]>
Commit | Line | Data |
---|---|---|
e905a9ed | 1 | /* Cluster IP hashmark target |
1da177e4 LT |
2 | * (C) 2003-2004 by Harald Welte <laforge@netfilter.org> |
3 | * based on ideas of Fabio Olive Leite <olive@unixforge.org> | |
4 | * | |
5 | * Development of this code funded by SuSE Linux AG, http://www.suse.com/ | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License version 2 as | |
9 | * published by the Free Software Foundation. | |
10 | * | |
11 | */ | |
12 | #include <linux/module.h> | |
1da177e4 LT |
13 | #include <linux/proc_fs.h> |
14 | #include <linux/jhash.h> | |
136e92bb | 15 | #include <linux/bitops.h> |
1da177e4 | 16 | #include <linux/skbuff.h> |
5a0e3ad6 | 17 | #include <linux/slab.h> |
1da177e4 LT |
18 | #include <linux/ip.h> |
19 | #include <linux/tcp.h> | |
20 | #include <linux/udp.h> | |
21 | #include <linux/icmp.h> | |
22 | #include <linux/if_arp.h> | |
1da177e4 | 23 | #include <linux/seq_file.h> |
1da177e4 | 24 | #include <linux/netfilter_arp.h> |
6709dbbb | 25 | #include <linux/netfilter/x_tables.h> |
1da177e4 LT |
26 | #include <linux/netfilter_ipv4/ip_tables.h> |
27 | #include <linux/netfilter_ipv4/ipt_CLUSTERIP.h> | |
587aa641 | 28 | #include <net/netfilter/nf_conntrack.h> |
457c4cbc | 29 | #include <net/net_namespace.h> |
587aa641 | 30 | #include <net/checksum.h> |
1da177e4 | 31 | |
136e92bb | 32 | #define CLUSTERIP_VERSION "0.8" |
1da177e4 | 33 | |
1da177e4 LT |
34 | MODULE_LICENSE("GPL"); |
35 | MODULE_AUTHOR("Harald Welte <laforge@netfilter.org>"); | |
2ae15b64 | 36 | MODULE_DESCRIPTION("Xtables: CLUSTERIP target"); |
1da177e4 LT |
37 | |
38 | struct clusterip_config { | |
39 | struct list_head list; /* list of all configs */ | |
40 | atomic_t refcount; /* reference count */ | |
44513624 KK |
41 | atomic_t entries; /* number of entries/rules |
42 | * referencing us */ | |
1da177e4 | 43 | |
6a19d614 | 44 | __be32 clusterip; /* the IP address */ |
1da177e4 LT |
45 | u_int8_t clustermac[ETH_ALEN]; /* the MAC address */ |
46 | struct net_device *dev; /* device */ | |
47 | u_int16_t num_total_nodes; /* total number of nodes */ | |
136e92bb | 48 | unsigned long local_nodes; /* node number array */ |
1da177e4 LT |
49 | |
50 | #ifdef CONFIG_PROC_FS | |
51 | struct proc_dir_entry *pde; /* proc dir entry */ | |
52 | #endif | |
53 | enum clusterip_hashmode hash_mode; /* which hashing mode */ | |
54 | u_int32_t hash_initval; /* hash initialization */ | |
55 | }; | |
56 | ||
57 | static LIST_HEAD(clusterip_configs); | |
58 | ||
136e92bb | 59 | /* clusterip_lock protects the clusterip_configs list */ |
e45b1be8 | 60 | static DEFINE_RWLOCK(clusterip_lock); |
1da177e4 LT |
61 | |
62 | #ifdef CONFIG_PROC_FS | |
9a32144e | 63 | static const struct file_operations clusterip_proc_fops; |
1da177e4 LT |
64 | static struct proc_dir_entry *clusterip_procdir; |
65 | #endif | |
66 | ||
67 | static inline void | |
44513624 KK |
68 | clusterip_config_get(struct clusterip_config *c) |
69 | { | |
1da177e4 LT |
70 | atomic_inc(&c->refcount); |
71 | } | |
72 | ||
73 | static inline void | |
44513624 KK |
74 | clusterip_config_put(struct clusterip_config *c) |
75 | { | |
76 | if (atomic_dec_and_test(&c->refcount)) | |
77 | kfree(c); | |
78 | } | |
79 | ||
44513624 KK |
80 | /* decrease the count of entries using/referencing this config. If last |
81 | * entry(rule) is removed, remove the config from lists, but don't free it | |
82 | * yet, since proc-files could still be holding references */ | |
83 | static inline void | |
84 | clusterip_config_entry_put(struct clusterip_config *c) | |
