | .procname = "gc_timeout", |
| 3161 | .data = &ip_rt_gc_timeout, |
| 3162 | .maxlen = sizeof(int), |
| 3163 | .mode = 0644, |
| 3164 | .proc_handler = &proc_dointvec_jiffies, |
| 3165 | .strategy = &sysctl_jiffies, |
| 3166 | }, |
| 3167 | { |
| 3168 | .ctl_name = NET_IPV4_ROUTE_GC_INTERVAL, |
| 3169 | .procname = "gc_interval", |
| 3170 | .data = &ip_rt_gc_interval, |
| 3171 | .maxlen = sizeof(int), |
| 3172 | .mode = 0644, |
| 3173 | .proc_handler = &proc_dointvec_jiffies, |
| 3174 | .strategy = &sysctl_jiffies, |
| 3175 | }, |
| 3176 | { |
| 3177 | .ctl_name = NET_IPV4_ROUTE_REDIRECT_LOAD, |
| 3178 | .procname = "redirect_load", |
| 3179 | .data = &ip_rt_redirect_load, |
| 3180 | .maxlen = sizeof(int), |
| 3181 | .mode = 0644, |
| 3182 | .proc_handler = &proc_dointvec, |
| 3183 | }, |
| 3184 | { |
| 3185 | .ctl_name = NET_IPV4_ROUTE_REDIRECT_NUMBER, |
| 3186 | .procname = "redirect_number", |
| 3187 | .data = &ip_rt_redirect_number, |
| 3188 | .maxlen = sizeof(int), |
| 3189 | .mode = 0644, |
| 3190 | .proc_handler = &proc_dointvec, |
| 3191 | }, |
| 3192 | { |
| 3193 | .ctl_name = NET_IPV4_ROUTE_REDIRECT_SILENCE, |
| 3194 | .procname = "redirect_silence", |
| 3195 | .data = &ip_rt_redirect_silence, |
| 3196 | .maxlen = sizeof(int), |
| 3197 | .mode = 0644, |
| 3198 | .proc_handler = &proc_dointvec, |
| 3199 | }, |
| 3200 | { |
| 3201 | .ctl_name = NET_IPV4_ROUTE_ERROR_COST, |
| 3202 | .procname = "error_cost", |
| 3203 | .data = &ip_rt_error_cost, |
| 3204 | .maxlen = sizeof(int), |
| 3205 | .mode = 0644, |
| 3206 | .proc_handler = &proc_dointvec, |
| 3207 | }, |
| 3208 | { |
| 3209 | .ctl_name = NET_IPV4_ROUTE_ERROR_BURST, |
| 3210 | .procname = "error_burst", |
| 3211 | .data = &ip_rt_error_burst, |
| 3212 | .maxlen = sizeof(int), |
| 3213 | .mode = 0644, |
| 3214 | .proc_handler = &proc_dointvec, |
| 3215 | }, |
| 3216 | { |
| 3217 | .ctl_name = NET_IPV4_ROUTE_GC_ELASTICITY, |
| 3218 | .procname = "gc_elasticity", |
| 3219 | .data = &ip_rt_gc_elasticity, |
| 3220 | .maxlen = sizeof(int), |
| 3221 | .mode = 0644, |
| 3222 | .proc_handler = &proc_dointvec, |
| 3223 | }, |
| 3224 | { |
| 3225 | .ctl_name = NET_IPV4_ROUTE_MTU_EXPIRES, |
| 3226 | .procname = "mtu_expires", |
| 3227 | .data = &ip_rt_mtu_expires, |
| 3228 | .maxlen = sizeof(int), |
| 3229 | .mode = 0644, |
| 3230 | .proc_handler = &proc_dointvec_jiffies, |
| 3231 | .strategy = &sysctl_jiffies, |
| 3232 | }, |
| 3233 | { |
| 3234 | .ctl_name = NET_IPV4_ROUTE_MIN_PMTU, |
| 3235 | .procname = "min_pmtu", |
| 3236 | .data = &ip_rt_min_pmtu, |
| 3237 | .maxlen = sizeof(int), |
| 3238 | .mode = 0644, |
| 3239 | .proc_handler = &proc_dointvec, |
| 3240 | }, |
| 3241 | { |
| 3242 | .ctl_name = NET_IPV4_ROUTE_MIN_ADVMSS, |
| 3243 | .procname = "min_adv_mss", |
| 3244 | .data = &ip_rt_min_advmss, |
| 3245 | .maxlen = sizeof(int), |
| 3246 | .mode = 0644, |
| 3247 | .proc_handler = &proc_dointvec, |
| 3248 | }, |
| 3249 | { |
| 3250 | .ctl_name = NET_IPV4_ROUTE_SECRET_INTERVAL, |
| 3251 | .procname = "secret_interval", |
| 3252 | .data = &ip_rt_secret_interval, |
| 3253 | .maxlen = sizeof(int), |
| 3254 | .mode = 0644, |
Herbert Xu | c6153b5 | 2008-08-15 13:44:31 -0700 | [diff] [blame] | 3255 | .proc_handler = &ipv4_sysctl_rt_secret_interval, |
| 3256 | .strategy = &ipv4_sysctl_rt_secret_interval_strategy, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3257 | }, |