85 | { | |
4dee9597 | 86 | write_lock_bh(&clusterip_lock); |
44513624 | 87 | if (atomic_dec_and_test(&c->entries)) { |
1da177e4 | 88 | list_del(&c->list); |
e45b1be8 | 89 | write_unlock_bh(&clusterip_lock); |
44513624 | 90 | |
1da177e4 LT |
91 | dev_mc_delete(c->dev, c->clustermac, ETH_ALEN, 0); |
92 | dev_put(c->dev); | |
44513624 KK |
93 | |
94 | /* In case anyone still accesses the file, the open/close | |
95 | * functions are also incrementing the refcount on their own, | |
96 | * so it's safe to remove the entry even if it's in use. */ | |
97 | #ifdef CONFIG_PROC_FS | |
98 | remove_proc_entry(c->pde->name, c->pde->parent); | |
99 | #endif | |
4dee9597 | 100 | return; |
1da177e4 | 101 | } |
4dee9597 | 102 | write_unlock_bh(&clusterip_lock); |
1da177e4 LT |
103 | } |
104 | ||
1da177e4 | 105 | static struct clusterip_config * |
6a19d614 | 106 | __clusterip_config_find(__be32 clusterip) |
1da177e4 | 107 | { |
4c610979 | 108 | struct clusterip_config *c; |
1da177e4 | 109 | |
4c610979 | 110 | list_for_each_entry(c, &clusterip_configs, list) { |
7c4e36bc | 111 | if (c->clusterip == clusterip) |
1da177e4 | 112 | return c; |
1da177e4 LT |
113 | } |
114 | ||
115 | return NULL; | |
116 | } | |
117 | ||
118 | static inline struct clusterip_config * | |
6a19d614 | 119 | clusterip_config_find_get(__be32 clusterip, int entry) |
1da177e4 LT |
120 | { |
121 | struct clusterip_config *c; | |
122 | ||
e45b1be8 | 123 | read_lock_bh(&clusterip_lock); |
1da177e4 LT |
124 | c = __clusterip_config_find(clusterip); |
125 | if (!c) { | |
e45b1be8 | 126 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
127 | return NULL; |
128 | } | |
129 | atomic_inc(&c->refcount); | |
44513624 KK |
130 | if (entry) |
131 | atomic_inc(&c->entries); | |
e45b1be8 | 132 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
133 | |
134 | return c; | |
135 | } | |
136 | ||
136e92bb KK |
137 | static void |
138 | clusterip_config_init_nodelist(struct clusterip_config *c, | |
139 | const struct ipt_clusterip_tgt_info *i) | |
140 | { | |
141 | int n; | |
142 | ||
7c4e36bc | 143 | for (n = 0; n < i->num_local_nodes; n++) |
136e92bb | 144 | set_bit(i->local_nodes[n] - 1, &c->local_nodes); |
136e92bb KK |
145 | } |
146 | ||
1da177e4 | 147 | static struct clusterip_config * |
3cf93c96 | 148 | clusterip_config_init(const struct ipt_clusterip_tgt_info *i, __be32 ip, |
1da177e4 LT |
149 | struct net_device *dev) |
150 | { | |
151 | struct clusterip_config *c; | |
1da177e4 | 152 | |
0da974f4 | 153 | c = kzalloc(sizeof(*c), GFP_ATOMIC); |
1da177e4 LT |
154 | if (!c) |
155 | return NULL; | |
156 | ||
1da177e4 LT |
157 | c->dev = dev; |
158 | c->clusterip = ip; | |
159 | memcpy(&c->clustermac, &i->clustermac, ETH_ALEN); | |
160 | c->num_total_nodes = i->num_total_nodes; | |
136e92bb | 161 | clusterip_config_init_nodelist(c, i); |
1da177e4 LT |
162 | c->hash_mode = i->hash_mode; |
163 | c->hash_initval = i->hash_initval; | |
164 | atomic_set(&c->refcount, 1); | |
44513624 | 165 | atomic_set(&c->entries, 1); |
1da177e4 LT |
166 | |
167 | #ifdef CONFIG_PROC_FS | |
76592584 PM |
168 | { |
169 | char buffer[16]; | |
170 | ||
171 | /* create proc dir entry */ | |
cffee385 | 172 | sprintf(buffer, "%pI4", &ip); |
6e79d85d DL |
173 | c->pde = proc_create_data(buffer, S_IWUSR|S_IRUSR, |
174 | clusterip_procdir, | |
175 | &clusterip_proc_fops, c); | |
76592584 PM |
176 | if (!c->pde) { |
177 | kfree(c); | |
178 | return NULL; | |
179 | } | |
1da177e4 | 180 | } |
1da177e4 LT |
181 | #endif |
182 | ||
e45b1be8 | 183 | write_lock_bh(&clusterip_lock); |
1da177e4 | 184 | list_add(&c->list, &clusterip_configs); |
e45b1be8 | 185 | write_unlock_bh(&clusterip_lock); |
1da177e4 LT |
186 | |
187 | return c; | |
188 | } | |
189 | ||
76592584 | 190 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
191 | static int |
192 | clusterip_add_node(struct clusterip_config *c, u_int16_t nodenum) | |
193 | { | |
1da177e4 | 194 | |
136e92bb KK |
195 | if (nodenum == 0 || |
196 | nodenum > c->num_total_nodes) | |
1da177e4 | 197 | return 1; |
1da177e4 | 198 | |
136e92bb KK |
199 | /* check if we already have this number in our bitfield */ |
200 | if (test_and_set_bit(nodenum - 1, &c->local_nodes)) | |
201 | return 1; | |
1da177e4 | 202 | |
1da177e4 LT |
203 | return 0; |
204 | } | |
205 | ||
e1931b78 | 206 | static bool |
1da177e4 LT |
207 | clusterip_del_node(struct clusterip_config *c, u_int16_t nodenum) |
208 | { | |
136e92bb KK |
209 | if (nodenum == 0 || |
210 | nodenum > c->num_total_nodes) | |
e1931b78 | 211 | return true; |
e905a9ed | 212 | |
136e92bb | 213 | if (test_and_clear_bit(nodenum - 1, &c->local_nodes)) |
e1931b78 | 214 | return false; |
1da177e4 | 215 | |
e1931b78 | 216 | return true; |
1da177e4 | 217 | } |
76592584 | 218 | #endif |
1da177e4 LT |
219 | |
220 | static inline u_int32_t | |
a47362a2 JE |
221 | clusterip_hashfn(const struct sk_buff *skb, |
222 | const struct clusterip_config *config) | |
1da177e4 | 223 | { |
a47362a2 | 224 | const struct iphdr *iph = ip_hdr(skb); |
1da177e4 LT |
225 | unsigned long hashval; |
226 | u_int16_t sport, dport; | |
a47362a2 | 227 | const u_int16_t *ports; |
1da177e4 LT |
228 | |
229 | switch (iph->protocol) { | |
230 | case IPPROTO_TCP: | |
1da177e4 | 231 | case IPPROTO_UDP: |
a8d0f952 | 232 | case IPPROTO_UDPLITE: |
957dc80a PM |
233 | case IPPROTO_SCTP: |
234 | case IPPROTO_DCCP: | |
1da177e4 | 235 | case IPPROTO_ICMP: |
a47362a2 | 236 | ports = (const void *)iph+iph->ihl*4; |
957dc80a PM |
237 | sport = ports[0]; |
238 | dport = ports[1]; | |
1da177e4 LT |
239 | break; |
240 | default: | |
7c4e36bc | 241 | if (net_ratelimit()) |
1da177e4 LT |
242 | printk(KERN_NOTICE "CLUSTERIP: unknown protocol `%u'\n", |
243 | iph->protocol); | |
1da177e4 LT |
244 | sport = dport = 0; |
245 | } | |
246 | ||
247 | switch (config->hash_mode) { | |
248 | case CLUSTERIP_HASHMODE_SIP: | |
249 | hashval = jhash_1word(ntohl(iph->saddr), | |
250 | config->hash_initval); | |
251 | break; | |
252 | case CLUSTERIP_HASHMODE_SIP_SPT: | |
e905a9ed | 253 | hashval = jhash_2words(ntohl(iph->saddr), sport, |
1da177e4 LT |
254 | config->hash_initval); |
255 | break; | |
256 | case CLUSTERIP_HASHMODE_SIP_SPT_DPT: | |
257 | hashval = jhash_3words(ntohl(iph->saddr), sport, dport, | |
258 | config->hash_initval); | |
259 | break; | |
260 | default: | |
261 | /* to make gcc happy */ | |
262 | hashval = 0; | |
263 | /* This cannot happen, unless the check function wasn't called | |
264 | * at rule load time */ | |
265 | printk("CLUSTERIP: unknown mode `%u'\n", config->hash_mode); | |
266 | BUG(); | |
267 | break; | |
268 | } | |
269 | ||
270 | /* node numbers are 1..n, not 0..n */ | |
34498825 | 271 | return (((u64)hashval * config->num_total_nodes) >> 32) + 1; |
1da177e4 LT |
272 | } |
273 | ||
274 | static inline int | |
a47362a2 | 275 | clusterip_responsible(const struct clusterip_config *config, u_int32_t hash) |
1da177e4 | 276 | { |
136e92bb | 277 | return test_bit(hash - 1, &config->local_nodes); |
1da177e4 LT |
278 | } |
279 | ||
e905a9ed YH |
280 | /*********************************************************************** |