| 3258 | { .ctl_name = 0 } |
| 3259 | }; |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3260 | |
Al Viro | 2f4520d | 2008-08-25 15:17:44 -0700 | [diff] [blame] | 3261 | static struct ctl_table empty[1]; |
| 3262 | |
| 3263 | static struct ctl_table ipv4_skeleton[] = |
| 3264 | { |
| 3265 | { .procname = "route", .ctl_name = NET_IPV4_ROUTE, |
Hugh Dickins | d994af0 | 2008-08-27 02:35:18 -0700 | [diff] [blame] | 3266 | .mode = 0555, .child = ipv4_route_table}, |
Al Viro | 2f4520d | 2008-08-25 15:17:44 -0700 | [diff] [blame] | 3267 | { .procname = "neigh", .ctl_name = NET_IPV4_NEIGH, |
Hugh Dickins | d994af0 | 2008-08-27 02:35:18 -0700 | [diff] [blame] | 3268 | .mode = 0555, .child = empty}, |
Al Viro | 2f4520d | 2008-08-25 15:17:44 -0700 | [diff] [blame] | 3269 | { } |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3270 | }; |
| 3271 | |
Al Viro | 2f4520d | 2008-08-25 15:17:44 -0700 | [diff] [blame] | 3272 | static __net_initdata struct ctl_path ipv4_path[] = { |
| 3273 | { .procname = "net", .ctl_name = CTL_NET, }, |
| 3274 | { .procname = "ipv4", .ctl_name = NET_IPV4, }, |
| 3275 | { }, |
| 3276 | }; |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3277 | |
| 3278 | static struct ctl_table ipv4_route_flush_table[] = { |
| 3279 | { |
| 3280 | .ctl_name = NET_IPV4_ROUTE_FLUSH, |
| 3281 | .procname = "flush", |
| 3282 | .maxlen = sizeof(int), |
| 3283 | .mode = 0200, |
| 3284 | .proc_handler = &ipv4_sysctl_rtcache_flush, |
| 3285 | .strategy = &ipv4_sysctl_rtcache_flush_strategy, |
| 3286 | }, |
| 3287 | { .ctl_name = 0 }, |
| 3288 | }; |
| 3289 | |
Al Viro | 2f4520d | 2008-08-25 15:17:44 -0700 | [diff] [blame] | 3290 | static __net_initdata struct ctl_path ipv4_route_path[] = { |
| 3291 | { .procname = "net", .ctl_name = CTL_NET, }, |
| 3292 | { .procname = "ipv4", .ctl_name = NET_IPV4, }, |
| 3293 | { .procname = "route", .ctl_name = NET_IPV4_ROUTE, }, |
| 3294 | { }, |
| 3295 | }; |
| 3296 | |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3297 | static __net_init int sysctl_route_net_init(struct net *net) |
| 3298 | { |
| 3299 | struct ctl_table *tbl; |
| 3300 | |
| 3301 | tbl = ipv4_route_flush_table; |
| 3302 | if (net != &init_net) { |
| 3303 | tbl = kmemdup(tbl, sizeof(ipv4_route_flush_table), GFP_KERNEL); |
| 3304 | if (tbl == NULL) |
| 3305 | goto err_dup; |
| 3306 | } |
| 3307 | tbl[0].extra1 = net; |
| 3308 | |
| 3309 | net->ipv4.route_hdr = |
| 3310 | register_net_sysctl_table(net, ipv4_route_path, tbl); |
| 3311 | if (net->ipv4.route_hdr == NULL) |
| 3312 | goto err_reg; |
| 3313 | return 0; |
| 3314 | |
| 3315 | err_reg: |
| 3316 | if (tbl != ipv4_route_flush_table) |
| 3317 | kfree(tbl); |
| 3318 | err_dup: |
| 3319 | return -ENOMEM; |
| 3320 | } |
| 3321 | |
| 3322 | static __net_exit void sysctl_route_net_exit(struct net *net) |
| 3323 | { |
| 3324 | struct ctl_table *tbl; |
| 3325 | |
| 3326 | tbl = net->ipv4.route_hdr->ctl_table_arg; |
| 3327 | unregister_net_sysctl_table(net->ipv4.route_hdr); |
| 3328 | BUG_ON(tbl == ipv4_route_flush_table); |
| 3329 | kfree(tbl); |
| 3330 | } |
| 3331 | |
| 3332 | static __net_initdata struct pernet_operations sysctl_route_ops = { |
| 3333 | .init = sysctl_route_net_init, |
| 3334 | .exit = sysctl_route_net_exit, |
| 3335 | }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3336 | #endif |