281 | * IPTABLES TARGET | |
1da177e4 LT |
282 | ***********************************************************************/ |
283 | ||
284 | static unsigned int | |
7eb35586 | 285 | clusterip_tg(struct sk_buff *skb, const struct xt_target_param *par) |
1da177e4 | 286 | { |
7eb35586 | 287 | const struct ipt_clusterip_tgt_info *cipinfo = par->targinfo; |
587aa641 | 288 | struct nf_conn *ct; |
1da177e4 | 289 | enum ip_conntrack_info ctinfo; |
587aa641 | 290 | u_int32_t hash; |
1da177e4 LT |
291 | |
292 | /* don't need to clusterip_config_get() here, since refcount | |
293 | * is only decremented by destroy() - and ip_tables guarantees | |
294 | * that the ->target() function isn't called after ->destroy() */ | |
295 | ||
3db05fea | 296 | ct = nf_ct_get(skb, &ctinfo); |
587aa641 | 297 | if (ct == NULL) { |
1da177e4 LT |
298 | printk(KERN_ERR "CLUSTERIP: no conntrack!\n"); |
299 | /* FIXME: need to drop invalid ones, since replies | |
e905a9ed | 300 | * to outgoing connections of other nodes will be |
1da177e4 LT |
301 | * marked as INVALID */ |
302 | return NF_DROP; | |
303 | } | |
304 | ||
305 | /* special case: ICMP error handling. conntrack distinguishes between | |
306 | * error messages (RELATED) and information requests (see below) */ | |
3666ed1c JP |
307 | if (ip_hdr(skb)->protocol == IPPROTO_ICMP && |
308 | (ctinfo == IP_CT_RELATED || | |
309 | ctinfo == IP_CT_RELATED + IP_CT_IS_REPLY)) | |
6709dbbb | 310 | return XT_CONTINUE; |
1da177e4 | 311 | |
e905a9ed | 312 | /* ip_conntrack_icmp guarantees us that we only have ICMP_ECHO, |
1da177e4 LT |
313 | * TIMESTAMP, INFO_REQUEST or ADDRESS type icmp packets from here |
314 | * on, which all have an ID field [relevant for hashing]. */ | |
315 | ||
3db05fea | 316 | hash = clusterip_hashfn(skb, cipinfo->config); |
1da177e4 LT |
317 | |
318 | switch (ctinfo) { | |
319 | case IP_CT_NEW: | |
587aa641 | 320 | ct->mark = hash; |
1da177e4 LT |
321 | break; |
322 | case IP_CT_RELATED: | |
323 | case IP_CT_RELATED+IP_CT_IS_REPLY: | |
324 | /* FIXME: we don't handle expectations at the | |
325 | * moment. they can arrive on a different node than | |
326 | * the master connection (e.g. FTP passive mode) */ | |
327 | case IP_CT_ESTABLISHED: | |
328 | case IP_CT_ESTABLISHED+IP_CT_IS_REPLY: | |
329 | break; | |
330 | default: | |
331 | break; | |
332 | } | |
333 | ||
0d53778e | 334 | #ifdef DEBUG |
3c9fba65 | 335 | nf_ct_dump_tuple_ip(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple); |
1da177e4 | 336 | #endif |
0d53778e | 337 | pr_debug("hash=%u ct_hash=%u ", hash, ct->mark); |
1da177e4 | 338 | if (!clusterip_responsible(cipinfo->config, hash)) { |
0d53778e | 339 | pr_debug("not responsible\n"); |
1da177e4 LT |
340 | return NF_DROP; |
341 | } | |
0d53778e | 342 | pr_debug("responsible\n"); |
1da177e4 LT |
343 | |
344 | /* despite being received via linklayer multicast, this is | |
345 | * actually a unicast IP packet. TCP doesn't like PACKET_MULTICAST */ | |
3db05fea | 346 | skb->pkt_type = PACKET_HOST; |
1da177e4 | 347 | |
6709dbbb | 348 | return XT_CONTINUE; |
1da177e4 LT |
349 | } |
350 | ||
af5d6dc2 | 351 | static bool clusterip_tg_check(const struct xt_tgchk_param *par) |
1da177e4 | 352 | { |
af5d6dc2 JE |
353 | struct ipt_clusterip_tgt_info *cipinfo = par->targinfo; |
354 | const struct ipt_entry *e = par->entryinfo; | |
1da177e4 LT |
355 | |
356 | struct clusterip_config *config; | |
357 | ||
1da177e4 LT |
358 | if (cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP && |
359 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT && | |
360 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT_DPT) { | |
361 | printk(KERN_WARNING "CLUSTERIP: unknown mode `%u'\n", | |
362 | cipinfo->hash_mode); | |
e1931b78 | 363 | return false; |
1da177e4 LT |
364 | |
365 | } | |
3666ed1c JP |
366 | if (e->ip.dmsk.s_addr != htonl(0xffffffff) || |
367 | e->ip.dst.s_addr == 0) { | |
1da177e4 | 368 | printk(KERN_ERR "CLUSTERIP: Please specify destination IP\n"); |
e1931b78 | 369 | return false; |
1da177e4 LT |
370 | } |
371 | ||
372 | /* FIXME: further sanity checks */ | |
373 | ||
44513624 | 374 | config = clusterip_config_find_get(e->ip.dst.s_addr, 1); |
d3c3f424 | 375 | if (!config) { |
1da177e4 | 376 | if (!(cipinfo->flags & CLUSTERIP_FLAG_NEW)) { |
cffee385 | 377 | printk(KERN_WARNING "CLUSTERIP: no config found for %pI4, need 'new'\n", &e->ip.dst.s_addr); |
e1931b78 | 378 | return false; |
1da177e4 LT |
379 | } else { |
380 | struct net_device *dev; | |
381 | ||
382 | if (e->ip.iniface[0] == '\0') { | |
383 | printk(KERN_WARNING "CLUSTERIP: Please specify an interface name\n"); | |
e1931b78 | 384 | return false; |
1da177e4 LT |
385 | } |
386 | ||
881d966b | 387 | dev = dev_get_by_name(&init_net, e->ip.iniface); |
1da177e4 LT |
388 | if (!dev) { |
389 | printk(KERN_WARNING "CLUSTERIP: no such interface %s\n", e->ip.iniface); | |
e1931b78 | 390 | return false; |
1da177e4 LT |
391 | } |
392 | ||
e905a9ed | 393 | config = clusterip_config_init(cipinfo, |
1da177e4 LT |
394 | e->ip.dst.s_addr, dev); |
395 | if (!config) { | |
396 | printk(KERN_WARNING "CLUSTERIP: cannot allocate config\n"); | |
397 | dev_put(dev); | |
e1931b78 | 398 | return false; |
1da177e4 LT |
399 | } |
400 | dev_mc_add(config->dev,config->clustermac, ETH_ALEN, 0); | |
401 | } | |
402 | } | |
d3c3f424 | 403 | cipinfo->config = config; |
1da177e4 | 404 | |
af5d6dc2 | 405 | if (nf_ct_l3proto_try_module_get(par->target->family) < 0) { |
11078c37 | 406 | printk(KERN_WARNING "can't load conntrack support for " |
af5d6dc2 | 407 | "proto=%u\n", par->target->family); |
e1931b78 | 408 | return false; |
11078c37 YK |
409 | } |
410 | ||
e1931b78 | 411 | return true; |
1da177e4 LT |
412 | } |
413 | ||
414 | /* drop reference count of cluster config when rule is deleted */ | |
a2df1648 | 415 | static void clusterip_tg_destroy(const struct xt_tgdtor_param *par) |
1da177e4 | 416 | { |
a2df1648 | 417 | const struct ipt_clusterip_tgt_info *cipinfo = par->targinfo; |
1da177e4 | 418 | |
44513624 KK |
419 | /* if no more entries are referencing the config, remove it |
420 | * from the list and destroy the proc entry */ | |
421 | clusterip_config_entry_put(cipinfo->config); | |
422 | ||
1da177e4 | 423 | clusterip_config_put(cipinfo->config); |
11078c37 | 424 | |
a2df1648 | 425 | nf_ct_l3proto_module_put(par->target->family); |
1da177e4 LT |
426 | } |
427 | ||
d3c3f424 PM |
428 | #ifdef CONFIG_COMPAT |
429 | struct compat_ipt_clusterip_tgt_info | |
430 | { | |
431 | u_int32_t flags; | |
432 | u_int8_t clustermac[6]; | |
433 | u_int16_t num_total_nodes; | |
434 | u_int16_t num_local_nodes; | |
435 | u_int16_t local_nodes[CLUSTERIP_MAX_NODES]; | |
436 | u_int32_t hash_mode; | |
437 | u_int32_t hash_initval; | |
438 | compat_uptr_t config; | |
439 | }; | |
440 | #endif /* CONFIG_COMPAT */ | |
441 | ||
d3c5ee6d | 442 | static struct xt_target clusterip_tg_reg __read_mostly = { |
1d5cd909 | 443 | .name = "CLUSTERIP", |
ee999d8b | 444 | .family = NFPROTO_IPV4, |
d3c5ee6d JE |
445 | .target = clusterip_tg, |
446 | .checkentry = clusterip_tg_check, | |