| 3337 | |
Denis V. Lunev | 9f5e97e | 2008-07-05 19:02:59 -0700 | [diff] [blame] | 3338 | |
| 3339 | static __net_init int rt_secret_timer_init(struct net *net) |
| 3340 | { |
Denis V. Lunev | e84f84f | 2008-07-05 19:04:32 -0700 | [diff] [blame] | 3341 | atomic_set(&net->ipv4.rt_genid, |
| 3342 | (int) ((num_physpages ^ (num_physpages>>8)) ^ |
| 3343 | (jiffies ^ (jiffies >> 7)))); |
| 3344 | |
Denis V. Lunev | 9f5e97e | 2008-07-05 19:02:59 -0700 | [diff] [blame] | 3345 | net->ipv4.rt_secret_timer.function = rt_secret_rebuild; |
| 3346 | net->ipv4.rt_secret_timer.data = (unsigned long)net; |
| 3347 | init_timer_deferrable(&net->ipv4.rt_secret_timer); |
| 3348 | |
Herbert Xu | c6153b5 | 2008-08-15 13:44:31 -0700 | [diff] [blame] | 3349 | if (ip_rt_secret_interval) { |
| 3350 | net->ipv4.rt_secret_timer.expires = |
| 3351 | jiffies + net_random() % ip_rt_secret_interval + |
| 3352 | ip_rt_secret_interval; |
| 3353 | add_timer(&net->ipv4.rt_secret_timer); |
| 3354 | } |
Denis V. Lunev | 9f5e97e | 2008-07-05 19:02:59 -0700 | [diff] [blame] | 3355 | return 0; |
| 3356 | } |
| 3357 | |
| 3358 | static __net_exit void rt_secret_timer_exit(struct net *net) |
| 3359 | { |
| 3360 | del_timer_sync(&net->ipv4.rt_secret_timer); |
| 3361 | } |
| 3362 | |
| 3363 | static __net_initdata struct pernet_operations rt_secret_timer_ops = { |
| 3364 | .init = rt_secret_timer_init, |
| 3365 | .exit = rt_secret_timer_exit, |
| 3366 | }; |
| 3367 | |
| 3368 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3369 | #ifdef CONFIG_NET_CLS_ROUTE |
Eric Dumazet | 8dbde28 | 2007-11-16 03:32:10 -0800 | [diff] [blame] | 3370 | struct ip_rt_acct *ip_rt_acct __read_mostly; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3371 | #endif /* CONFIG_NET_CLS_ROUTE */ |
| 3372 | |
| 3373 | static __initdata unsigned long rhash_entries; |
| 3374 | static int __init set_rhash_entries(char *str) |
| 3375 | { |
| 3376 | if (!str) |
| 3377 | return 0; |
| 3378 | rhash_entries = simple_strtoul(str, &str, 0); |
| 3379 | return 1; |
| 3380 | } |
| 3381 | __setup("rhash_entries=", set_rhash_entries); |
| 3382 | |
| 3383 | int __init ip_rt_init(void) |
| 3384 | { |
Eric Dumazet | 424c4b7 | 2005-07-05 14:58:19 -0700 | [diff] [blame] | 3385 | int rc = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3386 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3387 | #ifdef CONFIG_NET_CLS_ROUTE |
Eric Dumazet | 8dbde28 | 2007-11-16 03:32:10 -0800 | [diff] [blame] | 3388 | ip_rt_acct = __alloc_percpu(256 * sizeof(struct ip_rt_acct)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3389 | if (!ip_rt_acct) |
| 3390 | panic("IP: failed to allocate ip_rt_acct\n"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3391 | #endif |
| 3392 | |
Alexey Dobriyan | e5d679f33 | 2006-08-26 19:25:52 -0700 | [diff] [blame] | 3393 | ipv4_dst_ops.kmem_cachep = |
| 3394 | kmem_cache_create("ip_dst_cache", sizeof(struct rtable), 0, |
Paul Mundt | 20c2df8 | 2007-07-20 10:11:58 +0900 | [diff] [blame] | 3395 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3396 | |
David S. Miller | 14e50e5 | 2007-05-24 18:17:54 -0700 | [diff] [blame] | 3397 | ipv4_dst_blackhole_ops.kmem_cachep = ipv4_dst_ops.kmem_cachep; |
| 3398 | |
Eric Dumazet | 424c4b7 | 2005-07-05 14:58:19 -0700 | [diff] [blame] | 3399 | rt_hash_table = (struct rt_hash_bucket *) |