447 | .destroy = clusterip_tg_destroy, | |
d3c3f424 PM |
448 | .targetsize = sizeof(struct ipt_clusterip_tgt_info), |
449 | #ifdef CONFIG_COMPAT | |
450 | .compatsize = sizeof(struct compat_ipt_clusterip_tgt_info), | |
451 | #endif /* CONFIG_COMPAT */ | |
1d5cd909 | 452 | .me = THIS_MODULE |
1da177e4 LT |
453 | }; |
454 | ||
455 | ||
e905a9ed YH |
456 | /*********************************************************************** |
457 | * ARP MANGLING CODE | |
1da177e4 LT |
458 | ***********************************************************************/ |
459 | ||
460 | /* hardcoded for 48bit ethernet and 32bit ipv4 addresses */ | |
461 | struct arp_payload { | |
462 | u_int8_t src_hw[ETH_ALEN]; | |
6a19d614 | 463 | __be32 src_ip; |
1da177e4 | 464 | u_int8_t dst_hw[ETH_ALEN]; |
6a19d614 | 465 | __be32 dst_ip; |
1da177e4 LT |
466 | } __attribute__ ((packed)); |
467 | ||
0d53778e | 468 | #ifdef DEBUG |
e905a9ed | 469 | static void arp_print(struct arp_payload *payload) |
1da177e4 LT |
470 | { |
471 | #define HBUFFERLEN 30 | |
472 | char hbuffer[HBUFFERLEN]; | |
473 | int j,k; | |
1da177e4 LT |
474 | |
475 | for (k=0, j=0; k < HBUFFERLEN-3 && j < ETH_ALEN; j++) { | |
6a8341b6 HH |
476 | hbuffer[k++] = hex_asc_hi(payload->src_hw[j]); |
477 | hbuffer[k++] = hex_asc_lo(payload->src_hw[j]); | |
1da177e4 LT |
478 | hbuffer[k++]=':'; |
479 | } | |
480 | hbuffer[--k]='\0'; | |
481 | ||
cffee385 HH |
482 | printk("src %pI4@%s, dst %pI4\n", |
483 | &payload->src_ip, hbuffer, &payload->dst_ip); | |
1da177e4 LT |
484 | } |
485 | #endif | |
486 | ||
487 | static unsigned int | |
488 | arp_mangle(unsigned int hook, | |
3db05fea | 489 | struct sk_buff *skb, |
1da177e4 LT |
490 | const struct net_device *in, |
491 | const struct net_device *out, | |
492 | int (*okfn)(struct sk_buff *)) | |
493 | { | |
3db05fea | 494 | struct arphdr *arp = arp_hdr(skb); |
1da177e4 LT |
495 | struct arp_payload *payload; |
496 | struct clusterip_config *c; | |
497 | ||
498 | /* we don't care about non-ethernet and non-ipv4 ARP */ | |
3666ed1c JP |
499 | if (arp->ar_hrd != htons(ARPHRD_ETHER) || |
500 | arp->ar_pro != htons(ETH_P_IP) || | |
501 | arp->ar_pln != 4 || arp->ar_hln != ETH_ALEN) | |
1da177e4 LT |
502 | return NF_ACCEPT; |
503 | ||
4095ebf1 | 504 | /* we only want to mangle arp requests and replies */ |
3666ed1c JP |
505 | if (arp->ar_op != htons(ARPOP_REPLY) && |
506 | arp->ar_op != htons(ARPOP_REQUEST)) | |
1da177e4 LT |
507 | return NF_ACCEPT; |
508 | ||
509 | payload = (void *)(arp+1); | |
510 | ||
e905a9ed | 511 | /* if there is no clusterip configuration for the arp reply's |
1da177e4 | 512 | * source ip, we don't want to mangle it */ |
44513624 | 513 | c = clusterip_config_find_get(payload->src_ip, 0); |
1da177e4 LT |
514 | if (!c) |
515 | return NF_ACCEPT; | |
516 | ||
e905a9ed | 517 | /* normally the linux kernel always replies to arp queries of |
1da177e4 LT |
518 | * addresses on different interfacs. However, in the CLUSTERIP case |
519 | * this wouldn't work, since we didn't subscribe the mcast group on | |
520 | * other interfaces */ | |
521 | if (c->dev != out) { | |
0d53778e PM |
522 | pr_debug("CLUSTERIP: not mangling arp reply on different " |
523 | "interface: cip'%s'-skb'%s'\n", | |
524 | c->dev->name, out->name); | |
1da177e4 LT |
525 | clusterip_config_put(c); |
526 | return NF_ACCEPT; | |
527 | } | |
528 | ||
529 | /* mangle reply hardware address */ | |
530 | memcpy(payload->src_hw, c->clustermac, arp->ar_hln); | |
531 | ||
0d53778e PM |
532 | #ifdef DEBUG |