| 3400 | alloc_large_system_hash("IP route cache", |
| 3401 | sizeof(struct rt_hash_bucket), |
| 3402 | rhash_entries, |
| 3403 | (num_physpages >= 128 * 1024) ? |
Mike Stroyan | 18955cf | 2005-11-29 16:12:55 -0800 | [diff] [blame] | 3404 | 15 : 17, |
Kirill Korotaev | 8d1502d | 2006-08-07 20:44:22 -0700 | [diff] [blame] | 3405 | 0, |
Eric Dumazet | 424c4b7 | 2005-07-05 14:58:19 -0700 | [diff] [blame] | 3406 | &rt_hash_log, |
| 3407 | &rt_hash_mask, |
| 3408 | 0); |
Eric Dumazet | 22c047c | 2005-07-05 14:55:24 -0700 | [diff] [blame] | 3409 | memset(rt_hash_table, 0, (rt_hash_mask + 1) * sizeof(struct rt_hash_bucket)); |
| 3410 | rt_hash_lock_init(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3411 | |
| 3412 | ipv4_dst_ops.gc_thresh = (rt_hash_mask + 1); |
| 3413 | ip_rt_max_size = (rt_hash_mask + 1) * 16; |
| 3414 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3415 | devinet_init(); |
| 3416 | ip_fib_init(); |
| 3417 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3418 | /* All the timers, started at system startup tend |
| 3419 | to synchronize. Perturb it a bit. |
| 3420 | */ |
Eric Dumazet | 39c90ec | 2007-09-15 10:55:54 -0700 | [diff] [blame] | 3421 | schedule_delayed_work(&expires_work, |
| 3422 | net_random() % ip_rt_gc_interval + ip_rt_gc_interval); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3423 | |
Denis V. Lunev | 9f5e97e | 2008-07-05 19:02:59 -0700 | [diff] [blame] | 3424 | if (register_pernet_subsys(&rt_secret_timer_ops)) |
| 3425 | printk(KERN_ERR "Unable to setup rt_secret_timer\n"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3426 | |
Denis V. Lunev | 73b3871 | 2008-02-28 20:51:18 -0800 | [diff] [blame] | 3427 | if (ip_rt_proc_init()) |
Pavel Emelyanov | 107f163 | 2007-12-05 21:14:28 -0800 | [diff] [blame] | 3428 | printk(KERN_ERR "Unable to create route proc files\n"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3429 | #ifdef CONFIG_XFRM |
| 3430 | xfrm_init(); |
| 3431 | xfrm4_init(); |
| 3432 | #endif |
Thomas Graf | 63f3444 | 2007-03-22 11:55:17 -0700 | [diff] [blame] | 3433 | rtnl_register(PF_INET, RTM_GETROUTE, inet_rtm_getroute, NULL); |
| 3434 | |
Denis V. Lunev | 39a23e7 | 2008-07-05 19:02:33 -0700 | [diff] [blame] | 3435 | #ifdef CONFIG_SYSCTL |
| 3436 | register_pernet_subsys(&sysctl_route_ops); |
| 3437 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3438 | return rc; |
| 3439 | } |
| 3440 | |
Al Viro | a1bc6eb | 2008-07-30 06:32:52 -0400 | [diff] [blame] | 3441 | #ifdef CONFIG_SYSCTL |
Al Viro | eeb61f7 | 2008-07-27 08:59:33 +0100 | [diff] [blame] | 3442 | /* |
| 3443 | * We really need to sanitize the damn ipv4 init order, then all |
| 3444 | * this nonsense will go away. |
| 3445 | */ |
| 3446 | void __init ip_static_sysctl_init(void) |
| 3447 | { |
Al Viro | 2f4520d | 2008-08-25 15:17:44 -0700 | [diff] [blame] | 3448 | register_sysctl_paths(ipv4_path, ipv4_skeleton); |
Al Viro | eeb61f7 | 2008-07-27 08:59:33 +0100 | [diff] [blame] | 3449 | } |
Al Viro | a1bc6eb | 2008-07-30 06:32:52 -0400 | [diff] [blame] | 3450 | #endif |
Al Viro | eeb61f7 | 2008-07-27 08:59:33 +0100 | [diff] [blame] | 3451 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3452 | EXPORT_SYMBOL(__ip_select_ident); |
| 3453 | EXPORT_SYMBOL(ip_route_input); |
| 3454 | EXPORT_SYMBOL(ip_route_output_key); |