533 | pr_debug(KERN_DEBUG "CLUSTERIP mangled arp reply: "); | |
1da177e4 LT |
534 | arp_print(payload); |
535 | #endif | |
536 | ||
537 | clusterip_config_put(c); | |
538 | ||
539 | return NF_ACCEPT; | |
540 | } | |
541 | ||
1999414a | 542 | static struct nf_hook_ops cip_arp_ops __read_mostly = { |
1da177e4 | 543 | .hook = arp_mangle, |
ee999d8b | 544 | .pf = NFPROTO_ARP, |
1da177e4 LT |
545 | .hooknum = NF_ARP_OUT, |
546 | .priority = -1 | |
547 | }; | |
548 | ||
e905a9ed YH |
549 | /*********************************************************************** |
550 | * PROC DIR HANDLING | |
1da177e4 LT |
551 | ***********************************************************************/ |
552 | ||
553 | #ifdef CONFIG_PROC_FS | |
554 | ||
136e92bb KK |
555 | struct clusterip_seq_position { |
556 | unsigned int pos; /* position */ | |
557 | unsigned int weight; /* number of bits set == size */ | |
558 | unsigned int bit; /* current bit */ | |
559 | unsigned long val; /* current value */ | |
560 | }; | |
561 | ||
1da177e4 LT |
562 | static void *clusterip_seq_start(struct seq_file *s, loff_t *pos) |
563 | { | |
47778147 | 564 | struct clusterip_config *c = s->private; |
136e92bb KK |
565 | unsigned int weight; |
566 | u_int32_t local_nodes; | |
567 | struct clusterip_seq_position *idx; | |
568 | ||
569 | /* FIXME: possible race */ | |
570 | local_nodes = c->local_nodes; | |
571 | weight = hweight32(local_nodes); | |
572 | if (*pos >= weight) | |
1da177e4 LT |
573 | return NULL; |
574 | ||
136e92bb KK |
575 | idx = kmalloc(sizeof(struct clusterip_seq_position), GFP_KERNEL); |
576 | if (!idx) | |
1da177e4 LT |
577 | return ERR_PTR(-ENOMEM); |
578 | ||
136e92bb KK |
579 | idx->pos = *pos; |
580 | idx->weight = weight; | |
581 | idx->bit = ffs(local_nodes); | |
582 | idx->val = local_nodes; | |
583 | clear_bit(idx->bit - 1, &idx->val); | |
584 | ||
585 | return idx; | |
1da177e4 LT |
586 | } |
587 | ||
588 | static void *clusterip_seq_next(struct seq_file *s, void *v, loff_t *pos) | |
589 | { | |
3cf93c96 | 590 | struct clusterip_seq_position *idx = v; |
1da177e4 | 591 | |
136e92bb KK |
592 | *pos = ++idx->pos; |
593 | if (*pos >= idx->weight) { | |
1da177e4 LT |
594 | kfree(v); |
595 | return NULL; | |
596 | } | |
136e92bb KK |
597 | idx->bit = ffs(idx->val); |
598 | clear_bit(idx->bit - 1, &idx->val); | |
599 | return idx; | |
1da177e4 LT |
600 | } |
601 | ||
602 | static void clusterip_seq_stop(struct seq_file *s, void *v) | |
603 | { | |
604 | kfree(v); | |
1da177e4 LT |
605 | } |
606 | ||
607 | static int clusterip_seq_show(struct seq_file *s, void *v) | |
608 | { | |
3cf93c96 | 609 | struct clusterip_seq_position *idx = v; |
1da177e4 | 610 | |
e905a9ed | 611 | if (idx->pos != 0) |
1da177e4 | 612 | seq_putc(s, ','); |
1da177e4 | 613 | |
136e92bb KK |
614 | seq_printf(s, "%u", idx->bit); |
615 | ||
616 | if (idx->pos == idx->weight - 1) | |
1da177e4 LT |
617 | seq_putc(s, '\n'); |
618 | ||
619 | return 0; | |
620 | } | |
621 | ||
56b3d975 | 622 | static const struct seq_operations clusterip_seq_ops = { |
1da177e4 LT |
623 | .start = clusterip_seq_start, |
624 | .next = clusterip_seq_next, | |
625 | .stop = clusterip_seq_stop, | |
626 | .show = clusterip_seq_show, | |
627 | }; | |
628 | ||
629 | static int clusterip_proc_open(struct inode *inode, struct file *file) | |
630 | { | |
631 | int ret = seq_open(file, &clusterip_seq_ops); | |
632 | ||
633 | if (!ret) { | |
634 | struct seq_file *sf = file->private_data; | |
47778147 | 635 | struct clusterip_config *c = PDE(inode)->data; |
1da177e4 | 636 | |
47778147 | 637 | sf->private = c; |
1da177e4 LT |
638 | |
639 | clusterip_config_get(c); | |
640 | } | |
641 | ||
642 | return ret; | |
643 | } | |
644 | ||
645 | static int clusterip_proc_release(struct inode *inode, struct file *file) | |
646 | { | |
47778147 | 647 | struct clusterip_config *c = PDE(inode)->data; |
1da177e4 LT |
648 | int ret; |
649 | ||
650 | ret = seq_release(inode, file); | |
651 | ||
652 | if (!ret) | |
653 | clusterip_config_put(c); | |
654 | ||
655 | return ret; | |
656 | } | |
657 | ||
658 | static ssize_t clusterip_proc_write(struct file *file, const char __user *input, | |
659 | size_t size, loff_t *ofs) | |
660 | { | |
47778147 | 661 | struct clusterip_config *c = PDE(file->f_path.dentry->d_inode)->data; |
1da177e4 LT |
662 | #define PROC_WRITELEN 10 |
663 | char buffer[PROC_WRITELEN+1]; | |
1da177e4 LT |
664 | unsigned long nodenum; |
665 | ||
666 | if (copy_from_user(buffer, input, PROC_WRITELEN)) | |
667 | return -EFAULT; | |
668 | ||
669 | if (*buffer == '+') { | |
670 | nodenum = simple_strtoul(buffer+1, NULL, 10); | |
671 | if (clusterip_add_node(c, nodenum)) | |
672 | return -ENOMEM; | |
673 | } else if (*buffer == '-') { | |
674 | nodenum = simple_strtoul(buffer+1, NULL,10); | |
675 | if (clusterip_del_node(c, nodenum)) | |
676 | return -ENOENT; | |
677 | } else | |
678 | return -EIO; | |
679 | ||
680 | return size; | |
681 | } | |
682 | ||
9a32144e | 683 | static const struct file_operations clusterip_proc_fops = { |
1da177e4 LT |
684 | .owner = THIS_MODULE, |
685 | .open = clusterip_proc_open, | |
686 | .read = seq_read, | |
687 | .write = clusterip_proc_write, | |
688 | .llseek = seq_lseek, | |
689 | .release = clusterip_proc_release, | |
690 | }; | |
691 | ||
692 | #endif /* CONFIG_PROC_FS */ | |
693 | ||
d3c5ee6d | 694 | static int __init clusterip_tg_init(void) |
1da177e4 LT |
695 | { |
696 | int ret; | |
697 | ||
d3c5ee6d | 698 | ret = xt_register_target(&clusterip_tg_reg); |
32292a7f PM |
699 | if (ret < 0) |
700 | return ret; | |
1da177e4 | 701 | |
32292a7f PM |
702 | ret = nf_register_hook(&cip_arp_ops); |
703 | if (ret < 0) | |
1da177e4 | 704 | goto cleanup_target; |
1da177e4 LT |
705 | |
706 | #ifdef CONFIG_PROC_FS | |
457c4cbc | 707 | clusterip_procdir = proc_mkdir("ipt_CLUSTERIP", init_net.proc_net); |
1da177e4 LT |
708 | if (!clusterip_procdir) { |
709 | printk(KERN_ERR "CLUSTERIP: Unable to proc dir entry\n"); | |
710 | ret = -ENOMEM; | |
711 | goto cleanup_hook; | |
712 | } | |
713 | #endif /* CONFIG_PROC_FS */ | |
714 | ||
715 | printk(KERN_NOTICE "ClusterIP Version %s loaded successfully\n", | |
716 | CLUSTERIP_VERSION); | |
1da177e4 LT |
717 | return 0; |
718 | ||
76592584 | 719 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
720 | cleanup_hook: |
721 | nf_unregister_hook(&cip_arp_ops); | |
76592584 | 722 | #endif /* CONFIG_PROC_FS */ |
1da177e4 | 723 | cleanup_target: |
d3c5ee6d | 724 | xt_unregister_target(&clusterip_tg_reg); |
32292a7f | 725 | return ret; |
1da177e4 LT |
726 | } |
727 | ||
d3c5ee6d | 728 | static void __exit clusterip_tg_exit(void) |
1da177e4 | 729 | { |
32292a7f PM |
730 | printk(KERN_NOTICE "ClusterIP Version %s unloading\n", |
731 | CLUSTERIP_VERSION); | |
732 | #ifdef CONFIG_PROC_FS | |
733 | remove_proc_entry(clusterip_procdir->name, clusterip_procdir->parent); | |
734 | #endif | |
735 | nf_unregister_hook(&cip_arp_ops); | |
d3c5ee6d | 736 | xt_unregister_target(&clusterip_tg_reg); |
1da177e4 LT |
737 | } |
738 | ||
d3c5ee6d JE |
739 | module_init(clusterip_tg_init); |
740 | module_exit(clusterip_tg_exit); |